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_aio_read(struct kiocb *iocb, const struct iovec *iov,
 937                                   unsigned long nr_segs, loff_t pos)
 938 {
 939         struct inode *inode = iocb->ki_filp->f_mapping->host;
 940         struct fuse_conn *fc = get_fuse_conn(inode);
 941
 942         /*
 943          * In auto invalidate mode, always update attributes on read.
 944          * Otherwise, only update if we attempt to read past EOF (to ensure
 945          * i_size is up to date).
 946          */
 947         if (fc->auto_inval_data ||
 948             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 949                 int err;
 950                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 951                 if (err)
 952                         return err;
 953         }
 954
 955         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 956 }
 957
 958 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 959                             loff_t pos, size_t count)
 960 {
 961         struct fuse_write_in *inarg = &req->misc.write.in;
 962         struct fuse_write_out *outarg = &req->misc.write.out;
 963
 964         inarg->fh = ff->fh;
 965         inarg->offset = pos;
 966         inarg->size = count;
 967         req->in.h.opcode = FUSE_WRITE;
 968         req->in.h.nodeid = ff->nodeid;
 969         req->in.numargs = 2;
 970         if (ff->fc->minor < 9)
 971                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 972         else
 973                 req->in.args[0].size = sizeof(struct fuse_write_in);
 974         req->in.args[0].value = inarg;
 975         req->in.args[1].size = count;
 976         req->out.numargs = 1;
 977         req->out.args[0].size = sizeof(struct fuse_write_out);
 978         req->out.args[0].value = outarg;
 979 }
 980
 981 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 982                               loff_t pos, size_t count, fl_owner_t owner)
 983 {
 984         struct file *file = io->file;
 985         struct fuse_file *ff = file->private_data;
 986         struct fuse_conn *fc = ff->fc;
 987         struct fuse_write_in *inarg = &req->misc.write.in;
 988
 989         fuse_write_fill(req, ff, pos, count);
 990         inarg->flags = file->f_flags;
 991         if (owner != NULL) {
 992                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 993                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 994         }
 995
 996         if (io->async)
 997                 return fuse_async_req_send(fc, req, count, io);
 998
 999         fuse_request_send(fc, req);
1000         return req->misc.write.out.size;
1001 }
1002
1003 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1004 {
1005         struct fuse_conn *fc = get_fuse_conn(inode);
1006         struct fuse_inode *fi = get_fuse_inode(inode);
1007         bool ret = false;
1008
1009         spin_lock(&fc->lock);
1010         fi->attr_version = ++fc->attr_version;
1011         if (pos > inode->i_size) {
1012                 i_size_write(inode, pos);
1013                 ret = true;
1014         }
1015         spin_unlock(&fc->lock);
1016
1017         return ret;
1018 }
1019
1020 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1021                                     struct inode *inode, loff_t pos,
1022                                     size_t count)
1023 {
1024         size_t res;
1025         unsigned offset;
1026         unsigned i;
1027         struct fuse_io_priv io = { .async = 0, .file = file };
1028
1029         for (i = 0; i < req->num_pages; i++)
1030                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1031
1032         res = fuse_send_write(req, &io, pos, count, NULL);
1033
1034         offset = req->page_descs[0].offset;
1035         count = res;
1036         for (i = 0; i < req->num_pages; i++) {
1037                 struct page *page = req->pages[i];
1038
1039                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1040                         SetPageUptodate(page);
1041
1042                 if (count > PAGE_CACHE_SIZE - offset)
1043                         count -= PAGE_CACHE_SIZE - offset;
1044                 else
1045                         count = 0;
1046                 offset = 0;
1047
1048                 unlock_page(page);
1049                 page_cache_release(page);
1050         }
1051
1052         return res;
1053 }
1054
1055 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1056                                struct address_space *mapping,
1057                                struct iov_iter *ii, loff_t pos)
1058 {
1059         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1060         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1061         size_t count = 0;
1062         int err;
1063
1064         req->in.argpages = 1;
1065         req->page_descs[0].offset = offset;
1066
1067         do {
1068                 size_t tmp;
1069                 struct page *page;
1070                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1071                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1072                                      iov_iter_count(ii));
1073
1074                 bytes = min_t(size_t, bytes, fc->max_write - count);
1075
1076  again:
1077                 err = -EFAULT;
1078                 if (iov_iter_fault_in_readable(ii, bytes))
1079                         break;
1080
1081                 err = -ENOMEM;
1082                 page = grab_cache_page_write_begin(mapping, index, 0);
1083                 if (!page)
1084                         break;
1085
1086                 if (mapping_writably_mapped(mapping))
1087                         flush_dcache_page(page);
1088
1089                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1090                 flush_dcache_page(page);
1091
1092                 if (!tmp) {
1093                         unlock_page(page);
1094                         page_cache_release(page);
1095                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1096                         goto again;
1097                 }
1098
1099                 err = 0;
1100                 req->pages[req->num_pages] = page;
1101                 req->page_descs[req->num_pages].length = tmp;
1102                 req->num_pages++;
1103
1104                 iov_iter_advance(ii, tmp);
1105                 count += tmp;
1106                 pos += tmp;
1107                 offset += tmp;
1108                 if (offset == PAGE_CACHE_SIZE)
1109                         offset = 0;
1110
1111                 if (!fc->big_writes)
1112                         break;
1113         } while (iov_iter_count(ii) && count < fc->max_write &&
1114                  req->num_pages < req->max_pages && offset == 0);
1115
1116         return count > 0 ? count : err;
1117 }
1118
1119 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1120 {
1121         return min_t(unsigned,
1122                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1123                      (pos >> PAGE_CACHE_SHIFT) + 1,
1124                      FUSE_MAX_PAGES_PER_REQ);
1125 }
1126
1127 static ssize_t fuse_perform_write(struct file *file,
1128                                   struct address_space *mapping,
1129                                   struct iov_iter *ii, loff_t pos)
1130 {
1131         struct inode *inode = mapping->host;
1132         struct fuse_conn *fc = get_fuse_conn(inode);
1133         struct fuse_inode *fi = get_fuse_inode(inode);
1134         int err = 0;
1135         ssize_t res = 0;
1136
1137         if (is_bad_inode(inode))
1138                 return -EIO;
1139
1140         if (inode->i_size < pos + iov_iter_count(ii))
1141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1142
1143         do {
1144                 struct fuse_req *req;
1145                 ssize_t count;
1146                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1147
1148                 req = fuse_get_req(fc, nr_pages);
1149                 if (IS_ERR(req)) {
1150                         err = PTR_ERR(req);
1151                         break;
1152                 }
1153
1154                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1155                 if (count <= 0) {
1156                         err = count;
1157                 } else {
1158                         size_t num_written;
1159
1160                         num_written = fuse_send_write_pages(req, file, inode,
1161                                                             pos, count);
1162                         err = req->out.h.error;
1163                         if (!err) {
1164                                 res += num_written;
1165                                 pos += num_written;
1166
1167                                 /* break out of the loop on short write */
1168                                 if (num_written != count)
1169                                         err = -EIO;
1170                         }
1171                 }
1172                 fuse_put_request(fc, req);
1173         } while (!err && iov_iter_count(ii));
1174
1175         if (res > 0)
1176                 fuse_write_update_size(inode, pos);
1177
1178         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1179         fuse_invalidate_attr(inode);
1180
1181         return res > 0 ? res : err;
1182 }
1183
1184 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1185                                    unsigned long nr_segs, loff_t pos)
1186 {
1187         struct file *file = iocb->ki_filp;
1188         struct address_space *mapping = file->f_mapping;
1189         size_t count = 0;
1190         size_t ocount = 0;
1191         ssize_t written = 0;
1192         ssize_t written_buffered = 0;
1193         struct inode *inode = mapping->host;
1194         ssize_t err;
1195         struct iov_iter i;
1196         loff_t endbyte = 0;
1197
1198         if (get_fuse_conn(inode)->writeback_cache) {
1199                 /* Update size (EOF optimization) and mode (SUID clearing) */
1200                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1201                 if (err)
1202                         return err;
1203
1204                 return generic_file_aio_write(iocb, iov, nr_segs, pos);
1205         }
1206
1207         WARN_ON(iocb->ki_pos != pos);
1208
1209         ocount = 0;
1210         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1211         if (err)
