fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/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
  22 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  23                           int opcode, struct fuse_open_out *outargp)
  24 {
  25         struct fuse_open_in inarg;
  26         FUSE_ARGS(args);
  27
  28         memset(&inarg, 0, sizeof(inarg));
  29         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  30         if (!fc->atomic_o_trunc)
  31                 inarg.flags &= ~O_TRUNC;
  32         args.in.h.opcode = opcode;
  33         args.in.h.nodeid = nodeid;
  34         args.in.numargs = 1;
  35         args.in.args[0].size = sizeof(inarg);
  36         args.in.args[0].value = &inarg;
  37         args.out.numargs = 1;
  38         args.out.args[0].size = sizeof(*outargp);
  39         args.out.args[0].value = outargp;
  40
  41         return fuse_simple_request(fc, &args);
  42 }
  43
  44 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  45 {
  46         struct fuse_file *ff;
  47
  48         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  49         if (unlikely(!ff))
  50                 return NULL;
  51
  52         ff->fc = fc;
  53         ff->reserved_req = fuse_request_alloc(0);
  54         if (unlikely(!ff->reserved_req)) {
  55                 kfree(ff);
  56                 return NULL;
  57         }
  58
  59         INIT_LIST_HEAD(&ff->write_entry);
  60         mutex_init(&ff->readdir.lock);
  61         refcount_set(&ff->count, 1);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         ff->kh = atomic64_inc_return(&fc->khctr);
  66
  67         return ff;
  68 }
  69
  70 void fuse_file_free(struct fuse_file *ff)
  71 {
  72         fuse_request_free(ff->reserved_req);
  73         mutex_destroy(&ff->readdir.lock);
  74         kfree(ff);
  75 }
  76
  77 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  78 {
  79         refcount_inc(&ff->count);
  80         return ff;
  81 }
  82
  83 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  84 {
  85         iput(req->misc.release.inode);
  86 }
  87
  88 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
  89 {
  90         if (refcount_dec_and_test(&ff->count)) {
  91                 struct fuse_req *req = ff->reserved_req;
  92
  93                 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
  94                         /*
  95                          * Drop the release request when client does not
  96                          * implement 'open'
  97                          */
  98                         __clear_bit(FR_BACKGROUND, &req->flags);
  99                         iput(req->misc.release.inode);
 100                         fuse_put_request(ff->fc, req);
 101                 } else if (sync) {
 102                         __set_bit(FR_FORCE, &req->flags);
 103                         __clear_bit(FR_BACKGROUND, &req->flags);
 104                         fuse_request_send(ff->fc, req);
 105                         iput(req->misc.release.inode);
 106                         fuse_put_request(ff->fc, req);
 107                 } else {
 108                         req->end = fuse_release_end;
 109                         __set_bit(FR_BACKGROUND, &req->flags);
 110                         fuse_request_send_background(ff->fc, req);
 111                 }
 112                 kfree(ff);
 113         }
 114 }
 115
 116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 117                  bool isdir)
 118 {
 119         struct fuse_file *ff;
 120         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 121
 122         ff = fuse_file_alloc(fc);
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         ff->fh = 0;
 127         /* Default for no-open */
 128         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 129         if (isdir ? !fc->no_opendir : !fc->no_open) {
 130                 struct fuse_open_out outarg;
 131                 int err;
 132
 133                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 134                 if (!err) {
 135                         ff->fh = outarg.fh;
 136                         ff->open_flags = outarg.open_flags;
 137
 138                 } else if (err != -ENOSYS) {
 139                         fuse_file_free(ff);
 140                         return err;
 141                 } else {
 142                         if (isdir)
 143                                 fc->no_opendir = 1;
 144                         else
 145                                 fc->no_open = 1;
 146                 }
 147         }
 148
 149         if (isdir)
 150                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 151
 152         ff->nodeid = nodeid;
 153         file->private_data = ff;
 154
 155         return 0;
 156 }
 157 EXPORT_SYMBOL_GPL(fuse_do_open);
 158
 159 static void fuse_link_write_file(struct file *file)
 160 {
 161         struct inode *inode = file_inode(file);
 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(&fi->lock);
 169         if (list_empty(&ff->write_entry))
 170                 list_add(&ff->write_entry, &fi->write_files);
 171         spin_unlock(&fi->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_KEEP_CACHE))
 180                 invalidate_inode_pages2(inode->i_mapping);
 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         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 186                 struct fuse_inode *fi = get_fuse_inode(inode);
 187
 188                 spin_lock(&fi->lock);
 189                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 190                 i_size_write(inode, 0);
 191                 spin_unlock(&fi->lock);
 192                 fuse_invalidate_attr(inode);
 193                 if (fc->writeback_cache)
 194                         file_update_time(file);
 195         }
 196         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 197                 fuse_link_write_file(file);
 198 }
 199
 200 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 201 {
 202         struct fuse_conn *fc = get_fuse_conn(inode);
 203         int err;
 204         bool lock_inode = (file->f_flags & O_TRUNC) &&
 205                           fc->atomic_o_trunc &&
 206                           fc->writeback_cache;
 207
 208         err = generic_file_open(inode, file);
 209         if (err)
 210                 return err;
 211
 212         if (lock_inode)
 213                 inode_lock(inode);
 214
 215         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 216
 217         if (!err)
 218                 fuse_finish_open(inode, file);
 219
 220         if (lock_inode)
 221                 inode_unlock(inode);
 222
 223         return err;
 224 }
 225
 226 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 227                                  int flags, int opcode)
 228 {
 229         struct fuse_conn *fc = ff->fc;
 230         struct fuse_req *req = ff->reserved_req;
 231         struct fuse_release_in *inarg = &req->misc.release.in;
 232
 233         /* Inode is NULL on error path of fuse_create_open() */
 234         if (likely(fi)) {
 235                 spin_lock(&fi->lock);
 236                 list_del(&ff->write_entry);
 237                 spin_unlock(&fi->lock);
 238         }
 239         spin_lock(&fc->lock);
 240         if (!RB_EMPTY_NODE(&ff->polled_node))
 241                 rb_erase(&ff->polled_node, &fc->polled_files);
 242         spin_unlock(&fc->lock);
 243
 244         wake_up_interruptible_all(&ff->poll_wait);
 245
 246         inarg->fh = ff->fh;
 247         inarg->flags = flags;
 248         req->in.h.opcode = opcode;
 249         req->in.h.nodeid = ff->nodeid;
 250         req->in.numargs = 1;
 251         req->in.args[0].size = sizeof(struct fuse_release_in);
 252         req->in.args[0].value = inarg;
 253 }
 254
 255 void fuse_release_common(struct file *file, bool isdir)
 256 {
 257         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 258         struct fuse_file *ff = file->private_data;
 259         struct fuse_req *req = ff->reserved_req;
 260         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 261
 262         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 263
 264         if (ff->flock) {
 265                 struct fuse_release_in *inarg = &req->misc.release.in;
 266                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 267                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 268                                                        (fl_owner_t) file);
 269         }
 270         /* Hold inode until release is finished */
 271         req->misc.release.inode = igrab(file_inode(file));
 272
 273         /*
 274          * Normally this will send the RELEASE request, however if
 275          * some asynchronous READ or WRITE requests are outstanding,
 276          * the sending will be delayed.
 277          *
 278          * Make the release synchronous if this is a fuseblk mount,
 279          * synchronous RELEASE is allowed (and desirable) in this case
 280          * because the server can be trusted not to screw up.
 281          */
 282         fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
 283 }
 284
 285 static int fuse_open(struct inode *inode, struct file *file)
 286 {
 287         return fuse_open_common(inode, file, false);
 288 }
 289
 290 static int fuse_release(struct inode *inode, struct file *file)
 291 {
 292         struct fuse_conn *fc = get_fuse_conn(inode);
 293
 294         /* see fuse_vma_close() for !writeback_cache case */
 295         if (fc->writeback_cache)
 296                 write_inode_now(inode, 1);
 297
 298         fuse_release_common(file, false);
 299
 300         /* return value is ignored by VFS */
 301         return 0;
 302 }
 303
 304 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 305 {
 306         WARN_ON(refcount_read(&ff->count) > 1);
 307         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 308         /*
 309          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 310          * synchronous, we are fine with not doing igrab() here"
 311          */
 312         fuse_file_put(ff, true, false);
 313 }
 314 EXPORT_SYMBOL_GPL(fuse_sync_release);
 315
 316 /*
 317  * Scramble the ID space with XTEA, so that the value of the files_struct
 318  * pointer is not exposed to userspace.
 319  */
 320 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 321 {
 322         u32 *k = fc->scramble_key;
 323         u64 v = (unsigned long) id;
 324         u32 v0 = v;
 325         u32 v1 = v >> 32;
 326         u32 sum = 0;
 327         int i;
 328
 329         for (i = 0; i < 32; i++) {
 330                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 331                 sum += 0x9E3779B9;
 332                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 333         }
 334
 335         return (u64) v0 + ((u64) v1 << 32);
 336 }
 337
 338 static struct fuse_req *fuse_find_writeback(struct fuse_inode *fi,
 339                                             pgoff_t idx_from, pgoff_t idx_to)
 340 {
 341         struct fuse_req *req;
 342
 343         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 344                 pgoff_t curr_index;
 345
 346                 WARN_ON(get_fuse_inode(req->inode) != fi);
 347                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 348                 if (idx_from < curr_index + req->num_pages &&
 349                     curr_index <= idx_to) {
 350                         return req;
 351                 }
 352         }
 353         return NULL;
 354 }
 355
 356 /*
 357  * Check if any page in a range is under writeback
 358  *
 359  * This is currently done by walking the list of writepage requests
 360  * for the inode, which can be pretty inefficient.
 361  */
 362 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 363                                    pgoff_t idx_to)
 364 {
 365         struct fuse_inode *fi = get_fuse_inode(inode);
 366         bool found;
 367
 368         spin_lock(&fi->lock);
 369         found = fuse_find_writeback(fi, idx_from, idx_to);
 370         spin_unlock(&fi->lock);
 371
 372         return found;
 373 }
 374
 375 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 376 {
 377         return fuse_range_is_writeback(inode, index, index);
 378 }
 379
 380 /*
 381  * Wait for page writeback to be completed.
 382  *
 383  * Since fuse doesn't rely on the VM writeback tracking, this has to
 384  * use some other means.
 385  */
 386 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 387 {
 388         struct fuse_inode *fi = get_fuse_inode(inode);
 389
 390         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 391         return 0;
 392 }
 393
 394 /*
 395  * Wait for all pending writepages on the inode to finish.
 396  *
 397  * This is currently done by blocking further writes with FUSE_NOWRITE
 398  * and waiting for all sent writes to complete.
 399  *
 400  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 401  * could conflict with truncation.
