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
  17 static const struct file_operations fuse_direct_io_file_operations;
  18
  19 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  20                           int opcode, struct fuse_open_out *outargp)
  21 {
  22         struct fuse_open_in inarg;
  23         struct fuse_req *req;
  24         int err;
  25
  26         req = fuse_get_req(fc);
  27         if (IS_ERR(req))
  28                 return PTR_ERR(req);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34         req->in.h.opcode = opcode;
  35         req->in.h.nodeid = nodeid;
  36         req->in.numargs = 1;
  37         req->in.args[0].size = sizeof(inarg);
  38         req->in.args[0].value = &inarg;
  39         req->out.numargs = 1;
  40         req->out.args[0].size = sizeof(*outargp);
  41         req->out.args[0].value = outargp;
  42         fuse_request_send(fc, req);
  43         err = req->out.h.error;
  44         fuse_put_request(fc, req);
  45
  46         return err;
  47 }
  48
  49 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  50 {
  51         struct fuse_file *ff;
  52
  53         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  54         if (unlikely(!ff))
  55                 return NULL;
  56
  57         ff->fc = fc;
  58         ff->reserved_req = fuse_request_alloc();
  59         if (unlikely(!ff->reserved_req)) {
  60                 kfree(ff);
  61                 return NULL;
  62         }
  63
  64         INIT_LIST_HEAD(&ff->write_entry);
  65         atomic_set(&ff->count, 0);
  66         RB_CLEAR_NODE(&ff->polled_node);
  67         init_waitqueue_head(&ff->poll_wait);
  68
  69         spin_lock(&fc->lock);
  70         ff->kh = ++fc->khctr;
  71         spin_unlock(&fc->lock);
  72
  73         return ff;
  74 }
  75
  76 void fuse_file_free(struct fuse_file *ff)
  77 {
  78         fuse_request_free(ff->reserved_req);
  79         kfree(ff);
  80 }
  81
  82 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  83 {
  84         atomic_inc(&ff->count);
  85         return ff;
  86 }
  87
  88 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  89 {
  90         path_put(&req->misc.release.path);
  91 }
  92
  93 static void fuse_file_put(struct fuse_file *ff)
  94 {
  95         if (atomic_dec_and_test(&ff->count)) {
  96                 struct fuse_req *req = ff->reserved_req;
  97
  98                 req->end = fuse_release_end;
  99                 fuse_request_send_background(ff->fc, req);
 100                 kfree(ff);
 101         }
 102 }
 103
 104 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 105                  bool isdir)
 106 {
 107         struct fuse_open_out outarg;
 108         struct fuse_file *ff;
 109         int err;
 110         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 111
 112         ff = fuse_file_alloc(fc);
 113         if (!ff)
 114                 return -ENOMEM;
 115
 116         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 117         if (err) {
 118                 fuse_file_free(ff);
 119                 return err;
 120         }
 121
 122         if (isdir)
 123                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 124
 125         ff->fh = outarg.fh;
 126         ff->nodeid = nodeid;
 127         ff->open_flags = outarg.open_flags;
 128         file->private_data = fuse_file_get(ff);
 129
 130         return 0;
 131 }
 132 EXPORT_SYMBOL_GPL(fuse_do_open);
 133
 134 void fuse_finish_open(struct inode *inode, struct file *file)
 135 {
 136         struct fuse_file *ff = file->private_data;
 137
 138         if (ff->open_flags & FOPEN_DIRECT_IO)
 139                 file->f_op = &fuse_direct_io_file_operations;
 140         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 141                 invalidate_inode_pages2(inode->i_mapping);
 142         if (ff->open_flags & FOPEN_NONSEEKABLE)
 143                 nonseekable_open(inode, file);
 144 }
 145
 146 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 147 {
 148         struct fuse_conn *fc = get_fuse_conn(inode);
 149         int err;
 150
 151         /* VFS checks this, but only _after_ ->open() */
 152         if (file->f_flags & O_DIRECT)
 153                 return -EINVAL;
 154
 155         err = generic_file_open(inode, file);
 156         if (err)
 157                 return err;
 158
 159         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 160         if (err)
 161                 return err;
 162
 163         fuse_finish_open(inode, file);
 164
 165         return 0;
 166 }
 167
 168 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 169 {
 170         struct fuse_conn *fc = ff->fc;
 171         struct fuse_req *req = ff->reserved_req;
 172         struct fuse_release_in *inarg = &req->misc.release.in;
 173
 174         spin_lock(&fc->lock);
 175         list_del(&ff->write_entry);
 176         if (!RB_EMPTY_NODE(&ff->polled_node))
 177                 rb_erase(&ff->polled_node, &fc->polled_files);
 178         spin_unlock(&fc->lock);
 179
 180         wake_up_interruptible_sync(&ff->poll_wait);
 181
 182         inarg->fh = ff->fh;
 183         inarg->flags = flags;
 184         req->in.h.opcode = opcode;
 185         req->in.h.nodeid = ff->nodeid;
 186         req->in.numargs = 1;
 187         req->in.args[0].size = sizeof(struct fuse_release_in);
 188         req->in.args[0].value = inarg;
 189 }
 190
 191 void fuse_release_common(struct file *file, int opcode)
 192 {
 193         struct fuse_file *ff;
 194         struct fuse_req *req;
 195
 196         ff = file->private_data;
 197         if (unlikely(!ff))
 198                 return;
 199
 200         req = ff->reserved_req;
 201         fuse_prepare_release(ff, file->f_flags, opcode);
 202
 203         /* Hold vfsmount and dentry until release is finished */
 204         path_get(&file->f_path);
 205         req->misc.release.path = file->f_path;
 206
 207         /*
 208          * Normally this will send the RELEASE request, however if
 209          * some asynchronous READ or WRITE requests are outstanding,
 210          * the sending will be delayed.
 211          */
 212         fuse_file_put(ff);
 213 }
 214
 215 static int fuse_open(struct inode *inode, struct file *file)
 216 {
 217         return fuse_open_common(inode, file, false);
 218 }
 219
 220 static int fuse_release(struct inode *inode, struct file *file)
 221 {
 222         fuse_release_common(file, FUSE_RELEASE);
 223
 224         /* return value is ignored by VFS */
 225         return 0;
 226 }
 227
 228 void fuse_sync_release(struct fuse_file *ff, int flags)
 229 {
 230         WARN_ON(atomic_read(&ff->count) > 1);
 231         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 232         ff->reserved_req->force = 1;
 233         fuse_request_send(ff->fc, ff->reserved_req);
 234         fuse_put_request(ff->fc, ff->reserved_req);
 235         kfree(ff);
 236 }
 237 EXPORT_SYMBOL_GPL(fuse_sync_release);
 238
 239 /*
 240  * Scramble the ID space with XTEA, so that the value of the files_struct
 241  * pointer is not exposed to userspace.
 242  */
 243 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 244 {
 245         u32 *k = fc->scramble_key;
 246         u64 v = (unsigned long) id;
 247         u32 v0 = v;
 248         u32 v1 = v >> 32;
 249         u32 sum = 0;
 250         int i;
 251
 252         for (i = 0; i < 32; i++) {
 253                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 254                 sum += 0x9E3779B9;
 255                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 256         }
 257
 258         return (u64) v0 + ((u64) v1 << 32);
 259 }
 260
 261 /*
 262  * Check if page is under writeback
 263  *
 264  * This is currently done by walking the list of writepage requests
 265  * for the inode, which can be pretty inefficient.
