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
 707         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 708                 int err;
 709                 /*
 710                  * If trying to read past EOF, make sure the i_size
 711                  * attribute is up-to-date.
 712                  */
 713                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 714                 if (err)
 715                         return err;
 716         }
 717
 718         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 719 }
 720
 721 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 722                             loff_t pos, size_t count)
 723 {
 724         struct fuse_write_in *inarg = &req->misc.write.in;
 725         struct fuse_write_out *outarg = &req->misc.write.out;
 726
 727         inarg->fh = ff->fh;
 728         inarg->offset = pos;
 729         inarg->size = count;
 730         req->in.h.opcode = FUSE_WRITE;
 731         req->in.h.nodeid = ff->nodeid;
 732         req->in.numargs = 2;
 733         if (ff->fc->minor < 9)
 734                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 735         else
 736                 req->in.args[0].size = sizeof(struct fuse_write_in);
 737         req->in.args[0].value = inarg;
 738         req->in.args[1].size = count;
 739         req->out.numargs = 1;
 740         req->out.args[0].size = sizeof(struct fuse_write_out);
 741         req->out.args[0].value = outarg;
 742 }
 743
 744 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 745                               loff_t pos, size_t count, fl_owner_t owner)
 746 {
 747         struct fuse_file *ff = file->private_data;
 748         struct fuse_conn *fc = ff->fc;
 749         struct fuse_write_in *inarg = &req->misc.write.in;
 750
 751         fuse_write_fill(req, ff, pos, count);
 752         inarg->flags = file->f_flags;
 753         if (owner != NULL) {
 754                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 755                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 756         }
 757         fuse_request_send(fc, req);
 758         return req->misc.write.out.size;
 759 }
 760
 761 void fuse_write_update_size(struct inode *inode, loff_t pos)
 762 {
 763         struct fuse_conn *fc = get_fuse_conn(inode);
 764         struct fuse_inode *fi = get_fuse_inode(inode);
 765
 766         spin_lock(&fc->lock);
 767         fi->attr_version = ++fc->attr_version;
 768         if (pos > inode->i_size)
 769                 i_size_write(inode, pos);
 770         spin_unlock(&fc->lock);
 771 }
 772
 773 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 774                                     struct inode *inode, loff_t pos,
 775                                     size_t count)
 776 {
 777         size_t res;
 778         unsigned offset;
 779         unsigned i;
 780
 781         for (i = 0; i < req->num_pages; i++)
 782                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 783
 784         res = fuse_send_write(req, file, pos, count, NULL);
 785
 786         offset = req->page_offset;
 787         count = res;
 788         for (i = 0; i < req->num_pages; i++) {
 789                 struct page *page = req->pages[i];
 790
 791                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 792                         SetPageUptodate(page);
 793
 794                 if (count > PAGE_CACHE_SIZE - offset)
 795                         count -= PAGE_CACHE_SIZE - offset;
 796                 else
 797                         count = 0;
 798                 offset = 0;
 799
 800                 unlock_page(page);
 801                 page_cache_release(page);
 802         }
 803
 804         return res;
 805 }
 806
 807 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 808                                struct address_space *mapping,
 809                                struct iov_iter *ii, loff_t pos)
 810 {
 811         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 812         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 813         size_t count = 0;
 814         int err;
 815
 816         req->in.argpages = 1;
 817         req->page_offset = offset;
 818
 819         do {
 820                 size_t tmp;
 821                 struct page *page;
 822                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 823                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 824                                      iov_iter_count(ii));
 825
 826                 bytes = min_t(size_t, bytes, fc->max_write - count);
 827
 828  again:
 829                 err = -EFAULT;
 830                 if (iov_iter_fault_in_readable(ii, bytes))
 831                         break;
 832
 833                 err = -ENOMEM;
 834                 page = grab_cache_page_write_begin(mapping, index, 0);
 835                 if (!page)
 836                         break;
 837
 838                 if (mapping_writably_mapped(mapping))
 839                         flush_dcache_page(page);
 840
 841                 pagefault_disable();
 842                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 843                 pagefault_enable();
 844                 flush_dcache_page(page);
 845
 846                 mark_page_accessed(page);
 847
 848                 if (!tmp) {
 849                         unlock_page(page);
 850                         page_cache_release(page);
 851                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 852                         goto again;
 853                 }
 854
 855                 err = 0;
 856                 req->pages[req->num_pages] = page;
 857                 req->num_pages++;
 858
 859                 iov_iter_advance(ii, tmp);
 860                 count += tmp;
 861                 pos += tmp;
 862                 offset += tmp;
 863                 if (offset == PAGE_CACHE_SIZE)
 864                         offset = 0;
 865
 866                 if (!fc->big_writes)
 867                         break;
 868         } while (iov_iter_count(ii) && count < fc->max_write &&
 869                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 870
 871         return count > 0 ? count : err;
 872 }
 873
 874 static ssize_t fuse_perform_write(struct file *file,
 875                                   struct address_space *mapping,
 876                                   struct iov_iter *ii, loff_t pos)
 877 {
 878         struct inode *inode = mapping->host;
 879         struct fuse_conn *fc = get_fuse_conn(inode);
 880         int err = 0;
 881         ssize_t res = 0;
 882
 883         if (is_bad_inode(inode))
 884                 return -EIO;
 885
 886         do {
 887                 struct fuse_req *req;
 888                 ssize_t count;
 889
 890                 req = fuse_get_req(fc);
 891                 if (IS_ERR(req)) {
 892                         err = PTR_ERR(req);
 893                         break;
 894                 }
 895
 896                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 897                 if (count <= 0) {
 898                         err = count;
 899                 } else {
 900                         size_t num_written;
 901
 902                         num_written = fuse_send_write_pages(req, file, inode,
 903                                                             pos, count);
 904                         err = req->out.h.error;
 905                         if (!err) {
 906                                 res += num_written;
 907                                 pos += num_written;
 908