 402  */
 403 static void fuse_sync_writes(struct inode *inode)
 404 {
 405         fuse_set_nowrite(inode);
 406         fuse_release_nowrite(inode);
 407 }
 408
 409 static int fuse_flush(struct file *file, fl_owner_t id)
 410 {
 411         struct inode *inode = file_inode(file);
 412         struct fuse_conn *fc = get_fuse_conn(inode);
 413         struct fuse_file *ff = file->private_data;
 414         struct fuse_req *req;
 415         struct fuse_flush_in inarg;
 416         int err;
 417
 418         if (is_bad_inode(inode))
 419                 return -EIO;
 420
 421         if (fc->no_flush)
 422                 return 0;
 423
 424         err = write_inode_now(inode, 1);
 425         if (err)
 426                 return err;
 427
 428         inode_lock(inode);
 429         fuse_sync_writes(inode);
 430         inode_unlock(inode);
 431
 432         err = filemap_check_errors(file->f_mapping);
 433         if (err)
 434                 return err;
 435
 436         req = fuse_get_req_nofail_nopages(fc, file);
 437         memset(&inarg, 0, sizeof(inarg));
 438         inarg.fh = ff->fh;
 439         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 440         req->in.h.opcode = FUSE_FLUSH;
 441         req->in.h.nodeid = get_node_id(inode);
 442         req->in.numargs = 1;
 443         req->in.args[0].size = sizeof(inarg);
 444         req->in.args[0].value = &inarg;
 445         __set_bit(FR_FORCE, &req->flags);
 446         fuse_request_send(fc, req);
 447         err = req->out.h.error;
 448         fuse_put_request(fc, req);
 449         if (err == -ENOSYS) {
 450                 fc->no_flush = 1;
 451                 err = 0;
 452         }
 453         return err;
 454 }
 455
 456 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 457                       int datasync, int opcode)
 458 {
 459         struct inode *inode = file->f_mapping->host;
 460         struct fuse_conn *fc = get_fuse_conn(inode);
 461         struct fuse_file *ff = file->private_data;
 462         FUSE_ARGS(args);
 463         struct fuse_fsync_in inarg;
 464
 465         memset(&inarg, 0, sizeof(inarg));
 466         inarg.fh = ff->fh;
 467         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 468         args.in.h.opcode = opcode;
 469         args.in.h.nodeid = get_node_id(inode);
 470         args.in.numargs = 1;
 471         args.in.args[0].size = sizeof(inarg);
 472         args.in.args[0].value = &inarg;
 473         return fuse_simple_request(fc, &args);
 474 }
 475
 476 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 477                       int datasync)
 478 {
 479         struct inode *inode = file->f_mapping->host;
 480         struct fuse_conn *fc = get_fuse_conn(inode);
 481         int err;
 482
 483         if (is_bad_inode(inode))
 484                 return -EIO;
 485
 486         inode_lock(inode);
 487
 488         /*
 489          * Start writeback against all dirty pages of the inode, then
 490          * wait for all outstanding writes, before sending the FSYNC
 491          * request.
 492          */
 493         err = file_write_and_wait_range(file, start, end);
 494         if (err)
 495                 goto out;
 496
 497         fuse_sync_writes(inode);
 498
 499         /*
 500          * Due to implementation of fuse writeback
 501          * file_write_and_wait_range() does not catch errors.
 502          * We have to do this directly after fuse_sync_writes()
 503          */
 504         err = file_check_and_advance_wb_err(file);
 505         if (err)
 506                 goto out;
 507
 508         err = sync_inode_metadata(inode, 1);
 509         if (err)
 510                 goto out;
 511
 512         if (fc->no_fsync)
 513                 goto out;
 514
 515         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 516         if (err == -ENOSYS) {
 517                 fc->no_fsync = 1;
 518                 err = 0;
 519         }
 520 out:
 521         inode_unlock(inode);
 522
 523         return err;
 524 }
 525
 526 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 527                     size_t count, int opcode)
 528 {
 529         struct fuse_read_in *inarg = &req->misc.read.in;
 530         struct fuse_file *ff = file->private_data;
 531
 532         inarg->fh = ff->fh;
 533         inarg->offset = pos;
 534         inarg->size = count;
 535         inarg->flags = file->f_flags;
 536         req->in.h.opcode = opcode;
 537         req->in.h.nodeid = ff->nodeid;
 538         req->in.numargs = 1;
 539         req->in.args[0].size = sizeof(struct fuse_read_in);
 540         req->in.args[0].value = inarg;
 541         req->out.argvar = 1;
 542         req->out.numargs = 1;
 543         req->out.args[0].size = count;
 544 }
 545
 546 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 547 {
 548         unsigned i;
 549
 550         for (i = 0; i < req->num_pages; i++) {
 551                 struct page *page = req->pages[i];
 552                 if (should_dirty)
 553                         set_page_dirty_lock(page);
 554                 put_page(page);
 555         }
 556 }
 557
 558 static void fuse_io_release(struct kref *kref)
 559 {
 560         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 561 }
 562
 563 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 564 {
 565         if (io->err)
 566                 return io->err;
 567
 568         if (io->bytes >= 0 && io->write)
 569                 return -EIO;
 570
 571         return io->bytes < 0 ? io->size : io->bytes;
 572 }
 573
 574 /**
 575  * In case of short read, the caller sets 'pos' to the position of
 576  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 577  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 578  *
 579  * An example:
 580  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 581  * both submitted asynchronously. The first of them was ACKed by userspace as
 582  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 583  * second request was ACKed as short, e.g. only 1K was read, resulting in
 584  * pos == 33K.
 585  *
 586  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 587  * will be equal to the length of the longest contiguous fragment of
 588  * transferred data starting from the beginning of IO request.
 589  */
 590 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 591 {
 592         int left;
 593
 594         spin_lock(&io->lock);
 595         if (err)
 596                 io->err = io->err ? : err;
 597         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 598                 io->bytes = pos;
 599
 600         left = --io->reqs;
 601         if (!left && io->blocking)
 602                 complete(io->done);
 603         spin_unlock(&io->lock);
 604
 605         if (!left && !io->blocking) {
 606                 ssize_t res = fuse_get_res_by_io(io);
 607
 608                 if (res >= 0) {
 609                         struct inode *inode = file_inode(io->iocb->ki_filp);
 610                         struct fuse_conn *fc = get_fuse_conn(inode);
 611                         struct fuse_inode *fi = get_fuse_inode(inode);
 612
 613                         spin_lock(&fi->lock);
 614                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 615                         spin_unlock(&fi->lock);
 616                 }
 617
 618                 io->iocb->ki_complete(io->iocb, res, 0);
 619         }
 620
 621         kref_put(&io->refcnt, fuse_io_release);
 622 }
 623
 624 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 625 {
 626         struct fuse_io_priv *io = req->io;
 627         ssize_t pos = -1;
 628
 629         fuse_release_user_pages(req, io->should_dirty);
 630
 631         if (io->write) {
 632                 if (req->misc.write.in.size != req->misc.write.out.size)
 633                         pos = req->misc.write.in.offset - io->offset +
 634                                 req->misc.write.out.size;
 635         } else {
 636                 if (req->misc.read.in.size != req->out.args[0].size)
 637                         pos = req->misc.read.in.offset - io->offset +
 638                                 req->out.args[0].size;
 639         }
 640
 641         fuse_aio_complete(io, req->out.h.error, pos);
 642 }
 643
 644 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 645                 size_t num_bytes, struct fuse_io_priv *io)
 646 {
 647         spin_lock(&io->lock);
 648         kref_get(&io->refcnt);
 649         io->size += num_bytes;
 650         io->reqs++;
 651         spin_unlock(&io->lock);
 652
 653         req->io = io;
 654         req->end = fuse_aio_complete_req;
 655
 656         __fuse_get_request(req);
 657         fuse_request_send_background(fc, req);
 658
 659         return num_bytes;
 660 }
 661
 662 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 663                              loff_t pos, size_t count, fl_owner_t owner)
 664 {
 665         struct file *file = io->iocb->ki_filp;
 666         struct fuse_file *ff = file->private_data;
 667         struct fuse_conn *fc = ff->fc;
 668
 669         fuse_read_fill(req, file, pos, count, FUSE_READ);
 670         if (owner != NULL) {
 671                 struct fuse_read_in *inarg = &req->misc.read.in;
 672
 673                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 674                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 675         }
 676
 677         if (io->async)
 678                 return fuse_async_req_send(fc, req, count, io);
 679
 680         fuse_request_send(fc, req);
 681         return req->out.args[0].size;
 682 }
 683
 684 static void fuse_read_update_size(struct inode *inode, loff_t size,
 685                                   u64 attr_ver)
 686 {
 687         struct fuse_conn *fc = get_fuse_conn(inode);
 688         struct fuse_inode *fi = get_fuse_inode(inode);
 689
 690         spin_lock(&fi->lock);
 691         if (attr_ver == fi->attr_version && size < inode->i_size &&
 692             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 693                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 694                 i_size_write(inode, size);
 695         }
 696         spin_unlock(&fi->lock);
 697 }
 698
 699 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 700                             u64 attr_ver)
 701 {
 702         size_t num_read = req->out.args[0].size;
 703         struct fuse_conn *fc = get_fuse_conn(inode);
 704
 705         if (fc->writeback_cache) {
 706                 /*
 707                  * A hole in a file. Some data after the hole are in page cache,
 708                  * but have not reached the client fs yet. So, the hole is not
 709                  * present there.
 710                  */
 711                 int i;
 712                 int start_idx = num_read >> PAGE_SHIFT;
 713                 size_t off = num_read & (PAGE_SIZE - 1);
 714
 715                 for (i = start_idx; i < req->num_pages; i++) {
 716                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 717                         off = 0;
 718                 }
 719         } else {
 720                 loff_t pos = page_offset(req->pages[0]) + num_read;
 721                 fuse_read_update_size(inode, pos, attr_ver);
 722         }
 723 }
 724
 725 static int fuse_do_readpage(struct file *file, struct page *page)
 726 {
 727         struct kiocb iocb;
 728         struct fuse_io_priv io;
 729         struct inode *inode = page->mapping->host;
 730         struct fuse_conn *fc = get_fuse_conn(inode);
 731         struct fuse_req *req;
 732         size_t num_read;
 733         loff_t pos = page_offset(page);
 734         size_t count = PAGE_SIZE;
 735         u64 attr_ver;
 736         int err;
 737
 738         /*
 739          * Page writeback can extend beyond the lifetime of the
 740          * page-cache page, so make sure we read a properly synced
 741          * page.
