fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/signal.h>
  16 #include <linux/module.h>
  17 #include <linux/compat.h>
  18 #include <linux/swap.h>
  19 #include <linux/falloc.h>
  20 #include <linux/uio.h>
  21
  22 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  23                           int opcode, struct fuse_open_out *outargp)
  24 {
  25         struct fuse_open_in inarg;
  26         FUSE_ARGS(args);
  27
  28         memset(&inarg, 0, sizeof(inarg));
  29         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  30         if (!fc->atomic_o_trunc)
  31                 inarg.flags &= ~O_TRUNC;
  32         args.opcode = opcode;
  33         args.nodeid = nodeid;
  34         args.in_numargs = 1;
  35         args.in_args[0].size = sizeof(inarg);
  36         args.in_args[0].value = &inarg;
  37         args.out_numargs = 1;
  38         args.out_args[0].size = sizeof(*outargp);
  39         args.out_args[0].value = outargp;
  40
  41         return fuse_simple_request(fc, &args);
  42 }
  43
  44 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  45 {
  46         struct fuse_file *ff;
  47
  48         ff = kzalloc(sizeof(struct fuse_file), GFP_KERNEL);
  49         if (unlikely(!ff))
  50                 return NULL;
  51
  52         ff->fc = fc;
  53         ff->reserved_req = fuse_request_alloc(0);
  54         if (unlikely(!ff->reserved_req)) {
  55                 kfree(ff);
  56                 return NULL;
  57         }
  58
  59         INIT_LIST_HEAD(&ff->write_entry);
  60         mutex_init(&ff->readdir.lock);
  61         refcount_set(&ff->count, 1);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         ff->kh = atomic64_inc_return(&fc->khctr);
  66
  67         return ff;
  68 }
  69
  70 void fuse_file_free(struct fuse_file *ff)
  71 {
  72         fuse_request_free(ff->reserved_req);
  73         mutex_destroy(&ff->readdir.lock);
  74         kfree(ff);
  75 }
  76
  77 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  78 {
  79         refcount_inc(&ff->count);
  80         return ff;
  81 }
  82
  83 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  84 {
  85         iput(req->misc.release.inode);
  86 }
  87
  88 static void fuse_file_put(struct fuse_file *ff, bool sync, bool isdir)
  89 {
  90         if (refcount_dec_and_test(&ff->count)) {
  91                 struct fuse_req *req = ff->reserved_req;
  92
  93                 if (isdir ? ff->fc->no_opendir : ff->fc->no_open) {
  94                         /*
  95                          * Drop the release request when client does not
  96                          * implement 'open'
  97                          */
  98                         __clear_bit(FR_BACKGROUND, &req->flags);
  99                         iput(req->misc.release.inode);
 100                         fuse_put_request(ff->fc, req);
 101                 } else if (sync) {
 102                         __set_bit(FR_FORCE, &req->flags);
 103                         __clear_bit(FR_BACKGROUND, &req->flags);
 104                         fuse_request_send(ff->fc, req);
 105                         iput(req->misc.release.inode);
 106                         fuse_put_request(ff->fc, req);
 107                 } else {
 108                         req->end = fuse_release_end;
 109                         __set_bit(FR_BACKGROUND, &req->flags);
 110                         fuse_request_send_background(ff->fc, req);
 111                 }
 112                 kfree(ff);
 113         }
 114 }
 115
 116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 117                  bool isdir)
 118 {
 119         struct fuse_file *ff;
 120         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 121
 122         ff = fuse_file_alloc(fc);
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         ff->fh = 0;
 127         /* Default for no-open */
 128         ff->open_flags = FOPEN_KEEP_CACHE | (isdir ? FOPEN_CACHE_DIR : 0);
 129         if (isdir ? !fc->no_opendir : !fc->no_open) {
 130                 struct fuse_open_out outarg;
 131                 int err;
 132
 133                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 134                 if (!err) {
 135                         ff->fh = outarg.fh;
 136                         ff->open_flags = outarg.open_flags;
 137
 138                 } else if (err != -ENOSYS) {
 139                         fuse_file_free(ff);
 140                         return err;
 141                 } else {
 142                         if (isdir)
 143                                 fc->no_opendir = 1;
 144                         else
 145                                 fc->no_open = 1;
 146                 }
 147         }
 148
 149         if (isdir)
 150                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 151
 152         ff->nodeid = nodeid;
 153         file->private_data = ff;
 154
 155         return 0;
 156 }
 157 EXPORT_SYMBOL_GPL(fuse_do_open);
 158
 159 static void fuse_link_write_file(struct file *file)
 160 {
 161         struct inode *inode = file_inode(file);
 162         struct fuse_inode *fi = get_fuse_inode(inode);
 163         struct fuse_file *ff = file->private_data;
 164         /*
 165          * file may be written through mmap, so chain it onto the
 166          * inodes's write_file list
 167          */
 168         spin_lock(&fi->lock);
 169         if (list_empty(&ff->write_entry))
 170                 list_add(&ff->write_entry, &fi->write_files);
 171         spin_unlock(&fi->lock);
 172 }
 173
 174 void fuse_finish_open(struct inode *inode, struct file *file)
 175 {
 176         struct fuse_file *ff = file->private_data;
 177         struct fuse_conn *fc = get_fuse_conn(inode);
 178
 179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 180                 invalidate_inode_pages2(inode->i_mapping);
 181         if (ff->open_flags & FOPEN_STREAM)
 182                 stream_open(inode, file);
 183         else if (ff->open_flags & FOPEN_NONSEEKABLE)
 184                 nonseekable_open(inode, file);
 185         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 186                 struct fuse_inode *fi = get_fuse_inode(inode);
 187
 188                 spin_lock(&fi->lock);
 189                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 190                 i_size_write(inode, 0);
 191                 spin_unlock(&fi->lock);
 192                 fuse_invalidate_attr(inode);
 193                 if (fc->writeback_cache)
 194                         file_update_time(file);
 195         }
 196         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 197                 fuse_link_write_file(file);
 198 }
 199
 200 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 201 {
 202         struct fuse_conn *fc = get_fuse_conn(inode);
 203         int err;
 204         bool lock_inode = (file->f_flags & O_TRUNC) &&
 205                           fc->atomic_o_trunc &&
 206                           fc->writeback_cache;
 207
 208         err = generic_file_open(inode, file);
 209         if (err)
 210                 return err;
 211
 212         if (lock_inode)
 213                 inode_lock(inode);
 214
 215         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 216
 217         if (!err)
 218                 fuse_finish_open(inode, file);
 219
 220         if (lock_inode)
 221                 inode_unlock(inode);
 222
 223         return err;
 224 }
 225
 226 static void fuse_prepare_release(struct fuse_inode *fi, struct fuse_file *ff,
 227                                  int flags, int opcode)
 228 {
 229         struct fuse_conn *fc = ff->fc;
 230         struct fuse_req *req = ff->reserved_req;
 231         struct fuse_release_in *inarg = &req->misc.release.in;
 232
 233         /* Inode is NULL on error path of fuse_create_open() */
 234         if (likely(fi)) {
 235                 spin_lock(&fi->lock);
 236                 list_del(&ff->write_entry);
 237                 spin_unlock(&fi->lock);
 238         }
 239         spin_lock(&fc->lock);
 240         if (!RB_EMPTY_NODE(&ff->polled_node))
 241                 rb_erase(&ff->polled_node, &fc->polled_files);
 242         spin_unlock(&fc->lock);
 243
 244         wake_up_interruptible_all(&ff->poll_wait);
 245
 246         inarg->fh = ff->fh;
 247         inarg->flags = flags;
 248         req->in.h.opcode = opcode;
 249         req->in.h.nodeid = ff->nodeid;
 250         req->in.numargs = 1;
 251         req->in.args[0].size = sizeof(struct fuse_release_in);
 252         req->in.args[0].value = inarg;
 253 }
 254
 255 void fuse_release_common(struct file *file, bool isdir)
 256 {
 257         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 258         struct fuse_file *ff = file->private_data;
 259         struct fuse_req *req = ff->reserved_req;
 260         int opcode = isdir ? FUSE_RELEASEDIR : FUSE_RELEASE;
 261
 262         fuse_prepare_release(fi, ff, file->f_flags, opcode);
 263
 264         if (ff->flock) {
 265                 struct fuse_release_in *inarg = &req->misc.release.in;
 266                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 267                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 268                                                        (fl_owner_t) file);
 269         }
 270         /* Hold inode until release is finished */
 271         req->misc.release.inode = igrab(file_inode(file));
 272
 273         /*
 274          * Normally this will send the RELEASE request, however if
 275          * some asynchronous READ or WRITE requests are outstanding,
 276          * the sending will be delayed.
 277          *
 278          * Make the release synchronous if this is a fuseblk mount,
 279          * synchronous RELEASE is allowed (and desirable) in this case
 280          * because the server can be trusted not to screw up.
 281          */
 282         fuse_file_put(ff, ff->fc->destroy_req != NULL, isdir);
 283 }
 284
 285 static int fuse_open(struct inode *inode, struct file *file)
 286 {
 287         return fuse_open_common(inode, file, false);
 288 }
 289
 290 static int fuse_release(struct inode *inode, struct file *file)
 291 {
 292         struct fuse_conn *fc = get_fuse_conn(inode);
 293
 294         /* see fuse_vma_close() for !writeback_cache case */
 295         if (fc->writeback_cache)
 296                 write_inode_now(inode, 1);
 297
 298         fuse_release_common(file, false);
 299
 300         /* return value is ignored by VFS */
 301         return 0;
 302 }
 303
 304 void fuse_sync_release(struct fuse_inode *fi, struct fuse_file *ff, int flags)
 305 {
 306         WARN_ON(refcount_read(&ff->count) > 1);
 307         fuse_prepare_release(fi, ff, flags, FUSE_RELEASE);
 308         /*
 309          * iput(NULL) is a no-op and since the refcount is 1 and everything's
 310          * synchronous, we are fine with not doing igrab() here"
 311          */
 312         fuse_file_put(ff, true, false);
 313 }
 314 EXPORT_SYMBOL_GPL(fuse_sync_release);
 315
 316 /*
 317  * Scramble the ID space with XTEA, so that the value of the files_struct
 318  * pointer is not exposed to userspace.
 319  */
 320 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 321 {
 322         u32 *k = fc->scramble_key;
 323         u64 v = (unsigned long) id;
 324         u32 v0 = v;
 325         u32 v1 = v >> 32;
 326         u32 sum = 0;
 327         int i;
 328
 329         for (i = 0; i < 32; i++) {
 330                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 331                 sum += 0x9E3779B9;
 332                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 333         }
 334
 335         return (u64) v0 + ((u64) v1 << 32);
 336 }
 337
 338 static struct fuse_req *fuse_find_writeback(struct fuse_inode *fi,
 339                                             pgoff_t idx_from, pgoff_t idx_to)
 340 {
 341         struct fuse_req *req;
 342
 343         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 344                 pgoff_t curr_index;
 345
 346                 WARN_ON(get_fuse_inode(req->inode) != fi);
 347                 curr_index = req->misc.write.in.offset >> PAGE_SHIFT;
 348                 if (idx_from < curr_index + req->num_pages &&
 349                     curr_index <= idx_to) {
 350                         return req;
 351                 }
 352         }
 353         return NULL;
 354 }
 355
 356 /*
 357  * Check if any page in a range is under writeback
 358  *
 359  * This is currently done by walking the list of writepage requests
 360  * for the inode, which can be pretty inefficient.
 361  */
 362 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 363                                    pgoff_t idx_to)
 364 {
 365         struct fuse_inode *fi = get_fuse_inode(inode);
 366         bool found;
 367
 368         spin_lock(&fi->lock);
 369         found = fuse_find_writeback(fi, idx_from, idx_to);
 370         spin_unlock(&fi->lock);
 371
 372         return found;
 373 }
 374
 375 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 376 {
 377         return fuse_range_is_writeback(inode, index, index);
 378 }
 379
 380 /*
 381  * Wait for page writeback to be completed.
 382  *
 383  * Since fuse doesn't rely on the VM writeback tracking, this has to
 384  * use some other means.
 385  */
 386 static void fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 387 {
 388         struct fuse_inode *fi = get_fuse_inode(inode);
 389
 390         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 391 }
 392
 393 /*
 394  * Wait for all pending writepages on the inode to finish.
 395  *
 396  * This is currently done by blocking further writes with FUSE_NOWRITE
 397  * and waiting for all sent writes to complete.
 398  *
 399  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 400  * could conflict with truncation.
