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/aio.h>
  19 #include <linux/falloc.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                         req->background = 0;
 100                         iput(req->misc.release.inode);
 101                         fuse_put_request(ff->fc, req);
 102                 } else if (sync) {
 103                         req->background = 0;
 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                         req->background = 1;
 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         ff->reserved_req->force = 1;
 303         ff->reserved_req->background = 0;
 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         req->force = 1;
 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 /**
 532  * In case of short read, the caller sets 'pos' to the position of
 533  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 534  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 535  *
 536  * An example:
 537  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 538  * both submitted asynchronously. The first of them was ACKed by userspace as
 539  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 540  * second request was ACKed as short, e.g. only 1K was read, resulting in
 541  * pos == 33K.
 542  *
 543  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 544  * will be equal to the length of the longest contiguous fragment of
 545  * transferred data starting from the beginning of IO request.
 546  */
 547 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 548 {
 549         int left;
 550
 551         spin_lock(&io->lock);
 552         if (err)
 553                 io->err = io->err ? : err;
 554         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 555                 io->bytes = pos;
 556
 557         left = --io->reqs;
 558         spin_unlock(&io->lock);
 559
 560         if (!left) {
 561                 long res;
 562
 563                 if (io->err)
 564                         res = io->err;
 565                 else if (io->bytes >= 0 && io->write)
 566                         res = -EIO;
 567                 else {
 568                         res = io->bytes < 0 ? io->size : io->bytes;
 569
 570                         if (!is_sync_kiocb(io->iocb)) {
 571                                 struct inode *inode = file_inode(io->iocb->ki_filp);
 572                                 struct fuse_conn *fc = get_fuse_conn(inode);
 573                                 struct fuse_inode *fi = get_fuse_inode(inode);
 574
 575                                 spin_lock(&fc->lock);
 576                                 fi->attr_version = ++fc->attr_version;
 577                                 spin_unlock(&fc->lock);
 578                         }
 579                 }
 580
 581                 aio_complete(io->iocb, res, 0);
 582                 kfree(io);
 583         }
 584 }
 585
 586 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 587 {
 588         struct fuse_io_priv *io = req->io;
 589         ssize_t pos = -1;
 590
 591         fuse_release_user_pages(req, !io->write);
 592
 593         if (io->write) {
 594                 if (req->misc.write.in.size != req->misc.write.out.size)
 595                         pos = req->misc.write.in.offset - io->offset +
 596                                 req->misc.write.out.size;
 597         } else {
 598                 if (req->misc.read.in.size != req->out.args[0].size)
 599                         pos = req->misc.read.in.offset - io->offset +
 600                                 req->out.args[0].size;
 601         }
 602
 603         fuse_aio_complete(io, req->out.h.error, pos);
 604 }
 605
 606 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 607                 size_t num_bytes, struct fuse_io_priv *io)
 608 {
 609         spin_lock(&io->lock);
 610         io->size += num_bytes;
 611         io->reqs++;
 612         spin_unlock(&io->lock);
 613
 614         req->io = io;
 615         req->end = fuse_aio_complete_req;
 616
 617         __fuse_get_request(req);
 618         fuse_request_send_background(fc, req);
 619
 620         return num_bytes;
 621 }
 622
 623 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 624                              loff_t pos, size_t count, fl_owner_t owner)
 625 {
 626         struct file *file = io->file;
 627         struct fuse_file *ff = file->private_data;
 628         struct fuse_conn *fc = ff->fc;
 629
 630         fuse_read_fill(req, file, pos, count, FUSE_READ);
 631         if (owner != NULL) {
 632                 struct fuse_read_in *inarg = &req->misc.read.in;
 633
 634                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 635                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 636         }
 637
 638         if (io->async)
 639                 return fuse_async_req_send(fc, req, count, io);
 640
 641         fuse_request_send(fc, req);
 642         return req->out.args[0].size;
 643 }
 644
 645 static void fuse_read_update_size(struct inode *inode, loff_t size,
 646                                   u64 attr_ver)
 647 {
 648         struct fuse_conn *fc = get_fuse_conn(inode);
 649         struct fuse_inode *fi = get_fuse_inode(inode);
 650
 651         spin_lock(&fc->lock);
 652         if (attr_ver == fi->attr_version && size < inode->i_size &&
 653             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 654                 fi->attr_version = ++fc->attr_version;
 655                 i_size_write(inode, size);
 656         }
 657         spin_unlock(&fc->lock);
 658 }
 659
 660 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 661                             u64 attr_ver)
 662 {
 663         size_t num_read = req->out.args[0].size;
 664         struct fuse_conn *fc = get_fuse_conn(inode);
 665
 666         if (fc->writeback_cache) {
 667                 /*
 668                  * A hole in a file. Some data after the hole are in page cache,
 669                  * but have not reached the client fs yet. So, the hole is not
 670                  * present there.
 671                  */
 672                 int i;
 673                 int start_idx = num_read >> PAGE_CACHE_SHIFT;
 674                 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
 675
 676                 for (i = start_idx; i < req->num_pages; i++) {
 677                         zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
 678                         off = 0;
 679                 }
 680         } else {
 681                 loff_t pos = page_offset(req->pages[0]) + num_read;
 682                 fuse_read_update_size(inode, pos, attr_ver);
 683         }
 684 }
 685
 686 static int fuse_do_readpage(struct file *file, struct page *page)
 687 {
 688         struct fuse_io_priv io = { .async = 0, .file = file };
 689         struct inode *inode = page->mapping->host;
 690         struct fuse_conn *fc = get_fuse_conn(inode);
 691         struct fuse_req *req;
 692         size_t num_read;
 693         loff_t pos = page_offset(page);
 694         size_t count = PAGE_CACHE_SIZE;
 695         u64 attr_ver;
 696         int err;
 697
 698         /*
 699          * Page writeback can extend beyond the lifetime of the
 700          * page-cache page, so make sure we read a properly synced
 701          * page.
 702          */
 703         fuse_wait_on_page_writeback(inode, page->index);
 704
 705         req = fuse_get_req(fc, 1);
 706         if (IS_ERR(req))
 707                 return PTR_ERR(req);
 708
 709         attr_ver = fuse_get_attr_version(fc);
 710
 711         req->out.page_zeroing = 1;
 712         req->out.argpages = 1;
 713         req->num_pages = 1;
 714         req->pages[0] = page;
 715         req->page_descs[0].length = count;
 716         num_read = fuse_send_read(req, &io, pos, count, NULL);
 717         err = req->out.h.error;
 718
 719         if (!err) {
 720                 /*
 721                  * Short read means EOF.  If file size is larger, truncate it
 722                  */
 723                 if (num_read < count)
 724                         fuse_short_read(req, inode, attr_ver);
 725
 726                 SetPageUptodate(page);
 727         }
 728
 729         fuse_put_request(fc, req);
 730
 731         return err;
 732 }
 733
 734 static int fuse_readpage(struct file *file, struct page *page)
 735 {
 736         struct inode *inode = page->mapping->host;
 737         int err;
 738
 739         err = -EIO;
 740         if (is_bad_inode(inode))
 741                 goto out;
 742
 743         err = fuse_do_readpage(file, page);
 744         fuse_invalidate_atime(inode);
 745  out:
 746         unlock_page(page);
 747         return err;
 748 }
 749
 750 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 751 {
 752         int i;
 753         size_t count = req->misc.read.in.size;
 754         size_t num_read = req->out.args[0].size;
 755         struct address_space *mapping = NULL;
 756
 757         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 758                 mapping = req->pages[i]->mapping;
 759
 760         if (mapping) {
 761                 struct inode *inode = mapping->host;
 762
 763                 /*
 764                  * Short read means EOF. If file size is larger, truncate it
 765                  */
 766                 if (!req->out.h.error && num_read < count)
 767                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 768
 769                 fuse_invalidate_atime(inode);
 770         }
 771
 772         for (i = 0; i < req->num_pages; i++) {
 773                 struct page *page = req->pages[i];
 774                 if (!req->out.h.error)
 775                         SetPageUptodate(page);
 776                 else
 777                         SetPageError(page);
 778                 unlock_page(page);
 779                 page_cache_release(page);
 780         }
 781         if (req->ff)
 782                 fuse_file_put(req->ff, false);
 783 }
 784
 785 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 786 {
 787         struct fuse_file *ff = file->private_data;
 788         struct fuse_conn *fc = ff->fc;
 789         loff_t pos = page_offset(req->pages[0]);
 790         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 791
 792         req->out.argpages = 1;
 793         req->out.page_zeroing = 1;
 794         req->out.page_replace = 1;
 795         fuse_read_fill(req, file, pos, count, FUSE_READ);
 796         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 797         if (fc->async_read) {
 798                 req->ff = fuse_file_get(ff);
 799                 req->end = fuse_readpages_end;
 800                 fuse_request_send_background(fc, req);
 801         } else {
 802                 fuse_request_send(fc, req);
 803                 fuse_readpages_end(fc, req);
 804                 fuse_put_request(fc, req);
 805         }
 806 }
 807
 808 struct fuse_fill_data {
 809         struct fuse_req *req;
 810         struct file *file;
 811         struct inode *inode;
 812         unsigned nr_pages;
 813 };
 814
 815 static int fuse_readpages_fill(void *_data, struct page *page)
 816 {
 817         struct fuse_fill_data *data = _data;
 818         struct fuse_req *req = data->req;
 819         struct inode *inode = data->inode;
 820         struct fuse_conn *fc = get_fuse_conn(inode);
 821
 822         fuse_wait_on_page_writeback(inode, page->index);
 823
 824         if (req->num_pages &&
 825             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 826              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 827              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 828                 int nr_alloc = min_t(unsigned, data->nr_pages,
 829                                      FUSE_MAX_PAGES_PER_REQ);
 830                 fuse_send_readpages(req, data->file);
 831                 if (fc->async_read)
 832                         req = fuse_get_req_for_background(fc, nr_alloc);
 833                 else
 834                         req = fuse_get_req(fc, nr_alloc);
 835
 836                 data->req = req;
 837                 if (IS_ERR(req)) {
 838                         unlock_page(page);
 839                         return PTR_ERR(req);
 840                 }
 841         }
 842
 843         if (WARN_ON(req->num_pages >= req->max_pages)) {
 844                 fuse_put_request(fc, req);
 845                 return -EIO;
 846         }
 847
 848         page_cache_get(page);
 849         req->pages[req->num_pages] = page;
 850         req->page_descs[req->num_pages].length = PAGE_SIZE;
 851         req->num_pages++;
 852         data->nr_pages--;
 853         return 0;
 854 }
 855
 856 static int fuse_readpages(struct file *file, struct address_space *mapping,
 857                           struct list_head *pages, unsigned nr_pages)
 858 {
 859         struct inode *inode = mapping->host;
 860         struct fuse_conn *fc = get_fuse_conn(inode);
 861         struct fuse_fill_data data;
 862         int err;
 863         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 864
 865         err = -EIO;
 866         if (is_bad_inode(inode))
 867                 goto out;
 868
 869         data.file = file;
 870         data.inode = inode;
 871         if (fc->async_read)
 872                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 873         else
 874                 data.req = fuse_get_req(fc, nr_alloc);
 875         data.nr_pages = nr_pages;
 876         err = PTR_ERR(data.req);
 877         if (IS_ERR(data.req))
 878                 goto out;
 879
 880         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 881         if (!err) {
 882                 if (data.req->num_pages)
 883                         fuse_send_readpages(data.req, file);
 884                 else
 885                         fuse_put_request(fc, data.req);
 886         }
 887 out:
 888         return err;
 889 }
 890
 891 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 892 {
 893         struct inode *inode = iocb->ki_filp->f_mapping->host;
 894         struct fuse_conn *fc = get_fuse_conn(inode);
 895
 896         /*
 897          * In auto invalidate mode, always update attributes on read.
