SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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, int datasync, int isdir)
 355 {
 356         struct inode *inode = file->f_mapping->host;
 357         struct fuse_conn *fc = get_fuse_conn(inode);
 358         struct fuse_file *ff = file->private_data;
 359         struct fuse_req *req;
 360         struct fuse_fsync_in inarg;
 361         int err;
 362
 363         if (is_bad_inode(inode))
 364                 return -EIO;
 365
 366         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 367                 return 0;
 368
 369         /*
 370          * Start writeback against all dirty pages of the inode, then
 371          * wait for all outstanding writes, before sending the FSYNC
 372          * request.
 373          */
 374         err = write_inode_now(inode, 0);
 375         if (err)
 376                 return err;
 377
 378         fuse_sync_writes(inode);
 379
 380         req = fuse_get_req(fc);
 381         if (IS_ERR(req))
 382                 return PTR_ERR(req);
 383
 384         memset(&inarg, 0, sizeof(inarg));
 385         inarg.fh = ff->fh;
 386         inarg.fsync_flags = datasync ? 1 : 0;
 387         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 388         req->in.h.nodeid = get_node_id(inode);
 389         req->in.numargs = 1;
 390         req->in.args[0].size = sizeof(inarg);
 391         req->in.args[0].value = &inarg;
 392         fuse_request_send(fc, req);
 393         err = req->out.h.error;
 394         fuse_put_request(fc, req);
 395         if (err == -ENOSYS) {
 396                 if (isdir)
 397                         fc->no_fsyncdir = 1;
 398                 else
 399                         fc->no_fsync = 1;
 400                 err = 0;
 401         }
 402         return err;
 403 }
 404
 405 static int fuse_fsync(struct file *file, int datasync)
 406 {
 407         return fuse_fsync_common(file, datasync, 0);
 408 }
 409
 410 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 411                     size_t count, int opcode)
 412 {
 413         struct fuse_read_in *inarg = &req->misc.read.in;
 414         struct fuse_file *ff = file->private_data;
 415
 416         inarg->fh = ff->fh;
 417         inarg->offset = pos;
 418         inarg->size = count;
 419         inarg->flags = file->f_flags;
 420         req->in.h.opcode = opcode;
 421         req->in.h.nodeid = ff->nodeid;
 422         req->in.numargs = 1;
 423         req->in.args[0].size = sizeof(struct fuse_read_in);
 424         req->in.args[0].value = inarg;
 425         req->out.argvar = 1;
 426         req->out.numargs = 1;
 427         req->out.args[0].size = count;
 428 }
 429
 430 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 431                              loff_t pos, size_t count, fl_owner_t owner)
 432 {
 433         struct fuse_file *ff = file->private_data;
 434         struct fuse_conn *fc = ff->fc;
 435
 436         fuse_read_fill(req, file, pos, count, FUSE_READ);
 437         if (owner != NULL) {
 438                 struct fuse_read_in *inarg = &req->misc.read.in;
 439
 440                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 441                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 442         }
 443         fuse_request_send(fc, req);
 444         return req->out.args[0].size;
 445 }
 446
 447 static void fuse_read_update_size(struct inode *inode, loff_t size,
 448                                   u64 attr_ver)
 449 {
 450         struct fuse_conn *fc = get_fuse_conn(inode);
 451         struct fuse_inode *fi = get_fuse_inode(inode);
 452
 453         spin_lock(&fc->lock);
 454         if (attr_ver == fi->attr_version && size < inode->i_size) {
 455                 fi->attr_version = ++fc->attr_version;
 456                 i_size_write(inode, size);
 457         }
 458         spin_unlock(&fc->lock);
 459 }
 460
 461 static int fuse_readpage(struct file *file, struct page *page)
 462 {
 463         struct inode *inode = page->mapping->host;
 464         struct fuse_conn *fc = get_fuse_conn(inode);
 465         struct fuse_req *req;
 466         size_t num_read;
 467         loff_t pos = page_offset(page);
 468         size_t count = PAGE_CACHE_SIZE;
 469         u64 attr_ver;
 470         int err;
 471
 472         err = -EIO;
 473         if (is_bad_inode(inode))
 474                 goto out;
 475
 476         /*
 477          * Page writeback can extend beyond the liftime of the
 478          * page-cache page, so make sure we read a properly synced
 479          * page.
 480          */
 481         fuse_wait_on_page_writeback(inode, page->index);
 482
 483         req = fuse_get_req(fc);
 484         err = PTR_ERR(req);
 485         if (IS_ERR(req))
 486                 goto out;
 487
 488         attr_ver = fuse_get_attr_version(fc);
 489
 490         req->out.page_zeroing = 1;
 491         req->out.argpages = 1;
 492         req->num_pages = 1;
 493         req->pages[0] = page;
 494         num_read = fuse_send_read(req, file, pos, count, NULL);
 495         err = req->out.h.error;
 496         fuse_put_request(fc, req);
 497
 498         if (!err) {
 499                 /*
 500                  * Short read means EOF.  If file size is larger, truncate it
 501                  */
 502                 if (num_read < count)
 503                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 504
 505                 SetPageUptodate(page);
 506         }
 507
 508         fuse_invalidate_attr(inode); /* atime changed */
 509  out:
 510         unlock_page(page);
 511         return err;
 512 }
 513
 514 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 515 {
 516         int i;
 517         size_t count = req->misc.read.in.size;
 518         size_t num_read = req->out.args[0].size;
 519         struct inode *inode = req->pages[0]->mapping->host;
 520
 521         /*
 522          * Short read means EOF.  If file size is larger, truncate it
 523          */
 524         if (!req->out.h.error && num_read < count) {
 525                 loff_t pos = page_offset(req->pages[0]) + num_read;
 526                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
 527         }
 528
 529         fuse_invalidate_attr(inode); /* atime changed */
 530
 531         for (i = 0; i < req->num_pages; i++) {
 532                 struct page *page = req->pages[i];
 533                 if (!req->out.h.error)
 534                         SetPageUptodate(page);
 535                 else
 536                         SetPageError(page);
 537                 unlock_page(page);
 538         }
 539         if (req->ff)
 540                 fuse_file_put(req->ff);
 541 }
 542
 543 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 544 {
