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 struct page **fuse_pages_alloc(unsigned int npages, gfp_t flags,
  23                                struct fuse_page_desc **desc)
  24 {
  25         struct page **pages;
  26
  27         pages = kzalloc(npages * (sizeof(struct page *) +
  28                                   sizeof(struct fuse_page_desc)), flags);
  29         *desc = (void *) (pages + npages);
  30
  31         return pages;
  32 }
  33
  34 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  35                           int opcode, struct fuse_open_out *outargp)
  36 {
  37         struct fuse_open_in inarg;
  38         FUSE_ARGS(args);
  39
  40         memset(&inarg, 0, sizeof(inarg));
  41         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  42         if (!fc->atomic_o_trunc)
  43                 inarg.flags &= ~O_TRUNC;
  44         args.opcode = opcode;
  45         args.nodeid = nodeid;
  46         args.in_numargs = 1;
  47         args.in_args[0].size = sizeof(inarg);
  48         args.in_args[0].value = &inarg;
  49         args.out_numargs = 1;
  50         args.out_args[0].size = sizeof(*outargp);
  51         args.out_args[0].value = outargp;
  52
  53         return fuse_simple_request(fc, &args);
  54 }
  55
  56 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  57 {
  58         struct fuse_file *ff;
  59
  60         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  61         if (unlikely(!ff))
  62                 return NULL;
  63
  64         ff->fc = fc;
  65         ff->reserved_req = fuse_request_alloc(0);
  66         if (unlikely(!ff->reserved_req)) {
  67                 kfree(ff);
  68                 return NULL;
  69         }
  70
  71         INIT_LIST_HEAD(&ff->write_entry);
  72         mutex_init(&ff->readdir.lock);
  73         refcount_set(&ff->count, 1);
  74         RB_CLEAR_NODE(&ff->polled_node);
  75         init_waitqueue_head(&ff->poll_wait);
  76
  77         ff->kh = atomic64_inc_return(&fc->khctr);
  78
  79         return ff;
  80 }
  81
  82 void fuse_file_free(struct fuse_file *ff)
  83 {
  84         fuse_request_free(ff->reserved_req);
  85         mutex_destroy(&ff->readdir.lock);
  86         kfree(ff);
  87 }
  88
  89 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  90 {
  91         refcount_inc(&ff->count);
  92         return ff;
  93 }
  94
  95 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  96 {
  97         iput(req->misc.release.inode);
  98 }
  99
 100 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
 101 {
 102         if (refcount_dec_and_test(&ff->count)) {
 103                 struct fuse_req *req = ff->reserved_req;
 104
 105                 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
 106                         /*
 107                          * Drop the release request when client does not
 108                          * implement 'open'
 109                          */
 110                         __clear_bit(FR_BACKGROUND, &req->flags);
 111                         iput(req->misc.release.inode);
 112                         fuse_put_request(ff->fc, req);
 113                 } else if (sync) {
 114                         __set_bit(FR_FORCE, &req->flags);
 115                         __clear_bit(FR_BACKGROUND, &req->flags);
 116                         fuse_request_send(ff->fc, req);
 117                         iput(req->misc.release.inode);
 118                         fuse_put_request(ff->fc, req);
 119                 } else {
 120                         req->end = fuse_release_end;
 121                         __set_bit(FR_BACKGROUND, &req->flags);
 122                         fuse_request_send_background(ff->fc, req);
 123                 }
 124                 kfree(ff);
 125         }
 126 }
 127
 128 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 129                  bool isdir)
 130 {
 131         struct fuse_file *ff;
 132         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 133
 134         ff = fuse_file_alloc(fc);
 135         if (!ff)
 136                 return -ENOMEM;
 137
 138         ff->fh = 0;
 139         /* Default for no-open */
 140         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 141         if (isdir ? !fc->no_opendir : !fc->no_open) {
 142                 struct fuse_open_out outarg;
 143                 int err;
 144
 145                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 146                 if (!err) {
 147                         ff->fh = outarg.fh;
 148                         ff->open_flags = outarg.open_flags;
 149
 150                 } else if (err != -ENOSYS) {
 151                         fuse_file_free(ff);
 152                         return err;
 153                 } else {
 154                         if (isdir)
 155                                 fc->no_opendir = 1;
 156                         else
 157                                 fc->no_open = 1;
 158                 }
 159         }
 160
 161         if (isdir)
 162                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 163
 164         ff->nodeid = nodeid;
 165         file->private_data = ff;
 166
 167         return 0;
 168 }
 169 EXPORT_SYMBOL_GPL(fuse_do_open);
 170
 171 static void fuse_link_write_file(struct file *file)
 172 {
 173         struct inode *inode = file_inode(file);
 174         struct fuse_inode *fi = get_fuse_inode(inode);
 175         struct fuse_file *ff = file->private_data;
 176         /*
 177          * file may be written through mmap, so chain it onto the
 178          * inodes's write_file list
 179          */
 180         spin_lock(&fi->lock);
 181         if (list_empty(&ff->write_entry))
 182                 list_add(&ff->write_entry, &fi->write_files);
 183         spin_unlock(&fi->lock);
 184 }
 185
 186 void fuse_finish_open(struct inode *inode, struct file *file)
 187 {
 188         struct fuse_file *ff = file->private_data;
 189         struct fuse_conn *fc = get_fuse_conn(inode);
 190
 191         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 192                 invalidate_inode_pages2(inode->i_mapping);
 193         if (ff->open_flags & FOPEN_STREAM)
 194                 stream_open(inode, file);
 195         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 196                 nonseekable_open(inode, file);
 197         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 198                 struct fuse_inode *fi = get_fuse_inode(inode);
 199
 200                 spin_lock(&fi->lock);
 201                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 202                 i_size_write(inode, 0);
 203                 spin_unlock(&fi->lock);
 204                 fuse_invalidate_attr(inode);
 205                 if (fc->writeback_cache)
 206                         file_update_time(file);
 207         }
 208         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 209                 fuse_link_write_file(file);
 210 }
 211
 212 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 213 {
 214         struct fuse_conn *fc = get_fuse_conn(inode);
 215         int err;
 216         bool lock_inode = (file->f_flags & O_TRUNC) &&
 217                           fc->atomic_o_trunc &&
 218                           fc->writeback_cache;
 219
 220         err = generic_file_open(inode, file);
 221         if (err)
 222                 return err;
 223
 224         if (lock_inode)
 225                 inode_lock(inode);
 226
 227         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 228
 229         if (!err)
 230                 fuse_finish_open(inode, file);
 231
 232         if (lock_inode)
 233                 inode_unlock(inode);
 234
 235         return err;
 236 }
 237
 238 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 239                                  int flags, int opcode)
 240 {
 241         struct fuse_conn *fc = ff->fc;
 242         struct fuse_req *req = ff->reserved_req;
 243         struct fuse_release_in *inarg = &req->misc.release.in;
 244
 245         /* Inode is NULL on error path of fuse_create_open() */
 246         if (likely(fi)) {
 247                 spin_lock(&fi->lock);
 248                 list_del(&ff->write_entry);
 249                 spin_unlock(&fi->lock);
 250         }
 251         spin_lock(&fc->lock);
 252         if (!RB_EMPTY_NODE(&ff->polled_node))
 253                 rb_erase(&ff->polled_node, &fc->polled_files);
 254         spin_unlock(&fc->lock);
 255
 256         wake_up_interruptible_all(&ff->poll_wait);
 257
 258         inarg->fh = ff->fh;
 259         inarg->flags = flags;
 260         req->in.h.opcode = opcode;
 261         req->in.h.nodeid = ff->nodeid;
 262         req->in.numargs = 1;
 263         req->in.args[0].size = sizeof(struct fuse_release_in);
 264         req->in.args[0].value = inarg;
 265 }
 266
 267 void fuse_release_common(struct file *file, bool isdir)
 268 {
 269         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 270         struct fuse_file *ff = file->private_data;
 271         struct fuse_req *req = ff->reserved_req;
 272         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 273
 274         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 275
 276         if (ff->flock) {
 277                 struct fuse_release_in *inarg = &req->misc.release.in;
 278                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 279                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 280                                                        (fl_owner_t) file);
 281         }
 282         /* Hold inode until release is finished */
 283         req->misc.release.inode = igrab(file_inode(file));
 284
 285         /*
 286          * Normally this will send the RELEASE request, however if
 287          * some asynchronous READ or WRITE requests are outstanding,
 288          * the sending will be delayed.
 289          *
 290          * Make the release synchronous if this is a fuseblk mount,
 291          * synchronous RELEASE is allowed (and desirable) in this case
 292          * because the server can be trusted not to screw up.
 293          */
 294         fuse_file_put(ff, ff->fc->destroy, isdir);
 295 }
 296
 297 static int fuse_open(struct inode *inode, struct file *file)
 298 {
 299         return fuse_open_common(inode, file, false);
 300 }
 301
 302 static int fuse_release(struct inode *inode, struct file *file)
 303 {
 304         struct fuse_conn *fc = get_fuse_conn(inode);
 305
 306         /* see fuse_vma_close() for !writeback_cache case */
 307         if (fc->writeback_cache)
 308                 write_inode_now(inode, 1);
 309
 310         fuse_release_common(file, false);
 311
 312         /* return value is ignored by VFS */
 313         return 0;
 314 }
 315
 316 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 317 {
 318         WARN_ON(refcount_read(&ff->count) > 1);
 319         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 320         /*
 321          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 322          * synchronous, we are fine with not doing igrab() here"
 323          */
 324         fuse_file_put(ff, true, false);
 325 }
 326 EXPORT_SYMBOL_GPL(fuse_sync_release);
 327
 328 /*
 329  * Scramble the ID space with XTEA, so that the value of the files_struct
 330  * pointer is not exposed to userspace.
 331  */
 332 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 333 {
 334         u32 *k = fc->scramble_key;
 335         u64 v = (unsigned long) id;
 336         u32 v0 = v;
 337         u32 v1 = v >> 32;
 338         u32 sum = 0;
 339         int i;
 340
 341         for (i = 0; i < 32; i++) {
 342                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 343                 sum += 0x9E3779B9;
 344                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 345         }
 346
 347         return (u64) v0 + ((u64) v1 << 32);
 348 }
 349
 350 static struct fuse_req *fuse_find_writeback(struct fuse_inode *fi,
 351                                             pgoff_t idx_from, pgoff_t idx_to)
 352 {
 353         struct fuse_req *req;
 354
 355         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 356                 pgoff_t curr_index;
 357
 358                 WARN_ON(get_fuse_inode(req->inode) != fi);
 359                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 360                 if (idx_from < curr_index + req->num_pages &&
 361                     curr_index <= idx_to) {
 362                         return req;
 363                 }
 364         }
 365         return NULL;
 366 }
 367
 368 /*
 369  * Check if any page in a range is under writeback
 370  *
 371  * This is currently done by walking the list of writepage requests
 372  * for the inode, which can be pretty inefficient.
 373  */
 374 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 375                                    pgoff_t idx_to)
 376 {
 377         struct fuse_inode *fi = get_fuse_inode(inode);
 378         bool found;
 379
 380         spin_lock(&fi->lock);
 381         found = fuse_find_writeback(fi, idx_from, idx_to);
 382         spin_unlock(&fi->lock);
 383
 384         return found;
 385 }
 386
 387 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 388 {
 389         return fuse_range_is_writeback(inode, index, index);
 390 }
 391
 392 /*
 393  * Wait for page writeback to be completed.
 394  *
 395  * Since fuse doesn't rely on the VM writeback tracking, this has to
 396  * use some other means.
 397  */
 398 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 399 {
 400         struct fuse_inode *fi = get_fuse_inode(inode);
 401
 402         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 403 }
 404
 405 /*
 406  * Wait for all pending writepages on the inode to finish.
 407  *
 408  * This is currently done by blocking further writes with FUSE_NOWRITE
 409  * and waiting for all sent writes to complete.
 410  *
 411  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 412  * could conflict with truncation.
