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
  19 static const struct file_operations fuse_direct_io_file_operations;
  20
  21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  22                           int opcode, struct fuse_open_out *outargp)
  23 {
  24         struct fuse_open_in inarg;
  25         struct fuse_req *req;
  26         int err;
  27
  28         req = fuse_get_req(fc);
  29         if (IS_ERR(req))
  30                 return PTR_ERR(req);
  31
  32         memset(&inarg, 0, sizeof(inarg));
  33         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  34         if (!fc->atomic_o_trunc)
  35                 inarg.flags &= ~O_TRUNC;
  36         req->in.h.opcode = opcode;
  37         req->in.h.nodeid = nodeid;
  38         req->in.numargs = 1;
  39         req->in.args[0].size = sizeof(inarg);
  40         req->in.args[0].value = &inarg;
  41         req->out.numargs = 1;
  42         req->out.args[0].size = sizeof(*outargp);
  43         req->out.args[0].value = outargp;
  44         fuse_request_send(fc, req);
  45         err = req->out.h.error;
  46         fuse_put_request(fc, req);
  47
  48         return err;
  49 }
  50
  51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  52 {
  53         struct fuse_file *ff;
  54
  55         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  56         if (unlikely(!ff))
  57                 return NULL;
  58
  59         ff->fc = fc;
  60         ff->reserved_req = fuse_request_alloc();
  61         if (unlikely(!ff->reserved_req)) {
  62                 kfree(ff);
  63                 return NULL;
  64         }
  65
  66         INIT_LIST_HEAD(&ff->write_entry);
  67         atomic_set(&ff->count, 0);
  68         RB_CLEAR_NODE(&ff->polled_node);
  69         init_waitqueue_head(&ff->poll_wait);
  70
  71         spin_lock(&fc->lock);
  72         ff->kh = ++fc->khctr;
  73         spin_unlock(&fc->lock);
  74
  75         return ff;
  76 }
  77
  78 void fuse_file_free(struct fuse_file *ff)
  79 {
  80         fuse_request_free(ff->reserved_req);
  81         kfree(ff);
  82 }
  83
  84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  85 {
  86         atomic_inc(&ff->count);
  87         return ff;
  88 }
  89
  90 static void fuse_release_async(struct work_struct *work)
  91 {
  92         struct fuse_req *req;
  93         struct fuse_conn *fc;
  94         struct path path;
  95
  96         req = container_of(work, struct fuse_req, misc.release.work);
  97         path = req->misc.release.path;
  98         fc = get_fuse_conn(path.dentry->d_inode);
  99
 100         fuse_put_request(fc, req);
 101         path_put(&path);
 102 }
 103
 104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 105 {
 106         if (fc->destroy_req) {
 107                 /*
 108                  * If this is a fuseblk mount, then it's possible that
 109                  * releasing the path will result in releasing the
 110                  * super block and sending the DESTROY request.  If
 111                  * the server is single threaded, this would hang.
 112                  * For this reason do the path_put() in a separate
 113                  * thread.
 114                  */
 115                 atomic_inc(&req->count);
 116                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 117                 schedule_work(&req->misc.release.work);
 118         } else {
 119                 path_put(&req->misc.release.path);
 120         }
 121 }
 122
 123 static void fuse_file_put(struct fuse_file *ff, bool sync)
 124 {
 125         if (atomic_dec_and_test(&ff->count)) {
 126                 struct fuse_req *req = ff->reserved_req;
 127
 128                 if (sync) {
 129                         fuse_request_send(ff->fc, req);
 130                         path_put(&req->misc.release.path);
 131                         fuse_put_request(ff->fc, req);
 132                 } else {
 133                         req->end = fuse_release_end;
 134                         fuse_request_send_background(ff->fc, req);
 135                 }
 136                 kfree(ff);
 137         }
 138 }
 139
 140 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 141                  bool isdir)
 142 {
 143         struct fuse_open_out outarg;
 144         struct fuse_file *ff;
 145         int err;
 146         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 147
 148         ff = fuse_file_alloc(fc);
 149         if (!ff)
 150                 return -ENOMEM;
 151
 152         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 153         if (err) {
 154                 fuse_file_free(ff);
 155                 return err;
 156         }
 157
 158         if (isdir)
 159                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 160
 161         ff->fh = outarg.fh;
 162         ff->nodeid = nodeid;
 163         ff->open_flags = outarg.open_flags;
 164         file->private_data = fuse_file_get(ff);
 165
 166         return 0;
 167 }
 168 EXPORT_SYMBOL_GPL(fuse_do_open);
 169
 170 void fuse_finish_open(struct inode *inode, struct file *file)
 171 {
 172         struct fuse_file *ff = file->private_data;
 173         struct fuse_conn *fc = get_fuse_conn(inode);
 174
 175         if (ff->open_flags & FOPEN_DIRECT_IO)
 176                 file->f_op = &fuse_direct_io_file_operations;
 177         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 178                 invalidate_inode_pages2(inode->i_mapping);
 179         if (ff->open_flags & FOPEN_NONSEEKABLE)
 180                 nonseekable_open(inode, file);
 181         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 182                 struct fuse_inode *fi = get_fuse_inode(inode);
 183
 184                 spin_lock(&fc->lock);
 185                 fi->attr_version = ++fc->attr_version;
 186                 i_size_write(inode, 0);
 187                 spin_unlock(&fc->lock);
 188                 fuse_invalidate_attr(inode);
 189         }
 190 }
 191
 192 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 193 {
 194         struct fuse_conn *fc = get_fuse_conn(inode);
 195         int err;
 196
 197         err = generic_file_open(inode, file);
 198         if (err)
 199                 return err;
 200
 201         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 202         if (err)
 203                 return err;
 204
 205         fuse_finish_open(inode, file);
 206
 207         return 0;
 208 }
 209
 210 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 211 {
 212         struct fuse_conn *fc = ff->fc;
 213         struct fuse_req *req = ff->reserved_req;
 214         struct fuse_release_in *inarg = &req->misc.release.in;
 215
 216         spin_lock(&fc->lock);
 217         list_del(&ff->write_entry);
 218         if (!RB_EMPTY_NODE(&ff->polled_node))
 219                 rb_erase(&ff->polled_node, &fc->polled_files);
 220         spin_unlock(&fc->lock);
 221
 222         wake_up_interruptible_all(&ff->poll_wait);
 223
 224         inarg->fh = ff->fh;
 225         inarg->flags = flags;
 226         req->in.h.opcode = opcode;
 227         req->in.h.nodeid = ff->nodeid;
 228         req->in.numargs = 1;
 229         req->in.args[0].size = sizeof(struct fuse_release_in);
 230         req->in.args[0].value = inarg;
 231 }
 232
 233 void fuse_release_common(struct file *file, int opcode)
 234 {
 235         struct fuse_file *ff;
 236         struct fuse_req *req;
 237
 238         ff = file->private_data;
 239         if (unlikely(!ff))
 240                 return;
 241
 242         req = ff->reserved_req;
 243         fuse_prepare_release(ff, file->f_flags, opcode);
 244
 245         if (ff->flock) {
 246                 struct fuse_release_in *inarg = &req->misc.release.in;
 247                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 248                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 249                                                        (fl_owner_t) file);
 250         }
 251         /* Hold vfsmount and dentry until release is finished */
 252         path_get(&file->f_path);
 253         req->misc.release.path = file->f_path;
 254
 255         /*
 256          * Normally this will send the RELEASE request, however if
 257          * some asynchronous READ or WRITE requests are outstanding,
 258          * the sending will be delayed.
 259          *
 260          * Make the release synchronous if this is a fuseblk mount,
 261          * synchronous RELEASE is allowed (and desirable) in this case
 262          * because the server can be trusted not to screw up.
 263          */
 264         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 265 }
 266
 267 static int fuse_open(struct inode *inode, struct file *file)
 268 {
 269         return fuse_open_common(inode, file, false);
 270 }
 271
 272 static int fuse_release(struct inode *inode, struct file *file)
 273 {
 274         fuse_release_common(file, FUSE_RELEASE);
 275
 276         /* return value is ignored by VFS */
 277         return 0;
 278 }
 279
 280 void fuse_sync_release(struct fuse_file *ff, int flags)
 281 {
 282         WARN_ON(atomic_read(&ff->count) > 1);
 283         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 284         ff->reserved_req->force = 1;
 285         fuse_request_send(ff->fc, ff->reserved_req);
 286         fuse_put_request(ff->fc, ff->reserved_req);
 287         kfree(ff);
 288 }
 289 EXPORT_SYMBOL_GPL(fuse_sync_release);
 290
 291 /*
 292  * Scramble the ID space with XTEA, so that the value of the files_struct
 293  * pointer is not exposed to userspace.