 401  */
 402 static void fuse_sync_writes(struct inode *inode)
 403 {
 404         fuse_set_nowrite(inode);
 405         fuse_release_nowrite(inode);
 406 }
 407
 408 static int fuse_flush(struct file *file, fl_owner_t id)
 409 {
 410         struct inode *inode = file_inode(file);
 411         struct fuse_conn *fc = get_fuse_conn(inode);
 412         struct fuse_file *ff = file->private_data;
 413         struct fuse_req *req;
 414         struct fuse_flush_in inarg;
 415         int err;
 416
 417         if (is_bad_inode(inode))
 418                 return -EIO;
 419
 420         if (fc->no_flush)
 421                 return 0;
 422
 423         err = write_inode_now(inode, 1);
 424         if (err)
 425                 return err;
 426
 427         inode_lock(inode);
 428         fuse_sync_writes(inode);
 429         inode_unlock(inode);
 430
 431         err = filemap_check_errors(file->f_mapping);
 432         if (err)
 433                 return err;
 434
 435         req = fuse_get_req_nofail_nopages(fc);
 436         memset(&inarg, 0, sizeof(inarg));
 437         inarg.fh = ff->fh;
 438         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 439         req->in.h.opcode = FUSE_FLUSH;
 440         req->in.h.nodeid = get_node_id(inode);
 441         req->in.numargs = 1;
 442         req->in.args[0].size = sizeof(inarg);
 443         req->in.args[0].value = &inarg;
 444         __set_bit(FR_FORCE, &req->flags);
 445         fuse_request_send(fc, req);
 446         err = req->out.h.error;
 447         fuse_put_request(fc, req);
 448         if (err == -ENOSYS) {
 449                 fc->no_flush = 1;
 450                 err = 0;
 451         }
 452         return err;
 453 }
 454
 455 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 456                       int datasync, int opcode)
 457 {
 458         struct inode *inode = file->f_mapping->host;
 459         struct fuse_conn *fc = get_fuse_conn(inode);
 460         struct fuse_file *ff = file->private_data;
 461         FUSE_ARGS(args);
 462         struct fuse_fsync_in inarg;
 463
 464         memset(&inarg, 0, sizeof(inarg));
 465         inarg.fh = ff->fh;
 466         inarg.fsync_flags = datasync ? FUSE_FSYNC_FDATASYNC : 0;
 467         args.opcode = opcode;
 468         args.nodeid = get_node_id(inode);
 469         args.in_numargs = 1;
 470         args.in_args[0].size = sizeof(inarg);
 471         args.in_args[0].value = &inarg;
 472         return fuse_simple_request(fc, &args);
 473 }
 474
 475 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 476                       int datasync)
 477 {
 478         struct inode *inode = file->f_mapping->host;
 479         struct fuse_conn *fc = get_fuse_conn(inode);
 480         int err;
 481
 482         if (is_bad_inode(inode))
 483                 return -EIO;
 484
 485         inode_lock(inode);
 486
 487         /*
 488          * Start writeback against all dirty pages of the inode, then
 489          * wait for all outstanding writes, before sending the FSYNC
 490          * request.
 491          */
 492         err = file_write_and_wait_range(file, start, end);
 493         if (err)
 494                 goto out;
 495
 496         fuse_sync_writes(inode);
 497
 498         /*
 499          * Due to implementation of fuse writeback
 500          * file_write_and_wait_range() does not catch errors.
 501          * We have to do this directly after fuse_sync_writes()
 502          */
 503         err = file_check_and_advance_wb_err(file);
 504         if (err)
 505                 goto out;
 506
 507         err = sync_inode_metadata(inode, 1);
 508         if (err)
 509                 goto out;
 510
 511         if (fc->no_fsync)
 512                 goto out;
 513
 514         err = fuse_fsync_common(file, start, end, datasync, FUSE_FSYNC);
 515         if (err == -ENOSYS) {
 516                 fc->no_fsync = 1;
 517                 err = 0;
 518         }
 519 out:
 520         inode_unlock(inode);
 521
 522         return err;
 523 }
 524
 525 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 526                     size_t count, int opcode)
 527 {
 528         struct fuse_read_in *inarg = &req->misc.read.in;
 529         struct fuse_file *ff = file->private_data;
 530
 531         inarg->fh = ff->fh;
 532         inarg->offset = pos;
 533         inarg->size = count;
 534         inarg->flags = file->f_flags;
 535         req->in.h.opcode = opcode;
 536         req->in.h.nodeid = ff->nodeid;
 537         req->in.numargs = 1;
 538         req->in.args[0].size = sizeof(struct fuse_read_in);
 539         req->in.args[0].value = inarg;
 540         req->out.argvar = 1;
 541         req->out.numargs = 1;
 542         req->out.args[0].size = count;
 543 }
 544
 545 static void fuse_release_user_pages(struct fuse_req *req, bool should_dirty)
 546 {
 547         unsigned i;
 548
 549         for (i = 0; i < req->num_pages; i++) {
 550                 struct page *page = req->pages[i];
 551                 if (should_dirty)
 552                         set_page_dirty_lock(page);
 553                 put_page(page);
 554         }
 555 }
 556
 557 static void fuse_io_release(struct kref *kref)
 558 {
 559         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 560 }
 561
 562 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 563 {
 564         if (io->err)
 565                 return io->err;
 566
 567         if (io->bytes >= 0 && io->write)
 568                 return -EIO;
 569
 570         return io->bytes < 0 ? io->size : io->bytes;
 571 }
 572
 573 /**
 574  * In case of short read, the caller sets 'pos' to the position of
 575  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 576  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 577  *
 578  * An example:
 579  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 580  * both submitted asynchronously. The first of them was ACKed by userspace as
 581  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 582  * second request was ACKed as short, e.g. only 1K was read, resulting in
 583  * pos == 33K.
 584  *
 585  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 586  * will be equal to the length of the longest contiguous fragment of
 587  * transferred data starting from the beginning of IO request.
 588  */
 589 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 590 {
 591         int left;
 592
 593         spin_lock(&io->lock);
 594         if (err)
 595                 io->err = io->err ? : err;
 596         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 597                 io->bytes = pos;
 598
 599         left = --io->reqs;
 600         if (!left && io->blocking)
 601                 complete(io->done);
 602         spin_unlock(&io->lock);
 603
 604         if (!left && !io->blocking) {
 605                 ssize_t res = fuse_get_res_by_io(io);
 606
 607                 if (res >= 0) {
 608                         struct inode *inode = file_inode(io->iocb->ki_filp);
 609                         struct fuse_conn *fc = get_fuse_conn(inode);
 610                         struct fuse_inode *fi = get_fuse_inode(inode);
 611
 612                         spin_lock(&fi->lock);
 613                         fi->attr_version = atomic64_inc_return(&fc->attr_version);
 614                         spin_unlock(&fi->lock);
 615                 }
 616
 617                 io->iocb->ki_complete(io->iocb, res, 0);
 618         }
 619
 620         kref_put(&io->refcnt, fuse_io_release);
 621 }
 622
 623 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 624 {
 625         struct fuse_io_priv *io = req->io;
 626         ssize_t pos = -1;
 627
 628         fuse_release_user_pages(req, io->should_dirty);
 629
 630         if (io->write) {
 631                 if (req->misc.write.in.size != req->misc.write.out.size)
 632                         pos = req->misc.write.in.offset - io->offset +
 633                                 req->misc.write.out.size;
 634         } else {
 635                 if (req->misc.read.in.size != req->out.args[0].size)
 636                         pos = req->misc.read.in.offset - io->offset +
 637                                 req->out.args[0].size;
 638         }
 639
 640         fuse_aio_complete(io, req->out.h.error, pos);
 641 }
 642
 643 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 644                 size_t num_bytes, struct fuse_io_priv *io)
 645 {
 646         spin_lock(&io->lock);
 647         kref_get(&io->refcnt);
 648         io->size += num_bytes;
 649         io->reqs++;
 650         spin_unlock(&io->lock);
 651
 652         req->io = io;
 653         req->end = fuse_aio_complete_req;
 654
 655         __fuse_get_request(req);
 656         fuse_request_send_background(fc, req);
 657
 658         return num_bytes;
 659 }
 660
 661 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 662                              loff_t pos, size_t count, fl_owner_t owner)
 663 {
 664         struct file *file = io->iocb->ki_filp;
 665         struct fuse_file *ff = file->private_data;
 666         struct fuse_conn *fc = ff->fc;
 667
 668         fuse_read_fill(req, file, pos, count, FUSE_READ);
 669         if (owner != NULL) {
 670                 struct fuse_read_in *inarg = &req->misc.read.in;
 671
 672                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 673                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 674         }
 675
 676         if (io->async)
 677                 return fuse_async_req_send(fc, req, count, io);
 678
 679         fuse_request_send(fc, req);
 680         return req->out.args[0].size;
 681 }
 682
 683 static void fuse_read_update_size(struct inode *inode, loff_t size,
 684                                   u64 attr_ver)
 685 {
 686         struct fuse_conn *fc = get_fuse_conn(inode);
 687         struct fuse_inode *fi = get_fuse_inode(inode);
 688
 689         spin_lock(&fi->lock);
 690         if (attr_ver == fi->attr_version && size < inode->i_size &&
 691             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 692                 fi->attr_version = atomic64_inc_return(&fc->attr_version);
 693                 i_size_write(inode, size);
 694         }
 695         spin_unlock(&fi->lock);
 696 }
 697
 698 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 699                             u64 attr_ver)
 700 {
 701         size_t num_read = req->out.args[0].size;
 702         struct fuse_conn *fc = get_fuse_conn(inode);
 703
 704         if (fc->writeback_cache) {
 705                 /*
 706                  * A hole in a file. Some data after the hole are in page cache,
 707                  * but have not reached the client fs yet. So, the hole is not
 708                  * present there.
 709                  */
 710                 int i;
 711                 int start_idx = num_read >> PAGE_SHIFT;
 712                 size_t off = num_read & (PAGE_SIZE - 1);
 713
 714                 for (i = start_idx; i < req->num_pages; i++) {
 715                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 716                         off = 0;
 717                 }
 718         } else {
 719                 loff_t pos = page_offset(req->pages[0]) + num_read;
 720                 fuse_read_update_size(inode, pos, attr_ver);
 721         }
 722 }
 723
 724 static int fuse_do_readpage(struct file *file, struct page *page)
 725 {
 726         struct kiocb iocb;
 727         struct fuse_io_priv io;
 728         struct inode *inode = page->mapping->host;
 729         struct fuse_conn *fc = get_fuse_conn(inode);
 730         struct fuse_req *req;
 731         size_t num_read;
 732         loff_t pos = page_offset(page);
 733         size_t count = PAGE_SIZE;
 734         u64 attr_ver;
 735         int err;
 736
 737         /*
 738          * Page writeback can extend beyond the lifetime of the
 739          * page-cache page, so make sure we read a properly synced
 740          * page.