 413  */
 414 static void fuse_sync_writes(struct inode *inode)
 415 {
 416         fuse_set_nowrite(inode);
 417         fuse_release_nowrite(inode);
 418 }
 419
 420 static int fuse_flush(struct file *file, fl_owner_t id)
 421 {
 422         struct inode *inode = file_inode(file);
 423         struct fuse_conn *fc = get_fuse_conn(inode);
 424         struct fuse_file *ff = file->private_data;
 425         struct fuse_flush_in inarg;
 426         FUSE_ARGS(args);
 427         int err;
 428
 429         if (is_bad_inode(inode))
 430                 return -EIO;
 431
 432         if (fc->no_flush)
 433                 return 0;
 434
 435         err = write_inode_now(inode, 1);
 436         if (err)
 437                 return err;
 438
 439         inode_lock(inode);
 440         fuse_sync_writes(inode);
 441         inode_unlock(inode);
 442
 443         err = filemap_check_errors(file->f_mapping);
 444         if (err)
 445                 return err;
 446
 447         memset(&inarg, 0, sizeof(inarg));
 448         inarg.fh = ff->fh;
 449         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 450         args.opcode = FUSE_FLUSH;
 451         args.nodeid = get_node_id(inode);
 452         args.in_numargs = 1;
 453         args.in_args[0].size = sizeof(inarg);
 454         args.in_args[0].value = &inarg;
 455         args.force = true;
 456
 457         err = fuse_simple_request(fc, &args);
 458         if (err == -ENOSYS) {
 459                 fc->no_flush = 1;
 460                 err = 0;
 461         }
 462         return err;
 463 }
 464
 465 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 466                       int datasync, int opcode)
 467 {
 468         struct inode *inode = file->f_mapping->host;
 469         struct fuse_conn *fc = get_fuse_conn(inode);
 470         struct fuse_file *ff = file->private_data;
 471         FUSE_ARGS(args);
 472         struct fuse_fsync_in inarg;
 473
 474         memset(&inarg, 0, sizeof(inarg));
 475         inarg.fh = ff->fh;
 476         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 477         args.opcode = opcode;
 478         args.nodeid = get_node_id(inode);
 479         args.in_numargs = 1;
 480         args.in_args[0].size = sizeof(inarg);
 481         args.in_args[0].value = &inarg;
 482         return fuse_simple_request(fc, &args);
 483 }
 484
 485 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 486                       int datasync)
 487 {
 488         struct inode *inode = file->f_mapping->host;
 489         struct fuse_conn *fc = get_fuse_conn(inode);
 490         int err;
 491
 492         if (is_bad_inode(inode))
 493                 return -EIO;
 494
 495         inode_lock(inode);
 496
 497         /*
 498          * Start writeback against all dirty pages of the inode, then
 499          * wait for all outstanding writes, before sending the FSYNC
 500          * request.
 501          */
 502         err = file_write_and_wait_range(file, start, end);
 503         if (err)
 504                 goto out;
 505
 506         fuse_sync_writes(inode);
 507
 508         /*
 509          * Due to implementation of fuse writeback
 510          * file_write_and_wait_range() does not catch errors.
 511          * We have to do this directly after fuse_sync_writes()
 512          */
 513         err = file_check_and_advance_wb_err(file);
 514         if (err)
 515                 goto out;
 516
 517         err = sync_inode_metadata(inode, 1);
 518         if (err)
 519                 goto out;
 520
 521         if (fc->no_fsync)
 522                 goto out;
 523
 524         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 525         if (err == -ENOSYS) {
 526                 fc->no_fsync = 1;
 527                 err = 0;
 528         }
 529 out:
 530         inode_unlock(inode);
 531
 532         return err;
 533 }
 534
 535 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 536                     size_t count, int opcode)
 537 {
 538         struct fuse_read_in *inarg = &req->misc.read.in;
 539         struct fuse_file *ff = file->private_data;
 540
 541         inarg->fh = ff->fh;
 542         inarg->offset = pos;
 543         inarg->size = count;
 544         inarg->flags = file->f_flags;
 545         req->in.h.opcode = opcode;
 546         req->in.h.nodeid = ff->nodeid;
 547         req->in.numargs = 1;
 548         req->in.args[0].size = sizeof(struct fuse_read_in);
 549         req->in.args[0].value = inarg;
 550         req->out.argvar = 1;
 551         req->out.numargs = 1;
 552         req->out.args[0].size = count;
 553 }
 554
 555 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 556 {
 557         unsigned i;
 558
 559         for (i = 0; i < req->num_pages; i++) {
 560                 struct page *page = req->pages[i];
 561                 if (should_dirty)
 562                         set_page_dirty_lock(page);
 563                 put_page(page);
 564         }
 565 }
 566
 567 static void fuse_io_release(struct kref *kref)
 568 {
 569         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 570 }
 571
 572 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 573 {
 574         if (io->err)
 575                 return io->err;
 576
 577         if (io->bytes >= 0 && io->write)
 578                 return -EIO;
 579
 580         return io->bytes < 0 ? io->size : io->bytes;
 581 }
 582
 583 /**
 584  * In case of short read, the caller sets 'pos' to the position of
 585  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 586  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 587  *
 588  * An example:
 589  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 590  * both submitted asynchronously. The first of them was ACKed by userspace as
 591  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 592  * second request was ACKed as short, e.g. only 1K was read, resulting in
 593  * pos == 33K.
 594  *
 595  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 596  * will be equal to the length of the longest contiguous fragment of
 597  * transferred data starting from the beginning of IO request.
 598  */
 599 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 600 {
 601         int left;
 602
 603         spin_lock(&io->lock);
 604         if (err)
 605                 io->err = io->err ? : err;
 606         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 607                 io->bytes = pos;
 608
 609         left = --io->reqs;
 610         if (!left && io->blocking)
 611                 complete(io->done);
 612         spin_unlock(&io->lock);
 613
 614         if (!left && !io->blocking) {
 615                 ssize_t res = fuse_get_res_by_io(io);
 616
 617                 if (res >= 0) {
 618                         struct inode *inode = file_inode(io->iocb->ki_filp);
 619                         struct fuse_conn *fc = get_fuse_conn(inode);
 620                         struct fuse_inode *fi = get_fuse_inode(inode);
 621
 622                         spin_lock(&fi->lock);
 623                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 624                         spin_unlock(&fi->lock);
 625                 }
 626
 627                 io->iocb->ki_complete(io->iocb, res, 0);
 628         }
 629
 630         kref_put(&io->refcnt, fuse_io_release);
 631 }
 632
 633 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 634 {
 635         struct fuse_io_priv *io = req->io;
 636         ssize_t pos = -1;
 637
 638         fuse_release_user_pages(req, io->should_dirty);
 639
 640         if (io->write) {
 641                 if (req->misc.write.in.size != req->misc.write.out.size)
 642                         pos = req->misc.write.in.offset - io->offset +
 643                                 req->misc.write.out.size;
 644         } else {
 645                 if (req->misc.read.in.size != req->out.args[0].size)
 646                         pos = req->misc.read.in.offset - io->offset +
 647                                 req->out.args[0].size;
 648         }
 649
 650         fuse_aio_complete(io, req->out.h.error, pos);
 651 }
 652
 653 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 654                 size_t num_bytes, struct fuse_io_priv *io)
 655 {
 656         spin_lock(&io->lock);
 657         kref_get(&io->refcnt);
 658         io->size += num_bytes;
 659         io->reqs++;
 660         spin_unlock(&io->lock);
 661
 662         req->io = io;
 663         req->end = fuse_aio_complete_req;
 664
 665         __fuse_get_request(req);
 666         fuse_request_send_background(fc, req);
 667
 668         return num_bytes;
 669 }
 670
 671 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 672                              loff_t pos, size_t count, fl_owner_t owner)
 673 {
 674         struct file *file = io->iocb->ki_filp;
 675         struct fuse_file *ff = file->private_data;
 676         struct fuse_conn *fc = ff->fc;
 677
 678         fuse_read_fill(req, file, pos, count, FUSE_READ);
 679         if (owner != NULL) {
 680                 struct fuse_read_in *inarg = &req->misc.read.in;
 681
 682                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 683                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 684         }
 685
 686         if (io->async)
 687                 return fuse_async_req_send(fc, req, count, io);
 688
 689         fuse_request_send(fc, req);
 690         return req->out.args[0].size;
 691 }
 692
 693 static void fuse_read_update_size(struct inode *inode, loff_t size,
 694                                   u64 attr_ver)
 695 {
 696         struct fuse_conn *fc = get_fuse_conn(inode);
 697         struct fuse_inode *fi = get_fuse_inode(inode);
 698
 699         spin_lock(&fi->lock);
 700         if (attr_ver == fi->attr_version && size < inode->i_size &&
 701             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 702                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 703                 i_size_write(inode, size);
 704         }
 705         spin_unlock(&fi->lock);
 706 }
 707
 708 static void fuse_short_read(struct inode *inode, u64 attr_ver, size_t num_read,
 709                             struct page **pages, unsigned int num_pages)
 710 {
 711         struct fuse_conn *fc = get_fuse_conn(inode);
 712
 713         if (fc->writeback_cache) {
 714                 /*
 715                  * A hole in a file. Some data after the hole are in page cache,
 716                  * but have not reached the client fs yet. So, the hole is not
 717                  * present there.
 718                  */
 719                 int i;
 720                 int start_idx = num_read >> PAGE_SHIFT;
 721                 size_t off = num_read & (PAGE_SIZE - 1);
 722
 723                 for (i = start_idx; i < num_pages; i++) {
 724                         zero_user_segment(pages[i], off, PAGE_SIZE);
 725                         off = 0;
 726                 }
 727         } else {
 728                 loff_t pos = page_offset(pages[0]) + num_read;
 729                 fuse_read_update_size(inode, pos, attr_ver);
 730         }
 731 }
 732
 733 static int fuse_do_readpage(struct file *file, struct page *page)
 734 {
 735         struct kiocb iocb;
 736         struct fuse_io_priv io;
 737         struct inode *inode = page->mapping->host;
 738         struct fuse_conn *fc = get_fuse_conn(inode);
 739         struct fuse_req *req;
 740         size_t num_read;
 741         loff_t pos = page_offset(page);
 742         size_t count = PAGE_SIZE;
 743         u64 attr_ver;
 744         int err;
 745
 746         /*
 747          * Page writeback can extend beyond the lifetime of the
 748          * page-cache page, so make sure we read a properly synced
 749          * page.
