fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20 #include <linux/fs.h>
  21
  22 static const struct file_operations fuse_direct_io_file_operations;
  23
  24 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  25                           int opcode, struct fuse_open_out *outargp)
  26 {
  27         struct fuse_open_in inarg;
  28         FUSE_ARGS(args);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34         args.in.h.opcode = opcode;
  35         args.in.h.nodeid = nodeid;
  36         args.in.numargs = 1;
  37         args.in.args[0].size = sizeof(inarg);
  38         args.in.args[0].value = &inarg;
  39         args.out.numargs = 1;
  40         args.out.args[0].size = sizeof(*outargp);
  41         args.out.args[0].value = outargp;
  42
  43         return fuse_simple_request(fc, &args);
  44 }
  45
  46 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  47 {
  48         struct fuse_file *ff;
  49
  50         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  51         if (unlikely(!ff))
  52                 return NULL;
  53
  54         ff->fc = fc;
  55         ff->reserved_req = fuse_request_alloc(0);
  56         if (unlikely(!ff->reserved_req)) {
  57                 kfree(ff);
  58                 return NULL;
  59         }
  60
  61         INIT_LIST_HEAD(&ff->write_entry);
  62         atomic_set(&ff->count, 0);
  63         RB_CLEAR_NODE(&ff->polled_node);
  64         init_waitqueue_head(&ff->poll_wait);
  65
  66         spin_lock(&fc->lock);
  67         ff->kh = ++fc->khctr;
  68         spin_unlock(&fc->lock);
  69
  70         return ff;
  71 }
  72
  73 void fuse_file_free(struct fuse_file *ff)
  74 {
  75         fuse_request_free(ff->reserved_req);
  76         kfree(ff);
  77 }
  78
  79 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  80 {
  81         atomic_inc(&ff->count);
  82         return ff;
  83 }
  84
  85 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  86 {
  87         iput(req->misc.release.inode);
  88 }
  89
  90 static void fuse_file_put(struct fuse_file *ff, bool sync)
  91 {
  92         if (atomic_dec_and_test(&ff->count)) {
  93                 struct fuse_req *req = ff->reserved_req;
  94
  95                 if (ff->fc->no_open) {
  96                         /*
  97                          * Drop the release request when client does not
  98                          * implement 'open'
  99                          */
 100                         __clear_bit(FR_BACKGROUND, &req->flags);
 101                         iput(req->misc.release.inode);
 102                         fuse_put_request(ff->fc, req);
 103                 } else if (sync) {
 104                         __set_bit(FR_FORCE, &req->flags);
 105                         __clear_bit(FR_BACKGROUND, &req->flags);
 106                         fuse_request_send(ff->fc, req);
 107                         iput(req->misc.release.inode);
 108                         fuse_put_request(ff->fc, req);
 109                 } else {
 110                         req->end = fuse_release_end;
 111                         __set_bit(FR_BACKGROUND, &req->flags);
 112                         fuse_request_send_background(ff->fc, req);
 113                 }
 114                 kfree(ff);
 115         }
 116 }
 117
 118 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 119                  bool isdir)
 120 {
 121         struct fuse_file *ff;
 122         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 123
 124         ff = fuse_file_alloc(fc);
 125         if (!ff)
 126                 return -ENOMEM;
 127
 128         ff->fh = 0;
 129         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 130         if (!fc->no_open || isdir) {
 131                 struct fuse_open_out outarg;
 132                 int err;
 133
 134                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 135                 if (!err) {
 136                         ff->fh = outarg.fh;
 137                         ff->open_flags = outarg.open_flags;
 138
 139                 } else if (err != -ENOSYS || isdir) {
 140                         fuse_file_free(ff);
 141                         return err;
 142                 } else {
 143                         fc->no_open = 1;
 144                 }
 145         }
 146
 147         if (isdir)
 148                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 149
 150         ff->nodeid = nodeid;
 151         file->private_data = fuse_file_get(ff);
 152
 153         return 0;
 154 }
 155 EXPORT_SYMBOL_GPL(fuse_do_open);
 156
 157 static void fuse_link_write_file(struct file *file)
 158 {
 159         struct inode *inode = file_inode(file);
 160         struct fuse_conn *fc = get_fuse_conn(inode);
 161         struct fuse_inode *fi = get_fuse_inode(inode);
 162         struct fuse_file *ff = file->private_data;
 163         /*
 164          * file may be written through mmap, so chain it onto the
 165          * inodes's write_file list
 166          */
 167         spin_lock(&fc->lock);
 168         if (list_empty(&ff->write_entry))
 169                 list_add(&ff->write_entry, &fi->write_files);
 170         spin_unlock(&fc->lock);
 171 }
 172
 173 void fuse_finish_open(struct inode *inode, struct file *file)
 174 {
 175         struct fuse_file *ff = file->private_data;
 176         struct fuse_conn *fc = get_fuse_conn(inode);
 177
 178         if (ff->open_flags & FOPEN_DIRECT_IO)
 179                 file->f_op = &fuse_direct_io_file_operations;
 180         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 181                 invalidate_inode_pages2(inode->i_mapping);
 182         if (ff->open_flags & FOPEN_STREAM)
 183                 stream_open(inode, file);
 184         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 185                 nonseekable_open(inode, file);
 186         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 187                 struct fuse_inode *fi = get_fuse_inode(inode);
 188
 189                 spin_lock(&fc->lock);
 190                 fi->attr_version = ++fc->attr_version;
 191                 i_size_write(inode, 0);
 192                 spin_unlock(&fc->lock);
 193                 fuse_invalidate_attr(inode);
 194                 if (fc->writeback_cache)
 195                         file_update_time(file);
 196         }
 197         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 198                 fuse_link_write_file(file);
 199 }
 200
 201 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 202 {
 203         struct fuse_conn *fc = get_fuse_conn(inode);
 204         int err;
 205         bool is_wb_truncate = (file->f_flags & O_TRUNC) &&
 206                           fc->atomic_o_trunc &&
 207                           fc->writeback_cache;
 208
 209         err = generic_file_open(inode, file);
 210         if (err)
 211                 return err;
 212
 213         if (is_wb_truncate) {
 214                 mutex_lock(&inode->i_mutex);
 215                 fuse_set_nowrite(inode);
 216         }
 217
 218         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 219
 220         if (!err)
 221                 fuse_finish_open(inode, file);
 222
 223         if (is_wb_truncate) {
 224                 fuse_release_nowrite(inode);
 225                 mutex_unlock(&inode->i_mutex);
 226         }
 227
 228         return err;
 229 }
 230
 231 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 232 {
 233         struct fuse_conn *fc = ff->fc;
 234         struct fuse_req *req = ff->reserved_req;
 235         struct fuse_release_in *inarg = &req->misc.release.in;
 236
 237         spin_lock(&fc->lock);
 238         list_del(&ff->write_entry);
 239         if (!RB_EMPTY_NODE(&ff->polled_node))
 240                 rb_erase(&ff->polled_node, &fc->polled_files);
 241         spin_unlock(&fc->lock);
 242
 243         wake_up_interruptible_all(&ff->poll_wait);
 244
 245         inarg->fh = ff->fh;
 246         inarg->flags = flags;
 247         req->in.h.opcode = opcode;
 248         req->in.h.nodeid = ff->nodeid;
 249         req->in.numargs = 1;
 250         req->in.args[0].size = sizeof(struct fuse_release_in);
 251         req->in.args[0].value = inarg;
 252 }
 253
 254 void fuse_release_common(struct file *file, int opcode)
 255 {
 256         struct fuse_file *ff;
 257         struct fuse_req *req;
 258
 259         ff = file->private_data;
 260         if (unlikely(!ff))
 261                 return;
 262
 263         req = ff->reserved_req;
 264         fuse_prepare_release(ff, file->f_flags, opcode);
 265
 266         if (ff->flock) {
 267                 struct fuse_release_in *inarg = &req->misc.release.in;
 268                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 269                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 270                                                        (fl_owner_t) file);
 271         }
 272         /* Hold inode until release is finished */
 273         req->misc.release.inode = igrab(file_inode(file));
 274
 275         /*
 276          * Normally this will send the RELEASE request, however if
 277          * some asynchronous READ or WRITE requests are outstanding,
 278          * the sending will be delayed.
 279          *
 280          * Make the release synchronous if this is a fuseblk mount,
 281          * synchronous RELEASE is allowed (and desirable) in this case
 282          * because the server can be trusted not to screw up.
 283          */
 284         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 285 }
 286
 287 static int fuse_open(struct inode *inode, struct file *file)
 288 {
 289         return fuse_open_common(inode, file, false);
 290 }
 291
 292 static int fuse_release(struct inode *inode, struct file *file)
 293 {
 294         struct fuse_conn *fc = get_fuse_conn(inode);
 295
 296         /* see fuse_vma_close() for !writeback_cache case */
 297         if (fc->writeback_cache)
 298                 write_inode_now(inode, 1);
 299
 300         fuse_release_common(file, FUSE_RELEASE);
 301
 302         /* return value is ignored by VFS */
 303         return 0;
 304 }
 305
 306 void fuse_sync_release(struct fuse_file *ff, int flags)
 307 {
 308         WARN_ON(atomic_read(&ff->count) > 1);
 309         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 310         __set_bit(FR_FORCE, &ff->reserved_req->flags);
 311         __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
 312         fuse_request_send(ff->fc, ff->reserved_req);
 313         fuse_put_request(ff->fc, ff->reserved_req);
 314         kfree(ff);
 315 }
 316 EXPORT_SYMBOL_GPL(fuse_sync_release);
 317
 318 /*
 319  * Scramble the ID space with XTEA, so that the value of the files_struct
 320  * pointer is not exposed to userspace.
 321  */
 322 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 323 {
 324         u32 *k = fc->scramble_key;
 325         u64 v = (unsigned long) id;
 326         u32 v0 = v;
 327         u32 v1 = v >> 32;
 328         u32 sum = 0;
 329         int i;
 330
 331         for (i = 0; i < 32; i++) {
 332                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 333                 sum += 0x9E3779B9;
 334                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 335         }
 336
 337         return (u64) v0 + ((u64) v1 << 32);
 338 }
 339
 340 /*
 341  * Check if any page in a range is under writeback
 342  *
 343  * This is currently done by walking the list of writepage requests
 344  * for the inode, which can be pretty inefficient.
 345  */
 346 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 347                                    pgoff_t idx_to)
 348 {
 349         struct fuse_conn *fc = get_fuse_conn(inode);
 350         struct fuse_inode *fi = get_fuse_inode(inode);
 351         struct fuse_req *req;
 352         bool found = false;
 353
 354         spin_lock(&fc->lock);
 355         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 356                 pgoff_t curr_index;
 357
 358                 BUG_ON(req->inode != inode);
 359                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 360                 if (idx_from < curr_index + req->num_pages &&
 361                     curr_index <= idx_to) {
 362                         found = true;
 363                         break;
 364                 }
 365         }
 366         spin_unlock(&fc->lock);
 367
 368         return found;
 369 }
 370
 371 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 372 {
 373         return fuse_range_is_writeback(inode, index, index);
 374 }
 375
 376 /*
 377  * Wait for page writeback to be completed.
 378  *
 379  * Since fuse doesn't rely on the VM writeback tracking, this has to
 380  * use some other means.
 381  */
 382 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 383 {
 384         struct fuse_inode *fi = get_fuse_inode(inode);
 385
 386         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 387         return 0;
 388 }
 389
 390 /*
 391  * Wait for all pending writepages on the inode to finish.
