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 (ff->fc->no_open) {
 131                         /*
 132                          * Drop the release request when client does not
 133                          * implement 'open'
 134                          */
 135                         req->background = 0;
 136                         path_put(&req->misc.release.path);
 137                         fuse_put_request(ff->fc, req);
 138                 } else if (sync) {
 139                         req->background = 0;
 140                         fuse_request_send(ff->fc, req);
 141                         path_put(&req->misc.release.path);
 142                         fuse_put_request(ff->fc, req);
 143                 } else {
 144                         req->end = fuse_release_end;
 145                         req->background = 1;
 146                         fuse_request_send_background(ff->fc, req);
 147                 }
 148                 kfree(ff);
 149         }
 150 }
 151
 152 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 153                  bool isdir)
 154 {
 155         struct fuse_file *ff;
 156         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 157
 158         ff = fuse_file_alloc(fc);
 159         if (!ff)
 160                 return -ENOMEM;
 161
 162         ff->fh = 0;
 163         ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 164         if (!fc->no_open || isdir) {
 165                 struct fuse_open_out outarg;
 166                 int err;
 167
 168                 err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 169                 if (!err) {
 170                         ff->fh = outarg.fh;
 171                         ff->open_flags = outarg.open_flags;
 172
 173                 } else if (err != -ENOSYS || isdir) {
 174                         fuse_file_free(ff);
 175                         return err;
 176                 } else {
 177                         fc->no_open = 1;
 178                 }
 179         }
 180
 181         if (isdir)
 182                 ff->open_flags &= ~FOPEN_DIRECT_IO;
 183
 184         ff->nodeid = nodeid;
 185         file->private_data = fuse_file_get(ff);
 186
 187         return 0;
 188 }
 189 EXPORT_SYMBOL_GPL(fuse_do_open);
 190
 191 void fuse_finish_open(struct inode *inode, struct file *file)
 192 {
 193         struct fuse_file *ff = file->private_data;
 194         struct fuse_conn *fc = get_fuse_conn(inode);
 195
 196         if (ff->open_flags & FOPEN_DIRECT_IO)
 197                 file->f_op = &fuse_direct_io_file_operations;
 198         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 199                 invalidate_inode_pages2(inode->i_mapping);
 200         if (ff->open_flags & FOPEN_NONSEEKABLE)
 201                 nonseekable_open(inode, file);
 202         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 203                 struct fuse_inode *fi = get_fuse_inode(inode);
 204
 205                 spin_lock(&fc->lock);
 206                 fi->attr_version = ++fc->attr_version;
 207                 i_size_write(inode, 0);
 208                 spin_unlock(&fc->lock);
 209                 fuse_invalidate_attr(inode);
 210         }
 211 }
 212
 213 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 214 {
 215         struct fuse_conn *fc = get_fuse_conn(inode);
 216         int err;
 217
 218         err = generic_file_open(inode, file);
 219         if (err)
 220                 return err;
 221
 222         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 223         if (err)
 224                 return err;
 225
 226         fuse_finish_open(inode, file);
 227
 228         return 0;
 229 }
 230
 231 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 232 {
 233         struct fuse_conn *fc = ff->fc;
 234         struct fuse_req *req = ff->reserved_req;
 235         struct fuse_release_in *inarg = &req->misc.release.in;
 236
 237         spin_lock(&fc->lock);
 238         list_del(&ff->write_entry);
 239         if (!RB_EMPTY_NODE(&ff->polled_node))
 240                 rb_erase(&ff->polled_node, &fc->polled_files);
 241         spin_unlock(&fc->lock);
 242
 243         wake_up_interruptible_all(&ff->poll_wait);
 244
 245         inarg->fh = ff->fh;
 246         inarg->flags = flags;
 247         req->in.h.opcode = opcode;
 248         req->in.h.nodeid = ff->nodeid;
 249         req->in.numargs = 1;
 250         req->in.args[0].size = sizeof(struct fuse_release_in);
 251         req->in.args[0].value = inarg;
 252 }
 253
 254 void fuse_release_common(struct file *file, int opcode)
 255 {
 256         struct fuse_file *ff;
 257         struct fuse_req *req;
 258
 259         ff = file->private_data;
 260         if (unlikely(!ff))
 261                 return;
 262
 263         req = ff->reserved_req;
 264         fuse_prepare_release(ff, file->f_flags, opcode);
 265
 266         if (ff->flock) {
 267                 struct fuse_release_in *inarg = &req->misc.release.in;
 268                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 269                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 270                                                        (fl_owner_t) file);
 271         }
 272         /* Hold vfsmount and dentry until release is finished */
 273         path_get(&file->f_path);
 274         req->misc.release.path = file->f_path;
 275
 276         /*
 277          * Normally this will send the RELEASE request, however if
 278          * some asynchronous READ or WRITE requests are outstanding,
 279          * the sending will be delayed.
 280          *
 281          * Make the release synchronous if this is a fuseblk mount,
 282          * synchronous RELEASE is allowed (and desirable) in this case
 283          * because the server can be trusted not to screw up.
 284          */
 285         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 286 }
 287
 288 static int fuse_open(struct inode *inode, struct file *file)
 289 {
 290         return fuse_open_common(inode, file, false);
 291 }
 292
 293 static int fuse_release(struct inode *inode, struct file *file)
 294 {
 295         fuse_release_common(file, FUSE_RELEASE);
 296
 297         /* return value is ignored by VFS */
 298         return 0;
 299 }
 300
 301 void fuse_sync_release(struct fuse_file *ff, int flags)
 302 {
 303         WARN_ON(atomic_read(&ff->count) > 1);
 304         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 305         ff->reserved_req->force = 1;
 306         ff->reserved_req->background = 0;
 307         fuse_request_send(ff->fc, ff->reserved_req);
 308         fuse_put_request(ff->fc, ff->reserved_req);
 309         kfree(ff);
 310 }
 311 EXPORT_SYMBOL_GPL(fuse_sync_release);
 312
 313 /*
 314  * Scramble the ID space with XTEA, so that the value of the files_struct
 315  * pointer is not exposed to userspace.
 316  */
 317 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 318 {
 319         u32 *k = fc->scramble_key;
 320         u64 v = (unsigned long) id;
 321         u32 v0 = v;
 322         u32 v1 = v >> 32;
 323         u32 sum = 0;
 324         int i;
 325
 326         for (i = 0; i < 32; i++) {
 327                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 328                 sum += 0x9E3779B9;
 329                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 330         }
 331
 332         return (u64) v0 + ((u64) v1 << 32);
 333 }
 334
 335 /*
 336  * Check if page is under writeback
 337  *
 338  * This is currently done by walking the list of writepage requests
 339  * for the inode, which can be pretty inefficient.
 340  */
 341 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 342 {
 343         struct fuse_conn *fc = get_fuse_conn(inode);
 344         struct fuse_inode *fi = get_fuse_inode(inode);
 345         struct fuse_req *req;
 346         bool found = false;
 347
 348         spin_lock(&fc->lock);
 349         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 350                 pgoff_t curr_index;
 351
 352                 BUG_ON(req->inode != inode);
 353                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 354                 if (curr_index <= index &&
 355                     index < curr_index + req->num_pages) {
 356                         found = true;
 357                         break;
 358                 }
 359         }
 360         spin_unlock(&fc->lock);
 361
 362         return found;
 363 }
 364
 365 /*
 366  * Wait for page writeback to be completed.
 367  *
 368  * Since fuse doesn't rely on the VM writeback tracking, this has to
 369  * use some other means.
 370  */
 371 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 372 {
 373         struct fuse_inode *fi = get_fuse_inode(inode);
 374
 375         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 376         return 0;
 377 }
 378
 379 static int fuse_flush(struct file *file, fl_owner_t id)
 380 {
 381         struct inode *inode = file_inode(file);
 382         struct fuse_conn *fc = get_fuse_conn(inode);
 383         struct fuse_file *ff = file->private_data;
 384         struct fuse_req *req;
 385         struct fuse_flush_in inarg;
 386         int err;
 387
 388         if (is_bad_inode(inode))
 389                 return -EIO;
 390
 391         if (fc->no_flush)
 392                 return 0;
 393
 394         req = fuse_get_req_nofail_nopages(fc, file);
 395         memset(&inarg, 0, sizeof(inarg));
 396         inarg.fh = ff->fh;
 397         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 398         req->in.h.opcode = FUSE_FLUSH;
 399         req->in.h.nodeid = get_node_id(inode);
 400         req->in.numargs = 1;
 401         req->in.args[0].size = sizeof(inarg);
 402         req->in.args[0].value = &inarg;
 403         req->force = 1;
 404         fuse_request_send(fc, req);
 405         err = req->out.h.error;
 406         fuse_put_request(fc, req);
 407         if (err == -ENOSYS) {
 408                 fc->no_flush = 1;
 409                 err = 0;
 410         }
 411         return err;
 412 }
 413
 414 /*
 415  * Wait for all pending writepages on the inode to finish.
 416  *
 417  * This is currently done by blocking further writes with FUSE_NOWRITE
 418  * and waiting for all sent writes to complete.
 419  *
 420  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 421  * could conflict with truncation.
 422  */
 423 static void fuse_sync_writes(struct inode *inode)
 424 {
 425         fuse_set_nowrite(inode);
 426         fuse_release_nowrite(inode);
 427 }
 428
 429 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 430                       int datasync, int isdir)
 431 {
 432         struct inode *inode = file->f_mapping->host;
 433         struct fuse_conn *fc = get_fuse_conn(inode);
 434         struct fuse_file *ff = file->private_data;
 435         struct fuse_req *req;
 436         struct fuse_fsync_in inarg;
 437         int err;
 438
 439         if (is_bad_inode(inode))
 440                 return -EIO;
 441
 442         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 443         if (err)
 444                 return err;
 445
 446         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 447                 return 0;
 448
 449         mutex_lock(&inode->i_mutex);
 450
 451         /*
 452          * Start writeback against all dirty pages of the inode, then
 453          * wait for all outstanding writes, before sending the FSYNC
 454          * request.
