fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/falloc.h>
  19 #include <linux/uio.h>
  20
  21 static const struct file_operations fuse_direct_io_file_operations;
  22
  23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  24                           int opcode, struct fuse_open_out *outargp)
  25 {
  26         struct fuse_open_in inarg;
  27         FUSE_ARGS(args);
  28
  29         memset(&inarg, 0, sizeof(inarg));
  30         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  31         if (!fc->atomic_o_trunc)
  32                 inarg.flags &= ~O_TRUNC;
  33         args.in.h.opcode = opcode;
  34         args.in.h.nodeid = nodeid;
  35         args.in.numargs = 1;
  36         args.in.args[0].size = sizeof(inarg);
  37         args.in.args[0].value = &inarg;
  38         args.out.numargs = 1;
  39         args.out.args[0].size = sizeof(*outargp);
  40         args.out.args[0].value = outargp;
  41
  42         return fuse_simple_request(fc, &args);
  43 }
  44
  45 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  46 {
  47         struct fuse_file *ff;
  48
  49         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  50         if (unlikely(!ff))
  51                 return NULL;
  52
  53         ff->fc = fc;
  54         ff->reserved_req = fuse_request_alloc(0);
  55         if (unlikely(!ff->reserved_req)) {
  56                 kfree(ff);
  57                 return NULL;
  58         }
  59
  60         INIT_LIST_HEAD(&ff->write_entry);
  61         atomic_set(&ff->count, 0);
  62         RB_CLEAR_NODE(&ff->polled_node);
  63         init_waitqueue_head(&ff->poll_wait);
  64
  65         spin_lock(&fc->lock);
  66         ff->kh = ++fc->khctr;
  67         spin_unlock(&fc->lock);
  68
  69         return ff;
  70 }
  71
  72 void fuse_file_free(struct fuse_file *ff)
  73 {
  74         fuse_request_free(ff->reserved_req);
  75         kfree(ff);
  76 }
  77
  78 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  79 {
  80         atomic_inc(&ff->count);
  81         return ff;
  82 }
  83
  84 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  85 {
  86         iput(req->misc.release.inode);
  87 }
  88
  89 static void fuse_file_put(struct fuse_file *ff, bool sync)
  90 {
  91         if (atomic_dec_and_test(&ff->count)) {
  92                 struct fuse_req *req = ff->reserved_req;
  93
  94                 if (ff->fc->no_open) {
  95                         /*
  96                          * Drop the release request when client does not
  97                          * implement 'open'
  98                          */
  99                         __clear_bit(FR_BACKGROUND, &req->flags);
 100                         iput(req->misc.release.inode);
 101                         fuse_put_request(ff->fc, req);
 102                 } else if (sync) {
 103                         __clear_bit(FR_BACKGROUND, &req->flags);
 104                         fuse_request_send(ff->fc, req);
 105                         iput(req->misc.release.inode);
 106                         fuse_put_request(ff->fc, req);
 107                 } else {
 108                         req->end = fuse_release_end;
 109                         __set_bit(FR_BACKGROUND, &req->flags);
 110                         fuse_request_send_background(ff->fc, req);
 111                 }
 112                 kfree(ff);
 113         }
 114 }
 115
 116 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 117                  bool isdir)
 118 {
 119         struct fuse_file *ff;
 120         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 121
 122         ff = fuse_file_alloc(fc);
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         ff->fh = 0;
 127         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 128         if (!fc->no_open || isdir) {
 129                 struct fuse_open_out outarg;
 130                 int err;
 131
 132                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 133                 if (!err) {
 134                         ff->fh = outarg.fh;
 135                         ff->open_flags = outarg.open_flags;
 136
 137                 } else if (err != -ENOSYS || isdir) {
 138                         fuse_file_free(ff);
 139                         return err;
 140                 } else {
 141                         fc->no_open = 1;
 142                 }
 143         }
 144
 145         if (isdir)
 146                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 147
 148         ff->nodeid = nodeid;
 149         file->private_data = fuse_file_get(ff);
 150
 151         return 0;
 152 }
 153 EXPORT_SYMBOL_GPL(fuse_do_open);
 154
 155 static void fuse_link_write_file(struct file *file)
 156 {
 157         struct inode *inode = file_inode(file);
 158         struct fuse_conn *fc = get_fuse_conn(inode);
 159         struct fuse_inode *fi = get_fuse_inode(inode);
 160         struct fuse_file *ff = file->private_data;
 161         /*
 162          * file may be written through mmap, so chain it onto the
 163          * inodes's write_file list
 164          */
 165         spin_lock(&fc->lock);
 166         if (list_empty(&ff->write_entry))
 167                 list_add(&ff->write_entry, &fi->write_files);
 168         spin_unlock(&fc->lock);
 169 }
 170
 171 void fuse_finish_open(struct inode *inode, struct file *file)
 172 {
 173         struct fuse_file *ff = file->private_data;
 174         struct fuse_conn *fc = get_fuse_conn(inode);
 175
 176         if (ff->open_flags & FOPEN_DIRECT_IO)
 177                 file->f_op = &fuse_direct_io_file_operations;
 178         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 179                 invalidate_inode_pages2(inode->i_mapping);
 180         if (ff->open_flags & FOPEN_NONSEEKABLE)
 181                 nonseekable_open(inode, file);
 182         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 183                 struct fuse_inode *fi = get_fuse_inode(inode);
 184
 185                 spin_lock(&fc->lock);
 186                 fi->attr_version = ++fc->attr_version;
 187                 i_size_write(inode, 0);
 188                 spin_unlock(&fc->lock);
 189                 fuse_invalidate_attr(inode);
 190                 if (fc->writeback_cache)
 191                         file_update_time(file);
 192         }
 193         if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 194                 fuse_link_write_file(file);
 195 }
 196
 197 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 198 {
 199         struct fuse_conn *fc = get_fuse_conn(inode);
 200         int err;
 201         bool lock_inode = (file->f_flags & O_TRUNC) &&
 202                           fc->atomic_o_trunc &&
 203                           fc->writeback_cache;
 204
 205         err = generic_file_open(inode, file);
 206         if (err)
 207                 return err;
 208
 209         if (lock_inode)
 210                 mutex_lock(&inode->i_mutex);
 211
 212         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 213
 214         if (!err)
 215                 fuse_finish_open(inode, file);
 216
 217         if (lock_inode)
 218                 mutex_unlock(&inode->i_mutex);
 219
 220         return err;
 221 }
 222
 223 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 224 {
 225         struct fuse_conn *fc = ff->fc;
 226         struct fuse_req *req = ff->reserved_req;
 227         struct fuse_release_in *inarg = &req->misc.release.in;
 228
 229         spin_lock(&fc->lock);
 230         list_del(&ff->write_entry);
 231         if (!RB_EMPTY_NODE(&ff->polled_node))
 232                 rb_erase(&ff->polled_node, &fc->polled_files);
 233         spin_unlock(&fc->lock);
 234
 235         wake_up_interruptible_all(&ff->poll_wait);
 236
 237         inarg->fh = ff->fh;
 238         inarg->flags = flags;
 239         req->in.h.opcode = opcode;
 240         req->in.h.nodeid = ff->nodeid;
 241         req->in.numargs = 1;
 242         req->in.args[0].size = sizeof(struct fuse_release_in);
 243         req->in.args[0].value = inarg;
 244 }
 245
 246 void fuse_release_common(struct file *file, int opcode)
 247 {
 248         struct fuse_file *ff;
 249         struct fuse_req *req;
 250
 251         ff = file->private_data;
 252         if (unlikely(!ff))
 253                 return;
 254
 255         req = ff->reserved_req;
 256         fuse_prepare_release(ff, file->f_flags, opcode);
 257
 258         if (ff->flock) {
 259                 struct fuse_release_in *inarg = &req->misc.release.in;
 260                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 261                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 262                                                        (fl_owner_t) file);
 263         }
 264         /* Hold inode until release is finished */
 265         req->misc.release.inode = igrab(file_inode(file));
 266
 267         /*
 268          * Normally this will send the RELEASE request, however if
 269          * some asynchronous READ or WRITE requests are outstanding,
 270          * the sending will be delayed.
 271          *
 272          * Make the release synchronous if this is a fuseblk mount,
 273          * synchronous RELEASE is allowed (and desirable) in this case
 274          * because the server can be trusted not to screw up.
 275          */
 276         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 277 }
 278
 279 static int fuse_open(struct inode *inode, struct file *file)
 280 {
 281         return fuse_open_common(inode, file, false);
 282 }
 283
 284 static int fuse_release(struct inode *inode, struct file *file)
 285 {
 286         struct fuse_conn *fc = get_fuse_conn(inode);
 287
 288         /* see fuse_vma_close() for !writeback_cache case */
 289         if (fc->writeback_cache)
 290                 write_inode_now(inode, 1);
 291
 292         fuse_release_common(file, FUSE_RELEASE);
 293
 294         /* return value is ignored by VFS */
 295         return 0;
 296 }
 297
 298 void fuse_sync_release(struct fuse_file *ff, int flags)
 299 {
 300         WARN_ON(atomic_read(&ff->count) > 1);
 301         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 302         __set_bit(FR_FORCE, &ff->reserved_req->flags);
 303         __clear_bit(FR_BACKGROUND, &ff->reserved_req->flags);
 304         fuse_request_send(ff->fc, ff->reserved_req);
 305         fuse_put_request(ff->fc, ff->reserved_req);
 306         kfree(ff);
 307 }
 308 EXPORT_SYMBOL_GPL(fuse_sync_release);
 309
 310 /*
 311  * Scramble the ID space with XTEA, so that the value of the files_struct
 312  * pointer is not exposed to userspace.
 313  */
 314 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 315 {
 316         u32 *k = fc->scramble_key;
 317         u64 v = (unsigned long) id;
 318         u32 v0 = v;
 319         u32 v1 = v >> 32;
 320         u32 sum = 0;
 321         int i;
 322
 323         for (i = 0; i < 32; i++) {
 324                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 325                 sum += 0x9E3779B9;
 326                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 327         }
 328
 329         return (u64) v0 + ((u64) v1 << 32);
 330 }
 331
 332 /*
 333  * Check if any page in a range is under writeback
 334  *
 335  * This is currently done by walking the list of writepage requests
 336  * for the inode, which can be pretty inefficient.
 337  */
 338 static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 339                                    pgoff_t idx_to)
 340 {
 341         struct fuse_conn *fc = get_fuse_conn(inode);
 342         struct fuse_inode *fi = get_fuse_inode(inode);
 343         struct fuse_req *req;
 344         bool found = false;
 345
 346         spin_lock(&fc->lock);
 347         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 348                 pgoff_t curr_index;
 349
 350                 BUG_ON(req->inode != inode);
 351                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 352                 if (idx_from < curr_index + req->num_pages &&
 353                     curr_index <= idx_to) {
 354                         found = true;
 355                         break;
 356                 }
 357         }
 358         spin_unlock(&fc->lock);
 359
 360         return found;
 361 }
 362
 363 static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 364 {
 365         return fuse_range_is_writeback(inode, index, index);
 366 }
 367
 368 /*
 369  * Wait for page writeback to be completed.
 370  *
 371  * Since fuse doesn't rely on the VM writeback tracking, this has to
 372  * use some other means.
 373  */
 374 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 375 {
 376         struct fuse_inode *fi = get_fuse_inode(inode);
 377
 378         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 379         return 0;
 380 }
 381
 382 /*
 383  * Wait for all pending writepages on the inode to finish.
 384  *
 385  * This is currently done by blocking further writes with FUSE_NOWRITE
 386  * and waiting for all sent writes to complete.
 387  *
 388  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 389  * could conflict with truncation.