 898          * Otherwise, only update if we attempt to read past EOF (to ensure
 899          * i_size is up to date).
 900          */
 901         if (fc->auto_inval_data ||
 902             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 903                 int err;
 904                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 905                 if (err)
 906                         return err;
 907         }
 908
 909         return generic_file_read_iter(iocb, to);
 910 }
 911
 912 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 913                             loff_t pos, size_t count)
 914 {
 915         struct fuse_write_in *inarg = &req->misc.write.in;
 916         struct fuse_write_out *outarg = &req->misc.write.out;
 917
 918         inarg->fh = ff->fh;
 919         inarg->offset = pos;
 920         inarg->size = count;
 921         req->in.h.opcode = FUSE_WRITE;
 922         req->in.h.nodeid = ff->nodeid;
 923         req->in.numargs = 2;
 924         if (ff->fc->minor < 9)
 925                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 926         else
 927                 req->in.args[0].size = sizeof(struct fuse_write_in);
 928         req->in.args[0].value = inarg;
 929         req->in.args[1].size = count;
 930         req->out.numargs = 1;
 931         req->out.args[0].size = sizeof(struct fuse_write_out);
 932         req->out.args[0].value = outarg;
 933 }
 934
 935 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 936                               loff_t pos, size_t count, fl_owner_t owner)
 937 {
 938         struct file *file = io->file;
 939         struct fuse_file *ff = file->private_data;
 940         struct fuse_conn *fc = ff->fc;
 941         struct fuse_write_in *inarg = &req->misc.write.in;
 942
 943         fuse_write_fill(req, ff, pos, count);
 944         inarg->flags = file->f_flags;
 945         if (owner != NULL) {
 946                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 947                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 948         }
 949
 950         if (io->async)
 951                 return fuse_async_req_send(fc, req, count, io);
 952
 953         fuse_request_send(fc, req);
 954         return req->misc.write.out.size;
 955 }
 956
 957 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 958 {
 959         struct fuse_conn *fc = get_fuse_conn(inode);
 960         struct fuse_inode *fi = get_fuse_inode(inode);
 961         bool ret = false;
 962
 963         spin_lock(&fc->lock);
 964         fi->attr_version = ++fc->attr_version;
 965         if (pos > inode->i_size) {
 966                 i_size_write(inode, pos);
 967                 ret = true;
 968         }
 969         spin_unlock(&fc->lock);
 970
 971         return ret;
 972 }
 973
 974 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 975                                     struct inode *inode, loff_t pos,
 976                                     size_t count)
 977 {
 978         size_t res;
 979         unsigned offset;
 980         unsigned i;
 981         struct fuse_io_priv io = { .async = 0, .file = file };
 982
 983         for (i = 0; i < req->num_pages; i++)
 984                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 985
 986         res = fuse_send_write(req, &io, pos, count, NULL);
 987
 988         offset = req->page_descs[0].offset;
 989         count = res;
 990         for (i = 0; i < req->num_pages; i++) {
 991                 struct page *page = req->pages[i];
 992
 993                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 994                         SetPageUptodate(page);
 995
 996                 if (count > PAGE_CACHE_SIZE - offset)
 997                         count -= PAGE_CACHE_SIZE - offset;
 998                 else
 999                         count = 0;
1000                 offset = 0;
1001
1002                 unlock_page(page);
1003                 page_cache_release(page);
1004         }
1005
1006         return res;
1007 }
1008
1009 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1010                                struct address_space *mapping,
1011                                struct iov_iter *ii, loff_t pos)
1012 {
1013         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1014         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1015         size_t count = 0;
1016         int err;
1017
1018         req->in.argpages = 1;
1019         req->page_descs[0].offset = offset;
1020
1021         do {
1022                 size_t tmp;
1023                 struct page *page;
1024                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1025                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1026                                      iov_iter_count(ii));
1027
1028                 bytes = min_t(size_t, bytes, fc->max_write - count);
1029
1030  again:
1031                 err = -EFAULT;
1032                 if (iov_iter_fault_in_readable(ii, bytes))
1033                         break;
1034
1035                 err = -ENOMEM;
1036                 page = grab_cache_page_write_begin(mapping, index, 0);
1037                 if (!page)
1038                         break;
1039
1040                 if (mapping_writably_mapped(mapping))
1041                         flush_dcache_page(page);
1042
1043                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1044                 flush_dcache_page(page);
1045
1046                 if (!tmp) {
1047                         unlock_page(page);
1048                         page_cache_release(page);
1049                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1050                         goto again;
1051                 }
1052
1053                 err = 0;
1054                 req->pages[req->num_pages] = page;
1055                 req->page_descs[req->num_pages].length = tmp;
1056                 req->num_pages++;
1057
1058                 iov_iter_advance(ii, tmp);
1059                 count += tmp;
1060                 pos += tmp;
1061                 offset += tmp;
1062                 if (offset == PAGE_CACHE_SIZE)
1063                         offset = 0;
1064
1065                 if (!fc->big_writes)
1066                         break;
1067         } while (iov_iter_count(ii) && count < fc->max_write &&
1068                  req->num_pages < req->max_pages && offset == 0);
1069
1070         return count > 0 ? count : err;
1071 }
1072
1073 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1074 {
1075         return min_t(unsigned,
1076                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1077                      (pos >> PAGE_CACHE_SHIFT) + 1,
1078                      FUSE_MAX_PAGES_PER_REQ);
1079 }
1080
1081 static ssize_t fuse_perform_write(struct file *file,
1082                                   struct address_space *mapping,
1083                                   struct iov_iter *ii, loff_t pos)
1084 {
1085         struct inode *inode = mapping->host;
1086         struct fuse_conn *fc = get_fuse_conn(inode);
1087         struct fuse_inode *fi = get_fuse_inode(inode);
1088         int err = 0;
1089         ssize_t res = 0;
1090
1091         if (is_bad_inode(inode))
1092                 return -EIO;
1093
1094         if (inode->i_size < pos + iov_iter_count(ii))
1095                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1096
1097         do {
1098                 struct fuse_req *req;
1099                 ssize_t count;
1100                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1101
1102                 req = fuse_get_req(fc, nr_pages);
1103                 if (IS_ERR(req)) {
1104                         err = PTR_ERR(req);
1105                         break;
1106                 }
1107
1108                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1109                 if (count <= 0) {
1110                         err = count;
1111                 } else {
1112                         size_t num_written;
1113
1114                         num_written = fuse_send_write_pages(req, file, inode,
1115                                                             pos, count);
1116                         err = req->out.h.error;
1117                         if (!err) {
1118                                 res += num_written;
1119                                 pos += num_written;
1120
1121                                 /* break out of the loop on short write */
1122                                 if (num_written != count)
1123                                         err = -EIO;
1124                         }
1125                 }
1126                 fuse_put_request(fc, req);
1127         } while (!err && iov_iter_count(ii));
1128
1129         if (res > 0)
1130                 fuse_write_update_size(inode, pos);
1131
1132         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1133         fuse_invalidate_attr(inode);
1134
1135         return res > 0 ? res : err;
1136 }
1137
1138 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1139 {
1140         struct file *file = iocb->ki_filp;
1141         struct address_space *mapping = file->f_mapping;
1142         size_t count = iov_iter_count(from);
1143         ssize_t written = 0;
1144         ssize_t written_buffered = 0;
1145         struct inode *inode = mapping->host;
1146         ssize_t err;
1147         loff_t endbyte = 0;
1148         loff_t pos = iocb->ki_pos;
1149
1150         if (get_fuse_conn(inode)->writeback_cache) {
1151                 /* Update size (EOF optimization) and mode (SUID clearing) */
1152                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1153                 if (err)
1154                         return err;
1155
1156                 return generic_file_write_iter(iocb, from);
1157         }
1158
1159         mutex_lock(&inode->i_mutex);
1160
1161         /* We can write back this queue in page reclaim */
1162         current->backing_dev_info = inode_to_bdi(inode);
1163
1164         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1165         if (err)
1166                 goto out;
1167
1168         if (count == 0)
1169                 goto out;
1170
1171         iov_iter_truncate(from, count);
1172         err = file_remove_suid(file);
1173         if (err)
1174                 goto out;
1175
1176         err = file_update_time(file);
1177         if (err)
1178                 goto out;
1179
1180         if (file->f_flags & O_DIRECT) {
1181                 written = generic_file_direct_write(iocb, from, pos);
1182                 if (written < 0 || !iov_iter_count(from))
1183                         goto out;
1184
1185                 pos += written;
1186
1187                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1188                 if (written_buffered < 0) {
1189                         err = written_buffered;
1190                         goto out;
1191                 }
1192                 endbyte = pos + written_buffered - 1;
1193
1194                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1195                                                    endbyte);
1196                 if (err)
1197                         goto out;
1198
1199                 invalidate_mapping_pages(file->f_mapping,
1200                                          pos >> PAGE_CACHE_SHIFT,
1201                                          endbyte >> PAGE_CACHE_SHIFT);
1202