 750          */
 751         fuse_wait_on_page_writeback(inode, page->index);
 752
 753         req = fuse_get_req(fc, 1);
 754         if (IS_ERR(req))
 755                 return PTR_ERR(req);
 756
 757         attr_ver = fuse_get_attr_version(fc);
 758
 759         req->out.page_zeroing = 1;
 760         req->out.argpages = 1;
 761         req->num_pages = 1;
 762         req->pages[0] = page;
 763         req->page_descs[0].length = count;
 764         init_sync_kiocb(&iocb, file);
 765         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 766         num_read = fuse_send_read(req, &io, pos, count, NULL);
 767         err = req->out.h.error;
 768
 769         if (!err) {
 770                 /*
 771                  * Short read means EOF.  If file size is larger, truncate it
 772                  */
 773                 if (num_read < count)
 774                         fuse_short_read(inode, attr_ver, num_read, req->pages,
 775                                         req->num_pages);
 776
 777                 SetPageUptodate(page);
 778         }
 779
 780         fuse_put_request(fc, req);
 781
 782         return err;
 783 }
 784
 785 static int fuse_readpage(struct file *file, struct page *page)
 786 {
 787         struct inode *inode = page->mapping->host;
 788         int err;
 789
 790         err = -EIO;
 791         if (is_bad_inode(inode))
 792                 goto out;
 793
 794         err = fuse_do_readpage(file, page);
 795         fuse_invalidate_atime(inode);
 796  out:
 797         unlock_page(page);
 798         return err;
 799 }
 800
 801 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 802 {
 803         int i;
 804         size_t count = req->misc.read.in.size;
 805         size_t num_read = req->out.args[0].size;
 806         struct address_space *mapping = NULL;
 807
 808         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 809                 mapping = req->pages[i]->mapping;
 810
 811         if (mapping) {
 812                 struct inode *inode = mapping->host;
 813
 814                 /*
 815                  * Short read means EOF. If file size is larger, truncate it
 816                  */
 817                 if (!req->out.h.error && num_read < count)
 818                         fuse_short_read(inode, req->misc.read.attr_ver,
 819                                         num_read, req->pages, req->num_pages);
 820
 821                 fuse_invalidate_atime(inode);
 822         }
 823
 824         for (i = 0; i < req->num_pages; i++) {
 825                 struct page *page = req->pages[i];
 826                 if (!req->out.h.error)
 827                         SetPageUptodate(page);
 828                 else
 829                         SetPageError(page);
 830                 unlock_page(page);
 831                 put_page(page);
 832         }
 833         if (req->ff)
 834                 fuse_file_put(req->ff, false, false);
 835 }
 836
 837 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 838 {
 839         struct fuse_file *ff = file->private_data;
 840         struct fuse_conn *fc = ff->fc;
 841         loff_t pos = page_offset(req->pages[0]);
 842         size_t count = req->num_pages << PAGE_SHIFT;
 843
 844         req->out.argpages = 1;
 845         req->out.page_zeroing = 1;
 846         req->out.page_replace = 1;
 847         fuse_read_fill(req, file, pos, count, FUSE_READ);
 848         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 849         if (fc->async_read) {
 850                 req->ff = fuse_file_get(ff);
 851                 req->end = fuse_readpages_end;
 852                 fuse_request_send_background(fc, req);
 853         } else {
 854                 fuse_request_send(fc, req);
 855                 fuse_readpages_end(fc, req);
 856                 fuse_put_request(fc, req);
 857         }
 858 }
 859
 860 struct fuse_fill_data {
 861         struct fuse_req *req;
 862         struct file *file;
 863         struct inode *inode;
 864         unsigned nr_pages;
 865 };
 866
 867 static int fuse_readpages_fill(void *_data, struct page *page)
 868 {
 869         struct fuse_fill_data *data = _data;
 870         struct fuse_req *req = data->req;
 871         struct inode *inode = data->inode;
 872         struct fuse_conn *fc = get_fuse_conn(inode);
 873
 874         fuse_wait_on_page_writeback(inode, page->index);
 875
 876         if (req->num_pages &&
 877             (req->num_pages == fc->max_pages ||
 878              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 879              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 880                 unsigned int nr_alloc = min_t(unsigned int, data->nr_pages,
 881                                               fc->max_pages);
 882                 fuse_send_readpages(req, data->file);
 883                 if (fc->async_read)
 884                         req = fuse_get_req_for_background(fc, nr_alloc);
 885                 else
 886                         req = fuse_get_req(fc, nr_alloc);
 887
 888                 data->req = req;
 889                 if (IS_ERR(req)) {
 890                         unlock_page(page);
 891                         return PTR_ERR(req);
 892                 }
 893         }
 894
 895         if (WARN_ON(req->num_pages >= req->max_pages)) {
 896                 unlock_page(page);
 897                 fuse_put_request(fc, req);
 898                 return -EIO;
 899         }
 900
 901         get_page(page);
 902         req->pages[req->num_pages] = page;
 903         req->page_descs[req->num_pages].length = PAGE_SIZE;
 904         req->num_pages++;
 905         data->nr_pages--;
 906         return 0;
 907 }
 908
 909 static int fuse_readpages(struct file *file, struct address_space *mapping,
 910                           struct list_head *pages, unsigned nr_pages)
 911 {
 912         struct inode *inode = mapping->host;
 913         struct fuse_conn *fc = get_fuse_conn(inode);
 914         struct fuse_fill_data data;
 915         int err;
 916         unsigned int nr_alloc = min_t(unsigned int, nr_pages, fc->max_pages);
 917
 918         err = -EIO;
 919         if (is_bad_inode(inode))
 920                 goto out;
 921
 922         data.file = file;
 923         data.inode = inode;
 924         if (fc->async_read)
 925                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 926         else
 927                 data.req = fuse_get_req(fc, nr_alloc);
 928         data.nr_pages = nr_pages;
 929         err = PTR_ERR(data.req);
 930         if (IS_ERR(data.req))
 931                 goto out;
 932
 933         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 934         if (!err) {
 935                 if (data.req->num_pages)
 936                         fuse_send_readpages(data.req, file);
 937                 else
 938                         fuse_put_request(fc, data.req);
 939         }
 940 out:
 941         return err;
 942 }
 943
 944 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 945 {
 946         struct inode *inode = iocb->ki_filp->f_mapping->host;
 947         struct fuse_conn *fc = get_fuse_conn(inode);
 948
 949         /*
 950          * In auto invalidate mode, always update attributes on read.
 951          * Otherwise, only update if we attempt to read past EOF (to ensure
 952          * i_size is up to date).
 953          */
 954         if (fc->auto_inval_data ||
 955             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 956                 int err;
 957                 err = fuse_update_attributes(inode, iocb->ki_filp);
 958                 if (err)
 959                         return err;
 960         }
 961
 962         return generic_file_read_iter(iocb, to);
 963 }
 964
 965 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 966                             loff_t pos, size_t count)
 967 {
 968         struct fuse_write_in *inarg = &req->misc.write.in;
 969         struct fuse_write_out *outarg = &req->misc.write.out;
 970
 971         inarg->fh = ff->fh;
 972         inarg->offset = pos;
 973         inarg->size = count;
 974         req->in.h.opcode = FUSE_WRITE;
 975         req->in.h.nodeid = ff->nodeid;
 976         req->in.numargs = 2;
 977         if (ff->fc->minor < 9)
 978                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 979         else
 980                 req->in.args[0].size = sizeof(struct fuse_write_in);
 981         req->in.args[0].value = inarg;
 982         req->in.args[1].size = count;
 983         req->out.numargs = 1;
 984         req->out.args[0].size = sizeof(struct fuse_write_out);
 985         req->out.args[0].value = outarg;
 986 }
 987
 988 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 989                               loff_t pos, size_t count, fl_owner_t owner)
 990 {
 991         struct kiocb *iocb = io->iocb;
 992         struct file *file = iocb->ki_filp;
 993         struct fuse_file *ff = file->private_data;
 994         struct fuse_conn *fc = ff->fc;
 995         struct fuse_write_in *inarg = &req->misc.write.in;
 996
 997         fuse_write_fill(req, ff, pos, count);
 998         inarg->flags = file->f_flags;
 999         if (iocb->ki_flags & IOCB_DSYNC)
1000                 inarg->flags |= O_DSYNC;
1001         if (iocb->ki_flags & IOCB_SYNC)
1002                 inarg->flags |= O_SYNC;
1003         if (owner != NULL) {
1004                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
1005                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
1006         }
1007
1008         if (io->async)
1009                 return fuse_async_req_send(fc, req, count, io);
1010
1011         fuse_request_send(fc, req);
1012         return req->misc.write.out.size;
1013 }
1014
1015 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1016 {
1017         struct fuse_conn *fc = get_fuse_conn(inode);
1018         struct fuse_inode *fi = get_fuse_inode(inode);
1019         bool ret = false;
1020
1021         spin_lock(&fi->lock);
1022         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1023         if (pos > inode->i_size) {
1024                 i_size_write(inode, pos);
1025                 ret = true;
1026         }
1027         spin_unlock(&fi->lock);
1028
1029         return ret;
1030 }
1031
1032 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1033                                     struct inode *inode, loff_t pos,
1034                                     size_t count)
1035 {
1036         size_t res;
1037         unsigned offset;
1038         unsigned i;
1039         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1040
1041         for (i = 0; i < req->num_pages; i++)
1042                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1043
1044         res = fuse_send_write(req, &io, pos, count, NULL);
1045
1046         offset = req->page_descs[0].offset;
1047         count = res;
1048         for (i = 0; i < req->num_pages; i++) {
1049                 struct page *page = req->pages[i];
1050
1051                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1052                         SetPageUptodate(page);
1053
1054                 if (count > PAGE_SIZE - offset)
1055                         count -= PAGE_SIZE - offset;
1056                 else
1057                         count = 0;
1058                 offset = 0;
1059
1060                 unlock_page(page);
1061                 put_page(page);
1062         }
1063
1064         return res;
1065 }
1066
1067 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1068                                struct address_space *mapping,
1069                                struct iov_iter *ii, loff_t pos)
1070 {
1071         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1072         unsigned offset = pos & (PAGE_SIZE - 1);
1073         size_t count = 0;
1074         int err;
1075
1076         req->in.argpages = 1;
1077         req->page_descs[0].offset = offset;
1078
1079         do {
1080                 size_t tmp;
1081                 struct page *page;
1082                 pgoff_t index = pos >> PAGE_SHIFT;
1083                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1084                                      iov_iter_count(ii));
1085
1086                 bytes = min_t(size_t, bytes, fc->max_write - count);
1087
1088  again:
1089                 err = -EFAULT;
1090                 if (iov_iter_fault_in_readable(ii, bytes))
1091                         break;
1092
1093                 err = -ENOMEM;
1094                 page = grab_cache_page_write_begin(mapping, index, 0);
1095                 if (!page)
1096                         break;
1097
1098                 if (mapping_writably_mapped(mapping))
1099                         flush_dcache_page(page);
1100
1101                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1102                 flush_dcache_page(page);
1103
1104                 iov_iter_advance(ii, tmp);
1105                 if (!tmp) {
1106                         unlock_page(page);
1107                         put_page(page);
1108                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1109                         goto again;
1110                 }
1111
1112                 err = 0;
1113                 req->pages[req->num_pages] = page;
1114                 req->page_descs[req->num_pages].length = tmp;
1115                 req->num_pages++;
1116
1117                 count += tmp;
1118                 pos += tmp;
1119                 offset += tmp;
1120                 if (offset == PAGE_SIZE)
1121                         offset = 0;
1122
1123                 if (!fc->big_writes)
1124                         break;
1125         } while (iov_iter_count(ii) && count < fc->max_write &&
1126                  req->num_pages < req->max_pages && offset == 0);
1127
1128         return count > 0 ? count : err;
1129 }
1130
1131 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1132                                      unsigned int max_pages)
1133 {
1134         return min_t(unsigned int,
1135                      ((pos + len - 1) >> PAGE_SHIFT) -
1136                      (pos >> PAGE_SHIFT) + 1,
1137                      max_pages);
1138 }
1139
1140 static ssize_t fuse_perform_write(struct kiocb *iocb,
1141                                   struct address_space *mapping,
1142                                   struct iov_iter *ii, loff_t pos)
1143 {
1144         struct inode *inode = mapping->host;
1145         struct fuse_conn *fc = get_fuse_conn(inode);
1146         struct fuse_inode *fi = get_fuse_inode(inode);
1147         int err = 0;
1148         ssize_t res = 0;
1149
1150         if (inode->i_size < pos + iov_iter_count(ii))
1151                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1152
1153         do {
1154                 struct fuse_req *req;
1155                 ssize_t count;
1156                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1157                                                       fc->max_pages);
1158
1159                 req = fuse_get_req(fc, nr_pages);
1160                 if (IS_ERR(req)) {
1161                         err = PTR_ERR(req);
1162                         break;
1163                 }
1164
1165                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1166                 if (count <= 0) {
1167                         err = count;
1168                 } else {
1169                         size_t num_written;
1170
1171                         num_written = fuse_send_write_pages(req, iocb, inode,
1172                                                             pos, count);
1173                         err = req->out.h.error;
1174                         if (!err) {
1175                                 res += num_written;
1176                                 pos += num_written;
1177
1178                                 /* break out of the loop on short write */
1179                                 if (num_written != count)
1180                                         err = -EIO;
1181                         }
1182                 }
1183                 fuse_put_request(fc, req);
1184         } while (!err && iov_iter_count(ii));
1185
1186         if (res > 0)
1187                 fuse_write_update_size(inode, pos);
1188
1189         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1190         fuse_invalidate_attr(inode);
1191
1192         return res > 0 ? res : err;
1193 }
1194
1195 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1196 {
1197         struct file *file = iocb->ki_filp;
1198         struct address_space *mapping = file->f_mapping;
1199         ssize_t written = 0;
1200         ssize_t written_buffered = 0;
1201         struct inode *inode = mapping->host;
1202         ssize_t err;
1203         loff_t endbyte = 0;
1204
1205         if (get_fuse_conn(inode)->writeback_cache) {