1212                 return err;
1213
1214         count = ocount;
1215         mutex_lock(&inode->i_mutex);
1216
1217         /* We can write back this queue in page reclaim */
1218         current->backing_dev_info = mapping->backing_dev_info;
1219
1220         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1221         if (err)
1222                 goto out;
1223
1224         if (count == 0)
1225                 goto out;
1226
1227         err = file_remove_suid(file);
1228         if (err)
1229                 goto out;
1230
1231         err = file_update_time(file);
1232         if (err)
1233                 goto out;
1234
1235         if (file->f_flags & O_DIRECT) {
1236                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1237                                                     count, ocount);
1238                 if (written < 0 || written == count)
1239                         goto out;
1240
1241                 pos += written;
1242                 count -= written;
1243
1244                 iov_iter_init(&i, iov, nr_segs, count, written);
1245                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1246                 if (written_buffered < 0) {
1247                         err = written_buffered;
1248                         goto out;
1249                 }
1250                 endbyte = pos + written_buffered - 1;
1251
1252                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1253                                                    endbyte);
1254                 if (err)
1255                         goto out;
1256
1257                 invalidate_mapping_pages(file->f_mapping,
1258                                          pos >> PAGE_CACHE_SHIFT,
1259                                          endbyte >> PAGE_CACHE_SHIFT);
1260
1261                 written += written_buffered;
1262                 iocb->ki_pos = pos + written_buffered;
1263         } else {
1264                 iov_iter_init(&i, iov, nr_segs, count, 0);
1265                 written = fuse_perform_write(file, mapping, &i, pos);
1266                 if (written >= 0)
1267                         iocb->ki_pos = pos + written;
1268         }
1269 out:
1270         current->backing_dev_info = NULL;
1271         mutex_unlock(&inode->i_mutex);
1272
1273         return written ? written : err;
1274 }
1275
1276 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1277                 unsigned index, unsigned nr_pages)
1278 {
1279         int i;
1280
1281         for (i = index; i < index + nr_pages; i++)
1282                 req->page_descs[i].length = PAGE_SIZE -
1283                         req->page_descs[i].offset;
1284 }
1285
1286 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1287 {
1288         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1289 }
1290
1291 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1292                                         size_t max_size)
1293 {
1294         return min(iov_iter_single_seg_count(ii), max_size);
1295 }
1296
1297 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1298                                size_t *nbytesp, int write)
1299 {
1300         size_t nbytes = 0;  /* # bytes already packed in req */
1301
1302         /* Special case for kernel I/O: can copy directly into the buffer */
1303         if (segment_eq(get_fs(), KERNEL_DS)) {
1304                 unsigned long user_addr = fuse_get_user_addr(ii);
1305                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1306
1307                 if (write)
1308                         req->in.args[1].value = (void *) user_addr;
1309                 else
1310                         req->out.args[0].value = (void *) user_addr;
1311
1312                 iov_iter_advance(ii, frag_size);
1313                 *nbytesp = frag_size;
1314                 return 0;
1315         }
1316
1317         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1318                 unsigned npages;
1319                 unsigned long user_addr = fuse_get_user_addr(ii);
1320                 unsigned offset = user_addr & ~PAGE_MASK;
1321                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1322                 int ret;
1323
1324                 unsigned n = req->max_pages - req->num_pages;
1325                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1326
1327                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1328                 npages = clamp(npages, 1U, n);
1329
1330                 ret = get_user_pages_fast(user_addr, npages, !write,
1331                                           &req->pages[req->num_pages]);
1332                 if (ret < 0)
1333                         return ret;
1334
1335                 npages = ret;
1336                 frag_size = min_t(size_t, frag_size,
1337                                   (npages << PAGE_SHIFT) - offset);
1338                 iov_iter_advance(ii, frag_size);
1339
1340                 req->page_descs[req->num_pages].offset = offset;
1341                 fuse_page_descs_length_init(req, req->num_pages, npages);
1342
1343                 req->num_pages += npages;
1344                 req->page_descs[req->num_pages - 1].length -=
1345                         (npages << PAGE_SHIFT) - offset - frag_size;
1346
1347                 nbytes += frag_size;
1348         }
1349
1350         if (write)
1351                 req->in.argpages = 1;
1352         else
1353                 req->out.argpages = 1;
1354
1355         *nbytesp = nbytes;
1356
1357         return 0;
1358 }
1359
1360 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1361 {
1362         struct iov_iter ii = *ii_p;
1363         int npages = 0;
1364
1365         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1366                 unsigned long user_addr = fuse_get_user_addr(&ii);
1367                 unsigned offset = user_addr & ~PAGE_MASK;
1368                 size_t frag_size = iov_iter_single_seg_count(&ii);
1369
1370                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1371                 iov_iter_advance(&ii, frag_size);
1372         }
1373
1374         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1375 }
1376
1377 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1378                        unsigned long nr_segs, size_t count, loff_t *ppos,
1379                        int flags)
1380 {
1381         int write = flags & FUSE_DIO_WRITE;
1382         int cuse = flags & FUSE_DIO_CUSE;
1383         struct file *file = io->file;
1384         struct inode *inode = file->f_mapping->host;
1385         struct fuse_file *ff = file->private_data;
1386         struct fuse_conn *fc = ff->fc;
1387         size_t nmax = write ? fc->max_write : fc->max_read;
1388         loff_t pos = *ppos;
1389         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1390         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1391         ssize_t res = 0;
1392         struct fuse_req *req;
1393         struct iov_iter ii;
1394
1395         iov_iter_init(&ii, iov, nr_segs, count, 0);
1396
1397         if (io->async)
1398                 req = fuse_get_req_for_background(fc, fuse_iter_npages(&ii));
1399         else
1400                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1401         if (IS_ERR(req))
1402                 return PTR_ERR(req);
1403
1404         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1405                 if (!write)
1406                         mutex_lock(&inode->i_mutex);
1407                 fuse_sync_writes(inode);
1408                 if (!write)
1409                         mutex_unlock(&inode->i_mutex);
1410         }
1411
1412         while (count) {
1413                 size_t nres;
1414                 fl_owner_t owner = current->files;
1415                 size_t nbytes = min(count, nmax);
1416                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1417                 if (err) {
1418                         res = err;
1419                         break;
1420                 }
1421
1422                 if (write)
1423                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1424                 else
1425                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1426
1427                 if (!io->async)
1428                         fuse_release_user_pages(req, !write);
1429                 if (req->out.h.error) {
1430                         if (!res)
1431                                 res = req->out.h.error;
1432                         break;
1433                 } else if (nres > nbytes) {
1434                         res = -EIO;
1435                         break;
1436                 }
1437                 count -= nres;
1438                 res += nres;
1439                 pos += nres;
1440                 if (nres != nbytes)
1441                         break;
1442                 if (count) {
1443                         fuse_put_request(fc, req);
1444                         if (io->async)
1445                                 req = fuse_get_req_for_background(fc,
1446                                         fuse_iter_npages(&ii));
1447                         else
1448                                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1449                         if (IS_ERR(req))
1450                                 break;
1451                 }
1452         }
1453         if (!IS_ERR(req))
1454                 fuse_put_request(fc, req);
1455         if (res > 0)
1456                 *ppos = pos;
1457
1458         return res;
1459 }
1460 EXPORT_SYMBOL_GPL(fuse_direct_io);
1461
1462 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1463                                   const struct iovec *iov,
1464                                   unsigned long nr_segs, loff_t *ppos,
1465                                   size_t count)
1466 {
1467         ssize_t res;
1468         struct file *file = io->file;
1469         struct inode *inode = file_inode(file);
1470
1471         if (is_bad_inode(inode))
1472                 return -EIO;
1473
1474         res = fuse_direct_io(io, iov, nr_segs, count, ppos, 0);
1475
1476         fuse_invalidate_attr(inode);
1477
1478         return res;
1479 }
1480
1481 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1482                                      size_t count, loff_t *ppos)
1483 {
1484         struct fuse_io_priv io = { .async = 0, .file = file };
1485         struct iovec iov = { .iov_base = buf, .iov_len = count };
