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
  18 static const struct file_operations fuse_direct_io_file_operations;
  19
  20 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  21                           int opcode, struct fuse_open_out *outargp)
  22 {
  23         struct fuse_open_in inarg;
  24         struct fuse_req *req;
  25         int err;
  26
  27         req = fuse_get_req(fc);
  28         if (IS_ERR(req))
  29                 return PTR_ERR(req);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35         req->in.h.opcode = opcode;
  36         req->in.h.nodeid = nodeid;
  37         req->in.numargs = 1;
  38         req->in.args[0].size = sizeof(inarg);
  39         req->in.args[0].value = &inarg;
  40         req->out.numargs = 1;
  41         req->out.args[0].size = sizeof(*outargp);
  42         req->out.args[0].value = outargp;
  43         fuse_request_send(fc, req);
  44         err = req->out.h.error;
  45         fuse_put_request(fc, req);
  46
  47         return err;
  48 }
  49
  50 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  51 {
  52         struct fuse_file *ff;
  53
  54         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  55         if (unlikely(!ff))
  56                 return NULL;
  57
  58         ff->fc = fc;
  59         ff->reserved_req = fuse_request_alloc();
  60         if (unlikely(!ff->reserved_req)) {
  61                 kfree(ff);
  62                 return NULL;
  63         }
  64
  65         INIT_LIST_HEAD(&ff->write_entry);
  66         atomic_set(&ff->count, 0);
  67         RB_CLEAR_NODE(&ff->polled_node);
  68         init_waitqueue_head(&ff->poll_wait);
  69
  70         spin_lock(&fc->lock);
  71         ff->kh = ++fc->khctr;
  72         spin_unlock(&fc->lock);
  73
  74         return ff;
  75 }
  76
  77 void fuse_file_free(struct fuse_file *ff)
  78 {
  79         fuse_request_free(ff->reserved_req);
  80         kfree(ff);
  81 }
  82
  83 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  84 {
  85         atomic_inc(&ff->count);
  86         return ff;
  87 }
  88
  89 static void fuse_release_async(struct work_struct *work)
  90 {
  91         struct fuse_req *req;
  92         struct fuse_conn *fc;
  93         struct path path;
  94
  95         req = container_of(work, struct fuse_req, misc.release.work);
  96         path = req->misc.release.path;
  97         fc = get_fuse_conn(path.dentry->d_inode);
  98
  99         fuse_put_request(fc, req);
 100         path_put(&path);
 101 }
 102
 103 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 104 {
 105         if (fc->destroy_req) {
 106                 /*
 107                  * If this is a fuseblk mount, then it's possible that
 108                  * releasing the path will result in releasing the
 109                  * super block and sending the DESTROY request.  If
 110                  * the server is single threaded, this would hang.
 111                  * For this reason do the path_put() in a separate
 112                  * thread.
 113                  */
 114                 atomic_inc(&req->count);
 115                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 116                 schedule_work(&req->misc.release.work);
 117         } else {
 118                 path_put(&req->misc.release.path);
 119         }
 120 }
 121
 122 static void fuse_file_put(struct fuse_file *ff, bool sync)
 123 {
 124         if (atomic_dec_and_test(&ff->count)) {
 125                 struct fuse_req *req = ff->reserved_req;
 126
 127                 if (sync) {
 128                         fuse_request_send(ff->fc, req);
 129                         path_put(&req->misc.release.path);
 130                         fuse_put_request(ff->fc, req);
 131                 } else {
 132                         req->end = fuse_release_end;
 133                         fuse_request_send_background(ff->fc, req);
 134                 }
 135                 kfree(ff);
 136         }
 137 }
 138
 139 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 140                  bool isdir)
 141 {
 142         struct fuse_open_out outarg;
 143         struct fuse_file *ff;
 144         int err;
 145         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 146
 147         ff = fuse_file_alloc(fc);
 148         if (!ff)
 149                 return -ENOMEM;
 150
 151         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 152         if (err) {
 153                 fuse_file_free(ff);
 154                 return err;
 155         }
 156
 157         if (isdir)
 158                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 159
 160         ff->fh = outarg.fh;
 161         ff->nodeid = nodeid;
 162         ff->open_flags = outarg.open_flags;
 163         file->private_data = fuse_file_get(ff);
 164
 165         return 0;
 166 }
 167 EXPORT_SYMBOL_GPL(fuse_do_open);
 168
 169 void fuse_finish_open(struct inode *inode, struct file *file)
 170 {
 171         struct fuse_file *ff = file->private_data;
 172         struct fuse_conn *fc = get_fuse_conn(inode);
 173
 174         if (ff->open_flags & FOPEN_DIRECT_IO)
 175                 file->f_op = &fuse_direct_io_file_operations;
 176         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 177                 invalidate_inode_pages2(inode->i_mapping);
 178         if (ff->open_flags & FOPEN_NONSEEKABLE)
 179                 nonseekable_open(inode, file);
 180         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 181                 struct fuse_inode *fi = get_fuse_inode(inode);
 182
 183                 spin_lock(&fc->lock);
 184                 fi->attr_version = ++fc->attr_version;
 185                 i_size_write(inode, 0);
 186                 spin_unlock(&fc->lock);
 187                 fuse_invalidate_attr(inode);
 188         }
 189 }
 190
 191 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 192 {
 193         struct fuse_conn *fc = get_fuse_conn(inode);
 194         int err;
 195
 196         /* VFS checks this, but only _after_ ->open() */
 197         if (file->f_flags & O_DIRECT)
 198                 return -EINVAL;
 199
 200         err = generic_file_open(inode, file);
 201         if (err)
 202                 return err;
 203
 204         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 205         if (err)
 206                 return err;
 207
 208         fuse_finish_open(inode, file);
 209
 210         return 0;
 211 }
 212
 213 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 214 {
 215         struct fuse_conn *fc = ff->fc;
 216         struct fuse_req *req = ff->reserved_req;
 217         struct fuse_release_in *inarg = &req->misc.release.in;
 218
 219         spin_lock(&fc->lock);
 220         list_del(&ff->write_entry);
 221         if (!RB_EMPTY_NODE(&ff->polled_node))
 222                 rb_erase(&ff->polled_node, &fc->polled_files);
 223         spin_unlock(&fc->lock);
 224
 225         wake_up_interruptible_all(&ff->poll_wait);
 226
 227         inarg->fh = ff->fh;
 228         inarg->flags = flags;
 229         req->in.h.opcode = opcode;
 230         req->in.h.nodeid = ff->nodeid;
 231         req->in.numargs = 1;
 232         req->in.args[0].size = sizeof(struct fuse_release_in);
 233         req->in.args[0].value = inarg;
 234 }
 235
 236 void fuse_release_common(struct file *file, int opcode)
 237 {
 238         struct fuse_file *ff;
 239         struct fuse_req *req;
 240
 241         ff = file->private_data;
 242         if (unlikely(!ff))
 243                 return;
 244
 245         req = ff->reserved_req;
 246         fuse_prepare_release(ff, file->f_flags, opcode);
 247
 248         if (ff->flock) {
 249                 struct fuse_release_in *inarg = &req->misc.release.in;
 250                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 251                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 252                                                        (fl_owner_t) file);
 253         }
 254         /* Hold vfsmount and dentry until release is finished */
 255         path_get(&file->f_path);
 256         req->misc.release.path = file->f_path;
 257
 258         /*
 259          * Normally this will send the RELEASE request, however if
 260          * some asynchronous READ or WRITE requests are outstanding,
 261          * the sending will be delayed.
 262          *
 263          * Make the release synchronous if this is a fuseblk mount,
 264          * synchronous RELEASE is allowed (and desirable) in this case
 265          * because the server can be trusted not to screw up.
 266          */
 267         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 268 }
 269
 270 static int fuse_open(struct inode *inode, struct file *file)
 271 {
 272         return fuse_open_common(inode, file, false);
 273 }
 274
 275 static int fuse_release(struct inode *inode, struct file *file)
 276 {
 277         fuse_release_common(file, FUSE_RELEASE);
 278
 279         /* return value is ignored by VFS */
 280         return 0;
 281 }
 282
 283 void fuse_sync_release(struct fuse_file *ff, int flags)
 284 {
 285         WARN_ON(atomic_read(&ff->count) > 1);
 286         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 287         ff->reserved_req->force = 1;
 288         fuse_request_send(ff->fc, ff->reserved_req);
 289         fuse_put_request(ff->fc, ff->reserved_req);
 290         kfree(ff);
 291 }
 292 EXPORT_SYMBOL_GPL(fuse_sync_release);
 293
 294 /*
 295  * Scramble the ID space with XTEA, so that the value of the files_struct
 296  * pointer is not exposed to userspace.
