fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/compat.h>
  17 #include <linux/swap.h>
  18 #include <linux/aio.h>
  19
  20 static const struct file_operations fuse_direct_io_file_operations;
  21
  22 static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  23                           int opcode, struct fuse_open_out *outargp)
  24 {
  25         struct fuse_open_in inarg;
  26         struct fuse_req *req;
  27         int err;
  28
  29         req = fuse_get_req_nopages(fc);
  30         if (IS_ERR(req))
  31                 return PTR_ERR(req);
  32
  33         memset(&inarg, 0, sizeof(inarg));
  34         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  35         if (!fc->atomic_o_trunc)
  36                 inarg.flags &= ~O_TRUNC;
  37         req->in.h.opcode = opcode;
  38         req->in.h.nodeid = nodeid;
  39         req->in.numargs = 1;
  40         req->in.args[0].size = sizeof(inarg);
  41         req->in.args[0].value = &inarg;
  42         req->out.numargs = 1;
  43         req->out.args[0].size = sizeof(*outargp);
  44         req->out.args[0].value = outargp;
  45         fuse_request_send(fc, req);
  46         err = req->out.h.error;
  47         fuse_put_request(fc, req);
  48
  49         return err;
  50 }
  51
  52 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  53 {
  54         struct fuse_file *ff;
  55
  56         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  57         if (unlikely(!ff))
  58                 return NULL;
  59
  60         ff->fc = fc;
  61         ff->reserved_req = fuse_request_alloc(0);
  62         if (unlikely(!ff->reserved_req)) {
  63                 kfree(ff);
  64                 return NULL;
  65         }
  66
  67         INIT_LIST_HEAD(&ff->write_entry);
  68         atomic_set(&ff->count, 0);
  69         RB_CLEAR_NODE(&ff->polled_node);
  70         init_waitqueue_head(&ff->poll_wait);
  71
  72         spin_lock(&fc->lock);
  73         ff->kh = ++fc->khctr;
  74         spin_unlock(&fc->lock);
  75
  76         return ff;
  77 }
  78
  79 void fuse_file_free(struct fuse_file *ff)
  80 {
  81         fuse_request_free(ff->reserved_req);
  82         kfree(ff);
  83 }
  84
  85 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  86 {
  87         atomic_inc(&ff->count);
  88         return ff;
  89 }
  90
  91 static void fuse_release_async(struct work_struct *work)
  92 {
  93         struct fuse_req *req;
  94         struct fuse_conn *fc;
  95         struct path path;
  96
  97         req = container_of(work, struct fuse_req, misc.release.work);
  98         path = req->misc.release.path;
  99         fc = get_fuse_conn(path.dentry->d_inode);
 100
 101         fuse_put_request(fc, req);
 102         path_put(&path);
 103 }
 104
 105 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 106 {
 107         if (fc->destroy_req) {
 108                 /*
 109                  * If this is a fuseblk mount, then it's possible that
 110                  * releasing the path will result in releasing the
 111                  * super block and sending the DESTROY request.  If
 112                  * the server is single threaded, this would hang.
 113                  * For this reason do the path_put() in a separate
 114                  * thread.
 115                  */
 116                 atomic_inc(&req->count);
 117                 INIT_WORK(&req->misc.release.work, fuse_release_async);
 118                 schedule_work(&req->misc.release.work);
 119         } else {
 120                 path_put(&req->misc.release.path);
 121         }
 122 }
 123
 124 static void fuse_file_put(struct fuse_file *ff, bool sync)
 125 {
 126         if (atomic_dec_and_test(&ff->count)) {
 127                 struct fuse_req *req = ff->reserved_req;
 128
 129                 if (sync) {
 130                         fuse_request_send(ff->fc, req);
 131                         path_put(&req->misc.release.path);
 132                         fuse_put_request(ff->fc, req);
 133                 } else {
 134                         req->end = fuse_release_end;
 135                         fuse_request_send_background(ff->fc, req);
 136                 }
 137                 kfree(ff);
 138         }
 139 }
 140
 141 int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 142                  bool isdir)
 143 {
 144         struct fuse_open_out outarg;
 145         struct fuse_file *ff;
 146         int err;
 147         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 148
 149         ff = fuse_file_alloc(fc);
 150         if (!ff)
 151                 return -ENOMEM;
 152
 153         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 154         if (err) {
 155                 fuse_file_free(ff);
 156                 return err;
 157         }
 158
 159         if (isdir)
 160                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 161
 162         ff->fh = outarg.fh;
 163         ff->nodeid = nodeid;
 164         ff->open_flags = outarg.open_flags;
 165         file->private_data = fuse_file_get(ff);
 166
 167         return 0;
 168 }
 169 EXPORT_SYMBOL_GPL(fuse_do_open);
 170
 171 void fuse_finish_open(struct inode *inode, struct file *file)
 172 {