1486         return __fuse_direct_read(&io, &iov, 1, ppos, count);
1487 }
1488
1489 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1490                                    const struct iovec *iov,
1491                                    unsigned long nr_segs, loff_t *ppos)
1492 {
1493         struct file *file = io->file;
1494         struct inode *inode = file_inode(file);
1495         size_t count = iov_length(iov, nr_segs);
1496         ssize_t res;
1497
1498         res = generic_write_checks(file, ppos, &count, 0);
1499         if (!res)
1500                 res = fuse_direct_io(io, iov, nr_segs, count, ppos,
1501                                      FUSE_DIO_WRITE);
1502
1503         fuse_invalidate_attr(inode);
1504
1505         return res;
1506 }
1507
1508 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1509                                  size_t count, loff_t *ppos)
1510 {
1511         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1512         struct inode *inode = file_inode(file);
1513         ssize_t res;
1514         struct fuse_io_priv io = { .async = 0, .file = file };
1515
1516         if (is_bad_inode(inode))
1517                 return -EIO;
1518
1519         /* Don't allow parallel writes to the same file */
1520         mutex_lock(&inode->i_mutex);
1521         res = __fuse_direct_write(&io, &iov, 1, ppos);
1522         if (res > 0)
1523                 fuse_write_update_size(inode, *ppos);
1524         mutex_unlock(&inode->i_mutex);
1525
1526         return res;
1527 }
1528
1529 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1530 {
1531         int i;
1532
1533         for (i = 0; i < req->num_pages; i++)
1534                 __free_page(req->pages[i]);
1535
1536         if (req->ff)
1537                 fuse_file_put(req->ff, false);
1538 }
1539
1540 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1541 {
1542         struct inode *inode = req->inode;
1543         struct fuse_inode *fi = get_fuse_inode(inode);
1544         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1545         int i;
1546
1547         list_del(&req->writepages_entry);
1548         for (i = 0; i < req->num_pages; i++) {
1549                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1550                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1551                 bdi_writeout_inc(bdi);
1552         }
1553         wake_up(&fi->page_waitq);
1554 }
1555
1556 /* Called under fc->lock, may release and reacquire it */
1557 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1558                                 loff_t size)
1559 __releases(fc->lock)
1560 __acquires(fc->lock)
1561 {
1562         struct fuse_inode *fi = get_fuse_inode(req->inode);
1563         struct fuse_write_in *inarg = &req->misc.write.in;
1564         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1565
1566         if (!fc->connected)
1567                 goto out_free;
1568
1569         if (inarg->offset + data_size <= size) {
1570                 inarg->size = data_size;
1571         } else if (inarg->offset < size) {
1572                 inarg->size = size - inarg->offset;
1573         } else {
1574                 /* Got truncated off completely */
1575                 goto out_free;
1576         }
1577
1578         req->in.args[1].size = inarg->size;
1579         fi->writectr++;
1580         fuse_request_send_background_locked(fc, req);
1581         return;
1582
1583  out_free:
1584         fuse_writepage_finish(fc, req);
1585         spin_unlock(&fc->lock);
1586         fuse_writepage_free(fc, req);
1587         fuse_put_request(fc, req);
1588         spin_lock(&fc->lock);
1589 }
1590
1591 /*
1592  * If fi->writectr is positive (no truncate or fsync going on) send
1593  * all queued writepage requests.
1594  *
1595  * Called with fc->lock
1596  */
1597 void fuse_flush_writepages(struct inode *inode)
1598 __releases(fc->lock)
1599 __acquires(fc->lock)
1600 {
1601         struct fuse_conn *fc = get_fuse_conn(inode);
1602         struct fuse_inode *fi = get_fuse_inode(inode);
1603         size_t crop = i_size_read(inode);
1604         struct fuse_req *req;
1605
1606         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1607                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1608                 list_del_init(&req->list);
1609                 fuse_send_writepage(fc, req, crop);
1610         }
1611 }
1612
1613 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1614 {
1615         struct inode *inode = req->inode;
1616         struct fuse_inode *fi = get_fuse_inode(inode);
1617
1618         mapping_set_error(inode->i_mapping, req->out.h.error);
1619         spin_lock(&fc->lock);
1620         while (req->misc.write.next) {
1621                 struct fuse_conn *fc = get_fuse_conn(inode);
1622                 struct fuse_write_in *inarg = &req->misc.write.in;
1623                 struct fuse_req *next = req->misc.write.next;
1624                 req->misc.write.next = next->misc.write.next;
1625                 next->misc.write.next = NULL;
1626                 next->ff = fuse_file_get(req->ff);
1627                 list_add(&next->writepages_entry, &fi->writepages);
1628
1629                 /*
1630                  * Skip fuse_flush_writepages() to make it easy to crop requests
1631                  * based on primary request size.
1632                  *
1633                  * 1st case (trivial): there are no concurrent activities using
1634                  * fuse_set/release_nowrite.  Then we're on safe side because
1635                  * fuse_flush_writepages() would call fuse_send_writepage()
1636                  * anyway.
1637                  *
1638                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1639                  * now for completion of all in-flight requests.  This happens
1640                  * rarely and no more than once per page, so this should be
1641                  * okay.
1642                  *
1643                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1644                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1645                  * that fuse_set_nowrite returned implies that all in-flight
1646                  * requests were completed along with all of their secondary
1647                  * requests.  Further primary requests are blocked by negative
1648                  * writectr.  Hence there cannot be any in-flight requests and
1649                  * no invocations of fuse_writepage_end() while we're in
1650                  * fuse_set_nowrite..fuse_release_nowrite section.
1651                  */
1652                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1653         }
1654         fi->writectr--;
1655         fuse_writepage_finish(fc, req);
1656         spin_unlock(&fc->lock);
1657         fuse_writepage_free(fc, req);
1658 }
1659
1660 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1661                                                struct fuse_inode *fi)
1662 {
1663         struct fuse_file *ff = NULL;
1664
1665         spin_lock(&fc->lock);
1666         if (!list_empty(&fi->write_files)) {
1667                 ff = list_entry(fi->write_files.next, struct fuse_file,
1668                                 write_entry);
1669                 fuse_file_get(ff);
1670         }
1671         spin_unlock(&fc->lock);
1672
1673         return ff;
1674 }
1675
1676 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1677                                              struct fuse_inode *fi)
1678 {
1679         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1680         WARN_ON(!ff);
1681         return ff;
1682 }
1683
1684 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1685 {
1686         struct fuse_conn *fc = get_fuse_conn(inode);
1687         struct fuse_inode *fi = get_fuse_inode(inode);
1688         struct fuse_file *ff;
1689         int err;
1690
1691         ff = __fuse_write_file_get(fc, fi);
1692         err = fuse_flush_times(inode, ff);
1693         if (ff)
1694                 fuse_file_put(ff, 0);
1695
1696         return err;
1697 }
1698
1699 static int fuse_writepage_locked(struct page *page)
1700 {
1701         struct address_space *mapping = page->mapping;
1702         struct inode *inode = mapping->host;
1703         struct fuse_conn *fc = get_fuse_conn(inode);
1704         struct fuse_inode *fi = get_fuse_inode(inode);
1705         struct fuse_req *req;
1706         struct page *tmp_page;
1707         int error = -ENOMEM;
1708
1709         set_page_writeback(page);
1710
1711         req = fuse_request_alloc_nofs(1);
1712         if (!req)
1713                 goto err;
1714
1715         req->background = 1; /* writeback always goes to bg_queue */
1716         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1717         if (!tmp_page)
1718                 goto err_free;
1719
1720         error = -EIO;
1721         req->ff = fuse_write_file_get(fc, fi);
1722         if (!req->ff)
1723                 goto err_free;
1724
1725         fuse_write_fill(req, req->ff, page_offset(page), 0);
1726
1727         copy_highpage(tmp_page, page);
1728         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1729         req->misc.write.next = NULL;
1730         req->in.argpages = 1;
1731         req->num_pages = 1;
1732         req->pages[0] = tmp_page;
1733         req->page_descs[0].offset = 0;
1734         req->page_descs[0].length = PAGE_SIZE;
1735         req->end = fuse_writepage_end;
1736         req->inode = inode;
1737
1738         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1739         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1740
1741         spin_lock(&fc->lock);
1742         list_add(&req->writepages_entry, &fi->writepages);
1743         list_add_tail(&req->list, &fi->queued_writes);
1744         fuse_flush_writepages(inode);
1745         spin_unlock(&fc->lock);
1746
1747         end_page_writeback(page);
1748
1749         return 0;
1750
1751 err_free:
1752         fuse_request_free(req);
1753 err:
1754         end_page_writeback(page);
1755         return error;
1756 }
1757
1758 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1759 {
1760         int err;
1761
1762         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1763                 /*
1764                  * ->writepages() should be called for sync() and friends.  We
1765                  * should only get here on direct reclaim and then we are