 294  */
 295 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 296 {
 297         u32 *k = fc->scramble_key;
 298         u64 v = (unsigned long) id;
 299         u32 v0 = v;
 300         u32 v1 = v >> 32;
 301         u32 sum = 0;
 302         int i;
 303
 304         for (i = 0; i < 32; i++) {
 305                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 306                 sum += 0x9E3779B9;
 307                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 308         }
 309
 310         return (u64) v0 + ((u64) v1 << 32);
 311 }
 312
 313 /*
 314  * Check if page is under writeback
 315  *
 316  * This is currently done by walking the list of writepage requests
 317  * for the inode, which can be pretty inefficient.
 318  */
 319 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 320 {
 321         struct fuse_conn *fc = get_fuse_conn(inode);
 322         struct fuse_inode *fi = get_fuse_inode(inode);
 323         struct fuse_req *req;
 324         bool found = false;
 325
 326         spin_lock(&fc->lock);
 327         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 328                 pgoff_t curr_index;
 329
 330                 BUG_ON(req->inode != inode);
 331                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 332                 if (curr_index == index) {
 333                         found = true;
 334                         break;
 335                 }
 336         }
 337         spin_unlock(&fc->lock);
 338
 339         return found;
 340 }
 341
 342 /*
 343  * Wait for page writeback to be completed.
 344  *
 345  * Since fuse doesn't rely on the VM writeback tracking, this has to
 346  * use some other means.
 347  */
 348 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 349 {
 350         struct fuse_inode *fi = get_fuse_inode(inode);
 351
 352         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 353         return 0;
 354 }
 355
 356 static int fuse_flush(struct file *file, fl_owner_t id)
 357 {
 358         struct inode *inode = file->f_path.dentry->d_inode;
 359         struct fuse_conn *fc = get_fuse_conn(inode);
 360         struct fuse_file *ff = file->private_data;
 361         struct fuse_req *req;
 362         struct fuse_flush_in inarg;
 363         int err;
 364
 365         if (is_bad_inode(inode))
 366                 return -EIO;
 367
 368         if (fc->no_flush)
 369                 return 0;
 370
 371         req = fuse_get_req_nofail(fc, file);
 372         memset(&inarg, 0, sizeof(inarg));
 373         inarg.fh = ff->fh;
 374         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 375         req->in.h.opcode = FUSE_FLUSH;
 376         req->in.h.nodeid = get_node_id(inode);
 377         req->in.numargs = 1;
 378         req->in.args[0].size = sizeof(inarg);
 379         req->in.args[0].value = &inarg;
 380         req->force = 1;
 381         fuse_request_send(fc, req);
 382         err = req->out.h.error;
 383         fuse_put_request(fc, req);
 384         if (err == -ENOSYS) {
 385                 fc->no_flush = 1;
 386                 err = 0;
 387         }
 388         return err;
 389 }
 390
 391 /*
 392  * Wait for all pending writepages on the inode to finish.
 393  *
 394  * This is currently done by blocking further writes with FUSE_NOWRITE
 395  * and waiting for all sent writes to complete.
 396  *
 397  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 398  * could conflict with truncation.
 399  */
 400 static void fuse_sync_writes(struct inode *inode)
 401 {
 402         fuse_set_nowrite(inode);
 403         fuse_release_nowrite(inode);
 404 }
 405
 406 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 407                       int datasync, int isdir)
 408 {
 409         struct inode *inode = file->f_mapping->host;
 410         struct fuse_conn *fc = get_fuse_conn(inode);
 411         struct fuse_file *ff = file->private_data;
 412         struct fuse_req *req;
 413         struct fuse_fsync_in inarg;
 414         int err;
 415
 416         if (is_bad_inode(inode))
 417                 return -EIO;
 418
 419         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 420         if (err)
 421                 return err;
 422
 423         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 424                 return 0;
 425
 426         mutex_lock(&inode->i_mutex);
 427
 428         /*
 429          * Start writeback against all dirty pages of the inode, then
 430          * wait for all outstanding writes, before sending the FSYNC
 431          * request.
 432          */
 433         err = write_inode_now(inode, 0);
 434         if (err)
 435                 goto out;
 436
 437         fuse_sync_writes(inode);
 438
 439         req = fuse_get_req(fc);
 440         if (IS_ERR(req)) {
 441                 err = PTR_ERR(req);
 442                 goto out;
 443         }
 444
 445         memset(&inarg, 0, sizeof(inarg));
 446         inarg.fh = ff->fh;
 447         inarg.fsync_flags = datasync ? 1 : 0;
 448         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 449         req->in.h.nodeid = get_node_id(inode);
 450         req->in.numargs = 1;
 451         req->in.args[0].size = sizeof(inarg);
 452         req->in.args[0].value = &inarg;
 453         fuse_request_send(fc, req);
 454         err = req->out.h.error;
 455         fuse_put_request(fc, req);
 456         if (err == -ENOSYS) {
 457                 if (isdir)
 458                         fc->no_fsyncdir = 1;
 459                 else
 460                         fc->no_fsync = 1;
 461                 err = 0;
 462         }
 463 out:
 464         mutex_unlock(&inode->i_mutex);
 465         return err;
 466 }
 467
 468 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 469                       int datasync)
 470 {
 471         return fuse_fsync_common(file, start, end, datasync, 0);
 472 }
 473
 474 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 475                     size_t count, int opcode)
 476 {
 477         struct fuse_read_in *inarg = &req->misc.read.in;
 478         struct fuse_file *ff = file->private_data;
 479
 480         inarg->fh = ff->fh;
 481         inarg->offset = pos;
 482         inarg->size = count;
 483         inarg->flags = file->f_flags;
 484         req->in.h.opcode = opcode;
 485         req->in.h.nodeid = ff->nodeid;
 486         req->in.numargs = 1;
 487         req->in.args[0].size = sizeof(struct fuse_read_in);
 488         req->in.args[0].value = inarg;
 489         req->out.argvar = 1;
 490         req->out.numargs = 1;
 491         req->out.args[0].size = count;
 492 }
 493
 494 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 495                              loff_t pos, size_t count, fl_owner_t owner)
 496 {
 497         struct fuse_file *ff = file->private_data;
 498         struct fuse_conn *fc = ff->fc;
 499
 500         fuse_read_fill(req, file, pos, count, FUSE_READ);
 501         if (owner != NULL) {
 502                 struct fuse_read_in *inarg = &req->misc.read.in;
 503
 504                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 505                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 506         }
 507         fuse_request_send(fc, req);
 508         return req->out.args[0].size;
 509 }
 510
 511 static void fuse_read_update_size(struct inode *inode, loff_t size,
 512                                   u64 attr_ver)
 513 {
 514         struct fuse_conn *fc = get_fuse_conn(inode);
 515         struct fuse_inode *fi = get_fuse_inode(inode);
 516
 517         spin_lock(&fc->lock);
 518         if (attr_ver == fi->attr_version && size < inode->i_size) {
 519                 fi->attr_version = ++fc->attr_version;
 520                 i_size_write(inode, size);
 521         }
 522         spin_unlock(&fc->lock);
 523 }
 524
 525 static int fuse_readpage(struct file *file, struct page *page)
 526 {
 527         struct inode *inode = page->mapping->host;
 528         struct fuse_conn *fc = get_fuse_conn(inode);
 529         struct fuse_req *req;
 530         size_t num_read;
 531         loff_t pos = page_offset(page);
 532         size_t count = PAGE_CACHE_SIZE;
 533         u64 attr_ver;
 534         int err;
 535
 536         err = -EIO;
 537         if (is_bad_inode(inode))
 538                 goto out;
 539
 540         /*
 541          * Page writeback can extend beyond the lifetime of the
 542          * page-cache page, so make sure we read a properly synced
 543          * page.
