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 #include <linux/aio.h>
  19 #include <linux/falloc.h>
  20
  21 static const struct file_operations fuse_direct_io_file_operations;
  22
  23 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  24                           int opcode, struct fuse_open_out *outargp)
  25 {
  26         struct fuse_open_in inarg;
  27         struct fuse_req *req;
  28         int err;
  29
  30         req = fuse_get_req_nopages(fc);
  31         if (IS_ERR(req))
  32                 return PTR_ERR(req);
  33
  34         memset(&inarg, 0, sizeof(inarg));
  35         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  36         if (!fc->atomic_o_trunc)
  37                 inarg.flags &= ~O_TRUNC;
  38         req->in.h.opcode = opcode;
  39         req->in.h.nodeid = nodeid;
  40         req->in.numargs = 1;
  41         req->in.args[0].size = sizeof(inarg);
  42         req->in.args[0].value = &inarg;
  43         req->out.numargs = 1;
  44         req->out.args[0].size = sizeof(*outargp);
  45         req->out.args[0].value = outargp;
  46         fuse_request_send(fc, req);
  47         err = req->out.h.error;
  48         fuse_put_request(fc, req);
  49
  50         return err;
  51 }
  52
  53 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  54 {
  55         struct fuse_file *ff;
  56
  57         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  58         if (unlikely(!ff))
  59                 return NULL;
  60
  61         ff->fc = fc;
  62         ff->reserved_req = fuse_request_alloc(0);
  63         if (unlikely(!ff->reserved_req)) {
  64                 kfree(ff);
  65                 return NULL;
  66         }
  67
  68         INIT_LIST_HEAD(&ff->write_entry);
  69         atomic_set(&ff->count, 0);
  70         RB_CLEAR_NODE(&ff->polled_node);
  71         init_waitqueue_head(&ff->poll_wait);
  72
  73         spin_lock(&fc->lock);
  74         ff->kh = ++fc->khctr;
  75         spin_unlock(&fc->lock);
  76
  77         return ff;
  78 }
  79
  80 void fuse_file_free(struct fuse_file *ff)
  81 {
  82         fuse_request_free(ff->reserved_req);
  83         kfree(ff);
  84 }
  85
  86 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  87 {
  88         atomic_inc(&ff->count);
  89         return ff;
  90 }
  91
  92 static void fuse_release_async(struct work_struct *work)
  93 {
  94         struct fuse_req *req;
  95         struct fuse_conn *fc;
  96         struct path path;
  97
  98         req = container_of(work, struct fuse_req, misc.release.work);
  99         path = req->misc.release.path;
 100         fc = get_fuse_conn(path.dentry->d_inode);
 101
 102         fuse_put_request(fc, req);
 103         path_put(&path);
 104 }
 105
 106 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 107 {
 108         if (fc->destroy_req) {
 109                 /*
 110                  * If this is a fuseblk mount, then it's possible that
 111                  * releasing the path will result in releasing the
 112                  * super block and sending the DESTROY request.  If
 113                  * the server is single threaded, this would hang.
 114                  * For this reason do the path_put() in a separate
 115                  * thread.
 116                  */
 117                 atomic_inc(&req->count);
 118                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 119                 schedule_work(&req->misc.release.work);
 120         } else {
 121                 path_put(&req->misc.release.path);
 122         }
 123 }
 124
 125 static void fuse_file_put(struct fuse_file *ff, bool sync)
 126 {
 127         if (atomic_dec_and_test(&ff->count)) {
 128                 struct fuse_req *req = ff->reserved_req;
 129
 130                 if (sync) {
 131                         req->background = 0;
 132                         fuse_request_send(ff->fc, req);
 133                         path_put(&req->misc.release.path);
 134                         fuse_put_request(ff->fc, req);
 135                 } else {
 136                         req->end = fuse_release_end;
 137                         req->background = 1;
 138                         fuse_request_send_background(ff->fc, req);
 139                 }
 140                 kfree(ff);
 141         }
 142 }
 143
 144 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 145                  bool isdir)
 146 {
 147         struct fuse_open_out outarg;
 148         struct fuse_file *ff;
 149         int err;
 150         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 151
 152         ff = fuse_file_alloc(fc);
 153         if (!ff)
 154                 return -ENOMEM;
 155
 156         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 157         if (err) {
 158                 fuse_file_free(ff);
 159                 return err;
 160         }
 161
 162         if (isdir)
 163                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 164
 165         ff->fh = outarg.fh;
 166         ff->nodeid = nodeid;
 167         ff->open_flags = outarg.open_flags;
 168         file->private_data = fuse_file_get(ff);
 169
 170         return 0;
 171 }
 172 EXPORT_SYMBOL_GPL(fuse_do_open);
 173
 174 void fuse_finish_open(struct inode *inode, struct file *file)
 175 {
 176         struct fuse_file *ff = file->private_data;
 177         struct fuse_conn *fc = get_fuse_conn(inode);
 178
 179         if (ff->open_flags & FOPEN_DIRECT_IO)
 180                 file->f_op = &fuse_direct_io_file_operations;
 181         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 182                 invalidate_inode_pages2(inode->i_mapping);
 183         if (ff->open_flags & FOPEN_NONSEEKABLE)
 184                 nonseekable_open(inode, file);
 185         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 186                 struct fuse_inode *fi = get_fuse_inode(inode);
 187
 188                 spin_lock(&fc->lock);
 189                 fi->attr_version = ++fc->attr_version;
 190                 i_size_write(inode, 0);
 191                 spin_unlock(&fc->lock);
 192                 fuse_invalidate_attr(inode);
 193         }
 194 }
 195
 196 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 197 {
 198         struct fuse_conn *fc = get_fuse_conn(inode);
 199         int err;
 200
 201         err = generic_file_open(inode, file);
 202         if (err)
 203                 return err;
 204
 205         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 206         if (err)
 207                 return err;
 208
 209         fuse_finish_open(inode, file);
 210
 211         return 0;
 212 }
 213
 214 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 215 {
 216         struct fuse_conn *fc = ff->fc;
 217         struct fuse_req *req = ff->reserved_req;
 218         struct fuse_release_in *inarg = &req->misc.release.in;
 219
 220         spin_lock(&fc->lock);
 221         list_del(&ff->write_entry);
 222         if (!RB_EMPTY_NODE(&ff->polled_node))
 223                 rb_erase(&ff->polled_node, &fc->polled_files);
 224         spin_unlock(&fc->lock);
 225
 226         wake_up_interruptible_all(&ff->poll_wait);
 227
 228         inarg->fh = ff->fh;
 229         inarg->flags = flags;
 230         req->in.h.opcode = opcode;
 231         req->in.h.nodeid = ff->nodeid;
 232         req->in.numargs = 1;
 233         req->in.args[0].size = sizeof(struct fuse_release_in);
 234         req->in.args[0].value = inarg;
 235 }
 236
 237 void fuse_release_common(struct file *file, int opcode)
 238 {
 239         struct fuse_file *ff;
 240         struct fuse_req *req;
 241
 242         ff = file->private_data;
 243         if (unlikely(!ff))
 244                 return;
 245
 246         req = ff->reserved_req;
 247         fuse_prepare_release(ff, file->f_flags, opcode);
 248
 249         if (ff->flock) {
 250                 struct fuse_release_in *inarg = &req->misc.release.in;
 251                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 252                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 253                                                        (fl_owner_t) file);
 254         }
 255         /* Hold vfsmount and dentry until release is finished */
 256         path_get(&file->f_path);
 257         req->misc.release.path = file->f_path;
 258
 259         /*
 260          * Normally this will send the RELEASE request, however if
 261          * some asynchronous READ or WRITE requests are outstanding,
 262          * the sending will be delayed.
 263          *
 264          * Make the release synchronous if this is a fuseblk mount,
 265          * synchronous RELEASE is allowed (and desirable) in this case
 266          * because the server can be trusted not to screw up.
 267          */
 268         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 269 }
 270
 271 static int fuse_open(struct inode *inode, struct file *file)
 272 {
 273         return fuse_open_common(inode, file, false);
 274 }
 275
 276 static int fuse_release(struct inode *inode, struct file *file)
 277 {
 278         fuse_release_common(file, FUSE_RELEASE);
 279
 280         /* return value is ignored by VFS */
 281         return 0;
 282 }
 283
 284 void fuse_sync_release(struct fuse_file *ff, int flags)
 285 {
 286         WARN_ON(atomic_read(&ff->count) > 1);
 287         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 288         ff->reserved_req->force = 1;
 289         ff->reserved_req->background = 0;
 290         fuse_request_send(ff->fc, ff->reserved_req);
 291         fuse_put_request(ff->fc, ff->reserved_req);
 292         kfree(ff);
 293 }
 294 EXPORT_SYMBOL_GPL(fuse_sync_release);
 295
 296 /*
 297  * Scramble the ID space with XTEA, so that the value of the files_struct
 298  * pointer is not exposed to userspace.
 299  */
 300 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 301 {
 302         u32 *k = fc->scramble_key;
 303         u64 v = (unsigned long) id;
 304         u32 v0 = v;
 305         u32 v1 = v >> 32;
 306         u32 sum = 0;
 307         int i;
 308
 309         for (i = 0; i < 32; i++) {
 310                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 311                 sum += 0x9E3779B9;
 312                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 313         }
 314
 315         return (u64) v0 + ((u64) v1 << 32);
 316 }
 317
 318 /*
 319  * Check if page is under writeback
 320  *
 321  * This is currently done by walking the list of writepage requests
 322  * for the inode, which can be pretty inefficient.
 323  */
 324 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 325 {
 326         struct fuse_conn *fc = get_fuse_conn(inode);
 327         struct fuse_inode *fi = get_fuse_inode(inode);
 328         struct fuse_req *req;
 329         bool found = false;
 330
 331         spin_lock(&fc->lock);
 332         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 333                 pgoff_t curr_index;
 334
 335                 BUG_ON(req->inode != inode);
 336                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 337                 if (curr_index == index) {
 338                         found = true;
 339                         break;
 340                 }
 341         }
 342         spin_unlock(&fc->lock);
 343
 344         return found;
 345 }
 346
 347 /*
 348  * Wait for page writeback to be completed.
