fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17
  18 static const struct file_operations fuse_direct_io_file_operations;
  19
  20 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  21                           int opcode, struct fuse_open_out *outargp)
  22 {
  23         struct fuse_open_in inarg;
  24         struct fuse_req *req;
  25         int err;
  26
  27         req = fuse_get_req(fc);
  28         if (IS_ERR(req))
  29                 return PTR_ERR(req);
  30
  31         memset(&inarg, 0, sizeof(inarg));
  32         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  33         if (!fc->atomic_o_trunc)
  34                 inarg.flags &= ~O_TRUNC;
  35         req->in.h.opcode = opcode;
  36         req->in.h.nodeid = nodeid;
  37         req->in.numargs = 1;
  38         req->in.args[0].size = sizeof(inarg);
  39         req->in.args[0].value = &inarg;
  40         req->out.numargs = 1;
  41         req->out.args[0].size = sizeof(*outargp);
  42         req->out.args[0].value = outargp;
  43         fuse_request_send(fc, req);
  44         err = req->out.h.error;
  45         fuse_put_request(fc, req);
  46
  47         return err;
  48 }
  49
  50 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  51 {
  52         struct fuse_file *ff;
  53
  54         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  55         if (unlikely(!ff))
  56                 return NULL;
  57
  58         ff->fc = fc;
  59         ff->reserved_req = fuse_request_alloc();
  60         if (unlikely(!ff->reserved_req)) {
  61                 kfree(ff);
  62                 return NULL;
  63         }
  64
  65         INIT_LIST_HEAD(&ff->write_entry);
  66         atomic_set(&ff->count, 0);
  67         RB_CLEAR_NODE(&ff->polled_node);
  68         init_waitqueue_head(&ff->poll_wait);
  69
  70         spin_lock(&fc->lock);
  71         ff->kh = ++fc->khctr;
  72         spin_unlock(&fc->lock);
  73
  74         return ff;
  75 }
  76
  77 void fuse_file_free(struct fuse_file *ff)
  78 {
  79         fuse_request_free(ff->reserved_req);
  80         kfree(ff);
  81 }
  82
  83 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  84 {
  85         atomic_inc(&ff->count);
  86         return ff;
  87 }
  88
  89 static void fuse_release_async(struct work_struct *work)
  90 {
  91         struct fuse_req *req;
  92         struct fuse_conn *fc;
  93         struct path path;
  94
  95         req = container_of(work, struct fuse_req, misc.release.work);
  96         path = req->misc.release.path;
  97         fc = get_fuse_conn(path.dentry->d_inode);
  98
  99         fuse_put_request(fc, req);
 100         path_put(&path);
 101 }
 102
 103 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 104 {
 105         if (fc->destroy_req) {
 106                 /*
 107                  * If this is a fuseblk mount, then it's possible that
 108                  * releasing the path will result in releasing the
 109                  * super block and sending the DESTROY request.  If
 110                  * the server is single threaded, this would hang.
 111                  * For this reason do the path_put() in a separate
 112                  * thread.
 113                  */
 114                 atomic_inc(&req->count);
 115                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 116                 schedule_work(&req->misc.release.work);
 117         } else {
 118                 path_put(&req->misc.release.path);
 119         }
 120 }
 121
 122 static void fuse_file_put(struct fuse_file *ff, bool sync)
 123 {
 124         if (atomic_dec_and_test(&ff->count)) {
 125                 struct fuse_req *req = ff->reserved_req;
 126
 127                 if (sync) {
 128                         fuse_request_send(ff->fc, req);
 129                         path_put(&req->misc.release.path);
 130                         fuse_put_request(ff->fc, req);
 131                 } else {
 132                         req->end = fuse_release_end;
 133                         fuse_request_send_background(ff->fc, req);
 134                 }
 135                 kfree(ff);
 136         }
 137 }
 138
 139 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 140                  bool isdir)
 141 {
 142         struct fuse_open_out outarg;
 143         struct fuse_file *ff;
 144         int err;
 145         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 146
 147         ff = fuse_file_alloc(fc);
 148         if (!ff)
 149                 return -ENOMEM;
 150
 151         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 152         if (err) {
 153                 fuse_file_free(ff);
 154                 return err;
 155         }
 156
 157         if (isdir)
 158                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 159
 160         ff->fh = outarg.fh;
 161         ff->nodeid = nodeid;
 162         ff->open_flags = outarg.open_flags;
 163         file->private_data = fuse_file_get(ff);
 164
 165         return 0;
 166 }
 167 EXPORT_SYMBOL_GPL(fuse_do_open);
 168
 169 void fuse_finish_open(struct inode *inode, struct file *file)
 170 {
 171         struct fuse_file *ff = file->private_data;
 172         struct fuse_conn *fc = get_fuse_conn(inode);
 173
 174         if (ff->open_flags & FOPEN_DIRECT_IO)
 175                 file->f_op = &fuse_direct_io_file_operations;
 176         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 177                 invalidate_inode_pages2(inode->i_mapping);
 178         if (ff->open_flags & FOPEN_NONSEEKABLE)
 179                 nonseekable_open(inode, file);
 180         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 181                 struct fuse_inode *fi = get_fuse_inode(inode);
 182
 183                 spin_lock(&fc->lock);
 184                 fi->attr_version = ++fc->attr_version;
 185                 i_size_write(inode, 0);
 186                 spin_unlock(&fc->lock);
 187                 fuse_invalidate_attr(inode);
 188         }
 189 }
 190
 191 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 192 {
 193         struct fuse_conn *fc = get_fuse_conn(inode);
 194         int err;
 195
 196         /* VFS checks this, but only _after_ ->open() */
 197         if (file->f_flags & O_DIRECT)
 198                 return -EINVAL;
 199
 200         err = generic_file_open(inode, file);
 201         if (err)
 202                 return err;
 203
 204         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 205         if (err)
 206                 return err;
 207
 208         fuse_finish_open(inode, file);
 209
 210         return 0;
 211 }
 212
 213 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 214 {
 215         struct fuse_conn *fc = ff->fc;
 216         struct fuse_req *req = ff->reserved_req;
 217         struct fuse_release_in *inarg = &req->misc.release.in;
 218
 219         spin_lock(&fc->lock);
 220         list_del(&ff->write_entry);
 221         if (!RB_EMPTY_NODE(&ff->polled_node))
 222                 rb_erase(&ff->polled_node, &fc->polled_files);
 223         spin_unlock(&fc->lock);
 224
 225         wake_up_interruptible_all(&ff->poll_wait);
 226
 227         inarg->fh = ff->fh;
 228         inarg->flags = flags;
 229         req->in.h.opcode = opcode;
 230         req->in.h.nodeid = ff->nodeid;
 231         req->in.numargs = 1;
 232         req->in.args[0].size = sizeof(struct fuse_release_in);
 233         req->in.args[0].value = inarg;
 234 }
 235
 236 void fuse_release_common(struct file *file, int opcode)
 237 {
 238         struct fuse_file *ff;
 239         struct fuse_req *req;
 240
 241         ff = file->private_data;
 242         if (unlikely(!ff))
 243                 return;
 244
 245         req = ff->reserved_req;
 246         fuse_prepare_release(ff, file->f_flags, opcode);
 247
 248         /* Hold vfsmount and dentry until release is finished */
 249         path_get(&file->f_path);
 250         req->misc.release.path = file->f_path;
 251
 252         /*
 253          * Normally this will send the RELEASE request, however if
 254          * some asynchronous READ or WRITE requests are outstanding,
 255          * the sending will be delayed.
 256          *
 257          * Make the release synchronous if this is a fuseblk mount,
 258          * synchronous RELEASE is allowed (and desirable) in this case
 259          * because the server can be trusted not to screw up.
 260          */
 261         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 262 }
 263
 264 static int fuse_open(struct inode *inode, struct file *file)
 265 {
 266         return fuse_open_common(inode, file, false);
 267 }
 268
 269 static int fuse_release(struct inode *inode, struct file *file)
 270 {
 271         fuse_release_common(file, FUSE_RELEASE);
 272
 273         /* return value is ignored by VFS */
 274         return 0;
 275 }
 276
 277 void fuse_sync_release(struct fuse_file *ff, int flags)
 278 {
 279         WARN_ON(atomic_read(&ff->count) > 1);
 280         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 281         ff->reserved_req->force = 1;
 282         fuse_request_send(ff->fc, ff->reserved_req);
 283         fuse_put_request(ff->fc, ff->reserved_req);
 284         kfree(ff);
 285 }
 286 EXPORT_SYMBOL_GPL(fuse_sync_release);
 287
 288 /*
 289  * Scramble the ID space with XTEA, so that the value of the files_struct
 290  * pointer is not exposed to userspace.
