fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/compat.h>
  18 #include <linux/swap.h>
  19 #include <linux/falloc.h>
  20 #include <linux/uio.h>
  21 #include <linux/fs.h>
  22
  23 static const struct file_operations fuse_direct_io_file_operations;
  24
  25 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  26                           int opcode, struct fuse_open_out *outargp)
  27 {
  28         struct fuse_open_in inarg;
  29         FUSE_ARGS(args);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35         args.in.h.opcode = opcode;
  36         args.in.h.nodeid = nodeid;
  37         args.in.numargs = 1;
  38         args.in.args[0].size = sizeof(inarg);
  39         args.in.args[0].value = &inarg;
  40         args.out.numargs = 1;
  41         args.out.args[0].size = sizeof(*outargp);
  42         args.out.args[0].value = outargp;
  43
  44         return fuse_simple_request(fc, &args);
  45 }
  46
  47 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  48 {
  49         struct fuse_file *ff;
  50
  51         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  52         if (unlikely(!ff))
  53                 return NULL;
  54
  55         ff->fc = fc;
  56         ff->reserved_req = fuse_request_alloc(0);
  57         if (unlikely(!ff->reserved_req)) {
  58                 kfree(ff);
  59                 return NULL;
  60         }
  61
  62         INIT_LIST_HEAD(&ff->write_entry);
  63         refcount_set(&ff->count, 1);
  64         RB_CLEAR_NODE(&ff->polled_node);
  65         init_waitqueue_head(&ff->poll_wait);
  66
  67         spin_lock(&fc->lock);
  68         ff->kh = ++fc->khctr;
  69         spin_unlock(&fc->lock);
  70
  71         return ff;
  72 }
  73
  74 void fuse_file_free(struct fuse_file *ff)
  75 {
  76         fuse_request_free(ff->reserved_req);
  77         kfree(ff);
  78 }
  79
  80 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  81 {
  82         refcount_inc(&ff->count);
  83         return ff;
  84 }
  85
  86 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  87 {
  88         iput(req->misc.release.inode);
  89 }
  90
  91 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
  92 {
  93         if (refcount_dec_and_test(&ff->count)) {
  94                 struct fuse_req *req = ff->reserved_req;
  95
  96                 if (ff->fc->no_open && !isdir) {
  97                         /*
  98                          * Drop the release request when client does not
  99                          * implement 'open'
 100                          */
 101                         __clear_bit(FR_BACKGROUND, &req->flags);
 102                         iput(req->misc.release.inode);
 103                         fuse_put_request(ff->fc, req);
 104                 } else if (sync) {
 105                         __set_bit(FR_FORCE, &req->flags);
 106                         __clear_bit(FR_BACKGROUND, &req->flags);
 107                         fuse_request_send(ff->fc, req);
 108                         iput(req->misc.release.inode);
 109                         fuse_put_request(ff->fc, req);
 110                 } else {
 111                         req->end = fuse_release_end;
 112                         __set_bit(FR_BACKGROUND, &req->flags);
 113                         fuse_request_send_background(ff->fc, req);
 114                 }
 115                 kfree(ff);
 116         }
 117 }
 118
 119 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 120                  bool isdir)
 121 {
 122         struct fuse_file *ff;
 123         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 124
 125         ff = fuse_file_alloc(fc);
 126         if (!ff)
 127                 return -ENOMEM;
 128
 129         ff->fh = 0;
 130         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 131         if (!fc->no_open || isdir) {
 132                 struct fuse_open_out outarg;
 133                 int err;
 134
 135                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 136                 if (!err) {
 137                         ff->fh = outarg.fh;
 138                         ff->open_flags = outarg.open_flags;
 139
 140                 } else if (err != -ENOSYS || isdir) {
 141                         fuse_file_free(ff);
 142                         return err;
 143                 } else {
 144                         fc->no_open = 1;
 145                 }
 146         }
 147
 148         if (isdir)
 149                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 150
 151         ff->nodeid = nodeid;
 152         file->private_data = ff;
 153
 154         return 0;
 155 }
 156 EXPORT_SYMBOL_GPL(fuse_do_open);
 157
 158 static void fuse_link_write_file(struct file *file)
 159 {
 160         struct inode *inode = file_inode(file);
 161         struct fuse_conn *fc = get_fuse_conn(inode);
 162         struct fuse_inode *fi = get_fuse_inode(inode);
 163         struct fuse_file *ff = file->private_data;
 164         /*
 165          * file may be written through mmap, so chain it onto the
 166          * inodes's write_file list
 167          */
 168         spin_lock(&fc->lock);
 169         if (list_empty(&ff->write_entry))
 170                 list_add(&ff->write_entry, &fi->write_files);
 171         spin_unlock(&fc->lock);
 172 }
 173
 174 void fuse_finish_open(struct inode *inode, struct file *file)
 175 {
 176         struct fuse_file *ff = file->private_data;
 177         struct fuse_conn *fc = get_fuse_conn(inode);
 178
 179         if (ff->open_flags & FOPEN_DIRECT_IO)
 180                 file->f_op = &fuse_direct_io_file_operations;
 181         if (ff->open_flags & FOPEN_STREAM)
 182                 stream_open(inode, file);
 183         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 184                 nonseekable_open(inode, file);
 185
 186         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 187                 struct fuse_inode *fi = get_fuse_inode(inode);
 188
 189                 spin_lock(&fc->lock);
 190                 fi->attr_version = ++fc->attr_version;
 191                 i_size_write(inode, 0);
 192                 spin_unlock(&fc->lock);
 193                 truncate_pagecache(inode, 0);
 194                 fuse_invalidate_attr(inode);
 195                 if (fc->writeback_cache)
 196                         file_update_time(file);
 197         } else if (!(ff->open_flags & FOPEN_KEEP_CACHE)) {
 198                 invalidate_inode_pages2(inode->i_mapping);
 199         }
 200
 201         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 202                 fuse_link_write_file(file);
 203 }
 204
 205 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 206 {
 207         struct fuse_conn *fc = get_fuse_conn(inode);
 208         int err;
 209         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 210                           fc->atomic_o_trunc &&
 211                           fc->writeback_cache;
 212
 213         if (fuse_is_bad(inode))
 214                 return -EIO;
 215
 216         err = generic_file_open(inode, file);
 217         if (err)
 218                 return err;
 219
 220         if (is_wb_truncate) {
 221                 inode_lock(inode);
 222                 fuse_set_nowrite(inode);
 223         }
 224
 225         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 226
 227         if (!err)
 228                 fuse_finish_open(inode, file);
 229
 230         if (is_wb_truncate) {
 231                 fuse_release_nowrite(inode);
 232                 inode_unlock(inode);
 233         }
 234
 235         return err;
 236 }
 237
 238 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 239 {
 240         struct fuse_conn *fc = ff->fc;
 241         struct fuse_req *req = ff->reserved_req;
 242         struct fuse_release_in *inarg = &req->misc.release.in;
 243
 244         spin_lock(&fc->lock);
 245         list_del(&ff->write_entry);
 246         if (!RB_EMPTY_NODE(&ff->polled_node))
 247                 rb_erase(&ff->polled_node, &fc->polled_files);
 248         spin_unlock(&fc->lock);
 249
 250         wake_up_interruptible_all(&ff->poll_wait);
 251
 252         inarg->fh = ff->fh;
 253         inarg->flags = flags;
 254         req->in.h.opcode = opcode;
 255         req->in.h.nodeid = ff->nodeid;
 256         req->in.numargs = 1;
 257         req->in.args[0].size = sizeof(struct fuse_release_in);
 258         req->in.args[0].value = inarg;
 259 }
 260
 261 void fuse_release_common(struct file *file, bool isdir)
 262 {
 263         struct fuse_file *ff = file->private_data;
 264         struct fuse_req *req = ff->reserved_req;
 265         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 266
 267         fuse_prepare_release(ff, file->f_flags, opcode);
 268
 269         if (ff->flock) {
 270                 struct fuse_release_in *inarg = &req->misc.release.in;
 271                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 272                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 273                                                        (fl_owner_t) file);
 274         }
 275         /* Hold inode until release is finished */
 276         req->misc.release.inode = igrab(file_inode(file));
 277
 278         /*
 279          * Normally this will send the RELEASE request, however if
 280          * some asynchronous READ or WRITE requests are outstanding,
 281          * the sending will be delayed.
 282          *
 283          * Make the release synchronous if this is a fuseblk mount,
 284          * synchronous RELEASE is allowed (and desirable) in this case
 285          * because the server can be trusted not to screw up.
 286          */
 287         fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
 288 }
 289
 290 static int fuse_open(struct inode *inode, struct file *file)
 291 {
 292         return fuse_open_common(inode, file, false);
 293 }
 294
 295 static int fuse_release(struct inode *inode, struct file *file)
 296 {
 297         struct fuse_conn *fc = get_fuse_conn(inode);
 298
 299         /* see fuse_vma_close() for !writeback_cache case */
 300         if (fc->writeback_cache)
 301                 write_inode_now(inode, 1);
 302
 303         fuse_release_common(file, false);
 304
 305         /* return value is ignored by VFS */
 306         return 0;
 307 }
 308
 309 void fuse_sync_release(struct fuse_file *ff, int flags)
 310 {
 311         WARN_ON(refcount_read(&ff->count) > 1);
 312         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 313         /*
 314          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 315          * synchronous, we are fine with not doing igrab() here"
 316          */
 317         fuse_file_put(ff, true, false);
 318 }
 319 EXPORT_SYMBOL_GPL(fuse_sync_release);
 320
 321 /*
 322  * Scramble the ID space with XTEA, so that the value of the files_struct
 323  * pointer is not exposed to userspace.
 324  */
 325 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 326 {
 327         u32 *k = fc->scramble_key;
 328         u64 v = (unsigned long) id;
 329         u32 v0 = v;
 330         u32 v1 = v >> 32;
 331         u32 sum = 0;
 332         int i;
 333
 334         for (i = 0; i < 32; i++) {
 335                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 336                 sum += 0x9E3779B9;
 337                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 338         }
 339
 340         return (u64) v0 + ((u64) v1 << 32);
 341 }
 342
 343 /*
 344  * Check if any page in a range is under writeback
 345  *
 346  * This is currently done by walking the list of writepage requests
 347  * for the inode, which can be pretty inefficient.
 348  */
 349 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 350                                    pgoff_t idx_to)
 351 {
 352         struct fuse_conn *fc = get_fuse_conn(inode);
 353         struct fuse_inode *fi = get_fuse_inode(inode);
 354         struct fuse_req *req;
 355         bool found = false;
 356
 357         spin_lock(&fc->lock);
 358         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 359                 pgoff_t curr_index;
 360
 361                 BUG_ON(req->inode != inode);
 362                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 363                 if (idx_from < curr_index + req->num_pages &&
 364                     curr_index <= idx_to) {
 365                         found = true;
 366                         break;
 367                 }
 368         }
 369         spin_unlock(&fc->lock);
 370
 371         return found;
 372 }
 373
 374 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 375 {
 376         return fuse_range_is_writeback(inode, index, index);
 377 }
 378
 379 /*
 380  * Wait for page writeback to be completed.
 381  *
 382  * Since fuse doesn't rely on the VM writeback tracking, this has to
 383  * use some other means.
 384  */
 385 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 386 {
 387         struct fuse_inode *fi = get_fuse_inode(inode);
 388
 389         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 390         return 0;
 391 }
 392
 393 /*
 394  * Wait for all pending writepages on the inode to finish.
 395  *
 396  * This is currently done by blocking further writes with FUSE_NOWRITE
 397  * and waiting for all sent writes to complete.
 398  *
 399  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 400  * could conflict with truncation.