1766                  * allowed to skip a page which is already in flight
1767                  */
1768                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1769
1770                 redirty_page_for_writepage(wbc, page);
1771                 return 0;
1772         }
1773
1774         err = fuse_writepage_locked(page);
1775         unlock_page(page);
1776
1777         return err;
1778 }
1779
1780 struct fuse_fill_wb_data {
1781         struct fuse_req *req;
1782         struct fuse_file *ff;
1783         struct inode *inode;
1784         struct page **orig_pages;
1785 };
1786
1787 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1788 {
1789         struct fuse_req *req = data->req;
1790         struct inode *inode = data->inode;
1791         struct fuse_conn *fc = get_fuse_conn(inode);
1792         struct fuse_inode *fi = get_fuse_inode(inode);
1793         int num_pages = req->num_pages;
1794         int i;
1795
1796         req->ff = fuse_file_get(data->ff);
1797         spin_lock(&fc->lock);
1798         list_add_tail(&req->list, &fi->queued_writes);
1799         fuse_flush_writepages(inode);
1800         spin_unlock(&fc->lock);
1801
1802         for (i = 0; i < num_pages; i++)
1803                 end_page_writeback(data->orig_pages[i]);
1804 }
1805
1806 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1807                                      struct page *page)
1808 {
1809         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1810         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1811         struct fuse_req *tmp;
1812         struct fuse_req *old_req;
1813         bool found = false;
1814         pgoff_t curr_index;
1815
1816         BUG_ON(new_req->num_pages != 0);
1817
1818         spin_lock(&fc->lock);
1819         list_del(&new_req->writepages_entry);
1820         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1821                 BUG_ON(old_req->inode != new_req->inode);
1822                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1823                 if (curr_index <= page->index &&
1824                     page->index < curr_index + old_req->num_pages) {
1825                         found = true;
1826                         break;
1827                 }
1828         }
1829         if (!found) {
1830                 list_add(&new_req->writepages_entry, &fi->writepages);
1831                 goto out_unlock;
1832         }
1833
1834         new_req->num_pages = 1;
1835         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1836                 BUG_ON(tmp->inode != new_req->inode);
1837                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1838                 if (tmp->num_pages == 1 &&
1839                     curr_index == page->index) {
1840                         old_req = tmp;
1841                 }
1842         }
1843
1844         if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1845                                         old_req->state == FUSE_REQ_PENDING)) {
1846                 struct backing_dev_info *bdi = page->mapping->backing_dev_info;
1847
1848                 copy_highpage(old_req->pages[0], page);
1849                 spin_unlock(&fc->lock);
1850
1851                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1852                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1853                 bdi_writeout_inc(bdi);
1854                 fuse_writepage_free(fc, new_req);
1855                 fuse_request_free(new_req);
1856                 goto out;
1857         } else {
1858                 new_req->misc.write.next = old_req->misc.write.next;
1859                 old_req->misc.write.next = new_req;
1860         }
1861 out_unlock:
1862         spin_unlock(&fc->lock);
1863 out:
1864         return found;
1865 }
1866
1867 static int fuse_writepages_fill(struct page *page,
1868                 struct writeback_control *wbc, void *_data)
1869 {
1870         struct fuse_fill_wb_data *data = _data;
1871         struct fuse_req *req = data->req;
1872         struct inode *inode = data->inode;
1873         struct fuse_conn *fc = get_fuse_conn(inode);
1874         struct page *tmp_page;
1875         bool is_writeback;
1876         int err;
1877
1878         if (!data->ff) {
1879                 err = -EIO;
1880                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1881                 if (!data->ff)
1882                         goto out_unlock;
1883         }
1884
1885         /*
1886          * Being under writeback is unlikely but possible.  For example direct
1887          * read to an mmaped fuse file will set the page dirty twice; once when
1888          * the pages are faulted with get_user_pages(), and then after the read
1889          * completed.
1890          */
1891         is_writeback = fuse_page_is_writeback(inode, page->index);
1892
1893         if (req && req->num_pages &&
1894             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1895              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1896              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1897                 fuse_writepages_send(data);
1898                 data->req = NULL;
1899         }
1900         err = -ENOMEM;
1901         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1902         if (!tmp_page)
1903                 goto out_unlock;
1904
1905         /*
1906          * The page must not be redirtied until the writeout is completed
1907          * (i.e. userspace has sent a reply to the write request).  Otherwise
1908          * there could be more than one temporary page instance for each real
1909          * page.
1910          *
1911          * This is ensured by holding the page lock in page_mkwrite() while
1912          * checking fuse_page_is_writeback().  We already hold the page lock
1913          * since clear_page_dirty_for_io() and keep it held until we add the
1914          * request to the fi->writepages list and increment req->num_pages.
1915          * After this fuse_page_is_writeback() will indicate that the page is
1916          * under writeback, so we can release the page lock.
1917          */
1918         if (data->req == NULL) {
1919                 struct fuse_inode *fi = get_fuse_inode(inode);
1920
1921                 err = -ENOMEM;
1922                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1923                 if (!req) {
1924                         __free_page(tmp_page);
1925                         goto out_unlock;
1926                 }
1927
1928                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1929                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1930                 req->misc.write.next = NULL;
1931                 req->in.argpages = 1;
1932                 req->background = 1;
1933                 req->num_pages = 0;
1934                 req->end = fuse_writepage_end;
1935                 req->inode = inode;
1936
1937                 spin_lock(&fc->lock);
1938                 list_add(&req->writepages_entry, &fi->writepages);
1939                 spin_unlock(&fc->lock);
1940
1941                 data->req = req;
1942         }
1943         set_page_writeback(page);
1944
1945         copy_highpage(tmp_page, page);
1946         req->pages[req->num_pages] = tmp_page;
1947         req->page_descs[req->num_pages].offset = 0;
1948         req->page_descs[req->num_pages].length = PAGE_SIZE;
1949
1950         inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
1951         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1952
1953         err = 0;
1954         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1955                 end_page_writeback(page);
1956                 data->req = NULL;
1957                 goto out_unlock;
1958         }
1959         data->orig_pages[req->num_pages] = page;
1960
1961         /*
1962          * Protected by fc->lock against concurrent access by
1963          * fuse_page_is_writeback().
1964          */
1965         spin_lock(&fc->lock);
1966         req->num_pages++;
1967         spin_unlock(&fc->lock);
1968
1969 out_unlock:
1970         unlock_page(page);
1971
1972         return err;
1973 }
1974
1975 static int fuse_writepages(struct address_space *mapping,
1976                            struct writeback_control *wbc)
1977 {
1978         struct inode *inode = mapping->host;
1979         struct fuse_fill_wb_data data;
1980         int err;
1981
1982         err = -EIO;
1983         if (is_bad_inode(inode))
1984                 goto out;
1985
1986         data.inode = inode;
1987         data.req = NULL;
1988         data.ff = NULL;
1989
1990         err = -ENOMEM;
1991         data.orig_pages = kzalloc(sizeof(struct page *) *
1992                                   FUSE_MAX_PAGES_PER_REQ,
1993                                   GFP_NOFS);
1994         if (!data.orig_pages)
1995                 goto out;
1996
1997         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1998         if (data.req) {
1999                 /* Ignore errors if we can write at least one page */
2000                 BUG_ON(!data.req->num_pages);
2001                 fuse_writepages_send(&data);
2002                 err = 0;
2003         }
2004         if (data.ff)
2005                 fuse_file_put(data.ff, false);
2006
2007         kfree(data.orig_pages);
2008 out:
2009         return err;
2010 }
2011
2012 /*
2013  * It's worthy to make sure that space is reserved on disk for the write,
2014  * but how to implement it without killing performance need more thinking.
2015  */
2016 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2017                 loff_t pos, unsigned len, unsigned flags,
2018                 struct page **pagep, void **fsdata)
2019 {
2020         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2021         struct fuse_conn *fc = get_fuse_conn(file->f_dentry->d_inode);
2022         struct page *page;
2023         loff_t fsize;
2024         int err = -ENOMEM;
2025
2026         WARN_ON(!fc->writeback_cache);
2027
2028         page = grab_cache_page_write_begin(mapping, index, flags);
2029         if (!page)
2030                 goto error;
2031
2032         fuse_wait_on_page_writeback(mapping->host, page->index);
2033
2034         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
2035                 goto success;
2036         /*
2037          * Check if the start this page comes after the end of file, in which
2038          * case the readpage can be optimized away.