 173         struct fuse_file *ff = file->private_data;
 174         struct fuse_conn *fc = get_fuse_conn(inode);
 175
 176         if (ff->open_flags & FOPEN_DIRECT_IO)
 177                 file->f_op = &fuse_direct_io_file_operations;
 178         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 179                 invalidate_inode_pages2(inode->i_mapping);
 180         if (ff->open_flags & FOPEN_NONSEEKABLE)
 181                 nonseekable_open(inode, file);
 182         if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 183                 struct fuse_inode *fi = get_fuse_inode(inode);
 184
 185                 spin_lock(&fc->lock);
 186                 fi->attr_version = ++fc->attr_version;
 187                 i_size_write(inode, 0);
 188                 spin_unlock(&fc->lock);
 189                 fuse_invalidate_attr(inode);
 190         }
 191 }
 192
 193 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 194 {
 195         struct fuse_conn *fc = get_fuse_conn(inode);
 196         int err;
 197
 198         err = generic_file_open(inode, file);
 199         if (err)
 200                 return err;
 201
 202         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 203         if (err)
 204                 return err;
 205
 206         fuse_finish_open(inode, file);
 207
 208         return 0;
 209 }
 210
 211 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 212 {
 213         struct fuse_conn *fc = ff->fc;
 214         struct fuse_req *req = ff->reserved_req;
 215         struct fuse_release_in *inarg = &req->misc.release.in;
 216
 217         spin_lock(&fc->lock);
 218         list_del(&ff->write_entry);
 219         if (!RB_EMPTY_NODE(&ff->polled_node))
 220                 rb_erase(&ff->polled_node, &fc->polled_files);
 221         spin_unlock(&fc->lock);
 222
 223         wake_up_interruptible_all(&ff->poll_wait);
 224
 225         inarg->fh = ff->fh;
 226         inarg->flags = flags;
 227         req->in.h.opcode = opcode;
 228         req->in.h.nodeid = ff->nodeid;
 229         req->in.numargs = 1;
 230         req->in.args[0].size = sizeof(struct fuse_release_in);
 231         req->in.args[0].value = inarg;
 232 }
 233
 234 void fuse_release_common(struct file *file, int opcode)
 235 {
 236         struct fuse_file *ff;
 237         struct fuse_req *req;
 238
 239         ff = file->private_data;
 240         if (unlikely(!ff))
 241                 return;
 242
 243         req = ff->reserved_req;
 244         fuse_prepare_release(ff, file->f_flags, opcode);
 245
 246         if (ff->flock) {
 247                 struct fuse_release_in *inarg = &req->misc.release.in;
 248                 inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 249                 inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 250                                                        (fl_owner_t) file);
 251         }
 252         /* Hold vfsmount and dentry until release is finished */
 253         path_get(&file->f_path);
 254         req->misc.release.path = file->f_path;
 255
 256         /*
 257          * Normally this will send the RELEASE request, however if
 258          * some asynchronous READ or WRITE requests are outstanding,
 259          * the sending will be delayed.
 260          *
 261          * Make the release synchronous if this is a fuseblk mount,
 262          * synchronous RELEASE is allowed (and desirable) in this case
 263          * because the server can be trusted not to screw up.
 264          */
 265         fuse_file_put(ff, ff->fc->destroy_req != NULL);
 266 }
 267
 268 static int fuse_open(struct inode *inode, struct file *file)
 269 {
 270         return fuse_open_common(inode, file, false);
 271 }
 272
 273 static int fuse_release(struct inode *inode, struct file *file)
 274 {
 275         fuse_release_common(file, FUSE_RELEASE);
 276
 277         /* return value is ignored by VFS */
 278         return 0;
 279 }
 280
 281 void fuse_sync_release(struct fuse_file *ff, int flags)
 282 {
 283         WARN_ON(atomic_read(&ff->count) > 1);
 284         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 285         ff->reserved_req->force = 1;
 286         fuse_request_send(ff->fc, ff->reserved_req);
 287         fuse_put_request(ff->fc, ff->reserved_req);
 288         kfree(ff);
 289 }
 290 EXPORT_SYMBOL_GPL(fuse_sync_release);
 291
 292 /*
 293  * Scramble the ID space with XTEA, so that the value of the files_struct
 294  * pointer is not exposed to userspace.
 295  */
 296 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 297 {
 298         u32 *k = fc->scramble_key;
 299         u64 v = (unsigned long) id;
 300         u32 v0 = v;
 301         u32 v1 = v >> 32;
 302         u32 sum = 0;
 303         int i;
 304
 305         for (i = 0; i < 32; i++) {
 306                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 307                 sum += 0x9E3779B9;
 308                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 309         }
 310
 311         return (u64) v0 + ((u64) v1 << 32);
 312 }
 313
 314 /*
 315  * Check if page is under writeback
 316  *
 317  * This is currently done by walking the list of writepage requests
 318  * for the inode, which can be pretty inefficient.
