fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20
  21 static const struct file_operations fuse_direct_io_file_operations;
  22
  23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  24                           int opcode, struct fuse_open_out *outargp)
  25 {
  26         struct fuse_open_in inarg;
  27         FUSE_ARGS(args);
  28
  29         memset(&inarg, 0, sizeof(inarg));
  30         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  31         if (!fc->atomic_o_trunc)
  32                 inarg.flags &= ~O_TRUNC;
  33         args.in.h.opcode = opcode;
  34         args.in.h.nodeid = nodeid;
  35         args.in.numargs = 1;
  36         args.in.args[0].size = sizeof(inarg);
  37         args.in.args[0].value = &inarg;
  38         args.out.numargs = 1;
  39         args.out.args[0].size = sizeof(*outargp);
  40         args.out.args[0].value = outargp;
  41
  42         return fuse_simple_request(fc, &args);
  43 }
  44
  45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  46 {
  47         struct fuse_file *ff;
  48
  49         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  50         if (unlikely(!ff))
  51                 return NULL;
  52
  53         ff->fc = fc;
  54         ff->reserved_req = fuse_request_alloc(0);
  55         if (unlikely(!ff->reserved_req)) {
  56                 kfree(ff);
  57                 return NULL;
  58         }
  59
  60         INIT_LIST_HEAD(&ff->write_entry);
  61         atomic_set(&ff->count, 0);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         spin_lock(&fc->lock);
  66         ff->kh = ++fc->khctr;
  67         spin_unlock(&fc->lock);
  68
  69         return ff;
  70 }
  71
  72 void fuse_file_free(struct fuse_file *ff)
  73 {
  74         fuse_request_free(ff->reserved_req);
  75         kfree(ff);
  76 }
  77
  78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  79 {
  80         atomic_inc(&ff->count);
  81         return ff;
  82 }
  83
  84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  85 {
  86         iput(req->misc.release.inode);
  87 }
  88
  89 static void fuse_file_put(struct fuse_file *ff, bool sync)
  90 {
  91         if (atomic_dec_and_test(&ff->count)) {
  92                 struct fuse_req *req = ff->reserved_req;
  93
  94                 if (ff->fc->no_open) {
  95                         /*
  96                          * Drop the release request when client does not
  97                          * implement 'open'
  98                          */
  99                         __clear_bit(FR_BACKGROUND, &req->flags);
 100                         iput(req->misc.release.inode);
 101                         fuse_put_request(ff->fc, req);
 102                 } else if (sync) {
 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         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 128         if (!fc->no_open || isdir) {
 129                 struct fuse_open_out outarg;
 130                 int err;
 131
 132                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 133                 if (!err) {
 134                         ff->fh = outarg.fh;
 135                         ff->open_flags = outarg.open_flags;
 136
 137                 } else if (err != -ENOSYS || isdir) {
 138                         fuse_file_free(ff);
 139                         return err;
 140                 } else {
 141                         fc->no_open = 1;
 142                 }
 143         }
 144
 145         if (isdir)
 146                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 147
 148         ff->nodeid = nodeid;
 149         file->private_data = fuse_file_get(ff);
 150
 151         return 0;
 152 }
 153 EXPORT_SYMBOL_GPL(fuse_do_open);
 154
 155 static void fuse_link_write_file(struct file *file)
 156 {
 157         struct inode *inode = file_inode(file);
 158         struct fuse_conn *fc = get_fuse_conn(inode);
 159         struct fuse_inode *fi = get_fuse_inode(inode);
 160         struct fuse_file *ff = file->private_data;
 161         /*
 162          * file may be written through mmap, so chain it onto the
 163          * inodes's write_file list
 164          */
 165         spin_lock(&fc->lock);
 166         if (list_empty(&ff->write_entry))
 167                 list_add(&ff->write_entry, &fi->write_files);
 168         spin_unlock(&fc->lock);
 169 }
 170
 171 void fuse_finish_open(struct inode *inode, struct file *file)
 172 {
 173         struct fuse_file *ff = file->private_data;
 174         struct fuse_conn *fc = get_fuse_conn(inode);
 175
 176         if (ff->open_flags & FOPEN_DIRECT_IO)
 177                 file->f_op = &fuse_direct_io_file_operations;
 178         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 179                 invalidate_inode_pages2(inode->i_mapping);
 180         if (ff->open_flags & FOPEN_NONSEEKABLE)
 181                 nonseekable_open(inode, file);
 182         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 183                 struct fuse_inode *fi = get_fuse_inode(inode);
 184
 185                 spin_lock(&fc->lock);
 186                 fi->attr_version = ++fc->attr_version;
 187                 i_size_write(inode, 0);
 188                 spin_unlock(&fc->lock);
 189                 fuse_invalidate_attr(inode);
 190                 if (fc->writeback_cache)
 191                         file_update_time(file);
 192         }
 193         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 194                 fuse_link_write_file(file);
 195 }
 196
 197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 198 {
 199         struct fuse_conn *fc = get_fuse_conn(inode);
 200         int err;
 201         bool lock_inode = (file->f_flags & O_TRUNC) &&
 202                           fc->atomic_o_trunc &&
 203                           fc->writeback_cache;
 204
 205         err = generic_file_open(inode, file);
 206         if (err)
 207                 return err;
 208
 209         if (lock_inode)
 210                 mutex_lock(&inode->i_mutex);
 211
 212         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 213
 214         if (!err)
 215                 fuse_finish_open(inode, file);
 216
 217         if (lock_inode)
 218                 mutex_unlock(&inode->i_mutex);
 219
 220         return err;
 221 }
 222
 223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 224 {
 225         struct fuse_conn *fc = ff->fc;
 226         struct fuse_req *req = ff->reserved_req;
 227         struct fuse_release_in *inarg = &req->misc.release.in;
 228
 229         spin_lock(&fc->lock);
 230         list_del(&ff->write_entry);
 231         if (!RB_EMPTY_NODE(&ff->polled_node))
 232                 rb_erase(&ff->polled_node, &fc->polled_files);
 233         spin_unlock(&fc->lock);
 234
 235         wake_up_interruptible_all(&ff->poll_wait);
 236
 237         inarg->fh = ff->fh;
 238         inarg->flags = flags;
 239         req->in.h.opcode = opcode;
 240         req->in.h.nodeid = ff->nodeid;
 241         req->in.numargs = 1;
 242         req->in.args[0].size = sizeof(struct fuse_release_in);
 243         req->in.args[0].value = inarg;
 244 }
 245
 246 void fuse_release_common(struct file *file, int opcode)
 247 {
 248         struct fuse_file *ff;
 249         struct fuse_req *req;
 250
 251         ff = file->private_data;
 252         if (unlikely(!ff))
 253                 return;
 254
 255         req = ff->reserved_req;
 256         fuse_prepare_release(ff, file->f_flags, opcode);
 257
 258         if (ff->flock) {
 259                 struct fuse_release_in *inarg = &req->misc.release.in;
 260                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 261                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 262                                                        (fl_owner_t) file);
 263         }
 264         /* Hold inode until release is finished */
 265         req->misc.release.inode = igrab(file_inode(file));
 266
 267         /*
 268          * Normally this will send the RELEASE request, however if
 269          * some asynchronous READ or WRITE requests are outstanding,
 270          * the sending will be delayed.
 271          *
 272          * Make the release synchronous if this is a fuseblk mount,
 273          * synchronous RELEASE is allowed (and desirable) in this case
 274          * because the server can be trusted not to screw up.
 275          */
 276         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 277 }
 278
 279 static int fuse_open(struct inode *inode, struct file *file)
 280 {
 281         return fuse_open_common(inode, file, false);
 282 }
 283
 284 static int fuse_release(struct inode *inode, struct file *file)
 285 {
 286         struct fuse_conn *fc = get_fuse_conn(inode);
 287
 288         /* see fuse_vma_close() for !writeback_cache case */
 289         if (fc->writeback_cache)
 290                 write_inode_now(inode, 1);
 291
 292         fuse_release_common(file, FUSE_RELEASE);
 293
 294         /* return value is ignored by VFS */
 295         return 0;
 296 }
 297
 298 void fuse_sync_release(struct fuse_file *ff, int flags)
 299 {
 300         WARN_ON(atomic_read(&ff->count) > 1);
 301         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 302         __set_bit(FR_FORCE, &ff->reserved_req->flags);
 303         __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
 304         fuse_request_send(ff->fc, ff->reserved_req);
 305         fuse_put_request(ff->fc, ff->reserved_req);
 306         kfree(ff);
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_sync_release);
 309
 310 /*
 311  * Scramble the ID space with XTEA, so that the value of the files_struct
 312  * pointer is not exposed to userspace.
 313  */
 314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 315 {
 316         u32 *k = fc->scramble_key;
 317         u64 v = (unsigned long) id;
 318         u32 v0 = v;
 319         u32 v1 = v >> 32;
 320         u32 sum = 0;
 321         int i;
 322
 323         for (i = 0; i < 32; i++) {
 324                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 325                 sum += 0x9E3779B9;
 326                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 327         }
 328
 329         return (u64) v0 + ((u64) v1 << 32);
 330 }
 331
 332 /*
 333  * Check if any page in a range is under writeback
 334  *
 335  * This is currently done by walking the list of writepage requests
 336  * for the inode, which can be pretty inefficient.
 337  */
 338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 339                                    pgoff_t idx_to)
 340 {
 341         struct fuse_conn *fc = get_fuse_conn(inode);
 342         struct fuse_inode *fi = get_fuse_inode(inode);
 343         struct fuse_req *req;
 344         bool found = false;
 345
 346         spin_lock(&fc->lock);
 347         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 348                 pgoff_t curr_index;
 349
 350                 BUG_ON(req->inode != inode);
 351                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 352                 if (idx_from < curr_index + req->num_pages &&
 353                     curr_index <= idx_to) {
 354                         found = true;
 355                         break;
 356                 }
 357         }
 358         spin_unlock(&fc->lock);
 359
 360         return found;
 361 }
 362
 363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 364 {
 365         return fuse_range_is_writeback(inode, index, index);
 366 }
 367
 368 /*
 369  * Wait for page writeback to be completed.
 370  *
 371  * Since fuse doesn't rely on the VM writeback tracking, this has to
 372  * use some other means.
 373  */
 374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 375 {
 376         struct fuse_inode *fi = get_fuse_inode(inode);
 377
 378         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 379         return 0;
 380 }
 381
 382 /*
 383  * Wait for all pending writepages on the inode to finish.
 384  *
 385  * This is currently done by blocking further writes with FUSE_NOWRITE
 386  * and waiting for all sent writes to complete.
 387  *
 388  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 389  * could conflict with truncation.
