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_nopages(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(0);
  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                         req->background = 0;
 130                         fuse_request_send(ff->fc, req);
 131                         path_put(&req->misc.release.path);
 132                         fuse_put_request(ff->fc, req);
 133                 } else {
 134                         req->end = fuse_release_end;
 135                         req->background = 1;
 136                         fuse_request_send_background(ff->fc, req);
 137                 }
 138                 kfree(ff);
 139         }
 140 }
 141
 142 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 143                  bool isdir)
 144 {
 145         struct fuse_open_out outarg;
 146         struct fuse_file *ff;
 147         int err;
 148         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 149
 150         ff = fuse_file_alloc(fc);
 151         if (!ff)
 152                 return -ENOMEM;
 153
 154         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 155         if (err) {
 156                 fuse_file_free(ff);
 157                 return err;
 158         }
 159
 160         if (isdir)
 161                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 162
 163         ff->fh = outarg.fh;
 164         ff->nodeid = nodeid;
 165         ff->open_flags = outarg.open_flags;
 166         file->private_data = fuse_file_get(ff);
 167
 168         return 0;
 169 }
 170 EXPORT_SYMBOL_GPL(fuse_do_open);
 171
 172 void fuse_finish_open(struct inode *inode, struct file *file)
 173 {
 174         struct fuse_file *ff = file->private_data;
 175         struct fuse_conn *fc = get_fuse_conn(inode);
 176
 177         if (ff->open_flags & FOPEN_DIRECT_IO)
 178                 file->f_op = &fuse_direct_io_file_operations;
 179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 180                 invalidate_inode_pages2(inode->i_mapping);
 181         if (ff->open_flags & FOPEN_NONSEEKABLE)
 182                 nonseekable_open(inode, file);
 183         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 184                 struct fuse_inode *fi = get_fuse_inode(inode);
 185
 186                 spin_lock(&fc->lock);
 187                 fi->attr_version = ++fc->attr_version;
 188                 i_size_write(inode, 0);
 189                 spin_unlock(&fc->lock);
 190                 fuse_invalidate_attr(inode);
 191         }
 192 }
 193
 194 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 195 {
 196         struct fuse_conn *fc = get_fuse_conn(inode);
 197         int err;
 198
 199         err = generic_file_open(inode, file);
 200         if (err)
 201                 return err;
 202
 203         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 204         if (err)
 205                 return err;
 206
 207         fuse_finish_open(inode, file);
 208
 209         return 0;
 210 }
 211
 212 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 213 {
 214         struct fuse_conn *fc = ff->fc;
 215         struct fuse_req *req = ff->reserved_req;
 216         struct fuse_release_in *inarg = &req->misc.release.in;
 217
 218         spin_lock(&fc->lock);
 219         list_del(&ff->write_entry);
 220         if (!RB_EMPTY_NODE(&ff->polled_node))
 221                 rb_erase(&ff->polled_node, &fc->polled_files);
 222         spin_unlock(&fc->lock);
 223
 224         wake_up_interruptible_all(&ff->poll_wait);
 225
 226         inarg->fh = ff->fh;
 227         inarg->flags = flags;
 228         req->in.h.opcode = opcode;
 229         req->in.h.nodeid = ff->nodeid;
 230         req->in.numargs = 1;
 231         req->in.args[0].size = sizeof(struct fuse_release_in);
 232         req->in.args[0].value = inarg;
 233 }
 234
 235 void fuse_release_common(struct file *file, int opcode)
 236 {
 237         struct fuse_file *ff;
 238         struct fuse_req *req;
 239
 240         ff = file->private_data;
 241         if (unlikely(!ff))
 242                 return;
 243
 244         req = ff->reserved_req;
 245         fuse_prepare_release(ff, file->f_flags, opcode);
 246
 247         if (ff->flock) {
 248                 struct fuse_release_in *inarg = &req->misc.release.in;
 249                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 250                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 251                                                        (fl_owner_t) file);
 252         }
 253         /* Hold vfsmount and dentry until release is finished */
 254         path_get(&file->f_path);
 255         req->misc.release.path = file->f_path;
 256
 257         /*
 258          * Normally this will send the RELEASE request, however if
 259          * some asynchronous READ or WRITE requests are outstanding,
 260          * the sending will be delayed.
 261          *
 262          * Make the release synchronous if this is a fuseblk mount,
 263          * synchronous RELEASE is allowed (and desirable) in this case
 264          * because the server can be trusted not to screw up.
 265          */
 266         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 267 }
 268
 269 static int fuse_open(struct inode *inode, struct file *file)
 270 {
 271         return fuse_open_common(inode, file, false);
 272 }
 273
 274 static int fuse_release(struct inode *inode, struct file *file)
 275 {
 276         fuse_release_common(file, FUSE_RELEASE);
 277
 278         /* return value is ignored by VFS */
 279         return 0;
 280 }
 281
 282 void fuse_sync_release(struct fuse_file *ff, int flags)
 283 {
 284         WARN_ON(atomic_read(&ff->count) > 1);
 285         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 286         ff->reserved_req->force = 1;
 287         ff->reserved_req->background = 0;
 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_inode(file);
 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_nopages(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, loff_t start, loff_t end,
 410                       int datasync, int isdir)
 411 {
 412         struct inode *inode = file->f_mapping->host;
 413         struct fuse_conn *fc = get_fuse_conn(inode);
 414         struct fuse_file *ff = file->private_data;
 415         struct fuse_req *req;
 416         struct fuse_fsync_in inarg;
 417         int err;
 418
 419         if (is_bad_inode(inode))
 420                 return -EIO;
 421
 422         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 423         if (err)
 424                 return err;
 425
 426         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 427                 return 0;
 428
 429         mutex_lock(&inode->i_mutex);
 430
 431         /*
 432          * Start writeback against all dirty pages of the inode, then
 433          * wait for all outstanding writes, before sending the FSYNC
 434          * request.
 435          */
 436         err = write_inode_now(inode, 0);
 437         if (err)
 438                 goto out;
 439
 440         fuse_sync_writes(inode);
 441
 442         req = fuse_get_req_nopages(fc);
 443         if (IS_ERR(req)) {
 444                 err = PTR_ERR(req);
 445                 goto out;
 446         }
 447
 448         memset(&inarg, 0, sizeof(inarg));
 449         inarg.fh = ff->fh;
 450         inarg.fsync_flags = datasync ? 1 : 0;
 451         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 452         req->in.h.nodeid = get_node_id(inode);
 453         req->in.numargs = 1;
 454         req->in.args[0].size = sizeof(inarg);
 455         req->in.args[0].value = &inarg;
 456         fuse_request_send(fc, req);
 457         err = req->out.h.error;
 458         fuse_put_request(fc, req);
 459         if (err == -ENOSYS) {
 460                 if (isdir)
 461                         fc->no_fsyncdir = 1;
 462                 else
 463                         fc->no_fsync = 1;
 464                 err = 0;
 465         }
 466 out:
 467         mutex_unlock(&inode->i_mutex);
 468         return err;
 469 }
 470
 471 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 472                       int datasync)
 473 {
 474         return fuse_fsync_common(file, start, end, datasync, 0);
 475 }
 476
 477 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 478                     size_t count, int opcode)
 479 {
 480         struct fuse_read_in *inarg = &req->misc.read.in;
 481         struct fuse_file *ff = file->private_data;
 482
 483         inarg->fh = ff->fh;
 484         inarg->offset = pos;
 485         inarg->size = count;
 486         inarg->flags = file->f_flags;
 487         req->in.h.opcode = opcode;
 488         req->in.h.nodeid = ff->nodeid;
 489         req->in.numargs = 1;
 490         req->in.args[0].size = sizeof(struct fuse_read_in);
 491         req->in.args[0].value = inarg;
 492         req->out.argvar = 1;
 493         req->out.numargs = 1;
 494         req->out.args[0].size = count;
 495 }
 496
 497 static void fuse_release_user_pages(struct fuse_req *req, int write)
 498 {
 499         unsigned i;
 500
 501         for (i = 0; i < req->num_pages; i++) {
 502                 struct page *page = req->pages[i];
 503                 if (write)
 504                         set_page_dirty_lock(page);
 505                 put_page(page);
 506         }
 507 }
 508
 509 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 510                              loff_t pos, size_t count, fl_owner_t owner)
 511 {
 512         struct fuse_file *ff = file->private_data;
 513         struct fuse_conn *fc = ff->fc;
 514
 515         fuse_read_fill(req, file, pos, count, FUSE_READ);
 516         if (owner != NULL) {
 517                 struct fuse_read_in *inarg = &req->misc.read.in;
 518
 519                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 520                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 521         }
 522         fuse_request_send(fc, req);
 523         return req->out.args[0].size;
 524 }
 525
 526 static void fuse_read_update_size(struct inode *inode, loff_t size,
 527                                   u64 attr_ver)
 528 {
 529         struct fuse_conn *fc = get_fuse_conn(inode);
 530         struct fuse_inode *fi = get_fuse_inode(inode);
 531
 532         spin_lock(&fc->lock);
 533         if (attr_ver == fi->attr_version && size < inode->i_size) {
 534                 fi->attr_version = ++fc->attr_version;
 535                 i_size_write(inode, size);
 536         }
 537         spin_unlock(&fc->lock);
 538 }
 539
 540 static int fuse_readpage(struct file *file, struct page *page)
 541 {
 542         struct inode *inode = page->mapping->host;
 543         struct fuse_conn *fc = get_fuse_conn(inode);
 544         struct fuse_req *req;
 545         size_t num_read;
 546         loff_t pos = page_offset(page);
 547         size_t count = PAGE_CACHE_SIZE;
 548         u64 attr_ver;
 549         int err;
 550
 551         err = -EIO;
 552         if (is_bad_inode(inode))
 553                 goto out;
 554
 555         /*
 556          * Page writeback can extend beyond the lifetime of the
 557          * page-cache page, so make sure we read a properly synced
 558          * page.