 544          */
 545         fuse_wait_on_page_writeback(inode, page->index);
 546
 547         req = fuse_get_req(fc);
 548         err = PTR_ERR(req);
 549         if (IS_ERR(req))
 550                 goto out;
 551
 552         attr_ver = fuse_get_attr_version(fc);
 553
 554         req->out.page_zeroing = 1;
 555         req->out.argpages = 1;
 556         req->num_pages = 1;
 557         req->pages[0] = page;
 558         num_read = fuse_send_read(req, file, pos, count, NULL);
 559         err = req->out.h.error;
 560         fuse_put_request(fc, req);
 561
 562         if (!err) {
 563                 /*
 564                  * Short read means EOF.  If file size is larger, truncate it
 565                  */
 566                 if (num_read < count)
 567                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 568
 569                 SetPageUptodate(page);
 570         }
 571
 572         fuse_invalidate_attr(inode); /* atime changed */
 573  out:
 574         unlock_page(page);
 575         return err;
 576 }
 577
 578 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 579 {
 580         int i;
 581         size_t count = req->misc.read.in.size;
 582         size_t num_read = req->out.args[0].size;
 583         struct address_space *mapping = NULL;
 584
 585         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 586                 mapping = req->pages[i]->mapping;
 587
 588         if (mapping) {
 589                 struct inode *inode = mapping->host;
 590
 591                 /*
 592                  * Short read means EOF. If file size is larger, truncate it
 593                  */
 594                 if (!req->out.h.error && num_read < count) {
 595                         loff_t pos;
 596
 597                         pos = page_offset(req->pages[0]) + num_read;
 598                         fuse_read_update_size(inode, pos,
 599                                               req->misc.read.attr_ver);
 600                 }
 601                 fuse_invalidate_attr(inode); /* atime changed */
 602         }
 603
 604         for (i = 0; i < req->num_pages; i++) {
 605                 struct page *page = req->pages[i];
 606                 if (!req->out.h.error)
 607                         SetPageUptodate(page);
 608                 else
 609                         SetPageError(page);
 610                 unlock_page(page);
 611                 page_cache_release(page);
 612         }
 613         if (req->ff)
 614                 fuse_file_put(req->ff, false);
 615 }
 616
 617 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 618 {
 619         struct fuse_file *ff = file->private_data;
 620         struct fuse_conn *fc = ff->fc;
 621         loff_t pos = page_offset(req->pages[0]);
 622         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 623
 624         req->out.argpages = 1;
 625         req->out.page_zeroing = 1;
 626         req->out.page_replace = 1;
 627         fuse_read_fill(req, file, pos, count, FUSE_READ);
 628         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 629         if (fc->async_read) {
 630                 req->ff = fuse_file_get(ff);
 631                 req->end = fuse_readpages_end;
 632                 fuse_request_send_background(fc, req);
 633         } else {
 634                 fuse_request_send(fc, req);
 635                 fuse_readpages_end(fc, req);
 636                 fuse_put_request(fc, req);
 637         }
 638 }
 639
 640 struct fuse_fill_data {
 641         struct fuse_req *req;
 642         struct file *file;
 643         struct inode *inode;
 644 };
 645
 646 static int fuse_readpages_fill(void *_data, struct page *page)
 647 {
 648         struct fuse_fill_data *data = _data;
 649         struct fuse_req *req = data->req;
 650         struct inode *inode = data->inode;
 651         struct fuse_conn *fc = get_fuse_conn(inode);
 652
 653         fuse_wait_on_page_writeback(inode, page->index);
 654
 655         if (req->num_pages &&
 656             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 657              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 658              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 659                 fuse_send_readpages(req, data->file);
 660                 data->req = req = fuse_get_req(fc);
 661                 if (IS_ERR(req)) {
 662                         unlock_page(page);
 663                         return PTR_ERR(req);
 664                 }
 665         }
 666         page_cache_get(page);
 667         req->pages[req->num_pages] = page;
 668         req->num_pages++;
 669         return 0;
 670 }
 671
 672 static int fuse_readpages(struct file *file, struct address_space *mapping,
 673                           struct list_head *pages, unsigned nr_pages)
 674 {
 675         struct inode *inode = mapping->host;
 676         struct fuse_conn *fc = get_fuse_conn(inode);
 677         struct fuse_fill_data data;
 678         int err;
 679
 680         err = -EIO;
 681         if (is_bad_inode(inode))
 682                 goto out;
 683
 684         data.file = file;
 685         data.inode = inode;
 686         data.req = fuse_get_req(fc);
 687         err = PTR_ERR(data.req);
 688         if (IS_ERR(data.req))
 689                 goto out;
 690
 691         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 692         if (!err) {
 693                 if (data.req->num_pages)
 694                         fuse_send_readpages(data.req, file);
 695                 else
 696                         fuse_put_request(fc, data.req);
 697         }
 698 out:
 699         return err;
 700 }
 701
 702 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 703                                   unsigned long nr_segs, loff_t pos)
 704 {
 705         struct inode *inode = iocb->ki_filp->f_mapping->host;
 706         struct fuse_conn *fc = get_fuse_conn(inode);
 707
 708         /*
 709          * In auto invalidate mode, always update attributes on read.
 710          * Otherwise, only update if we attempt to read past EOF (to ensure
 711          * i_size is up to date).
 712          */
 713         if (fc->auto_inval_data ||
 714             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 715                 int err;
 716                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 717                 if (err)
 718                         return err;
 719         }
 720
 721         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 722 }
 723
 724 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 725                             loff_t pos, size_t count)
 726 {
 727         struct fuse_write_in *inarg = &req->misc.write.in;
 728         struct fuse_write_out *outarg = &req->misc.write.out;
 729
 730         inarg->fh = ff->fh;
 731         inarg->offset = pos;
 732         inarg->size = count;
 733         req->in.h.opcode = FUSE_WRITE;
 734         req->in.h.nodeid = ff->nodeid;
 735         req->in.numargs = 2;
 736         if (ff->fc->minor < 9)
 737                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 738         else
 739                 req->in.args[0].size = sizeof(struct fuse_write_in);
 740         req->in.args[0].value = inarg;
 741         req->in.args[1].size = count;
 742         req->out.numargs = 1;
 743         req->out.args[0].size = sizeof(struct fuse_write_out);
 744         req->out.args[0].value = outarg;
 745 }
 746
 747 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 748                               loff_t pos, size_t count, fl_owner_t owner)
 749 {
 750         struct fuse_file *ff = file->private_data;
 751         struct fuse_conn *fc = ff->fc;
 752         struct fuse_write_in *inarg = &req->misc.write.in;
 753
 754         fuse_write_fill(req, ff, pos, count);
 755         inarg->flags = file->f_flags;
 756         if (owner != NULL) {
 757                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 758                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 759         }
 760         fuse_request_send(fc, req);
 761         return req->misc.write.out.size;
 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 size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 777                                     struct inode *inode, loff_t pos,
 778                                     size_t count)
 779 {
 780         size_t res;
 781         unsigned offset;
 782         unsigned i;
 783
 784         for (i = 0; i < req->num_pages; i++)
 785                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 786
 787         res = fuse_send_write(req, file, pos, count, NULL);
 788
 789         offset = req->page_offset;
 790         count = res;
 791         for (i = 0; i < req->num_pages; i++) {
 792                 struct page *page = req->pages[i];
 793
 794                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 795                         SetPageUptodate(page);
 796
 797                 if (count > PAGE_CACHE_SIZE - offset)
 798                         count -= PAGE_CACHE_SIZE - offset;
 799                 else
 800                         count = 0;
 801                 offset = 0;
 802
 803                 unlock_page(page);
 804                 page_cache_release(page);
 805         }
 806
 807         return res;
 808 }
 809
 810 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 811                                struct address_space *mapping,
 812                                struct iov_iter *ii, loff_t pos)
 813 {
 814         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 815         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 816         size_t count = 0;
 817         int err;
 818
 819         req->in.argpages = 1;
 820         req->page_offset = offset;
 821
 822         do {
 823                 size_t tmp;
 824                 struct page *page;
 825                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 826                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 827                                      iov_iter_count(ii));
 828
 829                 bytes = min_t(size_t, bytes, fc->max_write - count);
 830
 831  again:
 832                 err = -EFAULT;
 833                 if (iov_iter_fault_in_readable(ii, bytes))
 834                         break;
 835
 836                 err = -ENOMEM;
 837                 page = grab_cache_page_write_begin(mapping, index, 0);
 838                 if (!page)
 839                         break;
 840
 841                 if (mapping_writably_mapped(mapping))
 842                         flush_dcache_page(page);
 843
 844                 pagefault_disable();
 845                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 846                 pagefault_enable();
 847                 flush_dcache_page(page);
 848
 849                 mark_page_accessed(page);
 850
 851                 if (!tmp) {
 852                         unlock_page(page);
 853                         page_cache_release(page);
 854                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 855                         goto again;
 856                 }
 857
 858                 err = 0;
 859                 req->pages[req->num_pages] = page;
 860                 req->num_pages++;
 861
 862                 iov_iter_advance(ii, tmp);
 863                 count += tmp;
 864                 pos += tmp;
 865                 offset += tmp;
 866                 if (offset == PAGE_CACHE_SIZE)
 867                         offset = 0;
 868
 869                 if (!fc->big_writes)
 870                         break;
 871         } while (iov_iter_count(ii) && count < fc->max_write &&
 872                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 873
 874         return count > 0 ? count : err;
 875 }
 876
 877 static ssize_t fuse_perform_write(struct file *file,
 878                                   struct address_space *mapping,
 879                                   struct iov_iter *ii, loff_t pos)
 880 {
 881         struct inode *inode = mapping->host;
 882         struct fuse_conn *fc = get_fuse_conn(inode);
 883         int err = 0;
 884         ssize_t res = 0;
 885
 886         if (is_bad_inode(inode))