 401  */
 402 static void fuse_sync_writes(struct inode *inode)
 403 {
 404         fuse_set_nowrite(inode);
 405         fuse_release_nowrite(inode);
 406 }
 407
 408 static int fuse_flush(struct file *file, fl_owner_t id)
 409 {
 410         struct inode *inode = file_inode(file);
 411         struct fuse_conn *fc = get_fuse_conn(inode);
 412         struct fuse_file *ff = file->private_data;
 413         struct fuse_req *req;
 414         struct fuse_flush_in inarg;
 415         int err;
 416
 417         if (fuse_is_bad(inode))
 418                 return -EIO;
 419
 420         if (fc->no_flush)
 421                 return 0;
 422
 423         err = write_inode_now(inode, 1);
 424         if (err)
 425                 return err;
 426
 427         inode_lock(inode);
 428         fuse_sync_writes(inode);
 429         inode_unlock(inode);
 430
 431         err = filemap_check_errors(file->f_mapping);
 432         if (err)
 433                 return err;
 434
 435         req = fuse_get_req_nofail_nopages(fc, file);
 436         memset(&inarg, 0, sizeof(inarg));
 437         inarg.fh = ff->fh;
 438         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 439         req->in.h.opcode = FUSE_FLUSH;
 440         req->in.h.nodeid = get_node_id(inode);
 441         req->in.numargs = 1;
 442         req->in.args[0].size = sizeof(inarg);
 443         req->in.args[0].value = &inarg;
 444         __set_bit(FR_FORCE, &req->flags);
 445         fuse_request_send(fc, req);
 446         err = req->out.h.error;
 447         fuse_put_request(fc, req);
 448         if (err == -ENOSYS) {
 449                 fc->no_flush = 1;
 450                 err = 0;
 451         }
 452         return err;
 453 }
 454
 455 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 456                       int datasync, int isdir)
 457 {
 458         struct inode *inode = file->f_mapping->host;
 459         struct fuse_conn *fc = get_fuse_conn(inode);
 460         struct fuse_file *ff = file->private_data;
 461         FUSE_ARGS(args);
 462         struct fuse_fsync_in inarg;
 463         int err;
 464
 465         if (fuse_is_bad(inode))
 466                 return -EIO;
 467
 468         inode_lock(inode);
 469
 470         /*
 471          * Start writeback against all dirty pages of the inode, then
 472          * wait for all outstanding writes, before sending the FSYNC
 473          * request.
 474          */
 475         err = file_write_and_wait_range(file, start, end);
 476         if (err)
 477                 goto out;
 478
 479         fuse_sync_writes(inode);
 480
 481         /*
 482          * Due to implementation of fuse writeback
 483          * file_write_and_wait_range() does not catch errors.
 484          * We have to do this directly after fuse_sync_writes()
 485          */
 486         err = file_check_and_advance_wb_err(file);
 487         if (err)
 488                 goto out;
 489
 490         err = sync_inode_metadata(inode, 1);
 491         if (err)
 492                 goto out;
 493
 494         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 495                 goto out;
 496
 497         memset(&inarg, 0, sizeof(inarg));
 498         inarg.fh = ff->fh;
 499         inarg.fsync_flags = datasync ? 1 : 0;
 500         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 501         args.in.h.nodeid = get_node_id(inode);
 502         args.in.numargs = 1;
 503         args.in.args[0].size = sizeof(inarg);
 504         args.in.args[0].value = &inarg;
 505         err = fuse_simple_request(fc, &args);
 506         if (err == -ENOSYS) {
 507                 if (isdir)
 508                         fc->no_fsyncdir = 1;
 509                 else
 510                         fc->no_fsync = 1;
 511                 err = 0;
 512         }
 513 out:
 514         inode_unlock(inode);
 515         return err;
 516 }
 517
 518 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 519                       int datasync)
 520 {
 521         return fuse_fsync_common(file, start, end, datasync, 0);
 522 }
 523
 524 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 525                     size_t count, int opcode)
 526 {
 527         struct fuse_read_in *inarg = &req->misc.read.in;
 528         struct fuse_file *ff = file->private_data;
 529
 530         inarg->fh = ff->fh;
 531         inarg->offset = pos;
 532         inarg->size = count;
 533         inarg->flags = file->f_flags;
 534         req->in.h.opcode = opcode;
 535         req->in.h.nodeid = ff->nodeid;
 536         req->in.numargs = 1;
 537         req->in.args[0].size = sizeof(struct fuse_read_in);
 538         req->in.args[0].value = inarg;
 539         req->out.argvar = 1;
 540         req->out.numargs = 1;
 541         req->out.args[0].size = count;
 542 }
 543
 544 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 545 {
 546         unsigned i;
 547
 548         for (i = 0; i < req->num_pages; i++) {
 549                 struct page *page = req->pages[i];
 550                 if (should_dirty)
 551                         set_page_dirty_lock(page);
 552                 put_page(page);
 553         }
 554 }
 555
 556 static void fuse_io_release(struct kref *kref)
 557 {
 558         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 559 }
 560
 561 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 562 {
 563         if (io->err)
 564                 return io->err;
 565
 566         if (io->bytes >= 0 && io->write)
 567                 return -EIO;
 568
 569         return io->bytes < 0 ? io->size : io->bytes;
 570 }
 571
 572 /**
 573  * In case of short read, the caller sets 'pos' to the position of
 574  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 575  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 576  *
 577  * An example:
 578  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 579  * both submitted asynchronously. The first of them was ACKed by userspace as
 580  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 581  * second request was ACKed as short, e.g. only 1K was read, resulting in
 582  * pos == 33K.
 583  *
 584  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 585  * will be equal to the length of the longest contiguous fragment of
 586  * transferred data starting from the beginning of IO request.
 587  */
 588 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 589 {
 590         int left;
 591
 592         spin_lock(&io->lock);
 593         if (err)
 594                 io->err = io->err ? : err;
 595         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 596                 io->bytes = pos;
 597
 598         left = --io->reqs;
 599         if (!left && io->blocking)
 600                 complete(io->done);
 601         spin_unlock(&io->lock);
 602
 603         if (!left && !io->blocking) {
 604                 ssize_t res = fuse_get_res_by_io(io);
 605
 606                 if (res >= 0) {
 607                         struct inode *inode = file_inode(io->iocb->ki_filp);
 608                         struct fuse_conn *fc = get_fuse_conn(inode);
 609                         struct fuse_inode *fi = get_fuse_inode(inode);
 610
 611                         spin_lock(&fc->lock);
 612                         fi->attr_version = ++fc->attr_version;
 613                         spin_unlock(&fc->lock);
 614                 }
 615
 616                 io->iocb->ki_complete(io->iocb, res, 0);
 617         }
 618
 619         kref_put(&io->refcnt, fuse_io_release);
 620 }
 621
 622 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 623 {
 624         struct fuse_io_priv *io = req->io;
 625         ssize_t pos = -1;
 626
 627         fuse_release_user_pages(req, io->should_dirty);
 628
 629         if (io->write) {
 630                 if (req->misc.write.in.size != req->misc.write.out.size)
 631                         pos = req->misc.write.in.offset - io->offset +
 632                                 req->misc.write.out.size;
 633         } else {
 634                 if (req->misc.read.in.size != req->out.args[0].size)
 635                         pos = req->misc.read.in.offset - io->offset +
 636                                 req->out.args[0].size;
 637         }
 638
 639         fuse_aio_complete(io, req->out.h.error, pos);
 640 }
 641
 642 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 643                 size_t num_bytes, struct fuse_io_priv *io)
 644 {
 645         spin_lock(&io->lock);
 646         kref_get(&io->refcnt);
 647         io->size += num_bytes;
 648         io->reqs++;
 649         spin_unlock(&io->lock);
 650
 651         req->io = io;
 652         req->end = fuse_aio_complete_req;
 653
 654         __fuse_get_request(req);
 655         fuse_request_send_background(fc, req);
 656
 657         return num_bytes;
 658 }
 659
 660 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 661                              loff_t pos, size_t count, fl_owner_t owner)
 662 {
 663         struct file *file = io->iocb->ki_filp;
 664         struct fuse_file *ff = file->private_data;
 665         struct fuse_conn *fc = ff->fc;
 666
 667         fuse_read_fill(req, file, pos, count, FUSE_READ);
 668         if (owner != NULL) {
 669                 struct fuse_read_in *inarg = &req->misc.read.in;
 670
 671                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 672                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 673         }
 674
 675         if (io->async)
 676                 return fuse_async_req_send(fc, req, count, io);
 677
 678         fuse_request_send(fc, req);
 679         return req->out.args[0].size;
 680 }
 681
 682 static void fuse_read_update_size(struct inode *inode, loff_t size,
 683                                   u64 attr_ver)
 684 {
 685         struct fuse_conn *fc = get_fuse_conn(inode);
 686         struct fuse_inode *fi = get_fuse_inode(inode);
 687
 688         spin_lock(&fc->lock);
 689         if (attr_ver == fi->attr_version && size < inode->i_size &&
 690             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 691                 fi->attr_version = ++fc->attr_version;
 692                 i_size_write(inode, size);
 693         }
 694         spin_unlock(&fc->lock);
 695 }
 696
 697 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 698                             u64 attr_ver)
 699 {
 700         size_t num_read = req->out.args[0].size;
 701         struct fuse_conn *fc = get_fuse_conn(inode);
 702
 703         if (fc->writeback_cache) {
 704                 /*
 705                  * A hole in a file. Some data after the hole are in page cache,
 706                  * but have not reached the client fs yet. So, the hole is not
 707                  * present there.
 708                  */
 709                 int i;
 710                 int start_idx = num_read >> PAGE_SHIFT;
 711                 size_t off = num_read & (PAGE_SIZE - 1);
 712
 713                 for (i = start_idx; i < req->num_pages; i++) {
 714                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 715                         off = 0;
 716                 }
 717         } else {
 718                 loff_t pos = page_offset(req->pages[0]) + num_read;
 719                 fuse_read_update_size(inode, pos, attr_ver);
 720         }
 721 }
 722
 723 static int fuse_do_readpage(struct file *file, struct page *page)
 724 {
 725         struct kiocb iocb;
 726         struct fuse_io_priv io;
 727         struct inode *inode = page->mapping->host;
 728         struct fuse_conn *fc = get_fuse_conn(inode);
 729         struct fuse_req *req;
 730         size_t num_read;
 731         loff_t pos = page_offset(page);
 732         size_t count = PAGE_SIZE;
 733         u64 attr_ver;
 734         int err;
 735
 736         /*
 737          * Page writeback can extend beyond the lifetime of the
 738          * page-cache page, so make sure we read a properly synced
 739          * page.
