fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18
  19 static const struct file_operations fuse_direct_io_file_operations;
  20
  21 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  22                           int opcode, struct fuse_open_out *outargp)
  23 {
  24         struct fuse_open_in inarg;
  25         struct fuse_req *req;
  26         int err;
  27
  28         req = fuse_get_req_nopages(fc);
  29         if (IS_ERR(req))
  30                 return PTR_ERR(req);
  31
  32         memset(&inarg, 0, sizeof(inarg));
  33         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  34         if (!fc->atomic_o_trunc)
  35                 inarg.flags &= ~O_TRUNC;
  36         req->in.h.opcode = opcode;
  37         req->in.h.nodeid = nodeid;
  38         req->in.numargs = 1;
  39         req->in.args[0].size = sizeof(inarg);
  40         req->in.args[0].value = &inarg;
  41         req->out.numargs = 1;
  42         req->out.args[0].size = sizeof(*outargp);
  43         req->out.args[0].value = outargp;
  44         fuse_request_send(fc, req);
  45         err = req->out.h.error;
  46         fuse_put_request(fc, req);
  47
  48         return err;
  49 }
  50
  51 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  52 {
  53         struct fuse_file *ff;
  54
  55         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  56         if (unlikely(!ff))
  57                 return NULL;
  58
  59         ff->fc = fc;
  60         ff->reserved_req = fuse_request_alloc(0);
  61         if (unlikely(!ff->reserved_req)) {
  62                 kfree(ff);
  63                 return NULL;
  64         }
  65
  66         INIT_LIST_HEAD(&ff->write_entry);
  67         atomic_set(&ff->count, 0);
  68         RB_CLEAR_NODE(&ff->polled_node);
  69         init_waitqueue_head(&ff->poll_wait);
  70
  71         spin_lock(&fc->lock);
  72         ff->kh = ++fc->khctr;
  73         spin_unlock(&fc->lock);
  74
  75         return ff;
  76 }
  77
  78 void fuse_file_free(struct fuse_file *ff)
  79 {
  80         fuse_request_free(ff->reserved_req);
  81         kfree(ff);
  82 }
  83
  84 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  85 {
  86         atomic_inc(&ff->count);
  87         return ff;
  88 }
  89
  90 static void fuse_release_async(struct work_struct *work)
  91 {
  92         struct fuse_req *req;
  93         struct fuse_conn *fc;
  94         struct path path;
  95
  96         req = container_of(work, struct fuse_req, misc.release.work);
  97         path = req->misc.release.path;
  98         fc = get_fuse_conn(path.dentry->d_inode);
  99
 100         fuse_put_request(fc, req);
 101         path_put(&path);
 102 }
 103
 104 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 105 {
 106         if (fc->destroy_req) {
 107                 /*
 108                  * If this is a fuseblk mount, then it's possible that
 109                  * releasing the path will result in releasing the
 110                  * super block and sending the DESTROY request.  If
 111                  * the server is single threaded, this would hang.
 112                  * For this reason do the path_put() in a separate
 113                  * thread.
 114                  */
 115                 atomic_inc(&req->count);
 116                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 117                 schedule_work(&req->misc.release.work);
 118         } else {
 119                 path_put(&req->misc.release.path);
 120         }
 121 }
 122
 123 static void fuse_file_put(struct fuse_file *ff, bool sync)
 124 {
 125         if (atomic_dec_and_test(&ff->count)) {
 126                 struct fuse_req *req = ff->reserved_req;
 127
 128                 if (sync) {
 129                         fuse_request_send(ff->fc, req);
 130                         path_put(&req->misc.release.path);
 131                         fuse_put_request(ff->fc, req);
 132                 } else {
 133                         req->end = fuse_release_end;
 134                         fuse_request_send_background(ff->fc, req);
 135                 }
 136                 kfree(ff);
 137         }
 138 }
 139
 140 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 141                  bool isdir)
 142 {
 143         struct fuse_open_out outarg;
 144         struct fuse_file *ff;
 145         int err;
 146         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 147
 148         ff = fuse_file_alloc(fc);
 149         if (!ff)
 150                 return -ENOMEM;
 151
 152         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 153         if (err) {
 154                 fuse_file_free(ff);
 155                 return err;
 156         }
 157
 158         if (isdir)
 159                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 160
 161         ff->fh = outarg.fh;
 162         ff->nodeid = nodeid;
 163         ff->open_flags = outarg.open_flags;
 164         file->private_data = fuse_file_get(ff);
 165
 166         return 0;
 167 }
 168 EXPORT_SYMBOL_GPL(fuse_do_open);
 169
 170 void fuse_finish_open(struct inode *inode, struct file *file)
 171 {
 172         struct fuse_file *ff = file->private_data;
 173         struct fuse_conn *fc = get_fuse_conn(inode);
 174
 175         if (ff->open_flags & FOPEN_DIRECT_IO)
 176                 file->f_op = &fuse_direct_io_file_operations;
 177         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 178                 invalidate_inode_pages2(inode->i_mapping);
 179         if (ff->open_flags & FOPEN_NONSEEKABLE)
 180                 nonseekable_open(inode, file);
 181         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 182                 struct fuse_inode *fi = get_fuse_inode(inode);
 183
 184                 spin_lock(&fc->lock);
 185                 fi->attr_version = ++fc->attr_version;
 186                 i_size_write(inode, 0);
 187                 spin_unlock(&fc->lock);
 188                 fuse_invalidate_attr(inode);
 189         }
 190 }
 191
 192 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 193 {
 194         struct fuse_conn *fc = get_fuse_conn(inode);
 195         int err;
 196
 197         err = generic_file_open(inode, file);
 198         if (err)
 199                 return err;
 200
 201         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 202         if (err)
 203                 return err;
 204
 205         fuse_finish_open(inode, file);
 206
 207         return 0;
 208 }
 209
 210 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 211 {
 212         struct fuse_conn *fc = ff->fc;
 213         struct fuse_req *req = ff->reserved_req;
 214         struct fuse_release_in *inarg = &req->misc.release.in;
 215
 216         spin_lock(&fc->lock);
 217         list_del(&ff->write_entry);
 218         if (!RB_EMPTY_NODE(&ff->polled_node))
 219                 rb_erase(&ff->polled_node, &fc->polled_files);
 220         spin_unlock(&fc->lock);
 221
 222         wake_up_interruptible_all(&ff->poll_wait);
 223
 224         inarg->fh = ff->fh;
 225         inarg->flags = flags;
 226         req->in.h.opcode = opcode;
 227         req->in.h.nodeid = ff->nodeid;
 228         req->in.numargs = 1;
 229         req->in.args[0].size = sizeof(struct fuse_release_in);
 230         req->in.args[0].value = inarg;
 231 }
 232
 233 void fuse_release_common(struct file *file, int opcode)
 234 {
 235         struct fuse_file *ff;
 236         struct fuse_req *req;
 237
 238         ff = file->private_data;
 239         if (unlikely(!ff))
 240                 return;
 241
 242         req = ff->reserved_req;
 243         fuse_prepare_release(ff, file->f_flags, opcode);
 244
 245         if (ff->flock) {
 246                 struct fuse_release_in *inarg = &req->misc.release.in;
 247                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 248                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 249                                                        (fl_owner_t) file);
 250         }
 251         /* Hold vfsmount and dentry until release is finished */
 252         path_get(&file->f_path);
 253         req->misc.release.path = file->f_path;
 254
 255         /*
 256          * Normally this will send the RELEASE request, however if
 257          * some asynchronous READ or WRITE requests are outstanding,
 258          * the sending will be delayed.
 259          *
 260          * Make the release synchronous if this is a fuseblk mount,
 261          * synchronous RELEASE is allowed (and desirable) in this case
 262          * because the server can be trusted not to screw up.
 263          */
 264         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 265 }
 266
 267 static int fuse_open(struct inode *inode, struct file *file)
 268 {
 269         return fuse_open_common(inode, file, false);
 270 }
 271
 272 static int fuse_release(struct inode *inode, struct file *file)
 273 {
 274         fuse_release_common(file, FUSE_RELEASE);
 275
 276         /* return value is ignored by VFS */
 277         return 0;
 278 }
 279
 280 void fuse_sync_release(struct fuse_file *ff, int flags)
 281 {
 282         WARN_ON(atomic_read(&ff->count) > 1);
 283         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 284         ff->reserved_req->force = 1;
 285         fuse_request_send(ff->fc, ff->reserved_req);
 286         fuse_put_request(ff->fc, ff->reserved_req);
 287         kfree(ff);
 288 }
 289 EXPORT_SYMBOL_GPL(fuse_sync_release);
 290
 291 /*
 292  * Scramble the ID space with XTEA, so that the value of the files_struct
 293  * pointer is not exposed to userspace.