2039          */
2040         fsize = i_size_read(mapping->host);
2041         if (fsize <= (pos & PAGE_CACHE_MASK)) {
2042                 size_t off = pos & ~PAGE_CACHE_MASK;
2043                 if (off)
2044                         zero_user_segment(page, 0, off);
2045                 goto success;
2046         }
2047         err = fuse_do_readpage(file, page);
2048         if (err)
2049                 goto cleanup;
2050 success:
2051         *pagep = page;
2052         return 0;
2053
2054 cleanup:
2055         unlock_page(page);
2056         page_cache_release(page);
2057 error:
2058         return err;
2059 }
2060
2061 static int fuse_write_end(struct file *file, struct address_space *mapping,
2062                 loff_t pos, unsigned len, unsigned copied,
2063                 struct page *page, void *fsdata)
2064 {
2065         struct inode *inode = page->mapping->host;
2066
2067         if (!PageUptodate(page)) {
2068                 /* Zero any unwritten bytes at the end of the page */
2069                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2070                 if (endoff)
2071                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2072                 SetPageUptodate(page);
2073         }
2074
2075         fuse_write_update_size(inode, pos + copied);
2076         set_page_dirty(page);
2077         unlock_page(page);
2078         page_cache_release(page);
2079
2080         return copied;
2081 }
2082
2083 static int fuse_launder_page(struct page *page)
2084 {
2085         int err = 0;
2086         if (clear_page_dirty_for_io(page)) {
2087                 struct inode *inode = page->mapping->host;
2088                 err = fuse_writepage_locked(page);
2089                 if (!err)
2090                         fuse_wait_on_page_writeback(inode, page->index);
2091         }
2092         return err;
2093 }
2094
2095 /*
2096  * Write back dirty pages now, because there may not be any suitable
2097  * open files later
2098  */
2099 static void fuse_vma_close(struct vm_area_struct *vma)
2100 {
2101         filemap_write_and_wait(vma->vm_file->f_mapping);
2102 }
2103
2104 /*
2105  * Wait for writeback against this page to complete before allowing it
2106  * to be marked dirty again, and hence written back again, possibly
2107  * before the previous writepage completed.
2108  *
2109  * Block here, instead of in ->writepage(), so that the userspace fs
2110  * can only block processes actually operating on the filesystem.
2111  *
2112  * Otherwise unprivileged userspace fs would be able to block
2113  * unrelated:
2114  *
2115  * - page migration
2116  * - sync(2)
2117  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2118  */
2119 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2120 {
2121         struct page *page = vmf->page;
2122         struct inode *inode = file_inode(vma->vm_file);
2123
2124         file_update_time(vma->vm_file);
2125         lock_page(page);
2126         if (page->mapping != inode->i_mapping) {
2127                 unlock_page(page);
2128                 return VM_FAULT_NOPAGE;
2129         }
2130
2131         fuse_wait_on_page_writeback(inode, page->index);
2132         return VM_FAULT_LOCKED;
2133 }
2134
2135 static const struct vm_operations_struct fuse_file_vm_ops = {
2136         .close          = fuse_vma_close,
2137         .fault          = filemap_fault,
2138         .map_pages      = filemap_map_pages,
2139         .page_mkwrite   = fuse_page_mkwrite,
2140         .remap_pages    = generic_file_remap_pages,
2141 };
2142
2143 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2144 {
2145         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2146                 fuse_link_write_file(file);
2147
2148         file_accessed(file);
2149         vma->vm_ops = &fuse_file_vm_ops;
2150         return 0;
2151 }
2152
2153 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2154 {
2155         /* Can't provide the coherency needed for MAP_SHARED */
2156         if (vma->vm_flags & VM_MAYSHARE)
2157                 return -ENODEV;
2158
2159         invalidate_inode_pages2(file->f_mapping);
2160
2161         return generic_file_mmap(file, vma);
2162 }
2163
2164 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2165                                   struct file_lock *fl)
2166 {
2167         switch (ffl->type) {
2168         case F_UNLCK:
2169                 break;
2170
2171         case F_RDLCK:
2172         case F_WRLCK:
2173                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2174                     ffl->end < ffl->start)
2175                         return -EIO;
2176
2177                 fl->fl_start = ffl->start;
2178                 fl->fl_end = ffl->end;
2179                 fl->fl_pid = ffl->pid;
2180                 break;
2181
2182         default:
2183                 return -EIO;
2184         }
2185         fl->fl_type = ffl->type;
2186         return 0;
2187 }
2188
2189 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
2190                          const struct file_lock *fl, int opcode, pid_t pid,
2191                          int flock)
2192 {
2193         struct inode *inode = file_inode(file);
2194         struct fuse_conn *fc = get_fuse_conn(inode);
2195         struct fuse_file *ff = file->private_data;
2196         struct fuse_lk_in *arg = &req->misc.lk_in;
2197
2198         arg->fh = ff->fh;
2199         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2200         arg->lk.start = fl->fl_start;
2201         arg->lk.end = fl->fl_end;
2202         arg->lk.type = fl->fl_type;
2203         arg->lk.pid = pid;
2204         if (flock)
2205                 arg->lk_flags |= FUSE_LK_FLOCK;
2206         req->in.h.opcode = opcode;
2207         req->in.h.nodeid = get_node_id(inode);
2208         req->in.numargs = 1;
2209         req->in.args[0].size = sizeof(*arg);
2210         req->in.args[0].value = arg;
2211 }
2212
2213 static int fuse_getlk(struct file *file, struct file_lock *fl)
2214 {
2215         struct inode *inode = file_inode(file);
2216         struct fuse_conn *fc = get_fuse_conn(inode);
2217         struct fuse_req *req;
2218         struct fuse_lk_out outarg;
2219         int err;
2220
2221         req = fuse_get_req_nopages(fc);
2222         if (IS_ERR(req))
2223                 return PTR_ERR(req);
2224
2225         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
2226         req->out.numargs = 1;
2227         req->out.args[0].size = sizeof(outarg);
2228         req->out.args[0].value = &outarg;
2229         fuse_request_send(fc, req);
2230         err = req->out.h.error;
2231         fuse_put_request(fc, req);
2232         if (!err)
2233                 err = convert_fuse_file_lock(&outarg.lk, fl);
2234
2235         return err;
2236 }
2237
2238 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2239 {
2240         struct inode *inode = file_inode(file);
2241         struct fuse_conn *fc = get_fuse_conn(inode);
2242         struct fuse_req *req;
2243         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2244         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2245         int err;
2246
2247         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2248                 /* NLM needs asynchronous locks, which we don't support yet */
2249                 return -ENOLCK;
2250         }
2251
2252         /* Unlock on close is handled by the flush method */
2253         if (fl->fl_flags & FL_CLOSE)
2254                 return 0;
2255
2256         req = fuse_get_req_nopages(fc);
2257         if (IS_ERR(req))
2258                 return PTR_ERR(req);
2259
2260         fuse_lk_fill(req, file, fl, opcode, pid, flock);
2261         fuse_request_send(fc, req);
2262         err = req->out.h.error;
2263         /* locking is restartable */
2264         if (err == -EINTR)
2265                 err = -ERESTARTSYS;
2266         fuse_put_request(fc, req);
2267         return err;
2268 }
2269
2270 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2271 {
2272         struct inode *inode = file_inode(file);
2273         struct fuse_conn *fc = get_fuse_conn(inode);
2274         int err;
2275
2276         if (cmd == F_CANCELLK) {
2277                 err = 0;
2278         } else if (cmd == F_GETLK) {
2279                 if (fc->no_lock) {
2280                         posix_test_lock(file, fl);
2281                         err = 0;
2282                 } else
2283                         err = fuse_getlk(file, fl);
2284         } else {
2285                 if (fc->no_lock)
2286                         err = posix_lock_file(file, fl, NULL);
2287                 else
2288                         err = fuse_setlk(file, fl, 0);
2289         }
2290         return err;
2291 }
2292
2293 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2294 {
2295         struct inode *inode = file_inode(file);
2296         struct fuse_conn *fc = get_fuse_conn(inode);
2297         int err;
2298
2299         if (fc->no_flock) {
2300                 err = flock_lock_file_wait(file, fl);
2301         } else {
2302                 struct fuse_file *ff = file->private_data;
2303
2304                 /* emulate flock with POSIX locks */
2305                 ff->flock = true;
2306                 err = fuse_setlk(file, fl, 1);
2307         }
2308
2309         return err;
2310 }
2311
2312 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2313 {
2314         struct inode *inode = mapping->host;
2315         struct fuse_conn *fc = get_fuse_conn(inode);
2316         struct fuse_req *req;
2317         struct fuse_bmap_in inarg;
2318         struct fuse_bmap_out outarg;
2319         int err;
2320
2321         if (!inode->i_sb->s_bdev || fc->no_bmap)
2322                 return 0;