 297  */
 298 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 299 {
 300         u32 *k = fc->scramble_key;
 301         u64 v = (unsigned long) id;
 302         u32 v0 = v;
 303         u32 v1 = v >> 32;
 304         u32 sum = 0;
 305         int i;
 306
 307         for (i = 0; i < 32; i++) {
 308                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 309                 sum += 0x9E3779B9;
 310                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 311         }
 312
 313         return (u64) v0 + ((u64) v1 << 32);
 314 }
 315
 316 /*
 317  * Check if page is under writeback
 318  *
 319  * This is currently done by walking the list of writepage requests
 320  * for the inode, which can be pretty inefficient.
 321  */
 322 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 323 {
 324         struct fuse_conn *fc = get_fuse_conn(inode);
 325         struct fuse_inode *fi = get_fuse_inode(inode);
 326         struct fuse_req *req;
 327         bool found = false;
 328
 329         spin_lock(&fc->lock);
 330         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 331                 pgoff_t curr_index;
 332
 333                 BUG_ON(req->inode != inode);
 334                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 335                 if (curr_index == index) {
 336                         found = true;
 337                         break;
 338                 }
 339         }
 340         spin_unlock(&fc->lock);
 341
 342         return found;
 343 }
 344
 345 /*
 346  * Wait for page writeback to be completed.
 347  *
 348  * Since fuse doesn't rely on the VM writeback tracking, this has to
 349  * use some other means.
 350  */
 351 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 352 {
 353         struct fuse_inode *fi = get_fuse_inode(inode);
 354
 355         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 356         return 0;
 357 }
 358
 359 static int fuse_flush(struct file *file, fl_owner_t id)
 360 {
 361         struct inode *inode = file->f_path.dentry->d_inode;
 362         struct fuse_conn *fc = get_fuse_conn(inode);
 363         struct fuse_file *ff = file->private_data;
 364         struct fuse_req *req;
 365         struct fuse_flush_in inarg;
 366         int err;
 367
 368         if (is_bad_inode(inode))
 369                 return -EIO;
 370
 371         if (fc->no_flush)
 372                 return 0;
 373
 374         req = fuse_get_req_nofail(fc, file);
 375         memset(&inarg, 0, sizeof(inarg));
 376         inarg.fh = ff->fh;
 377         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 378         req->in.h.opcode = FUSE_FLUSH;
 379         req->in.h.nodeid = get_node_id(inode);
 380         req->in.numargs = 1;
 381         req->in.args[0].size = sizeof(inarg);
 382         req->in.args[0].value = &inarg;
 383         req->force = 1;
 384         fuse_request_send(fc, req);
 385         err = req->out.h.error;
 386         fuse_put_request(fc, req);
 387         if (err == -ENOSYS) {
 388                 fc->no_flush = 1;
 389                 err = 0;
 390         }
 391         return err;
 392 }
 393
 394 /*
 395  * Wait for all pending writepages on the inode to finish.
 396  *
 397  * This is currently done by blocking further writes with FUSE_NOWRITE
 398  * and waiting for all sent writes to complete.
 399  *
 400  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 401  * could conflict with truncation.
 402  */
 403 static void fuse_sync_writes(struct inode *inode)
 404 {
 405         fuse_set_nowrite(inode);
 406         fuse_release_nowrite(inode);
 407 }
 408
 409 int fuse_fsync_common(struct file *file, int datasync, int isdir)
 410 {
 411         struct inode *inode = file->f_mapping->host;
 412         struct fuse_conn *fc = get_fuse_conn(inode);
 413         struct fuse_file *ff = file->private_data;
 414         struct fuse_req *req;
 415         struct fuse_fsync_in inarg;
 416         int err;
 417
 418         if (is_bad_inode(inode))
 419                 return -EIO;
 420
 421         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 422                 return 0;
 423
 424         /*
 425          * Start writeback against all dirty pages of the inode, then
 426          * wait for all outstanding writes, before sending the FSYNC
 427          * request.
 428          */
 429         err = write_inode_now(inode, 0);
 430         if (err)
 431                 return err;
 432
 433         fuse_sync_writes(inode);
 434
 435         req = fuse_get_req(fc);
 436         if (IS_ERR(req))
 437                 return PTR_ERR(req);
 438
 439         memset(&inarg, 0, sizeof(inarg));
 440         inarg.fh = ff->fh;
 441         inarg.fsync_flags = datasync ? 1 : 0;
 442         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 443         req->in.h.nodeid = get_node_id(inode);
 444         req->in.numargs = 1;
 445         req->in.args[0].size = sizeof(inarg);
 446         req->in.args[0].value = &inarg;
 447         fuse_request_send(fc, req);
 448         err = req->out.h.error;
 449         fuse_put_request(fc, req);
 450         if (err == -ENOSYS) {
 451                 if (isdir)
 452                         fc->no_fsyncdir = 1;
 453                 else
 454                         fc->no_fsync = 1;
 455                 err = 0;
 456         }
 457         return err;
 458 }
 459
 460 static int fuse_fsync(struct file *file, int datasync)
 461 {
 462         return fuse_fsync_common(file, datasync, 0);
 463 }
 464
 465 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 466                     size_t count, int opcode)
 467 {
 468         struct fuse_read_in *inarg = &req->misc.read.in;
 469         struct fuse_file *ff = file->private_data;
 470
 471         inarg->fh = ff->fh;
 472         inarg->offset = pos;
 473         inarg->size = count;
 474         inarg->flags = file->f_flags;
 475         req->in.h.opcode = opcode;
 476         req->in.h.nodeid = ff->nodeid;
 477         req->in.numargs = 1;
 478         req->in.args[0].size = sizeof(struct fuse_read_in);
 479         req->in.args[0].value = inarg;
 480         req->out.argvar = 1;
 481         req->out.numargs = 1;
 482         req->out.args[0].size = count;
 483 }
 484
 485 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 486                              loff_t pos, size_t count, fl_owner_t owner)
 487 {
 488         struct fuse_file *ff = file->private_data;
 489         struct fuse_conn *fc = ff->fc;
 490
 491         fuse_read_fill(req, file, pos, count, FUSE_READ);
 492         if (owner != NULL) {
 493                 struct fuse_read_in *inarg = &req->misc.read.in;
 494
 495                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 496                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 497         }
 498         fuse_request_send(fc, req);
 499         return req->out.args[0].size;
 500 }
 501
 502 static void fuse_read_update_size(struct inode *inode, loff_t size,
 503                                   u64 attr_ver)
 504 {
 505         struct fuse_conn *fc = get_fuse_conn(inode);
 506         struct fuse_inode *fi = get_fuse_inode(inode);
 507
 508         spin_lock(&fc->lock);
 509         if (attr_ver == fi->attr_version && size < inode->i_size) {
 510                 fi->attr_version = ++fc->attr_version;
 511                 i_size_write(inode, size);
 512         }
 513         spin_unlock(&fc->lock);
 514 }
 515
 516 static int fuse_readpage(struct file *file, struct page *page)
 517 {
 518         struct inode *inode = page->mapping->host;
 519         struct fuse_conn *fc = get_fuse_conn(inode);
 520         struct fuse_req *req;
 521         size_t num_read;
 522         loff_t pos = page_offset(page);
 523         size_t count = PAGE_CACHE_SIZE;
 524         u64 attr_ver;
 525         int err;
 526
 527         err = -EIO;
 528         if (is_bad_inode(inode))
 529                 goto out;
 530
 531         /*
 532          * Page writeback can extend beyond the lifetime of the
 533          * page-cache page, so make sure we read a properly synced
 534          * page.