 291  */
 292 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 293 {
 294         u32 *k = fc->scramble_key;
 295         u64 v = (unsigned long) id;
 296         u32 v0 = v;
 297         u32 v1 = v >> 32;
 298         u32 sum = 0;
 299         int i;
 300
 301         for (i = 0; i < 32; i++) {
 302                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 303                 sum += 0x9E3779B9;
 304                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 305         }
 306
 307         return (u64) v0 + ((u64) v1 << 32);
 308 }
 309
 310 /*
 311  * Check if page is under writeback
 312  *
 313  * This is currently done by walking the list of writepage requests
 314  * for the inode, which can be pretty inefficient.
 315  */
 316 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 317 {
 318         struct fuse_conn *fc = get_fuse_conn(inode);
 319         struct fuse_inode *fi = get_fuse_inode(inode);
 320         struct fuse_req *req;
 321         bool found = false;
 322
 323         spin_lock(&fc->lock);
 324         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 325                 pgoff_t curr_index;
 326
 327                 BUG_ON(req->inode != inode);
 328                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 329                 if (curr_index == index) {
 330                         found = true;
 331                         break;
 332                 }
 333         }
 334         spin_unlock(&fc->lock);
 335
 336         return found;
 337 }
 338
 339 /*
 340  * Wait for page writeback to be completed.
 341  *
 342  * Since fuse doesn't rely on the VM writeback tracking, this has to
 343  * use some other means.
 344  */
 345 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 346 {
 347         struct fuse_inode *fi = get_fuse_inode(inode);
 348
 349         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 350         return 0;
 351 }
 352
 353 static int fuse_flush(struct file *file, fl_owner_t id)
 354 {
 355         struct inode *inode = file->f_path.dentry->d_inode;
 356         struct fuse_conn *fc = get_fuse_conn(inode);
 357         struct fuse_file *ff = file->private_data;
 358         struct fuse_req *req;
 359         struct fuse_flush_in inarg;
 360         int err;
 361
 362         if (is_bad_inode(inode))
 363                 return -EIO;
 364
 365         if (fc->no_flush)
 366                 return 0;
 367
 368         req = fuse_get_req_nofail(fc, file);
 369         memset(&inarg, 0, sizeof(inarg));
 370         inarg.fh = ff->fh;
 371         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 372         req->in.h.opcode = FUSE_FLUSH;
 373         req->in.h.nodeid = get_node_id(inode);
 374         req->in.numargs = 1;
 375         req->in.args[0].size = sizeof(inarg);
 376         req->in.args[0].value = &inarg;
 377         req->force = 1;
 378         fuse_request_send(fc, req);
 379         err = req->out.h.error;
 380         fuse_put_request(fc, req);
 381         if (err == -ENOSYS) {
 382                 fc->no_flush = 1;
 383                 err = 0;
 384         }
 385         return err;
 386 }
 387
 388 /*
 389  * Wait for all pending writepages on the inode to finish.
 390  *
 391  * This is currently done by blocking further writes with FUSE_NOWRITE
 392  * and waiting for all sent writes to complete.
 393  *
 394  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 395  * could conflict with truncation.
 396  */
 397 static void fuse_sync_writes(struct inode *inode)
 398 {
 399         fuse_set_nowrite(inode);
 400         fuse_release_nowrite(inode);
 401 }
 402
 403 int fuse_fsync_common(struct file *file, int datasync, int isdir)
 404 {
 405         struct inode *inode = file->f_mapping->host;
 406         struct fuse_conn *fc = get_fuse_conn(inode);
 407         struct fuse_file *ff = file->private_data;
 408         struct fuse_req *req;
 409         struct fuse_fsync_in inarg;
 410         int err;
 411
 412         if (is_bad_inode(inode))
 413                 return -EIO;
 414
 415         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 416                 return 0;
 417
 418         /*
 419          * Start writeback against all dirty pages of the inode, then
 420          * wait for all outstanding writes, before sending the FSYNC
 421          * request.
 422          */
 423         err = write_inode_now(inode, 0);
 424         if (err)
 425                 return err;
 426
 427         fuse_sync_writes(inode);
 428
 429         req = fuse_get_req(fc);
 430         if (IS_ERR(req))
 431                 return PTR_ERR(req);
 432
 433         memset(&inarg, 0, sizeof(inarg));
 434         inarg.fh = ff->fh;
 435         inarg.fsync_flags = datasync ? 1 : 0;
 436         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 437         req->in.h.nodeid = get_node_id(inode);
 438         req->in.numargs = 1;
 439         req->in.args[0].size = sizeof(inarg);
 440         req->in.args[0].value = &inarg;
 441         fuse_request_send(fc, req);
 442         err = req->out.h.error;
 443         fuse_put_request(fc, req);
 444         if (err == -ENOSYS) {
 445                 if (isdir)
 446                         fc->no_fsyncdir = 1;
 447                 else
 448                         fc->no_fsync = 1;
 449                 err = 0;
 450         }
 451         return err;
 452 }
 453
 454 static int fuse_fsync(struct file *file, int datasync)
 455 {
 456         return fuse_fsync_common(file, datasync, 0);
 457 }
 458
 459 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 460                     size_t count, int opcode)
 461 {
 462         struct fuse_read_in *inarg = &req->misc.read.in;
 463         struct fuse_file *ff = file->private_data;
 464
 465         inarg->fh = ff->fh;
 466         inarg->offset = pos;
 467         inarg->size = count;
 468         inarg->flags = file->f_flags;
 469         req->in.h.opcode = opcode;
 470         req->in.h.nodeid = ff->nodeid;
 471         req->in.numargs = 1;
 472         req->in.args[0].size = sizeof(struct fuse_read_in);
 473         req->in.args[0].value = inarg;
 474         req->out.argvar = 1;
 475         req->out.numargs = 1;
 476         req->out.args[0].size = count;
 477 }
 478
 479 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 480                              loff_t pos, size_t count, fl_owner_t owner)
 481 {
 482         struct fuse_file *ff = file->private_data;
 483         struct fuse_conn *fc = ff->fc;
 484
 485         fuse_read_fill(req, file, pos, count, FUSE_READ);
 486         if (owner != NULL) {
 487                 struct fuse_read_in *inarg = &req->misc.read.in;
 488
 489                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 490                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 491         }
 492         fuse_request_send(fc, req);
 493         return req->out.args[0].size;
 494 }
 495
 496 static void fuse_read_update_size(struct inode *inode, loff_t size,
 497                                   u64 attr_ver)
 498 {
 499         struct fuse_conn *fc = get_fuse_conn(inode);
 500         struct fuse_inode *fi = get_fuse_inode(inode);
 501
 502         spin_lock(&fc->lock);
 503         if (attr_ver == fi->attr_version && size < inode->i_size) {
 504                 fi->attr_version = ++fc->attr_version;
 505                 i_size_write(inode, size);
 506         }
 507         spin_unlock(&fc->lock);
 508 }
 509
 510 static int fuse_readpage(struct file *file, struct page *page)
 511 {
 512         struct inode *inode = page->mapping->host;
 513         struct fuse_conn *fc = get_fuse_conn(inode);
 514         struct fuse_req *req;
 515         size_t num_read;
 516         loff_t pos = page_offset(page);
 517         size_t count = PAGE_CACHE_SIZE;
 518         u64 attr_ver;
 519         int err;
 520
 521         err = -EIO;
 522         if (is_bad_inode(inode))
 523                 goto out;
 524
 525         /*
 526          * Page writeback can extend beyond the lifetime of the
 527          * page-cache page, so make sure we read a properly synced
 528          * page.