 545         struct fuse_file *ff = file->private_data;
 546         struct fuse_conn *fc = ff->fc;
 547         loff_t pos = page_offset(req->pages[0]);
 548         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 549
 550         req->out.argpages = 1;
 551         req->out.page_zeroing = 1;
 552         fuse_read_fill(req, file, pos, count, FUSE_READ);
 553         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 554         if (fc->async_read) {
 555                 req->ff = fuse_file_get(ff);
 556                 req->end = fuse_readpages_end;
 557                 fuse_request_send_background(fc, req);
 558         } else {
 559                 fuse_request_send(fc, req);
 560                 fuse_readpages_end(fc, req);
 561                 fuse_put_request(fc, req);
 562         }
 563 }
 564
 565 struct fuse_fill_data {
 566         struct fuse_req *req;
 567         struct file *file;
 568         struct inode *inode;
 569 };
 570
 571 static int fuse_readpages_fill(void *_data, struct page *page)
 572 {
 573         struct fuse_fill_data *data = _data;
 574         struct fuse_req *req = data->req;
 575         struct inode *inode = data->inode;
 576         struct fuse_conn *fc = get_fuse_conn(inode);
 577
 578         fuse_wait_on_page_writeback(inode, page->index);
 579
 580         if (req->num_pages &&
 581             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 582              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 583              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 584                 fuse_send_readpages(req, data->file);
 585                 data->req = req = fuse_get_req(fc);
 586                 if (IS_ERR(req)) {
 587                         unlock_page(page);
 588                         return PTR_ERR(req);
 589                 }
 590         }
 591         req->pages[req->num_pages] = page;
 592         req->num_pages++;
 593         return 0;
 594 }
 595
 596 static int fuse_readpages(struct file *file, struct address_space *mapping,
 597                           struct list_head *pages, unsigned nr_pages)
 598 {
 599         struct inode *inode = mapping->host;
 600         struct fuse_conn *fc = get_fuse_conn(inode);
 601         struct fuse_fill_data data;
 602         int err;
 603
 604         err = -EIO;
 605         if (is_bad_inode(inode))
 606                 goto out;
 607
 608         data.file = file;
 609         data.inode = inode;
 610         data.req = fuse_get_req(fc);
 611         err = PTR_ERR(data.req);
 612         if (IS_ERR(data.req))
 613                 goto out;
 614
 615         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 616         if (!err) {
 617                 if (data.req->num_pages)
 618                         fuse_send_readpages(data.req, file);
 619                 else
 620                         fuse_put_request(fc, data.req);
 621         }
 622 out:
 623         return err;
 624 }
 625
 626 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 627                                   unsigned long nr_segs, loff_t pos)
 628 {
 629         struct inode *inode = iocb->ki_filp->f_mapping->host;
 630
 631         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 632                 int err;
 633                 /*
 634                  * If trying to read past EOF, make sure the i_size
 635                  * attribute is up-to-date.
 636                  */
 637                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 638                 if (err)
 639                         return err;
 640         }
 641
 642         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 643 }
 644
 645 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 646                             loff_t pos, size_t count)
 647 {
 648         struct fuse_write_in *inarg = &req->misc.write.in;
 649         struct fuse_write_out *outarg = &req->misc.write.out;
 650
 651         inarg->fh = ff->fh;
 652         inarg->offset = pos;
 653         inarg->size = count;
 654         req->in.h.opcode = FUSE_WRITE;
 655         req->in.h.nodeid = ff->nodeid;
 656         req->in.numargs = 2;
 657         if (ff->fc->minor < 9)
 658                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 659         else
 660                 req->in.args[0].size = sizeof(struct fuse_write_in);
 661         req->in.args[0].value = inarg;
 662         req->in.args[1].size = count;
 663         req->out.numargs = 1;
 664         req->out.args[0].size = sizeof(struct fuse_write_out);
 665         req->out.args[0].value = outarg;
 666 }
 667
 668 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 669                               loff_t pos, size_t count, fl_owner_t owner)
 670 {
 671         struct fuse_file *ff = file->private_data;
 672         struct fuse_conn *fc = ff->fc;
 673         struct fuse_write_in *inarg = &req->misc.write.in;
 674
 675         fuse_write_fill(req, ff, pos, count);
 676         inarg->flags = file->f_flags;
 677         if (owner != NULL) {
 678                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 679                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 680         }
 681         fuse_request_send(fc, req);
 682         return req->misc.write.out.size;
 683 }
 684
 685 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 686                         loff_t pos, unsigned len, unsigned flags,
 687                         struct page **pagep, void **fsdata)
 688 {
 689         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 690
 691         *pagep = grab_cache_page_write_begin(mapping, index, flags);
 692         if (!*pagep)
 693                 return -ENOMEM;
 694         return 0;
 695 }
 696
 697 static void fuse_write_update_size(struct inode *inode, loff_t pos)
 698 {
 699         struct fuse_conn *fc = get_fuse_conn(inode);
 700         struct fuse_inode *fi = get_fuse_inode(inode);
 701
 702         spin_lock(&fc->lock);
 703         fi->attr_version = ++fc->attr_version;
 704         if (pos > inode->i_size)
 705                 i_size_write(inode, pos);
 706         spin_unlock(&fc->lock);
 707 }
 708
 709 static int fuse_buffered_write(struct file *file, struct inode *inode,
 710                                loff_t pos, unsigned count, struct page *page)
 711 {
 712         int err;
 713         size_t nres;
 714         struct fuse_conn *fc = get_fuse_conn(inode);
 715         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 716         struct fuse_req *req;
 717
 718         if (is_bad_inode(inode))
 719                 return -EIO;
 720
 721         /*
 722          * Make sure writepages on the same page are not mixed up with
 723          * plain writes.