 390  */
 391 static void fuse_sync_writes(struct inode *inode)
 392 {
 393         fuse_set_nowrite(inode);
 394         fuse_release_nowrite(inode);
 395 }
 396
 397 static int fuse_flush(struct file *file, fl_owner_t id)
 398 {
 399         struct inode *inode = file_inode(file);
 400         struct fuse_conn *fc = get_fuse_conn(inode);
 401         struct fuse_file *ff = file->private_data;
 402         struct fuse_req *req;
 403         struct fuse_flush_in inarg;
 404         int err;
 405
 406         if (is_bad_inode(inode))
 407                 return -EIO;
 408
 409         if (fc->no_flush)
 410                 return 0;
 411
 412         err = write_inode_now(inode, 1);
 413         if (err)
 414                 return err;
 415
 416         mutex_lock(&inode->i_mutex);
 417         fuse_sync_writes(inode);
 418         mutex_unlock(&inode->i_mutex);
 419
 420         if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
 421             test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
 422                 err = -ENOSPC;
 423         if (test_bit(AS_EIO, &file->f_mapping->flags) &&
 424             test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
 425                 err = -EIO;
 426         if (err)
 427                 return err;
 428
 429         req = fuse_get_req_nofail_nopages(fc, file);
 430         memset(&inarg, 0, sizeof(inarg));
 431         inarg.fh = ff->fh;
 432         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 433         req->in.h.opcode = FUSE_FLUSH;
 434         req->in.h.nodeid = get_node_id(inode);
 435         req->in.numargs = 1;
 436         req->in.args[0].size = sizeof(inarg);
 437         req->in.args[0].value = &inarg;
 438         __set_bit(FR_FORCE, &req->flags);
 439         fuse_request_send(fc, req);
 440         err = req->out.h.error;
 441         fuse_put_request(fc, req);
 442         if (err == -ENOSYS) {
 443                 fc->no_flush = 1;
 444                 err = 0;
 445         }
 446         return err;
 447 }
 448
 449 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 450                       int datasync, int isdir)
 451 {
 452         struct inode *inode = file->f_mapping->host;
 453         struct fuse_conn *fc = get_fuse_conn(inode);
 454         struct fuse_file *ff = file->private_data;
 455         FUSE_ARGS(args);
 456         struct fuse_fsync_in inarg;
 457         int err;
 458
 459         if (is_bad_inode(inode))
 460                 return -EIO;
 461
 462         mutex_lock(&inode->i_mutex);
 463
 464         /*
 465          * Start writeback against all dirty pages of the inode, then
 466          * wait for all outstanding writes, before sending the FSYNC
 467          * request.
 468          */
 469         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 470         if (err)
 471                 goto out;
 472
 473         fuse_sync_writes(inode);
 474
 475         /*
 476          * Due to implementation of fuse writeback
 477          * filemap_write_and_wait_range() does not catch errors.
 478          * We have to do this directly after fuse_sync_writes()
 479          */
 480         if (test_bit(AS_ENOSPC, &file->f_mapping->flags) &&
 481             test_and_clear_bit(AS_ENOSPC, &file->f_mapping->flags))
 482                 err = -ENOSPC;
 483         if (test_bit(AS_EIO, &file->f_mapping->flags) &&
 484             test_and_clear_bit(AS_EIO, &file->f_mapping->flags))
 485                 err = -EIO;
 486         if (err)
 487                 goto out;
 488
 489         err = sync_inode_metadata(inode, 1);
 490         if (err)
 491                 goto out;
 492
 493         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 494                 goto out;
 495
 496         memset(&inarg, 0, sizeof(inarg));
 497         inarg.fh = ff->fh;
 498         inarg.fsync_flags = datasync ? 1 : 0;
 499         args.in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 500         args.in.h.nodeid = get_node_id(inode);
 501         args.in.numargs = 1;
 502         args.in.args[0].size = sizeof(inarg);
 503         args.in.args[0].value = &inarg;
 504         err = fuse_simple_request(fc, &args);
 505         if (err == -ENOSYS) {
 506                 if (isdir)
 507                         fc->no_fsyncdir = 1;
 508                 else
 509                         fc->no_fsync = 1;
 510                 err = 0;
 511         }
 512 out:
 513         mutex_unlock(&inode->i_mutex);
 514         return err;
 515 }
 516
 517 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 518                       int datasync)
 519 {
 520         return fuse_fsync_common(file, start, end, datasync, 0);
 521 }
 522
 523 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 524                     size_t count, int opcode)
 525 {
 526         struct fuse_read_in *inarg = &req->misc.read.in;
 527         struct fuse_file *ff = file->private_data;
 528
 529         inarg->fh = ff->fh;
 530         inarg->offset = pos;
 531         inarg->size = count;
 532         inarg->flags = file->f_flags;
 533         req->in.h.opcode = opcode;
 534         req->in.h.nodeid = ff->nodeid;
 535         req->in.numargs = 1;
 536         req->in.args[0].size = sizeof(struct fuse_read_in);
 537         req->in.args[0].value = inarg;
 538         req->out.argvar = 1;
 539         req->out.numargs = 1;
 540         req->out.args[0].size = count;
 541 }
 542
 543 static void fuse_release_user_pages(struct fuse_req *req, int write)
 544 {
 545         unsigned i;
 546
 547         for (i = 0; i < req->num_pages; i++) {
 548                 struct page *page = req->pages[i];
 549                 if (write)
 550                         set_page_dirty_lock(page);
 551                 put_page(page);
 552         }
 553 }
 554
 555 static void fuse_io_release(struct kref *kref)
 556 {
 557         kfree(container_of(kref, struct fuse_io_priv, refcnt));
 558 }
 559
 560 static ssize_t fuse_get_res_by_io(struct fuse_io_priv *io)
 561 {
 562         if (io->err)
 563                 return io->err;
 564
 565         if (io->bytes >= 0 && io->write)
 566                 return -EIO;
 567
 568         return io->bytes < 0 ? io->size : io->bytes;
 569 }
 570
 571 /**
 572  * In case of short read, the caller sets 'pos' to the position of
 573  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 574  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 575  *
 576  * An example:
 577  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 578  * both submitted asynchronously. The first of them was ACKed by userspace as
 579  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 580  * second request was ACKed as short, e.g. only 1K was read, resulting in
 581  * pos == 33K.
 582  *
 583  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 584  * will be equal to the length of the longest contiguous fragment of
 585  * transferred data starting from the beginning of IO request.
 586  */
 587 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 588 {
 589         bool is_sync = is_sync_kiocb(io->iocb);
 590         int left;
 591
 592         spin_lock(&io->lock);
 593         if (err)
 594                 io->err = io->err ? : err;
 595         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 596                 io->bytes = pos;
 597
 598         left = --io->reqs;
 599         if (!left && is_sync)
 600                 complete(io->done);
 601         spin_unlock(&io->lock);
 602
 603         if (!left && !is_sync) {
 604                 ssize_t res = fuse_get_res_by_io(io);
 605
 606                 if (res >= 0) {
 607                         struct inode *inode = file_inode(io->iocb->ki_filp);
 608                         struct fuse_conn *fc = get_fuse_conn(inode);
 609                         struct fuse_inode *fi = get_fuse_inode(inode);
 610
 611                         spin_lock(&fc->lock);
 612                         fi->attr_version = ++fc->attr_version;
 613                         spin_unlock(&fc->lock);
 614                 }
 615
 616                 io->iocb->ki_complete(io->iocb, res, 0);
 617         }
 618
 619         kref_put(&io->refcnt, fuse_io_release);
 620 }
 621
 622 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 623 {
 624         struct fuse_io_priv *io = req->io;
 625         ssize_t pos = -1;
 626
 627         fuse_release_user_pages(req, !io->write);
 628
 629         if (io->write) {
 630                 if (req->misc.write.in.size != req->misc.write.out.size)
 631                         pos = req->misc.write.in.offset - io->offset +
 632                                 req->misc.write.out.size;
 633         } else {
 634                 if (req->misc.read.in.size != req->out.args[0].size)
 635                         pos = req->misc.read.in.offset - io->offset +
 636                                 req->out.args[0].size;
 637         }
 638
 639         fuse_aio_complete(io, req->out.h.error, pos);
 640 }
 641
 642 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 643                 size_t num_bytes, struct fuse_io_priv *io)
 644 {
 645         spin_lock(&io->lock);
 646         kref_get(&io->refcnt);
 647         io->size += num_bytes;
 648         io->reqs++;
 649         spin_unlock(&io->lock);
 650
 651         req->io = io;
 652         req->end = fuse_aio_complete_req;
 653
 654         __fuse_get_request(req);
 655         fuse_request_send_background(fc, req);
 656
 657         return num_bytes;
 658 }
 659
 660 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 661                              loff_t pos, size_t count, fl_owner_t owner)
 662 {
 663         struct file *file = io->file;
 664         struct fuse_file *ff = file->private_data;
 665         struct fuse_conn *fc = ff->fc;
 666
 667         fuse_read_fill(req, file, pos, count, FUSE_READ);
 668         if (owner != NULL) {
 669                 struct fuse_read_in *inarg = &req->misc.read.in;
 670
 671                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 672                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 673         }
 674
 675         if (io->async)
 676                 return fuse_async_req_send(fc, req, count, io);
 677
 678         fuse_request_send(fc, req);
 679         return req->out.args[0].size;
 680 }
 681
 682 static void fuse_read_update_size(struct inode *inode, loff_t size,
 683                                   u64 attr_ver)
 684 {
 685         struct fuse_conn *fc = get_fuse_conn(inode);
 686         struct fuse_inode *fi = get_fuse_inode(inode);
 687
 688         spin_lock(&fc->lock);
 689         if (attr_ver == fi->attr_version && size < inode->i_size &&
 690             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 691                 fi->attr_version = ++fc->attr_version;
 692                 i_size_write(inode, size);
 693         }
 694         spin_unlock(&fc->lock);
 695 }
 696
 697 static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 698                             u64 attr_ver)
 699 {
 700         size_t num_read = req->out.args[0].size;
 701         struct fuse_conn *fc = get_fuse_conn(inode);
 702
 703         if (fc->writeback_cache) {
 704                 /*
 705                  * A hole in a file. Some data after the hole are in page cache,
 706                  * but have not reached the client fs yet. So, the hole is not
 707                  * present there.
 708                  */
 709                 int i;
 710                 int start_idx = num_read >> PAGE_CACHE_SHIFT;
 711                 size_t off = num_read & (PAGE_CACHE_SIZE - 1);
 712
 713                 for (i = start_idx; i < req->num_pages; i++) {
 714                         zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
 715                         off = 0;
 716                 }
 717         } else {
 718                 loff_t pos = page_offset(req->pages[0]) + num_read;
 719                 fuse_read_update_size(inode, pos, attr_ver);
 720         }
 721 }
 722
 723 static int fuse_do_readpage(struct file *file, struct page *page)
 724 {
 725         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
 726         struct inode *inode = page->mapping->host;
 727         struct fuse_conn *fc = get_fuse_conn(inode);
 728         struct fuse_req *req;
 729         size_t num_read;
 730         loff_t pos = page_offset(page);
 731         size_t count = PAGE_CACHE_SIZE;
 732         u64 attr_ver;
 733         int err;
 734
 735         /*
 736          * Page writeback can extend beyond the lifetime of the
 737          * page-cache page, so make sure we read a properly synced
 738          * page.