 909                                 /* break out of the loop on short write */
 910                                 if (num_written != count)
 911                                         err = -EIO;
 912                         }
 913                 }
 914                 fuse_put_request(fc, req);
 915         } while (!err && iov_iter_count(ii));
 916
 917         if (res > 0)
 918                 fuse_write_update_size(inode, pos);
 919
 920         fuse_invalidate_attr(inode);
 921
 922         return res > 0 ? res : err;
 923 }
 924
 925 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 926                                    unsigned long nr_segs, loff_t pos)
 927 {
 928         struct file *file = iocb->ki_filp;
 929         struct address_space *mapping = file->f_mapping;
 930         size_t count = 0;
 931         size_t ocount = 0;
 932         ssize_t written = 0;
 933         ssize_t written_buffered = 0;
 934         struct inode *inode = mapping->host;
 935         ssize_t err;
 936         struct iov_iter i;
 937         loff_t endbyte = 0;
 938
 939         WARN_ON(iocb->ki_pos != pos);
 940
 941         ocount = 0;
 942         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 943         if (err)
 944                 return err;
 945
 946         count = ocount;
 947         sb_start_write(inode->i_sb);
 948         mutex_lock(&inode->i_mutex);
 949
 950         /* We can write back this queue in page reclaim */
 951         current->backing_dev_info = mapping->backing_dev_info;
 952
 953         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 954         if (err)
 955                 goto out;
 956
 957         if (count == 0)
 958                 goto out;
 959
 960         err = file_remove_suid(file);
 961         if (err)
 962                 goto out;
 963
 964         err = file_update_time(file);
 965         if (err)
 966                 goto out;
 967
 968         if (file->f_flags & O_DIRECT) {
 969                 written = generic_file_direct_write(iocb, iov, &nr_segs,
 970                                                     pos, &iocb->ki_pos,
 971                                                     count, ocount);
 972                 if (written < 0 || written == count)
 973                         goto out;
 974
 975                 pos += written;
 976                 count -= written;
 977
 978                 iov_iter_init(&i, iov, nr_segs, count, written);
 979                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
 980                 if (written_buffered < 0) {
 981                         err = written_buffered;
 982                         goto out;
 983                 }
 984                 endbyte = pos + written_buffered - 1;
 985
 986                 err = filemap_write_and_wait_range(file->f_mapping, pos,
 987                                                    endbyte);
 988                 if (err)
 989                         goto out;
 990
 991                 invalidate_mapping_pages(file->f_mapping,
 992                                          pos >> PAGE_CACHE_SHIFT,
 993                                          endbyte >> PAGE_CACHE_SHIFT);
 994
 995                 written += written_buffered;
 996                 iocb->ki_pos = pos + written_buffered;
 997         } else {
 998                 iov_iter_init(&i, iov, nr_segs, count, 0);
 999                 written = fuse_perform_write(file, mapping, &i, pos);
1000                 if (written >= 0)
1001                         iocb->ki_pos = pos + written;
1002         }
1003 out:
1004         current->backing_dev_info = NULL;
1005         mutex_unlock(&inode->i_mutex);
1006         sb_end_write(inode->i_sb);
1007
1008         return written ? written : err;
1009 }
1010
1011 static void fuse_release_user_pages(struct fuse_req *req, int write)
1012 {
1013         unsigned i;
1014
1015         for (i = 0; i < req->num_pages; i++) {
1016                 struct page *page = req->pages[i];
1017                 if (write)
1018                         set_page_dirty_lock(page);
1019                 put_page(page);
1020         }
1021 }
1022
1023 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1024                                size_t *nbytesp, int write)
1025 {
1026         size_t nbytes = *nbytesp;
1027         unsigned long user_addr = (unsigned long) buf;
1028         unsigned offset = user_addr & ~PAGE_MASK;
1029         int npages;
1030
1031         /* Special case for kernel I/O: can copy directly into the buffer */
1032         if (segment_eq(get_fs(), KERNEL_DS)) {
1033                 if (write)
1034                         req->in.args[1].value = (void *) user_addr;
1035                 else
1036                         req->out.args[0].value = (void *) user_addr;
1037
1038                 return 0;
1039         }
1040
1041         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1042         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1043         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1044         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1045         if (npages < 0)
1046                 return npages;
1047
1048         req->num_pages = npages;
1049         req->page_offset = offset;
1050
1051         if (write)
1052                 req->in.argpages = 1;
1053         else
1054                 req->out.argpages = 1;
1055
1056         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1057         *nbytesp = min(*nbytesp, nbytes);
1058
1059         return 0;
1060 }
1061
1062 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1063                        size_t count, loff_t *ppos, int write)
1064 {
1065         struct fuse_file *ff = file->private_data;
1066         struct fuse_conn *fc = ff->fc;
1067         size_t nmax = write ? fc->max_write : fc->max_read;
1068         loff_t pos = *ppos;
1069         ssize_t res = 0;
1070         struct fuse_req *req;
1071
1072         req = fuse_get_req(fc);
1073         if (IS_ERR(req))
1074                 return PTR_ERR(req);
1075
1076         while (count) {
1077                 size_t nres;
1078                 fl_owner_t owner = current->files;
1079                 size_t nbytes = min(count, nmax);
1080                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1081                 if (err) {
1082                         res = err;
1083                         break;
1084                 }
1085
1086                 if (write)
1087                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1088                 else
1089                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1090
1091                 fuse_release_user_pages(req, !write);
1092                 if (req->out.h.error) {
1093                         if (!res)
1094                                 res = req->out.h.error;
1095                         break;
1096                 } else if (nres > nbytes) {
1097                         res = -EIO;
1098                         break;
1099                 }
1100                 count -= nres;
1101                 res += nres;
1102                 pos += nres;
1103                 buf += nres;
1104                 if (nres != nbytes)
1105                         break;
1106                 if (count) {
1107                         fuse_put_request(fc, req);
1108                         req = fuse_get_req(fc);
1109                         if (IS_ERR(req))
1110                                 break;
1111                 }
1112         }
1113         if (!IS_ERR(req))
1114                 fuse_put_request(fc, req);
1115         if (res > 0)
1116                 *ppos = pos;
1117
1118         return res;