 724          */
 725         fuse_wait_on_page_writeback(inode, page->index);
 726
 727         req = fuse_get_req(fc);
 728         if (IS_ERR(req))
 729                 return PTR_ERR(req);
 730
 731         req->in.argpages = 1;
 732         req->num_pages = 1;
 733         req->pages[0] = page;
 734         req->page_offset = offset;
 735         nres = fuse_send_write(req, file, pos, count, NULL);
 736         err = req->out.h.error;
 737         fuse_put_request(fc, req);
 738         if (!err && !nres)
 739                 err = -EIO;
 740         if (!err) {
 741                 pos += nres;
 742                 fuse_write_update_size(inode, pos);
 743                 if (count == PAGE_CACHE_SIZE)
 744                         SetPageUptodate(page);
 745         }
 746         fuse_invalidate_attr(inode);
 747         return err ? err : nres;
 748 }
 749
 750 static int fuse_write_end(struct file *file, struct address_space *mapping,
 751                         loff_t pos, unsigned len, unsigned copied,
 752                         struct page *page, void *fsdata)
 753 {
 754         struct inode *inode = mapping->host;
 755         int res = 0;
 756
 757         if (copied)
 758                 res = fuse_buffered_write(file, inode, pos, copied, page);
 759
 760         unlock_page(page);
 761         page_cache_release(page);
 762         return res;
 763 }
 764
 765 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 766                                     struct inode *inode, loff_t pos,
 767                                     size_t count)
 768 {
 769         size_t res;
 770         unsigned offset;
 771         unsigned i;
 772
 773         for (i = 0; i < req->num_pages; i++)
 774                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 775
 776         res = fuse_send_write(req, file, pos, count, NULL);
 777
 778         offset = req->page_offset;
 779         count = res;
 780         for (i = 0; i < req->num_pages; i++) {
 781                 struct page *page = req->pages[i];
 782
 783                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 784                         SetPageUptodate(page);
 785
 786                 if (count > PAGE_CACHE_SIZE - offset)
 787                         count -= PAGE_CACHE_SIZE - offset;
 788                 else
 789                         count = 0;
 790                 offset = 0;
 791
 792                 unlock_page(page);
 793                 page_cache_release(page);
 794         }
 795
 796         return res;
 797 }
 798
 799 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 800                                struct address_space *mapping,
 801                                struct iov_iter *ii, loff_t pos)
 802 {
 803         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 804         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 805         size_t count = 0;
 806         int err;
 807
 808         req->in.argpages = 1;
 809         req->page_offset = offset;
 810
 811         do {
 812                 size_t tmp;
 813                 struct page *page;
 814                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 815                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 816                                      iov_iter_count(ii));
 817
 818                 bytes = min_t(size_t, bytes, fc->max_write - count);
 819
 820  again:
 821                 err = -EFAULT;
 822                 if (iov_iter_fault_in_readable(ii, bytes))
 823                         break;
 824
 825                 err = -ENOMEM;
 826                 page = grab_cache_page_write_begin(mapping, index, 0);
 827                 if (!page)
 828                         break;
 829
 830                 if (mapping_writably_mapped(mapping))
 831                         flush_dcache_page(page);
 832
 833                 pagefault_disable();
 834                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 835                 pagefault_enable();
 836                 flush_dcache_page(page);
 837
 838                 if (!tmp) {
 839                         unlock_page(page);
 840                         page_cache_release(page);
 841                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 842                         goto again;
 843                 }
 844
 845                 err = 0;
 846                 req->pages[req->num_pages] = page;
 847                 req->num_pages++;
 848
 849                 iov_iter_advance(ii, tmp);
 850                 count += tmp;
 851                 pos += tmp;
 852                 offset += tmp;
 853                 if (offset == PAGE_CACHE_SIZE)
 854                         offset = 0;
 855
 856                 if (!fc->big_writes)
 857                         break;
 858         } while (iov_iter_count(ii) && count < fc->max_write &&
 859                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 860
 861         return count > 0 ? count : err;
 862 }
 863
 864 static ssize_t fuse_perform_write(struct file *file,
 865                                   struct address_space *mapping,
 866                                   struct iov_iter *ii, loff_t pos)
 867 {
 868         struct inode *inode = mapping->host;
 869         struct fuse_conn *fc = get_fuse_conn(inode);
 870         int err = 0;
 871         ssize_t res = 0;
 872
 873         if (is_bad_inode(inode))
 874                 return -EIO;
 875
 876         do {
 877                 struct fuse_req *req;
 878                 ssize_t count;
 879
 880                 req = fuse_get_req(fc);
 881                 if (IS_ERR(req)) {
 882                         err = PTR_ERR(req);
 883                         break;
 884                 }
 885
 886                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 887                 if (count <= 0) {
 888                         err = count;
 889                 } else {
 890                         size_t num_written;
 891
 892                         num_written = fuse_send_write_pages(req, file, inode,
 893                                                             pos, count);
 894                         err = req->out.h.error;
 895                         if (!err) {
 896                                 res += num_written;
 897                                 pos += num_written;
 898
 899                                 /* break out of the loop on short write */
 900                                 if (num_written != count)
 901                                         err = -EIO;
 902                         }
 903                 }
 904                 fuse_put_request(fc, req);
 905         } while (!err && iov_iter_count(ii));
 906
 907         if (res > 0)
 908                 fuse_write_update_size(inode, pos);
 909
 910         fuse_invalidate_attr(inode);
 911
 912         return res > 0 ? res : err;
 913 }
 914
 915 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 916                                    unsigned long nr_segs, loff_t pos)
 917 {
 918         struct file *file = iocb->ki_filp;
 919         struct address_space *mapping = file->f_mapping;
 920         size_t count = 0;
 921         ssize_t written = 0;
 922         struct inode *inode = mapping->host;
 923         ssize_t err;
 924         struct iov_iter i;
 925
 926         WARN_ON(iocb->ki_pos != pos);
 927
 928         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 929         if (err)
 930                 return err;
 931
 932         mutex_lock(&inode->i_mutex);
 933         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 934
 935         /* We can write back this queue in page reclaim */
 936         current->backing_dev_info = mapping->backing_dev_info;
 937
 938         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 939         if (err)
 940                 goto out;
 941
 942         if (count == 0)
 943                 goto out;
