fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20
  21 static const struct file_operations fuse_direct_io_file_operations;
  22
  23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  24                           int opcode, struct fuse_open_out *outargp)
  25 {
  26         struct fuse_open_in inarg;
  27         FUSE_ARGS(args);
  28
  29         memset(&inarg, 0, sizeof(inarg));
  30         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  31         if (!fc->atomic_o_trunc)
  32                 inarg.flags &= ~O_TRUNC;
  33         args.in.h.opcode = opcode;
  34         args.in.h.nodeid = nodeid;
  35         args.in.numargs = 1;
  36         args.in.args[0].size = sizeof(inarg);
  37         args.in.args[0].value = &inarg;
  38         args.out.numargs = 1;
  39         args.out.args[0].size = sizeof(*outargp);
  40         args.out.args[0].value = outargp;
  41
  42         return fuse_simple_request(fc, &args);
  43 }
  44
  45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  46 {
  47         struct fuse_file *ff;
  48
  49         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  50         if (unlikely(!ff))
  51                 return NULL;
  52
  53         ff->fc = fc;
  54         ff->reserved_req = fuse_request_alloc(0);
  55         if (unlikely(!ff->reserved_req)) {
  56                 kfree(ff);
  57                 return NULL;
  58         }
  59
  60         INIT_LIST_HEAD(&ff->write_entry);
  61         atomic_set(&ff->count, 0);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         spin_lock(&fc->lock);
  66         ff->kh = ++fc->khctr;
  67         spin_unlock(&fc->lock);
  68
  69         return ff;
  70 }
  71
  72 void fuse_file_free(struct fuse_file *ff)
  73 {
  74         fuse_request_free(ff->reserved_req);
  75         kfree(ff);
  76 }
  77
  78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  79 {
  80         atomic_inc(&ff->count);
  81         return ff;
  82 }
  83
  84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  85 {
  86         iput(req->misc.release.inode);
  87 }
  88
  89 static void fuse_file_put(struct fuse_file *ff, bool sync)
  90 {
  91         if (atomic_dec_and_test(&ff->count)) {
  92                 struct fuse_req *req = ff->reserved_req;
  93
  94                 if (ff->fc->no_open) {
  95                         /*
  96                          * Drop the release request when client does not
  97                          * implement 'open'
  98                          */
  99                         __clear_bit(FR_BACKGROUND, &req->flags);
 100                         iput(req->misc.release.inode);
 101                         fuse_put_request(ff->fc, req);
 102                 } else if (sync) {
 103                         __clear_bit(FR_BACKGROUND, &req->flags);
 104                         fuse_request_send(ff->fc, req);
 105                         iput(req->misc.release.inode);
 106                         fuse_put_request(ff->fc, req);
 107                 } else {
 108                         req->end = fuse_release_end;
 109                         __set_bit(FR_BACKGROUND, &req->flags);
 110                         fuse_request_send_background(ff->fc, req);
 111                 }
 112                 kfree(ff);
 113         }
 114 }
 115
 116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 117                  bool isdir)
 118 {
 119         struct fuse_file *ff;
 120         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 121
 122         ff = fuse_file_alloc(fc);
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         ff->fh = 0;
 127         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 128         if (!fc->no_open || isdir) {
 129                 struct fuse_open_out outarg;
 130                 int err;
 131
 132                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 133                 if (!err) {
 134                         ff->fh = outarg.fh;
 135                         ff->open_flags = outarg.open_flags;
 136
 137                 } else if (err != -ENOSYS || isdir) {
 138                         fuse_file_free(ff);
 139                         return err;
 140                 } else {
 141                         fc->no_open = 1;
 142                 }
 143         }
 144
 145         if (isdir)
 146                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 147
 148         ff->nodeid = nodeid;
 149         file->private_data = fuse_file_get(ff);
 150
 151         return 0;
 152 }
 153 EXPORT_SYMBOL_GPL(fuse_do_open);
 154
 155 static void fuse_link_write_file(struct file *file)
 156 {
 157         struct inode *inode = file_inode(file);
 158         struct fuse_conn *fc = get_fuse_conn(inode);
 159         struct fuse_inode *fi = get_fuse_inode(inode);
 160         struct fuse_file *ff = file->private_data;
 161         /*
 162          * file may be written through mmap, so chain it onto the
 163          * inodes's write_file list
 164          */
 165         spin_lock(&fc->lock);
 166         if (list_empty(&ff->write_entry))
 167                 list_add(&ff->write_entry, &fi->write_files);
 168         spin_unlock(&fc->lock);
 169 }
 170
 171 void fuse_finish_open(struct inode *inode, struct file *file)
 172 {
 173         struct fuse_file *ff = file->private_data;
 174         struct fuse_conn *fc = get_fuse_conn(inode);
 175
 176         if (ff->open_flags & FOPEN_DIRECT_IO)
 177                 file->f_op = &fuse_direct_io_file_operations;
 178         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 179                 invalidate_inode_pages2(inode->i_mapping);
 180         if (ff->open_flags & FOPEN_NONSEEKABLE)
 181                 nonseekable_open(inode, file);
 182         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 183                 struct fuse_inode *fi = get_fuse_inode(inode);
 184
 185                 spin_lock(&fc->lock);
 186                 fi->attr_version = ++fc->attr_version;
 187                 i_size_write(inode, 0);
 188                 spin_unlock(&fc->lock);
 189                 fuse_invalidate_attr(inode);
 190                 if (fc->writeback_cache)
 191                         file_update_time(file);
 192         }
 193         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 194                 fuse_link_write_file(file);
 195 }
 196
 197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 198 {
 199         struct fuse_conn *fc = get_fuse_conn(inode);
 200         int err;
 201         bool lock_inode = (file->f_flags & O_TRUNC) &&
 202                           fc->atomic_o_trunc &&
 203                           fc->writeback_cache;
 204
 205         err = generic_file_open(inode, file);
 206         if (err)
 207                 return err;
 208
 209         if (lock_inode)
 210                 inode_lock(inode);
 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                 inode_unlock(inode);
 219
 220         return err;
 221 }
 222
 223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 224 {
 225         struct fuse_conn *fc = ff->fc;
 226         struct fuse_req *req = ff->reserved_req;
 227         struct fuse_release_in *inarg = &req->misc.release.in;
 228
 229         spin_lock(&fc->lock);
 230         list_del(&ff->write_entry);
 231         if (!RB_EMPTY_NODE(&ff->polled_node))
 232                 rb_erase(&ff->polled_node, &fc->polled_files);
 233         spin_unlock(&fc->lock);
 234
 235         wake_up_interruptible_all(&ff->poll_wait);
 236
 237         inarg->fh = ff->fh;
 238         inarg->flags = flags;
 239         req->in.h.opcode = opcode;
 240         req->in.h.nodeid = ff->nodeid;
 241         req->in.numargs = 1;
 242         req->in.args[0].size = sizeof(struct fuse_release_in);
 243         req->in.args[0].value = inarg;
 244 }
 245
 246 void fuse_release_common(struct file *file, int opcode)
 247 {
 248         struct fuse_file *ff;
 249         struct fuse_req *req;
 250
 251         ff = file->private_data;
 252         if (unlikely(!ff))
 253                 return;
 254
 255         req = ff->reserved_req;
 256         fuse_prepare_release(ff, file->f_flags, opcode);
 257
 258         if (ff->flock) {
 259                 struct fuse_release_in *inarg = &req->misc.release.in;
 260                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 261                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 262                                                        (fl_owner_t) file);
 263         }
 264         /* Hold inode until release is finished */
 265         req->misc.release.inode = igrab(file_inode(file));
 266
 267         /*
 268          * Normally this will send the RELEASE request, however if
 269          * some asynchronous READ or WRITE requests are outstanding,
 270          * the sending will be delayed.
 271          *
 272          * Make the release synchronous if this is a fuseblk mount,
 273          * synchronous RELEASE is allowed (and desirable) in this case
 274          * because the server can be trusted not to screw up.
 275          */
 276         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 277 }
 278
 279 static int fuse_open(struct inode *inode, struct file *file)
 280 {
 281         return fuse_open_common(inode, file, false);
 282 }
 283
 284 static int fuse_release(struct inode *inode, struct file *file)
 285 {
 286         struct fuse_conn *fc = get_fuse_conn(inode);
 287
 288         /* see fuse_vma_close() for !writeback_cache case */
 289         if (fc->writeback_cache)
 290                 write_inode_now(inode, 1);
 291
 292         fuse_release_common(file, FUSE_RELEASE);
 293
 294         /* return value is ignored by VFS */
 295         return 0;
 296 }
 297
 298 void fuse_sync_release(struct fuse_file *ff, int flags)
 299 {
 300         WARN_ON(atomic_read(&ff->count) > 1);
 301         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 302         __set_bit(FR_FORCE, &ff->reserved_req->flags);
 303         __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
 304         fuse_request_send(ff->fc, ff->reserved_req);
 305         fuse_put_request(ff->fc, ff->reserved_req);
 306         kfree(ff);
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_sync_release);
 309
 310 /*
 311  * Scramble the ID space with XTEA, so that the value of the files_struct
 312  * pointer is not exposed to userspace.
 313  */
 314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 315 {
 316         u32 *k = fc->scramble_key;
 317         u64 v = (unsigned long) id;
 318         u32 v0 = v;
 319         u32 v1 = v >> 32;
 320         u32 sum = 0;
 321         int i;
 322
 323         for (i = 0; i < 32; i++) {
 324                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 325                 sum += 0x9E3779B9;
 326                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 327         }
 328
 329         return (u64) v0 + ((u64) v1 << 32);
 330 }
 331
 332 /*
 333  * Check if any page in a range is under writeback
 334  *
 335  * This is currently done by walking the list of writepage requests
 336  * for the inode, which can be pretty inefficient.
 337  */
 338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 339                                    pgoff_t idx_to)
 340 {
 341         struct fuse_conn *fc = get_fuse_conn(inode);
 342         struct fuse_inode *fi = get_fuse_inode(inode);
 343         struct fuse_req *req;
 344         bool found = false;
 345
 346         spin_lock(&fc->lock);
 347         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 348                 pgoff_t curr_index;
 349
 350                 BUG_ON(req->inode != inode);
 351                 curr_index = req->misc.write.in.offset >> PAGE_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         inode_lock(inode);
 417         fuse_sync_writes(inode);
 418         inode_unlock(inode);
 419
 420         err = filemap_check_errors(file->f_mapping);
 421         if (err)
 422                 return err;
 423
 424         req = fuse_get_req_nofail_nopages(fc, file);
 425         memset(&inarg, 0, sizeof(inarg));
 426         inarg.fh = ff->fh;
 427         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 428         req->in.h.opcode = FUSE_FLUSH;
 429         req->in.h.nodeid = get_node_id(inode);
 430         req->in.numargs = 1;
 431         req->in.args[0].size = sizeof(inarg);
 432         req->in.args[0].value = &inarg;
 433         __set_bit(FR_FORCE, &req->flags);
 434         fuse_request_send(fc, req);
 435         err = req->out.h.error;
 436         fuse_put_request(fc, req);
 437         if (err == -ENOSYS) {
 438                 fc->no_flush = 1;
 439                 err = 0;
 440         }
 441         return err;
 442 }
 443
 444 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 445                       int datasync, int isdir)
 446 {
 447         struct inode *inode = file->f_mapping->host;
 448         struct fuse_conn *fc = get_fuse_conn(inode);
 449         struct fuse_file *ff = file->private_data;
 450         FUSE_ARGS(args);
 451         struct fuse_fsync_in inarg;
 452         int err;
 453
 454         if (is_bad_inode(inode))
 455                 return -EIO;
 456
 457         inode_lock(inode);
 458
 459         /*
 460          * Start writeback against all dirty pages of the inode, then
 461          * wait for all outstanding writes, before sending the FSYNC
 462          * request.