 349  *
 350  * Since fuse doesn't rely on the VM writeback tracking, this has to
 351  * use some other means.
 352  */
 353 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 354 {
 355         struct fuse_inode *fi = get_fuse_inode(inode);
 356
 357         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 358         return 0;
 359 }
 360
 361 static int fuse_flush(struct file *file, fl_owner_t id)
 362 {
 363         struct inode *inode = file_inode(file);
 364         struct fuse_conn *fc = get_fuse_conn(inode);
 365         struct fuse_file *ff = file->private_data;
 366         struct fuse_req *req;
 367         struct fuse_flush_in inarg;
 368         int err;
 369
 370         if (is_bad_inode(inode))
 371                 return -EIO;
 372
 373         if (fc->no_flush)
 374                 return 0;
 375
 376         req = fuse_get_req_nofail_nopages(fc, file);
 377         memset(&inarg, 0, sizeof(inarg));
 378         inarg.fh = ff->fh;
 379         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 380         req->in.h.opcode = FUSE_FLUSH;
 381         req->in.h.nodeid = get_node_id(inode);
 382         req->in.numargs = 1;
 383         req->in.args[0].size = sizeof(inarg);
 384         req->in.args[0].value = &inarg;
 385         req->force = 1;
 386         fuse_request_send(fc, req);
 387         err = req->out.h.error;
 388         fuse_put_request(fc, req);
 389         if (err == -ENOSYS) {
 390                 fc->no_flush = 1;
 391                 err = 0;
 392         }
 393         return err;
 394 }
 395
 396 /*
 397  * Wait for all pending writepages on the inode to finish.
 398  *
 399  * This is currently done by blocking further writes with FUSE_NOWRITE
 400  * and waiting for all sent writes to complete.
 401  *
 402  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 403  * could conflict with truncation.
 404  */
 405 static void fuse_sync_writes(struct inode *inode)
 406 {
 407         fuse_set_nowrite(inode);
 408         fuse_release_nowrite(inode);
 409 }
 410
 411 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 412                       int datasync, int isdir)
 413 {
 414         struct inode *inode = file->f_mapping->host;
 415         struct fuse_conn *fc = get_fuse_conn(inode);
 416         struct fuse_file *ff = file->private_data;
 417         struct fuse_req *req;
 418         struct fuse_fsync_in inarg;
 419         int err;
 420
 421         if (is_bad_inode(inode))
 422                 return -EIO;
 423
 424         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 425         if (err)
 426                 return err;
 427
 428         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 429                 return 0;
 430
 431         mutex_lock(&inode->i_mutex);
 432
 433         /*
 434          * Start writeback against all dirty pages of the inode, then
 435          * wait for all outstanding writes, before sending the FSYNC
 436          * request.
 437          */
 438         err = write_inode_now(inode, 0);
 439         if (err)
 440                 goto out;
 441
 442         fuse_sync_writes(inode);
 443
 444         req = fuse_get_req_nopages(fc);
 445         if (IS_ERR(req)) {
 446                 err = PTR_ERR(req);
 447                 goto out;
 448         }
 449
 450         memset(&inarg, 0, sizeof(inarg));
 451         inarg.fh = ff->fh;
 452         inarg.fsync_flags = datasync ? 1 : 0;
 453         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 454         req->in.h.nodeid = get_node_id(inode);
 455         req->in.numargs = 1;
 456         req->in.args[0].size = sizeof(inarg);
 457         req->in.args[0].value = &inarg;
 458         fuse_request_send(fc, req);
 459         err = req->out.h.error;
 460         fuse_put_request(fc, req);
 461         if (err == -ENOSYS) {
 462                 if (isdir)
 463                         fc->no_fsyncdir = 1;
 464                 else
 465                         fc->no_fsync = 1;
 466                 err = 0;
 467         }
 468 out:
 469         mutex_unlock(&inode->i_mutex);
 470         return err;
 471 }
 472
 473 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 474                       int datasync)
 475 {
 476         return fuse_fsync_common(file, start, end, datasync, 0);
 477 }
 478
 479 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 480                     size_t count, int opcode)
 481 {
 482         struct fuse_read_in *inarg = &req->misc.read.in;
 483         struct fuse_file *ff = file->private_data;
 484
 485         inarg->fh = ff->fh;
 486         inarg->offset = pos;
 487         inarg->size = count;
 488         inarg->flags = file->f_flags;
 489         req->in.h.opcode = opcode;
 490         req->in.h.nodeid = ff->nodeid;
 491         req->in.numargs = 1;
 492         req->in.args[0].size = sizeof(struct fuse_read_in);
 493         req->in.args[0].value = inarg;
 494         req->out.argvar = 1;
 495         req->out.numargs = 1;
 496         req->out.args[0].size = count;
 497 }
 498
 499 static void fuse_release_user_pages(struct fuse_req *req, int write)
 500 {
 501         unsigned i;
 502
 503         for (i = 0; i < req->num_pages; i++) {
 504                 struct page *page = req->pages[i];
 505                 if (write)
 506                         set_page_dirty_lock(page);
 507                 put_page(page);
 508         }
 509 }
 510
 511 /**
 512  * In case of short read, the caller sets 'pos' to the position of
 513  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 514  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 515  *
 516  * An example:
 517  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 518  * both submitted asynchronously. The first of them was ACKed by userspace as
 519  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 520  * second request was ACKed as short, e.g. only 1K was read, resulting in
 521  * pos == 33K.
 522  *
 523  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 524  * will be equal to the length of the longest contiguous fragment of
 525  * transferred data starting from the beginning of IO request.
 526  */
 527 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 528 {
 529         int left;
 530
 531         spin_lock(&io->lock);
 532         if (err)
 533                 io->err = io->err ? : err;
 534         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 535                 io->bytes = pos;
 536
 537         left = --io->reqs;
 538         spin_unlock(&io->lock);
 539
 540         if (!left) {
 541                 long res;
 542
 543                 if (io->err)
 544                         res = io->err;
 545                 else if (io->bytes >= 0 && io->write)
 546                         res = -EIO;
 547                 else {
 548                         res = io->bytes < 0 ? io->size : io->bytes;
 549
 550                         if (!is_sync_kiocb(io->iocb)) {
 551                                 struct inode *inode = file_inode(io->iocb->ki_filp);
 552                                 struct fuse_conn *fc = get_fuse_conn(inode);
 553                                 struct fuse_inode *fi = get_fuse_inode(inode);
 554
 555                                 spin_lock(&fc->lock);
 556                                 fi->attr_version = ++fc->attr_version;
 557                                 spin_unlock(&fc->lock);
 558                         }
 559                 }
 560
 561                 aio_complete(io->iocb, res, 0);
 562                 kfree(io);
 563         }
 564 }
 565
 566 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 567 {
 568         struct fuse_io_priv *io = req->io;
 569         ssize_t pos = -1;
 570
 571         fuse_release_user_pages(req, !io->write);
 572
 573         if (io->write) {
 574                 if (req->misc.write.in.size != req->misc.write.out.size)
 575                         pos = req->misc.write.in.offset - io->offset +
 576                                 req->misc.write.out.size;
 577         } else {
 578                 if (req->misc.read.in.size != req->out.args[0].size)
 579                         pos = req->misc.read.in.offset - io->offset +
 580                                 req->out.args[0].size;
 581         }
 582
 583         fuse_aio_complete(io, req->out.h.error, pos);
 584 }
 585
 586 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 587                 size_t num_bytes, struct fuse_io_priv *io)
 588 {
 589         spin_lock(&io->lock);
 590         io->size += num_bytes;
 591         io->reqs++;
 592         spin_unlock(&io->lock);
 593
 594         req->io = io;
 595         req->end = fuse_aio_complete_req;
 596
 597         __fuse_get_request(req);
 598         fuse_request_send_background(fc, req);
 599
 600         return num_bytes;
 601 }
 602
 603 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 604                              loff_t pos, size_t count, fl_owner_t owner)
 605 {
 606         struct file *file = io->file;
 607         struct fuse_file *ff = file->private_data;
 608         struct fuse_conn *fc = ff->fc;
 609
 610         fuse_read_fill(req, file, pos, count, FUSE_READ);
 611         if (owner != NULL) {
 612                 struct fuse_read_in *inarg = &req->misc.read.in;
 613
 614                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 615                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 616         }
 617
 618         if (io->async)
 619                 return fuse_async_req_send(fc, req, count, io);
 620
 621         fuse_request_send(fc, req);
 622         return req->out.args[0].size;
 623 }
 624
 625 static void fuse_read_update_size(struct inode *inode, loff_t size,
 626                                   u64 attr_ver)
 627 {
 628         struct fuse_conn *fc = get_fuse_conn(inode);
 629         struct fuse_inode *fi = get_fuse_inode(inode);
 630
 631         spin_lock(&fc->lock);
 632         if (attr_ver == fi->attr_version && size < inode->i_size &&
 633             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 634                 fi->attr_version = ++fc->attr_version;
 635                 i_size_write(inode, size);
 636         }
 637         spin_unlock(&fc->lock);
 638 }
 639
 640 static int fuse_readpage(struct file *file, struct page *page)
 641 {
 642         struct fuse_io_priv io = { .async = 0, .file = file };
 643         struct inode *inode = page->mapping->host;
 644         struct fuse_conn *fc = get_fuse_conn(inode);
 645         struct fuse_req *req;
 646         size_t num_read;
 647         loff_t pos = page_offset(page);
 648         size_t count = PAGE_CACHE_SIZE;
 649         u64 attr_ver;
 650         int err;
 651
 652         err = -EIO;
 653         if (is_bad_inode(inode))
 654                 goto out;
 655
 656         /*
 657          * Page writeback can extend beyond the lifetime of the
 658          * page-cache page, so make sure we read a properly synced
 659          * page.