 944
 945         err = file_remove_suid(file);
 946         if (err)
 947                 goto out;
 948
 949         file_update_time(file);
 950
 951         iov_iter_init(&i, iov, nr_segs, count, 0);
 952         written = fuse_perform_write(file, mapping, &i, pos);
 953         if (written >= 0)
 954                 iocb->ki_pos = pos + written;
 955
 956 out:
 957         current->backing_dev_info = NULL;
 958         mutex_unlock(&inode->i_mutex);
 959
 960         return written ? written : err;
 961 }
 962
 963 static void fuse_release_user_pages(struct fuse_req *req, int write)
 964 {
 965         unsigned i;
 966
 967         for (i = 0; i < req->num_pages; i++) {
 968                 struct page *page = req->pages[i];
 969                 if (write)
 970                         set_page_dirty_lock(page);
 971                 put_page(page);
 972         }
 973 }
 974
 975 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 976                                size_t *nbytesp, int write)
 977 {
 978         size_t nbytes = *nbytesp;
 979         unsigned long user_addr = (unsigned long) buf;
 980         unsigned offset = user_addr & ~PAGE_MASK;
 981         int npages;
 982
 983         /* Special case for kernel I/O: can copy directly into the buffer */
 984         if (segment_eq(get_fs(), KERNEL_DS)) {
 985                 if (write)
 986                         req->in.args[1].value = (void *) user_addr;
 987                 else
 988                         req->out.args[0].value = (void *) user_addr;
 989
 990                 return 0;
 991         }
 992
 993         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
 994         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 995         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
 996         down_read(&current->mm->mmap_sem);
 997         npages = get_user_pages(current, current->mm, user_addr, npages, !write,
 998                                 0, req->pages, NULL);
 999         up_read(&current->mm->mmap_sem);
1000         if (npages < 0)
1001                 return npages;
1002
1003         req->num_pages = npages;
1004         req->page_offset = offset;
1005
1006         if (write)
1007                 req->in.argpages = 1;
1008         else
1009                 req->out.argpages = 1;
1010
1011         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1012         *nbytesp = min(*nbytesp, nbytes);
1013
1014         return 0;
1015 }
1016
1017 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1018                        size_t count, loff_t *ppos, int write)
1019 {
1020         struct fuse_file *ff = file->private_data;
1021         struct fuse_conn *fc = ff->fc;
1022         size_t nmax = write ? fc->max_write : fc->max_read;
1023         loff_t pos = *ppos;
1024         ssize_t res = 0;
1025         struct fuse_req *req;
1026
1027         req = fuse_get_req(fc);
1028         if (IS_ERR(req))
1029                 return PTR_ERR(req);
1030
1031         while (count) {
1032                 size_t nres;
1033                 fl_owner_t owner = current->files;
1034                 size_t nbytes = min(count, nmax);
1035                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1036                 if (err) {
1037                         res = err;
1038                         break;
1039                 }
1040
1041                 if (write)
1042                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1043                 else
1044                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1045
1046                 fuse_release_user_pages(req, !write);
1047                 if (req->out.h.error) {
1048                         if (!res)
1049                                 res = req->out.h.error;
1050                         break;
1051                 } else if (nres > nbytes) {
1052                         res = -EIO;
1053                         break;
1054                 }
1055                 count -= nres;
1056                 res += nres;
1057                 pos += nres;
1058                 buf += nres;
1059                 if (nres != nbytes)
1060                         break;
1061                 if (count) {
1062                         fuse_put_request(fc, req);
1063                         req = fuse_get_req(fc);
1064                         if (IS_ERR(req))
1065                                 break;
1066                 }
1067         }
1068         if (!IS_ERR(req))
1069                 fuse_put_request(fc, req);
1070         if (res > 0)
1071                 *ppos = pos;
1072
1073         return res;
1074 }
1075 EXPORT_SYMBOL_GPL(fuse_direct_io);
1076
1077 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1078                                      size_t count, loff_t *ppos)
1079 {
1080         ssize_t res;
1081         struct inode *inode = file->f_path.dentry->d_inode;
1082
1083         if (is_bad_inode(inode))
1084                 return -EIO;
1085
1086         res = fuse_direct_io(file, buf, count, ppos, 0);
1087
1088         fuse_invalidate_attr(inode);
1089
1090         return res;
1091 }
1092
1093 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1094                                  size_t count, loff_t *ppos)
1095 {
1096         struct inode *inode = file->f_path.dentry->d_inode;
1097         ssize_t res;
1098
1099         if (is_bad_inode(inode))
1100                 return -EIO;
1101
1102         /* Don't allow parallel writes to the same file */
1103         mutex_lock(&inode->i_mutex);
1104         res = generic_write_checks(file, ppos, &count, 0);
1105         if (!res) {
1106                 res = fuse_direct_io(file, buf, count, ppos, 1);
1107                 if (res > 0)
1108                         fuse_write_update_size(inode, *ppos);
1109         }
1110         mutex_unlock(&inode->i_mutex);
1111
1112         fuse_invalidate_attr(inode);
1113
1114         return res;
1115 }
1116
1117 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1118 {
1119         __free_page(req->pages[0]);
1120         fuse_file_put(req->ff);
1121 }
1122
1123 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1124 {
1125         struct inode *inode = req->inode;
1126         struct fuse_inode *fi = get_fuse_inode(inode);
1127         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1128
1129         list_del(&req->writepages_entry);
1130         dec_bdi_stat(bdi, BDI_WRITEBACK);
1131         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1132         bdi_writeout_inc(bdi);
1133         wake_up(&fi->page_waitq);
1134 }
1135
1136 /* Called under fc->lock, may release and reacquire it */
1137 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1138 __releases(&fc->lock)
1139 __acquires(&fc->lock)
1140 {
1141         struct fuse_inode *fi = get_fuse_inode(req->inode);
1142         loff_t size = i_size_read(req->inode);
1143         struct fuse_write_in *inarg = &req->misc.write.in;
1144
1145         if (!fc->connected)
1146                 goto out_free;
1147
1148         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1149                 inarg->size = PAGE_CACHE_SIZE;
1150         } else if (inarg->offset < size) {
1151                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1152         } else {
1153                 /* Got truncated off completely */
1154                 goto out_free;
1155         }
1156
1157         req->in.args[1].size = inarg->size;
1158         fi->writectr++;
1159         fuse_request_send_background_locked(fc, req);
1160         return;
1161
1162  out_free:
1163         fuse_writepage_finish(fc, req);
1164         spin_unlock(&fc->lock);
1165         fuse_writepage_free(fc, req);
1166         fuse_put_request(fc, req);
1167         spin_lock(&fc->lock);
1168 }
1169
1170 /*
1171  * If fi->writectr is positive (no truncate or fsync going on) send
1172  * all queued writepage requests.