 463          */
 464         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 465         if (err)
 466                 goto out;
 467
 468         fuse_sync_writes(inode);
 469
 470         /*
 471          * Due to implementation of fuse writeback
 472          * filemap_write_and_wait_range() does not catch errors.
 473          * We have to do this directly after fuse_sync_writes()
 474          */
 475         err = filemap_check_errors(file->f_mapping);
 476         if (err)
 477                 goto out;
 478
 479         err = sync_inode_metadata(inode, 1);
 480         if (err)
 481                 goto out;
 482
 483         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 484                 goto out;
 485
 486         memset(&inarg, 0, sizeof(inarg));
 487         inarg.fh = ff->fh;
 488         inarg.fsync_flags = datasync ? 1 : 0;
 489         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 490         args.in.h.nodeid = get_node_id(inode);
 491         args.in.numargs = 1;
 492         args.in.args[0].size = sizeof(inarg);
 493         args.in.args[0].value = &inarg;
 494         err = fuse_simple_request(fc, &args);
 495         if (err == -ENOSYS) {
 496                 if (isdir)
 497                         fc->no_fsyncdir = 1;
 498                 else
 499                         fc->no_fsync = 1;
 500                 err = 0;
 501         }
 502 out:
 503         inode_unlock(inode);
 504         return err;
 505 }
 506
 507 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 508                       int datasync)
 509 {
 510         return fuse_fsync_common(file, start, end, datasync, 0);
 511 }
 512
 513 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 514                     size_t count, int opcode)
 515 {
 516         struct fuse_read_in *inarg = &req->misc.read.in;
 517         struct fuse_file *ff = file->private_data;
 518
 519         inarg->fh = ff->fh;
 520         inarg->offset = pos;
 521         inarg->size = count;
 522         inarg->flags = file->f_flags;
 523         req->in.h.opcode = opcode;
 524         req->in.h.nodeid = ff->nodeid;
 525         req->in.numargs = 1;
 526         req->in.args[0].size = sizeof(struct fuse_read_in);
 527         req->in.args[0].value = inarg;
 528         req->out.argvar = 1;
 529         req->out.numargs = 1;
 530         req->out.args[0].size = count;
 531 }
 532
 533 static void fuse_release_user_pages(struct fuse_req *req, int write)
 534 {
 535         unsigned i;
 536
 537         for (i = 0; i < req->num_pages; i++) {
 538                 struct page *page = req->pages[i];
 539                 if (write)
 540                         set_page_dirty_lock(page);
 541                 put_page(page);
 542         }
 543 }
 544
 545 static void fuse_io_release(struct kref *kref)
 546 {
 547         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 548 }
 549
 550 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 551 {
 552         if (io->err)
 553                 return io->err;
 554
 555         if (io->bytes >= 0 && io->write)
 556                 return -EIO;
 557
 558         return io->bytes < 0 ? io->size : io->bytes;
 559 }
 560
 561 /**
 562  * In case of short read, the caller sets 'pos' to the position of
 563  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 564  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 565  *
 566  * An example:
 567  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 568  * both submitted asynchronously. The first of them was ACKed by userspace as
 569  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 570  * second request was ACKed as short, e.g. only 1K was read, resulting in
 571  * pos == 33K.
 572  *
 573  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 574  * will be equal to the length of the longest contiguous fragment of
 575  * transferred data starting from the beginning of IO request.
 576  */
 577 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 578 {
 579         int left;
 580
 581         spin_lock(&io->lock);
 582         if (err)
 583                 io->err = io->err ? : err;
 584         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 585                 io->bytes = pos;
 586
 587         left = --io->reqs;
 588         if (!left && io->blocking)
 589                 complete(io->done);
 590         spin_unlock(&io->lock);
 591
 592         if (!left && !io->blocking) {
 593                 ssize_t res = fuse_get_res_by_io(io);
 594
 595                 if (res >= 0) {
 596                         struct inode *inode = file_inode(io->iocb->ki_filp);
 597                         struct fuse_conn *fc = get_fuse_conn(inode);
 598                         struct fuse_inode *fi = get_fuse_inode(inode);
 599
 600                         spin_lock(&fc->lock);
 601                         fi->attr_version = ++fc->attr_version;
 602                         spin_unlock(&fc->lock);
 603                 }
 604
 605                 io->iocb->ki_complete(io->iocb, res, 0);
 606         }
 607
 608         kref_put(&io->refcnt, fuse_io_release);
 609 }
 610
 611 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 612 {
 613         struct fuse_io_priv *io = req->io;
 614         ssize_t pos = -1;
 615
 616         fuse_release_user_pages(req, !io->write);
 617
 618         if (io->write) {
 619                 if (req->misc.write.in.size != req->misc.write.out.size)
 620                         pos = req->misc.write.in.offset - io->offset +
 621                                 req->misc.write.out.size;
 622         } else {
 623                 if (req->misc.read.in.size != req->out.args[0].size)
 624                         pos = req->misc.read.in.offset - io->offset +
 625                                 req->out.args[0].size;
 626         }
 627
 628         fuse_aio_complete(io, req->out.h.error, pos);
 629 }
 630
 631 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 632                 size_t num_bytes, struct fuse_io_priv *io)
 633 {
 634         spin_lock(&io->lock);
 635         kref_get(&io->refcnt);
 636         io->size += num_bytes;
 637         io->reqs++;
 638         spin_unlock(&io->lock);
 639
 640         req->io = io;
 641         req->end = fuse_aio_complete_req;
 642
 643         __fuse_get_request(req);
 644         fuse_request_send_background(fc, req);
 645
 646         return num_bytes;
 647 }
 648
 649 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 650                              loff_t pos, size_t count, fl_owner_t owner)
 651 {
 652         struct file *file = io->file;
 653         struct fuse_file *ff = file->private_data;
 654         struct fuse_conn *fc = ff->fc;
 655
 656         fuse_read_fill(req, file, pos, count, FUSE_READ);
 657         if (owner != NULL) {
 658                 struct fuse_read_in *inarg = &req->misc.read.in;
 659
 660                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 661                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 662         }
 663
 664         if (io->async)
 665                 return fuse_async_req_send(fc, req, count, io);
 666
 667         fuse_request_send(fc, req);
 668         return req->out.args[0].size;
 669 }
 670
 671 static void fuse_read_update_size(struct inode *inode, loff_t size,
 672                                   u64 attr_ver)
 673 {
 674         struct fuse_conn *fc = get_fuse_conn(inode);
 675         struct fuse_inode *fi = get_fuse_inode(inode);
 676
 677         spin_lock(&fc->lock);
 678         if (attr_ver == fi->attr_version && size < inode->i_size &&
 679             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 680                 fi->attr_version = ++fc->attr_version;
 681                 i_size_write(inode, size);
 682         }
 683         spin_unlock(&fc->lock);
 684 }
 685
 686 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 687                             u64 attr_ver)
 688 {
 689         size_t num_read = req->out.args[0].size;
 690         struct fuse_conn *fc = get_fuse_conn(inode);
 691
 692         if (fc->writeback_cache) {
 693                 /*
 694                  * A hole in a file. Some data after the hole are in page cache,
 695                  * but have not reached the client fs yet. So, the hole is not
 696                  * present there.
 697                  */
 698                 int i;
 699                 int start_idx = num_read >> PAGE_SHIFT;
 700                 size_t off = num_read & (PAGE_SIZE - 1);
 701
 702                 for (i = start_idx; i < req->num_pages; i++) {
 703                         zero_user_segment(req->pages[i], off, PAGE_SIZE);
 704                         off = 0;
 705                 }
 706         } else {
 707                 loff_t pos = page_offset(req->pages[0]) + num_read;
 708                 fuse_read_update_size(inode, pos, attr_ver);
 709         }
 710 }
 711
 712 static int fuse_do_readpage(struct file *file, struct page *page)
 713 {
 714         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
 715         struct inode *inode = page->mapping->host;
 716         struct fuse_conn *fc = get_fuse_conn(inode);
 717         struct fuse_req *req;
 718         size_t num_read;
 719         loff_t pos = page_offset(page);
 720         size_t count = PAGE_SIZE;
 721         u64 attr_ver;
 722         int err;
 723
 724         /*
 725          * Page writeback can extend beyond the lifetime of the
 726          * page-cache page, so make sure we read a properly synced
 727          * page.