 660          */
 661         fuse_wait_on_page_writeback(inode, page->index);
 662
 663         req = fuse_get_req(fc, 1);
 664         err = PTR_ERR(req);
 665         if (IS_ERR(req))
 666                 goto out;
 667
 668         attr_ver = fuse_get_attr_version(fc);
 669
 670         req->out.page_zeroing = 1;
 671         req->out.argpages = 1;
 672         req->num_pages = 1;
 673         req->pages[0] = page;
 674         req->page_descs[0].length = count;
 675         num_read = fuse_send_read(req, &io, pos, count, NULL);
 676         err = req->out.h.error;
 677         fuse_put_request(fc, req);
 678
 679         if (!err) {
 680                 /*
 681                  * Short read means EOF.  If file size is larger, truncate it
 682                  */
 683                 if (num_read < count)
 684                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 685
 686                 SetPageUptodate(page);
 687         }
 688
 689         fuse_invalidate_attr(inode); /* atime changed */
 690  out:
 691         unlock_page(page);
 692         return err;
 693 }
 694
 695 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 696 {
 697         int i;
 698         size_t count = req->misc.read.in.size;
 699         size_t num_read = req->out.args[0].size;
 700         struct address_space *mapping = NULL;
 701
 702         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 703                 mapping = req->pages[i]->mapping;
 704
 705         if (mapping) {
 706                 struct inode *inode = mapping->host;
 707
 708                 /*
 709                  * Short read means EOF. If file size is larger, truncate it
 710                  */
 711                 if (!req->out.h.error && num_read < count) {
 712                         loff_t pos;
 713
 714                         pos = page_offset(req->pages[0]) + num_read;
 715                         fuse_read_update_size(inode, pos,
 716                                               req->misc.read.attr_ver);
 717                 }
 718                 fuse_invalidate_attr(inode); /* atime changed */
 719         }
 720
 721         for (i = 0; i < req->num_pages; i++) {
 722                 struct page *page = req->pages[i];
 723                 if (!req->out.h.error)
 724                         SetPageUptodate(page);
 725                 else
 726                         SetPageError(page);
 727                 unlock_page(page);
 728                 page_cache_release(page);
 729         }
 730         if (req->ff)
 731                 fuse_file_put(req->ff, false);
 732 }
 733
 734 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 735 {
 736         struct fuse_file *ff = file->private_data;
 737         struct fuse_conn *fc = ff->fc;
 738         loff_t pos = page_offset(req->pages[0]);
 739         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 740
 741         req->out.argpages = 1;
 742         req->out.page_zeroing = 1;
 743         req->out.page_replace = 1;
 744         fuse_read_fill(req, file, pos, count, FUSE_READ);
 745         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 746         if (fc->async_read) {
 747                 req->ff = fuse_file_get(ff);
 748                 req->end = fuse_readpages_end;
 749                 fuse_request_send_background(fc, req);
 750         } else {
 751                 fuse_request_send(fc, req);
 752                 fuse_readpages_end(fc, req);
 753                 fuse_put_request(fc, req);
 754         }
 755 }
 756
 757 struct fuse_fill_data {
 758         struct fuse_req *req;
 759         struct file *file;
 760         struct inode *inode;
 761         unsigned nr_pages;
 762 };
 763
 764 static int fuse_readpages_fill(void *_data, struct page *page)
 765 {
 766         struct fuse_fill_data *data = _data;
 767         struct fuse_req *req = data->req;
 768         struct inode *inode = data->inode;
 769         struct fuse_conn *fc = get_fuse_conn(inode);
 770
 771         fuse_wait_on_page_writeback(inode, page->index);
 772
 773         if (req->num_pages &&
 774             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 775              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 776              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 777                 int nr_alloc = min_t(unsigned, data->nr_pages,
 778                                      FUSE_MAX_PAGES_PER_REQ);
 779                 fuse_send_readpages(req, data->file);
 780                 if (fc->async_read)
 781                         req = fuse_get_req_for_background(fc, nr_alloc);
 782                 else
 783                         req = fuse_get_req(fc, nr_alloc);
 784
 785                 data->req = req;
 786                 if (IS_ERR(req)) {
 787                         unlock_page(page);
 788                         return PTR_ERR(req);
 789                 }
 790         }
 791
 792         if (WARN_ON(req->num_pages >= req->max_pages)) {
 793                 fuse_put_request(fc, req);
 794                 return -EIO;
 795         }
 796
 797         page_cache_get(page);
 798         req->pages[req->num_pages] = page;
 799         req->page_descs[req->num_pages].length = PAGE_SIZE;
 800         req->num_pages++;
 801         data->nr_pages--;
 802         return 0;
 803 }
 804
 805 static int fuse_readpages(struct file *file, struct address_space *mapping,
 806                           struct list_head *pages, unsigned nr_pages)
 807 {
 808         struct inode *inode = mapping->host;
 809         struct fuse_conn *fc = get_fuse_conn(inode);
 810         struct fuse_fill_data data;
 811         int err;
 812         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 813
 814         err = -EIO;
 815         if (is_bad_inode(inode))
 816                 goto out;
 817
 818         data.file = file;
 819         data.inode = inode;
 820         if (fc->async_read)
 821                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 822         else
 823                 data.req = fuse_get_req(fc, nr_alloc);
 824         data.nr_pages = nr_pages;
 825         err = PTR_ERR(data.req);
 826         if (IS_ERR(data.req))
 827                 goto out;
 828
 829         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 830         if (!err) {
 831                 if (data.req->num_pages)
 832                         fuse_send_readpages(data.req, file);
 833                 else
 834                         fuse_put_request(fc, data.req);
 835         }
 836 out:
 837         return err;
 838 }
 839
 840 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 841                                   unsigned long nr_segs, loff_t pos)
 842 {
 843         struct inode *inode = iocb->ki_filp->f_mapping->host;
 844         struct fuse_conn *fc = get_fuse_conn(inode);
 845
 846         /*
 847          * In auto invalidate mode, always update attributes on read.
 848          * Otherwise, only update if we attempt to read past EOF (to ensure
 849          * i_size is up to date).
 850          */
 851         if (fc->auto_inval_data ||
 852             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 853                 int err;
 854                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 855                 if (err)
 856                         return err;
 857         }
 858
 859         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 860 }
 861
 862 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 863                             loff_t pos, size_t count)
 864 {
 865         struct fuse_write_in *inarg = &req->misc.write.in;
 866         struct fuse_write_out *outarg = &req->misc.write.out;
 867
 868         inarg->fh = ff->fh;
 869         inarg->offset = pos;
 870         inarg->size = count;
 871         req->in.h.opcode = FUSE_WRITE;
 872         req->in.h.nodeid = ff->nodeid;
 873         req->in.numargs = 2;
 874         if (ff->fc->minor < 9)
 875                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 876         else
 877                 req->in.args[0].size = sizeof(struct fuse_write_in);
 878         req->in.args[0].value = inarg;
 879         req->in.args[1].size = count;
 880         req->out.numargs = 1;
 881         req->out.args[0].size = sizeof(struct fuse_write_out);
 882         req->out.args[0].value = outarg;
 883 }
 884
 885 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 886                               loff_t pos, size_t count, fl_owner_t owner)
 887 {
 888         struct file *file = io->file;
 889         struct fuse_file *ff = file->private_data;
 890         struct fuse_conn *fc = ff->fc;
 891         struct fuse_write_in *inarg = &req->misc.write.in;
 892
 893         fuse_write_fill(req, ff, pos, count);
 894         inarg->flags = file->f_flags;
 895         if (owner != NULL) {
 896                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 897                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 898         }
 899
 900         if (io->async)
 901                 return fuse_async_req_send(fc, req, count, io);
 902
 903         fuse_request_send(fc, req);
 904         return req->misc.write.out.size;
 905 }
 906
 907 void fuse_write_update_size(struct inode *inode, loff_t pos)
 908 {
 909         struct fuse_conn *fc = get_fuse_conn(inode);
 910         struct fuse_inode *fi = get_fuse_inode(inode);
 911
 912         spin_lock(&fc->lock);
 913         fi->attr_version = ++fc->attr_version;
 914         if (pos > inode->i_size)
 915                 i_size_write(inode, pos);
 916         spin_unlock(&fc->lock);
 917 }
 918
 919 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 920                                     struct inode *inode, loff_t pos,
 921                                     size_t count)
 922 {
 923         size_t res;
 924         unsigned offset;
 925         unsigned i;
 926         struct fuse_io_priv io = { .async = 0, .file = file };
 927
 928         for (i = 0; i < req->num_pages; i++)
 929                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 930
 931         res = fuse_send_write(req, &io, pos, count, NULL);
 932
 933         offset = req->page_descs[0].offset;
 934         count = res;
 935         for (i = 0; i < req->num_pages; i++) {
 936                 struct page *page = req->pages[i];
 937
 938                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 939                         SetPageUptodate(page);
 940
 941                 if (count > PAGE_CACHE_SIZE - offset)
 942                         count -= PAGE_CACHE_SIZE - offset;
 943                 else
 944                         count = 0;
 945                 offset = 0;
 946
 947                 unlock_page(page);
 948                 page_cache_release(page);
 949         }
 950
 951         return res;
 952 }
 953
 954 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 955                                struct address_space *mapping,
 956                                struct iov_iter *ii, loff_t pos)
 957 {
 958         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 959         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 960         size_t count = 0;
 961         int err;
 962
 963         req->in.argpages = 1;
 964         req->page_descs[0].offset = offset;
 965
 966         do {
 967                 size_t tmp;
 968                 struct page *page;
 969                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 970                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 971                                      iov_iter_count(ii));