 559          */
 560         fuse_wait_on_page_writeback(inode, page->index);
 561
 562         req = fuse_get_req(fc, 1);
 563         err = PTR_ERR(req);
 564         if (IS_ERR(req))
 565                 goto out;
 566
 567         attr_ver = fuse_get_attr_version(fc);
 568
 569         req->out.page_zeroing = 1;
 570         req->out.argpages = 1;
 571         req->num_pages = 1;
 572         req->pages[0] = page;
 573         req->page_descs[0].length = count;
 574         num_read = fuse_send_read(req, file, pos, count, NULL);
 575         err = req->out.h.error;
 576         fuse_put_request(fc, req);
 577
 578         if (!err) {
 579                 /*
 580                  * Short read means EOF.  If file size is larger, truncate it
 581                  */
 582                 if (num_read < count)
 583                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 584
 585                 SetPageUptodate(page);
 586         }
 587
 588         fuse_invalidate_attr(inode); /* atime changed */
 589  out:
 590         unlock_page(page);
 591         return err;
 592 }
 593
 594 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 595 {
 596         int i;
 597         size_t count = req->misc.read.in.size;
 598         size_t num_read = req->out.args[0].size;
 599         struct address_space *mapping = NULL;
 600
 601         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 602                 mapping = req->pages[i]->mapping;
 603
 604         if (mapping) {
 605                 struct inode *inode = mapping->host;
 606
 607                 /*
 608                  * Short read means EOF. If file size is larger, truncate it
 609                  */
 610                 if (!req->out.h.error && num_read < count) {
 611                         loff_t pos;
 612
 613                         pos = page_offset(req->pages[0]) + num_read;
 614                         fuse_read_update_size(inode, pos,
 615                                               req->misc.read.attr_ver);
 616                 }
 617                 fuse_invalidate_attr(inode); /* atime changed */
 618         }
 619
 620         for (i = 0; i < req->num_pages; i++) {
 621                 struct page *page = req->pages[i];
 622                 if (!req->out.h.error)
 623                         SetPageUptodate(page);
 624                 else
 625                         SetPageError(page);
 626                 unlock_page(page);
 627                 page_cache_release(page);
 628         }
 629         if (req->ff)
 630                 fuse_file_put(req->ff, false);
 631 }
 632
 633 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 634 {
 635         struct fuse_file *ff = file->private_data;
 636         struct fuse_conn *fc = ff->fc;
 637         loff_t pos = page_offset(req->pages[0]);
 638         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 639
 640         req->out.argpages = 1;
 641         req->out.page_zeroing = 1;
 642         req->out.page_replace = 1;
 643         fuse_read_fill(req, file, pos, count, FUSE_READ);
 644         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 645         if (fc->async_read) {
 646                 req->ff = fuse_file_get(ff);
 647                 req->end = fuse_readpages_end;
 648                 fuse_request_send_background(fc, req);
 649         } else {
 650                 fuse_request_send(fc, req);
 651                 fuse_readpages_end(fc, req);
 652                 fuse_put_request(fc, req);
 653         }
 654 }
 655
 656 struct fuse_fill_data {
 657         struct fuse_req *req;
 658         struct file *file;
 659         struct inode *inode;
 660         unsigned nr_pages;
 661 };
 662
 663 static int fuse_readpages_fill(void *_data, struct page *page)
 664 {
 665         struct fuse_fill_data *data = _data;
 666         struct fuse_req *req = data->req;
 667         struct inode *inode = data->inode;
 668         struct fuse_conn *fc = get_fuse_conn(inode);
 669
 670         fuse_wait_on_page_writeback(inode, page->index);
 671
 672         if (req->num_pages &&
 673             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 674              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 675              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 676                 int nr_alloc = min_t(unsigned, data->nr_pages,
 677                                      FUSE_MAX_PAGES_PER_REQ);
 678                 fuse_send_readpages(req, data->file);
 679                 if (fc->async_read)
 680                         req = fuse_get_req_for_background(fc, nr_alloc);
 681                 else
 682                         req = fuse_get_req(fc, nr_alloc);
 683
 684                 data->req = req;
 685                 if (IS_ERR(req)) {
 686                         unlock_page(page);
 687                         return PTR_ERR(req);
 688                 }
 689         }
 690
 691         if (WARN_ON(req->num_pages >= req->max_pages)) {
 692                 fuse_put_request(fc, req);
 693                 return -EIO;
 694         }
 695
 696         page_cache_get(page);
 697         req->pages[req->num_pages] = page;
 698         req->page_descs[req->num_pages].length = PAGE_SIZE;
 699         req->num_pages++;
 700         data->nr_pages--;
 701         return 0;
 702 }
 703
 704 static int fuse_readpages(struct file *file, struct address_space *mapping,
 705                           struct list_head *pages, unsigned nr_pages)
 706 {
 707         struct inode *inode = mapping->host;
 708         struct fuse_conn *fc = get_fuse_conn(inode);
 709         struct fuse_fill_data data;
 710         int err;
 711         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 712
 713         err = -EIO;
 714         if (is_bad_inode(inode))
 715                 goto out;
 716
 717         data.file = file;
 718         data.inode = inode;
 719         if (fc->async_read)
 720                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 721         else
 722                 data.req = fuse_get_req(fc, nr_alloc);
 723         data.nr_pages = nr_pages;
 724         err = PTR_ERR(data.req);
 725         if (IS_ERR(data.req))
 726                 goto out;
 727
 728         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 729         if (!err) {
 730                 if (data.req->num_pages)
 731                         fuse_send_readpages(data.req, file);
 732                 else
 733                         fuse_put_request(fc, data.req);
 734         }
 735 out:
 736         return err;
 737 }
 738
 739 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 740                                   unsigned long nr_segs, loff_t pos)
 741 {
 742         struct inode *inode = iocb->ki_filp->f_mapping->host;
 743         struct fuse_conn *fc = get_fuse_conn(inode);
 744
 745         /*
 746          * In auto invalidate mode, always update attributes on read.
 747          * Otherwise, only update if we attempt to read past EOF (to ensure
 748          * i_size is up to date).