1119 }
1120 EXPORT_SYMBOL_GPL(fuse_direct_io);
1121
1122 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1123                                      size_t count, loff_t *ppos)
1124 {
1125         ssize_t res;
1126         struct inode *inode = file->f_path.dentry->d_inode;
1127
1128         if (is_bad_inode(inode))
1129                 return -EIO;
1130
1131         res = fuse_direct_io(file, buf, count, ppos, 0);
1132
1133         fuse_invalidate_attr(inode);
1134
1135         return res;
1136 }
1137
1138 static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
1139                                    size_t count, loff_t *ppos)
1140 {
1141         struct inode *inode = file->f_path.dentry->d_inode;
1142         ssize_t res;
1143
1144         res = generic_write_checks(file, ppos, &count, 0);
1145         if (!res) {
1146                 res = fuse_direct_io(file, buf, count, ppos, 1);
1147                 if (res > 0)
1148                         fuse_write_update_size(inode, *ppos);
1149         }
1150
1151         fuse_invalidate_attr(inode);
1152
1153         return res;
1154 }
1155
1156 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1157                                  size_t count, loff_t *ppos)
1158 {
1159         struct inode *inode = file->f_path.dentry->d_inode;
1160         ssize_t res;
1161
1162         if (is_bad_inode(inode))
1163                 return -EIO;
1164
1165         /* Don't allow parallel writes to the same file */
1166         mutex_lock(&inode->i_mutex);
1167         res = __fuse_direct_write(file, buf, count, ppos);
1168         mutex_unlock(&inode->i_mutex);
1169
1170         return res;
1171 }
1172
1173 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1174 {
1175         __free_page(req->pages[0]);
1176         fuse_file_put(req->ff, false);
1177 }
1178
1179 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1180 {
1181         struct inode *inode = req->inode;
1182         struct fuse_inode *fi = get_fuse_inode(inode);
1183         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1184
1185         list_del(&req->writepages_entry);
1186         dec_bdi_stat(bdi, BDI_WRITEBACK);
1187         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1188         bdi_writeout_inc(bdi);
1189         wake_up(&fi->page_waitq);
1190 }
1191
1192 /* Called under fc->lock, may release and reacquire it */
1193 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1194 __releases(fc->lock)
1195 __acquires(fc->lock)
1196 {
1197         struct fuse_inode *fi = get_fuse_inode(req->inode);
1198         loff_t size = i_size_read(req->inode);
1199         struct fuse_write_in *inarg = &req->misc.write.in;
1200
1201         if (!fc->connected)
1202                 goto out_free;
1203
1204         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1205                 inarg->size = PAGE_CACHE_SIZE;
1206         } else if (inarg->offset < size) {
1207                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1208         } else {
1209                 /* Got truncated off completely */
1210                 goto out_free;
1211         }
1212
1213         req->in.args[1].size = inarg->size;
1214         fi->writectr++;
1215         fuse_request_send_background_locked(fc, req);
1216         return;
1217
1218  out_free:
1219         fuse_writepage_finish(fc, req);
1220         spin_unlock(&fc->lock);
1221         fuse_writepage_free(fc, req);
1222         fuse_put_request(fc, req);
1223         spin_lock(&fc->lock);
1224 }
1225
1226 /*
1227  * If fi->writectr is positive (no truncate or fsync going on) send
1228  * all queued writepage requests.
1229  *
1230  * Called with fc->lock
1231  */
1232 void fuse_flush_writepages(struct inode *inode)
1233 __releases(fc->lock)
1234 __acquires(fc->lock)
1235 {
1236         struct fuse_conn *fc = get_fuse_conn(inode);
1237         struct fuse_inode *fi = get_fuse_inode(inode);
1238         struct fuse_req *req;
1239
1240         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1241                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1242                 list_del_init(&req->list);
1243                 fuse_send_writepage(fc, req);
1244         }
1245 }
1246
1247 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1248 {
1249         struct inode *inode = req->inode;
1250         struct fuse_inode *fi = get_fuse_inode(inode);
1251
1252         mapping_set_error(inode->i_mapping, req->out.h.error);
1253         spin_lock(&fc->lock);
1254         fi->writectr--;
1255         fuse_writepage_finish(fc, req);
1256         spin_unlock(&fc->lock);
1257         fuse_writepage_free(fc, req);
1258 }
1259
1260 static int fuse_writepage_locked(struct page *page)
1261 {
1262         struct address_space *mapping = page->mapping;
1263         struct inode *inode = mapping->host;
1264         struct fuse_conn *fc = get_fuse_conn(inode);
1265         struct fuse_inode *fi = get_fuse_inode(inode);
1266         struct fuse_req *req;
1267         struct fuse_file *ff;
1268         struct page *tmp_page;
1269
1270         set_page_writeback(page);
1271
1272         req = fuse_request_alloc_nofs();
1273         if (!req)
1274                 goto err;
1275
1276         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1277         if (!tmp_page)
1278                 goto err_free;
1279
1280         spin_lock(&fc->lock);
1281         BUG_ON(list_empty(&fi->write_files));
1282         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1283         req->ff = fuse_file_get(ff);
1284         spin_unlock(&fc->lock);
1285
1286         fuse_write_fill(req, ff, page_offset(page), 0);
1287
1288         copy_highpage(tmp_page, page);
1289         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1290         req->in.argpages = 1;
1291         req->num_pages = 1;
1292         req->pages[0] = tmp_page;
1293         req->page_offset = 0;
1294         req->end = fuse_writepage_end;
1295         req->inode = inode;
1296
1297         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1298         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1299         end_page_writeback(page);
1300
1301         spin_lock(&fc->lock);
1302         list_add(&req->writepages_entry, &fi->writepages);
1303         list_add_tail(&req->list, &fi->queued_writes);
1304         fuse_flush_writepages(inode);
1305         spin_unlock(&fc->lock);
1306
1307         return 0;
1308
1309 err_free:
1310         fuse_request_free(req);
1311 err:
1312         end_page_writeback(page);
1313         return -ENOMEM;
1314 }
1315
1316 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1317 {
1318         int err;
1319
1320         err = fuse_writepage_locked(page);
1321         unlock_page(page);
1322
1323         return err;
1324 }
1325
1326 static int fuse_launder_page(struct page *page)
1327 {
1328         int err = 0;
1329         if (clear_page_dirty_for_io(page)) {
1330                 struct inode *inode = page->mapping->host;
1331                 err = fuse_writepage_locked(page);
1332                 if (!err)
1333                         fuse_wait_on_page_writeback(inode, page->index);
1334         }
1335         return err;
1336 }
1337
1338 /*
1339  * Write back dirty pages now, because there may not be any suitable
1340  * open files later
1341  */
1342 static void fuse_vma_close(struct vm_area_struct *vma)
1343 {
1344         filemap_write_and_wait(vma->vm_file->f_mapping);
1345 }
1346
1347 /*
1348  * Wait for writeback against this page to complete before allowing it
1349  * to be marked dirty again, and hence written back again, possibly
1350  * before the previous writepage completed.
