fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/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                 if (!tmp) {
1092                         unlock_page(page);
1093                         page_cache_release(page);
1094                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1095                         goto again;
1096                 }
1097
1098                 err = 0;
1099                 req->pages[req->num_pages] = page;
1100                 req->page_descs[req->num_pages].length = tmp;
1101                 req->num_pages++;
1102
1103                 iov_iter_advance(ii, tmp);
1104                 count += tmp;
1105                 pos += tmp;
1106                 offset += tmp;
1107                 if (offset == PAGE_CACHE_SIZE)
1108                         offset = 0;
1109
1110                 if (!fc->big_writes)
1111                         break;
1112         } while (iov_iter_count(ii) && count < fc->max_write &&
1113                  req->num_pages < req->max_pages && offset == 0);
1114
1115         return count > 0 ? count : err;
1116 }
1117
1118 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1119 {
1120         return min_t(unsigned,
1121                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1122                      (pos >> PAGE_CACHE_SHIFT) + 1,
1123                      FUSE_MAX_PAGES_PER_REQ);
1124 }
1125
1126 static ssize_t fuse_perform_write(struct file *file,
1127                                   struct address_space *mapping,
1128                                   struct iov_iter *ii, loff_t pos)
1129 {
1130         struct inode *inode = mapping->host;
1131         struct fuse_conn *fc = get_fuse_conn(inode);
1132         struct fuse_inode *fi = get_fuse_inode(inode);
1133         int err = 0;
1134         ssize_t res = 0;
1135
1136         if (is_bad_inode(inode))
1137                 return -EIO;
1138
1139         if (inode->i_size < pos + iov_iter_count(ii))
1140                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1141
1142         do {
1143                 struct fuse_req *req;
1144                 ssize_t count;
1145                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1146
1147                 req = fuse_get_req(fc, nr_pages);
1148                 if (IS_ERR(req)) {
1149                         err = PTR_ERR(req);
1150                         break;
1151                 }
1152
1153                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1154                 if (count <= 0) {
1155                         err = count;
1156                 } else {
1157                         size_t num_written;
1158
1159                         num_written = fuse_send_write_pages(req, file, inode,
1160                                                             pos, count);
1161                         err = req->out.h.error;
1162                         if (!err) {
1163                                 res += num_written;
1164                                 pos += num_written;
1165
1166                                 /* break out of the loop on short write */
1167                                 if (num_written != count)
1168                                         err = -EIO;
1169                         }
1170                 }
1171                 fuse_put_request(fc, req);
1172         } while (!err && iov_iter_count(ii));
1173
1174         if (res > 0)
1175                 fuse_write_update_size(inode, pos);
1176
1177         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1178         fuse_invalidate_attr(inode);
1179
1180         return res > 0 ? res : err;
1181 }
1182
1183 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1184 {
1185         struct file *file = iocb->ki_filp;
1186         struct address_space *mapping = file->f_mapping;
1187         size_t count = iov_iter_count(from);
1188         ssize_t written = 0;
1189         ssize_t written_buffered = 0;
1190         struct inode *inode = mapping->host;
1191         ssize_t err;
1192         loff_t endbyte = 0;
1193         loff_t pos = iocb->ki_pos;
1194
1195         if (get_fuse_conn(inode)->writeback_cache) {
1196                 /* Update size (EOF optimization) and mode (SUID clearing) */
1197                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1198                 if (err)
1199                         return err;
1200
1201                 return generic_file_write_iter(iocb, from);
1202         }
1203
1204         mutex_lock(&inode->i_mutex);
1205
1206         /* We can write back this queue in page reclaim */
1207         current->backing_dev_info = mapping->backing_dev_info;
1208
1209         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1210         if (err)
1211                 goto out;
1212
1213         if (count == 0)
1214                 goto out;
1215
1216         iov_iter_truncate(from, count);
1217         err = file_remove_suid(file);
1218         if (err)
1219                 goto out;
1220
1221         err = file_update_time(file);
1222         if (err)
1223                 goto out;
1224
1225         if (file->f_flags & O_DIRECT) {
1226                 written = generic_file_direct_write(iocb, from, pos);
1227                 if (written < 0 || !iov_iter_count(from))
1228                         goto out;
1229
1230                 pos += written;
1231
1232                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1233                 if (written_buffered < 0) {
1234                         err = written_buffered;
1235                         goto out;
1236                 }
1237                 endbyte = pos + written_buffered - 1;
1238
1239                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1240                                                    endbyte);
1241                 if (err)
1242                         goto out;
1243
1244                 invalidate_mapping_pages(file->f_mapping,
1245                                          pos >> PAGE_CACHE_SHIFT,
1246                                          endbyte >> PAGE_CACHE_SHIFT);
1247
1248                 written += written_buffered;
1249                 iocb->ki_pos = pos + written_buffered;
1250         } else {
1251                 written = fuse_perform_write(file, mapping, from, pos);
1252                 if (written >= 0)
1253                         iocb->ki_pos = pos + written;
1254         }
1255 out:
1256         current->backing_dev_info = NULL;
1257         mutex_unlock(&inode->i_mutex);
1258
1259         return written ? written : err;
1260 }
1261
1262 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1263                 unsigned index, unsigned nr_pages)
1264 {
1265         int i;
1266
1267         for (i = index; i < index + nr_pages; i++)
1268                 req->page_descs[i].length = PAGE_SIZE -
1269                         req->page_descs[i].offset;
1270 }
1271
1272 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1273 {
1274         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1275 }
1276
1277 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1278                                         size_t max_size)
1279 {
1280         return min(iov_iter_single_seg_count(ii), max_size);
1281 }
1282
1283 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1284                                size_t *nbytesp, int write)
1285 {
1286         size_t nbytes = 0;  /* # bytes already packed in req */
1287
1288         /* Special case for kernel I/O: can copy directly into the buffer */
1289         if (ii->type & ITER_KVEC) {
1290                 unsigned long user_addr = fuse_get_user_addr(ii);
1291                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1292
1293                 if (write)
1294                         req->in.args[1].value = (void *) user_addr;
1295                 else
1296                         req->out.args[0].value = (void *) user_addr;
1297
1298                 iov_iter_advance(ii, frag_size);
1299                 *nbytesp = frag_size;
1300                 return 0;
1301         }
1302
1303         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1304                 unsigned npages;
1305                 size_t start;
1306                 unsigned n = req->max_pages - req->num_pages;
1307                 ssize_t ret = iov_iter_get_pages(ii,
1308                                         &req->pages[req->num_pages],
1309                                         n * PAGE_SIZE, &start);
1310                 if (ret < 0)
1311                         return ret;
1312
1313                 iov_iter_advance(ii, ret);
1314                 nbytes += ret;
1315
1316                 ret += start;
1317                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1318
1319                 req->page_descs[req->num_pages].offset = start;
1320                 fuse_page_descs_length_init(req, req->num_pages, npages);
1321
1322                 req->num_pages += npages;
1323                 req->page_descs[req->num_pages - 1].length -=
1324                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1325         }
1326
1327         if (write)
1328                 req->in.argpages = 1;
1329         else
1330                 req->out.argpages = 1;
1331
1332         *nbytesp = nbytes;
1333
1334         return 0;
1335 }
1336
1337 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1338 {
1339         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1340 }
1341
1342 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1343                        loff_t *ppos, int flags)