 741          */
 742         fuse_wait_on_page_writeback(inode, page->index);
 743
 744         req = fuse_get_req(fc, 1);
 745         if (IS_ERR(req))
 746                 return PTR_ERR(req);
 747
 748         attr_ver = fuse_get_attr_version(fc);
 749
 750         req->out.page_zeroing = 1;
 751         req->out.argpages = 1;
 752         req->num_pages = 1;
 753         req->pages[0] = page;
 754         req->page_descs[0].length = count;
 755         init_sync_kiocb(&iocb, file);
 756         io = (struct fuse_io_priv) FUSE_IO_PRIV_SYNC(&iocb);
 757         num_read = fuse_send_read(req, &io, pos, count, NULL);
 758         err = req->out.h.error;
 759
 760         if (!err) {
 761                 /*
 762                  * Short read means EOF.  If file size is larger, truncate it
 763                  */
 764                 if (num_read < count)
 765                         fuse_short_read(req, inode, attr_ver);
 766
 767                 SetPageUptodate(page);
 768         }
 769
 770         fuse_put_request(fc, req);
 771
 772         return err;
 773 }
 774
 775 static int fuse_readpage(struct file *file, struct page *page)
 776 {
 777         struct inode *inode = page->mapping->host;
 778         int err;
 779
 780         err = -EIO;
 781         if (is_bad_inode(inode))
 782                 goto out;
 783
 784         err = fuse_do_readpage(file, page);
 785         fuse_invalidate_atime(inode);
 786  out:
 787         unlock_page(page);
 788         return err;
 789 }
 790
 791 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 792 {
 793         int i;
 794         size_t count = req->misc.read.in.size;
 795         size_t num_read = req->out.args[0].size;
 796         struct address_space *mapping = NULL;
 797
 798         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 799                 mapping = req->pages[i]->mapping;
 800
 801         if (mapping) {
 802                 struct inode *inode = mapping->host;
 803
 804                 /*
 805                  * Short read means EOF. If file size is larger, truncate it
 806                  */
 807                 if (!req->out.h.error && num_read < count)
 808                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 809
 810                 fuse_invalidate_atime(inode);
 811         }
 812
 813         for (i = 0; i < req->num_pages; i++) {
 814                 struct page *page = req->pages[i];
 815                 if (!req->out.h.error)
 816                         SetPageUptodate(page);
 817                 else
 818                         SetPageError(page);
 819                 unlock_page(page);
 820                 put_page(page);
 821         }
 822         if (req->ff)
 823                 fuse_file_put(req->ff, false, false);
 824 }
 825
 826 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 827 {
 828         struct fuse_file *ff = file->private_data;
 829         struct fuse_conn *fc = ff->fc;
 830         loff_t pos = page_offset(req->pages[0]);
 831         size_t count = req->num_pages << PAGE_SHIFT;
 832
 833         req->out.argpages = 1;
 834         req->out.page_zeroing = 1;
 835         req->out.page_replace = 1;
 836         fuse_read_fill(req, file, pos, count, FUSE_READ);
 837         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 838         if (fc->async_read) {
 839                 req->ff = fuse_file_get(ff);
 840                 req->end = fuse_readpages_end;
 841                 fuse_request_send_background(fc, req);
 842         } else {
 843                 fuse_request_send(fc, req);
 844                 fuse_readpages_end(fc, req);
 845                 fuse_put_request(fc, req);
 846         }
 847 }
 848
 849 struct fuse_fill_data {
 850         struct fuse_req *req;
 851         struct file *file;
 852         struct inode *inode;
 853         unsigned nr_pages;
 854 };
 855
 856 static int fuse_readpages_fill(void *_data, struct page *page)
 857 {
 858         struct fuse_fill_data *data = _data;
 859         struct fuse_req *req = data->req;
 860         struct inode *inode = data->inode;
 861         struct fuse_conn *fc = get_fuse_conn(inode);
 862
 863         fuse_wait_on_page_writeback(inode, page->index);
 864
 865         if (req->num_pages &&
 866             (req->num_pages == fc->max_pages ||
 867              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 868              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 869                 unsigned int nr_alloc = min_t(unsigned int, data->nr_pages,
 870                                               fc->max_pages);
 871                 fuse_send_readpages(req, data->file);
 872                 if (fc->async_read)
 873                         req = fuse_get_req_for_background(fc, nr_alloc);
 874                 else
 875                         req = fuse_get_req(fc, nr_alloc);
 876
 877                 data->req = req;
 878                 if (IS_ERR(req)) {
 879                         unlock_page(page);
 880                         return PTR_ERR(req);
 881                 }
 882         }
 883
 884         if (WARN_ON(req->num_pages >= req->max_pages)) {
 885                 unlock_page(page);
 886                 fuse_put_request(fc, req);
 887                 return -EIO;
 888         }
 889
 890         get_page(page);
 891         req->pages[req->num_pages] = page;
 892         req->page_descs[req->num_pages].length = PAGE_SIZE;
 893         req->num_pages++;
 894         data->nr_pages--;
 895         return 0;
 896 }
 897
 898 static int fuse_readpages(struct file *file, struct address_space *mapping,
 899                           struct list_head *pages, unsigned nr_pages)
 900 {
 901         struct inode *inode = mapping->host;
 902         struct fuse_conn *fc = get_fuse_conn(inode);
 903         struct fuse_fill_data data;
 904         int err;
 905         unsigned int nr_alloc = min_t(unsigned int, nr_pages, fc->max_pages);
 906
 907         err = -EIO;
 908         if (is_bad_inode(inode))
 909                 goto out;
 910
 911         data.file = file;
 912         data.inode = inode;
 913         if (fc->async_read)
 914                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 915         else
 916                 data.req = fuse_get_req(fc, nr_alloc);
 917         data.nr_pages = nr_pages;
 918         err = PTR_ERR(data.req);
 919         if (IS_ERR(data.req))
 920                 goto out;
 921
 922         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 923         if (!err) {
 924                 if (data.req->num_pages)
 925                         fuse_send_readpages(data.req, file);
 926                 else
 927                         fuse_put_request(fc, data.req);
 928         }
 929 out:
 930         return err;
 931 }
 932
 933 static ssize_t fuse_cache_read_iter(struct kiocb *iocb, struct iov_iter *to)
 934 {
 935         struct inode *inode = iocb->ki_filp->f_mapping->host;
 936         struct fuse_conn *fc = get_fuse_conn(inode);
 937
 938         /*
 939          * In auto invalidate mode, always update attributes on read.
 940          * Otherwise, only update if we attempt to read past EOF (to ensure
 941          * i_size is up to date).
 942          */
 943         if (fc->auto_inval_data ||
 944             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 945                 int err;
 946                 err = fuse_update_attributes(inode, iocb->ki_filp);
 947                 if (err)
 948                         return err;
 949         }
 950
 951         return generic_file_read_iter(iocb, to);
 952 }
 953
 954 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 955                             loff_t pos, size_t count)
 956 {
 957         struct fuse_write_in *inarg = &req->misc.write.in;
 958         struct fuse_write_out *outarg = &req->misc.write.out;
 959
 960         inarg->fh = ff->fh;
 961         inarg->offset = pos;
 962         inarg->size = count;
 963         req->in.h.opcode = FUSE_WRITE;
 964         req->in.h.nodeid = ff->nodeid;
 965         req->in.numargs = 2;
 966         if (ff->fc->minor < 9)
 967                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 968         else
 969                 req->in.args[0].size = sizeof(struct fuse_write_in);
 970         req->in.args[0].value = inarg;
 971         req->in.args[1].size = count;
 972         req->out.numargs = 1;
 973         req->out.args[0].size = sizeof(struct fuse_write_out);
 974         req->out.args[0].value = outarg;
 975 }
 976
 977 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 978                               loff_t pos, size_t count, fl_owner_t owner)
 979 {
 980         struct kiocb *iocb = io->iocb;
 981         struct file *file = iocb->ki_filp;
 982         struct fuse_file *ff = file->private_data;
 983         struct fuse_conn *fc = ff->fc;
 984         struct fuse_write_in *inarg = &req->misc.write.in;
 985
 986         fuse_write_fill(req, ff, pos, count);
 987         inarg->flags = file->f_flags;
 988         if (iocb->ki_flags & IOCB_DSYNC)
 989                 inarg->flags |= O_DSYNC;
 990         if (iocb->ki_flags & IOCB_SYNC)
 991                 inarg->flags |= O_SYNC;
 992         if (owner != NULL) {
 993                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 994                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 995         }
 996
 997         if (io->async)
 998                 return fuse_async_req_send(fc, req, count, io);
 999
1000         fuse_request_send(fc, req);
1001         return req->misc.write.out.size;
1002 }
1003
1004 bool fuse_write_update_size(struct inode *inode, loff_t pos)
1005 {
1006         struct fuse_conn *fc = get_fuse_conn(inode);
1007         struct fuse_inode *fi = get_fuse_inode(inode);
1008         bool ret = false;
1009
1010         spin_lock(&fi->lock);
1011         fi->attr_version = atomic64_inc_return(&fc->attr_version);
1012         if (pos > inode->i_size) {
1013                 i_size_write(inode, pos);
1014                 ret = true;
1015         }
1016         spin_unlock(&fi->lock);
1017
1018         return ret;
1019 }
1020
1021 static size_t fuse_send_write_pages(struct fuse_req *req, struct kiocb *iocb,
1022                                     struct inode *inode, loff_t pos,
1023                                     size_t count)
1024 {
1025         size_t res;
1026         unsigned offset;
1027         unsigned i;
1028         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1029
1030         for (i = 0; i < req->num_pages; i++)
1031                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1032
1033         res = fuse_send_write(req, &io, pos, count, NULL);
1034
1035         offset = req->page_descs[0].offset;
1036         count = res;
1037         for (i = 0; i < req->num_pages; i++) {
1038                 struct page *page = req->pages[i];
1039
1040                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1041                         SetPageUptodate(page);
1042
1043                 if (count > PAGE_SIZE - offset)
1044                         count -= PAGE_SIZE - offset;
1045                 else
1046                         count = 0;
1047                 offset = 0;
1048
1049                 unlock_page(page);
1050                 put_page(page);
1051         }
1052
1053         return res;
1054 }
1055
1056 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1057                                struct address_space *mapping,
1058                                struct iov_iter *ii, loff_t pos)
1059 {
1060         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1061         unsigned offset = pos & (PAGE_SIZE - 1);
1062         size_t count = 0;
1063         int err;
1064
1065         req->in.argpages = 1;
1066         req->page_descs[0].offset = offset;
1067
1068         do {
1069                 size_t tmp;
1070                 struct page *page;
1071                 pgoff_t index = pos >> PAGE_SHIFT;
1072                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1073                                      iov_iter_count(ii));
1074
1075                 bytes = min_t(size_t, bytes, fc->max_write - count);
1076
1077  again:
1078                 err = -EFAULT;
1079                 if (iov_iter_fault_in_readable(ii, bytes))
1080                         break;
1081
1082                 err = -ENOMEM;
1083                 page = grab_cache_page_write_begin(mapping, index, 0);
1084                 if (!page)
1085                         break;
1086
1087                 if (mapping_writably_mapped(mapping))
1088                         flush_dcache_page(page);
1089
1090                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1091                 flush_dcache_page(page);
1092
1093                 iov_iter_advance(ii, tmp);
1094                 if (!tmp) {
1095                         unlock_page(page);
1096                         put_page(page);
1097                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1098                         goto again;
1099                 }
1100
1101                 err = 0;
1102                 req->pages[req->num_pages] = page;
1103                 req->page_descs[req->num_pages].length = tmp;
1104                 req->num_pages++;
1105
1106                 count += tmp;
1107                 pos += tmp;
1108                 offset += tmp;
1109                 if (offset == PAGE_SIZE)
1110                         offset = 0;
1111
1112                 if (!fc->big_writes)
1113                         break;
1114         } while (iov_iter_count(ii) && count < fc->max_write &&
1115                  req->num_pages < req->max_pages && offset == 0);
1116
1117         return count > 0 ? count : err;
1118 }
1119
1120 static inline unsigned int fuse_wr_pages(loff_t pos, size_t len,
1121                                      unsigned int max_pages)
1122 {
1123         return min_t(unsigned int,
1124                      ((pos + len - 1) >> PAGE_SHIFT) -
1125                      (pos >> PAGE_SHIFT) + 1,
1126                      max_pages);
1127 }
1128
1129 static ssize_t fuse_perform_write(struct kiocb *iocb,
1130                                   struct address_space *mapping,
1131                                   struct iov_iter *ii, loff_t pos)
1132 {
1133         struct inode *inode = mapping->host;
1134         struct fuse_conn *fc = get_fuse_conn(inode);
1135         struct fuse_inode *fi = get_fuse_inode(inode);
1136         int err = 0;
1137         ssize_t res = 0;
1138
1139         if (inode->i_size < pos + iov_iter_count(ii))
1140                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1141
1142         do {
1143                 struct fuse_req *req;
1144                 ssize_t count;
1145                 unsigned int nr_pages = fuse_wr_pages(pos, iov_iter_count(ii),
1146                                                       fc->max_pages);
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, iocb, 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_cache_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, file);
1197                 if (err)
1198                         return err;
1199
1200                 return generic_file_write_iter(iocb, from);
1201         }
1202
1203         inode_lock(inode);
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);
1223                 if (written < 0 || !iov_iter_count(from))
1224                         goto out;
1225
1226                 pos += written;
1227
1228                 written_buffered = fuse_perform_write(iocb, 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_SHIFT,