 392  *
 393  * This is currently done by blocking further writes with FUSE_NOWRITE
 394  * and waiting for all sent writes to complete.
 395  *
 396  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 397  * could conflict with truncation.
 398  */
 399 static void fuse_sync_writes(struct inode *inode)
 400 {
 401         fuse_set_nowrite(inode);
 402         fuse_release_nowrite(inode);
 403 }
 404
 405 static int fuse_flush(struct file *file, fl_owner_t id)
 406 {
 407         struct inode *inode = file_inode(file);
 408         struct fuse_conn *fc = get_fuse_conn(inode);
 409         struct fuse_file *ff = file->private_data;
 410         struct fuse_req *req;
 411         struct fuse_flush_in inarg;
 412         int err;
 413
 414         if (is_bad_inode(inode))
 415                 return -EIO;
 416
 417         if (fc->no_flush)
 418                 return 0;
 419
 420         err = write_inode_now(inode, 1);
 421         if (err)
 422                 return err;
 423
 424         mutex_lock(&inode->i_mutex);
 425         fuse_sync_writes(inode);
 426         mutex_unlock(&inode->i_mutex);
 427
 428         if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
 429             test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
 430                 err = -ENOSPC;
 431         if (test_bit(AS_EIO, &file->f_mapping->flags) &&
 432             test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
 433                 err = -EIO;
 434         if (err)
 435                 return err;
 436
 437         req = fuse_get_req_nofail_nopages(fc, file);
 438         memset(&inarg, 0, sizeof(inarg));
 439         inarg.fh = ff->fh;
 440         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 441         req->in.h.opcode = FUSE_FLUSH;
 442         req->in.h.nodeid = get_node_id(inode);
 443         req->in.numargs = 1;
 444         req->in.args[0].size = sizeof(inarg);
 445         req->in.args[0].value = &inarg;
 446         __set_bit(FR_FORCE, &req->flags);
 447         fuse_request_send(fc, req);
 448         err = req->out.h.error;
 449         fuse_put_request(fc, req);
 450         if (err == -ENOSYS) {
 451                 fc->no_flush = 1;
 452                 err = 0;
 453         }
 454         return err;
 455 }
 456
 457 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 458                       int datasync, int isdir)
 459 {
 460         struct inode *inode = file->f_mapping->host;
 461         struct fuse_conn *fc = get_fuse_conn(inode);
 462         struct fuse_file *ff = file->private_data;
 463         FUSE_ARGS(args);
 464         struct fuse_fsync_in inarg;
 465         int err;
 466
 467         if (is_bad_inode(inode))
 468                 return -EIO;
 469
 470         mutex_lock(&inode->i_mutex);
 471
 472         /*
 473          * Start writeback against all dirty pages of the inode, then
 474          * wait for all outstanding writes, before sending the FSYNC
 475          * request.
 476          */
 477         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 478         if (err)
 479                 goto out;
 480
 481         fuse_sync_writes(inode);
 482
 483         /*
 484          * Due to implementation of fuse writeback
 485          * filemap_write_and_wait_range() does not catch errors.
 486          * We have to do this directly after fuse_sync_writes()
 487          */
 488         if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
 489             test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
 490                 err = -ENOSPC;
 491         if (test_bit(AS_EIO, &file->f_mapping->flags) &&
 492             test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
 493                 err = -EIO;
 494         if (err)
 495                 goto out;
 496
 497         err = sync_inode_metadata(inode, 1);
 498         if (err)
 499                 goto out;
 500
 501         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 502                 goto out;
 503
 504         memset(&inarg, 0, sizeof(inarg));
 505         inarg.fh = ff->fh;
 506         inarg.fsync_flags = datasync ? 1 : 0;
 507         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 508         args.in.h.nodeid = get_node_id(inode);
 509         args.in.numargs = 1;
 510         args.in.args[0].size = sizeof(inarg);
 511         args.in.args[0].value = &inarg;
 512         err = fuse_simple_request(fc, &args);
 513         if (err == -ENOSYS) {
 514                 if (isdir)
 515                         fc->no_fsyncdir = 1;
 516                 else
 517                         fc->no_fsync = 1;
 518                 err = 0;
 519         }
 520 out:
 521         mutex_unlock(&inode->i_mutex);
 522         return err;
 523 }
 524
 525 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 526                       int datasync)
 527 {
 528         return fuse_fsync_common(file, start, end, datasync, 0);
 529 }
 530
 531 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 532                     size_t count, int opcode)
 533 {
 534         struct fuse_read_in *inarg = &req->misc.read.in;
 535         struct fuse_file *ff = file->private_data;
 536
 537         inarg->fh = ff->fh;
 538         inarg->offset = pos;
 539         inarg->size = count;
 540         inarg->flags = file->f_flags;
 541         req->in.h.opcode = opcode;
 542         req->in.h.nodeid = ff->nodeid;
 543         req->in.numargs = 1;
 544         req->in.args[0].size = sizeof(struct fuse_read_in);
 545         req->in.args[0].value = inarg;
 546         req->out.argvar = 1;
 547         req->out.numargs = 1;
 548         req->out.args[0].size = count;
 549 }
 550
 551 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 552 {
 553         unsigned i;
 554
 555         for (i = 0; i < req->num_pages; i++) {
 556                 struct page *page = req->pages[i];
 557                 if (should_dirty)
 558                         set_page_dirty_lock(page);
 559                 put_page(page);
 560         }
 561 }
 562
 563 static void fuse_io_release(struct kref *kref)
 564 {
 565         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 566 }
 567
 568 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 569 {
 570         if (io->err)
 571                 return io->err;
 572
 573         if (io->bytes >= 0 && io->write)
 574                 return -EIO;
 575
 576         return io->bytes < 0 ? io->size : io->bytes;
 577 }
 578
 579 /**
 580  * In case of short read, the caller sets 'pos' to the position of
 581  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 582  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 583  *
 584  * An example:
 585  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 586  * both submitted asynchronously. The first of them was ACKed by userspace as
 587  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 588  * second request was ACKed as short, e.g. only 1K was read, resulting in
 589  * pos == 33K.
 590  *
 591  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 592  * will be equal to the length of the longest contiguous fragment of
 593  * transferred data starting from the beginning of IO request.
 594  */
 595 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 596 {
 597         bool is_sync = is_sync_kiocb(io->iocb);
 598         int left;
 599
 600         spin_lock(&io->lock);
 601         if (err)
 602                 io->err = io->err ? : err;
 603         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 604                 io->bytes = pos;
 605
 606         left = --io->reqs;
 607         if (!left && is_sync)
 608                 complete(io->done);
 609         spin_unlock(&io->lock);
 610
 611         if (!left && !is_sync) {
 612                 ssize_t res = fuse_get_res_by_io(io);
 613
 614                 if (res >= 0) {
 615                         struct inode *inode = file_inode(io->iocb->ki_filp);
 616                         struct fuse_conn *fc = get_fuse_conn(inode);
 617                         struct fuse_inode *fi = get_fuse_inode(inode);
 618
 619                         spin_lock(&fc->lock);
 620                         fi->attr_version = ++fc->attr_version;
 621                         spin_unlock(&fc->lock);
 622                 }
 623
 624                 io->iocb->ki_complete(io->iocb, res, 0);
 625         }
 626
 627         kref_put(&io->refcnt, fuse_io_release);
 628 }
 629
 630 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 631 {
 632         struct fuse_io_priv *io = req->io;
 633         ssize_t pos = -1;
 634
 635         fuse_release_user_pages(req, io->should_dirty);
 636
 637         if (io->write) {
 638                 if (req->misc.write.in.size != req->misc.write.out.size)
 639                         pos = req->misc.write.in.offset - io->offset +
 640                                 req->misc.write.out.size;
 641         } else {
 642                 if (req->misc.read.in.size != req->out.args[0].size)
 643                         pos = req->misc.read.in.offset - io->offset +
 644                                 req->out.args[0].size;
 645         }
 646
 647         fuse_aio_complete(io, req->out.h.error, pos);
 648 }
 649
 650 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 651                 size_t num_bytes, struct fuse_io_priv *io)
 652 {
 653         spin_lock(&io->lock);
 654         kref_get(&io->refcnt);
 655         io->size += num_bytes;
 656         io->reqs++;
 657         spin_unlock(&io->lock);
 658
 659         req->io = io;
 660         req->end = fuse_aio_complete_req;
 661
 662         __fuse_get_request(req);
 663         fuse_request_send_background(fc, req);
 664
 665         return num_bytes;
 666 }
 667
 668 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 669                              loff_t pos, size_t count, fl_owner_t owner)
 670 {
 671         struct file *file = io->file;
 672         struct fuse_file *ff = file->private_data;
 673         struct fuse_conn *fc = ff->fc;
 674
 675         fuse_read_fill(req, file, pos, count, FUSE_READ);
 676         if (owner != NULL) {
 677                 struct fuse_read_in *inarg = &req->misc.read.in;
 678
 679                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 680                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 681         }
 682
 683         if (io->async)
 684                 return fuse_async_req_send(fc, req, count, io);
 685
 686         fuse_request_send(fc, req);
 687         return req->out.args[0].size;
 688 }
 689
 690 static void fuse_read_update_size(struct inode *inode, loff_t size,
 691                                   u64 attr_ver)
 692 {
 693         struct fuse_conn *fc = get_fuse_conn(inode);
 694         struct fuse_inode *fi = get_fuse_inode(inode);
 695
 696         spin_lock(&fc->lock);
 697         if (attr_ver == fi->attr_version && size < inode->i_size &&
 698             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 699                 fi->attr_version = ++fc->attr_version;
 700                 i_size_write(inode, size);
 701         }
 702         spin_unlock(&fc->lock);
 703 }
 704
 705 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 706                             u64 attr_ver)
 707 {
 708         size_t num_read = req->out.args[0].size;
 709         struct fuse_conn *fc = get_fuse_conn(inode);
 710
 711         if (fc->writeback_cache) {
 712                 /*
 713                  * A hole in a file. Some data after the hole are in page cache,
 714                  * but have not reached the client fs yet. So, the hole is not
 715                  * present there.
 716                  */
 717                 int i;
 718                 int start_idx = num_read >> PAGE_CACHE_SHIFT;
 719                 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
 720
 721                 for (i = start_idx; i < req->num_pages; i++) {
 722                         zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
 723                         off = 0;
 724                 }
 725         } else {
 726                 loff_t pos = page_offset(req->pages[0]) + num_read;
 727                 fuse_read_update_size(inode, pos, attr_ver);
 728         }
 729 }
 730
 731 static int fuse_do_readpage(struct file *file, struct page *page)
 732 {
 733         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
 734         struct inode *inode = page->mapping->host;
 735         struct fuse_conn *fc = get_fuse_conn(inode);
 736         struct fuse_req *req;
 737         size_t num_read;
 738         loff_t pos = page_offset(page);
 739         size_t count = PAGE_CACHE_SIZE;
 740         u64 attr_ver;
 741         int err;
 742
 743         /*
 744          * Page writeback can extend beyond the lifetime of the
 745          * page-cache page, so make sure we read a properly synced
 746          * page.