1206                 /* Update size (EOF optimization) and mode (SUID clearing) */
1207                 err = fuse_update_attributes(mapping->host, file);
1208                 if (err)
1209                         return err;
1210
1211                 return generic_file_write_iter(iocb, from);
1212         }
1213
1214         inode_lock(inode);
1215
1216         /* We can write back this queue in page reclaim */
1217         current->backing_dev_info = inode_to_bdi(inode);
1218
1219         err = generic_write_checks(iocb, from);
1220         if (err <= 0)
1221                 goto out;
1222
1223         err = file_remove_privs(file);
1224         if (err)
1225                 goto out;
1226
1227         err = file_update_time(file);
1228         if (err)
1229                 goto out;
1230
1231         if (iocb->ki_flags & IOCB_DIRECT) {
1232                 loff_t pos = iocb->ki_pos;
1233                 written = generic_file_direct_write(iocb, from);
1234                 if (written < 0 || !iov_iter_count(from))
1235                         goto out;
1236
1237                 pos += written;
1238
1239                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1240                 if (written_buffered < 0) {
1241                         err = written_buffered;
1242                         goto out;
1243                 }
1244                 endbyte = pos + written_buffered - 1;
1245
1246                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1247                                                    endbyte);
1248                 if (err)
1249                         goto out;
1250
1251                 invalidate_mapping_pages(file->f_mapping,
1252                                          pos >> PAGE_SHIFT,
1253                                          endbyte >> PAGE_SHIFT);
1254
1255                 written += written_buffered;
1256                 iocb->ki_pos = pos + written_buffered;
1257         } else {
1258                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1259                 if (written >= 0)
1260                         iocb->ki_pos += written;
1261         }
1262 out:
1263         current->backing_dev_info = NULL;
1264         inode_unlock(inode);
1265         if (written > 0)
1266                 written = generic_write_sync(iocb, written);
1267
1268         return written ? written : err;
1269 }
1270
1271 static inline void fuse_page_descs_length_init(struct fuse_page_desc *descs,
1272                                                unsigned int index,
1273                                                unsigned int nr_pages)
1274 {
1275         int i;
1276
1277         for (i = index; i < index + nr_pages; i++)
1278                 descs[i].length = PAGE_SIZE - descs[i].offset;
1279 }
1280
1281 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1282 {
1283         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1284 }
1285
1286 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1287                                         size_t max_size)
1288 {
1289         return min(iov_iter_single_seg_count(ii), max_size);
1290 }
1291
1292 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1293                                size_t *nbytesp, int write)
1294 {
1295         size_t nbytes = 0;  /* # bytes already packed in req */
1296         ssize_t ret = 0;
1297
1298         /* Special case for kernel I/O: can copy directly into the buffer */
1299         if (iov_iter_is_kvec(ii)) {
1300                 unsigned long user_addr = fuse_get_user_addr(ii);
1301                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1302
1303                 if (write)
1304                         req->in.args[1].value = (void *) user_addr;
1305                 else
1306                         req->out.args[0].value = (void *) user_addr;
1307
1308                 iov_iter_advance(ii, frag_size);
1309                 *nbytesp = frag_size;
1310                 return 0;
1311         }
1312
1313         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1314                 unsigned npages;
1315                 size_t start;
1316                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1317                                         *nbytesp - nbytes,
1318                                         req->max_pages - req->num_pages,
1319                                         &start);
1320                 if (ret < 0)
1321                         break;
1322
1323                 iov_iter_advance(ii, ret);
1324                 nbytes += ret;
1325
1326                 ret += start;
1327                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1328
1329                 req->page_descs[req->num_pages].offset = start;
1330                 fuse_page_descs_length_init(req->page_descs, req->num_pages,
1331                                             npages);
1332
1333                 req->num_pages += npages;
1334                 req->page_descs[req->num_pages - 1].length -=
1335                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1336         }
1337
1338         if (write)
1339                 req->in.argpages = 1;
1340         else
1341                 req->out.argpages = 1;
1342
1343         *nbytesp = nbytes;
1344
1345         return ret < 0 ? ret : 0;
1346 }
1347
1348 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1349                        loff_t *ppos, int flags)
1350 {
1351         int write = flags & FUSE_DIO_WRITE;
1352         int cuse = flags & FUSE_DIO_CUSE;
1353         struct file *file = io->iocb->ki_filp;
1354         struct inode *inode = file->f_mapping->host;
1355         struct fuse_file *ff = file->private_data;
1356         struct fuse_conn *fc = ff->fc;
1357         size_t nmax = write ? fc->max_write : fc->max_read;
1358         loff_t pos = *ppos;
1359         size_t count = iov_iter_count(iter);
1360         pgoff_t idx_from = pos >> PAGE_SHIFT;
1361         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1362         ssize_t res = 0;
1363         struct fuse_req *req;
1364         int err = 0;
1365
1366         if (io->async)
1367                 req = fuse_get_req_for_background(fc, iov_iter_npages(iter,
1368                                                                 fc->max_pages));
1369         else
1370                 req = fuse_get_req(fc, iov_iter_npages(iter, fc->max_pages));
1371         if (IS_ERR(req))
1372                 return PTR_ERR(req);
1373
1374         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1375                 if (!write)
1376                         inode_lock(inode);
1377                 fuse_sync_writes(inode);
1378                 if (!write)
1379                         inode_unlock(inode);
1380         }
1381
1382         io->should_dirty = !write && iter_is_iovec(iter);
1383         while (count) {
1384                 size_t nres;
1385                 fl_owner_t owner = current->files;
1386                 size_t nbytes = min(count, nmax);
1387                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1388                 if (err && !nbytes)
1389                         break;
1390
1391                 if (write) {
1392                         if (!capable(CAP_FSETID)) {
1393                                 struct fuse_write_in *inarg;
1394
1395                                 inarg = &req->misc.write.in;
1396                                 inarg->write_flags |= FUSE_WRITE_KILL_PRIV;
1397                         }
1398                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1399                 } else {
1400                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1401                 }
1402
1403                 if (!io->async)
1404                         fuse_release_user_pages(req, io->should_dirty);
1405                 if (req->out.h.error) {
1406                         err = req->out.h.error;
1407                         break;
1408                 } else if (nres > nbytes) {
1409                         res = 0;
1410                         err = -EIO;
1411                         break;
1412                 }
1413                 count -= nres;
1414                 res += nres;
1415                 pos += nres;
1416                 if (nres != nbytes)
1417                         break;
1418                 if (count) {
1419                         fuse_put_request(fc, req);
1420                         if (io->async)
1421                                 req = fuse_get_req_for_background(fc,
1422                                         iov_iter_npages(iter, fc->max_pages));
1423                         else
1424                                 req = fuse_get_req(fc, iov_iter_npages(iter,
1425                                                                 fc->max_pages));
1426                         if (IS_ERR(req))
1427                                 break;
1428                 }
1429         }
1430         if (!IS_ERR(req))
1431                 fuse_put_request(fc, req);
1432         if (res > 0)
1433                 *ppos = pos;
1434
1435         return res > 0 ? res : err;
1436 }
1437 EXPORT_SYMBOL_GPL(fuse_direct_io);
1438
1439 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1440                                   struct iov_iter *iter,
1441                                   loff_t *ppos)
1442 {
1443         ssize_t res;
1444         struct inode *inode = file_inode(io->iocb->ki_filp);
1445
1446         res = fuse_direct_io(io, iter, ppos, 0);
1447
1448         fuse_invalidate_atime(inode);
1449
1450         return res;
1451 }
1452
1453 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1454
1455 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1456 {
1457         ssize_t res;
1458
1459         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1460                 res = fuse_direct_IO(iocb, to);
1461         } else {
1462                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1463
1464                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1465         }
1466
1467         return res;
1468 }
1469
1470 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1471 {
1472         struct inode *inode = file_inode(iocb->ki_filp);
1473         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1474         ssize_t res;
1475
1476         /* Don't allow parallel writes to the same file */
1477         inode_lock(inode);
1478         res = generic_write_checks(iocb, from);
1479         if (res > 0) {
1480                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1481                         res = fuse_direct_IO(iocb, from);
1482                 } else {
1483                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1484                                              FUSE_DIO_WRITE);
1485                 }
1486         }
1487         fuse_invalidate_attr(inode);
1488         if (res > 0)
1489                 fuse_write_update_size(inode, iocb->ki_pos);
1490         inode_unlock(inode);
1491
1492         return res;
1493 }
1494
1495 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1496 {
1497         struct file *file = iocb->ki_filp;
1498         struct fuse_file *ff = file->private_data;
1499
1500         if (is_bad_inode(file_inode(file)))
1501                 return -EIO;
1502
1503         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1504                 return fuse_cache_read_iter(iocb, to);
1505         else
1506                 return fuse_direct_read_iter(iocb, to);
1507 }
1508
1509 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1510 {
1511         struct file *file = iocb->ki_filp;
1512         struct fuse_file *ff = file->private_data;
1513
1514         if (is_bad_inode(file_inode(file)))
1515                 return -EIO;
1516
1517         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1518                 return fuse_cache_write_iter(iocb, from);
1519         else
1520                 return fuse_direct_write_iter(iocb, from);
1521 }
1522
1523 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1524 {
1525         int i;
1526
1527         for (i = 0; i < req->num_pages; i++)
1528                 __free_page(req->pages[i]);
1529
1530         if (req->ff)
1531                 fuse_file_put(req->ff, false, false);
1532 }
1533
1534 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1535 {
1536         struct inode *inode = req->inode;
1537         struct fuse_inode *fi = get_fuse_inode(inode);
1538         struct backing_dev_info *bdi = inode_to_bdi(inode);
1539         int i;
1540
1541         list_del(&req->writepages_entry);
1542         for (i = 0; i < req->num_pages; i++) {
1543                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1544                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1545                 wb_writeout_inc(&bdi->wb);
1546         }
1547         wake_up(&fi->page_waitq);
1548 }
1549
1550 /* Called under fi->lock, may release and reacquire it */
1551 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1552                                 loff_t size)
1553 __releases(fi->lock)
1554 __acquires(fi->lock)
1555 {
1556         struct fuse_req *aux, *next;
1557         struct fuse_inode *fi = get_fuse_inode(req->inode);
1558         struct fuse_write_in *inarg = &req->misc.write.in;
1559         __u64 data_size = req->num_pages * PAGE_SIZE;
1560         bool queued;
1561
1562         if (inarg->offset + data_size <= size) {
1563                 inarg->size = data_size;
1564         } else if (inarg->offset < size) {
1565                 inarg->size = size - inarg->offset;
1566         } else {
1567                 /* Got truncated off completely */
1568                 goto out_free;
1569         }
1570
1571         req->in.args[1].size = inarg->size;
1572         queued = fuse_request_queue_background(fc, req);
1573         /* Fails on broken connection only */
1574         if (unlikely(!queued))
1575                 goto out_free;
1576
1577         fi->writectr++;
1578         return;
1579
1580  out_free:
1581         fuse_writepage_finish(fc, req);
1582         spin_unlock(&fi->lock);
1583
1584         /* After fuse_writepage_finish() aux request list is private */
1585         for (aux = req->misc.write.next; aux; aux = next) {
1586                 next = aux->misc.write.next;
1587                 aux->misc.write.next = NULL;
1588                 fuse_writepage_free(fc, aux);
1589                 fuse_put_request(fc, aux);
1590         }
1591
1592         fuse_writepage_free(fc, req);
1593         fuse_put_request(fc, req);
1594         spin_lock(&fi->lock);
1595 }
1596
1597 /*
1598  * If fi->writectr is positive (no truncate or fsync going on) send
1599  * all queued writepage requests.
1600  *
1601  * Called with fi->lock
1602  */
1603 void fuse_flush_writepages(struct inode *inode)
1604 __releases(fi->lock)
1605 __acquires(fi->lock)
1606 {
1607         struct fuse_conn *fc = get_fuse_conn(inode);
1608         struct fuse_inode *fi = get_fuse_inode(inode);
1609         loff_t crop = i_size_read(inode);
1610         struct fuse_req *req;
1611
1612         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1613                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1614                 list_del_init(&req->list);
1615                 fuse_send_writepage(fc, req, crop);
1616         }
1617 }
1618
1619 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1620 {
1621         struct inode *inode = req->inode;
1622         struct fuse_inode *fi = get_fuse_inode(inode);
1623
1624         mapping_set_error(inode->i_mapping, req->out.h.error);
1625         spin_lock(&fi->lock);
1626         while (req->misc.write.next) {
1627                 struct fuse_conn *fc = get_fuse_conn(inode);
1628                 struct fuse_write_in *inarg = &req->misc.write.in;
1629                 struct fuse_req *next = req->misc.write.next;
1630                 req->misc.write.next = next->misc.write.next;
1631                 next->misc.write.next = NULL;
1632                 next->ff = fuse_file_get(req->ff);
1633                 list_add(&next->writepages_entry, &fi->writepages);
1634
1635                 /*
1636                  * Skip fuse_flush_writepages() to make it easy to crop requests
1637                  * based on primary request size.
1638                  *
1639                  * 1st case (trivial): there are no concurrent activities using
1640                  * fuse_set/release_nowrite.  Then we're on safe side because
1641                  * fuse_flush_writepages() would call fuse_send_writepage()
1642                  * anyway.
1643                  *
1644                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1645                  * now for completion of all in-flight requests.  This happens
1646                  * rarely and no more than once per page, so this should be
1647                  * okay.
