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/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20 #include <linux/fs.h>
  21
  22 static const struct file_operations fuse_direct_io_file_operations;
  23
  24 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  25                           int opcode, struct fuse_open_out *outargp)
  26 {
  27         struct fuse_open_in inarg;
  28         FUSE_ARGS(args);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34         args.in.h.opcode = opcode;
  35         args.in.h.nodeid = nodeid;
  36         args.in.numargs = 1;
  37         args.in.args[0].size = sizeof(inarg);
  38         args.in.args[0].value = &inarg;
  39         args.out.numargs = 1;
  40         args.out.args[0].size = sizeof(*outargp);
  41         args.out.args[0].value = outargp;
  42
  43         return fuse_simple_request(fc, &args);
  44 }
  45
  46 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  47 {
  48         struct fuse_file *ff;
  49
  50         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  51         if (unlikely(!ff))
  52                 return NULL;
  53
  54         ff->fc = fc;
  55         ff->reserved_req = fuse_request_alloc(0);
  56         if (unlikely(!ff->reserved_req)) {
  57                 kfree(ff);
  58                 return NULL;
  59         }
  60
  61         INIT_LIST_HEAD(&ff->write_entry);
  62         atomic_set(&ff->count, 0);
  63         RB_CLEAR_NODE(&ff->polled_node);
  64         init_waitqueue_head(&ff->poll_wait);
  65
  66         spin_lock(&fc->lock);
  67         ff->kh = ++fc->khctr;
  68         spin_unlock(&fc->lock);
  69
  70         return ff;
  71 }
  72
  73 void fuse_file_free(struct fuse_file *ff)
  74 {
  75         fuse_request_free(ff->reserved_req);
  76         kfree(ff);
  77 }
  78
  79 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  80 {
  81         atomic_inc(&ff->count);
  82         return ff;
  83 }
  84
  85 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  86 {
  87         iput(req->misc.release.inode);
  88 }
  89
  90 static void fuse_file_put(struct fuse_file *ff, bool sync)
  91 {
  92         if (atomic_dec_and_test(&ff->count)) {
  93                 struct fuse_req *req = ff->reserved_req;
  94
  95                 if (ff->fc->no_open) {
  96                         /*
  97                          * Drop the release request when client does not
  98                          * implement 'open'
  99                          */
 100                         __clear_bit(FR_BACKGROUND, &req->flags);
 101                         iput(req->misc.release.inode);
 102                         fuse_put_request(ff->fc, req);
 103                 } else if (sync) {
 104                         __set_bit(FR_FORCE, &req->flags);
 105                         __clear_bit(FR_BACKGROUND, &req->flags);
 106                         fuse_request_send(ff->fc, req);
 107                         iput(req->misc.release.inode);
 108                         fuse_put_request(ff->fc, req);
 109                 } else {
 110                         req->end = fuse_release_end;
 111                         __set_bit(FR_BACKGROUND, &req->flags);
 112                         fuse_request_send_background(ff->fc, req);
 113                 }
 114                 kfree(ff);
 115         }
 116 }
 117
 118 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 119                  bool isdir)
 120 {
 121         struct fuse_file *ff;
 122         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 123
 124         ff = fuse_file_alloc(fc);
 125         if (!ff)
 126                 return -ENOMEM;
 127
 128         ff->fh = 0;
 129         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 130         if (!fc->no_open || isdir) {
 131                 struct fuse_open_out outarg;
 132                 int err;
 133
 134                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 135                 if (!err) {
 136                         ff->fh = outarg.fh;
 137                         ff->open_flags = outarg.open_flags;
 138
 139                 } else if (err != -ENOSYS || isdir) {
 140                         fuse_file_free(ff);
 141                         return err;
 142                 } else {
 143                         fc->no_open = 1;
 144                 }
 145         }
 146
 147         if (isdir)
 148                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 149
 150         ff->nodeid = nodeid;
 151         file->private_data = fuse_file_get(ff);
 152
 153         return 0;
 154 }
 155 EXPORT_SYMBOL_GPL(fuse_do_open);
 156
 157 static void fuse_link_write_file(struct file *file)
 158 {
 159         struct inode *inode = file_inode(file);
 160         struct fuse_conn *fc = get_fuse_conn(inode);
 161         struct fuse_inode *fi = get_fuse_inode(inode);
 162         struct fuse_file *ff = file->private_data;
 163         /*
 164          * file may be written through mmap, so chain it onto the
 165          * inodes's write_file list
 166          */
 167         spin_lock(&fc->lock);
 168         if (list_empty(&ff->write_entry))
 169                 list_add(&ff->write_entry, &fi->write_files);
 170         spin_unlock(&fc->lock);
 171 }
 172
 173 void fuse_finish_open(struct inode *inode, struct file *file)
 174 {
 175         struct fuse_file *ff = file->private_data;
 176         struct fuse_conn *fc = get_fuse_conn(inode);
 177
 178         if (ff->open_flags & FOPEN_DIRECT_IO)
 179                 file->f_op = &fuse_direct_io_file_operations;
 180         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 181                 invalidate_inode_pages2(inode->i_mapping);
 182         if (ff->open_flags & FOPEN_STREAM)
 183                 stream_open(inode, file);
 184         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 185                 nonseekable_open(inode, file);
 186         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 187                 struct fuse_inode *fi = get_fuse_inode(inode);
 188
 189                 spin_lock(&fc->lock);
 190                 fi->attr_version = ++fc->attr_version;
 191                 i_size_write(inode, 0);
 192                 spin_unlock(&fc->lock);
 193                 fuse_invalidate_attr(inode);
 194                 if (fc->writeback_cache)
 195                         file_update_time(file);
 196         }
 197         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 198                 fuse_link_write_file(file);
 199 }
 200
 201 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 202 {
 203         struct fuse_conn *fc = get_fuse_conn(inode);
 204         int err;
 205         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 206                           fc->atomic_o_trunc &&
 207                           fc->writeback_cache;
 208
 209         if (fuse_is_bad(inode))
 210                 return -EIO;
 211
 212         err = generic_file_open(inode, file);
 213         if (err)
 214                 return err;
 215
 216         if (is_wb_truncate) {
 217                 inode_lock(inode);
 218                 fuse_set_nowrite(inode);
 219         }
 220
 221         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 222
 223         if (!err)
 224                 fuse_finish_open(inode, file);
 225
 226         if (is_wb_truncate) {
 227                 fuse_release_nowrite(inode);
 228                 inode_unlock(inode);
 229         }
 230
 231         return err;
 232 }
 233
 234 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 235 {
 236         struct fuse_conn *fc = ff->fc;
 237         struct fuse_req *req = ff->reserved_req;
 238         struct fuse_release_in *inarg = &req->misc.release.in;
 239
 240         spin_lock(&fc->lock);
 241         list_del(&ff->write_entry);
 242         if (!RB_EMPTY_NODE(&ff->polled_node))
 243                 rb_erase(&ff->polled_node, &fc->polled_files);
 244         spin_unlock(&fc->lock);
 245
 246         wake_up_interruptible_all(&ff->poll_wait);
 247
 248         inarg->fh = ff->fh;
 249         inarg->flags = flags;
 250         req->in.h.opcode = opcode;
 251         req->in.h.nodeid = ff->nodeid;
 252         req->in.numargs = 1;
 253         req->in.args[0].size = sizeof(struct fuse_release_in);
 254         req->in.args[0].value = inarg;
 255 }
 256
 257 void fuse_release_common(struct file *file, int opcode)
 258 {
 259         struct fuse_file *ff;
 260         struct fuse_req *req;
 261
 262         ff = file->private_data;
 263         if (unlikely(!ff))
 264                 return;
 265
 266         req = ff->reserved_req;
 267         fuse_prepare_release(ff, file->f_flags, opcode);
 268
 269         if (ff->flock) {
 270                 struct fuse_release_in *inarg = &req->misc.release.in;
 271                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 272                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 273                                                        (fl_owner_t) file);
 274         }
 275         /* Hold inode until release is finished */
 276         req->misc.release.inode = igrab(file_inode(file));
 277
 278         /*
 279          * Normally this will send the RELEASE request, however if
 280          * some asynchronous READ or WRITE requests are outstanding,
 281          * the sending will be delayed.
 282          *
 283          * Make the release synchronous if this is a fuseblk mount,
 284          * synchronous RELEASE is allowed (and desirable) in this case
 285          * because the server can be trusted not to screw up.
 286          */
 287         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 288 }
 289
 290 static int fuse_open(struct inode *inode, struct file *file)
 291 {
 292         return fuse_open_common(inode, file, false);
 293 }
 294
 295 static int fuse_release(struct inode *inode, struct file *file)
 296 {
 297         struct fuse_conn *fc = get_fuse_conn(inode);
 298
 299         /* see fuse_vma_close() for !writeback_cache case */
 300         if (fc->writeback_cache)
 301                 write_inode_now(inode, 1);
 302
 303         fuse_release_common(file, FUSE_RELEASE);
 304
 305         /* return value is ignored by VFS */
 306         return 0;
 307 }
 308
 309 void fuse_sync_release(struct fuse_file *ff, int flags)
 310 {
 311         WARN_ON(atomic_read(&ff->count) > 1);
 312         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 313         __set_bit(FR_FORCE, &ff->reserved_req->flags);
 314         __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
 315         fuse_request_send(ff->fc, ff->reserved_req);
 316         fuse_put_request(ff->fc, ff->reserved_req);
 317         kfree(ff);
 318 }
 319 EXPORT_SYMBOL_GPL(fuse_sync_release);
 320
 321 /*
 322  * Scramble the ID space with XTEA, so that the value of the files_struct
 323  * pointer is not exposed to userspace.
 324  */
 325 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 326 {
 327         u32 *k = fc->scramble_key;
 328         u64 v = (unsigned long) id;
 329         u32 v0 = v;
 330         u32 v1 = v >> 32;
 331         u32 sum = 0;
 332         int i;
 333
 334         for (i = 0; i < 32; i++) {
 335                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 336                 sum += 0x9E3779B9;
 337                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 338         }
 339
 340         return (u64) v0 + ((u64) v1 << 32);
 341 }
 342
 343 /*
 344  * Check if any page in a range is under writeback
 345  *
 346  * This is currently done by walking the list of writepage requests
 347  * for the inode, which can be pretty inefficient.
 348  */
 349 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 350                                    pgoff_t idx_to)
 351 {
 352         struct fuse_conn *fc = get_fuse_conn(inode);
 353         struct fuse_inode *fi = get_fuse_inode(inode);
 354         struct fuse_req *req;
 355         bool found = false;
 356
 357         spin_lock(&fc->lock);
 358         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 359                 pgoff_t curr_index;
 360
 361                 BUG_ON(req->inode != inode);
 362                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 363                 if (idx_from < curr_index + req->num_pages &&
 364                     curr_index <= idx_to) {
 365                         found = true;
 366                         break;
 367                 }
 368         }
 369         spin_unlock(&fc->lock);
 370
 371         return found;
 372 }
 373
 374 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 375 {
 376         return fuse_range_is_writeback(inode, index, index);
 377 }
 378
 379 /*
 380  * Wait for page writeback to be completed.
 381  *
 382  * Since fuse doesn't rely on the VM writeback tracking, this has to
 383  * use some other means.
 384  */
 385 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 386 {
 387         struct fuse_inode *fi = get_fuse_inode(inode);
 388
 389         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 390         return 0;
 391 }
 392
 393 /*
 394  * Wait for all pending writepages on the inode to finish.
 395  *
 396  * This is currently done by blocking further writes with FUSE_NOWRITE
 397  * and waiting for all sent writes to complete.
 398  *
 399  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 400  * could conflict with truncation.