 455          */
 456         err = write_inode_now(inode, 0);
 457         if (err)
 458                 goto out;
 459
 460         fuse_sync_writes(inode);
 461
 462         req = fuse_get_req_nopages(fc);
 463         if (IS_ERR(req)) {
 464                 err = PTR_ERR(req);
 465                 goto out;
 466         }
 467
 468         memset(&inarg, 0, sizeof(inarg));
 469         inarg.fh = ff->fh;
 470         inarg.fsync_flags = datasync ? 1 : 0;
 471         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 472         req->in.h.nodeid = get_node_id(inode);
 473         req->in.numargs = 1;
 474         req->in.args[0].size = sizeof(inarg);
 475         req->in.args[0].value = &inarg;
 476         fuse_request_send(fc, req);
 477         err = req->out.h.error;
 478         fuse_put_request(fc, req);
 479         if (err == -ENOSYS) {
 480                 if (isdir)
 481                         fc->no_fsyncdir = 1;
 482                 else
 483                         fc->no_fsync = 1;
 484                 err = 0;
 485         }
 486 out:
 487         mutex_unlock(&inode->i_mutex);
 488         return err;
 489 }
 490
 491 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 492                       int datasync)
 493 {
 494         return fuse_fsync_common(file, start, end, datasync, 0);
 495 }
 496
 497 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 498                     size_t count, int opcode)
 499 {
 500         struct fuse_read_in *inarg = &req->misc.read.in;
 501         struct fuse_file *ff = file->private_data;
 502
 503         inarg->fh = ff->fh;
 504         inarg->offset = pos;
 505         inarg->size = count;
 506         inarg->flags = file->f_flags;
 507         req->in.h.opcode = opcode;
 508         req->in.h.nodeid = ff->nodeid;
 509         req->in.numargs = 1;
 510         req->in.args[0].size = sizeof(struct fuse_read_in);
 511         req->in.args[0].value = inarg;
 512         req->out.argvar = 1;
 513         req->out.numargs = 1;
 514         req->out.args[0].size = count;
 515 }
 516
 517 static void fuse_release_user_pages(struct fuse_req *req, int write)
 518 {
 519         unsigned i;
 520
 521         for (i = 0; i < req->num_pages; i++) {
 522                 struct page *page = req->pages[i];
 523                 if (write)
 524                         set_page_dirty_lock(page);
 525                 put_page(page);
 526         }
 527 }
 528
 529 /**
 530  * In case of short read, the caller sets 'pos' to the position of
 531  * actual end of fuse request in IO request. Otherwise, if bytes_requested
 532  * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 533  *
 534  * An example:
 535  * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 536  * both submitted asynchronously. The first of them was ACKed by userspace as
 537  * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 538  * second request was ACKed as short, e.g. only 1K was read, resulting in
 539  * pos == 33K.
 540  *
 541  * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 542  * will be equal to the length of the longest contiguous fragment of
 543  * transferred data starting from the beginning of IO request.
 544  */
 545 static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 546 {
 547         int left;
 548
 549         spin_lock(&io->lock);
 550         if (err)
 551                 io->err = io->err ? : err;
 552         else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 553                 io->bytes = pos;
 554
 555         left = --io->reqs;
 556         spin_unlock(&io->lock);
 557
 558         if (!left) {
 559                 long res;
 560
 561                 if (io->err)
 562                         res = io->err;
 563                 else if (io->bytes >= 0 && io->write)
 564                         res = -EIO;
 565                 else {
 566                         res = io->bytes < 0 ? io->size : io->bytes;
 567
 568                         if (!is_sync_kiocb(io->iocb)) {
 569                                 struct inode *inode = file_inode(io->iocb->ki_filp);
 570                                 struct fuse_conn *fc = get_fuse_conn(inode);
 571                                 struct fuse_inode *fi = get_fuse_inode(inode);
 572
 573                                 spin_lock(&fc->lock);
 574                                 fi->attr_version = ++fc->attr_version;
 575                                 spin_unlock(&fc->lock);
 576                         }
 577                 }
 578
 579                 aio_complete(io->iocb, res, 0);
 580                 kfree(io);
 581         }
 582 }
 583
 584 static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 585 {
 586         struct fuse_io_priv *io = req->io;
 587         ssize_t pos = -1;
 588
 589         fuse_release_user_pages(req, !io->write);
 590
 591         if (io->write) {
 592                 if (req->misc.write.in.size != req->misc.write.out.size)
 593                         pos = req->misc.write.in.offset - io->offset +
 594                                 req->misc.write.out.size;
 595         } else {
 596                 if (req->misc.read.in.size != req->out.args[0].size)
 597                         pos = req->misc.read.in.offset - io->offset +
 598                                 req->out.args[0].size;
 599         }
 600
 601         fuse_aio_complete(io, req->out.h.error, pos);
 602 }
 603
 604 static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 605                 size_t num_bytes, struct fuse_io_priv *io)
 606 {
 607         spin_lock(&io->lock);
 608         io->size += num_bytes;
 609         io->reqs++;
 610         spin_unlock(&io->lock);
 611
 612         req->io = io;
 613         req->end = fuse_aio_complete_req;
 614
 615         __fuse_get_request(req);
 616         fuse_request_send_background(fc, req);
 617
 618         return num_bytes;
 619 }
 620
 621 static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 622                              loff_t pos, size_t count, fl_owner_t owner)
 623 {
 624         struct file *file = io->file;
 625         struct fuse_file *ff = file->private_data;
 626         struct fuse_conn *fc = ff->fc;
 627
 628         fuse_read_fill(req, file, pos, count, FUSE_READ);
 629         if (owner != NULL) {
 630                 struct fuse_read_in *inarg = &req->misc.read.in;
 631
 632                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 633                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 634         }
 635
 636         if (io->async)
 637                 return fuse_async_req_send(fc, req, count, io);
 638
 639         fuse_request_send(fc, req);
 640         return req->out.args[0].size;
 641 }
 642
 643 static void fuse_read_update_size(struct inode *inode, loff_t size,
 644                                   u64 attr_ver)
 645 {
 646         struct fuse_conn *fc = get_fuse_conn(inode);
 647         struct fuse_inode *fi = get_fuse_inode(inode);
 648
 649         spin_lock(&fc->lock);
 650         if (attr_ver == fi->attr_version && size < inode->i_size &&
 651             !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 652                 fi->attr_version = ++fc->attr_version;
 653                 i_size_write(inode, size);
 654         }
 655         spin_unlock(&fc->lock);
 656 }
 657
 658 static int fuse_readpage(struct file *file, struct page *page)
 659 {
 660         struct fuse_io_priv io = { .async = 0, .file = file };
 661         struct inode *inode = page->mapping->host;
 662         struct fuse_conn *fc = get_fuse_conn(inode);
 663         struct fuse_req *req;
 664         size_t num_read;
 665         loff_t pos = page_offset(page);
 666         size_t count = PAGE_CACHE_SIZE;
 667         u64 attr_ver;
 668         int err;
 669
 670         err = -EIO;
 671         if (is_bad_inode(inode))
 672                 goto out;
 673
 674         /*
 675          * Page writeback can extend beyond the lifetime of the
 676          * page-cache page, so make sure we read a properly synced
 677          * page.
 678          */
 679         fuse_wait_on_page_writeback(inode, page->index);
 680
 681         req = fuse_get_req(fc, 1);
 682         err = PTR_ERR(req);
 683         if (IS_ERR(req))
 684                 goto out;
 685
 686         attr_ver = fuse_get_attr_version(fc);
 687
 688         req->out.page_zeroing = 1;
 689         req->out.argpages = 1;
 690         req->num_pages = 1;
 691         req->pages[0] = page;
 692         req->page_descs[0].length = count;
 693         num_read = fuse_send_read(req, &io, pos, count, NULL);
 694         err = req->out.h.error;
 695         fuse_put_request(fc, req);
 696
 697         if (!err) {
 698                 /*
 699                  * Short read means EOF.  If file size is larger, truncate it
 700                  */
 701                 if (num_read < count)
 702                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 703
 704                 SetPageUptodate(page);
 705         }
 706
 707         fuse_invalidate_atime(inode);
 708  out:
 709         unlock_page(page);
 710         return err;
 711 }
 712
 713 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 714 {
 715         int i;
 716         size_t count = req->misc.read.in.size;
 717         size_t num_read = req->out.args[0].size;
 718         struct address_space *mapping = NULL;
 719
 720         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 721                 mapping = req->pages[i]->mapping;
 722
 723         if (mapping) {
 724                 struct inode *inode = mapping->host;
 725
 726                 /*
 727                  * Short read means EOF. If file size is larger, truncate it
 728                  */
 729                 if (!req->out.h.error && num_read < count) {
 730                         loff_t pos;
 731
 732                         pos = page_offset(req->pages[0]) + num_read;
 733                         fuse_read_update_size(inode, pos,
 734                                               req->misc.read.attr_ver);
 735                 }
 736                 fuse_invalidate_atime(inode);
 737         }
 738
 739         for (i = 0; i < req->num_pages; i++) {
 740                 struct page *page = req->pages[i];
 741                 if (!req->out.h.error)
 742                         SetPageUptodate(page);
 743                 else
 744                         SetPageError(page);
 745                 unlock_page(page);
 746                 page_cache_release(page);
 747         }
 748         if (req->ff)
 749                 fuse_file_put(req->ff, false);
 750 }
 751
 752 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 753 {
 754         struct fuse_file *ff = file->private_data;
 755         struct fuse_conn *fc = ff->fc;
 756         loff_t pos = page_offset(req->pages[0]);
 757         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 758
 759         req->out.argpages = 1;
 760         req->out.page_zeroing = 1;
 761         req->out.page_replace = 1;
 762         fuse_read_fill(req, file, pos, count, FUSE_READ);
 763         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 764         if (fc->async_read) {
 765                 req->ff = fuse_file_get(ff);
 766                 req->end = fuse_readpages_end;
 767                 fuse_request_send_background(fc, req);
 768         } else {
 769                 fuse_request_send(fc, req);
 770                 fuse_readpages_end(fc, req);
 771                 fuse_put_request(fc, req);
 772         }
 773 }
 774
 775 struct fuse_fill_data {
 776         struct fuse_req *req;
 777         struct file *file;
 778         struct inode *inode;
 779         unsigned nr_pages;
 780 };
 781
 782 static int fuse_readpages_fill(void *_data, struct page *page)
 783 {
 784         struct fuse_fill_data *data = _data;
 785         struct fuse_req *req = data->req;
 786         struct inode *inode = data->inode;
 787         struct fuse_conn *fc = get_fuse_conn(inode);
 788
 789         fuse_wait_on_page_writeback(inode, page->index);
 790
 791         if (req->num_pages &&
 792             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 793              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 794              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 795                 int nr_alloc = min_t(unsigned, data->nr_pages,
 796                                      FUSE_MAX_PAGES_PER_REQ);
 797                 fuse_send_readpages(req, data->file);
 798                 if (fc->async_read)
 799                         req = fuse_get_req_for_background(fc, nr_alloc);
 800                 else
 801                         req = fuse_get_req(fc, nr_alloc);
 802
 803                 data->req = req;
 804                 if (IS_ERR(req)) {
 805                         unlock_page(page);
 806                         return PTR_ERR(req);
 807                 }
 808         }
 809
 810         if (WARN_ON(req->num_pages >= req->max_pages)) {
 811                 fuse_put_request(fc, req);
 812                 return -EIO;
 813         }
 814
 815         page_cache_get(page);
 816         req->pages[req->num_pages] = page;
 817         req->page_descs[req->num_pages].length = PAGE_SIZE;
 818         req->num_pages++;
 819         data->nr_pages--;
 820         return 0;
 821 }
 822
 823 static int fuse_readpages(struct file *file, struct address_space *mapping,
 824                           struct list_head *pages, unsigned nr_pages)
 825 {
 826         struct inode *inode = mapping->host;
 827         struct fuse_conn *fc = get_fuse_conn(inode);
 828         struct fuse_fill_data data;
 829         int err;
 830         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 831
 832         err = -EIO;
 833         if (is_bad_inode(inode))
 834                 goto out;
 835
 836         data.file = file;
 837         data.inode = inode;
 838         if (fc->async_read)
 839                 data.req = fuse_get_req_for_background(fc, nr_alloc);
 840         else
 841                 data.req = fuse_get_req(fc, nr_alloc);
 842         data.nr_pages = nr_pages;
 843         err = PTR_ERR(data.req);
 844         if (IS_ERR(data.req))
 845                 goto out;
 846
 847         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 848         if (!err) {
 849                 if (data.req->num_pages)
 850                         fuse_send_readpages(data.req, file);
 851                 else
 852                         fuse_put_request(fc, data.req);
 853         }
 854 out:
 855         return err;
 856 }
 857
 858 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 859                                   unsigned long nr_segs, loff_t pos)
 860 {
 861         struct inode *inode = iocb->ki_filp->f_mapping->host;
 862         struct fuse_conn *fc = get_fuse_conn(inode);
 863
 864         /*
 865          * In auto invalidate mode, always update attributes on read.