 390  */
 391 static void fuse_sync_writes(struct inode *inode)
 392 {
 393         fuse_set_nowrite(inode);
 394         fuse_release_nowrite(inode);
 395 }
 396
 397 static int fuse_flush(struct file *file, fl_owner_t id)
 398 {
 399         struct inode *inode = file_inode(file);
 400         struct fuse_conn *fc = get_fuse_conn(inode);
 401         struct fuse_file *ff = file->private_data;
 402         struct fuse_req *req;
 403         struct fuse_flush_in inarg;
 404         int err;
 405
 406         if (is_bad_inode(inode))
 407                 return -EIO;
 408
 409         if (fc->no_flush)
 410                 return 0;
 411
 412         err = write_inode_now(inode, 1);
 413         if (err)
 414                 return err;
 415
 416         mutex_lock(&inode->i_mutex);
 417         fuse_sync_writes(inode);
 418         mutex_unlock(&inode->i_mutex);
 419
 420         req = fuse_get_req_nofail_nopages(fc, file);
 421         memset(&inarg, 0, sizeof(inarg));
 422         inarg.fh = ff->fh;
 423         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 424         req->in.h.opcode = FUSE_FLUSH;
 425         req->in.h.nodeid = get_node_id(inode);
 426         req->in.numargs = 1;
 427         req->in.args[0].size = sizeof(inarg);
 428         req->in.args[0].value = &inarg;
 429         __set_bit(FR_FORCE, &req->flags);
 430         fuse_request_send(fc, req);
 431         err = req->out.h.error;
 432         fuse_put_request(fc, req);
 433         if (err == -ENOSYS) {
 434                 fc->no_flush = 1;
 435                 err = 0;
 436         }
 437         return err;
 438 }
 439
 440 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 441                       int datasync, int isdir)
 442 {
 443         struct inode *inode = file->f_mapping->host;
 444         struct fuse_conn *fc = get_fuse_conn(inode);
 445         struct fuse_file *ff = file->private_data;
 446         FUSE_ARGS(args);
 447         struct fuse_fsync_in inarg;
 448         int err;
 449
 450         if (is_bad_inode(inode))
 451                 return -EIO;
 452
 453         mutex_lock(&inode->i_mutex);
 454
 455         /*
 456          * Start writeback against all dirty pages of the inode, then
 457          * wait for all outstanding writes, before sending the FSYNC
 458          * request.
 459          */
 460         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 461         if (err)
 462                 goto out;
 463
 464         fuse_sync_writes(inode);
 465         err = sync_inode_metadata(inode, 1);
 466         if (err)
 467                 goto out;
 468
 469         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 470                 goto out;
 471
 472         memset(&inarg, 0, sizeof(inarg));
 473         inarg.fh = ff->fh;
 474         inarg.fsync_flags = datasync ? 1 : 0;
 475         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 476         args.in.h.nodeid = get_node_id(inode);
 477         args.in.numargs = 1;
 478         args.in.args[0].size = sizeof(inarg);
 479         args.in.args[0].value = &inarg;
 480         err = fuse_simple_request(fc, &args);
 481         if (err == -ENOSYS) {
 482                 if (isdir)
 483                         fc->no_fsyncdir = 1;
 484                 else
 485                         fc->no_fsync = 1;
 486                 err = 0;
 487         }
 488 out:
 489         mutex_unlock(&inode->i_mutex);
 490         return err;
 491 }
 492
 493 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 494                       int datasync)
 495 {
 496         return fuse_fsync_common(file, start, end, datasync, 0);
 497 }
 498
 499 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 500                     size_t count, int opcode)
 501 {
 502         struct fuse_read_in *inarg = &req->misc.read.in;
 503         struct fuse_file *ff = file->private_data;
 504
 505         inarg->fh = ff->fh;
 506         inarg->offset = pos;
 507         inarg->size = count;
 508         inarg->flags = file->f_flags;
 509         req->in.h.opcode = opcode;
 510         req->in.h.nodeid = ff->nodeid;
 511         req->in.numargs = 1;
 512         req->in.args[0].size = sizeof(struct fuse_read_in);
 513         req->in.args[0].value = inarg;
 514         req->out.argvar = 1;
 515         req->out.numargs = 1;
 516         req->out.args[0].size = count;
 517 }
 518
 519 static void fuse_release_user_pages(struct fuse_req *req, int write)
 520 {
 521         unsigned i;
 522
 523         for (i = 0; i < req->num_pages; i++) {
 524                 struct page *page = req->pages[i];
 525                 if (write)
 526                         set_page_dirty_lock(page);
 527                 put_page(page);
 528         }
 529 }
 530
 531 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 532 {
 533         if (io->err)
 534                 return io->err;
 535
 536         if (io->bytes >= 0 && io->write)
 537                 return -EIO;
 538
 539         return io->bytes < 0 ? io->size : io->bytes;
 540 }
 541
 542 /**
 543  * In case of short read, the caller sets 'pos' to the position of
 544  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 545  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 546  *
 547  * An example:
 548  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 549  * both submitted asynchronously. The first of them was ACKed by userspace as
 550  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 551  * second request was ACKed as short, e.g. only 1K was read, resulting in
 552  * pos == 33K.
 553  *
 554  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 555  * will be equal to the length of the longest contiguous fragment of
 556  * transferred data starting from the beginning of IO request.
 557  */
 558 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 559 {
 560         bool is_sync = is_sync_kiocb(io->iocb);
 561         int left;
 562
 563         spin_lock(&io->lock);
 564         if (err)
 565                 io->err = io->err ? : err;
 566         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 567                 io->bytes = pos;
 568
 569         left = --io->reqs;
 570         if (!left && is_sync)
 571                 complete(io->done);
 572         spin_unlock(&io->lock);
 573
 574         if (!left && !is_sync) {
 575                 ssize_t res = fuse_get_res_by_io(io);
 576
 577                 if (res >= 0) {
 578                         struct inode *inode = file_inode(io->iocb->ki_filp);
 579                         struct fuse_conn *fc = get_fuse_conn(inode);
 580                         struct fuse_inode *fi = get_fuse_inode(inode);
 581
 582                         spin_lock(&fc->lock);
 583                         fi->attr_version = ++fc->attr_version;
 584                         spin_unlock(&fc->lock);
 585                 }
 586
 587                 io->iocb->ki_complete(io->iocb, res, 0);
 588                 kfree(io);
 589         }
 590 }
 591
 592 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 593 {
 594         struct fuse_io_priv *io = req->io;
 595         ssize_t pos = -1;
 596
 597         fuse_release_user_pages(req, !io->write);
 598
 599         if (io->write) {
 600                 if (req->misc.write.in.size != req->misc.write.out.size)
 601                         pos = req->misc.write.in.offset - io->offset +
 602                                 req->misc.write.out.size;
 603         } else {
 604                 if (req->misc.read.in.size != req->out.args[0].size)
 605                         pos = req->misc.read.in.offset - io->offset +
 606                                 req->out.args[0].size;
 607         }
 608
 609         fuse_aio_complete(io, req->out.h.error, pos);
 610 }
 611
 612 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 613                 size_t num_bytes, struct fuse_io_priv *io)
 614 {
 615         spin_lock(&io->lock);
 616         io->size += num_bytes;
 617         io->reqs++;
 618         spin_unlock(&io->lock);
 619
 620         req->io = io;
 621         req->end = fuse_aio_complete_req;
 622
 623         __fuse_get_request(req);
 624         fuse_request_send_background(fc, req);
 625
 626         return num_bytes;
 627 }
 628
 629 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 630                              loff_t pos, size_t count, fl_owner_t owner)
 631 {
 632         struct file *file = io->file;
 633         struct fuse_file *ff = file->private_data;
 634         struct fuse_conn *fc = ff->fc;
 635
 636         fuse_read_fill(req, file, pos, count, FUSE_READ);
 637         if (owner != NULL) {
 638                 struct fuse_read_in *inarg = &req->misc.read.in;
 639
 640                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 641                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 642         }
 643
 644         if (io->async)
 645                 return fuse_async_req_send(fc, req, count, io);
 646
 647         fuse_request_send(fc, req);
 648         return req->out.args[0].size;
 649 }
 650
 651 static void fuse_read_update_size(struct inode *inode, loff_t size,
 652                                   u64 attr_ver)
 653 {
 654         struct fuse_conn *fc = get_fuse_conn(inode);
 655         struct fuse_inode *fi = get_fuse_inode(inode);
 656
 657         spin_lock(&fc->lock);
 658         if (attr_ver == fi->attr_version && size < inode->i_size &&
 659             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 660                 fi->attr_version = ++fc->attr_version;
 661                 i_size_write(inode, size);
 662         }
 663         spin_unlock(&fc->lock);
 664 }
 665
 666 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 667                             u64 attr_ver)
 668 {
 669         size_t num_read = req->out.args[0].size;
 670         struct fuse_conn *fc = get_fuse_conn(inode);
 671
 672         if (fc->writeback_cache) {
 673                 /*
 674                  * A hole in a file. Some data after the hole are in page cache,
 675                  * but have not reached the client fs yet. So, the hole is not
 676                  * present there.
 677                  */
 678                 int i;
 679                 int start_idx = num_read >> PAGE_CACHE_SHIFT;
 680                 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
 681
 682                 for (i = start_idx; i < req->num_pages; i++) {
 683                         zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
 684                         off = 0;
 685                 }
 686         } else {
 687                 loff_t pos = page_offset(req->pages[0]) + num_read;
 688                 fuse_read_update_size(inode, pos, attr_ver);
 689         }
 690 }
 691
 692 static int fuse_do_readpage(struct file *file, struct page *page)
 693 {
 694         struct fuse_io_priv io = { .async = 0, .file = file };
 695         struct inode *inode = page->mapping->host;
 696         struct fuse_conn *fc = get_fuse_conn(inode);
 697         struct fuse_req *req;
 698         size_t num_read;
 699         loff_t pos = page_offset(page);
 700         size_t count = PAGE_CACHE_SIZE;
 701         u64 attr_ver;
 702         int err;
 703
 704         /*
 705          * Page writeback can extend beyond the lifetime of the
 706          * page-cache page, so make sure we read a properly synced
 707          * page.