 294  */
 295 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 296 {
 297         u32 *k = fc->scramble_key;
 298         u64 v = (unsigned long) id;
 299         u32 v0 = v;
 300         u32 v1 = v >> 32;
 301         u32 sum = 0;
 302         int i;
 303
 304         for (i = 0; i < 32; i++) {
 305                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 306                 sum += 0x9E3779B9;
 307                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 308         }
 309
 310         return (u64) v0 + ((u64) v1 << 32);
 311 }
 312
 313 /*
 314  * Check if page is under writeback
 315  *
 316  * This is currently done by walking the list of writepage requests
 317  * for the inode, which can be pretty inefficient.
 318  */
 319 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 320 {
 321         struct fuse_conn *fc = get_fuse_conn(inode);
 322         struct fuse_inode *fi = get_fuse_inode(inode);
 323         struct fuse_req *req;
 324         bool found = false;
 325
 326         spin_lock(&fc->lock);
 327         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 328                 pgoff_t curr_index;
 329
 330                 BUG_ON(req->inode != inode);
 331                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 332                 if (curr_index == index) {
 333                         found = true;
 334                         break;
 335                 }
 336         }
 337         spin_unlock(&fc->lock);
 338
 339         return found;
 340 }
 341
 342 /*
 343  * Wait for page writeback to be completed.
 344  *
 345  * Since fuse doesn't rely on the VM writeback tracking, this has to
 346  * use some other means.
 347  */
 348 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 349 {
 350         struct fuse_inode *fi = get_fuse_inode(inode);
 351
 352         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 353         return 0;
 354 }
 355
 356 static int fuse_flush(struct file *file, fl_owner_t id)
 357 {
 358         struct inode *inode = file->f_path.dentry->d_inode;
 359         struct fuse_conn *fc = get_fuse_conn(inode);
 360         struct fuse_file *ff = file->private_data;
 361         struct fuse_req *req;
 362         struct fuse_flush_in inarg;
 363         int err;
 364
 365         if (is_bad_inode(inode))
 366                 return -EIO;
 367
 368         if (fc->no_flush)
 369                 return 0;
 370
 371         req = fuse_get_req_nofail_nopages(fc, file);
 372         memset(&inarg, 0, sizeof(inarg));
 373         inarg.fh = ff->fh;
 374         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 375         req->in.h.opcode = FUSE_FLUSH;
 376         req->in.h.nodeid = get_node_id(inode);
 377         req->in.numargs = 1;
 378         req->in.args[0].size = sizeof(inarg);
 379         req->in.args[0].value = &inarg;
 380         req->force = 1;
 381         fuse_request_send(fc, req);
 382         err = req->out.h.error;
 383         fuse_put_request(fc, req);
 384         if (err == -ENOSYS) {
 385                 fc->no_flush = 1;
 386                 err = 0;
 387         }
 388         return err;
 389 }
 390
 391 /*
 392  * Wait for all pending writepages on the inode to finish.
 393  *
 394  * This is currently done by blocking further writes with FUSE_NOWRITE
 395  * and waiting for all sent writes to complete.
 396  *
 397  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 398  * could conflict with truncation.
 399  */
 400 static void fuse_sync_writes(struct inode *inode)
 401 {
 402         fuse_set_nowrite(inode);
 403         fuse_release_nowrite(inode);
 404 }
 405
 406 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 407                       int datasync, int isdir)
 408 {
 409         struct inode *inode = file->f_mapping->host;
 410         struct fuse_conn *fc = get_fuse_conn(inode);
 411         struct fuse_file *ff = file->private_data;
 412         struct fuse_req *req;
 413         struct fuse_fsync_in inarg;
 414         int err;
 415
 416         if (is_bad_inode(inode))
 417                 return -EIO;
 418
 419         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 420         if (err)
 421                 return err;
 422
 423         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 424                 return 0;
 425
 426         mutex_lock(&inode->i_mutex);
 427
 428         /*
 429          * Start writeback against all dirty pages of the inode, then
 430          * wait for all outstanding writes, before sending the FSYNC
 431          * request.
 432          */
 433         err = write_inode_now(inode, 0);
 434         if (err)
 435                 goto out;
 436
 437         fuse_sync_writes(inode);
 438
 439         req = fuse_get_req_nopages(fc);
 440         if (IS_ERR(req)) {
 441                 err = PTR_ERR(req);
 442                 goto out;
 443         }
 444
 445         memset(&inarg, 0, sizeof(inarg));
 446         inarg.fh = ff->fh;
 447         inarg.fsync_flags = datasync ? 1 : 0;
 448         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 449         req->in.h.nodeid = get_node_id(inode);
 450         req->in.numargs = 1;
 451         req->in.args[0].size = sizeof(inarg);
 452         req->in.args[0].value = &inarg;
 453         fuse_request_send(fc, req);
 454         err = req->out.h.error;
 455         fuse_put_request(fc, req);
 456         if (err == -ENOSYS) {
 457                 if (isdir)
 458                         fc->no_fsyncdir = 1;
 459                 else
 460                         fc->no_fsync = 1;
 461                 err = 0;
 462         }
 463 out:
 464         mutex_unlock(&inode->i_mutex);
 465         return err;
 466 }
 467
 468 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 469                       int datasync)
 470 {
 471         return fuse_fsync_common(file, start, end, datasync, 0);
 472 }
 473
 474 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 475                     size_t count, int opcode)
 476 {
 477         struct fuse_read_in *inarg = &req->misc.read.in;
 478         struct fuse_file *ff = file->private_data;
 479
 480         inarg->fh = ff->fh;
 481         inarg->offset = pos;
 482         inarg->size = count;
 483         inarg->flags = file->f_flags;
 484         req->in.h.opcode = opcode;
 485         req->in.h.nodeid = ff->nodeid;
 486         req->in.numargs = 1;
 487         req->in.args[0].size = sizeof(struct fuse_read_in);
 488         req->in.args[0].value = inarg;
 489         req->out.argvar = 1;
 490         req->out.numargs = 1;
 491         req->out.args[0].size = count;
 492 }
 493
 494 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 495                              loff_t pos, size_t count, fl_owner_t owner)
 496 {
 497         struct fuse_file *ff = file->private_data;
 498         struct fuse_conn *fc = ff->fc;
 499
 500         fuse_read_fill(req, file, pos, count, FUSE_READ);
 501         if (owner != NULL) {
 502                 struct fuse_read_in *inarg = &req->misc.read.in;
 503
 504                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 505                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 506         }
 507         fuse_request_send(fc, req);
 508         return req->out.args[0].size;
 509 }
 510
 511 static void fuse_read_update_size(struct inode *inode, loff_t size,
 512                                   u64 attr_ver)
 513 {
 514         struct fuse_conn *fc = get_fuse_conn(inode);
 515         struct fuse_inode *fi = get_fuse_inode(inode);
 516
 517         spin_lock(&fc->lock);
 518         if (attr_ver == fi->attr_version && size < inode->i_size) {
 519                 fi->attr_version = ++fc->attr_version;
 520                 i_size_write(inode, size);
 521         }
 522         spin_unlock(&fc->lock);
 523 }
 524
 525 static int fuse_readpage(struct file *file, struct page *page)
 526 {
 527         struct inode *inode = page->mapping->host;
 528         struct fuse_conn *fc = get_fuse_conn(inode);
 529         struct fuse_req *req;
 530         size_t num_read;
 531         loff_t pos = page_offset(page);
 532         size_t count = PAGE_CACHE_SIZE;
 533         u64 attr_ver;
 534         int err;
 535
 536         err = -EIO;
 537         if (is_bad_inode(inode))
 538                 goto out;
 539
 540         /*
 541          * Page writeback can extend beyond the lifetime of the
 542          * page-cache page, so make sure we read a properly synced
 543          * page.