2323
2324         req = fuse_get_req_nopages(fc);
2325         if (IS_ERR(req))
2326                 return 0;
2327
2328         memset(&inarg, 0, sizeof(inarg));
2329         inarg.block = block;
2330         inarg.blocksize = inode->i_sb->s_blocksize;
2331         req->in.h.opcode = FUSE_BMAP;
2332         req->in.h.nodeid = get_node_id(inode);
2333         req->in.numargs = 1;
2334         req->in.args[0].size = sizeof(inarg);
2335         req->in.args[0].value = &inarg;
2336         req->out.numargs = 1;
2337         req->out.args[0].size = sizeof(outarg);
2338         req->out.args[0].value = &outarg;
2339         fuse_request_send(fc, req);
2340         err = req->out.h.error;
2341         fuse_put_request(fc, req);
2342         if (err == -ENOSYS)
2343                 fc->no_bmap = 1;
2344
2345         return err ? 0 : outarg.block;
2346 }
2347
2348 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2349 {
2350         loff_t retval;
2351         struct inode *inode = file_inode(file);
2352
2353         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2354         if (whence == SEEK_CUR || whence == SEEK_SET)
2355                 return generic_file_llseek(file, offset, whence);
2356
2357         mutex_lock(&inode->i_mutex);
2358         retval = fuse_update_attributes(inode, NULL, file, NULL);
2359         if (!retval)
2360                 retval = generic_file_llseek(file, offset, whence);
2361         mutex_unlock(&inode->i_mutex);
2362
2363         return retval;
2364 }
2365
2366 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2367                         unsigned int nr_segs, size_t bytes, bool to_user)
2368 {
2369         struct iov_iter ii;
2370         int page_idx = 0;
2371
2372         if (!bytes)
2373                 return 0;
2374
2375         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
2376
2377         while (iov_iter_count(&ii)) {
2378                 struct page *page = pages[page_idx++];
2379                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2380                 void *kaddr;
2381
2382                 kaddr = kmap(page);
2383
2384                 while (todo) {
2385                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2386                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2387                         size_t copy = min(todo, iov_len);
2388                         size_t left;
2389
2390                         if (!to_user)
2391                                 left = copy_from_user(kaddr, uaddr, copy);
2392                         else
2393                                 left = copy_to_user(uaddr, kaddr, copy);
2394
2395                         if (unlikely(left))
2396                                 return -EFAULT;
2397
2398                         iov_iter_advance(&ii, copy);
2399                         todo -= copy;
2400                         kaddr += copy;
2401                 }
2402
2403                 kunmap(page);
2404         }
2405
2406         return 0;
2407 }
2408
2409 /*
2410  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2411  * ABI was defined to be 'struct iovec' which is different on 32bit
2412  * and 64bit.  Fortunately we can determine which structure the server
2413  * used from the size of the reply.
2414  */
2415 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2416                                      size_t transferred, unsigned count,
2417                                      bool is_compat)
2418 {
2419 #ifdef CONFIG_COMPAT
2420         if (count * sizeof(struct compat_iovec) == transferred) {
2421                 struct compat_iovec *ciov = src;
2422                 unsigned i;
2423
2424                 /*
2425                  * With this interface a 32bit server cannot support
2426                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2427                  * requests
2428                  */
2429                 if (!is_compat)
2430                         return -EINVAL;
2431
2432                 for (i = 0; i < count; i++) {
2433                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2434                         dst[i].iov_len = ciov[i].iov_len;
2435                 }
2436                 return 0;
2437         }
2438 #endif
2439
2440         if (count * sizeof(struct iovec) != transferred)
2441                 return -EIO;
2442
2443         memcpy(dst, src, transferred);
2444         return 0;
2445 }
2446
2447 /* Make sure iov_length() won't overflow */
2448 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2449 {
2450         size_t n;
2451         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2452
2453         for (n = 0; n < count; n++, iov++) {
2454                 if (iov->iov_len > (size_t) max)
2455                         return -ENOMEM;
2456                 max -= iov->iov_len;
2457         }
2458         return 0;
2459 }
2460
2461 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2462                                  void *src, size_t transferred, unsigned count,
2463                                  bool is_compat)
2464 {
2465         unsigned i;
2466         struct fuse_ioctl_iovec *fiov = src;
2467
2468         if (fc->minor < 16) {
2469                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2470                                                  count, is_compat);
2471         }
2472
2473         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2474                 return -EIO;
2475
2476         for (i = 0; i < count; i++) {
2477                 /* Did the server supply an inappropriate value? */
2478                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2479                     fiov[i].len != (unsigned long) fiov[i].len)
2480                         return -EIO;
2481
2482                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2483                 dst[i].iov_len = (size_t) fiov[i].len;
2484
2485 #ifdef CONFIG_COMPAT
2486                 if (is_compat &&
2487                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2488                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2489                         return -EIO;
2490 #endif
2491         }
2492
2493         return 0;
2494 }
2495
2496
2497 /*
2498  * For ioctls, there is no generic way to determine how much memory
2499  * needs to be read and/or written.  Furthermore, ioctls are allowed
2500  * to dereference the passed pointer, so the parameter requires deep
2501  * copying but FUSE has no idea whatsoever about what to copy in or
2502  * out.
2503  *
2504  * This is solved by allowing FUSE server to retry ioctl with
2505  * necessary in/out iovecs.  Let's assume the ioctl implementation
2506  * needs to read in the following structure.
2507  *
2508  * struct a {
2509  *      char    *buf;
2510  *      size_t  buflen;
2511  * }
2512  *
2513  * On the first callout to FUSE server, inarg->in_size and
2514  * inarg->out_size will be NULL; then, the server completes the ioctl
2515  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2516  * the actual iov array to
2517  *
2518  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2519  *
2520  * which tells FUSE to copy in the requested area and retry the ioctl.
2521  * On the second round, the server has access to the structure and
2522  * from that it can tell what to look for next, so on the invocation,
2523  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2524  *
2525  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2526  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2527  *
2528  * FUSE will copy both struct a and the pointed buffer from the
2529  * process doing the ioctl and retry ioctl with both struct a and the
2530  * buffer.
2531  *
2532  * This time, FUSE server has everything it needs and completes ioctl
2533  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2534  *
2535  * Copying data out works the same way.
2536  *
2537  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2538  * automatically initializes in and out iovs by decoding @cmd with
2539  * _IOC_* macros and the server is not allowed to request RETRY.  This
2540  * limits ioctl data transfers to well-formed ioctls and is the forced
2541  * behavior for all FUSE servers.
2542  */
2543 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2544                    unsigned int flags)
2545 {
2546         struct fuse_file *ff = file->private_data;
2547         struct fuse_conn *fc = ff->fc;
2548         struct fuse_ioctl_in inarg = {
2549                 .fh = ff->fh,
2550                 .cmd = cmd,
2551                 .arg = arg,
2552                 .flags = flags
2553         };
2554         struct fuse_ioctl_out outarg;
2555         struct fuse_req *req = NULL;
2556         struct page **pages = NULL;
2557         struct iovec *iov_page = NULL;
2558         struct iovec *in_iov = NULL, *out_iov = NULL;
2559         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2560         size_t in_size, out_size, transferred;
2561         int err;
2562
2563 #if BITS_PER_LONG == 32
2564         inarg.flags |= FUSE_IOCTL_32BIT;
2565 #else
2566         if (flags & FUSE_IOCTL_COMPAT)
2567                 inarg.flags |= FUSE_IOCTL_32BIT;
2568 #endif
2569
2570         /* assume all the iovs returned by client always fits in a page */
2571         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2572
2573         err = -ENOMEM;
2574         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2575         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2576         if (!pages || !iov_page)
2577                 goto out;
2578
2579         /*
2580          * If restricted, initialize IO parameters as encoded in @cmd.