 887                 return -EIO;
 888
 889         do {
 890                 struct fuse_req *req;
 891                 ssize_t count;
 892
 893                 req = fuse_get_req(fc);
 894                 if (IS_ERR(req)) {
 895                         err = PTR_ERR(req);
 896                         break;
 897                 }
 898
 899                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 900                 if (count <= 0) {
 901                         err = count;
 902                 } else {
 903                         size_t num_written;
 904
 905                         num_written = fuse_send_write_pages(req, file, inode,
 906                                                             pos, count);
 907                         err = req->out.h.error;
 908                         if (!err) {
 909                                 res += num_written;
 910                                 pos += num_written;
 911
 912                                 /* break out of the loop on short write */
 913                                 if (num_written != count)
 914                                         err = -EIO;
 915                         }
 916                 }
 917                 fuse_put_request(fc, req);
 918         } while (!err && iov_iter_count(ii));
 919
 920         if (res > 0)
 921                 fuse_write_update_size(inode, pos);
 922
 923         fuse_invalidate_attr(inode);
 924
 925         return res > 0 ? res : err;
 926 }
 927
 928 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 929                                    unsigned long nr_segs, loff_t pos)
 930 {
 931         struct file *file = iocb->ki_filp;
 932         struct address_space *mapping = file->f_mapping;
 933         size_t count = 0;
 934         size_t ocount = 0;
 935         ssize_t written = 0;
 936         ssize_t written_buffered = 0;
 937         struct inode *inode = mapping->host;
 938         ssize_t err;
 939         struct iov_iter i;
 940         loff_t endbyte = 0;
 941
 942         WARN_ON(iocb->ki_pos != pos);
 943
 944         ocount = 0;
 945         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 946         if (err)
 947                 return err;
 948
 949         count = ocount;
 950         sb_start_write(inode->i_sb);
 951         mutex_lock(&inode->i_mutex);
 952
 953         /* We can write back this queue in page reclaim */
 954         current->backing_dev_info = mapping->backing_dev_info;
 955
 956         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 957         if (err)
 958                 goto out;
 959
 960         if (count == 0)
 961                 goto out;
 962
 963         err = file_remove_suid(file);
 964         if (err)
 965                 goto out;
 966
 967         err = file_update_time(file);
 968         if (err)
 969                 goto out;
 970
 971         if (file->f_flags & O_DIRECT) {
 972                 written = generic_file_direct_write(iocb, iov, &nr_segs,
 973                                                     pos, &iocb->ki_pos,
 974                                                     count, ocount);
 975                 if (written < 0 || written == count)
 976                         goto out;
 977
 978                 pos += written;
 979                 count -= written;
 980
 981                 iov_iter_init(&i, iov, nr_segs, count, written);
 982                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
 983                 if (written_buffered < 0) {
 984                         err = written_buffered;
 985                         goto out;
 986                 }
 987                 endbyte = pos + written_buffered - 1;
 988
 989                 err = filemap_write_and_wait_range(file->f_mapping, pos,
 990                                                    endbyte);
 991                 if (err)
 992                         goto out;
 993
 994                 invalidate_mapping_pages(file->f_mapping,
 995                                          pos >> PAGE_CACHE_SHIFT,
 996                                          endbyte >> PAGE_CACHE_SHIFT);
 997
 998                 written += written_buffered;
 999                 iocb->ki_pos = pos + written_buffered;
1000         } else {
1001                 iov_iter_init(&i, iov, nr_segs, count, 0);
1002                 written = fuse_perform_write(file, mapping, &i, pos);
1003                 if (written >= 0)
1004                         iocb->ki_pos = pos + written;
1005         }
1006 out:
1007         current->backing_dev_info = NULL;
1008         mutex_unlock(&inode->i_mutex);
1009         sb_end_write(inode->i_sb);
1010
1011         return written ? written : err;
1012 }
1013
1014 static void fuse_release_user_pages(struct fuse_req *req, int write)
1015 {
1016         unsigned i;
1017
1018         for (i = 0; i < req->num_pages; i++) {
1019                 struct page *page = req->pages[i];
1020                 if (write)
1021                         set_page_dirty_lock(page);
1022                 put_page(page);
1023         }
1024 }
1025
1026 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1027                                size_t *nbytesp, int write)
1028 {
1029         size_t nbytes = *nbytesp;
1030         unsigned long user_addr = (unsigned long) buf;
1031         unsigned offset = user_addr & ~PAGE_MASK;
1032         int npages;
1033
1034         /* Special case for kernel I/O: can copy directly into the buffer */
1035         if (segment_eq(get_fs(), KERNEL_DS)) {
1036                 if (write)
1037                         req->in.args[1].value = (void *) user_addr;
1038                 else
1039                         req->out.args[0].value = (void *) user_addr;
1040
1041                 return 0;
1042         }
1043
1044         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1045         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1046         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1047         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1048         if (npages < 0)
1049                 return npages;
1050
1051         req->num_pages = npages;
1052         req->page_offset = offset;
1053
1054         if (write)
1055                 req->in.argpages = 1;
1056         else
1057                 req->out.argpages = 1;
1058
1059         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1060         *nbytesp = min(*nbytesp, nbytes);
1061
1062         return 0;
1063 }
1064
1065 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1066                        size_t count, loff_t *ppos, int write)
1067 {
1068         struct fuse_file *ff = file->private_data;
1069         struct fuse_conn *fc = ff->fc;
1070         size_t nmax = write ? fc->max_write : fc->max_read;
1071         loff_t pos = *ppos;
1072         ssize_t res = 0;
1073         struct fuse_req *req;
1074
1075         req = fuse_get_req(fc);
1076         if (IS_ERR(req))
1077                 return PTR_ERR(req);
1078
1079         while (count) {
1080                 size_t nres;
1081                 fl_owner_t owner = current->files;
1082                 size_t nbytes = min(count, nmax);
1083                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1084                 if (err) {
1085                         res = err;
1086                         break;
1087                 }
1088
1089                 if (write)
1090                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1091                 else
1092                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1093
1094                 fuse_release_user_pages(req, !write);
1095                 if (req->out.h.error) {
1096                         if (!res)
1097                                 res = req->out.h.error;
1098                         break;
1099                 } else if (nres > nbytes) {
1100                         res = -EIO;
1101                         break;
1102                 }
1103                 count -= nres;
1104                 res += nres;
1105                 pos += nres;
1106                 buf += nres;
1107                 if (nres != nbytes)
1108                         break;
1109                 if (count) {
1110                         fuse_put_request(fc, req);
1111                         req = fuse_get_req(fc);
1112                         if (IS_ERR(req))
1113                                 break;
1114                 }
1115         }
1116         if (!IS_ERR(req))
1117                 fuse_put_request(fc, req);
1118         if (res > 0)
1119                 *ppos = pos;
1120
1121         return res;
1122 }
1123 EXPORT_SYMBOL_GPL(fuse_direct_io);
1124
1125 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1126                                      size_t count, loff_t *ppos)
1127 {
1128         ssize_t res;
1129         struct inode *inode = file->f_path.dentry->d_inode;
1130
1131         if (is_bad_inode(inode))
1132                 return -EIO;
1133
1134         res = fuse_direct_io(file, buf, count, ppos, 0);
1135
1136         fuse_invalidate_attr(inode);
1137
1138         return res;
1139 }
1140
1141 static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
1142                                    size_t count, loff_t *ppos)
1143 {
1144         struct inode *inode = file->f_path.dentry->d_inode;
1145         ssize_t res;
1146
1147         res = generic_write_checks(file, ppos, &count, 0);
1148         if (!res) {
1149                 res = fuse_direct_io(file, buf, count, ppos, 1);
1150                 if (res > 0)
1151                         fuse_write_update_size(inode, *ppos);
1152         }
1153
1154         fuse_invalidate_attr(inode);
1155
1156         return res;
1157 }
1158
1159 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1160                                  size_t count, loff_t *ppos)
1161 {
1162         struct inode *inode = file->f_path.dentry->d_inode;
1163         ssize_t res;
1164
1165         if (is_bad_inode(inode))
1166                 return -EIO;
1167
1168         /* Don't allow parallel writes to the same file */
1169         mutex_lock(&inode->i_mutex);
1170         res = __fuse_direct_write(file, buf, count, ppos);
1171         mutex_unlock(&inode->i_mutex);
1172
1173         return res;
1174 }
1175
1176 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1177 {
1178         __free_page(req->pages[0]);
1179         fuse_file_put(req->ff, false);
1180 }
1181
1182 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1183 {
1184         struct inode *inode = req->inode;
1185         struct fuse_inode *fi = get_fuse_inode(inode);
1186         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1187
1188         list_del(&req->writepages_entry);
1189         dec_bdi_stat(bdi, BDI_WRITEBACK);
1190         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1191         bdi_writeout_inc(bdi);
1192         wake_up(&fi->page_waitq);
1193 }
1194
1195 /* Called under fc->lock, may release and reacquire it */
1196 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1197 __releases(fc->lock)
1198 __acquires(fc->lock)
1199 {
1200         struct fuse_inode *fi = get_fuse_inode(req->inode);
1201         loff_t size = i_size_read(req->inode);
1202         struct fuse_write_in *inarg = &req->misc.write.in;
1203
1204         if (!fc->connected)
1205                 goto out_free;
1206
1207         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1208                 inarg->size = PAGE_CACHE_SIZE;
1209         } else if (inarg->offset < size) {
1210                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1211         } else {
1212                 /* Got truncated off completely */
1213                 goto out_free;
1214         }
1215
1216         req->in.args[1].size = inarg->size;
1217         fi->writectr++;
1218         fuse_request_send_background_locked(fc, req);
1219         return;
1220
1221  out_free:
1222         fuse_writepage_finish(fc, req);
1223         spin_unlock(&fc->lock);
1224         fuse_writepage_free(fc, req);
1225         fuse_put_request(fc, req);
1226         spin_lock(&fc->lock);
1227 }
1228
1229 /*
1230  * If fi->writectr is positive (no truncate or fsync going on) send
1231  * all queued writepage requests.