 544          */
 545         fuse_wait_on_page_writeback(inode, page->index);
 546
 547         req = fuse_get_req(fc, 1);
 548         err = PTR_ERR(req);
 549         if (IS_ERR(req))
 550                 goto out;
 551
 552         attr_ver = fuse_get_attr_version(fc);
 553
 554         req->out.page_zeroing = 1;
 555         req->out.argpages = 1;
 556         req->num_pages = 1;
 557         req->pages[0] = page;
 558         num_read = fuse_send_read(req, file, pos, count, NULL);
 559         err = req->out.h.error;
 560         fuse_put_request(fc, req);
 561
 562         if (!err) {
 563                 /*
 564                  * Short read means EOF.  If file size is larger, truncate it
 565                  */
 566                 if (num_read < count)
 567                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 568
 569                 SetPageUptodate(page);
 570         }
 571
 572         fuse_invalidate_attr(inode); /* atime changed */
 573  out:
 574         unlock_page(page);
 575         return err;
 576 }
 577
 578 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 579 {
 580         int i;
 581         size_t count = req->misc.read.in.size;
 582         size_t num_read = req->out.args[0].size;
 583         struct address_space *mapping = NULL;
 584
 585         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 586                 mapping = req->pages[i]->mapping;
 587
 588         if (mapping) {
 589                 struct inode *inode = mapping->host;
 590
 591                 /*
 592                  * Short read means EOF. If file size is larger, truncate it
 593                  */
 594                 if (!req->out.h.error && num_read < count) {
 595                         loff_t pos;
 596
 597                         pos = page_offset(req->pages[0]) + num_read;
 598                         fuse_read_update_size(inode, pos,
 599                                               req->misc.read.attr_ver);
 600                 }
 601                 fuse_invalidate_attr(inode); /* atime changed */
 602         }
 603
 604         for (i = 0; i < req->num_pages; i++) {
 605                 struct page *page = req->pages[i];
 606                 if (!req->out.h.error)
 607                         SetPageUptodate(page);
 608                 else
 609                         SetPageError(page);
 610                 unlock_page(page);
 611                 page_cache_release(page);
 612         }
 613         if (req->ff)
 614                 fuse_file_put(req->ff, false);
 615 }
 616
 617 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 618 {
 619         struct fuse_file *ff = file->private_data;
 620         struct fuse_conn *fc = ff->fc;
 621         loff_t pos = page_offset(req->pages[0]);
 622         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 623
 624         req->out.argpages = 1;
 625         req->out.page_zeroing = 1;
 626         req->out.page_replace = 1;
 627         fuse_read_fill(req, file, pos, count, FUSE_READ);
 628         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 629         if (fc->async_read) {
 630                 req->ff = fuse_file_get(ff);
 631                 req->end = fuse_readpages_end;
 632                 fuse_request_send_background(fc, req);
 633         } else {
 634                 fuse_request_send(fc, req);
 635                 fuse_readpages_end(fc, req);
 636                 fuse_put_request(fc, req);
 637         }
 638 }
 639
 640 struct fuse_fill_data {
 641         struct fuse_req *req;
 642         struct file *file;
 643         struct inode *inode;
 644         unsigned nr_pages;
 645 };
 646
 647 static int fuse_readpages_fill(void *_data, struct page *page)
 648 {
 649         struct fuse_fill_data *data = _data;
 650         struct fuse_req *req = data->req;
 651         struct inode *inode = data->inode;
 652         struct fuse_conn *fc = get_fuse_conn(inode);
 653
 654         fuse_wait_on_page_writeback(inode, page->index);
 655
 656         if (req->num_pages &&
 657             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 658              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 659              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 660                 int nr_alloc = min_t(unsigned, data->nr_pages,
 661                                      FUSE_MAX_PAGES_PER_REQ);
 662                 fuse_send_readpages(req, data->file);
 663                 data->req = req = fuse_get_req(fc, nr_alloc);
 664                 if (IS_ERR(req)) {
 665                         unlock_page(page);
 666                         return PTR_ERR(req);
 667                 }
 668         }
 669
 670         if (WARN_ON(req->num_pages >= req->max_pages)) {
 671                 fuse_put_request(fc, req);
 672                 return -EIO;
 673         }
 674
 675         page_cache_get(page);
 676         req->pages[req->num_pages] = page;
 677         req->num_pages++;
 678         data->nr_pages--;
 679         return 0;
 680 }
 681
 682 static int fuse_readpages(struct file *file, struct address_space *mapping,
 683                           struct list_head *pages, unsigned nr_pages)
 684 {
 685         struct inode *inode = mapping->host;
 686         struct fuse_conn *fc = get_fuse_conn(inode);
 687         struct fuse_fill_data data;
 688         int err;
 689         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 690
 691         err = -EIO;
 692         if (is_bad_inode(inode))
 693                 goto out;
 694
 695         data.file = file;
 696         data.inode = inode;
 697         data.req = fuse_get_req(fc, nr_alloc);
 698         data.nr_pages = nr_pages;
 699         err = PTR_ERR(data.req);
 700         if (IS_ERR(data.req))
 701                 goto out;
 702
 703         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 704         if (!err) {
 705                 if (data.req->num_pages)
 706                         fuse_send_readpages(data.req, file);
 707                 else
 708                         fuse_put_request(fc, data.req);
 709         }
 710 out:
 711         return err;
 712 }
 713
 714 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 715                                   unsigned long nr_segs, loff_t pos)
 716 {
 717         struct inode *inode = iocb->ki_filp->f_mapping->host;
 718         struct fuse_conn *fc = get_fuse_conn(inode);
 719
 720         /*
 721          * In auto invalidate mode, always update attributes on read.
 722          * Otherwise, only update if we attempt to read past EOF (to ensure
 723          * i_size is up to date).
 724          */
 725         if (fc->auto_inval_data ||
 726             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 727                 int err;
 728                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 729                 if (err)
 730                         return err;
 731         }
 732
 733         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 734 }
 735
 736 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 737                             loff_t pos, size_t count)
 738 {
 739         struct fuse_write_in *inarg = &req->misc.write.in;
 740         struct fuse_write_out *outarg = &req->misc.write.out;
 741
 742         inarg->fh = ff->fh;
 743         inarg->offset = pos;
 744         inarg->size = count;
 745         req->in.h.opcode = FUSE_WRITE;
 746         req->in.h.nodeid = ff->nodeid;
 747         req->in.numargs = 2;
 748         if (ff->fc->minor < 9)
 749                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 750         else
 751                 req->in.args[0].size = sizeof(struct fuse_write_in);
 752         req->in.args[0].value = inarg;
 753         req->in.args[1].size = count;
 754         req->out.numargs = 1;
 755         req->out.args[0].size = sizeof(struct fuse_write_out);
 756         req->out.args[0].value = outarg;
 757 }
 758
 759 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 760                               loff_t pos, size_t count, fl_owner_t owner)
 761 {
 762         struct fuse_file *ff = file->private_data;
 763         struct fuse_conn *fc = ff->fc;
 764         struct fuse_write_in *inarg = &req->misc.write.in;
 765
 766         fuse_write_fill(req, ff, pos, count);
 767         inarg->flags = file->f_flags;
 768         if (owner != NULL) {
 769                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 770                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 771         }
 772         fuse_request_send(fc, req);
 773         return req->misc.write.out.size;
 774 }
 775
 776 void fuse_write_update_size(struct inode *inode, loff_t pos)
 777 {
 778         struct fuse_conn *fc = get_fuse_conn(inode);
 779         struct fuse_inode *fi = get_fuse_inode(inode);
 780
 781         spin_lock(&fc->lock);
 782         fi->attr_version = ++fc->attr_version;
 783         if (pos > inode->i_size)
 784                 i_size_write(inode, pos);
 785         spin_unlock(&fc->lock);
 786 }
 787
 788 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 789                                     struct inode *inode, loff_t pos,
 790                                     size_t count)
 791 {
 792         size_t res;
 793         unsigned offset;
 794         unsigned i;
 795
 796         for (i = 0; i < req->num_pages; i++)
 797                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 798
 799         res = fuse_send_write(req, file, pos, count, NULL);
 800
 801         offset = req->page_offset;
 802         count = res;
 803         for (i = 0; i < req->num_pages; i++) {
 804                 struct page *page = req->pages[i];
 805
 806                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 807                         SetPageUptodate(page);
 808
 809                 if (count > PAGE_CACHE_SIZE - offset)
 810                         count -= PAGE_CACHE_SIZE - offset;
 811                 else
 812                         count = 0;
 813                 offset = 0;
 814
 815                 unlock_page(page);
 816                 page_cache_release(page);