 535          */
 536         fuse_wait_on_page_writeback(inode, page->index);
 537
 538         req = fuse_get_req(fc);
 539         err = PTR_ERR(req);
 540         if (IS_ERR(req))
 541                 goto out;
 542
 543         attr_ver = fuse_get_attr_version(fc);
 544
 545         req->out.page_zeroing = 1;
 546         req->out.argpages = 1;
 547         req->num_pages = 1;
 548         req->pages[0] = page;
 549         num_read = fuse_send_read(req, file, pos, count, NULL);
 550         err = req->out.h.error;
 551         fuse_put_request(fc, req);
 552
 553         if (!err) {
 554                 /*
 555                  * Short read means EOF.  If file size is larger, truncate it
 556                  */
 557                 if (num_read < count)
 558                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 559
 560                 SetPageUptodate(page);
 561         }
 562
 563         fuse_invalidate_attr(inode); /* atime changed */
 564  out:
 565         unlock_page(page);
 566         return err;
 567 }
 568
 569 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 570 {
 571         int i;
 572         size_t count = req->misc.read.in.size;
 573         size_t num_read = req->out.args[0].size;
 574         struct address_space *mapping = NULL;
 575
 576         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 577                 mapping = req->pages[i]->mapping;
 578
 579         if (mapping) {
 580                 struct inode *inode = mapping->host;
 581
 582                 /*
 583                  * Short read means EOF. If file size is larger, truncate it
 584                  */
 585                 if (!req->out.h.error && num_read < count) {
 586                         loff_t pos;
 587
 588                         pos = page_offset(req->pages[0]) + num_read;
 589                         fuse_read_update_size(inode, pos,
 590                                               req->misc.read.attr_ver);
 591                 }
 592                 fuse_invalidate_attr(inode); /* atime changed */
 593         }
 594
 595         for (i = 0; i < req->num_pages; i++) {
 596                 struct page *page = req->pages[i];
 597                 if (!req->out.h.error)
 598                         SetPageUptodate(page);
 599                 else
 600                         SetPageError(page);
 601                 unlock_page(page);
 602                 page_cache_release(page);
 603         }
 604         if (req->ff)
 605                 fuse_file_put(req->ff, false);
 606 }
 607
 608 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 609 {
 610         struct fuse_file *ff = file->private_data;
 611         struct fuse_conn *fc = ff->fc;
 612         loff_t pos = page_offset(req->pages[0]);
 613         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 614
 615         req->out.argpages = 1;
 616         req->out.page_zeroing = 1;
 617         req->out.page_replace = 1;
 618         fuse_read_fill(req, file, pos, count, FUSE_READ);
 619         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 620         if (fc->async_read) {
 621                 req->ff = fuse_file_get(ff);
 622                 req->end = fuse_readpages_end;
 623                 fuse_request_send_background(fc, req);
 624         } else {
 625                 fuse_request_send(fc, req);
 626                 fuse_readpages_end(fc, req);
 627                 fuse_put_request(fc, req);
 628         }
 629 }
 630
 631 struct fuse_fill_data {
 632         struct fuse_req *req;
 633         struct file *file;
 634         struct inode *inode;
 635 };
 636
 637 static int fuse_readpages_fill(void *_data, struct page *page)
 638 {
 639         struct fuse_fill_data *data = _data;
 640         struct fuse_req *req = data->req;
 641         struct inode *inode = data->inode;
 642         struct fuse_conn *fc = get_fuse_conn(inode);
 643
 644         fuse_wait_on_page_writeback(inode, page->index);
 645
 646         if (req->num_pages &&
 647             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 648              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 649              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 650                 fuse_send_readpages(req, data->file);
 651                 data->req = req = fuse_get_req(fc);
 652                 if (IS_ERR(req)) {
 653                         unlock_page(page);
 654                         return PTR_ERR(req);
 655                 }
 656         }
 657         page_cache_get(page);
 658         req->pages[req->num_pages] = page;
 659         req->num_pages++;
 660         return 0;
 661 }
 662
 663 static int fuse_readpages(struct file *file, struct address_space *mapping,
 664                           struct list_head *pages, unsigned nr_pages)
 665 {
 666         struct inode *inode = mapping->host;
 667         struct fuse_conn *fc = get_fuse_conn(inode);
 668         struct fuse_fill_data data;
 669         int err;
 670
 671         err = -EIO;
 672         if (is_bad_inode(inode))
 673                 goto out;
 674
 675         data.file = file;
 676         data.inode = inode;
 677         data.req = fuse_get_req(fc);
 678         err = PTR_ERR(data.req);
 679         if (IS_ERR(data.req))
 680                 goto out;
 681
 682         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 683         if (!err) {
 684                 if (data.req->num_pages)
 685                         fuse_send_readpages(data.req, file);
 686                 else
 687                         fuse_put_request(fc, data.req);
 688         }
 689 out:
 690         return err;
 691 }
 692
 693 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 694                                   unsigned long nr_segs, loff_t pos)
 695 {
 696         struct inode *inode = iocb->ki_filp->f_mapping->host;
 697
 698         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 699                 int err;
 700                 /*
 701                  * If trying to read past EOF, make sure the i_size
 702                  * attribute is up-to-date.
 703                  */
 704                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 705                 if (err)
 706                         return err;
 707         }
 708
 709         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 710 }
 711
 712 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 713                             loff_t pos, size_t count)
 714 {
 715         struct fuse_write_in *inarg = &req->misc.write.in;
 716         struct fuse_write_out *outarg = &req->misc.write.out;
 717
 718         inarg->fh = ff->fh;
 719         inarg->offset = pos;
 720         inarg->size = count;
 721         req->in.h.opcode = FUSE_WRITE;
 722         req->in.h.nodeid = ff->nodeid;
 723         req->in.numargs = 2;
 724         if (ff->fc->minor < 9)
 725                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 726         else
 727                 req->in.args[0].size = sizeof(struct fuse_write_in);
 728         req->in.args[0].value = inarg;
 729         req->in.args[1].size = count;
 730         req->out.numargs = 1;
 731         req->out.args[0].size = sizeof(struct fuse_write_out);
 732         req->out.args[0].value = outarg;
 733 }
 734
 735 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 736                               loff_t pos, size_t count, fl_owner_t owner)
 737 {
 738         struct fuse_file *ff = file->private_data;
 739         struct fuse_conn *fc = ff->fc;
 740         struct fuse_write_in *inarg = &req->misc.write.in;
 741
 742         fuse_write_fill(req, ff, pos, count);
 743         inarg->flags = file->f_flags;
 744         if (owner != NULL) {
 745                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 746                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 747         }
 748         fuse_request_send(fc, req);
 749         return req->misc.write.out.size;
 750 }
 751
 752 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 753                         loff_t pos, unsigned len, unsigned flags,
 754                         struct page **pagep, void **fsdata)
 755 {
 756         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 757
 758         *pagep = grab_cache_page_write_begin(mapping, index, flags);
 759         if (!*pagep)
 760                 return -ENOMEM;
 761         return 0;
 762 }
 763
 764 void fuse_write_update_size(struct inode *inode, loff_t pos)
 765 {
 766         struct fuse_conn *fc = get_fuse_conn(inode);
 767         struct fuse_inode *fi = get_fuse_inode(inode);
 768
 769         spin_lock(&fc->lock);
 770         fi->attr_version = ++fc->attr_version;
 771         if (pos > inode->i_size)
 772                 i_size_write(inode, pos);
 773         spin_unlock(&fc->lock);
 774 }
 775
 776 static int fuse_buffered_write(struct file *file, struct inode *inode,
 777                                loff_t pos, unsigned count, struct page *page)
 778 {
 779         int err;
 780         size_t nres;
 781         struct fuse_conn *fc = get_fuse_conn(inode);
 782         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 783         struct fuse_req *req;
 784
 785         if (is_bad_inode(inode))
 786                 return -EIO;
 787
 788         /*
 789          * Make sure writepages on the same page are not mixed up with
 790          * plain writes.