1351  *
1352  * Block here, instead of in ->writepage(), so that the userspace fs
1353  * can only block processes actually operating on the filesystem.
1354  *
1355  * Otherwise unprivileged userspace fs would be able to block
1356  * unrelated:
1357  *
1358  * - page migration
1359  * - sync(2)
1360  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1361  */
1362 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1363 {
1364         struct page *page = vmf->page;
1365         /*
1366          * Don't use page->mapping as it may become NULL from a
1367          * concurrent truncate.
1368          */
1369         struct inode *inode = vma->vm_file->f_mapping->host;
1370
1371         fuse_wait_on_page_writeback(inode, page->index);
1372         return 0;
1373 }
1374
1375 static const struct vm_operations_struct fuse_file_vm_ops = {
1376         .close          = fuse_vma_close,
1377         .fault          = filemap_fault,
1378         .page_mkwrite   = fuse_page_mkwrite,
1379 };
1380
1381 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1382 {
1383         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1384                 struct inode *inode = file->f_dentry->d_inode;
1385                 struct fuse_conn *fc = get_fuse_conn(inode);
1386                 struct fuse_inode *fi = get_fuse_inode(inode);
1387                 struct fuse_file *ff = file->private_data;
1388                 /*
1389                  * file may be written through mmap, so chain it onto the
1390                  * inodes's write_file list
1391                  */
1392                 spin_lock(&fc->lock);
1393                 if (list_empty(&ff->write_entry))
1394                         list_add(&ff->write_entry, &fi->write_files);
1395                 spin_unlock(&fc->lock);
1396         }
1397         file_accessed(file);
1398         vma->vm_ops = &fuse_file_vm_ops;
1399         return 0;
1400 }
1401
1402 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1403 {
1404         /* Can't provide the coherency needed for MAP_SHARED */
1405         if (vma->vm_flags & VM_MAYSHARE)
1406                 return -ENODEV;
1407
1408         invalidate_inode_pages2(file->f_mapping);
1409
1410         return generic_file_mmap(file, vma);
1411 }
1412
1413 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1414                                   struct file_lock *fl)
1415 {
1416         switch (ffl->type) {
1417         case F_UNLCK:
1418                 break;
1419
1420         case F_RDLCK:
1421         case F_WRLCK:
1422                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1423                     ffl->end < ffl->start)
1424                         return -EIO;
1425
1426                 fl->fl_start = ffl->start;
1427                 fl->fl_end = ffl->end;
1428                 fl->fl_pid = ffl->pid;
1429                 break;
1430
1431         default:
1432                 return -EIO;
1433         }
1434         fl->fl_type = ffl->type;
1435         return 0;
1436 }
1437
1438 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1439                          const struct file_lock *fl, int opcode, pid_t pid,
1440                          int flock)
1441 {
1442         struct inode *inode = file->f_path.dentry->d_inode;
1443         struct fuse_conn *fc = get_fuse_conn(inode);
1444         struct fuse_file *ff = file->private_data;
1445         struct fuse_lk_in *arg = &req->misc.lk_in;
1446
1447         arg->fh = ff->fh;
1448         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1449         arg->lk.start = fl->fl_start;
1450         arg->lk.end = fl->fl_end;
1451         arg->lk.type = fl->fl_type;
1452         arg->lk.pid = pid;
1453         if (flock)
1454                 arg->lk_flags |= FUSE_LK_FLOCK;
1455         req->in.h.opcode = opcode;
1456         req->in.h.nodeid = get_node_id(inode);
1457         req->in.numargs = 1;
1458         req->in.args[0].size = sizeof(*arg);
1459         req->in.args[0].value = arg;
1460 }
1461
1462 static int fuse_getlk(struct file *file, struct file_lock *fl)
1463 {
1464         struct inode *inode = file->f_path.dentry->d_inode;
1465         struct fuse_conn *fc = get_fuse_conn(inode);
1466         struct fuse_req *req;
1467         struct fuse_lk_out outarg;
1468         int err;
1469
1470         req = fuse_get_req(fc);
1471         if (IS_ERR(req))
1472                 return PTR_ERR(req);
1473
1474         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1475         req->out.numargs = 1;
1476         req->out.args[0].size = sizeof(outarg);
1477         req->out.args[0].value = &outarg;
1478         fuse_request_send(fc, req);
1479         err = req->out.h.error;
1480         fuse_put_request(fc, req);
1481         if (!err)
1482                 err = convert_fuse_file_lock(&outarg.lk, fl);
1483
1484         return err;
1485 }
1486
1487 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1488 {
1489         struct inode *inode = file->f_path.dentry->d_inode;
1490         struct fuse_conn *fc = get_fuse_conn(inode);
1491         struct fuse_req *req;
1492         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1493         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1494         int err;
1495
1496         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1497                 /* NLM needs asynchronous locks, which we don't support yet */
1498                 return -ENOLCK;
1499         }
1500
1501         /* Unlock on close is handled by the flush method */
1502         if (fl->fl_flags & FL_CLOSE)
1503                 return 0;
1504