 817         }
 818
 819         return res;
 820 }
 821
 822 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 823                                struct address_space *mapping,
 824                                struct iov_iter *ii, loff_t pos)
 825 {
 826         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 827         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 828         size_t count = 0;
 829         int err;
 830
 831         req->in.argpages = 1;
 832         req->page_offset = offset;
 833
 834         do {
 835                 size_t tmp;
 836                 struct page *page;
 837                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 838                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 839                                      iov_iter_count(ii));
 840
 841                 bytes = min_t(size_t, bytes, fc->max_write - count);
 842
 843  again:
 844                 err = -EFAULT;
 845                 if (iov_iter_fault_in_readable(ii, bytes))
 846                         break;
 847
 848                 err = -ENOMEM;
 849                 page = grab_cache_page_write_begin(mapping, index, 0);
 850                 if (!page)
 851                         break;
 852
 853                 if (mapping_writably_mapped(mapping))
 854                         flush_dcache_page(page);
 855
 856                 pagefault_disable();
 857                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 858                 pagefault_enable();
 859                 flush_dcache_page(page);
 860
 861                 mark_page_accessed(page);
 862
 863                 if (!tmp) {
 864                         unlock_page(page);
 865                         page_cache_release(page);
 866                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 867                         goto again;
 868                 }
 869
 870                 err = 0;
 871                 req->pages[req->num_pages] = page;
 872                 req->num_pages++;
 873
 874                 iov_iter_advance(ii, tmp);
 875                 count += tmp;
 876                 pos += tmp;
 877                 offset += tmp;
 878                 if (offset == PAGE_CACHE_SIZE)
 879                         offset = 0;
 880
 881                 if (!fc->big_writes)
 882                         break;
 883         } while (iov_iter_count(ii) && count < fc->max_write &&
 884                  req->num_pages < req->max_pages && offset == 0);
 885
 886         return count > 0 ? count : err;
 887 }
 888
 889 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
 890 {
 891         return min_t(unsigned,
 892                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
 893                      (pos >> PAGE_CACHE_SHIFT) + 1,
 894                      FUSE_MAX_PAGES_PER_REQ);
 895 }
 896
 897 static ssize_t fuse_perform_write(struct file *file,
 898                                   struct address_space *mapping,
 899                                   struct iov_iter *ii, loff_t pos)
 900 {
 901         struct inode *inode = mapping->host;
 902         struct fuse_conn *fc = get_fuse_conn(inode);
 903         int err = 0;
 904         ssize_t res = 0;
 905
 906         if (is_bad_inode(inode))
 907                 return -EIO;
 908
 909         do {
 910                 struct fuse_req *req;
 911                 ssize_t count;
 912                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
 913
 914                 req = fuse_get_req(fc, nr_pages);
 915                 if (IS_ERR(req)) {
 916                         err = PTR_ERR(req);
 917                         break;
 918                 }
 919
 920                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 921                 if (count <= 0) {
 922                         err = count;
 923                 } else {
 924                         size_t num_written;
 925
 926                         num_written = fuse_send_write_pages(req, file, inode,
 927                                                             pos, count);
 928                         err = req->out.h.error;
 929                         if (!err) {
 930                                 res += num_written;
 931                                 pos += num_written;
 932
 933                                 /* break out of the loop on short write */
 934                                 if (num_written != count)
 935                                         err = -EIO;
 936                         }
 937                 }
 938                 fuse_put_request(fc, req);
 939         } while (!err && iov_iter_count(ii));
 940
 941         if (res > 0)
 942                 fuse_write_update_size(inode, pos);
 943
 944         fuse_invalidate_attr(inode);
 945
 946         return res > 0 ? res : err;
 947 }
 948
 949 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 950                                    unsigned long nr_segs, loff_t pos)
 951 {
 952         struct file *file = iocb->ki_filp;
 953         struct address_space *mapping = file->f_mapping;
 954         size_t count = 0;
 955         size_t ocount = 0;
 956         ssize_t written = 0;
 957         ssize_t written_buffered = 0;
 958         struct inode *inode = mapping->host;
 959         ssize_t err;
 960         struct iov_iter i;
 961         loff_t endbyte = 0;
 962
 963         WARN_ON(iocb->ki_pos != pos);
 964
 965         ocount = 0;
 966         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 967         if (err)
 968                 return err;
 969
 970         count = ocount;
 971         sb_start_write(inode->i_sb);
 972         mutex_lock(&inode->i_mutex);
 973
 974         /* We can write back this queue in page reclaim */
 975         current->backing_dev_info = mapping->backing_dev_info;
 976
 977         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 978         if (err)
 979                 goto out;
 980
 981         if (count == 0)
 982                 goto out;
 983
 984         err = file_remove_suid(file);
 985         if (err)
 986                 goto out;
 987
 988         err = file_update_time(file);
 989         if (err)
 990                 goto out;
 991
 992         if (file->f_flags & O_DIRECT) {
 993                 written = generic_file_direct_write(iocb, iov, &nr_segs,
 994                                                     pos, &iocb->ki_pos,
 995                                                     count, ocount);
 996                 if (written < 0 || written == count)
 997                         goto out;
 998
 999                 pos += written;
1000                 count -= written;
1001
1002                 iov_iter_init(&i, iov, nr_segs, count, written);
1003                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1004                 if (written_buffered < 0) {
1005                         err = written_buffered;
1006                         goto out;
1007                 }
1008                 endbyte = pos + written_buffered - 1;
1009
1010                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1011                                                    endbyte);
1012                 if (err)
1013                         goto out;
1014
1015                 invalidate_mapping_pages(file->f_mapping,
1016                                          pos >> PAGE_CACHE_SHIFT,
1017                                          endbyte >> PAGE_CACHE_SHIFT);
1018
1019                 written += written_buffered;
1020                 iocb->ki_pos = pos + written_buffered;
1021         } else {
1022                 iov_iter_init(&i, iov, nr_segs, count, 0);
1023                 written = fuse_perform_write(file, mapping, &i, pos);
1024                 if (written >= 0)
1025                         iocb->ki_pos = pos + written;
1026         }
1027 out:
1028         current->backing_dev_info = NULL;
1029         mutex_unlock(&inode->i_mutex);
1030         sb_end_write(inode->i_sb);
1031
1032         return written ? written : err;
1033 }
1034
1035 static void fuse_release_user_pages(struct fuse_req *req, int write)
1036 {
1037         unsigned i;
1038
1039         for (i = 0; i < req->num_pages; i++) {
1040                 struct page *page = req->pages[i];
1041                 if (write)
1042                         set_page_dirty_lock(page);
1043                 put_page(page);
1044         }
1045 }
1046
1047 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
1048                                size_t *nbytesp, int write)
1049 {
1050         size_t nbytes = *nbytesp;
1051         unsigned long user_addr = (unsigned long) buf;
1052         unsigned offset = user_addr & ~PAGE_MASK;
1053         int npages;
1054
1055         /* Special case for kernel I/O: can copy directly into the buffer */
1056         if (segment_eq(get_fs(), KERNEL_DS)) {
1057                 if (write)
1058                         req->in.args[1].value = (void *) user_addr;
1059                 else
1060                         req->out.args[0].value = (void *) user_addr;
1061
1062                 return 0;
1063         }
1064
1065         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
1066         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1067         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
1068         npages = get_user_pages_fast(user_addr, npages, !write, req->pages);
1069         if (npages < 0)
1070                 return npages;
1071
1072         req->num_pages = npages;
1073         req->page_offset = offset;
1074
1075         if (write)
1076                 req->in.argpages = 1;
1077         else
1078                 req->out.argpages = 1;
1079
1080         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1081         *nbytesp = min(*nbytesp, nbytes);
1082
1083         return 0;
1084 }
1085
1086 ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1087                        size_t count, loff_t *ppos, int write)
1088 {
1089         struct fuse_file *ff = file->private_data;
1090         struct fuse_conn *fc = ff->fc;
1091         size_t nmax = write ? fc->max_write : fc->max_read;