 739          */
 740         fuse_wait_on_page_writeback(inode, page->index);
 741
 742         req = fuse_get_req(fc, 1);
 743         if (IS_ERR(req))
 744                 return PTR_ERR(req);
 745
 746         attr_ver = fuse_get_attr_version(fc);
 747
 748         req->out.page_zeroing = 1;
 749         req->out.argpages = 1;
 750         req->num_pages = 1;
 751         req->pages[0] = page;
 752         req->page_descs[0].length = count;
 753         num_read = fuse_send_read(req, &io, pos, count, NULL);
 754         err = req->out.h.error;
 755
 756         if (!err) {
 757                 /*
 758                  * Short read means EOF.  If file size is larger, truncate it
 759                  */
 760                 if (num_read < count)
 761                         fuse_short_read(req, inode, attr_ver);
 762
 763                 SetPageUptodate(page);
 764         }
 765
 766         fuse_put_request(fc, req);
 767
 768         return err;
 769 }
 770
 771 static int fuse_readpage(struct file *file, struct page *page)
 772 {
 773         struct inode *inode = page->mapping->host;
 774         int err;
 775
 776         err = -EIO;
 777         if (is_bad_inode(inode))
 778                 goto out;
 779
 780         err = fuse_do_readpage(file, page);
 781         fuse_invalidate_atime(inode);
 782  out:
 783         unlock_page(page);
 784         return err;
 785 }
 786
 787 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 788 {
 789         int i;
 790         size_t count = req->misc.read.in.size;
 791         size_t num_read = req->out.args[0].size;
 792         struct address_space *mapping = NULL;
 793
 794         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 795                 mapping = req->pages[i]->mapping;
 796
 797         if (mapping) {
 798                 struct inode *inode = mapping->host;
 799
 800                 /*
 801                  * Short read means EOF. If file size is larger, truncate it
 802                  */
 803                 if (!req->out.h.error && num_read < count)
 804                         fuse_short_read(req, inode, req->misc.read.attr_ver);
 805
 806                 fuse_invalidate_atime(inode);
 807         }
 808
 809         for (i = 0; i < req->num_pages; i++) {
 810                 struct page *page = req->pages[i];
 811                 if (!req->out.h.error)
 812                         SetPageUptodate(page);
 813                 else
 814                         SetPageError(page);
 815                 unlock_page(page);
 816                 page_cache_release(page);
 817         }
 818         if (req->ff)
 819                 fuse_file_put(req->ff, false);
 820 }
 821
 822 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 823 {
 824         struct fuse_file *ff = file->private_data;
 825         struct fuse_conn *fc = ff->fc;
 826         loff_t pos = page_offset(req->pages[0]);
 827         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 828
 829         req->out.argpages = 1;
 830         req->out.page_zeroing = 1;
 831         req->out.page_replace = 1;
 832         fuse_read_fill(req, file, pos, count, FUSE_READ);
 833         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 834         if (fc->async_read) {
 835                 req->ff = fuse_file_get(ff);
 836                 req->end = fuse_readpages_end;
 837                 fuse_request_send_background(fc, req);
 838         } else {
 839                 fuse_request_send(fc, req);
 840                 fuse_readpages_end(fc, req);
 841                 fuse_put_request(fc, req);
 842         }
 843 }
 844
 845 struct fuse_fill_data {
 846         struct fuse_req *req;
 847         struct file *file;
 848         struct inode *inode;
 849         unsigned nr_pages;
 850 };
 851
 852 static int fuse_readpages_fill(void *_data, struct page *page)
 853 {
 854         struct fuse_fill_data *data = _data;
 855         struct fuse_req *req = data->req;
 856         struct inode *inode = data->inode;
 857         struct fuse_conn *fc = get_fuse_conn(inode);
 858
 859         fuse_wait_on_page_writeback(inode, page->index);
 860
 861         if (req->num_pages &&
 862             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 863              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 864              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 865                 int nr_alloc = min_t(unsigned, data->nr_pages,
 866                                      FUSE_MAX_PAGES_PER_REQ);
 867                 fuse_send_readpages(req, data->file);
 868                 if (fc->async_read)
 869                         req = fuse_get_req_for_background(fc, nr_alloc);
 870                 else
 871                         req = fuse_get_req(fc, nr_alloc);
 872
 873                 data->req = req;
 874                 if (IS_ERR(req)) {
 875                         unlock_page(page);
 876                         return PTR_ERR(req);
 877                 }
 878         }
 879
 880         if (WARN_ON(req->num_pages >= req->max_pages)) {
 881                 fuse_put_request(fc, req);
 882                 return -EIO;
 883         }
 884
 885         page_cache_get(page);
 886         req->pages[req->num_pages] = page;
 887         req->page_descs[req->num_pages].length = PAGE_SIZE;
 888         req->num_pages++;
 889         data->nr_pages--;
 890         return 0;
 891 }
 892
 893 static int fuse_readpages(struct file *file, struct address_space *mapping,
 894                           struct list_head *pages, unsigned nr_pages)
 895 {
 896         struct inode *inode = mapping->host;
 897         struct fuse_conn *fc = get_fuse_conn(inode);
 898         struct fuse_fill_data data;
 899         int err;
 900         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 901
 902         err = -EIO;
 903         if (is_bad_inode(inode))
 904                 goto out;
 905
 906         data.file = file;
 907         data.inode = inode;
 908         if (fc->async_read)
 909                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 910         else
 911                 data.req = fuse_get_req(fc, nr_alloc);
 912         data.nr_pages = nr_pages;
 913         err = PTR_ERR(data.req);
 914         if (IS_ERR(data.req))
 915                 goto out;
 916
 917         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 918         if (!err) {
 919                 if (data.req->num_pages)
 920                         fuse_send_readpages(data.req, file);
 921                 else
 922                         fuse_put_request(fc, data.req);
 923         }
 924 out:
 925         return err;
 926 }
 927
 928 static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 929 {
 930         struct inode *inode = iocb->ki_filp->f_mapping->host;
 931         struct fuse_conn *fc = get_fuse_conn(inode);
 932
 933         /*
 934          * In auto invalidate mode, always update attributes on read.
 935          * Otherwise, only update if we attempt to read past EOF (to ensure
 936          * i_size is up to date).
 937          */
 938         if (fc->auto_inval_data ||
 939             (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 940                 int err;
 941                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 942                 if (err)
 943                         return err;
 944         }
 945
 946         return generic_file_read_iter(iocb, to);
 947 }
 948
 949 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 950                             loff_t pos, size_t count)
 951 {
 952         struct fuse_write_in *inarg = &req->misc.write.in;
 953         struct fuse_write_out *outarg = &req->misc.write.out;
 954
 955         inarg->fh = ff->fh;
 956         inarg->offset = pos;
 957         inarg->size = count;
 958         req->in.h.opcode = FUSE_WRITE;
 959         req->in.h.nodeid = ff->nodeid;
 960         req->in.numargs = 2;
 961         if (ff->fc->minor < 9)
 962                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 963         else
 964                 req->in.args[0].size = sizeof(struct fuse_write_in);
 965         req->in.args[0].value = inarg;
 966         req->in.args[1].size = count;
 967         req->out.numargs = 1;
 968         req->out.args[0].size = sizeof(struct fuse_write_out);
 969         req->out.args[0].value = outarg;
 970 }
 971
 972 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 973                               loff_t pos, size_t count, fl_owner_t owner)
 974 {
 975         struct file *file = io->file;
 976         struct fuse_file *ff = file->private_data;
 977         struct fuse_conn *fc = ff->fc;
 978         struct fuse_write_in *inarg = &req->misc.write.in;
 979
 980         fuse_write_fill(req, ff, pos, count);
 981         inarg->flags = file->f_flags;
 982         if (owner != NULL) {
 983                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 984                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 985         }
 986
 987         if (io->async)
 988                 return fuse_async_req_send(fc, req, count, io);
 989
 990         fuse_request_send(fc, req);
 991         return req->misc.write.out.size;
 992 }
 993
 994 bool fuse_write_update_size(struct inode *inode, loff_t pos)
 995 {
 996         struct fuse_conn *fc = get_fuse_conn(inode);
 997         struct fuse_inode *fi = get_fuse_inode(inode);
 998         bool ret = false;
 999
1000         spin_lock(&fc->lock);
1001         fi->attr_version = ++fc->attr_version;
1002         if (pos > inode->i_size) {
1003                 i_size_write(inode, pos);
1004                 ret = true;
1005         }
1006         spin_unlock(&fc->lock);
1007
1008         return ret;
1009 }
1010
1011 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1012                                     struct inode *inode, loff_t pos,
1013                                     size_t count)
1014 {
1015         size_t res;
1016         unsigned offset;
1017         unsigned i;
1018         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1019
1020         for (i = 0; i < req->num_pages; i++)
1021                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1022
1023         res = fuse_send_write(req, &io, pos, count, NULL);
1024
1025         offset = req->page_descs[0].offset;
1026         count = res;
1027         for (i = 0; i < req->num_pages; i++) {
1028                 struct page *page = req->pages[i];
1029
1030                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1031                         SetPageUptodate(page);
1032
1033                 if (count > PAGE_CACHE_SIZE - offset)
1034                         count -= PAGE_CACHE_SIZE - offset;
1035                 else
1036                         count = 0;
1037                 offset = 0;
1038
1039                 unlock_page(page);
1040                 page_cache_release(page);
1041         }
1042
1043         return res;
1044 }
1045
1046 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1047                                struct address_space *mapping,
1048                                struct iov_iter *ii, loff_t pos)
1049 {
1050         struct fuse_conn *fc = get_fuse_conn(mapping->host);
1051         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1052         size_t count = 0;
1053         int err;
1054
1055         req->in.argpages = 1;
1056         req->page_descs[0].offset = offset;
1057
1058         do {
1059                 size_t tmp;
1060                 struct page *page;
1061                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1062                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1063                                      iov_iter_count(ii));
1064
1065                 bytes = min_t(size_t, bytes, fc->max_write - count);
1066
1067  again:
1068                 err = -EFAULT;
1069                 if (iov_iter_fault_in_readable(ii, bytes))
1070                         break;
1071
1072                 err = -ENOMEM;
1073                 page = grab_cache_page_write_begin(mapping, index, 0);
1074                 if (!page)
1075                         break;
1076
1077                 if (mapping_writably_mapped(mapping))
1078                         flush_dcache_page(page);
1079
1080                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1081                 flush_dcache_page(page);
1082
1083                 iov_iter_advance(ii, tmp);
1084                 if (!tmp) {
1085                         unlock_page(page);
1086                         page_cache_release(page);
1087                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1088                         goto again;
1089                 }
1090
1091                 err = 0;
1092                 req->pages[req->num_pages] = page;
1093                 req->page_descs[req->num_pages].length = tmp;
1094                 req->num_pages++;
1095
1096                 count += tmp;
1097                 pos += tmp;
1098                 offset += tmp;
1099                 if (offset == PAGE_CACHE_SIZE)
1100                         offset = 0;
1101
1102                 if (!fc->big_writes)
1103                         break;
1104         } while (iov_iter_count(ii) && count < fc->max_write &&
1105                  req->num_pages < req->max_pages && offset == 0);
1106
1107         return count > 0 ? count : err;
1108 }
1109
1110 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1111 {
1112         return min_t(unsigned,
1113                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1114                      (pos >> PAGE_CACHE_SHIFT) + 1,
1115                      FUSE_MAX_PAGES_PER_REQ);
1116 }
1117
1118 static ssize_t fuse_perform_write(struct file *file,
1119                                   struct address_space *mapping,
1120                                   struct iov_iter *ii, loff_t pos)
1121 {
1122         struct inode *inode = mapping->host;
1123         struct fuse_conn *fc = get_fuse_conn(inode);
1124         struct fuse_inode *fi = get_fuse_inode(inode);
1125         int err = 0;
1126         ssize_t res = 0;
1127
1128         if (is_bad_inode(inode))
1129                 return -EIO;
1130
1131         if (inode->i_size < pos + iov_iter_count(ii))
1132                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1133
1134         do {
1135                 struct fuse_req *req;
1136                 ssize_t count;