 266  */
 267 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 268 {
 269         struct fuse_conn *fc = get_fuse_conn(inode);
 270         struct fuse_inode *fi = get_fuse_inode(inode);
 271         struct fuse_req *req;
 272         bool found = false;
 273
 274         spin_lock(&fc->lock);
 275         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 276                 pgoff_t curr_index;
 277
 278                 BUG_ON(req->inode != inode);
 279                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 280                 if (curr_index == index) {
 281                         found = true;
 282                         break;
 283                 }
 284         }
 285         spin_unlock(&fc->lock);
 286
 287         return found;
 288 }
 289
 290 /*
 291  * Wait for page writeback to be completed.
 292  *
 293  * Since fuse doesn't rely on the VM writeback tracking, this has to
 294  * use some other means.
 295  */
 296 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 297 {
 298         struct fuse_inode *fi = get_fuse_inode(inode);
 299
 300         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 301         return 0;
 302 }
 303
 304 static int fuse_flush(struct file *file, fl_owner_t id)
 305 {
 306         struct inode *inode = file->f_path.dentry->d_inode;
 307         struct fuse_conn *fc = get_fuse_conn(inode);
 308         struct fuse_file *ff = file->private_data;
 309         struct fuse_req *req;
 310         struct fuse_flush_in inarg;
 311         int err;
 312
 313         if (is_bad_inode(inode))
 314                 return -EIO;
 315
 316         if (fc->no_flush)
 317                 return 0;
 318
 319         req = fuse_get_req_nofail(fc, file);
 320         memset(&inarg, 0, sizeof(inarg));
 321         inarg.fh = ff->fh;
 322         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 323         req->in.h.opcode = FUSE_FLUSH;
 324         req->in.h.nodeid = get_node_id(inode);
 325         req->in.numargs = 1;
 326         req->in.args[0].size = sizeof(inarg);
 327         req->in.args[0].value = &inarg;
 328         req->force = 1;
 329         fuse_request_send(fc, req);
 330         err = req->out.h.error;
 331         fuse_put_request(fc, req);
 332         if (err == -ENOSYS) {
 333                 fc->no_flush = 1;
 334                 err = 0;
 335         }
 336         return err;
 337 }
 338
 339 /*
 340  * Wait for all pending writepages on the inode to finish.
 341  *
 342  * This is currently done by blocking further writes with FUSE_NOWRITE
 343  * and waiting for all sent writes to complete.
 344  *
 345  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 346  * could conflict with truncation.
 347  */
 348 static void fuse_sync_writes(struct inode *inode)
 349 {
 350         fuse_set_nowrite(inode);
 351         fuse_release_nowrite(inode);
 352 }
 353
 354 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
 355                       int isdir)
 356 {
 357         struct inode *inode = de->d_inode;
 358         struct fuse_conn *fc = get_fuse_conn(inode);
 359         struct fuse_file *ff = file->private_data;
 360         struct fuse_req *req;
 361         struct fuse_fsync_in inarg;
 362         int err;
 363
 364         if (is_bad_inode(inode))
 365                 return -EIO;
 366
 367         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 368                 return 0;
 369
 370         /*
 371          * Start writeback against all dirty pages of the inode, then
 372          * wait for all outstanding writes, before sending the FSYNC
 373          * request.
 374          */
 375         err = write_inode_now(inode, 0);
 376         if (err)
 377                 return err;
 378
 379         fuse_sync_writes(inode);
 380
 381         req = fuse_get_req(fc);
 382         if (IS_ERR(req))
 383                 return PTR_ERR(req);
 384
 385         memset(&inarg, 0, sizeof(inarg));
 386         inarg.fh = ff->fh;
 387         inarg.fsync_flags = datasync ? 1 : 0;
 388         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 389         req->in.h.nodeid = get_node_id(inode);
 390         req->in.numargs = 1;
 391         req->in.args[0].size = sizeof(inarg);
 392         req->in.args[0].value = &inarg;
 393         fuse_request_send(fc, req);
 394         err = req->out.h.error;
 395         fuse_put_request(fc, req);
 396         if (err == -ENOSYS) {
 397                 if (isdir)
 398                         fc->no_fsyncdir = 1;
 399                 else
 400                         fc->no_fsync = 1;
 401                 err = 0;
 402         }
 403         return err;
 404 }
 405
 406 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
 407 {
 408         return fuse_fsync_common(file, de, datasync, 0);
 409 }
 410
 411 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 412                     size_t count, int opcode)
 413 {
 414         struct fuse_read_in *inarg = &req->misc.read.in;
 415         struct fuse_file *ff = file->private_data;
 416
 417         inarg->fh = ff->fh;
 418         inarg->offset = pos;
 419         inarg->size = count;
 420         inarg->flags = file->f_flags;
 421         req->in.h.opcode = opcode;
 422         req->in.h.nodeid = ff->nodeid;
 423         req->in.numargs = 1;
 424         req->in.args[0].size = sizeof(struct fuse_read_in);
 425         req->in.args[0].value = inarg;
 426         req->out.argvar = 1;
 427         req->out.numargs = 1;
 428         req->out.args[0].size = count;
 429 }
 430
 431 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 432                              loff_t pos, size_t count, fl_owner_t owner)
 433 {
 434         struct fuse_file *ff = file->private_data;
 435         struct fuse_conn *fc = ff->fc;
 436
 437         fuse_read_fill(req, file, pos, count, FUSE_READ);
 438         if (owner != NULL) {
 439                 struct fuse_read_in *inarg = &req->misc.read.in;
 440
 441                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 442                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 443         }
 444         fuse_request_send(fc, req);
 445         return req->out.args[0].size;
 446 }
 447
 448 static void fuse_read_update_size(struct inode *inode, loff_t size,
 449                                   u64 attr_ver)
 450 {
 451         struct fuse_conn *fc = get_fuse_conn(inode);
 452         struct fuse_inode *fi = get_fuse_inode(inode);
 453
 454         spin_lock(&fc->lock);
 455         if (attr_ver == fi->attr_version && size < inode->i_size) {
 456                 fi->attr_version = ++fc->attr_version;
 457                 i_size_write(inode, size);
 458         }
 459         spin_unlock(&fc->lock);
 460 }
 461
 462 static int fuse_readpage(struct file *file, struct page *page)
 463 {
 464         struct inode *inode = page->mapping->host;
 465         struct fuse_conn *fc = get_fuse_conn(inode);
 466         struct fuse_req *req;
 467         size_t num_read;
 468         loff_t pos = page_offset(page);
 469         size_t count = PAGE_CACHE_SIZE;
 470         u64 attr_ver;
 471         int err;
 472
 473         err = -EIO;
 474         if (is_bad_inode(inode))
 475                 goto out;
 476
 477         /*
 478          * Page writeback can extend beyond the liftime of the
 479          * page-cache page, so make sure we read a properly synced
 480          * page.
