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