1092         loff_t pos = *ppos;
1093         ssize_t res = 0;
1094         struct fuse_req *req;
1095
1096         req = fuse_get_req(fc, FUSE_MAX_PAGES_PER_REQ);
1097         if (IS_ERR(req))
1098                 return PTR_ERR(req);
1099
1100         while (count) {
1101                 size_t nres;
1102                 fl_owner_t owner = current->files;
1103                 size_t nbytes = min(count, nmax);
1104                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1105                 if (err) {
1106                         res = err;
1107                         break;
1108                 }
1109
1110                 if (write)
1111                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1112                 else
1113                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1114
1115                 fuse_release_user_pages(req, !write);
1116                 if (req->out.h.error) {
1117                         if (!res)
1118                                 res = req->out.h.error;
1119                         break;
1120                 } else if (nres > nbytes) {
1121                         res = -EIO;
1122                         break;
1123                 }
1124                 count -= nres;
1125                 res += nres;
1126                 pos += nres;
1127                 buf += nres;
1128                 if (nres != nbytes)
1129                         break;
1130                 if (count) {
1131                         fuse_put_request(fc, req);
1132                         req = fuse_get_req(fc, FUSE_MAX_PAGES_PER_REQ);
1133                         if (IS_ERR(req))
1134                                 break;
1135                 }
1136         }
1137         if (!IS_ERR(req))
1138                 fuse_put_request(fc, req);
1139         if (res > 0)
1140                 *ppos = pos;
1141
1142         return res;
1143 }
1144 EXPORT_SYMBOL_GPL(fuse_direct_io);
1145
1146 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1147                                      size_t count, loff_t *ppos)
1148 {
1149         ssize_t res;
1150         struct inode *inode = file->f_path.dentry->d_inode;
1151
1152         if (is_bad_inode(inode))
1153                 return -EIO;
1154
1155         res = fuse_direct_io(file, buf, count, ppos, 0);
1156
1157         fuse_invalidate_attr(inode);
1158
1159         return res;
1160 }
1161
1162 static ssize_t __fuse_direct_write(struct file *file, const char __user *buf,
1163                                    size_t count, loff_t *ppos)
1164 {
1165         struct inode *inode = file->f_path.dentry->d_inode;
1166         ssize_t res;
1167
1168         res = generic_write_checks(file, ppos, &count, 0);
1169         if (!res) {
1170                 res = fuse_direct_io(file, buf, count, ppos, 1);
1171                 if (res > 0)
1172                         fuse_write_update_size(inode, *ppos);
1173         }
1174
1175         fuse_invalidate_attr(inode);
1176
1177         return res;
1178 }
1179
1180 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1181                                  size_t count, loff_t *ppos)
1182 {
1183         struct inode *inode = file->f_path.dentry->d_inode;
1184         ssize_t res;
1185
1186         if (is_bad_inode(inode))
1187                 return -EIO;
1188
1189         /* Don't allow parallel writes to the same file */
1190         mutex_lock(&inode->i_mutex);
1191         res = __fuse_direct_write(file, buf, count, ppos);
1192         mutex_unlock(&inode->i_mutex);
1193
1194         return res;
1195 }
1196
1197 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1198 {
1199         __free_page(req->pages[0]);
1200         fuse_file_put(req->ff, false);
1201 }
1202
1203 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1204 {
1205         struct inode *inode = req->inode;
1206         struct fuse_inode *fi = get_fuse_inode(inode);
1207         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1208
1209         list_del(&req->writepages_entry);
1210         dec_bdi_stat(bdi, BDI_WRITEBACK);
1211         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1212         bdi_writeout_inc(bdi);
1213         wake_up(&fi->page_waitq);
1214 }
1215
1216 /* Called under fc->lock, may release and reacquire it */
1217 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1218 __releases(fc->lock)
1219 __acquires(fc->lock)
1220 {
1221         struct fuse_inode *fi = get_fuse_inode(req->inode);
1222         loff_t size = i_size_read(req->inode);
1223         struct fuse_write_in *inarg = &req->misc.write.in;
1224
1225         if (!fc->connected)
1226                 goto out_free;
1227
1228         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1229                 inarg->size = PAGE_CACHE_SIZE;
1230         } else if (inarg->offset < size) {
1231                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1232         } else {
1233                 /* Got truncated off completely */
1234                 goto out_free;
1235         }
1236
1237         req->in.args[1].size = inarg->size;
1238         fi->writectr++;
1239         fuse_request_send_background_locked(fc, req);
1240         return;
1241
1242  out_free:
1243         fuse_writepage_finish(fc, req);
1244         spin_unlock(&fc->lock);
1245         fuse_writepage_free(fc, req);
1246         fuse_put_request(fc, req);
1247         spin_lock(&fc->lock);
1248 }
1249
1250 /*
1251  * If fi->writectr is positive (no truncate or fsync going on) send
1252  * all queued writepage requests.
1253  *
1254  * Called with fc->lock
1255  */
1256 void fuse_flush_writepages(struct inode *inode)
1257 __releases(fc->lock)
1258 __acquires(fc->lock)
1259 {
1260         struct fuse_conn *fc = get_fuse_conn(inode);
1261         struct fuse_inode *fi = get_fuse_inode(inode);
1262         struct fuse_req *req;
1263
1264         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1265                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1266                 list_del_init(&req->list);
1267                 fuse_send_writepage(fc, req);
1268         }
1269 }
1270
1271 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1272 {
1273         struct inode *inode = req->inode;
1274         struct fuse_inode *fi = get_fuse_inode(inode);
1275
1276         mapping_set_error(inode->i_mapping, req->out.h.error);
1277         spin_lock(&fc->lock);
1278         fi->writectr--;
1279         fuse_writepage_finish(fc, req);
1280         spin_unlock(&fc->lock);
1281         fuse_writepage_free(fc, req);
1282 }
1283
1284 static int fuse_writepage_locked(struct page *page)
1285 {
1286         struct address_space *mapping = page->mapping;
1287         struct inode *inode = mapping->host;
1288         struct fuse_conn *fc = get_fuse_conn(inode);
1289         struct fuse_inode *fi = get_fuse_inode(inode);
1290         struct fuse_req *req;
1291         struct fuse_file *ff;
1292         struct page *tmp_page;
1293
1294         set_page_writeback(page);
1295
1296         req = fuse_request_alloc_nofs(1);
1297         if (!req)
1298                 goto err;
1299
1300         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1301         if (!tmp_page)
1302                 goto err_free;
1303
1304         spin_lock(&fc->lock);
1305         BUG_ON(list_empty(&fi->write_files));
1306         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1307         req->ff = fuse_file_get(ff);
1308         spin_unlock(&fc->lock);
1309
1310         fuse_write_fill(req, ff, page_offset(page), 0);
1311
1312         copy_highpage(tmp_page, page);
1313         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1314         req->in.argpages = 1;
1315         req->num_pages = 1;
1316         req->pages[0] = tmp_page;
1317         req->page_offset = 0;
1318         req->end = fuse_writepage_end;
1319         req->inode = inode;
1320
1321         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1322         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1323         end_page_writeback(page);
1324
1325         spin_lock(&fc->lock);
1326         list_add(&req->writepages_entry, &fi->writepages);
1327         list_add_tail(&req->list, &fi->queued_writes);
1328         fuse_flush_writepages(inode);
1329         spin_unlock(&fc->lock);
1330
1331         return 0;
1332
1333 err_free:
1334         fuse_request_free(req);
1335 err:
1336         end_page_writeback(page);
1337         return -ENOMEM;
1338 }
1339
1340 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1341 {
1342         int err;
1343
1344         err = fuse_writepage_locked(page);
1345         unlock_page(page);
1346
1347         return err;
1348 }
1349
1350 static int fuse_launder_page(struct page *page)
1351 {
1352         int err = 0;
1353         if (clear_page_dirty_for_io(page)) {
1354                 struct inode *inode = page->mapping->host;
1355                 err = fuse_writepage_locked(page);
1356                 if (!err)
1357                         fuse_wait_on_page_writeback(inode, page->index);
1358         }
1359         return err;
1360 }
1361
1362 /*
1363  * Write back dirty pages now, because there may not be any suitable
1364  * open files later
1365  */
1366 static void fuse_vma_close(struct vm_area_struct *vma)
1367 {
1368         filemap_write_and_wait(vma->vm_file->f_mapping);
1369 }
1370
1371 /*
1372  * Wait for writeback against this page to complete before allowing it
1373  * to be marked dirty again, and hence written back again, possibly
1374  * before the previous writepage completed.
1375  *
1376  * Block here, instead of in ->writepage(), so that the userspace fs
1377  * can only block processes actually operating on the filesystem.
1378  *
1379  * Otherwise unprivileged userspace fs would be able to block
1380  * unrelated:
1381  *
1382  * - page migration
1383  * - sync(2)
1384  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1385  */
1386 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1387 {
1388         struct page *page = vmf->page;
1389         /*
1390          * Don't use page->mapping as it may become NULL from a
1391          * concurrent truncate.