 972
 973                 bytes = min_t(size_t, bytes, fc->max_write - count);
 974
 975  again:
 976                 err = -EFAULT;
 977                 if (iov_iter_fault_in_readable(ii, bytes))
 978                         break;
 979
 980                 err = -ENOMEM;
 981                 page = grab_cache_page_write_begin(mapping, index, 0);
 982                 if (!page)
 983                         break;
 984
 985                 if (mapping_writably_mapped(mapping))
 986                         flush_dcache_page(page);
 987
 988                 pagefault_disable();
 989                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 990                 pagefault_enable();
 991                 flush_dcache_page(page);
 992
 993                 mark_page_accessed(page);
 994
 995                 if (!tmp) {
 996                         unlock_page(page);
 997                         page_cache_release(page);
 998                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 999                         goto again;
1000                 }
1001
1002                 err = 0;
1003                 req->pages[req->num_pages] = page;
1004                 req->page_descs[req->num_pages].length = tmp;
1005                 req->num_pages++;
1006
1007                 iov_iter_advance(ii, tmp);
1008                 count += tmp;
1009                 pos += tmp;
1010                 offset += tmp;
1011                 if (offset == PAGE_CACHE_SIZE)
1012                         offset = 0;
1013
1014                 if (!fc->big_writes)
1015                         break;
1016         } while (iov_iter_count(ii) && count < fc->max_write &&
1017                  req->num_pages < req->max_pages && offset == 0);
1018
1019         return count > 0 ? count : err;
1020 }
1021
1022 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1023 {
1024         return min_t(unsigned,
1025                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1026                      (pos >> PAGE_CACHE_SHIFT) + 1,
1027                      FUSE_MAX_PAGES_PER_REQ);
1028 }
1029
1030 static ssize_t fuse_perform_write(struct file *file,
1031                                   struct address_space *mapping,
1032                                   struct iov_iter *ii, loff_t pos)
1033 {
1034         struct inode *inode = mapping->host;
1035         struct fuse_conn *fc = get_fuse_conn(inode);
1036         struct fuse_inode *fi = get_fuse_inode(inode);
1037         int err = 0;
1038         ssize_t res = 0;
1039
1040         if (is_bad_inode(inode))
1041                 return -EIO;
1042
1043         if (inode->i_size < pos + iov_iter_count(ii))
1044                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1045
1046         do {
1047                 struct fuse_req *req;
1048                 ssize_t count;
1049                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1050
1051                 req = fuse_get_req(fc, nr_pages);
1052                 if (IS_ERR(req)) {
1053                         err = PTR_ERR(req);
1054                         break;
1055                 }
1056
1057                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1058                 if (count <= 0) {
1059                         err = count;
1060                 } else {
1061                         size_t num_written;
1062
1063                         num_written = fuse_send_write_pages(req, file, inode,
1064                                                             pos, count);
1065                         err = req->out.h.error;
1066                         if (!err) {
1067                                 res += num_written;
1068                                 pos += num_written;
1069
1070                                 /* break out of the loop on short write */
1071                                 if (num_written != count)
1072                                         err = -EIO;
1073                         }
1074                 }
1075                 fuse_put_request(fc, req);
1076         } while (!err && iov_iter_count(ii));
1077
1078         if (res > 0)
1079                 fuse_write_update_size(inode, pos);
1080
1081         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1082         fuse_invalidate_attr(inode);
1083
1084         return res > 0 ? res : err;
1085 }
1086
1087 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1088                                    unsigned long nr_segs, loff_t pos)
1089 {
1090         struct file *file = iocb->ki_filp;
1091         struct address_space *mapping = file->f_mapping;
1092         size_t count = 0;
1093         size_t ocount = 0;
1094         ssize_t written = 0;
1095         ssize_t written_buffered = 0;
1096         struct inode *inode = mapping->host;
1097         ssize_t err;
1098         struct iov_iter i;
1099         loff_t endbyte = 0;
1100
1101         WARN_ON(iocb->ki_pos != pos);
1102
1103         ocount = 0;
1104         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1105         if (err)
1106                 return err;
1107
1108         count = ocount;
1109         mutex_lock(&inode->i_mutex);
1110
1111         /* We can write back this queue in page reclaim */
1112         current->backing_dev_info = mapping->backing_dev_info;
1113
1114         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1115         if (err)
1116                 goto out;
1117
1118         if (count == 0)
1119                 goto out;
1120
1121         err = file_remove_suid(file);
1122         if (err)
1123                 goto out;
1124
1125         err = file_update_time(file);
1126         if (err)
1127                 goto out;
1128
1129         if (file->f_flags & O_DIRECT) {
1130                 written = generic_file_direct_write(iocb, iov, &nr_segs,
1131                                                     pos, &iocb->ki_pos,
1132                                                     count, ocount);
1133                 if (written < 0 || written == count)
1134                         goto out;
1135
1136                 pos += written;
1137                 count -= written;
1138
1139                 iov_iter_init(&i, iov, nr_segs, count, written);
1140                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1141                 if (written_buffered < 0) {
1142                         err = written_buffered;
1143                         goto out;
1144                 }
1145                 endbyte = pos + written_buffered - 1;
1146
1147                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1148                                                    endbyte);
1149                 if (err)
1150                         goto out;
1151
1152                 invalidate_mapping_pages(file->f_mapping,
1153                                          pos >> PAGE_CACHE_SHIFT,
1154                                          endbyte >> PAGE_CACHE_SHIFT);
1155
1156                 written += written_buffered;
1157                 iocb->ki_pos = pos + written_buffered;
1158         } else {
1159                 iov_iter_init(&i, iov, nr_segs, count, 0);
1160                 written = fuse_perform_write(file, mapping, &i, pos);
1161                 if (written >= 0)
1162                         iocb->ki_pos = pos + written;
1163         }
1164 out:
1165         current->backing_dev_info = NULL;
1166         mutex_unlock(&inode->i_mutex);
1167
1168         return written ? written : err;
1169 }
1170
1171 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1172                 unsigned index, unsigned nr_pages)
1173 {
1174         int i;
1175
1176         for (i = index; i < index + nr_pages; i++)
1177                 req->page_descs[i].length = PAGE_SIZE -
1178                         req->page_descs[i].offset;
1179 }
1180
1181 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1182 {
1183         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1184 }
1185
1186 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1187                                         size_t max_size)
1188 {
1189         return min(iov_iter_single_seg_count(ii), max_size);
1190 }
1191
1192 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1193                                size_t *nbytesp, int write)
1194 {
1195         size_t nbytes = 0;  /* # bytes already packed in req */
1196
1197         /* Special case for kernel I/O: can copy directly into the buffer */
1198         if (segment_eq(get_fs(), KERNEL_DS)) {
1199                 unsigned long user_addr = fuse_get_user_addr(ii);
1200                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1201
1202                 if (write)
1203                         req->in.args[1].value = (void *) user_addr;
1204                 else
1205                         req->out.args[0].value = (void *) user_addr;
1206
1207                 iov_iter_advance(ii, frag_size);
1208                 *nbytesp = frag_size;
1209                 return 0;
1210         }
1211
1212         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1213                 unsigned npages;
1214                 unsigned long user_addr = fuse_get_user_addr(ii);
1215                 unsigned offset = user_addr & ~PAGE_MASK;
1216                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1217                 int ret;
1218
1219                 unsigned n = req->max_pages - req->num_pages;
1220                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1221
1222                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1223                 npages = clamp(npages, 1U, n);
1224
1225                 ret = get_user_pages_fast(user_addr, npages, !write,
1226                                           &req->pages[req->num_pages]);
1227                 if (ret < 0)
1228                         return ret;
1229
1230                 npages = ret;
1231                 frag_size = min_t(size_t, frag_size,
1232                                   (npages << PAGE_SHIFT) - offset);
1233                 iov_iter_advance(ii, frag_size);
1234
1235                 req->page_descs[req->num_pages].offset = offset;
1236                 fuse_page_descs_length_init(req, req->num_pages, npages);
1237
1238                 req->num_pages += npages;
1239                 req->page_descs[req->num_pages - 1].length -=
1240                         (npages << PAGE_SHIFT) - offset - frag_size;
1241
1242                 nbytes += frag_size;
1243         }
1244
1245         if (write)
1246                 req->in.argpages = 1;
1247         else
1248                 req->out.argpages = 1;
1249
1250         *nbytesp = nbytes;
1251
1252         return 0;
1253 }
1254
1255 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1256 {
1257         struct iov_iter ii = *ii_p;
1258         int npages = 0;
1259
1260         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1261                 unsigned long user_addr = fuse_get_user_addr(&ii);
1262                 unsigned offset = user_addr & ~PAGE_MASK;
1263                 size_t frag_size = iov_iter_single_seg_count(&ii);
1264
1265                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1266                 iov_iter_advance(&ii, frag_size);
1267         }
1268
1269         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1270 }
1271
1272 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1273                        unsigned long nr_segs, size_t count, loff_t *ppos,
1274                        int write)
1275 {
1276         struct file *file = io->file;
1277         struct fuse_file *ff = file->private_data;