1173  *
1174  * Called with fc->lock
1175  */
1176 void fuse_flush_writepages(struct inode *inode)
1177 __releases(&fc->lock)
1178 __acquires(&fc->lock)
1179 {
1180         struct fuse_conn *fc = get_fuse_conn(inode);
1181         struct fuse_inode *fi = get_fuse_inode(inode);
1182         struct fuse_req *req;
1183
1184         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1185                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1186                 list_del_init(&req->list);
1187                 fuse_send_writepage(fc, req);
1188         }
1189 }
1190
1191 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1192 {
1193         struct inode *inode = req->inode;
1194         struct fuse_inode *fi = get_fuse_inode(inode);
1195
1196         mapping_set_error(inode->i_mapping, req->out.h.error);
1197         spin_lock(&fc->lock);
1198         fi->writectr--;
1199         fuse_writepage_finish(fc, req);
1200         spin_unlock(&fc->lock);
1201         fuse_writepage_free(fc, req);
1202 }
1203
1204 static int fuse_writepage_locked(struct page *page)
1205 {
1206         struct address_space *mapping = page->mapping;
1207         struct inode *inode = mapping->host;
1208         struct fuse_conn *fc = get_fuse_conn(inode);
1209         struct fuse_inode *fi = get_fuse_inode(inode);
1210         struct fuse_req *req;
1211         struct fuse_file *ff;
1212         struct page *tmp_page;
1213
1214         set_page_writeback(page);
1215
1216         req = fuse_request_alloc_nofs();
1217         if (!req)
1218                 goto err;
1219
1220         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1221         if (!tmp_page)
1222                 goto err_free;
1223
1224         spin_lock(&fc->lock);
1225         BUG_ON(list_empty(&fi->write_files));
1226         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1227         req->ff = fuse_file_get(ff);
1228         spin_unlock(&fc->lock);
1229
1230         fuse_write_fill(req, ff, page_offset(page), 0);
1231
1232         copy_highpage(tmp_page, page);
1233         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1234         req->in.argpages = 1;
1235         req->num_pages = 1;
1236         req->pages[0] = tmp_page;
1237         req->page_offset = 0;
1238         req->end = fuse_writepage_end;
1239         req->inode = inode;
1240
1241         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1242         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1243         end_page_writeback(page);
1244
1245         spin_lock(&fc->lock);
1246         list_add(&req->writepages_entry, &fi->writepages);
1247         list_add_tail(&req->list, &fi->queued_writes);
1248         fuse_flush_writepages(inode);
1249         spin_unlock(&fc->lock);
1250
1251         return 0;
1252
1253 err_free:
1254         fuse_request_free(req);
1255 err:
1256         end_page_writeback(page);
1257         return -ENOMEM;
1258 }
1259
1260 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1261 {
1262         int err;
1263
1264         err = fuse_writepage_locked(page);
1265         unlock_page(page);
1266
1267         return err;
1268 }
1269
1270 static int fuse_launder_page(struct page *page)
1271 {
1272         int err = 0;
1273         if (clear_page_dirty_for_io(page)) {
1274                 struct inode *inode = page->mapping->host;
1275                 err = fuse_writepage_locked(page);
1276                 if (!err)
1277                         fuse_wait_on_page_writeback(inode, page->index);
1278         }
1279         return err;
1280 }
1281
1282 /*
1283  * Write back dirty pages now, because there may not be any suitable
1284  * open files later
1285  */
1286 static void fuse_vma_close(struct vm_area_struct *vma)
1287 {
1288         filemap_write_and_wait(vma->vm_file->f_mapping);
1289 }
1290
1291 /*
1292  * Wait for writeback against this page to complete before allowing it
1293  * to be marked dirty again, and hence written back again, possibly
1294  * before the previous writepage completed.
1295  *
1296  * Block here, instead of in ->writepage(), so that the userspace fs
1297  * can only block processes actually operating on the filesystem.
1298  *
1299  * Otherwise unprivileged userspace fs would be able to block
1300  * unrelated:
1301  *
1302  * - page migration
1303  * - sync(2)
1304  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1305  */
1306 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1307 {
1308         struct page *page = vmf->page;
1309         /*
1310          * Don't use page->mapping as it may become NULL from a
1311          * concurrent truncate.
1312          */
1313         struct inode *inode = vma->vm_file->f_mapping->host;
1314
1315         fuse_wait_on_page_writeback(inode, page->index);
1316         return 0;
1317 }
1318
1319 static const struct vm_operations_struct fuse_file_vm_ops = {
1320         .close          = fuse_vma_close,
1321         .fault          = filemap_fault,
1322         .page_mkwrite   = fuse_page_mkwrite,
1323 };
1324
1325 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1326 {
1327         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1328                 struct inode *inode = file->f_dentry->d_inode;
1329                 struct fuse_conn *fc = get_fuse_conn(inode);
1330                 struct fuse_inode *fi = get_fuse_inode(inode);
1331                 struct fuse_file *ff = file->private_data;
1332                 /*
1333                  * file may be written through mmap, so chain it onto the
1334                  * inodes's write_file list
1335                  */
1336                 spin_lock(&fc->lock);
1337                 if (list_empty(&ff->write_entry))
1338                         list_add(&ff->write_entry, &fi->write_files);
1339                 spin_unlock(&fc->lock);
1340         }
1341         file_accessed(file);
1342         vma->vm_ops = &fuse_file_vm_ops;
1343         return 0;
1344 }
1345
1346 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1347 {
1348         /* Can't provide the coherency needed for MAP_SHARED */
1349         if (vma->vm_flags & VM_MAYSHARE)
1350                 return -ENODEV;
1351
1352         invalidate_inode_pages2(file->f_mapping);
1353
1354         return generic_file_mmap(file, vma);
1355 }
1356
1357 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1358                                   struct file_lock *fl)
1359 {
1360         switch (ffl->type) {
1361         case F_UNLCK:
1362                 break;
1363
1364         case F_RDLCK:
1365         case F_WRLCK:
1366                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1367                     ffl->end < ffl->start)
1368                         return -EIO;
1369
1370                 fl->fl_start = ffl->start;
1371                 fl->fl_end = ffl->end;
1372                 fl->fl_pid = ffl->pid;
1373                 break;
1374
1375         default:
1376                 return -EIO;
1377         }
1378         fl->fl_type = ffl->type;
1379         return 0;
1380 }
1381
1382 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1383                          const struct file_lock *fl, int opcode, pid_t pid,
1384                          int flock)
1385 {
1386         struct inode *inode = file->f_path.dentry->d_inode;
1387         struct fuse_conn *fc = get_fuse_conn(inode);
1388         struct fuse_file *ff = file->private_data;
1389         struct fuse_lk_in *arg = &req->misc.lk_in;
1390
1391         arg->fh = ff->fh;
1392         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1393         arg->lk.start = fl->fl_start;
1394         arg->lk.end = fl->fl_end;
1395         arg->lk.type = fl->fl_type;
1396         arg->lk.pid = pid;
1397         if (flock)
1398                 arg->lk_flags |= FUSE_LK_FLOCK;
1399         req->in.h.opcode = opcode;