 866          * Otherwise, only update if we attempt to read past EOF (to ensure
 867          * i_size is up to date).
 868          */
 869         if (fc->auto_inval_data ||
 870             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 871                 int err;
 872                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 873                 if (err)
 874                         return err;
 875         }
 876
 877         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 878 }
 879
 880 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 881                             loff_t pos, size_t count)
 882 {
 883         struct fuse_write_in *inarg = &req->misc.write.in;
 884         struct fuse_write_out *outarg = &req->misc.write.out;
 885
 886         inarg->fh = ff->fh;
 887         inarg->offset = pos;
 888         inarg->size = count;
 889         req->in.h.opcode = FUSE_WRITE;
 890         req->in.h.nodeid = ff->nodeid;
 891         req->in.numargs = 2;
 892         if (ff->fc->minor < 9)
 893                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 894         else
 895                 req->in.args[0].size = sizeof(struct fuse_write_in);
 896         req->in.args[0].value = inarg;
 897         req->in.args[1].size = count;
 898         req->out.numargs = 1;
 899         req->out.args[0].size = sizeof(struct fuse_write_out);
 900         req->out.args[0].value = outarg;
 901 }
 902
 903 static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 904                               loff_t pos, size_t count, fl_owner_t owner)
 905 {
 906         struct file *file = io->file;
 907         struct fuse_file *ff = file->private_data;
 908         struct fuse_conn *fc = ff->fc;
 909         struct fuse_write_in *inarg = &req->misc.write.in;
 910
 911         fuse_write_fill(req, ff, pos, count);
 912         inarg->flags = file->f_flags;
 913         if (owner != NULL) {
 914                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 915                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 916         }
 917
 918         if (io->async)
 919                 return fuse_async_req_send(fc, req, count, io);
 920
 921         fuse_request_send(fc, req);
 922         return req->misc.write.out.size;
 923 }
 924
 925 void fuse_write_update_size(struct inode *inode, loff_t pos)
 926 {
 927         struct fuse_conn *fc = get_fuse_conn(inode);
 928         struct fuse_inode *fi = get_fuse_inode(inode);
 929
 930         spin_lock(&fc->lock);
 931         fi->attr_version = ++fc->attr_version;
 932         if (pos > inode->i_size)
 933                 i_size_write(inode, pos);
 934         spin_unlock(&fc->lock);
 935 }
 936
 937 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 938                                     struct inode *inode, loff_t pos,
 939                                     size_t count)
 940 {
 941         size_t res;
 942         unsigned offset;
 943         unsigned i;
 944         struct fuse_io_priv io = { .async = 0, .file = file };
 945
 946         for (i = 0; i < req->num_pages; i++)
 947                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 948
 949         res = fuse_send_write(req, &io, pos, count, NULL);
 950
 951         offset = req->page_descs[0].offset;
 952         count = res;
 953         for (i = 0; i < req->num_pages; i++) {
 954                 struct page *page = req->pages[i];
 955
 956                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 957                         SetPageUptodate(page);
 958
 959                 if (count > PAGE_CACHE_SIZE - offset)
 960                         count -= PAGE_CACHE_SIZE - offset;
 961                 else
 962                         count = 0;
 963                 offset = 0;
 964
 965                 unlock_page(page);
 966                 page_cache_release(page);
 967         }
 968
 969         return res;
 970 }
 971
 972 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 973                                struct address_space *mapping,
 974                                struct iov_iter *ii, loff_t pos)
 975 {
 976         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 977         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 978         size_t count = 0;
 979         int err;
 980
 981         req->in.argpages = 1;
 982         req->page_descs[0].offset = offset;
 983
 984         do {
 985                 size_t tmp;
 986                 struct page *page;
 987                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 988                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 989                                      iov_iter_count(ii));
 990
 991                 bytes = min_t(size_t, bytes, fc->max_write - count);
 992
 993  again:
 994                 err = -EFAULT;
 995                 if (iov_iter_fault_in_readable(ii, bytes))
 996                         break;
 997
 998                 err = -ENOMEM;
 999                 page = grab_cache_page_write_begin(mapping, index, 0);
1000                 if (!page)
1001                         break;
1002
1003                 if (mapping_writably_mapped(mapping))
1004                         flush_dcache_page(page);
1005
1006                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1007                 flush_dcache_page(page);
1008
1009                 mark_page_accessed(page);
1010
1011                 if (!tmp) {
1012                         unlock_page(page);
1013                         page_cache_release(page);
1014                         bytes = min(bytes, iov_iter_single_seg_count(ii));
1015                         goto again;
1016                 }
1017
1018                 err = 0;
1019                 req->pages[req->num_pages] = page;
1020                 req->page_descs[req->num_pages].length = tmp;
1021                 req->num_pages++;
1022
1023                 iov_iter_advance(ii, tmp);
1024                 count += tmp;
1025                 pos += tmp;
1026                 offset += tmp;
1027                 if (offset == PAGE_CACHE_SIZE)
1028                         offset = 0;
1029
1030                 if (!fc->big_writes)
1031                         break;
1032         } while (iov_iter_count(ii) && count < fc->max_write &&
1033                  req->num_pages < req->max_pages && offset == 0);
1034
1035         return count > 0 ? count : err;
1036 }
1037
1038 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1039 {
1040         return min_t(unsigned,
1041                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1042                      (pos >> PAGE_CACHE_SHIFT) + 1,
1043                      FUSE_MAX_PAGES_PER_REQ);
1044 }
1045
1046 static ssize_t fuse_perform_write(struct file *file,
1047                                   struct address_space *mapping,
1048                                   struct iov_iter *ii, loff_t pos)
1049 {
1050         struct inode *inode = mapping->host;
1051         struct fuse_conn *fc = get_fuse_conn(inode);
1052         struct fuse_inode *fi = get_fuse_inode(inode);
1053         int err = 0;
1054         ssize_t res = 0;
1055
1056         if (is_bad_inode(inode))
1057                 return -EIO;
1058
1059         if (inode->i_size < pos + iov_iter_count(ii))
1060                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1061
1062         do {
1063                 struct fuse_req *req;
1064                 ssize_t count;
1065                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1066
1067                 req = fuse_get_req(fc, nr_pages);
1068                 if (IS_ERR(req)) {
1069                         err = PTR_ERR(req);
1070                         break;
1071                 }
1072
1073                 count = fuse_fill_write_pages(req, mapping, ii, pos);
1074                 if (count <= 0) {
1075                         err = count;
1076                 } else {
1077                         size_t num_written;
1078
1079                         num_written = fuse_send_write_pages(req, file, inode,
1080                                                             pos, count);
1081                         err = req->out.h.error;
1082                         if (!err) {
1083                                 res += num_written;
1084                                 pos += num_written;
1085
1086                                 /* break out of the loop on short write */
1087                                 if (num_written != count)
1088                                         err = -EIO;
1089                         }
1090                 }
1091                 fuse_put_request(fc, req);
1092         } while (!err && iov_iter_count(ii));
1093
1094         if (res > 0)
1095                 fuse_write_update_size(inode, pos);
1096
1097         clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1098         fuse_invalidate_attr(inode);
1099
1100         return res > 0 ? res : err;
1101 }
1102
1103 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
1104                                    unsigned long nr_segs, loff_t pos)
1105 {
1106         struct file *file = iocb->ki_filp;
1107         struct address_space *mapping = file->f_mapping;
1108         size_t count = 0;
1109         size_t ocount = 0;
1110         ssize_t written = 0;
1111         ssize_t written_buffered = 0;
1112         struct inode *inode = mapping->host;
1113         ssize_t err;
1114         struct iov_iter i;
1115         loff_t endbyte = 0;
1116
1117         WARN_ON(iocb->ki_pos != pos);
1118
1119         ocount = 0;
1120         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1121         if (err)
1122                 return err;
1123
1124         count = ocount;
1125         mutex_lock(&inode->i_mutex);
1126
1127         /* We can write back this queue in page reclaim */
1128         current->backing_dev_info = mapping->backing_dev_info;
1129
1130         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1131         if (err)
1132                 goto out;
1133
1134         if (count == 0)
1135                 goto out;
1136
1137         err = file_remove_suid(file);
1138         if (err)
1139                 goto out;
1140
1141         err = file_update_time(file);
1142         if (err)
1143                 goto out;
1144
1145         if (file->f_flags & O_DIRECT) {
1146                 written = generic_file_direct_write(iocb, iov, &nr_segs,
1147                                                     pos, &iocb->ki_pos,
1148                                                     count, ocount);
1149                 if (written < 0 || written == count)
1150                         goto out;
1151
1152                 pos += written;
1153                 count -= written;
1154
1155                 iov_iter_init(&i, iov, nr_segs, count, written);
1156                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1157                 if (written_buffered < 0) {
1158                         err = written_buffered;
1159                         goto out;
1160                 }
1161                 endbyte = pos + written_buffered - 1;
1162
1163                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1164                                                    endbyte);
1165                 if (err)
1166                         goto out;
1167
1168                 invalidate_mapping_pages(file->f_mapping,
1169                                          pos >> PAGE_CACHE_SHIFT,
1170                                          endbyte >> PAGE_CACHE_SHIFT);
1171
1172                 written += written_buffered;
1173                 iocb->ki_pos = pos + written_buffered;
1174         } else {
1175                 iov_iter_init(&i, iov, nr_segs, count, 0);
1176                 written = fuse_perform_write(file, mapping, &i, pos);
1177                 if (written >= 0)
1178                         iocb->ki_pos = pos + written;
1179         }
1180 out:
1181         current->backing_dev_info = NULL;
1182         mutex_unlock(&inode->i_mutex);
1183
1184         return written ? written : err;
1185 }
1186
1187 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1188                 unsigned index, unsigned nr_pages)
1189 {
1190         int i;
1191
1192         for (i = index; i < index + nr_pages; i++)
1193                 req->page_descs[i].length = PAGE_SIZE -
1194                         req->page_descs[i].offset;
1195 }
1196
1197 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1198 {
1199         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1200 }
1201
1202 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1203                                         size_t max_size)