1344 {
1345         int write = flags & FUSE_DIO_WRITE;
1346         int cuse = flags & FUSE_DIO_CUSE;
1347         struct file *file = io->file;
1348         struct inode *inode = file->f_mapping->host;
1349         struct fuse_file *ff = file->private_data;
1350         struct fuse_conn *fc = ff->fc;
1351         size_t nmax = write ? fc->max_write : fc->max_read;
1352         loff_t pos = *ppos;
1353         size_t count = iov_iter_count(iter);
1354         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1355         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1356         ssize_t res = 0;
1357         struct fuse_req *req;
1358
1359         if (io->async)
1360                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1361         else
1362                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1363         if (IS_ERR(req))
1364                 return PTR_ERR(req);
1365
1366         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1367                 if (!write)
1368                         mutex_lock(&inode->i_mutex);
1369                 fuse_sync_writes(inode);
1370                 if (!write)
1371                         mutex_unlock(&inode->i_mutex);
1372         }
1373
1374         while (count) {
1375                 size_t nres;
1376                 fl_owner_t owner = current->files;
1377                 size_t nbytes = min(count, nmax);
1378                 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1379                 if (err) {
1380                         res = err;
1381                         break;
1382                 }
1383
1384                 if (write)
1385                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1386                 else
1387                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1388
1389                 if (!io->async)
1390                         fuse_release_user_pages(req, !write);
1391                 if (req->out.h.error) {
1392                         if (!res)
1393                                 res = req->out.h.error;
1394                         break;
1395                 } else if (nres > nbytes) {
1396                         res = -EIO;
1397                         break;
1398                 }
1399                 count -= nres;
1400                 res += nres;
1401                 pos += nres;
1402                 if (nres != nbytes)
1403                         break;
1404                 if (count) {
1405                         fuse_put_request(fc, req);
1406                         if (io->async)
1407                                 req = fuse_get_req_for_background(fc,
1408                                         fuse_iter_npages(iter));
1409                         else
1410                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1411                         if (IS_ERR(req))
1412                                 break;
1413                 }
1414         }
1415         if (!IS_ERR(req))
1416                 fuse_put_request(fc, req);
1417         if (res > 0)
1418                 *ppos = pos;
1419
1420         return res;
1421 }
1422 EXPORT_SYMBOL_GPL(fuse_direct_io);
1423
1424 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1425                                   struct iov_iter *iter,
1426                                   loff_t *ppos)
1427 {
1428         ssize_t res;
1429         struct file *file = io->file;
1430         struct inode *inode = file_inode(file);
1431
1432         if (is_bad_inode(inode))
1433                 return -EIO;
1434
1435         res = fuse_direct_io(io, iter, ppos, 0);
1436
1437         fuse_invalidate_attr(inode);
1438
1439         return res;
1440 }
1441
1442 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1443                                      size_t count, loff_t *ppos)
1444 {
1445         struct fuse_io_priv io = { .async = 0, .file = file };
1446         struct iovec iov = { .iov_base = buf, .iov_len = count };
1447         struct iov_iter ii;
1448         iov_iter_init(&ii, READ, &iov, 1, count);
1449         return __fuse_direct_read(&io, &ii, ppos);
1450 }
1451
1452 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1453                                    struct iov_iter *iter,
1454                                    loff_t *ppos)
1455 {
1456         struct file *file = io->file;
1457         struct inode *inode = file_inode(file);
1458         size_t count = iov_iter_count(iter);
1459         ssize_t res;
1460
1461
1462         res = generic_write_checks(file, ppos, &count, 0);
1463         if (!res) {
1464                 iov_iter_truncate(iter, count);
1465                 res = fuse_direct_io(io, iter, ppos, FUSE_DIO_WRITE);
1466         }
1467
1468         fuse_invalidate_attr(inode);
1469
1470         return res;
1471 }
1472
1473 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1474                                  size_t count, loff_t *ppos)
1475 {
1476         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1477         struct inode *inode = file_inode(file);
1478         ssize_t res;
1479         struct fuse_io_priv io = { .async = 0, .file = file };
1480         struct iov_iter ii;
1481         iov_iter_init(&ii, WRITE, &iov, 1, count);
1482
1483         if (is_bad_inode(inode))
1484                 return -EIO;
1485
1486         /* Don't allow parallel writes to the same file */
1487         mutex_lock(&inode->i_mutex);
1488         res = __fuse_direct_write(&io, &ii, ppos);
1489         if (res > 0)
1490                 fuse_write_update_size(inode, *ppos);
1491         mutex_unlock(&inode->i_mutex);
1492
1493         return res;
1494 }
1495
1496 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1497 {
1498         int i;
1499
1500         for (i = 0; i < req->num_pages; i++)
1501                 __free_page(req->pages[i]);
1502
1503         if (req->ff)
1504                 fuse_file_put(req->ff, false);
1505 }
1506
1507 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1508 {
1509         struct inode *inode = req->inode;
1510         struct fuse_inode *fi = get_fuse_inode(inode);
1511         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1512         int i;
1513
1514         list_del(&req->writepages_entry);
1515         for (i = 0; i < req->num_pages; i++) {
1516                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1517                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1518                 bdi_writeout_inc(bdi);
1519         }
1520         wake_up(&fi->page_waitq);
1521 }
1522
1523 /* Called under fc->lock, may release and reacquire it */
1524 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1525                                 loff_t size)
1526 __releases(fc->lock)
1527 __acquires(fc->lock)
1528 {
1529         struct fuse_inode *fi = get_fuse_inode(req->inode);
1530         struct fuse_write_in *inarg = &req->misc.write.in;
1531         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1532
1533         if (!fc->connected)
1534                 goto out_free;
1535
1536         if (inarg->offset + data_size <= size) {
1537                 inarg->size = data_size;
1538         } else if (inarg->offset < size) {
1539                 inarg->size = size - inarg->offset;
1540         } else {
1541                 /* Got truncated off completely */
1542                 goto out_free;
1543         }
1544
1545         req->in.args[1].size = inarg->size;
1546         fi->writectr++;
1547         fuse_request_send_background_locked(fc, req);
1548         return;
1549
1550  out_free:
1551         fuse_writepage_finish(fc, req);
1552         spin_unlock(&fc->lock);
1553         fuse_writepage_free(fc, req);
1554         fuse_put_request(fc, req);
1555         spin_lock(&fc->lock);
1556 }
1557
1558 /*
1559  * If fi->writectr is positive (no truncate or fsync going on) send
1560  * all queued writepage requests.
1561  *
1562  * Called with fc->lock
1563  */
1564 void fuse_flush_writepages(struct inode *inode)
1565 __releases(fc->lock)
1566 __acquires(fc->lock)
1567 {
1568         struct fuse_conn *fc = get_fuse_conn(inode);
1569         struct fuse_inode *fi = get_fuse_inode(inode);
1570         size_t crop = i_size_read(inode);
1571         struct fuse_req *req;
1572
1573         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1574                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1575                 list_del_init(&req->list);
1576                 fuse_send_writepage(fc, req, crop);
1577         }
1578 }
1579
1580 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1581 {
1582         struct inode *inode = req->inode;
1583         struct fuse_inode *fi = get_fuse_inode(inode);
1584
1585         mapping_set_error(inode->i_mapping, req->out.h.error);
1586         spin_lock(&fc->lock);
1587         while (req->misc.write.next) {
1588                 struct fuse_conn *fc = get_fuse_conn(inode);
1589                 struct fuse_write_in *inarg = &req->misc.write.in;
1590                 struct fuse_req *next = req->misc.write.next;
1591                 req->misc.write.next = next->misc.write.next;
1592                 next->misc.write.next = NULL;
1593                 next->ff = fuse_file_get(req->ff);
1594                 list_add(&next->writepages_entry, &fi->writepages);
1595
1596                 /*
1597                  * Skip fuse_flush_writepages() to make it easy to crop requests
1598                  * based on primary request size.