 529          */
 530         fuse_wait_on_page_writeback(inode, page->index);
 531
 532         req = fuse_get_req(fc);
 533         err = PTR_ERR(req);
 534         if (IS_ERR(req))
 535                 goto out;
 536
 537         attr_ver = fuse_get_attr_version(fc);
 538
 539         req->out.page_zeroing = 1;
 540         req->out.argpages = 1;
 541         req->num_pages = 1;
 542         req->pages[0] = page;
 543         num_read = fuse_send_read(req, file, pos, count, NULL);
 544         err = req->out.h.error;
 545         fuse_put_request(fc, req);
 546
 547         if (!err) {
 548                 /*
 549                  * Short read means EOF.  If file size is larger, truncate it
 550                  */
 551                 if (num_read < count)
 552                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 553
 554                 SetPageUptodate(page);
 555         }
 556
 557         fuse_invalidate_attr(inode); /* atime changed */
 558  out:
 559         unlock_page(page);
 560         return err;
 561 }
 562
 563 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 564 {
 565         int i;
 566         size_t count = req->misc.read.in.size;
 567         size_t num_read = req->out.args[0].size;
 568         struct address_space *mapping = NULL;
 569
 570         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 571                 mapping = req->pages[i]->mapping;
 572
 573         if (mapping) {
 574                 struct inode *inode = mapping->host;
 575
 576                 /*
 577                  * Short read means EOF. If file size is larger, truncate it
 578                  */
 579                 if (!req->out.h.error && num_read < count) {
 580                         loff_t pos;
 581
 582                         pos = page_offset(req->pages[0]) + num_read;
 583                         fuse_read_update_size(inode, pos,
 584                                               req->misc.read.attr_ver);
 585                 }
 586                 fuse_invalidate_attr(inode); /* atime changed */
 587         }
 588
 589         for (i = 0; i < req->num_pages; i++) {
 590                 struct page *page = req->pages[i];
 591                 if (!req->out.h.error)
 592                         SetPageUptodate(page);
 593                 else
 594                         SetPageError(page);
 595                 unlock_page(page);
 596                 page_cache_release(page);
 597         }
 598         if (req->ff)
 599                 fuse_file_put(req->ff, false);
 600 }
 601
 602 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 603 {
 604         struct fuse_file *ff = file->private_data;
 605         struct fuse_conn *fc = ff->fc;
 606         loff_t pos = page_offset(req->pages[0]);
 607         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 608
 609         req->out.argpages = 1;
 610         req->out.page_zeroing = 1;
 611         req->out.page_replace = 1;
 612         fuse_read_fill(req, file, pos, count, FUSE_READ);
 613         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 614         if (fc->async_read) {
 615                 req->ff = fuse_file_get(ff);
 616                 req->end = fuse_readpages_end;
 617                 fuse_request_send_background(fc, req);
 618         } else {
 619                 fuse_request_send(fc, req);
 620                 fuse_readpages_end(fc, req);
 621                 fuse_put_request(fc, req);
 622         }
 623 }
 624
 625 struct fuse_fill_data {
 626         struct fuse_req *req;
 627         struct file *file;
 628         struct inode *inode;
 629 };
 630
 631 static int fuse_readpages_fill(void *_data, struct page *page)
 632 {
 633         struct fuse_fill_data *data = _data;
 634         struct fuse_req *req = data->req;
 635         struct inode *inode = data->inode;
 636         struct fuse_conn *fc = get_fuse_conn(inode);
 637
 638         fuse_wait_on_page_writeback(inode, page->index);
 639
 640         if (req->num_pages &&
 641             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 642              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 643              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 644                 fuse_send_readpages(req, data->file);
 645                 data->req = req = fuse_get_req(fc);
 646                 if (IS_ERR(req)) {
 647                         unlock_page(page);
 648                         return PTR_ERR(req);
 649                 }
 650         }
 651         page_cache_get(page);
 652         req->pages[req->num_pages] = page;
 653         req->num_pages++;
 654         return 0;
 655 }
 656
 657 static int fuse_readpages(struct file *file, struct address_space *mapping,
 658                           struct list_head *pages, unsigned nr_pages)
 659 {
 660         struct inode *inode = mapping->host;
 661         struct fuse_conn *fc = get_fuse_conn(inode);
 662         struct fuse_fill_data data;
 663         int err;
 664
 665         err = -EIO;
 666         if (is_bad_inode(inode))
 667                 goto out;
 668
 669         data.file = file;
 670         data.inode = inode;
 671         data.req = fuse_get_req(fc);
 672         err = PTR_ERR(data.req);
 673         if (IS_ERR(data.req))
 674                 goto out;
 675
 676         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 677         if (!err) {
 678                 if (data.req->num_pages)
 679                         fuse_send_readpages(data.req, file);
 680                 else
 681                         fuse_put_request(fc, data.req);
 682         }
 683 out:
 684         return err;
 685 }
 686
 687 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 688                                   unsigned long nr_segs, loff_t pos)
 689 {
 690         struct inode *inode = iocb->ki_filp->f_mapping->host;
 691
 692         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 693                 int err;
 694                 /*
 695                  * If trying to read past EOF, make sure the i_size
 696                  * attribute is up-to-date.
 697                  */
 698                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 699                 if (err)
 700                         return err;
 701         }
 702
 703         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 704 }
 705
 706 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 707                             loff_t pos, size_t count)
 708 {
 709         struct fuse_write_in *inarg = &req->misc.write.in;
 710         struct fuse_write_out *outarg = &req->misc.write.out;
 711
 712         inarg->fh = ff->fh;
 713         inarg->offset = pos;
 714         inarg->size = count;
 715         req->in.h.opcode = FUSE_WRITE;
 716         req->in.h.nodeid = ff->nodeid;
 717         req->in.numargs = 2;
 718         if (ff->fc->minor < 9)
 719                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 720         else
 721                 req->in.args[0].size = sizeof(struct fuse_write_in);
 722         req->in.args[0].value = inarg;
 723         req->in.args[1].size = count;
 724         req->out.numargs = 1;
 725         req->out.args[0].size = sizeof(struct fuse_write_out);
 726         req->out.args[0].value = outarg;
 727 }
 728
 729 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 730                               loff_t pos, size_t count, fl_owner_t owner)
 731 {
 732         struct fuse_file *ff = file->private_data;
 733         struct fuse_conn *fc = ff->fc;
 734         struct fuse_write_in *inarg = &req->misc.write.in;
 735
 736         fuse_write_fill(req, ff, pos, count);
 737         inarg->flags = file->f_flags;
 738         if (owner != NULL) {
 739                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 740                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 741         }
 742         fuse_request_send(fc, req);
 743         return req->misc.write.out.size;
 744 }
 745
 746 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 747                         loff_t pos, unsigned len, unsigned flags,
 748                         struct page **pagep, void **fsdata)
 749 {
 750         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 751
 752         *pagep = grab_cache_page_write_begin(mapping, index, flags);
 753         if (!*pagep)
 754                 return -ENOMEM;
 755         return 0;
 756 }
 757
 758 void fuse_write_update_size(struct inode *inode, loff_t pos)
 759 {
 760         struct fuse_conn *fc = get_fuse_conn(inode);
 761         struct fuse_inode *fi = get_fuse_inode(inode);
 762
 763         spin_lock(&fc->lock);
 764         fi->attr_version = ++fc->attr_version;
 765         if (pos > inode->i_size)
 766                 i_size_write(inode, pos);
 767         spin_unlock(&fc->lock);
 768 }
 769
 770 static int fuse_buffered_write(struct file *file, struct inode *inode,
 771                                loff_t pos, unsigned count, struct page *page)
 772 {
 773         int err;
 774         size_t nres;
 775         struct fuse_conn *fc = get_fuse_conn(inode);
 776         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 777         struct fuse_req *req;
 778
 779         if (is_bad_inode(inode))
 780                 return -EIO;
 781
 782         /*
 783          * Make sure writepages on the same page are not mixed up with
 784          * plain writes.