1505         req = fuse_get_req(fc);
1506         if (IS_ERR(req))
1507                 return PTR_ERR(req);
1508
1509         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1510         fuse_request_send(fc, req);
1511         err = req->out.h.error;
1512         /* locking is restartable */
1513         if (err == -EINTR)
1514                 err = -ERESTARTSYS;
1515         fuse_put_request(fc, req);
1516         return err;
1517 }
1518
1519 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1520 {
1521         struct inode *inode = file->f_path.dentry->d_inode;
1522         struct fuse_conn *fc = get_fuse_conn(inode);
1523         int err;
1524
1525         if (cmd == F_CANCELLK) {
1526                 err = 0;
1527         } else if (cmd == F_GETLK) {
1528                 if (fc->no_lock) {
1529                         posix_test_lock(file, fl);
1530                         err = 0;
1531                 } else
1532                         err = fuse_getlk(file, fl);
1533         } else {
1534                 if (fc->no_lock)
1535                         err = posix_lock_file(file, fl, NULL);
1536                 else
1537                         err = fuse_setlk(file, fl, 0);
1538         }
1539         return err;
1540 }
1541
1542 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1543 {
1544         struct inode *inode = file->f_path.dentry->d_inode;
1545         struct fuse_conn *fc = get_fuse_conn(inode);
1546         int err;
1547
1548         if (fc->no_flock) {
1549                 err = flock_lock_file_wait(file, fl);
1550         } else {
1551                 struct fuse_file *ff = file->private_data;
1552
1553                 /* emulate flock with POSIX locks */
1554                 fl->fl_owner = (fl_owner_t) file;
1555                 ff->flock = true;
1556                 err = fuse_setlk(file, fl, 1);
1557         }
1558
1559         return err;
1560 }
1561
1562 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1563 {
1564         struct inode *inode = mapping->host;
1565         struct fuse_conn *fc = get_fuse_conn(inode);
1566         struct fuse_req *req;
1567         struct fuse_bmap_in inarg;
1568         struct fuse_bmap_out outarg;
1569         int err;
1570
1571         if (!inode->i_sb->s_bdev || fc->no_bmap)
1572                 return 0;
1573
1574         req = fuse_get_req(fc);
1575         if (IS_ERR(req))
1576                 return 0;
1577
1578         memset(&inarg, 0, sizeof(inarg));
1579         inarg.block = block;
1580         inarg.blocksize = inode->i_sb->s_blocksize;
1581         req->in.h.opcode = FUSE_BMAP;
1582         req->in.h.nodeid = get_node_id(inode);
1583         req->in.numargs = 1;
1584         req->in.args[0].size = sizeof(inarg);
1585         req->in.args[0].value = &inarg;
1586         req->out.numargs = 1;
1587         req->out.args[0].size = sizeof(outarg);
1588         req->out.args[0].value = &outarg;
1589         fuse_request_send(fc, req);
1590         err = req->out.h.error;
1591         fuse_put_request(fc, req);
1592         if (err == -ENOSYS)
1593                 fc->no_bmap = 1;
1594
1595         return err ? 0 : outarg.block;
1596 }
1597
1598 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1599 {
1600         loff_t retval;
1601         struct inode *inode = file->f_path.dentry->d_inode;
1602
1603         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1604         if (origin == SEEK_CUR || origin == SEEK_SET)
1605                 return generic_file_llseek(file, offset, origin);
1606
1607         mutex_lock(&inode->i_mutex);
1608         retval = fuse_update_attributes(inode, NULL, file, NULL);
1609         if (!retval)
1610                 retval = generic_file_llseek(file, offset, origin);
1611         mutex_unlock(&inode->i_mutex);
1612
1613         return retval;
1614 }
1615
1616 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1617                         unsigned int nr_segs, size_t bytes, bool to_user)
1618 {
1619         struct iov_iter ii;
1620         int page_idx = 0;
1621
1622         if (!bytes)
1623                 return 0;
1624
1625         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1626
1627         while (iov_iter_count(&ii)) {
1628                 struct page *page = pages[page_idx++];
1629                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1630                 void *kaddr;
1631
1632                 kaddr = kmap(page);
1633
1634                 while (todo) {
1635                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1636                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1637                         size_t copy = min(todo, iov_len);
1638                         size_t left;
1639
1640                         if (!to_user)
1641                                 left = copy_from_user(kaddr, uaddr, copy);
1642                         else
1643                                 left = copy_to_user(uaddr, kaddr, copy);
1644
1645                         if (unlikely(left))
1646                                 return -EFAULT;
1647
1648                         iov_iter_advance(&ii, copy);
1649                         todo -= copy;
1650                         kaddr += copy;
1651                 }
1652
1653                 kunmap(page);
1654         }
1655
1656         return 0;
1657 }
1658
1659 /*
1660  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1661  * ABI was defined to be 'struct iovec' which is different on 32bit
1662  * and 64bit.  Fortunately we can determine which structure the server
1663  * used from the size of the reply.