1400         req->in.h.nodeid = get_node_id(inode);
1401         req->in.numargs = 1;
1402         req->in.args[0].size = sizeof(*arg);
1403         req->in.args[0].value = arg;
1404 }
1405
1406 static int fuse_getlk(struct file *file, struct file_lock *fl)
1407 {
1408         struct inode *inode = file->f_path.dentry->d_inode;
1409         struct fuse_conn *fc = get_fuse_conn(inode);
1410         struct fuse_req *req;
1411         struct fuse_lk_out outarg;
1412         int err;
1413
1414         req = fuse_get_req(fc);
1415         if (IS_ERR(req))
1416                 return PTR_ERR(req);
1417
1418         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1419         req->out.numargs = 1;
1420         req->out.args[0].size = sizeof(outarg);
1421         req->out.args[0].value = &outarg;
1422         fuse_request_send(fc, req);
1423         err = req->out.h.error;
1424         fuse_put_request(fc, req);
1425         if (!err)
1426                 err = convert_fuse_file_lock(&outarg.lk, fl);
1427
1428         return err;
1429 }
1430
1431 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1432 {
1433         struct inode *inode = file->f_path.dentry->d_inode;
1434         struct fuse_conn *fc = get_fuse_conn(inode);
1435         struct fuse_req *req;
1436         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1437         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1438         int err;
1439
1440         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1441                 /* NLM needs asynchronous locks, which we don't support yet */
1442                 return -ENOLCK;
1443         }
1444
1445         /* Unlock on close is handled by the flush method */
1446         if (fl->fl_flags & FL_CLOSE)
1447                 return 0;
1448
1449         req = fuse_get_req(fc);
1450         if (IS_ERR(req))
1451                 return PTR_ERR(req);
1452
1453         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1454         fuse_request_send(fc, req);
1455         err = req->out.h.error;
1456         /* locking is restartable */
1457         if (err == -EINTR)
1458                 err = -ERESTARTSYS;
1459         fuse_put_request(fc, req);
1460         return err;
1461 }
1462
1463 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1464 {
1465         struct inode *inode = file->f_path.dentry->d_inode;
1466         struct fuse_conn *fc = get_fuse_conn(inode);
1467         int err;
1468
1469         if (cmd == F_CANCELLK) {
1470                 err = 0;
1471         } else if (cmd == F_GETLK) {
1472                 if (fc->no_lock) {
1473                         posix_test_lock(file, fl);
1474                         err = 0;
1475                 } else
1476                         err = fuse_getlk(file, fl);
1477         } else {
1478                 if (fc->no_lock)
1479                         err = posix_lock_file(file, fl, NULL);
1480                 else
1481                         err = fuse_setlk(file, fl, 0);
1482         }
1483         return err;
1484 }
1485
1486 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1487 {
1488         struct inode *inode = file->f_path.dentry->d_inode;
1489         struct fuse_conn *fc = get_fuse_conn(inode);
1490         int err;
1491
1492         if (fc->no_lock) {
1493                 err = flock_lock_file_wait(file, fl);
1494         } else {
1495                 /* emulate flock with POSIX locks */
1496                 fl->fl_owner = (fl_owner_t) file;
1497                 err = fuse_setlk(file, fl, 1);
1498         }
1499
1500         return err;
1501 }
1502
1503 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1504 {
1505         struct inode *inode = mapping->host;
1506         struct fuse_conn *fc = get_fuse_conn(inode);
1507         struct fuse_req *req;
1508         struct fuse_bmap_in inarg;
1509         struct fuse_bmap_out outarg;
1510         int err;
1511
1512         if (!inode->i_sb->s_bdev || fc->no_bmap)
1513                 return 0;
1514
1515         req = fuse_get_req(fc);
1516         if (IS_ERR(req))
1517                 return 0;
1518
1519         memset(&inarg, 0, sizeof(inarg));
1520         inarg.block = block;
1521         inarg.blocksize = inode->i_sb->s_blocksize;
1522         req->in.h.opcode = FUSE_BMAP;
1523         req->in.h.nodeid = get_node_id(inode);
1524         req->in.numargs = 1;
1525         req->in.args[0].size = sizeof(inarg);
1526         req->in.args[0].value = &inarg;
1527         req->out.numargs = 1;
1528         req->out.args[0].size = sizeof(outarg);
1529         req->out.args[0].value = &outarg;
1530         fuse_request_send(fc, req);
1531         err = req->out.h.error;
1532         fuse_put_request(fc, req);
1533         if (err == -ENOSYS)
1534                 fc->no_bmap = 1;
1535
1536         return err ? 0 : outarg.block;
1537 }
1538
1539 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1540 {
1541         loff_t retval;
1542         struct inode *inode = file->f_path.dentry->d_inode;
1543
1544         mutex_lock(&inode->i_mutex);
1545         switch (origin) {
1546         case SEEK_END:
1547                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1548                 if (retval)
1549                         goto exit;
1550                 offset += i_size_read(inode);
1551                 break;
1552         case SEEK_CUR:
1553                 offset += file->f_pos;
1554         }
1555         retval = -EINVAL;
1556         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1557                 if (offset != file->f_pos) {
1558                         file->f_pos = offset;
1559                         file->f_version = 0;
1560                 }
1561                 retval = offset;
1562         }
1563 exit:
1564         mutex_unlock(&inode->i_mutex);
1565         return retval;
1566 }
1567
1568 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1569                         unsigned int nr_segs, size_t bytes, bool to_user)
1570 {
1571         struct iov_iter ii;
1572         int page_idx = 0;
1573
1574         if (!bytes)
1575                 return 0;
1576
1577         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1578
1579         while (iov_iter_count(&ii)) {
1580                 struct page *page = pages[page_idx++];
1581                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1582                 void *kaddr, *map;
1583
1584                 kaddr = map = kmap(page);
1585
1586                 while (todo) {
1587                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1588                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1589                         size_t copy = min(todo, iov_len);
1590                         size_t left;
1591
1592                         if (!to_user)
1593                                 left = copy_from_user(kaddr, uaddr, copy);
1594                         else
1595                                 left = copy_to_user(uaddr, kaddr, copy);
1596
1597                         if (unlikely(left))
1598                                 return -EFAULT;
1599
1600                         iov_iter_advance(&ii, copy);
1601                         todo -= copy;
1602                         kaddr += copy;
1603                 }
1604
1605                 kunmap(page);
1606         }
1607
1608         return 0;
1609 }
1610
1611 /*
1612  * For ioctls, there is no generic way to determine how much memory
1613  * needs to be read and/or written.  Furthermore, ioctls are allowed
1614  * to dereference the passed pointer, so the parameter requires deep
1615  * copying but FUSE has no idea whatsoever about what to copy in or
1616  * out.
1617  *
1618  * This is solved by allowing FUSE server to retry ioctl with
1619  * necessary in/out iovecs.  Let's assume the ioctl implementation
1620  * needs to read in the following structure.