1392          */
1393         struct inode *inode = vma->vm_file->f_mapping->host;
1394
1395         fuse_wait_on_page_writeback(inode, page->index);
1396         return 0;
1397 }
1398
1399 static const struct vm_operations_struct fuse_file_vm_ops = {
1400         .close          = fuse_vma_close,
1401         .fault          = filemap_fault,
1402         .page_mkwrite   = fuse_page_mkwrite,
1403         .remap_pages    = generic_file_remap_pages,
1404 };
1405
1406 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1407 {
1408         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1409                 struct inode *inode = file->f_dentry->d_inode;
1410                 struct fuse_conn *fc = get_fuse_conn(inode);
1411                 struct fuse_inode *fi = get_fuse_inode(inode);
1412                 struct fuse_file *ff = file->private_data;
1413                 /*
1414                  * file may be written through mmap, so chain it onto the
1415                  * inodes's write_file list
1416                  */
1417                 spin_lock(&fc->lock);
1418                 if (list_empty(&ff->write_entry))
1419                         list_add(&ff->write_entry, &fi->write_files);
1420                 spin_unlock(&fc->lock);
1421         }
1422         file_accessed(file);
1423         vma->vm_ops = &fuse_file_vm_ops;
1424         return 0;
1425 }
1426
1427 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1428 {
1429         /* Can't provide the coherency needed for MAP_SHARED */
1430         if (vma->vm_flags & VM_MAYSHARE)
1431                 return -ENODEV;
1432
1433         invalidate_inode_pages2(file->f_mapping);
1434
1435         return generic_file_mmap(file, vma);
1436 }
1437
1438 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1439                                   struct file_lock *fl)
1440 {
1441         switch (ffl->type) {
1442         case F_UNLCK:
1443                 break;
1444
1445         case F_RDLCK:
1446         case F_WRLCK:
1447                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1448                     ffl->end < ffl->start)
1449                         return -EIO;
1450
1451                 fl->fl_start = ffl->start;
1452                 fl->fl_end = ffl->end;
1453                 fl->fl_pid = ffl->pid;
1454                 break;
1455
1456         default:
1457                 return -EIO;
1458         }
1459         fl->fl_type = ffl->type;
1460         return 0;
1461 }
1462
1463 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1464                          const struct file_lock *fl, int opcode, pid_t pid,
1465                          int flock)
1466 {
1467         struct inode *inode = file->f_path.dentry->d_inode;
1468         struct fuse_conn *fc = get_fuse_conn(inode);
1469         struct fuse_file *ff = file->private_data;
1470         struct fuse_lk_in *arg = &req->misc.lk_in;
1471
1472         arg->fh = ff->fh;
1473         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1474         arg->lk.start = fl->fl_start;
1475         arg->lk.end = fl->fl_end;
1476         arg->lk.type = fl->fl_type;
1477         arg->lk.pid = pid;
1478         if (flock)
1479                 arg->lk_flags |= FUSE_LK_FLOCK;
1480         req->in.h.opcode = opcode;
1481         req->in.h.nodeid = get_node_id(inode);
1482         req->in.numargs = 1;
1483         req->in.args[0].size = sizeof(*arg);
1484         req->in.args[0].value = arg;
1485 }
1486
1487 static int fuse_getlk(struct file *file, struct file_lock *fl)
1488 {
1489         struct inode *inode = file->f_path.dentry->d_inode;
1490         struct fuse_conn *fc = get_fuse_conn(inode);
1491         struct fuse_req *req;
1492         struct fuse_lk_out outarg;
1493         int err;
1494
1495         req = fuse_get_req_nopages(fc);
1496         if (IS_ERR(req))
1497                 return PTR_ERR(req);
1498
1499         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1500         req->out.numargs = 1;
1501         req->out.args[0].size = sizeof(outarg);
1502         req->out.args[0].value = &outarg;
1503         fuse_request_send(fc, req);
1504         err = req->out.h.error;
1505         fuse_put_request(fc, req);
1506         if (!err)
1507                 err = convert_fuse_file_lock(&outarg.lk, fl);
1508
1509         return err;
1510 }
1511
1512 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1513 {
1514         struct inode *inode = file->f_path.dentry->d_inode;
1515         struct fuse_conn *fc = get_fuse_conn(inode);
1516         struct fuse_req *req;
1517         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1518         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1519         int err;
1520
1521         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1522                 /* NLM needs asynchronous locks, which we don't support yet */
1523                 return -ENOLCK;
1524         }
1525
1526         /* Unlock on close is handled by the flush method */
1527         if (fl->fl_flags & FL_CLOSE)
1528                 return 0;
1529
1530         req = fuse_get_req_nopages(fc);
1531         if (IS_ERR(req))
1532                 return PTR_ERR(req);
1533
1534         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1535         fuse_request_send(fc, req);
1536         err = req->out.h.error;
1537         /* locking is restartable */
1538         if (err == -EINTR)
1539                 err = -ERESTARTSYS;
1540         fuse_put_request(fc, req);
1541         return err;
1542 }
1543
1544 static int fuse_file_lock(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 (cmd == F_CANCELLK) {
1551                 err = 0;
1552         } else if (cmd == F_GETLK) {
1553                 if (fc->no_lock) {
1554                         posix_test_lock(file, fl);
1555                         err = 0;
1556                 } else
1557                         err = fuse_getlk(file, fl);
1558         } else {
1559                 if (fc->no_lock)
1560                         err = posix_lock_file(file, fl, NULL);
1561                 else
1562                         err = fuse_setlk(file, fl, 0);
1563         }
1564         return err;
1565 }
1566
1567 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1568 {
1569         struct inode *inode = file->f_path.dentry->d_inode;
1570         struct fuse_conn *fc = get_fuse_conn(inode);
1571         int err;
1572
1573         if (fc->no_flock) {
1574                 err = flock_lock_file_wait(file, fl);
1575         } else {
1576                 struct fuse_file *ff = file->private_data;
1577
1578                 /* emulate flock with POSIX locks */
1579                 fl->fl_owner = (fl_owner_t) file;
1580                 ff->flock = true;
1581                 err = fuse_setlk(file, fl, 1);
1582         }
1583
1584         return err;
1585 }
1586
1587 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1588 {
1589         struct inode *inode = mapping->host;
1590         struct fuse_conn *fc = get_fuse_conn(inode);
1591         struct fuse_req *req;
1592         struct fuse_bmap_in inarg;
1593         struct fuse_bmap_out outarg;
1594         int err;
1595
1596         if (!inode->i_sb->s_bdev || fc->no_bmap)
1597                 return 0;
1598
1599         req = fuse_get_req_nopages(fc);
1600         if (IS_ERR(req))
1601                 return 0;
1602
1603         memset(&inarg, 0, sizeof(inarg));
1604         inarg.block = block;
1605         inarg.blocksize = inode->i_sb->s_blocksize;
1606         req->in.h.opcode = FUSE_BMAP;
1607         req->in.h.nodeid = get_node_id(inode);
1608         req->in.numargs = 1;
1609         req->in.args[0].size = sizeof(inarg);
1610         req->in.args[0].value = &inarg;
1611         req->out.numargs = 1;
1612         req->out.args[0].size = sizeof(outarg);
1613         req->out.args[0].value = &outarg;
1614         fuse_request_send(fc, req);
1615         err = req->out.h.error;
1616         fuse_put_request(fc, req);
1617         if (err == -ENOSYS)
1618                 fc->no_bmap = 1;
1619
1620         return err ? 0 : outarg.block;
1621 }
1622
1623 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1624 {
1625         loff_t retval;
1626         struct inode *inode = file->f_path.dentry->d_inode;
1627
1628         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1629         if (whence == SEEK_CUR || whence == SEEK_SET)
1630                 return generic_file_llseek(file, offset, whence);
1631
1632         mutex_lock(&inode->i_mutex);
1633         retval = fuse_update_attributes(inode, NULL, file, NULL);
1634         if (!retval)
1635                 retval = generic_file_llseek(file, offset, whence);
1636         mutex_unlock(&inode->i_mutex);
1637
1638         return retval;
1639 }
1640
1641 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1642                         unsigned int nr_segs, size_t bytes, bool to_user)
1643 {
1644         struct iov_iter ii;
1645         int page_idx = 0;
1646
1647         if (!bytes)
1648                 return 0;
1649
1650         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1651
1652         while (iov_iter_count(&ii)) {
1653                 struct page *page = pages[page_idx++];
1654                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1655                 void *kaddr;
1656
1657                 kaddr = kmap(page);
1658
1659                 while (todo) {
1660                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1661                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1662                         size_t copy = min(todo, iov_len);
1663                         size_t left;
1664
1665                         if (!to_user)
1666                                 left = copy_from_user(kaddr, uaddr, copy);
1667                         else
1668                                 left = copy_to_user(uaddr, kaddr, copy);
1669
1670                         if (unlikely(left))
1671                                 return -EFAULT;
1672
1673                         iov_iter_advance(&ii, copy);
1674                         todo -= copy;
1675                         kaddr += copy;
1676                 }
1677
1678                 kunmap(page);
1679         }
1680
1681         return 0;
1682 }
1683
1684 /*
1685  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1686  * ABI was defined to be 'struct iovec' which is different on 32bit
1687  * and 64bit.  Fortunately we can determine which structure the server
1688  * used from the size of the reply.