 401  */
 402 static void fuse_sync_writes(struct inode *inode)
 403 {
 404         fuse_set_nowrite(inode);
 405         fuse_release_nowrite(inode);
 406 }
 407
 408 static int fuse_flush(struct file *file, fl_owner_t id)
 409 {
 410         struct inode *inode = file_inode(file);
 411         struct fuse_conn *fc = get_fuse_conn(inode);
 412         struct fuse_file *ff = file->private_data;
 413         struct fuse_req *req;
 414         struct fuse_flush_in inarg;
 415         int err;
 416
 417         if (fuse_is_bad(inode))
 418                 return -EIO;
 419
 420         if (fc->no_flush)
 421                 return 0;
 422
 423         err = write_inode_now(inode, 1);
 424         if (err)
 425                 return err;
 426
 427         inode_lock(inode);
 428         fuse_sync_writes(inode);
 429         inode_unlock(inode);
 430
 431         err = filemap_check_errors(file->f_mapping);
 432         if (err)
 433                 return err;
 434
 435         req = fuse_get_req_nofail_nopages(fc, file);
 436         memset(&inarg, 0, sizeof(inarg));
 437         inarg.fh = ff->fh;
 438         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 439         req->in.h.opcode = FUSE_FLUSH;
 440         req->in.h.nodeid = get_node_id(inode);
 441         req->in.numargs = 1;
 442         req->in.args[0].size = sizeof(inarg);
 443         req->in.args[0].value = &inarg;
 444         __set_bit(FR_FORCE, &req->flags);
 445         fuse_request_send(fc, req);
 446         err = req->out.h.error;
 447         fuse_put_request(fc, req);
 448         if (err == -ENOSYS) {
 449                 fc->no_flush = 1;
 450                 err = 0;
 451         }
 452         return err;
 453 }
 454
 455 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 456                       int datasync, int isdir)
 457 {
 458         struct inode *inode = file->f_mapping->host;
 459         struct fuse_conn *fc = get_fuse_conn(inode);
 460         struct fuse_file *ff = file->private_data;
 461         FUSE_ARGS(args);
 462         struct fuse_fsync_in inarg;
 463         int err;
 464
 465         if (fuse_is_bad(inode))
 466                 return -EIO;
 467
 468         inode_lock(inode);
 469
 470         /*
 471          * Start writeback against all dirty pages of the inode, then
 472          * wait for all outstanding writes, before sending the FSYNC
 473          * request.
 474          */
 475         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 476         if (err)
 477                 goto out;
 478
 479         fuse_sync_writes(inode);
 480
 481         /*
 482          * Due to implementation of fuse writeback
 483          * filemap_write_and_wait_range() does not catch errors.
 484          * We have to do this directly after fuse_sync_writes()
 485          */
 486         err = filemap_check_errors(file->f_mapping);
 487         if (err)
 488                 goto out;
 489
 490         err = sync_inode_metadata(inode, 1);
 491         if (err)
 492                 goto out;
 493
 494         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 495                 goto out;
 496
 497         memset(&inarg, 0, sizeof(inarg));
 498         inarg.fh = ff->fh;
 499         inarg.fsync_flags = datasync ? 1 : 0;
 500         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 501         args.in.h.nodeid = get_node_id(inode);
 502         args.in.numargs = 1;
 503         args.in.args[0].size = sizeof(inarg);
 504         args.in.args[0].value = &inarg;
 505         err = fuse_simple_request(fc, &args);
 506         if (err == -ENOSYS) {
 507                 if (isdir)
 508                         fc->no_fsyncdir = 1;
 509                 else
 510                         fc->no_fsync = 1;
 511                 err = 0;
 512         }
 513 out:
 514         inode_unlock(inode);
 515         return err;
 516 }
 517
 518 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 519                       int datasync)
 520 {
 521         return fuse_fsync_common(file, start, end, datasync, 0);
 522 }
 523
 524 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 525                     size_t count, int opcode)
 526 {
 527         struct fuse_read_in *inarg = &req->misc.read.in;
 528         struct fuse_file *ff = file->private_data;
 529
 530         inarg->fh = ff->fh;
 531         inarg->offset = pos;
 532         inarg->size = count;
 533         inarg->flags = file->f_flags;
 534         req->in.h.opcode = opcode;
 535         req->in.h.nodeid = ff->nodeid;
 536         req->in.numargs = 1;
 537         req->in.args[0].size = sizeof(struct fuse_read_in);
 538         req->in.args[0].value = inarg;
 539         req->out.argvar = 1;
 540         req->out.numargs = 1;
 541         req->out.args[0].size = count;
 542 }
 543
 544 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 545 {
 546         unsigned i;
 547
 548         for (i = 0; i < req->num_pages; i++) {
 549                 struct page *page = req->pages[i];
 550                 if (should_dirty)
 551                         set_page_dirty_lock(page);
 552                 put_page(page);
 553         }
 554 }
 555
 556 static void fuse_io_release(struct kref *kref)
 557 {
 558         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 559 }
 560
 561 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 562 {
 563         if (io->err)
 564                 return io->err;
 565
 566         if (io->bytes >= 0 && io->write)
 567                 return -EIO;
 568
 569         return io->bytes < 0 ? io->size : io->bytes;
 570 }
 571
 572 /**
 573  * In case of short read, the caller sets 'pos' to the position of
 574  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 575  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 576  *
 577  * An example:
 578  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 579  * both submitted asynchronously. The first of them was ACKed by userspace as
 580  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 581  * second request was ACKed as short, e.g. only 1K was read, resulting in
 582  * pos == 33K.
 583  *
 584  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 585  * will be equal to the length of the longest contiguous fragment of
 586  * transferred data starting from the beginning of IO request.
 587  */
 588 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 589 {
 590         int left;
 591
 592         spin_lock(&io->lock);
 593         if (err)
 594                 io->err = io->err ? : err;
 595         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 596                 io->bytes = pos;
 597
 598         left = --io->reqs;
 599         if (!left && io->blocking)
 600                 complete(io->done);
 601         spin_unlock(&io->lock);
 602
 603         if (!left && !io->blocking) {
 604                 ssize_t res = fuse_get_res_by_io(io);
 605
 606                 if (res >= 0) {
 607                         struct inode *inode = file_inode(io->iocb->ki_filp);
 608                         struct fuse_conn *fc = get_fuse_conn(inode);
 609                         struct fuse_inode *fi = get_fuse_inode(inode);
 610
 611                         spin_lock(&fc->lock);
 612                         fi->attr_version = ++fc->attr_version;
 613                         spin_unlock(&fc->lock);
 614                 }
 615
 616                 io->iocb->ki_complete(io->iocb, res, 0);
 617         }
 618
 619         kref_put(&io->refcnt, fuse_io_release);
 620 }
 621
 622 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 623 {
 624         struct fuse_io_priv *io = req->io;
 625         ssize_t pos = -1;
 626
 627         fuse_release_user_pages(req, !io->write);
 628
 629         if (io->write) {
 630                 if (req->misc.write.in.size != req->misc.write.out.size)
 631                         pos = req->misc.write.in.offset - io->offset +
 632                                 req->misc.write.out.size;
 633         } else {
 634                 if (req->misc.read.in.size != req->out.args[0].size)
 635                         pos = req->misc.read.in.offset - io->offset +
 636                                 req->out.args[0].size;
 637         }
 638
 639         fuse_aio_complete(io, req->out.h.error, pos);
 640 }
 641
 642 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 643                 size_t num_bytes, struct fuse_io_priv *io)
 644 {
 645         spin_lock(&io->lock);
 646         kref_get(&io->refcnt);
 647         io->size += num_bytes;
 648         io->reqs++;
 649         spin_unlock(&io->lock);
 650
 651         req->io = io;
 652         req->end = fuse_aio_complete_req;
 653
 654         __fuse_get_request(req);
 655         fuse_request_send_background(fc, req);
 656
 657         return num_bytes;
 658 }
 659
 660 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 661                              loff_t pos, size_t count, fl_owner_t owner)
 662 {
 663         struct file *file = io->file;
 664         struct fuse_file *ff = file->private_data;
 665         struct fuse_conn *fc = ff->fc;
 666
 667         fuse_read_fill(req, file, pos, count, FUSE_READ);
 668         if (owner != NULL) {
 669                 struct fuse_read_in *inarg = &req->misc.read.in;
 670
 671                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 672                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 673         }
 674
 675         if (io->async)
 676                 return fuse_async_req_send(fc, req, count, io);
 677
 678         fuse_request_send(fc, req);
 679         return req->out.args[0].size;
 680 }
 681
 682 static void fuse_read_update_size(struct inode *inode, loff_t size,
 683                                   u64 attr_ver)
 684 {
 685         struct fuse_conn *fc = get_fuse_conn(inode);
 686         struct fuse_inode *fi = get_fuse_inode(inode);
 687
 688         spin_lock(&fc->lock);
 689         if (attr_ver == fi->attr_version && size < inode->i_size &&
 690             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 691                 fi->attr_version = ++fc->attr_version;
 692                 i_size_write(inode, size);
 693         }
 694         spin_unlock(&fc->lock);
 695 }
 696
 697 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 698                             u64 attr_ver)
 699 {
 700         size_t num_read = req->out.args[0].size;
 701         struct fuse_conn *fc = get_fuse_conn(inode);
 702
 703         if (fc->writeback_cache) {
 704                 /*
 705                  * A hole in a file. Some data after the hole are in page cache,
 706                  * but have not reached the client fs yet. So, the hole is not
 707                  * present there.
 708                  */
 709                 int i;
 710                 int start_idx = num_read >> PAGE_SHIFT;
 711                 size_t off = num_read & (PAGE_SIZE - 1);
 712
 713                 for (i = start_idx; i < req->num_pages; i++) {
 714                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 715                         off = 0;
 716                 }
 717         } else {
 718                 loff_t pos = page_offset(req->pages[0]) + num_read;
 719                 fuse_read_update_size(inode, pos, attr_ver);
 720         }
 721 }
 722
 723 static int fuse_do_readpage(struct file *file, struct page *page)
 724 {
 725         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
 726         struct inode *inode = page->mapping->host;
 727         struct fuse_conn *fc = get_fuse_conn(inode);
 728         struct fuse_req *req;
 729         size_t num_read;
 730         loff_t pos = page_offset(page);
 731         size_t count = PAGE_SIZE;
 732         u64 attr_ver;
 733         int err;
 734
 735         /*
 736          * Page writeback can extend beyond the lifetime of the
 737          * page-cache page, so make sure we read a properly synced
 738          * page.