 319  */
 320 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 321 {
 322         struct fuse_conn *fc = get_fuse_conn(inode);
 323         struct fuse_inode *fi = get_fuse_inode(inode);
 324         struct fuse_req *req;
 325         bool found = false;
 326
 327         spin_lock(&fc->lock);
 328         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 329                 pgoff_t curr_index;
 330
 331                 BUG_ON(req->inode != inode);
 332                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 333                 if (curr_index == index) {
 334                         found = true;
 335                         break;
 336                 }
 337         }
 338         spin_unlock(&fc->lock);
 339
 340         return found;
 341 }
 342
 343 /*
 344  * Wait for page writeback to be completed.
 345  *
 346  * Since fuse doesn't rely on the VM writeback tracking, this has to
 347  * use some other means.
 348  */
 349 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 350 {
 351         struct fuse_inode *fi = get_fuse_inode(inode);
 352
 353         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 354         return 0;
 355 }
 356
 357 static int fuse_flush(struct file *file, fl_owner_t id)
 358 {
 359         struct inode *inode = file_inode(file);
 360         struct fuse_conn *fc = get_fuse_conn(inode);
 361         struct fuse_file *ff = file->private_data;
 362         struct fuse_req *req;
 363         struct fuse_flush_in inarg;
 364         int err;
 365
 366         if (is_bad_inode(inode))
 367                 return -EIO;
 368
 369         if (fc->no_flush)
 370                 return 0;
 371
 372         req = fuse_get_req_nofail_nopages(fc, file);
 373         memset(&inarg, 0, sizeof(inarg));
 374         inarg.fh = ff->fh;
 375         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 376         req->in.h.opcode = FUSE_FLUSH;
 377         req->in.h.nodeid = get_node_id(inode);
 378         req->in.numargs = 1;
 379         req->in.args[0].size = sizeof(inarg);
 380         req->in.args[0].value = &inarg;
 381         req->force = 1;
 382         fuse_request_send(fc, req);
 383         err = req->out.h.error;
 384         fuse_put_request(fc, req);
 385         if (err == -ENOSYS) {
 386                 fc->no_flush = 1;
 387                 err = 0;
 388         }
 389         return err;
 390 }
 391
 392 /*
 393  * Wait for all pending writepages on the inode to finish.
 394  *
 395  * This is currently done by blocking further writes with FUSE_NOWRITE
 396  * and waiting for all sent writes to complete.
 397  *
 398  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 399  * could conflict with truncation.
 400  */
 401 static void fuse_sync_writes(struct inode *inode)
 402 {
 403         fuse_set_nowrite(inode);
 404         fuse_release_nowrite(inode);
 405 }
 406
 407 int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 408                       int datasync, int isdir)
 409 {
 410         struct inode *inode = file->f_mapping->host;
 411         struct fuse_conn *fc = get_fuse_conn(inode);
 412         struct fuse_file *ff = file->private_data;
 413         struct fuse_req *req;
 414         struct fuse_fsync_in inarg;
 415         int err;
 416
 417         if (is_bad_inode(inode))
 418                 return -EIO;
 419
 420         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 421         if (err)
 422                 return err;
 423
 424         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 425                 return 0;
 426
 427         mutex_lock(&inode->i_mutex);
 428
 429         /*
 430          * Start writeback against all dirty pages of the inode, then
 431          * wait for all outstanding writes, before sending the FSYNC
 432          * request.
 433          */
 434         err = write_inode_now(inode, 0);
 435         if (err)
 436                 goto out;
 437
 438         fuse_sync_writes(inode);
 439
 440         req = fuse_get_req_nopages(fc);
 441         if (IS_ERR(req)) {
 442                 err = PTR_ERR(req);
 443                 goto out;
 444         }
 445
 446         memset(&inarg, 0, sizeof(inarg));
 447         inarg.fh = ff->fh;
 448         inarg.fsync_flags = datasync ? 1 : 0;
 449         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 450         req->in.h.nodeid = get_node_id(inode);
 451         req->in.numargs = 1;
 452         req->in.args[0].size = sizeof(inarg);
 453         req->in.args[0].value = &inarg;
 454         fuse_request_send(fc, req);
 455         err = req->out.h.error;
 456         fuse_put_request(fc, req);
 457         if (err == -ENOSYS) {
 458                 if (isdir)
 459                         fc->no_fsyncdir = 1;
 460                 else
 461                         fc->no_fsync = 1;
 462                 err = 0;
 463         }
 464 out:
 465         mutex_unlock(&inode->i_mutex);
 466         return err;
 467 }
 468
 469 static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 470                       int datasync)
 471 {
 472         return fuse_fsync_common(file, start, end, datasync, 0);
 473 }
 474
 475 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 476                     size_t count, int opcode)
 477 {
 478         struct fuse_read_in *inarg = &req->misc.read.in;
 479         struct fuse_file *ff = file->private_data;
 480
 481         inarg->fh = ff->fh;
 482         inarg->offset = pos;
 483         inarg->size = count;
 484         inarg->flags = file->f_flags;
 485         req->in.h.opcode = opcode;
 486         req->in.h.nodeid = ff->nodeid;
 487         req->in.numargs = 1;
 488         req->in.args[0].size = sizeof(struct fuse_read_in);
 489         req->in.args[0].value = inarg;
 490         req->out.argvar = 1;