 708          */
 709         fuse_wait_on_page_writeback(inode, page->index);
 710
 711         req = fuse_get_req(fc, 1);
 712         if (IS_ERR(req))
 713                 return PTR_ERR(req);
 714
 715         attr_ver = fuse_get_attr_version(fc);
 716
 717         req->out.page_zeroing = 1;
 718         req->out.argpages = 1;
 719         req->num_pages = 1;
 720         req->pages[0] = page;
 721         req->page_descs[0].length = count;
 722         num_read = fuse_send_read(req, &io, pos, count, NULL);
 723         err = req->out.h.error;
 724
 725         if (!err) {
 726                 /*
 727                  * Short read means EOF.  If file size is larger, truncate it
 728                  */
 729                 if (num_read < count)
 730                         fuse_short_read(req, inode, attr_ver);
 731
 732                 SetPageUptodate(page);
 733         }
 734
 735         fuse_put_request(fc, req);
 736
 737         return err;
 738 }
 739
 740 static int fuse_readpage(struct file *file, struct page *page)
 741 {
 742         struct inode *inode = page->mapping->host;
 743         int err;
 744
 745         err = -EIO;
 746         if (is_bad_inode(inode))
 747                 goto out;
 748
 749         err = fuse_do_readpage(file, page);
 750         fuse_invalidate_atime(inode);
 751  out:
 752         unlock_page(page);
 753         return err;
 754 }
 755
 756 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 757 {
 758         int i;
 759         size_t count = req->misc.read.in.size;
 760         size_t num_read = req->out.args[0].size;
 761         struct address_space *mapping = NULL;
 762
 763         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 764                 mapping = req->pages[i]->mapping;
 765
 766         if (mapping) {
 767                 struct inode *inode = mapping->host;
 768
 769                 /*
 770                  * Short read means EOF. If file size is larger, truncate it
 771                  */
 772                 if (!req->out.h.error && num_read < count)
 773                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 774
 775                 fuse_invalidate_atime(inode);
 776         }
 777
 778         for (i = 0; i < req->num_pages; i++) {
 779                 struct page *page = req->pages[i];
 780                 if (!req->out.h.error)
 781                         SetPageUptodate(page);
 782                 else
 783                         SetPageError(page);
 784                 unlock_page(page);
 785                 page_cache_release(page);
 786         }
 787         if (req->ff)
 788                 fuse_file_put(req->ff, false);
 789 }
 790
 791 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 792 {
 793         struct fuse_file *ff = file->private_data;
 794         struct fuse_conn *fc = ff->fc;
 795         loff_t pos = page_offset(req->pages[0]);
 796         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 797
 798         req->out.argpages = 1;
 799         req->out.page_zeroing = 1;
 800         req->out.page_replace = 1;
 801         fuse_read_fill(req, file, pos, count, FUSE_READ);
 802         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 803         if (fc->async_read) {
 804                 req->ff = fuse_file_get(ff);
 805                 req->end = fuse_readpages_end;
 806                 fuse_request_send_background(fc, req);
 807         } else {
 808                 fuse_request_send(fc, req);
 809                 fuse_readpages_end(fc, req);
 810                 fuse_put_request(fc, req);
 811         }
 812 }
 813
 814 struct fuse_fill_data {
 815         struct fuse_req *req;
 816         struct file *file;
 817         struct inode *inode;
 818         unsigned nr_pages;
 819 };
 820
 821 static int fuse_readpages_fill(void *_data, struct page *page)
 822 {
 823         struct fuse_fill_data *data = _data;
 824         struct fuse_req *req = data->req;
 825         struct inode *inode = data->inode;
 826         struct fuse_conn *fc = get_fuse_conn(inode);
 827
 828         fuse_wait_on_page_writeback(inode, page->index);
 829
 830         if (req->num_pages &&
 831             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 832              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 833              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 834                 int nr_alloc = min_t(unsigned, data->nr_pages,
 835                                      FUSE_MAX_PAGES_PER_REQ);
 836                 fuse_send_readpages(req, data->file);
 837                 if (fc->async_read)
 838                         req = fuse_get_req_for_background(fc, nr_alloc);
 839                 else
 840                         req = fuse_get_req(fc, nr_alloc);
 841
 842                 data->req = req;
 843                 if (IS_ERR(req)) {
 844                         unlock_page(page);
 845                         return PTR_ERR(req);
 846                 }
 847         }
 848
 849         if (WARN_ON(req->num_pages >= req->max_pages)) {
 850                 fuse_put_request(fc, req);
 851                 return -EIO;
 852         }
 853
 854         page_cache_get(page);
 855         req->pages[req->num_pages] = page;
 856         req->page_descs[req->num_pages].length = PAGE_SIZE;
 857         req->num_pages++;
 858         data->nr_pages--;
 859         return 0;
 860 }
 861
 862 static int fuse_readpages(struct file *file, struct address_space *mapping,
 863                           struct list_head *pages, unsigned nr_pages)
 864 {
 865         struct inode *inode = mapping->host;
 866         struct fuse_conn *fc = get_fuse_conn(inode);
 867         struct fuse_fill_data data;
 868         int err;
 869         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 870
 871         err = -EIO;
 872         if (is_bad_inode(inode))
 873                 goto out;
 874
 875         data.file = file;
 876         data.inode = inode;
 877         if (fc->async_read)
 878                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 879         else
 880                 data.req = fuse_get_req(fc, nr_alloc);
 881         data.nr_pages = nr_pages;
 882         err = PTR_ERR(data.req);
 883         if (IS_ERR(data.req))
 884                 goto out;
 885
 886         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 887         if (!err) {
 888                 if (data.req->num_pages)
 889                         fuse_send_readpages(data.req, file);
 890                 else
 891                         fuse_put_request(fc, data.req);
 892         }
 893 out:
 894         return err;
 895 }
 896
 897 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 898 {
 899         struct inode *inode = iocb->ki_filp->f_mapping->host;
 900         struct fuse_conn *fc = get_fuse_conn(inode);
 901
 902         /*
 903          * In auto invalidate mode, always update attributes on read.
 904          * Otherwise, only update if we attempt to read past EOF (to ensure
 905          * i_size is up to date).
 906          */
 907         if (fc->auto_inval_data ||
 908             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 909                 int err;
 910                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 911                 if (err)
 912                         return err;
 913         }
 914
 915         return generic_file_read_iter(iocb, to);
 916 }
 917
 918 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 919                             loff_t pos, size_t count)
 920 {
 921         struct fuse_write_in *inarg = &req->misc.write.in;
 922         struct fuse_write_out *outarg = &req->misc.write.out;
 923
 924         inarg->fh = ff->fh;
 925         inarg->offset = pos;
 926         inarg->size = count;
 927         req->in.h.opcode = FUSE_WRITE;
 928         req->in.h.nodeid = ff->nodeid;
 929         req->in.numargs = 2;
 930         if (ff->fc->minor < 9)
 931                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 932         else
 933                 req->in.args[0].size = sizeof(struct fuse_write_in);
 934         req->in.args[0].value = inarg;
 935         req->in.args[1].size = count;
 936         req->out.numargs = 1;
 937         req->out.args[0].size = sizeof(struct fuse_write_out);
 938         req->out.args[0].value = outarg;
 939 }
 940
 941 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 942                               loff_t pos, size_t count, fl_owner_t owner)
 943 {
 944         struct file *file = io->file;
 945         struct fuse_file *ff = file->private_data;
 946         struct fuse_conn *fc = ff->fc;
 947         struct fuse_write_in *inarg = &req->misc.write.in;
 948
 949         fuse_write_fill(req, ff, pos, count);
 950         inarg->flags = file->f_flags;
 951         if (owner != NULL) {
 952                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 953                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 954         }
 955
 956         if (io->async)
 957                 return fuse_async_req_send(fc, req, count, io);
 958
 959         fuse_request_send(fc, req);
 960         return req->misc.write.out.size;
 961 }
 962
 963 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 964 {
 965         struct fuse_conn *fc = get_fuse_conn(inode);
 966         struct fuse_inode *fi = get_fuse_inode(inode);
 967         bool ret = false;
 968
 969         spin_lock(&fc->lock);
 970         fi->attr_version = ++fc->attr_version;
 971         if (pos > inode->i_size) {
 972                 i_size_write(inode, pos);
 973                 ret = true;
 974         }
 975         spin_unlock(&fc->lock);
 976
 977         return ret;
 978 }
 979
 980 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 981                                     struct inode *inode, loff_t pos,
 982                                     size_t count)
 983 {
 984         size_t res;
 985         unsigned offset;
 986         unsigned i;
 987         struct fuse_io_priv io = { .async = 0, .file = file };
 988
 989         for (i = 0; i < req->num_pages; i++)
 990                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 991
 992         res = fuse_send_write(req, &io, pos, count, NULL);
 993
 994         offset = req->page_descs[0].offset;
 995         count = res;
 996         for (i = 0; i < req->num_pages; i++) {
 997                 struct page *page = req->pages[i];
 998
 999                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1000                         SetPageUptodate(page);
1001
1002                 if (count > PAGE_CACHE_SIZE - offset)
1003                         count -= PAGE_CACHE_SIZE - offset;
1004                 else
1005                         count = 0;
1006                 offset = 0;
1007
1008                 unlock_page(page);
1009                 page_cache_release(page);
1010         }
1011
1012         return res;
1013 }
1014
1015 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1016                                struct address_space *mapping,
1017                                struct iov_iter *ii, loff_t pos)
1018 {
1019         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1020         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1021         size_t count = 0;
1022         int err;
1023
1024         req->in.argpages = 1;
1025         req->page_descs[0].offset = offset;
1026
1027         do {
1028                 size_t tmp;
1029                 struct page *page;
1030                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1031                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1032                                      iov_iter_count(ii));
1033
1034                 bytes = min_t(size_t, bytes, fc->max_write - count);
1035
1036  again:
1037                 err = -EFAULT;
1038                 if (iov_iter_fault_in_readable(ii, bytes))
1039                         break;
1040
1041                 err = -ENOMEM;
1042                 page = grab_cache_page_write_begin(mapping, index, 0);
1043                 if (!page)
1044                         break;
1045
1046                 if (mapping_writably_mapped(mapping))
1047                         flush_dcache_page(page);
1048
1049                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1050                 flush_dcache_page(page);
1051
1052                 if (!tmp) {
1053                         unlock_page(page);
1054                         page_cache_release(page);
1055                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1056                         goto again;
1057                 }
1058
1059                 err = 0;
1060                 req->pages[req->num_pages] = page;
1061                 req->page_descs[req->num_pages].length = tmp;
1062                 req->num_pages++;
1063
1064                 iov_iter_advance(ii, tmp);
1065                 count += tmp;
1066                 pos += tmp;
1067                 offset += tmp;
1068                 if (offset == PAGE_CACHE_SIZE)
1069                         offset = 0;
1070
1071                 if (!fc->big_writes)
1072                         break;
1073         } while (iov_iter_count(ii) && count < fc->max_write &&
1074                  req->num_pages < req->max_pages && offset == 0);
1075
1076         return count > 0 ? count : err;
1077 }
1078
1079 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1080 {
1081         return min_t(unsigned,
1082                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1083                      (pos >> PAGE_CACHE_SHIFT) + 1,
1084                      FUSE_MAX_PAGES_PER_REQ);
1085 }
1086
1087 static ssize_t fuse_perform_write(struct file *file,
1088                                   struct address_space *mapping,
1089                                   struct iov_iter *ii, loff_t pos)
1090 {
1091         struct inode *inode = mapping->host;
1092         struct fuse_conn *fc = get_fuse_conn(inode);
1093         struct fuse_inode *fi = get_fuse_inode(inode);
1094         int err = 0;
1095         ssize_t res = 0;
1096
1097         if (is_bad_inode(inode))
1098                 return -EIO;
1099
1100         if (inode->i_size < pos + iov_iter_count(ii))
1101                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1102
1103         do {
1104                 struct fuse_req *req;
1105                 ssize_t count;
1106                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1107
1108                 req = fuse_get_req(fc, nr_pages);
1109                 if (IS_ERR(req)) {
1110                         err = PTR_ERR(req);
1111                         break;
1112                 }
1113
1114                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1115                 if (count <= 0) {
1116                         err = count;
1117                 } else {
1118                         size_t num_written;
1119
1120                         num_written = fuse_send_write_pages(req, file, inode,
1121                                                             pos, count);
1122                         err = req->out.h.error;