 739          */
 740         fuse_wait_on_page_writeback(inode, page->index);
 741
 742         req = fuse_get_req(fc, 1);
 743         if (IS_ERR(req))
 744                 return PTR_ERR(req);
 745
 746         attr_ver = fuse_get_attr_version(fc);
 747
 748         req->out.page_zeroing = 1;
 749         req->out.argpages = 1;
 750         req->num_pages = 1;
 751         req->pages[0] = page;
 752         req->page_descs[0].length = count;
 753         num_read = fuse_send_read(req, &io, pos, count, NULL);
 754         err = req->out.h.error;
 755
 756         if (!err) {
 757                 /*
 758                  * Short read means EOF.  If file size is larger, truncate it
 759                  */
 760                 if (num_read < count)
 761                         fuse_short_read(req, inode, attr_ver);
 762
 763                 SetPageUptodate(page);
 764         }
 765
 766         fuse_put_request(fc, req);
 767
 768         return err;
 769 }
 770
 771 static int fuse_readpage(struct file *file, struct page *page)
 772 {
 773         struct inode *inode = page->mapping->host;
 774         int err;
 775
 776         err = -EIO;
 777         if (fuse_is_bad(inode))
 778                 goto out;
 779
 780         err = fuse_do_readpage(file, page);
 781         fuse_invalidate_atime(inode);
 782  out:
 783         unlock_page(page);
 784         return err;
 785 }
 786
 787 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 788 {
 789         int i;
 790         size_t count = req->misc.read.in.size;
 791         size_t num_read = req->out.args[0].size;
 792         struct address_space *mapping = NULL;
 793
 794         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 795                 mapping = req->pages[i]->mapping;
 796
 797         if (mapping) {
 798                 struct inode *inode = mapping->host;
 799
 800                 /*
 801                  * Short read means EOF. If file size is larger, truncate it
 802                  */
 803                 if (!req->out.h.error && num_read < count)
 804                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 805
 806                 fuse_invalidate_atime(inode);
 807         }
 808
 809         for (i = 0; i < req->num_pages; i++) {
 810                 struct page *page = req->pages[i];
 811                 if (!req->out.h.error)
 812                         SetPageUptodate(page);
 813                 else
 814                         SetPageError(page);
 815                 unlock_page(page);
 816                 put_page(page);
 817         }
 818         if (req->ff)
 819                 fuse_file_put(req->ff, false);
 820 }
 821
 822 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 823 {
 824         struct fuse_file *ff = file->private_data;
 825         struct fuse_conn *fc = ff->fc;
 826         loff_t pos = page_offset(req->pages[0]);
 827         size_t count = req->num_pages << PAGE_SHIFT;
 828
 829         req->out.argpages = 1;
 830         req->out.page_zeroing = 1;
 831         req->out.page_replace = 1;
 832         fuse_read_fill(req, file, pos, count, FUSE_READ);
 833         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 834         if (fc->async_read) {
 835                 req->ff = fuse_file_get(ff);
 836                 req->end = fuse_readpages_end;
 837                 fuse_request_send_background(fc, req);
 838         } else {
 839                 fuse_request_send(fc, req);
 840                 fuse_readpages_end(fc, req);
 841                 fuse_put_request(fc, req);
 842         }
 843 }
 844
 845 struct fuse_fill_data {
 846         struct fuse_req *req;
 847         struct file *file;
 848         struct inode *inode;
 849         unsigned nr_pages;
 850 };
 851
 852 static int fuse_readpages_fill(void *_data, struct page *page)
 853 {
 854         struct fuse_fill_data *data = _data;
 855         struct fuse_req *req = data->req;
 856         struct inode *inode = data->inode;
 857         struct fuse_conn *fc = get_fuse_conn(inode);
 858
 859         fuse_wait_on_page_writeback(inode, page->index);
 860
 861         if (req->num_pages &&
 862             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 863              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 864              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 865                 int nr_alloc = min_t(unsigned, data->nr_pages,
 866                                      FUSE_MAX_PAGES_PER_REQ);
 867                 fuse_send_readpages(req, data->file);
 868                 if (fc->async_read)
 869                         req = fuse_get_req_for_background(fc, nr_alloc);
 870                 else
 871                         req = fuse_get_req(fc, nr_alloc);
 872
 873                 data->req = req;
 874                 if (IS_ERR(req)) {
 875                         unlock_page(page);
 876                         return PTR_ERR(req);
 877                 }
 878         }
 879
 880         if (WARN_ON(req->num_pages >= req->max_pages)) {
 881                 unlock_page(page);
 882                 fuse_put_request(fc, req);
 883                 return -EIO;
 884         }
 885
 886         get_page(page);
 887         req->pages[req->num_pages] = page;
 888         req->page_descs[req->num_pages].length = PAGE_SIZE;
 889         req->num_pages++;
 890         data->nr_pages--;
 891         return 0;
 892 }
 893
 894 static int fuse_readpages(struct file *file, struct address_space *mapping,
 895                           struct list_head *pages, unsigned nr_pages)
 896 {
 897         struct inode *inode = mapping->host;
 898         struct fuse_conn *fc = get_fuse_conn(inode);
 899         struct fuse_fill_data data;
 900         int err;
 901         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 902
 903         err = -EIO;
 904         if (fuse_is_bad(inode))
 905                 goto out;
 906
 907         data.file = file;
 908         data.inode = inode;
 909         if (fc->async_read)
 910                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 911         else
 912                 data.req = fuse_get_req(fc, nr_alloc);
 913         data.nr_pages = nr_pages;
 914         err = PTR_ERR(data.req);
 915         if (IS_ERR(data.req))
 916                 goto out;
 917
 918         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 919         if (!err) {
 920                 if (data.req->num_pages)
 921                         fuse_send_readpages(data.req, file);
 922                 else
 923                         fuse_put_request(fc, data.req);
 924         }
 925 out:
 926         return err;
 927 }
 928
 929 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 930 {
 931         struct inode *inode = iocb->ki_filp->f_mapping->host;
 932         struct fuse_conn *fc = get_fuse_conn(inode);
 933
 934         if (fuse_is_bad(inode))
 935                 return -EIO;
 936
 937         /*
 938          * In auto invalidate mode, always update attributes on read.
 939          * Otherwise, only update if we attempt to read past EOF (to ensure
 940          * i_size is up to date).
 941          */
 942         if (fc->auto_inval_data ||
 943             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 944                 int err;
 945                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 946                 if (err)
 947                         return err;
 948         }
 949
 950         return generic_file_read_iter(iocb, to);
 951 }
 952
 953 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 954                             loff_t pos, size_t count)
 955 {
 956         struct fuse_write_in *inarg = &req->misc.write.in;
 957         struct fuse_write_out *outarg = &req->misc.write.out;
 958
 959         inarg->fh = ff->fh;
 960         inarg->offset = pos;
 961         inarg->size = count;
 962         req->in.h.opcode = FUSE_WRITE;
 963         req->in.h.nodeid = ff->nodeid;
 964         req->in.numargs = 2;
 965         if (ff->fc->minor < 9)
 966                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 967         else
 968                 req->in.args[0].size = sizeof(struct fuse_write_in);
 969         req->in.args[0].value = inarg;
 970         req->in.args[1].size = count;
 971         req->out.numargs = 1;
 972         req->out.args[0].size = sizeof(struct fuse_write_out);
 973         req->out.args[0].value = outarg;
 974 }
 975
 976 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 977                               loff_t pos, size_t count, fl_owner_t owner)
 978 {
 979         struct file *file = io->file;
 980         struct fuse_file *ff = file->private_data;
 981         struct fuse_conn *fc = ff->fc;
 982         struct fuse_write_in *inarg = &req->misc.write.in;
 983
 984         fuse_write_fill(req, ff, pos, count);
 985         inarg->flags = file->f_flags;
 986         if (owner != NULL) {
 987                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 988                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 989         }
 990
 991         if (io->async)
 992                 return fuse_async_req_send(fc, req, count, io);
 993
 994         fuse_request_send(fc, req);
 995         return req->misc.write.out.size;
 996 }
 997
 998 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 999 {
1000         struct fuse_conn *fc = get_fuse_conn(inode);
1001         struct fuse_inode *fi = get_fuse_inode(inode);
1002         bool ret = false;
1003
1004         spin_lock(&fc->lock);
1005         fi->attr_version = ++fc->attr_version;
1006         if (pos > inode->i_size) {
1007                 i_size_write(inode, pos);
1008                 ret = true;
1009         }
1010         spin_unlock(&fc->lock);
1011
1012         return ret;
1013 }
1014
1015 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1016                                     struct inode *inode, loff_t pos,
1017                                     size_t count)
1018 {
1019         size_t res;
1020         unsigned offset;
1021         unsigned i;
1022         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1023
1024         for (i = 0; i < req->num_pages; i++)
1025                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1026
1027         res = fuse_send_write(req, &io, pos, count, NULL);
1028
1029         offset = req->page_descs[0].offset;
1030         count = res;
1031         for (i = 0; i < req->num_pages; i++) {
1032                 struct page *page = req->pages[i];
1033
1034                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1035                         SetPageUptodate(page);
1036
1037                 if (count > PAGE_SIZE - offset)
1038                         count -= PAGE_SIZE - offset;
1039                 else
1040                         count = 0;
1041                 offset = 0;
1042
1043                 unlock_page(page);
1044                 put_page(page);
1045         }
1046
1047         return res;
1048 }
1049
1050 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1051                                struct address_space *mapping,
1052                                struct iov_iter *ii, loff_t pos)
1053 {
1054         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1055         unsigned offset = pos & (PAGE_SIZE - 1);
1056         size_t count = 0;
1057         int err;
1058
1059         req->in.argpages = 1;
1060         req->page_descs[0].offset = offset;
1061
1062         do {
1063                 size_t tmp;
1064                 struct page *page;
1065                 pgoff_t index = pos >> PAGE_SHIFT;
1066                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1067                                      iov_iter_count(ii));
1068
1069                 bytes = min_t(size_t, bytes, fc->max_write - count);
1070
1071  again:
1072                 err = -EFAULT;
1073                 if (iov_iter_fault_in_readable(ii, bytes))
1074                         break;
1075
1076                 err = -ENOMEM;
1077                 page = grab_cache_page_write_begin(mapping, index, 0);
1078                 if (!page)
1079                         break;
1080
1081                 if (mapping_writably_mapped(mapping))
1082                         flush_dcache_page(page);
1083
1084                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1085                 flush_dcache_page(page);
1086
1087                 iov_iter_advance(ii, tmp);
1088                 if (!tmp) {
1089                         unlock_page(page);
1090                         put_page(page);
1091                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1092                         goto again;
1093                 }
1094
1095                 err = 0;
1096                 req->pages[req->num_pages] = page;
1097                 req->page_descs[req->num_pages].length = tmp;
1098                 req->num_pages++;
1099
1100                 count += tmp;
1101                 pos += tmp;
1102                 offset += tmp;
1103                 if (offset == PAGE_SIZE)
1104                         offset = 0;
1105
1106                 if (!fc->big_writes)
1107                         break;
1108         } while (iov_iter_count(ii) && count < fc->max_write &&
1109                  req->num_pages < req->max_pages && offset == 0);
1110
1111         return count > 0 ? count : err;
1112 }
1113
1114 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1115 {
1116         return min_t(unsigned,
1117                      ((pos + len - 1) >> PAGE_SHIFT) -
1118                      (pos >> PAGE_SHIFT) + 1,
1119                      FUSE_MAX_PAGES_PER_REQ);
1120 }
1121
1122 static ssize_t fuse_perform_write(struct file *file,
1123                                   struct address_space *mapping,
1124                                   struct iov_iter *ii, loff_t pos)
1125 {
1126         struct inode *inode = mapping->host;
1127         struct fuse_conn *fc = get_fuse_conn(inode);
1128         struct fuse_inode *fi = get_fuse_inode(inode);
1129         int err = 0;
1130         ssize_t res = 0;
1131
1132         if (fuse_is_bad(inode))
1133                 return -EIO;
1134
1135         if (inode->i_size < pos + iov_iter_count(ii))
1136                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1137
1138         do {
1139                 struct fuse_req *req;
1140                 ssize_t count;
1141                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1142
1143                 req = fuse_get_req(fc, nr_pages);
1144                 if (IS_ERR(req)) {
1145                         err = PTR_ERR(req);
1146                         break;
1147                 }
1148
1149                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1150                 if (count <= 0) {
1151                         err = count;
1152                 } else {
1153                         size_t num_written;
1154
1155                         num_written = fuse_send_write_pages(req, file, inode,
1156                                                             pos, count);
1157                         err = req->out.h.error;