 728          */
 729         fuse_wait_on_page_writeback(inode, page->index);
 730
 731         req = fuse_get_req(fc, 1);
 732         if (IS_ERR(req))
 733                 return PTR_ERR(req);
 734
 735         attr_ver = fuse_get_attr_version(fc);
 736
 737         req->out.page_zeroing = 1;
 738         req->out.argpages = 1;
 739         req->num_pages = 1;
 740         req->pages[0] = page;
 741         req->page_descs[0].length = count;
 742         num_read = fuse_send_read(req, &io, pos, count, NULL);
 743         err = req->out.h.error;
 744
 745         if (!err) {
 746                 /*
 747                  * Short read means EOF.  If file size is larger, truncate it
 748                  */
 749                 if (num_read < count)
 750                         fuse_short_read(req, inode, attr_ver);
 751
 752                 SetPageUptodate(page);
 753         }
 754
 755         fuse_put_request(fc, req);
 756
 757         return err;
 758 }
 759
 760 static int fuse_readpage(struct file *file, struct page *page)
 761 {
 762         struct inode *inode = page->mapping->host;
 763         int err;
 764
 765         err = -EIO;
 766         if (is_bad_inode(inode))
 767                 goto out;
 768
 769         err = fuse_do_readpage(file, page);
 770         fuse_invalidate_atime(inode);
 771  out:
 772         unlock_page(page);
 773         return err;
 774 }
 775
 776 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 777 {
 778         int i;
 779         size_t count = req->misc.read.in.size;
 780         size_t num_read = req->out.args[0].size;
 781         struct address_space *mapping = NULL;
 782
 783         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 784                 mapping = req->pages[i]->mapping;
 785
 786         if (mapping) {
 787                 struct inode *inode = mapping->host;
 788
 789                 /*
 790                  * Short read means EOF. If file size is larger, truncate it
 791                  */
 792                 if (!req->out.h.error && num_read < count)
 793                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 794
 795                 fuse_invalidate_atime(inode);
 796         }
 797
 798         for (i = 0; i < req->num_pages; i++) {
 799                 struct page *page = req->pages[i];
 800                 if (!req->out.h.error)
 801                         SetPageUptodate(page);
 802                 else
 803                         SetPageError(page);
 804                 unlock_page(page);
 805                 put_page(page);
 806         }
 807         if (req->ff)
 808                 fuse_file_put(req->ff, false);
 809 }
 810
 811 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 812 {
 813         struct fuse_file *ff = file->private_data;
 814         struct fuse_conn *fc = ff->fc;
 815         loff_t pos = page_offset(req->pages[0]);
 816         size_t count = req->num_pages << PAGE_SHIFT;
 817
 818         req->out.argpages = 1;
 819         req->out.page_zeroing = 1;
 820         req->out.page_replace = 1;
 821         fuse_read_fill(req, file, pos, count, FUSE_READ);
 822         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 823         if (fc->async_read) {
 824                 req->ff = fuse_file_get(ff);
 825                 req->end = fuse_readpages_end;
 826                 fuse_request_send_background(fc, req);
 827         } else {
 828                 fuse_request_send(fc, req);
 829                 fuse_readpages_end(fc, req);
 830                 fuse_put_request(fc, req);
 831         }
 832 }
 833
 834 struct fuse_fill_data {
 835         struct fuse_req *req;
 836         struct file *file;
 837         struct inode *inode;
 838         unsigned nr_pages;
 839 };
 840
 841 static int fuse_readpages_fill(void *_data, struct page *page)
 842 {
 843         struct fuse_fill_data *data = _data;
 844         struct fuse_req *req = data->req;
 845         struct inode *inode = data->inode;
 846         struct fuse_conn *fc = get_fuse_conn(inode);
 847
 848         fuse_wait_on_page_writeback(inode, page->index);
 849
 850         if (req->num_pages &&
 851             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 852              (req->num_pages + 1) * PAGE_SIZE > fc->max_read ||
 853              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 854                 int nr_alloc = min_t(unsigned, data->nr_pages,
 855                                      FUSE_MAX_PAGES_PER_REQ);
 856                 fuse_send_readpages(req, data->file);
 857                 if (fc->async_read)
 858                         req = fuse_get_req_for_background(fc, nr_alloc);
 859                 else
 860                         req = fuse_get_req(fc, nr_alloc);
 861
 862                 data->req = req;
 863                 if (IS_ERR(req)) {
 864                         unlock_page(page);
 865                         return PTR_ERR(req);
 866                 }
 867         }
 868
 869         if (WARN_ON(req->num_pages >= req->max_pages)) {
 870                 fuse_put_request(fc, req);
 871                 return -EIO;
 872         }
 873
 874         get_page(page);
 875         req->pages[req->num_pages] = page;
 876         req->page_descs[req->num_pages].length = PAGE_SIZE;
 877         req->num_pages++;
 878         data->nr_pages--;
 879         return 0;
 880 }
 881
 882 static int fuse_readpages(struct file *file, struct address_space *mapping,
 883                           struct list_head *pages, unsigned nr_pages)
 884 {
 885         struct inode *inode = mapping->host;
 886         struct fuse_conn *fc = get_fuse_conn(inode);
 887         struct fuse_fill_data data;
 888         int err;
 889         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 890
 891         err = -EIO;
 892         if (is_bad_inode(inode))
 893                 goto out;
 894
 895         data.file = file;
 896         data.inode = inode;
 897         if (fc->async_read)
 898                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 899         else
 900                 data.req = fuse_get_req(fc, nr_alloc);
 901         data.nr_pages = nr_pages;
 902         err = PTR_ERR(data.req);
 903         if (IS_ERR(data.req))
 904                 goto out;
 905
 906         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 907         if (!err) {
 908                 if (data.req->num_pages)
 909                         fuse_send_readpages(data.req, file);
 910                 else
 911                         fuse_put_request(fc, data.req);
 912         }
 913 out:
 914         return err;
 915 }
 916
 917 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 918 {
 919         struct inode *inode = iocb->ki_filp->f_mapping->host;
 920         struct fuse_conn *fc = get_fuse_conn(inode);
 921
 922         /*
 923          * In auto invalidate mode, always update attributes on read.
 924          * Otherwise, only update if we attempt to read past EOF (to ensure
 925          * i_size is up to date).
 926          */
 927         if (fc->auto_inval_data ||
 928             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 929                 int err;
 930                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 931                 if (err)
 932                         return err;
 933         }
 934
 935         return generic_file_read_iter(iocb, to);
 936 }
 937
 938 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 939                             loff_t pos, size_t count)
 940 {
 941         struct fuse_write_in *inarg = &req->misc.write.in;
 942         struct fuse_write_out *outarg = &req->misc.write.out;
 943
 944         inarg->fh = ff->fh;
 945         inarg->offset = pos;
 946         inarg->size = count;
 947         req->in.h.opcode = FUSE_WRITE;
 948         req->in.h.nodeid = ff->nodeid;
 949         req->in.numargs = 2;
 950         if (ff->fc->minor < 9)
 951                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 952         else
 953                 req->in.args[0].size = sizeof(struct fuse_write_in);
 954         req->in.args[0].value = inarg;
 955         req->in.args[1].size = count;
 956         req->out.numargs = 1;
 957         req->out.args[0].size = sizeof(struct fuse_write_out);
 958         req->out.args[0].value = outarg;
 959 }
 960
 961 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 962                               loff_t pos, size_t count, fl_owner_t owner)
 963 {
 964         struct file *file = io->file;
 965         struct fuse_file *ff = file->private_data;
 966         struct fuse_conn *fc = ff->fc;
 967         struct fuse_write_in *inarg = &req->misc.write.in;
 968
 969         fuse_write_fill(req, ff, pos, count);
 970         inarg->flags = file->f_flags;
 971         if (owner != NULL) {
 972                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 973                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 974         }
 975
 976         if (io->async)
 977                 return fuse_async_req_send(fc, req, count, io);
 978
 979         fuse_request_send(fc, req);
 980         return req->misc.write.out.size;
 981 }
 982
 983 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 984 {
 985         struct fuse_conn *fc = get_fuse_conn(inode);
 986         struct fuse_inode *fi = get_fuse_inode(inode);
 987         bool ret = false;
 988
 989         spin_lock(&fc->lock);
 990         fi->attr_version = ++fc->attr_version;
 991         if (pos > inode->i_size) {
 992                 i_size_write(inode, pos);
 993                 ret = true;
 994         }
 995         spin_unlock(&fc->lock);
 996
 997         return ret;
 998 }
 999
1000 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1001                                     struct inode *inode, loff_t pos,
1002                                     size_t count)
1003 {
1004         size_t res;
1005         unsigned offset;
1006         unsigned i;
1007         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1008
1009         for (i = 0; i < req->num_pages; i++)
1010                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1011
1012         res = fuse_send_write(req, &io, pos, count, NULL);
1013
1014         offset = req->page_descs[0].offset;
1015         count = res;
1016         for (i = 0; i < req->num_pages; i++) {
1017                 struct page *page = req->pages[i];
1018
1019                 if (!req->out.h.error && !offset && count >= PAGE_SIZE)
1020                         SetPageUptodate(page);
1021
1022                 if (count > PAGE_SIZE - offset)
1023                         count -= PAGE_SIZE - offset;
1024                 else
1025                         count = 0;
1026                 offset = 0;
1027
1028                 unlock_page(page);
1029                 put_page(page);
1030         }
1031
1032         return res;
1033 }
1034
1035 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1036                                struct address_space *mapping,
1037                                struct iov_iter *ii, loff_t pos)
1038 {
1039         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1040         unsigned offset = pos & (PAGE_SIZE - 1);
1041         size_t count = 0;
1042         int err;
1043
1044         req->in.argpages = 1;
1045         req->page_descs[0].offset = offset;
1046
1047         do {
1048                 size_t tmp;
1049                 struct page *page;
1050                 pgoff_t index = pos >> PAGE_SHIFT;
1051                 size_t bytes = min_t(size_t, PAGE_SIZE - offset,
1052                                      iov_iter_count(ii));
1053
1054                 bytes = min_t(size_t, bytes, fc->max_write - count);
1055
1056  again:
1057                 err = -EFAULT;
1058                 if (iov_iter_fault_in_readable(ii, bytes))
1059                         break;
1060
1061                 err = -ENOMEM;
1062                 page = grab_cache_page_write_begin(mapping, index, 0);
1063                 if (!page)
1064                         break;
1065
1066                 if (mapping_writably_mapped(mapping))
1067                         flush_dcache_page(page);
1068
1069                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1070                 flush_dcache_page(page);
1071
1072                 iov_iter_advance(ii, tmp);
1073                 if (!tmp) {
1074                         unlock_page(page);
1075                         put_page(page);
1076                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1077                         goto again;
1078                 }
1079
1080                 err = 0;
1081                 req->pages[req->num_pages] = page;
1082                 req->page_descs[req->num_pages].length = tmp;
1083                 req->num_pages++;
1084
1085                 count += tmp;
1086                 pos += tmp;
1087                 offset += tmp;
1088                 if (offset == PAGE_SIZE)
1089                         offset = 0;
1090
1091                 if (!fc->big_writes)
1092                         break;
1093         } while (iov_iter_count(ii) && count < fc->max_write &&
1094                  req->num_pages < req->max_pages && offset == 0);
1095
1096         return count > 0 ? count : err;
1097 }
1098
1099 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1100 {
1101         return min_t(unsigned,
1102                      ((pos + len - 1) >> PAGE_SHIFT) -
1103                      (pos >> PAGE_SHIFT) + 1,
1104                      FUSE_MAX_PAGES_PER_REQ);
1105 }
1106
1107 static ssize_t fuse_perform_write(struct file *file,
1108                                   struct address_space *mapping,
1109                                   struct iov_iter *ii, loff_t pos)
1110 {
1111         struct inode *inode = mapping->host;
1112         struct fuse_conn *fc = get_fuse_conn(inode);
1113         struct fuse_inode *fi = get_fuse_inode(inode);
1114         int err = 0;
1115         ssize_t res = 0;
1116
1117         if (is_bad_inode(inode))
1118                 return -EIO;
1119
1120         if (inode->i_size < pos + iov_iter_count(ii))
1121                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1122
1123         do {
1124                 struct fuse_req *req;
1125                 ssize_t count;
1126                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1127
1128                 req = fuse_get_req(fc, nr_pages);
1129                 if (IS_ERR(req)) {
1130                         err = PTR_ERR(req);
1131                         break;
1132                 }
1133
1134                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1135                 if (count <= 0) {
1136                         err = count;
1137                 } else {
1138                         size_t num_written;
1139
1140                         num_written = fuse_send_write_pages(req, file, inode,
1141                                                             pos, count);
1142                         err = req->out.h.error;
1143                         if (!err) {
1144                                 res += num_written;
1145                                 pos += num_written;
1146
1147                                 /* break out of the loop on short write */
1148                                 if (num_written != count)
1149                                         err = -EIO;
1150                         }
1151                 }
1152                 fuse_put_request(fc, req);
1153         } while (!err && iov_iter_count(ii));
1154
1155         if (res > 0)
1156                 fuse_write_update_size(inode, pos);