 747          */
 748         fuse_wait_on_page_writeback(inode, page->index);
 749
 750         req = fuse_get_req(fc, 1);
 751         if (IS_ERR(req))
 752                 return PTR_ERR(req);
 753
 754         attr_ver = fuse_get_attr_version(fc);
 755
 756         req->out.page_zeroing = 1;
 757         req->out.argpages = 1;
 758         req->num_pages = 1;
 759         req->pages[0] = page;
 760         req->page_descs[0].length = count;
 761         num_read = fuse_send_read(req, &io, pos, count, NULL);
 762         err = req->out.h.error;
 763
 764         if (!err) {
 765                 /*
 766                  * Short read means EOF.  If file size is larger, truncate it
 767                  */
 768                 if (num_read < count)
 769                         fuse_short_read(req, inode, attr_ver);
 770
 771                 SetPageUptodate(page);
 772         }
 773
 774         fuse_put_request(fc, req);
 775
 776         return err;
 777 }
 778
 779 static int fuse_readpage(struct file *file, struct page *page)
 780 {
 781         struct inode *inode = page->mapping->host;
 782         int err;
 783
 784         err = -EIO;
 785         if (is_bad_inode(inode))
 786                 goto out;
 787
 788         err = fuse_do_readpage(file, page);
 789         fuse_invalidate_atime(inode);
 790  out:
 791         unlock_page(page);
 792         return err;
 793 }
 794
 795 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 796 {
 797         int i;
 798         size_t count = req->misc.read.in.size;
 799         size_t num_read = req->out.args[0].size;
 800         struct address_space *mapping = NULL;
 801
 802         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 803                 mapping = req->pages[i]->mapping;
 804
 805         if (mapping) {
 806                 struct inode *inode = mapping->host;
 807
 808                 /*
 809                  * Short read means EOF. If file size is larger, truncate it
 810                  */
 811                 if (!req->out.h.error && num_read < count)
 812                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 813
 814                 fuse_invalidate_atime(inode);
 815         }
 816
 817         for (i = 0; i < req->num_pages; i++) {
 818                 struct page *page = req->pages[i];
 819                 if (!req->out.h.error)
 820                         SetPageUptodate(page);
 821                 else
 822                         SetPageError(page);
 823                 unlock_page(page);
 824                 page_cache_release(page);
 825         }
 826         if (req->ff)
 827                 fuse_file_put(req->ff, false);
 828 }
 829
 830 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 831 {
 832         struct fuse_file *ff = file->private_data;
 833         struct fuse_conn *fc = ff->fc;
 834         loff_t pos = page_offset(req->pages[0]);
 835         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 836
 837         req->out.argpages = 1;
 838         req->out.page_zeroing = 1;
 839         req->out.page_replace = 1;
 840         fuse_read_fill(req, file, pos, count, FUSE_READ);
 841         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 842         if (fc->async_read) {
 843                 req->ff = fuse_file_get(ff);
 844                 req->end = fuse_readpages_end;
 845                 fuse_request_send_background(fc, req);
 846         } else {
 847                 fuse_request_send(fc, req);
 848                 fuse_readpages_end(fc, req);
 849                 fuse_put_request(fc, req);
 850         }
 851 }
 852
 853 struct fuse_fill_data {
 854         struct fuse_req *req;
 855         struct file *file;
 856         struct inode *inode;
 857         unsigned nr_pages;
 858 };
 859
 860 static int fuse_readpages_fill(void *_data, struct page *page)
 861 {
 862         struct fuse_fill_data *data = _data;
 863         struct fuse_req *req = data->req;
 864         struct inode *inode = data->inode;
 865         struct fuse_conn *fc = get_fuse_conn(inode);
 866
 867         fuse_wait_on_page_writeback(inode, page->index);
 868
 869         if (req->num_pages &&
 870             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 871              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 872              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 873                 int nr_alloc = min_t(unsigned, data->nr_pages,
 874                                      FUSE_MAX_PAGES_PER_REQ);
 875                 fuse_send_readpages(req, data->file);
 876                 if (fc->async_read)
 877                         req = fuse_get_req_for_background(fc, nr_alloc);
 878                 else
 879                         req = fuse_get_req(fc, nr_alloc);
 880
 881                 data->req = req;
 882                 if (IS_ERR(req)) {
 883                         unlock_page(page);
 884                         return PTR_ERR(req);
 885                 }
 886         }
 887
 888         if (WARN_ON(req->num_pages >= req->max_pages)) {
 889                 unlock_page(page);
 890                 fuse_put_request(fc, req);
 891                 return -EIO;
 892         }
 893
 894         page_cache_get(page);
 895         req->pages[req->num_pages] = page;
 896         req->page_descs[req->num_pages].length = PAGE_SIZE;
 897         req->num_pages++;
 898         data->nr_pages--;
 899         return 0;
 900 }
 901
 902 static int fuse_readpages(struct file *file, struct address_space *mapping,
 903                           struct list_head *pages, unsigned nr_pages)
 904 {
 905         struct inode *inode = mapping->host;
 906         struct fuse_conn *fc = get_fuse_conn(inode);
 907         struct fuse_fill_data data;
 908         int err;
 909         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 910
 911         err = -EIO;
 912         if (is_bad_inode(inode))
 913                 goto out;
 914
 915         data.file = file;
 916         data.inode = inode;
 917         if (fc->async_read)
 918                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 919         else
 920                 data.req = fuse_get_req(fc, nr_alloc);
 921         data.nr_pages = nr_pages;
 922         err = PTR_ERR(data.req);
 923         if (IS_ERR(data.req))
 924                 goto out;
 925
 926         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 927         if (!err) {
 928                 if (data.req->num_pages)
 929                         fuse_send_readpages(data.req, file);
 930                 else
 931                         fuse_put_request(fc, data.req);
 932         }
 933 out:
 934         return err;
 935 }
 936
 937 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 938 {
 939         struct inode *inode = iocb->ki_filp->f_mapping->host;
 940         struct fuse_conn *fc = get_fuse_conn(inode);
 941
 942         /*
 943          * In auto invalidate mode, always update attributes on read.
 944          * Otherwise, only update if we attempt to read past EOF (to ensure
 945          * i_size is up to date).
 946          */
 947         if (fc->auto_inval_data ||
 948             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 949                 int err;
 950                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 951                 if (err)
 952                         return err;
 953         }
 954
 955         return generic_file_read_iter(iocb, to);
 956 }
 957
 958 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 959                             loff_t pos, size_t count)
 960 {
 961         struct fuse_write_in *inarg = &req->misc.write.in;
 962         struct fuse_write_out *outarg = &req->misc.write.out;
 963
 964         inarg->fh = ff->fh;
 965         inarg->offset = pos;
 966         inarg->size = count;
 967         req->in.h.opcode = FUSE_WRITE;
 968         req->in.h.nodeid = ff->nodeid;
 969         req->in.numargs = 2;
 970         if (ff->fc->minor < 9)
 971                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 972         else
 973                 req->in.args[0].size = sizeof(struct fuse_write_in);
 974         req->in.args[0].value = inarg;
 975         req->in.args[1].size = count;
 976         req->out.numargs = 1;
 977         req->out.args[0].size = sizeof(struct fuse_write_out);
 978         req->out.args[0].value = outarg;
 979 }
 980
 981 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 982                               loff_t pos, size_t count, fl_owner_t owner)
 983 {
 984         struct file *file = io->file;
 985         struct fuse_file *ff = file->private_data;
 986         struct fuse_conn *fc = ff->fc;
 987         struct fuse_write_in *inarg = &req->misc.write.in;
 988
 989         fuse_write_fill(req, ff, pos, count);
 990         inarg->flags = file->f_flags;
 991         if (owner != NULL) {
 992                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 993                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 994         }
 995
 996         if (io->async)
 997                 return fuse_async_req_send(fc, req, count, io);
 998
 999         fuse_request_send(fc, req);
1000         return req->misc.write.out.size;
1001 }
1002
1003 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1004 {
1005         struct fuse_conn *fc = get_fuse_conn(inode);
1006         struct fuse_inode *fi = get_fuse_inode(inode);
1007         bool ret = false;
1008
1009         spin_lock(&fc->lock);
1010         fi->attr_version = ++fc->attr_version;
1011         if (pos > inode->i_size) {
1012                 i_size_write(inode, pos);
1013                 ret = true;
1014         }
1015         spin_unlock(&fc->lock);
1016
1017         return ret;
1018 }
1019
1020 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1021                                     struct inode *inode, loff_t pos,
1022                                     size_t count)
1023 {
1024         size_t res;
1025         unsigned offset;
1026         unsigned i;
1027         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1028
1029         for (i = 0; i < req->num_pages; i++)
1030                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1031
1032         res = fuse_send_write(req, &io, pos, count, NULL);
1033
1034         offset = req->page_descs[0].offset;
1035         count = res;
1036         for (i = 0; i < req->num_pages; i++) {
1037                 struct page *page = req->pages[i];
1038
1039                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1040                         SetPageUptodate(page);
1041
1042                 if (count > PAGE_CACHE_SIZE - offset)
1043                         count -= PAGE_CACHE_SIZE - offset;
1044                 else
1045                         count = 0;
1046                 offset = 0;
1047
1048                 unlock_page(page);
1049                 page_cache_release(page);
1050         }
1051
1052         return res;
1053 }
1054
1055 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1056                                struct address_space *mapping,
1057                                struct iov_iter *ii, loff_t pos)
1058 {
1059         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1060         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1061         size_t count = 0;
1062         int err;
1063
1064         req->in.argpages = 1;
1065         req->page_descs[0].offset = offset;
1066
1067         do {
1068                 size_t tmp;
1069                 struct page *page;
1070                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1071                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1072                                      iov_iter_count(ii));
1073
1074                 bytes = min_t(size_t, bytes, fc->max_write - count);
1075
1076  again:
1077                 err = -EFAULT;
1078                 if (iov_iter_fault_in_readable(ii, bytes))
1079                         break;
1080
1081                 err = -ENOMEM;
1082                 page = grab_cache_page_write_begin(mapping, index, 0);
1083                 if (!page)
1084                         break;
1085
1086                 if (mapping_writably_mapped(mapping))
1087                         flush_dcache_page(page);
1088
1089                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1090                 flush_dcache_page(page);
1091
1092                 iov_iter_advance(ii, tmp);
1093                 if (!tmp) {
1094                         unlock_page(page);
1095                         page_cache_release(page);
1096                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1097                         goto again;
1098                 }
1099
1100                 err = 0;
1101                 req->pages[req->num_pages] = page;
1102                 req->page_descs[req->num_pages].length = tmp;
1103                 req->num_pages++;
1104
1105                 count += tmp;
1106                 pos += tmp;
1107                 offset += tmp;
1108                 if (offset == PAGE_CACHE_SIZE)
1109                         offset = 0;
1110
1111                 if (!fc->big_writes)
1112                         break;
1113         } while (iov_iter_count(ii) && count < fc->max_write &&
1114                  req->num_pages < req->max_pages && offset == 0);
1115
1116         return count > 0 ? count : err;
1117 }
1118
1119 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1120 {
1121         return min_t(unsigned,
1122                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1123                      (pos >> PAGE_CACHE_SHIFT) + 1,
1124                      FUSE_MAX_PAGES_PER_REQ);
1125 }
1126
1127 static ssize_t fuse_perform_write(struct file *file,
1128                                   struct address_space *mapping,
1129                                   struct iov_iter *ii, loff_t pos)
1130 {
1131         struct inode *inode = mapping->host;
1132         struct fuse_conn *fc = get_fuse_conn(inode);
1133         struct fuse_inode *fi = get_fuse_inode(inode);
1134         int err = 0;
1135         ssize_t res = 0;
1136
1137         if (is_bad_inode(inode))
1138                 return -EIO;
1139
1140         if (inode->i_size < pos + iov_iter_count(ii))
1141                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1142
1143         do {
1144                 struct fuse_req *req;
1145                 ssize_t count;
1146                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1147
1148                 req = fuse_get_req(fc, nr_pages);
1149                 if (IS_ERR(req)) {
1150                         err = PTR_ERR(req);
1151                         break;
1152                 }
1153
1154                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1155                 if (count <= 0) {
1156                         err = count;
1157                 } else {
1158                         size_t num_written;
1159
1160                         num_written = fuse_send_write_pages(req, file, inode,