 749          */
 750         if (fc->auto_inval_data ||
 751             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 752                 int err;
 753                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 754                 if (err)
 755                         return err;
 756         }
 757
 758         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 759 }
 760
 761 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 762                             loff_t pos, size_t count)
 763 {
 764         struct fuse_write_in *inarg = &req->misc.write.in;
 765         struct fuse_write_out *outarg = &req->misc.write.out;
 766
 767         inarg->fh = ff->fh;
 768         inarg->offset = pos;
 769         inarg->size = count;
 770         req->in.h.opcode = FUSE_WRITE;
 771         req->in.h.nodeid = ff->nodeid;
 772         req->in.numargs = 2;
 773         if (ff->fc->minor < 9)
 774                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 775         else
 776                 req->in.args[0].size = sizeof(struct fuse_write_in);
 777         req->in.args[0].value = inarg;
 778         req->in.args[1].size = count;
 779         req->out.numargs = 1;
 780         req->out.args[0].size = sizeof(struct fuse_write_out);
 781         req->out.args[0].value = outarg;
 782 }
 783
 784 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 785                               loff_t pos, size_t count, fl_owner_t owner)
 786 {
 787         struct fuse_file *ff = file->private_data;
 788         struct fuse_conn *fc = ff->fc;
 789         struct fuse_write_in *inarg = &req->misc.write.in;
 790
 791         fuse_write_fill(req, ff, pos, count);
 792         inarg->flags = file->f_flags;
 793         if (owner != NULL) {
 794                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 795                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 796         }
 797         fuse_request_send(fc, req);
 798         return req->misc.write.out.size;
 799 }
 800
 801 void fuse_write_update_size(struct inode *inode, loff_t pos)
 802 {
 803         struct fuse_conn *fc = get_fuse_conn(inode);
 804         struct fuse_inode *fi = get_fuse_inode(inode);
 805
 806         spin_lock(&fc->lock);
 807         fi->attr_version = ++fc->attr_version;
 808         if (pos > inode->i_size)
 809                 i_size_write(inode, pos);
 810         spin_unlock(&fc->lock);
 811 }
 812
 813 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 814                                     struct inode *inode, loff_t pos,
 815                                     size_t count)
 816 {
 817         size_t res;
 818         unsigned offset;
 819         unsigned i;
 820
 821         for (i = 0; i < req->num_pages; i++)
 822                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 823
 824         res = fuse_send_write(req, file, pos, count, NULL);
 825
 826         offset = req->page_descs[0].offset;
 827         count = res;
 828         for (i = 0; i < req->num_pages; i++) {
 829                 struct page *page = req->pages[i];
 830
 831                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 832                         SetPageUptodate(page);
 833
 834                 if (count > PAGE_CACHE_SIZE - offset)
 835                         count -= PAGE_CACHE_SIZE - offset;
 836                 else
 837                         count = 0;
 838                 offset = 0;
 839
 840                 unlock_page(page);
 841                 page_cache_release(page);
 842         }
 843
 844         return res;
 845 }
 846
 847 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 848                                struct address_space *mapping,
 849                                struct iov_iter *ii, loff_t pos)
 850 {
 851         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 852         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 853         size_t count = 0;
 854         int err;
 855
 856         req->in.argpages = 1;
 857         req->page_descs[0].offset = offset;
 858
 859         do {
 860                 size_t tmp;
 861                 struct page *page;
 862                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 863                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 864                                      iov_iter_count(ii));
 865
 866                 bytes = min_t(size_t, bytes, fc->max_write - count);
 867
 868  again:
 869                 err = -EFAULT;
 870                 if (iov_iter_fault_in_readable(ii, bytes))
 871                         break;
 872
 873                 err = -ENOMEM;
 874                 page = grab_cache_page_write_begin(mapping, index, 0);
 875                 if (!page)
 876                         break;
 877
 878                 if (mapping_writably_mapped(mapping))
 879                         flush_dcache_page(page);
 880
 881                 pagefault_disable();
 882                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 883                 pagefault_enable();
 884                 flush_dcache_page(page);
 885
 886                 mark_page_accessed(page);
 887
 888                 if (!tmp) {
 889                         unlock_page(page);
 890                         page_cache_release(page);
 891                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 892                         goto again;
 893                 }
 894
 895                 err = 0;
 896                 req->pages[req->num_pages] = page;
 897                 req->page_descs[req->num_pages].length = tmp;
 898                 req->num_pages++;
 899
 900                 iov_iter_advance(ii, tmp);
 901                 count += tmp;
 902                 pos += tmp;
 903                 offset += tmp;
 904                 if (offset == PAGE_CACHE_SIZE)
 905                         offset = 0;
 906
 907                 if (!fc->big_writes)
 908                         break;
 909         } while (iov_iter_count(ii) && count < fc->max_write &&
 910                  req->num_pages < req->max_pages && offset == 0);
 911
 912         return count > 0 ? count : err;
 913 }
 914
 915 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
 916 {
 917         return min_t(unsigned,
 918                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
 919                      (pos >> PAGE_CACHE_SHIFT) + 1,
 920                      FUSE_MAX_PAGES_PER_REQ);
 921 }
 922
 923 static ssize_t fuse_perform_write(struct file *file,
 924                                   struct address_space *mapping,
 925                                   struct iov_iter *ii, loff_t pos)
 926 {
 927         struct inode *inode = mapping->host;
 928         struct fuse_conn *fc = get_fuse_conn(inode);
 929         int err = 0;
 930         ssize_t res = 0;
 931
 932         if (is_bad_inode(inode))
 933                 return -EIO;
 934
 935         do {
 936                 struct fuse_req *req;
 937                 ssize_t count;
 938                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
 939
 940                 req = fuse_get_req(fc, nr_pages);
 941                 if (IS_ERR(req)) {
 942                         err = PTR_ERR(req);
 943                         break;
 944                 }
 945
 946                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 947                 if (count <= 0) {
 948                         err = count;
 949                 } else {
 950                         size_t num_written;
 951
 952                         num_written = fuse_send_write_pages(req, file, inode,
 953                                                             pos, count);
 954                         err = req->out.h.error;
 955                         if (!err) {
 956                                 res += num_written;
 957                                 pos += num_written;
 958
 959                                 /* break out of the loop on short write */
 960                                 if (num_written != count)
 961                                         err = -EIO;
 962                         }
 963                 }
 964                 fuse_put_request(fc, req);
 965         } while (!err && iov_iter_count(ii));
 966
 967         if (res > 0)
 968                 fuse_write_update_size(inode, pos);
 969
 970         fuse_invalidate_attr(inode);
 971
 972         return res > 0 ? res : err;
 973 }
 974
 975 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 976                                    unsigned long nr_segs, loff_t pos)
 977 {
 978         struct file *file = iocb->ki_filp;
 979         struct address_space *mapping = file->f_mapping;
 980         size_t count = 0;
 981         size_t ocount = 0;
 982         ssize_t written = 0;
 983         ssize_t written_buffered = 0;
 984         struct inode *inode = mapping->host;
 985         ssize_t err;
 986         struct iov_iter i;
 987         loff_t endbyte = 0;
 988
 989         WARN_ON(iocb->ki_pos != pos);
 990
 991         ocount = 0;
 992         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 993         if (err)
 994                 return err;
 995
 996         count = ocount;
 997         sb_start_write(inode->i_sb);
 998         mutex_lock(&inode->i_mutex);
 999
1000         /* We can write back this queue in page reclaim */
1001         current->backing_dev_info = mapping->backing_dev_info;
1002
1003         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1004         if (err)
1005                 goto out;
1006
1007         if (count == 0)
1008                 goto out;
1009
1010         err = file_remove_suid(file);
1011         if (err)
1012                 goto out;
1013
1014         err = file_update_time(file);
1015         if (err)
1016                 goto out;
1017
1018         if (file->f_flags & O_DIRECT) {
1019                 written = generic_file_direct_write(iocb, iov, &nr_segs,
1020                                                     pos, &iocb->ki_pos,
1021                                                     count, ocount);
1022                 if (written < 0 || written == count)
1023                         goto out;
1024
1025                 pos += written;
1026                 count -= written;
1027
1028                 iov_iter_init(&i, iov, nr_segs, count, written);
1029                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1030                 if (written_buffered < 0) {
1031                         err = written_buffered;
1032                         goto out;
1033                 }
1034                 endbyte = pos + written_buffered - 1;
1035
1036                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1037                                                    endbyte);
1038                 if (err)
1039                         goto out;
1040
1041                 invalidate_mapping_pages(file->f_mapping,
1042                                          pos >> PAGE_CACHE_SHIFT,