1648                  *
1649                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1650                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1651                  * that fuse_set_nowrite returned implies that all in-flight
1652                  * requests were completed along with all of their secondary
1653                  * requests.  Further primary requests are blocked by negative
1654                  * writectr.  Hence there cannot be any in-flight requests and
1655                  * no invocations of fuse_writepage_end() while we're in
1656                  * fuse_set_nowrite..fuse_release_nowrite section.
1657                  */
1658                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1659         }
1660         fi->writectr--;
1661         fuse_writepage_finish(fc, req);
1662         spin_unlock(&fi->lock);
1663         fuse_writepage_free(fc, req);
1664 }
1665
1666 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1667                                                struct fuse_inode *fi)
1668 {
1669         struct fuse_file *ff = NULL;
1670
1671         spin_lock(&fi->lock);
1672         if (!list_empty(&fi->write_files)) {
1673                 ff = list_entry(fi->write_files.next, struct fuse_file,
1674                                 write_entry);
1675                 fuse_file_get(ff);
1676         }
1677         spin_unlock(&fi->lock);
1678
1679         return ff;
1680 }
1681
1682 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1683                                              struct fuse_inode *fi)
1684 {
1685         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1686         WARN_ON(!ff);
1687         return ff;
1688 }
1689
1690 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1691 {
1692         struct fuse_conn *fc = get_fuse_conn(inode);
1693         struct fuse_inode *fi = get_fuse_inode(inode);
1694         struct fuse_file *ff;
1695         int err;
1696
1697         ff = __fuse_write_file_get(fc, fi);
1698         err = fuse_flush_times(inode, ff);
1699         if (ff)
1700                 fuse_file_put(ff, false, false);
1701
1702         return err;
1703 }
1704
1705 static int fuse_writepage_locked(struct page *page)
1706 {
1707         struct address_space *mapping = page->mapping;
1708         struct inode *inode = mapping->host;
1709         struct fuse_conn *fc = get_fuse_conn(inode);
1710         struct fuse_inode *fi = get_fuse_inode(inode);
1711         struct fuse_req *req;
1712         struct page *tmp_page;
1713         int error = -ENOMEM;
1714
1715         set_page_writeback(page);
1716
1717         req = fuse_request_alloc_nofs(1);
1718         if (!req)
1719                 goto err;
1720
1721         /* writeback always goes to bg_queue */
1722         __set_bit(FR_BACKGROUND, &req->flags);
1723         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1724         if (!tmp_page)
1725                 goto err_free;
1726
1727         error = -EIO;
1728         req->ff = fuse_write_file_get(fc, fi);
1729         if (!req->ff)
1730                 goto err_nofile;
1731
1732         fuse_write_fill(req, req->ff, page_offset(page), 0);
1733
1734         copy_highpage(tmp_page, page);
1735         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1736         req->misc.write.next = NULL;
1737         req->in.argpages = 1;
1738         req->num_pages = 1;
1739         req->pages[0] = tmp_page;
1740         req->page_descs[0].offset = 0;
1741         req->page_descs[0].length = PAGE_SIZE;
1742         req->end = fuse_writepage_end;
1743         req->inode = inode;
1744
1745         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1746         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1747
1748         spin_lock(&fi->lock);
1749         list_add(&req->writepages_entry, &fi->writepages);
1750         list_add_tail(&req->list, &fi->queued_writes);
1751         fuse_flush_writepages(inode);
1752         spin_unlock(&fi->lock);
1753
1754         end_page_writeback(page);
1755
1756         return 0;
1757
1758 err_nofile:
1759         __free_page(tmp_page);
1760 err_free:
1761         fuse_request_free(req);
1762 err:
1763         mapping_set_error(page->mapping, error);
1764         end_page_writeback(page);
1765         return error;
1766 }
1767
1768 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1769 {
1770         int err;
1771
1772         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1773                 /*
1774                  * ->writepages() should be called for sync() and friends.  We
1775                  * should only get here on direct reclaim and then we are
1776                  * allowed to skip a page which is already in flight
1777                  */
1778                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1779
1780                 redirty_page_for_writepage(wbc, page);
1781                 return 0;
1782         }
1783
1784         err = fuse_writepage_locked(page);
1785         unlock_page(page);
1786
1787         return err;
1788 }
1789
1790 struct fuse_fill_wb_data {
1791         struct fuse_req *req;
1792         struct fuse_file *ff;
1793         struct inode *inode;
1794         struct page **orig_pages;
1795 };
1796
1797 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1798 {
1799         struct fuse_req *req = data->req;
1800         struct inode *inode = data->inode;
1801         struct fuse_inode *fi = get_fuse_inode(inode);
1802         int num_pages = req->num_pages;
1803         int i;
1804
1805         req->ff = fuse_file_get(data->ff);
1806         spin_lock(&fi->lock);
1807         list_add_tail(&req->list, &fi->queued_writes);
1808         fuse_flush_writepages(inode);
1809         spin_unlock(&fi->lock);
1810
1811         for (i = 0; i < num_pages; i++)
1812                 end_page_writeback(data->orig_pages[i]);
1813 }
1814
1815 /*
1816  * First recheck under fi->lock if the offending offset is still under
1817  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1818  * one already added for a page at this offset.  If there's none, then insert
1819  * this new request onto the auxiliary list, otherwise reuse the existing one by
1820  * copying the new page contents over to the old temporary page.
1821  */
1822 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1823                                      struct page *page)
1824 {
1825         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1826         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1827         struct fuse_req *tmp;
1828         struct fuse_req *old_req;
1829
1830         WARN_ON(new_req->num_pages != 0);
1831
1832         spin_lock(&fi->lock);
1833         list_del(&new_req->writepages_entry);
1834         old_req = fuse_find_writeback(fi, page->index, page->index);
1835         if (!old_req) {
1836                 list_add(&new_req->writepages_entry, &fi->writepages);
1837                 spin_unlock(&fi->lock);
1838                 return false;
1839         }
1840
1841         new_req->num_pages = 1;
1842         for (tmp = old_req->misc.write.next; tmp; tmp = tmp->misc.write.next) {
1843                 pgoff_t curr_index;
1844
1845                 WARN_ON(tmp->inode != new_req->inode);
1846                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1847                 if (curr_index == page->index) {
1848                         WARN_ON(tmp->num_pages != 1);
1849                         WARN_ON(!test_bit(FR_PENDING, &tmp->flags));
1850                         swap(tmp->pages[0], new_req->pages[0]);
1851                         break;
1852                 }
1853         }
1854
1855         if (!tmp) {
1856                 new_req->misc.write.next = old_req->misc.write.next;
1857                 old_req->misc.write.next = new_req;
1858         }
1859
1860         spin_unlock(&fi->lock);
1861
1862         if (tmp) {
1863                 struct backing_dev_info *bdi = inode_to_bdi(new_req->inode);
1864
1865                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1866                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1867                 wb_writeout_inc(&bdi->wb);
1868                 fuse_writepage_free(fc, new_req);
1869                 fuse_request_free(new_req);
1870         }
1871
1872         return true;
1873 }
1874
1875 static int fuse_writepages_fill(struct page *page,
1876                 struct writeback_control *wbc, void *_data)
1877 {
1878         struct fuse_fill_wb_data *data = _data;
1879         struct fuse_req *req = data->req;
1880         struct inode *inode = data->inode;
1881         struct fuse_inode *fi = get_fuse_inode(inode);
1882         struct fuse_conn *fc = get_fuse_conn(inode);
1883         struct page *tmp_page;
1884         bool is_writeback;
1885         int err;
1886
1887         if (!data->ff) {
1888                 err = -EIO;
1889                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1890                 if (!data->ff)
1891                         goto out_unlock;
1892         }
1893
1894         /*
1895          * Being under writeback is unlikely but possible.  For example direct
1896          * read to an mmaped fuse file will set the page dirty twice; once when
1897          * the pages are faulted with get_user_pages(), and then after the read
1898          * completed.
1899          */
1900         is_writeback = fuse_page_is_writeback(inode, page->index);
1901
1902         if (req && req->num_pages &&
1903             (is_writeback || req->num_pages == fc->max_pages ||
1904              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1905              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1906                 fuse_writepages_send(data);
1907                 data->req = NULL;
1908         } else if (req && req->num_pages == req->max_pages) {
1909                 if (!fuse_req_realloc_pages(fc, req, GFP_NOFS)) {
1910                         fuse_writepages_send(data);
1911                         req = data->req = NULL;
1912                 }
1913         }
1914
1915         err = -ENOMEM;
1916         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1917         if (!tmp_page)
1918                 goto out_unlock;
1919
1920         /*
1921          * The page must not be redirtied until the writeout is completed
1922          * (i.e. userspace has sent a reply to the write request).  Otherwise
1923          * there could be more than one temporary page instance for each real
1924          * page.
1925          *
1926          * This is ensured by holding the page lock in page_mkwrite() while
1927          * checking fuse_page_is_writeback().  We already hold the page lock
1928          * since clear_page_dirty_for_io() and keep it held until we add the
1929          * request to the fi->writepages list and increment req->num_pages.
1930          * After this fuse_page_is_writeback() will indicate that the page is
1931          * under writeback, so we can release the page lock.
1932          */
1933         if (data->req == NULL) {
1934                 struct fuse_inode *fi = get_fuse_inode(inode);
1935
1936                 err = -ENOMEM;
1937                 req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
1938                 if (!req) {
1939                         __free_page(tmp_page);
1940                         goto out_unlock;
1941                 }
1942
1943                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1944                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1945                 req->misc.write.next = NULL;
1946                 req->in.argpages = 1;
1947                 __set_bit(FR_BACKGROUND, &req->flags);
1948                 req->num_pages = 0;
1949                 req->end = fuse_writepage_end;
1950                 req->inode = inode;
1951
1952                 spin_lock(&fi->lock);
1953                 list_add(&req->writepages_entry, &fi->writepages);
1954                 spin_unlock(&fi->lock);
1955
1956                 data->req = req;
1957         }
1958         set_page_writeback(page);
1959
1960         copy_highpage(tmp_page, page);
1961         req->pages[req->num_pages] = tmp_page;
1962         req->page_descs[req->num_pages].offset = 0;
1963         req->page_descs[req->num_pages].length = PAGE_SIZE;
1964
1965         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1966         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1967
1968         err = 0;
1969         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1970                 end_page_writeback(page);
1971                 data->req = NULL;
1972                 goto out_unlock;
1973         }
1974         data->orig_pages[req->num_pages] = page;
1975
1976         /*
1977          * Protected by fi->lock against concurrent access by
1978          * fuse_page_is_writeback().
1979          */
1980         spin_lock(&fi->lock);
1981         req->num_pages++;
1982         spin_unlock(&fi->lock);
1983
1984 out_unlock:
1985         unlock_page(page);
1986
1987         return err;
1988 }
1989
1990 static int fuse_writepages(struct address_space *mapping,
1991                            struct writeback_control *wbc)
1992 {
1993         struct inode *inode = mapping->host;
1994         struct fuse_conn *fc = get_fuse_conn(inode);
1995         struct fuse_fill_wb_data data;
1996         int err;
1997
1998         err = -EIO;
1999         if (is_bad_inode(inode))
2000                 goto out;
2001
2002         data.inode = inode;
2003         data.req = NULL;
2004         data.ff = NULL;
2005
2006         err = -ENOMEM;
2007         data.orig_pages = kcalloc(fc->max_pages,
2008                                   sizeof(struct page *),
2009                                   GFP_NOFS);
2010         if (!data.orig_pages)
2011                 goto out;
2012
2013         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2014         if (data.req) {
2015                 /* Ignore errors if we can write at least one page */
2016                 BUG_ON(!data.req->num_pages);
2017                 fuse_writepages_send(&data);
2018                 err = 0;
2019         }
2020         if (data.ff)
2021                 fuse_file_put(data.ff, false, false);
2022
2023         kfree(data.orig_pages);
2024 out:
2025         return err;
2026 }
2027
2028 /*
2029  * It's worthy to make sure that space is reserved on disk for the write,
2030  * but how to implement it without killing performance need more thinking.