1157
1158         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1159         fuse_invalidate_attr(inode);
1160
1161         return res > 0 ? res : err;
1162 }
1163
1164 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1165 {
1166         struct file *file = iocb->ki_filp;
1167         struct address_space *mapping = file->f_mapping;
1168         ssize_t written = 0;
1169         ssize_t written_buffered = 0;
1170         struct inode *inode = mapping->host;
1171         ssize_t err;
1172         loff_t endbyte = 0;
1173
1174         if (get_fuse_conn(inode)->writeback_cache) {
1175                 /* Update size (EOF optimization) and mode (SUID clearing) */
1176                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1177                 if (err)
1178                         return err;
1179
1180                 return generic_file_write_iter(iocb, from);
1181         }
1182
1183         inode_lock(inode);
1184
1185         /* We can write back this queue in page reclaim */
1186         current->backing_dev_info = inode_to_bdi(inode);
1187
1188         err = generic_write_checks(iocb, from);
1189         if (err <= 0)
1190                 goto out;
1191
1192         err = file_remove_privs(file);
1193         if (err)
1194                 goto out;
1195
1196         err = file_update_time(file);
1197         if (err)
1198                 goto out;
1199
1200         if (iocb->ki_flags & IOCB_DIRECT) {
1201                 loff_t pos = iocb->ki_pos;
1202                 written = generic_file_direct_write(iocb, from);
1203                 if (written < 0 || !iov_iter_count(from))
1204                         goto out;
1205
1206                 pos += written;
1207
1208                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1209                 if (written_buffered < 0) {
1210                         err = written_buffered;
1211                         goto out;
1212                 }
1213                 endbyte = pos + written_buffered - 1;
1214
1215                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1216                                                    endbyte);
1217                 if (err)
1218                         goto out;
1219
1220                 invalidate_mapping_pages(file->f_mapping,
1221                                          pos >> PAGE_SHIFT,
1222                                          endbyte >> PAGE_SHIFT);
1223
1224                 written += written_buffered;
1225                 iocb->ki_pos = pos + written_buffered;
1226         } else {
1227                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1228                 if (written >= 0)
1229                         iocb->ki_pos += written;
1230         }
1231 out:
1232         current->backing_dev_info = NULL;
1233         inode_unlock(inode);
1234
1235         return written ? written : err;
1236 }
1237
1238 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1239                 unsigned index, unsigned nr_pages)
1240 {
1241         int i;
1242
1243         for (i = index; i < index + nr_pages; i++)
1244                 req->page_descs[i].length = PAGE_SIZE -
1245                         req->page_descs[i].offset;
1246 }
1247
1248 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1249 {
1250         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1251 }
1252
1253 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1254                                         size_t max_size)
1255 {
1256         return min(iov_iter_single_seg_count(ii), max_size);
1257 }
1258
1259 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1260                                size_t *nbytesp, int write)
1261 {
1262         size_t nbytes = 0;  /* # bytes already packed in req */
1263         ssize_t ret = 0;
1264
1265         /* Special case for kernel I/O: can copy directly into the buffer */
1266         if (ii->type & ITER_KVEC) {
1267                 unsigned long user_addr = fuse_get_user_addr(ii);
1268                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1269
1270                 if (write)
1271                         req->in.args[1].value = (void *) user_addr;
1272                 else
1273                         req->out.args[0].value = (void *) user_addr;
1274
1275                 iov_iter_advance(ii, frag_size);
1276                 *nbytesp = frag_size;
1277                 return 0;
1278         }
1279
1280         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1281                 unsigned npages;
1282                 size_t start;
1283                 ret = iov_iter_get_pages(ii, &req->pages[req->num_pages],
1284                                         *nbytesp - nbytes,
1285                                         req->max_pages - req->num_pages,
1286                                         &start);
1287                 if (ret < 0)
1288                         break;
1289
1290                 iov_iter_advance(ii, ret);
1291                 nbytes += ret;
1292
1293                 ret += start;
1294                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1295
1296                 req->page_descs[req->num_pages].offset = start;
1297                 fuse_page_descs_length_init(req, req->num_pages, npages);
1298
1299                 req->num_pages += npages;
1300                 req->page_descs[req->num_pages - 1].length -=
1301                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1302         }
1303
1304         if (write)
1305                 req->in.argpages = 1;
1306         else
1307                 req->out.argpages = 1;
1308
1309         *nbytesp = nbytes;
1310
1311         return ret < 0 ? ret : 0;
1312 }
1313
1314 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1315 {
1316         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1317 }
1318
1319 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1320                        loff_t *ppos, int flags)
1321 {
1322         int write = flags & FUSE_DIO_WRITE;
1323         int cuse = flags & FUSE_DIO_CUSE;
1324         struct file *file = io->file;
1325         struct inode *inode = file->f_mapping->host;
1326         struct fuse_file *ff = file->private_data;
1327         struct fuse_conn *fc = ff->fc;
1328         size_t nmax = write ? fc->max_write : fc->max_read;
1329         loff_t pos = *ppos;
1330         size_t count = iov_iter_count(iter);
1331         pgoff_t idx_from = pos >> PAGE_SHIFT;
1332         pgoff_t idx_to = (pos + count - 1) >> PAGE_SHIFT;
1333         ssize_t res = 0;
1334         struct fuse_req *req;
1335         int err = 0;
1336
1337         if (io->async)
1338                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1339         else
1340                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1341         if (IS_ERR(req))
1342                 return PTR_ERR(req);
1343
1344         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1345                 if (!write)
1346                         inode_lock(inode);
1347                 fuse_sync_writes(inode);
1348                 if (!write)
1349                         inode_unlock(inode);
1350         }
1351
1352         while (count) {
1353                 size_t nres;
1354                 fl_owner_t owner = current->files;
1355                 size_t nbytes = min(count, nmax);
1356                 err = fuse_get_user_pages(req, iter, &nbytes, write);
1357                 if (err && !nbytes)
1358                         break;
1359
1360                 if (write)
1361                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1362                 else
1363                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1364
1365                 if (!io->async)
1366                         fuse_release_user_pages(req, !write);
1367                 if (req->out.h.error) {
1368                         err = req->out.h.error;
1369                         break;
1370                 } else if (nres > nbytes) {
1371                         res = 0;
1372                         err = -EIO;
1373                         break;
1374                 }
1375                 count -= nres;
1376                 res += nres;
1377                 pos += nres;
1378                 if (nres != nbytes)
1379                         break;
1380                 if (count) {
1381                         fuse_put_request(fc, req);
1382                         if (io->async)
1383                                 req = fuse_get_req_for_background(fc,
1384                                         fuse_iter_npages(iter));
1385                         else
1386                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1387                         if (IS_ERR(req))
1388                                 break;
1389                 }
1390         }
1391         if (!IS_ERR(req))
1392                 fuse_put_request(fc, req);
1393         if (res > 0)
1394                 *ppos = pos;
1395
1396         return res > 0 ? res : err;
1397 }
1398 EXPORT_SYMBOL_GPL(fuse_direct_io);
1399
1400 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1401                                   struct iov_iter *iter,
1402                                   loff_t *ppos)
1403 {
1404         ssize_t res;
1405         struct file *file = io->file;
1406         struct inode *inode = file_inode(file);
1407
1408         if (is_bad_inode(inode))
1409                 return -EIO;
1410
1411         res = fuse_direct_io(io, iter, ppos, 0);
1412
1413         fuse_invalidate_attr(inode);
1414
1415         return res;
1416 }
1417
1418 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1419 {
1420         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1421         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1422 }
1423
1424 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1425 {
1426         struct file *file = iocb->ki_filp;
1427         struct inode *inode = file_inode(file);
1428         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1429         ssize_t res;
1430
1431         if (is_bad_inode(inode))
1432                 return -EIO;
1433
1434         /* Don't allow parallel writes to the same file */
1435         inode_lock(inode);
1436         res = generic_write_checks(iocb, from);
1437         if (res > 0)
1438                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1439         fuse_invalidate_attr(inode);
1440         if (res > 0)
1441                 fuse_write_update_size(inode, iocb->ki_pos);
1442         inode_unlock(inode);
1443
1444         return res;
1445 }
1446
1447 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1448 {
1449         int i;
1450
1451         for (i = 0; i < req->num_pages; i++)
1452                 __free_page(req->pages[i]);
1453
1454         if (req->ff)
1455                 fuse_file_put(req->ff, false);
1456 }
1457
1458 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1459 {
1460         struct inode *inode = req->inode;
1461         struct fuse_inode *fi = get_fuse_inode(inode);
1462         struct backing_dev_info *bdi = inode_to_bdi(inode);
1463         int i;
1464
1465         list_del(&req->writepages_entry);
1466         for (i = 0; i < req->num_pages; i++) {
1467                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1468                 dec_node_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1469                 wb_writeout_inc(&bdi->wb);
1470         }
1471         wake_up(&fi->page_waitq);
1472 }
1473
1474 /* Called under fc->lock, may release and reacquire it */
1475 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1476                                 loff_t size)
1477 __releases(fc->lock)
1478 __acquires(fc->lock)
1479 {
1480         struct fuse_inode *fi = get_fuse_inode(req->inode);
1481         struct fuse_write_in *inarg = &req->misc.write.in;
1482         __u64 data_size = req->num_pages * PAGE_SIZE;
1483
1484         if (!fc->connected)
1485                 goto out_free;
1486
1487         if (inarg->offset + data_size <= size) {
1488                 inarg->size = data_size;
1489         } else if (inarg->offset < size) {
1490                 inarg->size = size - inarg->offset;
1491         } else {
1492                 /* Got truncated off completely */
1493                 goto out_free;
1494         }
1495
1496         req->in.args[1].size = inarg->size;
1497         fi->writectr++;
1498         fuse_request_send_background_locked(fc, req);
1499         return;
1500
1501  out_free:
1502         fuse_writepage_finish(fc, req);
1503         spin_unlock(&fc->lock);
1504         fuse_writepage_free(fc, req);
1505         fuse_put_request(fc, req);
1506         spin_lock(&fc->lock);
1507 }
1508
1509 /*
1510  * If fi->writectr is positive (no truncate or fsync going on) send
1511  * all queued writepage requests.
1512  *
1513  * Called with fc->lock
1514  */
1515 void fuse_flush_writepages(struct inode *inode)
1516 __releases(fc->lock)
1517 __acquires(fc->lock)
1518 {
1519         struct fuse_conn *fc = get_fuse_conn(inode);
1520         struct fuse_inode *fi = get_fuse_inode(inode);
1521         size_t crop = i_size_read(inode);
1522         struct fuse_req *req;
1523
1524         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1525                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1526                 list_del_init(&req->list);
1527                 fuse_send_writepage(fc, req, crop);
1528         }
1529 }
1530
1531 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1532 {
1533         struct inode *inode = req->inode;
1534         struct fuse_inode *fi = get_fuse_inode(inode);
1535
1536         mapping_set_error(inode->i_mapping, req->out.h.error);
1537         spin_lock(&fc->lock);
1538         while (req->misc.write.next) {
1539                 struct fuse_conn *fc = get_fuse_conn(inode);
1540                 struct fuse_write_in *inarg = &req->misc.write.in;
1541                 struct fuse_req *next = req->misc.write.next;
1542                 req->misc.write.next = next->misc.write.next;
1543                 next->misc.write.next = NULL;
1544                 next->ff = fuse_file_get(req->ff);
1545                 list_add(&next->writepages_entry, &fi->writepages);
1546
1547                 /*
1548                  * Skip fuse_flush_writepages() to make it easy to crop requests
1549                  * based on primary request size.
1550                  *
1551                  * 1st case (trivial): there are no concurrent activities using
1552                  * fuse_set/release_nowrite.  Then we're on safe side because
1553                  * fuse_flush_writepages() would call fuse_send_writepage()
1554                  * anyway.
1555                  *
1556                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1557                  * now for completion of all in-flight requests.  This happens
1558                  * rarely and no more than once per page, so this should be
1559                  * okay.
1560                  *
1561                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1562                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1563                  * that fuse_set_nowrite returned implies that all in-flight
1564                  * requests were completed along with all of their secondary
1565                  * requests.  Further primary requests are blocked by negative
1566                  * writectr.  Hence there cannot be any in-flight requests and
1567                  * no invocations of fuse_writepage_end() while we're in
1568                  * fuse_set_nowrite..fuse_release_nowrite section.