 740          */
 741         fuse_wait_on_page_writeback(inode, page->index);
 742
 743         req = fuse_get_req(fc, 1);
 744         if (IS_ERR(req))
 745                 return PTR_ERR(req);
 746
 747         attr_ver = fuse_get_attr_version(fc);
 748
 749         req->out.page_zeroing = 1;
 750         req->out.argpages = 1;
 751         req->num_pages = 1;
 752         req->pages[0] = page;
 753         req->page_descs[0].length = count;
 754         init_sync_kiocb(&iocb, file);
 755         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 756         num_read = fuse_send_read(req, &io, pos, count, NULL);
 757         err = req->out.h.error;
 758
 759         if (!err) {
 760                 /*
 761                  * Short read means EOF.  If file size is larger, truncate it
 762                  */
 763                 if (num_read < count)
 764                         fuse_short_read(req, inode, attr_ver);
 765
 766                 SetPageUptodate(page);
 767         }
 768
 769         fuse_put_request(fc, req);
 770
 771         return err;
 772 }
 773
 774 static int fuse_readpage(struct file *file, struct page *page)
 775 {
 776         struct inode *inode = page->mapping->host;
 777         int err;
 778
 779         err = -EIO;
 780         if (fuse_is_bad(inode))
 781                 goto out;
 782
 783         err = fuse_do_readpage(file, page);
 784         fuse_invalidate_atime(inode);
 785  out:
 786         unlock_page(page);
 787         return err;
 788 }
 789
 790 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 791 {
 792         int i;
 793         size_t count = req->misc.read.in.size;
 794         size_t num_read = req->out.args[0].size;
 795         struct address_space *mapping = NULL;
 796
 797         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 798                 mapping = req->pages[i]->mapping;
 799
 800         if (mapping) {
 801                 struct inode *inode = mapping->host;
 802
 803                 /*
 804                  * Short read means EOF. If file size is larger, truncate it
 805                  */
 806                 if (!req->out.h.error && num_read < count)
 807                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 808
 809                 fuse_invalidate_atime(inode);
 810         }
 811
 812         for (i = 0; i < req->num_pages; i++) {
 813                 struct page *page = req->pages[i];
 814                 if (!req->out.h.error)
 815                         SetPageUptodate(page);
 816                 else
 817                         SetPageError(page);
 818                 unlock_page(page);
 819                 put_page(page);
 820         }
 821         if (req->ff)
 822                 fuse_file_put(req->ff, false, false);
 823 }
 824
 825 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 826 {
 827         struct fuse_file *ff = file->private_data;
 828         struct fuse_conn *fc = ff->fc;
 829         loff_t pos = page_offset(req->pages[0]);
 830         size_t count = req->num_pages << PAGE_SHIFT;
 831
 832         req->out.argpages = 1;
 833         req->out.page_zeroing = 1;
 834         req->out.page_replace = 1;
 835         fuse_read_fill(req, file, pos, count, FUSE_READ);
 836         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 837         if (fc->async_read) {
 838                 req->ff = fuse_file_get(ff);
 839                 req->end = fuse_readpages_end;
 840                 fuse_request_send_background(fc, req);
 841         } else {
 842                 fuse_request_send(fc, req);
 843                 fuse_readpages_end(fc, req);
 844                 fuse_put_request(fc, req);
 845         }
 846 }
 847
 848 struct fuse_fill_data {
 849         struct fuse_req *req;
 850         struct file *file;
 851         struct inode *inode;
 852         unsigned nr_pages;
 853 };
 854
 855 static int fuse_readpages_fill(void *_data, struct page *page)
 856 {
 857         struct fuse_fill_data *data = _data;
 858         struct fuse_req *req = data->req;
 859         struct inode *inode = data->inode;
 860         struct fuse_conn *fc = get_fuse_conn(inode);
 861
 862         fuse_wait_on_page_writeback(inode, page->index);
 863
 864         if (req->num_pages &&
 865             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 866              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 867              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 868                 int nr_alloc = min_t(unsigned, data->nr_pages,
 869                                      FUSE_MAX_PAGES_PER_REQ);
 870                 fuse_send_readpages(req, data->file);
 871                 if (fc->async_read)
 872                         req = fuse_get_req_for_background(fc, nr_alloc);
 873                 else
 874                         req = fuse_get_req(fc, nr_alloc);
 875
 876                 data->req = req;
 877                 if (IS_ERR(req)) {
 878                         unlock_page(page);
 879                         return PTR_ERR(req);
 880                 }
 881         }
 882
 883         if (WARN_ON(req->num_pages >= req->max_pages)) {
 884                 unlock_page(page);
 885                 fuse_put_request(fc, req);
 886                 return -EIO;
 887         }
 888
 889         get_page(page);
 890         req->pages[req->num_pages] = page;
 891         req->page_descs[req->num_pages].length = PAGE_SIZE;
 892         req->num_pages++;
 893         data->nr_pages--;
 894         return 0;
 895 }
 896
 897 static int fuse_readpages(struct file *file, struct address_space *mapping,
 898                           struct list_head *pages, unsigned nr_pages)
 899 {
 900         struct inode *inode = mapping->host;
 901         struct fuse_conn *fc = get_fuse_conn(inode);
 902         struct fuse_fill_data data;
 903         int err;
 904         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 905
 906         err = -EIO;
 907         if (fuse_is_bad(inode))
 908                 goto out;
 909
 910         data.file = file;
 911         data.inode = inode;
 912         if (fc->async_read)
 913                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 914         else
 915                 data.req = fuse_get_req(fc, nr_alloc);
 916         data.nr_pages = nr_pages;
 917         err = PTR_ERR(data.req);
 918         if (IS_ERR(data.req))
 919                 goto out;
 920
 921         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 922         if (!err) {
 923                 if (data.req->num_pages)
 924                         fuse_send_readpages(data.req, file);
 925                 else
 926                         fuse_put_request(fc, data.req);
 927         }
 928 out:
 929         return err;
 930 }
 931
 932 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 933 {
 934         struct inode *inode = iocb->ki_filp->f_mapping->host;
 935         struct fuse_conn *fc = get_fuse_conn(inode);
 936
 937         if (fuse_is_bad(inode))
 938                 return -EIO;
 939
 940         /*
 941          * In auto invalidate mode, always update attributes on read.
 942          * Otherwise, only update if we attempt to read past EOF (to ensure
 943          * i_size is up to date).
 944          */
 945         if (fc->auto_inval_data ||
 946             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 947                 int err;
 948                 err = fuse_update_attributes(inode, iocb->ki_filp);
 949                 if (err)
 950                         return err;
 951         }
 952
 953         return generic_file_read_iter(iocb, to);
 954 }
 955
 956 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 957                             loff_t pos, size_t count)
 958 {
 959         struct fuse_write_in *inarg = &req->misc.write.in;
 960         struct fuse_write_out *outarg = &req->misc.write.out;
 961
 962         inarg->fh = ff->fh;
 963         inarg->offset = pos;
 964         inarg->size = count;
 965         req->in.h.opcode = FUSE_WRITE;
 966         req->in.h.nodeid = ff->nodeid;
 967         req->in.numargs = 2;
 968         if (ff->fc->minor < 9)
 969                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 970         else
 971                 req->in.args[0].size = sizeof(struct fuse_write_in);
 972         req->in.args[0].value = inarg;
 973         req->in.args[1].size = count;
 974         req->out.numargs = 1;
 975         req->out.args[0].size = sizeof(struct fuse_write_out);
 976         req->out.args[0].value = outarg;
 977 }
 978
 979 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 980                               loff_t pos, size_t count, fl_owner_t owner)
 981 {
 982         struct kiocb *iocb = io->iocb;
 983         struct file *file = iocb->ki_filp;
 984         struct fuse_file *ff = file->private_data;
 985         struct fuse_conn *fc = ff->fc;
 986         struct fuse_write_in *inarg = &req->misc.write.in;
 987
 988         fuse_write_fill(req, ff, pos, count);
 989         inarg->flags = file->f_flags;
 990         if (iocb->ki_flags & IOCB_DSYNC)
 991                 inarg->flags |= O_DSYNC;
 992         if (iocb->ki_flags & IOCB_SYNC)
 993                 inarg->flags |= O_SYNC;
 994         if (owner != NULL) {
 995                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 996                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 997         }
 998
 999         if (io->async)
1000                 return fuse_async_req_send(fc, req, count, io);
1001
1002         fuse_request_send(fc, req);
1003         return req->misc.write.out.size;
1004 }
1005
1006 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1007 {
1008         struct fuse_conn *fc = get_fuse_conn(inode);
1009         struct fuse_inode *fi = get_fuse_inode(inode);
1010         bool ret = false;
1011
1012         spin_lock(&fc->lock);
1013         fi->attr_version = ++fc->attr_version;
1014         if (pos > inode->i_size) {
1015                 i_size_write(inode, pos);
1016                 ret = true;
1017         }
1018         spin_unlock(&fc->lock);
1019
1020         return ret;
1021 }
1022
1023 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1024                                     struct inode *inode, loff_t pos,
1025                                     size_t count)
1026 {
1027         size_t res;
1028         unsigned offset;
1029         unsigned i;
1030         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1031
1032         for (i = 0; i < req->num_pages; i++)
1033                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1034
1035         res = fuse_send_write(req, &io, pos, count, NULL);
1036
1037         offset = req->page_descs[0].offset;
1038         count = res;
1039         for (i = 0; i < req->num_pages; i++) {
1040                 struct page *page = req->pages[i];
1041
1042                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1043                         SetPageUptodate(page);
1044
1045                 if (count > PAGE_SIZE - offset)
1046                         count -= PAGE_SIZE - offset;
1047                 else
1048                         count = 0;
1049                 offset = 0;
1050
1051                 unlock_page(page);
1052                 put_page(page);
1053         }
1054
1055         return res;
1056 }
1057
1058 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1059                                struct address_space *mapping,
1060                                struct iov_iter *ii, loff_t pos)
1061 {
1062         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1063         unsigned offset = pos & (PAGE_SIZE - 1);
1064         size_t count = 0;
1065         int err;
1066
1067         req->in.argpages = 1;
1068         req->page_descs[0].offset = offset;
1069
1070         do {
1071                 size_t tmp;
1072                 struct page *page;
1073                 pgoff_t index = pos >> PAGE_SHIFT;
1074                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1075                                      iov_iter_count(ii));
1076
1077                 bytes = min_t(size_t, bytes, fc->max_write - count);
1078
1079  again:
1080                 err = -EFAULT;
1081                 if (iov_iter_fault_in_readable(ii, bytes))
1082                         break;
1083
1084                 err = -ENOMEM;
1085                 page = grab_cache_page_write_begin(mapping, index, 0);
1086                 if (!page)
1087                         break;
1088
1089                 if (mapping_writably_mapped(mapping))
1090                         flush_dcache_page(page);
1091
1092                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1093                 flush_dcache_page(page);
1094
1095                 iov_iter_advance(ii, tmp);
1096                 if (!tmp) {
1097                         unlock_page(page);
1098                         put_page(page);
1099                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1100                         goto again;
1101                 }
1102
1103                 err = 0;
1104                 req->pages[req->num_pages] = page;
1105                 req->page_descs[req->num_pages].length = tmp;
1106                 req->num_pages++;
1107
1108                 count += tmp;
1109                 pos += tmp;
1110                 offset += tmp;
1111                 if (offset == PAGE_SIZE)
1112                         offset = 0;
1113
1114                 if (!fc->big_writes)
1115                         break;
1116         } while (iov_iter_count(ii) && count < fc->max_write &&
1117                  req->num_pages < req->max_pages && offset == 0);
1118
1119         return count > 0 ? count : err;
1120 }
1121
1122 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1123 {
1124         return min_t(unsigned,
1125                      ((pos + len - 1) >> PAGE_SHIFT) -
1126                      (pos >> PAGE_SHIFT) + 1,
1127                      FUSE_MAX_PAGES_PER_REQ);
1128 }
1129
1130 static ssize_t fuse_perform_write(struct kiocb *iocb,
1131                                   struct address_space *mapping,
1132                                   struct iov_iter *ii, loff_t pos)
1133 {
1134         struct inode *inode = mapping->host;
1135         struct fuse_conn *fc = get_fuse_conn(inode);
1136         struct fuse_inode *fi = get_fuse_inode(inode);
1137         int err = 0;
1138         ssize_t res = 0;
1139
1140         if (fuse_is_bad(inode))
1141                 return -EIO;
1142
1143         if (inode->i_size < pos + iov_iter_count(ii))
1144                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1145
1146         do {
1147                 struct fuse_req *req;
1148                 ssize_t count;
1149                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1150
1151                 req = fuse_get_req(fc, nr_pages);
1152                 if (IS_ERR(req)) {
1153                         err = PTR_ERR(req);
1154                         break;
1155                 }
1156
1157                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1158                 if (count <= 0) {
1159                         err = count;
1160                 } else {
1161                         size_t num_written;
1162
1163                         num_written = fuse_send_write_pages(req, iocb, inode,