 742          */
 743         fuse_wait_on_page_writeback(inode, page->index);
 744
 745         req = fuse_get_req(fc, 1);
 746         if (IS_ERR(req))
 747                 return PTR_ERR(req);
 748
 749         attr_ver = fuse_get_attr_version(fc);
 750
 751         req->out.page_zeroing = 1;
 752         req->out.argpages = 1;
 753         req->num_pages = 1;
 754         req->pages[0] = page;
 755         req->page_descs[0].length = count;
 756         init_sync_kiocb(&iocb, file);
 757         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 758         num_read = fuse_send_read(req, &io, pos, count, NULL);
 759         err = req->out.h.error;
 760
 761         if (!err) {
 762                 /*
 763                  * Short read means EOF.  If file size is larger, truncate it
 764                  */
 765                 if (num_read < count)
 766                         fuse_short_read(req, inode, attr_ver);
 767
 768                 SetPageUptodate(page);
 769         }
 770
 771         fuse_put_request(fc, req);
 772
 773         return err;
 774 }
 775
 776 static int fuse_readpage(struct file *file, struct page *page)
 777 {
 778         struct inode *inode = page->mapping->host;
 779         int err;
 780
 781         err = -EIO;
 782         if (is_bad_inode(inode))
 783                 goto out;
 784
 785         err = fuse_do_readpage(file, page);
 786         fuse_invalidate_atime(inode);
 787  out:
 788         unlock_page(page);
 789         return err;
 790 }
 791
 792 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 793 {
 794         int i;
 795         size_t count = req->misc.read.in.size;
 796         size_t num_read = req->out.args[0].size;
 797         struct address_space *mapping = NULL;
 798
 799         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 800                 mapping = req->pages[i]->mapping;
 801
 802         if (mapping) {
 803                 struct inode *inode = mapping->host;
 804
 805                 /*
 806                  * Short read means EOF. If file size is larger, truncate it
 807                  */
 808                 if (!req->out.h.error && num_read < count)
 809                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 810
 811                 fuse_invalidate_atime(inode);
 812         }
 813
 814         for (i = 0; i < req->num_pages; i++) {
 815                 struct page *page = req->pages[i];
 816                 if (!req->out.h.error)
 817                         SetPageUptodate(page);
 818                 else
 819                         SetPageError(page);
 820                 unlock_page(page);
 821                 put_page(page);
 822         }
 823         if (req->ff)
 824                 fuse_file_put(req->ff, false, false);
 825 }
 826
 827 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 828 {
 829         struct fuse_file *ff = file->private_data;
 830         struct fuse_conn *fc = ff->fc;
 831         loff_t pos = page_offset(req->pages[0]);
 832         size_t count = req->num_pages << PAGE_SHIFT;
 833
 834         req->out.argpages = 1;
 835         req->out.page_zeroing = 1;
 836         req->out.page_replace = 1;
 837         fuse_read_fill(req, file, pos, count, FUSE_READ);
 838         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 839         if (fc->async_read) {
 840                 req->ff = fuse_file_get(ff);
 841                 req->end = fuse_readpages_end;
 842                 fuse_request_send_background(fc, req);
 843         } else {
 844                 fuse_request_send(fc, req);
 845                 fuse_readpages_end(fc, req);
 846                 fuse_put_request(fc, req);
 847         }
 848 }
 849
 850 struct fuse_fill_data {
 851         struct fuse_req *req;
 852         struct file *file;
 853         struct inode *inode;
 854         unsigned nr_pages;
 855 };
 856
 857 static int fuse_readpages_fill(void *_data, struct page *page)
 858 {
 859         struct fuse_fill_data *data = _data;
 860         struct fuse_req *req = data->req;
 861         struct inode *inode = data->inode;
 862         struct fuse_conn *fc = get_fuse_conn(inode);
 863
 864         fuse_wait_on_page_writeback(inode, page->index);
 865
 866         if (req->num_pages &&
 867             (req->num_pages == fc->max_pages ||
 868              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 869              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 870                 unsigned int nr_alloc = min_t(unsigned int, data->nr_pages,
 871                                               fc->max_pages);
 872                 fuse_send_readpages(req, data->file);
 873                 if (fc->async_read)
 874                         req = fuse_get_req_for_background(fc, nr_alloc);
 875                 else
 876                         req = fuse_get_req(fc, nr_alloc);
 877
 878                 data->req = req;
 879                 if (IS_ERR(req)) {
 880                         unlock_page(page);
 881                         return PTR_ERR(req);
 882                 }
 883         }
 884
 885         if (WARN_ON(req->num_pages >= req->max_pages)) {
 886                 unlock_page(page);
 887                 fuse_put_request(fc, req);
 888                 return -EIO;
 889         }
 890
 891         get_page(page);
 892         req->pages[req->num_pages] = page;
 893         req->page_descs[req->num_pages].length = PAGE_SIZE;
 894         req->num_pages++;
 895         data->nr_pages--;
 896         return 0;
 897 }
 898
 899 static int fuse_readpages(struct file *file, struct address_space *mapping,
 900                           struct list_head *pages, unsigned nr_pages)
 901 {
 902         struct inode *inode = mapping->host;
 903         struct fuse_conn *fc = get_fuse_conn(inode);
 904         struct fuse_fill_data data;
 905         int err;
 906         unsigned int nr_alloc = min_t(unsigned int, nr_pages, fc->max_pages);
 907
 908         err = -EIO;
 909         if (is_bad_inode(inode))
 910                 goto out;
 911
 912         data.file = file;
 913         data.inode = inode;
 914         if (fc->async_read)
 915                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 916         else
 917                 data.req = fuse_get_req(fc, nr_alloc);
 918         data.nr_pages = nr_pages;
 919         err = PTR_ERR(data.req);
 920         if (IS_ERR(data.req))
 921                 goto out;
 922
 923         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 924         if (!err) {
 925                 if (data.req->num_pages)
 926                         fuse_send_readpages(data.req, file);
 927                 else
 928                         fuse_put_request(fc, data.req);
 929         }
 930 out:
 931         return err;
 932 }
 933
 934 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 935 {
 936         struct inode *inode = iocb->ki_filp->f_mapping->host;
 937         struct fuse_conn *fc = get_fuse_conn(inode);
 938
 939         /*
 940          * In auto invalidate mode, always update attributes on read.
 941          * Otherwise, only update if we attempt to read past EOF (to ensure
 942          * i_size is up to date).
 943          */
 944         if (fc->auto_inval_data ||
 945             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 946                 int err;
 947                 err = fuse_update_attributes(inode, iocb->ki_filp);
 948                 if (err)
 949                         return err;
 950         }
 951
 952         return generic_file_read_iter(iocb, to);
 953 }
 954
 955 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 956                             loff_t pos, size_t count)
 957 {
 958         struct fuse_write_in *inarg = &req->misc.write.in;
 959         struct fuse_write_out *outarg = &req->misc.write.out;
 960
 961         inarg->fh = ff->fh;
 962         inarg->offset = pos;
 963         inarg->size = count;
 964         req->in.h.opcode = FUSE_WRITE;
 965         req->in.h.nodeid = ff->nodeid;
 966         req->in.numargs = 2;
 967         if (ff->fc->minor < 9)
 968                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 969         else
 970                 req->in.args[0].size = sizeof(struct fuse_write_in);
 971         req->in.args[0].value = inarg;
 972         req->in.args[1].size = count;
 973         req->out.numargs = 1;
 974         req->out.args[0].size = sizeof(struct fuse_write_out);
 975         req->out.args[0].value = outarg;
 976 }
 977
 978 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 979                               loff_t pos, size_t count, fl_owner_t owner)
 980 {
 981         struct kiocb *iocb = io->iocb;
 982         struct file *file = iocb->ki_filp;
 983         struct fuse_file *ff = file->private_data;
 984         struct fuse_conn *fc = ff->fc;
 985         struct fuse_write_in *inarg = &req->misc.write.in;
 986
 987         fuse_write_fill(req, ff, pos, count);
 988         inarg->flags = file->f_flags;
 989         if (iocb->ki_flags & IOCB_DSYNC)
 990                 inarg->flags |= O_DSYNC;
 991         if (iocb->ki_flags & IOCB_SYNC)
 992                 inarg->flags |= O_SYNC;
 993         if (owner != NULL) {
 994                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 995                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 996         }
 997
 998         if (io->async)
 999                 return fuse_async_req_send(fc, req, count, io);
1000
1001         fuse_request_send(fc, req);
1002         return req->misc.write.out.size;
1003 }
1004
1005 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1006 {
1007         struct fuse_conn *fc = get_fuse_conn(inode);
1008         struct fuse_inode *fi = get_fuse_inode(inode);
1009         bool ret = false;
1010
1011         spin_lock(&fi->lock);
1012         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1013         if (pos > inode->i_size) {
1014                 i_size_write(inode, pos);
1015                 ret = true;
1016         }
1017         spin_unlock(&fi->lock);
1018
1019         return ret;
1020 }
1021
1022 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1023                                     struct inode *inode, loff_t pos,
1024                                     size_t count)
1025 {
1026         size_t res;
1027         unsigned offset;
1028         unsigned i;
1029         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1030
1031         for (i = 0; i < req->num_pages; i++)
1032                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1033
1034         res = fuse_send_write(req, &io, pos, count, NULL);
1035
1036         offset = req->page_descs[0].offset;
1037         count = res;
1038         for (i = 0; i < req->num_pages; i++) {
1039                 struct page *page = req->pages[i];
1040
1041                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1042                         SetPageUptodate(page);
1043
1044                 if (count > PAGE_SIZE - offset)
1045                         count -= PAGE_SIZE - offset;
1046                 else
1047                         count = 0;
1048                 offset = 0;
1049
1050                 unlock_page(page);
1051                 put_page(page);
1052         }
1053
1054         return res;
1055 }
1056
1057 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1058                                struct address_space *mapping,
1059                                struct iov_iter *ii, loff_t pos)
1060 {
1061         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1062         unsigned offset = pos & (PAGE_SIZE - 1);
1063         size_t count = 0;
1064         int err;
1065
1066         req->in.argpages = 1;
1067         req->page_descs[0].offset = offset;
1068
1069         do {
1070                 size_t tmp;
1071                 struct page *page;
1072                 pgoff_t index = pos >> PAGE_SHIFT;
1073                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1074                                      iov_iter_count(ii));
1075
1076                 bytes = min_t(size_t, bytes, fc->max_write - count);
1077
1078  again:
1079                 err = -EFAULT;
1080                 if (iov_iter_fault_in_readable(ii, bytes))
1081                         break;
1082
1083                 err = -ENOMEM;
1084                 page = grab_cache_page_write_begin(mapping, index, 0);
1085                 if (!page)
1086                         break;
1087
1088                 if (mapping_writably_mapped(mapping))
1089                         flush_dcache_page(page);
1090
1091                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1092                 flush_dcache_page(page);
1093
1094                 iov_iter_advance(ii, tmp);
1095                 if (!tmp) {
1096                         unlock_page(page);
1097                         put_page(page);
1098                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1099                         goto again;
1100                 }
1101
1102                 err = 0;
1103                 req->pages[req->num_pages] = page;
1104                 req->page_descs[req->num_pages].length = tmp;
1105                 req->num_pages++;
1106
1107                 count += tmp;
1108                 pos += tmp;
1109                 offset += tmp;
1110                 if (offset == PAGE_SIZE)
1111                         offset = 0;
1112
1113                 if (!fc->big_writes)
1114                         break;
1115         } while (iov_iter_count(ii) && count < fc->max_write &&
1116                  req->num_pages < req->max_pages && offset == 0);
1117
1118         return count > 0 ? count : err;
1119 }
1120
1121 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1122                                      unsigned int max_pages)
1123 {
1124         return min_t(unsigned int,
1125                      ((pos + len - 1) >> PAGE_SHIFT) -
1126                      (pos >> PAGE_SHIFT) + 1,
1127                      max_pages);
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 (inode->i_size < pos + iov_iter_count(ii))
1141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1142
1143         do {
1144                 struct fuse_req *req;
1145                 ssize_t count;
1146                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1147                                                       fc->max_pages);
1148
1149                 req = fuse_get_req(fc, nr_pages);
1150                 if (IS_ERR(req)) {
1151                         err = PTR_ERR(req);
1152                         break;
1153                 }
1154
1155                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1156                 if (count <= 0) {
1157                         err = count;
1158                 } else {
1159                         size_t num_written;
1160
1161                         num_written = fuse_send_write_pages(req, iocb, inode,
1162                                                             pos, count);
1163                         err = req->out.h.error;
1164                         if (!err) {
1165                                 res += num_written;
1166                                 pos += num_written;
1167
1168                                 /* break out of the loop on short write */
1169                                 if (num_written != count)
1170                                         err = -EIO;
1171                         }
1172                 }
1173                 fuse_put_request(fc, req);
1174         } while (!err && iov_iter_count(ii));
1175
1176         if (res > 0)
1177                 fuse_write_update_size(inode, pos);
1178
1179         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1180         fuse_invalidate_attr(inode);
1181
1182         return res > 0 ? res : err;
1183 }
1184
1185 static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
1186 {
1187         struct file *file = iocb->ki_filp;
1188         struct address_space *mapping = file->f_mapping;
1189         ssize_t written = 0;
1190         ssize_t written_buffered = 0;
1191         struct inode *inode = mapping->host;
1192         ssize_t err;
1193         loff_t endbyte = 0;
1194
1195         if (get_fuse_conn(inode)->writeback_cache) {
1196                 /* Update size (EOF optimization) and mode (SUID clearing) */
1197                 err = fuse_update_attributes(mapping->host, file);
1198                 if (err)
1199                         return err;
1200
1201                 return generic_file_write_iter(iocb, from);
1202         }
1203
1204         inode_lock(inode);
1205
1206         /* We can write back this queue in page reclaim */