1043                                          endbyte >> PAGE_CACHE_SHIFT);
1044
1045                 written += written_buffered;
1046                 iocb->ki_pos = pos + written_buffered;
1047         } else {
1048                 iov_iter_init(&i, iov, nr_segs, count, 0);
1049                 written = fuse_perform_write(file, mapping, &i, pos);
1050                 if (written >= 0)
1051                         iocb->ki_pos = pos + written;
1052         }
1053 out:
1054         current->backing_dev_info = NULL;
1055         mutex_unlock(&inode->i_mutex);
1056         sb_end_write(inode->i_sb);
1057
1058         return written ? written : err;
1059 }
1060
1061 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1062                 unsigned index, unsigned nr_pages)
1063 {
1064         int i;
1065
1066         for (i = index; i < index + nr_pages; i++)
1067                 req->page_descs[i].length = PAGE_SIZE -
1068                         req->page_descs[i].offset;
1069 }
1070
1071 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1072 {
1073         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1074 }
1075
1076 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1077                                         size_t max_size)
1078 {
1079         return min(iov_iter_single_seg_count(ii), max_size);
1080 }
1081
1082 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1083                                size_t *nbytesp, int write)
1084 {
1085         size_t nbytes = 0;  /* # bytes already packed in req */
1086
1087         /* Special case for kernel I/O: can copy directly into the buffer */
1088         if (segment_eq(get_fs(), KERNEL_DS)) {
1089                 unsigned long user_addr = fuse_get_user_addr(ii);
1090                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1091
1092                 if (write)
1093                         req->in.args[1].value = (void *) user_addr;
1094                 else
1095                         req->out.args[0].value = (void *) user_addr;
1096
1097                 iov_iter_advance(ii, frag_size);
1098                 *nbytesp = frag_size;
1099                 return 0;
1100         }
1101
1102         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1103                 unsigned npages;
1104                 unsigned long user_addr = fuse_get_user_addr(ii);
1105                 unsigned offset = user_addr & ~PAGE_MASK;
1106                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1107                 int ret;
1108
1109                 unsigned n = req->max_pages - req->num_pages;
1110                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1111
1112                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1113                 npages = clamp(npages, 1U, n);
1114
1115                 ret = get_user_pages_fast(user_addr, npages, !write,
1116                                           &req->pages[req->num_pages]);
1117                 if (ret < 0)
1118                         return ret;
1119
1120                 npages = ret;
1121                 frag_size = min_t(size_t, frag_size,
1122                                   (npages << PAGE_SHIFT) - offset);
1123                 iov_iter_advance(ii, frag_size);
1124
1125                 req->page_descs[req->num_pages].offset = offset;
1126                 fuse_page_descs_length_init(req, req->num_pages, npages);
1127
1128                 req->num_pages += npages;
1129                 req->page_descs[req->num_pages - 1].length -=
1130                         (npages << PAGE_SHIFT) - offset - frag_size;
1131
1132                 nbytes += frag_size;
1133         }
1134
1135         if (write)
1136                 req->in.argpages = 1;
1137         else
1138                 req->out.argpages = 1;
1139
1140         *nbytesp = nbytes;
1141
1142         return 0;
1143 }
1144
1145 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1146 {
1147         struct iov_iter ii = *ii_p;
1148         int npages = 0;
1149
1150         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1151                 unsigned long user_addr = fuse_get_user_addr(&ii);
1152                 unsigned offset = user_addr & ~PAGE_MASK;
1153                 size_t frag_size = iov_iter_single_seg_count(&ii);
1154
1155                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1156                 iov_iter_advance(&ii, frag_size);
1157         }
1158
1159         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1160 }
1161
1162 ssize_t fuse_direct_io(struct file *file, const struct iovec *iov,
1163                        unsigned long nr_segs, size_t count, loff_t *ppos,
1164                        int write)
1165 {
1166         struct fuse_file *ff = file->private_data;
1167         struct fuse_conn *fc = ff->fc;
1168         size_t nmax = write ? fc->max_write : fc->max_read;
1169         loff_t pos = *ppos;
1170         ssize_t res = 0;
1171         struct fuse_req *req;
1172         struct iov_iter ii;
1173
1174         iov_iter_init(&ii, iov, nr_segs, count, 0);
1175
1176         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1177         if (IS_ERR(req))
1178                 return PTR_ERR(req);
1179
1180         while (count) {
1181                 size_t nres;
1182                 fl_owner_t owner = current->files;
1183                 size_t nbytes = min(count, nmax);
1184                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1185                 if (err) {
1186                         res = err;
1187                         break;
1188                 }
1189
1190                 if (write)
1191                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1192                 else
1193                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1194
1195                 fuse_release_user_pages(req, !write);
1196                 if (req->out.h.error) {
1197                         if (!res)
1198                                 res = req->out.h.error;
1199                         break;
1200                 } else if (nres > nbytes) {
1201                         res = -EIO;
1202                         break;
1203                 }
1204                 count -= nres;
1205                 res += nres;
1206                 pos += nres;
1207                 if (nres != nbytes)
1208                         break;
1209                 if (count) {
1210                         fuse_put_request(fc, req);
1211                         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1212                         if (IS_ERR(req))
1213                                 break;
1214                 }
1215         }
1216         if (!IS_ERR(req))
1217                 fuse_put_request(fc, req);
1218         if (res > 0)
1219                 *ppos = pos;
1220
1221         return res;
1222 }
1223 EXPORT_SYMBOL_GPL(fuse_direct_io);
1224
1225 static ssize_t __fuse_direct_read(struct file *file, const struct iovec *iov,
1226                                   unsigned long nr_segs, loff_t *ppos)
1227 {
1228         ssize_t res;
1229         struct inode *inode = file_inode(file);
1230
1231         if (is_bad_inode(inode))
1232                 return -EIO;
1233
1234         res = fuse_direct_io(file, iov, nr_segs, iov_length(iov, nr_segs),
1235                              ppos, 0);
1236
1237         fuse_invalidate_attr(inode);
1238
1239         return res;
1240 }
1241
1242 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1243                                      size_t count, loff_t *ppos)
1244 {
1245         struct iovec iov = { .iov_base = buf, .iov_len = count };
1246         return __fuse_direct_read(file, &iov, 1, ppos);
1247 }
1248
1249 static ssize_t __fuse_direct_write(struct file *file, const struct iovec *iov,
1250                                    unsigned long nr_segs, loff_t *ppos)
1251 {
1252         struct inode *inode = file_inode(file);
1253         size_t count = iov_length(iov, nr_segs);
1254         ssize_t res;
1255
1256         res = generic_write_checks(file, ppos, &count, 0);
1257         if (!res) {
1258                 res = fuse_direct_io(file, iov, nr_segs, count, ppos, 1);
1259                 if (res > 0)
1260                         fuse_write_update_size(inode, *ppos);
1261         }
1262
1263         fuse_invalidate_attr(inode);
1264
1265         return res;
1266 }
1267
1268 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1269                                  size_t count, loff_t *ppos)
1270 {
1271         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1272         struct inode *inode = file_inode(file);
1273         ssize_t res;
1274
1275         if (is_bad_inode(inode))
1276                 return -EIO;
1277
1278         /* Don't allow parallel writes to the same file */
1279         mutex_lock(&inode->i_mutex);
1280         res = __fuse_direct_write(file, &iov, 1, ppos);
1281         mutex_unlock(&inode->i_mutex);
1282
1283         return res;
1284 }
1285
1286 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1287 {
1288         __free_page(req->pages[0]);
1289         fuse_file_put(req->ff, false);
1290 }
1291
1292 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1293 {
1294         struct inode *inode = req->inode;
1295         struct fuse_inode *fi = get_fuse_inode(inode);
1296         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1297
1298         list_del(&req->writepages_entry);
1299         dec_bdi_stat(bdi, BDI_WRITEBACK);
1300         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1301         bdi_writeout_inc(bdi);
1302         wake_up(&fi->page_waitq);
1303 }
1304
1305 /* Called under fc->lock, may release and reacquire it */
1306 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1307 __releases(fc->lock)
1308 __acquires(fc->lock)
1309 {
1310         struct fuse_inode *fi = get_fuse_inode(req->inode);
1311         loff_t size = i_size_read(req->inode);
1312         struct fuse_write_in *inarg = &req->misc.write.in;
1313
1314         if (!fc->connected)
1315                 goto out_free;
1316
1317         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1318                 inarg->size = PAGE_CACHE_SIZE;
1319         } else if (inarg->offset < size) {
1320                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1321         } else {
1322                 /* Got truncated off completely */
1323                 goto out_free;
1324         }
1325
1326         req->in.args[1].size = inarg->size;
1327         fi->writectr++;
1328         fuse_request_send_background_locked(fc, req);
1329         return;
1330
1331  out_free:
1332         fuse_writepage_finish(fc, req);
1333         spin_unlock(&fc->lock);
1334         fuse_writepage_free(fc, req);
1335         fuse_put_request(fc, req);
1336         spin_lock(&fc->lock);
1337 }
1338
1339 /*
1340  * If fi->writectr is positive (no truncate or fsync going on) send
1341  * all queued writepage requests.
