fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18
  19 static const struct file_operations fuse_direct_io_file_operations;
  20
  21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  22                           int opcode, struct fuse_open_out *outargp)
  23 {
  24         struct fuse_open_in inarg;
  25         struct fuse_req *req;
  26         int err;
  27
  28         req = fuse_get_req_nopages(fc);
  29         if (IS_ERR(req))
  30                 return PTR_ERR(req);
  31
  32         memset(&inarg, 0, sizeof(inarg));
  33         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  34         if (!fc->atomic_o_trunc)
  35                 inarg.flags &= ~O_TRUNC;
  36         req->in.h.opcode = opcode;
  37         req->in.h.nodeid = nodeid;
  38         req->in.numargs = 1;
  39         req->in.args[0].size = sizeof(inarg);
  40         req->in.args[0].value = &inarg;
  41         req->out.numargs = 1;
  42         req->out.args[0].size = sizeof(*outargp);
  43         req->out.args[0].value = outargp;
  44         fuse_request_send(fc, req);
  45         err = req->out.h.error;
  46         fuse_put_request(fc, req);
  47
  48         return err;
  49 }
  50
  51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  52 {
  53         struct fuse_file *ff;
  54
  55         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  56         if (unlikely(!ff))
  57                 return NULL;
  58
  59         ff->fc = fc;
  60         ff->reserved_req = fuse_request_alloc(0);
  61         if (unlikely(!ff->reserved_req)) {
  62                 kfree(ff);
  63                 return NULL;
  64         }
  65
  66         INIT_LIST_HEAD(&ff->write_entry);
  67         atomic_set(&ff->count, 0);
  68         RB_CLEAR_NODE(&ff->polled_node);
  69         init_waitqueue_head(&ff->poll_wait);
  70
  71         spin_lock(&fc->lock);
  72         ff->kh = ++fc->khctr;
  73         spin_unlock(&fc->lock);
  74
  75         return ff;
  76 }
  77
  78 void fuse_file_free(struct fuse_file *ff)
  79 {
  80         fuse_request_free(ff->reserved_req);
  81         kfree(ff);
  82 }
  83
  84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  85 {
  86         atomic_inc(&ff->count);
  87         return ff;
  88 }
  89
  90 static void fuse_release_async(struct work_struct *work)
  91 {
  92         struct fuse_req *req;
  93         struct fuse_conn *fc;
  94         struct path path;
  95
  96         req = container_of(work, struct fuse_req, misc.release.work);
  97         path = req->misc.release.path;
  98         fc = get_fuse_conn(path.dentry->d_inode);
  99
 100         fuse_put_request(fc, req);
 101         path_put(&path);
 102 }
 103
 104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 105 {
 106         if (fc->destroy_req) {
 107                 /*
 108                  * If this is a fuseblk mount, then it's possible that
 109                  * releasing the path will result in releasing the
 110                  * super block and sending the DESTROY request.  If
 111                  * the server is single threaded, this would hang.
 112                  * For this reason do the path_put() in a separate
 113                  * thread.
 114                  */
 115                 atomic_inc(&req->count);
 116                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 117                 schedule_work(&req->misc.release.work);
 118         } else {
 119                 path_put(&req->misc.release.path);
 120         }
 121 }
 122
 123 static void fuse_file_put(struct fuse_file *ff, bool sync)
 124 {
 125         if (atomic_dec_and_test(&ff->count)) {
 126                 struct fuse_req *req = ff->reserved_req;
 127
 128                 if (sync) {
 129                         req->background = 0;
 130                         fuse_request_send(ff->fc, req);
 131                         path_put(&req->misc.release.path);
 132                         fuse_put_request(ff->fc, req);
 133                 } else {
 134                         req->end = fuse_release_end;
 135                         req->background = 1;
 136                         fuse_request_send_background(ff->fc, req);
 137                 }
 138                 kfree(ff);
 139         }
 140 }
 141
 142 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 143                  bool isdir)
 144 {
 145         struct fuse_open_out outarg;
 146         struct fuse_file *ff;
 147         int err;
 148         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 149
 150         ff = fuse_file_alloc(fc);
 151         if (!ff)
 152                 return -ENOMEM;
 153
 154         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 155         if (err) {
 156                 fuse_file_free(ff);
 157                 return err;
 158         }
 159
 160         if (isdir)
 161                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 162
 163         ff->fh = outarg.fh;
 164         ff->nodeid = nodeid;
 165         ff->open_flags = outarg.open_flags;
 166         file->private_data = fuse_file_get(ff);
 167
 168         return 0;
 169 }
 170 EXPORT_SYMBOL_GPL(fuse_do_open);
 171
 172 void fuse_finish_open(struct inode *inode, struct file *file)
 173 {
 174         struct fuse_file *ff = file->private_data;
 175         struct fuse_conn *fc = get_fuse_conn(inode);
 176
 177         if (ff->open_flags & FOPEN_DIRECT_IO)
 178                 file->f_op = &fuse_direct_io_file_operations;
 179         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 180                 invalidate_inode_pages2(inode->i_mapping);
 181         if (ff->open_flags & FOPEN_NONSEEKABLE)
 182                 nonseekable_open(inode, file);
 183         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 184                 struct fuse_inode *fi = get_fuse_inode(inode);
 185
 186                 spin_lock(&fc->lock);
 187                 fi->attr_version = ++fc->attr_version;
 188                 i_size_write(inode, 0);
 189                 spin_unlock(&fc->lock);
 190                 fuse_invalidate_attr(inode);
 191         }
 192 }
 193
 194 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 195 {
 196         struct fuse_conn *fc = get_fuse_conn(inode);
 197         int err;
 198
 199         err = generic_file_open(inode, file);
 200         if (err)
 201                 return err;
 202
 203         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 204         if (err)
 205                 return err;
 206
 207         fuse_finish_open(inode, file);
 208
 209         return 0;
 210 }
 211
 212 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 213 {
 214         struct fuse_conn *fc = ff->fc;
 215         struct fuse_req *req = ff->reserved_req;
 216         struct fuse_release_in *inarg = &req->misc.release.in;
 217
 218         spin_lock(&fc->lock);
 219         list_del(&ff->write_entry);
 220         if (!RB_EMPTY_NODE(&ff->polled_node))
 221                 rb_erase(&ff->polled_node, &fc->polled_files);
 222         spin_unlock(&fc->lock);
 223
 224         wake_up_interruptible_all(&ff->poll_wait);
 225
 226         inarg->fh = ff->fh;
 227         inarg->flags = flags;
 228         req->in.h.opcode = opcode;
 229         req->in.h.nodeid = ff->nodeid;
 230         req->in.numargs = 1;
 231         req->in.args[0].size = sizeof(struct fuse_release_in);
 232         req->in.args[0].value = inarg;
 233 }
 234
 235 void fuse_release_common(struct file *file, int opcode)
 236 {
 237         struct fuse_file *ff;
 238         struct fuse_req *req;
 239
 240         ff = file->private_data;
 241         if (unlikely(!ff))
 242                 return;
 243
 244         req = ff->reserved_req;
 245         fuse_prepare_release(ff, file->f_flags, opcode);
 246
 247         if (ff->flock) {
 248                 struct fuse_release_in *inarg = &req->misc.release.in;
 249                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 250                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 251                                                        (fl_owner_t) file);
 252         }
 253         /* Hold vfsmount and dentry until release is finished */
 254         path_get(&file->f_path);
 255         req->misc.release.path = file->f_path;
 256
 257         /*
 258          * Normally this will send the RELEASE request, however if
 259          * some asynchronous READ or WRITE requests are outstanding,
 260          * the sending will be delayed.
 261          *
 262          * Make the release synchronous if this is a fuseblk mount,
 263          * synchronous RELEASE is allowed (and desirable) in this case
 264          * because the server can be trusted not to screw up.
 265          */
 266         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 267 }
 268
 269 static int fuse_open(struct inode *inode, struct file *file)
 270 {
 271         return fuse_open_common(inode, file, false);
 272 }
 273
 274 static int fuse_release(struct inode *inode, struct file *file)
 275 {
 276         fuse_release_common(file, FUSE_RELEASE);
 277
 278         /* return value is ignored by VFS */
 279         return 0;
 280 }
 281
 282 void fuse_sync_release(struct fuse_file *ff, int flags)
 283 {
 284         WARN_ON(atomic_read(&ff->count) > 1);
 285         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 286         ff->reserved_req->force = 1;
 287         ff->reserved_req->background = 0;
 288         fuse_request_send(ff->fc, ff->reserved_req);
 289         fuse_put_request(ff->fc, ff->reserved_req);
 290         kfree(ff);
 291 }
 292 EXPORT_SYMBOL_GPL(fuse_sync_release);
 293
 294 /*
 295  * Scramble the ID space with XTEA, so that the value of the files_struct
 296  * pointer is not exposed to userspace.
 297  */
 298 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 299 {
 300         u32 *k = fc->scramble_key;
 301         u64 v = (unsigned long) id;
 302         u32 v0 = v;
 303         u32 v1 = v >> 32;
 304         u32 sum = 0;
 305         int i;
 306
 307         for (i = 0; i < 32; i++) {
 308                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 309                 sum += 0x9E3779B9;
 310                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 311         }
 312
 313         return (u64) v0 + ((u64) v1 << 32);
 314 }
 315
 316 /*
 317  * Check if page is under writeback
 318  *
 319  * This is currently done by walking the list of writepage requests
 320  * for the inode, which can be pretty inefficient.