1204 {
1205         return min(iov_iter_single_seg_count(ii), max_size);
1206 }
1207
1208 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1209                                size_t *nbytesp, int write)
1210 {
1211         size_t nbytes = 0;  /* # bytes already packed in req */
1212
1213         /* Special case for kernel I/O: can copy directly into the buffer */
1214         if (segment_eq(get_fs(), KERNEL_DS)) {
1215                 unsigned long user_addr = fuse_get_user_addr(ii);
1216                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1217
1218                 if (write)
1219                         req->in.args[1].value = (void *) user_addr;
1220                 else
1221                         req->out.args[0].value = (void *) user_addr;
1222
1223                 iov_iter_advance(ii, frag_size);
1224                 *nbytesp = frag_size;
1225                 return 0;
1226         }
1227
1228         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1229                 unsigned npages;
1230                 unsigned long user_addr = fuse_get_user_addr(ii);
1231                 unsigned offset = user_addr & ~PAGE_MASK;
1232                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1233                 int ret;
1234
1235                 unsigned n = req->max_pages - req->num_pages;
1236                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1237
1238                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1239                 npages = clamp(npages, 1U, n);
1240
1241                 ret = get_user_pages_fast(user_addr, npages, !write,
1242                                           &req->pages[req->num_pages]);
1243                 if (ret < 0)
1244                         return ret;
1245
1246                 npages = ret;
1247                 frag_size = min_t(size_t, frag_size,
1248                                   (npages << PAGE_SHIFT) - offset);
1249                 iov_iter_advance(ii, frag_size);
1250
1251                 req->page_descs[req->num_pages].offset = offset;
1252                 fuse_page_descs_length_init(req, req->num_pages, npages);
1253
1254                 req->num_pages += npages;
1255                 req->page_descs[req->num_pages - 1].length -=
1256                         (npages << PAGE_SHIFT) - offset - frag_size;
1257
1258                 nbytes += frag_size;
1259         }
1260
1261         if (write)
1262                 req->in.argpages = 1;
1263         else
1264                 req->out.argpages = 1;
1265
1266         *nbytesp = nbytes;
1267
1268         return 0;
1269 }
1270
1271 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1272 {
1273         struct iov_iter ii = *ii_p;
1274         int npages = 0;
1275
1276         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1277                 unsigned long user_addr = fuse_get_user_addr(&ii);
1278                 unsigned offset = user_addr & ~PAGE_MASK;
1279                 size_t frag_size = iov_iter_single_seg_count(&ii);
1280
1281                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1282                 iov_iter_advance(&ii, frag_size);
1283         }
1284
1285         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1286 }
1287
1288 ssize_t fuse_direct_io(struct fuse_io_priv *io, const struct iovec *iov,
1289                        unsigned long nr_segs, size_t count, loff_t *ppos,
1290                        int write)
1291 {
1292         struct file *file = io->file;
1293         struct fuse_file *ff = file->private_data;
1294         struct fuse_conn *fc = ff->fc;
1295         size_t nmax = write ? fc->max_write : fc->max_read;
1296         loff_t pos = *ppos;
1297         ssize_t res = 0;
1298         struct fuse_req *req;
1299         struct iov_iter ii;
1300
1301         iov_iter_init(&ii, iov, nr_segs, count, 0);
1302
1303         if (io->async)
1304                 req = fuse_get_req_for_background(fc, fuse_iter_npages(&ii));
1305         else
1306                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1307         if (IS_ERR(req))
1308                 return PTR_ERR(req);
1309
1310         while (count) {
1311                 size_t nres;
1312                 fl_owner_t owner = current->files;
1313                 size_t nbytes = min(count, nmax);
1314                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1315                 if (err) {
1316                         res = err;
1317                         break;
1318                 }
1319
1320                 if (write)
1321                         nres = fuse_send_write(req, io, pos, nbytes, owner);
1322                 else
1323                         nres = fuse_send_read(req, io, pos, nbytes, owner);
1324
1325                 if (!io->async)
1326                         fuse_release_user_pages(req, !write);
1327                 if (req->out.h.error) {
1328                         if (!res)
1329                                 res = req->out.h.error;
1330                         break;
1331                 } else if (nres > nbytes) {
1332                         res = -EIO;
1333                         break;
1334                 }
1335                 count -= nres;
1336                 res += nres;
1337                 pos += nres;
1338                 if (nres != nbytes)
1339                         break;
1340                 if (count) {
1341                         fuse_put_request(fc, req);
1342                         if (io->async)
1343                                 req = fuse_get_req_for_background(fc,
1344                                         fuse_iter_npages(&ii));
1345                         else
1346                                 req = fuse_get_req(fc, fuse_iter_npages(&ii));
1347                         if (IS_ERR(req))
1348                                 break;
1349                 }
1350         }
1351         if (!IS_ERR(req))
1352                 fuse_put_request(fc, req);
1353         if (res > 0)
1354                 *ppos = pos;
1355
1356         return res;
1357 }
1358 EXPORT_SYMBOL_GPL(fuse_direct_io);
1359
1360 static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1361                                   const struct iovec *iov,
1362                                   unsigned long nr_segs, loff_t *ppos,
1363                                   size_t count)
1364 {
1365         ssize_t res;
1366         struct file *file = io->file;
1367         struct inode *inode = file_inode(file);
1368
1369         if (is_bad_inode(inode))
1370                 return -EIO;
1371
1372         res = fuse_direct_io(io, iov, nr_segs, count, ppos, 0);
1373
1374         fuse_invalidate_attr(inode);
1375
1376         return res;
1377 }
1378
1379 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1380                                      size_t count, loff_t *ppos)
1381 {
1382         struct fuse_io_priv io = { .async = 0, .file = file };
1383         struct iovec iov = { .iov_base = buf, .iov_len = count };
1384         return __fuse_direct_read(&io, &iov, 1, ppos, count);
1385 }
1386
1387 static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1388                                    const struct iovec *iov,
1389                                    unsigned long nr_segs, loff_t *ppos)
1390 {
1391         struct file *file = io->file;
1392         struct inode *inode = file_inode(file);
1393         size_t count = iov_length(iov, nr_segs);
1394         ssize_t res;
1395
1396         res = generic_write_checks(file, ppos, &count, 0);
1397         if (!res)
1398                 res = fuse_direct_io(io, iov, nr_segs, count, ppos, 1);
1399
1400         fuse_invalidate_attr(inode);
1401
1402         return res;
1403 }
1404
1405 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1406                                  size_t count, loff_t *ppos)
1407 {
1408         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1409         struct inode *inode = file_inode(file);
1410         ssize_t res;
1411         struct fuse_io_priv io = { .async = 0, .file = file };
1412
1413         if (is_bad_inode(inode))
1414                 return -EIO;
1415
1416         /* Don't allow parallel writes to the same file */
1417         mutex_lock(&inode->i_mutex);
1418         res = __fuse_direct_write(&io, &iov, 1, ppos);
1419         if (res > 0)
1420                 fuse_write_update_size(inode, *ppos);
1421         mutex_unlock(&inode->i_mutex);
1422
1423         return res;
1424 }
1425
1426 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1427 {
1428         int i;
1429
1430         for (i = 0; i < req->num_pages; i++)
1431                 __free_page(req->pages[i]);
1432
1433         if (req->ff)
1434                 fuse_file_put(req->ff, false);
1435 }
1436
1437 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1438 {
1439         struct inode *inode = req->inode;
1440         struct fuse_inode *fi = get_fuse_inode(inode);
1441         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1442         int i;
1443
1444         list_del(&req->writepages_entry);
1445         for (i = 0; i < req->num_pages; i++) {
1446                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1447                 dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1448                 bdi_writeout_inc(bdi);
1449         }
1450         wake_up(&fi->page_waitq);
1451 }
1452
1453 /* Called under fc->lock, may release and reacquire it */
1454 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1455                                 loff_t size)
1456 __releases(fc->lock)
1457 __acquires(fc->lock)
1458 {
1459         struct fuse_inode *fi = get_fuse_inode(req->inode);
1460         struct fuse_write_in *inarg = &req->misc.write.in;
1461         __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1462
1463         if (!fc->connected)
1464                 goto out_free;
1465
1466         if (inarg->offset + data_size <= size) {
1467                 inarg->size = data_size;
1468         } else if (inarg->offset < size) {
1469                 inarg->size = size - inarg->offset;
1470         } else {
1471                 /* Got truncated off completely */
1472                 goto out_free;
1473         }
1474
1475         req->in.args[1].size = inarg->size;
1476         fi->writectr++;
1477         fuse_request_send_background_locked(fc, req);
1478         return;
1479
1480  out_free:
1481         fuse_writepage_finish(fc, req);
1482         spin_unlock(&fc->lock);
1483         fuse_writepage_free(fc, req);
1484         fuse_put_request(fc, req);
1485         spin_lock(&fc->lock);
1486 }
1487
1488 /*
1489  * If fi->writectr is positive (no truncate or fsync going on) send
1490  * all queued writepage requests.
1491  *
1492  * Called with fc->lock
1493  */
1494 void fuse_flush_writepages(struct inode *inode)
1495 __releases(fc->lock)
1496 __acquires(fc->lock)
1497 {
1498         struct fuse_conn *fc = get_fuse_conn(inode);
1499         struct fuse_inode *fi = get_fuse_inode(inode);
1500         size_t crop = i_size_read(inode);
1501         struct fuse_req *req;
1502
1503         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1504                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1505                 list_del_init(&req->list);
1506                 fuse_send_writepage(fc, req, crop);
1507         }
1508 }
1509
1510 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1511 {
1512         struct inode *inode = req->inode;
1513         struct fuse_inode *fi = get_fuse_inode(inode);
1514
1515         mapping_set_error(inode->i_mapping, req->out.h.error);
1516         spin_lock(&fc->lock);
1517         while (req->misc.write.next) {
1518                 struct fuse_conn *fc = get_fuse_conn(inode);
1519                 struct fuse_write_in *inarg = &req->misc.write.in;
1520                 struct fuse_req *next = req->misc.write.next;
1521                 req->misc.write.next = next->misc.write.next;
1522                 next->misc.write.next = NULL;
1523                 next->ff = fuse_file_get(req->ff);
1524                 list_add(&next->writepages_entry, &fi->writepages);
1525
1526                 /*
1527                  * Skip fuse_flush_writepages() to make it easy to crop requests
1528                  * based on primary request size.
1529                  *
1530                  * 1st case (trivial): there are no concurrent activities using
1531                  * fuse_set/release_nowrite.  Then we're on safe side because
1532                  * fuse_flush_writepages() would call fuse_send_writepage()
1533                  * anyway.