1137                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1138
1139                 req = fuse_get_req(fc, nr_pages);
1140                 if (IS_ERR(req)) {
1141                         err = PTR_ERR(req);
1142                         break;
1143                 }
1144
1145                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1146                 if (count <= 0) {
1147                         err = count;
1148                 } else {
1149                         size_t num_written;
1150
1151                         num_written = fuse_send_write_pages(req, file, inode,
1152                                                             pos, count);
1153                         err = req->out.h.error;
1154                         if (!err) {
1155                                 res += num_written;
1156                                 pos += num_written;
1157
1158                                 /* break out of the loop on short write */
1159                                 if (num_written != count)
1160                                         err = -EIO;
1161                         }
1162                 }
1163                 fuse_put_request(fc, req);
1164         } while (!err && iov_iter_count(ii));
1165
1166         if (res > 0)
1167                 fuse_write_update_size(inode, pos);
1168
1169         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1170         fuse_invalidate_attr(inode);
1171
1172         return res > 0 ? res : err;
1173 }
1174
1175 static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1176 {
1177         struct file *file = iocb->ki_filp;
1178         struct address_space *mapping = file->f_mapping;
1179         ssize_t written = 0;
1180         ssize_t written_buffered = 0;
1181         struct inode *inode = mapping->host;
1182         ssize_t err;
1183         loff_t endbyte = 0;
1184
1185         if (get_fuse_conn(inode)->writeback_cache) {
1186                 /* Update size (EOF optimization) and mode (SUID clearing) */
1187                 err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1188                 if (err)
1189                         return err;
1190
1191                 return generic_file_write_iter(iocb, from);
1192         }
1193
1194         mutex_lock(&inode->i_mutex);
1195
1196         /* We can write back this queue in page reclaim */
1197         current->backing_dev_info = inode_to_bdi(inode);
1198
1199         err = generic_write_checks(iocb, from);
1200         if (err <= 0)
1201                 goto out;
1202
1203         err = file_remove_privs(file);
1204         if (err)
1205                 goto out;
1206
1207         err = file_update_time(file);
1208         if (err)
1209                 goto out;
1210
1211         if (iocb->ki_flags & IOCB_DIRECT) {
1212                 loff_t pos = iocb->ki_pos;
1213                 written = generic_file_direct_write(iocb, from, pos);
1214                 if (written < 0 || !iov_iter_count(from))
1215                         goto out;
1216
1217                 pos += written;
1218
1219                 written_buffered = fuse_perform_write(file, mapping, from, pos);
1220                 if (written_buffered < 0) {
1221                         err = written_buffered;
1222                         goto out;
1223                 }
1224                 endbyte = pos + written_buffered - 1;
1225
1226                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1227                                                    endbyte);
1228                 if (err)
1229                         goto out;
1230
1231                 invalidate_mapping_pages(file->f_mapping,
1232                                          pos >> PAGE_CACHE_SHIFT,
1233                                          endbyte >> PAGE_CACHE_SHIFT);
1234
1235                 written += written_buffered;
1236                 iocb->ki_pos = pos + written_buffered;
1237         } else {
1238                 written = fuse_perform_write(file, mapping, from, iocb->ki_pos);
1239                 if (written >= 0)
1240                         iocb->ki_pos += written;
1241         }
1242 out:
1243         current->backing_dev_info = NULL;
1244         mutex_unlock(&inode->i_mutex);
1245
1246         return written ? written : err;
1247 }
1248
1249 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1250                 unsigned index, unsigned nr_pages)
1251 {
1252         int i;
1253
1254         for (i = index; i < index + nr_pages; i++)
1255                 req->page_descs[i].length = PAGE_SIZE -
1256                         req->page_descs[i].offset;
1257 }
1258
1259 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1260 {
1261         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1262 }
1263
1264 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1265                                         size_t max_size)
1266 {
1267         return min(iov_iter_single_seg_count(ii), max_size);
1268 }
1269
1270 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1271                                size_t *nbytesp, int write)
1272 {
1273         size_t nbytes = 0;  /* # bytes already packed in req */
1274
1275         /* Special case for kernel I/O: can copy directly into the buffer */
1276         if (ii->type & ITER_KVEC) {
1277                 unsigned long user_addr = fuse_get_user_addr(ii);
1278                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1279
1280                 if (write)
1281                         req->in.args[1].value = (void *) user_addr;
1282                 else
1283                         req->out.args[0].value = (void *) user_addr;
1284
1285                 iov_iter_advance(ii, frag_size);
1286                 *nbytesp = frag_size;
1287                 return 0;
1288         }
1289
1290         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1291                 unsigned npages;
1292                 size_t start;
1293                 ssize_t ret = iov_iter_get_pages(ii,
1294                                         &req->pages[req->num_pages],
1295                                         *nbytesp - nbytes,
1296                                         req->max_pages - req->num_pages,
1297                                         &start);
1298                 if (ret < 0)
1299                         return ret;
1300
1301                 iov_iter_advance(ii, ret);
1302                 nbytes += ret;
1303
1304                 ret += start;
1305                 npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1306
1307                 req->page_descs[req->num_pages].offset = start;
1308                 fuse_page_descs_length_init(req, req->num_pages, npages);
1309
1310                 req->num_pages += npages;
1311                 req->page_descs[req->num_pages - 1].length -=
1312                         (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1313         }
1314
1315         if (write)
1316                 req->in.argpages = 1;
1317         else
1318                 req->out.argpages = 1;
1319
1320         *nbytesp = nbytes;
1321
1322         return 0;
1323 }
1324
1325 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1326 {
1327         return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1328 }
1329
1330 ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1331                        loff_t *ppos, int flags)
1332 {
1333         int write = flags & FUSE_DIO_WRITE;
1334         int cuse = flags & FUSE_DIO_CUSE;
1335         struct file *file = io->file;
1336         struct inode *inode = file->f_mapping->host;
1337         struct fuse_file *ff = file->private_data;
1338         struct fuse_conn *fc = ff->fc;
1339         size_t nmax = write ? fc->max_write : fc->max_read;
1340         loff_t pos = *ppos;
1341         size_t count = iov_iter_count(iter);
1342         pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1343         pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1344         ssize_t res = 0;
1345         struct fuse_req *req;
1346
1347         if (io->async)
1348                 req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1349         else
1350                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1351         if (IS_ERR(req))
1352                 return PTR_ERR(req);
1353
1354         if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1355                 if (!write)
1356                         mutex_lock(&inode->i_mutex);
1357                 fuse_sync_writes(inode);
1358                 if (!write)
1359                         mutex_unlock(&inode->i_mutex);
1360         }
1361
1362         while (count) {
1363                 size_t nres;
1364                 fl_owner_t owner = current->files;
1365                 size_t nbytes = min(count, nmax);
1366                 int err = fuse_get_user_pages(req, iter, &nbytes, write);
1367                 if (err) {
1368                         res = err;
1369                         break;
1370                 }
1371
1372                 if (write)
1373                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1374                 else
1375                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1376
1377                 if (!io->async)
1378                         fuse_release_user_pages(req, !write);
1379                 if (req->out.h.error) {
1380                         if (!res)
1381                                 res = req->out.h.error;
1382                         break;
1383                 } else if (nres > nbytes) {
1384                         res = -EIO;
1385                         break;
1386                 }
1387                 count -= nres;
1388                 res += nres;
1389                 pos += nres;
1390                 if (nres != nbytes)
1391                         break;
1392                 if (count) {
1393                         fuse_put_request(fc, req);
1394                         if (io->async)
1395                                 req = fuse_get_req_for_background(fc,
1396                                         fuse_iter_npages(iter));
1397                         else
1398                                 req = fuse_get_req(fc, fuse_iter_npages(iter));
1399                         if (IS_ERR(req))
1400                                 break;
1401                 }
1402         }
1403         if (!IS_ERR(req))
1404                 fuse_put_request(fc, req);
1405         if (res > 0)
1406                 *ppos = pos;
1407
1408         return res;
1409 }
1410 EXPORT_SYMBOL_GPL(fuse_direct_io);
1411
1412 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1413                                   struct iov_iter *iter,
1414                                   loff_t *ppos)
1415 {
1416         ssize_t res;
1417         struct file *file = io->file;
1418         struct inode *inode = file_inode(file);
1419
1420         if (is_bad_inode(inode))
1421                 return -EIO;
1422
1423         res = fuse_direct_io(io, iter, ppos, 0);
1424
1425         fuse_invalidate_attr(inode);
1426
1427         return res;
1428 }
1429
1430 static ssize_t fuse_direct_read_iter(struct kiocb *iocb, struct iov_iter *to)
1431 {
1432         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(iocb->ki_filp);
1433         return __fuse_direct_read(&io, to, &iocb->ki_pos);
1434 }
1435
1436 static ssize_t fuse_direct_write_iter(struct kiocb *iocb, struct iov_iter *from)
1437 {
1438         struct file *file = iocb->ki_filp;
1439         struct inode *inode = file_inode(file);
1440         struct fuse_io_priv io = FUSE_IO_PRIV_SYNC(file);
1441         ssize_t res;
1442
1443         if (is_bad_inode(inode))
1444                 return -EIO;
1445
1446         /* Don't allow parallel writes to the same file */
1447         mutex_lock(&inode->i_mutex);
1448         res = generic_write_checks(iocb, from);
1449         if (res > 0)
1450                 res = fuse_direct_io(&io, from, &iocb->ki_pos, FUSE_DIO_WRITE);
1451         fuse_invalidate_attr(inode);
1452         if (res > 0)
1453                 fuse_write_update_size(inode, iocb->ki_pos);
1454         mutex_unlock(&inode->i_mutex);
1455
1456         return res;
1457 }
1458
1459 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1460 {
1461         int i;
1462
1463         for (i = 0; i < req->num_pages; i++)
1464                 __free_page(req->pages[i]);
1465
1466         if (req->ff)
1467                 fuse_file_put(req->ff, false);
1468 }
1469
1470 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1471 {
1472         struct inode *inode = req->inode;
1473         struct fuse_inode *fi = get_fuse_inode(inode);
1474         struct backing_dev_info *bdi = inode_to_bdi(inode);
1475         int i;
1476
1477         list_del(&req->writepages_entry);
1478         for (i = 0; i < req->num_pages; i++) {
1479                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1480                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1481                 wb_writeout_inc(&bdi->wb);
1482         }
1483         wake_up(&fi->page_waitq);
1484 }
1485
1486 /* Called under fc->lock, may release and reacquire it */
1487 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1488                                 loff_t size)
1489 __releases(fc->lock)
1490 __acquires(fc->lock)
1491 {
1492         struct fuse_inode *fi = get_fuse_inode(req->inode);
1493         struct fuse_write_in *inarg = &req->misc.write.in;
1494         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1495
1496         if (!fc->connected)
1497                 goto out_free;
1498
1499         if (inarg->offset + data_size <= size) {
1500                 inarg->size = data_size;
1501         } else if (inarg->offset < size) {
1502                 inarg->size = size - inarg->offset;
1503         } else {
1504                 /* Got truncated off completely */
1505                 goto out_free;
1506         }
1507
1508         req->in.args[1].size = inarg->size;
1509         fi->writectr++;
1510         fuse_request_send_background_locked(fc, req);
1511         return;
1512
1513  out_free:
1514         fuse_writepage_finish(fc, req);
1515         spin_unlock(&fc->lock);
1516         fuse_writepage_free(fc, req);
1517         fuse_put_request(fc, req);
1518         spin_lock(&fc->lock);
1519 }
1520
1521 /*
1522  * If fi->writectr is positive (no truncate or fsync going on) send
1523  * all queued writepage requests.