1242                                          endbyte >> PAGE_SHIFT);
1243
1244                 written += written_buffered;
1245                 iocb->ki_pos = pos + written_buffered;
1246         } else {
1247                 written = fuse_perform_write(iocb, mapping, from, iocb->ki_pos);
1248                 if (written >= 0)
1249                         iocb->ki_pos += written;
1250         }
1251 out:
1252         current->backing_dev_info = NULL;
1253         inode_unlock(inode);
1254         if (written > 0)
1255                 written = generic_write_sync(iocb, written);
1256
1257         return written ? written : err;
1258 }
1259
1260 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1261                 unsigned index, unsigned nr_pages)
1262 {
1263         int i;
1264
1265         for (i = index; i < index + nr_pages; i++)
1266                 req->page_descs[i].length = PAGE_SIZE -
1267                         req->page_descs[i].offset;
1268 }
1269
1270 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1271 {
1272         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1273 }
1274
1275 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1276                                         size_t max_size)
1277 {
1278         return min(iov_iter_single_seg_count(ii), max_size);
1279 }
1280
1281 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1282                                size_t *nbytesp, int write)
1283 {
1284         size_t nbytes = 0;  /* # bytes already packed in req */
1285         ssize_t ret = 0;
1286
1287         /* Special case for kernel I/O: can copy directly into the buffer */
1288         if (iov_iter_is_kvec(ii)) {
1289                 unsigned long user_addr = fuse_get_user_addr(ii);
1290                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1291
1292                 if (write)
1293                         req->in.args[1].value = (void *) user_addr;
1294                 else
1295                         req->out.args[0].value = (void *) user_addr;
1296
1297                 iov_iter_advance(ii, frag_size);
1298                 *nbytesp = frag_size;
1299                 return 0;
1300         }
1301
1302         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1303                 unsigned npages;
1304                 size_t start;
1305                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1306                                         *nbytesp - nbytes,
1307                                         req->max_pages - req->num_pages,
1308                                         &start);
1309                 if (ret < 0)
1310                         break;
1311
1312                 iov_iter_advance(ii, ret);
1313                 nbytes += ret;
1314
1315                 ret += start;
1316                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1317
1318                 req->page_descs[req->num_pages].offset = start;
1319                 fuse_page_descs_length_init(req, req->num_pages, npages);
1320
1321                 req->num_pages += npages;
1322                 req->page_descs[req->num_pages - 1].length -=
1323                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1324         }
1325
1326         if (write)
1327                 req->in.argpages = 1;
1328         else
1329                 req->out.argpages = 1;
1330
1331         *nbytesp = nbytes;
1332
1333         return ret < 0 ? ret : 0;
1334 }
1335
1336 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1337                        loff_t *ppos, int flags)
1338 {
1339         int write = flags & FUSE_DIO_WRITE;
1340         int cuse = flags & FUSE_DIO_CUSE;
1341         struct file *file = io->iocb->ki_filp;
1342         struct inode *inode = file->f_mapping->host;
1343         struct fuse_file *ff = file->private_data;
1344         struct fuse_conn *fc = ff->fc;
1345         size_t nmax = write ? fc->max_write : fc->max_read;
1346         loff_t pos = *ppos;
1347         size_t count = iov_iter_count(iter);
1348         pgoff_t idx_from = pos >> PAGE_SHIFT;
1349         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1350         ssize_t res = 0;
1351         struct fuse_req *req;
1352         int err = 0;
1353
1354         if (io->async)
1355                 req = fuse_get_req_for_background(fc, iov_iter_npages(iter,
1356                                                                 fc->max_pages));
1357         else
1358                 req = fuse_get_req(fc, iov_iter_npages(iter, fc->max_pages));
1359         if (IS_ERR(req))
1360                 return PTR_ERR(req);
1361
1362         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1363                 if (!write)
1364                         inode_lock(inode);
1365                 fuse_sync_writes(inode);
1366                 if (!write)
1367                         inode_unlock(inode);
1368         }
1369
1370         io->should_dirty = !write && iter_is_iovec(iter);
1371         while (count) {
1372                 size_t nres;
1373                 fl_owner_t owner = current->files;
1374                 size_t nbytes = min(count, nmax);
1375                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1376                 if (err && !nbytes)
1377                         break;
1378
1379                 if (write) {
1380                         if (!capable(CAP_FSETID)) {
1381                                 struct fuse_write_in *inarg;
1382
1383                                 inarg = &req->misc.write.in;
1384                                 inarg->write_flags |= FUSE_WRITE_KILL_PRIV;
1385                         }
1386                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1387                 } else {
1388                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1389                 }
1390
1391                 if (!io->async)
1392                         fuse_release_user_pages(req, io->should_dirty);
1393                 if (req->out.h.error) {
1394                         err = req->out.h.error;
1395                         break;
1396                 } else if (nres > nbytes) {
1397                         res = 0;
1398                         err = -EIO;
1399                         break;
1400                 }
1401                 count -= nres;
1402                 res += nres;
1403                 pos += nres;
1404                 if (nres != nbytes)
1405                         break;
1406                 if (count) {
1407                         fuse_put_request(fc, req);
1408                         if (io->async)
1409                                 req = fuse_get_req_for_background(fc,
1410                                         iov_iter_npages(iter, fc->max_pages));
1411                         else
1412                                 req = fuse_get_req(fc, iov_iter_npages(iter,
1413                                                                 fc->max_pages));
1414                         if (IS_ERR(req))
1415                                 break;
1416                 }
1417         }
1418         if (!IS_ERR(req))
1419                 fuse_put_request(fc, req);
1420         if (res > 0)
1421                 *ppos = pos;
1422
1423         return res > 0 ? res : err;
1424 }
1425 EXPORT_SYMBOL_GPL(fuse_direct_io);
1426
1427 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1428                                   struct iov_iter *iter,
1429                                   loff_t *ppos)
1430 {
1431         ssize_t res;
1432         struct inode *inode = file_inode(io->iocb->ki_filp);
1433
1434         res = fuse_direct_io(io, iter, ppos, 0);
1435
1436         fuse_invalidate_atime(inode);
1437
1438         return res;
1439 }
1440
1441 static ssize_t fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter);
1442
1443 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1444 {
1445         ssize_t res;
1446
1447         if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1448                 res = fuse_direct_IO(iocb, to);
1449         } else {
1450                 struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1451
1452                 res = __fuse_direct_read(&io, to, &iocb->ki_pos);
1453         }
1454
1455         return res;
1456 }
1457
1458 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1459 {
1460         struct inode *inode = file_inode(iocb->ki_filp);
1461         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb);
1462         ssize_t res;
1463
1464         /* Don't allow parallel writes to the same file */
1465         inode_lock(inode);
1466         res = generic_write_checks(iocb, from);
1467         if (res > 0) {
1468                 if (!is_sync_kiocb(iocb) && iocb->ki_flags & IOCB_DIRECT) {
1469                         res = fuse_direct_IO(iocb, from);
1470                 } else {
1471                         res = fuse_direct_io(&io, from, &iocb->ki_pos,
1472                                              FUSE_DIO_WRITE);
1473                 }
1474         }
1475         fuse_invalidate_attr(inode);
1476         if (res > 0)
1477                 fuse_write_update_size(inode, iocb->ki_pos);
1478         inode_unlock(inode);
1479
1480         return res;
1481 }
1482
1483 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
1484 {
1485         struct file *file = iocb->ki_filp;
1486         struct fuse_file *ff = file->private_data;
1487
1488         if (is_bad_inode(file_inode(file)))
1489                 return -EIO;
1490
1491         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1492                 return fuse_cache_read_iter(iocb, to);
1493         else
1494                 return fuse_direct_read_iter(iocb, to);
1495 }
1496
1497 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1498 {
1499         struct file *file = iocb->ki_filp;
1500         struct fuse_file *ff = file->private_data;
1501
1502         if (is_bad_inode(file_inode(file)))
1503                 return -EIO;
1504
1505         if (!(ff->open_flags & FOPEN_DIRECT_IO))
1506                 return fuse_cache_write_iter(iocb, from);
1507         else
1508                 return fuse_direct_write_iter(iocb, from);
1509 }
1510
1511 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1512 {
1513         int i;
1514
1515         for (i = 0; i < req->num_pages; i++)
1516                 __free_page(req->pages[i]);
1517
1518         if (req->ff)
1519                 fuse_file_put(req->ff, false, false);
1520 }
1521
1522 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1523 {
1524         struct inode *inode = req->inode;
1525         struct fuse_inode *fi = get_fuse_inode(inode);
1526         struct backing_dev_info *bdi = inode_to_bdi(inode);
1527         int i;
1528
1529         list_del(&req->writepages_entry);
1530         for (i = 0; i < req->num_pages; i++) {
1531                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1532                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1533                 wb_writeout_inc(&bdi->wb);
1534         }
1535         wake_up(&fi->page_waitq);
1536 }
1537
1538 /* Called under fi->lock, may release and reacquire it */
1539 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1540                                 loff_t size)
1541 __releases(fi->lock)
1542 __acquires(fi->lock)
1543 {
1544         struct fuse_req *aux, *next;
1545         struct fuse_inode *fi = get_fuse_inode(req->inode);
1546         struct fuse_write_in *inarg = &req->misc.write.in;
1547         __u64 data_size = req->num_pages * PAGE_SIZE;
1548         bool queued;
1549
1550         if (inarg->offset + data_size <= size) {
1551                 inarg->size = data_size;
1552         } else if (inarg->offset < size) {
1553                 inarg->size = size - inarg->offset;
1554         } else {
1555                 /* Got truncated off completely */
1556                 goto out_free;
1557         }
1558
1559         req->in.args[1].size = inarg->size;
1560         queued = fuse_request_queue_background(fc, req);
1561         /* Fails on broken connection only */
1562         if (unlikely(!queued))
1563                 goto out_free;
1564
1565         fi->writectr++;
1566         return;
1567
1568  out_free:
1569         fuse_writepage_finish(fc, req);
1570         spin_unlock(&fi->lock);
1571
1572         /* After fuse_writepage_finish() aux request list is private */
1573         for (aux = req->misc.write.next; aux; aux = next) {
1574                 next = aux->misc.write.next;
1575                 aux->misc.write.next = NULL;
1576                 fuse_writepage_free(fc, aux);
1577                 fuse_put_request(fc, aux);
1578         }
1579
1580         fuse_writepage_free(fc, req);
1581         fuse_put_request(fc, req);
1582         spin_lock(&fi->lock);
1583 }
1584
1585 /*
1586  * If fi->writectr is positive (no truncate or fsync going on) send
1587  * all queued writepage requests.
1588  *
1589  * Called with fi->lock
1590  */
1591 void fuse_flush_writepages(struct inode *inode)
1592 __releases(fi->lock)
1593 __acquires(fi->lock)
1594 {
1595         struct fuse_conn *fc = get_fuse_conn(inode);
1596         struct fuse_inode *fi = get_fuse_inode(inode);
1597         loff_t crop = i_size_read(inode);
1598         struct fuse_req *req;
1599
1600         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1601                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1602                 list_del_init(&req->list);
1603                 fuse_send_writepage(fc, req, crop);
1604         }
1605 }
1606
1607 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1608 {
1609         struct inode *inode = req->inode;
1610         struct fuse_inode *fi = get_fuse_inode(inode);
1611
1612         mapping_set_error(inode->i_mapping, req->out.h.error);
1613         spin_lock(&fi->lock);
1614         while (req->misc.write.next) {
1615                 struct fuse_conn *fc = get_fuse_conn(inode);
1616                 struct fuse_write_in *inarg = &req->misc.write.in;
1617                 struct fuse_req *next = req->misc.write.next;
1618                 req->misc.write.next = next->misc.write.next;
1619                 next->misc.write.next = NULL;
1620                 next->ff = fuse_file_get(req->ff);
1621                 list_add(&next->writepages_entry, &fi->writepages);
1622
1623                 /*
1624                  * Skip fuse_flush_writepages() to make it easy to crop requests
1625                  * based on primary request size.
1626                  *
1627                  * 1st case (trivial): there are no concurrent activities using
1628                  * fuse_set/release_nowrite.  Then we're on safe side because
1629                  * fuse_flush_writepages() would call fuse_send_writepage()
1630                  * anyway.
1631                  *
1632                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1633                  * now for completion of all in-flight requests.  This happens
1634                  * rarely and no more than once per page, so this should be
1635                  * okay.
1636                  *
1637                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1638                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1639                  * that fuse_set_nowrite returned implies that all in-flight
1640                  * requests were completed along with all of their secondary
1641                  * requests.  Further primary requests are blocked by negative
1642                  * writectr.  Hence there cannot be any in-flight requests and
1643                  * no invocations of fuse_writepage_end() while we're in
1644                  * fuse_set_nowrite..fuse_release_nowrite section.