1164                                                             pos, count);
1165                         err = req->out.h.error;
1166                         if (!err) {
1167                                 res += num_written;
1168                                 pos += num_written;
1169
1170                                 /* break out of the loop on short write */
1171                                 if (num_written != count)
1172                                         err = -EIO;
1173                         }
1174                 }
1175                 fuse_put_request(fc, req);
1176         } while (!err && iov_iter_count(ii));
1177
1178         if (res > 0)
1179                 fuse_write_update_size(inode, pos);
1180
1181         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1182         fuse_invalidate_attr(inode);
1183
1184         return res > 0 ? res : err;
1185 }
1186
1187 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1188 {
1189         struct file *file = iocb->ki_filp;
1190         struct address_space *mapping = file->f_mapping;
1191         ssize_t written = 0;
1192         ssize_t written_buffered = 0;
1193         struct inode *inode = mapping->host;
1194         ssize_t err;
1195         loff_t endbyte = 0;
1196
1197         if (fuse_is_bad(inode))
1198                 return -EIO;
1199
1200         if (get_fuse_conn(inode)->writeback_cache) {
1201                 /* Update size (EOF optimization) and mode (SUID clearing) */
1202                 err = fuse_update_attributes(mapping->host, file);
1203                 if (err)
1204                         return err;
1205
1206                 return generic_file_write_iter(iocb, from);
1207         }
1208
1209         inode_lock(inode);
1210
1211         /* We can write back this queue in page reclaim */
1212         current->backing_dev_info = inode_to_bdi(inode);
1213
1214         err = generic_write_checks(iocb, from);
1215         if (err <= 0)
1216                 goto out;
1217
1218         err = file_remove_privs(file);
1219         if (err)
1220                 goto out;
1221
1222         err = file_update_time(file);
1223         if (err)
1224                 goto out;
1225
1226         if (iocb->ki_flags & IOCB_DIRECT) {
1227                 loff_t pos = iocb->ki_pos;
1228                 written = generic_file_direct_write(iocb, from);
1229                 if (written < 0 || !iov_iter_count(from))
1230                         goto out;
1231
1232                 pos += written;
1233
1234                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1235                 if (written_buffered < 0) {
1236                         err = written_buffered;
1237                         goto out;
1238                 }
1239                 endbyte = pos + written_buffered - 1;
1240
1241                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1242                                                    endbyte);
1243                 if (err)
1244                         goto out;
1245
1246                 invalidate_mapping_pages(file->f_mapping,
1247                                          pos >> PAGE_SHIFT,
1248                                          endbyte >> PAGE_SHIFT);
1249
1250                 written += written_buffered;
1251                 iocb->ki_pos = pos + written_buffered;
1252         } else {
1253                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1254                 if (written >= 0)
1255                         iocb->ki_pos += written;
1256         }
1257 out:
1258         current->backing_dev_info = NULL;
1259         inode_unlock(inode);
1260         if (written > 0)
1261                 written = generic_write_sync(iocb, written);
1262
1263         return written ? written : err;
1264 }
1265
1266 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1267                 unsigned index, unsigned nr_pages)
1268 {
1269         int i;
1270
1271         for (i = index; i < index + nr_pages; i++)
1272                 req->page_descs[i].length = PAGE_SIZE -
1273                         req->page_descs[i].offset;
1274 }
1275
1276 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1277 {
1278         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1279 }
1280
1281 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1282                                         size_t max_size)
1283 {
1284         return min(iov_iter_single_seg_count(ii), max_size);
1285 }
1286
1287 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1288                                size_t *nbytesp, int write)
1289 {
1290         size_t nbytes = 0;  /* # bytes already packed in req */
1291         ssize_t ret = 0;
1292
1293         /* Special case for kernel I/O: can copy directly into the buffer */
1294         if (ii->type & ITER_KVEC) {
1295                 unsigned long user_addr = fuse_get_user_addr(ii);
1296                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1297
1298                 if (write)
1299                         req->in.args[1].value = (void *) user_addr;
1300                 else
1301                         req->out.args[0].value = (void *) user_addr;
1302
1303                 iov_iter_advance(ii, frag_size);
1304                 *nbytesp = frag_size;
1305                 return 0;
1306         }
1307
1308         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1309                 unsigned npages;
1310                 size_t start;
1311                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1312                                         *nbytesp - nbytes,
1313                                         req->max_pages - req->num_pages,
1314                                         &start);
1315                 if (ret < 0)
1316                         break;
1317
1318                 iov_iter_advance(ii, ret);
1319                 nbytes += ret;
1320
1321                 ret += start;
1322                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1323
1324                 req->page_descs[req->num_pages].offset = start;
1325                 fuse_page_descs_length_init(req, req->num_pages, npages);
1326
1327                 req->num_pages += npages;
1328                 req->page_descs[req->num_pages - 1].length -=
1329                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1330         }
1331
1332         req->user_pages = true;
1333         if (write)
1334                 req->in.argpages = 1;
1335         else
1336                 req->out.argpages = 1;
1337
1338         *nbytesp = nbytes;
1339
1340         return ret < 0 ? ret : 0;
1341 }
1342
1343 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1344 {
1345         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1346 }
1347
1348 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1349                        loff_t *ppos, int flags)
1350 {
1351         int write = flags & FUSE_DIO_WRITE;
1352         int cuse = flags & FUSE_DIO_CUSE;
1353         struct file *file = io->iocb->ki_filp;
1354         struct inode *inode = file->f_mapping->host;
1355         struct fuse_file *ff = file->private_data;
1356         struct fuse_conn *fc = ff->fc;
1357         size_t nmax = write ? fc->max_write : fc->max_read;
1358         loff_t pos = *ppos;
1359         size_t count = iov_iter_count(iter);
1360         pgoff_t idx_from = pos >> PAGE_SHIFT;
1361         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1362         ssize_t res = 0;
1363         struct fuse_req *req;
1364         int err = 0;
1365
1366         if (io->async)
1367                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1368         else
1369                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1370         if (IS_ERR(req))
1371                 return PTR_ERR(req);
1372
1373         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1374                 if (!write)
1375                         inode_lock(inode);
1376                 fuse_sync_writes(inode);
1377                 if (!write)
1378                         inode_unlock(inode);
1379         }
1380
1381         io->should_dirty = !write && iter_is_iovec(iter);
1382         while (count) {
1383                 size_t nres;
1384                 fl_owner_t owner = current->files;
1385                 size_t nbytes = min(count, nmax);
1386                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1387                 if (err && !nbytes)
1388                         break;
1389
1390                 if (write)
1391                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1392                 else
1393                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1394
1395                 if (!io->async)
1396                         fuse_release_user_pages(req, io->should_dirty);
1397                 if (req->out.h.error) {
1398                         err = req->out.h.error;
1399                         break;
1400                 } else if (nres > nbytes) {
1401                         res = 0;
1402                         err = -EIO;
1403                         break;
1404                 }
1405                 count -= nres;
1406                 res += nres;
1407                 pos += nres;
1408                 if (nres != nbytes)
1409                         break;
1410                 if (count) {
1411                         fuse_put_request(fc, req);
1412                         if (io->async)
1413                                 req = fuse_get_req_for_background(fc,
1414                                         fuse_iter_npages(iter));
1415                         else
1416                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1417                         if (IS_ERR(req))
1418                                 break;
1419                 }
1420         }
1421         if (!IS_ERR(req))
1422                 fuse_put_request(fc, req);
1423         if (res > 0)
1424                 *ppos = pos;
1425
1426         return res > 0 ? res : err;
1427 }
1428 EXPORT_SYMBOL_GPL(fuse_direct_io);
1429
1430 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1431                                   struct iov_iter *iter,
1432                                   loff_t *ppos)
1433 {
1434         ssize_t res;
1435         struct inode *inode = file_inode(io->iocb->ki_filp);
1436
1437         if (fuse_is_bad(inode))
1438                 return -EIO;
1439
1440         res = fuse_direct_io(io, iter, ppos, 0);
1441
1442         fuse_invalidate_attr(inode);
1443
1444         return res;
1445 }
1446
1447 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1448 {
1449         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1450         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1451 }
1452
1453 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1454 {
1455         struct inode *inode = file_inode(iocb->ki_filp);
1456         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1457         ssize_t res;
1458
1459         if (fuse_is_bad(inode))
1460                 return -EIO;
1461
1462         /* Don't allow parallel writes to the same file */
1463         inode_lock(inode);
1464         res = generic_write_checks(iocb, from);
1465         if (res > 0)
1466                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1467         fuse_invalidate_attr(inode);
1468         if (res > 0)
1469                 fuse_write_update_size(inode, iocb->ki_pos);
1470         inode_unlock(inode);
1471
1472         return res;
1473 }
1474
1475 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1476 {
1477         int i;
1478
1479         for (i = 0; i < req->num_pages; i++)
1480                 __free_page(req->pages[i]);
1481
1482         if (req->ff)
1483                 fuse_file_put(req->ff, false, false);
1484 }
1485
1486 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1487 {
1488         struct inode *inode = req->inode;
1489         struct fuse_inode *fi = get_fuse_inode(inode);
1490         struct backing_dev_info *bdi = inode_to_bdi(inode);
1491         int i;
1492
1493         list_del(&req->writepages_entry);
1494         for (i = 0; i < req->num_pages; i++) {
1495                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1496                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1497                 wb_writeout_inc(&bdi->wb);
1498         }
1499         wake_up(&fi->page_waitq);
1500 }
1501
1502 /* Called under fc->lock, may release and reacquire it */
1503 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1504                                 loff_t size)
1505 __releases(fc->lock)
1506 __acquires(fc->lock)
1507 {
1508         struct fuse_inode *fi = get_fuse_inode(req->inode);
1509         struct fuse_write_in *inarg = &req->misc.write.in;
1510         __u64 data_size = req->num_pages * PAGE_SIZE;
1511
1512         if (!fc->connected)
1513                 goto out_free;
1514
1515         if (inarg->offset + data_size <= size) {
1516                 inarg->size = data_size;
1517         } else if (inarg->offset < size) {
1518                 inarg->size = size - inarg->offset;
1519         } else {
1520                 /* Got truncated off completely */
1521                 goto out_free;
1522         }
1523
1524         req->in.args[1].size = inarg->size;
1525         fi->writectr++;
1526         fuse_request_send_background_locked(fc, req);
1527         return;
1528
1529  out_free:
1530         fuse_writepage_finish(fc, req);
1531         spin_unlock(&fc->lock);
1532         fuse_writepage_free(fc, req);
1533         fuse_put_request(fc, req);
1534         spin_lock(&fc->lock);
1535 }
1536
1537 /*
1538  * If fi->writectr is positive (no truncate or fsync going on) send
1539  * all queued writepage requests.
1540  *
1541  * Called with fc->lock
1542  */
1543 void fuse_flush_writepages(struct inode *inode)
1544 __releases(fc->lock)
1545 __acquires(fc->lock)
1546 {
1547         struct fuse_conn *fc = get_fuse_conn(inode);
1548         struct fuse_inode *fi = get_fuse_inode(inode);
1549         loff_t crop = i_size_read(inode);
1550         struct fuse_req *req;
1551
1552         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1553                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1554                 list_del_init(&req->list);
1555                 fuse_send_writepage(fc, req, crop);
1556         }
1557 }
1558
1559 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1560 {
1561         struct inode *inode = req->inode;
1562         struct fuse_inode *fi = get_fuse_inode(inode);
1563
1564         mapping_set_error(inode->i_mapping, req->out.h.error);
1565         spin_lock(&fc->lock);
1566         while (req->misc.write.next) {
1567                 struct fuse_conn *fc = get_fuse_conn(inode);
1568                 struct fuse_write_in *inarg = &req->misc.write.in;
1569                 struct fuse_req *next = req->misc.write.next;
1570                 req->misc.write.next = next->misc.write.next;
1571                 next->misc.write.next = NULL;
1572                 next->ff = fuse_file_get(req->ff);
1573                 list_add(&next->writepages_entry, &fi->writepages);
1574
1575                 /*
1576                  * Skip fuse_flush_writepages() to make it easy to crop requests
1577                  * based on primary request size.