1524  *
1525  * Called with fc->lock
1526  */
1527 void fuse_flush_writepages(struct inode *inode)
1528 __releases(fc->lock)
1529 __acquires(fc->lock)
1530 {
1531         struct fuse_conn *fc = get_fuse_conn(inode);
1532         struct fuse_inode *fi = get_fuse_inode(inode);
1533         size_t crop = i_size_read(inode);
1534         struct fuse_req *req;
1535
1536         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1537                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1538                 list_del_init(&req->list);
1539                 fuse_send_writepage(fc, req, crop);
1540         }
1541 }
1542
1543 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1544 {
1545         struct inode *inode = req->inode;
1546         struct fuse_inode *fi = get_fuse_inode(inode);
1547
1548         mapping_set_error(inode->i_mapping, req->out.h.error);
1549         spin_lock(&fc->lock);
1550         while (req->misc.write.next) {
1551                 struct fuse_conn *fc = get_fuse_conn(inode);
1552                 struct fuse_write_in *inarg = &req->misc.write.in;
1553                 struct fuse_req *next = req->misc.write.next;
1554                 req->misc.write.next = next->misc.write.next;
1555                 next->misc.write.next = NULL;
1556                 next->ff = fuse_file_get(req->ff);
1557                 list_add(&next->writepages_entry, &fi->writepages);
1558
1559                 /*
1560                  * Skip fuse_flush_writepages() to make it easy to crop requests
1561                  * based on primary request size.
1562                  *
1563                  * 1st case (trivial): there are no concurrent activities using
1564                  * fuse_set/release_nowrite.  Then we're on safe side because
1565                  * fuse_flush_writepages() would call fuse_send_writepage()
1566                  * anyway.
1567                  *
1568                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1569                  * now for completion of all in-flight requests.  This happens
1570                  * rarely and no more than once per page, so this should be
1571                  * okay.
1572                  *
1573                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1574                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1575                  * that fuse_set_nowrite returned implies that all in-flight
1576                  * requests were completed along with all of their secondary
1577                  * requests.  Further primary requests are blocked by negative
1578                  * writectr.  Hence there cannot be any in-flight requests and
1579                  * no invocations of fuse_writepage_end() while we're in
1580                  * fuse_set_nowrite..fuse_release_nowrite section.
1581                  */
1582                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1583         }
1584         fi->writectr--;
1585         fuse_writepage_finish(fc, req);
1586         spin_unlock(&fc->lock);
1587         fuse_writepage_free(fc, req);
1588 }
1589
1590 static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1591                                                struct fuse_inode *fi)
1592 {
1593         struct fuse_file *ff = NULL;
1594
1595         spin_lock(&fc->lock);
1596         if (!list_empty(&fi->write_files)) {
1597                 ff = list_entry(fi->write_files.next, struct fuse_file,
1598                                 write_entry);
1599                 fuse_file_get(ff);
1600         }
1601         spin_unlock(&fc->lock);
1602
1603         return ff;
1604 }
1605
1606 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1607                                              struct fuse_inode *fi)
1608 {
1609         struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1610         WARN_ON(!ff);
1611         return ff;
1612 }
1613
1614 int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1615 {
1616         struct fuse_conn *fc = get_fuse_conn(inode);
1617         struct fuse_inode *fi = get_fuse_inode(inode);
1618         struct fuse_file *ff;
1619         int err;
1620
1621         ff = __fuse_write_file_get(fc, fi);
1622         err = fuse_flush_times(inode, ff);
1623         if (ff)
1624                 fuse_file_put(ff, 0);
1625
1626         return err;
1627 }
1628
1629 static int fuse_writepage_locked(struct page *page)
1630 {
1631         struct address_space *mapping = page->mapping;
1632         struct inode *inode = mapping->host;
1633         struct fuse_conn *fc = get_fuse_conn(inode);
1634         struct fuse_inode *fi = get_fuse_inode(inode);
1635         struct fuse_req *req;
1636         struct page *tmp_page;
1637         int error = -ENOMEM;
1638
1639         set_page_writeback(page);
1640
1641         req = fuse_request_alloc_nofs(1);
1642         if (!req)
1643                 goto err;
1644
1645         /* writeback always goes to bg_queue */
1646         __set_bit(FR_BACKGROUND, &req->flags);
1647         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1648         if (!tmp_page)
1649                 goto err_free;
1650
1651         error = -EIO;
1652         req->ff = fuse_write_file_get(fc, fi);
1653         if (!req->ff)
1654                 goto err_nofile;
1655
1656         fuse_write_fill(req, req->ff, page_offset(page), 0);
1657
1658         copy_highpage(tmp_page, page);
1659         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1660         req->misc.write.next = NULL;
1661         req->in.argpages = 1;
1662         req->num_pages = 1;
1663         req->pages[0] = tmp_page;
1664         req->page_descs[0].offset = 0;
1665         req->page_descs[0].length = PAGE_SIZE;
1666         req->end = fuse_writepage_end;
1667         req->inode = inode;
1668
1669         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1670         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1671
1672         spin_lock(&fc->lock);
1673         list_add(&req->writepages_entry, &fi->writepages);
1674         list_add_tail(&req->list, &fi->queued_writes);
1675         fuse_flush_writepages(inode);
1676         spin_unlock(&fc->lock);
1677
1678         end_page_writeback(page);
1679
1680         return 0;
1681
1682 err_nofile:
1683         __free_page(tmp_page);
1684 err_free:
1685         fuse_request_free(req);
1686 err:
1687         end_page_writeback(page);
1688         return error;
1689 }
1690
1691 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1692 {
1693         int err;
1694
1695         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1696                 /*
1697                  * ->writepages() should be called for sync() and friends.  We
1698                  * should only get here on direct reclaim and then we are
1699                  * allowed to skip a page which is already in flight
1700                  */
1701                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1702
1703                 redirty_page_for_writepage(wbc, page);
1704                 return 0;
1705         }
1706
1707         err = fuse_writepage_locked(page);
1708         unlock_page(page);
1709
1710         return err;
1711 }
1712
1713 struct fuse_fill_wb_data {
1714         struct fuse_req *req;
1715         struct fuse_file *ff;
1716         struct inode *inode;
1717         struct page **orig_pages;
1718 };
1719
1720 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1721 {
1722         struct fuse_req *req = data->req;
1723         struct inode *inode = data->inode;
1724         struct fuse_conn *fc = get_fuse_conn(inode);
1725         struct fuse_inode *fi = get_fuse_inode(inode);
1726         int num_pages = req->num_pages;
1727         int i;
1728
1729         req->ff = fuse_file_get(data->ff);
1730         spin_lock(&fc->lock);
1731         list_add_tail(&req->list, &fi->queued_writes);
1732         fuse_flush_writepages(inode);
1733         spin_unlock(&fc->lock);
1734
1735         for (i = 0; i < num_pages; i++)
1736                 end_page_writeback(data->orig_pages[i]);
1737 }
1738
1739 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1740                                      struct page *page)
1741 {
1742         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1743         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1744         struct fuse_req *tmp;
1745         struct fuse_req *old_req;
1746         bool found = false;
1747         pgoff_t curr_index;
1748
1749         BUG_ON(new_req->num_pages != 0);
1750
1751         spin_lock(&fc->lock);
1752         list_del(&new_req->writepages_entry);
1753         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1754                 BUG_ON(old_req->inode != new_req->inode);
1755                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1756                 if (curr_index <= page->index &&
1757                     page->index < curr_index + old_req->num_pages) {
1758                         found = true;
1759                         break;
1760                 }
1761         }
1762         if (!found) {
1763                 list_add(&new_req->writepages_entry, &fi->writepages);
1764                 goto out_unlock;
1765         }
1766
1767         new_req->num_pages = 1;
1768         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1769                 BUG_ON(tmp->inode != new_req->inode);
1770                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1771                 if (tmp->num_pages == 1 &&
1772                     curr_index == page->index) {
1773                         old_req = tmp;
1774                 }
1775         }
1776
1777         if (old_req->num_pages == 1 && test_bit(FR_PENDING, &old_req->flags)) {
1778                 struct backing_dev_info *bdi = inode_to_bdi(page->mapping->host);
1779
1780                 copy_highpage(old_req->pages[0], page);
1781                 spin_unlock(&fc->lock);
1782
1783                 dec_wb_stat(&bdi->wb, WB_WRITEBACK);
1784                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1785                 wb_writeout_inc(&bdi->wb);
1786                 fuse_writepage_free(fc, new_req);
1787                 fuse_request_free(new_req);
1788                 goto out;
1789         } else {
1790                 new_req->misc.write.next = old_req->misc.write.next;
1791                 old_req->misc.write.next = new_req;
1792         }
1793 out_unlock:
1794         spin_unlock(&fc->lock);
1795 out:
1796         return found;
1797 }
1798
1799 static int fuse_writepages_fill(struct page *page,
1800                 struct writeback_control *wbc, void *_data)
1801 {
1802         struct fuse_fill_wb_data *data = _data;
1803         struct fuse_req *req = data->req;
1804         struct inode *inode = data->inode;
1805         struct fuse_conn *fc = get_fuse_conn(inode);
1806         struct page *tmp_page;
1807         bool is_writeback;
1808         int err;
1809
1810         if (!data->ff) {
1811                 err = -EIO;
1812                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1813                 if (!data->ff)
1814                         goto out_unlock;
1815         }
1816
1817         /*
1818          * Being under writeback is unlikely but possible.  For example direct
1819          * read to an mmaped fuse file will set the page dirty twice; once when
1820          * the pages are faulted with get_user_pages(), and then after the read
1821          * completed.
1822          */
1823         is_writeback = fuse_page_is_writeback(inode, page->index);
1824
1825         if (req && req->num_pages &&
1826             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1827              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1828              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1829                 fuse_writepages_send(data);
1830                 data->req = NULL;
1831         }
1832         err = -ENOMEM;
1833         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1834         if (!tmp_page)
1835                 goto out_unlock;
1836
1837         /*
1838          * The page must not be redirtied until the writeout is completed
1839          * (i.e. userspace has sent a reply to the write request).  Otherwise
1840          * there could be more than one temporary page instance for each real
1841          * page.
1842          *
1843          * This is ensured by holding the page lock in page_mkwrite() while
1844          * checking fuse_page_is_writeback().  We already hold the page lock
1845          * since clear_page_dirty_for_io() and keep it held until we add the
1846          * request to the fi->writepages list and increment req->num_pages.
1847          * After this fuse_page_is_writeback() will indicate that the page is
1848          * under writeback, so we can release the page lock.
1849          */
1850         if (data->req == NULL) {
1851                 struct fuse_inode *fi = get_fuse_inode(inode);
1852
1853                 err = -ENOMEM;
1854                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1855                 if (!req) {
1856                         __free_page(tmp_page);
1857                         goto out_unlock;
1858                 }
1859
1860                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1861                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1862                 req->misc.write.next = NULL;
1863                 req->in.argpages = 1;
1864                 __set_bit(FR_BACKGROUND, &req->flags);
1865                 req->num_pages = 0;
1866                 req->end = fuse_writepage_end;
1867                 req->inode = inode;
1868
1869                 spin_lock(&fc->lock);
1870                 list_add(&req->writepages_entry, &fi->writepages);
1871                 spin_unlock(&fc->lock);
1872
1873                 data->req = req;
1874         }
1875         set_page_writeback(page);
1876
1877         copy_highpage(tmp_page, page);
1878         req->pages[req->num_pages] = tmp_page;
1879         req->page_descs[req->num_pages].offset = 0;
1880         req->page_descs[req->num_pages].length = PAGE_SIZE;
1881
1882         inc_wb_stat(&inode_to_bdi(inode)->wb, WB_WRITEBACK);
1883         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1884
1885         err = 0;
1886         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1887                 end_page_writeback(page);
1888                 data->req = NULL;
1889                 goto out_unlock;
1890         }
1891         data->orig_pages[req->num_pages] = page;
1892
1893         /*
1894          * Protected by fc->lock against concurrent access by
1895          * fuse_page_is_writeback().