1342  *
1343  * Called with fc->lock
1344  */
1345 void fuse_flush_writepages(struct inode *inode)
1346 __releases(fc->lock)
1347 __acquires(fc->lock)
1348 {
1349         struct fuse_conn *fc = get_fuse_conn(inode);
1350         struct fuse_inode *fi = get_fuse_inode(inode);
1351         struct fuse_req *req;
1352
1353         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1354                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1355                 list_del_init(&req->list);
1356                 fuse_send_writepage(fc, req);
1357         }
1358 }
1359
1360 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1361 {
1362         struct inode *inode = req->inode;
1363         struct fuse_inode *fi = get_fuse_inode(inode);
1364
1365         mapping_set_error(inode->i_mapping, req->out.h.error);
1366         spin_lock(&fc->lock);
1367         fi->writectr--;
1368         fuse_writepage_finish(fc, req);
1369         spin_unlock(&fc->lock);
1370         fuse_writepage_free(fc, req);
1371 }
1372
1373 static int fuse_writepage_locked(struct page *page)
1374 {
1375         struct address_space *mapping = page->mapping;
1376         struct inode *inode = mapping->host;
1377         struct fuse_conn *fc = get_fuse_conn(inode);
1378         struct fuse_inode *fi = get_fuse_inode(inode);
1379         struct fuse_req *req;
1380         struct fuse_file *ff;
1381         struct page *tmp_page;
1382
1383         set_page_writeback(page);
1384
1385         req = fuse_request_alloc_nofs(1);
1386         if (!req)
1387                 goto err;
1388
1389         req->background = 1; /* writeback always goes to bg_queue */
1390         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1391         if (!tmp_page)
1392                 goto err_free;
1393
1394         spin_lock(&fc->lock);
1395         BUG_ON(list_empty(&fi->write_files));
1396         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1397         req->ff = fuse_file_get(ff);
1398         spin_unlock(&fc->lock);
1399
1400         fuse_write_fill(req, ff, page_offset(page), 0);
1401
1402         copy_highpage(tmp_page, page);
1403         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1404         req->in.argpages = 1;
1405         req->num_pages = 1;
1406         req->pages[0] = tmp_page;
1407         req->page_descs[0].offset = 0;
1408         req->page_descs[0].length = PAGE_SIZE;
1409         req->end = fuse_writepage_end;
1410         req->inode = inode;
1411
1412         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1413         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1414         end_page_writeback(page);
1415
1416         spin_lock(&fc->lock);
1417         list_add(&req->writepages_entry, &fi->writepages);
1418         list_add_tail(&req->list, &fi->queued_writes);
1419         fuse_flush_writepages(inode);
1420         spin_unlock(&fc->lock);
1421
1422         return 0;
1423
1424 err_free:
1425         fuse_request_free(req);
1426 err:
1427         end_page_writeback(page);
1428         return -ENOMEM;
1429 }
1430
1431 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1432 {
1433         int err;
1434
1435         err = fuse_writepage_locked(page);
1436         unlock_page(page);
1437
1438         return err;
1439 }
1440
1441 static int fuse_launder_page(struct page *page)
1442 {
1443         int err = 0;
1444         if (clear_page_dirty_for_io(page)) {
1445                 struct inode *inode = page->mapping->host;
1446                 err = fuse_writepage_locked(page);
1447                 if (!err)
1448                         fuse_wait_on_page_writeback(inode, page->index);
1449         }
1450         return err;
1451 }
1452
1453 /*
1454  * Write back dirty pages now, because there may not be any suitable
1455  * open files later
1456  */
1457 static void fuse_vma_close(struct vm_area_struct *vma)
1458 {
1459         filemap_write_and_wait(vma->vm_file->f_mapping);
1460 }
1461
1462 /*
1463  * Wait for writeback against this page to complete before allowing it
1464  * to be marked dirty again, and hence written back again, possibly
1465  * before the previous writepage completed.
1466  *
1467  * Block here, instead of in ->writepage(), so that the userspace fs
1468  * can only block processes actually operating on the filesystem.
1469  *
1470  * Otherwise unprivileged userspace fs would be able to block
1471  * unrelated:
1472  *
1473  * - page migration
1474  * - sync(2)
1475  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1476  */
1477 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1478 {
1479         struct page *page = vmf->page;
1480         /*
1481          * Don't use page->mapping as it may become NULL from a
1482          * concurrent truncate.
1483          */
1484         struct inode *inode = vma->vm_file->f_mapping->host;
1485
1486         fuse_wait_on_page_writeback(inode, page->index);
1487         return 0;
1488 }
1489
1490 static const struct vm_operations_struct fuse_file_vm_ops = {
1491         .close          = fuse_vma_close,
1492         .fault          = filemap_fault,
1493         .page_mkwrite   = fuse_page_mkwrite,
1494         .remap_pages    = generic_file_remap_pages,
1495 };
1496
1497 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1498 {
1499         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1500                 struct inode *inode = file_inode(file);
1501                 struct fuse_conn *fc = get_fuse_conn(inode);
1502                 struct fuse_inode *fi = get_fuse_inode(inode);
1503                 struct fuse_file *ff = file->private_data;
1504                 /*
1505                  * file may be written through mmap, so chain it onto the
1506                  * inodes's write_file list
1507                  */
1508                 spin_lock(&fc->lock);
1509                 if (list_empty(&ff->write_entry))
1510                         list_add(&ff->write_entry, &fi->write_files);
1511                 spin_unlock(&fc->lock);
1512         }
1513         file_accessed(file);
1514         vma->vm_ops = &fuse_file_vm_ops;
1515         return 0;
1516 }
1517
1518 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1519 {
1520         /* Can't provide the coherency needed for MAP_SHARED */
1521         if (vma->vm_flags & VM_MAYSHARE)
1522                 return -ENODEV;
1523
1524         invalidate_inode_pages2(file->f_mapping);
1525
1526         return generic_file_mmap(file, vma);
1527 }
1528
1529 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1530                                   struct file_lock *fl)
1531 {
1532         switch (ffl->type) {
1533         case F_UNLCK:
1534                 break;
1535
1536         case F_RDLCK:
1537         case F_WRLCK:
1538                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1539                     ffl->end < ffl->start)
1540                         return -EIO;
1541
1542                 fl->fl_start = ffl->start;
1543                 fl->fl_end = ffl->end;
1544                 fl->fl_pid = ffl->pid;
1545                 break;
1546
1547         default:
1548                 return -EIO;
1549         }
1550         fl->fl_type = ffl->type;
1551         return 0;
1552 }
1553
1554 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1555                          const struct file_lock *fl, int opcode, pid_t pid,
1556                          int flock)
1557 {
1558         struct inode *inode = file_inode(file);
1559         struct fuse_conn *fc = get_fuse_conn(inode);
1560         struct fuse_file *ff = file->private_data;
1561         struct fuse_lk_in *arg = &req->misc.lk_in;
1562
1563         arg->fh = ff->fh;
1564         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1565         arg->lk.start = fl->fl_start;
1566         arg->lk.end = fl->fl_end;
1567         arg->lk.type = fl->fl_type;
1568         arg->lk.pid = pid;
1569         if (flock)
1570                 arg->lk_flags |= FUSE_LK_FLOCK;
1571         req->in.h.opcode = opcode;
1572         req->in.h.nodeid = get_node_id(inode);
1573         req->in.numargs = 1;
1574         req->in.args[0].size = sizeof(*arg);
1575         req->in.args[0].value = arg;
1576 }
1577
1578 static int fuse_getlk(struct file *file, struct file_lock *fl)
1579 {
1580         struct inode *inode = file_inode(file);
1581         struct fuse_conn *fc = get_fuse_conn(inode);
1582         struct fuse_req *req;
1583         struct fuse_lk_out outarg;
1584         int err;
1585
1586         req = fuse_get_req_nopages(fc);
1587         if (IS_ERR(req))
1588                 return PTR_ERR(req);
1589
1590         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1591         req->out.numargs = 1;
1592         req->out.args[0].size = sizeof(outarg);
1593         req->out.args[0].value = &outarg;
1594         fuse_request_send(fc, req);
1595         err = req->out.h.error;
1596         fuse_put_request(fc, req);
1597         if (!err)
1598                 err = convert_fuse_file_lock(&outarg.lk, fl);
1599
1600         return err;
1601 }
1602
1603 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1604 {
1605         struct inode *inode = file_inode(file);
1606         struct fuse_conn *fc = get_fuse_conn(inode);
1607         struct fuse_req *req;