1123                         if (!err) {
1124                                 res += num_written;
1125                                 pos += num_written;
1126
1127                                 /* break out of the loop on short write */
1128                                 if (num_written != count)
1129                                         err = -EIO;
1130                         }
1131                 }
1132                 fuse_put_request(fc, req);
1133         } while (!err && iov_iter_count(ii));
1134
1135         if (res > 0)
1136                 fuse_write_update_size(inode, pos);
1137
1138         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1139         fuse_invalidate_attr(inode);
1140
1141         return res > 0 ? res : err;
1142 }
1143
1144 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1145 {
1146         struct file *file = iocb->ki_filp;
1147         struct address_space *mapping = file->f_mapping;
1148         ssize_t written = 0;
1149         ssize_t written_buffered = 0;
1150         struct inode *inode = mapping->host;
1151         ssize_t err;
1152         loff_t endbyte = 0;
1153
1154         if (get_fuse_conn(inode)->writeback_cache) {
1155                 /* Update size (EOF optimization) and mode (SUID clearing) */
1156                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1157                 if (err)
1158                         return err;
1159
1160                 return generic_file_write_iter(iocb, from);
1161         }
1162
1163         mutex_lock(&inode->i_mutex);
1164
1165         /* We can write back this queue in page reclaim */
1166         current->backing_dev_info = inode_to_bdi(inode);
1167
1168         err = generic_write_checks(iocb, from);
1169         if (err <= 0)
1170                 goto out;
1171
1172         err = file_remove_privs(file);
1173         if (err)
1174                 goto out;
1175
1176         err = file_update_time(file);
1177         if (err)
1178                 goto out;
1179
1180         if (iocb->ki_flags & IOCB_DIRECT) {
1181                 loff_t pos = iocb->ki_pos;
1182                 written = generic_file_direct_write(iocb, from, pos);
1183                 if (written < 0 || !iov_iter_count(from))
1184                         goto out;
1185
1186                 pos += written;
1187
1188                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1189                 if (written_buffered < 0) {
1190                         err = written_buffered;
1191                         goto out;
1192                 }
1193                 endbyte = pos + written_buffered - 1;
1194
1195                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1196                                                    endbyte);
1197                 if (err)
1198                         goto out;
1199
1200                 invalidate_mapping_pages(file->f_mapping,
1201                                          pos >> PAGE_CACHE_SHIFT,
1202                                          endbyte >> PAGE_CACHE_SHIFT);
1203
1204                 written += written_buffered;
1205                 iocb->ki_pos = pos + written_buffered;
1206         } else {
1207                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1208                 if (written >= 0)
1209                         iocb->ki_pos += written;
1210         }
1211 out:
1212         current->backing_dev_info = NULL;
1213         mutex_unlock(&inode->i_mutex);
1214
1215         return written ? written : err;
1216 }
1217
1218 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1219                 unsigned index, unsigned nr_pages)
1220 {
1221         int i;
1222
1223         for (i = index; i < index + nr_pages; i++)
1224                 req->page_descs[i].length = PAGE_SIZE -
1225                         req->page_descs[i].offset;
1226 }
1227
1228 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1229 {
1230         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1231 }
1232
1233 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1234                                         size_t max_size)
1235 {
1236         return min(iov_iter_single_seg_count(ii), max_size);
1237 }
1238
1239 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1240                                size_t *nbytesp, int write)
1241 {
1242         size_t nbytes = 0;  /* # bytes already packed in req */
1243
1244         /* Special case for kernel I/O: can copy directly into the buffer */
1245         if (ii->type & ITER_KVEC) {
1246                 unsigned long user_addr = fuse_get_user_addr(ii);
1247                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1248
1249                 if (write)
1250                         req->in.args[1].value = (void *) user_addr;
1251                 else
1252                         req->out.args[0].value = (void *) user_addr;
1253
1254                 iov_iter_advance(ii, frag_size);
1255                 *nbytesp = frag_size;
1256                 return 0;
1257         }
1258
1259         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1260                 unsigned npages;
1261                 size_t start;
1262                 ssize_t ret = iov_iter_get_pages(ii,
1263                                         &req->pages[req->num_pages],
1264                                         *nbytesp - nbytes,
1265                                         req->max_pages - req->num_pages,
1266                                         &start);
1267                 if (ret < 0)
1268                         return ret;
1269
1270                 iov_iter_advance(ii, ret);
1271                 nbytes += ret;
1272
1273                 ret += start;
1274                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1275
1276                 req->page_descs[req->num_pages].offset = start;
1277                 fuse_page_descs_length_init(req, req->num_pages, npages);
1278
1279                 req->num_pages += npages;
1280                 req->page_descs[req->num_pages - 1].length -=
1281                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1282         }
1283
1284         if (write)
1285                 req->in.argpages = 1;
1286         else
1287                 req->out.argpages = 1;
1288
1289         *nbytesp = nbytes;
1290
1291         return 0;
1292 }
1293
1294 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1295 {
1296         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1297 }
1298
1299 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1300                        loff_t *ppos, int flags)
1301 {
1302         int write = flags & FUSE_DIO_WRITE;
1303         int cuse = flags & FUSE_DIO_CUSE;
1304         struct file *file = io->file;
1305         struct inode *inode = file->f_mapping->host;
1306         struct fuse_file *ff = file->private_data;
1307         struct fuse_conn *fc = ff->fc;
1308         size_t nmax = write ? fc->max_write : fc->max_read;
1309         loff_t pos = *ppos;
1310         size_t count = iov_iter_count(iter);
1311         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1312         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1313         ssize_t res = 0;
1314         struct fuse_req *req;
1315
1316         if (io->async)
1317                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1318         else
1319                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1320         if (IS_ERR(req))
1321                 return PTR_ERR(req);
1322
1323         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1324                 if (!write)
1325                         mutex_lock(&inode->i_mutex);
1326                 fuse_sync_writes(inode);
1327                 if (!write)
1328                         mutex_unlock(&inode->i_mutex);
1329         }
1330
1331         while (count) {
1332                 size_t nres;
1333                 fl_owner_t owner = current->files;
1334                 size_t nbytes = min(count, nmax);
1335                 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1336                 if (err) {
1337                         res = err;
1338                         break;
1339                 }
1340
1341                 if (write)
1342                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1343                 else
1344                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1345
1346                 if (!io->async)
1347                         fuse_release_user_pages(req, !write);
1348                 if (req->out.h.error) {
1349                         if (!res)
1350                                 res = req->out.h.error;
1351                         break;
1352                 } else if (nres > nbytes) {
1353                         res = -EIO;
1354                         break;
1355                 }
1356                 count -= nres;
1357                 res += nres;
1358                 pos += nres;
1359                 if (nres != nbytes)
1360                         break;
1361                 if (count) {
1362                         fuse_put_request(fc, req);
1363                         if (io->async)
1364                                 req = fuse_get_req_for_background(fc,
1365                                         fuse_iter_npages(iter));
1366                         else
1367                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1368                         if (IS_ERR(req))
1369                                 break;
1370                 }
1371         }
1372         if (!IS_ERR(req))
1373                 fuse_put_request(fc, req);
1374         if (res > 0)
1375                 *ppos = pos;
1376
1377         return res;
1378 }
1379 EXPORT_SYMBOL_GPL(fuse_direct_io);
1380
1381 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1382                                   struct iov_iter *iter,
1383                                   loff_t *ppos)
1384 {
1385         ssize_t res;
1386         struct file *file = io->file;
1387         struct inode *inode = file_inode(file);
1388
1389         if (is_bad_inode(inode))
1390                 return -EIO;
1391
1392         res = fuse_direct_io(io, iter, ppos, 0);
1393
1394         fuse_invalidate_attr(inode);
1395
1396         return res;
1397 }
1398
1399 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1400 {
1401         struct fuse_io_priv io = { .async = 0, .file = iocb->ki_filp };
1402         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1403 }
1404
1405 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1406 {
1407         struct file *file = iocb->ki_filp;
1408         struct inode *inode = file_inode(file);
1409         struct fuse_io_priv io = { .async = 0, .file = file };
1410         ssize_t res;
1411
1412         if (is_bad_inode(inode))
1413                 return -EIO;
1414
1415         /* Don't allow parallel writes to the same file */
1416         mutex_lock(&inode->i_mutex);
1417         res = generic_write_checks(iocb, from);
1418         if (res > 0)
1419                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1420         fuse_invalidate_attr(inode);
1421         if (res > 0)
1422                 fuse_write_update_size(inode, iocb->ki_pos);
1423         mutex_unlock(&inode->i_mutex);
1424
1425         return res;
1426 }
1427
1428 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1429 {
1430         int i;
1431
1432         for (i = 0; i < req->num_pages; i++)
1433                 __free_page(req->pages[i]);
1434
1435         if (req->ff)
1436                 fuse_file_put(req->ff, false);
1437 }
1438
1439 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1440 {
1441         struct inode *inode = req->inode;
1442         struct fuse_inode *fi = get_fuse_inode(inode);
1443         struct backing_dev_info *bdi = inode_to_bdi(inode);
1444         int i;
1445
1446         list_del(&req->writepages_entry);
1447         for (i = 0; i < req->num_pages; i++) {
1448                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1449                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1450                 wb_writeout_inc(&bdi->wb);
1451         }
1452         wake_up(&fi->page_waitq);
1453 }
1454
1455 /* Called under fc->lock, may release and reacquire it */
1456 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1457                                 loff_t size)
1458 __releases(fc->lock)
1459 __acquires(fc->lock)
1460 {
1461         struct fuse_inode *fi = get_fuse_inode(req->inode);
1462         struct fuse_write_in *inarg = &req->misc.write.in;
1463         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1464
1465         if (!fc->connected)
1466                 goto out_free;
1467
1468         if (inarg->offset + data_size <= size) {
1469                 inarg->size = data_size;
1470         } else if (inarg->offset < size) {
1471                 inarg->size = size - inarg->offset;
1472         } else {
1473                 /* Got truncated off completely */
1474                 goto out_free;
1475         }
1476
1477         req->in.args[1].size = inarg->size;
1478         fi->writectr++;
1479         fuse_request_send_background_locked(fc, req);
1480         return;
1481
1482  out_free:
1483         fuse_writepage_finish(fc, req);
1484         spin_unlock(&fc->lock);
1485         fuse_writepage_free(fc, req);
1486         fuse_put_request(fc, req);
1487         spin_lock(&fc->lock);
1488 }
1489
1490 /*
1491  * If fi->writectr is positive (no truncate or fsync going on) send
1492  * all queued writepage requests.
1493  *
1494  * Called with fc->lock
1495  */
1496 void fuse_flush_writepages(struct inode *inode)
1497 __releases(fc->lock)
1498 __acquires(fc->lock)
1499 {
1500         struct fuse_conn *fc = get_fuse_conn(inode);
1501         struct fuse_inode *fi = get_fuse_inode(inode);
1502         size_t crop = i_size_read(inode);
1503         struct fuse_req *req;
1504
1505         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1506                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1507                 list_del_init(&req->list);
1508                 fuse_send_writepage(fc, req, crop);
1509         }
1510 }
1511
1512 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1513 {
1514         struct inode *inode = req->inode;
1515         struct fuse_inode *fi = get_fuse_inode(inode);
1516
1517         mapping_set_error(inode->i_mapping, req->out.h.error);
1518         spin_lock(&fc->lock);
1519         while (req->misc.write.next) {
1520                 struct fuse_conn *fc = get_fuse_conn(inode);
1521                 struct fuse_write_in *inarg = &req->misc.write.in;
1522                 struct fuse_req *next = req->misc.write.next;
1523                 req->misc.write.next = next->misc.write.next;
1524                 next->misc.write.next = NULL;
1525                 next->ff = fuse_file_get(req->ff);
1526                 list_add(&next->writepages_entry, &fi->writepages);
1527
1528                 /*
1529                  * Skip fuse_flush_writepages() to make it easy to crop requests
1530                  * based on primary request size.