1534                  *
1535                  * 2nd case: someone called fuse_set_nowrite and it is waiting
1536                  * now for completion of all in-flight requests.  This happens
1537                  * rarely and no more than once per page, so this should be
1538                  * okay.
1539                  *
1540                  * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1541                  * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1542                  * that fuse_set_nowrite returned implies that all in-flight
1543                  * requests were completed along with all of their secondary
1544                  * requests.  Further primary requests are blocked by negative
1545                  * writectr.  Hence there cannot be any in-flight requests and
1546                  * no invocations of fuse_writepage_end() while we're in
1547                  * fuse_set_nowrite..fuse_release_nowrite section.
1548                  */
1549                 fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1550         }
1551         fi->writectr--;
1552         fuse_writepage_finish(fc, req);
1553         spin_unlock(&fc->lock);
1554         fuse_writepage_free(fc, req);
1555 }
1556
1557 static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1558                                              struct fuse_inode *fi)
1559 {
1560         struct fuse_file *ff = NULL;
1561
1562         spin_lock(&fc->lock);
1563         if (!WARN_ON(list_empty(&fi->write_files))) {
1564                 ff = list_entry(fi->write_files.next, struct fuse_file,
1565                                 write_entry);
1566                 fuse_file_get(ff);
1567         }
1568         spin_unlock(&fc->lock);
1569
1570         return ff;
1571 }
1572
1573 static int fuse_writepage_locked(struct page *page)
1574 {
1575         struct address_space *mapping = page->mapping;
1576         struct inode *inode = mapping->host;
1577         struct fuse_conn *fc = get_fuse_conn(inode);
1578         struct fuse_inode *fi = get_fuse_inode(inode);
1579         struct fuse_req *req;
1580         struct page *tmp_page;
1581         int error = -ENOMEM;
1582
1583         set_page_writeback(page);
1584
1585         req = fuse_request_alloc_nofs(1);
1586         if (!req)
1587                 goto err;
1588
1589         req->background = 1; /* writeback always goes to bg_queue */
1590         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1591         if (!tmp_page)
1592                 goto err_free;
1593
1594         error = -EIO;
1595         req->ff = fuse_write_file_get(fc, fi);
1596         if (!req->ff)
1597                 goto err_free;
1598
1599         fuse_write_fill(req, req->ff, page_offset(page), 0);
1600
1601         copy_highpage(tmp_page, page);
1602         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1603         req->misc.write.next = NULL;
1604         req->in.argpages = 1;
1605         req->num_pages = 1;
1606         req->pages[0] = tmp_page;
1607         req->page_descs[0].offset = 0;
1608         req->page_descs[0].length = PAGE_SIZE;
1609         req->end = fuse_writepage_end;
1610         req->inode = inode;
1611
1612         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1613         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1614
1615         spin_lock(&fc->lock);
1616         list_add(&req->writepages_entry, &fi->writepages);
1617         list_add_tail(&req->list, &fi->queued_writes);
1618         fuse_flush_writepages(inode);
1619         spin_unlock(&fc->lock);
1620
1621         end_page_writeback(page);
1622
1623         return 0;
1624
1625 err_free:
1626         fuse_request_free(req);
1627 err:
1628         end_page_writeback(page);
1629         return error;
1630 }
1631
1632 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1633 {
1634         int err;
1635
1636         if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1637                 /*
1638                  * ->writepages() should be called for sync() and friends.  We
1639                  * should only get here on direct reclaim and then we are
1640                  * allowed to skip a page which is already in flight
1641                  */
1642                 WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1643
1644                 redirty_page_for_writepage(wbc, page);
1645                 return 0;
1646         }
1647
1648         err = fuse_writepage_locked(page);
1649         unlock_page(page);
1650
1651         return err;
1652 }
1653
1654 struct fuse_fill_wb_data {
1655         struct fuse_req *req;
1656         struct fuse_file *ff;
1657         struct inode *inode;
1658         struct page **orig_pages;
1659 };
1660
1661 static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1662 {
1663         struct fuse_req *req = data->req;
1664         struct inode *inode = data->inode;
1665         struct fuse_conn *fc = get_fuse_conn(inode);
1666         struct fuse_inode *fi = get_fuse_inode(inode);
1667         int num_pages = req->num_pages;
1668         int i;
1669
1670         req->ff = fuse_file_get(data->ff);
1671         spin_lock(&fc->lock);
1672         list_add_tail(&req->list, &fi->queued_writes);
1673         fuse_flush_writepages(inode);
1674         spin_unlock(&fc->lock);
1675
1676         for (i = 0; i < num_pages; i++)
1677                 end_page_writeback(data->orig_pages[i]);
1678 }
1679
1680 static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1681                                      struct page *page)
1682 {
1683         struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1684         struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1685         struct fuse_req *tmp;
1686         struct fuse_req *old_req;
1687         bool found = false;
1688         pgoff_t curr_index;
1689
1690         BUG_ON(new_req->num_pages != 0);
1691
1692         spin_lock(&fc->lock);
1693         list_del(&new_req->writepages_entry);
1694         list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1695                 BUG_ON(old_req->inode != new_req->inode);
1696                 curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1697                 if (curr_index <= page->index &&
1698                     page->index < curr_index + old_req->num_pages) {
1699                         found = true;
1700                         break;
1701                 }
1702         }
1703         if (!found) {
1704                 list_add(&new_req->writepages_entry, &fi->writepages);
1705                 goto out_unlock;
1706         }
1707
1708         new_req->num_pages = 1;
1709         for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1710                 BUG_ON(tmp->inode != new_req->inode);
1711                 curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1712                 if (tmp->num_pages == 1 &&
1713                     curr_index == page->index) {
1714                         old_req = tmp;
1715                 }
1716         }
1717
1718         if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1719                                         old_req->state == FUSE_REQ_PENDING)) {
1720                 struct backing_dev_info *bdi = page->mapping->backing_dev_info;
1721
1722                 copy_highpage(old_req->pages[0], page);
1723                 spin_unlock(&fc->lock);
1724
1725                 dec_bdi_stat(bdi, BDI_WRITEBACK);
1726                 dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1727                 bdi_writeout_inc(bdi);
1728                 fuse_writepage_free(fc, new_req);
1729                 fuse_request_free(new_req);
1730                 goto out;
1731         } else {
1732                 new_req->misc.write.next = old_req->misc.write.next;
1733                 old_req->misc.write.next = new_req;
1734         }
1735 out_unlock:
1736         spin_unlock(&fc->lock);
1737 out:
1738         return found;
1739 }
1740
1741 static int fuse_writepages_fill(struct page *page,
1742                 struct writeback_control *wbc, void *_data)
1743 {
1744         struct fuse_fill_wb_data *data = _data;
1745         struct fuse_req *req = data->req;
1746         struct inode *inode = data->inode;
1747         struct fuse_conn *fc = get_fuse_conn(inode);
1748         struct page *tmp_page;
1749         bool is_writeback;
1750         int err;
1751
1752         if (!data->ff) {
1753                 err = -EIO;
1754                 data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1755                 if (!data->ff)
1756                         goto out_unlock;
1757         }
1758
1759         /*
1760          * Being under writeback is unlikely but possible.  For example direct
1761          * read to an mmaped fuse file will set the page dirty twice; once when
1762          * the pages are faulted with get_user_pages(), and then after the read
1763          * completed.
1764          */
1765         is_writeback = fuse_page_is_writeback(inode, page->index);
1766
1767         if (req && req->num_pages &&
1768             (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1769              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1770              data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1771                 fuse_writepages_send(data);
1772                 data->req = NULL;
1773         }
1774         err = -ENOMEM;
1775         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1776         if (!tmp_page)
1777                 goto out_unlock;
1778
1779         /*
1780          * The page must not be redirtied until the writeout is completed
1781          * (i.e. userspace has sent a reply to the write request).  Otherwise
1782          * there could be more than one temporary page instance for each real
1783          * page.
1784          *
1785          * This is ensured by holding the page lock in page_mkwrite() while
1786          * checking fuse_page_is_writeback().  We already hold the page lock
1787          * since clear_page_dirty_for_io() and keep it held until we add the
1788          * request to the fi->writepages list and increment req->num_pages.
1789          * After this fuse_page_is_writeback() will indicate that the page is
1790          * under writeback, so we can release the page lock.
1791          */
1792         if (data->req == NULL) {
1793                 struct fuse_inode *fi = get_fuse_inode(inode);
1794
1795                 err = -ENOMEM;
1796                 req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1797                 if (!req) {
1798                         __free_page(tmp_page);
1799                         goto out_unlock;
1800                 }
1801
1802                 fuse_write_fill(req, data->ff, page_offset(page), 0);
1803                 req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1804                 req->misc.write.next = NULL;
1805                 req->in.argpages = 1;
1806                 req->background = 1;
1807                 req->num_pages = 0;
1808                 req->end = fuse_writepage_end;
1809                 req->inode = inode;
1810
1811                 spin_lock(&fc->lock);
1812                 list_add(&req->writepages_entry, &fi->writepages);
1813                 spin_unlock(&fc->lock);
1814
1815                 data->req = req;
1816         }
1817         set_page_writeback(page);
1818
1819         copy_highpage(tmp_page, page);
1820         req->pages[req->num_pages] = tmp_page;
1821         req->page_descs[req->num_pages].offset = 0;
1822         req->page_descs[req->num_pages].length = PAGE_SIZE;
1823
1824         inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
1825         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1826
1827         err = 0;
1828         if (is_writeback && fuse_writepage_in_flight(req, page)) {
1829                 end_page_writeback(page);
1830                 data->req = NULL;
1831                 goto out_unlock;
1832         }
1833         data->orig_pages[req->num_pages] = page;
1834
1835         /*
1836          * Protected by fc->lock against concurrent access by
1837          * fuse_page_is_writeback().
1838          */
1839         spin_lock(&fc->lock);
1840         req->num_pages++;
1841         spin_unlock(&fc->lock);
1842
1843 out_unlock:
1844         unlock_page(page);
1845
1846         return err;
1847 }
1848
1849 static int fuse_writepages(struct address_space *mapping,
1850                            struct writeback_control *wbc)
1851 {
1852         struct inode *inode = mapping->host;
1853         struct fuse_fill_wb_data data;
1854         int err;
1855
1856         err = -EIO;
1857         if (is_bad_inode(inode))
1858                 goto out;
1859
1860         data.inode = inode;
1861         data.req = NULL;
1862         data.ff = NULL;
1863
1864         err = -ENOMEM;
1865         data.orig_pages = kzalloc(sizeof(struct page *) *
1866                                   FUSE_MAX_PAGES_PER_REQ,
1867                                   GFP_NOFS);
1868         if (!data.orig_pages)
1869                 goto out;
1870
1871         err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1872         if (data.req) {
1873                 /* Ignore errors if we can write at least one page */
1874                 BUG_ON(!data.req->num_pages);
1875                 fuse_writepages_send(&data);
1876                 err = 0;
1877         }
1878         if (data.ff)
1879                 fuse_file_put(data.ff, false);
1880
1881         kfree(data.orig_pages);
1882 out:
1883         return err;
1884 }
1885
1886 static int fuse_launder_page(struct page *page)
1887 {
1888         int err = 0;
1889         if (clear_page_dirty_for_io(page)) {
1890                 struct inode *inode = page->mapping->host;
1891                 err = fuse_writepage_locked(page);
1892                 if (!err)
1893                         fuse_wait_on_page_writeback(inode, page->index);
1894         }
1895         return err;
1896 }
1897
1898 /*
1899  * Write back dirty pages now, because there may not be any suitable
1900  * open files later
1901  */
1902 static void fuse_vma_close(struct vm_area_struct *vma)
1903 {
1904         filemap_write_and_wait(vma->vm_file->f_mapping);
1905 }
1906
1907 /*
1908  * Wait for writeback against this page to complete before allowing it
1909  * to be marked dirty again, and hence written back again, possibly
1910  * before the previous writepage completed.