1232  *
1233  * Called with fc->lock
1234  */
1235 void fuse_flush_writepages(struct inode *inode)
1236 __releases(fc->lock)
1237 __acquires(fc->lock)
1238 {
1239         struct fuse_conn *fc = get_fuse_conn(inode);
1240         struct fuse_inode *fi = get_fuse_inode(inode);
1241         struct fuse_req *req;
1242
1243         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1244                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1245                 list_del_init(&req->list);
1246                 fuse_send_writepage(fc, req);
1247         }
1248 }
1249
1250 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1251 {
1252         struct inode *inode = req->inode;
1253         struct fuse_inode *fi = get_fuse_inode(inode);
1254
1255         mapping_set_error(inode->i_mapping, req->out.h.error);
1256         spin_lock(&fc->lock);
1257         fi->writectr--;
1258         fuse_writepage_finish(fc, req);
1259         spin_unlock(&fc->lock);
1260         fuse_writepage_free(fc, req);
1261 }
1262
1263 static int fuse_writepage_locked(struct page *page)
1264 {
1265         struct address_space *mapping = page->mapping;
1266         struct inode *inode = mapping->host;
1267         struct fuse_conn *fc = get_fuse_conn(inode);
1268         struct fuse_inode *fi = get_fuse_inode(inode);
1269         struct fuse_req *req;
1270         struct fuse_file *ff;
1271         struct page *tmp_page;
1272
1273         set_page_writeback(page);
1274
1275         req = fuse_request_alloc_nofs();
1276         if (!req)
1277                 goto err;
1278
1279         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1280         if (!tmp_page)
1281                 goto err_free;
1282
1283         spin_lock(&fc->lock);
1284         BUG_ON(list_empty(&fi->write_files));
1285         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1286         req->ff = fuse_file_get(ff);
1287         spin_unlock(&fc->lock);
1288
1289         fuse_write_fill(req, ff, page_offset(page), 0);
1290
1291         copy_highpage(tmp_page, page);
1292         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1293         req->in.argpages = 1;
1294         req->num_pages = 1;
1295         req->pages[0] = tmp_page;
1296         req->page_offset = 0;
1297         req->end = fuse_writepage_end;
1298         req->inode = inode;
1299
1300         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1301         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1302         end_page_writeback(page);
1303
1304         spin_lock(&fc->lock);
1305         list_add(&req->writepages_entry, &fi->writepages);
1306         list_add_tail(&req->list, &fi->queued_writes);
1307         fuse_flush_writepages(inode);
1308         spin_unlock(&fc->lock);
1309
1310         return 0;
1311
1312 err_free:
1313         fuse_request_free(req);
1314 err:
1315         end_page_writeback(page);
1316         return -ENOMEM;
1317 }
1318
1319 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1320 {
1321         int err;
1322
1323         err = fuse_writepage_locked(page);
1324         unlock_page(page);
1325
1326         return err;
1327 }
1328
1329 static int fuse_launder_page(struct page *page)
1330 {
1331         int err = 0;
1332         if (clear_page_dirty_for_io(page)) {
1333                 struct inode *inode = page->mapping->host;
1334                 err = fuse_writepage_locked(page);
1335                 if (!err)
1336                         fuse_wait_on_page_writeback(inode, page->index);
1337         }
1338         return err;
1339 }
1340
1341 /*
1342  * Write back dirty pages now, because there may not be any suitable
1343  * open files later
1344  */
1345 static void fuse_vma_close(struct vm_area_struct *vma)
1346 {
1347         filemap_write_and_wait(vma->vm_file->f_mapping);
1348 }
1349
1350 /*
1351  * Wait for writeback against this page to complete before allowing it
1352  * to be marked dirty again, and hence written back again, possibly
1353  * before the previous writepage completed.
1354  *
1355  * Block here, instead of in ->writepage(), so that the userspace fs
1356  * can only block processes actually operating on the filesystem.
1357  *
1358  * Otherwise unprivileged userspace fs would be able to block
1359  * unrelated:
1360  *
1361  * - page migration
1362  * - sync(2)
1363  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1364  */
1365 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1366 {
1367         struct page *page = vmf->page;
1368         /*
1369          * Don't use page->mapping as it may become NULL from a
1370          * concurrent truncate.