 321  */
 322 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 323 {
 324         struct fuse_conn *fc = get_fuse_conn(inode);
 325         struct fuse_inode *fi = get_fuse_inode(inode);
 326         struct fuse_req *req;
 327         bool found = false;
 328
 329         spin_lock(&fc->lock);
 330         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 331                 pgoff_t curr_index;
 332
 333                 BUG_ON(req->inode != inode);
 334                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 335                 if (curr_index == index) {
 336                         found = true;
 337                         break;
 338                 }
 339         }
 340         spin_unlock(&fc->lock);
 341
 342         return found;
 343 }
 344
 345 /*
 346  * Wait for page writeback to be completed.
 347  *
 348  * Since fuse doesn't rely on the VM writeback tracking, this has to
 349  * use some other means.
 350  */
 351 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 352 {
 353         struct fuse_inode *fi = get_fuse_inode(inode);
 354
 355         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 356         return 0;
 357 }
 358
 359 static int fuse_flush(struct file *file, fl_owner_t id)
 360 {
 361         struct inode *inode = file_inode(file);
 362         struct fuse_conn *fc = get_fuse_conn(inode);
 363         struct fuse_file *ff = file->private_data;
 364         struct fuse_req *req;
 365         struct fuse_flush_in inarg;
 366         int err;
 367
 368         if (is_bad_inode(inode))
 369                 return -EIO;
 370
 371         if (fc->no_flush)
 372                 return 0;
 373
 374         req = fuse_get_req_nofail_nopages(fc, file);
 375         memset(&inarg, 0, sizeof(inarg));
 376         inarg.fh = ff->fh;
 377         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 378         req->in.h.opcode = FUSE_FLUSH;
 379         req->in.h.nodeid = get_node_id(inode);
 380         req->in.numargs = 1;
 381         req->in.args[0].size = sizeof(inarg);
 382         req->in.args[0].value = &inarg;
 383         req->force = 1;
 384         fuse_request_send(fc, req);
 385         err = req->out.h.error;
 386         fuse_put_request(fc, req);
 387         if (err == -ENOSYS) {
 388                 fc->no_flush = 1;
 389                 err = 0;
 390         }
 391         return err;
 392 }
 393
 394 /*
 395  * Wait for all pending writepages on the inode to finish.
 396  *
 397  * This is currently done by blocking further writes with FUSE_NOWRITE
 398  * and waiting for all sent writes to complete.
 399  *
 400  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 401  * could conflict with truncation.
 402  */
 403 static void fuse_sync_writes(struct inode *inode)
 404 {
 405         fuse_set_nowrite(inode);
 406         fuse_release_nowrite(inode);
 407 }
 408
 409 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 410                       int datasync, int isdir)
 411 {
 412         struct inode *inode = file->f_mapping->host;
 413         struct fuse_conn *fc = get_fuse_conn(inode);
 414         struct fuse_file *ff = file->private_data;
 415         struct fuse_req *req;
 416         struct fuse_fsync_in inarg;
 417         int err;
 418
 419         if (is_bad_inode(inode))
 420                 return -EIO;
 421
 422         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 423         if (err)
 424                 return err;
 425
 426         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 427                 return 0;
 428
 429         mutex_lock(&inode->i_mutex);
 430
 431         /*
 432          * Start writeback against all dirty pages of the inode, then
 433          * wait for all outstanding writes, before sending the FSYNC
 434          * request.
 435          */
 436         err = write_inode_now(inode, 0);
 437         if (err)
 438                 goto out;
 439
 440         fuse_sync_writes(inode);
 441
 442         req = fuse_get_req_nopages(fc);
 443         if (IS_ERR(req)) {
 444                 err = PTR_ERR(req);
 445                 goto out;
 446         }
 447
 448         memset(&inarg, 0, sizeof(inarg));
 449         inarg.fh = ff->fh;
 450         inarg.fsync_flags = datasync ? 1 : 0;
 451         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 452         req->in.h.nodeid = get_node_id(inode);
 453         req->in.numargs = 1;
 454         req->in.args[0].size = sizeof(inarg);
 455         req->in.args[0].value = &inarg;
 456         fuse_request_send(fc, req);
 457         err = req->out.h.error;
 458         fuse_put_request(fc, req);
 459         if (err == -ENOSYS) {
 460                 if (isdir)
 461                         fc->no_fsyncdir = 1;
 462                 else
 463                         fc->no_fsync = 1;
 464                 err = 0;
 465         }
 466 out:
 467         mutex_unlock(&inode->i_mutex);
 468         return err;
 469 }
 470
 471 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 472                       int datasync)
 473 {
 474         return fuse_fsync_common(file, start, end, datasync, 0);
 475 }
 476
 477 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 478                     size_t count, int opcode)
 479 {
 480         struct fuse_read_in *inarg = &req->misc.read.in;
 481         struct fuse_file *ff = file->private_data;
 482
 483         inarg->fh = ff->fh;
 484         inarg->offset = pos;
 485         inarg->size = count;
 486         inarg->flags = file->f_flags;
 487         req->in.h.opcode = opcode;
 488         req->in.h.nodeid = ff->nodeid;
 489         req->in.numargs = 1;
 490         req->in.args[0].size = sizeof(struct fuse_read_in);
 491         req->in.args[0].value = inarg;
 492         req->out.argvar = 1;
 493         req->out.numargs = 1;
 494         req->out.args[0].size = count;
 495 }
 496
 497 static void fuse_release_user_pages(struct fuse_req *req, int write)
 498 {
 499         unsigned i;
 500
 501         for (i = 0; i < req->num_pages; i++) {
 502                 struct page *page = req->pages[i];
 503                 if (write)
 504                         set_page_dirty_lock(page);
 505                 put_page(page);
 506         }
 507 }
 508
 509 /**
 510  * In case of short read, the caller sets 'pos' to the position of
 511  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 512  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 513  *
 514  * An example:
 515  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 516  * both submitted asynchronously. The first of them was ACKed by userspace as
 517  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 518  * second request was ACKed as short, e.g. only 1K was read, resulting in
 519  * pos == 33K.
 520  *
 521  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 522  * will be equal to the length of the longest contiguous fragment of
 523  * transferred data starting from the beginning of IO request.
 524  */
 525 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 526 {
 527         int left;
 528
 529         spin_lock(&io->lock);
 530         if (err)
 531                 io->err = io->err ? : err;
 532         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 533                 io->bytes = pos;
 534
 535         left = --io->reqs;
 536         spin_unlock(&io->lock);
 537
 538         if (!left) {
 539                 long res;
 540
 541                 if (io->err)
 542                         res = io->err;
 543                 else if (io->bytes >= 0 && io->write)
 544                         res = -EIO;
 545                 else {
 546                         res = io->bytes < 0 ? io->size : io->bytes;
 547
 548                         if (!is_sync_kiocb(io->iocb)) {
 549                                 struct path *path = &io->iocb->ki_filp->f_path;
 550                                 struct inode *inode = path->dentry->d_inode;
 551                                 struct fuse_conn *fc = get_fuse_conn(inode);
 552                                 struct fuse_inode *fi = get_fuse_inode(inode);
 553
 554                                 spin_lock(&fc->lock);
 555                                 fi->attr_version = ++fc->attr_version;
 556                                 spin_unlock(&fc->lock);
 557                         }
 558                 }
 559
 560                 aio_complete(io->iocb, res, 0);
 561                 kfree(io);
 562         }
 563 }
 564
 565 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 566 {
 567         struct fuse_io_priv *io = req->io;
 568         ssize_t pos = -1;
 569
 570         fuse_release_user_pages(req, !io->write);
 571
 572         if (io->write) {
 573                 if (req->misc.write.in.size != req->misc.write.out.size)
 574                         pos = req->misc.write.in.offset - io->offset +
 575                                 req->misc.write.out.size;
 576         } else {
 577                 if (req->misc.read.in.size != req->out.args[0].size)
 578                         pos = req->misc.read.in.offset - io->offset +
 579                                 req->out.args[0].size;
 580         }
 581
 582         fuse_aio_complete(io, req->out.h.error, pos);
 583 }
 584
 585 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 586                 size_t num_bytes, struct fuse_io_priv *io)
 587 {
 588         spin_lock(&io->lock);
 589         io->size += num_bytes;
 590         io->reqs++;
 591         spin_unlock(&io->lock);
 592
 593         req->io = io;
 594         req->end = fuse_aio_complete_req;
 595
 596         __fuse_get_request(req);
 597         fuse_request_send_background(fc, req);
 598
 599         return num_bytes;
 600 }
 601
 602 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 603                              loff_t pos, size_t count, fl_owner_t owner)
 604 {
 605         struct file *file = io->file;
 606         struct fuse_file *ff = file->private_data;
 607         struct fuse_conn *fc = ff->fc;
 608
 609         fuse_read_fill(req, file, pos, count, FUSE_READ);
 610         if (owner != NULL) {
 611                 struct fuse_read_in *inarg = &req->misc.read.in;
 612
 613                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 614                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 615         }
 616
 617         if (io->async)
 618                 return fuse_async_req_send(fc, req, count, io);
 619
 620         fuse_request_send(fc, req);
 621         return req->out.args[0].size;
 622 }
 623
 624 static void fuse_read_update_size(struct inode *inode, loff_t size,
 625                                   u64 attr_ver)
 626 {
 627         struct fuse_conn *fc = get_fuse_conn(inode);
 628         struct fuse_inode *fi = get_fuse_inode(inode);
 629
 630         spin_lock(&fc->lock);
 631         if (attr_ver == fi->attr_version && size < inode->i_size) {
 632                 fi->attr_version = ++fc->attr_version;
 633                 i_size_write(inode, size);
 634         }
 635         spin_unlock(&fc->lock);
 636 }
 637
 638 static int fuse_readpage(struct file *file, struct page *page)
 639 {
 640         struct fuse_io_priv io = { .async = 0, .file = file };
 641         struct inode *inode = page->mapping->host;
 642         struct fuse_conn *fc = get_fuse_conn(inode);
 643         struct fuse_req *req;
 644         size_t num_read;
 645         loff_t pos = page_offset(page);
 646         size_t count = PAGE_CACHE_SIZE;
 647         u64 attr_ver;
 648         int err;
 649
 650         err = -EIO;
 651         if (is_bad_inode(inode))
 652                 goto out;
 653
 654         /*
 655          * Page writeback can extend beyond the lifetime of the
 656          * page-cache page, so make sure we read a properly synced
 657          * page.