 481          */
 482         fuse_wait_on_page_writeback(inode, page->index);
 483
 484         req = fuse_get_req(fc);
 485         err = PTR_ERR(req);
 486         if (IS_ERR(req))
 487                 goto out;
 488
 489         attr_ver = fuse_get_attr_version(fc);
 490
 491         req->out.page_zeroing = 1;
 492         req->out.argpages = 1;
 493         req->num_pages = 1;
 494         req->pages[0] = page;
 495         num_read = fuse_send_read(req, file, pos, count, NULL);
 496         err = req->out.h.error;
 497         fuse_put_request(fc, req);
 498
 499         if (!err) {
 500                 /*
 501                  * Short read means EOF.  If file size is larger, truncate it
 502                  */
 503                 if (num_read < count)
 504                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 505
 506                 SetPageUptodate(page);
 507         }
 508
 509         fuse_invalidate_attr(inode); /* atime changed */
 510  out:
 511         unlock_page(page);
 512         return err;
 513 }
 514
 515 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 516 {
 517         int i;
 518         size_t count = req->misc.read.in.size;
 519         size_t num_read = req->out.args[0].size;
 520         struct address_space *mapping = NULL;
 521
 522         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 523                 mapping = req->pages[i]->mapping;
 524
 525         if (mapping) {
 526                 struct inode *inode = mapping->host;
 527
 528                 /*
 529                  * Short read means EOF. If file size is larger, truncate it
 530                  */
 531                 if (!req->out.h.error && num_read < count) {
 532                         loff_t pos;
 533
 534                         pos = page_offset(req->pages[0]) + num_read;
 535                         fuse_read_update_size(inode, pos,
 536                                               req->misc.read.attr_ver);
 537                 }
 538                 fuse_invalidate_attr(inode); /* atime changed */
 539         }
 540
 541         for (i = 0; i < req->num_pages; i++) {
 542                 struct page *page = req->pages[i];
 543                 if (!req->out.h.error)
 544                         SetPageUptodate(page);
 545                 else
 546                         SetPageError(page);
 547                 unlock_page(page);
 548                 page_cache_release(page);
 549         }
 550         if (req->ff)
 551                 fuse_file_put(req->ff);
 552 }
 553
 554 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 555 {
 556         struct fuse_file *ff = file->private_data;
 557         struct fuse_conn *fc = ff->fc;
 558         loff_t pos = page_offset(req->pages[0]);
 559         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 560
 561         req->out.argpages = 1;
 562         req->out.page_zeroing = 1;
 563         req->out.page_replace = 1;
 564         fuse_read_fill(req, file, pos, count, FUSE_READ);
 565         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 566         if (fc->async_read) {
 567                 req->ff = fuse_file_get(ff);
 568                 req->end = fuse_readpages_end;
 569                 fuse_request_send_background(fc, req);
 570         } else {
 571                 fuse_request_send(fc, req);
 572                 fuse_readpages_end(fc, req);
 573                 fuse_put_request(fc, req);
 574         }
 575 }
 576
 577 struct fuse_fill_data {
 578         struct fuse_req *req;
 579         struct file *file;
 580         struct inode *inode;
 581 };
 582
 583 static int fuse_readpages_fill(void *_data, struct page *page)
 584 {
 585         struct fuse_fill_data *data = _data;
 586         struct fuse_req *req = data->req;
 587         struct inode *inode = data->inode;
 588         struct fuse_conn *fc = get_fuse_conn(inode);
 589
 590         fuse_wait_on_page_writeback(inode, page->index);
 591
 592         if (req->num_pages &&
 593             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 594              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 595              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 596                 fuse_send_readpages(req, data->file);
 597                 data->req = req = fuse_get_req(fc);
 598                 if (IS_ERR(req)) {
 599                         unlock_page(page);
 600                         return PTR_ERR(req);
 601                 }
 602         }
 603         page_cache_get(page);
 604         req->pages[req->num_pages] = page;
 605         req->num_pages++;
 606         return 0;
 607 }
 608
 609 static int fuse_readpages(struct file *file, struct address_space *mapping,
 610                           struct list_head *pages, unsigned nr_pages)
 611 {
 612         struct inode *inode = mapping->host;
 613         struct fuse_conn *fc = get_fuse_conn(inode);
 614         struct fuse_fill_data data;
 615         int err;
 616
 617         err = -EIO;
 618         if (is_bad_inode(inode))
 619                 goto out;
 620
 621         data.file = file;
 622         data.inode = inode;
 623         data.req = fuse_get_req(fc);
 624         err = PTR_ERR(data.req);
 625         if (IS_ERR(data.req))
 626                 goto out;
 627
 628         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 629         if (!err) {
 630                 if (data.req->num_pages)
 631                         fuse_send_readpages(data.req, file);
 632                 else
 633                         fuse_put_request(fc, data.req);
 634         }
 635 out:
 636         return err;
 637 }
 638
 639 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 640                                   unsigned long nr_segs, loff_t pos)
 641 {
 642         struct inode *inode = iocb->ki_filp->f_mapping->host;
 643
 644         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 645                 int err;
 646                 /*
 647                  * If trying to read past EOF, make sure the i_size
 648                  * attribute is up-to-date.
 649                  */
 650                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 651                 if (err)
 652                         return err;
 653         }
 654
 655         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 656 }
 657
 658 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 659                             loff_t pos, size_t count)
 660 {
 661         struct fuse_write_in *inarg = &req->misc.write.in;
 662         struct fuse_write_out *outarg = &req->misc.write.out;
 663
 664         inarg->fh = ff->fh;
 665         inarg->offset = pos;
 666         inarg->size = count;
 667         req->in.h.opcode = FUSE_WRITE;
 668         req->in.h.nodeid = ff->nodeid;
 669         req->in.numargs = 2;
 670         if (ff->fc->minor < 9)
 671                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 672         else
 673                 req->in.args[0].size = sizeof(struct fuse_write_in);
 674         req->in.args[0].value = inarg;
 675         req->in.args[1].size = count;
 676         req->out.numargs = 1;
 677         req->out.args[0].size = sizeof(struct fuse_write_out);
 678         req->out.args[0].value = outarg;
 679 }
 680
 681 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 682                               loff_t pos, size_t count, fl_owner_t owner)
 683 {
 684         struct fuse_file *ff = file->private_data;
 685         struct fuse_conn *fc = ff->fc;
 686         struct fuse_write_in *inarg = &req->misc.write.in;
 687
 688         fuse_write_fill(req, ff, pos, count);
 689         inarg->flags = file->f_flags;
 690         if (owner != NULL) {
 691                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 692                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 693         }
 694         fuse_request_send(fc, req);
 695         return req->misc.write.out.size;
 696 }
 697
 698 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 699                         loff_t pos, unsigned len, unsigned flags,
 700                         struct page **pagep, void **fsdata)
 701 {
 702         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 703
 704         *pagep = grab_cache_page_write_begin(mapping, index, flags);
 705         if (!*pagep)
 706                 return -ENOMEM;
 707         return 0;
 708 }
 709
 710 static void fuse_write_update_size(struct inode *inode, loff_t pos)
 711 {
 712         struct fuse_conn *fc = get_fuse_conn(inode);
 713         struct fuse_inode *fi = get_fuse_inode(inode);
 714
 715         spin_lock(&fc->lock);
 716         fi->attr_version = ++fc->attr_version;
 717         if (pos > inode->i_size)
 718                 i_size_write(inode, pos);
 719         spin_unlock(&fc->lock);
 720 }
 721
 722 static int fuse_buffered_write(struct file *file, struct inode *inode,
 723                                loff_t pos, unsigned count, struct page *page)
 724 {
 725         int err;
 726         size_t nres;
 727         struct fuse_conn *fc = get_fuse_conn(inode);
 728         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 729         struct fuse_req *req;
 730
 731         if (is_bad_inode(inode))
 732                 return -EIO;
 733
 734         /*
 735          * Make sure writepages on the same page are not mixed up with
 736          * plain writes.