1207         current->backing_dev_info = inode_to_bdi(inode);
1208
1209         err = generic_write_checks(iocb, from);
1210         if (err <= 0)
1211                 goto out;
1212
1213         err = file_remove_privs(file);
1214         if (err)
1215                 goto out;
1216
1217         err = file_update_time(file);
1218         if (err)
1219                 goto out;
1220
1221         if (iocb->ki_flags & IOCB_DIRECT) {
1222                 loff_t pos = iocb->ki_pos;
1223                 written = generic_file_direct_write(iocb, from);
1224                 if (written < 0 || !iov_iter_count(from))
1225                         goto out;
1226
1227                 pos += written;
1228
1229                 written_buffered = fuse_perform_write(iocb, mapping, from, pos);
1230                 if (written_buffered < 0) {
1231                         err = written_buffered;
1232                         goto out;
1233                 }
1234                 endbyte = pos + written_buffered - 1;
1235
1236                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1237                                                    endbyte);
1238                 if (err)
1239                         goto out;
1240
1241                 invalidate_mapping_pages(file->f_mapping,
1242                                          pos >> PAGE_SHIFT,
1243                                          endbyte >> PAGE_SHIFT);
1244
1245                 written += written_buffered;
1246                 iocb->ki_pos = pos + written_buffered;
1247         } else {
1248                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1249                 if (written >= 0)
1250                         iocb->ki_pos += written;
1251         }
1252 out:
1253         current->backing_dev_info = NULL;
1254         inode_unlock(inode);
1255         if (written > 0)
1256                 written = generic_write_sync(iocb, written);
1257
1258         return written ? written : err;
1259 }
1260
1261 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1262                 unsigned index, unsigned nr_pages)
1263 {
1264         int i;
1265
1266         for (i = index; i < index + nr_pages; i++)
1267                 req->page_descs[i].length = PAGE_SIZE -
1268                         req->page_descs[i].offset;
1269 }
1270
1271 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1272 {
1273         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1274 }
1275
1276 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1277                                         size_t max_size)
1278 {
1279         return min(iov_iter_single_seg_count(ii), max_size);
1280 }
1281
1282 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1283                                size_t *nbytesp, int write)
1284 {
1285         size_t nbytes = 0;  /* # bytes already packed in req */
1286         ssize_t ret = 0;
1287
1288         /* Special case for kernel I/O: can copy directly into the buffer */
1289         if (iov_iter_is_kvec(ii)) {
1290                 unsigned long user_addr = fuse_get_user_addr(ii);
1291                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1292
1293                 if (write)
1294                         req->in.args[1].value = (void *) user_addr;
1295                 else
1296                         req->out.args[0].value = (void *) user_addr;
1297
1298                 iov_iter_advance(ii, frag_size);
1299                 *nbytesp = frag_size;
1300                 return 0;
1301         }
1302
1303         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1304                 unsigned npages;
1305                 size_t start;
1306                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1307                                         *nbytesp - nbytes,
1308                                         req->max_pages - req->num_pages,
1309                                         &start);
1310                 if (ret < 0)
1311                         break;
1312
1313                 iov_iter_advance(ii, ret);
1314                 nbytes += ret;
1315
1316                 ret += start;
1317                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1318
1319                 req->page_descs[req->num_pages].offset = start;
1320                 fuse_page_descs_length_init(req, req->num_pages, npages);
1321
1322                 req->num_pages += npages;
1323                 req->page_descs[req->num_pages - 1].length -=
1324                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1325         }
1326
1327         if (write)
1328                 req->in.argpages = 1;
1329         else
1330                 req->out.argpages = 1;
1331
1332         *nbytesp = nbytes;
1333
1334         return ret < 0 ? ret : 0;
1335 }
1336
1337 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1338                        loff_t *ppos, int flags)
1339 {
1340         int write = flags & FUSE_DIO_WRITE;
1341         int cuse = flags & FUSE_DIO_CUSE;
1342         struct file *file = io->iocb->ki_filp;
1343         struct inode *inode = file->f_mapping->host;
1344         struct fuse_file *ff = file->private_data;
1345         struct fuse_conn *fc = ff->fc;
1346         size_t nmax = write ? fc->max_write : fc->max_read;
1347         loff_t pos = *ppos;
1348         size_t count = iov_iter_count(iter);
1349         pgoff_t idx_from = pos >> PAGE_SHIFT;
1350         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1351         ssize_t res = 0;
1352         struct fuse_req *req;
1353         int err = 0;
1354
1355         if (io->async)
1356                 req = fuse_get_req_for_background(fc, iov_iter_npages(iter,
1357                                                                 fc->max_pages));
1358         else
1359                 req = fuse_get_req(fc, iov_iter_npages(iter, fc->max_pages));
1360         if (IS_ERR(req))
1361                 return PTR_ERR(req);
1362
1363         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1364                 if (!write)
1365                         inode_lock(inode);
1366                 fuse_sync_writes(inode);
1367                 if (!write)
1368                         inode_unlock(inode);
1369         }
1370
1371         io->should_dirty = !write && iter_is_iovec(iter);
1372         while (count) {
1373                 size_t nres;
1374                 fl_owner_t owner = current->files;
1375                 size_t nbytes = min(count, nmax);
1376                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1377                 if (err && !nbytes)
1378                         break;
1379
1380                 if (write)
1381                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1382                 else
1383                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1384
1385                 if (!io->async)
1386                         fuse_release_user_pages(req, io->should_dirty);
1387                 if (req->out.h.error) {
1388                         err = req->out.h.error;
1389                         break;
1390                 } else if (nres > nbytes) {
1391                         res = 0;
1392                         err = -EIO;
1393                         break;
1394                 }
1395                 count -= nres;
1396                 res += nres;
1397                 pos += nres;
1398                 if (nres != nbytes)
1399                         break;
1400                 if (count) {
1401                         fuse_put_request(fc, req);
1402                         if (io->async)
1403                                 req = fuse_get_req_for_background(fc,
1404                                         iov_iter_npages(iter, fc->max_pages));
1405                         else
1406                                 req = fuse_get_req(fc, iov_iter_npages(iter,
1407                                                                 fc->max_pages));
1408                         if (IS_ERR(req))
1409                                 break;
1410                 }
1411         }
1412         if (!IS_ERR(req))
1413                 fuse_put_request(fc, req);
1414         if (res > 0)
1415                 *ppos = pos;
1416
1417         return res > 0 ? res : err;
1418 }
1419 EXPORT_SYMBOL_GPL(fuse_direct_io);
1420
1421 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1422                                   struct iov_iter *iter,
1423                                   loff_t *ppos)
1424 {
1425         ssize_t res;
1426         struct inode *inode = file_inode(io->iocb->ki_filp);
1427
1428         res = fuse_direct_io(io, iter, ppos, 0);
1429
1430         fuse_invalidate_atime(inode);
1431
1432         return res;
1433 }
1434
1435 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1436
1437 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1438 {
1439         ssize_t res;
1440
1441         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1442                 res = fuse_direct_IO(iocb, to);
1443         } else {
1444                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1445
1446                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1447         }
1448
1449         return res;
1450 }
1451
1452 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1453 {
1454         struct inode *inode = file_inode(iocb->ki_filp);
1455         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1456         ssize_t res;
1457
1458         /* Don't allow parallel writes to the same file */
1459         inode_lock(inode);
1460         res = generic_write_checks(iocb, from);
1461         if (res > 0) {
1462                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1463                         res = fuse_direct_IO(iocb, from);
1464                 } else {
1465                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1466                                              FUSE_DIO_WRITE);
1467                 }
1468         }
1469         fuse_invalidate_attr(inode);
1470         if (res > 0)
1471                 fuse_write_update_size(inode, iocb->ki_pos);
1472         inode_unlock(inode);
1473
1474         return res;
1475 }
1476
1477 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1478 {
1479         struct file *file = iocb->ki_filp;
1480         struct fuse_file *ff = file->private_data;
1481
1482         if (is_bad_inode(file_inode(file)))
1483                 return -EIO;
1484
1485         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1486                 return fuse_cache_read_iter(iocb, to);
1487         else
1488                 return fuse_direct_read_iter(iocb, to);
1489 }
1490
1491 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1492 {
1493         struct file *file = iocb->ki_filp;
1494         struct fuse_file *ff = file->private_data;
1495
1496         if (is_bad_inode(file_inode(file)))
1497                 return -EIO;
1498
1499         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1500                 return fuse_cache_write_iter(iocb, from);
1501         else
1502                 return fuse_direct_write_iter(iocb, from);
1503 }
1504
1505 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1506 {
1507         int i;
1508
1509         for (i = 0; i < req->num_pages; i++)
1510                 __free_page(req->pages[i]);
1511
1512         if (req->ff)
1513                 fuse_file_put(req->ff, false, false);
1514 }
1515
1516 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1517 {
1518         struct inode *inode = req->inode;
1519         struct fuse_inode *fi = get_fuse_inode(inode);
1520         struct backing_dev_info *bdi = inode_to_bdi(inode);
1521         int i;
1522
1523         list_del(&req->writepages_entry);
1524         for (i = 0; i < req->num_pages; i++) {
1525                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1526                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1527                 wb_writeout_inc(&bdi->wb);
1528         }
1529         wake_up(&fi->page_waitq);
1530 }
1531
1532 /* Called under fi->lock, may release and reacquire it */
1533 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1534                                 loff_t size)
1535 __releases(fi->lock)
1536 __acquires(fi->lock)
1537 {
1538         struct fuse_req *aux, *next;
1539         struct fuse_inode *fi = get_fuse_inode(req->inode);
1540         struct fuse_write_in *inarg = &req->misc.write.in;
1541         __u64 data_size = req->num_pages * PAGE_SIZE;
1542         bool queued;
1543
1544         if (inarg->offset + data_size <= size) {
1545                 inarg->size = data_size;
1546         } else if (inarg->offset < size) {
1547                 inarg->size = size - inarg->offset;
1548         } else {
1549                 /* Got truncated off completely */
1550                 goto out_free;
1551         }
1552
1553         req->in.args[1].size = inarg->size;
1554         queued = fuse_request_queue_background(fc, req);
1555         /* Fails on broken connection only */
1556         if (unlikely(!queued))
1557                 goto out_free;
1558
1559         fi->writectr++;
1560         return;
1561
1562  out_free:
1563         fuse_writepage_finish(fc, req);
1564         spin_unlock(&fi->lock);
1565
1566         /* After fuse_writepage_finish() aux request list is private */
1567         for (aux = req->misc.write.next; aux; aux = next) {
1568                 next = aux->misc.write.next;
1569                 aux->misc.write.next = NULL;
1570                 fuse_writepage_free(fc, aux);
1571                 fuse_put_request(fc, aux);
1572         }
1573
1574         fuse_writepage_free(fc, req);
1575         fuse_put_request(fc, req);
1576         spin_lock(&fi->lock);
1577 }
1578
1579 /*
1580  * If fi->writectr is positive (no truncate or fsync going on) send
1581  * all queued writepage requests.
1582  *
1583  * Called with fi->lock
1584  */
1585 void fuse_flush_writepages(struct inode *inode)
1586 __releases(fi->lock)
1587 __acquires(fi->lock)
1588 {
1589         struct fuse_conn *fc = get_fuse_conn(inode);
1590         struct fuse_inode *fi = get_fuse_inode(inode);
1591         loff_t crop = i_size_read(inode);
1592         struct fuse_req *req;
1593
1594         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1595                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1596                 list_del_init(&req->list);
1597                 fuse_send_writepage(fc, req, crop);
1598         }
1599 }
1600
1601 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1602 {
1603         struct inode *inode = req->inode;
1604         struct fuse_inode *fi = get_fuse_inode(inode);
1605
1606         mapping_set_error(inode->i_mapping, req->out.h.error);
1607         spin_lock(&fi->lock);
1608         while (req->misc.write.next) {
1609                 struct fuse_conn *fc = get_fuse_conn(inode);
1610                 struct fuse_write_in *inarg = &req->misc.write.in;
1611                 struct fuse_req *next = req->misc.write.next;
1612                 req->misc.write.next = next->misc.write.next;
1613                 next->misc.write.next = NULL;
1614                 next->ff = fuse_file_get(req->ff);
1615                 list_add(&next->writepages_entry, &fi->writepages);
1616
1617                 /*
1618                  * Skip fuse_flush_writepages() to make it easy to crop requests
1619                  * based on primary request size.
1620                  *
1621                  * 1st case (trivial): there are no concurrent activities using
1622                  * fuse_set/release_nowrite.  Then we're on safe side because
1623                  * fuse_flush_writepages() would call fuse_send_writepage()
1624                  * anyway.
1625                  *
1626                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1627                  * now for completion of all in-flight requests.  This happens
1628                  * rarely and no more than once per page, so this should be
1629                  * okay.
1630                  *
1631                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1632                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1633                  * that fuse_set_nowrite returned implies that all in-flight
1634                  * requests were completed along with all of their secondary
1635                  * requests.  Further primary requests are blocked by negative
1636                  * writectr.  Hence there cannot be any in-flight requests and
1637                  * no invocations of fuse_writepage_end() while we're in
1638                  * fuse_set_nowrite..fuse_release_nowrite section.