1608         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1609         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1610         int err;
1611
1612         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1613                 /* NLM needs asynchronous locks, which we don't support yet */
1614                 return -ENOLCK;
1615         }
1616
1617         /* Unlock on close is handled by the flush method */
1618         if (fl->fl_flags & FL_CLOSE)
1619                 return 0;
1620
1621         req = fuse_get_req_nopages(fc);
1622         if (IS_ERR(req))
1623                 return PTR_ERR(req);
1624
1625         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1626         fuse_request_send(fc, req);
1627         err = req->out.h.error;
1628         /* locking is restartable */
1629         if (err == -EINTR)
1630                 err = -ERESTARTSYS;
1631         fuse_put_request(fc, req);
1632         return err;
1633 }
1634
1635 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1636 {
1637         struct inode *inode = file_inode(file);
1638         struct fuse_conn *fc = get_fuse_conn(inode);
1639         int err;
1640
1641         if (cmd == F_CANCELLK) {
1642                 err = 0;
1643         } else if (cmd == F_GETLK) {
1644                 if (fc->no_lock) {
1645                         posix_test_lock(file, fl);
1646                         err = 0;
1647                 } else
1648                         err = fuse_getlk(file, fl);
1649         } else {
1650                 if (fc->no_lock)
1651                         err = posix_lock_file(file, fl, NULL);
1652                 else
1653                         err = fuse_setlk(file, fl, 0);
1654         }
1655         return err;
1656 }
1657
1658 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1659 {
1660         struct inode *inode = file_inode(file);
1661         struct fuse_conn *fc = get_fuse_conn(inode);
1662         int err;
1663
1664         if (fc->no_flock) {
1665                 err = flock_lock_file_wait(file, fl);
1666         } else {
1667                 struct fuse_file *ff = file->private_data;
1668
1669                 /* emulate flock with POSIX locks */
1670                 fl->fl_owner = (fl_owner_t) file;
1671                 ff->flock = true;
1672                 err = fuse_setlk(file, fl, 1);
1673         }
1674
1675         return err;
1676 }
1677
1678 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1679 {
1680         struct inode *inode = mapping->host;
1681         struct fuse_conn *fc = get_fuse_conn(inode);
1682         struct fuse_req *req;
1683         struct fuse_bmap_in inarg;
1684         struct fuse_bmap_out outarg;
1685         int err;
1686
1687         if (!inode->i_sb->s_bdev || fc->no_bmap)
1688                 return 0;
1689
1690         req = fuse_get_req_nopages(fc);
1691         if (IS_ERR(req))
1692                 return 0;
1693
1694         memset(&inarg, 0, sizeof(inarg));
1695         inarg.block = block;
1696         inarg.blocksize = inode->i_sb->s_blocksize;
1697         req->in.h.opcode = FUSE_BMAP;
1698         req->in.h.nodeid = get_node_id(inode);
1699         req->in.numargs = 1;
1700         req->in.args[0].size = sizeof(inarg);
1701         req->in.args[0].value = &inarg;
1702         req->out.numargs = 1;
1703         req->out.args[0].size = sizeof(outarg);
1704         req->out.args[0].value = &outarg;
1705         fuse_request_send(fc, req);
1706         err = req->out.h.error;
1707         fuse_put_request(fc, req);
1708         if (err == -ENOSYS)
1709                 fc->no_bmap = 1;
1710
1711         return err ? 0 : outarg.block;
1712 }
1713
1714 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1715 {
1716         loff_t retval;
1717         struct inode *inode = file_inode(file);
1718
1719         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1720         if (whence == SEEK_CUR || whence == SEEK_SET)
1721                 return generic_file_llseek(file, offset, whence);
1722
1723         mutex_lock(&inode->i_mutex);
1724         retval = fuse_update_attributes(inode, NULL, file, NULL);
1725         if (!retval)
1726                 retval = generic_file_llseek(file, offset, whence);
1727         mutex_unlock(&inode->i_mutex);
1728
1729         return retval;
1730 }
1731
1732 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1733                         unsigned int nr_segs, size_t bytes, bool to_user)
1734 {
1735         struct iov_iter ii;
1736         int page_idx = 0;
1737
1738         if (!bytes)
1739                 return 0;
1740
1741         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1742
1743         while (iov_iter_count(&ii)) {
1744                 struct page *page = pages[page_idx++];
1745                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1746                 void *kaddr;
1747
1748                 kaddr = kmap(page);
1749
1750                 while (todo) {
1751                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1752                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1753                         size_t copy = min(todo, iov_len);
1754                         size_t left;
1755
1756                         if (!to_user)
1757                                 left = copy_from_user(kaddr, uaddr, copy);
1758                         else
1759                                 left = copy_to_user(uaddr, kaddr, copy);
1760
1761                         if (unlikely(left))
1762                                 return -EFAULT;
1763
1764                         iov_iter_advance(&ii, copy);
1765                         todo -= copy;
1766                         kaddr += copy;
1767                 }
1768
1769                 kunmap(page);
1770         }
1771
1772         return 0;
1773 }
1774
1775 /*
1776  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1777  * ABI was defined to be 'struct iovec' which is different on 32bit
1778  * and 64bit.  Fortunately we can determine which structure the server
1779  * used from the size of the reply.
1780  */
1781 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1782                                      size_t transferred, unsigned count,
1783                                      bool is_compat)
1784 {
1785 #ifdef CONFIG_COMPAT
1786         if (count * sizeof(struct compat_iovec) == transferred) {
1787                 struct compat_iovec *ciov = src;
1788                 unsigned i;
1789
1790                 /*
1791                  * With this interface a 32bit server cannot support
1792                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1793                  * requests
1794                  */
1795                 if (!is_compat)
1796                         return -EINVAL;
1797
1798                 for (i = 0; i < count; i++) {
1799                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1800                         dst[i].iov_len = ciov[i].iov_len;
1801                 }
1802                 return 0;
1803         }
1804 #endif
1805
1806         if (count * sizeof(struct iovec) != transferred)
1807                 return -EIO;
1808
1809         memcpy(dst, src, transferred);
1810         return 0;
1811 }
1812
1813 /* Make sure iov_length() won't overflow */
1814 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1815 {
1816         size_t n;
1817         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1818
1819         for (n = 0; n < count; n++, iov++) {
1820                 if (iov->iov_len > (size_t) max)
1821                         return -ENOMEM;
1822                 max -= iov->iov_len;
1823         }
1824         return 0;
1825 }
1826
1827 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1828                                  void *src, size_t transferred, unsigned count,
1829                                  bool is_compat)
1830 {
1831         unsigned i;
1832         struct fuse_ioctl_iovec *fiov = src;
1833
1834         if (fc->minor < 16) {
1835                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1836                                                  count, is_compat);
1837         }
1838
1839         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1840                 return -EIO;
1841
1842         for (i = 0; i < count; i++) {
1843                 /* Did the server supply an inappropriate value? */
1844                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1845                     fiov[i].len != (unsigned long) fiov[i].len)
1846                         return -EIO;
1847
1848                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1849                 dst[i].iov_len = (size_t) fiov[i].len;
1850
1851 #ifdef CONFIG_COMPAT
1852                 if (is_compat &&
1853                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1854                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1855                         return -EIO;
1856 #endif
1857         }
1858
1859         return 0;
1860 }
1861
1862
1863 /*
1864  * For ioctls, there is no generic way to determine how much memory
1865  * needs to be read and/or written.  Furthermore, ioctls are allowed
1866  * to dereference the passed pointer, so the parameter requires deep
1867  * copying but FUSE has no idea whatsoever about what to copy in or
1868  * out.
