fs/fuse/file.c

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