 491         req->out.numargs = 1;
 492         req->out.args[0].size = count;
 493 }
 494
 495 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 496                              loff_t pos, size_t count, fl_owner_t owner)
 497 {
 498         struct fuse_file *ff = file->private_data;
 499         struct fuse_conn *fc = ff->fc;
 500
 501         fuse_read_fill(req, file, pos, count, FUSE_READ);
 502         if (owner != NULL) {
 503                 struct fuse_read_in *inarg = &req->misc.read.in;
 504
 505                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 506                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 507         }
 508         fuse_request_send(fc, req);
 509         return req->out.args[0].size;
 510 }
 511
 512 static void fuse_read_update_size(struct inode *inode, loff_t size,
 513                                   u64 attr_ver)
 514 {
 515         struct fuse_conn *fc = get_fuse_conn(inode);
 516         struct fuse_inode *fi = get_fuse_inode(inode);
 517
 518         spin_lock(&fc->lock);
 519         if (attr_ver == fi->attr_version && size < inode->i_size) {
 520                 fi->attr_version = ++fc->attr_version;
 521                 i_size_write(inode, size);
 522         }
 523         spin_unlock(&fc->lock);
 524 }
 525
 526 static int fuse_readpage(struct file *file, struct page *page)
 527 {
 528         struct inode *inode = page->mapping->host;
 529         struct fuse_conn *fc = get_fuse_conn(inode);
 530         struct fuse_req *req;
 531         size_t num_read;
 532         loff_t pos = page_offset(page);
 533         size_t count = PAGE_CACHE_SIZE;
 534         u64 attr_ver;
 535         int err;
 536
 537         err = -EIO;
 538         if (is_bad_inode(inode))
 539                 goto out;
 540
 541         /*
 542          * Page writeback can extend beyond the lifetime of the
 543          * page-cache page, so make sure we read a properly synced
 544          * page.
 545          */
 546         fuse_wait_on_page_writeback(inode, page->index);
 547
 548         req = fuse_get_req(fc, 1);
 549         err = PTR_ERR(req);
 550         if (IS_ERR(req))
 551                 goto out;
 552
 553         attr_ver = fuse_get_attr_version(fc);
 554
 555         req->out.page_zeroing = 1;
 556         req->out.argpages = 1;
 557         req->num_pages = 1;
 558         req->pages[0] = page;
 559         req->page_descs[0].length = count;
 560         num_read = fuse_send_read(req, file, pos, count, NULL);
 561         err = req->out.h.error;
 562         fuse_put_request(fc, req);
 563
 564         if (!err) {
 565                 /*
 566                  * Short read means EOF.  If file size is larger, truncate it
 567                  */
 568                 if (num_read < count)
 569                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 570
 571                 SetPageUptodate(page);
 572         }
 573
 574         fuse_invalidate_attr(inode); /* atime changed */
 575  out:
 576         unlock_page(page);
 577         return err;
 578 }
 579
 580 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 581 {
 582         int i;
 583         size_t count = req->misc.read.in.size;
 584         size_t num_read = req->out.args[0].size;
 585         struct address_space *mapping = NULL;
 586
 587         for (i = 0; mapping == NULL && i < req->num_pages; i++)
 588                 mapping = req->pages[i]->mapping;
 589
 590         if (mapping) {
 591                 struct inode *inode = mapping->host;
 592
 593                 /*
 594                  * Short read means EOF. If file size is larger, truncate it
 595                  */
 596                 if (!req->out.h.error && num_read < count) {
 597                         loff_t pos;
 598
 599                         pos = page_offset(req->pages[0]) + num_read;
 600                         fuse_read_update_size(inode, pos,
 601                                               req->misc.read.attr_ver);
 602                 }
 603                 fuse_invalidate_attr(inode); /* atime changed */
 604         }
 605
 606         for (i = 0; i < req->num_pages; i++) {
 607                 struct page *page = req->pages[i];
 608                 if (!req->out.h.error)
 609                         SetPageUptodate(page);
 610                 else
 611                         SetPageError(page);
 612                 unlock_page(page);
 613                 page_cache_release(page);
 614         }
 615         if (req->ff)
 616                 fuse_file_put(req->ff, false);
 617 }
 618
 619 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 620 {
 621         struct fuse_file *ff = file->private_data;
 622         struct fuse_conn *fc = ff->fc;
 623         loff_t pos = page_offset(req->pages[0]);
 624         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 625
 626         req->out.argpages = 1;
 627         req->out.page_zeroing = 1;
 628         req->out.page_replace = 1;
 629         fuse_read_fill(req, file, pos, count, FUSE_READ);
 630         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 631         if (fc->async_read) {
 632                 req->ff = fuse_file_get(ff);
 633                 req->end = fuse_readpages_end;
 634                 fuse_request_send_background(fc, req);
 635         } else {
 636                 fuse_request_send(fc, req);
 637                 fuse_readpages_end(fc, req);
 638                 fuse_put_request(fc, req);
 639         }
 640 }
 641
 642 struct fuse_fill_data {
 643         struct fuse_req *req;
 644         struct file *file;
 645         struct inode *inode;
 646         unsigned nr_pages;
 647 };
 648
 649 static int fuse_readpages_fill(void *_data, struct page *page)
 650 {
 651         struct fuse_fill_data *data = _data;