1371          */
1372         struct inode *inode = vma->vm_file->f_mapping->host;
1373
1374         fuse_wait_on_page_writeback(inode, page->index);
1375         return 0;
1376 }
1377
1378 static const struct vm_operations_struct fuse_file_vm_ops = {
1379         .close          = fuse_vma_close,
1380         .fault          = filemap_fault,
1381         .page_mkwrite   = fuse_page_mkwrite,
1382 };
1383
1384 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1385 {
1386         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1387                 struct inode *inode = file->f_dentry->d_inode;
1388                 struct fuse_conn *fc = get_fuse_conn(inode);
1389                 struct fuse_inode *fi = get_fuse_inode(inode);
1390                 struct fuse_file *ff = file->private_data;
1391                 /*
1392                  * file may be written through mmap, so chain it onto the
1393                  * inodes's write_file list
1394                  */
1395                 spin_lock(&fc->lock);
1396                 if (list_empty(&ff->write_entry))
1397                         list_add(&ff->write_entry, &fi->write_files);
1398                 spin_unlock(&fc->lock);
1399         }
1400         file_accessed(file);
1401         vma->vm_ops = &fuse_file_vm_ops;
1402         return 0;
1403 }
1404
1405 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1406 {
1407         /* Can't provide the coherency needed for MAP_SHARED */
1408         if (vma->vm_flags & VM_MAYSHARE)
1409                 return -ENODEV;
1410
1411         invalidate_inode_pages2(file->f_mapping);
1412
1413         return generic_file_mmap(file, vma);
1414 }
1415
1416 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1417                                   struct file_lock *fl)
1418 {
1419         switch (ffl->type) {
1420         case F_UNLCK:
1421                 break;
1422
1423         case F_RDLCK:
1424         case F_WRLCK:
1425                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1426                     ffl->end < ffl->start)
1427                         return -EIO;
1428
1429                 fl->fl_start = ffl->start;
1430                 fl->fl_end = ffl->end;
1431                 fl->fl_pid = ffl->pid;
1432                 break;
1433
1434         default:
1435                 return -EIO;
1436         }
1437         fl->fl_type = ffl->type;
1438         return 0;
1439 }
1440
1441 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1442                          const struct file_lock *fl, int opcode, pid_t pid,
1443                          int flock)
1444 {
1445         struct inode *inode = file->f_path.dentry->d_inode;
1446         struct fuse_conn *fc = get_fuse_conn(inode);
1447         struct fuse_file *ff = file->private_data;
1448         struct fuse_lk_in *arg = &req->misc.lk_in;
1449
1450         arg->fh = ff->fh;
1451         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1452         arg->lk.start = fl->fl_start;
1453         arg->lk.end = fl->fl_end;
1454         arg->lk.type = fl->fl_type;
1455         arg->lk.pid = pid;
1456         if (flock)
1457                 arg->lk_flags |= FUSE_LK_FLOCK;
1458         req->in.h.opcode = opcode;
1459         req->in.h.nodeid = get_node_id(inode);
1460         req->in.numargs = 1;
1461         req->in.args[0].size = sizeof(*arg);
1462         req->in.args[0].value = arg;
1463 }
1464
1465 static int fuse_getlk(struct file *file, struct file_lock *fl)
1466 {
1467         struct inode *inode = file->f_path.dentry->d_inode;
1468         struct fuse_conn *fc = get_fuse_conn(inode);
1469         struct fuse_req *req;
1470         struct fuse_lk_out outarg;
1471         int err;
1472
1473         req = fuse_get_req(fc);
1474         if (IS_ERR(req))
1475                 return PTR_ERR(req);
1476
1477         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1478         req->out.numargs = 1;
1479         req->out.args[0].size = sizeof(outarg);
1480         req->out.args[0].value = &outarg;
1481         fuse_request_send(fc, req);
1482         err = req->out.h.error;
1483         fuse_put_request(fc, req);
1484         if (!err)
1485                 err = convert_fuse_file_lock(&outarg.lk, fl);
1486
1487         return err;
1488 }
1489
1490 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1491 {
1492         struct inode *inode = file->f_path.dentry->d_inode;
1493         struct fuse_conn *fc = get_fuse_conn(inode);
1494         struct fuse_req *req;
1495         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1496         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1497         int err;
1498
1499         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1500                 /* NLM needs asynchronous locks, which we don't support yet */
1501                 return -ENOLCK;
1502         }
1503
1504         /* Unlock on close is handled by the flush method */
1505         if (fl->fl_flags & FL_CLOSE)
1506                 return 0;
1507
1508         req = fuse_get_req(fc);
1509         if (IS_ERR(req))
1510                 return PTR_ERR(req);
1511
1512         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1513         fuse_request_send(fc, req);
1514         err = req->out.h.error;
1515         /* locking is restartable */
1516         if (err == -EINTR)
1517                 err = -ERESTARTSYS;
1518         fuse_put_request(fc, req);
1519         return err;
1520 }
1521
1522 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1523 {
1524         struct inode *inode = file->f_path.dentry->d_inode;
1525         struct fuse_conn *fc = get_fuse_conn(inode);
1526         int err;
1527
1528         if (cmd == F_CANCELLK) {
1529                 err = 0;
1530         } else if (cmd == F_GETLK) {
1531                 if (fc->no_lock) {
1532                         posix_test_lock(file, fl);
1533                         err = 0;
1534                 } else
1535                         err = fuse_getlk(file, fl);
1536         } else {
1537                 if (fc->no_lock)
1538                         err = posix_lock_file(file, fl, NULL);
1539                 else
1540                         err = fuse_setlk(file, fl, 0);
1541         }
1542         return err;
1543 }
1544
1545 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1546 {
1547         struct inode *inode = file->f_path.dentry->d_inode;
1548         struct fuse_conn *fc = get_fuse_conn(inode);
1549         int err;
1550
1551         if (fc->no_flock) {
1552                 err = flock_lock_file_wait(file, fl);
1553         } else {
1554                 struct fuse_file *ff = file->private_data;
1555
1556                 /* emulate flock with POSIX locks */
1557                 fl->fl_owner = (fl_owner_t) file;
1558                 ff->flock = true;
1559                 err = fuse_setlk(file, fl, 1);
1560         }
1561
1562         return err;
1563 }
1564
1565 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1566 {
1567         struct inode *inode = mapping->host;
1568         struct fuse_conn *fc = get_fuse_conn(inode);
1569         struct fuse_req *req;
1570         struct fuse_bmap_in inarg;
1571         struct fuse_bmap_out outarg;
1572         int err;
1573
1574         if (!inode->i_sb->s_bdev || fc->no_bmap)
1575                 return 0;
1576
1577         req = fuse_get_req(fc);
1578         if (IS_ERR(req))
1579                 return 0;
1580
1581         memset(&inarg, 0, sizeof(inarg));
1582         inarg.block = block;
1583         inarg.blocksize = inode->i_sb->s_blocksize;
1584         req->in.h.opcode = FUSE_BMAP;
1585         req->in.h.nodeid = get_node_id(inode);
1586         req->in.numargs = 1;
1587         req->in.args[0].size = sizeof(inarg);
1588         req->in.args[0].value = &inarg;
1589         req->out.numargs = 1;
1590         req->out.args[0].size = sizeof(outarg);
1591         req->out.args[0].value = &outarg;
1592         fuse_request_send(fc, req);
1593         err = req->out.h.error;
1594         fuse_put_request(fc, req);
1595         if (err == -ENOSYS)
1596                 fc->no_bmap = 1;
1597
1598         return err ? 0 : outarg.block;
1599 }
1600
1601 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1602 {
1603         loff_t retval;
1604         struct inode *inode = file->f_path.dentry->d_inode;
1605
1606         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1607         if (origin == SEEK_CUR || origin == SEEK_SET)
1608                 return generic_file_llseek(file, offset, origin);
1609
1610         mutex_lock(&inode->i_mutex);
1611         retval = fuse_update_attributes(inode, NULL, file, NULL);
1612         if (!retval)
1613                 retval = generic_file_llseek(file, offset, origin);
1614         mutex_unlock(&inode->i_mutex);
1615
1616         return retval;
1617 }
1618
1619 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1620                         unsigned int nr_segs, size_t bytes, bool to_user)
1621 {
1622         struct iov_iter ii;
1623         int page_idx = 0;
1624
1625         if (!bytes)
1626                 return 0;
1627
1628         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1629
1630         while (iov_iter_count(&ii)) {
1631                 struct page *page = pages[page_idx++];
1632                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1633                 void *kaddr;
1634
1635                 kaddr = kmap(page);
1636
1637                 while (todo) {
1638                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1639                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1640                         size_t copy = min(todo, iov_len);
1641                         size_t left;
1642
1643                         if (!to_user)
1644                                 left = copy_from_user(kaddr, uaddr, copy);
1645                         else
1646                                 left = copy_to_user(uaddr, kaddr, copy);
1647
1648                         if (unlikely(left))
1649                                 return -EFAULT;
1650
1651                         iov_iter_advance(&ii, copy);
1652                         todo -= copy;
1653                         kaddr += copy;
1654                 }
1655
1656                 kunmap(page);
1657         }
1658
1659         return 0;
1660 }
1661
1662 /*
1663  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1664  * ABI was defined to be 'struct iovec' which is different on 32bit
1665  * and 64bit.  Fortunately we can determine which structure the server
1666  * used from the size of the reply.
1667  */
1668 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1669                                      size_t transferred, unsigned count,
1670                                      bool is_compat)
1671 {
1672 #ifdef CONFIG_COMPAT
1673         if (count * sizeof(struct compat_iovec) == transferred) {
1674                 struct compat_iovec *ciov = src;
1675                 unsigned i;
1676
1677                 /*
1678                  * With this interface a 32bit server cannot support
1679                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1680                  * requests
1681                  */
1682                 if (!is_compat)
1683                         return -EINVAL;
1684
1685                 for (i = 0; i < count; i++) {
1686                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1687                         dst[i].iov_len = ciov[i].iov_len;
1688                 }
1689                 return 0;
1690         }
1691 #endif
1692
1693         if (count * sizeof(struct iovec) != transferred)
1694                 return -EIO;
1695
1696         memcpy(dst, src, transferred);
1697         return 0;
1698 }
1699
1700 /* Make sure iov_length() won't overflow */
1701 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1702 {
1703         size_t n;
1704         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1705
1706         for (n = 0; n < count; n++, iov++) {
1707                 if (iov->iov_len > (size_t) max)
1708                         return -ENOMEM;
1709                 max -= iov->iov_len;
1710         }
1711         return 0;
1712 }
1713
1714 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1715                                  void *src, size_t transferred, unsigned count,
1716                                  bool is_compat)
1717 {
1718         unsigned i;
1719         struct fuse_ioctl_iovec *fiov = src;
1720
1721         if (fc->minor < 16) {
1722                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1723                                                  count, is_compat);
1724         }
1725
1726         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1727                 return -EIO;
1728
1729         for (i = 0; i < count; i++) {
1730                 /* Did the server supply an inappropriate value? */
1731                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1732                     fiov[i].len != (unsigned long) fiov[i].len)
1733                         return -EIO;
1734
1735                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1736                 dst[i].iov_len = (size_t) fiov[i].len;
1737
1738 #ifdef CONFIG_COMPAT
1739                 if (is_compat &&
1740                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1741                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1742                         return -EIO;
1743 #endif
1744         }
1745
1746         return 0;
1747 }
1748
1749
1750 /*
1751  * For ioctls, there is no generic way to determine how much memory
1752  * needs to be read and/or written.  Furthermore, ioctls are allowed
1753  * to dereference the passed pointer, so the parameter requires deep
1754  * copying but FUSE has no idea whatsoever about what to copy in or
1755  * out.