1203                 written += written_buffered;
1204                 iocb->ki_pos = pos + written_buffered;
1205         } else {
1206                 written = fuse_perform_write(file, mapping, from, pos);
1207                 if (written >= 0)
1208                         iocb->ki_pos = pos + written;
1209         }
1210 out:
1211         current->backing_dev_info = NULL;
1212         mutex_unlock(&inode->i_mutex);
1213
1214         return written ? written : err;
1215 }
1216
1217 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1218                 unsigned index, unsigned nr_pages)
1219 {
1220         int i;
1221
1222         for (i = index; i < index + nr_pages; i++)
1223                 req->page_descs[i].length = PAGE_SIZE -
1224                         req->page_descs[i].offset;
1225 }
1226
1227 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1228 {
1229         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1230 }
1231
1232 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1233                                         size_t max_size)
1234 {
1235         return min(iov_iter_single_seg_count(ii), max_size);
1236 }
1237
1238 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1239                                size_t *nbytesp, int write)
1240 {
1241         size_t nbytes = 0;  /* # bytes already packed in req */
1242
1243         /* Special case for kernel I/O: can copy directly into the buffer */
1244         if (ii->type & ITER_KVEC) {
1245                 unsigned long user_addr = fuse_get_user_addr(ii);
1246                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1247
1248                 if (write)
1249                         req->in.args[1].value = (void *) user_addr;
1250                 else
1251                         req->out.args[0].value = (void *) user_addr;
1252
1253                 iov_iter_advance(ii, frag_size);
1254                 *nbytesp = frag_size;
1255                 return 0;
1256         }
1257
1258         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1259                 unsigned npages;
1260                 size_t start;
1261                 ssize_t ret = iov_iter_get_pages(ii,
1262                                         &req->pages[req->num_pages],
1263                                         *nbytesp - nbytes,
1264                                         req->max_pages - req->num_pages,
1265                                         &start);
1266                 if (ret < 0)
1267                         return ret;
1268
1269                 iov_iter_advance(ii, ret);
1270                 nbytes += ret;
1271
1272                 ret += start;
1273                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1274
1275                 req->page_descs[req->num_pages].offset = start;
1276                 fuse_page_descs_length_init(req, req->num_pages, npages);
1277
1278                 req->num_pages += npages;
1279                 req->page_descs[req->num_pages - 1].length -=
1280                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1281         }
1282
1283         if (write)
1284                 req->in.argpages = 1;
1285         else
1286                 req->out.argpages = 1;
1287
1288         *nbytesp = nbytes;
1289
1290         return 0;
1291 }
1292
1293 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1294 {
1295         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1296 }
1297
1298 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1299                        loff_t *ppos, int flags)
1300 {
1301         int write = flags & FUSE_DIO_WRITE;
1302         int cuse = flags & FUSE_DIO_CUSE;
1303         struct file *file = io->file;
1304         struct inode *inode = file->f_mapping->host;
1305         struct fuse_file *ff = file->private_data;
1306         struct fuse_conn *fc = ff->fc;
1307         size_t nmax = write ? fc->max_write : fc->max_read;
1308         loff_t pos = *ppos;
1309         size_t count = iov_iter_count(iter);
1310         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1311         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1312         ssize_t res = 0;
1313         struct fuse_req *req;
1314
1315         if (io->async)
1316                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1317         else
1318                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1319         if (IS_ERR(req))
1320                 return PTR_ERR(req);
1321
1322         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1323                 if (!write)
1324                         mutex_lock(&inode->i_mutex);
1325                 fuse_sync_writes(inode);
1326                 if (!write)
1327                         mutex_unlock(&inode->i_mutex);
1328         }
1329
1330         while (count) {
1331                 size_t nres;
1332                 fl_owner_t owner = current->files;
1333                 size_t nbytes = min(count, nmax);
1334                 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1335                 if (err) {
1336                         res = err;
1337                         break;
1338                 }
1339
1340                 if (write)
1341                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1342                 else
1343                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1344
1345                 if (!io->async)
1346                         fuse_release_user_pages(req, !write);
1347                 if (req->out.h.error) {
1348                         if (!res)
1349                                 res = req->out.h.error;
1350                         break;
1351                 } else if (nres > nbytes) {
1352                         res = -EIO;
1353                         break;
1354                 }
1355                 count -= nres;
1356                 res += nres;
1357                 pos += nres;
1358                 if (nres != nbytes)
1359                         break;
1360                 if (count) {
1361                         fuse_put_request(fc, req);
1362                         if (io->async)
1363                                 req = fuse_get_req_for_background(fc,
1364                                         fuse_iter_npages(iter));
1365                         else
1366                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1367                         if (IS_ERR(req))
1368                                 break;
1369                 }
1370         }
1371         if (!IS_ERR(req))
1372                 fuse_put_request(fc, req);
1373         if (res > 0)
1374                 *ppos = pos;
1375
1376         return res;
1377 }
1378 EXPORT_SYMBOL_GPL(fuse_direct_io);
1379
1380 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1381                                   struct iov_iter *iter,
1382                                   loff_t *ppos)
1383 {
1384         ssize_t res;
1385         struct file *file = io->file;
1386         struct inode *inode = file_inode(file);
1387
1388         if (is_bad_inode(inode))
1389                 return -EIO;
1390
1391         res = fuse_direct_io(io, iter, ppos, 0);
1392
1393         fuse_invalidate_attr(inode);
1394
1395         return res;
1396 }
1397
1398 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1399                                      size_t count, loff_t *ppos)
1400 {
1401         struct fuse_io_priv io = { .async = 0, .file = file };
1402         struct iovec iov = { .iov_base = buf, .iov_len = count };
1403         struct iov_iter ii;
1404         iov_iter_init(&ii, READ, &iov, 1, count);
1405         return __fuse_direct_read(&io, &ii, ppos);
1406 }
1407
1408 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1409                                    struct iov_iter *iter,
1410                                    loff_t *ppos)
1411 {
1412         struct file *file = io->file;
1413         struct inode *inode = file_inode(file);
1414         size_t count = iov_iter_count(iter);
1415         ssize_t res;
1416
1417
1418         res = generic_write_checks(file, ppos, &count, 0);
1419         if (!res) {
1420                 iov_iter_truncate(iter, count);
1421                 res = fuse_direct_io(io, iter, ppos, FUSE_DIO_WRITE);
1422         }
1423
1424         fuse_invalidate_attr(inode);
1425
1426         return res;
1427 }
1428
1429 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1430                                  size_t count, loff_t *ppos)
1431 {
1432         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1433         struct inode *inode = file_inode(file);
1434         ssize_t res;
1435         struct fuse_io_priv io = { .async = 0, .file = file };
1436         struct iov_iter ii;
1437         iov_iter_init(&ii, WRITE, &iov, 1, count);
1438
1439         if (is_bad_inode(inode))
1440                 return -EIO;
1441
1442         /* Don't allow parallel writes to the same file */
1443         mutex_lock(&inode->i_mutex);
1444         res = __fuse_direct_write(&io, &ii, ppos);
1445         if (res > 0)
1446                 fuse_write_update_size(inode, *ppos);
1447         mutex_unlock(&inode->i_mutex);
1448
1449         return res;
1450 }
1451
1452 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1453 {
1454         int i;
1455
1456         for (i = 0; i < req->num_pages; i++)
1457                 __free_page(req->pages[i]);
1458
1459         if (req->ff)
1460                 fuse_file_put(req->ff, false);
1461 }
1462
1463 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1464 {
1465         struct inode *inode = req->inode;
1466         struct fuse_inode *fi = get_fuse_inode(inode);
1467         struct backing_dev_info *bdi = inode_to_bdi(inode);
1468         int i;
1469
1470         list_del(&req->writepages_entry);
1471         for (i = 0; i < req->num_pages; i++) {
1472                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1473                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1474                 bdi_writeout_inc(bdi);
1475         }
1476         wake_up(&fi->page_waitq);
1477 }
1478
1479 /* Called under fc->lock, may release and reacquire it */
1480 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1481                                 loff_t size)
1482 __releases(fc->lock)
1483 __acquires(fc->lock)
1484 {
1485         struct fuse_inode *fi = get_fuse_inode(req->inode);
1486         struct fuse_write_in *inarg = &req->misc.write.in;
1487         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1488
1489         if (!fc->connected)
1490                 goto out_free;
1491
1492         if (inarg->offset + data_size <= size) {
1493                 inarg->size = data_size;
1494         } else if (inarg->offset < size) {
1495                 inarg->size = size - inarg->offset;
1496         } else {
1497                 /* Got truncated off completely */
1498                 goto out_free;
1499         }
1500
1501         req->in.args[1].size = inarg->size;
1502         fi->writectr++;
1503         fuse_request_send_background_locked(fc, req);
1504         return;
1505
1506  out_free:
1507         fuse_writepage_finish(fc, req);
1508         spin_unlock(&fc->lock);
1509         fuse_writepage_free(fc, req);
1510         fuse_put_request(fc, req);
1511         spin_lock(&fc->lock);
1512 }
1513
1514 /*
1515  * If fi->writectr is positive (no truncate or fsync going on) send
1516  * all queued writepage requests.