 652         struct fuse_req *req = data->req;
 653         struct inode *inode = data->inode;
 654         struct fuse_conn *fc = get_fuse_conn(inode);
 655
 656         fuse_wait_on_page_writeback(inode, page->index);
 657
 658         if (req->num_pages &&
 659             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 660              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 661              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 662                 int nr_alloc = min_t(unsigned, data->nr_pages,
 663                                      FUSE_MAX_PAGES_PER_REQ);
 664                 fuse_send_readpages(req, data->file);
 665                 data->req = req = fuse_get_req(fc, nr_alloc);
 666                 if (IS_ERR(req)) {
 667                         unlock_page(page);
 668                         return PTR_ERR(req);
 669                 }
 670         }
 671
 672         if (WARN_ON(req->num_pages >= req->max_pages)) {
 673                 fuse_put_request(fc, req);
 674                 return -EIO;
 675         }
 676
 677         page_cache_get(page);
 678         req->pages[req->num_pages] = page;
 679         req->page_descs[req->num_pages].length = PAGE_SIZE;
 680         req->num_pages++;
 681         data->nr_pages--;
 682         return 0;
 683 }
 684
 685 static int fuse_readpages(struct file *file, struct address_space *mapping,
 686                           struct list_head *pages, unsigned nr_pages)
 687 {
 688         struct inode *inode = mapping->host;
 689         struct fuse_conn *fc = get_fuse_conn(inode);
 690         struct fuse_fill_data data;
 691         int err;
 692         int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 693
 694         err = -EIO;
 695         if (is_bad_inode(inode))
 696                 goto out;
 697
 698         data.file = file;
 699         data.inode = inode;
 700         data.req = fuse_get_req(fc, nr_alloc);
 701         data.nr_pages = nr_pages;
 702         err = PTR_ERR(data.req);
 703         if (IS_ERR(data.req))
 704                 goto out;
 705
 706         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 707         if (!err) {
 708                 if (data.req->num_pages)
 709                         fuse_send_readpages(data.req, file);
 710                 else
 711                         fuse_put_request(fc, data.req);
 712         }
 713 out:
 714         return err;
 715 }
 716
 717 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 718                                   unsigned long nr_segs, loff_t pos)
 719 {
 720         struct inode *inode = iocb->ki_filp->f_mapping->host;
 721         struct fuse_conn *fc = get_fuse_conn(inode);
 722
 723         /*
 724          * In auto invalidate mode, always update attributes on read.
 725          * Otherwise, only update if we attempt to read past EOF (to ensure
 726          * i_size is up to date).
 727          */
 728         if (fc->auto_inval_data ||
 729             (pos + iov_length(iov, nr_segs) > i_size_read(inode))) {
 730                 int err;
 731                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 732                 if (err)
 733                         return err;
 734         }
 735
 736         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 737 }
 738
 739 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 740                             loff_t pos, size_t count)
 741 {
 742         struct fuse_write_in *inarg = &req->misc.write.in;
 743         struct fuse_write_out *outarg = &req->misc.write.out;
 744
 745         inarg->fh = ff->fh;
 746         inarg->offset = pos;
 747         inarg->size = count;
 748         req->in.h.opcode = FUSE_WRITE;
 749         req->in.h.nodeid = ff->nodeid;
 750         req->in.numargs = 2;
 751         if (ff->fc->minor < 9)
 752                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 753         else
 754                 req->in.args[0].size = sizeof(struct fuse_write_in);
 755         req->in.args[0].value = inarg;
 756         req->in.args[1].size = count;
 757         req->out.numargs = 1;
 758         req->out.args[0].size = sizeof(struct fuse_write_out);
 759         req->out.args[0].value = outarg;
 760 }
 761
 762 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 763                               loff_t pos, size_t count, fl_owner_t owner)
 764 {
 765         struct fuse_file *ff = file->private_data;
 766         struct fuse_conn *fc = ff->fc;
 767         struct fuse_write_in *inarg = &req->misc.write.in;
 768
 769         fuse_write_fill(req, ff, pos, count);
 770         inarg->flags = file->f_flags;
 771         if (owner != NULL) {
 772                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 773                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 774         }
 775         fuse_request_send(fc, req);
 776         return req->misc.write.out.size;
 777 }
 778
 779 void fuse_write_update_size(struct inode *inode, loff_t pos)
 780 {
 781         struct fuse_conn *fc = get_fuse_conn(inode);
 782         struct fuse_inode *fi = get_fuse_inode(inode);
 783
 784         spin_lock(&fc->lock);
 785         fi->attr_version = ++fc->attr_version;
 786         if (pos > inode->i_size)
 787                 i_size_write(inode, pos);
 788         spin_unlock(&fc->lock);
 789 }
 790
 791 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 792                                     struct inode *inode, loff_t pos,
 793                                     size_t count)
 794 {
 795         size_t res;
 796         unsigned offset;
 797         unsigned i;
 798
 799         for (i = 0; i < req->num_pages; i++)
 800                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 801
 802         res = fuse_send_write(req, file, pos, count, NULL);