1599                  *
1600                  * 1st case (trivial): there are no concurrent activities using
1601                  * fuse_set/release_nowrite.  Then we're on safe side because
1602                  * fuse_flush_writepages() would call fuse_send_writepage()
1603                  * anyway.
1604                  *
1605                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1606                  * now for completion of all in-flight requests.  This happens
1607                  * rarely and no more than once per page, so this should be
1608                  * okay.
1609                  *
1610                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1611                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1612                  * that fuse_set_nowrite returned implies that all in-flight
1613                  * requests were completed along with all of their secondary
1614                  * requests.  Further primary requests are blocked by negative
1615                  * writectr.  Hence there cannot be any in-flight requests and
1616                  * no invocations of fuse_writepage_end() while we're in
1617                  * fuse_set_nowrite..fuse_release_nowrite section.
1618                  */
1619                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1620         }
1621         fi->writectr--;
1622         fuse_writepage_finish(fc, req);
1623         spin_unlock(&fc->lock);
1624         fuse_writepage_free(fc, req);
1625 }
1626
1627 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1628                                                struct fuse_inode *fi)
1629 {
1630         struct fuse_file *ff = NULL;
1631
1632         spin_lock(&fc->lock);
1633         if (!list_empty(&fi->write_files)) {
1634                 ff = list_entry(fi->write_files.next, struct fuse_file,
1635                                 write_entry);
1636                 fuse_file_get(ff);
1637         }
1638         spin_unlock(&fc->lock);
1639
1640         return ff;
1641 }
1642
1643 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1644                                              struct fuse_inode *fi)
1645 {
1646         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1647         WARN_ON(!ff);
1648         return ff;
1649 }
1650
1651 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1652 {
1653         struct fuse_conn *fc = get_fuse_conn(inode);
1654         struct fuse_inode *fi = get_fuse_inode(inode);
1655         struct fuse_file *ff;
1656         int err;
1657
1658         ff = __fuse_write_file_get(fc, fi);
1659         err = fuse_flush_times(inode, ff);
1660         if (ff)
1661                 fuse_file_put(ff, 0);
1662
1663         return err;
1664 }
1665
1666 static int fuse_writepage_locked(struct page *page)
1667 {
1668         struct address_space *mapping = page->mapping;
1669         struct inode *inode = mapping->host;
1670         struct fuse_conn *fc = get_fuse_conn(inode);
1671         struct fuse_inode *fi = get_fuse_inode(inode);
1672         struct fuse_req *req;
1673         struct page *tmp_page;
1674         int error = -ENOMEM;
1675
1676         set_page_writeback(page);
1677
1678         req = fuse_request_alloc_nofs(1);
1679         if (!req)
1680                 goto err;
1681
1682         req->background = 1; /* writeback always goes to bg_queue */
1683         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1684         if (!tmp_page)
1685                 goto err_free;
1686
1687         error = -EIO;
1688         req->ff = fuse_write_file_get(fc, fi);
1689         if (!req->ff)
1690                 goto err_free;
1691
1692         fuse_write_fill(req, req->ff, page_offset(page), 0);
1693
1694         copy_highpage(tmp_page, page);
1695         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1696         req->misc.write.next = NULL;
1697         req->in.argpages = 1;
1698         req->num_pages = 1;
1699         req->pages[0] = tmp_page;
1700         req->page_descs[0].offset = 0;
1701         req->page_descs[0].length = PAGE_SIZE;
1702         req->end = fuse_writepage_end;
1703         req->inode = inode;
1704
1705         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1706         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1707
1708         spin_lock(&fc->lock);
1709         list_add(&req->writepages_entry, &fi->writepages);
1710         list_add_tail(&req->list, &fi->queued_writes);
1711         fuse_flush_writepages(inode);
1712         spin_unlock(&fc->lock);
1713
1714         end_page_writeback(page);
1715
1716         return 0;
1717
1718 err_free:
1719         fuse_request_free(req);
1720 err:
1721         end_page_writeback(page);
1722         return error;
1723 }
1724
1725 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1726 {
1727         int err;
1728
1729         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1730                 /*
1731                  * ->writepages() should be called for sync() and friends.  We
1732                  * should only get here on direct reclaim and then we are
1733                  * allowed to skip a page which is already in flight
1734                  */
1735                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1736
1737                 redirty_page_for_writepage(wbc, page);
1738                 return 0;
1739         }
1740
1741         err = fuse_writepage_locked(page);
1742         unlock_page(page);
1743
1744         return err;
1745 }
1746
1747 struct fuse_fill_wb_data {
1748         struct fuse_req *req;
1749         struct fuse_file *ff;
1750         struct inode *inode;
1751         struct page **orig_pages;
1752 };
1753
1754 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1755 {
1756         struct fuse_req *req = data->req;
1757         struct inode *inode = data->inode;
1758         struct fuse_conn *fc = get_fuse_conn(inode);
1759         struct fuse_inode *fi = get_fuse_inode(inode);
1760         int num_pages = req->num_pages;
1761         int i;
1762
1763         req->ff = fuse_file_get(data->ff);
1764         spin_lock(&fc->lock);
1765         list_add_tail(&req->list, &fi->queued_writes);
1766         fuse_flush_writepages(inode);
1767         spin_unlock(&fc->lock);
1768
1769         for (i = 0; i < num_pages; i++)
1770                 end_page_writeback(data->orig_pages[i]);
1771 }
1772
1773 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1774                                      struct page *page)
1775 {
1776         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1777         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1778         struct fuse_req *tmp;
1779         struct fuse_req *old_req;
1780         bool found = false;
1781         pgoff_t curr_index;
1782
1783         BUG_ON(new_req->num_pages != 0);
1784
1785         spin_lock(&fc->lock);
1786         list_del(&new_req->writepages_entry);
1787         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1788                 BUG_ON(old_req->inode != new_req->inode);
1789                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1790                 if (curr_index <= page->index &&
1791                     page->index < curr_index + old_req->num_pages) {
1792                         found = true;
1793                         break;
1794                 }
1795         }
1796         if (!found) {
1797                 list_add(&new_req->writepages_entry, &fi->writepages);
1798                 goto out_unlock;
1799         }
1800
1801         new_req->num_pages = 1;
1802         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1803                 BUG_ON(tmp->inode != new_req->inode);
1804                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1805                 if (tmp->num_pages == 1 &&
1806                     curr_index == page->index) {
1807                         old_req = tmp;
1808                 }
1809         }
1810
1811         if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1812                                         old_req->state == FUSE_REQ_PENDING)) {
1813                 struct backing_dev_info *bdi = page->mapping->backing_dev_info;
1814
1815                 copy_highpage(old_req->pages[0], page);
1816                 spin_unlock(&fc->lock);
1817
1818                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1819                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1820                 bdi_writeout_inc(bdi);
1821                 fuse_writepage_free(fc, new_req);
1822                 fuse_request_free(new_req);
1823                 goto out;
1824         } else {
1825                 new_req->misc.write.next = old_req->misc.write.next;
1826                 old_req->misc.write.next = new_req;
1827         }
1828 out_unlock:
1829         spin_unlock(&fc->lock);
1830 out:
1831         return found;
1832 }
1833
1834 static int fuse_writepages_fill(struct page *page,
1835                 struct writeback_control *wbc, void *_data)
1836 {
1837         struct fuse_fill_wb_data *data = _data;
1838         struct fuse_req *req = data->req;
1839         struct inode *inode = data->inode;
1840         struct fuse_conn *fc = get_fuse_conn(inode);
1841         struct page *tmp_page;
1842         bool is_writeback;
1843         int err;
1844
1845         if (!data->ff) {
1846                 err = -EIO;
1847                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1848                 if (!data->ff)
1849                         goto out_unlock;
1850         }
1851
1852         /*
1853          * Being under writeback is unlikely but possible.  For example direct
1854          * read to an mmaped fuse file will set the page dirty twice; once when
1855          * the pages are faulted with get_user_pages(), and then after the read
1856          * completed.