2031  */
2032 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2033                 loff_t pos, unsigned len, unsigned flags,
2034                 struct page **pagep, void **fsdata)
2035 {
2036         pgoff_t index = pos >> PAGE_SHIFT;
2037         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2038         struct page *page;
2039         loff_t fsize;
2040         int err = -ENOMEM;
2041
2042         WARN_ON(!fc->writeback_cache);
2043
2044         page = grab_cache_page_write_begin(mapping, index, flags);
2045         if (!page)
2046                 goto error;
2047
2048         fuse_wait_on_page_writeback(mapping->host, page->index);
2049
2050         if (PageUptodate(page) || len == PAGE_SIZE)
2051                 goto success;
2052         /*
2053          * Check if the start this page comes after the end of file, in which
2054          * case the readpage can be optimized away.
2055          */
2056         fsize = i_size_read(mapping->host);
2057         if (fsize <= (pos & PAGE_MASK)) {
2058                 size_t off = pos & ~PAGE_MASK;
2059                 if (off)
2060                         zero_user_segment(page, 0, off);
2061                 goto success;
2062         }
2063         err = fuse_do_readpage(file, page);
2064         if (err)
2065                 goto cleanup;
2066 success:
2067         *pagep = page;
2068         return 0;
2069
2070 cleanup:
2071         unlock_page(page);
2072         put_page(page);
2073 error:
2074         return err;
2075 }
2076
2077 static int fuse_write_end(struct file *file, struct address_space *mapping,
2078                 loff_t pos, unsigned len, unsigned copied,
2079                 struct page *page, void *fsdata)
2080 {
2081         struct inode *inode = page->mapping->host;
2082
2083         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2084         if (!copied)
2085                 goto unlock;
2086
2087         if (!PageUptodate(page)) {
2088                 /* Zero any unwritten bytes at the end of the page */
2089                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2090                 if (endoff)
2091                         zero_user_segment(page, endoff, PAGE_SIZE);
2092                 SetPageUptodate(page);
2093         }
2094
2095         fuse_write_update_size(inode, pos + copied);
2096         set_page_dirty(page);
2097
2098 unlock:
2099         unlock_page(page);
2100         put_page(page);
2101
2102         return copied;
2103 }
2104
2105 static int fuse_launder_page(struct page *page)
2106 {
2107         int err = 0;
2108         if (clear_page_dirty_for_io(page)) {
2109                 struct inode *inode = page->mapping->host;
2110                 err = fuse_writepage_locked(page);
2111                 if (!err)
2112                         fuse_wait_on_page_writeback(inode, page->index);
2113         }
2114         return err;
2115 }
2116
2117 /*
2118  * Write back dirty pages now, because there may not be any suitable
2119  * open files later
2120  */
2121 static void fuse_vma_close(struct vm_area_struct *vma)
2122 {
2123         filemap_write_and_wait(vma->vm_file->f_mapping);
2124 }
2125
2126 /*
2127  * Wait for writeback against this page to complete before allowing it
2128  * to be marked dirty again, and hence written back again, possibly
2129  * before the previous writepage completed.
2130  *
2131  * Block here, instead of in ->writepage(), so that the userspace fs
2132  * can only block processes actually operating on the filesystem.
2133  *
2134  * Otherwise unprivileged userspace fs would be able to block
2135  * unrelated:
2136  *
2137  * - page migration
2138  * - sync(2)
2139  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2140  */
2141 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2142 {
2143         struct page *page = vmf->page;
2144         struct inode *inode = file_inode(vmf->vma->vm_file);
2145
2146         file_update_time(vmf->vma->vm_file);
2147         lock_page(page);
2148         if (page->mapping != inode->i_mapping) {
2149                 unlock_page(page);
2150                 return VM_FAULT_NOPAGE;
2151         }
2152
2153         fuse_wait_on_page_writeback(inode, page->index);
2154         return VM_FAULT_LOCKED;
2155 }
2156
2157 static const struct vm_operations_struct fuse_file_vm_ops = {
2158         .close          = fuse_vma_close,
2159         .fault          = filemap_fault,
2160         .map_pages      = filemap_map_pages,
2161         .page_mkwrite   = fuse_page_mkwrite,
2162 };
2163
2164 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2165 {
2166         struct fuse_file *ff = file->private_data;
2167
2168         if (ff->open_flags & FOPEN_DIRECT_IO) {
2169                 /* Can't provide the coherency needed for MAP_SHARED */
2170                 if (vma->vm_flags & VM_MAYSHARE)
2171                         return -ENODEV;
2172
2173                 invalidate_inode_pages2(file->f_mapping);
2174
2175                 return generic_file_mmap(file, vma);
2176         }
2177
2178         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2179                 fuse_link_write_file(file);
2180
2181         file_accessed(file);
2182         vma->vm_ops = &fuse_file_vm_ops;
2183         return 0;
2184 }
2185
2186 static int convert_fuse_file_lock(struct fuse_conn *fc,
2187                                   const struct fuse_file_lock *ffl,
2188                                   struct file_lock *fl)
2189 {
2190         switch (ffl->type) {
2191         case F_UNLCK:
2192                 break;
2193
2194         case F_RDLCK:
2195         case F_WRLCK:
2196                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2197                     ffl->end < ffl->start)
2198                         return -EIO;
2199
2200                 fl->fl_start = ffl->start;
2201                 fl->fl_end = ffl->end;
2202
2203                 /*
2204                  * Convert pid into init's pid namespace.  The locks API will
2205                  * translate it into the caller's pid namespace.
2206                  */
2207                 rcu_read_lock();
2208                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2209                 rcu_read_unlock();
2210                 break;
2211
2212         default:
2213                 return -EIO;
2214         }
2215         fl->fl_type = ffl->type;
2216         return 0;
2217 }
2218
2219 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2220                          const struct file_lock *fl, int opcode, pid_t pid,
2221                          int flock, struct fuse_lk_in *inarg)
2222 {
2223         struct inode *inode = file_inode(file);
2224         struct fuse_conn *fc = get_fuse_conn(inode);
2225         struct fuse_file *ff = file->private_data;
2226
2227         memset(inarg, 0, sizeof(*inarg));
2228         inarg->fh = ff->fh;
2229         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2230         inarg->lk.start = fl->fl_start;
2231         inarg->lk.end = fl->fl_end;
2232         inarg->lk.type = fl->fl_type;
2233         inarg->lk.pid = pid;
2234         if (flock)
2235                 inarg->lk_flags |= FUSE_LK_FLOCK;
2236         args->opcode = opcode;
2237         args->nodeid = get_node_id(inode);
2238         args->in_numargs = 1;
2239         args->in_args[0].size = sizeof(*inarg);
2240         args->in_args[0].value = inarg;
2241 }
2242
2243 static int fuse_getlk(struct file *file, struct file_lock *fl)
2244 {
2245         struct inode *inode = file_inode(file);
2246         struct fuse_conn *fc = get_fuse_conn(inode);
2247         FUSE_ARGS(args);
2248         struct fuse_lk_in inarg;
2249         struct fuse_lk_out outarg;
2250         int err;
2251
2252         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2253         args.out_numargs = 1;
2254         args.out_args[0].size = sizeof(outarg);
2255         args.out_args[0].value = &outarg;
2256         err = fuse_simple_request(fc, &args);
2257         if (!err)
2258                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2259
2260         return err;
2261 }
2262
2263 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2264 {
2265         struct inode *inode = file_inode(file);
2266         struct fuse_conn *fc = get_fuse_conn(inode);
2267         FUSE_ARGS(args);
2268         struct fuse_lk_in inarg;
2269         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2270         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2271         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2272         int err;
2273
2274         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2275                 /* NLM needs asynchronous locks, which we don't support yet */
2276                 return -ENOLCK;
2277         }
2278
2279         /* Unlock on close is handled by the flush method */
2280         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2281                 return 0;
2282
2283         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2284         err = fuse_simple_request(fc, &args);
2285
2286         /* locking is restartable */
2287         if (err == -EINTR)
2288                 err = -ERESTARTSYS;
2289
2290         return err;
2291 }
2292
2293 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2294 {
2295         struct inode *inode = file_inode(file);
2296         struct fuse_conn *fc = get_fuse_conn(inode);
2297         int err;
2298
2299         if (cmd == F_CANCELLK) {
2300                 err = 0;
2301         } else if (cmd == F_GETLK) {
2302                 if (fc->no_lock) {
2303                         posix_test_lock(file, fl);
2304                         err = 0;
2305                 } else
2306                         err = fuse_getlk(file, fl);
2307         } else {
2308                 if (fc->no_lock)
2309                         err = posix_lock_file(file, fl, NULL);
2310                 else
2311                         err = fuse_setlk(file, fl, 0);
2312         }
2313         return err;
2314 }
2315
2316 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2317 {
2318         struct inode *inode = file_inode(file);
2319         struct fuse_conn *fc = get_fuse_conn(inode);
2320         int err;
2321
2322         if (fc->no_flock) {
2323                 err = locks_lock_file_wait(file, fl);
2324         } else {
2325                 struct fuse_file *ff = file->private_data;
2326
2327                 /* emulate flock with POSIX locks */
2328                 ff->flock = true;
2329                 err = fuse_setlk(file, fl, 1);
2330         }
2331
2332         return err;
2333 }
2334
2335 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2336 {
2337         struct inode *inode = mapping->host;
2338         struct fuse_conn *fc = get_fuse_conn(inode);
2339         FUSE_ARGS(args);
2340         struct fuse_bmap_in inarg;
2341         struct fuse_bmap_out outarg;
2342         int err;
2343
2344         if (!inode->i_sb->s_bdev || fc->no_bmap)
2345                 return 0;
2346
2347         memset(&inarg, 0, sizeof(inarg));
2348         inarg.block = block;
2349         inarg.blocksize = inode->i_sb->s_blocksize;
2350         args.opcode = FUSE_BMAP;
2351         args.nodeid = get_node_id(inode);
2352         args.in_numargs = 1;
2353         args.in_args[0].size = sizeof(inarg);
2354         args.in_args[0].value = &inarg;
2355         args.out_numargs = 1;
2356         args.out_args[0].size = sizeof(outarg);
2357         args.out_args[0].value = &outarg;
2358         err = fuse_simple_request(fc, &args);
2359         if (err == -ENOSYS)
2360                 fc->no_bmap = 1;
2361
2362         return err ? 0 : outarg.block;
2363 }
2364
2365 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2366 {
2367         struct inode *inode = file->f_mapping->host;
2368         struct fuse_conn *fc = get_fuse_conn(inode);
2369         struct fuse_file *ff = file->private_data;
2370         FUSE_ARGS(args);
2371         struct fuse_lseek_in inarg = {
2372                 .fh = ff->fh,
2373                 .offset = offset,
2374                 .whence = whence
2375         };
2376         struct fuse_lseek_out outarg;
2377         int err;
2378
2379         if (fc->no_lseek)
2380                 goto fallback;
2381
2382         args.opcode = FUSE_LSEEK;
2383         args.nodeid = ff->nodeid;
2384         args.in_numargs = 1;
2385         args.in_args[0].size = sizeof(inarg);
2386         args.in_args[0].value = &inarg;
2387         args.out_numargs = 1;
2388         args.out_args[0].size = sizeof(outarg);
2389         args.out_args[0].value = &outarg;
2390         err = fuse_simple_request(fc, &args);
2391         if (err) {
2392                 if (err == -ENOSYS) {
2393                         fc->no_lseek = 1;
2394                         goto fallback;
2395                 }
2396                 return err;
2397         }
2398
2399         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2400
2401 fallback:
2402         err = fuse_update_attributes(inode, file);
2403         if (!err)
2404                 return generic_file_llseek(file, offset, whence);
2405         else
2406                 return err;
2407 }
2408
2409 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2410 {
2411         loff_t retval;
2412         struct inode *inode = file_inode(file);
2413
2414         switch (whence) {
2415         case SEEK_SET:
2416         case SEEK_CUR:
2417                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2418                 retval = generic_file_llseek(file, offset, whence);
2419                 break;
2420         case SEEK_END:
2421                 inode_lock(inode);
2422                 retval = fuse_update_attributes(inode, file);
2423                 if (!retval)
2424                         retval = generic_file_llseek(file, offset, whence);
2425                 inode_unlock(inode);
2426                 break;
2427         case SEEK_HOLE:
2428         case SEEK_DATA:
2429                 inode_lock(inode);
2430                 retval = fuse_lseek(file, offset, whence);
2431                 inode_unlock(inode);
2432                 break;
2433         default:
2434                 retval = -EINVAL;
2435         }
2436
2437         return retval;
2438 }
2439
2440 /*
2441  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2442  * ABI was defined to be 'struct iovec' which is different on 32bit
2443  * and 64bit.  Fortunately we can determine which structure the server
2444  * used from the size of the reply.