1578                  *
1579                  * 1st case (trivial): there are no concurrent activities using
1580                  * fuse_set/release_nowrite.  Then we're on safe side because
1581                  * fuse_flush_writepages() would call fuse_send_writepage()
1582                  * anyway.
1583                  *
1584                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1585                  * now for completion of all in-flight requests.  This happens
1586                  * rarely and no more than once per page, so this should be
1587                  * okay.
1588                  *
1589                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1590                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1591                  * that fuse_set_nowrite returned implies that all in-flight
1592                  * requests were completed along with all of their secondary
1593                  * requests.  Further primary requests are blocked by negative
1594                  * writectr.  Hence there cannot be any in-flight requests and
1595                  * no invocations of fuse_writepage_end() while we're in
1596                  * fuse_set_nowrite..fuse_release_nowrite section.
1597                  */
1598                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1599         }
1600         fi->writectr--;
1601         fuse_writepage_finish(fc, req);
1602         spin_unlock(&fc->lock);
1603         fuse_writepage_free(fc, req);
1604 }
1605
1606 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1607                                                struct fuse_inode *fi)
1608 {
1609         struct fuse_file *ff = NULL;
1610
1611         spin_lock(&fc->lock);
1612         if (!list_empty(&fi->write_files)) {
1613                 ff = list_entry(fi->write_files.next, struct fuse_file,
1614                                 write_entry);
1615                 fuse_file_get(ff);
1616         }
1617         spin_unlock(&fc->lock);
1618
1619         return ff;
1620 }
1621
1622 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1623                                              struct fuse_inode *fi)
1624 {
1625         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1626         WARN_ON(!ff);
1627         return ff;
1628 }
1629
1630 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1631 {
1632         struct fuse_conn *fc = get_fuse_conn(inode);
1633         struct fuse_inode *fi = get_fuse_inode(inode);
1634         struct fuse_file *ff;
1635         int err;
1636
1637         ff = __fuse_write_file_get(fc, fi);
1638         err = fuse_flush_times(inode, ff);
1639         if (ff)
1640                 fuse_file_put(ff, false, false);
1641
1642         return err;
1643 }
1644
1645 static int fuse_writepage_locked(struct page *page)
1646 {
1647         struct address_space *mapping = page->mapping;
1648         struct inode *inode = mapping->host;
1649         struct fuse_conn *fc = get_fuse_conn(inode);
1650         struct fuse_inode *fi = get_fuse_inode(inode);
1651         struct fuse_req *req;
1652         struct page *tmp_page;
1653         int error = -ENOMEM;
1654
1655         set_page_writeback(page);
1656
1657         req = fuse_request_alloc_nofs(1);
1658         if (!req)
1659                 goto err;
1660
1661         /* writeback always goes to bg_queue */
1662         __set_bit(FR_BACKGROUND, &req->flags);
1663         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1664         if (!tmp_page)
1665                 goto err_free;
1666
1667         error = -EIO;
1668         req->ff = fuse_write_file_get(fc, fi);
1669         if (!req->ff)
1670                 goto err_nofile;
1671
1672         fuse_write_fill(req, req->ff, page_offset(page), 0);
1673
1674         copy_highpage(tmp_page, page);
1675         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1676         req->misc.write.next = NULL;
1677         req->in.argpages = 1;
1678         req->num_pages = 1;
1679         req->pages[0] = tmp_page;
1680         req->page_descs[0].offset = 0;
1681         req->page_descs[0].length = PAGE_SIZE;
1682         req->end = fuse_writepage_end;
1683         req->inode = inode;
1684
1685         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1686         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1687
1688         spin_lock(&fc->lock);
1689         list_add(&req->writepages_entry, &fi->writepages);
1690         list_add_tail(&req->list, &fi->queued_writes);
1691         fuse_flush_writepages(inode);
1692         spin_unlock(&fc->lock);
1693
1694         end_page_writeback(page);
1695
1696         return 0;
1697
1698 err_nofile:
1699         __free_page(tmp_page);
1700 err_free:
1701         fuse_request_free(req);
1702 err:
1703         mapping_set_error(page->mapping, error);
1704         end_page_writeback(page);
1705         return error;
1706 }
1707
1708 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1709 {
1710         int err;
1711
1712         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1713                 /*
1714                  * ->writepages() should be called for sync() and friends.  We
1715                  * should only get here on direct reclaim and then we are
1716                  * allowed to skip a page which is already in flight
1717                  */
1718                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1719
1720                 redirty_page_for_writepage(wbc, page);
1721                 unlock_page(page);
1722                 return 0;
1723         }
1724
1725         err = fuse_writepage_locked(page);
1726         unlock_page(page);
1727
1728         return err;
1729 }
1730
1731 struct fuse_fill_wb_data {
1732         struct fuse_req *req;
1733         struct fuse_file *ff;
1734         struct inode *inode;
1735         struct page **orig_pages;
1736 };
1737
1738 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1739 {
1740         struct fuse_req *req = data->req;
1741         struct inode *inode = data->inode;
1742         struct fuse_conn *fc = get_fuse_conn(inode);
1743         struct fuse_inode *fi = get_fuse_inode(inode);
1744         int num_pages = req->num_pages;
1745         int i;
1746
1747         req->ff = fuse_file_get(data->ff);
1748         spin_lock(&fc->lock);
1749         list_add_tail(&req->list, &fi->queued_writes);
1750         fuse_flush_writepages(inode);
1751         spin_unlock(&fc->lock);
1752
1753         for (i = 0; i < num_pages; i++)
1754                 end_page_writeback(data->orig_pages[i]);
1755 }
1756
1757 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1758                                      struct page *page)
1759 {
1760         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1761         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1762         struct fuse_req *tmp;
1763         struct fuse_req *old_req;
1764         bool found = false;
1765         pgoff_t curr_index;
1766
1767         BUG_ON(new_req->num_pages != 0);
1768
1769         spin_lock(&fc->lock);
1770         list_del(&new_req->writepages_entry);
1771         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1772                 BUG_ON(old_req->inode != new_req->inode);
1773                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1774                 if (curr_index <= page->index &&
1775                     page->index < curr_index + old_req->num_pages) {
1776                         found = true;
1777                         break;
1778                 }
1779         }
1780         if (!found) {
1781                 list_add(&new_req->writepages_entry, &fi->writepages);
1782                 goto out_unlock;
1783         }
1784
1785         new_req->num_pages = 1;
1786         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1787                 BUG_ON(tmp->inode != new_req->inode);
1788                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1789                 if (tmp->num_pages == 1 &&
1790                     curr_index == page->index) {
1791                         old_req = tmp;
1792                 }
1793         }
1794
1795         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1796                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1797
1798                 copy_highpage(old_req->pages[0], page);
1799                 spin_unlock(&fc->lock);
1800
1801                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1802                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1803                 wb_writeout_inc(&bdi->wb);
1804                 fuse_writepage_free(fc, new_req);
1805                 fuse_request_free(new_req);
1806                 goto out;
1807         } else {
1808                 new_req->misc.write.next = old_req->misc.write.next;
1809                 old_req->misc.write.next = new_req;
1810         }
1811 out_unlock:
1812         spin_unlock(&fc->lock);
1813 out:
1814         return found;
1815 }
1816
1817 static int fuse_writepages_fill(struct page *page,
1818                 struct writeback_control *wbc, void *_data)
1819 {
1820         struct fuse_fill_wb_data *data = _data;
1821         struct fuse_req *req = data->req;
1822         struct inode *inode = data->inode;
1823         struct fuse_conn *fc = get_fuse_conn(inode);
1824         struct page *tmp_page;
1825         bool is_writeback;
1826         int err;
1827
1828         if (!data->ff) {
1829                 err = -EIO;
1830                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1831                 if (!data->ff)
1832                         goto out_unlock;
1833         }
1834
1835         /*
1836          * Being under writeback is unlikely but possible.  For example direct
1837          * read to an mmaped fuse file will set the page dirty twice; once when
1838          * the pages are faulted with get_user_pages(), and then after the read
1839          * completed.
1840          */
1841         is_writeback = fuse_page_is_writeback(inode, page->index);
1842
1843         if (req && req->num_pages &&
1844             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1845              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1846              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1847                 fuse_writepages_send(data);
1848                 data->req = NULL;
1849         }
1850         err = -ENOMEM;
1851         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1852         if (!tmp_page)
1853                 goto out_unlock;
1854
1855         /*
1856          * The page must not be redirtied until the writeout is completed
1857          * (i.e. userspace has sent a reply to the write request).  Otherwise
1858          * there could be more than one temporary page instance for each real
1859          * page.
1860          *
1861          * This is ensured by holding the page lock in page_mkwrite() while
1862          * checking fuse_page_is_writeback().  We already hold the page lock
1863          * since clear_page_dirty_for_io() and keep it held until we add the
1864          * request to the fi->writepages list and increment req->num_pages.
1865          * After this fuse_page_is_writeback() will indicate that the page is
1866          * under writeback, so we can release the page lock.
1867          */
1868         if (data->req == NULL) {
1869                 struct fuse_inode *fi = get_fuse_inode(inode);
1870
1871                 err = -ENOMEM;
1872                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1873                 if (!req) {
1874                         __free_page(tmp_page);
1875                         goto out_unlock;
1876                 }
1877
1878                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1879                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1880                 req->misc.write.next = NULL;
1881                 req->in.argpages = 1;
1882                 __set_bit(FR_BACKGROUND, &req->flags);
1883                 req->num_pages = 0;
1884                 req->end = fuse_writepage_end;
1885                 req->inode = inode;
1886
1887                 spin_lock(&fc->lock);
1888                 list_add(&req->writepages_entry, &fi->writepages);
1889                 spin_unlock(&fc->lock);
1890
1891                 data->req = req;
1892         }
1893         set_page_writeback(page);
1894
1895         copy_highpage(tmp_page, page);
1896         req->pages[req->num_pages] = tmp_page;
1897         req->page_descs[req->num_pages].offset = 0;
1898         req->page_descs[req->num_pages].length = PAGE_SIZE;
1899
1900         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1901         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1902
1903         err = 0;
1904         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1905                 end_page_writeback(page);
1906                 data->req = NULL;
1907                 goto out_unlock;
1908         }
1909         data->orig_pages[req->num_pages] = page;
1910
1911         /*
1912          * Protected by fc->lock against concurrent access by
1913          * fuse_page_is_writeback().
1914          */
1915         spin_lock(&fc->lock);
1916         req->num_pages++;
1917         spin_unlock(&fc->lock);
1918
1919 out_unlock:
1920         unlock_page(page);
1921
1922         return err;
1923 }
1924
1925 static int fuse_writepages(struct address_space *mapping,
1926                            struct writeback_control *wbc)
1927 {
1928         struct inode *inode = mapping->host;
1929         struct fuse_fill_wb_data data;
1930         int err;
1931
1932         err = -EIO;
1933         if (fuse_is_bad(inode))
1934                 goto out;
1935
1936         data.inode = inode;
1937         data.req = NULL;
1938         data.ff = NULL;
1939
1940         err = -ENOMEM;
1941         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1942                                   sizeof(struct page *),
1943                                   GFP_NOFS);
1944         if (!data.orig_pages)
1945                 goto out;
1946
1947         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1948         if (data.req) {
1949                 /* Ignore errors if we can write at least one page */
1950                 BUG_ON(!data.req->num_pages);
1951                 fuse_writepages_send(&data);
1952                 err = 0;
1953         }
1954         if (data.ff)
1955                 fuse_file_put(data.ff, false, false);
1956
1957         kfree(data.orig_pages);
1958 out:
1959         return err;
1960 }
1961
1962 /*
1963  * It's worthy to make sure that space is reserved on disk for the write,
1964  * but how to implement it without killing performance need more thinking.