 658          */
 659         fuse_wait_on_page_writeback(inode, page->index);
 660
 661         req = fuse_get_req(fc, 1);
 662         err = PTR_ERR(req);
 663         if (IS_ERR(req))
 664                 goto out;
 665
 666         attr_ver = fuse_get_attr_version(fc);
 667
 668         req->out.page_zeroing = 1;
 669         req->out.argpages = 1;
 670         req->num_pages = 1;
 671         req->pages[0] = page;
 672         req->page_descs[0].length = count;
 673         num_read = fuse_send_read(req, &io, pos, count, NULL);
 674         err = req->out.h.error;
 675         fuse_put_request(fc, req);
 676
 677         if (!err) {
 678                 /*
 679                  * Short read means EOF.  If file size is larger, truncate it
 680                  */
 681                 if (num_read < count)
 682                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 683
 684                 SetPageUptodate(page);
 685         }
 686
 687         fuse_invalidate_attr(inode); /* atime changed */
 688  out:
 689         unlock_page(page);
 690         return err;
 691 }
 692
 693 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 694 {
 695         int i;
 696         size_t count = req->misc.read.in.size;
 697         size_t num_read = req->out.args[0].size;
 698         struct address_space *mapping = NULL;
 699
 700         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 701                 mapping = req->pages[i]->mapping;
 702
 703         if (mapping) {
 704                 struct inode *inode = mapping->host;
 705
 706                 /*
 707                  * Short read means EOF. If file size is larger, truncate it
 708                  */
 709                 if (!req->out.h.error && num_read < count) {
 710                         loff_t pos;
 711
 712                         pos = page_offset(req->pages[0]) + num_read;
 713                         fuse_read_update_size(inode, pos,
 714                                               req->misc.read.attr_ver);
 715                 }
 716                 fuse_invalidate_attr(inode); /* atime changed */
 717         }
 718
 719         for (i = 0; i < req->num_pages; i++) {
 720                 struct page *page = req->pages[i];
 721                 if (!req->out.h.error)
 722                         SetPageUptodate(page);
 723                 else
 724                         SetPageError(page);
 725                 unlock_page(page);
 726                 page_cache_release(page);
 727         }
 728         if (req->ff)
 729                 fuse_file_put(req->ff, false);
 730 }
 731
 732 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 733 {
 734         struct fuse_file *ff = file->private_data;
 735         struct fuse_conn *fc = ff->fc;
 736         loff_t pos = page_offset(req->pages[0]);
 737         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 738
 739         req->out.argpages = 1;
 740         req->out.page_zeroing = 1;
 741         req->out.page_replace = 1;
 742         fuse_read_fill(req, file, pos, count, FUSE_READ);
 743         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 744         if (fc->async_read) {
 745                 req->ff = fuse_file_get(ff);
 746                 req->end = fuse_readpages_end;
 747                 fuse_request_send_background(fc, req);
 748         } else {
 749                 fuse_request_send(fc, req);
 750                 fuse_readpages_end(fc, req);
 751                 fuse_put_request(fc, req);
 752         }
 753 }
 754
 755 struct fuse_fill_data {
 756         struct fuse_req *req;
 757         struct file *file;
 758         struct inode *inode;
 759         unsigned nr_pages;
 760 };
 761
 762 static int fuse_readpages_fill(void *_data, struct page *page)
 763 {
 764         struct fuse_fill_data *data = _data;
 765         struct fuse_req *req = data->req;
 766         struct inode *inode = data->inode;
 767         struct fuse_conn *fc = get_fuse_conn(inode);
 768
 769         fuse_wait_on_page_writeback(inode, page->index);
 770
 771         if (req->num_pages &&
 772             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 773              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 774              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 775                 int nr_alloc = min_t(unsigned, data->nr_pages,
 776                                      FUSE_MAX_PAGES_PER_REQ);
 777                 fuse_send_readpages(req, data->file);
 778                 if (fc->async_read)
 779                         req = fuse_get_req_for_background(fc, nr_alloc);
 780                 else
 781                         req = fuse_get_req(fc, nr_alloc);
 782
 783                 data->req = req;
 784                 if (IS_ERR(req)) {
 785                         unlock_page(page);
 786                         return PTR_ERR(req);
 787                 }
 788         }
 789
 790         if (WARN_ON(req->num_pages >= req->max_pages)) {
 791                 fuse_put_request(fc, req);
 792                 return -EIO;
 793         }
 794
 795         page_cache_get(page);
 796         req->pages[req->num_pages] = page;
 797         req->page_descs[req->num_pages].length = PAGE_SIZE;
 798         req->num_pages++;
 799         data->nr_pages--;
 800         return 0;
 801 }
 802
 803 static int fuse_readpages(struct file *file, struct address_space *mapping,
 804                           struct list_head *pages, unsigned nr_pages)
 805 {
 806         struct inode *inode = mapping->host;
 807         struct fuse_conn *fc = get_fuse_conn(inode);
 808         struct fuse_fill_data data;
 809         int err;
 810         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 811
 812         err = -EIO;
 813         if (is_bad_inode(inode))
 814                 goto out;
 815
 816         data.file = file;
 817         data.inode = inode;
 818         if (fc->async_read)
 819                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 820         else
 821                 data.req = fuse_get_req(fc, nr_alloc);
 822         data.nr_pages = nr_pages;
 823         err = PTR_ERR(data.req);
 824         if (IS_ERR(data.req))
 825                 goto out;
 826
 827         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 828         if (!err) {
 829                 if (data.req->num_pages)
 830                         fuse_send_readpages(data.req, file);
 831                 else
 832                         fuse_put_request(fc, data.req);
 833         }
 834 out:
 835         return err;
 836 }
 837
 838 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 839                                   unsigned long nr_segs, loff_t pos)
 840 {
 841         struct inode *inode = iocb->ki_filp->f_mapping->host;
 842         struct fuse_conn *fc = get_fuse_conn(inode);
 843
 844         /*
 845          * In auto invalidate mode, always update attributes on read.
 846          * Otherwise, only update if we attempt to read past EOF (to ensure
 847          * i_size is up to date).