2445  */
2446 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2447                                      size_t transferred, unsigned count,
2448                                      bool is_compat)
2449 {
2450 #ifdef CONFIG_COMPAT
2451         if (count * sizeof(struct compat_iovec) == transferred) {
2452                 struct compat_iovec *ciov = src;
2453                 unsigned i;
2454
2455                 /*
2456                  * With this interface a 32bit server cannot support
2457                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2458                  * requests
2459                  */
2460                 if (!is_compat)
2461                         return -EINVAL;
2462
2463                 for (i = 0; i < count; i++) {
2464                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2465                         dst[i].iov_len = ciov[i].iov_len;
2466                 }
2467                 return 0;
2468         }
2469 #endif
2470
2471         if (count * sizeof(struct iovec) != transferred)
2472                 return -EIO;
2473
2474         memcpy(dst, src, transferred);
2475         return 0;
2476 }
2477
2478 /* Make sure iov_length() won't overflow */
2479 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2480                                  size_t count)
2481 {
2482         size_t n;
2483         u32 max = fc->max_pages << PAGE_SHIFT;
2484
2485         for (n = 0; n < count; n++, iov++) {
2486                 if (iov->iov_len > (size_t) max)
2487                         return -ENOMEM;
2488                 max -= iov->iov_len;
2489         }
2490         return 0;
2491 }
2492
2493 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2494                                  void *src, size_t transferred, unsigned count,
2495                                  bool is_compat)
2496 {
2497         unsigned i;
2498         struct fuse_ioctl_iovec *fiov = src;
2499
2500         if (fc->minor < 16) {
2501                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2502                                                  count, is_compat);
2503         }
2504
2505         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2506                 return -EIO;
2507
2508         for (i = 0; i < count; i++) {
2509                 /* Did the server supply an inappropriate value? */
2510                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2511                     fiov[i].len != (unsigned long) fiov[i].len)
2512                         return -EIO;
2513
2514                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2515                 dst[i].iov_len = (size_t) fiov[i].len;
2516
2517 #ifdef CONFIG_COMPAT
2518                 if (is_compat &&
2519                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2520                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2521                         return -EIO;
2522 #endif
2523         }
2524
2525         return 0;
2526 }
2527
2528
2529 /*
2530  * For ioctls, there is no generic way to determine how much memory
2531  * needs to be read and/or written.  Furthermore, ioctls are allowed
2532  * to dereference the passed pointer, so the parameter requires deep
2533  * copying but FUSE has no idea whatsoever about what to copy in or
2534  * out.
2535  *
2536  * This is solved by allowing FUSE server to retry ioctl with
2537  * necessary in/out iovecs.  Let's assume the ioctl implementation
2538  * needs to read in the following structure.
2539  *
2540  * struct a {
2541  *      char    *buf;
2542  *      size_t  buflen;
2543  * }
2544  *
2545  * On the first callout to FUSE server, inarg->in_size and
2546  * inarg->out_size will be NULL; then, the server completes the ioctl
2547  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2548  * the actual iov array to
2549  *
2550  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2551  *
2552  * which tells FUSE to copy in the requested area and retry the ioctl.
2553  * On the second round, the server has access to the structure and
2554  * from that it can tell what to look for next, so on the invocation,
2555  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2556  *
2557  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2558  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2559  *
2560  * FUSE will copy both struct a and the pointed buffer from the
2561  * process doing the ioctl and retry ioctl with both struct a and the
2562  * buffer.
2563  *
2564  * This time, FUSE server has everything it needs and completes ioctl
2565  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2566  *
2567  * Copying data out works the same way.
2568  *
2569  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2570  * automatically initializes in and out iovs by decoding @cmd with
2571  * _IOC_* macros and the server is not allowed to request RETRY.  This
2572  * limits ioctl data transfers to well-formed ioctls and is the forced
2573  * behavior for all FUSE servers.
2574  */
2575 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2576                    unsigned int flags)
2577 {
2578         struct fuse_file *ff = file->private_data;
2579         struct fuse_conn *fc = ff->fc;
2580         struct fuse_ioctl_in inarg = {
2581                 .fh = ff->fh,
2582                 .cmd = cmd,
2583                 .arg = arg,
2584                 .flags = flags
2585         };
2586         struct fuse_ioctl_out outarg;
2587         struct iovec *iov_page = NULL;
2588         struct iovec *in_iov = NULL, *out_iov = NULL;
2589         unsigned int in_iovs = 0, out_iovs = 0, max_pages;
2590         size_t in_size, out_size, c;
2591         ssize_t transferred;
2592         int err, i;
2593         struct iov_iter ii;
2594         struct fuse_args_pages ap = {};
2595
2596 #if BITS_PER_LONG == 32
2597         inarg.flags |= FUSE_IOCTL_32BIT;
2598 #else
2599         if (flags & FUSE_IOCTL_COMPAT) {
2600                 inarg.flags |= FUSE_IOCTL_32BIT;
2601 #ifdef CONFIG_X86_X32
2602                 if (in_x32_syscall())
2603                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2604 #endif
2605         }
2606 #endif
2607
2608         /* assume all the iovs returned by client always fits in a page */
2609         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2610
2611         err = -ENOMEM;
2612         ap.pages = fuse_pages_alloc(fc->max_pages, GFP_KERNEL, &ap.descs);
2613         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2614         if (!ap.pages || !iov_page)
2615                 goto out;
2616
2617         fuse_page_descs_length_init(ap.descs, 0, fc->max_pages);
2618
2619         /*
2620          * If restricted, initialize IO parameters as encoded in @cmd.
2621          * RETRY from server is not allowed.
2622          */
2623         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2624                 struct iovec *iov = iov_page;
2625
2626                 iov->iov_base = (void __user *)arg;
2627                 iov->iov_len = _IOC_SIZE(cmd);
2628
2629                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2630                         in_iov = iov;
2631                         in_iovs = 1;
2632                 }
2633
2634                 if (_IOC_DIR(cmd) & _IOC_READ) {
2635                         out_iov = iov;
2636                         out_iovs = 1;
2637                 }
2638         }
2639
2640  retry:
2641         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2642         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2643
2644         /*
2645          * Out data can be used either for actual out data or iovs,
2646          * make sure there always is at least one page.
2647          */
2648         out_size = max_t(size_t, out_size, PAGE_SIZE);
2649         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2650
2651         /* make sure there are enough buffer pages and init request with them */
2652         err = -ENOMEM;
2653         if (max_pages > fc->max_pages)
2654                 goto out;
2655         while (ap.num_pages < max_pages) {
2656                 ap.pages[ap.num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2657                 if (!ap.pages[ap.num_pages])
2658                         goto out;
2659                 ap.num_pages++;
2660         }
2661
2662
2663         /* okay, let's send it to the client */
2664         ap.args.opcode = FUSE_IOCTL;
2665         ap.args.nodeid = ff->nodeid;
2666         ap.args.in_numargs = 1;
2667         ap.args.in_args[0].size = sizeof(inarg);
2668         ap.args.in_args[0].value = &inarg;
2669         if (in_size) {
2670                 ap.args.in_numargs++;
2671                 ap.args.in_args[1].size = in_size;
2672                 ap.args.in_pages = true;
2673
2674                 err = -EFAULT;
2675                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2676                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2677                         c = copy_page_from_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2678                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2679                                 goto out;
2680                 }
2681         }
2682
2683         ap.args.out_numargs = 2;
2684         ap.args.out_args[0].size = sizeof(outarg);
2685         ap.args.out_args[0].value = &outarg;
2686         ap.args.out_args[1].size = out_size;
2687         ap.args.out_pages = true;
2688         ap.args.out_argvar = true;
2689
2690         transferred = fuse_simple_request(fc, &ap.args);
2691         err = transferred;
2692         if (transferred < 0)
2693                 goto out;
2694
2695         /* did it ask for retry? */
2696         if (outarg.flags & FUSE_IOCTL_RETRY) {
2697                 void *vaddr;
2698
2699                 /* no retry if in restricted mode */
2700                 err = -EIO;
2701                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2702                         goto out;
2703
2704                 in_iovs = outarg.in_iovs;
2705                 out_iovs = outarg.out_iovs;
2706
2707                 /*
2708                  * Make sure things are in boundary, separate checks
2709                  * are to protect against overflow.
2710                  */
2711                 err = -ENOMEM;
2712                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2713                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2714                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2715                         goto out;
2716
2717                 vaddr = kmap_atomic(ap.pages[0]);
2718                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2719                                             transferred, in_iovs + out_iovs,
2720                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2721                 kunmap_atomic(vaddr);
2722                 if (err)
2723                         goto out;
2724
2725                 in_iov = iov_page;
2726                 out_iov = in_iov + in_iovs;
2727
2728                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2729                 if (err)
2730                         goto out;
2731
2732                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2733                 if (err)
2734                         goto out;
2735
2736                 goto retry;
2737         }
2738
2739         err = -EIO;
2740         if (transferred > inarg.out_size)
2741                 goto out;
2742
2743         err = -EFAULT;
2744         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2745         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= ap.num_pages); i++) {
2746                 c = copy_page_to_iter(ap.pages[i], 0, PAGE_SIZE, &ii);
2747                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2748                         goto out;
2749         }
2750         err = 0;
2751  out:
2752         free_page((unsigned long) iov_page);
2753         while (ap.num_pages)
2754                 __free_page(ap.pages[--ap.num_pages]);
2755         kfree(ap.pages);
2756
2757         return err ? err : outarg.result;
2758 }
2759 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2760
2761 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2762                        unsigned long arg, unsigned int flags)
2763 {
2764         struct inode *inode = file_inode(file);
2765         struct fuse_conn *fc = get_fuse_conn(inode);
2766
2767         if (!fuse_allow_current_process(fc))
2768                 return -EACCES;
2769
2770         if (is_bad_inode(inode))
2771                 return -EIO;
2772
2773         return fuse_do_ioctl(file, cmd, arg, flags);
2774 }
2775
2776 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2777                             unsigned long arg)
2778 {
2779         return fuse_ioctl_common(file, cmd, arg, 0);
2780 }
2781
2782 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2783                                    unsigned long arg)
2784 {
2785         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2786 }
2787
2788 /*
2789  * All files which have been polled are linked to RB tree
2790  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2791  * find the matching one.
2792  */
2793 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2794                                               struct rb_node **parent_out)
2795 {
2796         struct rb_node **link = &fc->polled_files.rb_node;
2797         struct rb_node *last = NULL;
2798
2799         while (*link) {
2800                 struct fuse_file *ff;
2801
2802                 last = *link;
2803                 ff = rb_entry(last, struct fuse_file, polled_node);
2804
2805                 if (kh < ff->kh)
2806                         link = &last->rb_left;
2807                 else if (kh > ff->kh)
2808                         link = &last->rb_right;
2809                 else
2810                         return link;
2811         }
2812
2813         if (parent_out)
2814                 *parent_out = last;
2815         return link;
2816 }
2817
2818 /*
2819  * The file is about to be polled.  Make sure it's on the polled_files
2820  * RB tree.  Note that files once added to the polled_files tree are
2821  * not removed before the file is released.  This is because a file
2822  * polled once is likely to be polled again.
2823  */
2824 static void fuse_register_polled_file(struct fuse_conn *fc,
2825                                       struct fuse_file *ff)
2826 {
2827         spin_lock(&fc->lock);
2828         if (RB_EMPTY_NODE(&ff->polled_node)) {
2829                 struct rb_node **link, *uninitialized_var(parent);
2830
2831                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2832                 BUG_ON(*link);
2833                 rb_link_node(&ff->polled_node, parent, link);
2834                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2835         }
2836         spin_unlock(&fc->lock);
2837 }
2838
2839 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2840 {
2841         struct fuse_file *ff = file->private_data;
2842         struct fuse_conn *fc = ff->fc;
2843         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2844         struct fuse_poll_out outarg;
2845         FUSE_ARGS(args);
2846         int err;
2847
2848         if (fc->no_poll)
2849                 return DEFAULT_POLLMASK;
2850
2851         poll_wait(file, &ff->poll_wait, wait);
2852         inarg.events = mangle_poll(poll_requested_events(wait));
2853
2854         /*
2855          * Ask for notification iff there's someone waiting for it.
2856          * The client may ignore the flag and always notify.