1621  *
1622  * struct a {
1623  *      char    *buf;
1624  *      size_t  buflen;
1625  * }
1626  *
1627  * On the first callout to FUSE server, inarg->in_size and
1628  * inarg->out_size will be NULL; then, the server completes the ioctl
1629  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1630  * the actual iov array to
1631  *
1632  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1633  *
1634  * which tells FUSE to copy in the requested area and retry the ioctl.
1635  * On the second round, the server has access to the structure and
1636  * from that it can tell what to look for next, so on the invocation,
1637  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1638  *
1639  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1640  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1641  *
1642  * FUSE will copy both struct a and the pointed buffer from the
1643  * process doing the ioctl and retry ioctl with both struct a and the
1644  * buffer.
1645  *
1646  * This time, FUSE server has everything it needs and completes ioctl
1647  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1648  *
1649  * Copying data out works the same way.
1650  *
1651  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1652  * automatically initializes in and out iovs by decoding @cmd with
1653  * _IOC_* macros and the server is not allowed to request RETRY.  This
1654  * limits ioctl data transfers to well-formed ioctls and is the forced
1655  * behavior for all FUSE servers.
1656  */
1657 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1658                    unsigned int flags)
1659 {
1660         struct fuse_file *ff = file->private_data;
1661         struct fuse_conn *fc = ff->fc;
1662         struct fuse_ioctl_in inarg = {
1663                 .fh = ff->fh,
1664                 .cmd = cmd,
1665                 .arg = arg,
1666                 .flags = flags
1667         };
1668         struct fuse_ioctl_out outarg;
1669         struct fuse_req *req = NULL;
1670         struct page **pages = NULL;
1671         struct page *iov_page = NULL;
1672         struct iovec *in_iov = NULL, *out_iov = NULL;
1673         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1674         size_t in_size, out_size, transferred;
1675         int err;
1676
1677         /* assume all the iovs returned by client always fits in a page */
1678         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1679
1680         err = -ENOMEM;
1681         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1682         iov_page = alloc_page(GFP_KERNEL);
1683         if (!pages || !iov_page)
1684                 goto out;
1685
1686         /*
1687          * If restricted, initialize IO parameters as encoded in @cmd.
1688          * RETRY from server is not allowed.
1689          */
1690         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1691                 struct iovec *iov = page_address(iov_page);
1692
1693                 iov->iov_base = (void __user *)arg;
1694                 iov->iov_len = _IOC_SIZE(cmd);
1695
1696                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1697                         in_iov = iov;
1698                         in_iovs = 1;
1699                 }
1700
1701                 if (_IOC_DIR(cmd) & _IOC_READ) {
1702                         out_iov = iov;
1703                         out_iovs = 1;
1704                 }
1705         }
1706
1707  retry:
1708         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1709         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1710
1711         /*
1712          * Out data can be used either for actual out data or iovs,
1713          * make sure there always is at least one page.
1714          */
1715         out_size = max_t(size_t, out_size, PAGE_SIZE);
1716         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1717
1718         /* make sure there are enough buffer pages and init request with them */
1719         err = -ENOMEM;
1720         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1721                 goto out;
1722         while (num_pages < max_pages) {
1723                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1724                 if (!pages[num_pages])
1725                         goto out;
1726                 num_pages++;
1727         }
1728
1729         req = fuse_get_req(fc);
1730         if (IS_ERR(req)) {
1731                 err = PTR_ERR(req);
1732                 req = NULL;
1733                 goto out;
1734         }
1735         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1736         req->num_pages = num_pages;
1737
1738         /* okay, let's send it to the client */
1739         req->in.h.opcode = FUSE_IOCTL;
1740         req->in.h.nodeid = ff->nodeid;
1741         req->in.numargs = 1;
1742         req->in.args[0].size = sizeof(inarg);
1743         req->in.args[0].value = &inarg;
1744         if (in_size) {
1745                 req->in.numargs++;
1746                 req->in.args[1].size = in_size;
1747                 req->in.argpages = 1;
1748
1749                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1750                                            false);
1751                 if (err)
1752                         goto out;
1753         }
1754
1755         req->out.numargs = 2;
1756         req->out.args[0].size = sizeof(outarg);
1757         req->out.args[0].value = &outarg;
1758         req->out.args[1].size = out_size;
1759         req->out.argpages = 1;
1760         req->out.argvar = 1;
1761
1762         fuse_request_send(fc, req);
1763         err = req->out.h.error;
1764         transferred = req->out.args[1].size;
1765         fuse_put_request(fc, req);
1766         req = NULL;
1767         if (err)
1768                 goto out;
1769
1770         /* did it ask for retry? */
1771         if (outarg.flags & FUSE_IOCTL_RETRY) {
1772                 char *vaddr;
1773
1774                 /* no retry if in restricted mode */
1775                 err = -EIO;
1776                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1777                         goto out;
1778
1779                 in_iovs = outarg.in_iovs;
1780                 out_iovs = outarg.out_iovs;
1781
1782                 /*
1783                  * Make sure things are in boundary, separate checks
1784                  * are to protect against overflow.
1785                  */
1786                 err = -ENOMEM;
1787                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1788                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1789                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1790                         goto out;
1791
1792                 err = -EIO;
1793                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1794                         goto out;
1795
1796                 /* okay, copy in iovs and retry */
1797                 vaddr = kmap_atomic(pages[0], KM_USER0);
1798                 memcpy(page_address(iov_page), vaddr, transferred);
1799                 kunmap_atomic(vaddr, KM_USER0);
1800
1801                 in_iov = page_address(iov_page);
1802                 out_iov = in_iov + in_iovs;
1803
1804                 goto retry;
1805         }
1806
1807         err = -EIO;
1808         if (transferred > inarg.out_size)
1809                 goto out;
1810
1811         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1812  out:
1813         if (req)
1814                 fuse_put_request(fc, req);
1815         if (iov_page)
1816                 __free_page(iov_page);
1817         while (num_pages)
1818                 __free_page(pages[--num_pages]);
1819         kfree(pages);
1820
1821         return err ? err : outarg.result;
1822 }
1823 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1824
1825 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1826                                    unsigned long arg, unsigned int flags)
1827 {
1828         struct inode *inode = file->f_dentry->d_inode;
1829         struct fuse_conn *fc = get_fuse_conn(inode);
1830
1831         if (!fuse_allow_task(fc, current))
1832                 return -EACCES;
1833
1834         if (is_bad_inode(inode))
1835                 return -EIO;
1836
1837         return fuse_do_ioctl(file, cmd, arg, flags);
1838 }
1839
1840 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1841                             unsigned long arg)
1842 {
1843         return fuse_file_ioctl_common(file, cmd, arg, 0);
1844 }
1845
1846 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1847                                    unsigned long arg)
1848 {
1849         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1850 }
1851
1852 /*
1853  * All files which have been polled are linked to RB tree
1854  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1855  * find the matching one.