 848          */
 849         if (fc->auto_inval_data ||
 850             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 851                 int err;
 852                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 853                 if (err)
 854                         return err;
 855         }
 856
 857         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 858 }
 859
 860 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 861                             loff_t pos, size_t count)
 862 {
 863         struct fuse_write_in *inarg = &req->misc.write.in;
 864         struct fuse_write_out *outarg = &req->misc.write.out;
 865
 866         inarg->fh = ff->fh;
 867         inarg->offset = pos;
 868         inarg->size = count;
 869         req->in.h.opcode = FUSE_WRITE;
 870         req->in.h.nodeid = ff->nodeid;
 871         req->in.numargs = 2;
 872         if (ff->fc->minor < 9)
 873                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 874         else
 875                 req->in.args[0].size = sizeof(struct fuse_write_in);
 876         req->in.args[0].value = inarg;
 877         req->in.args[1].size = count;
 878         req->out.numargs = 1;
 879         req->out.args[0].size = sizeof(struct fuse_write_out);
 880         req->out.args[0].value = outarg;
 881 }
 882
 883 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 884                               loff_t pos, size_t count, fl_owner_t owner)
 885 {
 886         struct file *file = io->file;
 887         struct fuse_file *ff = file->private_data;
 888         struct fuse_conn *fc = ff->fc;
 889         struct fuse_write_in *inarg = &req->misc.write.in;
 890
 891         fuse_write_fill(req, ff, pos, count);
 892         inarg->flags = file->f_flags;
 893         if (owner != NULL) {
 894                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 895                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 896         }
 897
 898         if (io->async)
 899                 return fuse_async_req_send(fc, req, count, io);
 900
 901         fuse_request_send(fc, req);
 902         return req->misc.write.out.size;
 903 }
 904
 905 void fuse_write_update_size(struct inode *inode, loff_t pos)
 906 {
 907         struct fuse_conn *fc = get_fuse_conn(inode);
 908         struct fuse_inode *fi = get_fuse_inode(inode);
 909
 910         spin_lock(&fc->lock);
 911         fi->attr_version = ++fc->attr_version;
 912         if (pos > inode->i_size)
 913                 i_size_write(inode, pos);
 914         spin_unlock(&fc->lock);
 915 }
 916
 917 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 918                                     struct inode *inode, loff_t pos,
 919                                     size_t count)
 920 {
 921         size_t res;
 922         unsigned offset;
 923         unsigned i;
 924         struct fuse_io_priv io = { .async = 0, .file = file };
 925
 926         for (i = 0; i < req->num_pages; i++)
 927                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 928
 929         res = fuse_send_write(req, &io, pos, count, NULL);
 930
 931         offset = req->page_descs[0].offset;
 932         count = res;
 933         for (i = 0; i < req->num_pages; i++) {
 934                 struct page *page = req->pages[i];
 935
 936                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 937                         SetPageUptodate(page);
 938
 939                 if (count > PAGE_CACHE_SIZE - offset)
 940                         count -= PAGE_CACHE_SIZE - offset;
 941                 else
 942                         count = 0;
 943                 offset = 0;
 944
 945                 unlock_page(page);
 946                 page_cache_release(page);
 947         }
 948
 949         return res;
 950 }
 951
 952 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 953                                struct address_space *mapping,
 954                                struct iov_iter *ii, loff_t pos)
 955 {
 956         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 957         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 958         size_t count = 0;
 959         int err;
 960
 961         req->in.argpages = 1;
 962         req->page_descs[0].offset = offset;
 963
 964         do {
 965                 size_t tmp;
 966                 struct page *page;
 967                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 968                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 969                                      iov_iter_count(ii));
 970
 971                 bytes = min_t(size_t, bytes, fc->max_write - count);
 972
 973  again:
 974                 err = -EFAULT;
 975                 if (iov_iter_fault_in_readable(ii, bytes))
 976                         break;
 977
 978                 err = -ENOMEM;
 979                 page = grab_cache_page_write_begin(mapping, index, 0);
 980                 if (!page)
 981                         break;
 982
 983                 if (mapping_writably_mapped(mapping))
 984                         flush_dcache_page(page);
 985
 986                 pagefault_disable();
 987                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 988                 pagefault_enable();
 989                 flush_dcache_page(page);
 990
 991                 mark_page_accessed(page);
 992
 993                 if (!tmp) {
 994                         unlock_page(page);
 995                         page_cache_release(page);
 996                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 997                         goto again;
 998                 }
 999
1000                 err = 0;
1001                 req->pages[req->num_pages] = page;
1002                 req->page_descs[req->num_pages].length = tmp;
1003                 req->num_pages++;
1004
1005                 iov_iter_advance(ii, tmp);
1006                 count += tmp;
1007                 pos += tmp;
1008                 offset += tmp;
1009                 if (offset == PAGE_CACHE_SIZE)
1010                         offset = 0;
1011
1012                 if (!fc->big_writes)
1013                         break;
1014         } while (iov_iter_count(ii) && count < fc->max_write &&
1015                  req->num_pages < req->max_pages && offset == 0);
1016
1017         return count > 0 ? count : err;
1018 }
1019
1020 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1021 {
1022         return min_t(unsigned,
1023                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1024                      (pos >> PAGE_CACHE_SHIFT) + 1,
1025                      FUSE_MAX_PAGES_PER_REQ);
1026 }
1027
1028 static ssize_t fuse_perform_write(struct file *file,
1029                                   struct address_space *mapping,
1030                                   struct iov_iter *ii, loff_t pos)
1031 {
1032         struct inode *inode = mapping->host;
1033         struct fuse_conn *fc = get_fuse_conn(inode);
1034         int err = 0;
1035         ssize_t res = 0;
1036
1037         if (is_bad_inode(inode))
1038                 return -EIO;
1039
1040         do {
1041                 struct fuse_req *req;
1042                 ssize_t count;
1043                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1044
1045                 req = fuse_get_req(fc, nr_pages);
1046                 if (IS_ERR(req)) {
1047                         err = PTR_ERR(req);
1048                         break;
1049                 }
1050
1051                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1052                 if (count <= 0) {
1053                         err = count;
1054                 } else {
1055                         size_t num_written;
1056
1057                         num_written = fuse_send_write_pages(req, file, inode,
1058                                                             pos, count);
1059                         err = req->out.h.error;
1060                         if (!err) {
1061                                 res += num_written;
1062                                 pos += num_written;
1063
1064                                 /* break out of the loop on short write */
1065                                 if (num_written != count)
1066                                         err = -EIO;
1067                         }
1068                 }
1069                 fuse_put_request(fc, req);
1070         } while (!err && iov_iter_count(ii));
1071
1072         if (res > 0)
1073                 fuse_write_update_size(inode, pos);
1074
1075         fuse_invalidate_attr(inode);
1076
1077         return res > 0 ? res : err;
1078 }
1079
1080 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1081                                    unsigned long nr_segs, loff_t pos)
1082 {
1083         struct file *file = iocb->ki_filp;
1084         struct address_space *mapping = file->f_mapping;
1085         size_t count = 0;
1086         size_t ocount = 0;
1087         ssize_t written = 0;
1088         ssize_t written_buffered = 0;
1089         struct inode *inode = mapping->host;
1090         ssize_t err;
1091         struct iov_iter i;
1092         loff_t endbyte = 0;
1093
1094         WARN_ON(iocb->ki_pos != pos);
1095
1096         ocount = 0;
1097         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1098         if (err)
1099                 return err;
1100
1101         count = ocount;
1102         sb_start_write(inode->i_sb);
1103         mutex_lock(&inode->i_mutex);
1104
1105         /* We can write back this queue in page reclaim */
1106         current->backing_dev_info = mapping->backing_dev_info;
1107
1108         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1109         if (err)
1110                 goto out;
1111
1112         if (count == 0)
1113                 goto out;
1114
1115         err = file_remove_suid(file);
1116         if (err)
1117                 goto out;
1118
1119         err = file_update_time(file);
1120         if (err)
1121                 goto out;
1122
1123         if (file->f_flags & O_DIRECT) {
1124                 written = generic_file_direct_write(iocb, iov, &nr_segs,
1125                                                     pos, &iocb->ki_pos,
1126                                                     count, ocount);
1127                 if (written < 0 || written == count)
1128                         goto out;
1129
1130                 pos += written;
1131                 count -= written;
1132
1133                 iov_iter_init(&i, iov, nr_segs, count, written);
1134                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1135                 if (written_buffered < 0) {
1136                         err = written_buffered;
1137                         goto out;
1138                 }
1139                 endbyte = pos + written_buffered - 1;
1140
1141                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1142                                                    endbyte);
1143                 if (err)
1144                         goto out;
1145
1146                 invalidate_mapping_pages(file->f_mapping,
1147                                          pos >> PAGE_CACHE_SHIFT,
1148                                          endbyte >> PAGE_CACHE_SHIFT);
1149
1150                 written += written_buffered;
1151                 iocb->ki_pos = pos + written_buffered;
1152         } else {
1153                 iov_iter_init(&i, iov, nr_segs, count, 0);
1154                 written = fuse_perform_write(file, mapping, &i, pos);
1155                 if (written >= 0)
1156                         iocb->ki_pos = pos + written;
1157         }
1158 out:
1159         current->backing_dev_info = NULL;
1160         mutex_unlock(&inode->i_mutex);
1161         sb_end_write(inode->i_sb);
1162
1163         return written ? written : err;
1164 }
1165
1166 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1167                 unsigned index, unsigned nr_pages)
1168 {
1169         int i;
1170
1171         for (i = index; i < index + nr_pages; i++)
1172                 req->page_descs[i].length = PAGE_SIZE -
1173                         req->page_descs[i].offset;
1174 }
1175
1176 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1177 {
1178         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1179 }
1180
1181 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1182                                         size_t max_size)
1183 {
1184         return min(iov_iter_single_seg_count(ii), max_size);
1185 }
1186
1187 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1188                                size_t *nbytesp, int write)
1189 {
1190         size_t nbytes = 0;  /* # bytes already packed in req */
1191
1192         /* Special case for kernel I/O: can copy directly into the buffer */
1193         if (segment_eq(get_fs(), KERNEL_DS)) {
1194                 unsigned long user_addr = fuse_get_user_addr(ii);
1195                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1196
1197                 if (write)
1198                         req->in.args[1].value = (void *) user_addr;
1199                 else
1200                         req->out.args[0].value = (void *) user_addr;
1201
1202                 iov_iter_advance(ii, frag_size);
1203                 *nbytesp = frag_size;