1531                  *
1532                  * 1st case (trivial): there are no concurrent activities using
1533                  * fuse_set/release_nowrite.  Then we're on safe side because
1534                  * fuse_flush_writepages() would call fuse_send_writepage()
1535                  * anyway.
1536                  *
1537                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1538                  * now for completion of all in-flight requests.  This happens
1539                  * rarely and no more than once per page, so this should be
1540                  * okay.
1541                  *
1542                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1543                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1544                  * that fuse_set_nowrite returned implies that all in-flight
1545                  * requests were completed along with all of their secondary
1546                  * requests.  Further primary requests are blocked by negative
1547                  * writectr.  Hence there cannot be any in-flight requests and
1548                  * no invocations of fuse_writepage_end() while we're in
1549                  * fuse_set_nowrite..fuse_release_nowrite section.
1550                  */
1551                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1552         }
1553         fi->writectr--;
1554         fuse_writepage_finish(fc, req);
1555         spin_unlock(&fc->lock);
1556         fuse_writepage_free(fc, req);
1557 }
1558
1559 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1560                                                struct fuse_inode *fi)
1561 {
1562         struct fuse_file *ff = NULL;
1563
1564         spin_lock(&fc->lock);
1565         if (!list_empty(&fi->write_files)) {
1566                 ff = list_entry(fi->write_files.next, struct fuse_file,
1567                                 write_entry);
1568                 fuse_file_get(ff);
1569         }
1570         spin_unlock(&fc->lock);
1571
1572         return ff;
1573 }
1574
1575 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1576                                              struct fuse_inode *fi)
1577 {
1578         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1579         WARN_ON(!ff);
1580         return ff;
1581 }
1582
1583 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1584 {
1585         struct fuse_conn *fc = get_fuse_conn(inode);
1586         struct fuse_inode *fi = get_fuse_inode(inode);
1587         struct fuse_file *ff;
1588         int err;
1589
1590         ff = __fuse_write_file_get(fc, fi);
1591         err = fuse_flush_times(inode, ff);
1592         if (ff)
1593                 fuse_file_put(ff, 0);
1594
1595         return err;
1596 }
1597
1598 static int fuse_writepage_locked(struct page *page)
1599 {
1600         struct address_space *mapping = page->mapping;
1601         struct inode *inode = mapping->host;
1602         struct fuse_conn *fc = get_fuse_conn(inode);
1603         struct fuse_inode *fi = get_fuse_inode(inode);
1604         struct fuse_req *req;
1605         struct page *tmp_page;
1606         int error = -ENOMEM;
1607
1608         set_page_writeback(page);
1609
1610         req = fuse_request_alloc_nofs(1);
1611         if (!req)
1612                 goto err;
1613
1614         /* writeback always goes to bg_queue */
1615         __set_bit(FR_BACKGROUND, &req->flags);
1616         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1617         if (!tmp_page)
1618                 goto err_free;
1619
1620         error = -EIO;
1621         req->ff = fuse_write_file_get(fc, fi);
1622         if (!req->ff)
1623                 goto err_nofile;
1624
1625         fuse_write_fill(req, req->ff, page_offset(page), 0);
1626
1627         copy_highpage(tmp_page, page);
1628         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1629         req->misc.write.next = NULL;
1630         req->in.argpages = 1;
1631         req->num_pages = 1;
1632         req->pages[0] = tmp_page;
1633         req->page_descs[0].offset = 0;
1634         req->page_descs[0].length = PAGE_SIZE;
1635         req->end = fuse_writepage_end;
1636         req->inode = inode;
1637
1638         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1639         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1640
1641         spin_lock(&fc->lock);
1642         list_add(&req->writepages_entry, &fi->writepages);
1643         list_add_tail(&req->list, &fi->queued_writes);
1644         fuse_flush_writepages(inode);
1645         spin_unlock(&fc->lock);
1646
1647         end_page_writeback(page);
1648
1649         return 0;
1650
1651 err_nofile:
1652         __free_page(tmp_page);
1653 err_free:
1654         fuse_request_free(req);
1655 err:
1656         end_page_writeback(page);
1657         return error;
1658 }
1659
1660 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1661 {
1662         int err;
1663
1664         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1665                 /*
1666                  * ->writepages() should be called for sync() and friends.  We
1667                  * should only get here on direct reclaim and then we are
1668                  * allowed to skip a page which is already in flight
1669                  */
1670                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1671
1672                 redirty_page_for_writepage(wbc, page);
1673                 return 0;
1674         }
1675
1676         err = fuse_writepage_locked(page);
1677         unlock_page(page);
1678
1679         return err;
1680 }
1681
1682 struct fuse_fill_wb_data {
1683         struct fuse_req *req;
1684         struct fuse_file *ff;
1685         struct inode *inode;
1686         struct page **orig_pages;
1687 };
1688
1689 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1690 {
1691         struct fuse_req *req = data->req;
1692         struct inode *inode = data->inode;
1693         struct fuse_conn *fc = get_fuse_conn(inode);
1694         struct fuse_inode *fi = get_fuse_inode(inode);
1695         int num_pages = req->num_pages;
1696         int i;
1697
1698         req->ff = fuse_file_get(data->ff);
1699         spin_lock(&fc->lock);
1700         list_add_tail(&req->list, &fi->queued_writes);
1701         fuse_flush_writepages(inode);
1702         spin_unlock(&fc->lock);
1703
1704         for (i = 0; i < num_pages; i++)
1705                 end_page_writeback(data->orig_pages[i]);
1706 }
1707
1708 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1709                                      struct page *page)
1710 {
1711         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1712         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1713         struct fuse_req *tmp;
1714         struct fuse_req *old_req;
1715         bool found = false;
1716         pgoff_t curr_index;
1717
1718         BUG_ON(new_req->num_pages != 0);
1719
1720         spin_lock(&fc->lock);
1721         list_del(&new_req->writepages_entry);
1722         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1723                 BUG_ON(old_req->inode != new_req->inode);
1724                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1725                 if (curr_index <= page->index &&
1726                     page->index < curr_index + old_req->num_pages) {
1727                         found = true;
1728                         break;
1729                 }
1730         }
1731         if (!found) {
1732                 list_add(&new_req->writepages_entry, &fi->writepages);
1733                 goto out_unlock;
1734         }
1735
1736         new_req->num_pages = 1;
1737         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1738                 BUG_ON(tmp->inode != new_req->inode);
1739                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1740                 if (tmp->num_pages == 1 &&
1741                     curr_index == page->index) {
1742                         old_req = tmp;
1743                 }
1744         }
1745
1746         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1747                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1748
1749                 copy_highpage(old_req->pages[0], page);
1750                 spin_unlock(&fc->lock);
1751
1752                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1753                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1754                 wb_writeout_inc(&bdi->wb);
1755                 fuse_writepage_free(fc, new_req);
1756                 fuse_request_free(new_req);
1757                 goto out;
1758         } else {
1759                 new_req->misc.write.next = old_req->misc.write.next;
1760                 old_req->misc.write.next = new_req;
1761         }
1762 out_unlock:
1763         spin_unlock(&fc->lock);
1764 out:
1765         return found;
1766 }
1767
1768 static int fuse_writepages_fill(struct page *page,
1769                 struct writeback_control *wbc, void *_data)
1770 {
1771         struct fuse_fill_wb_data *data = _data;
1772         struct fuse_req *req = data->req;
1773         struct inode *inode = data->inode;
1774         struct fuse_conn *fc = get_fuse_conn(inode);
1775         struct page *tmp_page;
1776         bool is_writeback;
1777         int err;
1778
1779         if (!data->ff) {
1780                 err = -EIO;
1781                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1782                 if (!data->ff)
1783                         goto out_unlock;
1784         }
1785
1786         /*
1787          * Being under writeback is unlikely but possible.  For example direct
1788          * read to an mmaped fuse file will set the page dirty twice; once when
1789          * the pages are faulted with get_user_pages(), and then after the read
1790          * completed.
1791          */
1792         is_writeback = fuse_page_is_writeback(inode, page->index);
1793
1794         if (req && req->num_pages &&
1795             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1796              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1797              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1798                 fuse_writepages_send(data);
1799                 data->req = NULL;
1800         }
1801         err = -ENOMEM;
1802         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1803         if (!tmp_page)
1804                 goto out_unlock;
1805
1806         /*
1807          * The page must not be redirtied until the writeout is completed
1808          * (i.e. userspace has sent a reply to the write request).  Otherwise
1809          * there could be more than one temporary page instance for each real
1810          * page.
1811          *
1812          * This is ensured by holding the page lock in page_mkwrite() while
1813          * checking fuse_page_is_writeback().  We already hold the page lock
1814          * since clear_page_dirty_for_io() and keep it held until we add the
1815          * request to the fi->writepages list and increment req->num_pages.
1816          * After this fuse_page_is_writeback() will indicate that the page is
1817          * under writeback, so we can release the page lock.
1818          */
1819         if (data->req == NULL) {
1820                 struct fuse_inode *fi = get_fuse_inode(inode);
1821
1822                 err = -ENOMEM;
1823                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1824                 if (!req) {
1825                         __free_page(tmp_page);
1826                         goto out_unlock;
1827                 }
1828
1829                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1830                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1831                 req->misc.write.next = NULL;
1832                 req->in.argpages = 1;
1833                 __set_bit(FR_BACKGROUND, &req->flags);
1834                 req->num_pages = 0;
1835                 req->end = fuse_writepage_end;
1836                 req->inode = inode;
1837
1838                 spin_lock(&fc->lock);
1839                 list_add(&req->writepages_entry, &fi->writepages);
1840                 spin_unlock(&fc->lock);
1841
1842                 data->req = req;
1843         }
1844         set_page_writeback(page);
1845
1846         copy_highpage(tmp_page, page);
1847         req->pages[req->num_pages] = tmp_page;
1848         req->page_descs[req->num_pages].offset = 0;
1849         req->page_descs[req->num_pages].length = PAGE_SIZE;
1850
1851         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1852         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1853
1854         err = 0;
1855         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1856                 end_page_writeback(page);
1857                 data->req = NULL;
1858                 goto out_unlock;
1859         }
1860         data->orig_pages[req->num_pages] = page;
1861
1862         /*
1863          * Protected by fc->lock against concurrent access by
1864          * fuse_page_is_writeback().
1865          */
1866         spin_lock(&fc->lock);
1867         req->num_pages++;
1868         spin_unlock(&fc->lock);
1869
1870 out_unlock:
1871         unlock_page(page);
1872
1873         return err;
1874 }
1875
1876 static int fuse_writepages(struct address_space *mapping,
1877                            struct writeback_control *wbc)
1878 {
1879         struct inode *inode = mapping->host;
1880         struct fuse_fill_wb_data data;
1881         int err;
1882
1883         err = -EIO;
1884         if (is_bad_inode(inode))
1885                 goto out;
1886
1887         data.inode = inode;
1888         data.req = NULL;
1889         data.ff = NULL;
1890
1891         err = -ENOMEM;
1892         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1893                                   sizeof(struct page *),
1894                                   GFP_NOFS);
1895         if (!data.orig_pages)
1896                 goto out;
1897
1898         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1899         if (data.req) {
1900                 /* Ignore errors if we can write at least one page */
1901                 BUG_ON(!data.req->num_pages);
1902                 fuse_writepages_send(&data);
1903                 err = 0;
1904         }
1905         if (data.ff)
1906                 fuse_file_put(data.ff, false);
1907
1908         kfree(data.orig_pages);
1909 out:
1910         return err;
1911 }
1912
1913 /*
1914  * It's worthy to make sure that space is reserved on disk for the write,
1915  * but how to implement it without killing performance need more thinking.