1857          */
1858         is_writeback = fuse_page_is_writeback(inode, page->index);
1859
1860         if (req && req->num_pages &&
1861             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1862              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1863              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1864                 fuse_writepages_send(data);
1865                 data->req = NULL;
1866         }
1867         err = -ENOMEM;
1868         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1869         if (!tmp_page)
1870                 goto out_unlock;
1871
1872         /*
1873          * The page must not be redirtied until the writeout is completed
1874          * (i.e. userspace has sent a reply to the write request).  Otherwise
1875          * there could be more than one temporary page instance for each real
1876          * page.
1877          *
1878          * This is ensured by holding the page lock in page_mkwrite() while
1879          * checking fuse_page_is_writeback().  We already hold the page lock
1880          * since clear_page_dirty_for_io() and keep it held until we add the
1881          * request to the fi->writepages list and increment req->num_pages.
1882          * After this fuse_page_is_writeback() will indicate that the page is
1883          * under writeback, so we can release the page lock.
1884          */
1885         if (data->req == NULL) {
1886                 struct fuse_inode *fi = get_fuse_inode(inode);
1887
1888                 err = -ENOMEM;
1889                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1890                 if (!req) {
1891                         __free_page(tmp_page);
1892                         goto out_unlock;
1893                 }
1894
1895                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1896                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1897                 req->misc.write.next = NULL;
1898                 req->in.argpages = 1;
1899                 req->background = 1;
1900                 req->num_pages = 0;
1901                 req->end = fuse_writepage_end;
1902                 req->inode = inode;
1903
1904                 spin_lock(&fc->lock);
1905                 list_add(&req->writepages_entry, &fi->writepages);
1906                 spin_unlock(&fc->lock);
1907
1908                 data->req = req;
1909         }
1910         set_page_writeback(page);
1911
1912         copy_highpage(tmp_page, page);
1913         req->pages[req->num_pages] = tmp_page;
1914         req->page_descs[req->num_pages].offset = 0;
1915         req->page_descs[req->num_pages].length = PAGE_SIZE;
1916
1917         inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
1918         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1919
1920         err = 0;
1921         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1922                 end_page_writeback(page);
1923                 data->req = NULL;
1924                 goto out_unlock;
1925         }
1926         data->orig_pages[req->num_pages] = page;
1927
1928         /*
1929          * Protected by fc->lock against concurrent access by
1930          * fuse_page_is_writeback().
1931          */
1932         spin_lock(&fc->lock);
1933         req->num_pages++;
1934         spin_unlock(&fc->lock);
1935
1936 out_unlock:
1937         unlock_page(page);
1938
1939         return err;
1940 }
1941
1942 static int fuse_writepages(struct address_space *mapping,
1943                            struct writeback_control *wbc)
1944 {
1945         struct inode *inode = mapping->host;
1946         struct fuse_fill_wb_data data;
1947         int err;
1948
1949         err = -EIO;
1950         if (is_bad_inode(inode))
1951                 goto out;
1952
1953         data.inode = inode;
1954         data.req = NULL;
1955         data.ff = NULL;
1956
1957         err = -ENOMEM;
1958         data.orig_pages = kzalloc(sizeof(struct page *) *
1959                                   FUSE_MAX_PAGES_PER_REQ,
1960                                   GFP_NOFS);
1961         if (!data.orig_pages)
1962                 goto out;
1963
1964         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1965         if (data.req) {
1966                 /* Ignore errors if we can write at least one page */
1967                 BUG_ON(!data.req->num_pages);
1968                 fuse_writepages_send(&data);
1969                 err = 0;
1970         }
1971         if (data.ff)
1972                 fuse_file_put(data.ff, false);
1973
1974         kfree(data.orig_pages);
1975 out:
1976         return err;
1977 }
1978
1979 /*
1980  * It's worthy to make sure that space is reserved on disk for the write,
1981  * but how to implement it without killing performance need more thinking.
1982  */
1983 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1984                 loff_t pos, unsigned len, unsigned flags,
1985                 struct page **pagep, void **fsdata)
1986 {
1987         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1988         struct fuse_conn *fc = get_fuse_conn(file->f_dentry->d_inode);
1989         struct page *page;
1990         loff_t fsize;
1991         int err = -ENOMEM;
1992
1993         WARN_ON(!fc->writeback_cache);
1994
1995         page = grab_cache_page_write_begin(mapping, index, flags);
1996         if (!page)
1997                 goto error;
1998
1999         fuse_wait_on_page_writeback(mapping->host, page->index);
2000
2001         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
2002                 goto success;
2003         /*
2004          * Check if the start this page comes after the end of file, in which
2005          * case the readpage can be optimized away.
2006          */
2007         fsize = i_size_read(mapping->host);
2008         if (fsize <= (pos & PAGE_CACHE_MASK)) {
2009                 size_t off = pos & ~PAGE_CACHE_MASK;
2010                 if (off)
2011                         zero_user_segment(page, 0, off);
2012                 goto success;
2013         }
2014         err = fuse_do_readpage(file, page);
2015         if (err)
2016                 goto cleanup;
2017 success:
2018         *pagep = page;
2019         return 0;
2020
2021 cleanup:
2022         unlock_page(page);
2023         page_cache_release(page);
2024 error:
2025         return err;
2026 }
2027
2028 static int fuse_write_end(struct file *file, struct address_space *mapping,
2029                 loff_t pos, unsigned len, unsigned copied,
2030                 struct page *page, void *fsdata)
2031 {
2032         struct inode *inode = page->mapping->host;
2033
2034         if (!PageUptodate(page)) {
2035                 /* Zero any unwritten bytes at the end of the page */
2036                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2037                 if (endoff)
2038                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2039                 SetPageUptodate(page);
2040         }
2041
2042         fuse_write_update_size(inode, pos + copied);
2043         set_page_dirty(page);
2044         unlock_page(page);
2045         page_cache_release(page);
2046
2047         return copied;
2048 }
2049
2050 static int fuse_launder_page(struct page *page)
2051 {
2052         int err = 0;
2053         if (clear_page_dirty_for_io(page)) {
2054                 struct inode *inode = page->mapping->host;
2055                 err = fuse_writepage_locked(page);
2056                 if (!err)
2057                         fuse_wait_on_page_writeback(inode, page->index);
2058         }
2059         return err;
2060 }
2061
2062 /*
2063  * Write back dirty pages now, because there may not be any suitable
2064  * open files later
2065  */
2066 static void fuse_vma_close(struct vm_area_struct *vma)
2067 {
2068         filemap_write_and_wait(vma->vm_file->f_mapping);
2069 }
2070
2071 /*
2072  * Wait for writeback against this page to complete before allowing it
2073  * to be marked dirty again, and hence written back again, possibly
2074  * before the previous writepage completed.
2075  *
2076  * Block here, instead of in ->writepage(), so that the userspace fs
2077  * can only block processes actually operating on the filesystem.