1689  */
1690 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1691                                      size_t transferred, unsigned count,
1692                                      bool is_compat)
1693 {
1694 #ifdef CONFIG_COMPAT
1695         if (count * sizeof(struct compat_iovec) == transferred) {
1696                 struct compat_iovec *ciov = src;
1697                 unsigned i;
1698
1699                 /*
1700                  * With this interface a 32bit server cannot support
1701                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1702                  * requests
1703                  */
1704                 if (!is_compat)
1705                         return -EINVAL;
1706
1707                 for (i = 0; i < count; i++) {
1708                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1709                         dst[i].iov_len = ciov[i].iov_len;
1710                 }
1711                 return 0;
1712         }
1713 #endif
1714
1715         if (count * sizeof(struct iovec) != transferred)
1716                 return -EIO;
1717
1718         memcpy(dst, src, transferred);
1719         return 0;
1720 }
1721
1722 /* Make sure iov_length() won't overflow */
1723 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1724 {
1725         size_t n;
1726         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1727
1728         for (n = 0; n < count; n++, iov++) {
1729                 if (iov->iov_len > (size_t) max)
1730                         return -ENOMEM;
1731                 max -= iov->iov_len;
1732         }
1733         return 0;
1734 }
1735
1736 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1737                                  void *src, size_t transferred, unsigned count,
1738                                  bool is_compat)
1739 {
1740         unsigned i;
1741         struct fuse_ioctl_iovec *fiov = src;
1742
1743         if (fc->minor < 16) {
1744                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1745                                                  count, is_compat);
1746         }
1747
1748         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1749                 return -EIO;
1750
1751         for (i = 0; i < count; i++) {
1752                 /* Did the server supply an inappropriate value? */
1753                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1754                     fiov[i].len != (unsigned long) fiov[i].len)
1755                         return -EIO;
1756
1757                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1758                 dst[i].iov_len = (size_t) fiov[i].len;
1759
1760 #ifdef CONFIG_COMPAT
1761                 if (is_compat &&
1762                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1763                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1764                         return -EIO;
1765 #endif
1766         }
1767
1768         return 0;
1769 }
1770
1771
1772 /*
1773  * For ioctls, there is no generic way to determine how much memory
1774  * needs to be read and/or written.  Furthermore, ioctls are allowed
1775  * to dereference the passed pointer, so the parameter requires deep
1776  * copying but FUSE has no idea whatsoever about what to copy in or
1777  * out.
1778  *
1779  * This is solved by allowing FUSE server to retry ioctl with
1780  * necessary in/out iovecs.  Let's assume the ioctl implementation
1781  * needs to read in the following structure.
1782  *
1783  * struct a {
1784  *      char    *buf;
1785  *      size_t  buflen;
1786  * }
1787  *
1788  * On the first callout to FUSE server, inarg->in_size and
1789  * inarg->out_size will be NULL; then, the server completes the ioctl
1790  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1791  * the actual iov array to
1792  *
1793  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1794  *
1795  * which tells FUSE to copy in the requested area and retry the ioctl.
1796  * On the second round, the server has access to the structure and
1797  * from that it can tell what to look for next, so on the invocation,
1798  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1799  *
1800  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1801  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1802  *
1803  * FUSE will copy both struct a and the pointed buffer from the
1804  * process doing the ioctl and retry ioctl with both struct a and the
1805  * buffer.
1806  *
1807  * This time, FUSE server has everything it needs and completes ioctl
1808  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1809  *
1810  * Copying data out works the same way.
1811  *
1812  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1813  * automatically initializes in and out iovs by decoding @cmd with
1814  * _IOC_* macros and the server is not allowed to request RETRY.  This
1815  * limits ioctl data transfers to well-formed ioctls and is the forced
1816  * behavior for all FUSE servers.
1817  */
1818 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1819                    unsigned int flags)
1820 {
1821         struct fuse_file *ff = file->private_data;
1822         struct fuse_conn *fc = ff->fc;
1823         struct fuse_ioctl_in inarg = {
1824                 .fh = ff->fh,
1825                 .cmd = cmd,
1826                 .arg = arg,
1827                 .flags = flags
1828         };
1829         struct fuse_ioctl_out outarg;
1830         struct fuse_req *req = NULL;
1831         struct page **pages = NULL;
1832         struct iovec *iov_page = NULL;
1833         struct iovec *in_iov = NULL, *out_iov = NULL;
1834         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1835         size_t in_size, out_size, transferred;
1836         int err;
1837
1838 #if BITS_PER_LONG == 32
1839         inarg.flags |= FUSE_IOCTL_32BIT;
1840 #else
1841         if (flags & FUSE_IOCTL_COMPAT)
1842                 inarg.flags |= FUSE_IOCTL_32BIT;
1843 #endif
1844
1845         /* assume all the iovs returned by client always fits in a page */
1846         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1847
1848         err = -ENOMEM;
1849         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1850         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1851         if (!pages || !iov_page)
1852                 goto out;
1853
1854         /*
1855          * If restricted, initialize IO parameters as encoded in @cmd.
1856          * RETRY from server is not allowed.
1857          */
1858         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1859                 struct iovec *iov = iov_page;
1860
1861                 iov->iov_base = (void __user *)arg;
1862                 iov->iov_len = _IOC_SIZE(cmd);
1863
1864                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1865                         in_iov = iov;
1866                         in_iovs = 1;
1867                 }
1868
1869                 if (_IOC_DIR(cmd) & _IOC_READ) {
1870                         out_iov = iov;
1871                         out_iovs = 1;
1872                 }
1873         }
1874
1875  retry:
1876         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1877         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1878
1879         /*
1880          * Out data can be used either for actual out data or iovs,
1881          * make sure there always is at least one page.
1882          */
1883         out_size = max_t(size_t, out_size, PAGE_SIZE);
1884         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1885
1886         /* make sure there are enough buffer pages and init request with them */
1887         err = -ENOMEM;
1888         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1889                 goto out;
1890         while (num_pages < max_pages) {
1891                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1892                 if (!pages[num_pages])
1893                         goto out;
1894                 num_pages++;
1895         }
1896
1897         req = fuse_get_req(fc, num_pages);
1898         if (IS_ERR(req)) {
1899                 err = PTR_ERR(req);
1900                 req = NULL;
1901                 goto out;
1902         }
1903         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1904         req->num_pages = num_pages;
1905
1906         /* okay, let's send it to the client */
1907         req->in.h.opcode = FUSE_IOCTL;
1908         req->in.h.nodeid = ff->nodeid;
1909         req->in.numargs = 1;
1910         req->in.args[0].size = sizeof(inarg);
1911         req->in.args[0].value = &inarg;
1912         if (in_size) {
1913                 req->in.numargs++;
1914                 req->in.args[1].size = in_size;
1915                 req->in.argpages = 1;
1916
1917                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1918                                            false);
1919                 if (err)
1920                         goto out;
1921         }
1922
1923         req->out.numargs = 2;
1924         req->out.args[0].size = sizeof(outarg);
1925         req->out.args[0].value = &outarg;
1926         req->out.args[1].size = out_size;
1927         req->out.argpages = 1;
1928         req->out.argvar = 1;
1929
1930         fuse_request_send(fc, req);
1931         err = req->out.h.error;
1932         transferred = req->out.args[1].size;
1933         fuse_put_request(fc, req);
1934         req = NULL;
1935         if (err)
1936                 goto out;
1937
1938         /* did it ask for retry? */
1939         if (outarg.flags & FUSE_IOCTL_RETRY) {
1940                 void *vaddr;
1941
1942                 /* no retry if in restricted mode */
1943                 err = -EIO;
1944                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1945                         goto out;
1946
1947                 in_iovs = outarg.in_iovs;
1948                 out_iovs = outarg.out_iovs;
1949
1950                 /*
1951                  * Make sure things are in boundary, separate checks
1952                  * are to protect against overflow.
1953                  */
1954                 err = -ENOMEM;
1955                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1956                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1957                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1958                         goto out;
1959
1960                 vaddr = kmap_atomic(pages[0]);
1961                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
1962                                             transferred, in_iovs + out_iovs,
1963                                             (flags & FUSE_IOCTL_COMPAT) != 0);
1964                 kunmap_atomic(vaddr);
1965                 if (err)
1966                         goto out;
1967
1968                 in_iov = iov_page;
1969                 out_iov = in_iov + in_iovs;
1970
1971                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
1972                 if (err)
1973                         goto out;
1974
1975                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
1976                 if (err)
1977                         goto out;
1978
1979                 goto retry;
1980         }
1981
1982         err = -EIO;
1983         if (transferred > inarg.out_size)
1984                 goto out;
1985
1986         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1987  out:
1988         if (req)
1989                 fuse_put_request(fc, req);
1990         free_page((unsigned long) iov_page);
1991         while (num_pages)
1992                 __free_page(pages[--num_pages]);
1993         kfree(pages);
1994
1995         return err ? err : outarg.result;
1996 }
1997 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
1998
1999 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2000                        unsigned long arg, unsigned int flags)
2001 {
2002         struct inode *inode = file->f_dentry->d_inode;
2003         struct fuse_conn *fc = get_fuse_conn(inode);
2004
2005         if (!fuse_allow_task(fc, current))
2006                 return -EACCES;
2007
2008         if (is_bad_inode(inode))
2009                 return -EIO;
2010
2011         return fuse_do_ioctl(file, cmd, arg, flags);
2012 }
2013
2014 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2015                             unsigned long arg)
2016 {
2017         return fuse_ioctl_common(file, cmd, arg, 0);
2018 }
2019
2020 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2021                                    unsigned long arg)
2022 {
2023         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2024 }
2025
2026 /*
2027  * All files which have been polled are linked to RB tree
2028  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2029  * find the matching one.