1517  *
1518  * Called with fc->lock
1519  */
1520 void fuse_flush_writepages(struct inode *inode)
1521 __releases(fc->lock)
1522 __acquires(fc->lock)
1523 {
1524         struct fuse_conn *fc = get_fuse_conn(inode);
1525         struct fuse_inode *fi = get_fuse_inode(inode);
1526         size_t crop = i_size_read(inode);
1527         struct fuse_req *req;
1528
1529         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1530                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1531                 list_del_init(&req->list);
1532                 fuse_send_writepage(fc, req, crop);
1533         }
1534 }
1535
1536 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1537 {
1538         struct inode *inode = req->inode;
1539         struct fuse_inode *fi = get_fuse_inode(inode);
1540
1541         mapping_set_error(inode->i_mapping, req->out.h.error);
1542         spin_lock(&fc->lock);
1543         while (req->misc.write.next) {
1544                 struct fuse_conn *fc = get_fuse_conn(inode);
1545                 struct fuse_write_in *inarg = &req->misc.write.in;
1546                 struct fuse_req *next = req->misc.write.next;
1547                 req->misc.write.next = next->misc.write.next;
1548                 next->misc.write.next = NULL;
1549                 next->ff = fuse_file_get(req->ff);
1550                 list_add(&next->writepages_entry, &fi->writepages);
1551
1552                 /*
1553                  * Skip fuse_flush_writepages() to make it easy to crop requests
1554                  * based on primary request size.
1555                  *
1556                  * 1st case (trivial): there are no concurrent activities using
1557                  * fuse_set/release_nowrite.  Then we're on safe side because
1558                  * fuse_flush_writepages() would call fuse_send_writepage()
1559                  * anyway.
1560                  *
1561                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1562                  * now for completion of all in-flight requests.  This happens
1563                  * rarely and no more than once per page, so this should be
1564                  * okay.
1565                  *
1566                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1567                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1568                  * that fuse_set_nowrite returned implies that all in-flight
1569                  * requests were completed along with all of their secondary
1570                  * requests.  Further primary requests are blocked by negative
1571                  * writectr.  Hence there cannot be any in-flight requests and
1572                  * no invocations of fuse_writepage_end() while we're in
1573                  * fuse_set_nowrite..fuse_release_nowrite section.
1574                  */
1575                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1576         }
1577         fi->writectr--;
1578         fuse_writepage_finish(fc, req);
1579         spin_unlock(&fc->lock);
1580         fuse_writepage_free(fc, req);
1581 }
1582
1583 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1584                                                struct fuse_inode *fi)
1585 {
1586         struct fuse_file *ff = NULL;
1587
1588         spin_lock(&fc->lock);
1589         if (!list_empty(&fi->write_files)) {
1590                 ff = list_entry(fi->write_files.next, struct fuse_file,
1591                                 write_entry);
1592                 fuse_file_get(ff);
1593         }
1594         spin_unlock(&fc->lock);
1595
1596         return ff;
1597 }
1598
1599 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1600                                              struct fuse_inode *fi)
1601 {
1602         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1603         WARN_ON(!ff);
1604         return ff;
1605 }
1606
1607 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1608 {
1609         struct fuse_conn *fc = get_fuse_conn(inode);
1610         struct fuse_inode *fi = get_fuse_inode(inode);
1611         struct fuse_file *ff;
1612         int err;
1613
1614         ff = __fuse_write_file_get(fc, fi);
1615         err = fuse_flush_times(inode, ff);
1616         if (ff)
1617                 fuse_file_put(ff, 0);
1618
1619         return err;
1620 }
1621
1622 static int fuse_writepage_locked(struct page *page)
1623 {
1624         struct address_space *mapping = page->mapping;
1625         struct inode *inode = mapping->host;
1626         struct fuse_conn *fc = get_fuse_conn(inode);
1627         struct fuse_inode *fi = get_fuse_inode(inode);
1628         struct fuse_req *req;
1629         struct page *tmp_page;
1630         int error = -ENOMEM;
1631
1632         set_page_writeback(page);
1633
1634         req = fuse_request_alloc_nofs(1);
1635         if (!req)
1636                 goto err;
1637
1638         req->background = 1; /* writeback always goes to bg_queue */
1639         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1640         if (!tmp_page)
1641                 goto err_free;
1642
1643         error = -EIO;
1644         req->ff = fuse_write_file_get(fc, fi);
1645         if (!req->ff)
1646                 goto err_nofile;
1647
1648         fuse_write_fill(req, req->ff, page_offset(page), 0);
1649
1650         copy_highpage(tmp_page, page);
1651         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1652         req->misc.write.next = NULL;
1653         req->in.argpages = 1;
1654         req->num_pages = 1;
1655         req->pages[0] = tmp_page;
1656         req->page_descs[0].offset = 0;
1657         req->page_descs[0].length = PAGE_SIZE;
1658         req->end = fuse_writepage_end;
1659         req->inode = inode;
1660
1661         inc_bdi_stat(inode_to_bdi(inode), BDI_WRITEBACK);
1662         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1663
1664         spin_lock(&fc->lock);
1665         list_add(&req->writepages_entry, &fi->writepages);
1666         list_add_tail(&req->list, &fi->queued_writes);
1667         fuse_flush_writepages(inode);
1668         spin_unlock(&fc->lock);
1669
1670         end_page_writeback(page);
1671
1672         return 0;
1673
1674 err_nofile:
1675         __free_page(tmp_page);
1676 err_free:
1677         fuse_request_free(req);
1678 err:
1679         end_page_writeback(page);
1680         return error;
1681 }
1682
1683 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1684 {
1685         int err;
1686
1687         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1688                 /*
1689                  * ->writepages() should be called for sync() and friends.  We
1690                  * should only get here on direct reclaim and then we are
1691                  * allowed to skip a page which is already in flight
1692                  */
1693                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1694
1695                 redirty_page_for_writepage(wbc, page);
1696                 return 0;
1697         }
1698
1699         err = fuse_writepage_locked(page);
1700         unlock_page(page);
1701
1702         return err;
1703 }
1704
1705 struct fuse_fill_wb_data {
1706         struct fuse_req *req;
1707         struct fuse_file *ff;
1708         struct inode *inode;
1709         struct page **orig_pages;
1710 };
1711
1712 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1713 {
1714         struct fuse_req *req = data->req;
1715         struct inode *inode = data->inode;
1716         struct fuse_conn *fc = get_fuse_conn(inode);
1717         struct fuse_inode *fi = get_fuse_inode(inode);
1718         int num_pages = req->num_pages;
1719         int i;
1720
1721         req->ff = fuse_file_get(data->ff);
1722         spin_lock(&fc->lock);
1723         list_add_tail(&req->list, &fi->queued_writes);
1724         fuse_flush_writepages(inode);
1725         spin_unlock(&fc->lock);
1726
1727         for (i = 0; i < num_pages; i++)
1728                 end_page_writeback(data->orig_pages[i]);
1729 }
1730
1731 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1732                                      struct page *page)
1733 {
1734         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1735         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1736         struct fuse_req *tmp;
1737         struct fuse_req *old_req;
1738         bool found = false;
1739         pgoff_t curr_index;
1740
1741         BUG_ON(new_req->num_pages != 0);
1742
1743         spin_lock(&fc->lock);
1744         list_del(&new_req->writepages_entry);
1745         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1746                 BUG_ON(old_req->inode != new_req->inode);
1747                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1748                 if (curr_index <= page->index &&
1749                     page->index < curr_index + old_req->num_pages) {
1750                         found = true;
1751                         break;
1752                 }
1753         }
1754         if (!found) {
1755                 list_add(&new_req->writepages_entry, &fi->writepages);
1756                 goto out_unlock;
1757         }
1758
1759         new_req->num_pages = 1;
1760         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1761                 BUG_ON(tmp->inode != new_req->inode);
1762                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1763                 if (tmp->num_pages == 1 &&
1764                     curr_index == page->index) {
1765                         old_req = tmp;
1766                 }
1767         }
1768
1769         if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1770                                         old_req->state == FUSE_REQ_PENDING)) {
1771                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1772
1773                 copy_highpage(old_req->pages[0], page);
1774                 spin_unlock(&fc->lock);
1775
1776                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1777                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1778                 bdi_writeout_inc(bdi);
1779                 fuse_writepage_free(fc, new_req);
1780                 fuse_request_free(new_req);
1781                 goto out;
1782         } else {
1783                 new_req->misc.write.next = old_req->misc.write.next;
1784                 old_req->misc.write.next = new_req;
1785         }
1786 out_unlock:
1787         spin_unlock(&fc->lock);
1788 out:
1789         return found;
1790 }
1791
1792 static int fuse_writepages_fill(struct page *page,
1793                 struct writeback_control *wbc, void *_data)
1794 {
1795         struct fuse_fill_wb_data *data = _data;
1796         struct fuse_req *req = data->req;
1797         struct inode *inode = data->inode;
1798         struct fuse_conn *fc = get_fuse_conn(inode);
1799         struct page *tmp_page;
1800         bool is_writeback;
1801         int err;
1802
1803         if (!data->ff) {
1804                 err = -EIO;
1805                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1806                 if (!data->ff)
1807                         goto out_unlock;
1808         }
1809
1810         /*
1811          * Being under writeback is unlikely but possible.  For example direct
1812          * read to an mmaped fuse file will set the page dirty twice; once when
1813          * the pages are faulted with get_user_pages(), and then after the read
1814          * completed.