1204                 return 0;
1205         }
1206
1207         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1208                 unsigned npages;
1209                 unsigned long user_addr = fuse_get_user_addr(ii);
1210                 unsigned offset = user_addr & ~PAGE_MASK;
1211                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1212                 int ret;
1213
1214                 unsigned n = req->max_pages - req->num_pages;
1215                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1216
1217                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1218                 npages = clamp(npages, 1U, n);
1219
1220                 ret = get_user_pages_fast(user_addr, npages, !write,
1221                                           &req->pages[req->num_pages]);
1222                 if (ret < 0)
1223                         return ret;
1224
1225                 npages = ret;
1226                 frag_size = min_t(size_t, frag_size,
1227                                   (npages << PAGE_SHIFT) - offset);
1228                 iov_iter_advance(ii, frag_size);
1229
1230                 req->page_descs[req->num_pages].offset = offset;
1231                 fuse_page_descs_length_init(req, req->num_pages, npages);
1232
1233                 req->num_pages += npages;
1234                 req->page_descs[req->num_pages - 1].length -=
1235                         (npages << PAGE_SHIFT) - offset - frag_size;
1236
1237                 nbytes += frag_size;
1238         }
1239
1240         if (write)
1241                 req->in.argpages = 1;
1242         else
1243                 req->out.argpages = 1;
1244
1245         *nbytesp = nbytes;
1246
1247         return 0;
1248 }
1249
1250 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1251 {
1252         struct iov_iter ii = *ii_p;
1253         int npages = 0;
1254
1255         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1256                 unsigned long user_addr = fuse_get_user_addr(&ii);
1257                 unsigned offset = user_addr & ~PAGE_MASK;
1258                 size_t frag_size = iov_iter_single_seg_count(&ii);
1259
1260                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1261                 iov_iter_advance(&ii, frag_size);
1262         }
1263
1264         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1265 }
1266
1267 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1268                        unsigned long nr_segs, size_t count, loff_t *ppos,
1269                        int write)
1270 {
1271         struct file *file = io->file;
1272         struct fuse_file *ff = file->private_data;
1273         struct fuse_conn *fc = ff->fc;
1274         size_t nmax = write ? fc->max_write : fc->max_read;
1275         loff_t pos = *ppos;
1276         ssize_t res = 0;
1277         struct fuse_req *req;
1278         struct iov_iter ii;
1279
1280         iov_iter_init(&ii, iov, nr_segs, count, 0);
1281
1282         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1283         if (IS_ERR(req))
1284                 return PTR_ERR(req);
1285
1286         while (count) {
1287                 size_t nres;
1288                 fl_owner_t owner = current->files;
1289                 size_t nbytes = min(count, nmax);
1290                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1291                 if (err) {
1292                         res = err;
1293                         break;
1294                 }
1295
1296                 if (write)
1297                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1298                 else
1299                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1300
1301                 if (!io->async)
1302                         fuse_release_user_pages(req, !write);
1303                 if (req->out.h.error) {
1304                         if (!res)
1305                                 res = req->out.h.error;
1306                         break;
1307                 } else if (nres > nbytes) {
1308                         res = -EIO;
1309                         break;
1310                 }
1311                 count -= nres;
1312                 res += nres;
1313                 pos += nres;
1314                 if (nres != nbytes)
1315                         break;
1316                 if (count) {
1317                         fuse_put_request(fc, req);
1318                         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1319                         if (IS_ERR(req))
1320                                 break;
1321                 }
1322         }
1323         if (!IS_ERR(req))
1324                 fuse_put_request(fc, req);
1325         if (res > 0)
1326                 *ppos = pos;
1327
1328         return res;
1329 }
1330 EXPORT_SYMBOL_GPL(fuse_direct_io);
1331
1332 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1333                                   const struct iovec *iov,
1334                                   unsigned long nr_segs, loff_t *ppos)
1335 {
1336         ssize_t res;
1337         struct file *file = io->file;
1338         struct inode *inode = file_inode(file);
1339
1340         if (is_bad_inode(inode))
1341                 return -EIO;
1342
1343         res = fuse_direct_io(io, iov, nr_segs, iov_length(iov, nr_segs),
1344                              ppos, 0);
1345
1346         fuse_invalidate_attr(inode);
1347
1348         return res;
1349 }
1350
1351 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1352                                      size_t count, loff_t *ppos)
1353 {
1354         struct fuse_io_priv io = { .async = 0, .file = file };
1355         struct iovec iov = { .iov_base = buf, .iov_len = count };
1356         return __fuse_direct_read(&io, &iov, 1, ppos);
1357 }
1358
1359 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1360                                    const struct iovec *iov,
1361                                    unsigned long nr_segs, loff_t *ppos)
1362 {
1363         struct file *file = io->file;
1364         struct inode *inode = file_inode(file);
1365         size_t count = iov_length(iov, nr_segs);
1366         ssize_t res;
1367
1368         res = generic_write_checks(file, ppos, &count, 0);
1369         if (!res) {
1370                 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 1);
1371                 if (!io->async && res > 0)
1372                         fuse_write_update_size(inode, *ppos);
1373         }
1374
1375         fuse_invalidate_attr(inode);
1376
1377         return res;
1378 }
1379
1380 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1381                                  size_t count, loff_t *ppos)
1382 {
1383         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1384         struct inode *inode = file_inode(file);
1385         ssize_t res;
1386         struct fuse_io_priv io = { .async = 0, .file = file };
1387
1388         if (is_bad_inode(inode))
1389                 return -EIO;
1390
1391         /* Don't allow parallel writes to the same file */
1392         mutex_lock(&inode->i_mutex);
1393         res = __fuse_direct_write(&io, &iov, 1, ppos);
1394         mutex_unlock(&inode->i_mutex);
1395
1396         return res;
1397 }
1398
1399 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1400 {
1401         __free_page(req->pages[0]);
1402         fuse_file_put(req->ff, false);
1403 }
1404
1405 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1406 {
1407         struct inode *inode = req->inode;
1408         struct fuse_inode *fi = get_fuse_inode(inode);
1409         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1410
1411         list_del(&req->writepages_entry);
1412         dec_bdi_stat(bdi, BDI_WRITEBACK);
1413         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1414         bdi_writeout_inc(bdi);
1415         wake_up(&fi->page_waitq);
1416 }
1417
1418 /* Called under fc->lock, may release and reacquire it */
1419 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1420 __releases(fc->lock)
1421 __acquires(fc->lock)
1422 {
1423         struct fuse_inode *fi = get_fuse_inode(req->inode);
1424         loff_t size = i_size_read(req->inode);
1425         struct fuse_write_in *inarg = &req->misc.write.in;
1426
1427         if (!fc->connected)
1428                 goto out_free;
1429
1430         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1431                 inarg->size = PAGE_CACHE_SIZE;
1432         } else if (inarg->offset < size) {
1433                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1434         } else {
1435                 /* Got truncated off completely */
1436                 goto out_free;
1437         }
1438
1439         req->in.args[1].size = inarg->size;
1440         fi->writectr++;
1441         fuse_request_send_background_locked(fc, req);
1442         return;
1443
1444  out_free:
1445         fuse_writepage_finish(fc, req);
1446         spin_unlock(&fc->lock);
1447         fuse_writepage_free(fc, req);
1448         fuse_put_request(fc, req);
1449         spin_lock(&fc->lock);
1450 }
1451
1452 /*
1453  * If fi->writectr is positive (no truncate or fsync going on) send
1454  * all queued writepage requests.
1455  *
1456  * Called with fc->lock
1457  */
1458 void fuse_flush_writepages(struct inode *inode)
1459 __releases(fc->lock)
1460 __acquires(fc->lock)
1461 {
1462         struct fuse_conn *fc = get_fuse_conn(inode);
1463         struct fuse_inode *fi = get_fuse_inode(inode);
1464         struct fuse_req *req;
1465
1466         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1467                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1468                 list_del_init(&req->list);
1469                 fuse_send_writepage(fc, req);
1470         }
1471 }
1472
1473 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1474 {
1475         struct inode *inode = req->inode;
1476         struct fuse_inode *fi = get_fuse_inode(inode);
1477
1478         mapping_set_error(inode->i_mapping, req->out.h.error);
1479         spin_lock(&fc->lock);
1480         fi->writectr--;
1481         fuse_writepage_finish(fc, req);
1482         spin_unlock(&fc->lock);
1483         fuse_writepage_free(fc, req);
1484 }
1485
1486 static int fuse_writepage_locked(struct page *page)
1487 {
1488         struct address_space *mapping = page->mapping;
1489         struct inode *inode = mapping->host;
1490         struct fuse_conn *fc = get_fuse_conn(inode);
1491         struct fuse_inode *fi = get_fuse_inode(inode);
1492         struct fuse_req *req;
1493         struct fuse_file *ff;
1494         struct page *tmp_page;
1495
1496         set_page_writeback(page);
1497
1498         req = fuse_request_alloc_nofs(1);
1499         if (!req)
1500                 goto err;
1501
1502         req->background = 1; /* writeback always goes to bg_queue */
1503         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1504         if (!tmp_page)
1505                 goto err_free;
1506
1507         spin_lock(&fc->lock);
1508         BUG_ON(list_empty(&fi->write_files));
1509         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1510         req->ff = fuse_file_get(ff);
1511         spin_unlock(&fc->lock);
1512
1513         fuse_write_fill(req, ff, page_offset(page), 0);
1514
1515         copy_highpage(tmp_page, page);
1516         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1517         req->in.argpages = 1;
1518         req->num_pages = 1;
1519         req->pages[0] = tmp_page;
1520         req->page_descs[0].offset = 0;
1521         req->page_descs[0].length = PAGE_SIZE;
1522         req->end = fuse_writepage_end;
1523         req->inode = inode;
1524
1525         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1526         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1527         end_page_writeback(page);
1528
1529         spin_lock(&fc->lock);
1530         list_add(&req->writepages_entry, &fi->writepages);
1531         list_add_tail(&req->list, &fi->queued_writes);
1532         fuse_flush_writepages(inode);
1533         spin_unlock(&fc->lock);
1534
1535         return 0;
1536
1537 err_free:
1538         fuse_request_free(req);
1539 err:
1540         end_page_writeback(page);
1541         return -ENOMEM;
1542 }
1543
1544 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1545 {
1546         int err;
1547
1548         err = fuse_writepage_locked(page);
1549         unlock_page(page);
1550
1551         return err;
1552 }
1553
1554 static int fuse_launder_page(struct page *page)
1555 {
1556         int err = 0;
1557         if (clear_page_dirty_for_io(page)) {
1558                 struct inode *inode = page->mapping->host;
1559                 err = fuse_writepage_locked(page);
1560                 if (!err)
1561                         fuse_wait_on_page_writeback(inode, page->index);
1562         }
1563         return err;
1564 }
1565
1566 /*
1567  * Write back dirty pages now, because there may not be any suitable
1568  * open files later
1569  */
1570 static void fuse_vma_close(struct vm_area_struct *vma)
1571 {
1572         filemap_write_and_wait(vma->vm_file->f_mapping);
1573 }
1574
1575 /*
1576  * Wait for writeback against this page to complete before allowing it
1577  * to be marked dirty again, and hence written back again, possibly
1578  * before the previous writepage completed.