1569                  */
1570                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1571         }
1572         fi->writectr--;
1573         fuse_writepage_finish(fc, req);
1574         spin_unlock(&fc->lock);
1575         fuse_writepage_free(fc, req);
1576 }
1577
1578 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1579                                                struct fuse_inode *fi)
1580 {
1581         struct fuse_file *ff = NULL;
1582
1583         spin_lock(&fc->lock);
1584         if (!list_empty(&fi->write_files)) {
1585                 ff = list_entry(fi->write_files.next, struct fuse_file,
1586                                 write_entry);
1587                 fuse_file_get(ff);
1588         }
1589         spin_unlock(&fc->lock);
1590
1591         return ff;
1592 }
1593
1594 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1595                                              struct fuse_inode *fi)
1596 {
1597         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1598         WARN_ON(!ff);
1599         return ff;
1600 }
1601
1602 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1603 {
1604         struct fuse_conn *fc = get_fuse_conn(inode);
1605         struct fuse_inode *fi = get_fuse_inode(inode);
1606         struct fuse_file *ff;
1607         int err;
1608
1609         ff = __fuse_write_file_get(fc, fi);
1610         err = fuse_flush_times(inode, ff);
1611         if (ff)
1612                 fuse_file_put(ff, 0);
1613
1614         return err;
1615 }
1616
1617 static int fuse_writepage_locked(struct page *page)
1618 {
1619         struct address_space *mapping = page->mapping;
1620         struct inode *inode = mapping->host;
1621         struct fuse_conn *fc = get_fuse_conn(inode);
1622         struct fuse_inode *fi = get_fuse_inode(inode);
1623         struct fuse_req *req;
1624         struct page *tmp_page;
1625         int error = -ENOMEM;
1626
1627         set_page_writeback(page);
1628
1629         req = fuse_request_alloc_nofs(1);
1630         if (!req)
1631                 goto err;
1632
1633         /* writeback always goes to bg_queue */
1634         __set_bit(FR_BACKGROUND, &req->flags);
1635         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1636         if (!tmp_page)
1637                 goto err_free;
1638
1639         error = -EIO;
1640         req->ff = fuse_write_file_get(fc, fi);
1641         if (!req->ff)
1642                 goto err_nofile;
1643
1644         fuse_write_fill(req, req->ff, page_offset(page), 0);
1645
1646         copy_highpage(tmp_page, page);
1647         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1648         req->misc.write.next = NULL;
1649         req->in.argpages = 1;
1650         req->num_pages = 1;
1651         req->pages[0] = tmp_page;
1652         req->page_descs[0].offset = 0;
1653         req->page_descs[0].length = PAGE_SIZE;
1654         req->end = fuse_writepage_end;
1655         req->inode = inode;
1656
1657         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1658         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1659
1660         spin_lock(&fc->lock);
1661         list_add(&req->writepages_entry, &fi->writepages);
1662         list_add_tail(&req->list, &fi->queued_writes);
1663         fuse_flush_writepages(inode);
1664         spin_unlock(&fc->lock);
1665
1666         end_page_writeback(page);
1667
1668         return 0;
1669
1670 err_nofile:
1671         __free_page(tmp_page);
1672 err_free:
1673         fuse_request_free(req);
1674 err:
1675         end_page_writeback(page);
1676         return error;
1677 }
1678
1679 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1680 {
1681         int err;
1682
1683         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1684                 /*
1685                  * ->writepages() should be called for sync() and friends.  We
1686                  * should only get here on direct reclaim and then we are
1687                  * allowed to skip a page which is already in flight
1688                  */
1689                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1690
1691                 redirty_page_for_writepage(wbc, page);
1692                 return 0;
1693         }
1694
1695         err = fuse_writepage_locked(page);
1696         unlock_page(page);
1697
1698         return err;
1699 }
1700
1701 struct fuse_fill_wb_data {
1702         struct fuse_req *req;
1703         struct fuse_file *ff;
1704         struct inode *inode;
1705         struct page **orig_pages;
1706 };
1707
1708 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1709 {
1710         struct fuse_req *req = data->req;
1711         struct inode *inode = data->inode;
1712         struct fuse_conn *fc = get_fuse_conn(inode);
1713         struct fuse_inode *fi = get_fuse_inode(inode);
1714         int num_pages = req->num_pages;
1715         int i;
1716
1717         req->ff = fuse_file_get(data->ff);
1718         spin_lock(&fc->lock);
1719         list_add_tail(&req->list, &fi->queued_writes);
1720         fuse_flush_writepages(inode);
1721         spin_unlock(&fc->lock);
1722
1723         for (i = 0; i < num_pages; i++)
1724                 end_page_writeback(data->orig_pages[i]);
1725 }
1726
1727 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1728                                      struct page *page)
1729 {
1730         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1731         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1732         struct fuse_req *tmp;
1733         struct fuse_req *old_req;
1734         bool found = false;
1735         pgoff_t curr_index;
1736
1737         BUG_ON(new_req->num_pages != 0);
1738
1739         spin_lock(&fc->lock);
1740         list_del(&new_req->writepages_entry);
1741         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1742                 BUG_ON(old_req->inode != new_req->inode);
1743                 curr_index = old_req->misc.write.in.offset >> PAGE_SHIFT;
1744                 if (curr_index <= page->index &&
1745                     page->index < curr_index + old_req->num_pages) {
1746                         found = true;
1747                         break;
1748                 }
1749         }
1750         if (!found) {
1751                 list_add(&new_req->writepages_entry, &fi->writepages);
1752                 goto out_unlock;
1753         }
1754
1755         new_req->num_pages = 1;
1756         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1757                 BUG_ON(tmp->inode != new_req->inode);
1758                 curr_index = tmp->misc.write.in.offset >> PAGE_SHIFT;
1759                 if (tmp->num_pages == 1 &&
1760                     curr_index == page->index) {
1761                         old_req = tmp;
1762                 }
1763         }
1764
1765         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1766                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1767
1768                 copy_highpage(old_req->pages[0], page);
1769                 spin_unlock(&fc->lock);
1770
1771                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1772                 dec_node_page_state(page, NR_WRITEBACK_TEMP);
1773                 wb_writeout_inc(&bdi->wb);
1774                 fuse_writepage_free(fc, new_req);
1775                 fuse_request_free(new_req);
1776                 goto out;
1777         } else {
1778                 new_req->misc.write.next = old_req->misc.write.next;
1779                 old_req->misc.write.next = new_req;
1780         }
1781 out_unlock:
1782         spin_unlock(&fc->lock);
1783 out:
1784         return found;
1785 }
1786
1787 static int fuse_writepages_fill(struct page *page,
1788                 struct writeback_control *wbc, void *_data)
1789 {
1790         struct fuse_fill_wb_data *data = _data;
1791         struct fuse_req *req = data->req;
1792         struct inode *inode = data->inode;
1793         struct fuse_conn *fc = get_fuse_conn(inode);
1794         struct page *tmp_page;
1795         bool is_writeback;
1796         int err;
1797
1798         if (!data->ff) {
1799                 err = -EIO;
1800                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1801                 if (!data->ff)
1802                         goto out_unlock;
1803         }
1804
1805         /*
1806          * Being under writeback is unlikely but possible.  For example direct
1807          * read to an mmaped fuse file will set the page dirty twice; once when
1808          * the pages are faulted with get_user_pages(), and then after the read
1809          * completed.
1810          */
1811         is_writeback = fuse_page_is_writeback(inode, page->index);
1812
1813         if (req && req->num_pages &&
1814             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1815              (req->num_pages + 1) * PAGE_SIZE > fc->max_write ||
1816              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1817                 fuse_writepages_send(data);
1818                 data->req = NULL;
1819         }
1820         err = -ENOMEM;
1821         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1822         if (!tmp_page)
1823                 goto out_unlock;
1824
1825         /*
1826          * The page must not be redirtied until the writeout is completed
1827          * (i.e. userspace has sent a reply to the write request).  Otherwise
1828          * there could be more than one temporary page instance for each real
1829          * page.
1830          *
1831          * This is ensured by holding the page lock in page_mkwrite() while
1832          * checking fuse_page_is_writeback().  We already hold the page lock
1833          * since clear_page_dirty_for_io() and keep it held until we add the
1834          * request to the fi->writepages list and increment req->num_pages.
1835          * After this fuse_page_is_writeback() will indicate that the page is
1836          * under writeback, so we can release the page lock.
1837          */
1838         if (data->req == NULL) {
1839                 struct fuse_inode *fi = get_fuse_inode(inode);
1840
1841                 err = -ENOMEM;
1842                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1843                 if (!req) {
1844                         __free_page(tmp_page);
1845                         goto out_unlock;
1846                 }
1847
1848                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1849                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1850                 req->misc.write.next = NULL;
1851                 req->in.argpages = 1;
1852                 __set_bit(FR_BACKGROUND, &req->flags);
1853                 req->num_pages = 0;
1854                 req->end = fuse_writepage_end;
1855                 req->inode = inode;
1856
1857                 spin_lock(&fc->lock);
1858                 list_add(&req->writepages_entry, &fi->writepages);
1859                 spin_unlock(&fc->lock);
1860
1861                 data->req = req;
1862         }
1863         set_page_writeback(page);
1864
1865         copy_highpage(tmp_page, page);
1866         req->pages[req->num_pages] = tmp_page;
1867         req->page_descs[req->num_pages].offset = 0;
1868         req->page_descs[req->num_pages].length = PAGE_SIZE;
1869
1870         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1871         inc_node_page_state(tmp_page, NR_WRITEBACK_TEMP);
1872
1873         err = 0;
1874         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1875                 end_page_writeback(page);
1876                 data->req = NULL;
1877                 goto out_unlock;
1878         }
1879         data->orig_pages[req->num_pages] = page;
1880
1881         /*
1882          * Protected by fc->lock against concurrent access by
1883          * fuse_page_is_writeback().
1884          */
1885         spin_lock(&fc->lock);
1886         req->num_pages++;
1887         spin_unlock(&fc->lock);
1888
1889 out_unlock:
1890         unlock_page(page);
1891
1892         return err;
1893 }
1894
1895 static int fuse_writepages(struct address_space *mapping,
1896                            struct writeback_control *wbc)
1897 {
1898         struct inode *inode = mapping->host;
1899         struct fuse_fill_wb_data data;
1900         int err;
1901
1902         err = -EIO;
1903         if (is_bad_inode(inode))
1904                 goto out;
1905
1906         data.inode = inode;
1907         data.req = NULL;
1908         data.ff = NULL;
1909
1910         err = -ENOMEM;
1911         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1912                                   sizeof(struct page *),
1913                                   GFP_NOFS);
1914         if (!data.orig_pages)
1915                 goto out;
1916
1917         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1918         if (data.req) {
1919                 /* Ignore errors if we can write at least one page */
1920                 BUG_ON(!data.req->num_pages);
1921                 fuse_writepages_send(&data);
1922                 err = 0;
1923         }
1924         if (data.ff)
1925                 fuse_file_put(data.ff, false);
1926
1927         kfree(data.orig_pages);
1928 out:
1929         return err;
1930 }
1931
1932 /*
1933  * It's worthy to make sure that space is reserved on disk for the write,
1934  * but how to implement it without killing performance need more thinking.
1935  */
1936 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1937                 loff_t pos, unsigned len, unsigned flags,
1938                 struct page **pagep, void **fsdata)
1939 {
1940         pgoff_t index = pos >> PAGE_SHIFT;
1941         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1942         struct page *page;
1943         loff_t fsize;
1944         int err = -ENOMEM;
1945
1946         WARN_ON(!fc->writeback_cache);
1947
1948         page = grab_cache_page_write_begin(mapping, index, flags);
1949         if (!page)
1950                 goto error;
1951
1952         fuse_wait_on_page_writeback(mapping->host, page->index);
1953
1954         if (PageUptodate(page) || len == PAGE_SIZE)
1955                 goto success;
1956         /*
1957          * Check if the start this page comes after the end of file, in which
1958          * case the readpage can be optimized away.