1645                  */
1646                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1647         }
1648         fi->writectr--;
1649         fuse_writepage_finish(fc, req);
1650         spin_unlock(&fi->lock);
1651         fuse_writepage_free(fc, req);
1652 }
1653
1654 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1655                                                struct fuse_inode *fi)
1656 {
1657         struct fuse_file *ff = NULL;
1658
1659         spin_lock(&fi->lock);
1660         if (!list_empty(&fi->write_files)) {
1661                 ff = list_entry(fi->write_files.next, struct fuse_file,
1662                                 write_entry);
1663                 fuse_file_get(ff);
1664         }
1665         spin_unlock(&fi->lock);
1666
1667         return ff;
1668 }
1669
1670 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1671                                              struct fuse_inode *fi)
1672 {
1673         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1674         WARN_ON(!ff);
1675         return ff;
1676 }
1677
1678 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1679 {
1680         struct fuse_conn *fc = get_fuse_conn(inode);
1681         struct fuse_inode *fi = get_fuse_inode(inode);
1682         struct fuse_file *ff;
1683         int err;
1684
1685         ff = __fuse_write_file_get(fc, fi);
1686         err = fuse_flush_times(inode, ff);
1687         if (ff)
1688                 fuse_file_put(ff, false, false);
1689
1690         return err;
1691 }
1692
1693 static int fuse_writepage_locked(struct page *page)
1694 {
1695         struct address_space *mapping = page->mapping;
1696         struct inode *inode = mapping->host;
1697         struct fuse_conn *fc = get_fuse_conn(inode);
1698         struct fuse_inode *fi = get_fuse_inode(inode);
1699         struct fuse_req *req;
1700         struct page *tmp_page;
1701         int error = -ENOMEM;
1702
1703         set_page_writeback(page);
1704
1705         req = fuse_request_alloc_nofs(1);
1706         if (!req)
1707                 goto err;
1708
1709         /* writeback always goes to bg_queue */
1710         __set_bit(FR_BACKGROUND, &req->flags);
1711         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1712         if (!tmp_page)
1713                 goto err_free;
1714
1715         error = -EIO;
1716         req->ff = fuse_write_file_get(fc, fi);
1717         if (!req->ff)
1718                 goto err_nofile;
1719
1720         fuse_write_fill(req, req->ff, page_offset(page), 0);
1721
1722         copy_highpage(tmp_page, page);
1723         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1724         req->misc.write.next = NULL;
1725         req->in.argpages = 1;
1726         req->num_pages = 1;
1727         req->pages[0] = tmp_page;
1728         req->page_descs[0].offset = 0;
1729         req->page_descs[0].length = PAGE_SIZE;
1730         req->end = fuse_writepage_end;
1731         req->inode = inode;
1732
1733         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1734         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1735
1736         spin_lock(&fi->lock);
1737         list_add(&req->writepages_entry, &fi->writepages);
1738         list_add_tail(&req->list, &fi->queued_writes);
1739         fuse_flush_writepages(inode);
1740         spin_unlock(&fi->lock);
1741
1742         end_page_writeback(page);
1743
1744         return 0;
1745
1746 err_nofile:
1747         __free_page(tmp_page);
1748 err_free:
1749         fuse_request_free(req);
1750 err:
1751         mapping_set_error(page->mapping, error);
1752         end_page_writeback(page);
1753         return error;
1754 }
1755
1756 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1757 {
1758         int err;
1759
1760         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1761                 /*
1762                  * ->writepages() should be called for sync() and friends.  We
1763                  * should only get here on direct reclaim and then we are
1764                  * allowed to skip a page which is already in flight
1765                  */
1766                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1767
1768                 redirty_page_for_writepage(wbc, page);
1769                 return 0;
1770         }
1771
1772         err = fuse_writepage_locked(page);
1773         unlock_page(page);
1774
1775         return err;
1776 }
1777
1778 struct fuse_fill_wb_data {
1779         struct fuse_req *req;
1780         struct fuse_file *ff;
1781         struct inode *inode;
1782         struct page **orig_pages;
1783 };
1784
1785 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1786 {
1787         struct fuse_req *req = data->req;
1788         struct inode *inode = data->inode;
1789         struct fuse_inode *fi = get_fuse_inode(inode);
1790         int num_pages = req->num_pages;
1791         int i;
1792
1793         req->ff = fuse_file_get(data->ff);
1794         spin_lock(&fi->lock);
1795         list_add_tail(&req->list, &fi->queued_writes);
1796         fuse_flush_writepages(inode);
1797         spin_unlock(&fi->lock);
1798
1799         for (i = 0; i < num_pages; i++)
1800                 end_page_writeback(data->orig_pages[i]);
1801 }
1802
1803 /*
1804  * First recheck under fi->lock if the offending offset is still under
1805  * writeback.  If yes, then iterate auxiliary write requests, to see if there's
1806  * one already added for a page at this offset.  If there's none, then insert
1807  * this new request onto the auxiliary list, otherwise reuse the existing one by
1808  * copying the new page contents over to the old temporary page.
1809  */
1810 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1811                                      struct page *page)
1812 {
1813         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1814         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1815         struct fuse_req *tmp;
1816         struct fuse_req *old_req;
1817
1818         WARN_ON(new_req->num_pages != 0);
1819
1820         spin_lock(&fi->lock);
1821         list_del(&new_req->writepages_entry);
1822         old_req = fuse_find_writeback(fi, page->index, page->index);
1823         if (!old_req) {
1824                 list_add(&new_req->writepages_entry, &fi->writepages);
1825                 spin_unlock(&fi->lock);
1826                 return false;
1827         }
1828
1829         new_req->num_pages = 1;
1830         for (tmp = old_req->misc.write.next; tmp; tmp = tmp->misc.write.next) {
1831                 pgoff_t curr_index;
1832
1833                 WARN_ON(tmp->inode != new_req->inode);
1834                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1835                 if (curr_index == page->index) {
1836                         WARN_ON(tmp->num_pages != 1);
1837                         WARN_ON(!test_bit(FR_PENDING, &tmp->flags));
1838                         swap(tmp->pages[0], new_req->pages[0]);
1839                         break;
1840                 }
1841         }
1842
1843         if (!tmp) {
1844                 new_req->misc.write.next = old_req->misc.write.next;
1845                 old_req->misc.write.next = new_req;
1846         }
1847
1848         spin_unlock(&fi->lock);
1849
1850         if (tmp) {
1851                 struct backing_dev_info *bdi = inode_to_bdi(new_req->inode);
1852
1853                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1854                 dec_node_page_state(new_req->pages[0], NR_WRITEBACK_TEMP);
1855                 wb_writeout_inc(&bdi->wb);
1856                 fuse_writepage_free(fc, new_req);
1857                 fuse_request_free(new_req);
1858         }
1859
1860         return true;
1861 }
1862
1863 static int fuse_writepages_fill(struct page *page,
1864                 struct writeback_control *wbc, void *_data)
1865 {
1866         struct fuse_fill_wb_data *data = _data;
1867         struct fuse_req *req = data->req;
1868         struct inode *inode = data->inode;
1869         struct fuse_inode *fi = get_fuse_inode(inode);
1870         struct fuse_conn *fc = get_fuse_conn(inode);
1871         struct page *tmp_page;
1872         bool is_writeback;
1873         int err;
1874
1875         if (!data->ff) {
1876                 err = -EIO;
1877                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1878                 if (!data->ff)
1879                         goto out_unlock;
1880         }
1881
1882         /*
1883          * Being under writeback is unlikely but possible.  For example direct
1884          * read to an mmaped fuse file will set the page dirty twice; once when
1885          * the pages are faulted with get_user_pages(), and then after the read
1886          * completed.
1887          */
1888         is_writeback = fuse_page_is_writeback(inode, page->index);
1889
1890         if (req && req->num_pages &&
1891             (is_writeback || req->num_pages == fc->max_pages ||
1892              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1893              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1894                 fuse_writepages_send(data);
1895                 data->req = NULL;
1896         } else if (req && req->num_pages == req->max_pages) {
1897                 if (!fuse_req_realloc_pages(fc, req, GFP_NOFS)) {
1898                         fuse_writepages_send(data);
1899                         req = data->req = NULL;
1900                 }
1901         }
1902
1903         err = -ENOMEM;
1904         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1905         if (!tmp_page)
1906                 goto out_unlock;
1907
1908         /*
1909          * The page must not be redirtied until the writeout is completed
1910          * (i.e. userspace has sent a reply to the write request).  Otherwise
1911          * there could be more than one temporary page instance for each real
1912          * page.
1913          *
1914          * This is ensured by holding the page lock in page_mkwrite() while
1915          * checking fuse_page_is_writeback().  We already hold the page lock
1916          * since clear_page_dirty_for_io() and keep it held until we add the
1917          * request to the fi->writepages list and increment req->num_pages.
1918          * After this fuse_page_is_writeback() will indicate that the page is
1919          * under writeback, so we can release the page lock.
1920          */
1921         if (data->req == NULL) {
1922                 struct fuse_inode *fi = get_fuse_inode(inode);
1923
1924                 err = -ENOMEM;
1925                 req = fuse_request_alloc_nofs(FUSE_REQ_INLINE_PAGES);
1926                 if (!req) {
1927                         __free_page(tmp_page);
1928                         goto out_unlock;
1929                 }
1930
1931                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1932                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1933                 req->misc.write.next = NULL;
1934                 req->in.argpages = 1;
1935                 __set_bit(FR_BACKGROUND, &req->flags);
1936                 req->num_pages = 0;
1937                 req->end = fuse_writepage_end;
1938                 req->inode = inode;
1939
1940                 spin_lock(&fi->lock);
1941                 list_add(&req->writepages_entry, &fi->writepages);
1942                 spin_unlock(&fi->lock);
1943
1944                 data->req = req;
1945         }
1946         set_page_writeback(page);
1947
1948         copy_highpage(tmp_page, page);
1949         req->pages[req->num_pages] = tmp_page;
1950         req->page_descs[req->num_pages].offset = 0;
1951         req->page_descs[req->num_pages].length = PAGE_SIZE;
1952
1953         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1954         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1955
1956         err = 0;
1957         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1958                 end_page_writeback(page);
1959                 data->req = NULL;
1960                 goto out_unlock;
1961         }
1962         data->orig_pages[req->num_pages] = page;
1963
1964         /*
1965          * Protected by fi->lock against concurrent access by
1966          * fuse_page_is_writeback().
1967          */
1968         spin_lock(&fi->lock);
1969         req->num_pages++;
1970         spin_unlock(&fi->lock);
1971
1972 out_unlock:
1973         unlock_page(page);
1974
1975         return err;
1976 }
1977
1978 static int fuse_writepages(struct address_space *mapping,
1979                            struct writeback_control *wbc)
1980 {
1981         struct inode *inode = mapping->host;
1982         struct fuse_conn *fc = get_fuse_conn(inode);
1983         struct fuse_fill_wb_data data;
1984         int err;
1985
1986         err = -EIO;
1987         if (is_bad_inode(inode))
1988                 goto out;
1989
1990         data.inode = inode;
1991         data.req = NULL;
1992         data.ff = NULL;
1993
1994         err = -ENOMEM;
1995         data.orig_pages = kcalloc(fc->max_pages,
1996                                   sizeof(struct page *),
1997                                   GFP_NOFS);
1998         if (!data.orig_pages)
1999                 goto out;
2000
2001         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
2002         if (data.req) {
2003                 /* Ignore errors if we can write at least one page */
2004                 BUG_ON(!data.req->num_pages);
2005                 fuse_writepages_send(&data);
2006                 err = 0;
2007         }
2008         if (data.ff)
2009                 fuse_file_put(data.ff, false, false);
2010
2011         kfree(data.orig_pages);
2012 out:
2013         return err;
2014 }
2015
2016 /*
2017  * It's worthy to make sure that space is reserved on disk for the write,
2018  * but how to implement it without killing performance need more thinking.
2019  */
2020 static int fuse_write_begin(struct file *file, struct address_space *mapping,
2021                 loff_t pos, unsigned len, unsigned flags,
2022                 struct page **pagep, void **fsdata)
2023 {
2024         pgoff_t index = pos >> PAGE_SHIFT;
2025         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
2026         struct page *page;
2027         loff_t fsize;
2028         int err = -ENOMEM;
2029
2030         WARN_ON(!fc->writeback_cache);
2031
2032         page = grab_cache_page_write_begin(mapping, index, flags);
2033         if (!page)
2034                 goto error;
2035
2036         fuse_wait_on_page_writeback(mapping->host, page->index);
2037
2038         if (PageUptodate(page) || len == PAGE_SIZE)
2039                 goto success;
2040         /*
2041          * Check if the start this page comes after the end of file, in which
2042          * case the readpage can be optimized away.
2043          */
2044         fsize = i_size_read(mapping->host);
2045         if (fsize <= (pos & PAGE_MASK)) {
2046                 size_t off = pos & ~PAGE_MASK;
2047                 if (off)
2048                         zero_user_segment(page, 0, off);
2049                 goto success;
2050         }
2051         err = fuse_do_readpage(file, page);
2052         if (err)
2053                 goto cleanup;
2054 success:
2055         *pagep = page;
2056         return 0;
2057
2058 cleanup:
2059         unlock_page(page);
2060         put_page(page);
2061 error:
2062         return err;
2063 }
2064
2065 static int fuse_write_end(struct file *file, struct address_space *mapping,
2066                 loff_t pos, unsigned len, unsigned copied,
2067                 struct page *page, void *fsdata)
2068 {
2069         struct inode *inode = page->mapping->host;
2070
2071         /* Haven't copied anything?  Skip zeroing, size extending, dirtying. */
2072         if (!copied)
2073                 goto unlock;
2074
2075         if (!PageUptodate(page)) {
2076                 /* Zero any unwritten bytes at the end of the page */
2077                 size_t endoff = (pos + copied) & ~PAGE_MASK;
2078                 if (endoff)
2079                         zero_user_segment(page, endoff, PAGE_SIZE);
2080                 SetPageUptodate(page);
2081         }
2082
2083         fuse_write_update_size(inode, pos + copied);
2084         set_page_dirty(page);
2085
2086 unlock:
2087         unlock_page(page);
2088         put_page(page);
2089
2090         return copied;
2091 }
2092
2093 static int fuse_launder_page(struct page *page)
2094 {
2095         int err = 0;
2096         if (clear_page_dirty_for_io(page)) {
2097                 struct inode *inode = page->mapping->host;
2098                 err = fuse_writepage_locked(page);
2099                 if (!err)
2100                         fuse_wait_on_page_writeback(inode, page->index);
2101         }
2102         return err;
2103 }
2104
2105 /*
2106  * Write back dirty pages now, because there may not be any suitable
2107  * open files later
2108  */
2109 static void fuse_vma_close(struct vm_area_struct *vma)
2110 {
2111         filemap_write_and_wait(vma->vm_file->f_mapping);
2112 }
2113
2114 /*
2115  * Wait for writeback against this page to complete before allowing it
2116  * to be marked dirty again, and hence written back again, possibly
2117  * before the previous writepage completed.