 785          */
 786         fuse_wait_on_page_writeback(inode, page->index);
 787
 788         req = fuse_get_req(fc);
 789         if (IS_ERR(req))
 790                 return PTR_ERR(req);
 791
 792         req->in.argpages = 1;
 793         req->num_pages = 1;
 794         req->pages[0] = page;
 795         req->page_offset = offset;
 796         nres = fuse_send_write(req, file, pos, count, NULL);
 797         err = req->out.h.error;
 798         fuse_put_request(fc, req);
 799         if (!err && !nres)
 800                 err = -EIO;
 801         if (!err) {
 802                 pos += nres;
 803                 fuse_write_update_size(inode, pos);
 804                 if (count == PAGE_CACHE_SIZE)
 805                         SetPageUptodate(page);
 806         }
 807         fuse_invalidate_attr(inode);
 808         return err ? err : nres;
 809 }
 810
 811 static int fuse_write_end(struct file *file, struct address_space *mapping,
 812                         loff_t pos, unsigned len, unsigned copied,
 813                         struct page *page, void *fsdata)
 814 {
 815         struct inode *inode = mapping->host;
 816         int res = 0;
 817
 818         if (copied)
 819                 res = fuse_buffered_write(file, inode, pos, copied, page);
 820
 821         unlock_page(page);
 822         page_cache_release(page);
 823         return res;
 824 }
 825
 826 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 827                                     struct inode *inode, loff_t pos,
 828                                     size_t count)
 829 {
 830         size_t res;
 831         unsigned offset;
 832         unsigned i;
 833
 834         for (i = 0; i < req->num_pages; i++)
 835                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 836
 837         res = fuse_send_write(req, file, pos, count, NULL);
 838
 839         offset = req->page_offset;
 840         count = res;
 841         for (i = 0; i < req->num_pages; i++) {
 842                 struct page *page = req->pages[i];
 843
 844                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 845                         SetPageUptodate(page);
 846
 847                 if (count > PAGE_CACHE_SIZE - offset)
 848                         count -= PAGE_CACHE_SIZE - offset;
 849                 else
 850                         count = 0;
 851                 offset = 0;
 852
 853                 unlock_page(page);
 854                 page_cache_release(page);
 855         }
 856
 857         return res;
 858 }
 859
 860 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 861                                struct address_space *mapping,
 862                                struct iov_iter *ii, loff_t pos)
 863 {
 864         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 865         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 866         size_t count = 0;
 867         int err;
 868
 869         req->in.argpages = 1;
 870         req->page_offset = offset;
 871
 872         do {
 873                 size_t tmp;
 874                 struct page *page;
 875                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 876                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 877                                      iov_iter_count(ii));
 878
 879                 bytes = min_t(size_t, bytes, fc->max_write - count);
 880
 881  again:
 882                 err = -EFAULT;
 883                 if (iov_iter_fault_in_readable(ii, bytes))
 884                         break;
 885
 886                 err = -ENOMEM;
 887                 page = grab_cache_page_write_begin(mapping, index, 0);
 888                 if (!page)
 889                         break;
 890
 891                 if (mapping_writably_mapped(mapping))
 892                         flush_dcache_page(page);
 893
 894                 pagefault_disable();
 895                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 896                 pagefault_enable();
 897                 flush_dcache_page(page);
 898
 899                 if (!tmp) {
 900                         unlock_page(page);
 901                         page_cache_release(page);
 902                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 903                         goto again;
 904                 }
 905
 906                 err = 0;
 907                 req->pages[req->num_pages] = page;
 908                 req->num_pages++;
 909
 910                 iov_iter_advance(ii, tmp);
 911                 count += tmp;
 912                 pos += tmp;
 913                 offset += tmp;
 914                 if (offset == PAGE_CACHE_SIZE)
 915                         offset = 0;
 916
 917                 if (!fc->big_writes)
 918                         break;
 919         } while (iov_iter_count(ii) && count < fc->max_write &&
 920                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 921
 922         return count > 0 ? count : err;
 923 }
 924
 925 static ssize_t fuse_perform_write(struct file *file,
 926                                   struct address_space *mapping,
 927                                   struct iov_iter *ii, loff_t pos)
 928 {
 929         struct inode *inode = mapping->host;
 930         struct fuse_conn *fc = get_fuse_conn(inode);
 931         int err = 0;
 932         ssize_t res = 0;
 933
 934         if (is_bad_inode(inode))
 935                 return -EIO;
 936
 937         do {
 938                 struct fuse_req *req;
 939                 ssize_t count;
 940
 941                 req = fuse_get_req(fc);
 942                 if (IS_ERR(req)) {
 943                         err = PTR_ERR(req);
 944                         break;
 945                 }
 946
 947                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 948                 if (count <= 0) {
 949                         err = count;
 950                 } else {
 951                         size_t num_written;
 952
 953                         num_written = fuse_send_write_pages(req, file, inode,
 954                                                             pos, count);
 955                         err = req->out.h.error;
 956                         if (!err) {
 957                                 res += num_written;
 958                                 pos += num_written;
 959
 960                                 /* break out of the loop on short write */
 961                                 if (num_written != count)
 962                                         err = -EIO;
 963                         }
 964                 }
 965                 fuse_put_request(fc, req);
 966         } while (!err && iov_iter_count(ii));
 967
 968         if (res > 0)
 969                 fuse_write_update_size(inode, pos);
 970
 971         fuse_invalidate_attr(inode);
 972
 973         return res > 0 ? res : err;
 974 }
 975
 976 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 977                                    unsigned long nr_segs, loff_t pos)
 978 {
 979         struct file *file = iocb->ki_filp;
 980         struct address_space *mapping = file->f_mapping;
 981         size_t count = 0;
 982         ssize_t written = 0;
 983         struct inode *inode = mapping->host;
 984         ssize_t err;
 985         struct iov_iter i;
 986
 987         WARN_ON(iocb->ki_pos != pos);
 988
 989         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 990         if (err)
 991                 return err;
 992
 993         mutex_lock(&inode->i_mutex);
 994         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 995
 996         /* We can write back this queue in page reclaim */
 997         current->backing_dev_info = mapping->backing_dev_info;
 998
 999         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1000         if (err)
1001                 goto out;
1002
1003         if (count == 0)
1004                 goto out;
1005
1006         err = file_remove_suid(file);
1007         if (err)
1008                 goto out;
1009
1010         file_update_time(file);
1011
1012         iov_iter_init(&i, iov, nr_segs, count, 0);
1013         written = fuse_perform_write(file, mapping, &i, pos);
1014         if (written >= 0)
1015                 iocb->ki_pos = pos + written;
1016
1017 out:
1018         current->backing_dev_info = NULL;
1019         mutex_unlock(&inode->i_mutex);
1020
1021         return written ? written : err;
1022 }
1023
1024 static void fuse_release_user_pages(struct fuse_req *req, int write)
1025 {
1026         unsigned i;
1027
1028         for (i = 0; i < req->num_pages; i++) {
1029                 struct page *page = req->pages[i];
1030                 if (write)
1031                         set_page_dirty_lock(page);
1032                 put_page(page);
1033         }
1034 }
1035
1036 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1037                                size_t *nbytesp, int write)
1038 {
1039         size_t nbytes = *nbytesp;
1040         unsigned long user_addr = (unsigned long) buf;
1041         unsigned offset = user_addr & ~PAGE_MASK;
1042         int npages;
1043
1044         /* Special case for kernel I/O: can copy directly into the buffer */
1045         if (segment_eq(get_fs(), KERNEL_DS)) {
1046                 if (write)
1047                         req->in.args[1].value = (void *) user_addr;
1048                 else
1049                         req->out.args[0].value = (void *) user_addr;
1050
1051                 return 0;
1052         }
1053
1054         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1055         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1056         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1057         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1058         if (npages < 0)
1059                 return npages;
1060
1061         req->num_pages = npages;
1062         req->page_offset = offset;
1063
1064         if (write)
1065                 req->in.argpages = 1;
1066         else
1067                 req->out.argpages = 1;
1068
1069         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1070         *nbytesp = min(*nbytesp, nbytes);