1959          */
1960         fsize = i_size_read(mapping->host);
1961         if (fsize <= (pos & PAGE_MASK)) {
1962                 size_t off = pos & ~PAGE_MASK;
1963                 if (off)
1964                         zero_user_segment(page, 0, off);
1965                 goto success;
1966         }
1967         err = fuse_do_readpage(file, page);
1968         if (err)
1969                 goto cleanup;
1970 success:
1971         *pagep = page;
1972         return 0;
1973
1974 cleanup:
1975         unlock_page(page);
1976         put_page(page);
1977 error:
1978         return err;
1979 }
1980
1981 static int fuse_write_end(struct file *file, struct address_space *mapping,
1982                 loff_t pos, unsigned len, unsigned copied,
1983                 struct page *page, void *fsdata)
1984 {
1985         struct inode *inode = page->mapping->host;
1986
1987         if (!PageUptodate(page)) {
1988                 /* Zero any unwritten bytes at the end of the page */
1989                 size_t endoff = (pos + copied) & ~PAGE_MASK;
1990                 if (endoff)
1991                         zero_user_segment(page, endoff, PAGE_SIZE);
1992                 SetPageUptodate(page);
1993         }
1994
1995         fuse_write_update_size(inode, pos + copied);
1996         set_page_dirty(page);
1997         unlock_page(page);
1998         put_page(page);
1999
2000         return copied;
2001 }
2002
2003 static int fuse_launder_page(struct page *page)
2004 {
2005         int err = 0;
2006         if (clear_page_dirty_for_io(page)) {
2007                 struct inode *inode = page->mapping->host;
2008                 err = fuse_writepage_locked(page);
2009                 if (!err)
2010                         fuse_wait_on_page_writeback(inode, page->index);
2011         }
2012         return err;
2013 }
2014
2015 /*
2016  * Write back dirty pages now, because there may not be any suitable
2017  * open files later
2018  */
2019 static void fuse_vma_close(struct vm_area_struct *vma)
2020 {
2021         filemap_write_and_wait(vma->vm_file->f_mapping);
2022 }
2023
2024 /*
2025  * Wait for writeback against this page to complete before allowing it
2026  * to be marked dirty again, and hence written back again, possibly
2027  * before the previous writepage completed.
2028  *
2029  * Block here, instead of in ->writepage(), so that the userspace fs
2030  * can only block processes actually operating on the filesystem.
2031  *
2032  * Otherwise unprivileged userspace fs would be able to block
2033  * unrelated:
2034  *
2035  * - page migration
2036  * - sync(2)
2037  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2038  */
2039 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2040 {
2041         struct page *page = vmf->page;
2042         struct inode *inode = file_inode(vma->vm_file);
2043
2044         file_update_time(vma->vm_file);
2045         lock_page(page);
2046         if (page->mapping != inode->i_mapping) {
2047                 unlock_page(page);
2048                 return VM_FAULT_NOPAGE;
2049         }
2050
2051         fuse_wait_on_page_writeback(inode, page->index);
2052         return VM_FAULT_LOCKED;
2053 }
2054
2055 static const struct vm_operations_struct fuse_file_vm_ops = {
2056         .close          = fuse_vma_close,
2057         .fault          = filemap_fault,
2058         .map_pages      = filemap_map_pages,
2059         .page_mkwrite   = fuse_page_mkwrite,
2060 };
2061
2062 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2063 {
2064         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2065                 fuse_link_write_file(file);
2066
2067         file_accessed(file);
2068         vma->vm_ops = &fuse_file_vm_ops;
2069         return 0;
2070 }
2071
2072 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2073 {
2074         /* Can't provide the coherency needed for MAP_SHARED */
2075         if (vma->vm_flags & VM_MAYSHARE)
2076                 return -ENODEV;
2077
2078         invalidate_inode_pages2(file->f_mapping);
2079
2080         return generic_file_mmap(file, vma);
2081 }
2082
2083 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2084                                   struct file_lock *fl)
2085 {
2086         switch (ffl->type) {
2087         case F_UNLCK:
2088                 break;
2089
2090         case F_RDLCK:
2091         case F_WRLCK:
2092                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2093                     ffl->end < ffl->start)
2094                         return -EIO;
2095
2096                 fl->fl_start = ffl->start;
2097                 fl->fl_end = ffl->end;
2098                 fl->fl_pid = ffl->pid;
2099                 break;
2100
2101         default:
2102                 return -EIO;
2103         }
2104         fl->fl_type = ffl->type;
2105         return 0;
2106 }
2107
2108 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2109                          const struct file_lock *fl, int opcode, pid_t pid,
2110                          int flock, struct fuse_lk_in *inarg)
2111 {
2112         struct inode *inode = file_inode(file);
2113         struct fuse_conn *fc = get_fuse_conn(inode);
2114         struct fuse_file *ff = file->private_data;
2115
2116         memset(inarg, 0, sizeof(*inarg));
2117         inarg->fh = ff->fh;
2118         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2119         inarg->lk.start = fl->fl_start;
2120         inarg->lk.end = fl->fl_end;
2121         inarg->lk.type = fl->fl_type;
2122         inarg->lk.pid = pid;
2123         if (flock)
2124                 inarg->lk_flags |= FUSE_LK_FLOCK;
2125         args->in.h.opcode = opcode;
2126         args->in.h.nodeid = get_node_id(inode);
2127         args->in.numargs = 1;
2128         args->in.args[0].size = sizeof(*inarg);
2129         args->in.args[0].value = inarg;
2130 }
2131
2132 static int fuse_getlk(struct file *file, struct file_lock *fl)
2133 {
2134         struct inode *inode = file_inode(file);
2135         struct fuse_conn *fc = get_fuse_conn(inode);
2136         FUSE_ARGS(args);
2137         struct fuse_lk_in inarg;
2138         struct fuse_lk_out outarg;
2139         int err;
2140
2141         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2142         args.out.numargs = 1;
2143         args.out.args[0].size = sizeof(outarg);
2144         args.out.args[0].value = &outarg;
2145         err = fuse_simple_request(fc, &args);
2146         if (!err)
2147                 err = convert_fuse_file_lock(&outarg.lk, fl);
2148
2149         return err;
2150 }
2151
2152 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2153 {
2154         struct inode *inode = file_inode(file);
2155         struct fuse_conn *fc = get_fuse_conn(inode);
2156         FUSE_ARGS(args);
2157         struct fuse_lk_in inarg;
2158         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2159         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2160         int err;
2161
2162         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2163                 /* NLM needs asynchronous locks, which we don't support yet */
2164                 return -ENOLCK;
2165         }
2166
2167         /* Unlock on close is handled by the flush method */
2168         if (fl->fl_flags & FL_CLOSE)
2169                 return 0;
2170
2171         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2172         err = fuse_simple_request(fc, &args);
2173
2174         /* locking is restartable */
2175         if (err == -EINTR)
2176                 err = -ERESTARTSYS;
2177
2178         return err;
2179 }
2180
2181 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2182 {
2183         struct inode *inode = file_inode(file);
2184         struct fuse_conn *fc = get_fuse_conn(inode);
2185         int err;
2186
2187         if (cmd == F_CANCELLK) {
2188                 err = 0;
2189         } else if (cmd == F_GETLK) {
2190                 if (fc->no_lock) {
2191                         posix_test_lock(file, fl);
2192                         err = 0;
2193                 } else
2194                         err = fuse_getlk(file, fl);
2195         } else {
2196                 if (fc->no_lock)
2197                         err = posix_lock_file(file, fl, NULL);
2198                 else
2199                         err = fuse_setlk(file, fl, 0);
2200         }
2201         return err;
2202 }
2203
2204 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2205 {
2206         struct inode *inode = file_inode(file);
2207         struct fuse_conn *fc = get_fuse_conn(inode);
2208         int err;
2209
2210         if (fc->no_flock) {
2211                 err = locks_lock_file_wait(file, fl);
2212         } else {
2213                 struct fuse_file *ff = file->private_data;
2214
2215                 /* emulate flock with POSIX locks */
2216                 ff->flock = true;
2217                 err = fuse_setlk(file, fl, 1);
2218         }
2219
2220         return err;
2221 }
2222
2223 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2224 {
2225         struct inode *inode = mapping->host;
2226         struct fuse_conn *fc = get_fuse_conn(inode);
2227         FUSE_ARGS(args);
2228         struct fuse_bmap_in inarg;
2229         struct fuse_bmap_out outarg;
2230         int err;
2231
2232         if (!inode->i_sb->s_bdev || fc->no_bmap)
2233                 return 0;
2234
2235         memset(&inarg, 0, sizeof(inarg));
2236         inarg.block = block;
2237         inarg.blocksize = inode->i_sb->s_blocksize;
2238         args.in.h.opcode = FUSE_BMAP;
2239         args.in.h.nodeid = get_node_id(inode);
2240         args.in.numargs = 1;
2241         args.in.args[0].size = sizeof(inarg);
2242         args.in.args[0].value = &inarg;
2243         args.out.numargs = 1;
2244         args.out.args[0].size = sizeof(outarg);
2245         args.out.args[0].value = &outarg;
2246         err = fuse_simple_request(fc, &args);
2247         if (err == -ENOSYS)
2248                 fc->no_bmap = 1;
2249
2250         return err ? 0 : outarg.block;
2251 }
2252
2253 static loff_t fuse_lseek(struct file *file, loff_t offset, int whence)
2254 {
2255         struct inode *inode = file->f_mapping->host;
2256         struct fuse_conn *fc = get_fuse_conn(inode);
2257         struct fuse_file *ff = file->private_data;
2258         FUSE_ARGS(args);
2259         struct fuse_lseek_in inarg = {
2260                 .fh = ff->fh,
2261                 .offset = offset,
2262                 .whence = whence
2263         };
2264         struct fuse_lseek_out outarg;
2265         int err;
2266
2267         if (fc->no_lseek)
2268                 goto fallback;
2269
2270         args.in.h.opcode = FUSE_LSEEK;
2271         args.in.h.nodeid = ff->nodeid;
2272         args.in.numargs = 1;
2273         args.in.args[0].size = sizeof(inarg);
2274         args.in.args[0].value = &inarg;
2275         args.out.numargs = 1;
2276         args.out.args[0].size = sizeof(outarg);
2277         args.out.args[0].value = &outarg;
2278         err = fuse_simple_request(fc, &args);
2279         if (err) {
2280                 if (err == -ENOSYS) {
2281                         fc->no_lseek = 1;
2282                         goto fallback;
2283                 }
2284                 return err;
2285         }
2286
2287         return vfs_setpos(file, outarg.offset, inode->i_sb->s_maxbytes);
2288
2289 fallback:
2290         err = fuse_update_attributes(inode, NULL, file, NULL);
2291         if (!err)
2292                 return generic_file_llseek(file, offset, whence);
2293         else
2294                 return err;
2295 }
2296
2297 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2298 {
2299         loff_t retval;
2300         struct inode *inode = file_inode(file);
2301
2302         switch (whence) {
2303         case SEEK_SET:
2304         case SEEK_CUR:
2305                  /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2306                 retval = generic_file_llseek(file, offset, whence);
2307                 break;
2308         case SEEK_END:
2309                 inode_lock(inode);
2310                 retval = fuse_update_attributes(inode, NULL, file, NULL);
2311                 if (!retval)
2312                         retval = generic_file_llseek(file, offset, whence);
2313                 inode_unlock(inode);
2314                 break;
2315         case SEEK_HOLE:
2316         case SEEK_DATA:
2317                 inode_lock(inode);
2318                 retval = fuse_lseek(file, offset, whence);
2319                 inode_unlock(inode);
2320                 break;
2321         default:
2322                 retval = -EINVAL;
2323         }
2324
2325         return retval;
2326 }
2327
2328 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2329                         unsigned int nr_segs, size_t bytes, bool to_user)
2330 {
2331         struct iov_iter ii;
2332         int page_idx = 0;
2333
2334         if (!bytes)
2335                 return 0;
2336
2337         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2338
2339         while (iov_iter_count(&ii)) {
2340                 struct page *page = pages[page_idx++];
2341                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2342                 void *kaddr;
2343
2344                 kaddr = kmap(page);
2345
2346                 while (todo) {
2347                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2348                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2349                         size_t copy = min(todo, iov_len);
2350                         size_t left;
2351
2352                         if (!to_user)
2353                                 left = copy_from_user(kaddr, uaddr, copy);
2354                         else
2355                                 left = copy_to_user(uaddr, kaddr, copy);
2356
2357                         if (unlikely(left))
2358                                 return -EFAULT;
2359
2360                         iov_iter_advance(&ii, copy);
2361                         todo -= copy;
2362                         kaddr += copy;
2363                 }
2364
2365                 kunmap(page);
2366         }
2367
2368         return 0;
2369 }
2370
2371 /*
2372  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2373  * ABI was defined to be 'struct iovec' which is different on 32bit
2374  * and 64bit.  Fortunately we can determine which structure the server
2375  * used from the size of the reply.