1664  */
1665 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1666                                      size_t transferred, unsigned count,
1667                                      bool is_compat)
1668 {
1669 #ifdef CONFIG_COMPAT
1670         if (count * sizeof(struct compat_iovec) == transferred) {
1671                 struct compat_iovec *ciov = src;
1672                 unsigned i;
1673
1674                 /*
1675                  * With this interface a 32bit server cannot support
1676                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1677                  * requests
1678                  */
1679                 if (!is_compat)
1680                         return -EINVAL;
1681
1682                 for (i = 0; i < count; i++) {
1683                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1684                         dst[i].iov_len = ciov[i].iov_len;
1685                 }
1686                 return 0;
1687         }
1688 #endif
1689
1690         if (count * sizeof(struct iovec) != transferred)
1691                 return -EIO;
1692
1693         memcpy(dst, src, transferred);
1694         return 0;
1695 }
1696
1697 /* Make sure iov_length() won't overflow */
1698 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1699 {
1700         size_t n;
1701         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1702
1703         for (n = 0; n < count; n++) {
1704                 if (iov->iov_len > (size_t) max)
1705                         return -ENOMEM;
1706                 max -= iov->iov_len;
1707         }
1708         return 0;
1709 }
1710
1711 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1712                                  void *src, size_t transferred, unsigned count,
1713                                  bool is_compat)
1714 {
1715         unsigned i;
1716         struct fuse_ioctl_iovec *fiov = src;
1717
1718         if (fc->minor < 16) {
1719                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1720                                                  count, is_compat);
1721         }
1722
1723         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1724                 return -EIO;
1725
1726         for (i = 0; i < count; i++) {
1727                 /* Did the server supply an inappropriate value? */
1728                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1729                     fiov[i].len != (unsigned long) fiov[i].len)
1730                         return -EIO;
1731
1732                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1733                 dst[i].iov_len = (size_t) fiov[i].len;
1734
1735 #ifdef CONFIG_COMPAT
1736                 if (is_compat &&
1737                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1738                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1739                         return -EIO;
1740 #endif
1741         }
1742
1743         return 0;
1744 }
1745
1746
1747 /*
1748  * For ioctls, there is no generic way to determine how much memory
1749  * needs to be read and/or written.  Furthermore, ioctls are allowed
1750  * to dereference the passed pointer, so the parameter requires deep
1751  * copying but FUSE has no idea whatsoever about what to copy in or
1752  * out.
1753  *
1754  * This is solved by allowing FUSE server to retry ioctl with
1755  * necessary in/out iovecs.  Let's assume the ioctl implementation
1756  * needs to read in the following structure.
1757  *
1758  * struct a {
1759  *      char    *buf;
1760  *      size_t  buflen;
1761  * }
1762  *
1763  * On the first callout to FUSE server, inarg->in_size and
1764  * inarg->out_size will be NULL; then, the server completes the ioctl
1765  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1766  * the actual iov array to
1767  *
1768  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1769  *
1770  * which tells FUSE to copy in the requested area and retry the ioctl.
1771  * On the second round, the server has access to the structure and
1772  * from that it can tell what to look for next, so on the invocation,
1773  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1774  *
1775  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1776  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1777  *
1778  * FUSE will copy both struct a and the pointed buffer from the
1779  * process doing the ioctl and retry ioctl with both struct a and the
1780  * buffer.
1781  *
1782  * This time, FUSE server has everything it needs and completes ioctl
1783  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1784  *
1785  * Copying data out works the same way.
1786  *
1787  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1788  * automatically initializes in and out iovs by decoding @cmd with
1789  * _IOC_* macros and the server is not allowed to request RETRY.  This
1790  * limits ioctl data transfers to well-formed ioctls and is the forced
1791  * behavior for all FUSE servers.
1792  */
1793 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1794                    unsigned int flags)
1795 {
1796         struct fuse_file *ff = file->private_data;
1797         struct fuse_conn *fc = ff->fc;
1798         struct fuse_ioctl_in inarg = {
1799                 .fh = ff->fh,
1800                 .cmd = cmd,
1801                 .arg = arg,
1802                 .flags = flags
1803         };
1804         struct fuse_ioctl_out outarg;
1805         struct fuse_req *req = NULL;
1806         struct page **pages = NULL;
1807         struct iovec *iov_page = NULL;
1808         struct iovec *in_iov = NULL, *out_iov = NULL;
1809         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1810         size_t in_size, out_size, transferred;
1811         int err;
1812
1813 #if BITS_PER_LONG == 32
1814         inarg.flags |= FUSE_IOCTL_32BIT;
1815 #else
1816         if (flags & FUSE_IOCTL_COMPAT)
1817                 inarg.flags |= FUSE_IOCTL_32BIT;
1818 #endif
1819
1820         /* assume all the iovs returned by client always fits in a page */
1821         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1822
1823         err = -ENOMEM;
1824         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1825         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1826         if (!pages || !iov_page)
1827                 goto out;
1828
1829         /*
1830          * If restricted, initialize IO parameters as encoded in @cmd.
1831          * RETRY from server is not allowed.
1832          */
1833         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1834                 struct iovec *iov = iov_page;
1835
1836                 iov->iov_base = (void __user *)arg;
1837                 iov->iov_len = _IOC_SIZE(cmd);
1838
1839                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1840                         in_iov = iov;
1841                         in_iovs = 1;
1842                 }
1843
1844                 if (_IOC_DIR(cmd) & _IOC_READ) {
1845                         out_iov = iov;
1846                         out_iovs = 1;
1847                 }
1848         }
1849
1850  retry:
1851         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1852         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1853
1854         /*
1855          * Out data can be used either for actual out data or iovs,
1856          * make sure there always is at least one page.