1916  */
1917 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1918                 loff_t pos, unsigned len, unsigned flags,
1919                 struct page **pagep, void **fsdata)
1920 {
1921         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1922         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1923         struct page *page;
1924         loff_t fsize;
1925         int err = -ENOMEM;
1926
1927         WARN_ON(!fc->writeback_cache);
1928
1929         page = grab_cache_page_write_begin(mapping, index, flags);
1930         if (!page)
1931                 goto error;
1932
1933         fuse_wait_on_page_writeback(mapping->host, page->index);
1934
1935         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1936                 goto success;
1937         /*
1938          * Check if the start this page comes after the end of file, in which
1939          * case the readpage can be optimized away.
1940          */
1941         fsize = i_size_read(mapping->host);
1942         if (fsize <= (pos & PAGE_CACHE_MASK)) {
1943                 size_t off = pos & ~PAGE_CACHE_MASK;
1944                 if (off)
1945                         zero_user_segment(page, 0, off);
1946                 goto success;
1947         }
1948         err = fuse_do_readpage(file, page);
1949         if (err)
1950                 goto cleanup;
1951 success:
1952         *pagep = page;
1953         return 0;
1954
1955 cleanup:
1956         unlock_page(page);
1957         page_cache_release(page);
1958 error:
1959         return err;
1960 }
1961
1962 static int fuse_write_end(struct file *file, struct address_space *mapping,
1963                 loff_t pos, unsigned len, unsigned copied,
1964                 struct page *page, void *fsdata)
1965 {
1966         struct inode *inode = page->mapping->host;
1967
1968         if (!PageUptodate(page)) {
1969                 /* Zero any unwritten bytes at the end of the page */
1970                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
1971                 if (endoff)
1972                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
1973                 SetPageUptodate(page);
1974         }
1975
1976         fuse_write_update_size(inode, pos + copied);
1977         set_page_dirty(page);
1978         unlock_page(page);
1979         page_cache_release(page);
1980
1981         return copied;
1982 }
1983
1984 static int fuse_launder_page(struct page *page)
1985 {
1986         int err = 0;
1987         if (clear_page_dirty_for_io(page)) {
1988                 struct inode *inode = page->mapping->host;
1989                 err = fuse_writepage_locked(page);
1990                 if (!err)
1991                         fuse_wait_on_page_writeback(inode, page->index);
1992         }
1993         return err;
1994 }
1995
1996 /*
1997  * Write back dirty pages now, because there may not be any suitable
1998  * open files later
1999  */
2000 static void fuse_vma_close(struct vm_area_struct *vma)
2001 {
2002         filemap_write_and_wait(vma->vm_file->f_mapping);
2003 }
2004
2005 /*
2006  * Wait for writeback against this page to complete before allowing it
2007  * to be marked dirty again, and hence written back again, possibly
2008  * before the previous writepage completed.
2009  *
2010  * Block here, instead of in ->writepage(), so that the userspace fs
2011  * can only block processes actually operating on the filesystem.
2012  *
2013  * Otherwise unprivileged userspace fs would be able to block
2014  * unrelated:
2015  *
2016  * - page migration
2017  * - sync(2)
2018  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2019  */
2020 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2021 {
2022         struct page *page = vmf->page;
2023         struct inode *inode = file_inode(vma->vm_file);
2024
2025         file_update_time(vma->vm_file);
2026         lock_page(page);
2027         if (page->mapping != inode->i_mapping) {
2028                 unlock_page(page);
2029                 return VM_FAULT_NOPAGE;
2030         }
2031
2032         fuse_wait_on_page_writeback(inode, page->index);
2033         return VM_FAULT_LOCKED;
2034 }
2035
2036 static const struct vm_operations_struct fuse_file_vm_ops = {
2037         .close          = fuse_vma_close,
2038         .fault          = filemap_fault,
2039         .map_pages      = filemap_map_pages,
2040         .page_mkwrite   = fuse_page_mkwrite,
2041 };
2042
2043 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2044 {
2045         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2046                 fuse_link_write_file(file);
2047
2048         file_accessed(file);
2049         vma->vm_ops = &fuse_file_vm_ops;
2050         return 0;
2051 }
2052
2053 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2054 {
2055         /* Can't provide the coherency needed for MAP_SHARED */
2056         if (vma->vm_flags & VM_MAYSHARE)
2057                 return -ENODEV;
2058
2059         invalidate_inode_pages2(file->f_mapping);
2060
2061         return generic_file_mmap(file, vma);
2062 }
2063
2064 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2065                                   struct file_lock *fl)
2066 {
2067         switch (ffl->type) {
2068         case F_UNLCK:
2069                 break;
2070
2071         case F_RDLCK:
2072         case F_WRLCK:
2073                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2074                     ffl->end < ffl->start)
2075                         return -EIO;
2076
2077                 fl->fl_start = ffl->start;
2078                 fl->fl_end = ffl->end;
2079                 fl->fl_pid = ffl->pid;
2080                 break;
2081
2082         default:
2083                 return -EIO;
2084         }
2085         fl->fl_type = ffl->type;
2086         return 0;
2087 }
2088
2089 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2090                          const struct file_lock *fl, int opcode, pid_t pid,
2091                          int flock, struct fuse_lk_in *inarg)
2092 {
2093         struct inode *inode = file_inode(file);
2094         struct fuse_conn *fc = get_fuse_conn(inode);
2095         struct fuse_file *ff = file->private_data;
2096
2097         memset(inarg, 0, sizeof(*inarg));
2098         inarg->fh = ff->fh;
2099         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2100         inarg->lk.start = fl->fl_start;
2101         inarg->lk.end = fl->fl_end;
2102         inarg->lk.type = fl->fl_type;
2103         inarg->lk.pid = pid;
2104         if (flock)
2105                 inarg->lk_flags |= FUSE_LK_FLOCK;
2106         args->in.h.opcode = opcode;
2107         args->in.h.nodeid = get_node_id(inode);
2108         args->in.numargs = 1;
2109         args->in.args[0].size = sizeof(*inarg);
2110         args->in.args[0].value = inarg;
2111 }
2112
2113 static int fuse_getlk(struct file *file, struct file_lock *fl)
2114 {
2115         struct inode *inode = file_inode(file);
2116         struct fuse_conn *fc = get_fuse_conn(inode);
2117         FUSE_ARGS(args);
2118         struct fuse_lk_in inarg;
2119         struct fuse_lk_out outarg;
2120         int err;
2121
2122         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2123         args.out.numargs = 1;
2124         args.out.args[0].size = sizeof(outarg);
2125         args.out.args[0].value = &outarg;
2126         err = fuse_simple_request(fc, &args);
2127         if (!err)
2128                 err = convert_fuse_file_lock(&outarg.lk, fl);
2129
2130         return err;
2131 }
2132
2133 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2134 {
2135         struct inode *inode = file_inode(file);
2136         struct fuse_conn *fc = get_fuse_conn(inode);
2137         FUSE_ARGS(args);
2138         struct fuse_lk_in inarg;
2139         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2140         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2141         int err;
2142
2143         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2144                 /* NLM needs asynchronous locks, which we don't support yet */
2145                 return -ENOLCK;
2146         }
2147
2148         /* Unlock on close is handled by the flush method */
2149         if (fl->fl_flags & FL_CLOSE)
2150                 return 0;
2151
2152         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2153         err = fuse_simple_request(fc, &args);
2154
2155         /* locking is restartable */
2156         if (err == -EINTR)
2157                 err = -ERESTARTSYS;
2158
2159         return err;
2160 }
2161
2162 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2163 {
2164         struct inode *inode = file_inode(file);
2165         struct fuse_conn *fc = get_fuse_conn(inode);
2166         int err;
2167
2168         if (cmd == F_CANCELLK) {
2169                 err = 0;
2170         } else if (cmd == F_GETLK) {
2171                 if (fc->no_lock) {
2172                         posix_test_lock(file, fl);
2173                         err = 0;
2174                 } else
2175                         err = fuse_getlk(file, fl);
2176         } else {
2177                 if (fc->no_lock)
2178                         err = posix_lock_file(file, fl, NULL);
2179                 else
2180                         err = fuse_setlk(file, fl, 0);
2181         }
2182         return err;
2183 }
2184
2185 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2186 {
2187         struct inode *inode = file_inode(file);
2188         struct fuse_conn *fc = get_fuse_conn(inode);
2189         int err;
2190
2191         if (fc->no_flock) {
2192                 err = flock_lock_file_wait(file, fl);
2193         } else {
2194                 struct fuse_file *ff = file->private_data;
2195
2196                 /* emulate flock with POSIX locks */
2197                 ff->flock = true;
2198                 err = fuse_setlk(file, fl, 1);
2199         }
2200
2201         return err;
2202 }
2203
2204 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2205 {
2206         struct inode *inode = mapping->host;
2207         struct fuse_conn *fc = get_fuse_conn(inode);
2208         FUSE_ARGS(args);
2209         struct fuse_bmap_in inarg;
2210         struct fuse_bmap_out outarg;
2211         int err;
2212
2213         if (!inode->i_sb->s_bdev || fc->no_bmap)
2214                 return 0;
2215
2216         memset(&inarg, 0, sizeof(inarg));
2217         inarg.block = block;
2218         inarg.blocksize = inode->i_sb->s_blocksize;
2219         args.in.h.opcode = FUSE_BMAP;
2220         args.in.h.nodeid = get_node_id(inode);
2221         args.in.numargs = 1;
2222         args.in.args[0].size = sizeof(inarg);
2223         args.in.args[0].value = &inarg;
2224         args.out.numargs = 1;
2225         args.out.args[0].size = sizeof(outarg);
2226         args.out.args[0].value = &outarg;
2227         err = fuse_simple_request(fc, &args);
2228         if (err == -ENOSYS)
2229                 fc->no_bmap = 1;
2230
2231         return err ? 0 : outarg.block;
2232 }
2233
2234 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2235 {
2236         loff_t retval;
2237         struct inode *inode = file_inode(file);
2238
2239         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2240         if (whence == SEEK_CUR || whence == SEEK_SET)
2241                 return generic_file_llseek(file, offset, whence);
2242
2243         mutex_lock(&inode->i_mutex);
2244         retval = fuse_update_attributes(inode, NULL, file, NULL);
2245         if (!retval)
2246                 retval = generic_file_llseek(file, offset, whence);
2247         mutex_unlock(&inode->i_mutex);
2248
2249         return retval;
2250 }
2251
2252 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2253                         unsigned int nr_segs, size_t bytes, bool to_user)
2254 {
2255         struct iov_iter ii;
2256         int page_idx = 0;
2257
2258         if (!bytes)
2259                 return 0;
2260
2261         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2262
2263         while (iov_iter_count(&ii)) {
2264                 struct page *page = pages[page_idx++];
2265                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2266                 void *kaddr;
2267
2268                 kaddr = kmap(page);
2269
2270                 while (todo) {
2271                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2272                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2273                         size_t copy = min(todo, iov_len);
2274                         size_t left;
2275
2276                         if (!to_user)
2277                                 left = copy_from_user(kaddr, uaddr, copy);
2278                         else
2279                                 left = copy_to_user(uaddr, kaddr, copy);
2280
2281                         if (unlikely(left))
2282                                 return -EFAULT;
2283
2284                         iov_iter_advance(&ii, copy);
2285                         todo -= copy;
2286                         kaddr += copy;
2287                 }
2288
2289                 kunmap(page);
2290         }
2291
2292         return 0;
2293 }
2294
2295 /*
2296  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2297  * ABI was defined to be 'struct iovec' which is different on 32bit
2298  * and 64bit.  Fortunately we can determine which structure the server
2299  * used from the size of the reply.