 737          */
 738         fuse_wait_on_page_writeback(inode, page->index);
 739
 740         req = fuse_get_req(fc);
 741         if (IS_ERR(req))
 742                 return PTR_ERR(req);
 743
 744         req->in.argpages = 1;
 745         req->num_pages = 1;
 746         req->pages[0] = page;
 747         req->page_offset = offset;
 748         nres = fuse_send_write(req, file, pos, count, NULL);
 749         err = req->out.h.error;
 750         fuse_put_request(fc, req);
 751         if (!err && !nres)
 752                 err = -EIO;
 753         if (!err) {
 754                 pos += nres;
 755                 fuse_write_update_size(inode, pos);
 756                 if (count == PAGE_CACHE_SIZE)
 757                         SetPageUptodate(page);
 758         }
 759         fuse_invalidate_attr(inode);
 760         return err ? err : nres;
 761 }
 762
 763 static int fuse_write_end(struct file *file, struct address_space *mapping,
 764                         loff_t pos, unsigned len, unsigned copied,
 765                         struct page *page, void *fsdata)
 766 {
 767         struct inode *inode = mapping->host;
 768         int res = 0;
 769
 770         if (copied)
 771                 res = fuse_buffered_write(file, inode, pos, copied, page);
 772
 773         unlock_page(page);
 774         page_cache_release(page);
 775         return res;
 776 }
 777
 778 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 779                                     struct inode *inode, loff_t pos,
 780                                     size_t count)
 781 {
 782         size_t res;
 783         unsigned offset;
 784         unsigned i;
 785
 786         for (i = 0; i < req->num_pages; i++)
 787                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 788
 789         res = fuse_send_write(req, file, pos, count, NULL);
 790
 791         offset = req->page_offset;
 792         count = res;
 793         for (i = 0; i < req->num_pages; i++) {
 794                 struct page *page = req->pages[i];
 795
 796                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 797                         SetPageUptodate(page);
 798
 799                 if (count > PAGE_CACHE_SIZE - offset)
 800                         count -= PAGE_CACHE_SIZE - offset;
 801                 else
 802                         count = 0;
 803                 offset = 0;
 804
 805                 unlock_page(page);
 806                 page_cache_release(page);
 807         }
 808
 809         return res;
 810 }
 811
 812 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 813                                struct address_space *mapping,
 814                                struct iov_iter *ii, loff_t pos)
 815 {
 816         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 817         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 818         size_t count = 0;
 819         int err;
 820
 821         req->in.argpages = 1;
 822         req->page_offset = offset;
 823
 824         do {
 825                 size_t tmp;
 826                 struct page *page;
 827                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 828                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 829                                      iov_iter_count(ii));
 830
 831                 bytes = min_t(size_t, bytes, fc->max_write - count);
 832
 833  again:
 834                 err = -EFAULT;
 835                 if (iov_iter_fault_in_readable(ii, bytes))
 836                         break;
 837
 838                 err = -ENOMEM;
 839                 page = grab_cache_page_write_begin(mapping, index, 0);
 840                 if (!page)
 841                         break;
 842
 843                 if (mapping_writably_mapped(mapping))
 844                         flush_dcache_page(page);
 845
 846                 pagefault_disable();
 847                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 848                 pagefault_enable();
 849                 flush_dcache_page(page);
 850
 851                 if (!tmp) {
 852                         unlock_page(page);
 853                         page_cache_release(page);
 854                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 855                         goto again;
 856                 }
 857
 858                 err = 0;
 859                 req->pages[req->num_pages] = page;
 860                 req->num_pages++;
 861
 862                 iov_iter_advance(ii, tmp);
 863                 count += tmp;
 864                 pos += tmp;
 865                 offset += tmp;
 866                 if (offset == PAGE_CACHE_SIZE)
 867                         offset = 0;
 868
 869                 if (!fc->big_writes)
 870                         break;
 871         } while (iov_iter_count(ii) && count < fc->max_write &&
 872                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 873
 874         return count > 0 ? count : err;
 875 }
 876
 877 static ssize_t fuse_perform_write(struct file *file,
 878                                   struct address_space *mapping,
 879                                   struct iov_iter *ii, loff_t pos)
 880 {
 881         struct inode *inode = mapping->host;
 882         struct fuse_conn *fc = get_fuse_conn(inode);
 883         int err = 0;
 884         ssize_t res = 0;
 885
 886         if (is_bad_inode(inode))
 887                 return -EIO;
 888
 889         do {
 890                 struct fuse_req *req;
 891                 ssize_t count;
 892
 893                 req = fuse_get_req(fc);
 894                 if (IS_ERR(req)) {
 895                         err = PTR_ERR(req);
 896                         break;
 897                 }
 898
 899                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 900                 if (count <= 0) {
 901                         err = count;
 902                 } else {
 903                         size_t num_written;
 904
 905                         num_written = fuse_send_write_pages(req, file, inode,
 906                                                             pos, count);
 907                         err = req->out.h.error;
 908                         if (!err) {
 909                                 res += num_written;
 910                                 pos += num_written;
 911
 912                                 /* break out of the loop on short write */
 913                                 if (num_written != count)
 914                                         err = -EIO;
 915                         }
 916                 }
 917                 fuse_put_request(fc, req);
 918         } while (!err && iov_iter_count(ii));
 919
 920         if (res > 0)
 921                 fuse_write_update_size(inode, pos);
 922
 923         fuse_invalidate_attr(inode);
 924
 925         return res > 0 ? res : err;
 926 }
 927
 928 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 929                                    unsigned long nr_segs, loff_t pos)
 930 {
 931         struct file *file = iocb->ki_filp;
 932         struct address_space *mapping = file->f_mapping;
 933         size_t count = 0;
 934         ssize_t written = 0;
 935         struct inode *inode = mapping->host;
 936         ssize_t err;
 937         struct iov_iter i;
 938
 939         WARN_ON(iocb->ki_pos != pos);
 940
 941         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 942         if (err)
 943                 return err;
 944
 945         mutex_lock(&inode->i_mutex);
 946         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 947
 948         /* We can write back this queue in page reclaim */
 949         current->backing_dev_info = mapping->backing_dev_info;
 950
 951         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 952         if (err)
 953                 goto out;
 954
 955         if (count == 0)
 956                 goto out;
 957
 958         err = file_remove_suid(file);
 959         if (err)
 960                 goto out;
 961
 962         file_update_time(file);
 963
 964         iov_iter_init(&i, iov, nr_segs, count, 0);
 965         written = fuse_perform_write(file, mapping, &i, pos);
 966         if (written >= 0)
 967                 iocb->ki_pos = pos + written;
 968
 969 out:
 970         current->backing_dev_info = NULL;
 971         mutex_unlock(&inode->i_mutex);
 972
 973         return written ? written : err;
 974 }
 975
 976 static void fuse_release_user_pages(struct fuse_req *req, int write)
 977 {
 978         unsigned i;
 979
 980         for (i = 0; i < req->num_pages; i++) {
 981                 struct page *page = req->pages[i];
 982                 if (write)
 983                         set_page_dirty_lock(page);
 984                 put_page(page);
 985         }
 986 }
 987
 988 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 989                                size_t *nbytesp, int write)
 990 {
 991         size_t nbytes = *nbytesp;
 992         unsigned long user_addr = (unsigned long) buf;
 993         unsigned offset = user_addr & ~PAGE_MASK;
 994         int npages;
 995
 996         /* Special case for kernel I/O: can copy directly into the buffer */
 997         if (segment_eq(get_fs(), KERNEL_DS)) {
 998                 if (write)
 999                         req->in.args[1].value = (void *) user_addr;
1000                 else
1001                         req->out.args[0].value = (void *) user_addr;
1002
1003                 return 0;
1004         }
1005
1006         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1007         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1008         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1009         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1010         if (npages < 0)
1011                 return npages;