1857          */
1858         out_size = max_t(size_t, out_size, PAGE_SIZE);
1859         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1860
1861         /* make sure there are enough buffer pages and init request with them */
1862         err = -ENOMEM;
1863         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1864                 goto out;
1865         while (num_pages < max_pages) {
1866                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1867                 if (!pages[num_pages])
1868                         goto out;
1869                 num_pages++;
1870         }
1871
1872         req = fuse_get_req(fc);
1873         if (IS_ERR(req)) {
1874                 err = PTR_ERR(req);
1875                 req = NULL;
1876                 goto out;
1877         }
1878         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1879         req->num_pages = num_pages;
1880
1881         /* okay, let's send it to the client */
1882         req->in.h.opcode = FUSE_IOCTL;
1883         req->in.h.nodeid = ff->nodeid;
1884         req->in.numargs = 1;
1885         req->in.args[0].size = sizeof(inarg);
1886         req->in.args[0].value = &inarg;
1887         if (in_size) {
1888                 req->in.numargs++;
1889                 req->in.args[1].size = in_size;
1890                 req->in.argpages = 1;
1891
1892                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1893                                            false);
1894                 if (err)
1895                         goto out;
1896         }
1897
1898         req->out.numargs = 2;
1899         req->out.args[0].size = sizeof(outarg);
1900         req->out.args[0].value = &outarg;
1901         req->out.args[1].size = out_size;
1902         req->out.argpages = 1;
1903         req->out.argvar = 1;
1904
1905         fuse_request_send(fc, req);
1906         err = req->out.h.error;
1907         transferred = req->out.args[1].size;
1908         fuse_put_request(fc, req);
1909         req = NULL;
1910         if (err)
1911                 goto out;
1912
1913         /* did it ask for retry? */
1914         if (outarg.flags & FUSE_IOCTL_RETRY) {
1915                 void *vaddr;
1916
1917                 /* no retry if in restricted mode */
1918                 err = -EIO;
1919                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1920                         goto out;
1921
1922                 in_iovs = outarg.in_iovs;
1923                 out_iovs = outarg.out_iovs;
1924
1925                 /*
1926                  * Make sure things are in boundary, separate checks
1927                  * are to protect against overflow.
1928                  */
1929                 err = -ENOMEM;
1930                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1931                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1932                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1933                         goto out;
1934
1935                 vaddr = kmap_atomic(pages[0]);
1936                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1937                                             transferred, in_iovs + out_iovs,
1938                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1939                 kunmap_atomic(vaddr);
1940                 if (err)
1941                         goto out;
1942
1943                 in_iov = iov_page;
1944                 out_iov = in_iov + in_iovs;
1945
1946                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1947                 if (err)
1948                         goto out;
1949
1950                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1951                 if (err)
1952                         goto out;
1953
1954                 goto retry;
1955         }
1956
1957         err = -EIO;
1958         if (transferred > inarg.out_size)
1959                 goto out;
1960
1961         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1962  out:
1963         if (req)
1964                 fuse_put_request(fc, req);
1965         free_page((unsigned long) iov_page);
1966         while (num_pages)
1967                 __free_page(pages[--num_pages]);
1968         kfree(pages);
1969
1970         return err ? err : outarg.result;
1971 }
1972 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1973
1974 long fuse_ioctl_common(struct file *file, unsigned int cmd,
1975                        unsigned long arg, unsigned int flags)
1976 {
1977         struct inode *inode = file->f_dentry->d_inode;
1978         struct fuse_conn *fc = get_fuse_conn(inode);
1979
1980         if (!fuse_allow_task(fc, current))
1981                 return -EACCES;
1982
1983         if (is_bad_inode(inode))
1984                 return -EIO;
1985
1986         return fuse_do_ioctl(file, cmd, arg, flags);
1987 }
1988
1989 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1990                             unsigned long arg)
1991 {
1992         return fuse_ioctl_common(file, cmd, arg, 0);
1993 }
1994
1995 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1996                                    unsigned long arg)
1997 {
1998         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1999 }
2000
2001 /*
2002  * All files which have been polled are linked to RB tree
2003  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2004  * find the matching one.
2005  */
2006 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2007                                               struct rb_node **parent_out)
2008 {
2009         struct rb_node **link = &fc->polled_files.rb_node;
2010         struct rb_node *last = NULL;
2011
2012         while (*link) {
2013                 struct fuse_file *ff;
2014
2015                 last = *link;
2016                 ff = rb_entry(last, struct fuse_file, polled_node);
2017
2018                 if (kh < ff->kh)
2019                         link = &last->rb_left;
2020                 else if (kh > ff->kh)
2021                         link = &last->rb_right;
2022                 else
2023                         return link;
2024         }
2025
2026         if (parent_out)
2027                 *parent_out = last;
2028         return link;
2029 }
2030
2031 /*
2032  * The file is about to be polled.  Make sure it's on the polled_files
2033  * RB tree.  Note that files once added to the polled_files tree are
2034  * not removed before the file is released.  This is because a file
2035  * polled once is likely to be polled again.
2036  */
2037 static void fuse_register_polled_file(struct fuse_conn *fc,
2038                                       struct fuse_file *ff)
2039 {
2040         spin_lock(&fc->lock);
2041         if (RB_EMPTY_NODE(&ff->polled_node)) {
2042                 struct rb_node **link, *parent;
2043
2044                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2045                 BUG_ON(*link);
2046                 rb_link_node(&ff->polled_node, parent, link);
2047                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2048         }
2049         spin_unlock(&fc->lock);
2050 }
2051
2052 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2053 {
2054         struct fuse_file *ff = file->private_data;
2055         struct fuse_conn *fc = ff->fc;
2056         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2057         struct fuse_poll_out outarg;
2058         struct fuse_req *req;
2059         int err;
2060
2061         if (fc->no_poll)
2062                 return DEFAULT_POLLMASK;
2063
2064         poll_wait(file, &ff->poll_wait, wait);
2065
2066         /*
2067          * Ask for notification iff there's someone waiting for it.
2068          * The client may ignore the flag and always notify.