2118  *
2119  * Block here, instead of in ->writepage(), so that the userspace fs
2120  * can only block processes actually operating on the filesystem.
2121  *
2122  * Otherwise unprivileged userspace fs would be able to block
2123  * unrelated:
2124  *
2125  * - page migration
2126  * - sync(2)
2127  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2128  */
2129 static vm_fault_t fuse_page_mkwrite(struct vm_fault *vmf)
2130 {
2131         struct page *page = vmf->page;
2132         struct inode *inode = file_inode(vmf->vma->vm_file);
2133
2134         file_update_time(vmf->vma->vm_file);
2135         lock_page(page);
2136         if (page->mapping != inode->i_mapping) {
2137                 unlock_page(page);
2138                 return VM_FAULT_NOPAGE;
2139         }
2140
2141         fuse_wait_on_page_writeback(inode, page->index);
2142         return VM_FAULT_LOCKED;
2143 }
2144
2145 static const struct vm_operations_struct fuse_file_vm_ops = {
2146         .close          = fuse_vma_close,
2147         .fault          = filemap_fault,
2148         .map_pages      = filemap_map_pages,
2149         .page_mkwrite   = fuse_page_mkwrite,
2150 };
2151
2152 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2153 {
2154         struct fuse_file *ff = file->private_data;
2155
2156         if (ff->open_flags & FOPEN_DIRECT_IO) {
2157                 /* Can't provide the coherency needed for MAP_SHARED */
2158                 if (vma->vm_flags & VM_MAYSHARE)
2159                         return -ENODEV;
2160
2161                 invalidate_inode_pages2(file->f_mapping);
2162
2163                 return generic_file_mmap(file, vma);
2164         }
2165
2166         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2167                 fuse_link_write_file(file);
2168
2169         file_accessed(file);
2170         vma->vm_ops = &fuse_file_vm_ops;
2171         return 0;
2172 }
2173
2174 static int convert_fuse_file_lock(struct fuse_conn *fc,
2175                                   const struct fuse_file_lock *ffl,
2176                                   struct file_lock *fl)
2177 {
2178         switch (ffl->type) {
2179         case F_UNLCK:
2180                 break;
2181
2182         case F_RDLCK:
2183         case F_WRLCK:
2184                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2185                     ffl->end < ffl->start)
2186                         return -EIO;
2187
2188                 fl->fl_start = ffl->start;
2189                 fl->fl_end = ffl->end;
2190
2191                 /*
2192                  * Convert pid into init's pid namespace.  The locks API will
2193                  * translate it into the caller's pid namespace.
2194                  */
2195                 rcu_read_lock();
2196                 fl->fl_pid = pid_nr_ns(find_pid_ns(ffl->pid, fc->pid_ns), &init_pid_ns);
2197                 rcu_read_unlock();
2198                 break;
2199
2200         default:
2201                 return -EIO;
2202         }
2203         fl->fl_type = ffl->type;
2204         return 0;
2205 }
2206
2207 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2208                          const struct file_lock *fl, int opcode, pid_t pid,
2209                          int flock, struct fuse_lk_in *inarg)
2210 {
2211         struct inode *inode = file_inode(file);
2212         struct fuse_conn *fc = get_fuse_conn(inode);
2213         struct fuse_file *ff = file->private_data;
2214
2215         memset(inarg, 0, sizeof(*inarg));
2216         inarg->fh = ff->fh;
2217         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2218         inarg->lk.start = fl->fl_start;
2219         inarg->lk.end = fl->fl_end;
2220         inarg->lk.type = fl->fl_type;
2221         inarg->lk.pid = pid;
2222         if (flock)
2223                 inarg->lk_flags |= FUSE_LK_FLOCK;
2224         args->opcode = opcode;
2225         args->nodeid = get_node_id(inode);
2226         args->in_numargs = 1;
2227         args->in_args[0].size = sizeof(*inarg);
2228         args->in_args[0].value = inarg;
2229 }
2230
2231 static int fuse_getlk(struct file *file, struct file_lock *fl)
2232 {
2233         struct inode *inode = file_inode(file);
2234         struct fuse_conn *fc = get_fuse_conn(inode);
2235         FUSE_ARGS(args);
2236         struct fuse_lk_in inarg;
2237         struct fuse_lk_out outarg;
2238         int err;
2239
2240         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2241         args.out_numargs = 1;
2242         args.out_args[0].size = sizeof(outarg);
2243         args.out_args[0].value = &outarg;
2244         err = fuse_simple_request(fc, &args);
2245         if (!err)
2246                 err = convert_fuse_file_lock(fc, &outarg.lk, fl);
2247
2248         return err;
2249 }
2250
2251 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2252 {
2253         struct inode *inode = file_inode(file);
2254         struct fuse_conn *fc = get_fuse_conn(inode);
2255         FUSE_ARGS(args);
2256         struct fuse_lk_in inarg;
2257         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2258         struct pid *pid = fl->fl_type != F_UNLCK ? task_tgid(current) : NULL;
2259         pid_t pid_nr = pid_nr_ns(pid, fc->pid_ns);
2260         int err;
2261
2262         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2263                 /* NLM needs asynchronous locks, which we don't support yet */
2264                 return -ENOLCK;
2265         }
2266
2267         /* Unlock on close is handled by the flush method */
2268         if ((fl->fl_flags & FL_CLOSE_POSIX) == FL_CLOSE_POSIX)
2269                 return 0;
2270
2271         fuse_lk_fill(&args, file, fl, opcode, pid_nr, flock, &inarg);
2272         err = fuse_simple_request(fc, &args);
2273
2274         /* locking is restartable */
2275         if (err == -EINTR)
2276                 err = -ERESTARTSYS;
2277
2278         return err;
2279 }
2280
2281 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2282 {
2283         struct inode *inode = file_inode(file);
2284         struct fuse_conn *fc = get_fuse_conn(inode);
2285         int err;
2286
2287         if (cmd == F_CANCELLK) {
2288                 err = 0;
2289         } else if (cmd == F_GETLK) {
2290                 if (fc->no_lock) {
2291                         posix_test_lock(file, fl);
2292                         err = 0;
2293                 } else
2294                         err = fuse_getlk(file, fl);
2295         } else {
2296                 if (fc->no_lock)
2297                         err = posix_lock_file(file, fl, NULL);
2298                 else
2299                         err = fuse_setlk(file, fl, 0);
2300         }
2301         return err;
2302 }
2303
2304 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2305 {
2306         struct inode *inode = file_inode(file);
2307         struct fuse_conn *fc = get_fuse_conn(inode);
2308         int err;
2309
2310         if (fc->no_flock) {
2311                 err = locks_lock_file_wait(file, fl);
2312         } else {
2313                 struct fuse_file *ff = file->private_data;
2314
2315                 /* emulate flock with POSIX locks */
2316                 ff->flock = true;
2317                 err = fuse_setlk(file, fl, 1);
2318         }
2319
2320         return err;
2321 }
2322
2323 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2324 {
2325         struct inode *inode = mapping->host;
2326         struct fuse_conn *fc = get_fuse_conn(inode);
2327         FUSE_ARGS(args);
2328         struct fuse_bmap_in inarg;
2329         struct fuse_bmap_out outarg;
2330         int err;
2331
2332         if (!inode->i_sb->s_bdev || fc->no_bmap)
2333                 return 0;
2334
2335         memset(&inarg, 0, sizeof(inarg));
2336         inarg.block = block;
2337         inarg.blocksize = inode->i_sb->s_blocksize;
2338         args.opcode = FUSE_BMAP;
2339         args.nodeid = get_node_id(inode);
2340         args.in_numargs = 1;
2341         args.in_args[0].size = sizeof(inarg);
2342         args.in_args[0].value = &inarg;
2343         args.out_numargs = 1;
2344         args.out_args[0].size = sizeof(outarg);
2345         args.out_args[0].value = &outarg;
2346         err = fuse_simple_request(fc, &args);
2347         if (err == -ENOSYS)
2348                 fc->no_bmap = 1;
2349
2350         return err ? 0 : outarg.block;
2351 }
2352
2353 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2354 {
2355         struct inode *inode = file->f_mapping->host;
2356         struct fuse_conn *fc = get_fuse_conn(inode);
2357         struct fuse_file *ff = file->private_data;
2358         FUSE_ARGS(args);
2359         struct fuse_lseek_in inarg = {
2360                 .fh = ff->fh,
2361                 .offset = offset,
2362                 .whence = whence
2363         };
2364         struct fuse_lseek_out outarg;
2365         int err;
2366
2367         if (fc->no_lseek)
2368                 goto fallback;
2369
2370         args.opcode = FUSE_LSEEK;
2371         args.nodeid = ff->nodeid;
2372         args.in_numargs = 1;
2373         args.in_args[0].size = sizeof(inarg);
2374         args.in_args[0].value = &inarg;
2375         args.out_numargs = 1;
2376         args.out_args[0].size = sizeof(outarg);
2377         args.out_args[0].value = &outarg;
2378         err = fuse_simple_request(fc, &args);
2379         if (err) {
2380                 if (err == -ENOSYS) {
2381                         fc->no_lseek = 1;
2382                         goto fallback;
2383                 }
2384                 return err;
2385         }
2386
2387         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2388
2389 fallback:
2390         err = fuse_update_attributes(inode, file);
2391         if (!err)
2392                 return generic_file_llseek(file, offset, whence);
2393         else
2394                 return err;
2395 }
2396
2397 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2398 {
2399         loff_t retval;
2400         struct inode *inode = file_inode(file);
2401
2402         switch (whence) {
2403         case SEEK_SET:
2404         case SEEK_CUR:
2405                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2406                 retval = generic_file_llseek(file, offset, whence);
2407                 break;
2408         case SEEK_END:
2409                 inode_lock(inode);
2410                 retval = fuse_update_attributes(inode, file);
2411                 if (!retval)
2412                         retval = generic_file_llseek(file, offset, whence);
2413                 inode_unlock(inode);
2414                 break;
2415         case SEEK_HOLE:
2416         case SEEK_DATA:
2417                 inode_lock(inode);
2418                 retval = fuse_lseek(file, offset, whence);
2419                 inode_unlock(inode);
2420                 break;
2421         default:
2422                 retval = -EINVAL;
2423         }
2424
2425         return retval;
2426 }
2427
2428 /*
2429  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2430  * ABI was defined to be 'struct iovec' which is different on 32bit
2431  * and 64bit.  Fortunately we can determine which structure the server
2432  * used from the size of the reply.
2433  */
2434 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2435                                      size_t transferred, unsigned count,
2436                                      bool is_compat)
2437 {
2438 #ifdef CONFIG_COMPAT
2439         if (count * sizeof(struct compat_iovec) == transferred) {
2440                 struct compat_iovec *ciov = src;
2441                 unsigned i;
2442
2443                 /*
2444                  * With this interface a 32bit server cannot support
2445                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2446                  * requests
2447                  */
2448                 if (!is_compat)
2449                         return -EINVAL;
2450
2451                 for (i = 0; i < count; i++) {
2452                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2453                         dst[i].iov_len = ciov[i].iov_len;
2454                 }
2455                 return 0;
2456         }
2457 #endif
2458
2459         if (count * sizeof(struct iovec) != transferred)
2460                 return -EIO;
2461
2462         memcpy(dst, src, transferred);
2463         return 0;
2464 }
2465
2466 /* Make sure iov_length() won't overflow */
2467 static int fuse_verify_ioctl_iov(struct fuse_conn *fc, struct iovec *iov,
2468                                  size_t count)
2469 {
2470         size_t n;
2471         u32 max = fc->max_pages << PAGE_SHIFT;
2472
2473         for (n = 0; n < count; n++, iov++) {
2474                 if (iov->iov_len > (size_t) max)
2475                         return -ENOMEM;
2476                 max -= iov->iov_len;
2477         }
2478         return 0;
2479 }
2480
2481 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2482                                  void *src, size_t transferred, unsigned count,
2483                                  bool is_compat)
2484 {
2485         unsigned i;
2486         struct fuse_ioctl_iovec *fiov = src;
2487
2488         if (fc->minor < 16) {
2489                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2490                                                  count, is_compat);
2491         }
2492
2493         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2494                 return -EIO;
2495
2496         for (i = 0; i < count; i++) {
2497                 /* Did the server supply an inappropriate value? */
2498                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2499                     fiov[i].len != (unsigned long) fiov[i].len)
2500                         return -EIO;
2501
2502                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2503                 dst[i].iov_len = (size_t) fiov[i].len;
2504
2505 #ifdef CONFIG_COMPAT
2506                 if (is_compat &&
2507                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2508                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2509                         return -EIO;
2510 #endif
2511         }
2512
2513         return 0;
2514 }
2515
2516
2517 /*
2518  * For ioctls, there is no generic way to determine how much memory
2519  * needs to be read and/or written.  Furthermore, ioctls are allowed
2520  * to dereference the passed pointer, so the parameter requires deep
2521  * copying but FUSE has no idea whatsoever about what to copy in or
2522  * out.