 791          */
 792         fuse_wait_on_page_writeback(inode, page->index);
 793
 794         req = fuse_get_req(fc);
 795         if (IS_ERR(req))
 796                 return PTR_ERR(req);
 797
 798         req->in.argpages = 1;
 799         req->num_pages = 1;
 800         req->pages[0] = page;
 801         req->page_offset = offset;
 802         nres = fuse_send_write(req, file, pos, count, NULL);
 803         err = req->out.h.error;
 804         fuse_put_request(fc, req);
 805         if (!err && !nres)
 806                 err = -EIO;
 807         if (!err) {
 808                 pos += nres;
 809                 fuse_write_update_size(inode, pos);
 810                 if (count == PAGE_CACHE_SIZE)
 811                         SetPageUptodate(page);
 812         }
 813         fuse_invalidate_attr(inode);
 814         return err ? err : nres;
 815 }
 816
 817 static int fuse_write_end(struct file *file, struct address_space *mapping,
 818                         loff_t pos, unsigned len, unsigned copied,
 819                         struct page *page, void *fsdata)
 820 {
 821         struct inode *inode = mapping->host;
 822         int res = 0;
 823
 824         if (copied)
 825                 res = fuse_buffered_write(file, inode, pos, copied, page);
 826
 827         unlock_page(page);
 828         page_cache_release(page);
 829         return res;
 830 }
 831
 832 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 833                                     struct inode *inode, loff_t pos,
 834                                     size_t count)
 835 {
 836         size_t res;
 837         unsigned offset;
 838         unsigned i;
 839
 840         for (i = 0; i < req->num_pages; i++)
 841                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 842
 843         res = fuse_send_write(req, file, pos, count, NULL);
 844
 845         offset = req->page_offset;
 846         count = res;
 847         for (i = 0; i < req->num_pages; i++) {
 848                 struct page *page = req->pages[i];
 849
 850                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 851                         SetPageUptodate(page);
 852
 853                 if (count > PAGE_CACHE_SIZE - offset)
 854                         count -= PAGE_CACHE_SIZE - offset;
 855                 else
 856                         count = 0;
 857                 offset = 0;
 858
 859                 unlock_page(page);
 860                 page_cache_release(page);
 861         }
 862
 863         return res;
 864 }
 865
 866 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 867                                struct address_space *mapping,
 868                                struct iov_iter *ii, loff_t pos)
 869 {
 870         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 871         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 872         size_t count = 0;
 873         int err;
 874
 875         req->in.argpages = 1;
 876         req->page_offset = offset;
 877
 878         do {
 879                 size_t tmp;
 880                 struct page *page;
 881                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 882                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 883                                      iov_iter_count(ii));
 884
 885                 bytes = min_t(size_t, bytes, fc->max_write - count);
 886
 887  again:
 888                 err = -EFAULT;
 889                 if (iov_iter_fault_in_readable(ii, bytes))
 890                         break;
 891
 892                 err = -ENOMEM;
 893                 page = grab_cache_page_write_begin(mapping, index, 0);
 894                 if (!page)
 895                         break;
 896
 897                 if (mapping_writably_mapped(mapping))
 898                         flush_dcache_page(page);
 899
 900                 pagefault_disable();
 901                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 902                 pagefault_enable();
 903                 flush_dcache_page(page);
 904
 905                 if (!tmp) {
 906                         unlock_page(page);
 907                         page_cache_release(page);
 908                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 909                         goto again;
 910                 }
 911
 912                 err = 0;
 913                 req->pages[req->num_pages] = page;
 914                 req->num_pages++;
 915
 916                 iov_iter_advance(ii, tmp);
 917                 count += tmp;
 918                 pos += tmp;
 919                 offset += tmp;
 920                 if (offset == PAGE_CACHE_SIZE)
 921                         offset = 0;
 922
 923                 if (!fc->big_writes)
 924                         break;
 925         } while (iov_iter_count(ii) && count < fc->max_write &&
 926                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 927
 928         return count > 0 ? count : err;
 929 }
 930
 931 static ssize_t fuse_perform_write(struct file *file,
 932                                   struct address_space *mapping,
 933                                   struct iov_iter *ii, loff_t pos)
 934 {
 935         struct inode *inode = mapping->host;
 936         struct fuse_conn *fc = get_fuse_conn(inode);
 937         int err = 0;
 938         ssize_t res = 0;
 939
 940         if (is_bad_inode(inode))
 941                 return -EIO;
 942
 943         do {
 944                 struct fuse_req *req;
 945                 ssize_t count;
 946
 947                 req = fuse_get_req(fc);
 948                 if (IS_ERR(req)) {
 949                         err = PTR_ERR(req);
 950                         break;
 951                 }
 952
 953                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 954                 if (count <= 0) {
 955                         err = count;
 956                 } else {
 957                         size_t num_written;
 958
 959                         num_written = fuse_send_write_pages(req, file, inode,
 960                                                             pos, count);
 961                         err = req->out.h.error;
 962                         if (!err) {
 963                                 res += num_written;
 964                                 pos += num_written;
 965
 966                                 /* break out of the loop on short write */
 967                                 if (num_written != count)
 968                                         err = -EIO;
 969                         }
 970                 }
 971                 fuse_put_request(fc, req);
 972         } while (!err && iov_iter_count(ii));
 973
 974         if (res > 0)
 975                 fuse_write_update_size(inode, pos);
 976
 977         fuse_invalidate_attr(inode);
 978
 979         return res > 0 ? res : err;
 980 }
 981
 982 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 983                                    unsigned long nr_segs, loff_t pos)
 984 {
 985         struct file *file = iocb->ki_filp;
 986         struct address_space *mapping = file->f_mapping;
 987         size_t count = 0;
 988         ssize_t written = 0;
 989         struct inode *inode = mapping->host;
 990         ssize_t err;
 991         struct iov_iter i;
 992
 993         WARN_ON(iocb->ki_pos != pos);
 994
 995         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 996         if (err)
 997                 return err;
 998
 999         mutex_lock(&inode->i_mutex);
1000         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1001
1002         /* We can write back this queue in page reclaim */
1003         current->backing_dev_info = mapping->backing_dev_info;
1004
1005         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1006         if (err)
1007                 goto out;
1008
1009         if (count == 0)
1010                 goto out;
1011
1012         err = file_remove_suid(file);
1013         if (err)
1014                 goto out;
1015
1016         file_update_time(file);
1017
1018         iov_iter_init(&i, iov, nr_segs, count, 0);
1019         written = fuse_perform_write(file, mapping, &i, pos);
1020         if (written >= 0)
1021                 iocb->ki_pos = pos + written;
1022
1023 out:
1024         current->backing_dev_info = NULL;
1025         mutex_unlock(&inode->i_mutex);
1026
1027         return written ? written : err;
1028 }
1029
1030 static void fuse_release_user_pages(struct fuse_req *req, int write)
1031 {
1032         unsigned i;
1033
1034         for (i = 0; i < req->num_pages; i++) {
1035                 struct page *page = req->pages[i];
1036                 if (write)
1037                         set_page_dirty_lock(page);
1038                 put_page(page);
1039         }
1040 }
1041
1042 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1043                                size_t *nbytesp, int write)
1044 {
1045         size_t nbytes = *nbytesp;
1046         unsigned long user_addr = (unsigned long) buf;
1047         unsigned offset = user_addr & ~PAGE_MASK;
1048         int npages;
1049
1050         /* Special case for kernel I/O: can copy directly into the buffer */
1051         if (segment_eq(get_fs(), KERNEL_DS)) {
1052                 if (write)
1053                         req->in.args[1].value = (void *) user_addr;
1054                 else
1055                         req->out.args[0].value = (void *) user_addr;
1056
1057                 return 0;
1058         }
1059
1060         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1061         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1062         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1063         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1064         if (npages < 0)
1065                 return npages;
1066
1067         req->num_pages = npages;
1068         req->page_offset = offset;
1069
1070         if (write)
1071                 req->in.argpages = 1;
1072         else
1073                 req->out.argpages = 1;
1074
1075         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1076         *nbytesp = min(*nbytesp, nbytes);