1965  */
1966 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1967                 loff_t pos, unsigned len, unsigned flags,
1968                 struct page **pagep, void **fsdata)
1969 {
1970         pgoff_t index = pos >> PAGE_SHIFT;
1971         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1972         struct page *page;
1973         loff_t fsize;
1974         int err = -ENOMEM;
1975
1976         WARN_ON(!fc->writeback_cache);
1977
1978         page = grab_cache_page_write_begin(mapping, index, flags);
1979         if (!page)
1980                 goto error;
1981
1982         fuse_wait_on_page_writeback(mapping->host, page->index);
1983
1984         if (PageUptodate(page) || len == PAGE_SIZE)
1985                 goto success;
1986         /*
1987          * Check if the start this page comes after the end of file, in which
1988          * case the readpage can be optimized away.
1989          */
1990         fsize = i_size_read(mapping->host);
1991         if (fsize <= (pos & PAGE_MASK)) {
1992                 size_t off = pos & ~PAGE_MASK;
1993                 if (off)
1994                         zero_user_segment(page, 0, off);
1995                 goto success;
1996         }
1997         err = fuse_do_readpage(file, page);
1998         if (err)
1999                 goto cleanup;
2000 success:
2001         *pagep = page;
2002         return 0;
2003
2004 cleanup:
2005         unlock_page(page);
2006         put_page(page);
2007 error:
2008         return err;
2009 }
2010
2011 static int fuse_write_end(struct file *file, struct address_space *mapping,
2012                 loff_t pos, unsigned len, unsigned copied,
2013                 struct page *page, void *fsdata)
2014 {
2015         struct inode *inode = page->mapping->host;
2016
2017         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2018         if (!copied)
2019                 goto unlock;
2020
2021         if (!PageUptodate(page)) {
2022                 /* Zero any unwritten bytes at the end of the page */
2023                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2024                 if (endoff)
2025                         zero_user_segment(page, endoff, PAGE_SIZE);
2026                 SetPageUptodate(page);
2027         }
2028
2029         fuse_write_update_size(inode, pos + copied);
2030         set_page_dirty(page);
2031
2032 unlock:
2033         unlock_page(page);
2034         put_page(page);
2035
2036         return copied;
2037 }
2038
2039 static int fuse_launder_page(struct page *page)
2040 {
2041         int err = 0;
2042         if (clear_page_dirty_for_io(page)) {
2043                 struct inode *inode = page->mapping->host;
2044                 err = fuse_writepage_locked(page);
2045                 if (!err)
2046                         fuse_wait_on_page_writeback(inode, page->index);
2047         }
2048         return err;
2049 }
2050
2051 /*
2052  * Write back dirty pages now, because there may not be any suitable
2053  * open files later
2054  */
2055 static void fuse_vma_close(struct vm_area_struct *vma)
2056 {
2057         filemap_write_and_wait(vma->vm_file->f_mapping);
2058 }
2059
2060 /*
2061  * Wait for writeback against this page to complete before allowing it
2062  * to be marked dirty again, and hence written back again, possibly
2063  * before the previous writepage completed.
2064  *
2065  * Block here, instead of in ->writepage(), so that the userspace fs
2066  * can only block processes actually operating on the filesystem.
2067  *
2068  * Otherwise unprivileged userspace fs would be able to block
2069  * unrelated:
2070  *
2071  * - page migration
2072  * - sync(2)
2073  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2074  */
2075 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2076 {
2077         struct page *page = vmf->page;
2078         struct inode *inode = file_inode(vmf->vma->vm_file);
2079
2080         file_update_time(vmf->vma->vm_file);
2081         lock_page(page);
2082         if (page->mapping != inode->i_mapping) {
2083                 unlock_page(page);
2084                 return VM_FAULT_NOPAGE;
2085         }
2086
2087         fuse_wait_on_page_writeback(inode, page->index);
2088         return VM_FAULT_LOCKED;
2089 }
2090
2091 static const struct vm_operations_struct fuse_file_vm_ops = {
2092         .close          = fuse_vma_close,
2093         .fault          = filemap_fault,
2094         .map_pages      = filemap_map_pages,
2095         .page_mkwrite   = fuse_page_mkwrite,
2096 };
2097
2098 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2099 {
2100         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2101                 fuse_link_write_file(file);
2102
2103         file_accessed(file);
2104         vma->vm_ops = &fuse_file_vm_ops;
2105         return 0;
2106 }
2107
2108 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2109 {
2110         /* Can't provide the coherency needed for MAP_SHARED */
2111         if (vma->vm_flags & VM_MAYSHARE)
2112                 return -ENODEV;
2113
2114         invalidate_inode_pages2(file->f_mapping);
2115
2116         return generic_file_mmap(file, vma);
2117 }
2118
2119 static int convert_fuse_file_lock(struct fuse_conn *fc,
2120                                   const struct fuse_file_lock *ffl,
2121                                   struct file_lock *fl)
2122 {
2123         switch (ffl->type) {
2124         case F_UNLCK:
2125                 break;
2126
2127         case F_RDLCK:
2128         case F_WRLCK:
2129                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2130                     ffl->end < ffl->start)
2131                         return -EIO;
2132
2133                 fl->fl_start = ffl->start;
2134                 fl->fl_end = ffl->end;
2135
2136                 /*
2137                  * Convert pid into init's pid namespace.  The locks API will
2138                  * translate it into the caller's pid namespace.
2139                  */
2140                 rcu_read_lock();
2141                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2142                 rcu_read_unlock();
2143                 break;
2144
2145         default:
2146                 return -EIO;
2147         }
2148         fl->fl_type = ffl->type;
2149         return 0;
2150 }
2151
2152 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2153                          const struct file_lock *fl, int opcode, pid_t pid,
2154                          int flock, struct fuse_lk_in *inarg)
2155 {
2156         struct inode *inode = file_inode(file);
2157         struct fuse_conn *fc = get_fuse_conn(inode);
2158         struct fuse_file *ff = file->private_data;
2159
2160         memset(inarg, 0, sizeof(*inarg));
2161         inarg->fh = ff->fh;
2162         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2163         inarg->lk.start = fl->fl_start;
2164         inarg->lk.end = fl->fl_end;
2165         inarg->lk.type = fl->fl_type;
2166         inarg->lk.pid = pid;
2167         if (flock)
2168                 inarg->lk_flags |= FUSE_LK_FLOCK;
2169         args->in.h.opcode = opcode;
2170         args->in.h.nodeid = get_node_id(inode);
2171         args->in.numargs = 1;
2172         args->in.args[0].size = sizeof(*inarg);
2173         args->in.args[0].value = inarg;
2174 }
2175
2176 static int fuse_getlk(struct file *file, struct file_lock *fl)
2177 {
2178         struct inode *inode = file_inode(file);
2179         struct fuse_conn *fc = get_fuse_conn(inode);
2180         FUSE_ARGS(args);
2181         struct fuse_lk_in inarg;
2182         struct fuse_lk_out outarg;
2183         int err;
2184
2185         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2186         args.out.numargs = 1;
2187         args.out.args[0].size = sizeof(outarg);
2188         args.out.args[0].value = &outarg;
2189         err = fuse_simple_request(fc, &args);
2190         if (!err)
2191                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2192
2193         return err;
2194 }
2195
2196 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2197 {
2198         struct inode *inode = file_inode(file);
2199         struct fuse_conn *fc = get_fuse_conn(inode);
2200         FUSE_ARGS(args);
2201         struct fuse_lk_in inarg;
2202         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2203         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2204         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2205         int err;
2206
2207         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2208                 /* NLM needs asynchronous locks, which we don't support yet */
2209                 return -ENOLCK;
2210         }
2211
2212         /* Unlock on close is handled by the flush method */
2213         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2214                 return 0;
2215
2216         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2217         err = fuse_simple_request(fc, &args);
2218
2219         /* locking is restartable */
2220         if (err == -EINTR)
2221                 err = -ERESTARTSYS;
2222
2223         return err;
2224 }
2225
2226 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2227 {
2228         struct inode *inode = file_inode(file);
2229         struct fuse_conn *fc = get_fuse_conn(inode);
2230         int err;
2231
2232         if (cmd == F_CANCELLK) {
2233                 err = 0;
2234         } else if (cmd == F_GETLK) {
2235                 if (fc->no_lock) {
2236                         posix_test_lock(file, fl);
2237                         err = 0;
2238                 } else
2239                         err = fuse_getlk(file, fl);
2240         } else {
2241                 if (fc->no_lock)
2242                         err = posix_lock_file(file, fl, NULL);
2243                 else
2244                         err = fuse_setlk(file, fl, 0);
2245         }
2246         return err;
2247 }
2248
2249 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2250 {
2251         struct inode *inode = file_inode(file);
2252         struct fuse_conn *fc = get_fuse_conn(inode);
2253         int err;
2254
2255         if (fc->no_flock) {
2256                 err = locks_lock_file_wait(file, fl);
2257         } else {
2258                 struct fuse_file *ff = file->private_data;
2259
2260                 /* emulate flock with POSIX locks */
2261                 ff->flock = true;
2262                 err = fuse_setlk(file, fl, 1);
2263         }
2264
2265         return err;
2266 }
2267
2268 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2269 {
2270         struct inode *inode = mapping->host;
2271         struct fuse_conn *fc = get_fuse_conn(inode);
2272         FUSE_ARGS(args);
2273         struct fuse_bmap_in inarg;
2274         struct fuse_bmap_out outarg;
2275         int err;
2276
2277         if (!inode->i_sb->s_bdev || fc->no_bmap)
2278                 return 0;
2279
2280         memset(&inarg, 0, sizeof(inarg));
2281         inarg.block = block;
2282         inarg.blocksize = inode->i_sb->s_blocksize;
2283         args.in.h.opcode = FUSE_BMAP;
2284         args.in.h.nodeid = get_node_id(inode);
2285         args.in.numargs = 1;
2286         args.in.args[0].size = sizeof(inarg);
2287         args.in.args[0].value = &inarg;
2288         args.out.numargs = 1;
2289         args.out.args[0].size = sizeof(outarg);
2290         args.out.args[0].value = &outarg;
2291         err = fuse_simple_request(fc, &args);
2292         if (err == -ENOSYS)
2293                 fc->no_bmap = 1;
2294
2295         return err ? 0 : outarg.block;
2296 }
2297
2298 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2299 {
2300         struct inode *inode = file->f_mapping->host;
2301         struct fuse_conn *fc = get_fuse_conn(inode);
2302         struct fuse_file *ff = file->private_data;
2303         FUSE_ARGS(args);
2304         struct fuse_lseek_in inarg = {
2305                 .fh = ff->fh,
2306                 .offset = offset,
2307                 .whence = whence
2308         };
2309         struct fuse_lseek_out outarg;
2310         int err;
2311
2312         if (fc->no_lseek)
2313                 goto fallback;
2314
2315         args.in.h.opcode = FUSE_LSEEK;
2316         args.in.h.nodeid = ff->nodeid;
2317         args.in.numargs = 1;
2318         args.in.args[0].size = sizeof(inarg);
2319         args.in.args[0].value = &inarg;
2320         args.out.numargs = 1;
2321         args.out.args[0].size = sizeof(outarg);
2322         args.out.args[0].value = &outarg;
2323         err = fuse_simple_request(fc, &args);
2324         if (err) {
2325                 if (err == -ENOSYS) {
2326                         fc->no_lseek = 1;
2327                         goto fallback;
2328                 }
2329                 return err;
2330         }
2331
2332         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2333
2334 fallback:
2335         err = fuse_update_attributes(inode, file);
2336         if (!err)
2337                 return generic_file_llseek(file, offset, whence);
2338         else
2339                 return err;
2340 }
2341
2342 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2343 {
2344         loff_t retval;
2345         struct inode *inode = file_inode(file);
2346
2347         switch (whence) {
2348         case SEEK_SET:
2349         case SEEK_CUR:
2350                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2351                 retval = generic_file_llseek(file, offset, whence);
2352                 break;
2353         case SEEK_END:
2354                 inode_lock(inode);
2355                 retval = fuse_update_attributes(inode, file);
2356                 if (!retval)
2357                         retval = generic_file_llseek(file, offset, whence);
2358                 inode_unlock(inode);
2359                 break;
2360         case SEEK_HOLE:
2361         case SEEK_DATA:
2362                 inode_lock(inode);
2363                 retval = fuse_lseek(file, offset, whence);
2364                 inode_unlock(inode);
2365                 break;
2366         default:
2367                 retval = -EINVAL;
2368         }
2369
2370         return retval;
2371 }
2372
2373 /*
2374  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2375  * ABI was defined to be 'struct iovec' which is different on 32bit
2376  * and 64bit.  Fortunately we can determine which structure the server
2377  * used from the size of the reply.