2078  *
2079  * Otherwise unprivileged userspace fs would be able to block
2080  * unrelated:
2081  *
2082  * - page migration
2083  * - sync(2)
2084  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2085  */
2086 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2087 {
2088         struct page *page = vmf->page;
2089         struct inode *inode = file_inode(vma->vm_file);
2090
2091         file_update_time(vma->vm_file);
2092         lock_page(page);
2093         if (page->mapping != inode->i_mapping) {
2094                 unlock_page(page);
2095                 return VM_FAULT_NOPAGE;
2096         }
2097
2098         fuse_wait_on_page_writeback(inode, page->index);
2099         return VM_FAULT_LOCKED;
2100 }
2101
2102 static const struct vm_operations_struct fuse_file_vm_ops = {
2103         .close          = fuse_vma_close,
2104         .fault          = filemap_fault,
2105         .map_pages      = filemap_map_pages,
2106         .page_mkwrite   = fuse_page_mkwrite,
2107         .remap_pages    = generic_file_remap_pages,
2108 };
2109
2110 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2111 {
2112         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2113                 fuse_link_write_file(file);
2114
2115         file_accessed(file);
2116         vma->vm_ops = &fuse_file_vm_ops;
2117         return 0;
2118 }
2119
2120 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2121 {
2122         /* Can't provide the coherency needed for MAP_SHARED */
2123         if (vma->vm_flags & VM_MAYSHARE)
2124                 return -ENODEV;
2125
2126         invalidate_inode_pages2(file->f_mapping);
2127
2128         return generic_file_mmap(file, vma);
2129 }
2130
2131 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2132                                   struct file_lock *fl)
2133 {
2134         switch (ffl->type) {
2135         case F_UNLCK:
2136                 break;
2137
2138         case F_RDLCK:
2139         case F_WRLCK:
2140                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2141                     ffl->end < ffl->start)
2142                         return -EIO;
2143
2144                 fl->fl_start = ffl->start;
2145                 fl->fl_end = ffl->end;
2146                 fl->fl_pid = ffl->pid;
2147                 break;
2148
2149         default:
2150                 return -EIO;
2151         }
2152         fl->fl_type = ffl->type;
2153         return 0;
2154 }
2155
2156 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
2157                          const struct file_lock *fl, int opcode, pid_t pid,
2158                          int flock)
2159 {
2160         struct inode *inode = file_inode(file);
2161         struct fuse_conn *fc = get_fuse_conn(inode);
2162         struct fuse_file *ff = file->private_data;
2163         struct fuse_lk_in *arg = &req->misc.lk_in;
2164
2165         arg->fh = ff->fh;
2166         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2167         arg->lk.start = fl->fl_start;
2168         arg->lk.end = fl->fl_end;
2169         arg->lk.type = fl->fl_type;
2170         arg->lk.pid = pid;
2171         if (flock)
2172                 arg->lk_flags |= FUSE_LK_FLOCK;
2173         req->in.h.opcode = opcode;
2174         req->in.h.nodeid = get_node_id(inode);
2175         req->in.numargs = 1;
2176         req->in.args[0].size = sizeof(*arg);
2177         req->in.args[0].value = arg;
2178 }
2179
2180 static int fuse_getlk(struct file *file, struct file_lock *fl)
2181 {
2182         struct inode *inode = file_inode(file);
2183         struct fuse_conn *fc = get_fuse_conn(inode);
2184         struct fuse_req *req;
2185         struct fuse_lk_out outarg;
2186         int err;
2187
2188         req = fuse_get_req_nopages(fc);
2189         if (IS_ERR(req))
2190                 return PTR_ERR(req);
2191
2192         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
2193         req->out.numargs = 1;
2194         req->out.args[0].size = sizeof(outarg);
2195         req->out.args[0].value = &outarg;
2196         fuse_request_send(fc, req);
2197         err = req->out.h.error;
2198         fuse_put_request(fc, req);
2199         if (!err)
2200                 err = convert_fuse_file_lock(&outarg.lk, fl);
2201
2202         return err;
2203 }
2204
2205 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2206 {
2207         struct inode *inode = file_inode(file);
2208         struct fuse_conn *fc = get_fuse_conn(inode);
2209         struct fuse_req *req;
2210         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2211         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2212         int err;
2213
2214         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2215                 /* NLM needs asynchronous locks, which we don't support yet */
2216                 return -ENOLCK;
2217         }
2218
2219         /* Unlock on close is handled by the flush method */
2220         if (fl->fl_flags & FL_CLOSE)
2221                 return 0;
2222
2223         req = fuse_get_req_nopages(fc);
2224         if (IS_ERR(req))
2225                 return PTR_ERR(req);
2226
2227         fuse_lk_fill(req, file, fl, opcode, pid, flock);
2228         fuse_request_send(fc, req);
2229         err = req->out.h.error;
2230         /* locking is restartable */
2231         if (err == -EINTR)
2232                 err = -ERESTARTSYS;
2233         fuse_put_request(fc, req);
2234         return err;
2235 }
2236
2237 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2238 {
2239         struct inode *inode = file_inode(file);
2240         struct fuse_conn *fc = get_fuse_conn(inode);
2241         int err;
2242
2243         if (cmd == F_CANCELLK) {
2244                 err = 0;
2245         } else if (cmd == F_GETLK) {
2246                 if (fc->no_lock) {
2247                         posix_test_lock(file, fl);
2248                         err = 0;
2249                 } else
2250                         err = fuse_getlk(file, fl);
2251         } else {
2252                 if (fc->no_lock)
2253                         err = posix_lock_file(file, fl, NULL);
2254                 else
2255                         err = fuse_setlk(file, fl, 0);
2256         }
2257         return err;
2258 }
2259
2260 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2261 {
2262         struct inode *inode = file_inode(file);
2263         struct fuse_conn *fc = get_fuse_conn(inode);
2264         int err;
2265
2266         if (fc->no_flock) {
2267                 err = flock_lock_file_wait(file, fl);
2268         } else {
2269                 struct fuse_file *ff = file->private_data;
2270
2271                 /* emulate flock with POSIX locks */
2272                 ff->flock = true;
2273                 err = fuse_setlk(file, fl, 1);
2274         }
2275
2276         return err;
2277 }
2278
2279 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2280 {
2281         struct inode *inode = mapping->host;
2282         struct fuse_conn *fc = get_fuse_conn(inode);
2283         struct fuse_req *req;
2284         struct fuse_bmap_in inarg;
2285         struct fuse_bmap_out outarg;
2286         int err;
2287
2288         if (!inode->i_sb->s_bdev || fc->no_bmap)
2289                 return 0;
2290
2291         req = fuse_get_req_nopages(fc);
2292         if (IS_ERR(req))
2293                 return 0;
2294
2295         memset(&inarg, 0, sizeof(inarg));
2296         inarg.block = block;
2297         inarg.blocksize = inode->i_sb->s_blocksize;
2298         req->in.h.opcode = FUSE_BMAP;
2299         req->in.h.nodeid = get_node_id(inode);
2300         req->in.numargs = 1;
2301         req->in.args[0].size = sizeof(inarg);
2302         req->in.args[0].value = &inarg;
2303         req->out.numargs = 1;
2304         req->out.args[0].size = sizeof(outarg);
2305         req->out.args[0].value = &outarg;
2306         fuse_request_send(fc, req);
2307         err = req->out.h.error;
2308         fuse_put_request(fc, req);
2309         if (err == -ENOSYS)
2310                 fc->no_bmap = 1;
2311
2312         return err ? 0 : outarg.block;
2313 }
2314
2315 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2316 {
2317         loff_t retval;
2318         struct inode *inode = file_inode(file);
2319
2320         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2321         if (whence == SEEK_CUR || whence == SEEK_SET)
2322                 return generic_file_llseek(file, offset, whence);
2323
2324         mutex_lock(&inode->i_mutex);
2325         retval = fuse_update_attributes(inode, NULL, file, NULL);
2326         if (!retval)
2327                 retval = generic_file_llseek(file, offset, whence);
2328         mutex_unlock(&inode->i_mutex);
2329
2330         return retval;
2331 }
2332
2333 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2334                         unsigned int nr_segs, size_t bytes, bool to_user)
2335 {
2336         struct iov_iter ii;
2337         int page_idx = 0;
2338
2339         if (!bytes)
2340                 return 0;
2341
2342         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2343
2344         while (iov_iter_count(&ii)) {
2345                 struct page *page = pages[page_idx++];
2346                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2347                 void *kaddr;
2348
2349                 kaddr = kmap(page);
2350
2351                 while (todo) {
2352                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2353                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2354                         size_t copy = min(todo, iov_len);
2355                         size_t left;
2356
2357                         if (!to_user)
2358                                 left = copy_from_user(kaddr, uaddr, copy);
2359                         else
2360                                 left = copy_to_user(uaddr, kaddr, copy);
2361
2362                         if (unlikely(left))
2363                                 return -EFAULT;
2364
2365                         iov_iter_advance(&ii, copy);
2366                         todo -= copy;
2367                         kaddr += copy;
2368                 }
2369
2370                 kunmap(page);
2371         }
2372
2373         return 0;
2374 }
2375
2376 /*
2377  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2378  * ABI was defined to be 'struct iovec' which is different on 32bit
2379  * and 64bit.  Fortunately we can determine which structure the server
2380  * used from the size of the reply.