1077
1078         return 0;
1079 }
1080
1081 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1082                        size_t count, loff_t *ppos, int write)
1083 {
1084         struct fuse_file *ff = file->private_data;
1085         struct fuse_conn *fc = ff->fc;
1086         size_t nmax = write ? fc->max_write : fc->max_read;
1087         loff_t pos = *ppos;
1088         ssize_t res = 0;
1089         struct fuse_req *req;
1090
1091         req = fuse_get_req(fc);
1092         if (IS_ERR(req))
1093                 return PTR_ERR(req);
1094
1095         while (count) {
1096                 size_t nres;
1097                 fl_owner_t owner = current->files;
1098                 size_t nbytes = min(count, nmax);
1099                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1100                 if (err) {
1101                         res = err;
1102                         break;
1103                 }
1104
1105                 if (write)
1106                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1107                 else
1108                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1109
1110                 fuse_release_user_pages(req, !write);
1111                 if (req->out.h.error) {
1112                         if (!res)
1113                                 res = req->out.h.error;
1114                         break;
1115                 } else if (nres > nbytes) {
1116                         res = -EIO;
1117                         break;
1118                 }
1119                 count -= nres;
1120                 res += nres;
1121                 pos += nres;
1122                 buf += nres;
1123                 if (nres != nbytes)
1124                         break;
1125                 if (count) {
1126                         fuse_put_request(fc, req);
1127                         req = fuse_get_req(fc);
1128                         if (IS_ERR(req))
1129                                 break;
1130                 }
1131         }
1132         if (!IS_ERR(req))
1133                 fuse_put_request(fc, req);
1134         if (res > 0)
1135                 *ppos = pos;
1136
1137         return res;
1138 }
1139 EXPORT_SYMBOL_GPL(fuse_direct_io);
1140
1141 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1142                                      size_t count, loff_t *ppos)
1143 {
1144         ssize_t res;
1145         struct inode *inode = file->f_path.dentry->d_inode;
1146
1147         if (is_bad_inode(inode))
1148                 return -EIO;
1149
1150         res = fuse_direct_io(file, buf, count, ppos, 0);
1151
1152         fuse_invalidate_attr(inode);
1153
1154         return res;
1155 }
1156
1157 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1158                                  size_t count, loff_t *ppos)
1159 {
1160         struct inode *inode = file->f_path.dentry->d_inode;
1161         ssize_t res;
1162
1163         if (is_bad_inode(inode))
1164                 return -EIO;
1165
1166         /* Don't allow parallel writes to the same file */
1167         mutex_lock(&inode->i_mutex);
1168         res = generic_write_checks(file, ppos, &count, 0);
1169         if (!res) {
1170                 res = fuse_direct_io(file, buf, count, ppos, 1);
1171                 if (res > 0)
1172                         fuse_write_update_size(inode, *ppos);
1173         }
1174         mutex_unlock(&inode->i_mutex);
1175
1176         fuse_invalidate_attr(inode);
1177
1178         return res;
1179 }
1180
1181 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1182 {
1183         __free_page(req->pages[0]);
1184         fuse_file_put(req->ff, false);
1185 }
1186
1187 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1188 {
1189         struct inode *inode = req->inode;
1190         struct fuse_inode *fi = get_fuse_inode(inode);
1191         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1192
1193         list_del(&req->writepages_entry);
1194         dec_bdi_stat(bdi, BDI_WRITEBACK);
1195         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1196         bdi_writeout_inc(bdi);
1197         wake_up(&fi->page_waitq);
1198 }
1199
1200 /* Called under fc->lock, may release and reacquire it */
1201 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1202 __releases(fc->lock)
1203 __acquires(fc->lock)
1204 {
1205         struct fuse_inode *fi = get_fuse_inode(req->inode);
1206         loff_t size = i_size_read(req->inode);
1207         struct fuse_write_in *inarg = &req->misc.write.in;
1208
1209         if (!fc->connected)
1210                 goto out_free;
1211
1212         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1213                 inarg->size = PAGE_CACHE_SIZE;
1214         } else if (inarg->offset < size) {
1215                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1216         } else {
1217                 /* Got truncated off completely */
1218                 goto out_free;
1219         }
1220
1221         req->in.args[1].size = inarg->size;
1222         fi->writectr++;
1223         fuse_request_send_background_locked(fc, req);
1224         return;
1225
1226  out_free:
1227         fuse_writepage_finish(fc, req);
1228         spin_unlock(&fc->lock);
1229         fuse_writepage_free(fc, req);
1230         fuse_put_request(fc, req);
1231         spin_lock(&fc->lock);
1232 }
1233
1234 /*
1235  * If fi->writectr is positive (no truncate or fsync going on) send
1236  * all queued writepage requests.
1237  *
1238  * Called with fc->lock
1239  */
1240 void fuse_flush_writepages(struct inode *inode)
1241 __releases(fc->lock)
1242 __acquires(fc->lock)
1243 {
1244         struct fuse_conn *fc = get_fuse_conn(inode);
1245         struct fuse_inode *fi = get_fuse_inode(inode);
1246         struct fuse_req *req;
1247
1248         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1249                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1250                 list_del_init(&req->list);
1251                 fuse_send_writepage(fc, req);
1252         }
1253 }
1254
1255 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1256 {
1257         struct inode *inode = req->inode;
1258         struct fuse_inode *fi = get_fuse_inode(inode);
1259
1260         mapping_set_error(inode->i_mapping, req->out.h.error);
1261         spin_lock(&fc->lock);
1262         fi->writectr--;
1263         fuse_writepage_finish(fc, req);
1264         spin_unlock(&fc->lock);
1265         fuse_writepage_free(fc, req);
1266 }
1267
1268 static int fuse_writepage_locked(struct page *page)
1269 {
1270         struct address_space *mapping = page->mapping;
1271         struct inode *inode = mapping->host;
1272         struct fuse_conn *fc = get_fuse_conn(inode);
1273         struct fuse_inode *fi = get_fuse_inode(inode);
1274         struct fuse_req *req;
1275         struct fuse_file *ff;
1276         struct page *tmp_page;
1277
1278         set_page_writeback(page);
1279
1280         req = fuse_request_alloc_nofs();
1281         if (!req)
1282                 goto err;
1283
1284         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1285         if (!tmp_page)
1286                 goto err_free;
1287
1288         spin_lock(&fc->lock);
1289         BUG_ON(list_empty(&fi->write_files));
1290         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1291         req->ff = fuse_file_get(ff);
1292         spin_unlock(&fc->lock);
1293
1294         fuse_write_fill(req, ff, page_offset(page), 0);
1295
1296         copy_highpage(tmp_page, page);
1297         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1298         req->in.argpages = 1;
1299         req->num_pages = 1;
1300         req->pages[0] = tmp_page;
1301         req->page_offset = 0;
1302         req->end = fuse_writepage_end;
1303         req->inode = inode;
1304
1305         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1306         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1307         end_page_writeback(page);
1308
1309         spin_lock(&fc->lock);
1310         list_add(&req->writepages_entry, &fi->writepages);
1311         list_add_tail(&req->list, &fi->queued_writes);
1312         fuse_flush_writepages(inode);
1313         spin_unlock(&fc->lock);
1314
1315         return 0;
1316
1317 err_free:
1318         fuse_request_free(req);
1319 err:
1320         end_page_writeback(page);
1321         return -ENOMEM;
1322 }
1323
1324 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1325 {
1326         int err;
1327
1328         err = fuse_writepage_locked(page);
1329         unlock_page(page);
1330
1331         return err;
1332 }
1333
1334 static int fuse_launder_page(struct page *page)
1335 {
1336         int err = 0;
1337         if (clear_page_dirty_for_io(page)) {
1338                 struct inode *inode = page->mapping->host;
1339                 err = fuse_writepage_locked(page);
1340                 if (!err)
1341                         fuse_wait_on_page_writeback(inode, page->index);
1342         }
1343         return err;
1344 }
1345
1346 /*
1347  * Write back dirty pages now, because there may not be any suitable
1348  * open files later
1349  */
1350 static void fuse_vma_close(struct vm_area_struct *vma)
1351 {
1352         filemap_write_and_wait(vma->vm_file->f_mapping);
1353 }
1354
1355 /*
1356  * Wait for writeback against this page to complete before allowing it
1357  * to be marked dirty again, and hence written back again, possibly
1358  * before the previous writepage completed.
1359  *
1360  * Block here, instead of in ->writepage(), so that the userspace fs
1361  * can only block processes actually operating on the filesystem.
1362  *
1363  * Otherwise unprivileged userspace fs would be able to block
1364  * unrelated:
1365  *
1366  * - page migration
1367  * - sync(2)
1368  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1369  */
1370 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1371 {
1372         struct page *page = vmf->page;
1373         /*
1374          * Don't use page->mapping as it may become NULL from a
1375          * concurrent truncate.