1815          */
1816         is_writeback = fuse_page_is_writeback(inode, page->index);
1817
1818         if (req && req->num_pages &&
1819             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1820              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1821              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1822                 fuse_writepages_send(data);
1823                 data->req = NULL;
1824         }
1825         err = -ENOMEM;
1826         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1827         if (!tmp_page)
1828                 goto out_unlock;
1829
1830         /*
1831          * The page must not be redirtied until the writeout is completed
1832          * (i.e. userspace has sent a reply to the write request).  Otherwise
1833          * there could be more than one temporary page instance for each real
1834          * page.
1835          *
1836          * This is ensured by holding the page lock in page_mkwrite() while
1837          * checking fuse_page_is_writeback().  We already hold the page lock
1838          * since clear_page_dirty_for_io() and keep it held until we add the
1839          * request to the fi->writepages list and increment req->num_pages.
1840          * After this fuse_page_is_writeback() will indicate that the page is
1841          * under writeback, so we can release the page lock.
1842          */
1843         if (data->req == NULL) {
1844                 struct fuse_inode *fi = get_fuse_inode(inode);
1845
1846                 err = -ENOMEM;
1847                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1848                 if (!req) {
1849                         __free_page(tmp_page);
1850                         goto out_unlock;
1851                 }
1852
1853                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1854                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1855                 req->misc.write.next = NULL;
1856                 req->in.argpages = 1;
1857                 req->background = 1;
1858                 req->num_pages = 0;
1859                 req->end = fuse_writepage_end;
1860                 req->inode = inode;
1861
1862                 spin_lock(&fc->lock);
1863                 list_add(&req->writepages_entry, &fi->writepages);
1864                 spin_unlock(&fc->lock);
1865
1866                 data->req = req;
1867         }
1868         set_page_writeback(page);
1869
1870         copy_highpage(tmp_page, page);
1871         req->pages[req->num_pages] = tmp_page;
1872         req->page_descs[req->num_pages].offset = 0;
1873         req->page_descs[req->num_pages].length = PAGE_SIZE;
1874
1875         inc_bdi_stat(inode_to_bdi(inode), BDI_WRITEBACK);
1876         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1877
1878         err = 0;
1879         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1880                 end_page_writeback(page);
1881                 data->req = NULL;
1882                 goto out_unlock;
1883         }
1884         data->orig_pages[req->num_pages] = page;
1885
1886         /*
1887          * Protected by fc->lock against concurrent access by
1888          * fuse_page_is_writeback().
1889          */
1890         spin_lock(&fc->lock);
1891         req->num_pages++;
1892         spin_unlock(&fc->lock);
1893
1894 out_unlock:
1895         unlock_page(page);
1896
1897         return err;
1898 }
1899
1900 static int fuse_writepages(struct address_space *mapping,
1901                            struct writeback_control *wbc)
1902 {
1903         struct inode *inode = mapping->host;
1904         struct fuse_fill_wb_data data;
1905         int err;
1906
1907         err = -EIO;
1908         if (is_bad_inode(inode))
1909                 goto out;
1910
1911         data.inode = inode;
1912         data.req = NULL;
1913         data.ff = NULL;
1914
1915         err = -ENOMEM;
1916         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1917                                   sizeof(struct page *),
1918                                   GFP_NOFS);
1919         if (!data.orig_pages)
1920                 goto out;
1921
1922         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1923         if (data.req) {
1924                 /* Ignore errors if we can write at least one page */
1925                 BUG_ON(!data.req->num_pages);
1926                 fuse_writepages_send(&data);
1927                 err = 0;
1928         }
1929         if (data.ff)
1930                 fuse_file_put(data.ff, false);
1931
1932         kfree(data.orig_pages);
1933 out:
1934         return err;
1935 }
1936
1937 /*
1938  * It's worthy to make sure that space is reserved on disk for the write,
1939  * but how to implement it without killing performance need more thinking.
1940  */
1941 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1942                 loff_t pos, unsigned len, unsigned flags,
1943                 struct page **pagep, void **fsdata)
1944 {
1945         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1946         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1947         struct page *page;
1948         loff_t fsize;
1949         int err = -ENOMEM;
1950
1951         WARN_ON(!fc->writeback_cache);
1952
1953         page = grab_cache_page_write_begin(mapping, index, flags);
1954         if (!page)
1955                 goto error;
1956
1957         fuse_wait_on_page_writeback(mapping->host, page->index);
1958
1959         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1960                 goto success;
1961         /*
1962          * Check if the start this page comes after the end of file, in which
1963          * case the readpage can be optimized away.
1964          */
1965         fsize = i_size_read(mapping->host);
1966         if (fsize <= (pos & PAGE_CACHE_MASK)) {
1967                 size_t off = pos & ~PAGE_CACHE_MASK;
1968                 if (off)
1969                         zero_user_segment(page, 0, off);
1970                 goto success;
1971         }
1972         err = fuse_do_readpage(file, page);
1973         if (err)
1974                 goto cleanup;
1975 success:
1976         *pagep = page;
1977         return 0;
1978
1979 cleanup:
1980         unlock_page(page);
1981         page_cache_release(page);
1982 error:
1983         return err;
1984 }
1985
1986 static int fuse_write_end(struct file *file, struct address_space *mapping,
1987                 loff_t pos, unsigned len, unsigned copied,
1988                 struct page *page, void *fsdata)
1989 {
1990         struct inode *inode = page->mapping->host;
1991
1992         if (!PageUptodate(page)) {
1993                 /* Zero any unwritten bytes at the end of the page */
1994                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
1995                 if (endoff)
1996                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
1997                 SetPageUptodate(page);
1998         }
1999
2000         fuse_write_update_size(inode, pos + copied);
2001         set_page_dirty(page);
2002         unlock_page(page);
2003         page_cache_release(page);
2004
2005         return copied;
2006 }
2007
2008 static int fuse_launder_page(struct page *page)
2009 {
2010         int err = 0;
2011         if (clear_page_dirty_for_io(page)) {
2012                 struct inode *inode = page->mapping->host;
2013                 err = fuse_writepage_locked(page);
2014                 if (!err)
2015                         fuse_wait_on_page_writeback(inode, page->index);
2016         }
2017         return err;
2018 }
2019
2020 /*
2021  * Write back dirty pages now, because there may not be any suitable
2022  * open files later
2023  */
2024 static void fuse_vma_close(struct vm_area_struct *vma)
2025 {
2026         filemap_write_and_wait(vma->vm_file->f_mapping);
2027 }
2028
2029 /*
2030  * Wait for writeback against this page to complete before allowing it
2031  * to be marked dirty again, and hence written back again, possibly
2032  * before the previous writepage completed.
2033  *
2034  * Block here, instead of in ->writepage(), so that the userspace fs
2035  * can only block processes actually operating on the filesystem.