2381  */
2382 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2383                                      size_t transferred, unsigned count,
2384                                      bool is_compat)
2385 {
2386 #ifdef CONFIG_COMPAT
2387         if (count * sizeof(struct compat_iovec) == transferred) {
2388                 struct compat_iovec *ciov = src;
2389                 unsigned i;
2390
2391                 /*
2392                  * With this interface a 32bit server cannot support
2393                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2394                  * requests
2395                  */
2396                 if (!is_compat)
2397                         return -EINVAL;
2398
2399                 for (i = 0; i < count; i++) {
2400                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2401                         dst[i].iov_len = ciov[i].iov_len;
2402                 }
2403                 return 0;
2404         }
2405 #endif
2406
2407         if (count * sizeof(struct iovec) != transferred)
2408                 return -EIO;
2409
2410         memcpy(dst, src, transferred);
2411         return 0;
2412 }
2413
2414 /* Make sure iov_length() won't overflow */
2415 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2416 {
2417         size_t n;
2418         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2419
2420         for (n = 0; n < count; n++, iov++) {
2421                 if (iov->iov_len > (size_t) max)
2422                         return -ENOMEM;
2423                 max -= iov->iov_len;
2424         }
2425         return 0;
2426 }
2427
2428 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2429                                  void *src, size_t transferred, unsigned count,
2430                                  bool is_compat)
2431 {
2432         unsigned i;
2433         struct fuse_ioctl_iovec *fiov = src;
2434
2435         if (fc->minor < 16) {
2436                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2437                                                  count, is_compat);
2438         }
2439
2440         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2441                 return -EIO;
2442
2443         for (i = 0; i < count; i++) {
2444                 /* Did the server supply an inappropriate value? */
2445                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2446                     fiov[i].len != (unsigned long) fiov[i].len)
2447                         return -EIO;
2448
2449                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2450                 dst[i].iov_len = (size_t) fiov[i].len;
2451
2452 #ifdef CONFIG_COMPAT
2453                 if (is_compat &&
2454                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2455                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2456                         return -EIO;
2457 #endif
2458         }
2459
2460         return 0;
2461 }
2462
2463
2464 /*
2465  * For ioctls, there is no generic way to determine how much memory
2466  * needs to be read and/or written.  Furthermore, ioctls are allowed
2467  * to dereference the passed pointer, so the parameter requires deep
2468  * copying but FUSE has no idea whatsoever about what to copy in or
2469  * out.
2470  *
2471  * This is solved by allowing FUSE server to retry ioctl with
2472  * necessary in/out iovecs.  Let's assume the ioctl implementation
2473  * needs to read in the following structure.
2474  *
2475  * struct a {
2476  *      char    *buf;
2477  *      size_t  buflen;
2478  * }
2479  *
2480  * On the first callout to FUSE server, inarg->in_size and
2481  * inarg->out_size will be NULL; then, the server completes the ioctl
2482  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2483  * the actual iov array to
2484  *
2485  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2486  *
2487  * which tells FUSE to copy in the requested area and retry the ioctl.
2488  * On the second round, the server has access to the structure and
2489  * from that it can tell what to look for next, so on the invocation,
2490  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2491  *
2492  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2493  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2494  *
2495  * FUSE will copy both struct a and the pointed buffer from the
2496  * process doing the ioctl and retry ioctl with both struct a and the
2497  * buffer.
2498  *
2499  * This time, FUSE server has everything it needs and completes ioctl
2500  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2501  *
2502  * Copying data out works the same way.
2503  *
2504  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2505  * automatically initializes in and out iovs by decoding @cmd with
2506  * _IOC_* macros and the server is not allowed to request RETRY.  This
2507  * limits ioctl data transfers to well-formed ioctls and is the forced
2508  * behavior for all FUSE servers.
2509  */
2510 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2511                    unsigned int flags)
2512 {
2513         struct fuse_file *ff = file->private_data;
2514         struct fuse_conn *fc = ff->fc;
2515         struct fuse_ioctl_in inarg = {
2516                 .fh = ff->fh,
2517                 .cmd = cmd,
2518                 .arg = arg,
2519                 .flags = flags
2520         };
2521         struct fuse_ioctl_out outarg;
2522         struct fuse_req *req = NULL;
2523         struct page **pages = NULL;
2524         struct iovec *iov_page = NULL;
2525         struct iovec *in_iov = NULL, *out_iov = NULL;
2526         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2527         size_t in_size, out_size, transferred;
2528         int err;
2529
2530 #if BITS_PER_LONG == 32
2531         inarg.flags |= FUSE_IOCTL_32BIT;
2532 #else
2533         if (flags & FUSE_IOCTL_COMPAT)
2534                 inarg.flags |= FUSE_IOCTL_32BIT;
2535 #endif
2536
2537         /* assume all the iovs returned by client always fits in a page */
2538         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2539
2540         err = -ENOMEM;
2541         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2542         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2543         if (!pages || !iov_page)
2544                 goto out;
2545
2546         /*
2547          * If restricted, initialize IO parameters as encoded in @cmd.
2548          * RETRY from server is not allowed.
2549          */
2550         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2551                 struct iovec *iov = iov_page;
2552
2553                 iov->iov_base = (void __user *)arg;
2554                 iov->iov_len = _IOC_SIZE(cmd);
2555
2556                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2557                         in_iov = iov;
2558                         in_iovs = 1;
2559                 }
2560
2561                 if (_IOC_DIR(cmd) & _IOC_READ) {
2562                         out_iov = iov;
2563                         out_iovs = 1;
2564                 }
2565         }
2566
2567  retry:
2568         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2569         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2570
2571         /*
2572          * Out data can be used either for actual out data or iovs,
2573          * make sure there always is at least one page.