1376          */
1377         struct inode *inode = vma->vm_file->f_mapping->host;
1378
1379         fuse_wait_on_page_writeback(inode, page->index);
1380         return 0;
1381 }
1382
1383 static const struct vm_operations_struct fuse_file_vm_ops = {
1384         .close          = fuse_vma_close,
1385         .fault          = filemap_fault,
1386         .page_mkwrite   = fuse_page_mkwrite,
1387 };
1388
1389 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1390 {
1391         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1392                 struct inode *inode = file->f_dentry->d_inode;
1393                 struct fuse_conn *fc = get_fuse_conn(inode);
1394                 struct fuse_inode *fi = get_fuse_inode(inode);
1395                 struct fuse_file *ff = file->private_data;
1396                 /*
1397                  * file may be written through mmap, so chain it onto the
1398                  * inodes's write_file list
1399                  */
1400                 spin_lock(&fc->lock);
1401                 if (list_empty(&ff->write_entry))
1402                         list_add(&ff->write_entry, &fi->write_files);
1403                 spin_unlock(&fc->lock);
1404         }
1405         file_accessed(file);
1406         vma->vm_ops = &fuse_file_vm_ops;
1407         return 0;
1408 }
1409
1410 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1411 {
1412         /* Can't provide the coherency needed for MAP_SHARED */
1413         if (vma->vm_flags & VM_MAYSHARE)
1414                 return -ENODEV;
1415
1416         invalidate_inode_pages2(file->f_mapping);
1417
1418         return generic_file_mmap(file, vma);
1419 }
1420
1421 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1422                                   struct file_lock *fl)
1423 {
1424         switch (ffl->type) {
1425         case F_UNLCK:
1426                 break;
1427
1428         case F_RDLCK:
1429         case F_WRLCK:
1430                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1431                     ffl->end < ffl->start)
1432                         return -EIO;
1433
1434                 fl->fl_start = ffl->start;
1435                 fl->fl_end = ffl->end;
1436                 fl->fl_pid = ffl->pid;
1437                 break;
1438
1439         default:
1440                 return -EIO;
1441         }
1442         fl->fl_type = ffl->type;
1443         return 0;
1444 }
1445
1446 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1447                          const struct file_lock *fl, int opcode, pid_t pid,
1448                          int flock)
1449 {
1450         struct inode *inode = file->f_path.dentry->d_inode;
1451         struct fuse_conn *fc = get_fuse_conn(inode);
1452         struct fuse_file *ff = file->private_data;
1453         struct fuse_lk_in *arg = &req->misc.lk_in;
1454
1455         arg->fh = ff->fh;
1456         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1457         arg->lk.start = fl->fl_start;
1458         arg->lk.end = fl->fl_end;
1459         arg->lk.type = fl->fl_type;
1460         arg->lk.pid = pid;
1461         if (flock)
1462                 arg->lk_flags |= FUSE_LK_FLOCK;
1463         req->in.h.opcode = opcode;
1464         req->in.h.nodeid = get_node_id(inode);
1465         req->in.numargs = 1;
1466         req->in.args[0].size = sizeof(*arg);
1467         req->in.args[0].value = arg;
1468 }
1469
1470 static int fuse_getlk(struct file *file, struct file_lock *fl)
1471 {
1472         struct inode *inode = file->f_path.dentry->d_inode;
1473         struct fuse_conn *fc = get_fuse_conn(inode);
1474         struct fuse_req *req;
1475         struct fuse_lk_out outarg;
1476         int err;
1477
1478         req = fuse_get_req(fc);
1479         if (IS_ERR(req))
1480                 return PTR_ERR(req);
1481
1482         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1483         req->out.numargs = 1;
1484         req->out.args[0].size = sizeof(outarg);
1485         req->out.args[0].value = &outarg;
1486         fuse_request_send(fc, req);
1487         err = req->out.h.error;
1488         fuse_put_request(fc, req);
1489         if (!err)
1490                 err = convert_fuse_file_lock(&outarg.lk, fl);
1491
1492         return err;
1493 }
1494
1495 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1496 {
1497         struct inode *inode = file->f_path.dentry->d_inode;
1498         struct fuse_conn *fc = get_fuse_conn(inode);
1499         struct fuse_req *req;
1500         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1501         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1502         int err;
1503
1504         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1505                 /* NLM needs asynchronous locks, which we don't support yet */
1506                 return -ENOLCK;
1507         }
1508
1509         /* Unlock on close is handled by the flush method */
1510         if (fl->fl_flags & FL_CLOSE)
1511                 return 0;
1512
1513         req = fuse_get_req(fc);
1514         if (IS_ERR(req))
1515                 return PTR_ERR(req);
1516
1517         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1518         fuse_request_send(fc, req);
1519         err = req->out.h.error;
1520         /* locking is restartable */
1521         if (err == -EINTR)
1522                 err = -ERESTARTSYS;
1523         fuse_put_request(fc, req);
1524         return err;
1525 }
1526
1527 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1528 {
1529         struct inode *inode = file->f_path.dentry->d_inode;
1530         struct fuse_conn *fc = get_fuse_conn(inode);
1531         int err;
1532
1533         if (cmd == F_CANCELLK) {
1534                 err = 0;
1535         } else if (cmd == F_GETLK) {
1536                 if (fc->no_lock) {
1537                         posix_test_lock(file, fl);
1538                         err = 0;
1539                 } else
1540                         err = fuse_getlk(file, fl);
1541         } else {
1542                 if (fc->no_lock)
1543                         err = posix_lock_file(file, fl, NULL);
1544                 else
1545                         err = fuse_setlk(file, fl, 0);
1546         }
1547         return err;
1548 }
1549
1550 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1551 {
1552         struct inode *inode = file->f_path.dentry->d_inode;
1553         struct fuse_conn *fc = get_fuse_conn(inode);
1554         int err;
1555
1556         if (fc->no_flock) {
1557                 err = flock_lock_file_wait(file, fl);
1558         } else {
1559                 struct fuse_file *ff = file->private_data;
1560
1561                 /* emulate flock with POSIX locks */
1562                 fl->fl_owner = (fl_owner_t) file;
1563                 ff->flock = true;
1564                 err = fuse_setlk(file, fl, 1);
1565         }
1566
1567         return err;
1568 }
1569
1570 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1571 {
1572         struct inode *inode = mapping->host;
1573         struct fuse_conn *fc = get_fuse_conn(inode);
1574         struct fuse_req *req;
1575         struct fuse_bmap_in inarg;
1576         struct fuse_bmap_out outarg;
1577         int err;
1578
1579         if (!inode->i_sb->s_bdev || fc->no_bmap)
1580                 return 0;
1581
1582         req = fuse_get_req(fc);
1583         if (IS_ERR(req))
1584                 return 0;
1585
1586         memset(&inarg, 0, sizeof(inarg));
1587         inarg.block = block;
1588         inarg.blocksize = inode->i_sb->s_blocksize;
1589         req->in.h.opcode = FUSE_BMAP;
1590         req->in.h.nodeid = get_node_id(inode);
1591         req->in.numargs = 1;
1592         req->in.args[0].size = sizeof(inarg);
1593         req->in.args[0].value = &inarg;
1594         req->out.numargs = 1;
1595         req->out.args[0].size = sizeof(outarg);
1596         req->out.args[0].value = &outarg;
1597         fuse_request_send(fc, req);
1598         err = req->out.h.error;
1599         fuse_put_request(fc, req);
1600         if (err == -ENOSYS)
1601                 fc->no_bmap = 1;
1602
1603         return err ? 0 : outarg.block;
1604 }
1605
1606 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1607 {
1608         loff_t retval;
1609         struct inode *inode = file->f_path.dentry->d_inode;
1610
1611         mutex_lock(&inode->i_mutex);
1612         switch (origin) {
1613         case SEEK_END:
1614                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1615                 if (retval)
1616                         goto exit;
1617                 offset += i_size_read(inode);
1618                 break;
1619         case SEEK_CUR:
1620                 offset += file->f_pos;
1621         }
1622         retval = -EINVAL;
1623         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1624                 if (offset != file->f_pos) {
1625                         file->f_pos = offset;
1626                         file->f_version = 0;
1627                 }
1628                 retval = offset;
1629         }
1630 exit:
1631         mutex_unlock(&inode->i_mutex);
1632         return retval;
1633 }
1634
1635 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1636                         unsigned int nr_segs, size_t bytes, bool to_user)
1637 {
1638         struct iov_iter ii;
1639         int page_idx = 0;
1640
1641         if (!bytes)
1642                 return 0;
1643
1644         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1645
1646         while (iov_iter_count(&ii)) {
1647                 struct page *page = pages[page_idx++];
1648                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1649                 void *kaddr;
1650
1651                 kaddr = kmap(page);
1652
1653                 while (todo) {
1654                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1655                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1656                         size_t copy = min(todo, iov_len);
1657                         size_t left;
1658
1659                         if (!to_user)
1660                                 left = copy_from_user(kaddr, uaddr, copy);
1661                         else
1662                                 left = copy_to_user(uaddr, kaddr, copy);
1663
1664                         if (unlikely(left))
1665                                 return -EFAULT;
1666
1667                         iov_iter_advance(&ii, copy);
1668                         todo -= copy;
1669                         kaddr += copy;
1670                 }
1671
1672                 kunmap(page);
1673         }
1674
1675         return 0;
1676 }
1677
1678 /*
1679  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1680  * ABI was defined to be 'struct iovec' which is different on 32bit
1681  * and 64bit.  Fortunately we can determine which structure the server
1682  * used from the size of the reply.