2036  *
2037  * Otherwise unprivileged userspace fs would be able to block
2038  * unrelated:
2039  *
2040  * - page migration
2041  * - sync(2)
2042  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2043  */
2044 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2045 {
2046         struct page *page = vmf->page;
2047         struct inode *inode = file_inode(vma->vm_file);
2048
2049         file_update_time(vma->vm_file);
2050         lock_page(page);
2051         if (page->mapping != inode->i_mapping) {
2052                 unlock_page(page);
2053                 return VM_FAULT_NOPAGE;
2054         }
2055
2056         fuse_wait_on_page_writeback(inode, page->index);
2057         return VM_FAULT_LOCKED;
2058 }
2059
2060 static const struct vm_operations_struct fuse_file_vm_ops = {
2061         .close          = fuse_vma_close,
2062         .fault          = filemap_fault,
2063         .map_pages      = filemap_map_pages,
2064         .page_mkwrite   = fuse_page_mkwrite,
2065 };
2066
2067 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2068 {
2069         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2070                 fuse_link_write_file(file);
2071
2072         file_accessed(file);
2073         vma->vm_ops = &fuse_file_vm_ops;
2074         return 0;
2075 }
2076
2077 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2078 {
2079         /* Can't provide the coherency needed for MAP_SHARED */
2080         if (vma->vm_flags & VM_MAYSHARE)
2081                 return -ENODEV;
2082
2083         invalidate_inode_pages2(file->f_mapping);
2084
2085         return generic_file_mmap(file, vma);
2086 }
2087
2088 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2089                                   struct file_lock *fl)
2090 {
2091         switch (ffl->type) {
2092         case F_UNLCK:
2093                 break;
2094
2095         case F_RDLCK:
2096         case F_WRLCK:
2097                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2098                     ffl->end < ffl->start)
2099                         return -EIO;
2100
2101                 fl->fl_start = ffl->start;
2102                 fl->fl_end = ffl->end;
2103                 fl->fl_pid = ffl->pid;
2104                 break;
2105
2106         default:
2107                 return -EIO;
2108         }
2109         fl->fl_type = ffl->type;
2110         return 0;
2111 }
2112
2113 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2114                          const struct file_lock *fl, int opcode, pid_t pid,
2115                          int flock, struct fuse_lk_in *inarg)
2116 {
2117         struct inode *inode = file_inode(file);
2118         struct fuse_conn *fc = get_fuse_conn(inode);
2119         struct fuse_file *ff = file->private_data;
2120
2121         memset(inarg, 0, sizeof(*inarg));
2122         inarg->fh = ff->fh;
2123         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2124         inarg->lk.start = fl->fl_start;
2125         inarg->lk.end = fl->fl_end;
2126         inarg->lk.type = fl->fl_type;
2127         inarg->lk.pid = pid;
2128         if (flock)
2129                 inarg->lk_flags |= FUSE_LK_FLOCK;
2130         args->in.h.opcode = opcode;
2131         args->in.h.nodeid = get_node_id(inode);
2132         args->in.numargs = 1;
2133         args->in.args[0].size = sizeof(*inarg);
2134         args->in.args[0].value = inarg;
2135 }
2136
2137 static int fuse_getlk(struct file *file, struct file_lock *fl)
2138 {
2139         struct inode *inode = file_inode(file);
2140         struct fuse_conn *fc = get_fuse_conn(inode);
2141         FUSE_ARGS(args);
2142         struct fuse_lk_in inarg;
2143         struct fuse_lk_out outarg;
2144         int err;
2145
2146         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2147         args.out.numargs = 1;
2148         args.out.args[0].size = sizeof(outarg);
2149         args.out.args[0].value = &outarg;
2150         err = fuse_simple_request(fc, &args);
2151         if (!err)
2152                 err = convert_fuse_file_lock(&outarg.lk, fl);
2153
2154         return err;
2155 }
2156
2157 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2158 {
2159         struct inode *inode = file_inode(file);
2160         struct fuse_conn *fc = get_fuse_conn(inode);
2161         FUSE_ARGS(args);
2162         struct fuse_lk_in inarg;
2163         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2164         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2165         int err;
2166
2167         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2168                 /* NLM needs asynchronous locks, which we don't support yet */
2169                 return -ENOLCK;
2170         }
2171
2172         /* Unlock on close is handled by the flush method */
2173         if (fl->fl_flags & FL_CLOSE)
2174                 return 0;
2175
2176         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2177         err = fuse_simple_request(fc, &args);
2178
2179         /* locking is restartable */
2180         if (err == -EINTR)
2181                 err = -ERESTARTSYS;
2182
2183         return err;
2184 }
2185
2186 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2187 {
2188         struct inode *inode = file_inode(file);
2189         struct fuse_conn *fc = get_fuse_conn(inode);
2190         int err;
2191
2192         if (cmd == F_CANCELLK) {
2193                 err = 0;
2194         } else if (cmd == F_GETLK) {
2195                 if (fc->no_lock) {
2196                         posix_test_lock(file, fl);
2197                         err = 0;
2198                 } else
2199                         err = fuse_getlk(file, fl);
2200         } else {
2201                 if (fc->no_lock)
2202                         err = posix_lock_file(file, fl, NULL);
2203                 else
2204                         err = fuse_setlk(file, fl, 0);
2205         }
2206         return err;
2207 }
2208
2209 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2210 {
2211         struct inode *inode = file_inode(file);
2212         struct fuse_conn *fc = get_fuse_conn(inode);
2213         int err;
2214
2215         if (fc->no_flock) {
2216                 err = flock_lock_file_wait(file, fl);
2217         } else {
2218                 struct fuse_file *ff = file->private_data;
2219
2220                 /* emulate flock with POSIX locks */
2221                 ff->flock = true;
2222                 err = fuse_setlk(file, fl, 1);
2223         }
2224
2225         return err;
2226 }
2227
2228 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2229 {
2230         struct inode *inode = mapping->host;
2231         struct fuse_conn *fc = get_fuse_conn(inode);
2232         FUSE_ARGS(args);
2233         struct fuse_bmap_in inarg;
2234         struct fuse_bmap_out outarg;
2235         int err;
2236
2237         if (!inode->i_sb->s_bdev || fc->no_bmap)
2238                 return 0;
2239
2240         memset(&inarg, 0, sizeof(inarg));
2241         inarg.block = block;
2242         inarg.blocksize = inode->i_sb->s_blocksize;
2243         args.in.h.opcode = FUSE_BMAP;
2244         args.in.h.nodeid = get_node_id(inode);
2245         args.in.numargs = 1;
2246         args.in.args[0].size = sizeof(inarg);
2247         args.in.args[0].value = &inarg;
2248         args.out.numargs = 1;
2249         args.out.args[0].size = sizeof(outarg);
2250         args.out.args[0].value = &outarg;
2251         err = fuse_simple_request(fc, &args);
2252         if (err == -ENOSYS)
2253                 fc->no_bmap = 1;
2254
2255         return err ? 0 : outarg.block;
2256 }
2257
2258 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2259 {
2260         loff_t retval;
2261         struct inode *inode = file_inode(file);
2262
2263         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2264         if (whence == SEEK_CUR || whence == SEEK_SET)
2265                 return generic_file_llseek(file, offset, whence);
2266
2267         mutex_lock(&inode->i_mutex);
2268         retval = fuse_update_attributes(inode, NULL, file, NULL);
2269         if (!retval)
2270                 retval = generic_file_llseek(file, offset, whence);
2271         mutex_unlock(&inode->i_mutex);
2272
2273         return retval;
2274 }
2275
2276 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2277                         unsigned int nr_segs, size_t bytes, bool to_user)
2278 {
2279         struct iov_iter ii;
2280         int page_idx = 0;
2281
2282         if (!bytes)
2283                 return 0;
2284
2285         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2286
2287         while (iov_iter_count(&ii)) {
2288                 struct page *page = pages[page_idx++];
2289                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2290                 void *kaddr;
2291
2292                 kaddr = kmap(page);
2293
2294                 while (todo) {
2295                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2296                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2297                         size_t copy = min(todo, iov_len);
2298                         size_t left;
2299
2300                         if (!to_user)
2301                                 left = copy_from_user(kaddr, uaddr, copy);
2302                         else
2303                                 left = copy_to_user(uaddr, kaddr, copy);
2304
2305                         if (unlikely(left))
2306                                 return -EFAULT;
2307
2308                         iov_iter_advance(&ii, copy);
2309                         todo -= copy;
2310                         kaddr += copy;
2311                 }
2312
2313                 kunmap(page);
2314         }
2315
2316         return 0;
2317 }
2318
2319 /*
2320  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2321  * ABI was defined to be 'struct iovec' which is different on 32bit
2322  * and 64bit.  Fortunately we can determine which structure the server
2323  * used from the size of the reply.
2324  */
2325 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2326                                      size_t transferred, unsigned count,
2327                                      bool is_compat)
2328 {
2329 #ifdef CONFIG_COMPAT
2330         if (count * sizeof(struct compat_iovec) == transferred) {
2331                 struct compat_iovec *ciov = src;
2332                 unsigned i;
2333
2334                 /*
2335                  * With this interface a 32bit server cannot support
2336                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2337                  * requests
2338                  */
2339                 if (!is_compat)
2340                         return -EINVAL;
2341
2342                 for (i = 0; i < count; i++) {
2343                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2344                         dst[i].iov_len = ciov[i].iov_len;
2345                 }
2346                 return 0;
2347         }
2348 #endif
2349
2350         if (count * sizeof(struct iovec) != transferred)
2351                 return -EIO;
2352
2353         memcpy(dst, src, transferred);
2354         return 0;
2355 }
2356
2357 /* Make sure iov_length() won't overflow */
2358 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2359 {
2360         size_t n;
2361         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2362
2363         for (n = 0; n < count; n++, iov++) {
2364                 if (iov->iov_len > (size_t) max)
2365                         return -ENOMEM;
2366                 max -= iov->iov_len;
2367         }
2368         return 0;
2369 }
2370
2371 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2372                                  void *src, size_t transferred, unsigned count,
2373                                  bool is_compat)
2374 {
2375         unsigned i;
2376         struct fuse_ioctl_iovec *fiov = src;
2377
2378         if (fc->minor < 16) {
2379                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2380                                                  count, is_compat);
2381         }
2382
2383         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2384                 return -EIO;
2385
2386         for (i = 0; i < count; i++) {
2387                 /* Did the server supply an inappropriate value? */
2388                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2389                     fiov[i].len != (unsigned long) fiov[i].len)
2390                         return -EIO;
2391
2392                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2393                 dst[i].iov_len = (size_t) fiov[i].len;
2394
2395 #ifdef CONFIG_COMPAT
2396                 if (is_compat &&
2397                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2398                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2399                         return -EIO;
2400 #endif
2401         }
2402
2403         return 0;
2404 }
2405
2406
2407 /*
2408  * For ioctls, there is no generic way to determine how much memory
2409  * needs to be read and/or written.  Furthermore, ioctls are allowed
2410  * to dereference the passed pointer, so the parameter requires deep
2411  * copying but FUSE has no idea whatsoever about what to copy in or
2412  * out.