1012
1013         req->num_pages = npages;
1014         req->page_offset = offset;
1015
1016         if (write)
1017                 req->in.argpages = 1;
1018         else
1019                 req->out.argpages = 1;
1020
1021         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1022         *nbytesp = min(*nbytesp, nbytes);
1023
1024         return 0;
1025 }
1026
1027 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1028                        size_t count, loff_t *ppos, int write)
1029 {
1030         struct fuse_file *ff = file->private_data;
1031         struct fuse_conn *fc = ff->fc;
1032         size_t nmax = write ? fc->max_write : fc->max_read;
1033         loff_t pos = *ppos;
1034         ssize_t res = 0;
1035         struct fuse_req *req;
1036
1037         req = fuse_get_req(fc);
1038         if (IS_ERR(req))
1039                 return PTR_ERR(req);
1040
1041         while (count) {
1042                 size_t nres;
1043                 fl_owner_t owner = current->files;
1044                 size_t nbytes = min(count, nmax);
1045                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1046                 if (err) {
1047                         res = err;
1048                         break;
1049                 }
1050
1051                 if (write)
1052                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1053                 else
1054                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1055
1056                 fuse_release_user_pages(req, !write);
1057                 if (req->out.h.error) {
1058                         if (!res)
1059                                 res = req->out.h.error;
1060                         break;
1061                 } else if (nres > nbytes) {
1062                         res = -EIO;
1063                         break;
1064                 }
1065                 count -= nres;
1066                 res += nres;
1067                 pos += nres;
1068                 buf += nres;
1069                 if (nres != nbytes)
1070                         break;
1071                 if (count) {
1072                         fuse_put_request(fc, req);
1073                         req = fuse_get_req(fc);
1074                         if (IS_ERR(req))
1075                                 break;
1076                 }
1077         }
1078         if (!IS_ERR(req))
1079                 fuse_put_request(fc, req);
1080         if (res > 0)
1081                 *ppos = pos;
1082
1083         return res;
1084 }
1085 EXPORT_SYMBOL_GPL(fuse_direct_io);
1086
1087 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1088                                      size_t count, loff_t *ppos)
1089 {
1090         ssize_t res;
1091         struct inode *inode = file->f_path.dentry->d_inode;
1092
1093         if (is_bad_inode(inode))
1094                 return -EIO;
1095
1096         res = fuse_direct_io(file, buf, count, ppos, 0);
1097
1098         fuse_invalidate_attr(inode);
1099
1100         return res;
1101 }
1102
1103 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1104                                  size_t count, loff_t *ppos)
1105 {
1106         struct inode *inode = file->f_path.dentry->d_inode;
1107         ssize_t res;
1108
1109         if (is_bad_inode(inode))
1110                 return -EIO;
1111
1112         /* Don't allow parallel writes to the same file */
1113         mutex_lock(&inode->i_mutex);
1114         res = generic_write_checks(file, ppos, &count, 0);
1115         if (!res) {
1116                 res = fuse_direct_io(file, buf, count, ppos, 1);
1117                 if (res > 0)
1118                         fuse_write_update_size(inode, *ppos);
1119         }
1120         mutex_unlock(&inode->i_mutex);
1121
1122         fuse_invalidate_attr(inode);
1123
1124         return res;
1125 }
1126
1127 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1128 {
1129         __free_page(req->pages[0]);
1130         fuse_file_put(req->ff);
1131 }
1132
1133 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1134 {
1135         struct inode *inode = req->inode;
1136         struct fuse_inode *fi = get_fuse_inode(inode);
1137         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1138
1139         list_del(&req->writepages_entry);
1140         dec_bdi_stat(bdi, BDI_WRITEBACK);
1141         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1142         bdi_writeout_inc(bdi);
1143         wake_up(&fi->page_waitq);
1144 }
1145
1146 /* Called under fc->lock, may release and reacquire it */
1147 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1148 __releases(&fc->lock)
1149 __acquires(&fc->lock)
1150 {
1151         struct fuse_inode *fi = get_fuse_inode(req->inode);
1152         loff_t size = i_size_read(req->inode);
1153         struct fuse_write_in *inarg = &req->misc.write.in;
1154
1155         if (!fc->connected)
1156                 goto out_free;
1157
1158         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1159                 inarg->size = PAGE_CACHE_SIZE;
1160         } else if (inarg->offset < size) {
1161                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1162         } else {
1163                 /* Got truncated off completely */
1164                 goto out_free;
1165         }
1166
1167         req->in.args[1].size = inarg->size;
1168         fi->writectr++;
1169         fuse_request_send_background_locked(fc, req);
1170         return;
1171
1172  out_free:
1173         fuse_writepage_finish(fc, req);
1174         spin_unlock(&fc->lock);
1175         fuse_writepage_free(fc, req);
1176         fuse_put_request(fc, req);
1177         spin_lock(&fc->lock);
1178 }
1179
1180 /*
1181  * If fi->writectr is positive (no truncate or fsync going on) send
1182  * all queued writepage requests.
1183  *
1184  * Called with fc->lock
1185  */
1186 void fuse_flush_writepages(struct inode *inode)
1187 __releases(&fc->lock)
1188 __acquires(&fc->lock)
1189 {
1190         struct fuse_conn *fc = get_fuse_conn(inode);
1191         struct fuse_inode *fi = get_fuse_inode(inode);
1192         struct fuse_req *req;
1193
1194         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1195                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1196                 list_del_init(&req->list);
1197                 fuse_send_writepage(fc, req);
1198         }
1199 }
1200
1201 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1202 {
1203         struct inode *inode = req->inode;
1204         struct fuse_inode *fi = get_fuse_inode(inode);
1205
1206         mapping_set_error(inode->i_mapping, req->out.h.error);
1207         spin_lock(&fc->lock);
1208         fi->writectr--;
1209         fuse_writepage_finish(fc, req);
1210         spin_unlock(&fc->lock);
1211         fuse_writepage_free(fc, req);
1212 }
1213
1214 static int fuse_writepage_locked(struct page *page)
1215 {
1216         struct address_space *mapping = page->mapping;
1217         struct inode *inode = mapping->host;
1218         struct fuse_conn *fc = get_fuse_conn(inode);
1219         struct fuse_inode *fi = get_fuse_inode(inode);
1220         struct fuse_req *req;
1221         struct fuse_file *ff;
1222         struct page *tmp_page;
1223
1224         set_page_writeback(page);
1225
1226         req = fuse_request_alloc_nofs();
1227         if (!req)
1228                 goto err;
1229
1230         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1231         if (!tmp_page)
1232                 goto err_free;
1233
1234         spin_lock(&fc->lock);
1235         BUG_ON(list_empty(&fi->write_files));
1236         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1237         req->ff = fuse_file_get(ff);
1238         spin_unlock(&fc->lock);
1239
1240         fuse_write_fill(req, ff, page_offset(page), 0);
1241
1242         copy_highpage(tmp_page, page);
1243         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1244         req->in.argpages = 1;
1245         req->num_pages = 1;
1246         req->pages[0] = tmp_page;
1247         req->page_offset = 0;
1248         req->end = fuse_writepage_end;
1249         req->inode = inode;
1250
1251         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1252         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1253         end_page_writeback(page);
1254
1255         spin_lock(&fc->lock);
1256         list_add(&req->writepages_entry, &fi->writepages);
1257         list_add_tail(&req->list, &fi->queued_writes);
1258         fuse_flush_writepages(inode);
1259         spin_unlock(&fc->lock);
1260
1261         return 0;
1262
1263 err_free:
1264         fuse_request_free(req);
1265 err:
1266         end_page_writeback(page);
1267         return -ENOMEM;
1268 }
1269
1270 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1271 {
1272         int err;
1273
1274         err = fuse_writepage_locked(page);
1275         unlock_page(page);
1276
1277         return err;
1278 }
1279
1280 static int fuse_launder_page(struct page *page)
1281 {
1282         int err = 0;
1283         if (clear_page_dirty_for_io(page)) {
1284                 struct inode *inode = page->mapping->host;
1285                 err = fuse_writepage_locked(page);
1286                 if (!err)
1287                         fuse_wait_on_page_writeback(inode, page->index);
1288         }
1289         return err;
1290 }
1291
1292 /*
1293  * Write back dirty pages now, because there may not be any suitable
1294  * open files later
1295  */
1296 static void fuse_vma_close(struct vm_area_struct *vma)
1297 {
1298         filemap_write_and_wait(vma->vm_file->f_mapping);
1299 }
1300
1301 /*
1302  * Wait for writeback against this page to complete before allowing it
1303  * to be marked dirty again, and hence written back again, possibly
1304  * before the previous writepage completed.