1683  */
1684 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1685                                      size_t transferred, unsigned count,
1686                                      bool is_compat)
1687 {
1688 #ifdef CONFIG_COMPAT
1689         if (count * sizeof(struct compat_iovec) == transferred) {
1690                 struct compat_iovec *ciov = src;
1691                 unsigned i;
1692
1693                 /*
1694                  * With this interface a 32bit server cannot support
1695                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1696                  * requests
1697                  */
1698                 if (!is_compat)
1699                         return -EINVAL;
1700
1701                 for (i = 0; i < count; i++) {
1702                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1703                         dst[i].iov_len = ciov[i].iov_len;
1704                 }
1705                 return 0;
1706         }
1707 #endif
1708
1709         if (count * sizeof(struct iovec) != transferred)
1710                 return -EIO;
1711
1712         memcpy(dst, src, transferred);
1713         return 0;
1714 }
1715
1716 /* Make sure iov_length() won't overflow */
1717 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1718 {
1719         size_t n;
1720         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1721
1722         for (n = 0; n < count; n++) {
1723                 if (iov->iov_len > (size_t) max)
1724                         return -ENOMEM;
1725                 max -= iov->iov_len;
1726         }
1727         return 0;
1728 }
1729
1730 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1731                                  void *src, size_t transferred, unsigned count,
1732                                  bool is_compat)
1733 {
1734         unsigned i;
1735         struct fuse_ioctl_iovec *fiov = src;
1736
1737         if (fc->minor < 16) {
1738                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1739                                                  count, is_compat);
1740         }
1741
1742         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1743                 return -EIO;
1744
1745         for (i = 0; i < count; i++) {
1746                 /* Did the server supply an inappropriate value? */
1747                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1748                     fiov[i].len != (unsigned long) fiov[i].len)
1749                         return -EIO;
1750
1751                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1752                 dst[i].iov_len = (size_t) fiov[i].len;
1753
1754 #ifdef CONFIG_COMPAT
1755                 if (is_compat &&
1756                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1757                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1758                         return -EIO;
1759 #endif
1760         }
1761
1762         return 0;
1763 }
1764
1765
1766 /*
1767  * For ioctls, there is no generic way to determine how much memory
1768  * needs to be read and/or written.  Furthermore, ioctls are allowed
1769  * to dereference the passed pointer, so the parameter requires deep
1770  * copying but FUSE has no idea whatsoever about what to copy in or
1771  * out.
1772  *
1773  * This is solved by allowing FUSE server to retry ioctl with
1774  * necessary in/out iovecs.  Let's assume the ioctl implementation
1775  * needs to read in the following structure.
1776  *
1777  * struct a {
1778  *      char    *buf;
1779  *      size_t  buflen;
1780  * }
1781  *
1782  * On the first callout to FUSE server, inarg->in_size and
1783  * inarg->out_size will be NULL; then, the server completes the ioctl
1784  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1785  * the actual iov array to
1786  *
1787  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1788  *
1789  * which tells FUSE to copy in the requested area and retry the ioctl.
1790  * On the second round, the server has access to the structure and
1791  * from that it can tell what to look for next, so on the invocation,
1792  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1793  *
1794  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1795  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1796  *
1797  * FUSE will copy both struct a and the pointed buffer from the
1798  * process doing the ioctl and retry ioctl with both struct a and the
1799  * buffer.
1800  *
1801  * This time, FUSE server has everything it needs and completes ioctl
1802  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1803  *
1804  * Copying data out works the same way.
1805  *
1806  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1807  * automatically initializes in and out iovs by decoding @cmd with
1808  * _IOC_* macros and the server is not allowed to request RETRY.  This
1809  * limits ioctl data transfers to well-formed ioctls and is the forced
1810  * behavior for all FUSE servers.
1811  */
1812 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1813                    unsigned int flags)
1814 {
1815         struct fuse_file *ff = file->private_data;
1816         struct fuse_conn *fc = ff->fc;
1817         struct fuse_ioctl_in inarg = {
1818                 .fh = ff->fh,
1819                 .cmd = cmd,
1820                 .arg = arg,
1821                 .flags = flags
1822         };
1823         struct fuse_ioctl_out outarg;
1824         struct fuse_req *req = NULL;
1825         struct page **pages = NULL;
1826         struct iovec *iov_page = NULL;
1827         struct iovec *in_iov = NULL, *out_iov = NULL;
1828         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1829         size_t in_size, out_size, transferred;
1830         int err;
1831
1832 #if BITS_PER_LONG == 32
1833         inarg.flags |= FUSE_IOCTL_32BIT;
1834 #else
1835         if (flags & FUSE_IOCTL_COMPAT)
1836                 inarg.flags |= FUSE_IOCTL_32BIT;
1837 #endif
1838
1839         /* assume all the iovs returned by client always fits in a page */
1840         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1841
1842         err = -ENOMEM;
1843         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1844         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1845         if (!pages || !iov_page)
1846                 goto out;
1847
1848         /*
1849          * If restricted, initialize IO parameters as encoded in @cmd.
1850          * RETRY from server is not allowed.
1851          */
1852         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1853                 struct iovec *iov = iov_page;
1854
1855                 iov->iov_base = (void __user *)arg;
1856                 iov->iov_len = _IOC_SIZE(cmd);
1857
1858                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1859                         in_iov = iov;
1860                         in_iovs = 1;
1861                 }
1862
1863                 if (_IOC_DIR(cmd) & _IOC_READ) {
1864                         out_iov = iov;
1865                         out_iovs = 1;
1866                 }
1867         }
1868
1869  retry:
1870         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1871         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1872
1873         /*
1874          * Out data can be used either for actual out data or iovs,
1875          * make sure there always is at least one page.
1876          */
1877         out_size = max_t(size_t, out_size, PAGE_SIZE);
1878         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1879
1880         /* make sure there are enough buffer pages and init request with them */
1881         err = -ENOMEM;
1882         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1883                 goto out;
1884         while (num_pages < max_pages) {
1885                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1886                 if (!pages[num_pages])
1887                         goto out;
1888                 num_pages++;
1889         }
1890
1891         req = fuse_get_req(fc);
1892         if (IS_ERR(req)) {
1893                 err = PTR_ERR(req);
1894                 req = NULL;
1895                 goto out;
1896         }
1897         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1898         req->num_pages = num_pages;
1899
1900         /* okay, let's send it to the client */
1901         req->in.h.opcode = FUSE_IOCTL;
1902         req->in.h.nodeid = ff->nodeid;
1903         req->in.numargs = 1;
1904         req->in.args[0].size = sizeof(inarg);
1905         req->in.args[0].value = &inarg;
1906         if (in_size) {
1907                 req->in.numargs++;
1908                 req->in.args[1].size = in_size;
1909                 req->in.argpages = 1;
1910
1911                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1912                                            false);
1913                 if (err)
1914                         goto out;
1915         }
1916
1917         req->out.numargs = 2;
1918         req->out.args[0].size = sizeof(outarg);
1919         req->out.args[0].value = &outarg;
1920         req->out.args[1].size = out_size;
1921         req->out.argpages = 1;
1922         req->out.argvar = 1;
1923
1924         fuse_request_send(fc, req);
1925         err = req->out.h.error;
1926         transferred = req->out.args[1].size;
1927         fuse_put_request(fc, req);
1928         req = NULL;
1929         if (err)
1930                 goto out;
1931
1932         /* did it ask for retry? */
1933         if (outarg.flags & FUSE_IOCTL_RETRY) {
1934                 void *vaddr;
1935
1936                 /* no retry if in restricted mode */
1937                 err = -EIO;
1938                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1939                         goto out;
1940
1941                 in_iovs = outarg.in_iovs;
1942                 out_iovs = outarg.out_iovs;
1943
1944                 /*
1945                  * Make sure things are in boundary, separate checks
1946                  * are to protect against overflow.