1161                                                             pos, count);
1162                         err = req->out.h.error;
1163                         if (!err) {
1164                                 res += num_written;
1165                                 pos += num_written;
1166
1167                                 /* break out of the loop on short write */
1168                                 if (num_written != count)
1169                                         err = -EIO;
1170                         }
1171                 }
1172                 fuse_put_request(fc, req);
1173         } while (!err && iov_iter_count(ii));
1174
1175         if (res > 0)
1176                 fuse_write_update_size(inode, pos);
1177
1178         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1179         fuse_invalidate_attr(inode);
1180
1181         return res > 0 ? res : err;
1182 }
1183
1184 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1185 {
1186         struct file *file = iocb->ki_filp;
1187         struct address_space *mapping = file->f_mapping;
1188         ssize_t written = 0;
1189         ssize_t written_buffered = 0;
1190         struct inode *inode = mapping->host;
1191         ssize_t err;
1192         loff_t endbyte = 0;
1193
1194         if (get_fuse_conn(inode)->writeback_cache) {
1195                 /* Update size (EOF optimization) and mode (SUID clearing) */
1196                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1197                 if (err)
1198                         return err;
1199
1200                 return generic_file_write_iter(iocb, from);
1201         }
1202
1203         mutex_lock(&inode->i_mutex);
1204
1205         /* We can write back this queue in page reclaim */
1206         current->backing_dev_info = inode_to_bdi(inode);
1207
1208         err = generic_write_checks(iocb, from);
1209         if (err <= 0)
1210                 goto out;
1211
1212         err = file_remove_privs(file);
1213         if (err)
1214                 goto out;
1215
1216         err = file_update_time(file);
1217         if (err)
1218                 goto out;
1219
1220         if (iocb->ki_flags & IOCB_DIRECT) {
1221                 loff_t pos = iocb->ki_pos;
1222                 written = generic_file_direct_write(iocb, from, pos);
1223                 if (written < 0 || !iov_iter_count(from))
1224                         goto out;
1225
1226                 pos += written;
1227
1228                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1229                 if (written_buffered < 0) {
1230                         err = written_buffered;
1231                         goto out;
1232                 }
1233                 endbyte = pos + written_buffered - 1;
1234
1235                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1236                                                    endbyte);
1237                 if (err)
1238                         goto out;
1239
1240                 invalidate_mapping_pages(file->f_mapping,
1241                                          pos >> PAGE_CACHE_SHIFT,
1242                                          endbyte >> PAGE_CACHE_SHIFT);
1243
1244                 written += written_buffered;
1245                 iocb->ki_pos = pos + written_buffered;
1246         } else {
1247                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1248                 if (written >= 0)
1249                         iocb->ki_pos += written;
1250         }
1251 out:
1252         current->backing_dev_info = NULL;
1253         mutex_unlock(&inode->i_mutex);
1254
1255         return written ? written : err;
1256 }
1257
1258 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1259                 unsigned index, unsigned nr_pages)
1260 {
1261         int i;
1262
1263         for (i = index; i < index + nr_pages; i++)
1264                 req->page_descs[i].length = PAGE_SIZE -
1265                         req->page_descs[i].offset;
1266 }
1267
1268 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1269 {
1270         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1271 }
1272
1273 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1274                                         size_t max_size)
1275 {
1276         return min(iov_iter_single_seg_count(ii), max_size);
1277 }
1278
1279 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1280                                size_t *nbytesp, int write)
1281 {
1282         size_t nbytes = 0;  /* # bytes already packed in req */
1283
1284         /* Special case for kernel I/O: can copy directly into the buffer */
1285         if (ii->type & ITER_KVEC) {
1286                 unsigned long user_addr = fuse_get_user_addr(ii);
1287                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1288
1289                 if (write)
1290                         req->in.args[1].value = (void *) user_addr;
1291                 else
1292                         req->out.args[0].value = (void *) user_addr;
1293
1294                 iov_iter_advance(ii, frag_size);
1295                 *nbytesp = frag_size;
1296                 return 0;
1297         }
1298
1299         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1300                 unsigned npages;
1301                 size_t start;
1302                 ssize_t ret = iov_iter_get_pages(ii,
1303                                         &req->pages[req->num_pages],
1304                                         *nbytesp - nbytes,
1305                                         req->max_pages - req->num_pages,
1306                                         &start);
1307                 if (ret < 0)
1308                         return ret;
1309
1310                 iov_iter_advance(ii, ret);
1311                 nbytes += ret;
1312
1313                 ret += start;
1314                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1315
1316                 req->page_descs[req->num_pages].offset = start;
1317                 fuse_page_descs_length_init(req, req->num_pages, npages);
1318
1319                 req->num_pages += npages;
1320                 req->page_descs[req->num_pages - 1].length -=
1321                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1322         }
1323
1324         if (write)
1325                 req->in.argpages = 1;
1326         else
1327                 req->out.argpages = 1;
1328
1329         *nbytesp = nbytes;
1330
1331         return 0;
1332 }
1333
1334 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1335 {
1336         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1337 }
1338
1339 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1340                        loff_t *ppos, int flags)
1341 {
1342         int write = flags & FUSE_DIO_WRITE;
1343         int cuse = flags & FUSE_DIO_CUSE;
1344         struct file *file = io->file;
1345         struct inode *inode = file->f_mapping->host;
1346         struct fuse_file *ff = file->private_data;
1347         struct fuse_conn *fc = ff->fc;
1348         size_t nmax = write ? fc->max_write : fc->max_read;
1349         loff_t pos = *ppos;
1350         size_t count = iov_iter_count(iter);
1351         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1352         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1353         ssize_t res = 0;
1354         struct fuse_req *req;
1355
1356         if (io->async)
1357                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1358         else
1359                 req = fuse_get_req(fc, fuse_iter_npages(iter));
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                         mutex_lock(&inode->i_mutex);
1366                 fuse_sync_writes(inode);
1367                 if (!write)
1368                         mutex_unlock(&inode->i_mutex);
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                 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1377                 if (err) {
1378                         res = err;
1379                         break;
1380                 }
1381
1382                 if (write)
1383                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1384                 else
1385                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1386
1387                 if (!io->async)
1388                         fuse_release_user_pages(req, io->should_dirty);
1389                 if (req->out.h.error) {
1390                         if (!res)
1391                                 res = req->out.h.error;
1392                         break;
1393                 } else if (nres > nbytes) {
1394                         res = -EIO;
1395                         break;
1396                 }
1397                 count -= nres;
1398                 res += nres;
1399                 pos += nres;
1400                 if (nres != nbytes)
1401                         break;
1402                 if (count) {
1403                         fuse_put_request(fc, req);
1404                         if (io->async)
1405                                 req = fuse_get_req_for_background(fc,
1406                                         fuse_iter_npages(iter));
1407                         else
1408                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1409                         if (IS_ERR(req))
1410                                 break;
1411                 }
1412         }
1413         if (!IS_ERR(req))
1414                 fuse_put_request(fc, req);
1415         if (res > 0)
1416                 *ppos = pos;
1417
1418         return res;
1419 }
1420 EXPORT_SYMBOL_GPL(fuse_direct_io);
1421
1422 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1423                                   struct iov_iter *iter,
1424                                   loff_t *ppos)
1425 {
1426         ssize_t res;
1427         struct file *file = io->file;
1428         struct inode *inode = file_inode(file);
1429
1430         if (is_bad_inode(inode))
1431                 return -EIO;
1432
1433         res = fuse_direct_io(io, iter, ppos, 0);
1434
1435         fuse_invalidate_attr(inode);
1436
1437         return res;
1438 }
1439
1440 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1441 {
1442         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1443         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1444 }
1445
1446 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1447 {
1448         struct file *file = iocb->ki_filp;
1449         struct inode *inode = file_inode(file);
1450         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1451         ssize_t res;
1452
1453         if (is_bad_inode(inode))
1454                 return -EIO;
1455
1456         /* Don't allow parallel writes to the same file */
1457         mutex_lock(&inode->i_mutex);
1458         res = generic_write_checks(iocb, from);
1459         if (res > 0)
1460                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1461         fuse_invalidate_attr(inode);
1462         if (res > 0)
1463                 fuse_write_update_size(inode, iocb->ki_pos);
1464         mutex_unlock(&inode->i_mutex);
1465
1466         return res;
1467 }
1468
1469 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1470 {
1471         int i;
1472
1473         for (i = 0; i < req->num_pages; i++)
1474                 __free_page(req->pages[i]);
1475
1476         if (req->ff)
1477                 fuse_file_put(req->ff, false);
1478 }
1479
1480 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1481 {
1482         struct inode *inode = req->inode;
1483         struct fuse_inode *fi = get_fuse_inode(inode);
1484         struct backing_dev_info *bdi = inode_to_bdi(inode);
1485         int i;
1486
1487         list_del(&req->writepages_entry);
1488         for (i = 0; i < req->num_pages; i++) {
1489                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1490                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1491                 wb_writeout_inc(&bdi->wb);
1492         }
1493         wake_up(&fi->page_waitq);
1494 }
1495
1496 /* Called under fc->lock, may release and reacquire it */
1497 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1498                                 loff_t size)
1499 __releases(fc->lock)
1500 __acquires(fc->lock)
1501 {
1502         struct fuse_inode *fi = get_fuse_inode(req->inode);
1503         struct fuse_write_in *inarg = &req->misc.write.in;
1504         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1505
1506         if (!fc->connected)
1507                 goto out_free;
1508
1509         if (inarg->offset + data_size <= size) {
1510                 inarg->size = data_size;
1511         } else if (inarg->offset < size) {
1512                 inarg->size = size - inarg->offset;
1513         } else {
1514                 /* Got truncated off completely */
1515                 goto out_free;
1516         }
1517
1518         req->in.args[1].size = inarg->size;
1519         fi->writectr++;
1520         fuse_request_send_background_locked(fc, req);
1521         return;
1522
1523  out_free:
1524         fuse_writepage_finish(fc, req);
1525         spin_unlock(&fc->lock);
1526         fuse_writepage_free(fc, req);
1527         fuse_put_request(fc, req);
1528         spin_lock(&fc->lock);
1529 }
1530
1531 /*
1532  * If fi->writectr is positive (no truncate or fsync going on) send
1533  * all queued writepage requests.
1534  *
1535  * Called with fc->lock
1536  */
1537 void fuse_flush_writepages(struct inode *inode)
1538 __releases(fc->lock)
1539 __acquires(fc->lock)
1540 {
1541         struct fuse_conn *fc = get_fuse_conn(inode);
1542         struct fuse_inode *fi = get_fuse_inode(inode);
1543         loff_t crop = i_size_read(inode);
1544         struct fuse_req *req;
1545
1546         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1547                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1548                 list_del_init(&req->list);
1549                 fuse_send_writepage(fc, req, crop);
1550         }
1551 }
1552
1553 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1554 {
1555         struct inode *inode = req->inode;
1556         struct fuse_inode *fi = get_fuse_inode(inode);
1557
1558         mapping_set_error(inode->i_mapping, req->out.h.error);
1559         spin_lock(&fc->lock);
1560         while (req->misc.write.next) {
1561                 struct fuse_conn *fc = get_fuse_conn(inode);
1562                 struct fuse_write_in *inarg = &req->misc.write.in;
1563                 struct fuse_req *next = req->misc.write.next;
1564                 req->misc.write.next = next->misc.write.next;
1565                 next->misc.write.next = NULL;
1566                 next->ff = fuse_file_get(req->ff);
1567                 list_add(&next->writepages_entry, &fi->writepages);
1568
1569                 /*
1570                  * Skip fuse_flush_writepages() to make it easy to crop requests
1571                  * based on primary request size.