2069          */
2070         if (waitqueue_active(&ff->poll_wait)) {
2071                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2072                 fuse_register_polled_file(fc, ff);
2073         }
2074
2075         req = fuse_get_req(fc);
2076         if (IS_ERR(req))
2077                 return POLLERR;
2078
2079         req->in.h.opcode = FUSE_POLL;
2080         req->in.h.nodeid = ff->nodeid;
2081         req->in.numargs = 1;
2082         req->in.args[0].size = sizeof(inarg);
2083         req->in.args[0].value = &inarg;
2084         req->out.numargs = 1;
2085         req->out.args[0].size = sizeof(outarg);
2086         req->out.args[0].value = &outarg;
2087         fuse_request_send(fc, req);
2088         err = req->out.h.error;
2089         fuse_put_request(fc, req);
2090
2091         if (!err)
2092                 return outarg.revents;
2093         if (err == -ENOSYS) {
2094                 fc->no_poll = 1;
2095                 return DEFAULT_POLLMASK;
2096         }
2097         return POLLERR;
2098 }
2099 EXPORT_SYMBOL_GPL(fuse_file_poll);
2100
2101 /*
2102  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2103  * wakes up the poll waiters.
2104  */
2105 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2106                             struct fuse_notify_poll_wakeup_out *outarg)
2107 {
2108         u64 kh = outarg->kh;
2109         struct rb_node **link;
2110
2111         spin_lock(&fc->lock);
2112
2113         link = fuse_find_polled_node(fc, kh, NULL);
2114         if (*link) {
2115                 struct fuse_file *ff;
2116
2117                 ff = rb_entry(*link, struct fuse_file, polled_node);
2118                 wake_up_interruptible_sync(&ff->poll_wait);
2119         }
2120
2121         spin_unlock(&fc->lock);
2122         return 0;
2123 }
2124
2125 static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
2126                              unsigned long nr_segs, loff_t *ppos, int rw)
2127 {
2128         const struct iovec *vector = iov;
2129         ssize_t ret = 0;
2130
2131         while (nr_segs > 0) {
2132                 void __user *base;
2133                 size_t len;
2134                 ssize_t nr;
2135
2136                 base = vector->iov_base;
2137                 len = vector->iov_len;
2138                 vector++;
2139                 nr_segs--;
2140
2141                 if (rw == WRITE)
2142                         nr = __fuse_direct_write(filp, base, len, ppos);
2143                 else
2144                         nr = fuse_direct_read(filp, base, len, ppos);
2145
2146                 if (nr < 0) {
2147                         if (!ret)
2148                                 ret = nr;
2149                         break;
2150                 }
2151                 ret += nr;
2152                 if (nr != len)
2153                         break;
2154         }
2155
2156         return ret;
2157 }
2158
2159
2160 static ssize_t
2161 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2162                         loff_t offset, unsigned long nr_segs)
2163 {
2164         ssize_t ret = 0;
2165         struct file *file = NULL;
2166         loff_t pos = 0;
2167
2168         file = iocb->ki_filp;
2169         pos = offset;
2170
2171         ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
2172
2173         return ret;
2174 }
2175
2176 long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2177                             loff_t length)
2178 {
2179         struct fuse_file *ff = file->private_data;
2180         struct fuse_conn *fc = ff->fc;
2181         struct fuse_req *req;
2182         struct fuse_fallocate_in inarg = {
2183                 .fh = ff->fh,
2184                 .offset = offset,
2185                 .length = length,
2186                 .mode = mode
2187         };
2188         int err;
2189
2190         if (fc->no_fallocate)
2191                 return -EOPNOTSUPP;
2192
2193         req = fuse_get_req(fc);
2194         if (IS_ERR(req))
2195                 return PTR_ERR(req);
2196
2197         req->in.h.opcode = FUSE_FALLOCATE;
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         fuse_request_send(fc, req);
2203         err = req->out.h.error;
2204         if (err == -ENOSYS) {
2205                 fc->no_fallocate = 1;
2206                 err = -EOPNOTSUPP;
2207         }
2208         fuse_put_request(fc, req);
2209
2210         return err;
2211 }
2212 EXPORT_SYMBOL_GPL(fuse_file_fallocate);
2213
2214 static const struct file_operations fuse_file_operations = {
2215         .llseek         = fuse_file_llseek,
2216         .read           = do_sync_read,
2217         .aio_read       = fuse_file_aio_read,
2218         .write          = do_sync_write,
2219         .aio_write      = fuse_file_aio_write,
2220         .mmap           = fuse_file_mmap,
2221         .open           = fuse_open,
2222         .flush          = fuse_flush,
2223         .release        = fuse_release,
2224         .fsync          = fuse_fsync,
2225         .lock           = fuse_file_lock,
2226         .flock          = fuse_file_flock,
2227         .splice_read    = generic_file_splice_read,
2228         .unlocked_ioctl = fuse_file_ioctl,
2229         .compat_ioctl   = fuse_file_compat_ioctl,
2230         .poll           = fuse_file_poll,
2231         .fallocate      = fuse_file_fallocate,
2232 };
2233
2234 static const struct file_operations fuse_direct_io_file_operations = {
2235         .llseek         = fuse_file_llseek,
2236         .read           = fuse_direct_read,
2237         .write          = fuse_direct_write,
2238         .mmap           = fuse_direct_mmap,
2239         .open           = fuse_open,
2240         .flush          = fuse_flush,
2241         .release        = fuse_release,
2242         .fsync          = fuse_fsync,
2243         .lock           = fuse_file_lock,
2244         .flock          = fuse_file_flock,
2245         .unlocked_ioctl = fuse_file_ioctl,
2246         .compat_ioctl   = fuse_file_compat_ioctl,
2247         .poll           = fuse_file_poll,
2248         .fallocate      = fuse_file_fallocate,
2249         /* no splice_read */
2250 };
2251
2252 static const struct address_space_operations fuse_file_aops  = {
2253         .readpage       = fuse_readpage,
2254         .writepage      = fuse_writepage,
2255         .launder_page   = fuse_launder_page,
2256         .readpages      = fuse_readpages,
2257         .set_page_dirty = __set_page_dirty_nobuffers,
2258         .bmap           = fuse_bmap,
2259         .direct_IO      = fuse_direct_IO,
2260 };
2261
2262 void fuse_init_file_inode(struct inode *inode)
2263 {
2264         inode->i_fop = &fuse_file_operations;
2265         inode->i_data.a_ops = &fuse_file_aops;
2266 }