 803
 804         offset = req->page_descs[0].offset;
 805         count = res;
 806         for (i = 0; i < req->num_pages; i++) {
 807                 struct page *page = req->pages[i];
 808
 809                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 810                         SetPageUptodate(page);
 811
 812                 if (count > PAGE_CACHE_SIZE - offset)
 813                         count -= PAGE_CACHE_SIZE - offset;
 814                 else
 815                         count = 0;
 816                 offset = 0;
 817
 818                 unlock_page(page);
 819                 page_cache_release(page);
 820         }
 821
 822         return res;
 823 }
 824
 825 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 826                                struct address_space *mapping,
 827                                struct iov_iter *ii, loff_t pos)
 828 {
 829         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 830         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 831         size_t count = 0;
 832         int err;
 833
 834         req->in.argpages = 1;
 835         req->page_descs[0].offset = offset;
 836
 837         do {
 838                 size_t tmp;
 839                 struct page *page;
 840                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 841                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 842                                      iov_iter_count(ii));
 843
 844                 bytes = min_t(size_t, bytes, fc->max_write - count);
 845
 846  again:
 847                 err = -EFAULT;
 848                 if (iov_iter_fault_in_readable(ii, bytes))
 849                         break;
 850
 851                 err = -ENOMEM;
 852                 page = grab_cache_page_write_begin(mapping, index, 0);
 853                 if (!page)
 854                         break;
 855
 856                 if (mapping_writably_mapped(mapping))
 857                         flush_dcache_page(page);
 858
 859                 pagefault_disable();
 860                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 861                 pagefault_enable();
 862                 flush_dcache_page(page);
 863
 864                 mark_page_accessed(page);
 865
 866                 if (!tmp) {
 867                         unlock_page(page);
 868                         page_cache_release(page);
 869                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 870                         goto again;
 871                 }
 872
 873                 err = 0;
 874                 req->pages[req->num_pages] = page;
 875                 req->page_descs[req->num_pages].length = tmp;
 876                 req->num_pages++;
 877
 878                 iov_iter_advance(ii, tmp);
 879                 count += tmp;
 880                 pos += tmp;
 881                 offset += tmp;
 882                 if (offset == PAGE_CACHE_SIZE)
 883                         offset = 0;
 884
 885                 if (!fc->big_writes)
 886                         break;
 887         } while (iov_iter_count(ii) && count < fc->max_write &&
 888                  req->num_pages < req->max_pages && offset == 0);
 889
 890         return count > 0 ? count : err;
 891 }
 892
 893 static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
 894 {
 895         return min_t(unsigned,
 896                      ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
 897                      (pos >> PAGE_CACHE_SHIFT) + 1,
 898                      FUSE_MAX_PAGES_PER_REQ);
 899 }
 900
 901 static ssize_t fuse_perform_write(struct file *file,
 902                                   struct address_space *mapping,
 903                                   struct iov_iter *ii, loff_t pos)
 904 {
 905         struct inode *inode = mapping->host;
 906         struct fuse_conn *fc = get_fuse_conn(inode);
 907         int err = 0;
 908         ssize_t res = 0;
 909
 910         if (is_bad_inode(inode))
 911                 return -EIO;
 912
 913         do {
 914                 struct fuse_req *req;
 915                 ssize_t count;
 916                 unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
 917
 918                 req = fuse_get_req(fc, nr_pages);
 919                 if (IS_ERR(req)) {
 920                         err = PTR_ERR(req);
 921                         break;
 922                 }
 923
 924                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 925                 if (count <= 0) {
 926                         err = count;
 927                 } else {
 928                         size_t num_written;
 929
 930                         num_written = fuse_send_write_pages(req, file, inode,
 931                                                             pos, count);
 932                         err = req->out.h.error;
 933                         if (!err) {
 934                                 res += num_written;
 935                                 pos += num_written;
 936
 937                                 /* break out of the loop on short write */
 938                                 if (num_written != count)
 939                                         err = -EIO;
 940                         }
 941                 }
 942                 fuse_put_request(fc, req);
 943         } while (!err && iov_iter_count(ii));
 944
 945         if (res > 0)
 946                 fuse_write_update_size(inode, pos);
 947
 948         fuse_invalidate_attr(inode);
 949
 950         return res > 0 ? res : err;
 951 }
 952
 953 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 954                                    unsigned long nr_segs, loff_t pos)
 955 {
 956         struct file *file = iocb->ki_filp;
 957         struct address_space *mapping = file->f_mapping;