1896          */
1897         spin_lock(&fc->lock);
1898         req->num_pages++;
1899         spin_unlock(&fc->lock);
1900
1901 out_unlock:
1902         unlock_page(page);
1903
1904         return err;
1905 }
1906
1907 static int fuse_writepages(struct address_space *mapping,
1908                            struct writeback_control *wbc)
1909 {
1910         struct inode *inode = mapping->host;
1911         struct fuse_fill_wb_data data;
1912         int err;
1913
1914         err = -EIO;
1915         if (is_bad_inode(inode))
1916                 goto out;
1917
1918         data.inode = inode;
1919         data.req = NULL;
1920         data.ff = NULL;
1921
1922         err = -ENOMEM;
1923         data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1924                                   sizeof(struct page *),
1925                                   GFP_NOFS);
1926         if (!data.orig_pages)
1927                 goto out;
1928
1929         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1930         if (data.req) {
1931                 /* Ignore errors if we can write at least one page */
1932                 BUG_ON(!data.req->num_pages);
1933                 fuse_writepages_send(&data);
1934                 err = 0;
1935         }
1936         if (data.ff)
1937                 fuse_file_put(data.ff, false);
1938
1939         kfree(data.orig_pages);
1940 out:
1941         return err;
1942 }
1943
1944 /*
1945  * It's worthy to make sure that space is reserved on disk for the write,
1946  * but how to implement it without killing performance need more thinking.
1947  */
1948 static int fuse_write_begin(struct file *file, struct address_space *mapping,
1949                 loff_t pos, unsigned len, unsigned flags,
1950                 struct page **pagep, void **fsdata)
1951 {
1952         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1953         struct fuse_conn *fc = get_fuse_conn(file_inode(file));
1954         struct page *page;
1955         loff_t fsize;
1956         int err = -ENOMEM;
1957
1958         WARN_ON(!fc->writeback_cache);
1959
1960         page = grab_cache_page_write_begin(mapping, index, flags);
1961         if (!page)
1962                 goto error;
1963
1964         fuse_wait_on_page_writeback(mapping->host, page->index);
1965
1966         if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
1967                 goto success;
1968         /*
1969          * Check if the start this page comes after the end of file, in which
1970          * case the readpage can be optimized away.
1971          */
1972         fsize = i_size_read(mapping->host);
1973         if (fsize <= (pos & PAGE_CACHE_MASK)) {
1974                 size_t off = pos & ~PAGE_CACHE_MASK;
1975                 if (off)
1976                         zero_user_segment(page, 0, off);
1977                 goto success;
1978         }
1979         err = fuse_do_readpage(file, page);
1980         if (err)
1981                 goto cleanup;
1982 success:
1983         *pagep = page;
1984         return 0;
1985
1986 cleanup:
1987         unlock_page(page);
1988         page_cache_release(page);
1989 error:
1990         return err;
1991 }
1992
1993 static int fuse_write_end(struct file *file, struct address_space *mapping,
1994                 loff_t pos, unsigned len, unsigned copied,
1995                 struct page *page, void *fsdata)
1996 {
1997         struct inode *inode = page->mapping->host;
1998
1999         if (!PageUptodate(page)) {
2000                 /* Zero any unwritten bytes at the end of the page */
2001                 size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2002                 if (endoff)
2003                         zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2004                 SetPageUptodate(page);
2005         }
2006
2007         fuse_write_update_size(inode, pos + copied);
2008         set_page_dirty(page);
2009         unlock_page(page);
2010         page_cache_release(page);
2011
2012         return copied;
2013 }
2014
2015 static int fuse_launder_page(struct page *page)
2016 {
2017         int err = 0;
2018         if (clear_page_dirty_for_io(page)) {
2019                 struct inode *inode = page->mapping->host;
2020                 err = fuse_writepage_locked(page);
2021                 if (!err)
2022                         fuse_wait_on_page_writeback(inode, page->index);
2023         }
2024         return err;
2025 }
2026
2027 /*
2028  * Write back dirty pages now, because there may not be any suitable
2029  * open files later
2030  */
2031 static void fuse_vma_close(struct vm_area_struct *vma)
2032 {
2033         filemap_write_and_wait(vma->vm_file->f_mapping);
2034 }
2035
2036 /*
2037  * Wait for writeback against this page to complete before allowing it
2038  * to be marked dirty again, and hence written back again, possibly
2039  * before the previous writepage completed.
2040  *
2041  * Block here, instead of in ->writepage(), so that the userspace fs
2042  * can only block processes actually operating on the filesystem.
2043  *
2044  * Otherwise unprivileged userspace fs would be able to block
2045  * unrelated:
2046  *
2047  * - page migration
2048  * - sync(2)
2049  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2050  */
2051 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2052 {
2053         struct page *page = vmf->page;
2054         struct inode *inode = file_inode(vma->vm_file);
2055
2056         file_update_time(vma->vm_file);
2057         lock_page(page);
2058         if (page->mapping != inode->i_mapping) {
2059                 unlock_page(page);
2060                 return VM_FAULT_NOPAGE;
2061         }
2062
2063         fuse_wait_on_page_writeback(inode, page->index);
2064         return VM_FAULT_LOCKED;
2065 }
2066
2067 static const struct vm_operations_struct fuse_file_vm_ops = {
2068         .close          = fuse_vma_close,
2069         .fault          = filemap_fault,
2070         .map_pages      = filemap_map_pages,
2071         .page_mkwrite   = fuse_page_mkwrite,
2072 };
2073
2074 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2075 {
2076         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2077                 fuse_link_write_file(file);
2078
2079         file_accessed(file);
2080         vma->vm_ops = &fuse_file_vm_ops;
2081         return 0;
2082 }
2083
2084 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2085 {
2086         /* Can't provide the coherency needed for MAP_SHARED */
2087         if (vma->vm_flags & VM_MAYSHARE)
2088                 return -ENODEV;
2089
2090         invalidate_inode_pages2(file->f_mapping);
2091
2092         return generic_file_mmap(file, vma);
2093 }
2094
2095 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2096                                   struct file_lock *fl)
2097 {
2098         switch (ffl->type) {
2099         case F_UNLCK:
2100                 break;
2101
2102         case F_RDLCK:
2103         case F_WRLCK:
2104                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2105                     ffl->end < ffl->start)
2106                         return -EIO;
2107
2108                 fl->fl_start = ffl->start;
2109                 fl->fl_end = ffl->end;
2110                 fl->fl_pid = ffl->pid;
2111                 break;
2112
2113         default:
2114                 return -EIO;
2115         }
2116         fl->fl_type = ffl->type;
2117         return 0;
2118 }
2119
2120 static void fuse_lk_fill(struct fuse_args *args, struct file *file,
2121                          const struct file_lock *fl, int opcode, pid_t pid,
2122                          int flock, struct fuse_lk_in *inarg)
2123 {
2124         struct inode *inode = file_inode(file);
2125         struct fuse_conn *fc = get_fuse_conn(inode);
2126         struct fuse_file *ff = file->private_data;
2127
2128         memset(inarg, 0, sizeof(*inarg));
2129         inarg->fh = ff->fh;
2130         inarg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2131         inarg->lk.start = fl->fl_start;
2132         inarg->lk.end = fl->fl_end;
2133         inarg->lk.type = fl->fl_type;
2134         inarg->lk.pid = pid;
2135         if (flock)
2136                 inarg->lk_flags |= FUSE_LK_FLOCK;
2137         args->in.h.opcode = opcode;
2138         args->in.h.nodeid = get_node_id(inode);
2139         args->in.numargs = 1;
2140         args->in.args[0].size = sizeof(*inarg);
2141         args->in.args[0].value = inarg;
2142 }
2143
2144 static int fuse_getlk(struct file *file, struct file_lock *fl)
2145 {
2146         struct inode *inode = file_inode(file);
2147         struct fuse_conn *fc = get_fuse_conn(inode);
2148         FUSE_ARGS(args);
2149         struct fuse_lk_in inarg;
2150         struct fuse_lk_out outarg;
2151         int err;
2152
2153         fuse_lk_fill(&args, file, fl, FUSE_GETLK, 0, 0, &inarg);
2154         args.out.numargs = 1;
2155         args.out.args[0].size = sizeof(outarg);
2156         args.out.args[0].value = &outarg;
2157         err = fuse_simple_request(fc, &args);
2158         if (!err)
2159                 err = convert_fuse_file_lock(&outarg.lk, fl);
2160
2161         return err;
2162 }
2163
2164 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2165 {
2166         struct inode *inode = file_inode(file);
2167         struct fuse_conn *fc = get_fuse_conn(inode);
2168         FUSE_ARGS(args);
2169         struct fuse_lk_in inarg;
2170         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2171         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2172         int err;
2173
2174         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2175                 /* NLM needs asynchronous locks, which we don't support yet */
2176                 return -ENOLCK;
2177         }
2178
2179         /* Unlock on close is handled by the flush method */
2180         if (fl->fl_flags & FL_CLOSE)
2181                 return 0;
2182
2183         fuse_lk_fill(&args, file, fl, opcode, pid, flock, &inarg);
2184         err = fuse_simple_request(fc, &args);
2185
2186         /* locking is restartable */
2187         if (err == -EINTR)
2188                 err = -ERESTARTSYS;
2189
2190         return err;
2191 }
2192
2193 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2194 {
2195         struct inode *inode = file_inode(file);
2196         struct fuse_conn *fc = get_fuse_conn(inode);
2197         int err;
2198
2199         if (cmd == F_CANCELLK) {
2200                 err = 0;
2201         } else if (cmd == F_GETLK) {
2202                 if (fc->no_lock) {
2203                         posix_test_lock(file, fl);
2204                         err = 0;
2205                 } else
2206                         err = fuse_getlk(file, fl);
2207         } else {
2208                 if (fc->no_lock)
2209                         err = posix_lock_file(file, fl, NULL);
2210                 else
2211                         err = fuse_setlk(file, fl, 0);
2212         }
2213         return err;
2214 }
2215
2216 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2217 {
2218         struct inode *inode = file_inode(file);
2219         struct fuse_conn *fc = get_fuse_conn(inode);
2220         int err;
2221
2222         if (fc->no_flock) {
2223                 err = locks_lock_file_wait(file, fl);
2224         } else {
2225                 struct fuse_file *ff = file->private_data;
2226
2227                 /* emulate flock with POSIX locks */
2228                 ff->flock = true;
2229                 err = fuse_setlk(file, fl, 1);
2230         }
2231
2232         return err;
2233 }
2234
2235 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2236 {
2237         struct inode *inode = mapping->host;
2238         struct fuse_conn *fc = get_fuse_conn(inode);
2239         FUSE_ARGS(args);
2240         struct fuse_bmap_in inarg;
2241         struct fuse_bmap_out outarg;
2242         int err;
2243
2244         if (!inode->i_sb->s_bdev || fc->no_bmap)
2245                 return 0;
2246
2247         memset(&inarg, 0, sizeof(inarg));
2248         inarg.block = block;
2249         inarg.blocksize = inode->i_sb->s_blocksize;
2250         args.in.h.opcode = FUSE_BMAP;
2251         args.in.h.nodeid = get_node_id(inode);
2252         args.in.numargs = 1;
2253         args.in.args[0].size = sizeof(inarg);
2254         args.in.args[0].value = &inarg;
2255         args.out.numargs = 1;
2256         args.out.args[0].size = sizeof(outarg);
2257         args.out.args[0].value = &outarg;
2258         err = fuse_simple_request(fc, &args);
2259         if (err == -ENOSYS)
2260                 fc->no_bmap = 1;
2261
2262         return err ? 0 : outarg.block;
2263 }
2264
2265 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2266 {
2267         loff_t retval;
2268         struct inode *inode = file_inode(file);
2269
2270         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2271         if (whence == SEEK_CUR || whence == SEEK_SET)
2272                 return generic_file_llseek(file, offset, whence);
2273
2274         mutex_lock(&inode->i_mutex);
2275         retval = fuse_update_attributes(inode, NULL, file, NULL);
2276         if (!retval)
2277                 retval = generic_file_llseek(file, offset, whence);
2278         mutex_unlock(&inode->i_mutex);
2279
2280         return retval;
2281 }
2282
2283 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2284                         unsigned int nr_segs, size_t bytes, bool to_user)
2285 {
2286         struct iov_iter ii;
2287         int page_idx = 0;
2288
2289         if (!bytes)
2290                 return 0;
2291
2292         iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2293
2294         while (iov_iter_count(&ii)) {
2295                 struct page *page = pages[page_idx++];
2296                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2297                 void *kaddr;
2298
2299                 kaddr = kmap(page);
2300
2301                 while (todo) {
2302                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2303                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2304                         size_t copy = min(todo, iov_len);
2305                         size_t left;
2306
2307                         if (!to_user)
2308                                 left = copy_from_user(kaddr, uaddr, copy);
2309                         else
2310                                 left = copy_to_user(uaddr, kaddr, copy);
2311
2312                         if (unlikely(left))
2313                                 return -EFAULT;
2314
2315                         iov_iter_advance(&ii, copy);
2316                         todo -= copy;
2317                         kaddr += copy;
2318                 }
2319
2320                 kunmap(page);
2321         }
2322
2323         return 0;
2324 }
2325
2326 /*
2327  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2328  * ABI was defined to be 'struct iovec' which is different on 32bit
2329  * and 64bit.  Fortunately we can determine which structure the server
2330  * used from the size of the reply.