2413  *
2414  * This is solved by allowing FUSE server to retry ioctl with
2415  * necessary in/out iovecs.  Let's assume the ioctl implementation
2416  * needs to read in the following structure.
2417  *
2418  * struct a {
2419  *      char    *buf;
2420  *      size_t  buflen;
2421  * }
2422  *
2423  * On the first callout to FUSE server, inarg->in_size and
2424  * inarg->out_size will be NULL; then, the server completes the ioctl
2425  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2426  * the actual iov array to
2427  *
2428  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2429  *
2430  * which tells FUSE to copy in the requested area and retry the ioctl.
2431  * On the second round, the server has access to the structure and
2432  * from that it can tell what to look for next, so on the invocation,
2433  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2434  *
2435  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2436  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2437  *
2438  * FUSE will copy both struct a and the pointed buffer from the
2439  * process doing the ioctl and retry ioctl with both struct a and the
2440  * buffer.
2441  *
2442  * This time, FUSE server has everything it needs and completes ioctl
2443  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2444  *
2445  * Copying data out works the same way.
2446  *
2447  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2448  * automatically initializes in and out iovs by decoding @cmd with
2449  * _IOC_* macros and the server is not allowed to request RETRY.  This
2450  * limits ioctl data transfers to well-formed ioctls and is the forced
2451  * behavior for all FUSE servers.
2452  */
2453 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2454                    unsigned int flags)
2455 {
2456         struct fuse_file *ff = file->private_data;
2457         struct fuse_conn *fc = ff->fc;
2458         struct fuse_ioctl_in inarg = {
2459                 .fh = ff->fh,
2460                 .cmd = cmd,
2461                 .arg = arg,
2462                 .flags = flags
2463         };
2464         struct fuse_ioctl_out outarg;
2465         struct fuse_req *req = NULL;
2466         struct page **pages = NULL;
2467         struct iovec *iov_page = NULL;
2468         struct iovec *in_iov = NULL, *out_iov = NULL;
2469         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2470         size_t in_size, out_size, transferred;
2471         int err;
2472
2473 #if BITS_PER_LONG == 32
2474         inarg.flags |= FUSE_IOCTL_32BIT;
2475 #else
2476         if (flags & FUSE_IOCTL_COMPAT)
2477                 inarg.flags |= FUSE_IOCTL_32BIT;
2478 #endif
2479
2480         /* assume all the iovs returned by client always fits in a page */
2481         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2482
2483         err = -ENOMEM;
2484         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2485         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2486         if (!pages || !iov_page)
2487                 goto out;
2488
2489         /*
2490          * If restricted, initialize IO parameters as encoded in @cmd.
2491          * RETRY from server is not allowed.
2492          */
2493         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2494                 struct iovec *iov = iov_page;
2495
2496                 iov->iov_base = (void __user *)arg;
2497                 iov->iov_len = _IOC_SIZE(cmd);
2498
2499                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2500                         in_iov = iov;
2501                         in_iovs = 1;
2502                 }
2503
2504                 if (_IOC_DIR(cmd) & _IOC_READ) {
2505                         out_iov = iov;
2506                         out_iovs = 1;
2507                 }
2508         }
2509
2510  retry:
2511         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2512         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2513
2514         /*
2515          * Out data can be used either for actual out data or iovs,
2516          * make sure there always is at least one page.
2517          */
2518         out_size = max_t(size_t, out_size, PAGE_SIZE);
2519         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2520
2521         /* make sure there are enough buffer pages and init request with them */
2522         err = -ENOMEM;
2523         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2524                 goto out;
2525         while (num_pages < max_pages) {
2526                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2527                 if (!pages[num_pages])
2528                         goto out;
2529                 num_pages++;
2530         }
2531
2532         req = fuse_get_req(fc, num_pages);
2533         if (IS_ERR(req)) {
2534                 err = PTR_ERR(req);
2535                 req = NULL;
2536                 goto out;
2537         }
2538         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2539         req->num_pages = num_pages;
2540         fuse_page_descs_length_init(req, 0, req->num_pages);
2541
2542         /* okay, let's send it to the client */
2543         req->in.h.opcode = FUSE_IOCTL;
2544         req->in.h.nodeid = ff->nodeid;
2545         req->in.numargs = 1;
2546         req->in.args[0].size = sizeof(inarg);
2547         req->in.args[0].value = &inarg;
2548         if (in_size) {
2549                 req->in.numargs++;
2550                 req->in.args[1].size = in_size;
2551                 req->in.argpages = 1;
2552
2553                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2554                                            false);
2555                 if (err)
2556                         goto out;
2557         }
2558
2559         req->out.numargs = 2;
2560         req->out.args[0].size = sizeof(outarg);
2561         req->out.args[0].value = &outarg;
2562         req->out.args[1].size = out_size;
2563         req->out.argpages = 1;
2564         req->out.argvar = 1;
2565
2566         fuse_request_send(fc, req);
2567         err = req->out.h.error;
2568         transferred = req->out.args[1].size;
2569         fuse_put_request(fc, req);
2570         req = NULL;
2571         if (err)
2572                 goto out;
2573
2574         /* did it ask for retry? */
2575         if (outarg.flags & FUSE_IOCTL_RETRY) {
2576                 void *vaddr;
2577
2578                 /* no retry if in restricted mode */
2579                 err = -EIO;
2580                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2581                         goto out;
2582
2583                 in_iovs = outarg.in_iovs;
2584                 out_iovs = outarg.out_iovs;
2585
2586                 /*
2587                  * Make sure things are in boundary, separate checks
2588                  * are to protect against overflow.
2589                  */
2590                 err = -ENOMEM;
2591                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2592                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2593                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2594                         goto out;
2595
2596                 vaddr = kmap_atomic(pages[0]);
2597                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2598                                             transferred, in_iovs + out_iovs,
2599                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2600                 kunmap_atomic(vaddr);
2601                 if (err)
2602                         goto out;
2603
2604                 in_iov = iov_page;
2605                 out_iov = in_iov + in_iovs;
2606
2607                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2608                 if (err)
2609                         goto out;
2610
2611                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2612                 if (err)
2613                         goto out;
2614
2615                 goto retry;
2616         }
2617
2618         err = -EIO;
2619         if (transferred > inarg.out_size)
2620                 goto out;
2621
2622         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2623  out:
2624         if (req)
2625                 fuse_put_request(fc, req);
2626         free_page((unsigned long) iov_page);
2627         while (num_pages)
2628                 __free_page(pages[--num_pages]);
2629         kfree(pages);
2630
2631         return err ? err : outarg.result;
2632 }
2633 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2634
2635 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2636                        unsigned long arg, unsigned int flags)
2637 {
2638         struct inode *inode = file_inode(file);
2639         struct fuse_conn *fc = get_fuse_conn(inode);
2640
2641         if (!fuse_allow_current_process(fc))
2642                 return -EACCES;
2643
2644         if (is_bad_inode(inode))
2645                 return -EIO;
2646
2647         return fuse_do_ioctl(file, cmd, arg, flags);
2648 }
2649
2650 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2651                             unsigned long arg)
2652 {
2653         return fuse_ioctl_common(file, cmd, arg, 0);
2654 }
2655
2656 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2657                                    unsigned long arg)
2658 {
2659         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2660 }
2661
2662 /*
2663  * All files which have been polled are linked to RB tree
2664  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2665  * find the matching one.
2666  */
2667 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2668                                               struct rb_node **parent_out)
2669 {
2670         struct rb_node **link = &fc->polled_files.rb_node;
2671         struct rb_node *last = NULL;
2672
2673         while (*link) {
2674                 struct fuse_file *ff;
2675
2676                 last = *link;
2677                 ff = rb_entry(last, struct fuse_file, polled_node);
2678
2679                 if (kh < ff->kh)
2680                         link = &last->rb_left;
2681                 else if (kh > ff->kh)
2682                         link = &last->rb_right;
2683                 else
2684                         return link;
2685         }
2686
2687         if (parent_out)
2688                 *parent_out = last;
2689         return link;
2690 }
2691
2692 /*
2693  * The file is about to be polled.  Make sure it's on the polled_files
2694  * RB tree.  Note that files once added to the polled_files tree are
2695  * not removed before the file is released.  This is because a file
2696  * polled once is likely to be polled again.
2697  */
2698 static void fuse_register_polled_file(struct fuse_conn *fc,
2699                                       struct fuse_file *ff)
2700 {
2701         spin_lock(&fc->lock);
2702         if (RB_EMPTY_NODE(&ff->polled_node)) {
2703                 struct rb_node **link, *uninitialized_var(parent);
2704
2705                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2706                 BUG_ON(*link);
2707                 rb_link_node(&ff->polled_node, parent, link);
2708                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2709         }
2710         spin_unlock(&fc->lock);
2711 }
2712
2713 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2714 {
2715         struct fuse_file *ff = file->private_data;
2716         struct fuse_conn *fc = ff->fc;
2717         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2718         struct fuse_poll_out outarg;
2719         FUSE_ARGS(args);
2720         int err;
2721
2722         if (fc->no_poll)
2723                 return DEFAULT_POLLMASK;
2724
2725         poll_wait(file, &ff->poll_wait, wait);
2726         inarg.events = (__u32)poll_requested_events(wait);
2727
2728         /*
2729          * Ask for notification iff there's someone waiting for it.