1158                         if (!err) {
1159                                 res += num_written;
1160                                 pos += num_written;
1161
1162                                 /* break out of the loop on short write */
1163                                 if (num_written != count)
1164                                         err = -EIO;
1165                         }
1166                 }
1167                 fuse_put_request(fc, req);
1168         } while (!err && iov_iter_count(ii));
1169
1170         if (res > 0)
1171                 fuse_write_update_size(inode, pos);
1172
1173         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1174         fuse_invalidate_attr(inode);
1175
1176         return res > 0 ? res : err;
1177 }
1178
1179 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1180 {
1181         struct file *file = iocb->ki_filp;
1182         struct address_space *mapping = file->f_mapping;
1183         ssize_t written = 0;
1184         ssize_t written_buffered = 0;
1185         struct inode *inode = mapping->host;
1186         ssize_t err;
1187         loff_t endbyte = 0;
1188
1189         if (fuse_is_bad(inode))
1190                 return -EIO;
1191
1192         if (get_fuse_conn(inode)->writeback_cache) {
1193                 /* Update size (EOF optimization) and mode (SUID clearing) */
1194                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1195                 if (err)
1196                         return err;
1197
1198                 return generic_file_write_iter(iocb, from);
1199         }
1200
1201         inode_lock(inode);
1202
1203         /* We can write back this queue in page reclaim */
1204         current->backing_dev_info = inode_to_bdi(inode);
1205
1206         err = generic_write_checks(iocb, from);
1207         if (err <= 0)
1208                 goto out;
1209
1210         err = file_remove_privs(file);
1211         if (err)
1212                 goto out;
1213
1214         err = file_update_time(file);
1215         if (err)
1216                 goto out;
1217
1218         if (iocb->ki_flags & IOCB_DIRECT) {
1219                 loff_t pos = iocb->ki_pos;
1220                 written = generic_file_direct_write(iocb, from);
1221                 if (written < 0 || !iov_iter_count(from))
1222                         goto out;
1223
1224                 pos += written;
1225
1226                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1227                 if (written_buffered < 0) {
1228                         err = written_buffered;
1229                         goto out;
1230                 }
1231                 endbyte = pos + written_buffered - 1;
1232
1233                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1234                                                    endbyte);
1235                 if (err)
1236                         goto out;
1237
1238                 invalidate_mapping_pages(file->f_mapping,
1239                                          pos >> PAGE_SHIFT,
1240                                          endbyte >> PAGE_SHIFT);
1241
1242                 written += written_buffered;
1243                 iocb->ki_pos = pos + written_buffered;
1244         } else {
1245                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1246                 if (written >= 0)
1247                         iocb->ki_pos += written;
1248         }
1249 out:
1250         current->backing_dev_info = NULL;
1251         inode_unlock(inode);
1252
1253         return written ? written : err;
1254 }
1255
1256 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1257                 unsigned index, unsigned nr_pages)
1258 {
1259         int i;
1260
1261         for (i = index; i < index + nr_pages; i++)
1262                 req->page_descs[i].length = PAGE_SIZE -
1263                         req->page_descs[i].offset;
1264 }
1265
1266 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1267 {
1268         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1269 }
1270
1271 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1272                                         size_t max_size)
1273 {
1274         return min(iov_iter_single_seg_count(ii), max_size);
1275 }
1276
1277 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1278                                size_t *nbytesp, int write)
1279 {
1280         size_t nbytes = 0;  /* # bytes already packed in req */
1281         ssize_t ret = 0;
1282
1283         /* Special case for kernel I/O: can copy directly into the buffer */
1284         if (ii->type & ITER_KVEC) {
1285                 unsigned long user_addr = fuse_get_user_addr(ii);
1286                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1287
1288                 if (write)
1289                         req->in.args[1].value = (void *) user_addr;
1290                 else
1291                         req->out.args[0].value = (void *) user_addr;
1292
1293                 iov_iter_advance(ii, frag_size);
1294                 *nbytesp = frag_size;
1295                 return 0;
1296         }
1297
1298         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1299                 unsigned npages;
1300                 size_t start;
1301                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1302                                         *nbytesp - nbytes,
1303                                         req->max_pages - req->num_pages,
1304                                         &start);
1305                 if (ret < 0)
1306                         break;
1307
1308                 iov_iter_advance(ii, ret);
1309                 nbytes += ret;
1310
1311                 ret += start;
1312                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1313
1314                 req->page_descs[req->num_pages].offset = start;
1315                 fuse_page_descs_length_init(req, req->num_pages, npages);
1316
1317                 req->num_pages += npages;
1318                 req->page_descs[req->num_pages - 1].length -=
1319                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1320         }
1321
1322         req->user_pages = true;
1323         if (write)
1324                 req->in.argpages = 1;
1325         else
1326                 req->out.argpages = 1;
1327
1328         *nbytesp = nbytes;
1329
1330         return ret < 0 ? ret : 0;
1331 }
1332
1333 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1334 {
1335         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1336 }
1337
1338 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1339                        loff_t *ppos, int flags)
1340 {
1341         int write = flags & FUSE_DIO_WRITE;
1342         bool should_dirty = !write && iter_is_iovec(iter);
1343         int cuse = flags & FUSE_DIO_CUSE;
1344         struct file *file = io->file;
1345         struct inode *inode = file->f_mapping->host;
1346         struct fuse_file *ff = file->private_data;
1347         struct fuse_conn *fc = ff->fc;
1348         size_t nmax = write ? fc->max_write : fc->max_read;
1349         loff_t pos = *ppos;
1350         size_t count = iov_iter_count(iter);
1351         pgoff_t idx_from = pos >> PAGE_SHIFT;
1352         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1353         ssize_t res = 0;
1354         struct fuse_req *req;
1355         int err = 0;
1356
1357         if (io->async)
1358                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1359         else
1360                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1361         if (IS_ERR(req))
1362                 return PTR_ERR(req);
1363
1364         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1365                 if (!write)
1366                         inode_lock(inode);
1367                 fuse_sync_writes(inode);
1368                 if (!write)
1369                         inode_unlock(inode);
1370         }
1371
1372         while (count) {
1373                 size_t nres;
1374                 fl_owner_t owner = current->files;
1375                 size_t nbytes = min(count, nmax);
1376                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1377                 if (err && !nbytes)
1378                         break;
1379
1380                 if (write)
1381                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1382                 else
1383                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1384
1385                 if (!io->async)
1386                         fuse_release_user_pages(req, should_dirty);
1387                 if (req->out.h.error) {
1388                         err = req->out.h.error;
1389                         break;
1390                 } else if (nres > nbytes) {
1391                         res = 0;
1392                         err = -EIO;
1393                         break;
1394                 }
1395                 count -= nres;
1396                 res += nres;
1397                 pos += nres;
1398                 if (nres != nbytes)
1399                         break;
1400                 if (count) {
1401                         fuse_put_request(fc, req);
1402                         if (io->async)
1403                                 req = fuse_get_req_for_background(fc,
1404                                         fuse_iter_npages(iter));
1405                         else
1406                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1407                         if (IS_ERR(req))
1408                                 break;
1409                 }
1410         }
1411         if (!IS_ERR(req))
1412                 fuse_put_request(fc, req);
1413         if (res > 0)
1414                 *ppos = pos;
1415
1416         return res > 0 ? res : err;
1417 }
1418 EXPORT_SYMBOL_GPL(fuse_direct_io);
1419
1420 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1421                                   struct iov_iter *iter,
1422                                   loff_t *ppos)
1423 {
1424         ssize_t res;
1425         struct file *file = io->file;
1426         struct inode *inode = file_inode(file);
1427
1428         if (fuse_is_bad(inode))
1429                 return -EIO;
1430
1431         res = fuse_direct_io(io, iter, ppos, 0);
1432
1433         fuse_invalidate_attr(inode);
1434
1435         return res;
1436 }
1437
1438 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1439 {
1440         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1441         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1442 }
1443
1444 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1445 {
1446         struct file *file = iocb->ki_filp;
1447         struct inode *inode = file_inode(file);
1448         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1449         ssize_t res;
1450
1451         if (fuse_is_bad(inode))
1452                 return -EIO;
1453
1454         /* Don't allow parallel writes to the same file */
1455         inode_lock(inode);
1456         res = generic_write_checks(iocb, from);
1457         if (res > 0)
1458                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1459         fuse_invalidate_attr(inode);
1460         if (res > 0)
1461                 fuse_write_update_size(inode, iocb->ki_pos);
1462         inode_unlock(inode);
1463
1464         return res;
1465 }
1466
1467 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1468 {
1469         int i;
1470
1471         for (i = 0; i < req->num_pages; i++)
1472                 __free_page(req->pages[i]);
1473
1474         if (req->ff)
1475                 fuse_file_put(req->ff, false);
1476 }
1477
1478 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1479 {
1480         struct inode *inode = req->inode;
1481         struct fuse_inode *fi = get_fuse_inode(inode);
1482         struct backing_dev_info *bdi = inode_to_bdi(inode);
1483         int i;
1484
1485         list_del(&req->writepages_entry);
1486         for (i = 0; i < req->num_pages; i++) {
1487                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1488                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1489                 wb_writeout_inc(&bdi->wb);
1490         }
1491         wake_up(&fi->page_waitq);
1492 }
1493
1494 /* Called under fc->lock, may release and reacquire it */
1495 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1496                                 loff_t size)
1497 __releases(fc->lock)
1498 __acquires(fc->lock)
1499 {
1500         struct fuse_inode *fi = get_fuse_inode(req->inode);
1501         struct fuse_write_in *inarg = &req->misc.write.in;
1502         __u64 data_size = req->num_pages * PAGE_SIZE;
1503
1504         if (!fc->connected)
1505                 goto out_free;
1506
1507         if (inarg->offset + data_size <= size) {
1508                 inarg->size = data_size;
1509         } else if (inarg->offset < size) {
1510                 inarg->size = size - inarg->offset;
1511         } else {
1512                 /* Got truncated off completely */
1513                 goto out_free;
1514         }
1515
1516         req->in.args[1].size = inarg->size;
1517         fi->writectr++;
1518         fuse_request_send_background_locked(fc, req);
1519         return;
1520
1521  out_free:
1522         fuse_writepage_finish(fc, req);
1523         spin_unlock(&fc->lock);
1524         fuse_writepage_free(fc, req);
1525         fuse_put_request(fc, req);
1526         spin_lock(&fc->lock);
1527 }
1528
1529 /*
1530  * If fi->writectr is positive (no truncate or fsync going on) send
1531  * all queued writepage requests.
1532  *
1533  * Called with fc->lock
1534  */
1535 void fuse_flush_writepages(struct inode *inode)
1536 __releases(fc->lock)
1537 __acquires(fc->lock)
1538 {
1539         struct fuse_conn *fc = get_fuse_conn(inode);
1540         struct fuse_inode *fi = get_fuse_inode(inode);
1541         loff_t crop = i_size_read(inode);
1542         struct fuse_req *req;
1543
1544         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1545                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1546                 list_del_init(&req->list);
1547                 fuse_send_writepage(fc, req, crop);
1548         }
1549 }
1550
1551 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1552 {
1553         struct inode *inode = req->inode;
1554         struct fuse_inode *fi = get_fuse_inode(inode);
1555
1556         mapping_set_error(inode->i_mapping, req->out.h.error);
1557         spin_lock(&fc->lock);
1558         while (req->misc.write.next) {
1559                 struct fuse_conn *fc = get_fuse_conn(inode);
1560                 struct fuse_write_in *inarg = &req->misc.write.in;
1561                 struct fuse_req *next = req->misc.write.next;
1562                 req->misc.write.next = next->misc.write.next;
1563                 next->misc.write.next = NULL;
1564                 next->ff = fuse_file_get(req->ff);
1565                 list_add(&next->writepages_entry, &fi->writepages);
1566
1567                 /*
1568                  * Skip fuse_flush_writepages() to make it easy to crop requests
1569                  * based on primary request size.