1278         struct fuse_conn *fc = ff->fc;
1279         size_t nmax = write ? fc->max_write : fc->max_read;
1280         loff_t pos = *ppos;
1281         ssize_t res = 0;
1282         struct fuse_req *req;
1283         struct iov_iter ii;
1284
1285         iov_iter_init(&ii, iov, nr_segs, count, 0);
1286
1287         if (io->async)
1288                 req = fuse_get_req_for_background(fc, fuse_iter_npages(&ii));
1289         else
1290                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1291         if (IS_ERR(req))
1292                 return PTR_ERR(req);
1293
1294         while (count) {
1295                 size_t nres;
1296                 fl_owner_t owner = current->files;
1297                 size_t nbytes = min(count, nmax);
1298                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1299                 if (err) {
1300                         res = err;
1301                         break;
1302                 }
1303
1304                 if (write)
1305                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1306                 else
1307                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1308
1309                 if (!io->async)
1310                         fuse_release_user_pages(req, !write);
1311                 if (req->out.h.error) {
1312                         if (!res)
1313                                 res = req->out.h.error;
1314                         break;
1315                 } else if (nres > nbytes) {
1316                         res = -EIO;
1317                         break;
1318                 }
1319                 count -= nres;
1320                 res += nres;
1321                 pos += nres;
1322                 if (nres != nbytes)
1323                         break;
1324                 if (count) {
1325                         fuse_put_request(fc, req);
1326                         if (io->async)
1327                                 req = fuse_get_req_for_background(fc,
1328                                         fuse_iter_npages(&ii));
1329                         else
1330                                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1331                         if (IS_ERR(req))
1332                                 break;
1333                 }
1334         }
1335         if (!IS_ERR(req))
1336                 fuse_put_request(fc, req);
1337         if (res > 0)
1338                 *ppos = pos;
1339
1340         return res;
1341 }
1342 EXPORT_SYMBOL_GPL(fuse_direct_io);
1343
1344 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1345                                   const struct iovec *iov,
1346                                   unsigned long nr_segs, loff_t *ppos,
1347                                   size_t count)
1348 {
1349         ssize_t res;
1350         struct file *file = io->file;
1351         struct inode *inode = file_inode(file);
1352
1353         if (is_bad_inode(inode))
1354                 return -EIO;
1355
1356         res = fuse_direct_io(io, iov, nr_segs, count, ppos, 0);
1357
1358         fuse_invalidate_attr(inode);
1359
1360         return res;
1361 }
1362
1363 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1364                                      size_t count, loff_t *ppos)
1365 {
1366         struct fuse_io_priv io = { .async = 0, .file = file };
1367         struct iovec iov = { .iov_base = buf, .iov_len = count };
1368         return __fuse_direct_read(&io, &iov, 1, ppos, count);
1369 }
1370
1371 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1372                                    const struct iovec *iov,
1373                                    unsigned long nr_segs, loff_t *ppos)
1374 {
1375         struct file *file = io->file;
1376         struct inode *inode = file_inode(file);
1377         size_t count = iov_length(iov, nr_segs);
1378         ssize_t res;
1379
1380         res = generic_write_checks(file, ppos, &count, 0);
1381         if (!res)
1382                 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 1);
1383
1384         fuse_invalidate_attr(inode);
1385
1386         return res;
1387 }
1388
1389 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1390                                  size_t count, loff_t *ppos)
1391 {
1392         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1393         struct inode *inode = file_inode(file);
1394         ssize_t res;
1395         struct fuse_io_priv io = { .async = 0, .file = file };
1396
1397         if (is_bad_inode(inode))
1398                 return -EIO;
1399
1400         /* Don't allow parallel writes to the same file */
1401         mutex_lock(&inode->i_mutex);
1402         res = __fuse_direct_write(&io, &iov, 1, ppos);
1403         if (res > 0)
1404                 fuse_write_update_size(inode, *ppos);
1405         mutex_unlock(&inode->i_mutex);
1406
1407         return res;
1408 }
1409
1410 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1411 {
1412         __free_page(req->pages[0]);
1413         fuse_file_put(req->ff, false);
1414 }
1415
1416 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1417 {
1418         struct inode *inode = req->inode;
1419         struct fuse_inode *fi = get_fuse_inode(inode);
1420         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1421
1422         list_del(&req->writepages_entry);
1423         dec_bdi_stat(bdi, BDI_WRITEBACK);
1424         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1425         bdi_writeout_inc(bdi);
1426         wake_up(&fi->page_waitq);
1427 }
1428
1429 /* Called under fc->lock, may release and reacquire it */
1430 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1431 __releases(fc->lock)
1432 __acquires(fc->lock)
1433 {
1434         struct fuse_inode *fi = get_fuse_inode(req->inode);
1435         loff_t size = i_size_read(req->inode);
1436         struct fuse_write_in *inarg = &req->misc.write.in;
1437
1438         if (!fc->connected)
1439                 goto out_free;
1440
1441         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1442                 inarg->size = PAGE_CACHE_SIZE;
1443         } else if (inarg->offset < size) {
1444                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1445         } else {
1446                 /* Got truncated off completely */
1447                 goto out_free;
1448         }
1449
1450         req->in.args[1].size = inarg->size;
1451         fi->writectr++;
1452         fuse_request_send_background_locked(fc, req);
1453         return;
1454
1455  out_free:
1456         fuse_writepage_finish(fc, req);
1457         spin_unlock(&fc->lock);
1458         fuse_writepage_free(fc, req);
1459         fuse_put_request(fc, req);
1460         spin_lock(&fc->lock);
1461 }
1462
1463 /*
1464  * If fi->writectr is positive (no truncate or fsync going on) send
1465  * all queued writepage requests.
1466  *
1467  * Called with fc->lock
1468  */
1469 void fuse_flush_writepages(struct inode *inode)
1470 __releases(fc->lock)
1471 __acquires(fc->lock)
1472 {
1473         struct fuse_conn *fc = get_fuse_conn(inode);
1474         struct fuse_inode *fi = get_fuse_inode(inode);
1475         struct fuse_req *req;
1476
1477         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1478                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1479                 list_del_init(&req->list);
1480                 fuse_send_writepage(fc, req);
1481         }
1482 }
1483
1484 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1485 {
1486         struct inode *inode = req->inode;
1487         struct fuse_inode *fi = get_fuse_inode(inode);
1488
1489         mapping_set_error(inode->i_mapping, req->out.h.error);
1490         spin_lock(&fc->lock);
1491         fi->writectr--;
1492         fuse_writepage_finish(fc, req);
1493         spin_unlock(&fc->lock);
1494         fuse_writepage_free(fc, req);
1495 }
1496
1497 static int fuse_writepage_locked(struct page *page)
1498 {
1499         struct address_space *mapping = page->mapping;
1500         struct inode *inode = mapping->host;
1501         struct fuse_conn *fc = get_fuse_conn(inode);
1502         struct fuse_inode *fi = get_fuse_inode(inode);
1503         struct fuse_req *req;
1504         struct fuse_file *ff;
1505         struct page *tmp_page;
1506
1507         set_page_writeback(page);
1508
1509         req = fuse_request_alloc_nofs(1);
1510         if (!req)
1511                 goto err;
1512
1513         req->background = 1; /* writeback always goes to bg_queue */
1514         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1515         if (!tmp_page)
1516                 goto err_free;
1517
1518         spin_lock(&fc->lock);
1519         BUG_ON(list_empty(&fi->write_files));
1520         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1521         req->ff = fuse_file_get(ff);
1522         spin_unlock(&fc->lock);
1523
1524         fuse_write_fill(req, ff, page_offset(page), 0);
1525
1526         copy_highpage(tmp_page, page);
1527         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1528         req->in.argpages = 1;
1529         req->num_pages = 1;
1530         req->pages[0] = tmp_page;
1531         req->page_descs[0].offset = 0;
1532         req->page_descs[0].length = PAGE_SIZE;
1533         req->end = fuse_writepage_end;
1534         req->inode = inode;
1535
1536         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1537         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1538
1539         spin_lock(&fc->lock);
1540         list_add(&req->writepages_entry, &fi->writepages);
1541         list_add_tail(&req->list, &fi->queued_writes);
1542         fuse_flush_writepages(inode);
1543         spin_unlock(&fc->lock);
1544
1545         end_page_writeback(page);
1546
1547         return 0;
1548
1549 err_free:
1550         fuse_request_free(req);
1551 err:
1552         end_page_writeback(page);
1553         return -ENOMEM;
1554 }
1555
1556 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1557 {
1558         int err;
1559
1560         err = fuse_writepage_locked(page);
1561         unlock_page(page);
1562
1563         return err;
1564 }
1565
1566 static int fuse_launder_page(struct page *page)
1567 {
1568         int err = 0;
1569         if (clear_page_dirty_for_io(page)) {
1570                 struct inode *inode = page->mapping->host;
1571                 err = fuse_writepage_locked(page);
1572                 if (!err)
1573                         fuse_wait_on_page_writeback(inode, page->index);
1574         }
1575         return err;
1576 }
1577
1578 /*
1579  * Write back dirty pages now, because there may not be any suitable
1580  * open files later
1581  */
1582 static void fuse_vma_close(struct vm_area_struct *vma)
1583 {
1584         filemap_write_and_wait(vma->vm_file->f_mapping);
1585 }
1586
1587 /*
1588  * Wait for writeback against this page to complete before allowing it
1589  * to be marked dirty again, and hence written back again, possibly
1590  * before the previous writepage completed.
1591  *
1592  * Block here, instead of in ->writepage(), so that the userspace fs
1593  * can only block processes actually operating on the filesystem.
1594  *
1595  * Otherwise unprivileged userspace fs would be able to block
1596  * unrelated:
1597  *
1598  * - page migration
1599  * - sync(2)
1600  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1601  */
1602 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1603 {
1604         struct page *page = vmf->page;
1605         /*
1606          * Don't use page->mapping as it may become NULL from a
1607          * concurrent truncate.