2574          */
2575         out_size = max_t(size_t, out_size, PAGE_SIZE);
2576         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2577
2578         /* make sure there are enough buffer pages and init request with them */
2579         err = -ENOMEM;
2580         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2581                 goto out;
2582         while (num_pages < max_pages) {
2583                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2584                 if (!pages[num_pages])
2585                         goto out;
2586                 num_pages++;
2587         }
2588
2589         req = fuse_get_req(fc, num_pages);
2590         if (IS_ERR(req)) {
2591                 err = PTR_ERR(req);
2592                 req = NULL;
2593                 goto out;
2594         }
2595         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2596         req->num_pages = num_pages;
2597         fuse_page_descs_length_init(req, 0, req->num_pages);
2598
2599         /* okay, let's send it to the client */
2600         req->in.h.opcode = FUSE_IOCTL;
2601         req->in.h.nodeid = ff->nodeid;
2602         req->in.numargs = 1;
2603         req->in.args[0].size = sizeof(inarg);
2604         req->in.args[0].value = &inarg;
2605         if (in_size) {
2606                 req->in.numargs++;
2607                 req->in.args[1].size = in_size;
2608                 req->in.argpages = 1;
2609
2610                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2611                                            false);
2612                 if (err)
2613                         goto out;
2614         }
2615
2616         req->out.numargs = 2;
2617         req->out.args[0].size = sizeof(outarg);
2618         req->out.args[0].value = &outarg;
2619         req->out.args[1].size = out_size;
2620         req->out.argpages = 1;
2621         req->out.argvar = 1;
2622
2623         fuse_request_send(fc, req);
2624         err = req->out.h.error;
2625         transferred = req->out.args[1].size;
2626         fuse_put_request(fc, req);
2627         req = NULL;
2628         if (err)
2629                 goto out;
2630
2631         /* did it ask for retry? */
2632         if (outarg.flags & FUSE_IOCTL_RETRY) {
2633                 void *vaddr;
2634
2635                 /* no retry if in restricted mode */
2636                 err = -EIO;
2637                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2638                         goto out;
2639
2640                 in_iovs = outarg.in_iovs;
2641                 out_iovs = outarg.out_iovs;
2642
2643                 /*
2644                  * Make sure things are in boundary, separate checks
2645                  * are to protect against overflow.
2646                  */
2647                 err = -ENOMEM;
2648                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2649                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2650                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2651                         goto out;
2652
2653                 vaddr = kmap_atomic(pages[0]);
2654                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2655                                             transferred, in_iovs + out_iovs,
2656                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2657                 kunmap_atomic(vaddr);
2658                 if (err)
2659                         goto out;
2660
2661                 in_iov = iov_page;
2662                 out_iov = in_iov + in_iovs;
2663
2664                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2665                 if (err)
2666                         goto out;
2667
2668                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2669                 if (err)
2670                         goto out;
2671
2672                 goto retry;
2673         }
2674
2675         err = -EIO;
2676         if (transferred > inarg.out_size)
2677                 goto out;
2678
2679         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2680  out:
2681         if (req)
2682                 fuse_put_request(fc, req);
2683         free_page((unsigned long) iov_page);
2684         while (num_pages)
2685                 __free_page(pages[--num_pages]);
2686         kfree(pages);
2687
2688         return err ? err : outarg.result;
2689 }
2690 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2691
2692 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2693                        unsigned long arg, unsigned int flags)
2694 {
2695         struct inode *inode = file_inode(file);
2696         struct fuse_conn *fc = get_fuse_conn(inode);
2697
2698         if (!fuse_allow_current_process(fc))
2699                 return -EACCES;
2700
2701         if (is_bad_inode(inode))
2702                 return -EIO;
2703
2704         return fuse_do_ioctl(file, cmd, arg, flags);
2705 }
2706
2707 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2708                             unsigned long arg)
2709 {
2710         return fuse_ioctl_common(file, cmd, arg, 0);
2711 }
2712
2713 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2714                                    unsigned long arg)
2715 {
2716         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2717 }
2718
2719 /*
2720  * All files which have been polled are linked to RB tree
2721  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2722  * find the matching one.
2723  */
2724 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2725                                               struct rb_node **parent_out)
2726 {
2727         struct rb_node **link = &fc->polled_files.rb_node;
2728         struct rb_node *last = NULL;
2729
2730         while (*link) {
2731                 struct fuse_file *ff;
2732
2733                 last = *link;
2734                 ff = rb_entry(last, struct fuse_file, polled_node);
2735
2736                 if (kh < ff->kh)
2737                         link = &last->rb_left;
2738                 else if (kh > ff->kh)
2739                         link = &last->rb_right;
2740                 else
2741                         return link;
2742         }
2743
2744         if (parent_out)
2745                 *parent_out = last;
2746         return link;
2747 }
2748
2749 /*
2750  * The file is about to be polled.  Make sure it's on the polled_files
2751  * RB tree.  Note that files once added to the polled_files tree are
2752  * not removed before the file is released.  This is because a file
2753  * polled once is likely to be polled again.
2754  */
2755 static void fuse_register_polled_file(struct fuse_conn *fc,
2756                                       struct fuse_file *ff)
2757 {
2758         spin_lock(&fc->lock);
2759         if (RB_EMPTY_NODE(&ff->polled_node)) {
2760                 struct rb_node **link, *uninitialized_var(parent);
2761
2762                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2763                 BUG_ON(*link);
2764                 rb_link_node(&ff->polled_node, parent, link);
2765                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2766         }
2767         spin_unlock(&fc->lock);
2768 }
2769
2770 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2771 {
2772         struct fuse_file *ff = file->private_data;
2773         struct fuse_conn *fc = ff->fc;
2774         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2775         struct fuse_poll_out outarg;
2776         struct fuse_req *req;
2777         int err;
2778
2779         if (fc->no_poll)
2780                 return DEFAULT_POLLMASK;
2781
2782         poll_wait(file, &ff->poll_wait, wait);
2783         inarg.events = (__u32)poll_requested_events(wait);
2784
2785         /*
2786          * Ask for notification iff there's someone waiting for it.
2787          * The client may ignore the flag and always notify.
2788          */
2789         if (waitqueue_active(&ff->poll_wait)) {
2790                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2791                 fuse_register_polled_file(fc, ff);
2792         }
2793
2794         req = fuse_get_req_nopages(fc);
2795         if (IS_ERR(req))
2796                 return POLLERR;
2797
2798         req->in.h.opcode = FUSE_POLL;
2799         req->in.h.nodeid = ff->nodeid;
2800         req->in.numargs = 1;
2801         req->in.args[0].size = sizeof(inarg);
2802         req->in.args[0].value = &inarg;
2803         req->out.numargs = 1;
2804         req->out.args[0].size = sizeof(outarg);
2805         req->out.args[0].value = &outarg;
2806         fuse_request_send(fc, req);
2807         err = req->out.h.error;
2808         fuse_put_request(fc, req);
2809
2810         if (!err)
2811                 return outarg.revents;
2812         if (err == -ENOSYS) {
2813                 fc->no_poll = 1;
2814                 return DEFAULT_POLLMASK;
2815         }
2816         return POLLERR;
2817 }
2818 EXPORT_SYMBOL_GPL(fuse_file_poll);
2819
2820 /*
2821  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2822  * wakes up the poll waiters.