1071
1072         return 0;
1073 }
1074
1075 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1076                        size_t count, loff_t *ppos, int write)
1077 {
1078         struct fuse_file *ff = file->private_data;
1079         struct fuse_conn *fc = ff->fc;
1080         size_t nmax = write ? fc->max_write : fc->max_read;
1081         loff_t pos = *ppos;
1082         ssize_t res = 0;
1083         struct fuse_req *req;
1084
1085         req = fuse_get_req(fc);
1086         if (IS_ERR(req))
1087                 return PTR_ERR(req);
1088
1089         while (count) {
1090                 size_t nres;
1091                 fl_owner_t owner = current->files;
1092                 size_t nbytes = min(count, nmax);
1093                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1094                 if (err) {
1095                         res = err;
1096                         break;
1097                 }
1098
1099                 if (write)
1100                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1101                 else
1102                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1103
1104                 fuse_release_user_pages(req, !write);
1105                 if (req->out.h.error) {
1106                         if (!res)
1107                                 res = req->out.h.error;
1108                         break;
1109                 } else if (nres > nbytes) {
1110                         res = -EIO;
1111                         break;
1112                 }
1113                 count -= nres;
1114                 res += nres;
1115                 pos += nres;
1116                 buf += nres;
1117                 if (nres != nbytes)
1118                         break;
1119                 if (count) {
1120                         fuse_put_request(fc, req);
1121                         req = fuse_get_req(fc);
1122                         if (IS_ERR(req))
1123                                 break;
1124                 }
1125         }
1126         if (!IS_ERR(req))
1127                 fuse_put_request(fc, req);
1128         if (res > 0)
1129                 *ppos = pos;
1130
1131         return res;
1132 }
1133 EXPORT_SYMBOL_GPL(fuse_direct_io);
1134
1135 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1136                                      size_t count, loff_t *ppos)
1137 {
1138         ssize_t res;
1139         struct inode *inode = file->f_path.dentry->d_inode;
1140
1141         if (is_bad_inode(inode))
1142                 return -EIO;
1143
1144         res = fuse_direct_io(file, buf, count, ppos, 0);
1145
1146         fuse_invalidate_attr(inode);
1147
1148         return res;
1149 }
1150
1151 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1152                                  size_t count, loff_t *ppos)
1153 {
1154         struct inode *inode = file->f_path.dentry->d_inode;
1155         ssize_t res;
1156
1157         if (is_bad_inode(inode))
1158                 return -EIO;
1159
1160         /* Don't allow parallel writes to the same file */
1161         mutex_lock(&inode->i_mutex);
1162         res = generic_write_checks(file, ppos, &count, 0);
1163         if (!res) {
1164                 res = fuse_direct_io(file, buf, count, ppos, 1);
1165                 if (res > 0)
1166                         fuse_write_update_size(inode, *ppos);
1167         }
1168         mutex_unlock(&inode->i_mutex);
1169
1170         fuse_invalidate_attr(inode);
1171
1172         return res;
1173 }
1174
1175 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1176 {
1177         __free_page(req->pages[0]);
1178         fuse_file_put(req->ff, false);
1179 }
1180
1181 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1182 {
1183         struct inode *inode = req->inode;
1184         struct fuse_inode *fi = get_fuse_inode(inode);
1185         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1186
1187         list_del(&req->writepages_entry);
1188         dec_bdi_stat(bdi, BDI_WRITEBACK);
1189         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1190         bdi_writeout_inc(bdi);
1191         wake_up(&fi->page_waitq);
1192 }
1193
1194 /* Called under fc->lock, may release and reacquire it */
1195 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1196 __releases(fc->lock)
1197 __acquires(fc->lock)
1198 {
1199         struct fuse_inode *fi = get_fuse_inode(req->inode);
1200         loff_t size = i_size_read(req->inode);
1201         struct fuse_write_in *inarg = &req->misc.write.in;
1202
1203         if (!fc->connected)
1204                 goto out_free;
1205
1206         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1207                 inarg->size = PAGE_CACHE_SIZE;
1208         } else if (inarg->offset < size) {
1209                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1210         } else {
1211                 /* Got truncated off completely */
1212                 goto out_free;
1213         }
1214
1215         req->in.args[1].size = inarg->size;
1216         fi->writectr++;
1217         fuse_request_send_background_locked(fc, req);
1218         return;
1219
1220  out_free:
1221         fuse_writepage_finish(fc, req);
1222         spin_unlock(&fc->lock);
1223         fuse_writepage_free(fc, req);
1224         fuse_put_request(fc, req);
1225         spin_lock(&fc->lock);
1226 }
1227
1228 /*
1229  * If fi->writectr is positive (no truncate or fsync going on) send
1230  * all queued writepage requests.
1231  *
1232  * Called with fc->lock
1233  */
1234 void fuse_flush_writepages(struct inode *inode)
1235 __releases(fc->lock)
1236 __acquires(fc->lock)
1237 {
1238         struct fuse_conn *fc = get_fuse_conn(inode);
1239         struct fuse_inode *fi = get_fuse_inode(inode);
1240         struct fuse_req *req;
1241
1242         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1243                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1244                 list_del_init(&req->list);
1245                 fuse_send_writepage(fc, req);
1246         }
1247 }
1248
1249 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1250 {
1251         struct inode *inode = req->inode;
1252         struct fuse_inode *fi = get_fuse_inode(inode);
1253
1254         mapping_set_error(inode->i_mapping, req->out.h.error);
1255         spin_lock(&fc->lock);
1256         fi->writectr--;
1257         fuse_writepage_finish(fc, req);
1258         spin_unlock(&fc->lock);
1259         fuse_writepage_free(fc, req);
1260 }
1261
1262 static int fuse_writepage_locked(struct page *page)
1263 {
1264         struct address_space *mapping = page->mapping;
1265         struct inode *inode = mapping->host;
1266         struct fuse_conn *fc = get_fuse_conn(inode);
1267         struct fuse_inode *fi = get_fuse_inode(inode);
1268         struct fuse_req *req;
1269         struct fuse_file *ff;
1270         struct page *tmp_page;
1271
1272         set_page_writeback(page);
1273
1274         req = fuse_request_alloc_nofs();
1275         if (!req)
1276                 goto err;
1277
1278         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1279         if (!tmp_page)
1280                 goto err_free;
1281
1282         spin_lock(&fc->lock);
1283         BUG_ON(list_empty(&fi->write_files));
1284         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1285         req->ff = fuse_file_get(ff);
1286         spin_unlock(&fc->lock);
1287
1288         fuse_write_fill(req, ff, page_offset(page), 0);
1289
1290         copy_highpage(tmp_page, page);
1291         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1292         req->in.argpages = 1;
1293         req->num_pages = 1;
1294         req->pages[0] = tmp_page;
1295         req->page_offset = 0;
1296         req->end = fuse_writepage_end;
1297         req->inode = inode;
1298
1299         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1300         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1301         end_page_writeback(page);
1302
1303         spin_lock(&fc->lock);
1304         list_add(&req->writepages_entry, &fi->writepages);
1305         list_add_tail(&req->list, &fi->queued_writes);
1306         fuse_flush_writepages(inode);
1307         spin_unlock(&fc->lock);
1308
1309         return 0;
1310
1311 err_free:
1312         fuse_request_free(req);
1313 err:
1314         end_page_writeback(page);
1315         return -ENOMEM;
1316 }
1317
1318 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1319 {
1320         int err;
1321
1322         err = fuse_writepage_locked(page);
1323         unlock_page(page);
1324
1325         return err;
1326 }
1327
1328 static int fuse_launder_page(struct page *page)
1329 {
1330         int err = 0;
1331         if (clear_page_dirty_for_io(page)) {
1332                 struct inode *inode = page->mapping->host;
1333                 err = fuse_writepage_locked(page);
1334                 if (!err)
1335                         fuse_wait_on_page_writeback(inode, page->index);
1336         }
1337         return err;
1338 }
1339
1340 /*
1341  * Write back dirty pages now, because there may not be any suitable
1342  * open files later
1343  */
1344 static void fuse_vma_close(struct vm_area_struct *vma)
1345 {
1346         filemap_write_and_wait(vma->vm_file->f_mapping);
1347 }
1348
1349 /*
1350  * Wait for writeback against this page to complete before allowing it
1351  * to be marked dirty again, and hence written back again, possibly
1352  * before the previous writepage completed.
1353  *
1354  * Block here, instead of in ->writepage(), so that the userspace fs
1355  * can only block processes actually operating on the filesystem.
1356  *
1357  * Otherwise unprivileged userspace fs would be able to block
1358  * unrelated:
1359  *
1360  * - page migration
1361  * - sync(2)
1362  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1363  */
1364 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1365 {
1366         struct page *page = vmf->page;
1367         /*
1368          * Don't use page->mapping as it may become NULL from a
1369          * concurrent truncate.