1579  *
1580  * Block here, instead of in ->writepage(), so that the userspace fs
1581  * can only block processes actually operating on the filesystem.
1582  *
1583  * Otherwise unprivileged userspace fs would be able to block
1584  * unrelated:
1585  *
1586  * - page migration
1587  * - sync(2)
1588  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1589  */
1590 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1591 {
1592         struct page *page = vmf->page;
1593         /*
1594          * Don't use page->mapping as it may become NULL from a
1595          * concurrent truncate.
1596          */
1597         struct inode *inode = vma->vm_file->f_mapping->host;
1598
1599         fuse_wait_on_page_writeback(inode, page->index);
1600         return 0;
1601 }
1602
1603 static const struct vm_operations_struct fuse_file_vm_ops = {
1604         .close          = fuse_vma_close,
1605         .fault          = filemap_fault,
1606         .page_mkwrite   = fuse_page_mkwrite,
1607         .remap_pages    = generic_file_remap_pages,
1608 };
1609
1610 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1611 {
1612         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1613                 struct inode *inode = file_inode(file);
1614                 struct fuse_conn *fc = get_fuse_conn(inode);
1615                 struct fuse_inode *fi = get_fuse_inode(inode);
1616                 struct fuse_file *ff = file->private_data;
1617                 /*
1618                  * file may be written through mmap, so chain it onto the
1619                  * inodes's write_file list
1620                  */
1621                 spin_lock(&fc->lock);
1622                 if (list_empty(&ff->write_entry))
1623                         list_add(&ff->write_entry, &fi->write_files);
1624                 spin_unlock(&fc->lock);
1625         }
1626         file_accessed(file);
1627         vma->vm_ops = &fuse_file_vm_ops;
1628         return 0;
1629 }
1630
1631 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1632 {
1633         /* Can't provide the coherency needed for MAP_SHARED */
1634         if (vma->vm_flags & VM_MAYSHARE)
1635                 return -ENODEV;
1636
1637         invalidate_inode_pages2(file->f_mapping);
1638
1639         return generic_file_mmap(file, vma);
1640 }
1641
1642 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1643                                   struct file_lock *fl)
1644 {
1645         switch (ffl->type) {
1646         case F_UNLCK:
1647                 break;
1648
1649         case F_RDLCK:
1650         case F_WRLCK:
1651                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1652                     ffl->end < ffl->start)
1653                         return -EIO;
1654
1655                 fl->fl_start = ffl->start;
1656                 fl->fl_end = ffl->end;
1657                 fl->fl_pid = ffl->pid;
1658                 break;
1659
1660         default:
1661                 return -EIO;
1662         }
1663         fl->fl_type = ffl->type;
1664         return 0;
1665 }
1666
1667 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1668                          const struct file_lock *fl, int opcode, pid_t pid,
1669                          int flock)
1670 {
1671         struct inode *inode = file_inode(file);
1672         struct fuse_conn *fc = get_fuse_conn(inode);
1673         struct fuse_file *ff = file->private_data;
1674         struct fuse_lk_in *arg = &req->misc.lk_in;
1675
1676         arg->fh = ff->fh;
1677         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1678         arg->lk.start = fl->fl_start;
1679         arg->lk.end = fl->fl_end;
1680         arg->lk.type = fl->fl_type;
1681         arg->lk.pid = pid;
1682         if (flock)
1683                 arg->lk_flags |= FUSE_LK_FLOCK;
1684         req->in.h.opcode = opcode;
1685         req->in.h.nodeid = get_node_id(inode);
1686         req->in.numargs = 1;
1687         req->in.args[0].size = sizeof(*arg);
1688         req->in.args[0].value = arg;
1689 }
1690
1691 static int fuse_getlk(struct file *file, struct file_lock *fl)
1692 {
1693         struct inode *inode = file_inode(file);
1694         struct fuse_conn *fc = get_fuse_conn(inode);
1695         struct fuse_req *req;
1696         struct fuse_lk_out outarg;
1697         int err;
1698
1699         req = fuse_get_req_nopages(fc);
1700         if (IS_ERR(req))
1701                 return PTR_ERR(req);
1702
1703         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1704         req->out.numargs = 1;
1705         req->out.args[0].size = sizeof(outarg);
1706         req->out.args[0].value = &outarg;
1707         fuse_request_send(fc, req);
1708         err = req->out.h.error;
1709         fuse_put_request(fc, req);
1710         if (!err)
1711                 err = convert_fuse_file_lock(&outarg.lk, fl);
1712
1713         return err;
1714 }
1715
1716 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1717 {
1718         struct inode *inode = file_inode(file);
1719         struct fuse_conn *fc = get_fuse_conn(inode);
1720         struct fuse_req *req;
1721         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1722         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1723         int err;
1724
1725         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1726                 /* NLM needs asynchronous locks, which we don't support yet */
1727                 return -ENOLCK;
1728         }
1729
1730         /* Unlock on close is handled by the flush method */
1731         if (fl->fl_flags & FL_CLOSE)
1732                 return 0;
1733
1734         req = fuse_get_req_nopages(fc);
1735         if (IS_ERR(req))
1736                 return PTR_ERR(req);
1737
1738         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1739         fuse_request_send(fc, req);
1740         err = req->out.h.error;
1741         /* locking is restartable */
1742         if (err == -EINTR)
1743                 err = -ERESTARTSYS;
1744         fuse_put_request(fc, req);
1745         return err;
1746 }
1747
1748 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1749 {
1750         struct inode *inode = file_inode(file);
1751         struct fuse_conn *fc = get_fuse_conn(inode);
1752         int err;
1753
1754         if (cmd == F_CANCELLK) {
1755                 err = 0;
1756         } else if (cmd == F_GETLK) {
1757                 if (fc->no_lock) {
1758                         posix_test_lock(file, fl);
1759                         err = 0;
1760                 } else
1761                         err = fuse_getlk(file, fl);
1762         } else {
1763                 if (fc->no_lock)
1764                         err = posix_lock_file(file, fl, NULL);
1765                 else
1766                         err = fuse_setlk(file, fl, 0);
1767         }
1768         return err;
1769 }
1770
1771 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1772 {
1773         struct inode *inode = file_inode(file);
1774         struct fuse_conn *fc = get_fuse_conn(inode);
1775         int err;
1776
1777         if (fc->no_flock) {
1778                 err = flock_lock_file_wait(file, fl);
1779         } else {
1780                 struct fuse_file *ff = file->private_data;
1781
1782                 /* emulate flock with POSIX locks */
1783                 fl->fl_owner = (fl_owner_t) file;
1784                 ff->flock = true;
1785                 err = fuse_setlk(file, fl, 1);
1786         }
1787
1788         return err;
1789 }
1790
1791 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1792 {
1793         struct inode *inode = mapping->host;
1794         struct fuse_conn *fc = get_fuse_conn(inode);
1795         struct fuse_req *req;
1796         struct fuse_bmap_in inarg;
1797         struct fuse_bmap_out outarg;
1798         int err;
1799
1800         if (!inode->i_sb->s_bdev || fc->no_bmap)
1801                 return 0;
1802
1803         req = fuse_get_req_nopages(fc);
1804         if (IS_ERR(req))
1805                 return 0;
1806
1807         memset(&inarg, 0, sizeof(inarg));
1808         inarg.block = block;
1809         inarg.blocksize = inode->i_sb->s_blocksize;
1810         req->in.h.opcode = FUSE_BMAP;
1811         req->in.h.nodeid = get_node_id(inode);
1812         req->in.numargs = 1;
1813         req->in.args[0].size = sizeof(inarg);
1814         req->in.args[0].value = &inarg;
1815         req->out.numargs = 1;
1816         req->out.args[0].size = sizeof(outarg);
1817         req->out.args[0].value = &outarg;
1818         fuse_request_send(fc, req);
1819         err = req->out.h.error;
1820         fuse_put_request(fc, req);
1821         if (err == -ENOSYS)
1822                 fc->no_bmap = 1;
1823
1824         return err ? 0 : outarg.block;
1825 }
1826
1827 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1828 {
1829         loff_t retval;
1830         struct inode *inode = file_inode(file);
1831
1832         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1833         if (whence == SEEK_CUR || whence == SEEK_SET)
1834                 return generic_file_llseek(file, offset, whence);
1835
1836         mutex_lock(&inode->i_mutex);
1837         retval = fuse_update_attributes(inode, NULL, file, NULL);
1838         if (!retval)
1839                 retval = generic_file_llseek(file, offset, whence);
1840         mutex_unlock(&inode->i_mutex);
1841
1842         return retval;
1843 }
1844
1845 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1846                         unsigned int nr_segs, size_t bytes, bool to_user)
1847 {
1848         struct iov_iter ii;
1849         int page_idx = 0;
1850
1851         if (!bytes)
1852                 return 0;
1853
1854         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1855
1856         while (iov_iter_count(&ii)) {
1857                 struct page *page = pages[page_idx++];
1858                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1859                 void *kaddr;
1860
1861                 kaddr = kmap(page);
1862
1863                 while (todo) {
1864                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1865                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1866                         size_t copy = min(todo, iov_len);
1867                         size_t left;
1868
1869                         if (!to_user)
1870                                 left = copy_from_user(kaddr, uaddr, copy);
1871                         else
1872                                 left = copy_to_user(uaddr, kaddr, copy);
1873
1874                         if (unlikely(left))
1875                                 return -EFAULT;
1876
1877                         iov_iter_advance(&ii, copy);
1878                         todo -= copy;
1879                         kaddr += copy;
1880                 }
1881
1882                 kunmap(page);
1883         }
1884
1885         return 0;
1886 }
1887
1888 /*
1889  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1890  * ABI was defined to be 'struct iovec' which is different on 32bit
1891  * and 64bit.  Fortunately we can determine which structure the server
1892  * used from the size of the reply.