2523  *
2524  * This is solved by allowing FUSE server to retry ioctl with
2525  * necessary in/out iovecs.  Let's assume the ioctl implementation
2526  * needs to read in the following structure.
2527  *
2528  * struct a {
2529  *      char    *buf;
2530  *      size_t  buflen;
2531  * }
2532  *
2533  * On the first callout to FUSE server, inarg->in_size and
2534  * inarg->out_size will be NULL; then, the server completes the ioctl
2535  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2536  * the actual iov array to
2537  *
2538  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2539  *
2540  * which tells FUSE to copy in the requested area and retry the ioctl.
2541  * On the second round, the server has access to the structure and
2542  * from that it can tell what to look for next, so on the invocation,
2543  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2544  *
2545  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2546  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2547  *
2548  * FUSE will copy both struct a and the pointed buffer from the
2549  * process doing the ioctl and retry ioctl with both struct a and the
2550  * buffer.
2551  *
2552  * This time, FUSE server has everything it needs and completes ioctl
2553  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2554  *
2555  * Copying data out works the same way.
2556  *
2557  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2558  * automatically initializes in and out iovs by decoding @cmd with
2559  * _IOC_* macros and the server is not allowed to request RETRY.  This
2560  * limits ioctl data transfers to well-formed ioctls and is the forced
2561  * behavior for all FUSE servers.
2562  */
2563 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2564                    unsigned int flags)
2565 {
2566         struct fuse_file *ff = file->private_data;
2567         struct fuse_conn *fc = ff->fc;
2568         struct fuse_ioctl_in inarg = {
2569                 .fh = ff->fh,
2570                 .cmd = cmd,
2571                 .arg = arg,
2572                 .flags = flags
2573         };
2574         struct fuse_ioctl_out outarg;
2575         struct fuse_req *req = NULL;
2576         struct page **pages = NULL;
2577         struct iovec *iov_page = NULL;
2578         struct iovec *in_iov = NULL, *out_iov = NULL;
2579         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2580         size_t in_size, out_size, transferred, c;
2581         int err, i;
2582         struct iov_iter ii;
2583
2584 #if BITS_PER_LONG == 32
2585         inarg.flags |= FUSE_IOCTL_32BIT;
2586 #else
2587         if (flags & FUSE_IOCTL_COMPAT) {
2588                 inarg.flags |= FUSE_IOCTL_32BIT;
2589 #ifdef CONFIG_X86_X32
2590                 if (in_x32_syscall())
2591                         inarg.flags |= FUSE_IOCTL_COMPAT_X32;
2592 #endif
2593         }
2594 #endif
2595
2596         /* assume all the iovs returned by client always fits in a page */
2597         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2598
2599         err = -ENOMEM;
2600         pages = kcalloc(fc->max_pages, sizeof(pages[0]), GFP_KERNEL);
2601         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2602         if (!pages || !iov_page)
2603                 goto out;
2604
2605         /*
2606          * If restricted, initialize IO parameters as encoded in @cmd.
2607          * RETRY from server is not allowed.
2608          */
2609         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2610                 struct iovec *iov = iov_page;
2611
2612                 iov->iov_base = (void __user *)arg;
2613                 iov->iov_len = _IOC_SIZE(cmd);
2614
2615                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2616                         in_iov = iov;
2617                         in_iovs = 1;
2618                 }
2619
2620                 if (_IOC_DIR(cmd) & _IOC_READ) {
2621                         out_iov = iov;
2622                         out_iovs = 1;
2623                 }
2624         }
2625
2626  retry:
2627         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2628         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2629
2630         /*
2631          * Out data can be used either for actual out data or iovs,
2632          * make sure there always is at least one page.
2633          */
2634         out_size = max_t(size_t, out_size, PAGE_SIZE);
2635         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2636
2637         /* make sure there are enough buffer pages and init request with them */
2638         err = -ENOMEM;
2639         if (max_pages > fc->max_pages)
2640                 goto out;
2641         while (num_pages < max_pages) {
2642                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2643                 if (!pages[num_pages])
2644                         goto out;
2645                 num_pages++;
2646         }
2647
2648         req = fuse_get_req(fc, num_pages);
2649         if (IS_ERR(req)) {
2650                 err = PTR_ERR(req);
2651                 req = NULL;
2652                 goto out;
2653         }
2654         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2655         req->num_pages = num_pages;
2656         fuse_page_descs_length_init(req, 0, req->num_pages);
2657
2658         /* okay, let's send it to the client */
2659         req->in.h.opcode = FUSE_IOCTL;
2660         req->in.h.nodeid = ff->nodeid;
2661         req->in.numargs = 1;
2662         req->in.args[0].size = sizeof(inarg);
2663         req->in.args[0].value = &inarg;
2664         if (in_size) {
2665                 req->in.numargs++;
2666                 req->in.args[1].size = in_size;
2667                 req->in.argpages = 1;
2668
2669                 err = -EFAULT;
2670                 iov_iter_init(&ii, WRITE, in_iov, in_iovs, in_size);
2671                 for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2672                         c = copy_page_from_iter(pages[i], 0, PAGE_SIZE, &ii);
2673                         if (c != PAGE_SIZE && iov_iter_count(&ii))
2674                                 goto out;
2675                 }
2676         }
2677
2678         req->out.numargs = 2;
2679         req->out.args[0].size = sizeof(outarg);
2680         req->out.args[0].value = &outarg;
2681         req->out.args[1].size = out_size;
2682         req->out.argpages = 1;
2683         req->out.argvar = 1;
2684
2685         fuse_request_send(fc, req);
2686         err = req->out.h.error;
2687         transferred = req->out.args[1].size;
2688         fuse_put_request(fc, req);
2689         req = NULL;
2690         if (err)
2691                 goto out;
2692
2693         /* did it ask for retry? */
2694         if (outarg.flags & FUSE_IOCTL_RETRY) {
2695                 void *vaddr;
2696
2697                 /* no retry if in restricted mode */
2698                 err = -EIO;
2699                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2700                         goto out;
2701
2702                 in_iovs = outarg.in_iovs;
2703                 out_iovs = outarg.out_iovs;
2704
2705                 /*
2706                  * Make sure things are in boundary, separate checks
2707                  * are to protect against overflow.
2708                  */
2709                 err = -ENOMEM;
2710                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2711                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2712                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2713                         goto out;
2714
2715                 vaddr = kmap_atomic(pages[0]);
2716                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2717                                             transferred, in_iovs + out_iovs,
2718                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2719                 kunmap_atomic(vaddr);
2720                 if (err)
2721                         goto out;
2722
2723                 in_iov = iov_page;
2724                 out_iov = in_iov + in_iovs;
2725
2726                 err = fuse_verify_ioctl_iov(fc, in_iov, in_iovs);
2727                 if (err)
2728                         goto out;
2729
2730                 err = fuse_verify_ioctl_iov(fc, out_iov, out_iovs);
2731                 if (err)
2732                         goto out;
2733
2734                 goto retry;
2735         }
2736
2737         err = -EIO;
2738         if (transferred > inarg.out_size)
2739                 goto out;
2740
2741         err = -EFAULT;
2742         iov_iter_init(&ii, READ, out_iov, out_iovs, transferred);
2743         for (i = 0; iov_iter_count(&ii) && !WARN_ON(i >= num_pages); i++) {
2744                 c = copy_page_to_iter(pages[i], 0, PAGE_SIZE, &ii);
2745                 if (c != PAGE_SIZE && iov_iter_count(&ii))
2746                         goto out;
2747         }
2748         err = 0;
2749  out:
2750         if (req)
2751                 fuse_put_request(fc, req);
2752         free_page((unsigned long) iov_page);
2753         while (num_pages)
2754                 __free_page(pages[--num_pages]);
2755         kfree(pages);
2756
2757         return err ? err : outarg.result;
2758 }
2759 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2760
2761 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2762                        unsigned long arg, unsigned int flags)
2763 {
2764         struct inode *inode = file_inode(file);
2765         struct fuse_conn *fc = get_fuse_conn(inode);
2766
2767         if (!fuse_allow_current_process(fc))
2768                 return -EACCES;
2769
2770         if (is_bad_inode(inode))
2771                 return -EIO;
2772
2773         return fuse_do_ioctl(file, cmd, arg, flags);
2774 }
2775
2776 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2777                             unsigned long arg)
2778 {
2779         return fuse_ioctl_common(file, cmd, arg, 0);
2780 }
2781
2782 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2783                                    unsigned long arg)
2784 {
2785         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2786 }
2787
2788 /*
2789  * All files which have been polled are linked to RB tree
2790  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2791  * find the matching one.
2792  */
2793 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2794                                               struct rb_node **parent_out)
2795 {
2796         struct rb_node **link = &fc->polled_files.rb_node;
2797         struct rb_node *last = NULL;
2798
2799         while (*link) {
2800                 struct fuse_file *ff;
2801
2802                 last = *link;
2803                 ff = rb_entry(last, struct fuse_file, polled_node);
2804
2805                 if (kh < ff->kh)
2806                         link = &last->rb_left;
2807                 else if (kh > ff->kh)
2808                         link = &last->rb_right;
2809                 else
2810                         return link;
2811         }
2812
2813         if (parent_out)
2814                 *parent_out = last;
2815         return link;
2816 }
2817
2818 /*
2819  * The file is about to be polled.  Make sure it's on the polled_files
2820  * RB tree.  Note that files once added to the polled_files tree are
2821  * not removed before the file is released.  This is because a file
2822  * polled once is likely to be polled again.
2823  */
2824 static void fuse_register_polled_file(struct fuse_conn *fc,
2825                                       struct fuse_file *ff)
2826 {
2827         spin_lock(&fc->lock);
2828         if (RB_EMPTY_NODE(&ff->polled_node)) {
2829                 struct rb_node **link, *uninitialized_var(parent);
2830
2831                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2832                 BUG_ON(*link);
2833                 rb_link_node(&ff->polled_node, parent, link);
2834                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2835         }
2836         spin_unlock(&fc->lock);
2837 }
2838
2839 __poll_t fuse_file_poll(struct file *file, poll_table *wait)
2840 {
2841         struct fuse_file *ff = file->private_data;
2842         struct fuse_conn *fc = ff->fc;
2843         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2844         struct fuse_poll_out outarg;
2845         FUSE_ARGS(args);
2846         int err;
2847
2848         if (fc->no_poll)
2849                 return DEFAULT_POLLMASK;
2850
2851         poll_wait(file, &ff->poll_wait, wait);
2852         inarg.events = mangle_poll(poll_requested_events(wait));
2853
2854         /*
2855          * Ask for notification iff there's someone waiting for it.
2856          * The client may ignore the flag and always notify.
2857          */
2858         if (waitqueue_active(&ff->poll_wait)) {
2859                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2860                 fuse_register_polled_file(fc, ff);
2861         }
2862
2863         args.opcode = FUSE_POLL;
2864         args.nodeid = ff->nodeid;
2865         args.in_numargs = 1;
2866         args.in_args[0].size = sizeof(inarg);
2867         args.in_args[0].value = &inarg;
2868         args.out_numargs = 1;
2869         args.out_args[0].size = sizeof(outarg);
2870         args.out_args[0].value = &outarg;
2871         err = fuse_simple_request(fc, &args);
2872
2873         if (!err)
2874                 return demangle_poll(outarg.revents);
2875         if (err == -ENOSYS) {
2876                 fc->no_poll = 1;
2877                 return DEFAULT_POLLMASK;
2878         }
2879         return EPOLLERR;
2880 }
2881 EXPORT_SYMBOL_GPL(fuse_file_poll);
2882
2883 /*
2884  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2885  * wakes up the poll waiters.