2857          */
2858         if (waitqueue_active(&ff->poll_wait)) {
2859                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2860                 fuse_register_polled_file(fc, ff);
2861         }
2862
2863         args.opcode = FUSE_POLL;
2864         args.nodeid = ff->nodeid;
2865         args.in_numargs = 1;
2866         args.in_args[0].size = sizeof(inarg);
2867         args.in_args[0].value = &inarg;
2868         args.out_numargs = 1;
2869         args.out_args[0].size = sizeof(outarg);
2870         args.out_args[0].value = &outarg;
2871         err = fuse_simple_request(fc, &args);
2872
2873         if (!err)
2874                 return demangle_poll(outarg.revents);
2875         if (err == -ENOSYS) {
2876                 fc->no_poll = 1;
2877                 return DEFAULT_POLLMASK;
2878         }
2879         return EPOLLERR;
2880 }
2881 EXPORT_SYMBOL_GPL(fuse_file_poll);
2882
2883 /*
2884  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2885  * wakes up the poll waiters.
2886  */
2887 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2888                             struct fuse_notify_poll_wakeup_out *outarg)
2889 {
2890         u64 kh = outarg->kh;
2891         struct rb_node **link;
2892
2893         spin_lock(&fc->lock);
2894
2895         link = fuse_find_polled_node(fc, kh, NULL);
2896         if (*link) {
2897                 struct fuse_file *ff;
2898
2899                 ff = rb_entry(*link, struct fuse_file, polled_node);
2900                 wake_up_interruptible_sync(&ff->poll_wait);
2901         }
2902
2903         spin_unlock(&fc->lock);
2904         return 0;
2905 }
2906
2907 static void fuse_do_truncate(struct file *file)
2908 {
2909         struct inode *inode = file->f_mapping->host;
2910         struct iattr attr;
2911
2912         attr.ia_valid = ATTR_SIZE;
2913         attr.ia_size = i_size_read(inode);
2914
2915         attr.ia_file = file;
2916         attr.ia_valid |= ATTR_FILE;
2917
2918         fuse_do_setattr(file_dentry(file), &attr, file);
2919 }
2920
2921 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2922 {
2923         return round_up(off, fc->max_pages << PAGE_SHIFT);
2924 }
2925
2926 static ssize_t
2927 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2928 {
2929         DECLARE_COMPLETION_ONSTACK(wait);
2930         ssize_t ret = 0;
2931         struct file *file = iocb->ki_filp;
2932         struct fuse_file *ff = file->private_data;
2933         bool async_dio = ff->fc->async_dio;
2934         loff_t pos = 0;
2935         struct inode *inode;
2936         loff_t i_size;
2937         size_t count = iov_iter_count(iter);
2938         loff_t offset = iocb->ki_pos;
2939         struct fuse_io_priv *io;
2940
2941         pos = offset;
2942         inode = file->f_mapping->host;
2943         i_size = i_size_read(inode);
2944
2945         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2946                 return 0;
2947
2948         /* optimization for short read */
2949         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2950                 if (offset >= i_size)
2951                         return 0;
2952                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
2953                 count = iov_iter_count(iter);
2954         }
2955
2956         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2957         if (!io)
2958                 return -ENOMEM;
2959         spin_lock_init(&io->lock);
2960         kref_init(&io->refcnt);
2961         io->reqs = 1;
2962         io->bytes = -1;
2963         io->size = 0;
2964         io->offset = offset;
2965         io->write = (iov_iter_rw(iter) == WRITE);
2966         io->err = 0;
2967         /*
2968          * By default, we want to optimize all I/Os with async request
2969          * submission to the client filesystem if supported.
2970          */
2971         io->async = async_dio;
2972         io->iocb = iocb;
2973         io->blocking = is_sync_kiocb(iocb);
2974
2975         /*
2976          * We cannot asynchronously extend the size of a file.
2977          * In such case the aio will behave exactly like sync io.
2978          */
2979         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2980                 io->blocking = true;
2981
2982         if (io->async && io->blocking) {
2983                 /*
2984                  * Additional reference to keep io around after
2985                  * calling fuse_aio_complete()
2986                  */
2987                 kref_get(&io->refcnt);
2988                 io->done = &wait;
2989         }
2990
2991         if (iov_iter_rw(iter) == WRITE) {
2992                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2993                 fuse_invalidate_attr(inode);
2994         } else {
2995                 ret = __fuse_direct_read(io, iter, &pos);
2996         }
2997
2998         if (io->async) {
2999                 bool blocking = io->blocking;
3000
3001                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3002
3003                 /* we have a non-extending, async request, so return */
3004                 if (!blocking)
3005                         return -EIOCBQUEUED;
3006
3007                 wait_for_completion(&wait);
3008                 ret = fuse_get_res_by_io(io);
3009         }
3010
3011         kref_put(&io->refcnt, fuse_io_release);
3012
3013         if (iov_iter_rw(iter) == WRITE) {
3014                 if (ret > 0)
3015                         fuse_write_update_size(inode, pos);
3016                 else if (ret < 0 && offset + count > i_size)
3017                         fuse_do_truncate(file);
3018         }
3019
3020         return ret;
3021 }
3022
3023 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3024 {
3025         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3026
3027         if (!err)
3028                 fuse_sync_writes(inode);
3029
3030         return err;
3031 }
3032
3033 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3034                                 loff_t length)
3035 {
3036         struct fuse_file *ff = file->private_data;
3037         struct inode *inode = file_inode(file);
3038         struct fuse_inode *fi = get_fuse_inode(inode);
3039         struct fuse_conn *fc = ff->fc;
3040         FUSE_ARGS(args);
3041         struct fuse_fallocate_in inarg = {
3042                 .fh = ff->fh,
3043                 .offset = offset,
3044                 .length = length,
3045                 .mode = mode
3046         };
3047         int err;
3048         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3049                            (mode & FALLOC_FL_PUNCH_HOLE);
3050
3051         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3052                 return -EOPNOTSUPP;
3053
3054         if (fc->no_fallocate)
3055                 return -EOPNOTSUPP;
3056
3057         if (lock_inode) {
3058                 inode_lock(inode);
3059                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3060                         loff_t endbyte = offset + length - 1;
3061
3062                         err = fuse_writeback_range(inode, offset, endbyte);
3063                         if (err)
3064                                 goto out;
3065                 }
3066         }
3067
3068         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3069             offset + length > i_size_read(inode)) {
3070                 err = inode_newsize_ok(inode, offset + length);
3071                 if (err)
3072                         goto out;
3073         }
3074
3075         if (!(mode & FALLOC_FL_KEEP_SIZE))
3076                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3077
3078         args.opcode = FUSE_FALLOCATE;
3079         args.nodeid = ff->nodeid;
3080         args.in_numargs = 1;
3081         args.in_args[0].size = sizeof(inarg);
3082         args.in_args[0].value = &inarg;
3083         err = fuse_simple_request(fc, &args);
3084         if (err == -ENOSYS) {
3085                 fc->no_fallocate = 1;
3086                 err = -EOPNOTSUPP;
3087         }
3088         if (err)
3089                 goto out;
3090
3091         /* we could have extended the file */
3092         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3093                 bool changed = fuse_write_update_size(inode, offset + length);
3094
3095                 if (changed && fc->writeback_cache)
3096                         file_update_time(file);
3097         }
3098
3099         if (mode & FALLOC_FL_PUNCH_HOLE)
3100                 truncate_pagecache_range(inode, offset, offset + length - 1);
3101
3102         fuse_invalidate_attr(inode);
3103
3104 out:
3105         if (!(mode & FALLOC_FL_KEEP_SIZE))
3106                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3107
3108         if (lock_inode)
3109                 inode_unlock(inode);
3110
3111         return err;
3112 }
3113
3114 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3115                                       struct file *file_out, loff_t pos_out,
3116                                       size_t len, unsigned int flags)
3117 {
3118         struct fuse_file *ff_in = file_in->private_data;
3119         struct fuse_file *ff_out = file_out->private_data;
3120         struct inode *inode_in = file_inode(file_in);
3121         struct inode *inode_out = file_inode(file_out);
3122         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3123         struct fuse_conn *fc = ff_in->fc;
3124         FUSE_ARGS(args);
3125         struct fuse_copy_file_range_in inarg = {
3126                 .fh_in = ff_in->fh,
3127                 .off_in = pos_in,
3128                 .nodeid_out = ff_out->nodeid,
3129                 .fh_out = ff_out->fh,
3130                 .off_out = pos_out,
3131                 .len = len,
3132                 .flags = flags
3133         };
3134         struct fuse_write_out outarg;
3135         ssize_t err;
3136         /* mark unstable when write-back is not used, and file_out gets
3137          * extended */
3138         bool is_unstable = (!fc->writeback_cache) &&
3139                            ((pos_out + len) > inode_out->i_size);
3140
3141         if (fc->no_copy_file_range)
3142                 return -EOPNOTSUPP;
3143
3144         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3145                 return -EXDEV;
3146
3147         if (fc->writeback_cache) {
3148                 inode_lock(inode_in);
3149                 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3150                 inode_unlock(inode_in);
3151                 if (err)
3152                         return err;
3153         }
3154
3155         inode_lock(inode_out);
3156
3157         err = file_modified(file_out);
3158         if (err)
3159                 goto out;
3160
3161         if (fc->writeback_cache) {
3162                 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3163                 if (err)
3164                         goto out;
3165         }
3166
3167         if (is_unstable)
3168                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3169
3170         args.opcode = FUSE_COPY_FILE_RANGE;
3171         args.nodeid = ff_in->nodeid;
3172         args.in_numargs = 1;
3173         args.in_args[0].size = sizeof(inarg);
3174         args.in_args[0].value = &inarg;
3175         args.out_numargs = 1;
3176         args.out_args[0].size = sizeof(outarg);
3177         args.out_args[0].value = &outarg;
3178         err = fuse_simple_request(fc, &args);
3179         if (err == -ENOSYS) {
3180                 fc->no_copy_file_range = 1;
3181                 err = -EOPNOTSUPP;
3182         }
3183         if (err)
3184                 goto out;
3185
3186         if (fc->writeback_cache) {
3187                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3188                 file_update_time(file_out);
3189         }
3190
3191         fuse_invalidate_attr(inode_out);
3192
3193         err = outarg.size;
3194 out:
3195         if (is_unstable)
3196                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3197
3198         inode_unlock(inode_out);
3199         file_accessed(file_in);
3200
3201         return err;
3202 }
3203
3204 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3205                                     struct file *dst_file, loff_t dst_off,
3206                                     size_t len, unsigned int flags)
3207 {
3208         ssize_t ret;
3209
3210         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3211                                      len, flags);
3212
3213         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3214                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3215                                               dst_off, len, flags);
3216         return ret;
3217 }
3218
3219 static const struct file_operations fuse_file_operations = {
3220         .llseek         = fuse_file_llseek,
3221         .read_iter      = fuse_file_read_iter,
3222         .write_iter     = fuse_file_write_iter,
3223         .mmap           = fuse_file_mmap,
3224         .open           = fuse_open,
3225         .flush          = fuse_flush,
3226         .release        = fuse_release,
3227         .fsync          = fuse_fsync,
3228         .lock           = fuse_file_lock,
3229         .flock          = fuse_file_flock,
3230         .splice_read    = generic_file_splice_read,
3231         .splice_write   = iter_file_splice_write,
3232         .unlocked_ioctl = fuse_file_ioctl,
3233         .compat_ioctl   = fuse_file_compat_ioctl,
3234         .poll           = fuse_file_poll,
3235         .fallocate      = fuse_file_fallocate,
3236         .copy_file_range = fuse_copy_file_range,
3237 };
3238
3239 static const struct address_space_operations fuse_file_aops  = {
3240         .readpage       = fuse_readpage,
3241         .writepage      = fuse_writepage,
3242         .writepages     = fuse_writepages,
3243         .launder_page   = fuse_launder_page,
3244         .readpages      = fuse_readpages,
3245         .set_page_dirty = __set_page_dirty_nobuffers,
3246         .bmap           = fuse_bmap,
3247         .direct_IO      = fuse_direct_IO,
3248         .write_begin    = fuse_write_begin,
3249         .write_end      = fuse_write_end,
3250 };
3251
3252 void fuse_init_file_inode(struct inode *inode)
3253 {
3254         struct fuse_inode *fi = get_fuse_inode(inode);
3255
3256         inode->i_fop = &fuse_file_operations;
3257         inode->i_data.a_ops = &fuse_file_aops;
3258
3259         INIT_LIST_HEAD(&fi->write_files);
3260         INIT_LIST_HEAD(&fi->queued_writes);
3261         fi->writectr = 0;
3262         init_waitqueue_head(&fi->page_waitq);
3263         INIT_LIST_HEAD(&fi->writepages);
3264 }