1893  */
1894 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1895                                      size_t transferred, unsigned count,
1896                                      bool is_compat)
1897 {
1898 #ifdef CONFIG_COMPAT
1899         if (count * sizeof(struct compat_iovec) == transferred) {
1900                 struct compat_iovec *ciov = src;
1901                 unsigned i;
1902
1903                 /*
1904                  * With this interface a 32bit server cannot support
1905                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1906                  * requests
1907                  */
1908                 if (!is_compat)
1909                         return -EINVAL;
1910
1911                 for (i = 0; i < count; i++) {
1912                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1913                         dst[i].iov_len = ciov[i].iov_len;
1914                 }
1915                 return 0;
1916         }
1917 #endif
1918
1919         if (count * sizeof(struct iovec) != transferred)
1920                 return -EIO;
1921
1922         memcpy(dst, src, transferred);
1923         return 0;
1924 }
1925
1926 /* Make sure iov_length() won't overflow */
1927 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1928 {
1929         size_t n;
1930         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1931
1932         for (n = 0; n < count; n++, iov++) {
1933                 if (iov->iov_len > (size_t) max)
1934                         return -ENOMEM;
1935                 max -= iov->iov_len;
1936         }
1937         return 0;
1938 }
1939
1940 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1941                                  void *src, size_t transferred, unsigned count,
1942                                  bool is_compat)
1943 {
1944         unsigned i;
1945         struct fuse_ioctl_iovec *fiov = src;
1946
1947         if (fc->minor < 16) {
1948                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1949                                                  count, is_compat);
1950         }
1951
1952         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1953                 return -EIO;
1954
1955         for (i = 0; i < count; i++) {
1956                 /* Did the server supply an inappropriate value? */
1957                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1958                     fiov[i].len != (unsigned long) fiov[i].len)
1959                         return -EIO;
1960
1961                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1962                 dst[i].iov_len = (size_t) fiov[i].len;
1963
1964 #ifdef CONFIG_COMPAT
1965                 if (is_compat &&
1966                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1967                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1968                         return -EIO;
1969 #endif
1970         }
1971
1972         return 0;
1973 }
1974
1975
1976 /*
1977  * For ioctls, there is no generic way to determine how much memory
1978  * needs to be read and/or written.  Furthermore, ioctls are allowed
1979  * to dereference the passed pointer, so the parameter requires deep
1980  * copying but FUSE has no idea whatsoever about what to copy in or
1981  * out.
1982  *
1983  * This is solved by allowing FUSE server to retry ioctl with
1984  * necessary in/out iovecs.  Let's assume the ioctl implementation
1985  * needs to read in the following structure.
1986  *
1987  * struct a {
1988  *      char    *buf;
1989  *      size_t  buflen;
1990  * }
1991  *
1992  * On the first callout to FUSE server, inarg->in_size and
1993  * inarg->out_size will be NULL; then, the server completes the ioctl
1994  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1995  * the actual iov array to
1996  *
1997  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1998  *
1999  * which tells FUSE to copy in the requested area and retry the ioctl.
2000  * On the second round, the server has access to the structure and
2001  * from that it can tell what to look for next, so on the invocation,
2002  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2003  *
2004  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2005  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2006  *
2007  * FUSE will copy both struct a and the pointed buffer from the
2008  * process doing the ioctl and retry ioctl with both struct a and the
2009  * buffer.
2010  *
2011  * This time, FUSE server has everything it needs and completes ioctl
2012  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2013  *
2014  * Copying data out works the same way.
2015  *
2016  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2017  * automatically initializes in and out iovs by decoding @cmd with
2018  * _IOC_* macros and the server is not allowed to request RETRY.  This
2019  * limits ioctl data transfers to well-formed ioctls and is the forced
2020  * behavior for all FUSE servers.
2021  */
2022 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2023                    unsigned int flags)
2024 {
2025         struct fuse_file *ff = file->private_data;
2026         struct fuse_conn *fc = ff->fc;
2027         struct fuse_ioctl_in inarg = {
2028                 .fh = ff->fh,
2029                 .cmd = cmd,
2030                 .arg = arg,
2031                 .flags = flags
2032         };
2033         struct fuse_ioctl_out outarg;
2034         struct fuse_req *req = NULL;
2035         struct page **pages = NULL;
2036         struct iovec *iov_page = NULL;
2037         struct iovec *in_iov = NULL, *out_iov = NULL;
2038         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2039         size_t in_size, out_size, transferred;
2040         int err;
2041
2042 #if BITS_PER_LONG == 32
2043         inarg.flags |= FUSE_IOCTL_32BIT;
2044 #else
2045         if (flags & FUSE_IOCTL_COMPAT)
2046                 inarg.flags |= FUSE_IOCTL_32BIT;
2047 #endif
2048
2049         /* assume all the iovs returned by client always fits in a page */
2050         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2051
2052         err = -ENOMEM;
2053         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2054         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2055         if (!pages || !iov_page)
2056                 goto out;
2057
2058         /*
2059          * If restricted, initialize IO parameters as encoded in @cmd.
2060          * RETRY from server is not allowed.
2061          */
2062         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2063                 struct iovec *iov = iov_page;
2064
2065                 iov->iov_base = (void __user *)arg;
2066                 iov->iov_len = _IOC_SIZE(cmd);
2067
2068                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2069                         in_iov = iov;
2070                         in_iovs = 1;
2071                 }
2072
2073                 if (_IOC_DIR(cmd) & _IOC_READ) {
2074                         out_iov = iov;
2075                         out_iovs = 1;
2076                 }
2077         }
2078
2079  retry:
2080         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2081         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2082
2083         /*
2084          * Out data can be used either for actual out data or iovs,
2085          * make sure there always is at least one page.
2086          */
2087         out_size = max_t(size_t, out_size, PAGE_SIZE);
2088         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2089
2090         /* make sure there are enough buffer pages and init request with them */
2091         err = -ENOMEM;
2092         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2093                 goto out;
2094         while (num_pages < max_pages) {
2095                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2096                 if (!pages[num_pages])
2097                         goto out;
2098                 num_pages++;
2099         }
2100
2101         req = fuse_get_req(fc, num_pages);
2102         if (IS_ERR(req)) {
2103                 err = PTR_ERR(req);
2104                 req = NULL;
2105                 goto out;
2106         }
2107         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2108         req->num_pages = num_pages;
2109         fuse_page_descs_length_init(req, 0, req->num_pages);
2110
2111         /* okay, let's send it to the client */
2112         req->in.h.opcode = FUSE_IOCTL;
2113         req->in.h.nodeid = ff->nodeid;
2114         req->in.numargs = 1;
2115         req->in.args[0].size = sizeof(inarg);
2116         req->in.args[0].value = &inarg;
2117         if (in_size) {
2118                 req->in.numargs++;
2119                 req->in.args[1].size = in_size;
2120                 req->in.argpages = 1;
2121
2122                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2123                                            false);
2124                 if (err)
2125                         goto out;
2126         }
2127
2128         req->out.numargs = 2;
2129         req->out.args[0].size = sizeof(outarg);
2130         req->out.args[0].value = &outarg;
2131         req->out.args[1].size = out_size;
2132         req->out.argpages = 1;
2133         req->out.argvar = 1;
2134
2135         fuse_request_send(fc, req);
2136         err = req->out.h.error;
2137         transferred = req->out.args[1].size;
2138         fuse_put_request(fc, req);
2139         req = NULL;
2140         if (err)
2141                 goto out;
2142
2143         /* did it ask for retry? */
2144         if (outarg.flags & FUSE_IOCTL_RETRY) {
2145                 void *vaddr;
2146
2147                 /* no retry if in restricted mode */
2148                 err = -EIO;
2149                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2150                         goto out;
2151
2152                 in_iovs = outarg.in_iovs;
2153                 out_iovs = outarg.out_iovs;
2154
2155                 /*
2156                  * Make sure things are in boundary, separate checks
2157                  * are to protect against overflow.