2581          * RETRY from server is not allowed.
2582          */
2583         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2584                 struct iovec *iov = iov_page;
2585
2586                 iov->iov_base = (void __user *)arg;
2587                 iov->iov_len = _IOC_SIZE(cmd);
2588
2589                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2590                         in_iov = iov;
2591                         in_iovs = 1;
2592                 }
2593
2594                 if (_IOC_DIR(cmd) & _IOC_READ) {
2595                         out_iov = iov;
2596                         out_iovs = 1;
2597                 }
2598         }
2599
2600  retry:
2601         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2602         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2603
2604         /*
2605          * Out data can be used either for actual out data or iovs,
2606          * make sure there always is at least one page.
2607          */
2608         out_size = max_t(size_t, out_size, PAGE_SIZE);
2609         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2610
2611         /* make sure there are enough buffer pages and init request with them */
2612         err = -ENOMEM;
2613         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2614                 goto out;
2615         while (num_pages < max_pages) {
2616                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2617                 if (!pages[num_pages])
2618                         goto out;
2619                 num_pages++;
2620         }
2621
2622         req = fuse_get_req(fc, num_pages);
2623         if (IS_ERR(req)) {
2624                 err = PTR_ERR(req);
2625                 req = NULL;
2626                 goto out;
2627         }
2628         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2629         req->num_pages = num_pages;
2630         fuse_page_descs_length_init(req, 0, req->num_pages);
2631
2632         /* okay, let's send it to the client */
2633         req->in.h.opcode = FUSE_IOCTL;
2634         req->in.h.nodeid = ff->nodeid;
2635         req->in.numargs = 1;
2636         req->in.args[0].size = sizeof(inarg);
2637         req->in.args[0].value = &inarg;
2638         if (in_size) {
2639                 req->in.numargs++;
2640                 req->in.args[1].size = in_size;
2641                 req->in.argpages = 1;
2642
2643                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2644                                            false);
2645                 if (err)
2646                         goto out;
2647         }
2648
2649         req->out.numargs = 2;
2650         req->out.args[0].size = sizeof(outarg);
2651         req->out.args[0].value = &outarg;
2652         req->out.args[1].size = out_size;
2653         req->out.argpages = 1;
2654         req->out.argvar = 1;
2655
2656         fuse_request_send(fc, req);
2657         err = req->out.h.error;
2658         transferred = req->out.args[1].size;
2659         fuse_put_request(fc, req);
2660         req = NULL;
2661         if (err)
2662                 goto out;
2663
2664         /* did it ask for retry? */
2665         if (outarg.flags & FUSE_IOCTL_RETRY) {
2666                 void *vaddr;
2667
2668                 /* no retry if in restricted mode */
2669                 err = -EIO;
2670                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2671                         goto out;
2672
2673                 in_iovs = outarg.in_iovs;
2674                 out_iovs = outarg.out_iovs;
2675
2676                 /*
2677                  * Make sure things are in boundary, separate checks
2678                  * are to protect against overflow.
2679                  */
2680                 err = -ENOMEM;
2681                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2682                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2683                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2684                         goto out;
2685
2686                 vaddr = kmap_atomic(pages[0]);
2687                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2688                                             transferred, in_iovs + out_iovs,
2689                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2690                 kunmap_atomic(vaddr);
2691                 if (err)
2692                         goto out;
2693
2694                 in_iov = iov_page;
2695                 out_iov = in_iov + in_iovs;
2696
2697                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2698                 if (err)
2699                         goto out;
2700
2701                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2702                 if (err)
2703                         goto out;
2704
2705                 goto retry;
2706         }
2707
2708         err = -EIO;
2709         if (transferred > inarg.out_size)
2710                 goto out;
2711
2712         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2713  out:
2714         if (req)
2715                 fuse_put_request(fc, req);
2716         free_page((unsigned long) iov_page);
2717         while (num_pages)
2718                 __free_page(pages[--num_pages]);
2719         kfree(pages);
2720
2721         return err ? err : outarg.result;
2722 }
2723 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2724
2725 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2726                        unsigned long arg, unsigned int flags)
2727 {
2728         struct inode *inode = file_inode(file);
2729         struct fuse_conn *fc = get_fuse_conn(inode);
2730
2731         if (!fuse_allow_current_process(fc))
2732                 return -EACCES;
2733
2734         if (is_bad_inode(inode))
2735                 return -EIO;
2736
2737         return fuse_do_ioctl(file, cmd, arg, flags);
2738 }
2739
2740 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2741                             unsigned long arg)
2742 {
2743         return fuse_ioctl_common(file, cmd, arg, 0);
2744 }
2745
2746 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2747                                    unsigned long arg)
2748 {
2749         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2750 }
2751
2752 /*
2753  * All files which have been polled are linked to RB tree
2754  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2755  * find the matching one.
2756  */
2757 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2758                                               struct rb_node **parent_out)
2759 {
2760         struct rb_node **link = &fc->polled_files.rb_node;
2761         struct rb_node *last = NULL;
2762
2763         while (*link) {
2764                 struct fuse_file *ff;
2765
2766                 last = *link;
2767                 ff = rb_entry(last, struct fuse_file, polled_node);
2768
2769                 if (kh < ff->kh)
2770                         link = &last->rb_left;
2771                 else if (kh > ff->kh)
2772                         link = &last->rb_right;
2773                 else
2774                         return link;
2775         }
2776
2777         if (parent_out)
2778                 *parent_out = last;
2779         return link;
2780 }
2781
2782 /*
2783  * The file is about to be polled.  Make sure it's on the polled_files
2784  * RB tree.  Note that files once added to the polled_files tree are
2785  * not removed before the file is released.  This is because a file
2786  * polled once is likely to be polled again.
2787  */
2788 static void fuse_register_polled_file(struct fuse_conn *fc,
2789                                       struct fuse_file *ff)
2790 {
2791         spin_lock(&fc->lock);
2792         if (RB_EMPTY_NODE(&ff->polled_node)) {
2793                 struct rb_node **link, *uninitialized_var(parent);
2794
2795                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2796                 BUG_ON(*link);
2797                 rb_link_node(&ff->polled_node, parent, link);
2798                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2799         }
2800         spin_unlock(&fc->lock);
2801 }
2802
2803 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2804 {
2805         struct fuse_file *ff = file->private_data;
2806         struct fuse_conn *fc = ff->fc;
2807         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2808         struct fuse_poll_out outarg;
2809         struct fuse_req *req;
2810         int err;
2811
2812         if (fc->no_poll)
2813                 return DEFAULT_POLLMASK;
2814
2815         poll_wait(file, &ff->poll_wait, wait);
2816         inarg.events = (__u32)poll_requested_events(wait);
2817
2818         /*
2819          * Ask for notification iff there's someone waiting for it.
2820          * The client may ignore the flag and always notify.
2821          */
2822         if (waitqueue_active(&ff->poll_wait)) {
2823                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2824                 fuse_register_polled_file(fc, ff);
2825         }
2826
2827         req = fuse_get_req_nopages(fc);
2828         if (IS_ERR(req))
2829                 return POLLERR;
2830
2831         req->in.h.opcode = FUSE_POLL;
2832         req->in.h.nodeid = ff->nodeid;
2833         req->in.numargs = 1;
2834         req->in.args[0].size = sizeof(inarg);
2835         req->in.args[0].value = &inarg;
2836         req->out.numargs = 1;
2837         req->out.args[0].size = sizeof(outarg);
2838         req->out.args[0].value = &outarg;
2839         fuse_request_send(fc, req);
2840         err = req->out.h.error;
2841         fuse_put_request(fc, req);
2842
2843         if (!err)
2844                 return outarg.revents;
2845         if (err == -ENOSYS) {
2846                 fc->no_poll = 1;
2847                 return DEFAULT_POLLMASK;
2848         }
2849         return POLLERR;
2850 }
2851 EXPORT_SYMBOL_GPL(fuse_file_poll);
2852
2853 /*
2854  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2855  * wakes up the poll waiters.