1639                  */
1640                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1641         }
1642         fi->writectr--;
1643         fuse_writepage_finish(fc, req);
1644         spin_unlock(&fi->lock);
1645         fuse_writepage_free(fc, req);
1646 }
1647
1648 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1649                                                struct fuse_inode *fi)
1650 {
1651         struct fuse_file *ff = NULL;
1652
1653         spin_lock(&fi->lock);
1654         if (!list_empty(&fi->write_files)) {
1655                 ff = list_entry(fi->write_files.next, struct fuse_file,
1656                                 write_entry);
1657                 fuse_file_get(ff);
1658         }
1659         spin_unlock(&fi->lock);
1660
1661         return ff;
1662 }
1663
1664 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1665                                              struct fuse_inode *fi)
1666 {
1667         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1668         WARN_ON(!ff);
1669         return ff;
1670 }
1671
1672 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1673 {
1674         struct fuse_conn *fc = get_fuse_conn(inode);
1675         struct fuse_inode *fi = get_fuse_inode(inode);
1676         struct fuse_file *ff;
1677         int err;
1678
1679         ff = __fuse_write_file_get(fc, fi);
1680         err = fuse_flush_times(inode, ff);
1681         if (ff)
1682                 fuse_file_put(ff, false, false);
1683
1684         return err;
1685 }
1686
1687 static int fuse_writepage_locked(struct page *page)
1688 {
1689         struct address_space *mapping = page->mapping;
1690         struct inode *inode = mapping->host;
1691         struct fuse_conn *fc = get_fuse_conn(inode);
1692         struct fuse_inode *fi = get_fuse_inode(inode);
1693         struct fuse_req *req;
1694         struct page *tmp_page;
1695         int error = -ENOMEM;
1696
1697         set_page_writeback(page);
1698
1699         req = fuse_request_alloc_nofs(1);
1700         if (!req)
1701                 goto err;
1702
1703         /* writeback always goes to bg_queue */
1704         __set_bit(FR_BACKGROUND, &req->flags);
1705         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1706         if (!tmp_page)
1707                 goto err_free;
1708
1709         error = -EIO;
1710         req->ff = fuse_write_file_get(fc, fi);
1711         if (!req->ff)
1712                 goto err_nofile;
1713
1714         fuse_write_fill(req, req->ff, page_offset(page), 0);
1715
1716         copy_highpage(tmp_page, page);
1717         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1718         req->misc.write.next = NULL;
1719         req->in.argpages = 1;
1720         req->num_pages = 1;
1721         req->pages[0] = tmp_page;
1722         req->page_descs[0].offset = 0;
1723         req->page_descs[0].length = PAGE_SIZE;
1724         req->end = fuse_writepage_end;
1725         req->inode = inode;
1726
1727         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1728         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1729
1730         spin_lock(&fi->lock);
1731         list_add(&req->writepages_entry, &fi->writepages);
1732         list_add_tail(&req->list, &fi->queued_writes);
1733         fuse_flush_writepages(inode);
1734         spin_unlock(&fi->lock);
1735
1736         end_page_writeback(page);
1737
1738         return 0;
1739
1740 err_nofile:
1741         __free_page(tmp_page);
1742 err_free:
1743         fuse_request_free(req);
1744 err:
1745         mapping_set_error(page->mapping, error);
1746         end_page_writeback(page);
1747         return error;
1748 }
1749
1750 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1751 {
1752         int err;
1753
1754         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1755                 /*
1756                  * ->writepages() should be called for sync() and friends.  We
1757                  * should only get here on direct reclaim and then we are
1758                  * allowed to skip a page which is already in flight
1759                  */
1760                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1761
1762                 redirty_page_for_writepage(wbc, page);
1763                 return 0;
1764         }
1765
1766         err = fuse_writepage_locked(page);
1767         unlock_page(page);
1768
1769         return err;
1770 }
1771
1772 struct fuse_fill_wb_data {
1773         struct fuse_req *req;
1774         struct fuse_file *ff;
1775         struct inode *inode;
1776         struct page **orig_pages;
1777 };
1778
1779 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1780 {
1781         struct fuse_req *req = data->req;
1782         struct inode *inode = data->inode;
1783         struct fuse_inode *fi = get_fuse_inode(inode);
1784         int num_pages = req->num_pages;
1785         int i;
1786
1787         req->ff = fuse_file_get(data->ff);
1788         spin_lock(&fi->lock);
1789         list_add_tail(&req->list, &fi->queued_writes);
1790         fuse_flush_writepages(inode);
1791         spin_unlock(&fi->lock);
1792
1793         for (i = 0; i < num_pages; i++)
1794                 end_page_writeback(data->orig_pages[i]);
1795 }
1796
1797 /*
1798  * First recheck under fi->lock if the offending offset is still under
1799  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1800  * one already added for a page at this offset.  If there's none, then insert
1801  * this new request onto the auxiliary list, otherwise reuse the existing one by
1802  * copying the new page contents over to the old temporary page.
1803  */
1804 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1805                                      struct page *page)
1806 {
1807         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1808         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1809         struct fuse_req *tmp;
1810         struct fuse_req *old_req;
1811
1812         WARN_ON(new_req->num_pages != 0);
1813
1814         spin_lock(&fi->lock);
1815         list_del(&new_req->writepages_entry);
1816         old_req = fuse_find_writeback(fi, page->index, page->index);
1817         if (!old_req) {
1818                 list_add(&new_req->writepages_entry, &fi->writepages);
1819                 spin_unlock(&fi->lock);
1820                 return false;
1821         }
1822
1823         new_req->num_pages = 1;
1824         for (tmp = old_req->misc.write.next; tmp; tmp = tmp->misc.write.next) {
1825                 pgoff_t curr_index;
1826
1827                 WARN_ON(tmp->inode != new_req->inode);
1828                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1829                 if (curr_index == page->index) {
1830                         WARN_ON(tmp->num_pages != 1);
1831                         WARN_ON(!test_bit(FR_PENDING, &tmp->flags));
1832                         swap(tmp->pages[0], new_req->pages[0]);
1833                         break;
1834                 }
1835         }
1836
1837         if (!tmp) {
1838                 new_req->misc.write.next = old_req->misc.write.next;
1839                 old_req->misc.write.next = new_req;
1840         }
1841
1842         spin_unlock(&fi->lock);
1843
1844         if (tmp) {
1845                 struct backing_dev_info *bdi = inode_to_bdi(new_req->inode);
1846
1847                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1848                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1849                 wb_writeout_inc(&bdi->wb);
1850                 fuse_writepage_free(fc, new_req);
1851                 fuse_request_free(new_req);
1852         }
1853
1854         return true;
1855 }
1856
1857 static int fuse_writepages_fill(struct page *page,
1858                 struct writeback_control *wbc, void *_data)
1859 {
1860         struct fuse_fill_wb_data *data = _data;
1861         struct fuse_req *req = data->req;
1862         struct inode *inode = data->inode;
1863         struct fuse_inode *fi = get_fuse_inode(inode);
1864         struct fuse_conn *fc = get_fuse_conn(inode);
1865         struct page *tmp_page;
1866         bool is_writeback;
1867         int err;
1868
1869         if (!data->ff) {
1870                 err = -EIO;
1871                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1872                 if (!data->ff)
1873                         goto out_unlock;
1874         }
1875
1876         /*
1877          * Being under writeback is unlikely but possible.  For example direct
1878          * read to an mmaped fuse file will set the page dirty twice; once when
1879          * the pages are faulted with get_user_pages(), and then after the read
1880          * completed.
1881          */
1882         is_writeback = fuse_page_is_writeback(inode, page->index);
1883
1884         if (req && req->num_pages &&
1885             (is_writeback || req->num_pages == fc->max_pages ||
1886              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1887              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1888                 fuse_writepages_send(data);
1889                 data->req = NULL;
1890         } else if (req && req->num_pages == req->max_pages) {
1891                 if (!fuse_req_realloc_pages(fc, req, GFP_NOFS)) {
1892                         fuse_writepages_send(data);
1893                         req = data->req = NULL;
1894                 }
1895         }
1896
1897         err = -ENOMEM;
1898         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1899         if (!tmp_page)
1900                 goto out_unlock;
1901
1902         /*
1903          * The page must not be redirtied until the writeout is completed
1904          * (i.e. userspace has sent a reply to the write request).  Otherwise
1905          * there could be more than one temporary page instance for each real
1906          * page.
1907          *
1908          * This is ensured by holding the page lock in page_mkwrite() while
1909          * checking fuse_page_is_writeback().  We already hold the page lock
1910          * since clear_page_dirty_for_io() and keep it held until we add the
1911          * request to the fi->writepages list and increment req->num_pages.
1912          * After this fuse_page_is_writeback() will indicate that the page is
1913          * under writeback, so we can release the page lock.
1914          */
1915         if (data->req == NULL) {
1916                 struct fuse_inode *fi = get_fuse_inode(inode);
1917
1918                 err = -ENOMEM;
1919                 req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
1920                 if (!req) {
1921                         __free_page(tmp_page);
1922                         goto out_unlock;
1923                 }
1924
1925                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1926                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1927                 req->misc.write.next = NULL;
1928                 req->in.argpages = 1;
1929                 __set_bit(FR_BACKGROUND, &req->flags);
1930                 req->num_pages = 0;
1931                 req->end = fuse_writepage_end;
1932                 req->inode = inode;
1933
1934                 spin_lock(&fi->lock);
1935                 list_add(&req->writepages_entry, &fi->writepages);
1936                 spin_unlock(&fi->lock);
1937
1938                 data->req = req;
1939         }
1940         set_page_writeback(page);
1941
1942         copy_highpage(tmp_page, page);
1943         req->pages[req->num_pages] = tmp_page;
1944         req->page_descs[req->num_pages].offset = 0;
1945         req->page_descs[req->num_pages].length = PAGE_SIZE;
1946
1947         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1948         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1949
1950         err = 0;
1951         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1952                 end_page_writeback(page);
1953                 data->req = NULL;
1954                 goto out_unlock;
1955         }
1956         data->orig_pages[req->num_pages] = page;
1957
1958         /*
1959          * Protected by fi->lock against concurrent access by
1960          * fuse_page_is_writeback().
1961          */
1962         spin_lock(&fi->lock);
1963         req->num_pages++;
1964         spin_unlock(&fi->lock);
1965
1966 out_unlock:
1967         unlock_page(page);
1968
1969         return err;
1970 }
1971
1972 static int fuse_writepages(struct address_space *mapping,
1973                            struct writeback_control *wbc)
1974 {
1975         struct inode *inode = mapping->host;
1976         struct fuse_conn *fc = get_fuse_conn(inode);
1977         struct fuse_fill_wb_data data;
1978         int err;
1979
1980         err = -EIO;
1981         if (is_bad_inode(inode))
1982                 goto out;
1983
1984         data.inode = inode;
1985         data.req = NULL;
1986         data.ff = NULL;
1987
1988         err = -ENOMEM;
1989         data.orig_pages = kcalloc(fc->max_pages,
1990                                   sizeof(struct page *),
1991                                   GFP_NOFS);
1992         if (!data.orig_pages)
1993                 goto out;
1994
1995         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1996         if (data.req) {
1997                 /* Ignore errors if we can write at least one page */
1998                 BUG_ON(!data.req->num_pages);
1999                 fuse_writepages_send(&data);
2000                 err = 0;
2001         }
2002         if (data.ff)
2003                 fuse_file_put(data.ff, false, false);
2004
2005         kfree(data.orig_pages);
2006 out:
2007         return err;
2008 }
2009
2010 /*
2011  * It's worthy to make sure that space is reserved on disk for the write,
2012  * but how to implement it without killing performance need more thinking.
2013  */
2014 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2015                 loff_t pos, unsigned len, unsigned flags,
2016                 struct page **pagep, void **fsdata)
2017 {
2018         pgoff_t index = pos >> PAGE_SHIFT;
2019         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2020         struct page *page;
2021         loff_t fsize;
2022         int err = -ENOMEM;
2023
2024         WARN_ON(!fc->writeback_cache);
2025
2026         page = grab_cache_page_write_begin(mapping, index, flags);
2027         if (!page)
2028                 goto error;
2029
2030         fuse_wait_on_page_writeback(mapping->host, page->index);
2031
2032         if (PageUptodate(page) || len == PAGE_SIZE)
2033                 goto success;
2034         /*
2035          * Check if the start this page comes after the end of file, in which
2036          * case the readpage can be optimized away.