2886  */
2887 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2888                             struct fuse_notify_poll_wakeup_out *outarg)
2889 {
2890         u64 kh = outarg->kh;
2891         struct rb_node **link;
2892
2893         spin_lock(&fc->lock);
2894
2895         link = fuse_find_polled_node(fc, kh, NULL);
2896         if (*link) {
2897                 struct fuse_file *ff;
2898
2899                 ff = rb_entry(*link, struct fuse_file, polled_node);
2900                 wake_up_interruptible_sync(&ff->poll_wait);
2901         }
2902
2903         spin_unlock(&fc->lock);
2904         return 0;
2905 }
2906
2907 static void fuse_do_truncate(struct file *file)
2908 {
2909         struct inode *inode = file->f_mapping->host;
2910         struct iattr attr;
2911
2912         attr.ia_valid = ATTR_SIZE;
2913         attr.ia_size = i_size_read(inode);
2914
2915         attr.ia_file = file;
2916         attr.ia_valid |= ATTR_FILE;
2917
2918         fuse_do_setattr(file_dentry(file), &attr, file);
2919 }
2920
2921 static inline loff_t fuse_round_up(struct fuse_conn *fc, loff_t off)
2922 {
2923         return round_up(off, fc->max_pages << PAGE_SHIFT);
2924 }
2925
2926 static ssize_t
2927 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2928 {
2929         DECLARE_COMPLETION_ONSTACK(wait);
2930         ssize_t ret = 0;
2931         struct file *file = iocb->ki_filp;
2932         struct fuse_file *ff = file->private_data;
2933         bool async_dio = ff->fc->async_dio;
2934         loff_t pos = 0;
2935         struct inode *inode;
2936         loff_t i_size;
2937         size_t count = iov_iter_count(iter);
2938         loff_t offset = iocb->ki_pos;
2939         struct fuse_io_priv *io;
2940
2941         pos = offset;
2942         inode = file->f_mapping->host;
2943         i_size = i_size_read(inode);
2944
2945         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2946                 return 0;
2947
2948         /* optimization for short read */
2949         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2950                 if (offset >= i_size)
2951                         return 0;
2952                 iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset));
2953                 count = iov_iter_count(iter);
2954         }
2955
2956         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2957         if (!io)
2958                 return -ENOMEM;
2959         spin_lock_init(&io->lock);
2960         kref_init(&io->refcnt);
2961         io->reqs = 1;
2962         io->bytes = -1;
2963         io->size = 0;
2964         io->offset = offset;
2965         io->write = (iov_iter_rw(iter) == WRITE);
2966         io->err = 0;
2967         /*
2968          * By default, we want to optimize all I/Os with async request
2969          * submission to the client filesystem if supported.
2970          */
2971         io->async = async_dio;
2972         io->iocb = iocb;
2973         io->blocking = is_sync_kiocb(iocb);
2974
2975         /*
2976          * We cannot asynchronously extend the size of a file.
2977          * In such case the aio will behave exactly like sync io.
2978          */
2979         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2980                 io->blocking = true;
2981
2982         if (io->async && io->blocking) {
2983                 /*
2984                  * Additional reference to keep io around after
2985                  * calling fuse_aio_complete()
2986                  */
2987                 kref_get(&io->refcnt);
2988                 io->done = &wait;
2989         }
2990
2991         if (iov_iter_rw(iter) == WRITE) {
2992                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2993                 fuse_invalidate_attr(inode);
2994         } else {
2995                 ret = __fuse_direct_read(io, iter, &pos);
2996         }
2997
2998         if (io->async) {
2999                 bool blocking = io->blocking;
3000
3001                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
3002
3003                 /* we have a non-extending, async request, so return */
3004                 if (!blocking)
3005                         return -EIOCBQUEUED;
3006
3007                 wait_for_completion(&wait);
3008                 ret = fuse_get_res_by_io(io);
3009         }
3010
3011         kref_put(&io->refcnt, fuse_io_release);
3012
3013         if (iov_iter_rw(iter) == WRITE) {
3014                 if (ret > 0)
3015                         fuse_write_update_size(inode, pos);
3016                 else if (ret < 0 && offset + count > i_size)
3017                         fuse_do_truncate(file);
3018         }
3019
3020         return ret;
3021 }
3022
3023 static int fuse_writeback_range(struct inode *inode, loff_t start, loff_t end)
3024 {
3025         int err = filemap_write_and_wait_range(inode->i_mapping, start, end);
3026
3027         if (!err)
3028                 fuse_sync_writes(inode);
3029
3030         return err;
3031 }
3032
3033 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
3034                                 loff_t length)
3035 {
3036         struct fuse_file *ff = file->private_data;
3037         struct inode *inode = file_inode(file);
3038         struct fuse_inode *fi = get_fuse_inode(inode);
3039         struct fuse_conn *fc = ff->fc;
3040         FUSE_ARGS(args);
3041         struct fuse_fallocate_in inarg = {
3042                 .fh = ff->fh,
3043                 .offset = offset,
3044                 .length = length,
3045                 .mode = mode
3046         };
3047         int err;
3048         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
3049                            (mode & FALLOC_FL_PUNCH_HOLE);
3050
3051         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
3052                 return -EOPNOTSUPP;
3053
3054         if (fc->no_fallocate)
3055                 return -EOPNOTSUPP;
3056
3057         if (lock_inode) {
3058                 inode_lock(inode);
3059                 if (mode & FALLOC_FL_PUNCH_HOLE) {
3060                         loff_t endbyte = offset + length - 1;
3061
3062                         err = fuse_writeback_range(inode, offset, endbyte);
3063                         if (err)
3064                                 goto out;
3065                 }
3066         }
3067
3068         if (!(mode & FALLOC_FL_KEEP_SIZE) &&
3069             offset + length > i_size_read(inode)) {
3070                 err = inode_newsize_ok(inode, offset + length);
3071                 if (err)
3072                         goto out;
3073         }
3074
3075         if (!(mode & FALLOC_FL_KEEP_SIZE))
3076                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3077
3078         args.opcode = FUSE_FALLOCATE;
3079         args.nodeid = ff->nodeid;
3080         args.in_numargs = 1;
3081         args.in_args[0].size = sizeof(inarg);
3082         args.in_args[0].value = &inarg;
3083         err = fuse_simple_request(fc, &args);
3084         if (err == -ENOSYS) {
3085                 fc->no_fallocate = 1;
3086                 err = -EOPNOTSUPP;
3087         }
3088         if (err)
3089                 goto out;
3090
3091         /* we could have extended the file */
3092         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3093                 bool changed = fuse_write_update_size(inode, offset + length);
3094
3095                 if (changed && fc->writeback_cache)
3096                         file_update_time(file);
3097         }
3098
3099         if (mode & FALLOC_FL_PUNCH_HOLE)
3100                 truncate_pagecache_range(inode, offset, offset + length - 1);
3101
3102         fuse_invalidate_attr(inode);
3103
3104 out:
3105         if (!(mode & FALLOC_FL_KEEP_SIZE))
3106                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3107
3108         if (lock_inode)
3109                 inode_unlock(inode);
3110
3111         return err;
3112 }
3113
3114 static ssize_t __fuse_copy_file_range(struct file *file_in, loff_t pos_in,
3115                                       struct file *file_out, loff_t pos_out,
3116                                       size_t len, unsigned int flags)
3117 {
3118         struct fuse_file *ff_in = file_in->private_data;
3119         struct fuse_file *ff_out = file_out->private_data;
3120         struct inode *inode_in = file_inode(file_in);
3121         struct inode *inode_out = file_inode(file_out);
3122         struct fuse_inode *fi_out = get_fuse_inode(inode_out);
3123         struct fuse_conn *fc = ff_in->fc;
3124         FUSE_ARGS(args);
3125         struct fuse_copy_file_range_in inarg = {
3126                 .fh_in = ff_in->fh,
3127                 .off_in = pos_in,
3128                 .nodeid_out = ff_out->nodeid,
3129                 .fh_out = ff_out->fh,
3130                 .off_out = pos_out,
3131                 .len = len,
3132                 .flags = flags
3133         };
3134         struct fuse_write_out outarg;
3135         ssize_t err;
3136         /* mark unstable when write-back is not used, and file_out gets
3137          * extended */
3138         bool is_unstable = (!fc->writeback_cache) &&
3139                            ((pos_out + len) > inode_out->i_size);
3140
3141         if (fc->no_copy_file_range)
3142                 return -EOPNOTSUPP;
3143
3144         if (file_inode(file_in)->i_sb != file_inode(file_out)->i_sb)
3145                 return -EXDEV;
3146
3147         if (fc->writeback_cache) {
3148                 inode_lock(inode_in);
3149                 err = fuse_writeback_range(inode_in, pos_in, pos_in + len);
3150                 inode_unlock(inode_in);
3151                 if (err)
3152                         return err;
3153         }
3154
3155         inode_lock(inode_out);
3156
3157         err = file_modified(file_out);
3158         if (err)
3159                 goto out;
3160
3161         if (fc->writeback_cache) {
3162                 err = fuse_writeback_range(inode_out, pos_out, pos_out + len);
3163                 if (err)
3164                         goto out;
3165         }
3166
3167         if (is_unstable)
3168                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3169
3170         args.opcode = FUSE_COPY_FILE_RANGE;
3171         args.nodeid = ff_in->nodeid;
3172         args.in_numargs = 1;
3173         args.in_args[0].size = sizeof(inarg);
3174         args.in_args[0].value = &inarg;
3175         args.out_numargs = 1;
3176         args.out_args[0].size = sizeof(outarg);
3177         args.out_args[0].value = &outarg;
3178         err = fuse_simple_request(fc, &args);
3179         if (err == -ENOSYS) {
3180                 fc->no_copy_file_range = 1;
3181                 err = -EOPNOTSUPP;
3182         }
3183         if (err)
3184                 goto out;
3185
3186         if (fc->writeback_cache) {
3187                 fuse_write_update_size(inode_out, pos_out + outarg.size);
3188                 file_update_time(file_out);
3189         }
3190
3191         fuse_invalidate_attr(inode_out);
3192
3193         err = outarg.size;
3194 out:
3195         if (is_unstable)
3196                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi_out->state);
3197
3198         inode_unlock(inode_out);
3199         file_accessed(file_in);
3200
3201         return err;
3202 }
3203
3204 static ssize_t fuse_copy_file_range(struct file *src_file, loff_t src_off,
3205                                     struct file *dst_file, loff_t dst_off,
3206                                     size_t len, unsigned int flags)
3207 {
3208         ssize_t ret;
3209
3210         ret = __fuse_copy_file_range(src_file, src_off, dst_file, dst_off,
3211                                      len, flags);
3212
3213         if (ret == -EOPNOTSUPP || ret == -EXDEV)
3214                 ret = generic_copy_file_range(src_file, src_off, dst_file,
3215                                               dst_off, len, flags);
3216         return ret;
3217 }
3218
3219 static const struct file_operations fuse_file_operations = {
3220         .llseek         = fuse_file_llseek,
3221         .read_iter      = fuse_file_read_iter,
3222         .write_iter     = fuse_file_write_iter,
3223         .mmap           = fuse_file_mmap,
3224         .open           = fuse_open,
3225         .flush          = fuse_flush,
3226         .release        = fuse_release,
3227         .fsync          = fuse_fsync,
3228         .lock           = fuse_file_lock,
3229         .flock          = fuse_file_flock,
3230         .splice_read    = generic_file_splice_read,
3231         .splice_write   = iter_file_splice_write,
3232         .unlocked_ioctl = fuse_file_ioctl,
3233         .compat_ioctl   = fuse_file_compat_ioctl,
3234         .poll           = fuse_file_poll,
3235         .fallocate      = fuse_file_fallocate,
3236         .copy_file_range = fuse_copy_file_range,
3237 };
3238
3239 static const struct address_space_operations fuse_file_aops  = {
3240         .readpage       = fuse_readpage,
3241         .writepage      = fuse_writepage,
3242         .writepages     = fuse_writepages,
3243         .launder_page   = fuse_launder_page,
3244         .readpages      = fuse_readpages,
3245         .set_page_dirty = __set_page_dirty_nobuffers,
3246         .bmap           = fuse_bmap,
3247         .direct_IO      = fuse_direct_IO,
3248         .write_begin    = fuse_write_begin,
3249         .write_end      = fuse_write_end,
3250 };
3251
3252 void fuse_init_file_inode(struct inode *inode)
3253 {
3254         struct fuse_inode *fi = get_fuse_inode(inode);
3255
3256         inode->i_fop = &fuse_file_operations;
3257         inode->i_data.a_ops = &fuse_file_aops;
3258
3259         INIT_LIST_HEAD(&fi->write_files);
3260         INIT_LIST_HEAD(&fi->queued_writes);
3261         fi->writectr = 0;
3262         init_waitqueue_head(&fi->page_waitq);
3263         INIT_LIST_HEAD(&fi->writepages);
3264 }