2823  */
2824 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2825                             struct fuse_notify_poll_wakeup_out *outarg)
2826 {
2827         u64 kh = outarg->kh;
2828         struct rb_node **link;
2829
2830         spin_lock(&fc->lock);
2831
2832         link = fuse_find_polled_node(fc, kh, NULL);
2833         if (*link) {
2834                 struct fuse_file *ff;
2835
2836                 ff = rb_entry(*link, struct fuse_file, polled_node);
2837                 wake_up_interruptible_sync(&ff->poll_wait);
2838         }
2839
2840         spin_unlock(&fc->lock);
2841         return 0;
2842 }
2843
2844 static void fuse_do_truncate(struct file *file)
2845 {
2846         struct inode *inode = file->f_mapping->host;
2847         struct iattr attr;
2848
2849         attr.ia_valid = ATTR_SIZE;
2850         attr.ia_size = i_size_read(inode);
2851
2852         attr.ia_file = file;
2853         attr.ia_valid |= ATTR_FILE;
2854
2855         fuse_do_setattr(inode, &attr, file);
2856 }
2857
2858 static inline loff_t fuse_round_up(loff_t off)
2859 {
2860         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2861 }
2862
2863 static ssize_t
2864 fuse_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter,
2865                         loff_t offset)
2866 {
2867         ssize_t ret = 0;
2868         struct file *file = iocb->ki_filp;
2869         struct fuse_file *ff = file->private_data;
2870         bool async_dio = ff->fc->async_dio;
2871         loff_t pos = 0;
2872         struct inode *inode;
2873         loff_t i_size;
2874         size_t count = iov_iter_count(iter);
2875         struct fuse_io_priv *io;
2876
2877         pos = offset;
2878         inode = file->f_mapping->host;
2879         i_size = i_size_read(inode);
2880
2881         if ((rw == READ) && (offset > i_size))
2882                 return 0;
2883
2884         /* optimization for short read */
2885         if (async_dio && rw != WRITE && offset + count > i_size) {
2886                 if (offset >= i_size)
2887                         return 0;
2888                 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2889                 iov_iter_truncate(iter, count);
2890         }
2891
2892         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2893         if (!io)
2894                 return -ENOMEM;
2895         spin_lock_init(&io->lock);
2896         io->reqs = 1;
2897         io->bytes = -1;
2898         io->size = 0;
2899         io->offset = offset;
2900         io->write = (rw == WRITE);
2901         io->err = 0;
2902         io->file = file;
2903         /*
2904          * By default, we want to optimize all I/Os with async request
2905          * submission to the client filesystem if supported.
2906          */
2907         io->async = async_dio;
2908         io->iocb = iocb;
2909
2910         /*
2911          * We cannot asynchronously extend the size of a file. We have no method
2912          * to wait on real async I/O requests, so we must submit this request
2913          * synchronously.
2914          */
2915         if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2916                 io->async = false;
2917
2918         if (rw == WRITE)
2919                 ret = __fuse_direct_write(io, iter, &pos);
2920         else
2921                 ret = __fuse_direct_read(io, iter, &pos);
2922
2923         if (io->async) {
2924                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2925
2926                 /* we have a non-extending, async request, so return */
2927                 if (!is_sync_kiocb(iocb))
2928                         return -EIOCBQUEUED;
2929
2930                 ret = wait_on_sync_kiocb(iocb);
2931         } else {
2932                 kfree(io);
2933         }
2934
2935         if (rw == WRITE) {
2936                 if (ret > 0)
2937                         fuse_write_update_size(inode, pos);
2938                 else if (ret < 0 && offset + count > i_size)
2939                         fuse_do_truncate(file);
2940         }
2941
2942         return ret;
2943 }
2944
2945 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2946                                 loff_t length)
2947 {
2948         struct fuse_file *ff = file->private_data;
2949         struct inode *inode = file->f_inode;
2950         struct fuse_inode *fi = get_fuse_inode(inode);
2951         struct fuse_conn *fc = ff->fc;
2952         struct fuse_req *req;
2953         struct fuse_fallocate_in inarg = {
2954                 .fh = ff->fh,
2955                 .offset = offset,
2956                 .length = length,
2957                 .mode = mode
2958         };
2959         int err;
2960         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2961                            (mode & FALLOC_FL_PUNCH_HOLE);
2962
2963         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2964                 return -EOPNOTSUPP;
2965
2966         if (fc->no_fallocate)
2967                 return -EOPNOTSUPP;
2968
2969         if (lock_inode) {
2970                 mutex_lock(&inode->i_mutex);
2971                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2972                         loff_t endbyte = offset + length - 1;
2973                         err = filemap_write_and_wait_range(inode->i_mapping,
2974                                                            offset, endbyte);
2975                         if (err)
2976                                 goto out;
2977
2978                         fuse_sync_writes(inode);
2979                 }
2980         }
2981
2982         if (!(mode & FALLOC_FL_KEEP_SIZE))
2983                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2984
2985         req = fuse_get_req_nopages(fc);
2986         if (IS_ERR(req)) {
2987                 err = PTR_ERR(req);
2988                 goto out;
2989         }
2990
2991         req->in.h.opcode = FUSE_FALLOCATE;
2992         req->in.h.nodeid = ff->nodeid;
2993         req->in.numargs = 1;
2994         req->in.args[0].size = sizeof(inarg);
2995         req->in.args[0].value = &inarg;
2996         fuse_request_send(fc, req);
2997         err = req->out.h.error;
2998         if (err == -ENOSYS) {
2999                 fc->no_fallocate = 1;
3000                 err = -EOPNOTSUPP;
3001         }
3002         fuse_put_request(fc, req);
3003
3004         if (err)
3005                 goto out;
3006
3007         /* we could have extended the file */
3008         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3009                 bool changed = fuse_write_update_size(inode, offset + length);
3010
3011                 if (changed && fc->writeback_cache)
3012                         file_update_time(file);
3013         }
3014
3015         if (mode & FALLOC_FL_PUNCH_HOLE)
3016                 truncate_pagecache_range(inode, offset, offset + length - 1);
3017
3018         fuse_invalidate_attr(inode);
3019
3020 out:
3021         if (!(mode & FALLOC_FL_KEEP_SIZE))
3022                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3023
3024         if (lock_inode)
3025                 mutex_unlock(&inode->i_mutex);
3026
3027         return err;
3028 }
3029
3030 static const struct file_operations fuse_file_operations = {
3031         .llseek         = fuse_file_llseek,
3032         .read           = new_sync_read,
3033         .read_iter      = fuse_file_read_iter,
3034         .write          = new_sync_write,
3035         .write_iter     = fuse_file_write_iter,
3036         .mmap           = fuse_file_mmap,
3037         .open           = fuse_open,
3038         .flush          = fuse_flush,
3039         .release        = fuse_release,
3040         .fsync          = fuse_fsync,
3041         .lock           = fuse_file_lock,
3042         .flock          = fuse_file_flock,
3043         .splice_read    = generic_file_splice_read,
3044         .unlocked_ioctl = fuse_file_ioctl,
3045         .compat_ioctl   = fuse_file_compat_ioctl,
3046         .poll           = fuse_file_poll,
3047         .fallocate      = fuse_file_fallocate,
3048 };
3049
3050 static const struct file_operations fuse_direct_io_file_operations = {
3051         .llseek         = fuse_file_llseek,
3052         .read           = fuse_direct_read,
3053         .write          = fuse_direct_write,
3054         .mmap           = fuse_direct_mmap,
3055         .open           = fuse_open,
3056         .flush          = fuse_flush,
3057         .release        = fuse_release,
3058         .fsync          = fuse_fsync,
3059         .lock           = fuse_file_lock,
3060         .flock          = fuse_file_flock,
3061         .unlocked_ioctl = fuse_file_ioctl,
3062         .compat_ioctl   = fuse_file_compat_ioctl,
3063         .poll           = fuse_file_poll,
3064         .fallocate      = fuse_file_fallocate,
3065         /* no splice_read */
3066 };
3067
3068 static const struct address_space_operations fuse_file_aops  = {
3069         .readpage       = fuse_readpage,
3070         .writepage      = fuse_writepage,
3071         .writepages     = fuse_writepages,
3072         .launder_page   = fuse_launder_page,
3073         .readpages      = fuse_readpages,
3074         .set_page_dirty = __set_page_dirty_nobuffers,
3075         .bmap           = fuse_bmap,
3076         .direct_IO      = fuse_direct_IO,
3077         .write_begin    = fuse_write_begin,
3078         .write_end      = fuse_write_end,
3079 };
3080
3081 void fuse_init_file_inode(struct inode *inode)
3082 {
3083         inode->i_fop = &fuse_file_operations;
3084         inode->i_data.a_ops = &fuse_file_aops;
3085 }