1570                  *
1571                  * 1st case (trivial): there are no concurrent activities using
1572                  * fuse_set/release_nowrite.  Then we're on safe side because
1573                  * fuse_flush_writepages() would call fuse_send_writepage()
1574                  * anyway.
1575                  *
1576                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1577                  * now for completion of all in-flight requests.  This happens
1578                  * rarely and no more than once per page, so this should be
1579                  * okay.
1580                  *
1581                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1582                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1583                  * that fuse_set_nowrite returned implies that all in-flight
1584                  * requests were completed along with all of their secondary
1585                  * requests.  Further primary requests are blocked by negative
1586                  * writectr.  Hence there cannot be any in-flight requests and
1587                  * no invocations of fuse_writepage_end() while we're in
1588                  * fuse_set_nowrite..fuse_release_nowrite section.
1589                  */
1590                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1591         }
1592         fi->writectr--;
1593         fuse_writepage_finish(fc, req);
1594         spin_unlock(&fc->lock);
1595         fuse_writepage_free(fc, req);
1596 }
1597
1598 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1599                                                struct fuse_inode *fi)
1600 {
1601         struct fuse_file *ff = NULL;
1602
1603         spin_lock(&fc->lock);
1604         if (!list_empty(&fi->write_files)) {
1605                 ff = list_entry(fi->write_files.next, struct fuse_file,
1606                                 write_entry);
1607                 fuse_file_get(ff);
1608         }
1609         spin_unlock(&fc->lock);
1610
1611         return ff;
1612 }
1613
1614 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1615                                              struct fuse_inode *fi)
1616 {
1617         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1618         WARN_ON(!ff);
1619         return ff;
1620 }
1621
1622 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1623 {
1624         struct fuse_conn *fc = get_fuse_conn(inode);
1625         struct fuse_inode *fi = get_fuse_inode(inode);
1626         struct fuse_file *ff;
1627         int err;
1628
1629         ff = __fuse_write_file_get(fc, fi);
1630         err = fuse_flush_times(inode, ff);
1631         if (ff)
1632                 fuse_file_put(ff, 0);
1633
1634         return err;
1635 }
1636
1637 static int fuse_writepage_locked(struct page *page)
1638 {
1639         struct address_space *mapping = page->mapping;
1640         struct inode *inode = mapping->host;
1641         struct fuse_conn *fc = get_fuse_conn(inode);
1642         struct fuse_inode *fi = get_fuse_inode(inode);
1643         struct fuse_req *req;
1644         struct page *tmp_page;
1645         int error = -ENOMEM;
1646
1647         set_page_writeback(page);
1648
1649         req = fuse_request_alloc_nofs(1);
1650         if (!req)
1651                 goto err;
1652
1653         /* writeback always goes to bg_queue */
1654         __set_bit(FR_BACKGROUND, &req->flags);
1655         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1656         if (!tmp_page)
1657                 goto err_free;
1658
1659         error = -EIO;
1660         req->ff = fuse_write_file_get(fc, fi);
1661         if (!req->ff)
1662                 goto err_nofile;
1663
1664         fuse_write_fill(req, req->ff, page_offset(page), 0);
1665
1666         copy_highpage(tmp_page, page);
1667         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1668         req->misc.write.next = NULL;
1669         req->in.argpages = 1;
1670         req->num_pages = 1;
1671         req->pages[0] = tmp_page;
1672         req->page_descs[0].offset = 0;
1673         req->page_descs[0].length = PAGE_SIZE;
1674         req->end = fuse_writepage_end;
1675         req->inode = inode;
1676
1677         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1678         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1679
1680         spin_lock(&fc->lock);
1681         list_add(&req->writepages_entry, &fi->writepages);
1682         list_add_tail(&req->list, &fi->queued_writes);
1683         fuse_flush_writepages(inode);
1684         spin_unlock(&fc->lock);
1685
1686         end_page_writeback(page);
1687
1688         return 0;
1689
1690 err_nofile:
1691         __free_page(tmp_page);
1692 err_free:
1693         fuse_request_free(req);
1694 err:
1695         end_page_writeback(page);
1696         return error;
1697 }
1698
1699 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1700 {
1701         int err;
1702
1703         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1704                 /*
1705                  * ->writepages() should be called for sync() and friends.  We
1706                  * should only get here on direct reclaim and then we are
1707                  * allowed to skip a page which is already in flight
1708                  */
1709                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1710
1711                 redirty_page_for_writepage(wbc, page);
1712                 unlock_page(page);
1713                 return 0;
1714         }
1715
1716         err = fuse_writepage_locked(page);
1717         unlock_page(page);
1718
1719         return err;
1720 }
1721
1722 struct fuse_fill_wb_data {
1723         struct fuse_req *req;
1724         struct fuse_file *ff;
1725         struct inode *inode;
1726         struct page **orig_pages;
1727 };
1728
1729 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1730 {
1731         struct fuse_req *req = data->req;
1732         struct inode *inode = data->inode;
1733         struct fuse_conn *fc = get_fuse_conn(inode);
1734         struct fuse_inode *fi = get_fuse_inode(inode);
1735         int num_pages = req->num_pages;
1736         int i;
1737
1738         req->ff = fuse_file_get(data->ff);
1739         spin_lock(&fc->lock);
1740         list_add_tail(&req->list, &fi->queued_writes);
1741         fuse_flush_writepages(inode);
1742         spin_unlock(&fc->lock);
1743
1744         for (i = 0; i < num_pages; i++)
1745                 end_page_writeback(data->orig_pages[i]);
1746 }
1747
1748 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1749                                      struct page *page)
1750 {
1751         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1752         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1753         struct fuse_req *tmp;
1754         struct fuse_req *old_req;
1755         bool found = false;
1756         pgoff_t curr_index;
1757
1758         BUG_ON(new_req->num_pages != 0);
1759
1760         spin_lock(&fc->lock);
1761         list_del(&new_req->writepages_entry);
1762         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1763                 BUG_ON(old_req->inode != new_req->inode);
1764                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1765                 if (curr_index <= page->index &&
1766                     page->index < curr_index + old_req->num_pages) {
1767                         found = true;
1768                         break;
1769                 }
1770         }
1771         if (!found) {
1772                 list_add(&new_req->writepages_entry, &fi->writepages);
1773                 goto out_unlock;
1774         }
1775
1776         new_req->num_pages = 1;
1777         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1778                 BUG_ON(tmp->inode != new_req->inode);
1779                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1780                 if (tmp->num_pages == 1 &&
1781                     curr_index == page->index) {
1782                         old_req = tmp;
1783                 }
1784         }
1785
1786         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1787                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1788
1789                 copy_highpage(old_req->pages[0], page);
1790                 spin_unlock(&fc->lock);
1791
1792                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1793                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1794                 wb_writeout_inc(&bdi->wb);
1795                 fuse_writepage_free(fc, new_req);
1796                 fuse_request_free(new_req);
1797                 goto out;
1798         } else {
1799                 new_req->misc.write.next = old_req->misc.write.next;
1800                 old_req->misc.write.next = new_req;
1801         }
1802 out_unlock:
1803         spin_unlock(&fc->lock);
1804 out:
1805         return found;
1806 }
1807
1808 static int fuse_writepages_fill(struct page *page,
1809                 struct writeback_control *wbc, void *_data)
1810 {
1811         struct fuse_fill_wb_data *data = _data;
1812         struct fuse_req *req = data->req;
1813         struct inode *inode = data->inode;
1814         struct fuse_conn *fc = get_fuse_conn(inode);
1815         struct page *tmp_page;
1816         bool is_writeback;
1817         int err;
1818
1819         if (!data->ff) {
1820                 err = -EIO;
1821                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1822                 if (!data->ff)
1823                         goto out_unlock;
1824         }
1825
1826         /*
1827          * Being under writeback is unlikely but possible.  For example direct
1828          * read to an mmaped fuse file will set the page dirty twice; once when
1829          * the pages are faulted with get_user_pages(), and then after the read
1830          * completed.
1831          */
1832         is_writeback = fuse_page_is_writeback(inode, page->index);
1833
1834         if (req && req->num_pages &&
1835             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1836              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1837              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1838                 fuse_writepages_send(data);
1839                 data->req = NULL;
1840         }
1841         err = -ENOMEM;
1842         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1843         if (!tmp_page)
1844                 goto out_unlock;
1845
1846         /*
1847          * The page must not be redirtied until the writeout is completed
1848          * (i.e. userspace has sent a reply to the write request).  Otherwise
1849          * there could be more than one temporary page instance for each real
1850          * page.
1851          *
1852          * This is ensured by holding the page lock in page_mkwrite() while
1853          * checking fuse_page_is_writeback().  We already hold the page lock
1854          * since clear_page_dirty_for_io() and keep it held until we add the
1855          * request to the fi->writepages list and increment req->num_pages.
1856          * After this fuse_page_is_writeback() will indicate that the page is
1857          * under writeback, so we can release the page lock.
1858          */
1859         if (data->req == NULL) {
1860                 struct fuse_inode *fi = get_fuse_inode(inode);
1861
1862                 err = -ENOMEM;
1863                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1864                 if (!req) {
1865                         __free_page(tmp_page);
1866                         goto out_unlock;
1867                 }
1868
1869                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1870                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1871                 req->misc.write.next = NULL;
1872                 req->in.argpages = 1;
1873                 __set_bit(FR_BACKGROUND, &req->flags);
1874                 req->num_pages = 0;
1875                 req->end = fuse_writepage_end;
1876                 req->inode = inode;
1877
1878                 spin_lock(&fc->lock);
1879                 list_add(&req->writepages_entry, &fi->writepages);
1880                 spin_unlock(&fc->lock);
1881
1882                 data->req = req;
1883         }
1884         set_page_writeback(page);
1885
1886         copy_highpage(tmp_page, page);
1887         req->pages[req->num_pages] = tmp_page;
1888         req->page_descs[req->num_pages].offset = 0;
1889         req->page_descs[req->num_pages].length = PAGE_SIZE;
1890
1891         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1892         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1893
1894         err = 0;
1895         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1896                 end_page_writeback(page);
1897                 data->req = NULL;
1898                 goto out_unlock;
1899         }
1900         data->orig_pages[req->num_pages] = page;
1901
1902         /*
1903          * Protected by fc->lock against concurrent access by
1904          * fuse_page_is_writeback().
1905          */
1906         spin_lock(&fc->lock);
1907         req->num_pages++;
1908         spin_unlock(&fc->lock);
1909
1910 out_unlock:
1911         unlock_page(page);
1912
1913         return err;
1914 }
1915
1916 static int fuse_writepages(struct address_space *mapping,
1917                            struct writeback_control *wbc)
1918 {
1919         struct inode *inode = mapping->host;
1920         struct fuse_fill_wb_data data;
1921         int err;
1922
1923         err = -EIO;
1924         if (fuse_is_bad(inode))
1925                 goto out;
1926
1927         data.inode = inode;
1928         data.req = NULL;
1929         data.ff = NULL;
1930
1931         err = -ENOMEM;
1932         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1933                                   sizeof(struct page *),
1934                                   GFP_NOFS);
1935         if (!data.orig_pages)
1936                 goto out;
1937
1938         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1939         if (data.req) {
1940                 /* Ignore errors if we can write at least one page */
1941                 BUG_ON(!data.req->num_pages);
1942                 fuse_writepages_send(&data);
1943                 err = 0;
1944         }
1945         if (data.ff)
1946                 fuse_file_put(data.ff, false);
1947
1948         kfree(data.orig_pages);
1949 out:
1950         return err;
1951 }
1952
1953 /*
1954  * It's worthy to make sure that space is reserved on disk for the write,
1955  * but how to implement it without killing performance need more thinking.