2331  */
2332 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2333                                      size_t transferred, unsigned count,
2334                                      bool is_compat)
2335 {
2336 #ifdef CONFIG_COMPAT
2337         if (count * sizeof(struct compat_iovec) == transferred) {
2338                 struct compat_iovec *ciov = src;
2339                 unsigned i;
2340
2341                 /*
2342                  * With this interface a 32bit server cannot support
2343                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2344                  * requests
2345                  */
2346                 if (!is_compat)
2347                         return -EINVAL;
2348
2349                 for (i = 0; i < count; i++) {
2350                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2351                         dst[i].iov_len = ciov[i].iov_len;
2352                 }
2353                 return 0;
2354         }
2355 #endif
2356
2357         if (count * sizeof(struct iovec) != transferred)
2358                 return -EIO;
2359
2360         memcpy(dst, src, transferred);
2361         return 0;
2362 }
2363
2364 /* Make sure iov_length() won't overflow */
2365 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2366 {
2367         size_t n;
2368         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2369
2370         for (n = 0; n < count; n++, iov++) {
2371                 if (iov->iov_len > (size_t) max)
2372                         return -ENOMEM;
2373                 max -= iov->iov_len;
2374         }
2375         return 0;
2376 }
2377
2378 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2379                                  void *src, size_t transferred, unsigned count,
2380                                  bool is_compat)
2381 {
2382         unsigned i;
2383         struct fuse_ioctl_iovec *fiov = src;
2384
2385         if (fc->minor < 16) {
2386                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2387                                                  count, is_compat);
2388         }
2389
2390         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2391                 return -EIO;
2392
2393         for (i = 0; i < count; i++) {
2394                 /* Did the server supply an inappropriate value? */
2395                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2396                     fiov[i].len != (unsigned long) fiov[i].len)
2397                         return -EIO;
2398
2399                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2400                 dst[i].iov_len = (size_t) fiov[i].len;
2401
2402 #ifdef CONFIG_COMPAT
2403                 if (is_compat &&
2404                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2405                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2406                         return -EIO;
2407 #endif
2408         }
2409
2410         return 0;
2411 }
2412
2413
2414 /*
2415  * For ioctls, there is no generic way to determine how much memory
2416  * needs to be read and/or written.  Furthermore, ioctls are allowed
2417  * to dereference the passed pointer, so the parameter requires deep
2418  * copying but FUSE has no idea whatsoever about what to copy in or
2419  * out.
2420  *
2421  * This is solved by allowing FUSE server to retry ioctl with
2422  * necessary in/out iovecs.  Let's assume the ioctl implementation
2423  * needs to read in the following structure.
2424  *
2425  * struct a {
2426  *      char    *buf;
2427  *      size_t  buflen;
2428  * }
2429  *
2430  * On the first callout to FUSE server, inarg->in_size and
2431  * inarg->out_size will be NULL; then, the server completes the ioctl
2432  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2433  * the actual iov array to
2434  *
2435  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2436  *
2437  * which tells FUSE to copy in the requested area and retry the ioctl.
2438  * On the second round, the server has access to the structure and
2439  * from that it can tell what to look for next, so on the invocation,
2440  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2441  *
2442  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2443  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2444  *
2445  * FUSE will copy both struct a and the pointed buffer from the
2446  * process doing the ioctl and retry ioctl with both struct a and the
2447  * buffer.
2448  *
2449  * This time, FUSE server has everything it needs and completes ioctl
2450  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2451  *
2452  * Copying data out works the same way.
2453  *
2454  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2455  * automatically initializes in and out iovs by decoding @cmd with
2456  * _IOC_* macros and the server is not allowed to request RETRY.  This
2457  * limits ioctl data transfers to well-formed ioctls and is the forced
2458  * behavior for all FUSE servers.
2459  */
2460 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2461                    unsigned int flags)
2462 {
2463         struct fuse_file *ff = file->private_data;
2464         struct fuse_conn *fc = ff->fc;
2465         struct fuse_ioctl_in inarg = {
2466                 .fh = ff->fh,
2467                 .cmd = cmd,
2468                 .arg = arg,
2469                 .flags = flags
2470         };
2471         struct fuse_ioctl_out outarg;
2472         struct fuse_req *req = NULL;
2473         struct page **pages = NULL;
2474         struct iovec *iov_page = NULL;
2475         struct iovec *in_iov = NULL, *out_iov = NULL;
2476         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2477         size_t in_size, out_size, transferred;
2478         int err;
2479
2480 #if BITS_PER_LONG == 32
2481         inarg.flags |= FUSE_IOCTL_32BIT;
2482 #else
2483         if (flags & FUSE_IOCTL_COMPAT)
2484                 inarg.flags |= FUSE_IOCTL_32BIT;
2485 #endif
2486
2487         /* assume all the iovs returned by client always fits in a page */
2488         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2489
2490         err = -ENOMEM;
2491         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2492         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2493         if (!pages || !iov_page)
2494                 goto out;
2495
2496         /*
2497          * If restricted, initialize IO parameters as encoded in @cmd.
2498          * RETRY from server is not allowed.
2499          */
2500         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2501                 struct iovec *iov = iov_page;
2502
2503                 iov->iov_base = (void __user *)arg;
2504                 iov->iov_len = _IOC_SIZE(cmd);
2505
2506                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2507                         in_iov = iov;
2508                         in_iovs = 1;
2509                 }
2510
2511                 if (_IOC_DIR(cmd) & _IOC_READ) {
2512                         out_iov = iov;
2513                         out_iovs = 1;
2514                 }
2515         }
2516
2517  retry:
2518         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2519         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2520
2521         /*
2522          * Out data can be used either for actual out data or iovs,
2523          * make sure there always is at least one page.
2524          */
2525         out_size = max_t(size_t, out_size, PAGE_SIZE);
2526         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2527
2528         /* make sure there are enough buffer pages and init request with them */
2529         err = -ENOMEM;
2530         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2531                 goto out;
2532         while (num_pages < max_pages) {
2533                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2534                 if (!pages[num_pages])
2535                         goto out;
2536                 num_pages++;
2537         }
2538
2539         req = fuse_get_req(fc, num_pages);
2540         if (IS_ERR(req)) {
2541                 err = PTR_ERR(req);
2542                 req = NULL;
2543                 goto out;
2544         }
2545         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2546         req->num_pages = num_pages;
2547         fuse_page_descs_length_init(req, 0, req->num_pages);
2548
2549         /* okay, let's send it to the client */
2550         req->in.h.opcode = FUSE_IOCTL;
2551         req->in.h.nodeid = ff->nodeid;
2552         req->in.numargs = 1;
2553         req->in.args[0].size = sizeof(inarg);
2554         req->in.args[0].value = &inarg;
2555         if (in_size) {
2556                 req->in.numargs++;
2557                 req->in.args[1].size = in_size;
2558                 req->in.argpages = 1;
2559
2560                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2561                                            false);
2562                 if (err)
2563                         goto out;
2564         }
2565
2566         req->out.numargs = 2;
2567         req->out.args[0].size = sizeof(outarg);
2568         req->out.args[0].value = &outarg;
2569         req->out.args[1].size = out_size;
2570         req->out.argpages = 1;
2571         req->out.argvar = 1;
2572
2573         fuse_request_send(fc, req);
2574         err = req->out.h.error;
2575         transferred = req->out.args[1].size;
2576         fuse_put_request(fc, req);
2577         req = NULL;
2578         if (err)
2579                 goto out;
2580
2581         /* did it ask for retry? */
2582         if (outarg.flags & FUSE_IOCTL_RETRY) {
2583                 void *vaddr;
2584
2585                 /* no retry if in restricted mode */
2586                 err = -EIO;
2587                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2588                         goto out;
2589
2590                 in_iovs = outarg.in_iovs;
2591                 out_iovs = outarg.out_iovs;
2592
2593                 /*
2594                  * Make sure things are in boundary, separate checks
2595                  * are to protect against overflow.
2596                  */
2597                 err = -ENOMEM;
2598                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2599                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2600                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2601                         goto out;
2602
2603                 vaddr = kmap_atomic(pages[0]);
2604                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2605                                             transferred, in_iovs + out_iovs,
2606                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2607                 kunmap_atomic(vaddr);
2608                 if (err)
2609                         goto out;
2610
2611                 in_iov = iov_page;
2612                 out_iov = in_iov + in_iovs;
2613
2614                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2615                 if (err)
2616                         goto out;
2617
2618                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2619                 if (err)
2620                         goto out;
2621
2622                 goto retry;
2623         }
2624
2625         err = -EIO;
2626         if (transferred > inarg.out_size)
2627                 goto out;
2628
2629         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2630  out:
2631         if (req)
2632                 fuse_put_request(fc, req);
2633         free_page((unsigned long) iov_page);
2634         while (num_pages)
2635                 __free_page(pages[--num_pages]);
2636         kfree(pages);
2637
2638         return err ? err : outarg.result;
2639 }
2640 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2641
2642 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2643                        unsigned long arg, unsigned int flags)
2644 {
2645         struct inode *inode = file_inode(file);
2646         struct fuse_conn *fc = get_fuse_conn(inode);
2647
2648         if (!fuse_allow_current_process(fc))
2649                 return -EACCES;
2650
2651         if (is_bad_inode(inode))
2652                 return -EIO;
2653
2654         return fuse_do_ioctl(file, cmd, arg, flags);
2655 }
2656
2657 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2658                             unsigned long arg)
2659 {
2660         return fuse_ioctl_common(file, cmd, arg, 0);
2661 }
2662
2663 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2664                                    unsigned long arg)
2665 {
2666         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2667 }
2668
2669 /*
2670  * All files which have been polled are linked to RB tree
2671  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2672  * find the matching one.
2673  */
2674 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2675                                               struct rb_node **parent_out)
2676 {
2677         struct rb_node **link = &fc->polled_files.rb_node;
2678         struct rb_node *last = NULL;
2679
2680         while (*link) {
2681                 struct fuse_file *ff;
2682
2683                 last = *link;
2684                 ff = rb_entry(last, struct fuse_file, polled_node);
2685
2686                 if (kh < ff->kh)
2687                         link = &last->rb_left;
2688                 else if (kh > ff->kh)
2689                         link = &last->rb_right;
2690                 else
2691                         return link;
2692         }
2693
2694         if (parent_out)
2695                 *parent_out = last;
2696         return link;
2697 }
2698
2699 /*
2700  * The file is about to be polled.  Make sure it's on the polled_files
2701  * RB tree.  Note that files once added to the polled_files tree are
2702  * not removed before the file is released.  This is because a file
2703  * polled once is likely to be polled again.