2158                  */
2159                 err = -ENOMEM;
2160                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2161                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2162                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2163                         goto out;
2164
2165                 vaddr = kmap_atomic(pages[0]);
2166                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2167                                             transferred, in_iovs + out_iovs,
2168                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2169                 kunmap_atomic(vaddr);
2170                 if (err)
2171                         goto out;
2172
2173                 in_iov = iov_page;
2174                 out_iov = in_iov + in_iovs;
2175
2176                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2177                 if (err)
2178                         goto out;
2179
2180                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2181                 if (err)
2182                         goto out;
2183
2184                 goto retry;
2185         }
2186
2187         err = -EIO;
2188         if (transferred > inarg.out_size)
2189                 goto out;
2190
2191         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2192  out:
2193         if (req)
2194                 fuse_put_request(fc, req);
2195         free_page((unsigned long) iov_page);
2196         while (num_pages)
2197                 __free_page(pages[--num_pages]);
2198         kfree(pages);
2199
2200         return err ? err : outarg.result;
2201 }
2202 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2203
2204 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2205                        unsigned long arg, unsigned int flags)
2206 {
2207         struct inode *inode = file_inode(file);
2208         struct fuse_conn *fc = get_fuse_conn(inode);
2209
2210         if (!fuse_allow_current_process(fc))
2211                 return -EACCES;
2212
2213         if (is_bad_inode(inode))
2214                 return -EIO;
2215
2216         return fuse_do_ioctl(file, cmd, arg, flags);
2217 }
2218
2219 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2220                             unsigned long arg)
2221 {
2222         return fuse_ioctl_common(file, cmd, arg, 0);
2223 }
2224
2225 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2226                                    unsigned long arg)
2227 {
2228         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2229 }
2230
2231 /*
2232  * All files which have been polled are linked to RB tree
2233  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2234  * find the matching one.
2235  */
2236 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2237                                               struct rb_node **parent_out)
2238 {
2239         struct rb_node **link = &fc->polled_files.rb_node;
2240         struct rb_node *last = NULL;
2241
2242         while (*link) {
2243                 struct fuse_file *ff;
2244
2245                 last = *link;
2246                 ff = rb_entry(last, struct fuse_file, polled_node);
2247
2248                 if (kh < ff->kh)
2249                         link = &last->rb_left;
2250                 else if (kh > ff->kh)
2251                         link = &last->rb_right;
2252                 else
2253                         return link;
2254         }
2255
2256         if (parent_out)
2257                 *parent_out = last;
2258         return link;
2259 }
2260
2261 /*
2262  * The file is about to be polled.  Make sure it's on the polled_files
2263  * RB tree.  Note that files once added to the polled_files tree are
2264  * not removed before the file is released.  This is because a file
2265  * polled once is likely to be polled again.
2266  */
2267 static void fuse_register_polled_file(struct fuse_conn *fc,
2268                                       struct fuse_file *ff)
2269 {
2270         spin_lock(&fc->lock);
2271         if (RB_EMPTY_NODE(&ff->polled_node)) {
2272                 struct rb_node **link, *parent;
2273
2274                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2275                 BUG_ON(*link);
2276                 rb_link_node(&ff->polled_node, parent, link);
2277                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2278         }
2279         spin_unlock(&fc->lock);
2280 }
2281
2282 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2283 {
2284         struct fuse_file *ff = file->private_data;
2285         struct fuse_conn *fc = ff->fc;
2286         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2287         struct fuse_poll_out outarg;
2288         struct fuse_req *req;
2289         int err;
2290
2291         if (fc->no_poll)
2292                 return DEFAULT_POLLMASK;
2293
2294         poll_wait(file, &ff->poll_wait, wait);
2295         inarg.events = (__u32)poll_requested_events(wait);
2296
2297         /*
2298          * Ask for notification iff there's someone waiting for it.
2299          * The client may ignore the flag and always notify.
2300          */
2301         if (waitqueue_active(&ff->poll_wait)) {
2302                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2303                 fuse_register_polled_file(fc, ff);
2304         }
2305
2306         req = fuse_get_req_nopages(fc);
2307         if (IS_ERR(req))
2308                 return POLLERR;
2309
2310         req->in.h.opcode = FUSE_POLL;
2311         req->in.h.nodeid = ff->nodeid;
2312         req->in.numargs = 1;
2313         req->in.args[0].size = sizeof(inarg);
2314         req->in.args[0].value = &inarg;
2315         req->out.numargs = 1;
2316         req->out.args[0].size = sizeof(outarg);
2317         req->out.args[0].value = &outarg;
2318         fuse_request_send(fc, req);
2319         err = req->out.h.error;
2320         fuse_put_request(fc, req);
2321
2322         if (!err)
2323                 return outarg.revents;
2324         if (err == -ENOSYS) {
2325                 fc->no_poll = 1;
2326                 return DEFAULT_POLLMASK;
2327         }
2328         return POLLERR;
2329 }
2330 EXPORT_SYMBOL_GPL(fuse_file_poll);
2331
2332 /*
2333  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2334  * wakes up the poll waiters.
2335  */
2336 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2337                             struct fuse_notify_poll_wakeup_out *outarg)
2338 {
2339         u64 kh = outarg->kh;
2340         struct rb_node **link;
2341
2342         spin_lock(&fc->lock);
2343
2344         link = fuse_find_polled_node(fc, kh, NULL);
2345         if (*link) {
2346                 struct fuse_file *ff;
2347
2348                 ff = rb_entry(*link, struct fuse_file, polled_node);
2349                 wake_up_interruptible_sync(&ff->poll_wait);
2350         }
2351
2352         spin_unlock(&fc->lock);
2353         return 0;
2354 }
2355
2356 static ssize_t
2357 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2358                         loff_t offset, unsigned long nr_segs)
2359 {
2360         ssize_t ret = 0;
2361         struct file *file = NULL;
2362         loff_t pos = 0;
2363         struct fuse_io_priv *io;
2364
2365         file = iocb->ki_filp;
2366         pos = offset;
2367
2368         io = kzalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2369         if (!io)
2370                 return -ENOMEM;
2371
2372         io->file = file;
2373
2374         if (rw == WRITE)
2375                 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2376         else
2377                 ret = __fuse_direct_read(io, iov, nr_segs, &pos);
2378
2379         kfree(io);
2380
2381         return ret;
2382 }
2383
2384 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2385                                 loff_t length)
2386 {
2387         struct fuse_file *ff = file->private_data;
2388         struct fuse_conn *fc = ff->fc;
2389         struct fuse_req *req;
2390         struct fuse_fallocate_in inarg = {
2391                 .fh = ff->fh,
2392                 .offset = offset,
2393                 .length = length,
2394                 .mode = mode
2395         };
2396         int err;
2397
2398         if (fc->no_fallocate)
2399                 return -EOPNOTSUPP;
2400
2401         req = fuse_get_req_nopages(fc);
2402         if (IS_ERR(req))
2403                 return PTR_ERR(req);
2404
2405         req->in.h.opcode = FUSE_FALLOCATE;
2406         req->in.h.nodeid = ff->nodeid;
2407         req->in.numargs = 1;
2408         req->in.args[0].size = sizeof(inarg);
2409         req->in.args[0].value = &inarg;
2410         fuse_request_send(fc, req);
2411         err = req->out.h.error;
2412         if (err == -ENOSYS) {
2413                 fc->no_fallocate = 1;
2414                 err = -EOPNOTSUPP;
2415         }
2416         fuse_put_request(fc, req);
2417
2418         return err;
2419 }
2420
2421 static const struct file_operations fuse_file_operations = {
2422         .llseek         = fuse_file_llseek,
2423         .read           = do_sync_read,
2424         .aio_read       = fuse_file_aio_read,
2425         .write          = do_sync_write,
2426         .aio_write      = fuse_file_aio_write,
2427         .mmap           = fuse_file_mmap,
2428         .open           = fuse_open,
2429         .flush          = fuse_flush,
2430         .release        = fuse_release,
2431         .fsync          = fuse_fsync,
2432         .lock           = fuse_file_lock,
2433         .flock          = fuse_file_flock,
2434         .splice_read    = generic_file_splice_read,
2435         .unlocked_ioctl = fuse_file_ioctl,
2436         .compat_ioctl   = fuse_file_compat_ioctl,
2437         .poll           = fuse_file_poll,
2438         .fallocate      = fuse_file_fallocate,
2439 };
2440
2441 static const struct file_operations fuse_direct_io_file_operations = {
2442         .llseek         = fuse_file_llseek,
2443         .read           = fuse_direct_read,
2444         .write          = fuse_direct_write,
2445         .mmap           = fuse_direct_mmap,
2446         .open           = fuse_open,
2447         .flush          = fuse_flush,
2448         .release        = fuse_release,
2449         .fsync          = fuse_fsync,
2450         .lock           = fuse_file_lock,
2451         .flock          = fuse_file_flock,
2452         .unlocked_ioctl = fuse_file_ioctl,
2453         .compat_ioctl   = fuse_file_compat_ioctl,
2454         .poll           = fuse_file_poll,
2455         .fallocate      = fuse_file_fallocate,
2456         /* no splice_read */
2457 };
2458
2459 static const struct address_space_operations fuse_file_aops  = {
2460         .readpage       = fuse_readpage,
2461         .writepage      = fuse_writepage,
2462         .launder_page   = fuse_launder_page,
2463         .readpages      = fuse_readpages,
2464         .set_page_dirty = __set_page_dirty_nobuffers,
2465         .bmap           = fuse_bmap,
2466         .direct_IO      = fuse_direct_IO,
2467 };
2468
2469 void fuse_init_file_inode(struct inode *inode)
2470 {
2471         inode->i_fop = &fuse_file_operations;
2472         inode->i_data.a_ops = &fuse_file_aops;
2473 }