 958         size_t count = 0;
 959         size_t ocount = 0;
 960         ssize_t written = 0;
 961         ssize_t written_buffered = 0;
 962         struct inode *inode = mapping->host;
 963         ssize_t err;
 964         struct iov_iter i;
 965         loff_t endbyte = 0;
 966
 967         WARN_ON(iocb->ki_pos != pos);
 968
 969         ocount = 0;
 970         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
 971         if (err)
 972                 return err;
 973
 974         count = ocount;
 975         mutex_lock(&inode->i_mutex);
 976
 977         /* We can write back this queue in page reclaim */
 978         current->backing_dev_info = mapping->backing_dev_info;
 979
 980         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 981         if (err)
 982                 goto out;
 983
 984         if (count == 0)
 985                 goto out;
 986
 987         err = file_remove_suid(file);
 988         if (err)
 989                 goto out;
 990
 991         err = file_update_time(file);
 992         if (err)
 993                 goto out;
 994
 995         if (file->f_flags & O_DIRECT) {
 996                 written = generic_file_direct_write(iocb, iov, &nr_segs,
 997                                                     pos, &iocb->ki_pos,
 998                                                     count, ocount);
 999                 if (written < 0 || written == count)
1000                         goto out;
1001
1002                 pos += written;
1003                 count -= written;
1004
1005                 iov_iter_init(&i, iov, nr_segs, count, written);
1006                 written_buffered = fuse_perform_write(file, mapping, &i, pos);
1007                 if (written_buffered < 0) {
1008                         err = written_buffered;
1009                         goto out;
1010                 }
1011                 endbyte = pos + written_buffered - 1;
1012
1013                 err = filemap_write_and_wait_range(file->f_mapping, pos,
1014                                                    endbyte);
1015                 if (err)
1016                         goto out;
1017
1018                 invalidate_mapping_pages(file->f_mapping,
1019                                          pos >> PAGE_CACHE_SHIFT,
1020                                          endbyte >> PAGE_CACHE_SHIFT);
1021
1022                 written += written_buffered;
1023                 iocb->ki_pos = pos + written_buffered;
1024         } else {
1025                 iov_iter_init(&i, iov, nr_segs, count, 0);
1026                 written = fuse_perform_write(file, mapping, &i, pos);
1027                 if (written >= 0)
1028                         iocb->ki_pos = pos + written;
1029         }
1030 out:
1031         current->backing_dev_info = NULL;
1032         mutex_unlock(&inode->i_mutex);
1033
1034         return written ? written : err;
1035 }
1036
1037 static void fuse_release_user_pages(struct fuse_req *req, int write)
1038 {
1039         unsigned i;
1040
1041         for (i = 0; i < req->num_pages; i++) {
1042                 struct page *page = req->pages[i];
1043                 if (write)
1044                         set_page_dirty_lock(page);
1045                 put_page(page);
1046         }
1047 }
1048
1049 static inline void fuse_page_descs_length_init(struct fuse_req *req,
1050                 unsigned index, unsigned nr_pages)
1051 {
1052         int i;
1053
1054         for (i = index; i < index + nr_pages; i++)
1055                 req->page_descs[i].length = PAGE_SIZE -
1056                         req->page_descs[i].offset;
1057 }
1058
1059 static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1060 {
1061         return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1062 }
1063
1064 static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1065                                         size_t max_size)
1066 {
1067         return min(iov_iter_single_seg_count(ii), max_size);
1068 }
1069
1070 static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1071                                size_t *nbytesp, int write)
1072 {
1073         size_t nbytes = 0;  /* # bytes already packed in req */
1074
1075         /* Special case for kernel I/O: can copy directly into the buffer */
1076         if (segment_eq(get_fs(), KERNEL_DS)) {
1077                 unsigned long user_addr = fuse_get_user_addr(ii);
1078                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1079
1080                 if (write)
1081                         req->in.args[1].value = (void *) user_addr;
1082                 else
1083                         req->out.args[0].value = (void *) user_addr;
1084
1085                 iov_iter_advance(ii, frag_size);
1086                 *nbytesp = frag_size;
1087                 return 0;
1088         }
1089
1090         while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1091                 unsigned npages;
1092                 unsigned long user_addr = fuse_get_user_addr(ii);
1093                 unsigned offset = user_addr & ~PAGE_MASK;
1094                 size_t frag_size = fuse_get_frag_size(ii, *nbytesp - nbytes);
1095                 int ret;
1096
1097                 unsigned n = req->max_pages - req->num_pages;
1098                 frag_size = min_t(size_t, frag_size, n << PAGE_SHIFT);
1099
1100                 npages = (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1101                 npages = clamp(npages, 1U, n);
1102
1103                 ret = get_user_pages_fast(user_addr, npages, !write,
1104                                           &req->pages[req->num_pages]);
1105                 if (ret < 0)
1106                         return ret;
1107
1108                 npages = ret;
1109                 frag_size = min_t(size_t, frag_size,