1956  */
1957 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1958                 loff_t pos, unsigned len, unsigned flags,
1959                 struct page **pagep, void **fsdata)
1960 {
1961         pgoff_t index = pos >> PAGE_SHIFT;
1962         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1963         struct page *page;
1964         loff_t fsize;
1965         int err = -ENOMEM;
1966
1967         WARN_ON(!fc->writeback_cache);
1968
1969         page = grab_cache_page_write_begin(mapping, index, flags);
1970         if (!page)
1971                 goto error;
1972
1973         fuse_wait_on_page_writeback(mapping->host, page->index);
1974
1975         if (PageUptodate(page) || len == PAGE_SIZE)
1976                 goto success;
1977         /*
1978          * Check if the start this page comes after the end of file, in which
1979          * case the readpage can be optimized away.
1980          */
1981         fsize = i_size_read(mapping->host);
1982         if (fsize <= (pos & PAGE_MASK)) {
1983                 size_t off = pos & ~PAGE_MASK;
1984                 if (off)
1985                         zero_user_segment(page, 0, off);
1986                 goto success;
1987         }
1988         err = fuse_do_readpage(file, page);
1989         if (err)
1990                 goto cleanup;
1991 success:
1992         *pagep = page;
1993         return 0;
1994
1995 cleanup:
1996         unlock_page(page);
1997         put_page(page);
1998 error:
1999         return err;
2000 }
2001
2002 static int fuse_write_end(struct file *file, struct address_space *mapping,
2003                 loff_t pos, unsigned len, unsigned copied,
2004                 struct page *page, void *fsdata)
2005 {
2006         struct inode *inode = page->mapping->host;
2007
2008         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2009         if (!copied)
2010                 goto unlock;
2011
2012         if (!PageUptodate(page)) {
2013                 /* Zero any unwritten bytes at the end of the page */
2014                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2015                 if (endoff)
2016                         zero_user_segment(page, endoff, PAGE_SIZE);
2017                 SetPageUptodate(page);
2018         }
2019
2020         fuse_write_update_size(inode, pos + copied);
2021         set_page_dirty(page);
2022
2023 unlock:
2024         unlock_page(page);
2025         put_page(page);
2026
2027         return copied;
2028 }
2029
2030 static int fuse_launder_page(struct page *page)
2031 {
2032         int err = 0;
2033         if (clear_page_dirty_for_io(page)) {
2034                 struct inode *inode = page->mapping->host;
2035                 err = fuse_writepage_locked(page);
2036                 if (!err)
2037                         fuse_wait_on_page_writeback(inode, page->index);
2038         }
2039         return err;
2040 }
2041
2042 /*
2043  * Write back dirty pages now, because there may not be any suitable
2044  * open files later
2045  */
2046 static void fuse_vma_close(struct vm_area_struct *vma)
2047 {
2048         filemap_write_and_wait(vma->vm_file->f_mapping);
2049 }
2050
2051 /*
2052  * Wait for writeback against this page to complete before allowing it
2053  * to be marked dirty again, and hence written back again, possibly
2054  * before the previous writepage completed.
2055  *
2056  * Block here, instead of in ->writepage(), so that the userspace fs
2057  * can only block processes actually operating on the filesystem.
2058  *
2059  * Otherwise unprivileged userspace fs would be able to block
2060  * unrelated:
2061  *
2062  * - page migration
2063  * - sync(2)
2064  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2065  */
2066 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2067 {
2068         struct page *page = vmf->page;
2069         struct inode *inode = file_inode(vma->vm_file);
2070
2071         file_update_time(vma->vm_file);
2072         lock_page(page);
2073         if (page->mapping != inode->i_mapping) {
2074                 unlock_page(page);
2075                 return VM_FAULT_NOPAGE;
2076         }
2077
2078         fuse_wait_on_page_writeback(inode, page->index);
2079         return VM_FAULT_LOCKED;
2080 }
2081
2082 static const struct vm_operations_struct fuse_file_vm_ops = {
2083         .close          = fuse_vma_close,
2084         .fault          = filemap_fault,
2085         .map_pages      = filemap_map_pages,
2086         .page_mkwrite   = fuse_page_mkwrite,
2087 };
2088
2089 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2090 {
2091         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2092                 fuse_link_write_file(file);
2093
2094         file_accessed(file);
2095         vma->vm_ops = &fuse_file_vm_ops;
2096         return 0;
2097 }
2098
2099 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2100 {
2101         /* Can't provide the coherency needed for MAP_SHARED */
2102         if (vma->vm_flags & VM_MAYSHARE)
2103                 return -ENODEV;
2104
2105         invalidate_inode_pages2(file->f_mapping);
2106
2107         return generic_file_mmap(file, vma);
2108 }
2109
2110 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2111                                   struct file_lock *fl)
2112 {
2113         switch (ffl->type) {
2114         case F_UNLCK:
2115                 break;
2116
2117         case F_RDLCK:
2118         case F_WRLCK:
2119                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2120                     ffl->end < ffl->start)
2121                         return -EIO;
2122
2123                 fl->fl_start = ffl->start;
2124                 fl->fl_end = ffl->end;
2125                 fl->fl_pid = ffl->pid;
2126                 break;
2127
2128         default:
2129                 return -EIO;
2130         }
2131         fl->fl_type = ffl->type;
2132         return 0;
2133 }
2134
2135 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2136                          const struct file_lock *fl, int opcode, pid_t pid,
2137                          int flock, struct fuse_lk_in *inarg)
2138 {
2139         struct inode *inode = file_inode(file);
2140         struct fuse_conn *fc = get_fuse_conn(inode);
2141         struct fuse_file *ff = file->private_data;
2142
2143         memset(inarg, 0, sizeof(*inarg));
2144         inarg->fh = ff->fh;
2145         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2146         inarg->lk.start = fl->fl_start;
2147         inarg->lk.end = fl->fl_end;
2148         inarg->lk.type = fl->fl_type;
2149         inarg->lk.pid = pid;
2150         if (flock)
2151                 inarg->lk_flags |= FUSE_LK_FLOCK;
2152         args->in.h.opcode = opcode;
2153         args->in.h.nodeid = get_node_id(inode);
2154         args->in.numargs = 1;
2155         args->in.args[0].size = sizeof(*inarg);
2156         args->in.args[0].value = inarg;
2157 }
2158
2159 static int fuse_getlk(struct file *file, struct file_lock *fl)
2160 {
2161         struct inode *inode = file_inode(file);
2162         struct fuse_conn *fc = get_fuse_conn(inode);
2163         FUSE_ARGS(args);
2164         struct fuse_lk_in inarg;
2165         struct fuse_lk_out outarg;
2166         int err;
2167
2168         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2169         args.out.numargs = 1;
2170         args.out.args[0].size = sizeof(outarg);
2171         args.out.args[0].value = &outarg;
2172         err = fuse_simple_request(fc, &args);
2173         if (!err)
2174                 err = convert_fuse_file_lock(&outarg.lk, fl);
2175
2176         return err;
2177 }
2178
2179 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2180 {
2181         struct inode *inode = file_inode(file);
2182         struct fuse_conn *fc = get_fuse_conn(inode);
2183         FUSE_ARGS(args);
2184         struct fuse_lk_in inarg;
2185         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2186         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2187         int err;
2188
2189         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2190                 /* NLM needs asynchronous locks, which we don't support yet */
2191                 return -ENOLCK;
2192         }
2193
2194         /* Unlock on close is handled by the flush method */
2195         if (fl->fl_flags & FL_CLOSE)
2196                 return 0;
2197
2198         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2199         err = fuse_simple_request(fc, &args);
2200
2201         /* locking is restartable */
2202         if (err == -EINTR)
2203                 err = -ERESTARTSYS;
2204
2205         return err;
2206 }
2207
2208 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2209 {
2210         struct inode *inode = file_inode(file);
2211         struct fuse_conn *fc = get_fuse_conn(inode);
2212         int err;
2213
2214         if (cmd == F_CANCELLK) {
2215                 err = 0;
2216         } else if (cmd == F_GETLK) {
2217                 if (fc->no_lock) {
2218                         posix_test_lock(file, fl);
2219                         err = 0;
2220                 } else
2221                         err = fuse_getlk(file, fl);
2222         } else {
2223                 if (fc->no_lock)
2224                         err = posix_lock_file(file, fl, NULL);
2225                 else
2226                         err = fuse_setlk(file, fl, 0);
2227         }
2228         return err;
2229 }
2230
2231 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2232 {
2233         struct inode *inode = file_inode(file);
2234         struct fuse_conn *fc = get_fuse_conn(inode);
2235         int err;
2236
2237         if (fc->no_flock) {
2238                 err = locks_lock_file_wait(file, fl);
2239         } else {
2240                 struct fuse_file *ff = file->private_data;
2241
2242                 /* emulate flock with POSIX locks */
2243                 ff->flock = true;
2244                 err = fuse_setlk(file, fl, 1);
2245         }
2246
2247         return err;
2248 }
2249
2250 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2251 {
2252         struct inode *inode = mapping->host;
2253         struct fuse_conn *fc = get_fuse_conn(inode);
2254         FUSE_ARGS(args);
2255         struct fuse_bmap_in inarg;
2256         struct fuse_bmap_out outarg;
2257         int err;
2258
2259         if (!inode->i_sb->s_bdev || fc->no_bmap)
2260                 return 0;
2261
2262         memset(&inarg, 0, sizeof(inarg));
2263         inarg.block = block;
2264         inarg.blocksize = inode->i_sb->s_blocksize;
2265         args.in.h.opcode = FUSE_BMAP;
2266         args.in.h.nodeid = get_node_id(inode);
2267         args.in.numargs = 1;
2268         args.in.args[0].size = sizeof(inarg);
2269         args.in.args[0].value = &inarg;
2270         args.out.numargs = 1;
2271         args.out.args[0].size = sizeof(outarg);
2272         args.out.args[0].value = &outarg;
2273         err = fuse_simple_request(fc, &args);
2274         if (err == -ENOSYS)
2275                 fc->no_bmap = 1;
2276
2277         return err ? 0 : outarg.block;
2278 }
2279
2280 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2281 {
2282         struct inode *inode = file->f_mapping->host;
2283         struct fuse_conn *fc = get_fuse_conn(inode);
2284         struct fuse_file *ff = file->private_data;
2285         FUSE_ARGS(args);
2286         struct fuse_lseek_in inarg = {
2287                 .fh = ff->fh,
2288                 .offset = offset,
2289                 .whence = whence
2290         };
2291         struct fuse_lseek_out outarg;
2292         int err;
2293
2294         if (fc->no_lseek)
2295                 goto fallback;
2296
2297         args.in.h.opcode = FUSE_LSEEK;
2298         args.in.h.nodeid = ff->nodeid;
2299         args.in.numargs = 1;
2300         args.in.args[0].size = sizeof(inarg);
2301         args.in.args[0].value = &inarg;
2302         args.out.numargs = 1;
2303         args.out.args[0].size = sizeof(outarg);
2304         args.out.args[0].value = &outarg;
2305         err = fuse_simple_request(fc, &args);
2306         if (err) {
2307                 if (err == -ENOSYS) {
2308                         fc->no_lseek = 1;
2309                         goto fallback;
2310                 }
2311                 return err;
2312         }
2313
2314         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2315
2316 fallback:
2317         err = fuse_update_attributes(inode, NULL, file, NULL);
2318         if (!err)
2319                 return generic_file_llseek(file, offset, whence);
2320         else
2321                 return err;
2322 }
2323
2324 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2325 {
2326         loff_t retval;
2327         struct inode *inode = file_inode(file);
2328
2329         switch (whence) {
2330         case SEEK_SET:
2331         case SEEK_CUR:
2332                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2333                 retval = generic_file_llseek(file, offset, whence);
2334                 break;
2335         case SEEK_END:
2336                 inode_lock(inode);
2337                 retval = fuse_update_attributes(inode, NULL, file, NULL);
2338                 if (!retval)
2339                         retval = generic_file_llseek(file, offset, whence);
2340                 inode_unlock(inode);
2341                 break;
2342         case SEEK_HOLE:
2343         case SEEK_DATA:
2344                 inode_lock(inode);
2345                 retval = fuse_lseek(file, offset, whence);
2346                 inode_unlock(inode);
2347                 break;
2348         default:
2349                 retval = -EINVAL;
2350         }
2351
2352         return retval;
2353 }
2354
2355 /*
2356  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2357  * ABI was defined to be 'struct iovec' which is different on 32bit
2358  * and 64bit.  Fortunately we can determine which structure the server
2359  * used from the size of the reply.