1911  *
1912  * Block here, instead of in ->writepage(), so that the userspace fs
1913  * can only block processes actually operating on the filesystem.
1914  *
1915  * Otherwise unprivileged userspace fs would be able to block
1916  * unrelated:
1917  *
1918  * - page migration
1919  * - sync(2)
1920  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1921  */
1922 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1923 {
1924         struct page *page = vmf->page;
1925         struct inode *inode = file_inode(vma->vm_file);
1926
1927         file_update_time(vma->vm_file);
1928         lock_page(page);
1929         if (page->mapping != inode->i_mapping) {
1930                 unlock_page(page);
1931                 return VM_FAULT_NOPAGE;
1932         }
1933
1934         fuse_wait_on_page_writeback(inode, page->index);
1935         return VM_FAULT_LOCKED;
1936 }
1937
1938 static const struct vm_operations_struct fuse_file_vm_ops = {
1939         .close          = fuse_vma_close,
1940         .fault          = filemap_fault,
1941         .page_mkwrite   = fuse_page_mkwrite,
1942         .remap_pages    = generic_file_remap_pages,
1943 };
1944
1945 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1946 {
1947         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1948                 struct inode *inode = file_inode(file);
1949                 struct fuse_conn *fc = get_fuse_conn(inode);
1950                 struct fuse_inode *fi = get_fuse_inode(inode);
1951                 struct fuse_file *ff = file->private_data;
1952                 /*
1953                  * file may be written through mmap, so chain it onto the
1954                  * inodes's write_file list
1955                  */
1956                 spin_lock(&fc->lock);
1957                 if (list_empty(&ff->write_entry))
1958                         list_add(&ff->write_entry, &fi->write_files);
1959                 spin_unlock(&fc->lock);
1960         }
1961         file_accessed(file);
1962         vma->vm_ops = &fuse_file_vm_ops;
1963         return 0;
1964 }
1965
1966 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1967 {
1968         /* Can't provide the coherency needed for MAP_SHARED */
1969         if (vma->vm_flags & VM_MAYSHARE)
1970                 return -ENODEV;
1971
1972         invalidate_inode_pages2(file->f_mapping);
1973
1974         return generic_file_mmap(file, vma);
1975 }
1976
1977 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1978                                   struct file_lock *fl)
1979 {
1980         switch (ffl->type) {
1981         case F_UNLCK:
1982                 break;
1983
1984         case F_RDLCK:
1985         case F_WRLCK:
1986                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1987                     ffl->end < ffl->start)
1988                         return -EIO;
1989
1990                 fl->fl_start = ffl->start;
1991                 fl->fl_end = ffl->end;
1992                 fl->fl_pid = ffl->pid;
1993                 break;
1994
1995         default:
1996                 return -EIO;
1997         }
1998         fl->fl_type = ffl->type;
1999         return 0;
2000 }
2001
2002 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
2003                          const struct file_lock *fl, int opcode, pid_t pid,
2004                          int flock)
2005 {
2006         struct inode *inode = file_inode(file);
2007         struct fuse_conn *fc = get_fuse_conn(inode);
2008         struct fuse_file *ff = file->private_data;
2009         struct fuse_lk_in *arg = &req->misc.lk_in;
2010
2011         arg->fh = ff->fh;
2012         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2013         arg->lk.start = fl->fl_start;
2014         arg->lk.end = fl->fl_end;
2015         arg->lk.type = fl->fl_type;
2016         arg->lk.pid = pid;
2017         if (flock)
2018                 arg->lk_flags |= FUSE_LK_FLOCK;
2019         req->in.h.opcode = opcode;
2020         req->in.h.nodeid = get_node_id(inode);
2021         req->in.numargs = 1;
2022         req->in.args[0].size = sizeof(*arg);
2023         req->in.args[0].value = arg;
2024 }
2025
2026 static int fuse_getlk(struct file *file, struct file_lock *fl)
2027 {
2028         struct inode *inode = file_inode(file);
2029         struct fuse_conn *fc = get_fuse_conn(inode);
2030         struct fuse_req *req;
2031         struct fuse_lk_out outarg;
2032         int err;
2033
2034         req = fuse_get_req_nopages(fc);
2035         if (IS_ERR(req))
2036                 return PTR_ERR(req);
2037
2038         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
2039         req->out.numargs = 1;
2040         req->out.args[0].size = sizeof(outarg);
2041         req->out.args[0].value = &outarg;
2042         fuse_request_send(fc, req);
2043         err = req->out.h.error;
2044         fuse_put_request(fc, req);
2045         if (!err)
2046                 err = convert_fuse_file_lock(&outarg.lk, fl);
2047
2048         return err;
2049 }
2050
2051 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2052 {
2053         struct inode *inode = file_inode(file);
2054         struct fuse_conn *fc = get_fuse_conn(inode);
2055         struct fuse_req *req;
2056         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2057         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2058         int err;
2059
2060         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2061                 /* NLM needs asynchronous locks, which we don't support yet */
2062                 return -ENOLCK;
2063         }
2064
2065         /* Unlock on close is handled by the flush method */
2066         if (fl->fl_flags & FL_CLOSE)
2067                 return 0;
2068
2069         req = fuse_get_req_nopages(fc);
2070         if (IS_ERR(req))
2071                 return PTR_ERR(req);
2072
2073         fuse_lk_fill(req, file, fl, opcode, pid, flock);
2074         fuse_request_send(fc, req);
2075         err = req->out.h.error;
2076         /* locking is restartable */
2077         if (err == -EINTR)
2078                 err = -ERESTARTSYS;
2079         fuse_put_request(fc, req);
2080         return err;
2081 }
2082
2083 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2084 {
2085         struct inode *inode = file_inode(file);
2086         struct fuse_conn *fc = get_fuse_conn(inode);
2087         int err;
2088
2089         if (cmd == F_CANCELLK) {
2090                 err = 0;
2091         } else if (cmd == F_GETLK) {
2092                 if (fc->no_lock) {
2093                         posix_test_lock(file, fl);
2094                         err = 0;
2095                 } else
2096                         err = fuse_getlk(file, fl);
2097         } else {
2098                 if (fc->no_lock)
2099                         err = posix_lock_file(file, fl, NULL);
2100                 else
2101                         err = fuse_setlk(file, fl, 0);
2102         }
2103         return err;
2104 }
2105
2106 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2107 {
2108         struct inode *inode = file_inode(file);
2109         struct fuse_conn *fc = get_fuse_conn(inode);
2110         int err;
2111
2112         if (fc->no_flock) {
2113                 err = flock_lock_file_wait(file, fl);
2114         } else {
2115                 struct fuse_file *ff = file->private_data;
2116
2117                 /* emulate flock with POSIX locks */
2118                 fl->fl_owner = (fl_owner_t) file;
2119                 ff->flock = true;
2120                 err = fuse_setlk(file, fl, 1);
2121         }
2122
2123         return err;
2124 }
2125
2126 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2127 {
2128         struct inode *inode = mapping->host;
2129         struct fuse_conn *fc = get_fuse_conn(inode);
2130         struct fuse_req *req;
2131         struct fuse_bmap_in inarg;
2132         struct fuse_bmap_out outarg;
2133         int err;
2134
2135         if (!inode->i_sb->s_bdev || fc->no_bmap)
2136                 return 0;
2137
2138         req = fuse_get_req_nopages(fc);
2139         if (IS_ERR(req))
2140                 return 0;
2141
2142         memset(&inarg, 0, sizeof(inarg));
2143         inarg.block = block;
2144         inarg.blocksize = inode->i_sb->s_blocksize;
2145         req->in.h.opcode = FUSE_BMAP;
2146         req->in.h.nodeid = get_node_id(inode);
2147         req->in.numargs = 1;
2148         req->in.args[0].size = sizeof(inarg);
2149         req->in.args[0].value = &inarg;
2150         req->out.numargs = 1;
2151         req->out.args[0].size = sizeof(outarg);
2152         req->out.args[0].value = &outarg;
2153         fuse_request_send(fc, req);
2154         err = req->out.h.error;
2155         fuse_put_request(fc, req);
2156         if (err == -ENOSYS)
2157                 fc->no_bmap = 1;
2158
2159         return err ? 0 : outarg.block;
2160 }
2161
2162 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2163 {
2164         loff_t retval;
2165         struct inode *inode = file_inode(file);
2166
2167         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2168         if (whence == SEEK_CUR || whence == SEEK_SET)
2169                 return generic_file_llseek(file, offset, whence);
2170
2171         mutex_lock(&inode->i_mutex);
2172         retval = fuse_update_attributes(inode, NULL, file, NULL);
2173         if (!retval)
2174                 retval = generic_file_llseek(file, offset, whence);
2175         mutex_unlock(&inode->i_mutex);
2176
2177         return retval;
2178 }
2179
2180 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2181                         unsigned int nr_segs, size_t bytes, bool to_user)
2182 {
2183         struct iov_iter ii;
2184         int page_idx = 0;
2185
2186         if (!bytes)
2187                 return 0;
2188
2189         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
2190
2191         while (iov_iter_count(&ii)) {
2192                 struct page *page = pages[page_idx++];
2193                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2194                 void *kaddr;
2195
2196                 kaddr = kmap(page);
2197
2198                 while (todo) {
2199                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2200                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2201                         size_t copy = min(todo, iov_len);
2202                         size_t left;
2203
2204                         if (!to_user)
2205                                 left = copy_from_user(kaddr, uaddr, copy);
2206                         else
2207                                 left = copy_to_user(uaddr, kaddr, copy);
2208
2209                         if (unlikely(left))
2210                                 return -EFAULT;
2211
2212                         iov_iter_advance(&ii, copy);
2213                         todo -= copy;
2214                         kaddr += copy;
2215                 }
2216
2217                 kunmap(page);
2218         }
2219
2220         return 0;
2221 }
2222
2223 /*
2224  * CUSE servers compiled on 32bit broke on 64bit kernels because the
2225  * ABI was defined to be 'struct iovec' which is different on 32bit
2226  * and 64bit.  Fortunately we can determine which structure the server
2227  * used from the size of the reply.