2378  */
2379 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2380                                      size_t transferred, unsigned count,
2381                                      bool is_compat)
2382 {
2383 #ifdef CONFIG_COMPAT
2384         if (count * sizeof(struct compat_iovec) == transferred) {
2385                 struct compat_iovec *ciov = src;
2386                 unsigned i;
2387
2388                 /*
2389                  * With this interface a 32bit server cannot support
2390                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2391                  * requests
2392                  */
2393                 if (!is_compat)
2394                         return -EINVAL;
2395
2396                 for (i = 0; i < count; i++) {
2397                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2398                         dst[i].iov_len = ciov[i].iov_len;
2399                 }
2400                 return 0;
2401         }
2402 #endif
2403
2404         if (count * sizeof(struct iovec) != transferred)
2405                 return -EIO;
2406
2407         memcpy(dst, src, transferred);
2408         return 0;
2409 }
2410
2411 /* Make sure iov_length() won't overflow */
2412 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2413 {
2414         size_t n;
2415         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2416
2417         for (n = 0; n < count; n++, iov++) {
2418                 if (iov->iov_len > (size_t) max)
2419                         return -ENOMEM;
2420                 max -= iov->iov_len;
2421         }
2422         return 0;
2423 }
2424
2425 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2426                                  void *src, size_t transferred, unsigned count,
2427                                  bool is_compat)
2428 {
2429         unsigned i;
2430         struct fuse_ioctl_iovec *fiov = src;
2431
2432         if (fc->minor < 16) {
2433                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2434                                                  count, is_compat);
2435         }
2436
2437         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2438                 return -EIO;
2439
2440         for (i = 0; i < count; i++) {
2441                 /* Did the server supply an inappropriate value? */
2442                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2443                     fiov[i].len != (unsigned long) fiov[i].len)
2444                         return -EIO;
2445
2446                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2447                 dst[i].iov_len = (size_t) fiov[i].len;
2448
2449 #ifdef CONFIG_COMPAT
2450                 if (is_compat &&
2451                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2452                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2453                         return -EIO;
2454 #endif
2455         }
2456
2457         return 0;
2458 }
2459
2460
2461 /*
2462  * For ioctls, there is no generic way to determine how much memory
2463  * needs to be read and/or written.  Furthermore, ioctls are allowed
2464  * to dereference the passed pointer, so the parameter requires deep
2465  * copying but FUSE has no idea whatsoever about what to copy in or
2466  * out.
2467  *
2468  * This is solved by allowing FUSE server to retry ioctl with
2469  * necessary in/out iovecs.  Let's assume the ioctl implementation
2470  * needs to read in the following structure.
2471  *
2472  * struct a {
2473  *      char    *buf;
2474  *      size_t  buflen;
2475  * }
2476  *
2477  * On the first callout to FUSE server, inarg->in_size and
2478  * inarg->out_size will be NULL; then, the server completes the ioctl
2479  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2480  * the actual iov array to
2481  *
2482  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2483  *
2484  * which tells FUSE to copy in the requested area and retry the ioctl.
2485  * On the second round, the server has access to the structure and
2486  * from that it can tell what to look for next, so on the invocation,
2487  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2488  *
2489  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2490  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2491  *
2492  * FUSE will copy both struct a and the pointed buffer from the
2493  * process doing the ioctl and retry ioctl with both struct a and the
2494  * buffer.
2495  *
2496  * This time, FUSE server has everything it needs and completes ioctl
2497  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2498  *
2499  * Copying data out works the same way.
2500  *
2501  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2502  * automatically initializes in and out iovs by decoding @cmd with
2503  * _IOC_* macros and the server is not allowed to request RETRY.  This
2504  * limits ioctl data transfers to well-formed ioctls and is the forced
2505  * behavior for all FUSE servers.
2506  */
2507 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2508                    unsigned int flags)
2509 {
2510         struct fuse_file *ff = file->private_data;
2511         struct fuse_conn *fc = ff->fc;
2512         struct fuse_ioctl_in inarg = {
2513                 .fh = ff->fh,
2514                 .cmd = cmd,
2515                 .arg = arg,
2516                 .flags = flags
2517         };
2518         struct fuse_ioctl_out outarg;
2519         struct fuse_req *req = NULL;
2520         struct page **pages = NULL;
2521         struct iovec *iov_page = NULL;
2522         struct iovec *in_iov = NULL, *out_iov = NULL;
2523         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2524         size_t in_size, out_size, transferred, c;
2525         int err, i;
2526         struct iov_iter ii;
2527
2528 #if BITS_PER_LONG == 32
2529         inarg.flags |= FUSE_IOCTL_32BIT;
2530 #else
2531         if (flags & FUSE_IOCTL_COMPAT)
2532                 inarg.flags |= FUSE_IOCTL_32BIT;
2533 #endif
2534
2535         /* assume all the iovs returned by client always fits in a page */
2536         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2537
2538         err = -ENOMEM;
2539         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2540         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2541         if (!pages || !iov_page)
2542                 goto out;
2543
2544         /*
2545          * If restricted, initialize IO parameters as encoded in @cmd.
2546          * RETRY from server is not allowed.
2547          */
2548         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2549                 struct iovec *iov = iov_page;
2550
2551                 iov->iov_base = (void __user *)arg;
2552
2553                 switch (cmd) {
2554                 case FS_IOC_GETFLAGS:
2555                 case FS_IOC_SETFLAGS:
2556                         iov->iov_len = sizeof(int);
2557                         break;
2558                 default:
2559                         iov->iov_len = _IOC_SIZE(cmd);
2560                         break;
2561                 }
2562
2563                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2564                         in_iov = iov;
2565                         in_iovs = 1;
2566                 }
2567
2568                 if (_IOC_DIR(cmd) & _IOC_READ) {
2569                         out_iov = iov;
2570                         out_iovs = 1;
2571                 }
2572         }
2573
2574  retry:
2575         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2576         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2577
2578         /*
2579          * Out data can be used either for actual out data or iovs,
2580          * make sure there always is at least one page.
2581          */
2582         out_size = max_t(size_t, out_size, PAGE_SIZE);
2583         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2584
2585         /* make sure there are enough buffer pages and init request with them */
2586         err = -ENOMEM;
2587         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2588                 goto out;
2589         while (num_pages < max_pages) {
2590                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2591                 if (!pages[num_pages])
2592                         goto out;
2593                 num_pages++;
2594         }
2595
2596         req = fuse_get_req(fc, num_pages);
2597         if (IS_ERR(req)) {
2598                 err = PTR_ERR(req);
2599                 req = NULL;
2600                 goto out;
2601         }
2602         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2603         req->num_pages = num_pages;
2604         fuse_page_descs_length_init(req, 0, req->num_pages);
2605
2606         /* okay, let's send it to the client */
2607         req->in.h.opcode = FUSE_IOCTL;
2608         req->in.h.nodeid = ff->nodeid;
2609         req->in.numargs = 1;
2610         req->in.args[0].size = sizeof(inarg);
2611         req->in.args[0].value = &inarg;
2612         if (in_size) {
2613                 req->in.numargs++;
2614                 req->in.args[1].size = in_size;
2615                 req->in.argpages = 1;
2616
2617                 err = -EFAULT;
2618                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2619                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2620                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2621                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2622                                 goto out;
2623                 }
2624         }
2625
2626         req->out.numargs = 2;
2627         req->out.args[0].size = sizeof(outarg);
2628         req->out.args[0].value = &outarg;
2629         req->out.args[1].size = out_size;
2630         req->out.argpages = 1;
2631         req->out.argvar = 1;
2632
2633         fuse_request_send(fc, req);
2634         err = req->out.h.error;
2635         transferred = req->out.args[1].size;
2636         fuse_put_request(fc, req);
2637         req = NULL;
2638         if (err)
2639                 goto out;
2640
2641         /* did it ask for retry? */
2642         if (outarg.flags & FUSE_IOCTL_RETRY) {
2643                 void *vaddr;
2644
2645                 /* no retry if in restricted mode */
2646                 err = -EIO;
2647                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2648                         goto out;
2649
2650                 in_iovs = outarg.in_iovs;
2651                 out_iovs = outarg.out_iovs;
2652
2653                 /*
2654                  * Make sure things are in boundary, separate checks
2655                  * are to protect against overflow.
2656                  */
2657                 err = -ENOMEM;
2658                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2659                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2660                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2661                         goto out;
2662
2663                 vaddr = kmap_atomic(pages[0]);
2664                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2665                                             transferred, in_iovs + out_iovs,
2666                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2667                 kunmap_atomic(vaddr);
2668                 if (err)
2669                         goto out;
2670
2671                 in_iov = iov_page;
2672                 out_iov = in_iov + in_iovs;
2673
2674                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2675                 if (err)
2676                         goto out;
2677
2678                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2679                 if (err)
2680                         goto out;
2681
2682                 goto retry;
2683         }
2684
2685         err = -EIO;
2686         if (transferred > inarg.out_size)
2687                 goto out;
2688
2689         err = -EFAULT;
2690         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2691         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2692                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2693                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2694                         goto out;
2695         }
2696         err = 0;
2697  out:
2698         if (req)
2699                 fuse_put_request(fc, req);
2700         free_page((unsigned long) iov_page);
2701         while (num_pages)
2702                 __free_page(pages[--num_pages]);
2703         kfree(pages);
2704
2705         return err ? err : outarg.result;
2706 }
2707 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2708
2709 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2710                        unsigned long arg, unsigned int flags)
2711 {
2712         struct inode *inode = file_inode(file);
2713         struct fuse_conn *fc = get_fuse_conn(inode);
2714
2715         if (!fuse_allow_current_process(fc))
2716                 return -EACCES;
2717
2718         if (fuse_is_bad(inode))
2719                 return -EIO;
2720
2721         return fuse_do_ioctl(file, cmd, arg, flags);
2722 }
2723
2724 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2725                             unsigned long arg)
2726 {
2727         return fuse_ioctl_common(file, cmd, arg, 0);
2728 }
2729
2730 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2731                                    unsigned long arg)
2732 {
2733         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2734 }
2735
2736 /*
2737  * All files which have been polled are linked to RB tree
2738  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2739  * find the matching one.