1370          */
1371         struct inode *inode = vma->vm_file->f_mapping->host;
1372
1373         fuse_wait_on_page_writeback(inode, page->index);
1374         return 0;
1375 }
1376
1377 static const struct vm_operations_struct fuse_file_vm_ops = {
1378         .close          = fuse_vma_close,
1379         .fault          = filemap_fault,
1380         .page_mkwrite   = fuse_page_mkwrite,
1381 };
1382
1383 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1384 {
1385         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1386                 struct inode *inode = file->f_dentry->d_inode;
1387                 struct fuse_conn *fc = get_fuse_conn(inode);
1388                 struct fuse_inode *fi = get_fuse_inode(inode);
1389                 struct fuse_file *ff = file->private_data;
1390                 /*
1391                  * file may be written through mmap, so chain it onto the
1392                  * inodes's write_file list
1393                  */
1394                 spin_lock(&fc->lock);
1395                 if (list_empty(&ff->write_entry))
1396                         list_add(&ff->write_entry, &fi->write_files);
1397                 spin_unlock(&fc->lock);
1398         }
1399         file_accessed(file);
1400         vma->vm_ops = &fuse_file_vm_ops;
1401         return 0;
1402 }
1403
1404 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1405 {
1406         /* Can't provide the coherency needed for MAP_SHARED */
1407         if (vma->vm_flags & VM_MAYSHARE)
1408                 return -ENODEV;
1409
1410         invalidate_inode_pages2(file->f_mapping);
1411
1412         return generic_file_mmap(file, vma);
1413 }
1414
1415 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1416                                   struct file_lock *fl)
1417 {
1418         switch (ffl->type) {
1419         case F_UNLCK:
1420                 break;
1421
1422         case F_RDLCK:
1423         case F_WRLCK:
1424                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1425                     ffl->end < ffl->start)
1426                         return -EIO;
1427
1428                 fl->fl_start = ffl->start;
1429                 fl->fl_end = ffl->end;
1430                 fl->fl_pid = ffl->pid;
1431                 break;
1432
1433         default:
1434                 return -EIO;
1435         }
1436         fl->fl_type = ffl->type;
1437         return 0;
1438 }
1439
1440 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1441                          const struct file_lock *fl, int opcode, pid_t pid,
1442                          int flock)
1443 {
1444         struct inode *inode = file->f_path.dentry->d_inode;
1445         struct fuse_conn *fc = get_fuse_conn(inode);
1446         struct fuse_file *ff = file->private_data;
1447         struct fuse_lk_in *arg = &req->misc.lk_in;
1448
1449         arg->fh = ff->fh;
1450         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1451         arg->lk.start = fl->fl_start;
1452         arg->lk.end = fl->fl_end;
1453         arg->lk.type = fl->fl_type;
1454         arg->lk.pid = pid;
1455         if (flock)
1456                 arg->lk_flags |= FUSE_LK_FLOCK;
1457         req->in.h.opcode = opcode;
1458         req->in.h.nodeid = get_node_id(inode);
1459         req->in.numargs = 1;
1460         req->in.args[0].size = sizeof(*arg);
1461         req->in.args[0].value = arg;
1462 }
1463
1464 static int fuse_getlk(struct file *file, struct file_lock *fl)
1465 {
1466         struct inode *inode = file->f_path.dentry->d_inode;
1467         struct fuse_conn *fc = get_fuse_conn(inode);
1468         struct fuse_req *req;
1469         struct fuse_lk_out outarg;
1470         int err;
1471
1472         req = fuse_get_req(fc);
1473         if (IS_ERR(req))
1474                 return PTR_ERR(req);
1475
1476         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1477         req->out.numargs = 1;
1478         req->out.args[0].size = sizeof(outarg);
1479         req->out.args[0].value = &outarg;
1480         fuse_request_send(fc, req);
1481         err = req->out.h.error;
1482         fuse_put_request(fc, req);
1483         if (!err)
1484                 err = convert_fuse_file_lock(&outarg.lk, fl);
1485
1486         return err;
1487 }
1488
1489 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1490 {
1491         struct inode *inode = file->f_path.dentry->d_inode;
1492         struct fuse_conn *fc = get_fuse_conn(inode);
1493         struct fuse_req *req;
1494         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1495         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1496         int err;
1497
1498         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1499                 /* NLM needs asynchronous locks, which we don't support yet */
1500                 return -ENOLCK;
1501         }
1502
1503         /* Unlock on close is handled by the flush method */
1504         if (fl->fl_flags & FL_CLOSE)
1505                 return 0;
1506
1507         req = fuse_get_req(fc);
1508         if (IS_ERR(req))
1509                 return PTR_ERR(req);
1510
1511         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1512         fuse_request_send(fc, req);
1513         err = req->out.h.error;
1514         /* locking is restartable */
1515         if (err == -EINTR)
1516                 err = -ERESTARTSYS;
1517         fuse_put_request(fc, req);
1518         return err;
1519 }
1520
1521 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1522 {
1523         struct inode *inode = file->f_path.dentry->d_inode;
1524         struct fuse_conn *fc = get_fuse_conn(inode);
1525         int err;
1526
1527         if (cmd == F_CANCELLK) {
1528                 err = 0;
1529         } else if (cmd == F_GETLK) {
1530                 if (fc->no_lock) {
1531                         posix_test_lock(file, fl);
1532                         err = 0;
1533                 } else
1534                         err = fuse_getlk(file, fl);
1535         } else {
1536                 if (fc->no_lock)
1537                         err = posix_lock_file(file, fl, NULL);
1538                 else
1539                         err = fuse_setlk(file, fl, 0);
1540         }
1541         return err;
1542 }
1543
1544 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1545 {
1546         struct inode *inode = file->f_path.dentry->d_inode;
1547         struct fuse_conn *fc = get_fuse_conn(inode);
1548         int err;
1549
1550         if (fc->no_lock) {
1551                 err = flock_lock_file_wait(file, fl);
1552         } else {
1553                 /* emulate flock with POSIX locks */
1554                 fl->fl_owner = (fl_owner_t) file;
1555                 err = fuse_setlk(file, fl, 1);
1556         }
1557
1558         return err;
1559 }
1560
1561 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1562 {
1563         struct inode *inode = mapping->host;
1564         struct fuse_conn *fc = get_fuse_conn(inode);
1565         struct fuse_req *req;
1566         struct fuse_bmap_in inarg;
1567         struct fuse_bmap_out outarg;
1568         int err;
1569
1570         if (!inode->i_sb->s_bdev || fc->no_bmap)
1571                 return 0;
1572
1573         req = fuse_get_req(fc);
1574         if (IS_ERR(req))
1575                 return 0;
1576
1577         memset(&inarg, 0, sizeof(inarg));
1578         inarg.block = block;
1579         inarg.blocksize = inode->i_sb->s_blocksize;
1580         req->in.h.opcode = FUSE_BMAP;
1581         req->in.h.nodeid = get_node_id(inode);
1582         req->in.numargs = 1;
1583         req->in.args[0].size = sizeof(inarg);
1584         req->in.args[0].value = &inarg;
1585         req->out.numargs = 1;
1586         req->out.args[0].size = sizeof(outarg);
1587         req->out.args[0].value = &outarg;
1588         fuse_request_send(fc, req);
1589         err = req->out.h.error;
1590         fuse_put_request(fc, req);
1591         if (err == -ENOSYS)
1592                 fc->no_bmap = 1;
1593
1594         return err ? 0 : outarg.block;
1595 }
1596
1597 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1598 {
1599         loff_t retval;
1600         struct inode *inode = file->f_path.dentry->d_inode;
1601
1602         mutex_lock(&inode->i_mutex);
1603         if (origin != SEEK_CUR || origin != SEEK_SET) {
1604                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1605                 if (retval)
1606                         goto exit;
1607         }
1608
1609         switch (origin) {
1610         case SEEK_END:
1611                 offset += i_size_read(inode);
1612                 break;
1613         case SEEK_CUR:
1614                 offset += file->f_pos;
1615                 break;
1616         case SEEK_DATA:
1617                 if (offset >= i_size_read(inode)) {
1618                         retval = -ENXIO;
1619                         goto exit;
1620                 }
1621                 break;
1622         case SEEK_HOLE:
1623                 if (offset >= i_size_read(inode)) {
1624                         retval = -ENXIO;
1625                         goto exit;
1626                 }
1627                 offset = i_size_read(inode);
1628                 break;
1629         }
1630         retval = -EINVAL;
1631         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1632                 if (offset != file->f_pos) {
1633                         file->f_pos = offset;
1634                         file->f_version = 0;
1635                 }
1636                 retval = offset;
1637         }
1638 exit:
1639         mutex_unlock(&inode->i_mutex);
1640         return retval;
1641 }
1642
1643 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1644                         unsigned int nr_segs, size_t bytes, bool to_user)
1645 {
1646         struct iov_iter ii;
1647         int page_idx = 0;
1648
1649         if (!bytes)
1650                 return 0;
1651
1652         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1653
1654         while (iov_iter_count(&ii)) {
1655                 struct page *page = pages[page_idx++];
1656                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1657                 void *kaddr;
1658
1659                 kaddr = kmap(page);
1660
1661                 while (todo) {
1662                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1663                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1664                         size_t copy = min(todo, iov_len);
1665                         size_t left;
1666
1667                         if (!to_user)
1668                                 left = copy_from_user(kaddr, uaddr, copy);
1669                         else
1670                                 left = copy_to_user(uaddr, kaddr, copy);
1671
1672                         if (unlikely(left))
1673                                 return -EFAULT;
1674
1675                         iov_iter_advance(&ii, copy);
1676                         todo -= copy;
1677                         kaddr += copy;
1678                 }
1679
1680                 kunmap(page);
1681         }
1682
1683         return 0;
1684 }
1685
1686 /*
1687  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1688  * ABI was defined to be 'struct iovec' which is different on 32bit
1689  * and 64bit.  Fortunately we can determine which structure the server
1690  * used from the size of the reply.