1756  *
1757  * This is solved by allowing FUSE server to retry ioctl with
1758  * necessary in/out iovecs.  Let's assume the ioctl implementation
1759  * needs to read in the following structure.
1760  *
1761  * struct a {
1762  *      char    *buf;
1763  *      size_t  buflen;
1764  * }
1765  *
1766  * On the first callout to FUSE server, inarg->in_size and
1767  * inarg->out_size will be NULL; then, the server completes the ioctl
1768  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1769  * the actual iov array to
1770  *
1771  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1772  *
1773  * which tells FUSE to copy in the requested area and retry the ioctl.
1774  * On the second round, the server has access to the structure and
1775  * from that it can tell what to look for next, so on the invocation,
1776  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1777  *
1778  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1779  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1780  *
1781  * FUSE will copy both struct a and the pointed buffer from the
1782  * process doing the ioctl and retry ioctl with both struct a and the
1783  * buffer.
1784  *
1785  * This time, FUSE server has everything it needs and completes ioctl
1786  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1787  *
1788  * Copying data out works the same way.
1789  *
1790  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1791  * automatically initializes in and out iovs by decoding @cmd with
1792  * _IOC_* macros and the server is not allowed to request RETRY.  This
1793  * limits ioctl data transfers to well-formed ioctls and is the forced
1794  * behavior for all FUSE servers.
1795  */
1796 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1797                    unsigned int flags)
1798 {
1799         struct fuse_file *ff = file->private_data;
1800         struct fuse_conn *fc = ff->fc;
1801         struct fuse_ioctl_in inarg = {
1802                 .fh = ff->fh,
1803                 .cmd = cmd,
1804                 .arg = arg,
1805                 .flags = flags
1806         };
1807         struct fuse_ioctl_out outarg;
1808         struct fuse_req *req = NULL;
1809         struct page **pages = NULL;
1810         struct iovec *iov_page = NULL;
1811         struct iovec *in_iov = NULL, *out_iov = NULL;
1812         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1813         size_t in_size, out_size, transferred;
1814         int err;
1815
1816 #if BITS_PER_LONG == 32
1817         inarg.flags |= FUSE_IOCTL_32BIT;
1818 #else
1819         if (flags & FUSE_IOCTL_COMPAT)
1820                 inarg.flags |= FUSE_IOCTL_32BIT;
1821 #endif
1822
1823         /* assume all the iovs returned by client always fits in a page */
1824         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1825
1826         err = -ENOMEM;
1827         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1828         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1829         if (!pages || !iov_page)
1830                 goto out;
1831
1832         /*
1833          * If restricted, initialize IO parameters as encoded in @cmd.
1834          * RETRY from server is not allowed.
1835          */
1836         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1837                 struct iovec *iov = iov_page;
1838
1839                 iov->iov_base = (void __user *)arg;
1840                 iov->iov_len = _IOC_SIZE(cmd);
1841
1842                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1843                         in_iov = iov;
1844                         in_iovs = 1;
1845                 }
1846
1847                 if (_IOC_DIR(cmd) & _IOC_READ) {
1848                         out_iov = iov;
1849                         out_iovs = 1;
1850                 }
1851         }
1852
1853  retry:
1854         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1855         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1856
1857         /*
1858          * Out data can be used either for actual out data or iovs,
1859          * make sure there always is at least one page.
1860          */
1861         out_size = max_t(size_t, out_size, PAGE_SIZE);
1862         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1863
1864         /* make sure there are enough buffer pages and init request with them */
1865         err = -ENOMEM;
1866         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1867                 goto out;
1868         while (num_pages < max_pages) {
1869                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1870                 if (!pages[num_pages])
1871                         goto out;
1872                 num_pages++;
1873         }
1874
1875         req = fuse_get_req(fc);
1876         if (IS_ERR(req)) {
1877                 err = PTR_ERR(req);
1878                 req = NULL;
1879                 goto out;
1880         }
1881         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1882         req->num_pages = num_pages;
1883
1884         /* okay, let's send it to the client */
1885         req->in.h.opcode = FUSE_IOCTL;
1886         req->in.h.nodeid = ff->nodeid;
1887         req->in.numargs = 1;
1888         req->in.args[0].size = sizeof(inarg);
1889         req->in.args[0].value = &inarg;
1890         if (in_size) {
1891                 req->in.numargs++;
1892                 req->in.args[1].size = in_size;
1893                 req->in.argpages = 1;
1894
1895                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1896                                            false);
1897                 if (err)
1898                         goto out;
1899         }
1900
1901         req->out.numargs = 2;
1902         req->out.args[0].size = sizeof(outarg);
1903         req->out.args[0].value = &outarg;
1904         req->out.args[1].size = out_size;
1905         req->out.argpages = 1;
1906         req->out.argvar = 1;
1907
1908         fuse_request_send(fc, req);
1909         err = req->out.h.error;
1910         transferred = req->out.args[1].size;
1911         fuse_put_request(fc, req);
1912         req = NULL;
1913         if (err)
1914                 goto out;
1915
1916         /* did it ask for retry? */
1917         if (outarg.flags & FUSE_IOCTL_RETRY) {
1918                 void *vaddr;
1919
1920                 /* no retry if in restricted mode */
1921                 err = -EIO;
1922                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1923                         goto out;
1924
1925                 in_iovs = outarg.in_iovs;
1926                 out_iovs = outarg.out_iovs;
1927
1928                 /*
1929                  * Make sure things are in boundary, separate checks
1930                  * are to protect against overflow.
1931                  */
1932                 err = -ENOMEM;
1933                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1934                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1935                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1936                         goto out;
1937
1938                 vaddr = kmap_atomic(pages[0]);
1939                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1940                                             transferred, in_iovs + out_iovs,
1941                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1942                 kunmap_atomic(vaddr);
1943                 if (err)
1944                         goto out;
1945
1946                 in_iov = iov_page;
1947                 out_iov = in_iov + in_iovs;
1948
1949                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1950                 if (err)
1951                         goto out;
1952
1953                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1954                 if (err)
1955                         goto out;
1956
1957                 goto retry;
1958         }
1959
1960         err = -EIO;
1961         if (transferred > inarg.out_size)
1962                 goto out;
1963
1964         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1965  out:
1966         if (req)
1967                 fuse_put_request(fc, req);
1968         free_page((unsigned long) iov_page);
1969         while (num_pages)
1970                 __free_page(pages[--num_pages]);
1971         kfree(pages);
1972
1973         return err ? err : outarg.result;
1974 }
1975 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1976
1977 long fuse_ioctl_common(struct file *file, unsigned int cmd,
1978                        unsigned long arg, unsigned int flags)
1979 {
1980         struct inode *inode = file->f_dentry->d_inode;
1981         struct fuse_conn *fc = get_fuse_conn(inode);
1982
1983         if (!fuse_allow_task(fc, current))
1984                 return -EACCES;
1985
1986         if (is_bad_inode(inode))
1987                 return -EIO;
1988
1989         return fuse_do_ioctl(file, cmd, arg, flags);
1990 }
1991
1992 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1993                             unsigned long arg)
1994 {
1995         return fuse_ioctl_common(file, cmd, arg, 0);
1996 }
1997
1998 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1999                                    unsigned long arg)
2000 {
2001         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2002 }
2003
2004 /*
2005  * All files which have been polled are linked to RB tree
2006  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2007  * find the matching one.