2730          * The client may ignore the flag and always notify.
2731          */
2732         if (waitqueue_active(&ff->poll_wait)) {
2733                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2734                 fuse_register_polled_file(fc, ff);
2735         }
2736
2737         args.in.h.opcode = FUSE_POLL;
2738         args.in.h.nodeid = ff->nodeid;
2739         args.in.numargs = 1;
2740         args.in.args[0].size = sizeof(inarg);
2741         args.in.args[0].value = &inarg;
2742         args.out.numargs = 1;
2743         args.out.args[0].size = sizeof(outarg);
2744         args.out.args[0].value = &outarg;
2745         err = fuse_simple_request(fc, &args);
2746
2747         if (!err)
2748                 return outarg.revents;
2749         if (err == -ENOSYS) {
2750                 fc->no_poll = 1;
2751                 return DEFAULT_POLLMASK;
2752         }
2753         return POLLERR;
2754 }
2755 EXPORT_SYMBOL_GPL(fuse_file_poll);
2756
2757 /*
2758  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2759  * wakes up the poll waiters.
2760  */
2761 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2762                             struct fuse_notify_poll_wakeup_out *outarg)
2763 {
2764         u64 kh = outarg->kh;
2765         struct rb_node **link;
2766
2767         spin_lock(&fc->lock);
2768
2769         link = fuse_find_polled_node(fc, kh, NULL);
2770         if (*link) {
2771                 struct fuse_file *ff;
2772
2773                 ff = rb_entry(*link, struct fuse_file, polled_node);
2774                 wake_up_interruptible_sync(&ff->poll_wait);
2775         }
2776
2777         spin_unlock(&fc->lock);
2778         return 0;
2779 }
2780
2781 static void fuse_do_truncate(struct file *file)
2782 {
2783         struct inode *inode = file->f_mapping->host;
2784         struct iattr attr;
2785
2786         attr.ia_valid = ATTR_SIZE;
2787         attr.ia_size = i_size_read(inode);
2788
2789         attr.ia_file = file;
2790         attr.ia_valid |= ATTR_FILE;
2791
2792         fuse_do_setattr(inode, &attr, file);
2793 }
2794
2795 static inline loff_t fuse_round_up(loff_t off)
2796 {
2797         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2798 }
2799
2800 static ssize_t
2801 fuse_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter,
2802                         loff_t offset)
2803 {
2804         ssize_t ret = 0;
2805         struct file *file = iocb->ki_filp;
2806         struct fuse_file *ff = file->private_data;
2807         bool async_dio = ff->fc->async_dio;
2808         loff_t pos = 0;
2809         struct inode *inode;
2810         loff_t i_size;
2811         size_t count = iov_iter_count(iter);
2812         struct fuse_io_priv *io;
2813
2814         pos = offset;
2815         inode = file->f_mapping->host;
2816         i_size = i_size_read(inode);
2817
2818         if ((rw == READ) && (offset > i_size))
2819                 return 0;
2820
2821         /* optimization for short read */
2822         if (async_dio && rw != WRITE && offset + count > i_size) {
2823                 if (offset >= i_size)
2824                         return 0;
2825                 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2826                 iov_iter_truncate(iter, count);
2827         }
2828
2829         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2830         if (!io)
2831                 return -ENOMEM;
2832         spin_lock_init(&io->lock);
2833         io->reqs = 1;
2834         io->bytes = -1;
2835         io->size = 0;
2836         io->offset = offset;
2837         io->write = (rw == WRITE);
2838         io->err = 0;
2839         io->file = file;
2840         /*
2841          * By default, we want to optimize all I/Os with async request
2842          * submission to the client filesystem if supported.
2843          */
2844         io->async = async_dio;
2845         io->iocb = iocb;
2846
2847         /*
2848          * We cannot asynchronously extend the size of a file. We have no method
2849          * to wait on real async I/O requests, so we must submit this request
2850          * synchronously.
2851          */
2852         if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2853                 io->async = false;
2854
2855         if (rw == WRITE)
2856                 ret = __fuse_direct_write(io, iter, &pos);
2857         else
2858                 ret = __fuse_direct_read(io, iter, &pos);
2859
2860         if (io->async) {
2861                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2862
2863                 /* we have a non-extending, async request, so return */
2864                 if (!is_sync_kiocb(iocb))
2865                         return -EIOCBQUEUED;
2866
2867                 ret = wait_on_sync_kiocb(iocb);
2868         } else {
2869                 kfree(io);
2870         }
2871
2872         if (rw == WRITE) {
2873                 if (ret > 0)
2874                         fuse_write_update_size(inode, pos);
2875                 else if (ret < 0 && offset + count > i_size)
2876                         fuse_do_truncate(file);
2877         }
2878
2879         return ret;
2880 }
2881
2882 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2883                                 loff_t length)
2884 {
2885         struct fuse_file *ff = file->private_data;
2886         struct inode *inode = file_inode(file);
2887         struct fuse_inode *fi = get_fuse_inode(inode);
2888         struct fuse_conn *fc = ff->fc;
2889         FUSE_ARGS(args);
2890         struct fuse_fallocate_in inarg = {
2891                 .fh = ff->fh,
2892                 .offset = offset,
2893                 .length = length,
2894                 .mode = mode
2895         };
2896         int err;
2897         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2898                            (mode & FALLOC_FL_PUNCH_HOLE);
2899
2900         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2901                 return -EOPNOTSUPP;
2902
2903         if (fc->no_fallocate)
2904                 return -EOPNOTSUPP;
2905
2906         if (lock_inode) {
2907                 mutex_lock(&inode->i_mutex);
2908                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2909                         loff_t endbyte = offset + length - 1;
2910                         err = filemap_write_and_wait_range(inode->i_mapping,
2911                                                            offset, endbyte);
2912                         if (err)
2913                                 goto out;
2914
2915                         fuse_sync_writes(inode);
2916                 }
2917         }
2918
2919         if (!(mode & FALLOC_FL_KEEP_SIZE))
2920                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2921
2922         args.in.h.opcode = FUSE_FALLOCATE;
2923         args.in.h.nodeid = ff->nodeid;
2924         args.in.numargs = 1;
2925         args.in.args[0].size = sizeof(inarg);
2926         args.in.args[0].value = &inarg;
2927         err = fuse_simple_request(fc, &args);
2928         if (err == -ENOSYS) {
2929                 fc->no_fallocate = 1;
2930                 err = -EOPNOTSUPP;
2931         }
2932         if (err)
2933                 goto out;
2934
2935         /* we could have extended the file */
2936         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2937                 bool changed = fuse_write_update_size(inode, offset + length);
2938
2939                 if (changed && fc->writeback_cache)
2940                         file_update_time(file);
2941         }
2942
2943         if (mode & FALLOC_FL_PUNCH_HOLE)
2944                 truncate_pagecache_range(inode, offset, offset + length - 1);
2945
2946         fuse_invalidate_attr(inode);
2947
2948 out:
2949         if (!(mode & FALLOC_FL_KEEP_SIZE))
2950                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2951
2952         if (lock_inode)
2953                 mutex_unlock(&inode->i_mutex);
2954
2955         return err;
2956 }
2957
2958 static const struct file_operations fuse_file_operations = {
2959         .llseek         = fuse_file_llseek,
2960         .read           = new_sync_read,
2961         .read_iter      = fuse_file_read_iter,
2962         .write          = new_sync_write,
2963         .write_iter     = fuse_file_write_iter,
2964         .mmap           = fuse_file_mmap,
2965         .open           = fuse_open,
2966         .flush          = fuse_flush,
2967         .release        = fuse_release,
2968         .fsync          = fuse_fsync,
2969         .lock           = fuse_file_lock,
2970         .flock          = fuse_file_flock,
2971         .splice_read    = generic_file_splice_read,
2972         .unlocked_ioctl = fuse_file_ioctl,
2973         .compat_ioctl   = fuse_file_compat_ioctl,
2974         .poll           = fuse_file_poll,
2975         .fallocate      = fuse_file_fallocate,
2976 };
2977
2978 static const struct file_operations fuse_direct_io_file_operations = {
2979         .llseek         = fuse_file_llseek,
2980         .read           = fuse_direct_read,
2981         .write          = fuse_direct_write,
2982         .mmap           = fuse_direct_mmap,
2983         .open           = fuse_open,
2984         .flush          = fuse_flush,
2985         .release        = fuse_release,
2986         .fsync          = fuse_fsync,
2987         .lock           = fuse_file_lock,
2988         .flock          = fuse_file_flock,
2989         .unlocked_ioctl = fuse_file_ioctl,
2990         .compat_ioctl   = fuse_file_compat_ioctl,
2991         .poll           = fuse_file_poll,
2992         .fallocate      = fuse_file_fallocate,
2993         /* no splice_read */
2994 };
2995
2996 static const struct address_space_operations fuse_file_aops  = {
2997         .readpage       = fuse_readpage,
2998         .writepage      = fuse_writepage,
2999         .writepages     = fuse_writepages,
3000         .launder_page   = fuse_launder_page,
3001         .readpages      = fuse_readpages,
3002         .set_page_dirty = __set_page_dirty_nobuffers,
3003         .bmap           = fuse_bmap,
3004         .direct_IO      = fuse_direct_IO,
3005         .write_begin    = fuse_write_begin,
3006         .write_end      = fuse_write_end,
3007 };
3008
3009 void fuse_init_file_inode(struct inode *inode)
3010 {
3011         inode->i_fop = &fuse_file_operations;
3012         inode->i_data.a_ops = &fuse_file_aops;
3013 }