1691  */
1692 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1693                                      size_t transferred, unsigned count,
1694                                      bool is_compat)
1695 {
1696 #ifdef CONFIG_COMPAT
1697         if (count * sizeof(struct compat_iovec) == transferred) {
1698                 struct compat_iovec *ciov = src;
1699                 unsigned i;
1700
1701                 /*
1702                  * With this interface a 32bit server cannot support
1703                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1704                  * requests
1705                  */
1706                 if (!is_compat)
1707                         return -EINVAL;
1708
1709                 for (i = 0; i < count; i++) {
1710                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1711                         dst[i].iov_len = ciov[i].iov_len;
1712                 }
1713                 return 0;
1714         }
1715 #endif
1716
1717         if (count * sizeof(struct iovec) != transferred)
1718                 return -EIO;
1719
1720         memcpy(dst, src, transferred);
1721         return 0;
1722 }
1723
1724 /* Make sure iov_length() won't overflow */
1725 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1726 {
1727         size_t n;
1728         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1729
1730         for (n = 0; n < count; n++) {
1731                 if (iov->iov_len > (size_t) max)
1732                         return -ENOMEM;
1733                 max -= iov->iov_len;
1734         }
1735         return 0;
1736 }
1737
1738 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1739                                  void *src, size_t transferred, unsigned count,
1740                                  bool is_compat)
1741 {
1742         unsigned i;
1743         struct fuse_ioctl_iovec *fiov = src;
1744
1745         if (fc->minor < 16) {
1746                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1747                                                  count, is_compat);
1748         }
1749
1750         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1751                 return -EIO;
1752
1753         for (i = 0; i < count; i++) {
1754                 /* Did the server supply an inappropriate value? */
1755                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1756                     fiov[i].len != (unsigned long) fiov[i].len)
1757                         return -EIO;
1758
1759                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1760                 dst[i].iov_len = (size_t) fiov[i].len;
1761
1762 #ifdef CONFIG_COMPAT
1763                 if (is_compat &&
1764                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1765                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1766                         return -EIO;
1767 #endif
1768         }
1769
1770         return 0;
1771 }
1772
1773
1774 /*
1775  * For ioctls, there is no generic way to determine how much memory
1776  * needs to be read and/or written.  Furthermore, ioctls are allowed
1777  * to dereference the passed pointer, so the parameter requires deep
1778  * copying but FUSE has no idea whatsoever about what to copy in or
1779  * out.
1780  *
1781  * This is solved by allowing FUSE server to retry ioctl with
1782  * necessary in/out iovecs.  Let's assume the ioctl implementation
1783  * needs to read in the following structure.
1784  *
1785  * struct a {
1786  *      char    *buf;
1787  *      size_t  buflen;
1788  * }
1789  *
1790  * On the first callout to FUSE server, inarg->in_size and
1791  * inarg->out_size will be NULL; then, the server completes the ioctl
1792  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1793  * the actual iov array to
1794  *
1795  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1796  *
1797  * which tells FUSE to copy in the requested area and retry the ioctl.
1798  * On the second round, the server has access to the structure and
1799  * from that it can tell what to look for next, so on the invocation,
1800  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1801  *
1802  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1803  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1804  *
1805  * FUSE will copy both struct a and the pointed buffer from the
1806  * process doing the ioctl and retry ioctl with both struct a and the
1807  * buffer.
1808  *
1809  * This time, FUSE server has everything it needs and completes ioctl
1810  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1811  *
1812  * Copying data out works the same way.
1813  *
1814  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1815  * automatically initializes in and out iovs by decoding @cmd with
1816  * _IOC_* macros and the server is not allowed to request RETRY.  This
1817  * limits ioctl data transfers to well-formed ioctls and is the forced
1818  * behavior for all FUSE servers.
1819  */
1820 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1821                    unsigned int flags)
1822 {
1823         struct fuse_file *ff = file->private_data;
1824         struct fuse_conn *fc = ff->fc;
1825         struct fuse_ioctl_in inarg = {
1826                 .fh = ff->fh,
1827                 .cmd = cmd,
1828                 .arg = arg,
1829                 .flags = flags
1830         };
1831         struct fuse_ioctl_out outarg;
1832         struct fuse_req *req = NULL;
1833         struct page **pages = NULL;
1834         struct iovec *iov_page = NULL;
1835         struct iovec *in_iov = NULL, *out_iov = NULL;
1836         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1837         size_t in_size, out_size, transferred;
1838         int err;
1839
1840 #if BITS_PER_LONG == 32
1841         inarg.flags |= FUSE_IOCTL_32BIT;
1842 #else
1843         if (flags & FUSE_IOCTL_COMPAT)
1844                 inarg.flags |= FUSE_IOCTL_32BIT;
1845 #endif
1846
1847         /* assume all the iovs returned by client always fits in a page */
1848         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1849
1850         err = -ENOMEM;
1851         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1852         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1853         if (!pages || !iov_page)
1854                 goto out;
1855
1856         /*
1857          * If restricted, initialize IO parameters as encoded in @cmd.
1858          * RETRY from server is not allowed.
1859          */
1860         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1861                 struct iovec *iov = iov_page;
1862
1863                 iov->iov_base = (void __user *)arg;
1864                 iov->iov_len = _IOC_SIZE(cmd);
1865
1866                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1867                         in_iov = iov;
1868                         in_iovs = 1;
1869                 }
1870
1871                 if (_IOC_DIR(cmd) & _IOC_READ) {
1872                         out_iov = iov;
1873                         out_iovs = 1;
1874                 }
1875         }
1876
1877  retry:
1878         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1879         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1880
1881         /*
1882          * Out data can be used either for actual out data or iovs,
1883          * make sure there always is at least one page.
1884          */
1885         out_size = max_t(size_t, out_size, PAGE_SIZE);
1886         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1887
1888         /* make sure there are enough buffer pages and init request with them */
1889         err = -ENOMEM;
1890         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1891                 goto out;
1892         while (num_pages < max_pages) {
1893                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1894                 if (!pages[num_pages])
1895                         goto out;
1896                 num_pages++;
1897         }
1898
1899         req = fuse_get_req(fc);
1900         if (IS_ERR(req)) {
1901                 err = PTR_ERR(req);
1902                 req = NULL;
1903                 goto out;
1904         }
1905         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1906         req->num_pages = num_pages;
1907
1908         /* okay, let's send it to the client */
1909         req->in.h.opcode = FUSE_IOCTL;
1910         req->in.h.nodeid = ff->nodeid;
1911         req->in.numargs = 1;
1912         req->in.args[0].size = sizeof(inarg);
1913         req->in.args[0].value = &inarg;
1914         if (in_size) {
1915                 req->in.numargs++;
1916                 req->in.args[1].size = in_size;
1917                 req->in.argpages = 1;
1918
1919                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1920                                            false);
1921                 if (err)
1922                         goto out;
1923         }
1924
1925         req->out.numargs = 2;
1926         req->out.args[0].size = sizeof(outarg);
1927         req->out.args[0].value = &outarg;
1928         req->out.args[1].size = out_size;
1929         req->out.argpages = 1;
1930         req->out.argvar = 1;
1931
1932         fuse_request_send(fc, req);
1933         err = req->out.h.error;
1934         transferred = req->out.args[1].size;
1935         fuse_put_request(fc, req);
1936         req = NULL;
1937         if (err)
1938                 goto out;
1939
1940         /* did it ask for retry? */
1941         if (outarg.flags & FUSE_IOCTL_RETRY) {
1942                 void *vaddr;
1943
1944                 /* no retry if in restricted mode */
1945                 err = -EIO;
1946                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1947                         goto out;
1948
1949                 in_iovs = outarg.in_iovs;
1950                 out_iovs = outarg.out_iovs;
1951
1952                 /*
1953                  * Make sure things are in boundary, separate checks
1954                  * are to protect against overflow.