1305  *
1306  * Block here, instead of in ->writepage(), so that the userspace fs
1307  * can only block processes actually operating on the filesystem.
1308  *
1309  * Otherwise unprivileged userspace fs would be able to block
1310  * unrelated:
1311  *
1312  * - page migration
1313  * - sync(2)
1314  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1315  */
1316 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1317 {
1318         struct page *page = vmf->page;
1319         /*
1320          * Don't use page->mapping as it may become NULL from a
1321          * concurrent truncate.
1322          */
1323         struct inode *inode = vma->vm_file->f_mapping->host;
1324
1325         fuse_wait_on_page_writeback(inode, page->index);
1326         return 0;
1327 }
1328
1329 static const struct vm_operations_struct fuse_file_vm_ops = {
1330         .close          = fuse_vma_close,
1331         .fault          = filemap_fault,
1332         .page_mkwrite   = fuse_page_mkwrite,
1333 };
1334
1335 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1336 {
1337         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1338                 struct inode *inode = file->f_dentry->d_inode;
1339                 struct fuse_conn *fc = get_fuse_conn(inode);
1340                 struct fuse_inode *fi = get_fuse_inode(inode);
1341                 struct fuse_file *ff = file->private_data;
1342                 /*
1343                  * file may be written through mmap, so chain it onto the
1344                  * inodes's write_file list
1345                  */
1346                 spin_lock(&fc->lock);
1347                 if (list_empty(&ff->write_entry))
1348                         list_add(&ff->write_entry, &fi->write_files);
1349                 spin_unlock(&fc->lock);
1350         }
1351         file_accessed(file);
1352         vma->vm_ops = &fuse_file_vm_ops;
1353         return 0;
1354 }
1355
1356 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1357 {
1358         /* Can't provide the coherency needed for MAP_SHARED */
1359         if (vma->vm_flags & VM_MAYSHARE)
1360                 return -ENODEV;
1361
1362         invalidate_inode_pages2(file->f_mapping);
1363
1364         return generic_file_mmap(file, vma);
1365 }
1366
1367 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1368                                   struct file_lock *fl)
1369 {
1370         switch (ffl->type) {
1371         case F_UNLCK:
1372                 break;
1373
1374         case F_RDLCK:
1375         case F_WRLCK:
1376                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1377                     ffl->end < ffl->start)
1378                         return -EIO;
1379
1380                 fl->fl_start = ffl->start;
1381                 fl->fl_end = ffl->end;
1382                 fl->fl_pid = ffl->pid;
1383                 break;
1384
1385         default:
1386                 return -EIO;
1387         }
1388         fl->fl_type = ffl->type;
1389         return 0;
1390 }
1391
1392 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1393                          const struct file_lock *fl, int opcode, pid_t pid,
1394                          int flock)
1395 {
1396         struct inode *inode = file->f_path.dentry->d_inode;
1397         struct fuse_conn *fc = get_fuse_conn(inode);
1398         struct fuse_file *ff = file->private_data;
1399         struct fuse_lk_in *arg = &req->misc.lk_in;
1400
1401         arg->fh = ff->fh;
1402         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1403         arg->lk.start = fl->fl_start;
1404         arg->lk.end = fl->fl_end;
1405         arg->lk.type = fl->fl_type;
1406         arg->lk.pid = pid;
1407         if (flock)
1408                 arg->lk_flags |= FUSE_LK_FLOCK;
1409         req->in.h.opcode = opcode;
1410         req->in.h.nodeid = get_node_id(inode);
1411         req->in.numargs = 1;
1412         req->in.args[0].size = sizeof(*arg);
1413         req->in.args[0].value = arg;
1414 }
1415
1416 static int fuse_getlk(struct file *file, struct file_lock *fl)
1417 {
1418         struct inode *inode = file->f_path.dentry->d_inode;
1419         struct fuse_conn *fc = get_fuse_conn(inode);
1420         struct fuse_req *req;
1421         struct fuse_lk_out outarg;
1422         int err;
1423
1424         req = fuse_get_req(fc);
1425         if (IS_ERR(req))
1426                 return PTR_ERR(req);
1427
1428         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1429         req->out.numargs = 1;
1430         req->out.args[0].size = sizeof(outarg);
1431         req->out.args[0].value = &outarg;
1432         fuse_request_send(fc, req);
1433         err = req->out.h.error;
1434         fuse_put_request(fc, req);
1435         if (!err)
1436                 err = convert_fuse_file_lock(&outarg.lk, fl);
1437
1438         return err;
1439 }
1440
1441 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1442 {
1443         struct inode *inode = file->f_path.dentry->d_inode;
1444         struct fuse_conn *fc = get_fuse_conn(inode);
1445         struct fuse_req *req;
1446         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1447         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1448         int err;
1449
1450         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1451                 /* NLM needs asynchronous locks, which we don't support yet */
1452                 return -ENOLCK;
1453         }
1454
1455         /* Unlock on close is handled by the flush method */
1456         if (fl->fl_flags & FL_CLOSE)
1457                 return 0;
1458
1459         req = fuse_get_req(fc);
1460         if (IS_ERR(req))
1461                 return PTR_ERR(req);
1462
1463         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1464         fuse_request_send(fc, req);
1465         err = req->out.h.error;
1466         /* locking is restartable */
1467         if (err == -EINTR)
1468                 err = -ERESTARTSYS;
1469         fuse_put_request(fc, req);
1470         return err;
1471 }
1472
1473 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1474 {
1475         struct inode *inode = file->f_path.dentry->d_inode;
1476         struct fuse_conn *fc = get_fuse_conn(inode);
1477         int err;
1478
1479         if (cmd == F_CANCELLK) {
1480                 err = 0;
1481         } else if (cmd == F_GETLK) {
1482                 if (fc->no_lock) {
1483                         posix_test_lock(file, fl);
1484                         err = 0;
1485                 } else
1486                         err = fuse_getlk(file, fl);
1487         } else {
1488                 if (fc->no_lock)
1489                         err = posix_lock_file(file, fl, NULL);
1490                 else
1491                         err = fuse_setlk(file, fl, 0);
1492         }
1493         return err;
1494 }
1495
1496 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1497 {
1498         struct inode *inode = file->f_path.dentry->d_inode;
1499         struct fuse_conn *fc = get_fuse_conn(inode);
1500         int err;
1501
1502         if (fc->no_lock) {
1503                 err = flock_lock_file_wait(file, fl);
1504         } else {
1505                 /* emulate flock with POSIX locks */
1506                 fl->fl_owner = (fl_owner_t) file;
1507                 err = fuse_setlk(file, fl, 1);
1508         }
1509
1510         return err;
1511 }
1512
1513 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1514 {
1515         struct inode *inode = mapping->host;
1516         struct fuse_conn *fc = get_fuse_conn(inode);
1517         struct fuse_req *req;
1518         struct fuse_bmap_in inarg;
1519         struct fuse_bmap_out outarg;
1520         int err;
1521
1522         if (!inode->i_sb->s_bdev || fc->no_bmap)
1523                 return 0;
1524
1525         req = fuse_get_req(fc);
1526         if (IS_ERR(req))
1527                 return 0;
1528
1529         memset(&inarg, 0, sizeof(inarg));
1530         inarg.block = block;
1531         inarg.blocksize = inode->i_sb->s_blocksize;
1532         req->in.h.opcode = FUSE_BMAP;
1533         req->in.h.nodeid = get_node_id(inode);
1534         req->in.numargs = 1;
1535         req->in.args[0].size = sizeof(inarg);
1536         req->in.args[0].value = &inarg;
1537         req->out.numargs = 1;
1538         req->out.args[0].size = sizeof(outarg);
1539         req->out.args[0].value = &outarg;
1540         fuse_request_send(fc, req);
1541         err = req->out.h.error;
1542         fuse_put_request(fc, req);
1543         if (err == -ENOSYS)
1544                 fc->no_bmap = 1;
1545
1546         return err ? 0 : outarg.block;
1547 }
1548
1549 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1550 {
1551         loff_t retval;
1552         struct inode *inode = file->f_path.dentry->d_inode;
1553
1554         mutex_lock(&inode->i_mutex);
1555         switch (origin) {
1556         case SEEK_END:
1557                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1558                 if (retval)