1869  *
1870  * This is solved by allowing FUSE server to retry ioctl with
1871  * necessary in/out iovecs.  Let's assume the ioctl implementation
1872  * needs to read in the following structure.
1873  *
1874  * struct a {
1875  *      char    *buf;
1876  *      size_t  buflen;
1877  * }
1878  *
1879  * On the first callout to FUSE server, inarg->in_size and
1880  * inarg->out_size will be NULL; then, the server completes the ioctl
1881  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1882  * the actual iov array to
1883  *
1884  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1885  *
1886  * which tells FUSE to copy in the requested area and retry the ioctl.
1887  * On the second round, the server has access to the structure and
1888  * from that it can tell what to look for next, so on the invocation,
1889  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1890  *
1891  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1892  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1893  *
1894  * FUSE will copy both struct a and the pointed buffer from the
1895  * process doing the ioctl and retry ioctl with both struct a and the
1896  * buffer.
1897  *
1898  * This time, FUSE server has everything it needs and completes ioctl
1899  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1900  *
1901  * Copying data out works the same way.
1902  *
1903  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1904  * automatically initializes in and out iovs by decoding @cmd with
1905  * _IOC_* macros and the server is not allowed to request RETRY.  This
1906  * limits ioctl data transfers to well-formed ioctls and is the forced
1907  * behavior for all FUSE servers.
1908  */
1909 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1910                    unsigned int flags)
1911 {
1912         struct fuse_file *ff = file->private_data;
1913         struct fuse_conn *fc = ff->fc;
1914         struct fuse_ioctl_in inarg = {
1915                 .fh = ff->fh,
1916                 .cmd = cmd,
1917                 .arg = arg,
1918                 .flags = flags
1919         };
1920         struct fuse_ioctl_out outarg;
1921         struct fuse_req *req = NULL;
1922         struct page **pages = NULL;
1923         struct iovec *iov_page = NULL;
1924         struct iovec *in_iov = NULL, *out_iov = NULL;
1925         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1926         size_t in_size, out_size, transferred;
1927         int err;
1928
1929 #if BITS_PER_LONG == 32
1930         inarg.flags |= FUSE_IOCTL_32BIT;
1931 #else
1932         if (flags & FUSE_IOCTL_COMPAT)
1933                 inarg.flags |= FUSE_IOCTL_32BIT;
1934 #endif
1935
1936         /* assume all the iovs returned by client always fits in a page */
1937         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1938
1939         err = -ENOMEM;
1940         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1941         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1942         if (!pages || !iov_page)
1943                 goto out;
1944
1945         /*
1946          * If restricted, initialize IO parameters as encoded in @cmd.
1947          * RETRY from server is not allowed.
1948          */
1949         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1950                 struct iovec *iov = iov_page;
1951
1952                 iov->iov_base = (void __user *)arg;
1953                 iov->iov_len = _IOC_SIZE(cmd);
1954
1955                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1956                         in_iov = iov;
1957                         in_iovs = 1;
1958                 }
1959
1960                 if (_IOC_DIR(cmd) & _IOC_READ) {
1961                         out_iov = iov;
1962                         out_iovs = 1;
1963                 }
1964         }
1965
1966  retry:
1967         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1968         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1969
1970         /*
1971          * Out data can be used either for actual out data or iovs,
1972          * make sure there always is at least one page.
1973          */
1974         out_size = max_t(size_t, out_size, PAGE_SIZE);
1975         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1976
1977         /* make sure there are enough buffer pages and init request with them */
1978         err = -ENOMEM;
1979         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1980                 goto out;
1981         while (num_pages < max_pages) {
1982                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1983                 if (!pages[num_pages])
1984                         goto out;
1985                 num_pages++;
1986         }
1987
1988         req = fuse_get_req(fc, num_pages);
1989         if (IS_ERR(req)) {
1990                 err = PTR_ERR(req);
1991                 req = NULL;
1992                 goto out;
1993         }
1994         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1995         req->num_pages = num_pages;
1996         fuse_page_descs_length_init(req, 0, req->num_pages);
1997
1998         /* okay, let's send it to the client */
1999         req->in.h.opcode = FUSE_IOCTL;
2000         req->in.h.nodeid = ff->nodeid;
2001         req->in.numargs = 1;
2002         req->in.args[0].size = sizeof(inarg);
2003         req->in.args[0].value = &inarg;
2004         if (in_size) {
2005                 req->in.numargs++;
2006                 req->in.args[1].size = in_size;
2007                 req->in.argpages = 1;
2008
2009                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2010                                            false);
2011                 if (err)
2012                         goto out;
2013         }
2014
2015         req->out.numargs = 2;
2016         req->out.args[0].size = sizeof(outarg);
2017         req->out.args[0].value = &outarg;
2018         req->out.args[1].size = out_size;
2019         req->out.argpages = 1;
2020         req->out.argvar = 1;
2021
2022         fuse_request_send(fc, req);
2023         err = req->out.h.error;
2024         transferred = req->out.args[1].size;
2025         fuse_put_request(fc, req);
2026         req = NULL;
2027         if (err)
2028                 goto out;
2029
2030         /* did it ask for retry? */
2031         if (outarg.flags & FUSE_IOCTL_RETRY) {
2032                 void *vaddr;
2033
2034                 /* no retry if in restricted mode */
2035                 err = -EIO;
2036                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2037                         goto out;
2038
2039                 in_iovs = outarg.in_iovs;
2040                 out_iovs = outarg.out_iovs;
2041
2042                 /*
2043                  * Make sure things are in boundary, separate checks
2044                  * are to protect against overflow.
2045                  */
2046                 err = -ENOMEM;
2047                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2048                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2049                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2050                         goto out;
2051
2052                 vaddr = kmap_atomic(pages[0]);
2053                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2054                                             transferred, in_iovs + out_iovs,
2055                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2056                 kunmap_atomic(vaddr);
2057                 if (err)
2058                         goto out;
2059
2060                 in_iov = iov_page;
2061                 out_iov = in_iov + in_iovs;
2062
2063                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2064                 if (err)
2065                         goto out;
2066
2067                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2068                 if (err)
2069                         goto out;
2070
2071                 goto retry;
2072         }
2073
2074         err = -EIO;
2075         if (transferred > inarg.out_size)
2076                 goto out;
2077
2078         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2079  out:
2080         if (req)
2081                 fuse_put_request(fc, req);
2082         free_page((unsigned long) iov_page);
2083         while (num_pages)
2084                 __free_page(pages[--num_pages]);
2085         kfree(pages);
2086
2087         return err ? err : outarg.result;
2088 }
2089 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2090
2091 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2092                        unsigned long arg, unsigned int flags)
2093 {
2094         struct inode *inode = file_inode(file);
2095         struct fuse_conn *fc = get_fuse_conn(inode);
2096
2097         if (!fuse_allow_current_process(fc))
2098                 return -EACCES;
2099
2100         if (is_bad_inode(inode))
2101                 return -EIO;
2102
2103         return fuse_do_ioctl(file, cmd, arg, flags);
2104 }
2105
2106 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2107                             unsigned long arg)
2108 {
2109         return fuse_ioctl_common(file, cmd, arg, 0);
2110 }
2111
2112 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2113                                    unsigned long arg)
2114 {
2115         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2116 }
2117
2118 /*
2119  * All files which have been polled are linked to RB tree
2120  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2121  * find the matching one.