2300  */
2301 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2302                                      size_t transferred, unsigned count,
2303                                      bool is_compat)
2304 {
2305 #ifdef CONFIG_COMPAT
2306         if (count * sizeof(struct compat_iovec) == transferred) {
2307                 struct compat_iovec *ciov = src;
2308                 unsigned i;
2309
2310                 /*
2311                  * With this interface a 32bit server cannot support
2312                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2313                  * requests
2314                  */
2315                 if (!is_compat)
2316                         return -EINVAL;
2317
2318                 for (i = 0; i < count; i++) {
2319                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2320                         dst[i].iov_len = ciov[i].iov_len;
2321                 }
2322                 return 0;
2323         }
2324 #endif
2325
2326         if (count * sizeof(struct iovec) != transferred)
2327                 return -EIO;
2328
2329         memcpy(dst, src, transferred);
2330         return 0;
2331 }
2332
2333 /* Make sure iov_length() won't overflow */
2334 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2335 {
2336         size_t n;
2337         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2338
2339         for (n = 0; n < count; n++, iov++) {
2340                 if (iov->iov_len > (size_t) max)
2341                         return -ENOMEM;
2342                 max -= iov->iov_len;
2343         }
2344         return 0;
2345 }
2346
2347 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2348                                  void *src, size_t transferred, unsigned count,
2349                                  bool is_compat)
2350 {
2351         unsigned i;
2352         struct fuse_ioctl_iovec *fiov = src;
2353
2354         if (fc->minor < 16) {
2355                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2356                                                  count, is_compat);
2357         }
2358
2359         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2360                 return -EIO;
2361
2362         for (i = 0; i < count; i++) {
2363                 /* Did the server supply an inappropriate value? */
2364                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2365                     fiov[i].len != (unsigned long) fiov[i].len)
2366                         return -EIO;
2367
2368                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2369                 dst[i].iov_len = (size_t) fiov[i].len;
2370
2371 #ifdef CONFIG_COMPAT
2372                 if (is_compat &&
2373                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2374                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2375                         return -EIO;
2376 #endif
2377         }
2378
2379         return 0;
2380 }
2381
2382
2383 /*
2384  * For ioctls, there is no generic way to determine how much memory
2385  * needs to be read and/or written.  Furthermore, ioctls are allowed
2386  * to dereference the passed pointer, so the parameter requires deep
2387  * copying but FUSE has no idea whatsoever about what to copy in or
2388  * out.
2389  *
2390  * This is solved by allowing FUSE server to retry ioctl with
2391  * necessary in/out iovecs.  Let's assume the ioctl implementation
2392  * needs to read in the following structure.
2393  *
2394  * struct a {
2395  *      char    *buf;
2396  *      size_t  buflen;
2397  * }
2398  *
2399  * On the first callout to FUSE server, inarg->in_size and
2400  * inarg->out_size will be NULL; then, the server completes the ioctl
2401  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2402  * the actual iov array to
2403  *
2404  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2405  *
2406  * which tells FUSE to copy in the requested area and retry the ioctl.
2407  * On the second round, the server has access to the structure and
2408  * from that it can tell what to look for next, so on the invocation,
2409  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2410  *
2411  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2412  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2413  *
2414  * FUSE will copy both struct a and the pointed buffer from the
2415  * process doing the ioctl and retry ioctl with both struct a and the
2416  * buffer.
2417  *
2418  * This time, FUSE server has everything it needs and completes ioctl
2419  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2420  *
2421  * Copying data out works the same way.
2422  *
2423  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2424  * automatically initializes in and out iovs by decoding @cmd with
2425  * _IOC_* macros and the server is not allowed to request RETRY.  This
2426  * limits ioctl data transfers to well-formed ioctls and is the forced
2427  * behavior for all FUSE servers.
2428  */
2429 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2430                    unsigned int flags)
2431 {
2432         struct fuse_file *ff = file->private_data;
2433         struct fuse_conn *fc = ff->fc;
2434         struct fuse_ioctl_in inarg = {
2435                 .fh = ff->fh,
2436                 .cmd = cmd,
2437                 .arg = arg,
2438                 .flags = flags
2439         };
2440         struct fuse_ioctl_out outarg;
2441         struct fuse_req *req = NULL;
2442         struct page **pages = NULL;
2443         struct iovec *iov_page = NULL;
2444         struct iovec *in_iov = NULL, *out_iov = NULL;
2445         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2446         size_t in_size, out_size, transferred;
2447         int err;
2448
2449 #if BITS_PER_LONG == 32
2450         inarg.flags |= FUSE_IOCTL_32BIT;
2451 #else
2452         if (flags & FUSE_IOCTL_COMPAT)
2453                 inarg.flags |= FUSE_IOCTL_32BIT;
2454 #endif
2455
2456         /* assume all the iovs returned by client always fits in a page */
2457         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2458
2459         err = -ENOMEM;
2460         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2461         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2462         if (!pages || !iov_page)
2463                 goto out;
2464
2465         /*
2466          * If restricted, initialize IO parameters as encoded in @cmd.
2467          * RETRY from server is not allowed.
2468          */
2469         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2470                 struct iovec *iov = iov_page;
2471
2472                 iov->iov_base = (void __user *)arg;
2473                 iov->iov_len = _IOC_SIZE(cmd);
2474
2475                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2476                         in_iov = iov;
2477                         in_iovs = 1;
2478                 }
2479
2480                 if (_IOC_DIR(cmd) & _IOC_READ) {
2481                         out_iov = iov;
2482                         out_iovs = 1;
2483                 }
2484         }
2485
2486  retry:
2487         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2488         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2489
2490         /*
2491          * Out data can be used either for actual out data or iovs,
2492          * make sure there always is at least one page.
2493          */
2494         out_size = max_t(size_t, out_size, PAGE_SIZE);
2495         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2496
2497         /* make sure there are enough buffer pages and init request with them */
2498         err = -ENOMEM;
2499         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2500                 goto out;
2501         while (num_pages < max_pages) {
2502                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2503                 if (!pages[num_pages])
2504                         goto out;
2505                 num_pages++;
2506         }
2507
2508         req = fuse_get_req(fc, num_pages);
2509         if (IS_ERR(req)) {
2510                 err = PTR_ERR(req);
2511                 req = NULL;
2512                 goto out;
2513         }
2514         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2515         req->num_pages = num_pages;
2516         fuse_page_descs_length_init(req, 0, req->num_pages);
2517
2518         /* okay, let's send it to the client */
2519         req->in.h.opcode = FUSE_IOCTL;
2520         req->in.h.nodeid = ff->nodeid;
2521         req->in.numargs = 1;
2522         req->in.args[0].size = sizeof(inarg);
2523         req->in.args[0].value = &inarg;
2524         if (in_size) {
2525                 req->in.numargs++;
2526                 req->in.args[1].size = in_size;
2527                 req->in.argpages = 1;
2528
2529                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2530                                            false);
2531                 if (err)
2532                         goto out;
2533         }
2534
2535         req->out.numargs = 2;
2536         req->out.args[0].size = sizeof(outarg);
2537         req->out.args[0].value = &outarg;
2538         req->out.args[1].size = out_size;
2539         req->out.argpages = 1;
2540         req->out.argvar = 1;
2541
2542         fuse_request_send(fc, req);
2543         err = req->out.h.error;
2544         transferred = req->out.args[1].size;
2545         fuse_put_request(fc, req);
2546         req = NULL;
2547         if (err)
2548                 goto out;
2549
2550         /* did it ask for retry? */
2551         if (outarg.flags & FUSE_IOCTL_RETRY) {
2552                 void *vaddr;
2553
2554                 /* no retry if in restricted mode */
2555                 err = -EIO;
2556                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2557                         goto out;
2558
2559                 in_iovs = outarg.in_iovs;
2560                 out_iovs = outarg.out_iovs;
2561
2562                 /*
2563                  * Make sure things are in boundary, separate checks
2564                  * are to protect against overflow.
2565                  */
2566                 err = -ENOMEM;
2567                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2568                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2569                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2570                         goto out;
2571
2572                 vaddr = kmap_atomic(pages[0]);
2573                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2574                                             transferred, in_iovs + out_iovs,
2575                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2576                 kunmap_atomic(vaddr);
2577                 if (err)
2578                         goto out;
2579
2580                 in_iov = iov_page;
2581                 out_iov = in_iov + in_iovs;
2582
2583                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2584                 if (err)
2585                         goto out;
2586
2587                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2588                 if (err)
2589                         goto out;
2590
2591                 goto retry;
2592         }
2593
2594         err = -EIO;
2595         if (transferred > inarg.out_size)
2596                 goto out;
2597
2598         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2599  out:
2600         if (req)
2601                 fuse_put_request(fc, req);
2602         free_page((unsigned long) iov_page);
2603         while (num_pages)
2604                 __free_page(pages[--num_pages]);
2605         kfree(pages);
2606
2607         return err ? err : outarg.result;
2608 }
2609 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2610
2611 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2612                        unsigned long arg, unsigned int flags)
2613 {
2614         struct inode *inode = file_inode(file);
2615         struct fuse_conn *fc = get_fuse_conn(inode);
2616
2617         if (!fuse_allow_current_process(fc))
2618                 return -EACCES;
2619
2620         if (is_bad_inode(inode))
2621                 return -EIO;
2622
2623         return fuse_do_ioctl(file, cmd, arg, flags);
2624 }
2625
2626 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2627                             unsigned long arg)
2628 {
2629         return fuse_ioctl_common(file, cmd, arg, 0);
2630 }
2631
2632 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2633                                    unsigned long arg)
2634 {
2635         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2636 }
2637
2638 /*
2639  * All files which have been polled are linked to RB tree
2640  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2641  * find the matching one.