1955                  */
1956                 err = -ENOMEM;
1957                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1958                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1959                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1960                         goto out;
1961
1962                 vaddr = kmap_atomic(pages[0], KM_USER0);
1963                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1964                                             transferred, in_iovs + out_iovs,
1965                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1966                 kunmap_atomic(vaddr, KM_USER0);
1967                 if (err)
1968                         goto out;
1969
1970                 in_iov = iov_page;
1971                 out_iov = in_iov + in_iovs;
1972
1973                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1974                 if (err)
1975                         goto out;
1976
1977                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1978                 if (err)
1979                         goto out;
1980
1981                 goto retry;
1982         }
1983
1984         err = -EIO;
1985         if (transferred > inarg.out_size)
1986                 goto out;
1987
1988         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1989  out:
1990         if (req)
1991                 fuse_put_request(fc, req);
1992         free_page((unsigned long) iov_page);
1993         while (num_pages)
1994                 __free_page(pages[--num_pages]);
1995         kfree(pages);
1996
1997         return err ? err : outarg.result;
1998 }
1999 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2000
2001 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
2002                                    unsigned long arg, unsigned int flags)
2003 {
2004         struct inode *inode = file->f_dentry->d_inode;
2005         struct fuse_conn *fc = get_fuse_conn(inode);
2006
2007         if (!fuse_allow_task(fc, current))
2008                 return -EACCES;
2009
2010         if (is_bad_inode(inode))
2011                 return -EIO;
2012
2013         return fuse_do_ioctl(file, cmd, arg, flags);
2014 }
2015
2016 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2017                             unsigned long arg)
2018 {
2019         return fuse_file_ioctl_common(file, cmd, arg, 0);
2020 }
2021
2022 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2023                                    unsigned long arg)
2024 {
2025         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2026 }
2027
2028 /*
2029  * All files which have been polled are linked to RB tree
2030  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2031  * find the matching one.
2032  */
2033 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2034                                               struct rb_node **parent_out)
2035 {
2036         struct rb_node **link = &fc->polled_files.rb_node;
2037         struct rb_node *last = NULL;
2038
2039         while (*link) {
2040                 struct fuse_file *ff;
2041
2042                 last = *link;
2043                 ff = rb_entry(last, struct fuse_file, polled_node);
2044
2045                 if (kh < ff->kh)
2046                         link = &last->rb_left;
2047                 else if (kh > ff->kh)
2048                         link = &last->rb_right;
2049                 else
2050                         return link;
2051         }
2052
2053         if (parent_out)
2054                 *parent_out = last;
2055         return link;
2056 }
2057
2058 /*
2059  * The file is about to be polled.  Make sure it's on the polled_files
2060  * RB tree.  Note that files once added to the polled_files tree are
2061  * not removed before the file is released.  This is because a file
2062  * polled once is likely to be polled again.
2063  */
2064 static void fuse_register_polled_file(struct fuse_conn *fc,
2065                                       struct fuse_file *ff)
2066 {
2067         spin_lock(&fc->lock);
2068         if (RB_EMPTY_NODE(&ff->polled_node)) {
2069                 struct rb_node **link, *parent;
2070
2071                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2072                 BUG_ON(*link);
2073                 rb_link_node(&ff->polled_node, parent, link);
2074                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2075         }
2076         spin_unlock(&fc->lock);
2077 }
2078
2079 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2080 {
2081         struct fuse_file *ff = file->private_data;
2082         struct fuse_conn *fc = ff->fc;
2083         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2084         struct fuse_poll_out outarg;
2085         struct fuse_req *req;
2086         int err;
2087
2088         if (fc->no_poll)
2089                 return DEFAULT_POLLMASK;
2090
2091         poll_wait(file, &ff->poll_wait, wait);
2092
2093         /*
2094          * Ask for notification iff there's someone waiting for it.
2095          * The client may ignore the flag and always notify.
2096          */
2097         if (waitqueue_active(&ff->poll_wait)) {
2098                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2099                 fuse_register_polled_file(fc, ff);
2100         }
2101
2102         req = fuse_get_req(fc);
2103         if (IS_ERR(req))
2104                 return POLLERR;
2105
2106         req->in.h.opcode = FUSE_POLL;
2107         req->in.h.nodeid = ff->nodeid;
2108         req->in.numargs = 1;
2109         req->in.args[0].size = sizeof(inarg);
2110         req->in.args[0].value = &inarg;
2111         req->out.numargs = 1;
2112         req->out.args[0].size = sizeof(outarg);
2113         req->out.args[0].value = &outarg;
2114         fuse_request_send(fc, req);
2115         err = req->out.h.error;
2116         fuse_put_request(fc, req);
2117
2118         if (!err)
2119                 return outarg.revents;
2120         if (err == -ENOSYS) {
2121                 fc->no_poll = 1;
2122                 return DEFAULT_POLLMASK;
2123         }
2124         return POLLERR;
2125 }
2126 EXPORT_SYMBOL_GPL(fuse_file_poll);
2127
2128 /*
2129  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2130  * wakes up the poll waiters.
2131  */
2132 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2133                             struct fuse_notify_poll_wakeup_out *outarg)
2134 {
2135         u64 kh = outarg->kh;
2136         struct rb_node **link;
2137
2138         spin_lock(&fc->lock);
2139
2140         link = fuse_find_polled_node(fc, kh, NULL);
2141         if (*link) {
2142                 struct fuse_file *ff;
2143
2144                 ff = rb_entry(*link, struct fuse_file, polled_node);
2145                 wake_up_interruptible_sync(&ff->poll_wait);
2146         }
2147
2148         spin_unlock(&fc->lock);
2149         return 0;
2150 }
2151
2152 static const struct file_operations fuse_file_operations = {
2153         .llseek         = fuse_file_llseek,
2154         .read           = do_sync_read,
2155         .aio_read       = fuse_file_aio_read,
2156         .write          = do_sync_write,
2157         .aio_write      = fuse_file_aio_write,
2158         .mmap           = fuse_file_mmap,
2159         .open           = fuse_open,
2160         .flush          = fuse_flush,
2161         .release        = fuse_release,
2162         .fsync          = fuse_fsync,
2163         .lock           = fuse_file_lock,
2164         .flock          = fuse_file_flock,
2165         .splice_read    = generic_file_splice_read,
2166         .unlocked_ioctl = fuse_file_ioctl,
2167         .compat_ioctl   = fuse_file_compat_ioctl,
2168         .poll           = fuse_file_poll,
2169 };
2170
2171 static const struct file_operations fuse_direct_io_file_operations = {
2172         .llseek         = fuse_file_llseek,
2173         .read           = fuse_direct_read,
2174         .write          = fuse_direct_write,
2175         .mmap           = fuse_direct_mmap,
2176         .open           = fuse_open,
2177         .flush          = fuse_flush,
2178         .release        = fuse_release,
2179         .fsync          = fuse_fsync,
2180         .lock           = fuse_file_lock,
2181         .flock          = fuse_file_flock,
2182         .unlocked_ioctl = fuse_file_ioctl,
2183         .compat_ioctl   = fuse_file_compat_ioctl,
2184         .poll           = fuse_file_poll,
2185         /* no splice_read */
2186 };
2187
2188 static const struct address_space_operations fuse_file_aops  = {
2189         .readpage       = fuse_readpage,
2190         .writepage      = fuse_writepage,
2191         .launder_page   = fuse_launder_page,
2192         .write_begin    = fuse_write_begin,
2193         .write_end      = fuse_write_end,
2194         .readpages      = fuse_readpages,
2195         .set_page_dirty = __set_page_dirty_nobuffers,
2196         .bmap           = fuse_bmap,
2197 };
2198
2199 void fuse_init_file_inode(struct inode *inode)
2200 {
2201         inode->i_fop = &fuse_file_operations;
2202         inode->i_data.a_ops = &fuse_file_aops;
2203 }