1608          */
1609         struct inode *inode = vma->vm_file->f_mapping->host;
1610
1611         fuse_wait_on_page_writeback(inode, page->index);
1612         return 0;
1613 }
1614
1615 static const struct vm_operations_struct fuse_file_vm_ops = {
1616         .close          = fuse_vma_close,
1617         .fault          = filemap_fault,
1618         .page_mkwrite   = fuse_page_mkwrite,
1619         .remap_pages    = generic_file_remap_pages,
1620 };
1621
1622 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1623 {
1624         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1625                 struct inode *inode = file_inode(file);
1626                 struct fuse_conn *fc = get_fuse_conn(inode);
1627                 struct fuse_inode *fi = get_fuse_inode(inode);
1628                 struct fuse_file *ff = file->private_data;
1629                 /*
1630                  * file may be written through mmap, so chain it onto the
1631                  * inodes's write_file list
1632                  */
1633                 spin_lock(&fc->lock);
1634                 if (list_empty(&ff->write_entry))
1635                         list_add(&ff->write_entry, &fi->write_files);
1636                 spin_unlock(&fc->lock);
1637         }
1638         file_accessed(file);
1639         vma->vm_ops = &fuse_file_vm_ops;
1640         return 0;
1641 }
1642
1643 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1644 {
1645         /* Can't provide the coherency needed for MAP_SHARED */
1646         if (vma->vm_flags & VM_MAYSHARE)
1647                 return -ENODEV;
1648
1649         invalidate_inode_pages2(file->f_mapping);
1650
1651         return generic_file_mmap(file, vma);
1652 }
1653
1654 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1655                                   struct file_lock *fl)
1656 {
1657         switch (ffl->type) {
1658         case F_UNLCK:
1659                 break;
1660
1661         case F_RDLCK:
1662         case F_WRLCK:
1663                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1664                     ffl->end < ffl->start)
1665                         return -EIO;
1666
1667                 fl->fl_start = ffl->start;
1668                 fl->fl_end = ffl->end;
1669                 fl->fl_pid = ffl->pid;
1670                 break;
1671
1672         default:
1673                 return -EIO;
1674         }
1675         fl->fl_type = ffl->type;
1676         return 0;
1677 }
1678
1679 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1680                          const struct file_lock *fl, int opcode, pid_t pid,
1681                          int flock)
1682 {
1683         struct inode *inode = file_inode(file);
1684         struct fuse_conn *fc = get_fuse_conn(inode);
1685         struct fuse_file *ff = file->private_data;
1686         struct fuse_lk_in *arg = &req->misc.lk_in;
1687
1688         arg->fh = ff->fh;
1689         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1690         arg->lk.start = fl->fl_start;
1691         arg->lk.end = fl->fl_end;
1692         arg->lk.type = fl->fl_type;
1693         arg->lk.pid = pid;
1694         if (flock)
1695                 arg->lk_flags |= FUSE_LK_FLOCK;
1696         req->in.h.opcode = opcode;
1697         req->in.h.nodeid = get_node_id(inode);
1698         req->in.numargs = 1;
1699         req->in.args[0].size = sizeof(*arg);
1700         req->in.args[0].value = arg;
1701 }
1702
1703 static int fuse_getlk(struct file *file, struct file_lock *fl)
1704 {
1705         struct inode *inode = file_inode(file);
1706         struct fuse_conn *fc = get_fuse_conn(inode);
1707         struct fuse_req *req;
1708         struct fuse_lk_out outarg;
1709         int err;
1710
1711         req = fuse_get_req_nopages(fc);
1712         if (IS_ERR(req))
1713                 return PTR_ERR(req);
1714
1715         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1716         req->out.numargs = 1;
1717         req->out.args[0].size = sizeof(outarg);
1718         req->out.args[0].value = &outarg;
1719         fuse_request_send(fc, req);
1720         err = req->out.h.error;
1721         fuse_put_request(fc, req);
1722         if (!err)
1723                 err = convert_fuse_file_lock(&outarg.lk, fl);
1724
1725         return err;
1726 }
1727
1728 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1729 {
1730         struct inode *inode = file_inode(file);
1731         struct fuse_conn *fc = get_fuse_conn(inode);
1732         struct fuse_req *req;
1733         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1734         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1735         int err;
1736
1737         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1738                 /* NLM needs asynchronous locks, which we don't support yet */
1739                 return -ENOLCK;
1740         }
1741
1742         /* Unlock on close is handled by the flush method */
1743         if (fl->fl_flags & FL_CLOSE)
1744                 return 0;
1745
1746         req = fuse_get_req_nopages(fc);
1747         if (IS_ERR(req))
1748                 return PTR_ERR(req);
1749
1750         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1751         fuse_request_send(fc, req);
1752         err = req->out.h.error;
1753         /* locking is restartable */
1754         if (err == -EINTR)
1755                 err = -ERESTARTSYS;
1756         fuse_put_request(fc, req);
1757         return err;
1758 }
1759
1760 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1761 {
1762         struct inode *inode = file_inode(file);
1763         struct fuse_conn *fc = get_fuse_conn(inode);
1764         int err;
1765
1766         if (cmd == F_CANCELLK) {
1767                 err = 0;
1768         } else if (cmd == F_GETLK) {
1769                 if (fc->no_lock) {
1770                         posix_test_lock(file, fl);
1771                         err = 0;
1772                 } else
1773                         err = fuse_getlk(file, fl);
1774         } else {
1775                 if (fc->no_lock)
1776                         err = posix_lock_file(file, fl, NULL);
1777                 else
1778                         err = fuse_setlk(file, fl, 0);
1779         }
1780         return err;
1781 }
1782
1783 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1784 {
1785         struct inode *inode = file_inode(file);
1786         struct fuse_conn *fc = get_fuse_conn(inode);
1787         int err;
1788
1789         if (fc->no_flock) {
1790                 err = flock_lock_file_wait(file, fl);
1791         } else {
1792                 struct fuse_file *ff = file->private_data;
1793
1794                 /* emulate flock with POSIX locks */
1795                 fl->fl_owner = (fl_owner_t) file;
1796                 ff->flock = true;
1797                 err = fuse_setlk(file, fl, 1);
1798         }
1799
1800         return err;
1801 }
1802
1803 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1804 {
1805         struct inode *inode = mapping->host;
1806         struct fuse_conn *fc = get_fuse_conn(inode);
1807         struct fuse_req *req;
1808         struct fuse_bmap_in inarg;
1809         struct fuse_bmap_out outarg;
1810         int err;
1811
1812         if (!inode->i_sb->s_bdev || fc->no_bmap)
1813                 return 0;
1814
1815         req = fuse_get_req_nopages(fc);
1816         if (IS_ERR(req))
1817                 return 0;
1818
1819         memset(&inarg, 0, sizeof(inarg));
1820         inarg.block = block;
1821         inarg.blocksize = inode->i_sb->s_blocksize;
1822         req->in.h.opcode = FUSE_BMAP;
1823         req->in.h.nodeid = get_node_id(inode);
1824         req->in.numargs = 1;
1825         req->in.args[0].size = sizeof(inarg);
1826         req->in.args[0].value = &inarg;
1827         req->out.numargs = 1;
1828         req->out.args[0].size = sizeof(outarg);
1829         req->out.args[0].value = &outarg;
1830         fuse_request_send(fc, req);
1831         err = req->out.h.error;
1832         fuse_put_request(fc, req);
1833         if (err == -ENOSYS)
1834                 fc->no_bmap = 1;
1835
1836         return err ? 0 : outarg.block;
1837 }
1838
1839 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1840 {
1841         loff_t retval;
1842         struct inode *inode = file_inode(file);
1843
1844         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1845         if (whence == SEEK_CUR || whence == SEEK_SET)
1846                 return generic_file_llseek(file, offset, whence);
1847
1848         mutex_lock(&inode->i_mutex);
1849         retval = fuse_update_attributes(inode, NULL, file, NULL);
1850         if (!retval)
1851                 retval = generic_file_llseek(file, offset, whence);
1852         mutex_unlock(&inode->i_mutex);
1853
1854         return retval;
1855 }
1856
1857 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1858                         unsigned int nr_segs, size_t bytes, bool to_user)
1859 {
1860         struct iov_iter ii;
1861         int page_idx = 0;
1862
1863         if (!bytes)
1864                 return 0;
1865
1866         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1867
1868         while (iov_iter_count(&ii)) {
1869                 struct page *page = pages[page_idx++];
1870                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1871                 void *kaddr;
1872
1873                 kaddr = kmap(page);
1874
1875                 while (todo) {
1876                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1877                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1878                         size_t copy = min(todo, iov_len);
1879                         size_t left;
1880
1881                         if (!to_user)
1882                                 left = copy_from_user(kaddr, uaddr, copy);
1883                         else
1884                                 left = copy_to_user(uaddr, kaddr, copy);
1885
1886                         if (unlikely(left))
1887                                 return -EFAULT;
1888
1889                         iov_iter_advance(&ii, copy);
1890                         todo -= copy;
1891                         kaddr += copy;
1892                 }
1893
1894                 kunmap(page);
1895         }
1896
1897         return 0;
1898 }
1899
1900 /*
1901  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1902  * ABI was defined to be 'struct iovec' which is different on 32bit
1903  * and 64bit.  Fortunately we can determine which structure the server
1904  * used from the size of the reply.
1905  */
1906 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1907                                      size_t transferred, unsigned count,
1908                                      bool is_compat)
1909 {
1910 #ifdef CONFIG_COMPAT
1911         if (count * sizeof(struct compat_iovec) == transferred) {
1912                 struct compat_iovec *ciov = src;
1913                 unsigned i;
1914
1915                 /*
1916                  * With this interface a 32bit server cannot support
1917                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1918                  * requests
1919                  */
1920                 if (!is_compat)
1921                         return -EINVAL;
1922
1923                 for (i = 0; i < count; i++) {
1924                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1925                         dst[i].iov_len = ciov[i].iov_len;
1926                 }
1927                 return 0;
1928         }
1929 #endif
1930
1931         if (count * sizeof(struct iovec) != transferred)
1932                 return -EIO;
1933
1934         memcpy(dst, src, transferred);
1935         return 0;
1936 }
1937
1938 /* Make sure iov_length() won't overflow */
1939 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1940 {
1941         size_t n;
1942         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1943
1944         for (n = 0; n < count; n++, iov++) {
1945                 if (iov->iov_len > (size_t) max)
1946                         return -ENOMEM;
1947                 max -= iov->iov_len;
1948         }
1949         return 0;
1950 }
1951
1952 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1953                                  void *src, size_t transferred, unsigned count,
1954                                  bool is_compat)
1955 {
1956         unsigned i;
1957         struct fuse_ioctl_iovec *fiov = src;
1958
1959         if (fc->minor < 16) {
1960                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1961                                                  count, is_compat);
1962         }
1963
1964         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1965                 return -EIO;
1966
1967         for (i = 0; i < count; i++) {
1968                 /* Did the server supply an inappropriate value? */
1969                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1970                     fiov[i].len != (unsigned long) fiov[i].len)
1971                         return -EIO;
1972
1973                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1974                 dst[i].iov_len = (size_t) fiov[i].len;
1975
1976 #ifdef CONFIG_COMPAT
1977                 if (is_compat &&
1978                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1979                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1980                         return -EIO;
1981 #endif
1982         }
1983
1984         return 0;
1985 }
1986
1987
1988 /*
1989  * For ioctls, there is no generic way to determine how much memory
1990  * needs to be read and/or written.  Furthermore, ioctls are allowed
1991  * to dereference the passed pointer, so the parameter requires deep
1992  * copying but FUSE has no idea whatsoever about what to copy in or
1993  * out.