2740  */
2741 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2742                                               struct rb_node **parent_out)
2743 {
2744         struct rb_node **link = &fc->polled_files.rb_node;
2745         struct rb_node *last = NULL;
2746
2747         while (*link) {
2748                 struct fuse_file *ff;
2749
2750                 last = *link;
2751                 ff = rb_entry(last, struct fuse_file, polled_node);
2752
2753                 if (kh < ff->kh)
2754                         link = &last->rb_left;
2755                 else if (kh > ff->kh)
2756                         link = &last->rb_right;
2757                 else
2758                         return link;
2759         }
2760
2761         if (parent_out)
2762                 *parent_out = last;
2763         return link;
2764 }
2765
2766 /*
2767  * The file is about to be polled.  Make sure it's on the polled_files
2768  * RB tree.  Note that files once added to the polled_files tree are
2769  * not removed before the file is released.  This is because a file
2770  * polled once is likely to be polled again.
2771  */
2772 static void fuse_register_polled_file(struct fuse_conn *fc,
2773                                       struct fuse_file *ff)
2774 {
2775         spin_lock(&fc->lock);
2776         if (RB_EMPTY_NODE(&ff->polled_node)) {
2777                 struct rb_node **link, *uninitialized_var(parent);
2778
2779                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2780                 BUG_ON(*link);
2781                 rb_link_node(&ff->polled_node, parent, link);
2782                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2783         }
2784         spin_unlock(&fc->lock);
2785 }
2786
2787 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2788 {
2789         struct fuse_file *ff = file->private_data;
2790         struct fuse_conn *fc = ff->fc;
2791         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2792         struct fuse_poll_out outarg;
2793         FUSE_ARGS(args);
2794         int err;
2795
2796         if (fc->no_poll)
2797                 return DEFAULT_POLLMASK;
2798
2799         poll_wait(file, &ff->poll_wait, wait);
2800         inarg.events = mangle_poll(poll_requested_events(wait));
2801
2802         /*
2803          * Ask for notification iff there's someone waiting for it.
2804          * The client may ignore the flag and always notify.
2805          */
2806         if (waitqueue_active(&ff->poll_wait)) {
2807                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2808                 fuse_register_polled_file(fc, ff);
2809         }
2810
2811         args.in.h.opcode = FUSE_POLL;
2812         args.in.h.nodeid = ff->nodeid;
2813         args.in.numargs = 1;
2814         args.in.args[0].size = sizeof(inarg);
2815         args.in.args[0].value = &inarg;
2816         args.out.numargs = 1;
2817         args.out.args[0].size = sizeof(outarg);
2818         args.out.args[0].value = &outarg;
2819         err = fuse_simple_request(fc, &args);
2820
2821         if (!err)
2822                 return demangle_poll(outarg.revents);
2823         if (err == -ENOSYS) {
2824                 fc->no_poll = 1;
2825                 return DEFAULT_POLLMASK;
2826         }
2827         return EPOLLERR;
2828 }
2829 EXPORT_SYMBOL_GPL(fuse_file_poll);
2830
2831 /*
2832  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2833  * wakes up the poll waiters.
2834  */
2835 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2836                             struct fuse_notify_poll_wakeup_out *outarg)
2837 {
2838         u64 kh = outarg->kh;
2839         struct rb_node **link;
2840
2841         spin_lock(&fc->lock);
2842
2843         link = fuse_find_polled_node(fc, kh, NULL);
2844         if (*link) {
2845                 struct fuse_file *ff;
2846
2847                 ff = rb_entry(*link, struct fuse_file, polled_node);
2848                 wake_up_interruptible_sync(&ff->poll_wait);
2849         }
2850
2851         spin_unlock(&fc->lock);
2852         return 0;
2853 }
2854
2855 static void fuse_do_truncate(struct file *file)
2856 {
2857         struct inode *inode = file->f_mapping->host;
2858         struct iattr attr;
2859
2860         attr.ia_valid = ATTR_SIZE;
2861         attr.ia_size = i_size_read(inode);
2862
2863         attr.ia_file = file;
2864         attr.ia_valid |= ATTR_FILE;
2865
2866         fuse_do_setattr(file_dentry(file), &attr, file);
2867 }
2868
2869 static inline loff_t fuse_round_up(loff_t off)
2870 {
2871         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2872 }
2873
2874 static ssize_t
2875 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2876 {
2877         DECLARE_COMPLETION_ONSTACK(wait);
2878         ssize_t ret = 0;
2879         struct file *file = iocb->ki_filp;
2880         struct fuse_file *ff = file->private_data;
2881         bool async_dio = ff->fc->async_dio;
2882         loff_t pos = 0;
2883         struct inode *inode;
2884         loff_t i_size;
2885         size_t count = iov_iter_count(iter);
2886         loff_t offset = iocb->ki_pos;
2887         struct fuse_io_priv *io;
2888
2889         pos = offset;
2890         inode = file->f_mapping->host;
2891         i_size = i_size_read(inode);
2892
2893         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2894                 return 0;
2895
2896         /* optimization for short read */
2897         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2898                 if (offset >= i_size)
2899                         return 0;
2900                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2901                 count = iov_iter_count(iter);
2902         }
2903
2904         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2905         if (!io)
2906                 return -ENOMEM;
2907         spin_lock_init(&io->lock);
2908         kref_init(&io->refcnt);
2909         io->reqs = 1;
2910         io->bytes = -1;
2911         io->size = 0;
2912         io->offset = offset;
2913         io->write = (iov_iter_rw(iter) == WRITE);
2914         io->err = 0;
2915         /*
2916          * By default, we want to optimize all I/Os with async request
2917          * submission to the client filesystem if supported.
2918          */
2919         io->async = async_dio;
2920         io->iocb = iocb;
2921         io->blocking = is_sync_kiocb(iocb);
2922
2923         /*
2924          * We cannot asynchronously extend the size of a file.
2925          * In such case the aio will behave exactly like sync io.
2926          */
2927         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2928                 io->blocking = true;
2929
2930         if (io->async && io->blocking) {
2931                 /*
2932                  * Additional reference to keep io around after
2933                  * calling fuse_aio_complete()
2934                  */
2935                 kref_get(&io->refcnt);
2936                 io->done = &wait;
2937         }
2938
2939         if (iov_iter_rw(iter) == WRITE) {
2940                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2941                 fuse_invalidate_attr(inode);
2942         } else {
2943                 ret = __fuse_direct_read(io, iter, &pos);
2944         }
2945
2946         if (io->async) {
2947                 bool blocking = io->blocking;
2948
2949                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2950
2951                 /* we have a non-extending, async request, so return */
2952                 if (!blocking)
2953                         return -EIOCBQUEUED;
2954
2955                 wait_for_completion(&wait);
2956                 ret = fuse_get_res_by_io(io);
2957         }
2958
2959         kref_put(&io->refcnt, fuse_io_release);
2960
2961         if (iov_iter_rw(iter) == WRITE) {
2962                 if (ret > 0)
2963                         fuse_write_update_size(inode, pos);
2964                 else if (ret < 0 && offset + count > i_size)
2965                         fuse_do_truncate(file);
2966         }
2967
2968         return ret;
2969 }
2970
2971 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2972                                 loff_t length)
2973 {
2974         struct fuse_file *ff = file->private_data;
2975         struct inode *inode = file_inode(file);
2976         struct fuse_inode *fi = get_fuse_inode(inode);
2977         struct fuse_conn *fc = ff->fc;
2978         FUSE_ARGS(args);
2979         struct fuse_fallocate_in inarg = {
2980                 .fh = ff->fh,
2981                 .offset = offset,
2982                 .length = length,
2983                 .mode = mode
2984         };
2985         int err;
2986         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2987                            (mode & FALLOC_FL_PUNCH_HOLE);
2988
2989         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2990                 return -EOPNOTSUPP;
2991
2992         if (fc->no_fallocate)
2993                 return -EOPNOTSUPP;
2994
2995         if (lock_inode) {
2996                 inode_lock(inode);
2997                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2998                         loff_t endbyte = offset + length - 1;
2999                         err = filemap_write_and_wait_range(inode->i_mapping,
3000                                                            offset, endbyte);
3001                         if (err)
3002                                 goto out;
3003
3004                         fuse_sync_writes(inode);
3005                 }
3006         }
3007
3008         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3009             offset + length > i_size_read(inode)) {
3010                 err = inode_newsize_ok(inode, offset + length);
3011                 if (err)
3012                         goto out;
3013         }
3014
3015         if (!(mode & FALLOC_FL_KEEP_SIZE))
3016                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3017
3018         args.in.h.opcode = FUSE_FALLOCATE;
3019         args.in.h.nodeid = ff->nodeid;
3020         args.in.numargs = 1;
3021         args.in.args[0].size = sizeof(inarg);
3022         args.in.args[0].value = &inarg;
3023         err = fuse_simple_request(fc, &args);
3024         if (err == -ENOSYS) {
3025                 fc->no_fallocate = 1;
3026                 err = -EOPNOTSUPP;
3027         }
3028         if (err)
3029                 goto out;
3030
3031         /* we could have extended the file */
3032         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3033                 bool changed = fuse_write_update_size(inode, offset + length);
3034
3035                 if (changed && fc->writeback_cache)
3036                         file_update_time(file);
3037         }
3038
3039         if (mode & FALLOC_FL_PUNCH_HOLE)
3040                 truncate_pagecache_range(inode, offset, offset + length - 1);
3041
3042         fuse_invalidate_attr(inode);
3043
3044 out:
3045         if (!(mode & FALLOC_FL_KEEP_SIZE))
3046                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3047
3048         if (lock_inode)
3049                 inode_unlock(inode);
3050
3051         return err;
3052 }
3053
3054 static const struct file_operations fuse_file_operations = {
3055         .llseek         = fuse_file_llseek,
3056         .read_iter      = fuse_file_read_iter,
3057         .write_iter     = fuse_file_write_iter,
3058         .mmap           = fuse_file_mmap,
3059         .open           = fuse_open,
3060         .flush          = fuse_flush,
3061         .release        = fuse_release,
3062         .fsync          = fuse_fsync,
3063         .lock           = fuse_file_lock,
3064         .flock          = fuse_file_flock,
3065         .splice_read    = generic_file_splice_read,
3066         .unlocked_ioctl = fuse_file_ioctl,
3067         .compat_ioctl   = fuse_file_compat_ioctl,
3068         .poll           = fuse_file_poll,
3069         .fallocate      = fuse_file_fallocate,
3070 };
3071
3072 static const struct file_operations fuse_direct_io_file_operations = {
3073         .llseek         = fuse_file_llseek,
3074         .read_iter      = fuse_direct_read_iter,
3075         .write_iter     = fuse_direct_write_iter,
3076         .mmap           = fuse_direct_mmap,
3077         .open           = fuse_open,
3078         .flush          = fuse_flush,
3079         .release        = fuse_release,
3080         .fsync          = fuse_fsync,
3081         .lock           = fuse_file_lock,
3082         .flock          = fuse_file_flock,
3083         .unlocked_ioctl = fuse_file_ioctl,
3084         .compat_ioctl   = fuse_file_compat_ioctl,
3085         .poll           = fuse_file_poll,
3086         .fallocate      = fuse_file_fallocate,
3087         /* no splice_read */
3088 };
3089
3090 static const struct address_space_operations fuse_file_aops  = {
3091         .readpage       = fuse_readpage,
3092         .writepage      = fuse_writepage,
3093         .writepages     = fuse_writepages,
3094         .launder_page   = fuse_launder_page,
3095         .readpages      = fuse_readpages,
3096         .set_page_dirty = __set_page_dirty_nobuffers,
3097         .bmap           = fuse_bmap,
3098         .direct_IO      = fuse_direct_IO,
3099         .write_begin    = fuse_write_begin,
3100         .write_end      = fuse_write_end,
3101 };
3102
3103 void fuse_init_file_inode(struct inode *inode)
3104 {
3105         inode->i_fop = &fuse_file_operations;
3106         inode->i_data.a_ops = &fuse_file_aops;
3107 }