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