1559                         goto exit;
1560                 offset += i_size_read(inode);
1561                 break;
1562         case SEEK_CUR:
1563                 offset += file->f_pos;
1564         }
1565         retval = -EINVAL;
1566         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1567                 if (offset != file->f_pos) {
1568                         file->f_pos = offset;
1569                         file->f_version = 0;
1570                 }
1571                 retval = offset;
1572         }
1573 exit:
1574         mutex_unlock(&inode->i_mutex);
1575         return retval;
1576 }
1577
1578 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1579                         unsigned int nr_segs, size_t bytes, bool to_user)
1580 {
1581         struct iov_iter ii;
1582         int page_idx = 0;
1583
1584         if (!bytes)
1585                 return 0;
1586
1587         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1588
1589         while (iov_iter_count(&ii)) {
1590                 struct page *page = pages[page_idx++];
1591                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1592                 void *kaddr;
1593
1594                 kaddr = kmap(page);
1595
1596                 while (todo) {
1597                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1598                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1599                         size_t copy = min(todo, iov_len);
1600                         size_t left;
1601
1602                         if (!to_user)
1603                                 left = copy_from_user(kaddr, uaddr, copy);
1604                         else
1605                                 left = copy_to_user(uaddr, kaddr, copy);
1606
1607                         if (unlikely(left))
1608                                 return -EFAULT;
1609
1610                         iov_iter_advance(&ii, copy);
1611                         todo -= copy;
1612                         kaddr += copy;
1613                 }
1614
1615                 kunmap(page);
1616         }
1617
1618         return 0;
1619 }
1620
1621 /*
1622  * For ioctls, there is no generic way to determine how much memory
1623  * needs to be read and/or written.  Furthermore, ioctls are allowed
1624  * to dereference the passed pointer, so the parameter requires deep
1625  * copying but FUSE has no idea whatsoever about what to copy in or
1626  * out.
1627  *
1628  * This is solved by allowing FUSE server to retry ioctl with
1629  * necessary in/out iovecs.  Let's assume the ioctl implementation
1630  * needs to read in the following structure.
1631  *
1632  * struct a {
1633  *      char    *buf;
1634  *      size_t  buflen;
1635  * }
1636  *
1637  * On the first callout to FUSE server, inarg->in_size and
1638  * inarg->out_size will be NULL; then, the server completes the ioctl
1639  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1640  * the actual iov array to
1641  *
1642  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1643  *
1644  * which tells FUSE to copy in the requested area and retry the ioctl.
1645  * On the second round, the server has access to the structure and
1646  * from that it can tell what to look for next, so on the invocation,
1647  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1648  *
1649  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1650  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1651  *
1652  * FUSE will copy both struct a and the pointed buffer from the
1653  * process doing the ioctl and retry ioctl with both struct a and the
1654  * buffer.
1655  *
1656  * This time, FUSE server has everything it needs and completes ioctl
1657  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1658  *
1659  * Copying data out works the same way.
1660  *
1661  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1662  * automatically initializes in and out iovs by decoding @cmd with
1663  * _IOC_* macros and the server is not allowed to request RETRY.  This
1664  * limits ioctl data transfers to well-formed ioctls and is the forced
1665  * behavior for all FUSE servers.
1666  */
1667 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1668                    unsigned int flags)
1669 {
1670         struct fuse_file *ff = file->private_data;
1671         struct fuse_conn *fc = ff->fc;
1672         struct fuse_ioctl_in inarg = {
1673                 .fh = ff->fh,
1674                 .cmd = cmd,
1675                 .arg = arg,
1676                 .flags = flags
1677         };
1678         struct fuse_ioctl_out outarg;
1679         struct fuse_req *req = NULL;
1680         struct page **pages = NULL;
1681         struct page *iov_page = NULL;
1682         struct iovec *in_iov = NULL, *out_iov = NULL;
1683         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1684         size_t in_size, out_size, transferred;
1685         int err;
1686
1687         /* assume all the iovs returned by client always fits in a page */
1688         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1689
1690         err = -ENOMEM;
1691         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1692         iov_page = alloc_page(GFP_KERNEL);
1693         if (!pages || !iov_page)
1694                 goto out;
1695
1696         /*
1697          * If restricted, initialize IO parameters as encoded in @cmd.
1698          * RETRY from server is not allowed.
1699          */
1700         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1701                 struct iovec *iov = page_address(iov_page);
1702
1703                 iov->iov_base = (void __user *)arg;
1704                 iov->iov_len = _IOC_SIZE(cmd);
1705
1706                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1707                         in_iov = iov;
1708                         in_iovs = 1;
1709                 }
1710
1711                 if (_IOC_DIR(cmd) & _IOC_READ) {
1712                         out_iov = iov;
1713                         out_iovs = 1;
1714                 }
1715         }
1716
1717  retry:
1718         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1719         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1720
1721         /*
1722          * Out data can be used either for actual out data or iovs,
1723          * make sure there always is at least one page.
1724          */
1725         out_size = max_t(size_t, out_size, PAGE_SIZE);
1726         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1727
1728         /* make sure there are enough buffer pages and init request with them */
1729         err = -ENOMEM;
1730         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1731                 goto out;
1732         while (num_pages < max_pages) {
1733                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1734                 if (!pages[num_pages])
1735                         goto out;
1736                 num_pages++;
1737         }
1738
1739         req = fuse_get_req(fc);
1740         if (IS_ERR(req)) {
1741                 err = PTR_ERR(req);
1742                 req = NULL;
1743                 goto out;
1744         }
1745         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1746         req->num_pages = num_pages;
1747
1748         /* okay, let's send it to the client */
1749         req->in.h.opcode = FUSE_IOCTL;
1750         req->in.h.nodeid = ff->nodeid;
1751         req->in.numargs = 1;
1752         req->in.args[0].size = sizeof(inarg);
1753         req->in.args[0].value = &inarg;
1754         if (in_size) {
1755                 req->in.numargs++;
1756                 req->in.args[1].size = in_size;
1757                 req->in.argpages = 1;
1758
1759                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1760                                            false);
1761                 if (err)
1762                         goto out;
1763         }
1764
1765         req->out.numargs = 2;
1766         req->out.args[0].size = sizeof(outarg);
1767         req->out.args[0].value = &outarg;
1768         req->out.args[1].size = out_size;
1769         req->out.argpages = 1;
1770         req->out.argvar = 1;
1771
1772         fuse_request_send(fc, req);
1773         err = req->out.h.error;
1774         transferred = req->out.args[1].size;
1775         fuse_put_request(fc, req);
1776         req = NULL;
1777         if (err)
1778                 goto out;
1779
1780         /* did it ask for retry? */
1781         if (outarg.flags & FUSE_IOCTL_RETRY) {
1782                 char *vaddr;
1783
1784                 /* no retry if in restricted mode */
1785                 err = -EIO;
1786                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1787                         goto out;
1788
1789                 in_iovs = outarg.in_iovs;
1790                 out_iovs = outarg.out_iovs;
1791
1792                 /*
1793                  * Make sure things are in boundary, separate checks
1794                  * are to protect against overflow.