2037          */
2038         fsize = i_size_read(mapping->host);
2039         if (fsize <= (pos & PAGE_MASK)) {
2040                 size_t off = pos & ~PAGE_MASK;
2041                 if (off)
2042                         zero_user_segment(page, 0, off);
2043                 goto success;
2044         }
2045         err = fuse_do_readpage(file, page);
2046         if (err)
2047                 goto cleanup;
2048 success:
2049         *pagep = page;
2050         return 0;
2051
2052 cleanup:
2053         unlock_page(page);
2054         put_page(page);
2055 error:
2056         return err;
2057 }
2058
2059 static int fuse_write_end(struct file *file, struct address_space *mapping,
2060                 loff_t pos, unsigned len, unsigned copied,
2061                 struct page *page, void *fsdata)
2062 {
2063         struct inode *inode = page->mapping->host;
2064
2065         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2066         if (!copied)
2067                 goto unlock;
2068
2069         if (!PageUptodate(page)) {
2070                 /* Zero any unwritten bytes at the end of the page */
2071                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2072                 if (endoff)
2073                         zero_user_segment(page, endoff, PAGE_SIZE);
2074                 SetPageUptodate(page);
2075         }
2076
2077         fuse_write_update_size(inode, pos + copied);
2078         set_page_dirty(page);
2079
2080 unlock:
2081         unlock_page(page);
2082         put_page(page);
2083
2084         return copied;
2085 }
2086
2087 static int fuse_launder_page(struct page *page)
2088 {
2089         int err = 0;
2090         if (clear_page_dirty_for_io(page)) {
2091                 struct inode *inode = page->mapping->host;
2092                 err = fuse_writepage_locked(page);
2093                 if (!err)
2094                         fuse_wait_on_page_writeback(inode, page->index);
2095         }
2096         return err;
2097 }
2098
2099 /*
2100  * Write back dirty pages now, because there may not be any suitable
2101  * open files later
2102  */
2103 static void fuse_vma_close(struct vm_area_struct *vma)
2104 {
2105         filemap_write_and_wait(vma->vm_file->f_mapping);
2106 }
2107
2108 /*
2109  * Wait for writeback against this page to complete before allowing it
2110  * to be marked dirty again, and hence written back again, possibly
2111  * before the previous writepage completed.
2112  *
2113  * Block here, instead of in ->writepage(), so that the userspace fs
2114  * can only block processes actually operating on the filesystem.
2115  *
2116  * Otherwise unprivileged userspace fs would be able to block
2117  * unrelated:
2118  *
2119  * - page migration
2120  * - sync(2)
2121  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2122  */
2123 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2124 {
2125         struct page *page = vmf->page;
2126         struct inode *inode = file_inode(vmf->vma->vm_file);
2127
2128         file_update_time(vmf->vma->vm_file);
2129         lock_page(page);
2130         if (page->mapping != inode->i_mapping) {
2131                 unlock_page(page);
2132                 return VM_FAULT_NOPAGE;
2133         }
2134
2135         fuse_wait_on_page_writeback(inode, page->index);
2136         return VM_FAULT_LOCKED;
2137 }
2138
2139 static const struct vm_operations_struct fuse_file_vm_ops = {
2140         .close          = fuse_vma_close,
2141         .fault          = filemap_fault,
2142         .map_pages      = filemap_map_pages,
2143         .page_mkwrite   = fuse_page_mkwrite,
2144 };
2145
2146 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2147 {
2148         struct fuse_file *ff = file->private_data;
2149
2150         if (ff->open_flags & FOPEN_DIRECT_IO) {
2151                 /* Can't provide the coherency needed for MAP_SHARED */
2152                 if (vma->vm_flags & VM_MAYSHARE)
2153                         return -ENODEV;
2154
2155                 invalidate_inode_pages2(file->f_mapping);
2156
2157                 return generic_file_mmap(file, vma);
2158         }
2159
2160         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2161                 fuse_link_write_file(file);
2162
2163         file_accessed(file);
2164         vma->vm_ops = &fuse_file_vm_ops;
2165         return 0;
2166 }
2167
2168 static int convert_fuse_file_lock(struct fuse_conn *fc,
2169                                   const struct fuse_file_lock *ffl,
2170                                   struct file_lock *fl)
2171 {
2172         switch (ffl->type) {
2173         case F_UNLCK:
2174                 break;
2175
2176         case F_RDLCK:
2177         case F_WRLCK:
2178                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2179                     ffl->end < ffl->start)
2180                         return -EIO;
2181
2182                 fl->fl_start = ffl->start;
2183                 fl->fl_end = ffl->end;
2184
2185                 /*
2186                  * Convert pid into init's pid namespace.  The locks API will
2187                  * translate it into the caller's pid namespace.
2188                  */
2189                 rcu_read_lock();
2190                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2191                 rcu_read_unlock();
2192                 break;
2193
2194         default:
2195                 return -EIO;
2196         }
2197         fl->fl_type = ffl->type;
2198         return 0;
2199 }
2200
2201 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2202                          const struct file_lock *fl, int opcode, pid_t pid,
2203                          int flock, struct fuse_lk_in *inarg)
2204 {
2205         struct inode *inode = file_inode(file);
2206         struct fuse_conn *fc = get_fuse_conn(inode);
2207         struct fuse_file *ff = file->private_data;
2208
2209         memset(inarg, 0, sizeof(*inarg));
2210         inarg->fh = ff->fh;
2211         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2212         inarg->lk.start = fl->fl_start;
2213         inarg->lk.end = fl->fl_end;
2214         inarg->lk.type = fl->fl_type;
2215         inarg->lk.pid = pid;
2216         if (flock)
2217                 inarg->lk_flags |= FUSE_LK_FLOCK;
2218         args->in.h.opcode = opcode;
2219         args->in.h.nodeid = get_node_id(inode);
2220         args->in.numargs = 1;
2221         args->in.args[0].size = sizeof(*inarg);
2222         args->in.args[0].value = inarg;
2223 }
2224
2225 static int fuse_getlk(struct file *file, struct file_lock *fl)
2226 {
2227         struct inode *inode = file_inode(file);
2228         struct fuse_conn *fc = get_fuse_conn(inode);
2229         FUSE_ARGS(args);
2230         struct fuse_lk_in inarg;
2231         struct fuse_lk_out outarg;
2232         int err;
2233
2234         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2235         args.out.numargs = 1;
2236         args.out.args[0].size = sizeof(outarg);
2237         args.out.args[0].value = &outarg;
2238         err = fuse_simple_request(fc, &args);
2239         if (!err)
2240                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2241
2242         return err;
2243 }
2244
2245 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2246 {
2247         struct inode *inode = file_inode(file);
2248         struct fuse_conn *fc = get_fuse_conn(inode);
2249         FUSE_ARGS(args);
2250         struct fuse_lk_in inarg;
2251         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2252         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2253         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2254         int err;
2255
2256         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2257                 /* NLM needs asynchronous locks, which we don't support yet */
2258                 return -ENOLCK;
2259         }
2260
2261         /* Unlock on close is handled by the flush method */
2262         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2263                 return 0;
2264
2265         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2266         err = fuse_simple_request(fc, &args);
2267
2268         /* locking is restartable */
2269         if (err == -EINTR)
2270                 err = -ERESTARTSYS;
2271
2272         return err;
2273 }
2274
2275 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2276 {
2277         struct inode *inode = file_inode(file);
2278         struct fuse_conn *fc = get_fuse_conn(inode);
2279         int err;
2280
2281         if (cmd == F_CANCELLK) {
2282                 err = 0;
2283         } else if (cmd == F_GETLK) {
2284                 if (fc->no_lock) {
2285                         posix_test_lock(file, fl);
2286                         err = 0;
2287                 } else
2288                         err = fuse_getlk(file, fl);
2289         } else {
2290                 if (fc->no_lock)
2291                         err = posix_lock_file(file, fl, NULL);
2292                 else
2293                         err = fuse_setlk(file, fl, 0);
2294         }
2295         return err;
2296 }
2297
2298 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2299 {
2300         struct inode *inode = file_inode(file);
2301         struct fuse_conn *fc = get_fuse_conn(inode);
2302         int err;
2303
2304         if (fc->no_flock) {
2305                 err = locks_lock_file_wait(file, fl);
2306         } else {
2307                 struct fuse_file *ff = file->private_data;
2308
2309                 /* emulate flock with POSIX locks */
2310                 ff->flock = true;
2311                 err = fuse_setlk(file, fl, 1);
2312         }
2313
2314         return err;
2315 }
2316
2317 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2318 {
2319         struct inode *inode = mapping->host;
2320         struct fuse_conn *fc = get_fuse_conn(inode);
2321         FUSE_ARGS(args);
2322         struct fuse_bmap_in inarg;
2323         struct fuse_bmap_out outarg;
2324         int err;
2325
2326         if (!inode->i_sb->s_bdev || fc->no_bmap)
2327                 return 0;
2328
2329         memset(&inarg, 0, sizeof(inarg));
2330         inarg.block = block;
2331         inarg.blocksize = inode->i_sb->s_blocksize;
2332         args.in.h.opcode = FUSE_BMAP;
2333         args.in.h.nodeid = get_node_id(inode);
2334         args.in.numargs = 1;
2335         args.in.args[0].size = sizeof(inarg);
2336         args.in.args[0].value = &inarg;
2337         args.out.numargs = 1;
2338         args.out.args[0].size = sizeof(outarg);
2339         args.out.args[0].value = &outarg;
2340         err = fuse_simple_request(fc, &args);
2341         if (err == -ENOSYS)
2342                 fc->no_bmap = 1;
2343
2344         return err ? 0 : outarg.block;
2345 }
2346
2347 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2348 {
2349         struct inode *inode = file->f_mapping->host;
2350         struct fuse_conn *fc = get_fuse_conn(inode);
2351         struct fuse_file *ff = file->private_data;
2352         FUSE_ARGS(args);
2353         struct fuse_lseek_in inarg = {
2354                 .fh = ff->fh,
2355                 .offset = offset,
2356                 .whence = whence
2357         };
2358         struct fuse_lseek_out outarg;
2359         int err;
2360
2361         if (fc->no_lseek)
2362                 goto fallback;
2363
2364         args.in.h.opcode = FUSE_LSEEK;
2365         args.in.h.nodeid = ff->nodeid;
2366         args.in.numargs = 1;
2367         args.in.args[0].size = sizeof(inarg);
2368         args.in.args[0].value = &inarg;
2369         args.out.numargs = 1;
2370         args.out.args[0].size = sizeof(outarg);
2371         args.out.args[0].value = &outarg;
2372         err = fuse_simple_request(fc, &args);
2373         if (err) {
2374                 if (err == -ENOSYS) {
2375                         fc->no_lseek = 1;
2376                         goto fallback;
2377                 }
2378                 return err;
2379         }
2380
2381         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2382
2383 fallback:
2384         err = fuse_update_attributes(inode, file);
2385         if (!err)
2386                 return generic_file_llseek(file, offset, whence);
2387         else
2388                 return err;
2389 }
2390
2391 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2392 {
2393         loff_t retval;
2394         struct inode *inode = file_inode(file);
2395
2396         switch (whence) {
2397         case SEEK_SET:
2398         case SEEK_CUR:
2399                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2400                 retval = generic_file_llseek(file, offset, whence);
2401                 break;
2402         case SEEK_END:
2403                 inode_lock(inode);
2404                 retval = fuse_update_attributes(inode, file);
2405                 if (!retval)
2406                         retval = generic_file_llseek(file, offset, whence);
2407                 inode_unlock(inode);
2408                 break;
2409         case SEEK_HOLE:
2410         case SEEK_DATA:
2411                 inode_lock(inode);
2412                 retval = fuse_lseek(file, offset, whence);
2413                 inode_unlock(inode);
2414                 break;
2415         default:
2416                 retval = -EINVAL;
2417         }
2418
2419         return retval;
2420 }
2421
2422 /*
2423  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2424  * ABI was defined to be 'struct iovec' which is different on 32bit
2425  * and 64bit.  Fortunately we can determine which structure the server
2426  * used from the size of the reply.