1572                  *
1573                  * 1st case (trivial): there are no concurrent activities using
1574                  * fuse_set/release_nowrite.  Then we're on safe side because
1575                  * fuse_flush_writepages() would call fuse_send_writepage()
1576                  * anyway.
1577                  *
1578                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1579                  * now for completion of all in-flight requests.  This happens
1580                  * rarely and no more than once per page, so this should be
1581                  * okay.
1582                  *
1583                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1584                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1585                  * that fuse_set_nowrite returned implies that all in-flight
1586                  * requests were completed along with all of their secondary
1587                  * requests.  Further primary requests are blocked by negative
1588                  * writectr.  Hence there cannot be any in-flight requests and
1589                  * no invocations of fuse_writepage_end() while we're in
1590                  * fuse_set_nowrite..fuse_release_nowrite section.
1591                  */
1592                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1593         }
1594         fi->writectr--;
1595         fuse_writepage_finish(fc, req);
1596         spin_unlock(&fc->lock);
1597         fuse_writepage_free(fc, req);
1598 }
1599
1600 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1601                                                struct fuse_inode *fi)
1602 {
1603         struct fuse_file *ff = NULL;
1604
1605         spin_lock(&fc->lock);
1606         if (!list_empty(&fi->write_files)) {
1607                 ff = list_entry(fi->write_files.next, struct fuse_file,
1608                                 write_entry);
1609                 fuse_file_get(ff);
1610         }
1611         spin_unlock(&fc->lock);
1612
1613         return ff;
1614 }
1615
1616 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1617                                              struct fuse_inode *fi)
1618 {
1619         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1620         WARN_ON(!ff);
1621         return ff;
1622 }
1623
1624 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1625 {
1626         struct fuse_conn *fc = get_fuse_conn(inode);
1627         struct fuse_inode *fi = get_fuse_inode(inode);
1628         struct fuse_file *ff;
1629         int err;
1630
1631         ff = __fuse_write_file_get(fc, fi);
1632         err = fuse_flush_times(inode, ff);
1633         if (ff)
1634                 fuse_file_put(ff, 0);
1635
1636         return err;
1637 }
1638
1639 static int fuse_writepage_locked(struct page *page)
1640 {
1641         struct address_space *mapping = page->mapping;
1642         struct inode *inode = mapping->host;
1643         struct fuse_conn *fc = get_fuse_conn(inode);
1644         struct fuse_inode *fi = get_fuse_inode(inode);
1645         struct fuse_req *req;
1646         struct page *tmp_page;
1647         int error = -ENOMEM;
1648
1649         set_page_writeback(page);
1650
1651         req = fuse_request_alloc_nofs(1);
1652         if (!req)
1653                 goto err;
1654
1655         /* writeback always goes to bg_queue */
1656         __set_bit(FR_BACKGROUND, &req->flags);
1657         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1658         if (!tmp_page)
1659                 goto err_free;
1660
1661         error = -EIO;
1662         req->ff = fuse_write_file_get(fc, fi);
1663         if (!req->ff)
1664                 goto err_nofile;
1665
1666         fuse_write_fill(req, req->ff, page_offset(page), 0);
1667
1668         copy_highpage(tmp_page, page);
1669         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1670         req->misc.write.next = NULL;
1671         req->in.argpages = 1;
1672         req->num_pages = 1;
1673         req->pages[0] = tmp_page;
1674         req->page_descs[0].offset = 0;
1675         req->page_descs[0].length = PAGE_SIZE;
1676         req->end = fuse_writepage_end;
1677         req->inode = inode;
1678
1679         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1680         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1681
1682         spin_lock(&fc->lock);
1683         list_add(&req->writepages_entry, &fi->writepages);
1684         list_add_tail(&req->list, &fi->queued_writes);
1685         fuse_flush_writepages(inode);
1686         spin_unlock(&fc->lock);
1687
1688         end_page_writeback(page);
1689
1690         return 0;
1691
1692 err_nofile:
1693         __free_page(tmp_page);
1694 err_free:
1695         fuse_request_free(req);
1696 err:
1697         end_page_writeback(page);
1698         return error;
1699 }
1700
1701 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1702 {
1703         int err;
1704
1705         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1706                 /*
1707                  * ->writepages() should be called for sync() and friends.  We
1708                  * should only get here on direct reclaim and then we are
1709                  * allowed to skip a page which is already in flight
1710                  */
1711                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1712
1713                 redirty_page_for_writepage(wbc, page);
1714                 unlock_page(page);
1715                 return 0;
1716         }
1717
1718         err = fuse_writepage_locked(page);
1719         unlock_page(page);
1720
1721         return err;
1722 }
1723
1724 struct fuse_fill_wb_data {
1725         struct fuse_req *req;
1726         struct fuse_file *ff;
1727         struct inode *inode;
1728         struct page **orig_pages;
1729 };
1730
1731 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1732 {
1733         struct fuse_req *req = data->req;
1734         struct inode *inode = data->inode;
1735         struct fuse_conn *fc = get_fuse_conn(inode);
1736         struct fuse_inode *fi = get_fuse_inode(inode);
1737         int num_pages = req->num_pages;
1738         int i;
1739
1740         req->ff = fuse_file_get(data->ff);
1741         spin_lock(&fc->lock);
1742         list_add_tail(&req->list, &fi->queued_writes);
1743         fuse_flush_writepages(inode);
1744         spin_unlock(&fc->lock);
1745
1746         for (i = 0; i < num_pages; i++)
1747                 end_page_writeback(data->orig_pages[i]);
1748 }
1749
1750 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1751                                      struct page *page)
1752 {
1753         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1754         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1755         struct fuse_req *tmp;
1756         struct fuse_req *old_req;
1757         bool found = false;
1758         pgoff_t curr_index;
1759
1760         BUG_ON(new_req->num_pages != 0);
1761
1762         spin_lock(&fc->lock);
1763         list_del(&new_req->writepages_entry);
1764         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1765                 BUG_ON(old_req->inode != new_req->inode);
1766                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1767                 if (curr_index <= page->index &&
1768                     page->index < curr_index + old_req->num_pages) {
1769                         found = true;
1770                         break;
1771                 }
1772         }
1773         if (!found) {
1774                 list_add(&new_req->writepages_entry, &fi->writepages);
1775                 goto out_unlock;
1776         }
1777
1778         new_req->num_pages = 1;
1779         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1780                 BUG_ON(tmp->inode != new_req->inode);
1781                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1782                 if (tmp->num_pages == 1 &&
1783                     curr_index == page->index) {
1784                         old_req = tmp;
1785                 }
1786         }
1787
1788         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1789                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1790
1791                 copy_highpage(old_req->pages[0], page);
1792                 spin_unlock(&fc->lock);
1793
1794                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1795                 dec_zone_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1796                 wb_writeout_inc(&bdi->wb);
1797                 fuse_writepage_free(fc, new_req);
1798                 fuse_request_free(new_req);
1799                 goto out;
1800         } else {
1801                 new_req->misc.write.next = old_req->misc.write.next;
1802                 old_req->misc.write.next = new_req;
1803         }
1804 out_unlock:
1805         spin_unlock(&fc->lock);
1806 out:
1807         return found;
1808 }
1809
1810 static int fuse_writepages_fill(struct page *page,
1811                 struct writeback_control *wbc, void *_data)
1812 {
1813         struct fuse_fill_wb_data *data = _data;
1814         struct fuse_req *req = data->req;
1815         struct inode *inode = data->inode;
1816         struct fuse_conn *fc = get_fuse_conn(inode);
1817         struct page *tmp_page;
1818         bool is_writeback;
1819         int err;
1820
1821         if (!data->ff) {
1822                 err = -EIO;
1823                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1824                 if (!data->ff)
1825                         goto out_unlock;
1826         }
1827
1828         /*
1829          * Being under writeback is unlikely but possible.  For example direct
1830          * read to an mmaped fuse file will set the page dirty twice; once when
1831          * the pages are faulted with get_user_pages(), and then after the read
1832          * completed.
1833          */
1834         is_writeback = fuse_page_is_writeback(inode, page->index);
1835
1836         if (req && req->num_pages &&
1837             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1838              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1839              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1840                 fuse_writepages_send(data);
1841                 data->req = NULL;
1842         }
1843         err = -ENOMEM;
1844         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1845         if (!tmp_page)
1846                 goto out_unlock;
1847
1848         /*
1849          * The page must not be redirtied until the writeout is completed
1850          * (i.e. userspace has sent a reply to the write request).  Otherwise
1851          * there could be more than one temporary page instance for each real
1852          * page.
1853          *
1854          * This is ensured by holding the page lock in page_mkwrite() while
1855          * checking fuse_page_is_writeback().  We already hold the page lock
1856          * since clear_page_dirty_for_io() and keep it held until we add the
1857          * request to the fi->writepages list and increment req->num_pages.
1858          * After this fuse_page_is_writeback() will indicate that the page is
1859          * under writeback, so we can release the page lock.
1860          */
1861         if (data->req == NULL) {
1862                 struct fuse_inode *fi = get_fuse_inode(inode);
1863
1864                 err = -ENOMEM;
1865                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1866                 if (!req) {
1867                         __free_page(tmp_page);
1868                         goto out_unlock;
1869                 }
1870
1871                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1872                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1873                 req->misc.write.next = NULL;
1874                 req->in.argpages = 1;
1875                 __set_bit(FR_BACKGROUND, &req->flags);
1876                 req->num_pages = 0;
1877                 req->end = fuse_writepage_end;
1878                 req->inode = inode;
1879
1880                 spin_lock(&fc->lock);
1881                 list_add(&req->writepages_entry, &fi->writepages);
1882                 spin_unlock(&fc->lock);
1883
1884                 data->req = req;
1885         }
1886         set_page_writeback(page);
1887
1888         copy_highpage(tmp_page, page);
1889         req->pages[req->num_pages] = tmp_page;
1890         req->page_descs[req->num_pages].offset = 0;
1891         req->page_descs[req->num_pages].length = PAGE_SIZE;
1892
1893         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1894         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1895
1896         err = 0;
1897         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1898                 end_page_writeback(page);
1899                 data->req = NULL;
1900                 goto out_unlock;
1901         }
1902         data->orig_pages[req->num_pages] = page;
1903
1904         /*
1905          * Protected by fc->lock against concurrent access by
1906          * fuse_page_is_writeback().