1795                  */
1796                 err = -ENOMEM;
1797                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1798                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1799                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1800                         goto out;
1801
1802                 err = -EIO;
1803                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1804                         goto out;
1805
1806                 /* okay, copy in iovs and retry */
1807                 vaddr = kmap_atomic(pages[0], KM_USER0);
1808                 memcpy(page_address(iov_page), vaddr, transferred);
1809                 kunmap_atomic(vaddr, KM_USER0);
1810
1811                 in_iov = page_address(iov_page);
1812                 out_iov = in_iov + in_iovs;
1813
1814                 goto retry;
1815         }
1816
1817         err = -EIO;
1818         if (transferred > inarg.out_size)
1819                 goto out;
1820
1821         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1822  out:
1823         if (req)
1824                 fuse_put_request(fc, req);
1825         if (iov_page)
1826                 __free_page(iov_page);
1827         while (num_pages)
1828                 __free_page(pages[--num_pages]);
1829         kfree(pages);
1830
1831         return err ? err : outarg.result;
1832 }
1833 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1834
1835 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1836                                    unsigned long arg, unsigned int flags)
1837 {
1838         struct inode *inode = file->f_dentry->d_inode;
1839         struct fuse_conn *fc = get_fuse_conn(inode);
1840
1841         if (!fuse_allow_task(fc, current))
1842                 return -EACCES;
1843
1844         if (is_bad_inode(inode))
1845                 return -EIO;
1846
1847         return fuse_do_ioctl(file, cmd, arg, flags);
1848 }
1849
1850 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1851                             unsigned long arg)
1852 {
1853         return fuse_file_ioctl_common(file, cmd, arg, 0);
1854 }
1855
1856 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1857                                    unsigned long arg)
1858 {
1859         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1860 }
1861
1862 /*
1863  * All files which have been polled are linked to RB tree
1864  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1865  * find the matching one.
1866  */
1867 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1868                                               struct rb_node **parent_out)
1869 {
1870         struct rb_node **link = &fc->polled_files.rb_node;
1871         struct rb_node *last = NULL;
1872
1873         while (*link) {
1874                 struct fuse_file *ff;
1875
1876                 last = *link;
1877                 ff = rb_entry(last, struct fuse_file, polled_node);
1878
1879                 if (kh < ff->kh)
1880                         link = &last->rb_left;
1881                 else if (kh > ff->kh)
1882                         link = &last->rb_right;
1883                 else
1884                         return link;
1885         }
1886
1887         if (parent_out)
1888                 *parent_out = last;
1889         return link;
1890 }
1891
1892 /*
1893  * The file is about to be polled.  Make sure it's on the polled_files
1894  * RB tree.  Note that files once added to the polled_files tree are
1895  * not removed before the file is released.  This is because a file
1896  * polled once is likely to be polled again.
1897  */
1898 static void fuse_register_polled_file(struct fuse_conn *fc,
1899                                       struct fuse_file *ff)
1900 {
1901         spin_lock(&fc->lock);
1902         if (RB_EMPTY_NODE(&ff->polled_node)) {
1903                 struct rb_node **link, *parent;
1904
1905                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1906                 BUG_ON(*link);
1907                 rb_link_node(&ff->polled_node, parent, link);
1908                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1909         }
1910         spin_unlock(&fc->lock);
1911 }
1912
1913 unsigned fuse_file_poll(struct file *file, poll_table *wait)
1914 {
1915         struct fuse_file *ff = file->private_data;
1916         struct fuse_conn *fc = ff->fc;
1917         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1918         struct fuse_poll_out outarg;
1919         struct fuse_req *req;
1920         int err;
1921
1922         if (fc->no_poll)
1923                 return DEFAULT_POLLMASK;
1924
1925         poll_wait(file, &ff->poll_wait, wait);
1926
1927         /*
1928          * Ask for notification iff there's someone waiting for it.
1929          * The client may ignore the flag and always notify.
1930          */
1931         if (waitqueue_active(&ff->poll_wait)) {
1932                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1933                 fuse_register_polled_file(fc, ff);
1934         }
1935
1936         req = fuse_get_req(fc);
1937         if (IS_ERR(req))
1938                 return POLLERR;
1939
1940         req->in.h.opcode = FUSE_POLL;
1941         req->in.h.nodeid = ff->nodeid;
1942         req->in.numargs = 1;
1943         req->in.args[0].size = sizeof(inarg);
1944         req->in.args[0].value = &inarg;
1945         req->out.numargs = 1;
1946         req->out.args[0].size = sizeof(outarg);
1947         req->out.args[0].value = &outarg;
1948         fuse_request_send(fc, req);
1949         err = req->out.h.error;
1950         fuse_put_request(fc, req);
1951
1952         if (!err)
1953                 return outarg.revents;
1954         if (err == -ENOSYS) {
1955                 fc->no_poll = 1;
1956                 return DEFAULT_POLLMASK;
1957         }
1958         return POLLERR;
1959 }
1960 EXPORT_SYMBOL_GPL(fuse_file_poll);
1961
1962 /*
1963  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1964  * wakes up the poll waiters.
1965  */
1966 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1967                             struct fuse_notify_poll_wakeup_out *outarg)
1968 {
1969         u64 kh = outarg->kh;
1970         struct rb_node **link;
1971
1972         spin_lock(&fc->lock);
1973
1974         link = fuse_find_polled_node(fc, kh, NULL);
1975         if (*link) {
1976                 struct fuse_file *ff;
1977
1978                 ff = rb_entry(*link, struct fuse_file, polled_node);
1979                 wake_up_interruptible_sync(&ff->poll_wait);
1980         }
1981
1982         spin_unlock(&fc->lock);
1983         return 0;
1984 }
1985
1986 static const struct file_operations fuse_file_operations = {
1987         .llseek         = fuse_file_llseek,
1988         .read           = do_sync_read,
1989         .aio_read       = fuse_file_aio_read,
1990         .write          = do_sync_write,
1991         .aio_write      = fuse_file_aio_write,
1992         .mmap           = fuse_file_mmap,
1993         .open           = fuse_open,
1994         .flush          = fuse_flush,
1995         .release        = fuse_release,
1996         .fsync          = fuse_fsync,
1997         .lock           = fuse_file_lock,
1998         .flock          = fuse_file_flock,
1999         .splice_read    = generic_file_splice_read,
2000         .unlocked_ioctl = fuse_file_ioctl,
2001         .compat_ioctl   = fuse_file_compat_ioctl,
2002         .poll           = fuse_file_poll,
2003 };
2004
2005 static const struct file_operations fuse_direct_io_file_operations = {
2006         .llseek         = fuse_file_llseek,
2007         .read           = fuse_direct_read,
2008         .write          = fuse_direct_write,
2009         .mmap           = fuse_direct_mmap,
2010         .open           = fuse_open,
2011         .flush          = fuse_flush,
2012         .release        = fuse_release,
2013         .fsync          = fuse_fsync,
2014         .lock           = fuse_file_lock,
2015         .flock          = fuse_file_flock,
2016         .unlocked_ioctl = fuse_file_ioctl,
2017         .compat_ioctl   = fuse_file_compat_ioctl,
2018         .poll           = fuse_file_poll,
2019         /* no splice_read */
2020 };
2021
2022 static const struct address_space_operations fuse_file_aops  = {
2023         .readpage       = fuse_readpage,
2024         .writepage      = fuse_writepage,
2025         .launder_page   = fuse_launder_page,
2026         .write_begin    = fuse_write_begin,
2027         .write_end      = fuse_write_end,
2028         .readpages      = fuse_readpages,
2029         .set_page_dirty = __set_page_dirty_nobuffers,
2030         .bmap           = fuse_bmap,
2031 };
2032
2033 void fuse_init_file_inode(struct inode *inode)
2034 {
2035         inode->i_fop = &fuse_file_operations;
2036         inode->i_data.a_ops = &fuse_file_aops;
2037 }