2642  */
2643 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2644                                               struct rb_node **parent_out)
2645 {
2646         struct rb_node **link = &fc->polled_files.rb_node;
2647         struct rb_node *last = NULL;
2648
2649         while (*link) {
2650                 struct fuse_file *ff;
2651
2652                 last = *link;
2653                 ff = rb_entry(last, struct fuse_file, polled_node);
2654
2655                 if (kh < ff->kh)
2656                         link = &last->rb_left;
2657                 else if (kh > ff->kh)
2658                         link = &last->rb_right;
2659                 else
2660                         return link;
2661         }
2662
2663         if (parent_out)
2664                 *parent_out = last;
2665         return link;
2666 }
2667
2668 /*
2669  * The file is about to be polled.  Make sure it's on the polled_files
2670  * RB tree.  Note that files once added to the polled_files tree are
2671  * not removed before the file is released.  This is because a file
2672  * polled once is likely to be polled again.
2673  */
2674 static void fuse_register_polled_file(struct fuse_conn *fc,
2675                                       struct fuse_file *ff)
2676 {
2677         spin_lock(&fc->lock);
2678         if (RB_EMPTY_NODE(&ff->polled_node)) {
2679                 struct rb_node **link, *uninitialized_var(parent);
2680
2681                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2682                 BUG_ON(*link);
2683                 rb_link_node(&ff->polled_node, parent, link);
2684                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2685         }
2686         spin_unlock(&fc->lock);
2687 }
2688
2689 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2690 {
2691         struct fuse_file *ff = file->private_data;
2692         struct fuse_conn *fc = ff->fc;
2693         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2694         struct fuse_poll_out outarg;
2695         FUSE_ARGS(args);
2696         int err;
2697
2698         if (fc->no_poll)
2699                 return DEFAULT_POLLMASK;
2700
2701         poll_wait(file, &ff->poll_wait, wait);
2702         inarg.events = (__u32)poll_requested_events(wait);
2703
2704         /*
2705          * Ask for notification iff there's someone waiting for it.
2706          * The client may ignore the flag and always notify.
2707          */
2708         if (waitqueue_active(&ff->poll_wait)) {
2709                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2710                 fuse_register_polled_file(fc, ff);
2711         }
2712
2713         args.in.h.opcode = FUSE_POLL;
2714         args.in.h.nodeid = ff->nodeid;
2715         args.in.numargs = 1;
2716         args.in.args[0].size = sizeof(inarg);
2717         args.in.args[0].value = &inarg;
2718         args.out.numargs = 1;
2719         args.out.args[0].size = sizeof(outarg);
2720         args.out.args[0].value = &outarg;
2721         err = fuse_simple_request(fc, &args);
2722
2723         if (!err)
2724                 return outarg.revents;
2725         if (err == -ENOSYS) {
2726                 fc->no_poll = 1;
2727                 return DEFAULT_POLLMASK;
2728         }
2729         return POLLERR;
2730 }
2731 EXPORT_SYMBOL_GPL(fuse_file_poll);
2732
2733 /*
2734  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2735  * wakes up the poll waiters.
2736  */
2737 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2738                             struct fuse_notify_poll_wakeup_out *outarg)
2739 {
2740         u64 kh = outarg->kh;
2741         struct rb_node **link;
2742
2743         spin_lock(&fc->lock);
2744
2745         link = fuse_find_polled_node(fc, kh, NULL);
2746         if (*link) {
2747                 struct fuse_file *ff;
2748
2749                 ff = rb_entry(*link, struct fuse_file, polled_node);
2750                 wake_up_interruptible_sync(&ff->poll_wait);
2751         }
2752
2753         spin_unlock(&fc->lock);
2754         return 0;
2755 }
2756
2757 static void fuse_do_truncate(struct file *file)
2758 {
2759         struct inode *inode = file->f_mapping->host;
2760         struct iattr attr;
2761
2762         attr.ia_valid = ATTR_SIZE;
2763         attr.ia_size = i_size_read(inode);
2764
2765         attr.ia_file = file;
2766         attr.ia_valid |= ATTR_FILE;
2767
2768         fuse_do_setattr(inode, &attr, file);
2769 }
2770
2771 static inline loff_t fuse_round_up(loff_t off)
2772 {
2773         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2774 }
2775
2776 static ssize_t
2777 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2778 {
2779         DECLARE_COMPLETION_ONSTACK(wait);
2780         ssize_t ret = 0;
2781         struct file *file = iocb->ki_filp;
2782         struct fuse_file *ff = file->private_data;
2783         bool async_dio = ff->fc->async_dio;
2784         loff_t pos = 0;
2785         struct inode *inode;
2786         loff_t i_size;
2787         size_t count = iov_iter_count(iter);
2788         struct fuse_io_priv *io;
2789
2790         pos = offset;
2791         inode = file->f_mapping->host;
2792         i_size = i_size_read(inode);
2793
2794         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2795                 return 0;
2796
2797         /* optimization for short read */
2798         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2799                 if (offset >= i_size)
2800                         return 0;
2801                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2802                 count = iov_iter_count(iter);
2803         }
2804
2805         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2806         if (!io)
2807                 return -ENOMEM;
2808         spin_lock_init(&io->lock);
2809         io->reqs = 1;
2810         io->bytes = -1;
2811         io->size = 0;
2812         io->offset = offset;
2813         io->write = (iov_iter_rw(iter) == WRITE);
2814         io->err = 0;
2815         io->file = file;
2816         /*
2817          * By default, we want to optimize all I/Os with async request
2818          * submission to the client filesystem if supported.
2819          */
2820         io->async = async_dio;
2821         io->iocb = iocb;
2822
2823         /*
2824          * We cannot asynchronously extend the size of a file. We have no method
2825          * to wait on real async I/O requests, so we must submit this request
2826          * synchronously.
2827          */
2828         if (!is_sync_kiocb(iocb) && (offset + count > i_size) &&
2829             iov_iter_rw(iter) == WRITE)
2830                 io->async = false;
2831
2832         if (io->async && is_sync_kiocb(iocb))
2833                 io->done = &wait;
2834
2835         if (iov_iter_rw(iter) == WRITE) {
2836                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2837                 fuse_invalidate_attr(inode);
2838         } else {
2839                 ret = __fuse_direct_read(io, iter, &pos);
2840         }
2841
2842         if (io->async) {
2843                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2844
2845                 /* we have a non-extending, async request, so return */
2846                 if (!is_sync_kiocb(iocb))
2847                         return -EIOCBQUEUED;
2848
2849                 wait_for_completion(&wait);
2850                 ret = fuse_get_res_by_io(io);
2851         }
2852
2853         kfree(io);
2854
2855         if (iov_iter_rw(iter) == WRITE) {
2856                 if (ret > 0)
2857                         fuse_write_update_size(inode, pos);
2858                 else if (ret < 0 && offset + count > i_size)
2859                         fuse_do_truncate(file);
2860         }
2861
2862         return ret;
2863 }
2864
2865 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2866                                 loff_t length)
2867 {
2868         struct fuse_file *ff = file->private_data;
2869         struct inode *inode = file_inode(file);
2870         struct fuse_inode *fi = get_fuse_inode(inode);
2871         struct fuse_conn *fc = ff->fc;
2872         FUSE_ARGS(args);
2873         struct fuse_fallocate_in inarg = {
2874                 .fh = ff->fh,
2875                 .offset = offset,
2876                 .length = length,
2877                 .mode = mode
2878         };
2879         int err;
2880         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2881                            (mode & FALLOC_FL_PUNCH_HOLE);
2882
2883         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2884                 return -EOPNOTSUPP;
2885
2886         if (fc->no_fallocate)
2887                 return -EOPNOTSUPP;
2888
2889         if (lock_inode) {
2890                 mutex_lock(&inode->i_mutex);
2891                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2892                         loff_t endbyte = offset + length - 1;
2893                         err = filemap_write_and_wait_range(inode->i_mapping,
2894                                                            offset, endbyte);
2895                         if (err)
2896                                 goto out;
2897
2898                         fuse_sync_writes(inode);
2899                 }
2900         }
2901
2902         if (!(mode & FALLOC_FL_KEEP_SIZE))
2903                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2904
2905         args.in.h.opcode = FUSE_FALLOCATE;
2906         args.in.h.nodeid = ff->nodeid;
2907         args.in.numargs = 1;
2908         args.in.args[0].size = sizeof(inarg);
2909         args.in.args[0].value = &inarg;
2910         err = fuse_simple_request(fc, &args);
2911         if (err == -ENOSYS) {
2912                 fc->no_fallocate = 1;
2913                 err = -EOPNOTSUPP;
2914         }
2915         if (err)
2916                 goto out;
2917
2918         /* we could have extended the file */
2919         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2920                 bool changed = fuse_write_update_size(inode, offset + length);
2921
2922                 if (changed && fc->writeback_cache)
2923                         file_update_time(file);
2924         }
2925
2926         if (mode & FALLOC_FL_PUNCH_HOLE)
2927                 truncate_pagecache_range(inode, offset, offset + length - 1);
2928
2929         fuse_invalidate_attr(inode);
2930
2931 out:
2932         if (!(mode & FALLOC_FL_KEEP_SIZE))
2933                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2934
2935         if (lock_inode)
2936                 mutex_unlock(&inode->i_mutex);
2937
2938         return err;
2939 }
2940
2941 static const struct file_operations fuse_file_operations = {
2942         .llseek         = fuse_file_llseek,
2943         .read_iter      = fuse_file_read_iter,
2944         .write_iter     = fuse_file_write_iter,
2945         .mmap           = fuse_file_mmap,
2946         .open           = fuse_open,
2947         .flush          = fuse_flush,
2948         .release        = fuse_release,
2949         .fsync          = fuse_fsync,
2950         .lock           = fuse_file_lock,
2951         .flock          = fuse_file_flock,
2952         .splice_read    = generic_file_splice_read,
2953         .unlocked_ioctl = fuse_file_ioctl,
2954         .compat_ioctl   = fuse_file_compat_ioctl,
2955         .poll           = fuse_file_poll,
2956         .fallocate      = fuse_file_fallocate,
2957 };
2958
2959 static const struct file_operations fuse_direct_io_file_operations = {
2960         .llseek         = fuse_file_llseek,
2961         .read_iter      = fuse_direct_read_iter,
2962         .write_iter     = fuse_direct_write_iter,
2963         .mmap           = fuse_direct_mmap,
2964         .open           = fuse_open,
2965         .flush          = fuse_flush,
2966         .release        = fuse_release,
2967         .fsync          = fuse_fsync,
2968         .lock           = fuse_file_lock,
2969         .flock          = fuse_file_flock,
2970         .unlocked_ioctl = fuse_file_ioctl,
2971         .compat_ioctl   = fuse_file_compat_ioctl,
2972         .poll           = fuse_file_poll,
2973         .fallocate      = fuse_file_fallocate,
2974         /* no splice_read */
2975 };
2976
2977 static const struct address_space_operations fuse_file_aops  = {
2978         .readpage       = fuse_readpage,
2979         .writepage      = fuse_writepage,
2980         .writepages     = fuse_writepages,
2981         .launder_page   = fuse_launder_page,
2982         .readpages      = fuse_readpages,
2983         .set_page_dirty = __set_page_dirty_nobuffers,
2984         .bmap           = fuse_bmap,
2985         .direct_IO      = fuse_direct_IO,
2986         .write_begin    = fuse_write_begin,
2987         .write_end      = fuse_write_end,
2988 };
2989
2990 void fuse_init_file_inode(struct inode *inode)
2991 {
2992         inode->i_fop = &fuse_file_operations;
2993         inode->i_data.a_ops = &fuse_file_aops;
2994 }