2360  */
2361 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2362                                      size_t transferred, unsigned count,
2363                                      bool is_compat)
2364 {
2365 #ifdef CONFIG_COMPAT
2366         if (count * sizeof(struct compat_iovec) == transferred) {
2367                 struct compat_iovec *ciov = src;
2368                 unsigned i;
2369
2370                 /*
2371                  * With this interface a 32bit server cannot support
2372                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2373                  * requests
2374                  */
2375                 if (!is_compat)
2376                         return -EINVAL;
2377
2378                 for (i = 0; i < count; i++) {
2379                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2380                         dst[i].iov_len = ciov[i].iov_len;
2381                 }
2382                 return 0;
2383         }
2384 #endif
2385
2386         if (count * sizeof(struct iovec) != transferred)
2387                 return -EIO;
2388
2389         memcpy(dst, src, transferred);
2390         return 0;
2391 }
2392
2393 /* Make sure iov_length() won't overflow */
2394 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2395 {
2396         size_t n;
2397         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2398
2399         for (n = 0; n < count; n++, iov++) {
2400                 if (iov->iov_len > (size_t) max)
2401                         return -ENOMEM;
2402                 max -= iov->iov_len;
2403         }
2404         return 0;
2405 }
2406
2407 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2408                                  void *src, size_t transferred, unsigned count,
2409                                  bool is_compat)
2410 {
2411         unsigned i;
2412         struct fuse_ioctl_iovec *fiov = src;
2413
2414         if (fc->minor < 16) {
2415                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2416                                                  count, is_compat);
2417         }
2418
2419         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2420                 return -EIO;
2421
2422         for (i = 0; i < count; i++) {
2423                 /* Did the server supply an inappropriate value? */
2424                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2425                     fiov[i].len != (unsigned long) fiov[i].len)
2426                         return -EIO;
2427
2428                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2429                 dst[i].iov_len = (size_t) fiov[i].len;
2430
2431 #ifdef CONFIG_COMPAT
2432                 if (is_compat &&
2433                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2434                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2435                         return -EIO;
2436 #endif
2437         }
2438
2439         return 0;
2440 }
2441
2442
2443 /*
2444  * For ioctls, there is no generic way to determine how much memory
2445  * needs to be read and/or written.  Furthermore, ioctls are allowed
2446  * to dereference the passed pointer, so the parameter requires deep
2447  * copying but FUSE has no idea whatsoever about what to copy in or
2448  * out.
2449  *
2450  * This is solved by allowing FUSE server to retry ioctl with
2451  * necessary in/out iovecs.  Let's assume the ioctl implementation
2452  * needs to read in the following structure.
2453  *
2454  * struct a {
2455  *      char    *buf;
2456  *      size_t  buflen;
2457  * }
2458  *
2459  * On the first callout to FUSE server, inarg->in_size and
2460  * inarg->out_size will be NULL; then, the server completes the ioctl
2461  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2462  * the actual iov array to
2463  *
2464  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2465  *
2466  * which tells FUSE to copy in the requested area and retry the ioctl.
2467  * On the second round, the server has access to the structure and
2468  * from that it can tell what to look for next, so on the invocation,
2469  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2470  *
2471  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2472  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2473  *
2474  * FUSE will copy both struct a and the pointed buffer from the
2475  * process doing the ioctl and retry ioctl with both struct a and the
2476  * buffer.
2477  *
2478  * This time, FUSE server has everything it needs and completes ioctl
2479  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2480  *
2481  * Copying data out works the same way.
2482  *
2483  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2484  * automatically initializes in and out iovs by decoding @cmd with
2485  * _IOC_* macros and the server is not allowed to request RETRY.  This
2486  * limits ioctl data transfers to well-formed ioctls and is the forced
2487  * behavior for all FUSE servers.
2488  */
2489 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2490                    unsigned int flags)
2491 {
2492         struct fuse_file *ff = file->private_data;
2493         struct fuse_conn *fc = ff->fc;
2494         struct fuse_ioctl_in inarg = {
2495                 .fh = ff->fh,
2496                 .cmd = cmd,
2497                 .arg = arg,
2498                 .flags = flags
2499         };
2500         struct fuse_ioctl_out outarg;
2501         struct fuse_req *req = NULL;
2502         struct page **pages = NULL;
2503         struct iovec *iov_page = NULL;
2504         struct iovec *in_iov = NULL, *out_iov = NULL;
2505         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2506         size_t in_size, out_size, transferred, c;
2507         int err, i;
2508         struct iov_iter ii;
2509
2510 #if BITS_PER_LONG == 32
2511         inarg.flags |= FUSE_IOCTL_32BIT;
2512 #else
2513         if (flags & FUSE_IOCTL_COMPAT)
2514                 inarg.flags |= FUSE_IOCTL_32BIT;
2515 #endif
2516
2517         /* assume all the iovs returned by client always fits in a page */
2518         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2519
2520         err = -ENOMEM;
2521         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2522         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2523         if (!pages || !iov_page)
2524                 goto out;
2525
2526         /*
2527          * If restricted, initialize IO parameters as encoded in @cmd.
2528          * RETRY from server is not allowed.
2529          */
2530         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2531                 struct iovec *iov = iov_page;
2532
2533                 iov->iov_base = (void __user *)arg;
2534
2535                 switch (cmd) {
2536                 case FS_IOC_GETFLAGS:
2537                 case FS_IOC_SETFLAGS:
2538                         iov->iov_len = sizeof(int);
2539                         break;
2540                 default:
2541                         iov->iov_len = _IOC_SIZE(cmd);
2542                         break;
2543                 }
2544
2545                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2546                         in_iov = iov;
2547                         in_iovs = 1;
2548                 }
2549
2550                 if (_IOC_DIR(cmd) & _IOC_READ) {
2551                         out_iov = iov;
2552                         out_iovs = 1;
2553                 }
2554         }
2555
2556  retry:
2557         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2558         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2559
2560         /*
2561          * Out data can be used either for actual out data or iovs,
2562          * make sure there always is at least one page.
2563          */
2564         out_size = max_t(size_t, out_size, PAGE_SIZE);
2565         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2566
2567         /* make sure there are enough buffer pages and init request with them */
2568         err = -ENOMEM;
2569         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2570                 goto out;
2571         while (num_pages < max_pages) {
2572                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2573                 if (!pages[num_pages])
2574                         goto out;
2575                 num_pages++;
2576         }
2577
2578         req = fuse_get_req(fc, num_pages);
2579         if (IS_ERR(req)) {
2580                 err = PTR_ERR(req);
2581                 req = NULL;
2582                 goto out;
2583         }
2584         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2585         req->num_pages = num_pages;
2586         fuse_page_descs_length_init(req, 0, req->num_pages);
2587
2588         /* okay, let's send it to the client */
2589         req->in.h.opcode = FUSE_IOCTL;
2590         req->in.h.nodeid = ff->nodeid;
2591         req->in.numargs = 1;
2592         req->in.args[0].size = sizeof(inarg);
2593         req->in.args[0].value = &inarg;
2594         if (in_size) {
2595                 req->in.numargs++;
2596                 req->in.args[1].size = in_size;
2597                 req->in.argpages = 1;
2598
2599                 err = -EFAULT;
2600                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2601                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2602                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2603                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2604                                 goto out;
2605                 }
2606         }
2607
2608         req->out.numargs = 2;
2609         req->out.args[0].size = sizeof(outarg);
2610         req->out.args[0].value = &outarg;
2611         req->out.args[1].size = out_size;
2612         req->out.argpages = 1;
2613         req->out.argvar = 1;
2614
2615         fuse_request_send(fc, req);
2616         err = req->out.h.error;
2617         transferred = req->out.args[1].size;
2618         fuse_put_request(fc, req);
2619         req = NULL;
2620         if (err)
2621                 goto out;
2622
2623         /* did it ask for retry? */
2624         if (outarg.flags & FUSE_IOCTL_RETRY) {
2625                 void *vaddr;
2626
2627                 /* no retry if in restricted mode */
2628                 err = -EIO;
2629                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2630                         goto out;
2631
2632                 in_iovs = outarg.in_iovs;
2633                 out_iovs = outarg.out_iovs;
2634
2635                 /*
2636                  * Make sure things are in boundary, separate checks
2637                  * are to protect against overflow.
2638                  */
2639                 err = -ENOMEM;
2640                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2641                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2642                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2643                         goto out;
2644
2645                 vaddr = kmap_atomic(pages[0]);
2646                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2647                                             transferred, in_iovs + out_iovs,
2648                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2649                 kunmap_atomic(vaddr);
2650                 if (err)
2651                         goto out;
2652
2653                 in_iov = iov_page;
2654                 out_iov = in_iov + in_iovs;
2655
2656                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2657                 if (err)
2658                         goto out;
2659
2660                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2661                 if (err)
2662                         goto out;
2663
2664                 goto retry;
2665         }
2666
2667         err = -EIO;
2668         if (transferred > inarg.out_size)
2669                 goto out;
2670
2671         err = -EFAULT;
2672         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2673         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2674                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2675                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2676                         goto out;
2677         }
2678         err = 0;
2679  out:
2680         if (req)
2681                 fuse_put_request(fc, req);
2682         free_page((unsigned long) iov_page);
2683         while (num_pages)
2684                 __free_page(pages[--num_pages]);
2685         kfree(pages);
2686
2687         return err ? err : outarg.result;
2688 }
2689 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2690
2691 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2692                        unsigned long arg, unsigned int flags)
2693 {
2694         struct inode *inode = file_inode(file);
2695         struct fuse_conn *fc = get_fuse_conn(inode);
2696
2697         if (!fuse_allow_current_process(fc))
2698                 return -EACCES;
2699
2700         if (fuse_is_bad(inode))
2701                 return -EIO;
2702
2703         return fuse_do_ioctl(file, cmd, arg, flags);
2704 }
2705
2706 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2707                             unsigned long arg)
2708 {
2709         return fuse_ioctl_common(file, cmd, arg, 0);
2710 }
2711
2712 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2713                                    unsigned long arg)
2714 {
2715         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2716 }
2717
2718 /*
2719  * All files which have been polled are linked to RB tree
2720  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2721  * find the matching one.