2122  */
2123 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2124                                               struct rb_node **parent_out)
2125 {
2126         struct rb_node **link = &fc->polled_files.rb_node;
2127         struct rb_node *last = NULL;
2128
2129         while (*link) {
2130                 struct fuse_file *ff;
2131
2132                 last = *link;
2133                 ff = rb_entry(last, struct fuse_file, polled_node);
2134
2135                 if (kh < ff->kh)
2136                         link = &last->rb_left;
2137                 else if (kh > ff->kh)
2138                         link = &last->rb_right;
2139                 else
2140                         return link;
2141         }
2142
2143         if (parent_out)
2144                 *parent_out = last;
2145         return link;
2146 }
2147
2148 /*
2149  * The file is about to be polled.  Make sure it's on the polled_files
2150  * RB tree.  Note that files once added to the polled_files tree are
2151  * not removed before the file is released.  This is because a file
2152  * polled once is likely to be polled again.
2153  */
2154 static void fuse_register_polled_file(struct fuse_conn *fc,
2155                                       struct fuse_file *ff)
2156 {
2157         spin_lock(&fc->lock);
2158         if (RB_EMPTY_NODE(&ff->polled_node)) {
2159                 struct rb_node **link, *parent;
2160
2161                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2162                 BUG_ON(*link);
2163                 rb_link_node(&ff->polled_node, parent, link);
2164                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2165         }
2166         spin_unlock(&fc->lock);
2167 }
2168
2169 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2170 {
2171         struct fuse_file *ff = file->private_data;
2172         struct fuse_conn *fc = ff->fc;
2173         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2174         struct fuse_poll_out outarg;
2175         struct fuse_req *req;
2176         int err;
2177
2178         if (fc->no_poll)
2179                 return DEFAULT_POLLMASK;
2180
2181         poll_wait(file, &ff->poll_wait, wait);
2182         inarg.events = (__u32)poll_requested_events(wait);
2183
2184         /*
2185          * Ask for notification iff there's someone waiting for it.
2186          * The client may ignore the flag and always notify.
2187          */
2188         if (waitqueue_active(&ff->poll_wait)) {
2189                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2190                 fuse_register_polled_file(fc, ff);
2191         }
2192
2193         req = fuse_get_req_nopages(fc);
2194         if (IS_ERR(req))
2195                 return POLLERR;
2196
2197         req->in.h.opcode = FUSE_POLL;
2198         req->in.h.nodeid = ff->nodeid;
2199         req->in.numargs = 1;
2200         req->in.args[0].size = sizeof(inarg);
2201         req->in.args[0].value = &inarg;
2202         req->out.numargs = 1;
2203         req->out.args[0].size = sizeof(outarg);
2204         req->out.args[0].value = &outarg;
2205         fuse_request_send(fc, req);
2206         err = req->out.h.error;
2207         fuse_put_request(fc, req);
2208
2209         if (!err)
2210                 return outarg.revents;
2211         if (err == -ENOSYS) {
2212                 fc->no_poll = 1;
2213                 return DEFAULT_POLLMASK;
2214         }
2215         return POLLERR;
2216 }
2217 EXPORT_SYMBOL_GPL(fuse_file_poll);
2218
2219 /*
2220  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2221  * wakes up the poll waiters.
2222  */
2223 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2224                             struct fuse_notify_poll_wakeup_out *outarg)
2225 {
2226         u64 kh = outarg->kh;
2227         struct rb_node **link;
2228
2229         spin_lock(&fc->lock);
2230
2231         link = fuse_find_polled_node(fc, kh, NULL);
2232         if (*link) {
2233                 struct fuse_file *ff;
2234
2235                 ff = rb_entry(*link, struct fuse_file, polled_node);
2236                 wake_up_interruptible_sync(&ff->poll_wait);
2237         }
2238
2239         spin_unlock(&fc->lock);
2240         return 0;
2241 }
2242
2243 static ssize_t
2244 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2245                         loff_t offset, unsigned long nr_segs)
2246 {
2247         ssize_t ret = 0;
2248         struct file *file = NULL;
2249         loff_t pos = 0;
2250
2251         file = iocb->ki_filp;
2252         pos = offset;
2253
2254         if (rw == WRITE)
2255                 ret = __fuse_direct_write(file, iov, nr_segs, &pos);
2256         else
2257                 ret = __fuse_direct_read(file, iov, nr_segs, &pos);
2258
2259         return ret;
2260 }
2261
2262 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2263                                 loff_t length)
2264 {
2265         struct fuse_file *ff = file->private_data;
2266         struct fuse_conn *fc = ff->fc;
2267         struct fuse_req *req;
2268         struct fuse_fallocate_in inarg = {
2269                 .fh = ff->fh,
2270                 .offset = offset,
2271                 .length = length,
2272                 .mode = mode
2273         };
2274         int err;
2275
2276         if (fc->no_fallocate)
2277                 return -EOPNOTSUPP;
2278
2279         req = fuse_get_req_nopages(fc);
2280         if (IS_ERR(req))
2281                 return PTR_ERR(req);
2282
2283         req->in.h.opcode = FUSE_FALLOCATE;
2284         req->in.h.nodeid = ff->nodeid;
2285         req->in.numargs = 1;
2286         req->in.args[0].size = sizeof(inarg);
2287         req->in.args[0].value = &inarg;
2288         fuse_request_send(fc, req);
2289         err = req->out.h.error;
2290         if (err == -ENOSYS) {
2291                 fc->no_fallocate = 1;
2292                 err = -EOPNOTSUPP;
2293         }
2294         fuse_put_request(fc, req);
2295
2296         return err;
2297 }
2298
2299 static const struct file_operations fuse_file_operations = {
2300         .llseek         = fuse_file_llseek,
2301         .read           = do_sync_read,
2302         .aio_read       = fuse_file_aio_read,
2303         .write          = do_sync_write,
2304         .aio_write      = fuse_file_aio_write,
2305         .mmap           = fuse_file_mmap,
2306         .open           = fuse_open,
2307         .flush          = fuse_flush,
2308         .release        = fuse_release,
2309         .fsync          = fuse_fsync,
2310         .lock           = fuse_file_lock,
2311         .flock          = fuse_file_flock,
2312         .splice_read    = generic_file_splice_read,
2313         .unlocked_ioctl = fuse_file_ioctl,
2314         .compat_ioctl   = fuse_file_compat_ioctl,
2315         .poll           = fuse_file_poll,
2316         .fallocate      = fuse_file_fallocate,
2317 };
2318
2319 static const struct file_operations fuse_direct_io_file_operations = {
2320         .llseek         = fuse_file_llseek,
2321         .read           = fuse_direct_read,
2322         .write          = fuse_direct_write,
2323         .mmap           = fuse_direct_mmap,
2324         .open           = fuse_open,
2325         .flush          = fuse_flush,
2326         .release        = fuse_release,
2327         .fsync          = fuse_fsync,
2328         .lock           = fuse_file_lock,
2329         .flock          = fuse_file_flock,
2330         .unlocked_ioctl = fuse_file_ioctl,
2331         .compat_ioctl   = fuse_file_compat_ioctl,
2332         .poll           = fuse_file_poll,
2333         .fallocate      = fuse_file_fallocate,
2334         /* no splice_read */
2335 };
2336
2337 static const struct address_space_operations fuse_file_aops  = {
2338         .readpage       = fuse_readpage,
2339         .writepage      = fuse_writepage,
2340         .launder_page   = fuse_launder_page,
2341         .readpages      = fuse_readpages,
2342         .set_page_dirty = __set_page_dirty_nobuffers,
2343         .bmap           = fuse_bmap,
2344         .direct_IO      = fuse_direct_IO,
2345 };
2346
2347 void fuse_init_file_inode(struct inode *inode)
2348 {
2349         inode->i_fop = &fuse_file_operations;
2350         inode->i_data.a_ops = &fuse_file_aops;
2351 }