1907          */
1908         spin_lock(&fc->lock);
1909         req->num_pages++;
1910         spin_unlock(&fc->lock);
1911
1912 out_unlock:
1913         unlock_page(page);
1914
1915         return err;
1916 }
1917
1918 static int fuse_writepages(struct address_space *mapping,
1919                            struct writeback_control *wbc)
1920 {
1921         struct inode *inode = mapping->host;
1922         struct fuse_fill_wb_data data;
1923         int err;
1924
1925         err = -EIO;
1926         if (is_bad_inode(inode))
1927                 goto out;
1928
1929         data.inode = inode;
1930         data.req = NULL;
1931         data.ff = NULL;
1932
1933         err = -ENOMEM;
1934         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1935                                   sizeof(struct page *),
1936                                   GFP_NOFS);
1937         if (!data.orig_pages)
1938                 goto out;
1939
1940         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1941         if (data.req) {
1942                 /* Ignore errors if we can write at least one page */
1943                 BUG_ON(!data.req->num_pages);
1944                 fuse_writepages_send(&data);
1945                 err = 0;
1946         }
1947         if (data.ff)
1948                 fuse_file_put(data.ff, false);
1949
1950         kfree(data.orig_pages);
1951 out:
1952         return err;
1953 }
1954
1955 /*
1956  * It's worthy to make sure that space is reserved on disk for the write,
1957  * but how to implement it without killing performance need more thinking.
1958  */
1959 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1960                 loff_t pos, unsigned len, unsigned flags,
1961                 struct page **pagep, void **fsdata)
1962 {
1963         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1964         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1965         struct page *page;
1966         loff_t fsize;
1967         int err = -ENOMEM;
1968
1969         WARN_ON(!fc->writeback_cache);
1970
1971         page = grab_cache_page_write_begin(mapping, index, flags);
1972         if (!page)
1973                 goto error;
1974
1975         fuse_wait_on_page_writeback(mapping->host, page->index);
1976
1977         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1978                 goto success;
1979         /*
1980          * Check if the start this page comes after the end of file, in which
1981          * case the readpage can be optimized away.
1982          */
1983         fsize = i_size_read(mapping->host);
1984         if (fsize <= (pos & PAGE_CACHE_MASK)) {
1985                 size_t off = pos & ~PAGE_CACHE_MASK;
1986                 if (off)
1987                         zero_user_segment(page, 0, off);
1988                 goto success;
1989         }
1990         err = fuse_do_readpage(file, page);
1991         if (err)
1992                 goto cleanup;
1993 success:
1994         *pagep = page;
1995         return 0;
1996
1997 cleanup:
1998         unlock_page(page);
1999         page_cache_release(page);
2000 error:
2001         return err;
2002 }
2003
2004 static int fuse_write_end(struct file *file, struct address_space *mapping,
2005                 loff_t pos, unsigned len, unsigned copied,
2006                 struct page *page, void *fsdata)
2007 {
2008         struct inode *inode = page->mapping->host;
2009
2010         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2011         if (!copied)
2012                 goto unlock;
2013
2014         if (!PageUptodate(page)) {
2015                 /* Zero any unwritten bytes at the end of the page */
2016                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2017                 if (endoff)
2018                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2019                 SetPageUptodate(page);
2020         }
2021
2022         fuse_write_update_size(inode, pos + copied);
2023         set_page_dirty(page);
2024
2025 unlock:
2026         unlock_page(page);
2027         page_cache_release(page);
2028
2029         return copied;
2030 }
2031
2032 static int fuse_launder_page(struct page *page)
2033 {
2034         int err = 0;
2035         if (clear_page_dirty_for_io(page)) {
2036                 struct inode *inode = page->mapping->host;
2037                 err = fuse_writepage_locked(page);
2038                 if (!err)
2039                         fuse_wait_on_page_writeback(inode, page->index);
2040         }
2041         return err;
2042 }
2043
2044 /*
2045  * Write back dirty pages now, because there may not be any suitable
2046  * open files later
2047  */
2048 static void fuse_vma_close(struct vm_area_struct *vma)
2049 {
2050         filemap_write_and_wait(vma->vm_file->f_mapping);
2051 }
2052
2053 /*
2054  * Wait for writeback against this page to complete before allowing it
2055  * to be marked dirty again, and hence written back again, possibly
2056  * before the previous writepage completed.
2057  *
2058  * Block here, instead of in ->writepage(), so that the userspace fs
2059  * can only block processes actually operating on the filesystem.
2060  *
2061  * Otherwise unprivileged userspace fs would be able to block
2062  * unrelated:
2063  *
2064  * - page migration
2065  * - sync(2)
2066  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2067  */
2068 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2069 {
2070         struct page *page = vmf->page;
2071         struct inode *inode = file_inode(vma->vm_file);
2072
2073         file_update_time(vma->vm_file);
2074         lock_page(page);
2075         if (page->mapping != inode->i_mapping) {
2076                 unlock_page(page);
2077                 return VM_FAULT_NOPAGE;
2078         }
2079
2080         fuse_wait_on_page_writeback(inode, page->index);
2081         return VM_FAULT_LOCKED;
2082 }
2083
2084 static const struct vm_operations_struct fuse_file_vm_ops = {
2085         .close          = fuse_vma_close,
2086         .fault          = filemap_fault,
2087         .map_pages      = filemap_map_pages,
2088         .page_mkwrite   = fuse_page_mkwrite,
2089 };
2090
2091 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2092 {
2093         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2094                 fuse_link_write_file(file);
2095
2096         file_accessed(file);
2097         vma->vm_ops = &fuse_file_vm_ops;
2098         return 0;
2099 }
2100
2101 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2102 {
2103         /* Can't provide the coherency needed for MAP_SHARED */
2104         if (vma->vm_flags & VM_MAYSHARE)
2105                 return -ENODEV;
2106
2107         invalidate_inode_pages2(file->f_mapping);
2108
2109         return generic_file_mmap(file, vma);
2110 }
2111
2112 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2113                                   struct file_lock *fl)
2114 {
2115         switch (ffl->type) {
2116         case F_UNLCK:
2117                 break;
2118
2119         case F_RDLCK:
2120         case F_WRLCK:
2121                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2122                     ffl->end < ffl->start)
2123                         return -EIO;
2124
2125                 fl->fl_start = ffl->start;
2126                 fl->fl_end = ffl->end;
2127                 fl->fl_pid = ffl->pid;
2128                 break;
2129
2130         default:
2131                 return -EIO;
2132         }
2133         fl->fl_type = ffl->type;
2134         return 0;
2135 }
2136
2137 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2138                          const struct file_lock *fl, int opcode, pid_t pid,
2139                          int flock, struct fuse_lk_in *inarg)
2140 {
2141         struct inode *inode = file_inode(file);
2142         struct fuse_conn *fc = get_fuse_conn(inode);
2143         struct fuse_file *ff = file->private_data;
2144
2145         memset(inarg, 0, sizeof(*inarg));
2146         inarg->fh = ff->fh;
2147         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2148         inarg->lk.start = fl->fl_start;
2149         inarg->lk.end = fl->fl_end;
2150         inarg->lk.type = fl->fl_type;
2151         inarg->lk.pid = pid;
2152         if (flock)
2153                 inarg->lk_flags |= FUSE_LK_FLOCK;
2154         args->in.h.opcode = opcode;
2155         args->in.h.nodeid = get_node_id(inode);
2156         args->in.numargs = 1;
2157         args->in.args[0].size = sizeof(*inarg);
2158         args->in.args[0].value = inarg;
2159 }
2160
2161 static int fuse_getlk(struct file *file, struct file_lock *fl)
2162 {
2163         struct inode *inode = file_inode(file);
2164         struct fuse_conn *fc = get_fuse_conn(inode);
2165         FUSE_ARGS(args);
2166         struct fuse_lk_in inarg;
2167         struct fuse_lk_out outarg;
2168         int err;
2169
2170         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2171         args.out.numargs = 1;
2172         args.out.args[0].size = sizeof(outarg);
2173         args.out.args[0].value = &outarg;
2174         err = fuse_simple_request(fc, &args);
2175         if (!err)
2176                 err = convert_fuse_file_lock(&outarg.lk, fl);
2177
2178         return err;
2179 }
2180
2181 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2182 {
2183         struct inode *inode = file_inode(file);
2184         struct fuse_conn *fc = get_fuse_conn(inode);
2185         FUSE_ARGS(args);
2186         struct fuse_lk_in inarg;
2187         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2188         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2189         int err;
2190
2191         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2192                 /* NLM needs asynchronous locks, which we don't support yet */
2193                 return -ENOLCK;
2194         }
2195
2196         /* Unlock on close is handled by the flush method */
2197         if (fl->fl_flags & FL_CLOSE)
2198                 return 0;
2199
2200         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2201         err = fuse_simple_request(fc, &args);
2202
2203         /* locking is restartable */
2204         if (err == -EINTR)
2205                 err = -ERESTARTSYS;
2206
2207         return err;
2208 }
2209
2210 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2211 {
2212         struct inode *inode = file_inode(file);
2213         struct fuse_conn *fc = get_fuse_conn(inode);
2214         int err;
2215
2216         if (cmd == F_CANCELLK) {
2217                 err = 0;
2218         } else if (cmd == F_GETLK) {
2219                 if (fc->no_lock) {
2220                         posix_test_lock(file, fl);
2221                         err = 0;
2222                 } else
2223                         err = fuse_getlk(file, fl);
2224         } else {
2225                 if (fc->no_lock)
2226                         err = posix_lock_file(file, fl, NULL);
2227                 else
2228                         err = fuse_setlk(file, fl, 0);
2229         }
2230         return err;
2231 }
2232
2233 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2234 {
2235         struct inode *inode = file_inode(file);
2236         struct fuse_conn *fc = get_fuse_conn(inode);
2237         int err;
2238
2239         if (fc->no_flock) {
2240                 err = locks_lock_file_wait(file, fl);
2241         } else {
2242                 struct fuse_file *ff = file->private_data;
2243
2244                 /* emulate flock with POSIX locks */
2245                 ff->flock = true;
2246                 err = fuse_setlk(file, fl, 1);
2247         }
2248
2249         return err;
2250 }
2251
2252 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2253 {
2254         struct inode *inode = mapping->host;
2255         struct fuse_conn *fc = get_fuse_conn(inode);
2256         FUSE_ARGS(args);
2257         struct fuse_bmap_in inarg;
2258         struct fuse_bmap_out outarg;
2259         int err;
2260
2261         if (!inode->i_sb->s_bdev || fc->no_bmap)
2262                 return 0;
2263
2264         memset(&inarg, 0, sizeof(inarg));
2265         inarg.block = block;
2266         inarg.blocksize = inode->i_sb->s_blocksize;
2267         args.in.h.opcode = FUSE_BMAP;
2268         args.in.h.nodeid = get_node_id(inode);
2269         args.in.numargs = 1;
2270         args.in.args[0].size = sizeof(inarg);
2271         args.in.args[0].value = &inarg;
2272         args.out.numargs = 1;
2273         args.out.args[0].size = sizeof(outarg);
2274         args.out.args[0].value = &outarg;
2275         err = fuse_simple_request(fc, &args);
2276         if (err == -ENOSYS)
2277                 fc->no_bmap = 1;
2278
2279         return err ? 0 : outarg.block;
2280 }
2281
2282 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2283 {
2284         loff_t retval;
2285         struct inode *inode = file_inode(file);
2286
2287         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2288         if (whence == SEEK_CUR || whence == SEEK_SET)
2289                 return generic_file_llseek(file, offset, whence);
2290
2291         mutex_lock(&inode->i_mutex);
2292         retval = fuse_update_attributes(inode, NULL, file, NULL);
2293         if (!retval)
2294                 retval = generic_file_llseek(file, offset, whence);
2295         mutex_unlock(&inode->i_mutex);
2296
2297         return retval;
2298 }
2299
2300 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2301                         unsigned int nr_segs, size_t bytes, bool to_user)
2302 {
2303         struct iov_iter ii;
2304         int page_idx = 0;
2305
2306         if (!bytes)
2307                 return 0;
2308
2309         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2310
2311         while (iov_iter_count(&ii)) {
2312                 struct page *page = pages[page_idx++];
2313                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2314                 void *kaddr;
2315
2316                 kaddr = kmap(page);
2317
2318                 while (todo) {
2319                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2320                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2321                         size_t copy = min(todo, iov_len);
2322                         size_t left;
2323
2324                         if (!to_user)
2325                                 left = copy_from_user(kaddr, uaddr, copy);
2326                         else
2327                                 left = copy_to_user(uaddr, kaddr, copy);
2328
2329                         if (unlikely(left))
2330                                 return -EFAULT;
2331
2332                         iov_iter_advance(&ii, copy);
2333                         todo -= copy;
2334                         kaddr += copy;
2335                 }
2336
2337                 kunmap(page);
2338         }
2339
2340         return 0;
2341 }
2342
2343 /*
2344  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2345  * ABI was defined to be 'struct iovec' which is different on 32bit
2346  * and 64bit.  Fortunately we can determine which structure the server
2347  * used from the size of the reply.