1947                  */
1948                 err = -ENOMEM;
1949                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1950                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1951                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1952                         goto out;
1953
1954                 vaddr = kmap_atomic(pages[0], KM_USER0);
1955                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1956                                             transferred, in_iovs + out_iovs,
1957                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1958                 kunmap_atomic(vaddr, KM_USER0);
1959                 if (err)
1960                         goto out;
1961
1962                 in_iov = iov_page;
1963                 out_iov = in_iov + in_iovs;
1964
1965                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1966                 if (err)
1967                         goto out;
1968
1969                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1970                 if (err)
1971                         goto out;
1972
1973                 goto retry;
1974         }
1975
1976         err = -EIO;
1977         if (transferred > inarg.out_size)
1978                 goto out;
1979
1980         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1981  out:
1982         if (req)
1983                 fuse_put_request(fc, req);
1984         free_page((unsigned long) iov_page);
1985         while (num_pages)
1986                 __free_page(pages[--num_pages]);
1987         kfree(pages);
1988
1989         return err ? err : outarg.result;
1990 }
1991 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1992
1993 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1994                                    unsigned long arg, unsigned int flags)
1995 {
1996         struct inode *inode = file->f_dentry->d_inode;
1997         struct fuse_conn *fc = get_fuse_conn(inode);
1998
1999         if (!fuse_allow_task(fc, current))
2000                 return -EACCES;
2001
2002         if (is_bad_inode(inode))
2003                 return -EIO;
2004
2005         return fuse_do_ioctl(file, cmd, arg, flags);
2006 }
2007
2008 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2009                             unsigned long arg)
2010 {
2011         return fuse_file_ioctl_common(file, cmd, arg, 0);
2012 }
2013
2014 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2015                                    unsigned long arg)
2016 {
2017         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2018 }
2019
2020 /*
2021  * All files which have been polled are linked to RB tree
2022  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2023  * find the matching one.
2024  */
2025 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2026                                               struct rb_node **parent_out)
2027 {
2028         struct rb_node **link = &fc->polled_files.rb_node;
2029         struct rb_node *last = NULL;
2030
2031         while (*link) {
2032                 struct fuse_file *ff;
2033
2034                 last = *link;
2035                 ff = rb_entry(last, struct fuse_file, polled_node);
2036
2037                 if (kh < ff->kh)
2038                         link = &last->rb_left;
2039                 else if (kh > ff->kh)
2040                         link = &last->rb_right;
2041                 else
2042                         return link;
2043         }
2044
2045         if (parent_out)
2046                 *parent_out = last;
2047         return link;
2048 }
2049
2050 /*
2051  * The file is about to be polled.  Make sure it's on the polled_files
2052  * RB tree.  Note that files once added to the polled_files tree are
2053  * not removed before the file is released.  This is because a file
2054  * polled once is likely to be polled again.
2055  */
2056 static void fuse_register_polled_file(struct fuse_conn *fc,
2057                                       struct fuse_file *ff)
2058 {
2059         spin_lock(&fc->lock);
2060         if (RB_EMPTY_NODE(&ff->polled_node)) {
2061                 struct rb_node **link, *parent;
2062
2063                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2064                 BUG_ON(*link);
2065                 rb_link_node(&ff->polled_node, parent, link);
2066                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2067         }
2068         spin_unlock(&fc->lock);
2069 }
2070
2071 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2072 {
2073         struct fuse_file *ff = file->private_data;
2074         struct fuse_conn *fc = ff->fc;
2075         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2076         struct fuse_poll_out outarg;
2077         struct fuse_req *req;
2078         int err;
2079
2080         if (fc->no_poll)
2081                 return DEFAULT_POLLMASK;
2082
2083         poll_wait(file, &ff->poll_wait, wait);
2084
2085         /*
2086          * Ask for notification iff there's someone waiting for it.
2087          * The client may ignore the flag and always notify.
2088          */
2089         if (waitqueue_active(&ff->poll_wait)) {
2090                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2091                 fuse_register_polled_file(fc, ff);
2092         }
2093
2094         req = fuse_get_req(fc);
2095         if (IS_ERR(req))
2096                 return POLLERR;
2097
2098         req->in.h.opcode = FUSE_POLL;
2099         req->in.h.nodeid = ff->nodeid;
2100         req->in.numargs = 1;
2101         req->in.args[0].size = sizeof(inarg);
2102         req->in.args[0].value = &inarg;
2103         req->out.numargs = 1;
2104         req->out.args[0].size = sizeof(outarg);
2105         req->out.args[0].value = &outarg;
2106         fuse_request_send(fc, req);
2107         err = req->out.h.error;
2108         fuse_put_request(fc, req);
2109
2110         if (!err)
2111                 return outarg.revents;
2112         if (err == -ENOSYS) {
2113                 fc->no_poll = 1;
2114                 return DEFAULT_POLLMASK;
2115         }
2116         return POLLERR;
2117 }
2118 EXPORT_SYMBOL_GPL(fuse_file_poll);
2119
2120 /*
2121  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2122  * wakes up the poll waiters.
2123  */
2124 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2125                             struct fuse_notify_poll_wakeup_out *outarg)
2126 {
2127         u64 kh = outarg->kh;
2128         struct rb_node **link;
2129
2130         spin_lock(&fc->lock);
2131
2132         link = fuse_find_polled_node(fc, kh, NULL);
2133         if (*link) {
2134                 struct fuse_file *ff;
2135
2136                 ff = rb_entry(*link, struct fuse_file, polled_node);
2137                 wake_up_interruptible_sync(&ff->poll_wait);
2138         }
2139
2140         spin_unlock(&fc->lock);
2141         return 0;
2142 }
2143
2144 static const struct file_operations fuse_file_operations = {
2145         .llseek         = fuse_file_llseek,
2146         .read           = do_sync_read,
2147         .aio_read       = fuse_file_aio_read,
2148         .write          = do_sync_write,
2149         .aio_write      = fuse_file_aio_write,
2150         .mmap           = fuse_file_mmap,
2151         .open           = fuse_open,
2152         .flush          = fuse_flush,
2153         .release        = fuse_release,
2154         .fsync          = fuse_fsync,
2155         .lock           = fuse_file_lock,
2156         .flock          = fuse_file_flock,
2157         .splice_read    = generic_file_splice_read,
2158         .unlocked_ioctl = fuse_file_ioctl,
2159         .compat_ioctl   = fuse_file_compat_ioctl,
2160         .poll           = fuse_file_poll,
2161 };
2162
2163 static const struct file_operations fuse_direct_io_file_operations = {
2164         .llseek         = fuse_file_llseek,
2165         .read           = fuse_direct_read,
2166         .write          = fuse_direct_write,
2167         .mmap           = fuse_direct_mmap,
2168         .open           = fuse_open,
2169         .flush          = fuse_flush,
2170         .release        = fuse_release,
2171         .fsync          = fuse_fsync,
2172         .lock           = fuse_file_lock,
2173         .flock          = fuse_file_flock,
2174         .unlocked_ioctl = fuse_file_ioctl,
2175         .compat_ioctl   = fuse_file_compat_ioctl,
2176         .poll           = fuse_file_poll,
2177         /* no splice_read */
2178 };
2179
2180 static const struct address_space_operations fuse_file_aops  = {
2181         .readpage       = fuse_readpage,
2182         .writepage      = fuse_writepage,
2183         .launder_page   = fuse_launder_page,
2184         .write_begin    = fuse_write_begin,
2185         .write_end      = fuse_write_end,
2186         .readpages      = fuse_readpages,
2187         .set_page_dirty = __set_page_dirty_nobuffers,
2188         .bmap           = fuse_bmap,
2189 };
2190
2191 void fuse_init_file_inode(struct inode *inode)
2192 {
2193         inode->i_fop = &fuse_file_operations;
2194         inode->i_data.a_ops = &fuse_file_aops;
2195 }