2228  */
2229 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2230                                      size_t transferred, unsigned count,
2231                                      bool is_compat)
2232 {
2233 #ifdef CONFIG_COMPAT
2234         if (count * sizeof(struct compat_iovec) == transferred) {
2235                 struct compat_iovec *ciov = src;
2236                 unsigned i;
2237
2238                 /*
2239                  * With this interface a 32bit server cannot support
2240                  * non-compat (i.e. ones coming from 64bit apps) ioctl
2241                  * requests
2242                  */
2243                 if (!is_compat)
2244                         return -EINVAL;
2245
2246                 for (i = 0; i < count; i++) {
2247                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2248                         dst[i].iov_len = ciov[i].iov_len;
2249                 }
2250                 return 0;
2251         }
2252 #endif
2253
2254         if (count * sizeof(struct iovec) != transferred)
2255                 return -EIO;
2256
2257         memcpy(dst, src, transferred);
2258         return 0;
2259 }
2260
2261 /* Make sure iov_length() won't overflow */
2262 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2263 {
2264         size_t n;
2265         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2266
2267         for (n = 0; n < count; n++, iov++) {
2268                 if (iov->iov_len > (size_t) max)
2269                         return -ENOMEM;
2270                 max -= iov->iov_len;
2271         }
2272         return 0;
2273 }
2274
2275 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2276                                  void *src, size_t transferred, unsigned count,
2277                                  bool is_compat)
2278 {
2279         unsigned i;
2280         struct fuse_ioctl_iovec *fiov = src;
2281
2282         if (fc->minor < 16) {
2283                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2284                                                  count, is_compat);
2285         }
2286
2287         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2288                 return -EIO;
2289
2290         for (i = 0; i < count; i++) {
2291                 /* Did the server supply an inappropriate value? */
2292                 if (fiov[i].base != (unsigned long) fiov[i].base ||
2293                     fiov[i].len != (unsigned long) fiov[i].len)
2294                         return -EIO;
2295
2296                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2297                 dst[i].iov_len = (size_t) fiov[i].len;
2298
2299 #ifdef CONFIG_COMPAT
2300                 if (is_compat &&
2301                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2302                      (compat_size_t) dst[i].iov_len != fiov[i].len))
2303                         return -EIO;
2304 #endif
2305         }
2306
2307         return 0;
2308 }
2309
2310
2311 /*
2312  * For ioctls, there is no generic way to determine how much memory
2313  * needs to be read and/or written.  Furthermore, ioctls are allowed
2314  * to dereference the passed pointer, so the parameter requires deep
2315  * copying but FUSE has no idea whatsoever about what to copy in or
2316  * out.
2317  *
2318  * This is solved by allowing FUSE server to retry ioctl with
2319  * necessary in/out iovecs.  Let's assume the ioctl implementation
2320  * needs to read in the following structure.
2321  *
2322  * struct a {
2323  *      char    *buf;
2324  *      size_t  buflen;
2325  * }
2326  *
2327  * On the first callout to FUSE server, inarg->in_size and
2328  * inarg->out_size will be NULL; then, the server completes the ioctl
2329  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2330  * the actual iov array to
2331  *
2332  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2333  *
2334  * which tells FUSE to copy in the requested area and retry the ioctl.
2335  * On the second round, the server has access to the structure and
2336  * from that it can tell what to look for next, so on the invocation,
2337  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2338  *
2339  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2340  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2341  *
2342  * FUSE will copy both struct a and the pointed buffer from the
2343  * process doing the ioctl and retry ioctl with both struct a and the
2344  * buffer.
2345  *
2346  * This time, FUSE server has everything it needs and completes ioctl
2347  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2348  *
2349  * Copying data out works the same way.
2350  *
2351  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2352  * automatically initializes in and out iovs by decoding @cmd with
2353  * _IOC_* macros and the server is not allowed to request RETRY.  This
2354  * limits ioctl data transfers to well-formed ioctls and is the forced
2355  * behavior for all FUSE servers.
2356  */
2357 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2358                    unsigned int flags)
2359 {
2360         struct fuse_file *ff = file->private_data;
2361         struct fuse_conn *fc = ff->fc;
2362         struct fuse_ioctl_in inarg = {
2363                 .fh = ff->fh,
2364                 .cmd = cmd,
2365                 .arg = arg,
2366                 .flags = flags
2367         };
2368         struct fuse_ioctl_out outarg;
2369         struct fuse_req *req = NULL;
2370         struct page **pages = NULL;
2371         struct iovec *iov_page = NULL;
2372         struct iovec *in_iov = NULL, *out_iov = NULL;
2373         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2374         size_t in_size, out_size, transferred;
2375         int err;
2376
2377 #if BITS_PER_LONG == 32
2378         inarg.flags |= FUSE_IOCTL_32BIT;
2379 #else
2380         if (flags & FUSE_IOCTL_COMPAT)
2381                 inarg.flags |= FUSE_IOCTL_32BIT;
2382 #endif
2383
2384         /* assume all the iovs returned by client always fits in a page */
2385         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2386
2387         err = -ENOMEM;
2388         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2389         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2390         if (!pages || !iov_page)
2391                 goto out;
2392
2393         /*
2394          * If restricted, initialize IO parameters as encoded in @cmd.
2395          * RETRY from server is not allowed.
2396          */
2397         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2398                 struct iovec *iov = iov_page;
2399
2400                 iov->iov_base = (void __user *)arg;
2401                 iov->iov_len = _IOC_SIZE(cmd);
2402
2403                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
2404                         in_iov = iov;
2405                         in_iovs = 1;
2406                 }
2407
2408                 if (_IOC_DIR(cmd) & _IOC_READ) {
2409                         out_iov = iov;
2410                         out_iovs = 1;
2411                 }
2412         }
2413
2414  retry:
2415         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2416         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2417
2418         /*
2419          * Out data can be used either for actual out data or iovs,
2420          * make sure there always is at least one page.
2421          */
2422         out_size = max_t(size_t, out_size, PAGE_SIZE);
2423         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2424
2425         /* make sure there are enough buffer pages and init request with them */
2426         err = -ENOMEM;
2427         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2428                 goto out;
2429         while (num_pages < max_pages) {
2430                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2431                 if (!pages[num_pages])
2432                         goto out;
2433                 num_pages++;
2434         }
2435
2436         req = fuse_get_req(fc, num_pages);
2437         if (IS_ERR(req)) {
2438                 err = PTR_ERR(req);
2439                 req = NULL;
2440                 goto out;
2441         }
2442         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2443         req->num_pages = num_pages;
2444         fuse_page_descs_length_init(req, 0, req->num_pages);
2445
2446         /* okay, let's send it to the client */
2447         req->in.h.opcode = FUSE_IOCTL;
2448         req->in.h.nodeid = ff->nodeid;
2449         req->in.numargs = 1;
2450         req->in.args[0].size = sizeof(inarg);
2451         req->in.args[0].value = &inarg;
2452         if (in_size) {
2453                 req->in.numargs++;
2454                 req->in.args[1].size = in_size;
2455                 req->in.argpages = 1;
2456
2457                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2458                                            false);
2459                 if (err)
2460                         goto out;
2461         }
2462
2463         req->out.numargs = 2;
2464         req->out.args[0].size = sizeof(outarg);
2465         req->out.args[0].value = &outarg;
2466         req->out.args[1].size = out_size;
2467         req->out.argpages = 1;
2468         req->out.argvar = 1;
2469
2470         fuse_request_send(fc, req);
2471         err = req->out.h.error;
2472         transferred = req->out.args[1].size;
2473         fuse_put_request(fc, req);
2474         req = NULL;
2475         if (err)
2476                 goto out;
2477
2478         /* did it ask for retry? */
2479         if (outarg.flags & FUSE_IOCTL_RETRY) {
2480                 void *vaddr;
2481
2482                 /* no retry if in restricted mode */
2483                 err = -EIO;
2484                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2485                         goto out;
2486
2487                 in_iovs = outarg.in_iovs;
2488                 out_iovs = outarg.out_iovs;
2489
2490                 /*
2491                  * Make sure things are in boundary, separate checks
2492                  * are to protect against overflow.
2493                  */
2494                 err = -ENOMEM;
2495                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2496                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2497                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2498                         goto out;
2499
2500                 vaddr = kmap_atomic(pages[0]);
2501                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2502                                             transferred, in_iovs + out_iovs,
2503                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2504                 kunmap_atomic(vaddr);
2505                 if (err)
2506                         goto out;
2507
2508                 in_iov = iov_page;
2509                 out_iov = in_iov + in_iovs;
2510
2511                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2512                 if (err)
2513                         goto out;
2514
2515                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2516                 if (err)
2517                         goto out;
2518
2519                 goto retry;
2520         }
2521
2522         err = -EIO;
2523         if (transferred > inarg.out_size)
2524                 goto out;
2525
2526         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2527  out:
2528         if (req)
2529                 fuse_put_request(fc, req);
2530         free_page((unsigned long) iov_page);
2531         while (num_pages)
2532                 __free_page(pages[--num_pages]);
2533         kfree(pages);
2534
2535         return err ? err : outarg.result;
2536 }
2537 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2538
2539 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2540                        unsigned long arg, unsigned int flags)
2541 {
2542         struct inode *inode = file_inode(file);
2543         struct fuse_conn *fc = get_fuse_conn(inode);
2544
2545         if (!fuse_allow_current_process(fc))
2546                 return -EACCES;
2547
2548         if (is_bad_inode(inode))
2549                 return -EIO;
2550
2551         return fuse_do_ioctl(file, cmd, arg, flags);
2552 }
2553
2554 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2555                             unsigned long arg)
2556 {
2557         return fuse_ioctl_common(file, cmd, arg, 0);
2558 }
2559
2560 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2561                                    unsigned long arg)
2562 {
2563         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2564 }
2565
2566 /*
2567  * All files which have been polled are linked to RB tree
2568  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2569  * find the matching one.
2570  */
2571 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2572                                               struct rb_node **parent_out)
2573 {
2574         struct rb_node **link = &fc->polled_files.rb_node;
2575         struct rb_node *last = NULL;
2576
2577         while (*link) {
2578                 struct fuse_file *ff;
2579
2580                 last = *link;
2581                 ff = rb_entry(last, struct fuse_file, polled_node);
2582
2583                 if (kh < ff->kh)
2584                         link = &last->rb_left;
2585                 else if (kh > ff->kh)
2586                         link = &last->rb_right;
2587                 else
2588                         return link;
2589         }
2590
2591         if (parent_out)
2592                 *parent_out = last;
2593         return link;
2594 }
2595
2596 /*
2597  * The file is about to be polled.  Make sure it's on the polled_files
2598  * RB tree.  Note that files once added to the polled_files tree are
2599  * not removed before the file is released.  This is because a file
2600  * polled once is likely to be polled again.