2348  */
2349 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2350                                      size_t transferred, unsigned count,
2351                                      bool is_compat)
2352 {
2353 #ifdef CONFIG_COMPAT
2354         if (count * sizeof(struct compat_iovec) == transferred) {
2355                 struct compat_iovec *ciov = src;
2356                 unsigned i;
2357
2358                 /*
2359                  * With this interface a 32bit server cannot support
2360                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2361                  * requests
2362                  */
2363                 if (!is_compat)
2364                         return -EINVAL;
2365
2366                 for (i = 0; i < count; i++) {
2367                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2368                         dst[i].iov_len = ciov[i].iov_len;
2369                 }
2370                 return 0;
2371         }
2372 #endif
2373
2374         if (count * sizeof(struct iovec) != transferred)
2375                 return -EIO;
2376
2377         memcpy(dst, src, transferred);
2378         return 0;
2379 }
2380
2381 /* Make sure iov_length() won't overflow */
2382 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2383 {
2384         size_t n;
2385         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2386
2387         for (n = 0; n < count; n++, iov++) {
2388                 if (iov->iov_len > (size_t) max)
2389                         return -ENOMEM;
2390                 max -= iov->iov_len;
2391         }
2392         return 0;
2393 }
2394
2395 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2396                                  void *src, size_t transferred, unsigned count,
2397                                  bool is_compat)
2398 {
2399         unsigned i;
2400         struct fuse_ioctl_iovec *fiov = src;
2401
2402         if (fc->minor < 16) {
2403                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2404                                                  count, is_compat);
2405         }
2406
2407         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2408                 return -EIO;
2409
2410         for (i = 0; i < count; i++) {
2411                 /* Did the server supply an inappropriate value? */
2412                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2413                     fiov[i].len != (unsigned long) fiov[i].len)
2414                         return -EIO;
2415
2416                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2417                 dst[i].iov_len = (size_t) fiov[i].len;
2418
2419 #ifdef CONFIG_COMPAT
2420                 if (is_compat &&
2421                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2422                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2423                         return -EIO;
2424 #endif
2425         }
2426
2427         return 0;
2428 }
2429
2430
2431 /*
2432  * For ioctls, there is no generic way to determine how much memory
2433  * needs to be read and/or written.  Furthermore, ioctls are allowed
2434  * to dereference the passed pointer, so the parameter requires deep
2435  * copying but FUSE has no idea whatsoever about what to copy in or
2436  * out.
2437  *
2438  * This is solved by allowing FUSE server to retry ioctl with
2439  * necessary in/out iovecs.  Let's assume the ioctl implementation
2440  * needs to read in the following structure.
2441  *
2442  * struct a {
2443  *      char    *buf;
2444  *      size_t  buflen;
2445  * }
2446  *
2447  * On the first callout to FUSE server, inarg->in_size and
2448  * inarg->out_size will be NULL; then, the server completes the ioctl
2449  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2450  * the actual iov array to
2451  *
2452  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2453  *
2454  * which tells FUSE to copy in the requested area and retry the ioctl.
2455  * On the second round, the server has access to the structure and
2456  * from that it can tell what to look for next, so on the invocation,
2457  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2458  *
2459  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2460  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2461  *
2462  * FUSE will copy both struct a and the pointed buffer from the
2463  * process doing the ioctl and retry ioctl with both struct a and the
2464  * buffer.
2465  *
2466  * This time, FUSE server has everything it needs and completes ioctl
2467  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2468  *
2469  * Copying data out works the same way.
2470  *
2471  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2472  * automatically initializes in and out iovs by decoding @cmd with
2473  * _IOC_* macros and the server is not allowed to request RETRY.  This
2474  * limits ioctl data transfers to well-formed ioctls and is the forced
2475  * behavior for all FUSE servers.
2476  */
2477 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2478                    unsigned int flags)
2479 {
2480         struct fuse_file *ff = file->private_data;
2481         struct fuse_conn *fc = ff->fc;
2482         struct fuse_ioctl_in inarg = {
2483                 .fh = ff->fh,
2484                 .cmd = cmd,
2485                 .arg = arg,
2486                 .flags = flags
2487         };
2488         struct fuse_ioctl_out outarg;
2489         struct fuse_req *req = NULL;
2490         struct page **pages = NULL;
2491         struct iovec *iov_page = NULL;
2492         struct iovec *in_iov = NULL, *out_iov = NULL;
2493         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2494         size_t in_size, out_size, transferred;
2495         int err;
2496
2497 #if BITS_PER_LONG == 32
2498         inarg.flags |= FUSE_IOCTL_32BIT;
2499 #else
2500         if (flags & FUSE_IOCTL_COMPAT)
2501                 inarg.flags |= FUSE_IOCTL_32BIT;
2502 #endif
2503
2504         /* assume all the iovs returned by client always fits in a page */
2505         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2506
2507         err = -ENOMEM;
2508         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2509         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2510         if (!pages || !iov_page)
2511                 goto out;
2512
2513         /*
2514          * If restricted, initialize IO parameters as encoded in @cmd.
2515          * RETRY from server is not allowed.
2516          */
2517         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2518                 struct iovec *iov = iov_page;
2519
2520                 iov->iov_base = (void __user *)arg;
2521
2522                 switch (cmd) {
2523                 case FS_IOC_GETFLAGS:
2524                 case FS_IOC_SETFLAGS:
2525                         iov->iov_len = sizeof(int);
2526                         break;
2527                 default:
2528                         iov->iov_len = _IOC_SIZE(cmd);
2529                         break;
2530                 }
2531
2532                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2533                         in_iov = iov;
2534                         in_iovs = 1;
2535                 }
2536
2537                 if (_IOC_DIR(cmd) & _IOC_READ) {
2538                         out_iov = iov;
2539                         out_iovs = 1;
2540                 }
2541         }
2542
2543  retry:
2544         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2545         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2546
2547         /*
2548          * Out data can be used either for actual out data or iovs,
2549          * make sure there always is at least one page.
2550          */
2551         out_size = max_t(size_t, out_size, PAGE_SIZE);
2552         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2553
2554         /* make sure there are enough buffer pages and init request with them */
2555         err = -ENOMEM;
2556         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2557                 goto out;
2558         while (num_pages < max_pages) {
2559                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2560                 if (!pages[num_pages])
2561                         goto out;
2562                 num_pages++;
2563         }
2564
2565         req = fuse_get_req(fc, num_pages);
2566         if (IS_ERR(req)) {
2567                 err = PTR_ERR(req);
2568                 req = NULL;
2569                 goto out;
2570         }
2571         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2572         req->num_pages = num_pages;
2573         fuse_page_descs_length_init(req, 0, req->num_pages);
2574
2575         /* okay, let's send it to the client */
2576         req->in.h.opcode = FUSE_IOCTL;
2577         req->in.h.nodeid = ff->nodeid;
2578         req->in.numargs = 1;
2579         req->in.args[0].size = sizeof(inarg);
2580         req->in.args[0].value = &inarg;
2581         if (in_size) {
2582                 req->in.numargs++;
2583                 req->in.args[1].size = in_size;
2584                 req->in.argpages = 1;
2585
2586                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2587                                            false);
2588                 if (err)
2589                         goto out;
2590         }
2591
2592         req->out.numargs = 2;
2593         req->out.args[0].size = sizeof(outarg);
2594         req->out.args[0].value = &outarg;
2595         req->out.args[1].size = out_size;
2596         req->out.argpages = 1;
2597         req->out.argvar = 1;
2598
2599         fuse_request_send(fc, req);
2600         err = req->out.h.error;
2601         transferred = req->out.args[1].size;
2602         fuse_put_request(fc, req);
2603         req = NULL;
2604         if (err)
2605                 goto out;
2606
2607         /* did it ask for retry? */
2608         if (outarg.flags & FUSE_IOCTL_RETRY) {
2609                 void *vaddr;
2610
2611                 /* no retry if in restricted mode */
2612                 err = -EIO;
2613                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2614                         goto out;
2615
2616                 in_iovs = outarg.in_iovs;
2617                 out_iovs = outarg.out_iovs;
2618
2619                 /*
2620                  * Make sure things are in boundary, separate checks
2621                  * are to protect against overflow.
2622                  */
2623                 err = -ENOMEM;
2624                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2625                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2626                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2627                         goto out;
2628
2629                 vaddr = kmap_atomic(pages[0]);
2630                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2631                                             transferred, in_iovs + out_iovs,
2632                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2633                 kunmap_atomic(vaddr);
2634                 if (err)
2635                         goto out;
2636
2637                 in_iov = iov_page;
2638                 out_iov = in_iov + in_iovs;
2639
2640                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2641                 if (err)
2642                         goto out;
2643
2644                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2645                 if (err)
2646                         goto out;
2647
2648                 goto retry;
2649         }
2650
2651         err = -EIO;
2652         if (transferred > inarg.out_size)
2653                 goto out;
2654
2655         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2656  out:
2657         if (req)
2658                 fuse_put_request(fc, req);
2659         free_page((unsigned long) iov_page);
2660         while (num_pages)
2661                 __free_page(pages[--num_pages]);
2662         kfree(pages);
2663
2664         return err ? err : outarg.result;
2665 }
2666 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2667
2668 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2669                        unsigned long arg, unsigned int flags)
2670 {
2671         struct inode *inode = file_inode(file);
2672         struct fuse_conn *fc = get_fuse_conn(inode);
2673
2674         if (!fuse_allow_current_process(fc))
2675                 return -EACCES;
2676
2677         if (is_bad_inode(inode))
2678                 return -EIO;
2679
2680         return fuse_do_ioctl(file, cmd, arg, flags);
2681 }
2682
2683 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2684                             unsigned long arg)
2685 {
2686         return fuse_ioctl_common(file, cmd, arg, 0);
2687 }
2688
2689 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2690                                    unsigned long arg)
2691 {
2692         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2693 }
2694
2695 /*
2696  * All files which have been polled are linked to RB tree
2697  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2698  * find the matching one.
2699  */
2700 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2701                                               struct rb_node **parent_out)
2702 {
2703         struct rb_node **link = &fc->polled_files.rb_node;
2704         struct rb_node *last = NULL;
2705
2706         while (*link) {
2707                 struct fuse_file *ff;
2708
2709                 last = *link;
2710                 ff = rb_entry(last, struct fuse_file, polled_node);
2711
2712                 if (kh < ff->kh)
2713                         link = &last->rb_left;
2714                 else if (kh > ff->kh)
2715                         link = &last->rb_right;
2716                 else
2717                         return link;
2718         }
2719
2720         if (parent_out)
2721                 *parent_out = last;
2722         return link;
2723 }
2724
2725 /*
2726  * The file is about to be polled.  Make sure it's on the polled_files
2727  * RB tree.  Note that files once added to the polled_files tree are
2728  * not removed before the file is released.  This is because a file
2729  * polled once is likely to be polled again.