1994  *
1995  * This is solved by allowing FUSE server to retry ioctl with
1996  * necessary in/out iovecs.  Let's assume the ioctl implementation
1997  * needs to read in the following structure.
1998  *
1999  * struct a {
2000  *      char    *buf;
2001  *      size_t  buflen;
2002  * }
2003  *
2004  * On the first callout to FUSE server, inarg->in_size and
2005  * inarg->out_size will be NULL; then, the server completes the ioctl
2006  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2007  * the actual iov array to
2008  *
2009  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2010  *
2011  * which tells FUSE to copy in the requested area and retry the ioctl.
2012  * On the second round, the server has access to the structure and
2013  * from that it can tell what to look for next, so on the invocation,
2014  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2015  *
2016  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2017  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2018  *
2019  * FUSE will copy both struct a and the pointed buffer from the
2020  * process doing the ioctl and retry ioctl with both struct a and the
2021  * buffer.
2022  *
2023  * This time, FUSE server has everything it needs and completes ioctl
2024  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2025  *
2026  * Copying data out works the same way.
2027  *
2028  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2029  * automatically initializes in and out iovs by decoding @cmd with
2030  * _IOC_* macros and the server is not allowed to request RETRY.  This
2031  * limits ioctl data transfers to well-formed ioctls and is the forced
2032  * behavior for all FUSE servers.
2033  */
2034 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2035                    unsigned int flags)
2036 {
2037         struct fuse_file *ff = file->private_data;
2038         struct fuse_conn *fc = ff->fc;
2039         struct fuse_ioctl_in inarg = {
2040                 .fh = ff->fh,
2041                 .cmd = cmd,
2042                 .arg = arg,
2043                 .flags = flags
2044         };
2045         struct fuse_ioctl_out outarg;
2046         struct fuse_req *req = NULL;
2047         struct page **pages = NULL;
2048         struct iovec *iov_page = NULL;
2049         struct iovec *in_iov = NULL, *out_iov = NULL;
2050         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2051         size_t in_size, out_size, transferred;
2052         int err;
2053
2054 #if BITS_PER_LONG == 32
2055         inarg.flags |= FUSE_IOCTL_32BIT;
2056 #else
2057         if (flags & FUSE_IOCTL_COMPAT)
2058                 inarg.flags |= FUSE_IOCTL_32BIT;
2059 #endif
2060
2061         /* assume all the iovs returned by client always fits in a page */
2062         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2063
2064         err = -ENOMEM;
2065         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2066         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2067         if (!pages || !iov_page)
2068                 goto out;
2069
2070         /*
2071          * If restricted, initialize IO parameters as encoded in @cmd.
2072          * RETRY from server is not allowed.
2073          */
2074         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2075                 struct iovec *iov = iov_page;
2076
2077                 iov->iov_base = (void __user *)arg;
2078                 iov->iov_len = _IOC_SIZE(cmd);
2079
2080                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2081                         in_iov = iov;
2082                         in_iovs = 1;
2083                 }
2084
2085                 if (_IOC_DIR(cmd) & _IOC_READ) {
2086                         out_iov = iov;
2087                         out_iovs = 1;
2088                 }
2089         }
2090
2091  retry:
2092         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2093         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2094
2095         /*
2096          * Out data can be used either for actual out data or iovs,
2097          * make sure there always is at least one page.
2098          */
2099         out_size = max_t(size_t, out_size, PAGE_SIZE);
2100         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2101
2102         /* make sure there are enough buffer pages and init request with them */
2103         err = -ENOMEM;
2104         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2105                 goto out;
2106         while (num_pages < max_pages) {
2107                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2108                 if (!pages[num_pages])
2109                         goto out;
2110                 num_pages++;
2111         }
2112
2113         req = fuse_get_req(fc, num_pages);
2114         if (IS_ERR(req)) {
2115                 err = PTR_ERR(req);
2116                 req = NULL;
2117                 goto out;
2118         }
2119         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2120         req->num_pages = num_pages;
2121         fuse_page_descs_length_init(req, 0, req->num_pages);
2122
2123         /* okay, let's send it to the client */
2124         req->in.h.opcode = FUSE_IOCTL;
2125         req->in.h.nodeid = ff->nodeid;
2126         req->in.numargs = 1;
2127         req->in.args[0].size = sizeof(inarg);
2128         req->in.args[0].value = &inarg;
2129         if (in_size) {
2130                 req->in.numargs++;
2131                 req->in.args[1].size = in_size;
2132                 req->in.argpages = 1;
2133
2134                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2135                                            false);
2136                 if (err)
2137                         goto out;
2138         }
2139
2140         req->out.numargs = 2;
2141         req->out.args[0].size = sizeof(outarg);
2142         req->out.args[0].value = &outarg;
2143         req->out.args[1].size = out_size;
2144         req->out.argpages = 1;
2145         req->out.argvar = 1;
2146
2147         fuse_request_send(fc, req);
2148         err = req->out.h.error;
2149         transferred = req->out.args[1].size;
2150         fuse_put_request(fc, req);
2151         req = NULL;
2152         if (err)
2153                 goto out;
2154
2155         /* did it ask for retry? */
2156         if (outarg.flags & FUSE_IOCTL_RETRY) {
2157                 void *vaddr;
2158
2159                 /* no retry if in restricted mode */
2160                 err = -EIO;
2161                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2162                         goto out;
2163
2164                 in_iovs = outarg.in_iovs;
2165                 out_iovs = outarg.out_iovs;
2166
2167                 /*
2168                  * Make sure things are in boundary, separate checks
2169                  * are to protect against overflow.
2170                  */
2171                 err = -ENOMEM;
2172                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2173                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2174                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2175                         goto out;
2176
2177                 vaddr = kmap_atomic(pages[0]);
2178                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2179                                             transferred, in_iovs + out_iovs,
2180                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2181                 kunmap_atomic(vaddr);
2182                 if (err)
2183                         goto out;
2184
2185                 in_iov = iov_page;
2186                 out_iov = in_iov + in_iovs;
2187
2188                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2189                 if (err)
2190                         goto out;
2191
2192                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2193                 if (err)
2194                         goto out;
2195
2196                 goto retry;
2197         }
2198
2199         err = -EIO;
2200         if (transferred > inarg.out_size)
2201                 goto out;
2202
2203         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2204  out:
2205         if (req)
2206                 fuse_put_request(fc, req);
2207         free_page((unsigned long) iov_page);
2208         while (num_pages)
2209                 __free_page(pages[--num_pages]);
2210         kfree(pages);
2211
2212         return err ? err : outarg.result;
2213 }
2214 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2215
2216 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2217                        unsigned long arg, unsigned int flags)
2218 {
2219         struct inode *inode = file_inode(file);
2220         struct fuse_conn *fc = get_fuse_conn(inode);
2221
2222         if (!fuse_allow_current_process(fc))
2223                 return -EACCES;
2224
2225         if (is_bad_inode(inode))
2226                 return -EIO;
2227
2228         return fuse_do_ioctl(file, cmd, arg, flags);
2229 }
2230
2231 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2232                             unsigned long arg)
2233 {
2234         return fuse_ioctl_common(file, cmd, arg, 0);
2235 }
2236
2237 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2238                                    unsigned long arg)
2239 {
2240         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2241 }
2242
2243 /*
2244  * All files which have been polled are linked to RB tree
2245  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2246  * find the matching one.
2247  */
2248 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2249                                               struct rb_node **parent_out)
2250 {
2251         struct rb_node **link = &fc->polled_files.rb_node;
2252         struct rb_node *last = NULL;
2253
2254         while (*link) {
2255                 struct fuse_file *ff;
2256
2257                 last = *link;
2258                 ff = rb_entry(last, struct fuse_file, polled_node);
2259
2260                 if (kh < ff->kh)
2261                         link = &last->rb_left;
2262                 else if (kh > ff->kh)
2263                         link = &last->rb_right;
2264                 else
2265                         return link;
2266         }
2267
2268         if (parent_out)
2269                 *parent_out = last;
2270         return link;
2271 }
2272
2273 /*
2274  * The file is about to be polled.  Make sure it's on the polled_files
2275  * RB tree.  Note that files once added to the polled_files tree are
2276  * not removed before the file is released.  This is because a file
2277  * polled once is likely to be polled again.
2278  */
2279 static void fuse_register_polled_file(struct fuse_conn *fc,
2280                                       struct fuse_file *ff)
2281 {
2282         spin_lock(&fc->lock);
2283         if (RB_EMPTY_NODE(&ff->polled_node)) {
2284                 struct rb_node **link, *parent;
2285
2286                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2287                 BUG_ON(*link);
2288                 rb_link_node(&ff->polled_node, parent, link);
2289                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2290         }
2291         spin_unlock(&fc->lock);
2292 }
2293
2294 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2295 {
2296         struct fuse_file *ff = file->private_data;
2297         struct fuse_conn *fc = ff->fc;
2298         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2299         struct fuse_poll_out outarg;
2300         struct fuse_req *req;
2301         int err;
2302
2303         if (fc->no_poll)
2304                 return DEFAULT_POLLMASK;
2305
2306         poll_wait(file, &ff->poll_wait, wait);
2307         inarg.events = (__u32)poll_requested_events(wait);
2308
2309         /*
2310          * Ask for notification iff there's someone waiting for it.