2601  */
2602 static void fuse_register_polled_file(struct fuse_conn *fc,
2603                                       struct fuse_file *ff)
2604 {
2605         spin_lock(&fc->lock);
2606         if (RB_EMPTY_NODE(&ff->polled_node)) {
2607                 struct rb_node **link, *parent;
2608
2609                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2610                 BUG_ON(*link);
2611                 rb_link_node(&ff->polled_node, parent, link);
2612                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2613         }
2614         spin_unlock(&fc->lock);
2615 }
2616
2617 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2618 {
2619         struct fuse_file *ff = file->private_data;
2620         struct fuse_conn *fc = ff->fc;
2621         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2622         struct fuse_poll_out outarg;
2623         struct fuse_req *req;
2624         int err;
2625
2626         if (fc->no_poll)
2627                 return DEFAULT_POLLMASK;
2628
2629         poll_wait(file, &ff->poll_wait, wait);
2630         inarg.events = (__u32)poll_requested_events(wait);
2631
2632         /*
2633          * Ask for notification iff there's someone waiting for it.
2634          * The client may ignore the flag and always notify.
2635          */
2636         if (waitqueue_active(&ff->poll_wait)) {
2637                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2638                 fuse_register_polled_file(fc, ff);
2639         }
2640
2641         req = fuse_get_req_nopages(fc);
2642         if (IS_ERR(req))
2643                 return POLLERR;
2644
2645         req->in.h.opcode = FUSE_POLL;
2646         req->in.h.nodeid = ff->nodeid;
2647         req->in.numargs = 1;
2648         req->in.args[0].size = sizeof(inarg);
2649         req->in.args[0].value = &inarg;
2650         req->out.numargs = 1;
2651         req->out.args[0].size = sizeof(outarg);
2652         req->out.args[0].value = &outarg;
2653         fuse_request_send(fc, req);
2654         err = req->out.h.error;
2655         fuse_put_request(fc, req);
2656
2657         if (!err)
2658                 return outarg.revents;
2659         if (err == -ENOSYS) {
2660                 fc->no_poll = 1;
2661                 return DEFAULT_POLLMASK;
2662         }
2663         return POLLERR;
2664 }
2665 EXPORT_SYMBOL_GPL(fuse_file_poll);
2666
2667 /*
2668  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2669  * wakes up the poll waiters.
2670  */
2671 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2672                             struct fuse_notify_poll_wakeup_out *outarg)
2673 {
2674         u64 kh = outarg->kh;
2675         struct rb_node **link;
2676
2677         spin_lock(&fc->lock);
2678
2679         link = fuse_find_polled_node(fc, kh, NULL);
2680         if (*link) {
2681                 struct fuse_file *ff;
2682
2683                 ff = rb_entry(*link, struct fuse_file, polled_node);
2684                 wake_up_interruptible_sync(&ff->poll_wait);
2685         }
2686
2687         spin_unlock(&fc->lock);
2688         return 0;
2689 }
2690
2691 static void fuse_do_truncate(struct file *file)
2692 {
2693         struct inode *inode = file->f_mapping->host;
2694         struct iattr attr;
2695
2696         attr.ia_valid = ATTR_SIZE;
2697         attr.ia_size = i_size_read(inode);
2698
2699         attr.ia_file = file;
2700         attr.ia_valid |= ATTR_FILE;
2701
2702         fuse_do_setattr(inode, &attr, file);
2703 }
2704
2705 static inline loff_t fuse_round_up(loff_t off)
2706 {
2707         return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2708 }
2709
2710 static ssize_t
2711 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2712                         loff_t offset, unsigned long nr_segs)
2713 {
2714         ssize_t ret = 0;
2715         struct file *file = iocb->ki_filp;
2716         struct fuse_file *ff = file->private_data;
2717         bool async_dio = ff->fc->async_dio;
2718         loff_t pos = 0;
2719         struct inode *inode;
2720         loff_t i_size;
2721         size_t count = iov_length(iov, nr_segs);
2722         struct fuse_io_priv *io;
2723
2724         pos = offset;
2725         inode = file->f_mapping->host;
2726         i_size = i_size_read(inode);
2727
2728         if ((rw == READ) && (offset > i_size))
2729                 return 0;
2730
2731         /* optimization for short read */
2732         if (async_dio && rw != WRITE && offset + count > i_size) {
2733                 if (offset >= i_size)
2734                         return 0;
2735                 count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2736         }
2737
2738         io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2739         if (!io)
2740                 return -ENOMEM;
2741         spin_lock_init(&io->lock);
2742         io->reqs = 1;
2743         io->bytes = -1;
2744         io->size = 0;
2745         io->offset = offset;
2746         io->write = (rw == WRITE);
2747         io->err = 0;
2748         io->file = file;
2749         /*
2750          * By default, we want to optimize all I/Os with async request
2751          * submission to the client filesystem if supported.
2752          */
2753         io->async = async_dio;
2754         io->iocb = iocb;
2755
2756         /*
2757          * We cannot asynchronously extend the size of a file. We have no method
2758          * to wait on real async I/O requests, so we must submit this request
2759          * synchronously.
2760          */
2761         if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2762                 io->async = false;
2763
2764         if (rw == WRITE)
2765                 ret = __fuse_direct_write(io, iov, nr_segs, &pos);
2766         else
2767                 ret = __fuse_direct_read(io, iov, nr_segs, &pos, count);
2768
2769         if (io->async) {
2770                 fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2771
2772                 /* we have a non-extending, async request, so return */
2773                 if (!is_sync_kiocb(iocb))
2774                         return -EIOCBQUEUED;
2775
2776                 ret = wait_on_sync_kiocb(iocb);
2777         } else {
2778                 kfree(io);
2779         }
2780
2781         if (rw == WRITE) {
2782                 if (ret > 0)
2783                         fuse_write_update_size(inode, pos);
2784                 else if (ret < 0 && offset + count > i_size)
2785                         fuse_do_truncate(file);
2786         }
2787
2788         return ret;
2789 }
2790
2791 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2792                                 loff_t length)
2793 {
2794         struct fuse_file *ff = file->private_data;
2795         struct inode *inode = file->f_inode;
2796         struct fuse_inode *fi = get_fuse_inode(inode);
2797         struct fuse_conn *fc = ff->fc;
2798         struct fuse_req *req;
2799         struct fuse_fallocate_in inarg = {
2800                 .fh = ff->fh,
2801                 .offset = offset,
2802                 .length = length,
2803                 .mode = mode
2804         };
2805         int err;
2806         bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2807                            (mode & FALLOC_FL_PUNCH_HOLE);
2808
2809         if (fc->no_fallocate)
2810                 return -EOPNOTSUPP;
2811
2812         if (lock_inode) {
2813                 mutex_lock(&inode->i_mutex);
2814                 if (mode & FALLOC_FL_PUNCH_HOLE) {
2815                         loff_t endbyte = offset + length - 1;
2816                         err = filemap_write_and_wait_range(inode->i_mapping,
2817                                                            offset, endbyte);
2818                         if (err)
2819                                 goto out;
2820
2821                         fuse_sync_writes(inode);
2822                 }
2823         }
2824
2825         if (!(mode & FALLOC_FL_KEEP_SIZE))
2826                 set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2827
2828         req = fuse_get_req_nopages(fc);
2829         if (IS_ERR(req)) {
2830                 err = PTR_ERR(req);
2831                 goto out;
2832         }
2833
2834         req->in.h.opcode = FUSE_FALLOCATE;
2835         req->in.h.nodeid = ff->nodeid;
2836         req->in.numargs = 1;
2837         req->in.args[0].size = sizeof(inarg);
2838         req->in.args[0].value = &inarg;
2839         fuse_request_send(fc, req);
2840         err = req->out.h.error;
2841         if (err == -ENOSYS) {
2842                 fc->no_fallocate = 1;
2843                 err = -EOPNOTSUPP;
2844         }
2845         fuse_put_request(fc, req);
2846
2847         if (err)
2848                 goto out;
2849
2850         /* we could have extended the file */
2851         if (!(mode & FALLOC_FL_KEEP_SIZE))
2852                 fuse_write_update_size(inode, offset + length);
2853
2854         if (mode & FALLOC_FL_PUNCH_HOLE)
2855                 truncate_pagecache_range(inode, offset, offset + length - 1);
2856
2857         fuse_invalidate_attr(inode);
2858
2859 out:
2860         if (!(mode & FALLOC_FL_KEEP_SIZE))
2861                 clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2862
2863         if (lock_inode)
2864                 mutex_unlock(&inode->i_mutex);
2865
2866         return err;
2867 }
2868
2869 static const struct file_operations fuse_file_operations = {
2870         .llseek         = fuse_file_llseek,
2871         .read           = do_sync_read,
2872         .aio_read       = fuse_file_aio_read,
2873         .write          = do_sync_write,
2874         .aio_write      = fuse_file_aio_write,
2875         .mmap           = fuse_file_mmap,
2876         .open           = fuse_open,
2877         .flush          = fuse_flush,
2878         .release        = fuse_release,
2879         .fsync          = fuse_fsync,
2880         .lock           = fuse_file_lock,
2881         .flock          = fuse_file_flock,
2882         .splice_read    = generic_file_splice_read,
2883         .unlocked_ioctl = fuse_file_ioctl,
2884         .compat_ioctl   = fuse_file_compat_ioctl,
2885         .poll           = fuse_file_poll,
2886         .fallocate      = fuse_file_fallocate,
2887 };
2888
2889 static const struct file_operations fuse_direct_io_file_operations = {
2890         .llseek         = fuse_file_llseek,
2891         .read           = fuse_direct_read,
2892         .write          = fuse_direct_write,
2893         .mmap           = fuse_direct_mmap,
2894         .open           = fuse_open,
2895         .flush          = fuse_flush,
2896         .release        = fuse_release,
2897         .fsync          = fuse_fsync,
2898         .lock           = fuse_file_lock,
2899         .flock          = fuse_file_flock,
2900         .unlocked_ioctl = fuse_file_ioctl,
2901         .compat_ioctl   = fuse_file_compat_ioctl,
2902         .poll           = fuse_file_poll,
2903         .fallocate      = fuse_file_fallocate,
2904         /* no splice_read */
2905 };
2906
2907 static const struct address_space_operations fuse_file_aops  = {
2908         .readpage       = fuse_readpage,
2909         .writepage      = fuse_writepage,
2910         .writepages     = fuse_writepages,
2911         .launder_page   = fuse_launder_page,
2912         .readpages      = fuse_readpages,
2913         .set_page_dirty = __set_page_dirty_nobuffers,
2914         .bmap           = fuse_bmap,
2915         .direct_IO      = fuse_direct_IO,
2916 };
2917
2918 void fuse_init_file_inode(struct inode *inode)
2919 {
2920         inode->i_fop = &fuse_file_operations;
2921         inode->i_data.a_ops = &fuse_file_aops;
2922 }