1856  */
1857 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1858                                               struct rb_node **parent_out)
1859 {
1860         struct rb_node **link = &fc->polled_files.rb_node;
1861         struct rb_node *last = NULL;
1862
1863         while (*link) {
1864                 struct fuse_file *ff;
1865
1866                 last = *link;
1867                 ff = rb_entry(last, struct fuse_file, polled_node);
1868
1869                 if (kh < ff->kh)
1870                         link = &last->rb_left;
1871                 else if (kh > ff->kh)
1872                         link = &last->rb_right;
1873                 else
1874                         return link;
1875         }
1876
1877         if (parent_out)
1878                 *parent_out = last;
1879         return link;
1880 }
1881
1882 /*
1883  * The file is about to be polled.  Make sure it's on the polled_files
1884  * RB tree.  Note that files once added to the polled_files tree are
1885  * not removed before the file is released.  This is because a file
1886  * polled once is likely to be polled again.
1887  */
1888 static void fuse_register_polled_file(struct fuse_conn *fc,
1889                                       struct fuse_file *ff)
1890 {
1891         spin_lock(&fc->lock);
1892         if (RB_EMPTY_NODE(&ff->polled_node)) {
1893                 struct rb_node **link, *parent;
1894
1895                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1896                 BUG_ON(*link);
1897                 rb_link_node(&ff->polled_node, parent, link);
1898                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1899         }
1900         spin_unlock(&fc->lock);
1901 }
1902
1903 unsigned fuse_file_poll(struct file *file, poll_table *wait)
1904 {
1905         struct fuse_file *ff = file->private_data;
1906         struct fuse_conn *fc = ff->fc;
1907         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1908         struct fuse_poll_out outarg;
1909         struct fuse_req *req;
1910         int err;
1911
1912         if (fc->no_poll)
1913                 return DEFAULT_POLLMASK;
1914
1915         poll_wait(file, &ff->poll_wait, wait);
1916
1917         /*
1918          * Ask for notification iff there's someone waiting for it.
1919          * The client may ignore the flag and always notify.
1920          */
1921         if (waitqueue_active(&ff->poll_wait)) {
1922                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1923                 fuse_register_polled_file(fc, ff);
1924         }
1925
1926         req = fuse_get_req(fc);
1927         if (IS_ERR(req))
1928                 return POLLERR;
1929
1930         req->in.h.opcode = FUSE_POLL;
1931         req->in.h.nodeid = ff->nodeid;
1932         req->in.numargs = 1;
1933         req->in.args[0].size = sizeof(inarg);
1934         req->in.args[0].value = &inarg;
1935         req->out.numargs = 1;
1936         req->out.args[0].size = sizeof(outarg);
1937         req->out.args[0].value = &outarg;
1938         fuse_request_send(fc, req);
1939         err = req->out.h.error;
1940         fuse_put_request(fc, req);
1941
1942         if (!err)
1943                 return outarg.revents;
1944         if (err == -ENOSYS) {
1945                 fc->no_poll = 1;
1946                 return DEFAULT_POLLMASK;
1947         }
1948         return POLLERR;
1949 }
1950 EXPORT_SYMBOL_GPL(fuse_file_poll);
1951
1952 /*
1953  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1954  * wakes up the poll waiters.
1955  */
1956 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1957                             struct fuse_notify_poll_wakeup_out *outarg)
1958 {
1959         u64 kh = outarg->kh;
1960         struct rb_node **link;
1961
1962         spin_lock(&fc->lock);
1963
1964         link = fuse_find_polled_node(fc, kh, NULL);
1965         if (*link) {
1966                 struct fuse_file *ff;
1967
1968                 ff = rb_entry(*link, struct fuse_file, polled_node);
1969                 wake_up_interruptible_sync(&ff->poll_wait);
1970         }
1971
1972         spin_unlock(&fc->lock);
1973         return 0;
1974 }
1975
1976 static const struct file_operations fuse_file_operations = {
1977         .llseek         = fuse_file_llseek,
1978         .read           = do_sync_read,
1979         .aio_read       = fuse_file_aio_read,
1980         .write          = do_sync_write,
1981         .aio_write      = fuse_file_aio_write,
1982         .mmap           = fuse_file_mmap,
1983         .open           = fuse_open,
1984         .flush          = fuse_flush,
1985         .release        = fuse_release,
1986         .fsync          = fuse_fsync,
1987         .lock           = fuse_file_lock,
1988         .flock          = fuse_file_flock,
1989         .splice_read    = generic_file_splice_read,
1990         .unlocked_ioctl = fuse_file_ioctl,
1991         .compat_ioctl   = fuse_file_compat_ioctl,
1992         .poll           = fuse_file_poll,
1993 };
1994
1995 static const struct file_operations fuse_direct_io_file_operations = {
1996         .llseek         = fuse_file_llseek,
1997         .read           = fuse_direct_read,
1998         .write          = fuse_direct_write,
1999         .mmap           = fuse_direct_mmap,
2000         .open           = fuse_open,
2001         .flush          = fuse_flush,
2002         .release        = fuse_release,
2003         .fsync          = fuse_fsync,
2004         .lock           = fuse_file_lock,
2005         .flock          = fuse_file_flock,
2006         .unlocked_ioctl = fuse_file_ioctl,
2007         .compat_ioctl   = fuse_file_compat_ioctl,
2008         .poll           = fuse_file_poll,
2009         /* no splice_read */
2010 };
2011
2012 static const struct address_space_operations fuse_file_aops  = {
2013         .readpage       = fuse_readpage,
2014         .writepage      = fuse_writepage,
2015         .launder_page   = fuse_launder_page,
2016         .write_begin    = fuse_write_begin,
2017         .write_end      = fuse_write_end,
2018         .readpages      = fuse_readpages,
2019         .set_page_dirty = __set_page_dirty_nobuffers,
2020         .bmap           = fuse_bmap,
2021 };
2022
2023 void fuse_init_file_inode(struct inode *inode)
2024 {
2025         inode->i_fop = &fuse_file_operations;
2026         inode->i_data.a_ops = &fuse_file_aops;
2027 }