2311          * The client may ignore the flag and always notify.
2312          */
2313         if (waitqueue_active(&ff->poll_wait)) {
2314                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2315                 fuse_register_polled_file(fc, ff);
2316         }
2317
2318         req = fuse_get_req_nopages(fc);
2319         if (IS_ERR(req))
2320                 return POLLERR;
2321
2322         req->in.h.opcode = FUSE_POLL;
2323         req->in.h.nodeid = ff->nodeid;
2324         req->in.numargs = 1;
2325         req->in.args[0].size = sizeof(inarg);
2326         req->in.args[0].value = &inarg;
2327         req->out.numargs = 1;
2328         req->out.args[0].size = sizeof(outarg);
2329         req->out.args[0].value = &outarg;
2330         fuse_request_send(fc, req);
2331         err = req->out.h.error;
2332         fuse_put_request(fc, req);
2333
2334         if (!err)
2335                 return outarg.revents;
2336         if (err == -ENOSYS) {
2337                 fc->no_poll = 1;
2338                 return DEFAULT_POLLMASK;
2339         }
2340         return POLLERR;
2341 }
2342 EXPORT_SYMBOL_GPL(fuse_file_poll);
2343
2344 /*
2345  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2346  * wakes up the poll waiters.
2347  */
2348 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2349                             struct fuse_notify_poll_wakeup_out *outarg)
2350 {
2351         u64 kh = outarg->kh;
2352         struct rb_node **link;
2353
2354         spin_lock(&fc->lock);
2355
2356         link = fuse_find_polled_node(fc, kh, NULL);
2357         if (*link) {
2358                 struct fuse_file *ff;
2359
2360                 ff = rb_entry(*link, struct fuse_file, polled_node);
2361                 wake_up_interruptible_sync(&ff->poll_wait);
2362         }
2363
2364         spin_unlock(&fc->lock);
2365         return 0;
2366 }
2367
2368 static void fuse_do_truncate(struct file *file)
2369 {
2370         struct inode *inode = file->f_mapping->host;
2371         struct iattr attr;
2372
2373         attr.ia_valid = ATTR_SIZE;
2374         attr.ia_size = i_size_read(inode);
2375
2376         attr.ia_file = file;
2377         attr.ia_valid |= ATTR_FILE;
2378
2379         fuse_do_setattr(inode, &attr, file);
2380 }
2381
2382 static inline loff_t fuse_round_up(loff_t off)
2383 {
2384         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2385 }
2386
2387 static ssize_t
2388 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2389                         loff_t offset, unsigned long nr_segs)
2390 {
2391         ssize_t ret = 0;
2392         struct file *file = iocb->ki_filp;
2393         struct fuse_file *ff = file->private_data;
2394         bool async_dio = ff->fc->async_dio;
2395         loff_t pos = 0;
2396         struct inode *inode;
2397         loff_t i_size;
2398         size_t count = iov_length(iov, nr_segs);
2399         struct fuse_io_priv *io;
2400
2401         pos = offset;
2402         inode = file->f_mapping->host;
2403         i_size = i_size_read(inode);
2404
2405         /* optimization for short read */
2406         if (async_dio && rw != WRITE && offset + count > i_size) {
2407                 if (offset >= i_size)
2408                         return 0;
2409                 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2410         }
2411
2412         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2413         if (!io)
2414                 return -ENOMEM;
2415         spin_lock_init(&io->lock);
2416         io->reqs = 1;
2417         io->bytes = -1;
2418         io->size = 0;
2419         io->offset = offset;
2420         io->write = (rw == WRITE);
2421         io->err = 0;
2422         io->file = file;
2423         /*
2424          * By default, we want to optimize all I/Os with async request
2425          * submission to the client filesystem if supported.
2426          */
2427         io->async = async_dio;
2428         io->iocb = iocb;
2429
2430         /*
2431          * We cannot asynchronously extend the size of a file. We have no method
2432          * to wait on real async I/O requests, so we must submit this request
2433          * synchronously.
2434          */
2435         if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2436                 io->async = false;
2437
2438         if (rw == WRITE)
2439                 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2440         else
2441                 ret = __fuse_direct_read(io, iov, nr_segs, &pos, count);
2442
2443         if (io->async) {
2444                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2445
2446                 /* we have a non-extending, async request, so return */
2447                 if (!is_sync_kiocb(iocb))
2448                         return -EIOCBQUEUED;
2449
2450                 ret = wait_on_sync_kiocb(iocb);
2451         } else {
2452                 kfree(io);
2453         }
2454
2455         if (rw == WRITE) {
2456                 if (ret > 0)
2457                         fuse_write_update_size(inode, pos);
2458                 else if (ret < 0 && offset + count > i_size)
2459                         fuse_do_truncate(file);
2460         }
2461
2462         return ret;
2463 }
2464
2465 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2466                                 loff_t length)
2467 {
2468         struct fuse_file *ff = file->private_data;
2469         struct inode *inode = file->f_inode;
2470         struct fuse_inode *fi = get_fuse_inode(inode);
2471         struct fuse_conn *fc = ff->fc;
2472         struct fuse_req *req;
2473         struct fuse_fallocate_in inarg = {
2474                 .fh = ff->fh,
2475                 .offset = offset,
2476                 .length = length,
2477                 .mode = mode
2478         };
2479         int err;
2480         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2481                            (mode & FALLOC_FL_PUNCH_HOLE);
2482
2483         if (fc->no_fallocate)
2484                 return -EOPNOTSUPP;
2485
2486         if (lock_inode) {
2487                 mutex_lock(&inode->i_mutex);
2488                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2489                         loff_t endbyte = offset + length - 1;
2490                         err = filemap_write_and_wait_range(inode->i_mapping,
2491                                                            offset, endbyte);
2492                         if (err)
2493                                 goto out;
2494
2495                         fuse_sync_writes(inode);
2496                 }
2497         }
2498
2499         if (!(mode & FALLOC_FL_KEEP_SIZE))
2500                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2501
2502         req = fuse_get_req_nopages(fc);
2503         if (IS_ERR(req)) {
2504                 err = PTR_ERR(req);
2505                 goto out;
2506         }
2507
2508         req->in.h.opcode = FUSE_FALLOCATE;
2509         req->in.h.nodeid = ff->nodeid;
2510         req->in.numargs = 1;
2511         req->in.args[0].size = sizeof(inarg);
2512         req->in.args[0].value = &inarg;
2513         fuse_request_send(fc, req);
2514         err = req->out.h.error;
2515         if (err == -ENOSYS) {
2516                 fc->no_fallocate = 1;
2517                 err = -EOPNOTSUPP;
2518         }
2519         fuse_put_request(fc, req);
2520
2521         if (err)
2522                 goto out;
2523
2524         /* we could have extended the file */
2525         if (!(mode & FALLOC_FL_KEEP_SIZE))
2526                 fuse_write_update_size(inode, offset + length);
2527
2528         if (mode & FALLOC_FL_PUNCH_HOLE)
2529                 truncate_pagecache_range(inode, offset, offset + length - 1);
2530
2531         fuse_invalidate_attr(inode);
2532
2533 out:
2534         if (!(mode & FALLOC_FL_KEEP_SIZE))
2535                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2536
2537         if (lock_inode)
2538                 mutex_unlock(&inode->i_mutex);
2539
2540         return err;
2541 }
2542
2543 static const struct file_operations fuse_file_operations = {
2544         .llseek         = fuse_file_llseek,
2545         .read           = do_sync_read,
2546         .aio_read       = fuse_file_aio_read,
2547         .write          = do_sync_write,
2548         .aio_write      = fuse_file_aio_write,
2549         .mmap           = fuse_file_mmap,
2550         .open           = fuse_open,
2551         .flush          = fuse_flush,
2552         .release        = fuse_release,
2553         .fsync          = fuse_fsync,
2554         .lock           = fuse_file_lock,
2555         .flock          = fuse_file_flock,
2556         .splice_read    = generic_file_splice_read,
2557         .unlocked_ioctl = fuse_file_ioctl,
2558         .compat_ioctl   = fuse_file_compat_ioctl,
2559         .poll           = fuse_file_poll,
2560         .fallocate      = fuse_file_fallocate,
2561 };
2562
2563 static const struct file_operations fuse_direct_io_file_operations = {
2564         .llseek         = fuse_file_llseek,
2565         .read           = fuse_direct_read,
2566         .write          = fuse_direct_write,
2567         .mmap           = fuse_direct_mmap,
2568         .open           = fuse_open,
2569         .flush          = fuse_flush,
2570         .release        = fuse_release,
2571         .fsync          = fuse_fsync,
2572         .lock           = fuse_file_lock,
2573         .flock          = fuse_file_flock,
2574         .unlocked_ioctl = fuse_file_ioctl,
2575         .compat_ioctl   = fuse_file_compat_ioctl,
2576         .poll           = fuse_file_poll,
2577         .fallocate      = fuse_file_fallocate,
2578         /* no splice_read */
2579 };
2580
2581 static const struct address_space_operations fuse_file_aops  = {
2582         .readpage       = fuse_readpage,
2583         .writepage      = fuse_writepage,
2584         .launder_page   = fuse_launder_page,
2585         .readpages      = fuse_readpages,
2586         .set_page_dirty = __set_page_dirty_nobuffers,
2587         .bmap           = fuse_bmap,
2588         .direct_IO      = fuse_direct_IO,
2589 };
2590
2591 void fuse_init_file_inode(struct inode *inode)
2592 {
2593         inode->i_fop = &fuse_file_operations;
2594         inode->i_data.a_ops = &fuse_file_aops;
2595 }