2008  */
2009 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2010                                               struct rb_node **parent_out)
2011 {
2012         struct rb_node **link = &fc->polled_files.rb_node;
2013         struct rb_node *last = NULL;
2014
2015         while (*link) {
2016                 struct fuse_file *ff;
2017
2018                 last = *link;
2019                 ff = rb_entry(last, struct fuse_file, polled_node);
2020
2021                 if (kh < ff->kh)
2022                         link = &last->rb_left;
2023                 else if (kh > ff->kh)
2024                         link = &last->rb_right;
2025                 else
2026                         return link;
2027         }
2028
2029         if (parent_out)
2030                 *parent_out = last;
2031         return link;
2032 }
2033
2034 /*
2035  * The file is about to be polled.  Make sure it's on the polled_files
2036  * RB tree.  Note that files once added to the polled_files tree are
2037  * not removed before the file is released.  This is because a file
2038  * polled once is likely to be polled again.
2039  */
2040 static void fuse_register_polled_file(struct fuse_conn *fc,
2041                                       struct fuse_file *ff)
2042 {
2043         spin_lock(&fc->lock);
2044         if (RB_EMPTY_NODE(&ff->polled_node)) {
2045                 struct rb_node **link, *parent;
2046
2047                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2048                 BUG_ON(*link);
2049                 rb_link_node(&ff->polled_node, parent, link);
2050                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2051         }
2052         spin_unlock(&fc->lock);
2053 }
2054
2055 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2056 {
2057         struct fuse_file *ff = file->private_data;
2058         struct fuse_conn *fc = ff->fc;
2059         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2060         struct fuse_poll_out outarg;
2061         struct fuse_req *req;
2062         int err;
2063
2064         if (fc->no_poll)
2065                 return DEFAULT_POLLMASK;
2066
2067         poll_wait(file, &ff->poll_wait, wait);
2068
2069         /*
2070          * Ask for notification iff there's someone waiting for it.
2071          * The client may ignore the flag and always notify.
2072          */
2073         if (waitqueue_active(&ff->poll_wait)) {
2074                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2075                 fuse_register_polled_file(fc, ff);
2076         }
2077
2078         req = fuse_get_req(fc);
2079         if (IS_ERR(req))
2080                 return POLLERR;
2081
2082         req->in.h.opcode = FUSE_POLL;
2083         req->in.h.nodeid = ff->nodeid;
2084         req->in.numargs = 1;
2085         req->in.args[0].size = sizeof(inarg);
2086         req->in.args[0].value = &inarg;
2087         req->out.numargs = 1;
2088         req->out.args[0].size = sizeof(outarg);
2089         req->out.args[0].value = &outarg;
2090         fuse_request_send(fc, req);
2091         err = req->out.h.error;
2092         fuse_put_request(fc, req);
2093
2094         if (!err)
2095                 return outarg.revents;
2096         if (err == -ENOSYS) {
2097                 fc->no_poll = 1;
2098                 return DEFAULT_POLLMASK;
2099         }
2100         return POLLERR;
2101 }
2102 EXPORT_SYMBOL_GPL(fuse_file_poll);
2103
2104 /*
2105  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2106  * wakes up the poll waiters.
2107  */
2108 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2109                             struct fuse_notify_poll_wakeup_out *outarg)
2110 {
2111         u64 kh = outarg->kh;
2112         struct rb_node **link;
2113
2114         spin_lock(&fc->lock);
2115
2116         link = fuse_find_polled_node(fc, kh, NULL);
2117         if (*link) {
2118                 struct fuse_file *ff;
2119
2120                 ff = rb_entry(*link, struct fuse_file, polled_node);
2121                 wake_up_interruptible_sync(&ff->poll_wait);
2122         }
2123
2124         spin_unlock(&fc->lock);
2125         return 0;
2126 }
2127
2128 static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
2129                              unsigned long nr_segs, loff_t *ppos, int rw)
2130 {
2131         const struct iovec *vector = iov;
2132         ssize_t ret = 0;
2133
2134         while (nr_segs > 0) {
2135                 void __user *base;
2136                 size_t len;
2137                 ssize_t nr;
2138
2139                 base = vector->iov_base;
2140                 len = vector->iov_len;
2141                 vector++;
2142                 nr_segs--;
2143
2144                 if (rw == WRITE)
2145                         nr = __fuse_direct_write(filp, base, len, ppos);
2146                 else
2147                         nr = fuse_direct_read(filp, base, len, ppos);
2148
2149                 if (nr < 0) {
2150                         if (!ret)
2151                                 ret = nr;
2152                         break;
2153                 }
2154                 ret += nr;
2155                 if (nr != len)
2156                         break;
2157         }
2158
2159         return ret;
2160 }
2161
2162
2163 static ssize_t
2164 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2165                         loff_t offset, unsigned long nr_segs)
2166 {
2167         ssize_t ret = 0;
2168         struct file *file = NULL;
2169         loff_t pos = 0;
2170
2171         file = iocb->ki_filp;
2172         pos = offset;
2173
2174         ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
2175
2176         return ret;
2177 }
2178
2179 long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2180                             loff_t length)
2181 {
2182         struct fuse_file *ff = file->private_data;
2183         struct fuse_conn *fc = ff->fc;
2184         struct fuse_req *req;
2185         struct fuse_fallocate_in inarg = {
2186                 .fh = ff->fh,
2187                 .offset = offset,
2188                 .length = length,
2189                 .mode = mode
2190         };
2191         int err;
2192
2193         if (fc->no_fallocate)
2194                 return -EOPNOTSUPP;
2195
2196         req = fuse_get_req(fc);
2197         if (IS_ERR(req))
2198                 return PTR_ERR(req);
2199
2200         req->in.h.opcode = FUSE_FALLOCATE;
2201         req->in.h.nodeid = ff->nodeid;
2202         req->in.numargs = 1;
2203         req->in.args[0].size = sizeof(inarg);
2204         req->in.args[0].value = &inarg;
2205         fuse_request_send(fc, req);
2206         err = req->out.h.error;
2207         if (err == -ENOSYS) {
2208                 fc->no_fallocate = 1;
2209                 err = -EOPNOTSUPP;
2210         }
2211         fuse_put_request(fc, req);
2212
2213         return err;
2214 }
2215 EXPORT_SYMBOL_GPL(fuse_file_fallocate);
2216
2217 static const struct file_operations fuse_file_operations = {
2218         .llseek         = fuse_file_llseek,
2219         .read           = do_sync_read,
2220         .aio_read       = fuse_file_aio_read,
2221         .write          = do_sync_write,
2222         .aio_write      = fuse_file_aio_write,
2223         .mmap           = fuse_file_mmap,
2224         .open           = fuse_open,
2225         .flush          = fuse_flush,
2226         .release        = fuse_release,
2227         .fsync          = fuse_fsync,
2228         .lock           = fuse_file_lock,
2229         .flock          = fuse_file_flock,
2230         .splice_read    = generic_file_splice_read,
2231         .unlocked_ioctl = fuse_file_ioctl,
2232         .compat_ioctl   = fuse_file_compat_ioctl,
2233         .poll           = fuse_file_poll,
2234         .fallocate      = fuse_file_fallocate,
2235 };
2236
2237 static const struct file_operations fuse_direct_io_file_operations = {
2238         .llseek         = fuse_file_llseek,
2239         .read           = fuse_direct_read,
2240         .write          = fuse_direct_write,
2241         .mmap           = fuse_direct_mmap,
2242         .open           = fuse_open,
2243         .flush          = fuse_flush,
2244         .release        = fuse_release,
2245         .fsync          = fuse_fsync,
2246         .lock           = fuse_file_lock,
2247         .flock          = fuse_file_flock,
2248         .unlocked_ioctl = fuse_file_ioctl,
2249         .compat_ioctl   = fuse_file_compat_ioctl,
2250         .poll           = fuse_file_poll,
2251         .fallocate      = fuse_file_fallocate,
2252         /* no splice_read */
2253 };
2254
2255 static const struct address_space_operations fuse_file_aops  = {
2256         .readpage       = fuse_readpage,
2257         .writepage      = fuse_writepage,
2258         .launder_page   = fuse_launder_page,
2259         .readpages      = fuse_readpages,
2260         .set_page_dirty = __set_page_dirty_nobuffers,
2261         .bmap           = fuse_bmap,
2262         .direct_IO      = fuse_direct_IO,
2263 };
2264
2265 void fuse_init_file_inode(struct inode *inode)
2266 {
2267         inode->i_fop = &fuse_file_operations;
2268         inode->i_data.a_ops = &fuse_file_aops;
2269 }