2376  */
2377 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2378                                      size_t transferred, unsigned count,
2379                                      bool is_compat)
2380 {
2381 #ifdef CONFIG_COMPAT
2382         if (count * sizeof(struct compat_iovec) == transferred) {
2383                 struct compat_iovec *ciov = src;
2384                 unsigned i;
2385
2386                 /*
2387                  * With this interface a 32bit server cannot support
2388                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2389                  * requests
2390                  */
2391                 if (!is_compat)
2392                         return -EINVAL;
2393
2394                 for (i = 0; i < count; i++) {
2395                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2396                         dst[i].iov_len = ciov[i].iov_len;
2397                 }
2398                 return 0;
2399         }
2400 #endif
2401
2402         if (count * sizeof(struct iovec) != transferred)
2403                 return -EIO;
2404
2405         memcpy(dst, src, transferred);
2406         return 0;
2407 }
2408
2409 /* Make sure iov_length() won't overflow */
2410 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2411 {
2412         size_t n;
2413         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2414
2415         for (n = 0; n < count; n++, iov++) {
2416                 if (iov->iov_len > (size_t) max)
2417                         return -ENOMEM;
2418                 max -= iov->iov_len;
2419         }
2420         return 0;
2421 }
2422
2423 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2424                                  void *src, size_t transferred, unsigned count,
2425                                  bool is_compat)
2426 {
2427         unsigned i;
2428         struct fuse_ioctl_iovec *fiov = src;
2429
2430         if (fc->minor < 16) {
2431                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2432                                                  count, is_compat);
2433         }
2434
2435         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2436                 return -EIO;
2437
2438         for (i = 0; i < count; i++) {
2439                 /* Did the server supply an inappropriate value? */
2440                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2441                     fiov[i].len != (unsigned long) fiov[i].len)
2442                         return -EIO;
2443
2444                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2445                 dst[i].iov_len = (size_t) fiov[i].len;
2446
2447 #ifdef CONFIG_COMPAT
2448                 if (is_compat &&
2449                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2450                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2451                         return -EIO;
2452 #endif
2453         }
2454
2455         return 0;
2456 }
2457
2458
2459 /*
2460  * For ioctls, there is no generic way to determine how much memory
2461  * needs to be read and/or written.  Furthermore, ioctls are allowed
2462  * to dereference the passed pointer, so the parameter requires deep
2463  * copying but FUSE has no idea whatsoever about what to copy in or
2464  * out.
2465  *
2466  * This is solved by allowing FUSE server to retry ioctl with
2467  * necessary in/out iovecs.  Let's assume the ioctl implementation
2468  * needs to read in the following structure.
2469  *
2470  * struct a {
2471  *      char    *buf;
2472  *      size_t  buflen;
2473  * }
2474  *
2475  * On the first callout to FUSE server, inarg->in_size and
2476  * inarg->out_size will be NULL; then, the server completes the ioctl
2477  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2478  * the actual iov array to
2479  *
2480  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2481  *
2482  * which tells FUSE to copy in the requested area and retry the ioctl.
2483  * On the second round, the server has access to the structure and
2484  * from that it can tell what to look for next, so on the invocation,
2485  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2486  *
2487  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2488  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2489  *
2490  * FUSE will copy both struct a and the pointed buffer from the
2491  * process doing the ioctl and retry ioctl with both struct a and the
2492  * buffer.
2493  *
2494  * This time, FUSE server has everything it needs and completes ioctl
2495  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2496  *
2497  * Copying data out works the same way.
2498  *
2499  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2500  * automatically initializes in and out iovs by decoding @cmd with
2501  * _IOC_* macros and the server is not allowed to request RETRY.  This
2502  * limits ioctl data transfers to well-formed ioctls and is the forced
2503  * behavior for all FUSE servers.
2504  */
2505 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2506                    unsigned int flags)
2507 {
2508         struct fuse_file *ff = file->private_data;
2509         struct fuse_conn *fc = ff->fc;
2510         struct fuse_ioctl_in inarg = {
2511                 .fh = ff->fh,
2512                 .cmd = cmd,
2513                 .arg = arg,
2514                 .flags = flags
2515         };
2516         struct fuse_ioctl_out outarg;
2517         struct fuse_req *req = NULL;
2518         struct page **pages = NULL;
2519         struct iovec *iov_page = NULL;
2520         struct iovec *in_iov = NULL, *out_iov = NULL;
2521         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2522         size_t in_size, out_size, transferred;
2523         int err;
2524
2525 #if BITS_PER_LONG == 32
2526         inarg.flags |= FUSE_IOCTL_32BIT;
2527 #else
2528         if (flags & FUSE_IOCTL_COMPAT)
2529                 inarg.flags |= FUSE_IOCTL_32BIT;
2530 #endif
2531
2532         /* assume all the iovs returned by client always fits in a page */
2533         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2534
2535         err = -ENOMEM;
2536         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2537         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2538         if (!pages || !iov_page)
2539                 goto out;
2540
2541         /*
2542          * If restricted, initialize IO parameters as encoded in @cmd.
2543          * RETRY from server is not allowed.
2544          */
2545         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2546                 struct iovec *iov = iov_page;
2547
2548                 iov->iov_base = (void __user *)arg;
2549                 iov->iov_len = _IOC_SIZE(cmd);
2550
2551                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2552                         in_iov = iov;
2553                         in_iovs = 1;
2554                 }
2555
2556                 if (_IOC_DIR(cmd) & _IOC_READ) {
2557                         out_iov = iov;
2558                         out_iovs = 1;
2559                 }
2560         }
2561
2562  retry:
2563         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2564         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2565
2566         /*
2567          * Out data can be used either for actual out data or iovs,
2568          * make sure there always is at least one page.
2569          */
2570         out_size = max_t(size_t, out_size, PAGE_SIZE);
2571         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2572
2573         /* make sure there are enough buffer pages and init request with them */
2574         err = -ENOMEM;
2575         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2576                 goto out;
2577         while (num_pages < max_pages) {
2578                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2579                 if (!pages[num_pages])
2580                         goto out;
2581                 num_pages++;
2582         }
2583
2584         req = fuse_get_req(fc, num_pages);
2585         if (IS_ERR(req)) {
2586                 err = PTR_ERR(req);
2587                 req = NULL;
2588                 goto out;
2589         }
2590         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2591         req->num_pages = num_pages;
2592         fuse_page_descs_length_init(req, 0, req->num_pages);
2593
2594         /* okay, let's send it to the client */
2595         req->in.h.opcode = FUSE_IOCTL;
2596         req->in.h.nodeid = ff->nodeid;
2597         req->in.numargs = 1;
2598         req->in.args[0].size = sizeof(inarg);
2599         req->in.args[0].value = &inarg;
2600         if (in_size) {
2601                 req->in.numargs++;
2602                 req->in.args[1].size = in_size;
2603                 req->in.argpages = 1;
2604
2605                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2606                                            false);
2607                 if (err)
2608                         goto out;
2609         }
2610
2611         req->out.numargs = 2;
2612         req->out.args[0].size = sizeof(outarg);
2613         req->out.args[0].value = &outarg;
2614         req->out.args[1].size = out_size;
2615         req->out.argpages = 1;
2616         req->out.argvar = 1;
2617
2618         fuse_request_send(fc, req);
2619         err = req->out.h.error;
2620         transferred = req->out.args[1].size;
2621         fuse_put_request(fc, req);
2622         req = NULL;
2623         if (err)
2624                 goto out;
2625
2626         /* did it ask for retry? */
2627         if (outarg.flags & FUSE_IOCTL_RETRY) {
2628                 void *vaddr;
2629
2630                 /* no retry if in restricted mode */
2631                 err = -EIO;
2632                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2633                         goto out;
2634
2635                 in_iovs = outarg.in_iovs;
2636                 out_iovs = outarg.out_iovs;
2637
2638                 /*
2639                  * Make sure things are in boundary, separate checks
2640                  * are to protect against overflow.
2641                  */
2642                 err = -ENOMEM;
2643                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2644                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2645                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2646                         goto out;
2647
2648                 vaddr = kmap_atomic(pages[0]);
2649                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2650                                             transferred, in_iovs + out_iovs,
2651                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2652                 kunmap_atomic(vaddr);
2653                 if (err)
2654                         goto out;
2655
2656                 in_iov = iov_page;
2657                 out_iov = in_iov + in_iovs;
2658
2659                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2660                 if (err)
2661                         goto out;
2662
2663                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2664                 if (err)
2665                         goto out;
2666
2667                 goto retry;
2668         }
2669
2670         err = -EIO;
2671         if (transferred > inarg.out_size)
2672                 goto out;
2673
2674         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2675  out:
2676         if (req)
2677                 fuse_put_request(fc, req);
2678         free_page((unsigned long) iov_page);
2679         while (num_pages)
2680                 __free_page(pages[--num_pages]);
2681         kfree(pages);
2682
2683         return err ? err : outarg.result;
2684 }
2685 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2686
2687 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2688                        unsigned long arg, unsigned int flags)
2689 {
2690         struct inode *inode = file_inode(file);
2691         struct fuse_conn *fc = get_fuse_conn(inode);
2692
2693         if (!fuse_allow_current_process(fc))
2694                 return -EACCES;
2695
2696         if (is_bad_inode(inode))
2697                 return -EIO;
2698
2699         return fuse_do_ioctl(file, cmd, arg, flags);
2700 }
2701
2702 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2703                             unsigned long arg)
2704 {
2705         return fuse_ioctl_common(file, cmd, arg, 0);
2706 }
2707
2708 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2709                                    unsigned long arg)
2710 {
2711         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2712 }
2713
2714 /*
2715  * All files which have been polled are linked to RB tree
2716  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2717  * find the matching one.