2730  */
2731 static void fuse_register_polled_file(struct fuse_conn *fc,
2732                                       struct fuse_file *ff)
2733 {
2734         spin_lock(&fc->lock);
2735         if (RB_EMPTY_NODE(&ff->polled_node)) {
2736                 struct rb_node **link, *uninitialized_var(parent);
2737
2738                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2739                 BUG_ON(*link);
2740                 rb_link_node(&ff->polled_node, parent, link);
2741                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2742         }
2743         spin_unlock(&fc->lock);
2744 }
2745
2746 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2747 {
2748         struct fuse_file *ff = file->private_data;
2749         struct fuse_conn *fc = ff->fc;
2750         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2751         struct fuse_poll_out outarg;
2752         FUSE_ARGS(args);
2753         int err;
2754
2755         if (fc->no_poll)
2756                 return DEFAULT_POLLMASK;
2757
2758         poll_wait(file, &ff->poll_wait, wait);
2759         inarg.events = (__u32)poll_requested_events(wait);
2760
2761         /*
2762          * Ask for notification iff there's someone waiting for it.
2763          * The client may ignore the flag and always notify.
2764          */
2765         if (waitqueue_active(&ff->poll_wait)) {
2766                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2767                 fuse_register_polled_file(fc, ff);
2768         }
2769
2770         args.in.h.opcode = FUSE_POLL;
2771         args.in.h.nodeid = ff->nodeid;
2772         args.in.numargs = 1;
2773         args.in.args[0].size = sizeof(inarg);
2774         args.in.args[0].value = &inarg;
2775         args.out.numargs = 1;
2776         args.out.args[0].size = sizeof(outarg);
2777         args.out.args[0].value = &outarg;
2778         err = fuse_simple_request(fc, &args);
2779
2780         if (!err)
2781                 return outarg.revents;
2782         if (err == -ENOSYS) {
2783                 fc->no_poll = 1;
2784                 return DEFAULT_POLLMASK;
2785         }
2786         return POLLERR;
2787 }
2788 EXPORT_SYMBOL_GPL(fuse_file_poll);
2789
2790 /*
2791  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2792  * wakes up the poll waiters.
2793  */
2794 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2795                             struct fuse_notify_poll_wakeup_out *outarg)
2796 {
2797         u64 kh = outarg->kh;
2798         struct rb_node **link;
2799
2800         spin_lock(&fc->lock);
2801
2802         link = fuse_find_polled_node(fc, kh, NULL);
2803         if (*link) {
2804                 struct fuse_file *ff;
2805
2806                 ff = rb_entry(*link, struct fuse_file, polled_node);
2807                 wake_up_interruptible_sync(&ff->poll_wait);
2808         }
2809
2810         spin_unlock(&fc->lock);
2811         return 0;
2812 }
2813
2814 static void fuse_do_truncate(struct file *file)
2815 {
2816         struct inode *inode = file->f_mapping->host;
2817         struct iattr attr;
2818
2819         attr.ia_valid = ATTR_SIZE;
2820         attr.ia_size = i_size_read(inode);
2821
2822         attr.ia_file = file;
2823         attr.ia_valid |= ATTR_FILE;
2824
2825         fuse_do_setattr(inode, &attr, file);
2826 }
2827
2828 static inline loff_t fuse_round_up(loff_t off)
2829 {
2830         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2831 }
2832
2833 static ssize_t
2834 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2835 {
2836         DECLARE_COMPLETION_ONSTACK(wait);
2837         ssize_t ret = 0;
2838         struct file *file = iocb->ki_filp;
2839         struct fuse_file *ff = file->private_data;
2840         bool async_dio = ff->fc->async_dio;
2841         loff_t pos = 0;
2842         struct inode *inode;
2843         loff_t i_size;
2844         size_t count = iov_iter_count(iter);
2845         struct fuse_io_priv *io;
2846         bool is_sync = is_sync_kiocb(iocb);
2847
2848         pos = offset;
2849         inode = file->f_mapping->host;
2850         i_size = i_size_read(inode);
2851
2852         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2853                 return 0;
2854
2855         /* optimization for short read */
2856         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2857                 if (offset >= i_size)
2858                         return 0;
2859                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2860                 count = iov_iter_count(iter);
2861         }
2862
2863         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2864         if (!io)
2865                 return -ENOMEM;
2866         spin_lock_init(&io->lock);
2867         kref_init(&io->refcnt);
2868         io->reqs = 1;
2869         io->bytes = -1;
2870         io->size = 0;
2871         io->offset = offset;
2872         io->write = (iov_iter_rw(iter) == WRITE);
2873         io->err = 0;
2874         io->file = file;
2875         /*
2876          * By default, we want to optimize all I/Os with async request
2877          * submission to the client filesystem if supported.
2878          */
2879         io->async = async_dio;
2880         io->iocb = iocb;
2881
2882         /*
2883          * We cannot asynchronously extend the size of a file. We have no method
2884          * to wait on real async I/O requests, so we must submit this request
2885          * synchronously.
2886          */
2887         if (!is_sync && (offset + count > i_size) &&
2888             iov_iter_rw(iter) == WRITE)
2889                 io->async = false;
2890
2891         if (io->async && is_sync) {
2892                 /*
2893                  * Additional reference to keep io around after
2894                  * calling fuse_aio_complete()
2895                  */
2896                 kref_get(&io->refcnt);
2897                 io->done = &wait;
2898         }
2899
2900         if (iov_iter_rw(iter) == WRITE) {
2901                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2902                 fuse_invalidate_attr(inode);
2903         } else {
2904                 ret = __fuse_direct_read(io, iter, &pos);
2905         }
2906
2907         if (io->async) {
2908                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2909
2910                 /* we have a non-extending, async request, so return */
2911                 if (!is_sync)
2912                         return -EIOCBQUEUED;
2913
2914                 wait_for_completion(&wait);
2915                 ret = fuse_get_res_by_io(io);
2916         }
2917
2918         kref_put(&io->refcnt, fuse_io_release);
2919
2920         if (iov_iter_rw(iter) == WRITE) {
2921                 if (ret > 0)
2922                         fuse_write_update_size(inode, pos);
2923                 else if (ret < 0 && offset + count > i_size)
2924                         fuse_do_truncate(file);
2925         }
2926
2927         return ret;
2928 }
2929
2930 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2931                                 loff_t length)
2932 {
2933         struct fuse_file *ff = file->private_data;
2934         struct inode *inode = file_inode(file);
2935         struct fuse_inode *fi = get_fuse_inode(inode);
2936         struct fuse_conn *fc = ff->fc;
2937         FUSE_ARGS(args);
2938         struct fuse_fallocate_in inarg = {
2939                 .fh = ff->fh,
2940                 .offset = offset,
2941                 .length = length,
2942                 .mode = mode
2943         };
2944         int err;
2945         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2946                            (mode & FALLOC_FL_PUNCH_HOLE);
2947
2948         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2949                 return -EOPNOTSUPP;
2950
2951         if (fc->no_fallocate)
2952                 return -EOPNOTSUPP;
2953
2954         if (lock_inode) {
2955                 mutex_lock(&inode->i_mutex);
2956                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2957                         loff_t endbyte = offset + length - 1;
2958                         err = filemap_write_and_wait_range(inode->i_mapping,
2959                                                            offset, endbyte);
2960                         if (err)
2961                                 goto out;
2962
2963                         fuse_sync_writes(inode);
2964                 }
2965         }
2966
2967         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
2968             offset + length > i_size_read(inode)) {
2969                 err = inode_newsize_ok(inode, offset + length);
2970                 if (err)
2971                         goto out;
2972         }
2973
2974         if (!(mode & FALLOC_FL_KEEP_SIZE))
2975                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2976
2977         args.in.h.opcode = FUSE_FALLOCATE;
2978         args.in.h.nodeid = ff->nodeid;
2979         args.in.numargs = 1;
2980         args.in.args[0].size = sizeof(inarg);
2981         args.in.args[0].value = &inarg;
2982         err = fuse_simple_request(fc, &args);
2983         if (err == -ENOSYS) {
2984                 fc->no_fallocate = 1;
2985                 err = -EOPNOTSUPP;
2986         }
2987         if (err)
2988                 goto out;
2989
2990         /* we could have extended the file */
2991         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2992                 bool changed = fuse_write_update_size(inode, offset + length);
2993
2994                 if (changed && fc->writeback_cache)
2995                         file_update_time(file);
2996         }
2997
2998         if (mode & FALLOC_FL_PUNCH_HOLE)
2999                 truncate_pagecache_range(inode, offset, offset + length - 1);
3000
3001         fuse_invalidate_attr(inode);
3002
3003 out:
3004         if (!(mode & FALLOC_FL_KEEP_SIZE))
3005                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3006
3007         if (lock_inode)
3008                 mutex_unlock(&inode->i_mutex);
3009
3010         return err;
3011 }
3012
3013 static const struct file_operations fuse_file_operations = {
3014         .llseek         = fuse_file_llseek,
3015         .read_iter      = fuse_file_read_iter,
3016         .write_iter     = fuse_file_write_iter,
3017         .mmap           = fuse_file_mmap,
3018         .open           = fuse_open,
3019         .flush          = fuse_flush,
3020         .release        = fuse_release,
3021         .fsync          = fuse_fsync,
3022         .lock           = fuse_file_lock,
3023         .flock          = fuse_file_flock,
3024         .splice_read    = generic_file_splice_read,
3025         .unlocked_ioctl = fuse_file_ioctl,
3026         .compat_ioctl   = fuse_file_compat_ioctl,
3027         .poll           = fuse_file_poll,
3028         .fallocate      = fuse_file_fallocate,
3029 };
3030
3031 static const struct file_operations fuse_direct_io_file_operations = {
3032         .llseek         = fuse_file_llseek,
3033         .read_iter      = fuse_direct_read_iter,
3034         .write_iter     = fuse_direct_write_iter,
3035         .mmap           = fuse_direct_mmap,
3036         .open           = fuse_open,
3037         .flush          = fuse_flush,
3038         .release        = fuse_release,
3039         .fsync          = fuse_fsync,
3040         .lock           = fuse_file_lock,
3041         .flock          = fuse_file_flock,
3042         .unlocked_ioctl = fuse_file_ioctl,
3043         .compat_ioctl   = fuse_file_compat_ioctl,
3044         .poll           = fuse_file_poll,
3045         .fallocate      = fuse_file_fallocate,
3046         /* no splice_read */
3047 };
3048
3049 static const struct address_space_operations fuse_file_aops  = {
3050         .readpage       = fuse_readpage,
3051         .writepage      = fuse_writepage,
3052         .writepages     = fuse_writepages,
3053         .launder_page   = fuse_launder_page,
3054         .readpages      = fuse_readpages,
3055         .set_page_dirty = __set_page_dirty_nobuffers,
3056         .bmap           = fuse_bmap,
3057         .direct_IO      = fuse_direct_IO,
3058         .write_begin    = fuse_write_begin,
3059         .write_end      = fuse_write_end,
3060 };
3061
3062 void fuse_init_file_inode(struct inode *inode)
3063 {
3064         inode->i_fop = &fuse_file_operations;
3065         inode->i_data.a_ops = &fuse_file_aops;
3066 }