2856  */
2857 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2858                             struct fuse_notify_poll_wakeup_out *outarg)
2859 {
2860         u64 kh = outarg->kh;
2861         struct rb_node **link;
2862
2863         spin_lock(&fc->lock);
2864
2865         link = fuse_find_polled_node(fc, kh, NULL);
2866         if (*link) {
2867                 struct fuse_file *ff;
2868
2869                 ff = rb_entry(*link, struct fuse_file, polled_node);
2870                 wake_up_interruptible_sync(&ff->poll_wait);
2871         }
2872
2873         spin_unlock(&fc->lock);
2874         return 0;
2875 }
2876
2877 static void fuse_do_truncate(struct file *file)
2878 {
2879         struct inode *inode = file->f_mapping->host;
2880         struct iattr attr;
2881
2882         attr.ia_valid = ATTR_SIZE;
2883         attr.ia_size = i_size_read(inode);
2884
2885         attr.ia_file = file;
2886         attr.ia_valid |= ATTR_FILE;
2887
2888         fuse_do_setattr(inode, &attr, file);
2889 }
2890
2891 static inline loff_t fuse_round_up(loff_t off)
2892 {
2893         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2894 }
2895
2896 static ssize_t
2897 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2898                         loff_t offset, unsigned long nr_segs)
2899 {
2900         ssize_t ret = 0;
2901         struct file *file = iocb->ki_filp;
2902         struct fuse_file *ff = file->private_data;
2903         bool async_dio = ff->fc->async_dio;
2904         loff_t pos = 0;
2905         struct inode *inode;
2906         loff_t i_size;
2907         size_t count = iov_length(iov, nr_segs);
2908         struct fuse_io_priv *io;
2909
2910         pos = offset;
2911         inode = file->f_mapping->host;
2912         i_size = i_size_read(inode);
2913
2914         if ((rw == READ) && (offset > i_size))
2915                 return 0;
2916
2917         /* optimization for short read */
2918         if (async_dio && rw != WRITE && offset + count > i_size) {
2919                 if (offset >= i_size)
2920                         return 0;
2921                 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2922         }
2923
2924         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2925         if (!io)
2926                 return -ENOMEM;
2927         spin_lock_init(&io->lock);
2928         io->reqs = 1;
2929         io->bytes = -1;
2930         io->size = 0;
2931         io->offset = offset;
2932         io->write = (rw == WRITE);
2933         io->err = 0;
2934         io->file = file;
2935         /*
2936          * By default, we want to optimize all I/Os with async request
2937          * submission to the client filesystem if supported.
2938          */
2939         io->async = async_dio;
2940         io->iocb = iocb;
2941
2942         /*
2943          * We cannot asynchronously extend the size of a file. We have no method
2944          * to wait on real async I/O requests, so we must submit this request
2945          * synchronously.
2946          */
2947         if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2948                 io->async = false;
2949
2950         if (rw == WRITE)
2951                 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2952         else
2953                 ret = __fuse_direct_read(io, iov, nr_segs, &pos, count);
2954
2955         if (io->async) {
2956                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2957
2958                 /* we have a non-extending, async request, so return */
2959                 if (!is_sync_kiocb(iocb))
2960                         return -EIOCBQUEUED;
2961
2962                 ret = wait_on_sync_kiocb(iocb);
2963         } else {
2964                 kfree(io);
2965         }
2966
2967         if (rw == WRITE) {
2968                 if (ret > 0)
2969                         fuse_write_update_size(inode, pos);
2970                 else if (ret < 0 && offset + count > i_size)
2971                         fuse_do_truncate(file);
2972         }
2973
2974         return ret;
2975 }
2976
2977 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2978                                 loff_t length)
2979 {
2980         struct fuse_file *ff = file->private_data;
2981         struct inode *inode = file->f_inode;
2982         struct fuse_inode *fi = get_fuse_inode(inode);
2983         struct fuse_conn *fc = ff->fc;
2984         struct fuse_req *req;
2985         struct fuse_fallocate_in inarg = {
2986                 .fh = ff->fh,
2987                 .offset = offset,
2988                 .length = length,
2989                 .mode = mode
2990         };
2991         int err;
2992         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2993                            (mode & FALLOC_FL_PUNCH_HOLE);
2994
2995         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2996                 return -EOPNOTSUPP;
2997
2998         if (fc->no_fallocate)
2999                 return -EOPNOTSUPP;
3000
3001         if (lock_inode) {
3002                 mutex_lock(&inode->i_mutex);
3003                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3004                         loff_t endbyte = offset + length - 1;
3005                         err = filemap_write_and_wait_range(inode->i_mapping,
3006                                                            offset, endbyte);
3007                         if (err)
3008                                 goto out;
3009
3010                         fuse_sync_writes(inode);
3011                 }
3012         }
3013
3014         if (!(mode & FALLOC_FL_KEEP_SIZE))
3015                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3016
3017         req = fuse_get_req_nopages(fc);
3018         if (IS_ERR(req)) {
3019                 err = PTR_ERR(req);
3020                 goto out;
3021         }
3022
3023         req->in.h.opcode = FUSE_FALLOCATE;
3024         req->in.h.nodeid = ff->nodeid;
3025         req->in.numargs = 1;
3026         req->in.args[0].size = sizeof(inarg);
3027         req->in.args[0].value = &inarg;
3028         fuse_request_send(fc, req);
3029         err = req->out.h.error;
3030         if (err == -ENOSYS) {
3031                 fc->no_fallocate = 1;
3032                 err = -EOPNOTSUPP;
3033         }
3034         fuse_put_request(fc, req);
3035
3036         if (err)
3037                 goto out;
3038
3039         /* we could have extended the file */
3040         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3041                 bool changed = fuse_write_update_size(inode, offset + length);
3042
3043                 if (changed && fc->writeback_cache)
3044                         file_update_time(file);
3045         }
3046
3047         if (mode & FALLOC_FL_PUNCH_HOLE)
3048                 truncate_pagecache_range(inode, offset, offset + length - 1);
3049
3050         fuse_invalidate_attr(inode);
3051
3052 out:
3053         if (!(mode & FALLOC_FL_KEEP_SIZE))
3054                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3055
3056         if (lock_inode)
3057                 mutex_unlock(&inode->i_mutex);
3058
3059         return err;
3060 }
3061
3062 static const struct file_operations fuse_file_operations = {
3063         .llseek         = fuse_file_llseek,
3064         .read           = do_sync_read,
3065         .aio_read       = fuse_file_aio_read,
3066         .write          = do_sync_write,
3067         .aio_write      = fuse_file_aio_write,
3068         .mmap           = fuse_file_mmap,
3069         .open           = fuse_open,
3070         .flush          = fuse_flush,
3071         .release        = fuse_release,
3072         .fsync          = fuse_fsync,
3073         .lock           = fuse_file_lock,
3074         .flock          = fuse_file_flock,
3075         .splice_read    = generic_file_splice_read,
3076         .unlocked_ioctl = fuse_file_ioctl,
3077         .compat_ioctl   = fuse_file_compat_ioctl,
3078         .poll           = fuse_file_poll,
3079         .fallocate      = fuse_file_fallocate,
3080 };
3081
3082 static const struct file_operations fuse_direct_io_file_operations = {
3083         .llseek         = fuse_file_llseek,
3084         .read           = fuse_direct_read,
3085         .write          = fuse_direct_write,
3086         .mmap           = fuse_direct_mmap,
3087         .open           = fuse_open,
3088         .flush          = fuse_flush,
3089         .release        = fuse_release,
3090         .fsync          = fuse_fsync,
3091         .lock           = fuse_file_lock,
3092         .flock          = fuse_file_flock,
3093         .unlocked_ioctl = fuse_file_ioctl,
3094         .compat_ioctl   = fuse_file_compat_ioctl,
3095         .poll           = fuse_file_poll,
3096         .fallocate      = fuse_file_fallocate,
3097         /* no splice_read */
3098 };
3099
3100 static const struct address_space_operations fuse_file_aops  = {
3101         .readpage       = fuse_readpage,
3102         .writepage      = fuse_writepage,
3103         .writepages     = fuse_writepages,
3104         .launder_page   = fuse_launder_page,
3105         .readpages      = fuse_readpages,
3106         .set_page_dirty = __set_page_dirty_nobuffers,
3107         .bmap           = fuse_bmap,
3108         .direct_IO      = fuse_direct_IO,
3109         .write_begin    = fuse_write_begin,
3110         .write_end      = fuse_write_end,
3111 };
3112
3113 void fuse_init_file_inode(struct inode *inode)
3114 {
3115         inode->i_fop = &fuse_file_operations;
3116         inode->i_data.a_ops = &fuse_file_aops;
3117 }