2718  */
2719 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2720                                               struct rb_node **parent_out)
2721 {
2722         struct rb_node **link = &fc->polled_files.rb_node;
2723         struct rb_node *last = NULL;
2724
2725         while (*link) {
2726                 struct fuse_file *ff;
2727
2728                 last = *link;
2729                 ff = rb_entry(last, struct fuse_file, polled_node);
2730
2731                 if (kh < ff->kh)
2732                         link = &last->rb_left;
2733                 else if (kh > ff->kh)
2734                         link = &last->rb_right;
2735                 else
2736                         return link;
2737         }
2738
2739         if (parent_out)
2740                 *parent_out = last;
2741         return link;
2742 }
2743
2744 /*
2745  * The file is about to be polled.  Make sure it's on the polled_files
2746  * RB tree.  Note that files once added to the polled_files tree are
2747  * not removed before the file is released.  This is because a file
2748  * polled once is likely to be polled again.
2749  */
2750 static void fuse_register_polled_file(struct fuse_conn *fc,
2751                                       struct fuse_file *ff)
2752 {
2753         spin_lock(&fc->lock);
2754         if (RB_EMPTY_NODE(&ff->polled_node)) {
2755                 struct rb_node **link, *uninitialized_var(parent);
2756
2757                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2758                 BUG_ON(*link);
2759                 rb_link_node(&ff->polled_node, parent, link);
2760                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2761         }
2762         spin_unlock(&fc->lock);
2763 }
2764
2765 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2766 {
2767         struct fuse_file *ff = file->private_data;
2768         struct fuse_conn *fc = ff->fc;
2769         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2770         struct fuse_poll_out outarg;
2771         FUSE_ARGS(args);
2772         int err;
2773
2774         if (fc->no_poll)
2775                 return DEFAULT_POLLMASK;
2776
2777         poll_wait(file, &ff->poll_wait, wait);
2778         inarg.events = (__u32)poll_requested_events(wait);
2779
2780         /*
2781          * Ask for notification iff there's someone waiting for it.
2782          * The client may ignore the flag and always notify.
2783          */
2784         if (waitqueue_active(&ff->poll_wait)) {
2785                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2786                 fuse_register_polled_file(fc, ff);
2787         }
2788
2789         args.in.h.opcode = FUSE_POLL;
2790         args.in.h.nodeid = ff->nodeid;
2791         args.in.numargs = 1;
2792         args.in.args[0].size = sizeof(inarg);
2793         args.in.args[0].value = &inarg;
2794         args.out.numargs = 1;
2795         args.out.args[0].size = sizeof(outarg);
2796         args.out.args[0].value = &outarg;
2797         err = fuse_simple_request(fc, &args);
2798
2799         if (!err)
2800                 return outarg.revents;
2801         if (err == -ENOSYS) {
2802                 fc->no_poll = 1;
2803                 return DEFAULT_POLLMASK;
2804         }
2805         return POLLERR;
2806 }
2807 EXPORT_SYMBOL_GPL(fuse_file_poll);
2808
2809 /*
2810  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2811  * wakes up the poll waiters.
2812  */
2813 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2814                             struct fuse_notify_poll_wakeup_out *outarg)
2815 {
2816         u64 kh = outarg->kh;
2817         struct rb_node **link;
2818
2819         spin_lock(&fc->lock);
2820
2821         link = fuse_find_polled_node(fc, kh, NULL);
2822         if (*link) {
2823                 struct fuse_file *ff;
2824
2825                 ff = rb_entry(*link, struct fuse_file, polled_node);
2826                 wake_up_interruptible_sync(&ff->poll_wait);
2827         }
2828
2829         spin_unlock(&fc->lock);
2830         return 0;
2831 }
2832
2833 static void fuse_do_truncate(struct file *file)
2834 {
2835         struct inode *inode = file->f_mapping->host;
2836         struct iattr attr;
2837
2838         attr.ia_valid = ATTR_SIZE;
2839         attr.ia_size = i_size_read(inode);
2840
2841         attr.ia_file = file;
2842         attr.ia_valid |= ATTR_FILE;
2843
2844         fuse_do_setattr(inode, &attr, file);
2845 }
2846
2847 static inline loff_t fuse_round_up(loff_t off)
2848 {
2849         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2850 }
2851
2852 static ssize_t
2853 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2854 {
2855         DECLARE_COMPLETION_ONSTACK(wait);
2856         ssize_t ret = 0;
2857         struct file *file = iocb->ki_filp;
2858         struct fuse_file *ff = file->private_data;
2859         bool async_dio = ff->fc->async_dio;
2860         loff_t pos = 0;
2861         struct inode *inode;
2862         loff_t i_size;
2863         size_t count = iov_iter_count(iter);
2864         loff_t offset = iocb->ki_pos;
2865         struct fuse_io_priv *io;
2866
2867         pos = offset;
2868         inode = file->f_mapping->host;
2869         i_size = i_size_read(inode);
2870
2871         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2872                 return 0;
2873
2874         /* optimization for short read */
2875         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2876                 if (offset >= i_size)
2877                         return 0;
2878                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2879                 count = iov_iter_count(iter);
2880         }
2881
2882         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2883         if (!io)
2884                 return -ENOMEM;
2885         spin_lock_init(&io->lock);
2886         kref_init(&io->refcnt);
2887         io->reqs = 1;
2888         io->bytes = -1;
2889         io->size = 0;
2890         io->offset = offset;
2891         io->write = (iov_iter_rw(iter) == WRITE);
2892         io->err = 0;
2893         io->file = file;
2894         /*
2895          * By default, we want to optimize all I/Os with async request
2896          * submission to the client filesystem if supported.
2897          */
2898         io->async = async_dio;
2899         io->iocb = iocb;
2900         io->blocking = is_sync_kiocb(iocb);
2901
2902         /*
2903          * We cannot asynchronously extend the size of a file.
2904          * In such case the aio will behave exactly like sync io.
2905          */
2906         if ((offset + count > i_size) && iov_iter_rw(iter) == WRITE)
2907                 io->blocking = true;
2908
2909         if (io->async && io->blocking) {
2910                 /*
2911                  * Additional reference to keep io around after
2912                  * calling fuse_aio_complete()
2913                  */
2914                 kref_get(&io->refcnt);
2915                 io->done = &wait;
2916         }
2917
2918         if (iov_iter_rw(iter) == WRITE) {
2919                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2920                 fuse_invalidate_attr(inode);
2921         } else {
2922                 ret = __fuse_direct_read(io, iter, &pos);
2923         }
2924
2925         if (io->async) {
2926                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2927
2928                 /* we have a non-extending, async request, so return */
2929                 if (!io->blocking)
2930                         return -EIOCBQUEUED;
2931
2932                 wait_for_completion(&wait);
2933                 ret = fuse_get_res_by_io(io);
2934         }
2935
2936         kref_put(&io->refcnt, fuse_io_release);
2937
2938         if (iov_iter_rw(iter) == WRITE) {
2939                 if (ret > 0)
2940                         fuse_write_update_size(inode, pos);
2941                 else if (ret < 0 && offset + count > i_size)
2942                         fuse_do_truncate(file);
2943         }
2944
2945         return ret;
2946 }
2947
2948 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2949                                 loff_t length)
2950 {
2951         struct fuse_file *ff = file->private_data;
2952         struct inode *inode = file_inode(file);
2953         struct fuse_inode *fi = get_fuse_inode(inode);
2954         struct fuse_conn *fc = ff->fc;
2955         FUSE_ARGS(args);
2956         struct fuse_fallocate_in inarg = {
2957                 .fh = ff->fh,
2958                 .offset = offset,
2959                 .length = length,
2960                 .mode = mode
2961         };
2962         int err;
2963         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2964                            (mode & FALLOC_FL_PUNCH_HOLE);
2965
2966         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2967                 return -EOPNOTSUPP;
2968
2969         if (fc->no_fallocate)
2970                 return -EOPNOTSUPP;
2971
2972         if (lock_inode) {
2973                 inode_lock(inode);
2974                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2975                         loff_t endbyte = offset + length - 1;
2976                         err = filemap_write_and_wait_range(inode->i_mapping,
2977                                                            offset, endbyte);
2978                         if (err)
2979                                 goto out;
2980
2981                         fuse_sync_writes(inode);
2982                 }
2983         }
2984
2985         if (!(mode & FALLOC_FL_KEEP_SIZE))
2986                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2987
2988         args.in.h.opcode = FUSE_FALLOCATE;
2989         args.in.h.nodeid = ff->nodeid;
2990         args.in.numargs = 1;
2991         args.in.args[0].size = sizeof(inarg);
2992         args.in.args[0].value = &inarg;
2993         err = fuse_simple_request(fc, &args);
2994         if (err == -ENOSYS) {
2995                 fc->no_fallocate = 1;
2996                 err = -EOPNOTSUPP;
2997         }
2998         if (err)
2999                 goto out;
3000
3001         /* we could have extended the file */
3002         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3003                 bool changed = fuse_write_update_size(inode, offset + length);
3004
3005                 if (changed && fc->writeback_cache)
3006                         file_update_time(file);
3007         }
3008
3009         if (mode & FALLOC_FL_PUNCH_HOLE)
3010                 truncate_pagecache_range(inode, offset, offset + length - 1);
3011
3012         fuse_invalidate_attr(inode);
3013
3014 out:
3015         if (!(mode & FALLOC_FL_KEEP_SIZE))
3016                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3017
3018         if (lock_inode)
3019                 inode_unlock(inode);
3020
3021         return err;
3022 }
3023
3024 static const struct file_operations fuse_file_operations = {
3025         .llseek         = fuse_file_llseek,
3026         .read_iter      = fuse_file_read_iter,
3027         .write_iter     = fuse_file_write_iter,
3028         .mmap           = fuse_file_mmap,
3029         .open           = fuse_open,
3030         .flush          = fuse_flush,
3031         .release        = fuse_release,
3032         .fsync          = fuse_fsync,
3033         .lock           = fuse_file_lock,
3034         .flock          = fuse_file_flock,
3035         .splice_read    = generic_file_splice_read,
3036         .unlocked_ioctl = fuse_file_ioctl,
3037         .compat_ioctl   = fuse_file_compat_ioctl,
3038         .poll           = fuse_file_poll,
3039         .fallocate      = fuse_file_fallocate,
3040 };
3041
3042 static const struct file_operations fuse_direct_io_file_operations = {
3043         .llseek         = fuse_file_llseek,
3044         .read_iter      = fuse_direct_read_iter,
3045         .write_iter     = fuse_direct_write_iter,
3046         .mmap           = fuse_direct_mmap,
3047         .open           = fuse_open,
3048         .flush          = fuse_flush,
3049         .release        = fuse_release,
3050         .fsync          = fuse_fsync,
3051         .lock           = fuse_file_lock,
3052         .flock          = fuse_file_flock,
3053         .unlocked_ioctl = fuse_file_ioctl,
3054         .compat_ioctl   = fuse_file_compat_ioctl,
3055         .poll           = fuse_file_poll,
3056         .fallocate      = fuse_file_fallocate,
3057         /* no splice_read */
3058 };
3059
3060 static const struct address_space_operations fuse_file_aops  = {
3061         .readpage       = fuse_readpage,
3062         .writepage      = fuse_writepage,
3063         .writepages     = fuse_writepages,
3064         .launder_page   = fuse_launder_page,
3065         .readpages      = fuse_readpages,
3066         .set_page_dirty = __set_page_dirty_nobuffers,
3067         .bmap           = fuse_bmap,
3068         .direct_IO      = fuse_direct_IO,
3069         .write_begin    = fuse_write_begin,
3070         .write_end      = fuse_write_end,
3071 };
3072
3073 void fuse_init_file_inode(struct inode *inode)
3074 {
3075         inode->i_fop = &fuse_file_operations;
3076         inode->i_data.a_ops = &fuse_file_aops;
3077 }