2722  */
2723 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2724                                               struct rb_node **parent_out)
2725 {
2726         struct rb_node **link = &fc->polled_files.rb_node;
2727         struct rb_node *last = NULL;
2728
2729         while (*link) {
2730                 struct fuse_file *ff;
2731
2732                 last = *link;
2733                 ff = rb_entry(last, struct fuse_file, polled_node);
2734
2735                 if (kh < ff->kh)
2736                         link = &last->rb_left;
2737                 else if (kh > ff->kh)
2738                         link = &last->rb_right;
2739                 else
2740                         return link;
2741         }
2742
2743         if (parent_out)
2744                 *parent_out = last;
2745         return link;
2746 }
2747
2748 /*
2749  * The file is about to be polled.  Make sure it's on the polled_files
2750  * RB tree.  Note that files once added to the polled_files tree are
2751  * not removed before the file is released.  This is because a file
2752  * polled once is likely to be polled again.
2753  */
2754 static void fuse_register_polled_file(struct fuse_conn *fc,
2755                                       struct fuse_file *ff)
2756 {
2757         spin_lock(&fc->lock);
2758         if (RB_EMPTY_NODE(&ff->polled_node)) {
2759                 struct rb_node **link, *uninitialized_var(parent);
2760
2761                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2762                 BUG_ON(*link);
2763                 rb_link_node(&ff->polled_node, parent, link);
2764                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2765         }
2766         spin_unlock(&fc->lock);
2767 }
2768
2769 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2770 {
2771         struct fuse_file *ff = file->private_data;
2772         struct fuse_conn *fc = ff->fc;
2773         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2774         struct fuse_poll_out outarg;
2775         FUSE_ARGS(args);
2776         int err;
2777
2778         if (fc->no_poll)
2779                 return DEFAULT_POLLMASK;
2780
2781         poll_wait(file, &ff->poll_wait, wait);
2782         inarg.events = (__u32)poll_requested_events(wait);
2783
2784         /*
2785          * Ask for notification iff there's someone waiting for it.
2786          * The client may ignore the flag and always notify.
2787          */
2788         if (waitqueue_active(&ff->poll_wait)) {
2789                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2790                 fuse_register_polled_file(fc, ff);
2791         }
2792
2793         args.in.h.opcode = FUSE_POLL;
2794         args.in.h.nodeid = ff->nodeid;
2795         args.in.numargs = 1;
2796         args.in.args[0].size = sizeof(inarg);
2797         args.in.args[0].value = &inarg;
2798         args.out.numargs = 1;
2799         args.out.args[0].size = sizeof(outarg);
2800         args.out.args[0].value = &outarg;
2801         err = fuse_simple_request(fc, &args);
2802
2803         if (!err)
2804                 return outarg.revents;
2805         if (err == -ENOSYS) {
2806                 fc->no_poll = 1;
2807                 return DEFAULT_POLLMASK;
2808         }
2809         return POLLERR;
2810 }
2811 EXPORT_SYMBOL_GPL(fuse_file_poll);
2812
2813 /*
2814  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2815  * wakes up the poll waiters.
2816  */
2817 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2818                             struct fuse_notify_poll_wakeup_out *outarg)
2819 {
2820         u64 kh = outarg->kh;
2821         struct rb_node **link;
2822
2823         spin_lock(&fc->lock);
2824
2825         link = fuse_find_polled_node(fc, kh, NULL);
2826         if (*link) {
2827                 struct fuse_file *ff;
2828
2829                 ff = rb_entry(*link, struct fuse_file, polled_node);
2830                 wake_up_interruptible_sync(&ff->poll_wait);
2831         }
2832
2833         spin_unlock(&fc->lock);
2834         return 0;
2835 }
2836
2837 static void fuse_do_truncate(struct file *file)
2838 {
2839         struct inode *inode = file->f_mapping->host;
2840         struct iattr attr;
2841
2842         attr.ia_valid = ATTR_SIZE;
2843         attr.ia_size = i_size_read(inode);
2844
2845         attr.ia_file = file;
2846         attr.ia_valid |= ATTR_FILE;
2847
2848         fuse_do_setattr(file_dentry(file), &attr, file);
2849 }
2850
2851 static inline loff_t fuse_round_up(loff_t off)
2852 {
2853         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2854 }
2855
2856 static ssize_t
2857 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2858 {
2859         DECLARE_COMPLETION_ONSTACK(wait);
2860         ssize_t ret = 0;
2861         struct file *file = iocb->ki_filp;
2862         struct fuse_file *ff = file->private_data;
2863         bool async_dio = ff->fc->async_dio;
2864         loff_t pos = 0;
2865         struct inode *inode;
2866         loff_t i_size;
2867         size_t count = iov_iter_count(iter);
2868         loff_t offset = iocb->ki_pos;
2869         struct fuse_io_priv *io;
2870
2871         pos = offset;
2872         inode = file->f_mapping->host;
2873         i_size = i_size_read(inode);
2874
2875         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2876                 return 0;
2877
2878         /* optimization for short read */
2879         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2880                 if (offset >= i_size)
2881                         return 0;
2882                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2883                 count = iov_iter_count(iter);
2884         }
2885
2886         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2887         if (!io)
2888                 return -ENOMEM;
2889         spin_lock_init(&io->lock);
2890         kref_init(&io->refcnt);
2891         io->reqs = 1;
2892         io->bytes = -1;
2893         io->size = 0;
2894         io->offset = offset;
2895         io->write = (iov_iter_rw(iter) == WRITE);
2896         io->err = 0;
2897         io->file = file;
2898         /*
2899          * By default, we want to optimize all I/Os with async request
2900          * submission to the client filesystem if supported.
2901          */
2902         io->async = async_dio;
2903         io->iocb = iocb;
2904         io->blocking = is_sync_kiocb(iocb);
2905
2906         /*
2907          * We cannot asynchronously extend the size of a file.
2908          * In such case the aio will behave exactly like sync io.
2909          */
2910         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2911                 io->blocking = true;
2912
2913         if (io->async && io->blocking) {
2914                 /*
2915                  * Additional reference to keep io around after
2916                  * calling fuse_aio_complete()
2917                  */
2918                 kref_get(&io->refcnt);
2919                 io->done = &wait;
2920         }
2921
2922         if (iov_iter_rw(iter) == WRITE) {
2923                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2924                 fuse_invalidate_attr(inode);
2925         } else {
2926                 ret = __fuse_direct_read(io, iter, &pos);
2927         }
2928
2929         if (io->async) {
2930                 bool blocking = io->blocking;
2931
2932                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2933
2934                 /* we have a non-extending, async request, so return */
2935                 if (!blocking)
2936                         return -EIOCBQUEUED;
2937
2938                 wait_for_completion(&wait);
2939                 ret = fuse_get_res_by_io(io);
2940         }
2941
2942         kref_put(&io->refcnt, fuse_io_release);
2943
2944         if (iov_iter_rw(iter) == WRITE) {
2945                 if (ret > 0)
2946                         fuse_write_update_size(inode, pos);
2947                 else if (ret < 0 && offset + count > i_size)
2948                         fuse_do_truncate(file);
2949         }
2950
2951         return ret;
2952 }
2953
2954 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2955                                 loff_t length)
2956 {
2957         struct fuse_file *ff = file->private_data;
2958         struct inode *inode = file_inode(file);
2959         struct fuse_inode *fi = get_fuse_inode(inode);
2960         struct fuse_conn *fc = ff->fc;
2961         FUSE_ARGS(args);
2962         struct fuse_fallocate_in inarg = {
2963                 .fh = ff->fh,
2964                 .offset = offset,
2965                 .length = length,
2966                 .mode = mode
2967         };
2968         int err;
2969         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2970                            (mode & FALLOC_FL_PUNCH_HOLE);
2971
2972         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2973                 return -EOPNOTSUPP;
2974
2975         if (fc->no_fallocate)
2976                 return -EOPNOTSUPP;
2977
2978         if (lock_inode) {
2979                 inode_lock(inode);
2980                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2981                         loff_t endbyte = offset + length - 1;
2982                         err = filemap_write_and_wait_range(inode->i_mapping,
2983                                                            offset, endbyte);
2984                         if (err)
2985                                 goto out;
2986
2987                         fuse_sync_writes(inode);
2988                 }
2989         }
2990
2991         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2992             offset + length > i_size_read(inode)) {
2993                 err = inode_newsize_ok(inode, offset + length);
2994                 if (err)
2995                         goto out;
2996         }
2997
2998         if (!(mode & FALLOC_FL_KEEP_SIZE))
2999                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3000
3001         args.in.h.opcode = FUSE_FALLOCATE;
3002         args.in.h.nodeid = ff->nodeid;
3003         args.in.numargs = 1;
3004         args.in.args[0].size = sizeof(inarg);
3005         args.in.args[0].value = &inarg;
3006         err = fuse_simple_request(fc, &args);
3007         if (err == -ENOSYS) {
3008                 fc->no_fallocate = 1;
3009                 err = -EOPNOTSUPP;
3010         }
3011         if (err)
3012                 goto out;
3013
3014         /* we could have extended the file */
3015         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3016                 bool changed = fuse_write_update_size(inode, offset + length);
3017
3018                 if (changed && fc->writeback_cache)
3019                         file_update_time(file);
3020         }
3021
3022         if (mode & FALLOC_FL_PUNCH_HOLE)
3023                 truncate_pagecache_range(inode, offset, offset + length - 1);
3024
3025         fuse_invalidate_attr(inode);
3026
3027 out:
3028         if (!(mode & FALLOC_FL_KEEP_SIZE))
3029                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3030
3031         if (lock_inode)
3032                 inode_unlock(inode);
3033
3034         return err;
3035 }
3036
3037 static const struct file_operations fuse_file_operations = {
3038         .llseek         = fuse_file_llseek,
3039         .read_iter      = fuse_file_read_iter,
3040         .write_iter     = fuse_file_write_iter,
3041         .mmap           = fuse_file_mmap,
3042         .open           = fuse_open,
3043         .flush          = fuse_flush,
3044         .release        = fuse_release,
3045         .fsync          = fuse_fsync,
3046         .lock           = fuse_file_lock,
3047         .flock          = fuse_file_flock,
3048         .splice_read    = generic_file_splice_read,
3049         .unlocked_ioctl = fuse_file_ioctl,
3050         .compat_ioctl   = fuse_file_compat_ioctl,
3051         .poll           = fuse_file_poll,
3052         .fallocate      = fuse_file_fallocate,
3053 };
3054
3055 static const struct file_operations fuse_direct_io_file_operations = {
3056         .llseek         = fuse_file_llseek,
3057         .read_iter      = fuse_direct_read_iter,
3058         .write_iter     = fuse_direct_write_iter,
3059         .mmap           = fuse_direct_mmap,
3060         .open           = fuse_open,
3061         .flush          = fuse_flush,
3062         .release        = fuse_release,
3063         .fsync          = fuse_fsync,
3064         .lock           = fuse_file_lock,
3065         .flock          = fuse_file_flock,
3066         .unlocked_ioctl = fuse_file_ioctl,
3067         .compat_ioctl   = fuse_file_compat_ioctl,
3068         .poll           = fuse_file_poll,
3069         .fallocate      = fuse_file_fallocate,
3070         /* no splice_read */
3071 };
3072
3073 static const struct address_space_operations fuse_file_aops  = {
3074         .readpage       = fuse_readpage,
3075         .writepage      = fuse_writepage,
3076         .writepages     = fuse_writepages,
3077         .launder_page   = fuse_launder_page,
3078         .readpages      = fuse_readpages,
3079         .set_page_dirty = __set_page_dirty_nobuffers,
3080         .bmap           = fuse_bmap,
3081         .direct_IO      = fuse_direct_IO,
3082         .write_begin    = fuse_write_begin,
3083         .write_end      = fuse_write_end,
3084 };
3085
3086 void fuse_init_file_inode(struct inode *inode)
3087 {
3088         inode->i_fop = &fuse_file_operations;
3089         inode->i_data.a_ops = &fuse_file_aops;
3090 }