2030  */
2031 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2032                                               struct rb_node **parent_out)
2033 {
2034         struct rb_node **link = &fc->polled_files.rb_node;
2035         struct rb_node *last = NULL;
2036
2037         while (*link) {
2038                 struct fuse_file *ff;
2039
2040                 last = *link;
2041                 ff = rb_entry(last, struct fuse_file, polled_node);
2042
2043                 if (kh < ff->kh)
2044                         link = &last->rb_left;
2045                 else if (kh > ff->kh)
2046                         link = &last->rb_right;
2047                 else
2048                         return link;
2049         }
2050
2051         if (parent_out)
2052                 *parent_out = last;
2053         return link;
2054 }
2055
2056 /*
2057  * The file is about to be polled.  Make sure it's on the polled_files
2058  * RB tree.  Note that files once added to the polled_files tree are
2059  * not removed before the file is released.  This is because a file
2060  * polled once is likely to be polled again.
2061  */
2062 static void fuse_register_polled_file(struct fuse_conn *fc,
2063                                       struct fuse_file *ff)
2064 {
2065         spin_lock(&fc->lock);
2066         if (RB_EMPTY_NODE(&ff->polled_node)) {
2067                 struct rb_node **link, *parent;
2068
2069                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2070                 BUG_ON(*link);
2071                 rb_link_node(&ff->polled_node, parent, link);
2072                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2073         }
2074         spin_unlock(&fc->lock);
2075 }
2076
2077 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2078 {
2079         struct fuse_file *ff = file->private_data;
2080         struct fuse_conn *fc = ff->fc;
2081         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2082         struct fuse_poll_out outarg;
2083         struct fuse_req *req;
2084         int err;
2085
2086         if (fc->no_poll)
2087                 return DEFAULT_POLLMASK;
2088
2089         poll_wait(file, &ff->poll_wait, wait);
2090
2091         /*
2092          * Ask for notification iff there's someone waiting for it.
2093          * The client may ignore the flag and always notify.
2094          */
2095         if (waitqueue_active(&ff->poll_wait)) {
2096                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2097                 fuse_register_polled_file(fc, ff);
2098         }
2099
2100         req = fuse_get_req_nopages(fc);
2101         if (IS_ERR(req))
2102                 return POLLERR;
2103
2104         req->in.h.opcode = FUSE_POLL;
2105         req->in.h.nodeid = ff->nodeid;
2106         req->in.numargs = 1;
2107         req->in.args[0].size = sizeof(inarg);
2108         req->in.args[0].value = &inarg;
2109         req->out.numargs = 1;
2110         req->out.args[0].size = sizeof(outarg);
2111         req->out.args[0].value = &outarg;
2112         fuse_request_send(fc, req);
2113         err = req->out.h.error;
2114         fuse_put_request(fc, req);
2115
2116         if (!err)
2117                 return outarg.revents;
2118         if (err == -ENOSYS) {
2119                 fc->no_poll = 1;
2120                 return DEFAULT_POLLMASK;
2121         }
2122         return POLLERR;
2123 }
2124 EXPORT_SYMBOL_GPL(fuse_file_poll);
2125
2126 /*
2127  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2128  * wakes up the poll waiters.
2129  */
2130 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2131                             struct fuse_notify_poll_wakeup_out *outarg)
2132 {
2133         u64 kh = outarg->kh;
2134         struct rb_node **link;
2135
2136         spin_lock(&fc->lock);
2137
2138         link = fuse_find_polled_node(fc, kh, NULL);
2139         if (*link) {
2140                 struct fuse_file *ff;
2141
2142                 ff = rb_entry(*link, struct fuse_file, polled_node);
2143                 wake_up_interruptible_sync(&ff->poll_wait);
2144         }
2145
2146         spin_unlock(&fc->lock);
2147         return 0;
2148 }
2149
2150 static ssize_t fuse_loop_dio(struct file *filp, const struct iovec *iov,
2151                              unsigned long nr_segs, loff_t *ppos, int rw)
2152 {
2153         const struct iovec *vector = iov;
2154         ssize_t ret = 0;
2155
2156         while (nr_segs > 0) {
2157                 void __user *base;
2158                 size_t len;
2159                 ssize_t nr;
2160
2161                 base = vector->iov_base;
2162                 len = vector->iov_len;
2163                 vector++;
2164                 nr_segs--;
2165
2166                 if (rw == WRITE)
2167                         nr = __fuse_direct_write(filp, base, len, ppos);
2168                 else
2169                         nr = fuse_direct_read(filp, base, len, ppos);
2170
2171                 if (nr < 0) {
2172                         if (!ret)
2173                                 ret = nr;
2174                         break;
2175                 }
2176                 ret += nr;
2177                 if (nr != len)
2178                         break;
2179         }
2180
2181         return ret;
2182 }
2183
2184
2185 static ssize_t
2186 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2187                         loff_t offset, unsigned long nr_segs)
2188 {
2189         ssize_t ret = 0;
2190         struct file *file = NULL;
2191         loff_t pos = 0;
2192
2193         file = iocb->ki_filp;
2194         pos = offset;
2195
2196         ret = fuse_loop_dio(file, iov, nr_segs, &pos, rw);
2197
2198         return ret;
2199 }
2200
2201 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2202                                 loff_t length)
2203 {
2204         struct fuse_file *ff = file->private_data;
2205         struct fuse_conn *fc = ff->fc;
2206         struct fuse_req *req;
2207         struct fuse_fallocate_in inarg = {
2208                 .fh = ff->fh,
2209                 .offset = offset,
2210                 .length = length,
2211                 .mode = mode
2212         };
2213         int err;
2214
2215         if (fc->no_fallocate)
2216                 return -EOPNOTSUPP;
2217
2218         req = fuse_get_req_nopages(fc);
2219         if (IS_ERR(req))
2220                 return PTR_ERR(req);
2221
2222         req->in.h.opcode = FUSE_FALLOCATE;
2223         req->in.h.nodeid = ff->nodeid;
2224         req->in.numargs = 1;
2225         req->in.args[0].size = sizeof(inarg);
2226         req->in.args[0].value = &inarg;
2227         fuse_request_send(fc, req);
2228         err = req->out.h.error;
2229         if (err == -ENOSYS) {
2230                 fc->no_fallocate = 1;
2231                 err = -EOPNOTSUPP;
2232         }
2233         fuse_put_request(fc, req);
2234
2235         return err;
2236 }
2237
2238 static const struct file_operations fuse_file_operations = {
2239         .llseek         = fuse_file_llseek,
2240         .read           = do_sync_read,
2241         .aio_read       = fuse_file_aio_read,
2242         .write          = do_sync_write,
2243         .aio_write      = fuse_file_aio_write,
2244         .mmap           = fuse_file_mmap,
2245         .open           = fuse_open,
2246         .flush          = fuse_flush,
2247         .release        = fuse_release,
2248         .fsync          = fuse_fsync,
2249         .lock           = fuse_file_lock,
2250         .flock          = fuse_file_flock,
2251         .splice_read    = generic_file_splice_read,
2252         .unlocked_ioctl = fuse_file_ioctl,
2253         .compat_ioctl   = fuse_file_compat_ioctl,
2254         .poll           = fuse_file_poll,
2255         .fallocate      = fuse_file_fallocate,
2256 };
2257
2258 static const struct file_operations fuse_direct_io_file_operations = {
2259         .llseek         = fuse_file_llseek,
2260         .read           = fuse_direct_read,
2261         .write          = fuse_direct_write,
2262         .mmap           = fuse_direct_mmap,
2263         .open           = fuse_open,
2264         .flush          = fuse_flush,
2265         .release        = fuse_release,
2266         .fsync          = fuse_fsync,
2267         .lock           = fuse_file_lock,
2268         .flock          = fuse_file_flock,
2269         .unlocked_ioctl = fuse_file_ioctl,
2270         .compat_ioctl   = fuse_file_compat_ioctl,
2271         .poll           = fuse_file_poll,
2272         .fallocate      = fuse_file_fallocate,
2273         /* no splice_read */
2274 };
2275
2276 static const struct address_space_operations fuse_file_aops  = {
2277         .readpage       = fuse_readpage,
2278         .writepage      = fuse_writepage,
2279         .launder_page   = fuse_launder_page,
2280         .readpages      = fuse_readpages,
2281         .set_page_dirty = __set_page_dirty_nobuffers,
2282         .bmap           = fuse_bmap,
2283         .direct_IO      = fuse_direct_IO,
2284 };
2285
2286 void fuse_init_file_inode(struct inode *inode)
2287 {
2288         inode->i_fop = &fuse_file_operations;
2289         inode->i_data.a_ops = &fuse_file_aops;
2290 }