2427  */
2428 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2429                                      size_t transferred, unsigned count,
2430                                      bool is_compat)
2431 {
2432 #ifdef CONFIG_COMPAT
2433         if (count * sizeof(struct compat_iovec) == transferred) {
2434                 struct compat_iovec *ciov = src;
2435                 unsigned i;
2436
2437                 /*
2438                  * With this interface a 32bit server cannot support
2439                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2440                  * requests
2441                  */
2442                 if (!is_compat)
2443                         return -EINVAL;
2444
2445                 for (i = 0; i < count; i++) {
2446                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2447                         dst[i].iov_len = ciov[i].iov_len;
2448                 }
2449                 return 0;
2450         }
2451 #endif
2452
2453         if (count * sizeof(struct iovec) != transferred)
2454                 return -EIO;
2455
2456         memcpy(dst, src, transferred);
2457         return 0;
2458 }
2459
2460 /* Make sure iov_length() won't overflow */
2461 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2462                                  size_t count)
2463 {
2464         size_t n;
2465         u32 max = fc->max_pages << PAGE_SHIFT;
2466
2467         for (n = 0; n < count; n++, iov++) {
2468                 if (iov->iov_len > (size_t) max)
2469                         return -ENOMEM;
2470                 max -= iov->iov_len;
2471         }
2472         return 0;
2473 }
2474
2475 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2476                                  void *src, size_t transferred, unsigned count,
2477                                  bool is_compat)
2478 {
2479         unsigned i;
2480         struct fuse_ioctl_iovec *fiov = src;
2481
2482         if (fc->minor < 16) {
2483                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2484                                                  count, is_compat);
2485         }
2486
2487         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2488                 return -EIO;
2489
2490         for (i = 0; i < count; i++) {
2491                 /* Did the server supply an inappropriate value? */
2492                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2493                     fiov[i].len != (unsigned long) fiov[i].len)
2494                         return -EIO;
2495
2496                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2497                 dst[i].iov_len = (size_t) fiov[i].len;
2498
2499 #ifdef CONFIG_COMPAT
2500                 if (is_compat &&
2501                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2502                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2503                         return -EIO;
2504 #endif
2505         }
2506
2507         return 0;
2508 }
2509
2510
2511 /*
2512  * For ioctls, there is no generic way to determine how much memory
2513  * needs to be read and/or written.  Furthermore, ioctls are allowed
2514  * to dereference the passed pointer, so the parameter requires deep
2515  * copying but FUSE has no idea whatsoever about what to copy in or
2516  * out.
2517  *
2518  * This is solved by allowing FUSE server to retry ioctl with
2519  * necessary in/out iovecs.  Let's assume the ioctl implementation
2520  * needs to read in the following structure.
2521  *
2522  * struct a {
2523  *      char    *buf;
2524  *      size_t  buflen;
2525  * }
2526  *
2527  * On the first callout to FUSE server, inarg->in_size and
2528  * inarg->out_size will be NULL; then, the server completes the ioctl
2529  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2530  * the actual iov array to
2531  *
2532  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2533  *
2534  * which tells FUSE to copy in the requested area and retry the ioctl.
2535  * On the second round, the server has access to the structure and
2536  * from that it can tell what to look for next, so on the invocation,
2537  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2538  *
2539  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2540  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2541  *
2542  * FUSE will copy both struct a and the pointed buffer from the
2543  * process doing the ioctl and retry ioctl with both struct a and the
2544  * buffer.
2545  *
2546  * This time, FUSE server has everything it needs and completes ioctl
2547  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2548  *
2549  * Copying data out works the same way.
2550  *
2551  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2552  * automatically initializes in and out iovs by decoding @cmd with
2553  * _IOC_* macros and the server is not allowed to request RETRY.  This
2554  * limits ioctl data transfers to well-formed ioctls and is the forced
2555  * behavior for all FUSE servers.
2556  */
2557 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2558                    unsigned int flags)
2559 {
2560         struct fuse_file *ff = file->private_data;
2561         struct fuse_conn *fc = ff->fc;
2562         struct fuse_ioctl_in inarg = {
2563                 .fh = ff->fh,
2564                 .cmd = cmd,
2565                 .arg = arg,
2566                 .flags = flags
2567         };
2568         struct fuse_ioctl_out outarg;
2569         struct fuse_req *req = NULL;
2570         struct page **pages = NULL;
2571         struct iovec *iov_page = NULL;
2572         struct iovec *in_iov = NULL, *out_iov = NULL;
2573         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2574         size_t in_size, out_size, transferred, c;
2575         int err, i;
2576         struct iov_iter ii;
2577
2578 #if BITS_PER_LONG == 32
2579         inarg.flags |= FUSE_IOCTL_32BIT;
2580 #else
2581         if (flags & FUSE_IOCTL_COMPAT) {
2582                 inarg.flags |= FUSE_IOCTL_32BIT;
2583 #ifdef CONFIG_X86_X32
2584                 if (in_x32_syscall())
2585                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2586 #endif
2587         }
2588 #endif
2589
2590         /* assume all the iovs returned by client always fits in a page */
2591         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2592
2593         err = -ENOMEM;
2594         pages = kcalloc(fc->max_pages, sizeof(pages[0]), GFP_KERNEL);
2595         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2596         if (!pages || !iov_page)
2597                 goto out;
2598
2599         /*
2600          * If restricted, initialize IO parameters as encoded in @cmd.
2601          * RETRY from server is not allowed.
2602          */
2603         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2604                 struct iovec *iov = iov_page;
2605
2606                 iov->iov_base = (void __user *)arg;
2607                 iov->iov_len = _IOC_SIZE(cmd);
2608
2609                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2610                         in_iov = iov;
2611                         in_iovs = 1;
2612                 }
2613
2614                 if (_IOC_DIR(cmd) & _IOC_READ) {
2615                         out_iov = iov;
2616                         out_iovs = 1;
2617                 }
2618         }
2619
2620  retry:
2621         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2622         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2623
2624         /*
2625          * Out data can be used either for actual out data or iovs,
2626          * make sure there always is at least one page.
2627          */
2628         out_size = max_t(size_t, out_size, PAGE_SIZE);
2629         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2630
2631         /* make sure there are enough buffer pages and init request with them */
2632         err = -ENOMEM;
2633         if (max_pages > fc->max_pages)
2634                 goto out;
2635         while (num_pages < max_pages) {
2636                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2637                 if (!pages[num_pages])
2638                         goto out;
2639                 num_pages++;
2640         }
2641
2642         req = fuse_get_req(fc, num_pages);
2643         if (IS_ERR(req)) {
2644                 err = PTR_ERR(req);
2645                 req = NULL;
2646                 goto out;
2647         }
2648         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2649         req->num_pages = num_pages;
2650         fuse_page_descs_length_init(req, 0, req->num_pages);
2651
2652         /* okay, let's send it to the client */
2653         req->in.h.opcode = FUSE_IOCTL;
2654         req->in.h.nodeid = ff->nodeid;
2655         req->in.numargs = 1;
2656         req->in.args[0].size = sizeof(inarg);
2657         req->in.args[0].value = &inarg;
2658         if (in_size) {
2659                 req->in.numargs++;
2660                 req->in.args[1].size = in_size;
2661                 req->in.argpages = 1;
2662
2663                 err = -EFAULT;
2664                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2665                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2666                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2667                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2668                                 goto out;
2669                 }
2670         }
2671
2672         req->out.numargs = 2;
2673         req->out.args[0].size = sizeof(outarg);
2674         req->out.args[0].value = &outarg;
2675         req->out.args[1].size = out_size;
2676         req->out.argpages = 1;
2677         req->out.argvar = 1;
2678
2679         fuse_request_send(fc, req);
2680         err = req->out.h.error;
2681         transferred = req->out.args[1].size;
2682         fuse_put_request(fc, req);
2683         req = NULL;
2684         if (err)
2685                 goto out;
2686
2687         /* did it ask for retry? */
2688         if (outarg.flags & FUSE_IOCTL_RETRY) {
2689                 void *vaddr;
2690
2691                 /* no retry if in restricted mode */
2692                 err = -EIO;
2693                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2694                         goto out;
2695
2696                 in_iovs = outarg.in_iovs;
2697                 out_iovs = outarg.out_iovs;
2698
2699                 /*
2700                  * Make sure things are in boundary, separate checks
2701                  * are to protect against overflow.
2702                  */
2703                 err = -ENOMEM;
2704                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2705                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2706                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2707                         goto out;
2708
2709                 vaddr = kmap_atomic(pages[0]);
2710                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2711                                             transferred, in_iovs + out_iovs,
2712                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2713                 kunmap_atomic(vaddr);
2714                 if (err)
2715                         goto out;
2716
2717                 in_iov = iov_page;
2718                 out_iov = in_iov + in_iovs;
2719
2720                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2721                 if (err)
2722                         goto out;
2723
2724                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2725                 if (err)
2726                         goto out;
2727
2728                 goto retry;
2729         }
2730
2731         err = -EIO;
2732         if (transferred > inarg.out_size)
2733                 goto out;
2734
2735         err = -EFAULT;
2736         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2737         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2738                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2739                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2740                         goto out;
2741         }
2742         err = 0;
2743  out:
2744         if (req)
2745                 fuse_put_request(fc, req);
2746         free_page((unsigned long) iov_page);
2747         while (num_pages)
2748                 __free_page(pages[--num_pages]);
2749         kfree(pages);
2750
2751         return err ? err : outarg.result;
2752 }
2753 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2754
2755 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2756                        unsigned long arg, unsigned int flags)
2757 {
2758         struct inode *inode = file_inode(file);
2759         struct fuse_conn *fc = get_fuse_conn(inode);
2760
2761         if (!fuse_allow_current_process(fc))
2762                 return -EACCES;
2763
2764         if (is_bad_inode(inode))
2765                 return -EIO;
2766
2767         return fuse_do_ioctl(file, cmd, arg, flags);
2768 }
2769
2770 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2771                             unsigned long arg)
2772 {
2773         return fuse_ioctl_common(file, cmd, arg, 0);
2774 }
2775
2776 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2777                                    unsigned long arg)
2778 {
2779         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2780 }
2781
2782 /*
2783  * All files which have been polled are linked to RB tree
2784  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2785  * find the matching one.
2786  */
2787 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2788                                               struct rb_node **parent_out)
2789 {
2790         struct rb_node **link = &fc->polled_files.rb_node;
2791         struct rb_node *last = NULL;
2792
2793         while (*link) {
2794                 struct fuse_file *ff;
2795
2796                 last = *link;
2797                 ff = rb_entry(last, struct fuse_file, polled_node);
2798
2799                 if (kh < ff->kh)
2800                         link = &last->rb_left;
2801                 else if (kh > ff->kh)
2802                         link = &last->rb_right;
2803                 else
2804                         return link;
2805         }
2806
2807         if (parent_out)
2808                 *parent_out = last;
2809         return link;
2810 }
2811
2812 /*
2813  * The file is about to be polled.  Make sure it's on the polled_files
2814  * RB tree.  Note that files once added to the polled_files tree are
2815  * not removed before the file is released.  This is because a file
2816  * polled once is likely to be polled again.
2817  */
2818 static void fuse_register_polled_file(struct fuse_conn *fc,
2819                                       struct fuse_file *ff)
2820 {
2821         spin_lock(&fc->lock);
2822         if (RB_EMPTY_NODE(&ff->polled_node)) {
2823                 struct rb_node **link, *uninitialized_var(parent);
2824
2825                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2826                 BUG_ON(*link);
2827                 rb_link_node(&ff->polled_node, parent, link);
2828                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2829         }
2830         spin_unlock(&fc->lock);
2831 }
2832
2833 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2834 {
2835         struct fuse_file *ff = file->private_data;
2836         struct fuse_conn *fc = ff->fc;
2837         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2838         struct fuse_poll_out outarg;
2839         FUSE_ARGS(args);
2840         int err;
2841
2842         if (fc->no_poll)
2843                 return DEFAULT_POLLMASK;
2844
2845         poll_wait(file, &ff->poll_wait, wait);
2846         inarg.events = mangle_poll(poll_requested_events(wait));
2847
2848         /*
2849          * Ask for notification iff there's someone waiting for it.
2850          * The client may ignore the flag and always notify.