2704  */
2705 static void fuse_register_polled_file(struct fuse_conn *fc,
2706                                       struct fuse_file *ff)
2707 {
2708         spin_lock(&fc->lock);
2709         if (RB_EMPTY_NODE(&ff->polled_node)) {
2710                 struct rb_node **link, *uninitialized_var(parent);
2711
2712                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2713                 BUG_ON(*link);
2714                 rb_link_node(&ff->polled_node, parent, link);
2715                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2716         }
2717         spin_unlock(&fc->lock);
2718 }
2719
2720 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2721 {
2722         struct fuse_file *ff = file->private_data;
2723         struct fuse_conn *fc = ff->fc;
2724         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2725         struct fuse_poll_out outarg;
2726         FUSE_ARGS(args);
2727         int err;
2728
2729         if (fc->no_poll)
2730                 return DEFAULT_POLLMASK;
2731
2732         poll_wait(file, &ff->poll_wait, wait);
2733         inarg.events = (__u32)poll_requested_events(wait);
2734
2735         /*
2736          * Ask for notification iff there's someone waiting for it.
2737          * The client may ignore the flag and always notify.
2738          */
2739         if (waitqueue_active(&ff->poll_wait)) {
2740                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2741                 fuse_register_polled_file(fc, ff);
2742         }
2743
2744         args.in.h.opcode = FUSE_POLL;
2745         args.in.h.nodeid = ff->nodeid;
2746         args.in.numargs = 1;
2747         args.in.args[0].size = sizeof(inarg);
2748         args.in.args[0].value = &inarg;
2749         args.out.numargs = 1;
2750         args.out.args[0].size = sizeof(outarg);
2751         args.out.args[0].value = &outarg;
2752         err = fuse_simple_request(fc, &args);
2753
2754         if (!err)
2755                 return outarg.revents;
2756         if (err == -ENOSYS) {
2757                 fc->no_poll = 1;
2758                 return DEFAULT_POLLMASK;
2759         }
2760         return POLLERR;
2761 }
2762 EXPORT_SYMBOL_GPL(fuse_file_poll);
2763
2764 /*
2765  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2766  * wakes up the poll waiters.
2767  */
2768 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2769                             struct fuse_notify_poll_wakeup_out *outarg)
2770 {
2771         u64 kh = outarg->kh;
2772         struct rb_node **link;
2773
2774         spin_lock(&fc->lock);
2775
2776         link = fuse_find_polled_node(fc, kh, NULL);
2777         if (*link) {
2778                 struct fuse_file *ff;
2779
2780                 ff = rb_entry(*link, struct fuse_file, polled_node);
2781                 wake_up_interruptible_sync(&ff->poll_wait);
2782         }
2783
2784         spin_unlock(&fc->lock);
2785         return 0;
2786 }
2787
2788 static void fuse_do_truncate(struct file *file)
2789 {
2790         struct inode *inode = file->f_mapping->host;
2791         struct iattr attr;
2792
2793         attr.ia_valid = ATTR_SIZE;
2794         attr.ia_size = i_size_read(inode);
2795
2796         attr.ia_file = file;
2797         attr.ia_valid |= ATTR_FILE;
2798
2799         fuse_do_setattr(inode, &attr, file);
2800 }
2801
2802 static inline loff_t fuse_round_up(loff_t off)
2803 {
2804         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2805 }
2806
2807 static ssize_t
2808 fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t offset)
2809 {
2810         DECLARE_COMPLETION_ONSTACK(wait);
2811         ssize_t ret = 0;
2812         struct file *file = iocb->ki_filp;
2813         struct fuse_file *ff = file->private_data;
2814         bool async_dio = ff->fc->async_dio;
2815         loff_t pos = 0;
2816         struct inode *inode;
2817         loff_t i_size;
2818         size_t count = iov_iter_count(iter);
2819         struct fuse_io_priv *io;
2820         bool is_sync = is_sync_kiocb(iocb);
2821
2822         pos = offset;
2823         inode = file->f_mapping->host;
2824         i_size = i_size_read(inode);
2825
2826         if ((iov_iter_rw(iter) == READ) && (offset > i_size))
2827                 return 0;
2828
2829         /* optimization for short read */
2830         if (async_dio && iov_iter_rw(iter) != WRITE && offset + count > i_size) {
2831                 if (offset >= i_size)
2832                         return 0;
2833                 iov_iter_truncate(iter, fuse_round_up(i_size - offset));
2834                 count = iov_iter_count(iter);
2835         }
2836
2837         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2838         if (!io)
2839                 return -ENOMEM;
2840         spin_lock_init(&io->lock);
2841         kref_init(&io->refcnt);
2842         io->reqs = 1;
2843         io->bytes = -1;
2844         io->size = 0;
2845         io->offset = offset;
2846         io->write = (iov_iter_rw(iter) == WRITE);
2847         io->err = 0;
2848         io->file = file;
2849         /*
2850          * By default, we want to optimize all I/Os with async request
2851          * submission to the client filesystem if supported.
2852          */
2853         io->async = async_dio;
2854         io->iocb = iocb;
2855
2856         /*
2857          * We cannot asynchronously extend the size of a file. We have no method
2858          * to wait on real async I/O requests, so we must submit this request
2859          * synchronously.
2860          */
2861         if (!is_sync && (offset + count > i_size) &&
2862             iov_iter_rw(iter) == WRITE)
2863                 io->async = false;
2864
2865         if (io->async && is_sync) {
2866                 /*
2867                  * Additional reference to keep io around after
2868                  * calling fuse_aio_complete()
2869                  */
2870                 kref_get(&io->refcnt);
2871                 io->done = &wait;
2872         }
2873
2874         if (iov_iter_rw(iter) == WRITE) {
2875                 ret = fuse_direct_io(io, iter, &pos, FUSE_DIO_WRITE);
2876                 fuse_invalidate_attr(inode);
2877         } else {
2878                 ret = __fuse_direct_read(io, iter, &pos);
2879         }
2880
2881         if (io->async) {
2882                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2883
2884                 /* we have a non-extending, async request, so return */
2885                 if (!is_sync)
2886                         return -EIOCBQUEUED;
2887
2888                 wait_for_completion(&wait);
2889                 ret = fuse_get_res_by_io(io);
2890         }
2891
2892         kref_put(&io->refcnt, fuse_io_release);
2893
2894         if (iov_iter_rw(iter) == WRITE) {
2895                 if (ret > 0)
2896                         fuse_write_update_size(inode, pos);
2897                 else if (ret < 0 && offset + count > i_size)
2898                         fuse_do_truncate(file);
2899         }
2900
2901         return ret;
2902 }
2903
2904 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2905                                 loff_t length)
2906 {
2907         struct fuse_file *ff = file->private_data;
2908         struct inode *inode = file_inode(file);
2909         struct fuse_inode *fi = get_fuse_inode(inode);
2910         struct fuse_conn *fc = ff->fc;
2911         FUSE_ARGS(args);
2912         struct fuse_fallocate_in inarg = {
2913                 .fh = ff->fh,
2914                 .offset = offset,
2915                 .length = length,
2916                 .mode = mode
2917         };
2918         int err;
2919         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2920                            (mode & FALLOC_FL_PUNCH_HOLE);
2921
2922         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2923                 return -EOPNOTSUPP;
2924
2925         if (fc->no_fallocate)
2926                 return -EOPNOTSUPP;
2927
2928         if (lock_inode) {
2929                 mutex_lock(&inode->i_mutex);
2930                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2931                         loff_t endbyte = offset + length - 1;
2932                         err = filemap_write_and_wait_range(inode->i_mapping,
2933                                                            offset, endbyte);
2934                         if (err)
2935                                 goto out;
2936
2937                         fuse_sync_writes(inode);
2938                 }
2939         }
2940
2941         if (!(mode & FALLOC_FL_KEEP_SIZE))
2942                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2943
2944         args.in.h.opcode = FUSE_FALLOCATE;
2945         args.in.h.nodeid = ff->nodeid;
2946         args.in.numargs = 1;
2947         args.in.args[0].size = sizeof(inarg);
2948         args.in.args[0].value = &inarg;
2949         err = fuse_simple_request(fc, &args);
2950         if (err == -ENOSYS) {
2951                 fc->no_fallocate = 1;
2952                 err = -EOPNOTSUPP;
2953         }
2954         if (err)
2955                 goto out;
2956
2957         /* we could have extended the file */
2958         if (!(mode & FALLOC_FL_KEEP_SIZE)) {
2959                 bool changed = fuse_write_update_size(inode, offset + length);
2960
2961                 if (changed && fc->writeback_cache)
2962                         file_update_time(file);
2963         }
2964
2965         if (mode & FALLOC_FL_PUNCH_HOLE)
2966                 truncate_pagecache_range(inode, offset, offset + length - 1);
2967
2968         fuse_invalidate_attr(inode);
2969
2970 out:
2971         if (!(mode & FALLOC_FL_KEEP_SIZE))
2972                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2973
2974         if (lock_inode)
2975                 mutex_unlock(&inode->i_mutex);
2976
2977         return err;
2978 }
2979
2980 static const struct file_operations fuse_file_operations = {
2981         .llseek         = fuse_file_llseek,
2982         .read_iter      = fuse_file_read_iter,
2983         .write_iter     = fuse_file_write_iter,
2984         .mmap           = fuse_file_mmap,
2985         .open           = fuse_open,
2986         .flush          = fuse_flush,
2987         .release        = fuse_release,
2988         .fsync          = fuse_fsync,
2989         .lock           = fuse_file_lock,
2990         .flock          = fuse_file_flock,
2991         .splice_read    = generic_file_splice_read,
2992         .unlocked_ioctl = fuse_file_ioctl,
2993         .compat_ioctl   = fuse_file_compat_ioctl,
2994         .poll           = fuse_file_poll,
2995         .fallocate      = fuse_file_fallocate,
2996 };
2997
2998 static const struct file_operations fuse_direct_io_file_operations = {
2999         .llseek         = fuse_file_llseek,
3000         .read_iter      = fuse_direct_read_iter,
3001         .write_iter     = fuse_direct_write_iter,
3002         .mmap           = fuse_direct_mmap,
3003         .open           = fuse_open,
3004         .flush          = fuse_flush,
3005         .release        = fuse_release,
3006         .fsync          = fuse_fsync,
3007         .lock           = fuse_file_lock,
3008         .flock          = fuse_file_flock,
3009         .unlocked_ioctl = fuse_file_ioctl,
3010         .compat_ioctl   = fuse_file_compat_ioctl,
3011         .poll           = fuse_file_poll,
3012         .fallocate      = fuse_file_fallocate,
3013         /* no splice_read */
3014 };
3015
3016 static const struct address_space_operations fuse_file_aops  = {
3017         .readpage       = fuse_readpage,
3018         .writepage      = fuse_writepage,
3019         .writepages     = fuse_writepages,
3020         .launder_page   = fuse_launder_page,
3021         .readpages      = fuse_readpages,
3022         .set_page_dirty = __set_page_dirty_nobuffers,
3023         .bmap           = fuse_bmap,
3024         .direct_IO      = fuse_direct_IO,
3025         .write_begin    = fuse_write_begin,
3026         .write_end      = fuse_write_end,
3027 };
3028
3029 void fuse_init_file_inode(struct inode *inode)
3030 {
3031         inode->i_fop = &fuse_file_operations;
3032         inode->i_data.a_ops = &fuse_file_aops;
3033 }