2851          */
2852         if (waitqueue_active(&ff->poll_wait)) {
2853                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2854                 fuse_register_polled_file(fc, ff);
2855         }
2856
2857         args.in.h.opcode = FUSE_POLL;
2858         args.in.h.nodeid = ff->nodeid;
2859         args.in.numargs = 1;
2860         args.in.args[0].size = sizeof(inarg);
2861         args.in.args[0].value = &inarg;
2862         args.out.numargs = 1;
2863         args.out.args[0].size = sizeof(outarg);
2864         args.out.args[0].value = &outarg;
2865         err = fuse_simple_request(fc, &args);
2866
2867         if (!err)
2868                 return demangle_poll(outarg.revents);
2869         if (err == -ENOSYS) {
2870                 fc->no_poll = 1;
2871                 return DEFAULT_POLLMASK;
2872         }
2873         return EPOLLERR;
2874 }
2875 EXPORT_SYMBOL_GPL(fuse_file_poll);
2876
2877 /*
2878  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2879  * wakes up the poll waiters.
2880  */
2881 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2882                             struct fuse_notify_poll_wakeup_out *outarg)
2883 {
2884         u64 kh = outarg->kh;
2885         struct rb_node **link;
2886
2887         spin_lock(&fc->lock);
2888
2889         link = fuse_find_polled_node(fc, kh, NULL);
2890         if (*link) {
2891                 struct fuse_file *ff;
2892
2893                 ff = rb_entry(*link, struct fuse_file, polled_node);
2894                 wake_up_interruptible_sync(&ff->poll_wait);
2895         }
2896
2897         spin_unlock(&fc->lock);
2898         return 0;
2899 }
2900
2901 static void fuse_do_truncate(struct file *file)
2902 {
2903         struct inode *inode = file->f_mapping->host;
2904         struct iattr attr;
2905
2906         attr.ia_valid = ATTR_SIZE;
2907         attr.ia_size = i_size_read(inode);
2908
2909         attr.ia_file = file;
2910         attr.ia_valid |= ATTR_FILE;
2911
2912         fuse_do_setattr(file_dentry(file), &attr, file);
2913 }
2914
2915 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2916 {
2917         return round_up(off, fc->max_pages << PAGE_SHIFT);
2918 }
2919
2920 static ssize_t
2921 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2922 {
2923         DECLARE_COMPLETION_ONSTACK(wait);
2924         ssize_t ret = 0;
2925         struct file *file = iocb->ki_filp;
2926         struct fuse_file *ff = file->private_data;
2927         bool async_dio = ff->fc->async_dio;
2928         loff_t pos = 0;
2929         struct inode *inode;
2930         loff_t i_size;
2931         size_t count = iov_iter_count(iter);
2932         loff_t offset = iocb->ki_pos;
2933         struct fuse_io_priv *io;
2934
2935         pos = offset;
2936         inode = file->f_mapping->host;
2937         i_size = i_size_read(inode);
2938
2939         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2940                 return 0;
2941
2942         /* optimization for short read */
2943         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2944                 if (offset >= i_size)
2945                         return 0;
2946                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
2947                 count = iov_iter_count(iter);
2948         }
2949
2950         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2951         if (!io)
2952                 return -ENOMEM;
2953         spin_lock_init(&io->lock);
2954         kref_init(&io->refcnt);
2955         io->reqs = 1;
2956         io->bytes = -1;
2957         io->size = 0;
2958         io->offset = offset;
2959         io->write = (iov_iter_rw(iter) == WRITE);
2960         io->err = 0;
2961         /*
2962          * By default, we want to optimize all I/Os with async request
2963          * submission to the client filesystem if supported.
2964          */
2965         io->async = async_dio;
2966         io->iocb = iocb;
2967         io->blocking = is_sync_kiocb(iocb);
2968
2969         /*
2970          * We cannot asynchronously extend the size of a file.
2971          * In such case the aio will behave exactly like sync io.
2972          */
2973         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2974                 io->blocking = true;
2975
2976         if (io->async && io->blocking) {
2977                 /*
2978                  * Additional reference to keep io around after
2979                  * calling fuse_aio_complete()
2980                  */
2981                 kref_get(&io->refcnt);
2982                 io->done = &wait;
2983         }
2984
2985         if (iov_iter_rw(iter) == WRITE) {
2986                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2987                 fuse_invalidate_attr(inode);
2988         } else {
2989                 ret = __fuse_direct_read(io, iter, &pos);
2990         }
2991
2992         if (io->async) {
2993                 bool blocking = io->blocking;
2994
2995                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2996
2997                 /* we have a non-extending, async request, so return */
2998                 if (!blocking)
2999                         return -EIOCBQUEUED;
3000
3001                 wait_for_completion(&wait);
3002                 ret = fuse_get_res_by_io(io);
3003         }
3004
3005         kref_put(&io->refcnt, fuse_io_release);
3006
3007         if (iov_iter_rw(iter) == WRITE) {
3008                 if (ret > 0)
3009                         fuse_write_update_size(inode, pos);
3010                 else if (ret < 0 && offset + count > i_size)
3011                         fuse_do_truncate(file);
3012         }
3013
3014         return ret;
3015 }
3016
3017 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3018                                 loff_t length)
3019 {
3020         struct fuse_file *ff = file->private_data;
3021         struct inode *inode = file_inode(file);
3022         struct fuse_inode *fi = get_fuse_inode(inode);
3023         struct fuse_conn *fc = ff->fc;
3024         FUSE_ARGS(args);
3025         struct fuse_fallocate_in inarg = {
3026                 .fh = ff->fh,
3027                 .offset = offset,
3028                 .length = length,
3029                 .mode = mode
3030         };
3031         int err;
3032         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3033                            (mode & FALLOC_FL_PUNCH_HOLE);
3034
3035         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3036                 return -EOPNOTSUPP;
3037
3038         if (fc->no_fallocate)
3039                 return -EOPNOTSUPP;
3040
3041         if (lock_inode) {
3042                 inode_lock(inode);
3043                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3044                         loff_t endbyte = offset + length - 1;
3045                         err = filemap_write_and_wait_range(inode->i_mapping,
3046                                                            offset, endbyte);
3047                         if (err)
3048                                 goto out;
3049
3050                         fuse_sync_writes(inode);
3051                 }
3052         }
3053
3054         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3055             offset + length > i_size_read(inode)) {
3056                 err = inode_newsize_ok(inode, offset + length);
3057                 if (err)
3058                         return err;
3059         }
3060
3061         if (!(mode & FALLOC_FL_KEEP_SIZE))
3062                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3063
3064         args.in.h.opcode = FUSE_FALLOCATE;
3065         args.in.h.nodeid = ff->nodeid;
3066         args.in.numargs = 1;
3067         args.in.args[0].size = sizeof(inarg);
3068         args.in.args[0].value = &inarg;
3069         err = fuse_simple_request(fc, &args);
3070         if (err == -ENOSYS) {
3071                 fc->no_fallocate = 1;
3072                 err = -EOPNOTSUPP;
3073         }
3074         if (err)
3075                 goto out;
3076
3077         /* we could have extended the file */
3078         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3079                 bool changed = fuse_write_update_size(inode, offset + length);
3080
3081                 if (changed && fc->writeback_cache)
3082                         file_update_time(file);
3083         }
3084
3085         if (mode & FALLOC_FL_PUNCH_HOLE)
3086                 truncate_pagecache_range(inode, offset, offset + length - 1);
3087
3088         fuse_invalidate_attr(inode);
3089
3090 out:
3091         if (!(mode & FALLOC_FL_KEEP_SIZE))
3092                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3093
3094         if (lock_inode)
3095                 inode_unlock(inode);
3096
3097         return err;
3098 }
3099
3100 static ssize_t fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3101                                     struct file *file_out, loff_t pos_out,
3102                                     size_t len, unsigned int flags)
3103 {
3104         struct fuse_file *ff_in = file_in->private_data;
3105         struct fuse_file *ff_out = file_out->private_data;
3106         struct inode *inode_out = file_inode(file_out);
3107         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3108         struct fuse_conn *fc = ff_in->fc;
3109         FUSE_ARGS(args);
3110         struct fuse_copy_file_range_in inarg = {
3111                 .fh_in = ff_in->fh,
3112                 .off_in = pos_in,
3113                 .nodeid_out = ff_out->nodeid,
3114                 .fh_out = ff_out->fh,
3115                 .off_out = pos_out,
3116                 .len = len,
3117                 .flags = flags
3118         };
3119         struct fuse_write_out outarg;
3120         ssize_t err;
3121         /* mark unstable when write-back is not used, and file_out gets
3122          * extended */
3123         bool is_unstable = (!fc->writeback_cache) &&
3124                            ((pos_out + len) > inode_out->i_size);
3125
3126         if (fc->no_copy_file_range)
3127                 return -EOPNOTSUPP;
3128
3129         inode_lock(inode_out);
3130
3131         if (fc->writeback_cache) {
3132                 err = filemap_write_and_wait_range(inode_out->i_mapping,
3133                                                    pos_out, pos_out + len);
3134                 if (err)
3135                         goto out;
3136
3137                 fuse_sync_writes(inode_out);
3138         }
3139
3140         if (is_unstable)
3141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3142
3143         args.in.h.opcode = FUSE_COPY_FILE_RANGE;
3144         args.in.h.nodeid = ff_in->nodeid;
3145         args.in.numargs = 1;
3146         args.in.args[0].size = sizeof(inarg);
3147         args.in.args[0].value = &inarg;
3148         args.out.numargs = 1;
3149         args.out.args[0].size = sizeof(outarg);
3150         args.out.args[0].value = &outarg;
3151         err = fuse_simple_request(fc, &args);
3152         if (err == -ENOSYS) {
3153                 fc->no_copy_file_range = 1;
3154                 err = -EOPNOTSUPP;
3155         }
3156         if (err)
3157                 goto out;
3158
3159         if (fc->writeback_cache) {
3160                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3161                 file_update_time(file_out);
3162         }
3163
3164         fuse_invalidate_attr(inode_out);
3165
3166         err = outarg.size;
3167 out:
3168         if (is_unstable)
3169                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3170
3171         inode_unlock(inode_out);
3172
3173         return err;
3174 }
3175
3176 static const struct file_operations fuse_file_operations = {
3177         .llseek         = fuse_file_llseek,
3178         .read_iter      = fuse_file_read_iter,
3179         .write_iter     = fuse_file_write_iter,
3180         .mmap           = fuse_file_mmap,
3181         .open           = fuse_open,
3182         .flush          = fuse_flush,
3183         .release        = fuse_release,
3184         .fsync          = fuse_fsync,
3185         .lock           = fuse_file_lock,
3186         .flock          = fuse_file_flock,
3187         .splice_read    = generic_file_splice_read,
3188         .splice_write   = iter_file_splice_write,
3189         .unlocked_ioctl = fuse_file_ioctl,
3190         .compat_ioctl   = fuse_file_compat_ioctl,
3191         .poll           = fuse_file_poll,
3192         .fallocate      = fuse_file_fallocate,
3193         .copy_file_range = fuse_copy_file_range,
3194 };
3195
3196 static const struct address_space_operations fuse_file_aops  = {
3197         .readpage       = fuse_readpage,
3198         .writepage      = fuse_writepage,
3199         .writepages     = fuse_writepages,
3200         .launder_page   = fuse_launder_page,
3201         .readpages      = fuse_readpages,
3202         .set_page_dirty = __set_page_dirty_nobuffers,
3203         .bmap           = fuse_bmap,
3204         .direct_IO      = fuse_direct_IO,
3205         .write_begin    = fuse_write_begin,
3206         .write_end      = fuse_write_end,
3207 };
3208
3209 void fuse_init_file_inode(struct inode *inode)
3210 {
3211         struct fuse_inode *fi = get_fuse_inode(inode);
3212
3213         inode->i_fop = &fuse_file_operations;
3214         inode->i_data.a_ops = &fuse_file_aops;
3215
3216         INIT_LIST_HEAD(&fi->write_files);
3217         INIT_LIST_HEAD(&fi->queued_writes);
3218         fi->writectr = 0;
3219         init_waitqueue_head(&fi->page_waitq);
3220         INIT_LIST_HEAD(&fi->writepages);
3221 }