1110                                   (npages << PAGE_SHIFT) - offset);
1111                 iov_iter_advance(ii, frag_size);
1112
1113                 req->page_descs[req->num_pages].offset = offset;
1114                 fuse_page_descs_length_init(req, req->num_pages, npages);
1115
1116                 req->num_pages += npages;
1117                 req->page_descs[req->num_pages - 1].length -=
1118                         (npages << PAGE_SHIFT) - offset - frag_size;
1119
1120                 nbytes += frag_size;
1121         }
1122
1123         if (write)
1124                 req->in.argpages = 1;
1125         else
1126                 req->out.argpages = 1;
1127
1128         *nbytesp = nbytes;
1129
1130         return 0;
1131 }
1132
1133 static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1134 {
1135         struct iov_iter ii = *ii_p;
1136         int npages = 0;
1137
1138         while (iov_iter_count(&ii) && npages < FUSE_MAX_PAGES_PER_REQ) {
1139                 unsigned long user_addr = fuse_get_user_addr(&ii);
1140                 unsigned offset = user_addr & ~PAGE_MASK;
1141                 size_t frag_size = iov_iter_single_seg_count(&ii);
1142
1143                 npages += (frag_size + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1144                 iov_iter_advance(&ii, frag_size);
1145         }
1146
1147         return min(npages, FUSE_MAX_PAGES_PER_REQ);
1148 }
1149
1150 ssize_t fuse_direct_io(struct file *file, const struct iovec *iov,
1151                        unsigned long nr_segs, size_t count, loff_t *ppos,
1152                        int write)
1153 {
1154         struct fuse_file *ff = file->private_data;
1155         struct fuse_conn *fc = ff->fc;
1156         size_t nmax = write ? fc->max_write : fc->max_read;
1157         loff_t pos = *ppos;
1158         ssize_t res = 0;
1159         struct fuse_req *req;
1160         struct iov_iter ii;
1161
1162         iov_iter_init(&ii, iov, nr_segs, count, 0);
1163
1164         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1165         if (IS_ERR(req))
1166                 return PTR_ERR(req);
1167
1168         while (count) {
1169                 size_t nres;
1170                 fl_owner_t owner = current->files;
1171                 size_t nbytes = min(count, nmax);
1172                 int err = fuse_get_user_pages(req, &ii, &nbytes, write);
1173                 if (err) {
1174                         res = err;
1175                         break;
1176                 }
1177
1178                 if (write)
1179                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1180                 else
1181                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1182
1183                 fuse_release_user_pages(req, !write);
1184                 if (req->out.h.error) {
1185                         if (!res)
1186                                 res = req->out.h.error;
1187                         break;
1188                 } else if (nres > nbytes) {
1189                         res = -EIO;
1190                         break;
1191                 }
1192                 count -= nres;
1193                 res += nres;
1194                 pos += nres;
1195                 if (nres != nbytes)
1196                         break;
1197                 if (count) {
1198                         fuse_put_request(fc, req);
1199                         req = fuse_get_req(fc, fuse_iter_npages(&ii));
1200                         if (IS_ERR(req))
1201                                 break;
1202                 }
1203         }
1204         if (!IS_ERR(req))
1205                 fuse_put_request(fc, req);
1206         if (res > 0)
1207                 *ppos = pos;
1208
1209         return res;
1210 }
1211 EXPORT_SYMBOL_GPL(fuse_direct_io);
1212
1213 static ssize_t __fuse_direct_read(struct file *file, const struct iovec *iov,
1214                                   unsigned long nr_segs, loff_t *ppos)
1215 {
1216         ssize_t res;
1217         struct inode *inode = file_inode(file);
1218
1219         if (is_bad_inode(inode))
1220                 return -EIO;
1221
1222         res = fuse_direct_io(file, iov, nr_segs, iov_length(iov, nr_segs),
1223                              ppos, 0);
1224
1225         fuse_invalidate_attr(inode);
1226
1227         return res;
1228 }
1229
1230 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1231                                      size_t count, loff_t *ppos)
1232 {
1233         struct iovec iov = { .iov_base = buf, .iov_len = count };
1234         return __fuse_direct_read(file, &iov, 1, ppos);
1235 }
1236
1237 static ssize_t __fuse_direct_write(struct file *file, const struct iovec *iov,
1238                                    unsigned long nr_segs, loff_t *ppos)
1239 {
1240         struct inode *inode = file_inode(file);
1241         size_t count = iov_length(iov, nr_segs);
1242         ssize_t res;
1243
1244         res = generic_write_checks(file, ppos, &count, 0);
1245         if (!res) {
1246                 res = fuse_direct_io(file, iov, nr_segs, count, ppos, 1);
1247                 if (res > 0)
1248                         fuse_write_update_size(inode, *ppos);
1249         }
1250
1251         fuse_invalidate_attr(inode);
1252
1253         return res;
1254 }
1255
1256 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1257                                  size_t count, loff_t *ppos)
1258 {
1259         struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1260         struct inode *inode = file_inode(file);
1261         ssize_t res;
1262
1263         if (is_bad_inode(inode))
1264                 return -EIO;
1265
1266         /* Don't allow parallel writes to the same file */
1267         mutex_lock(&inode->i_mutex);
1268         res = __fuse_direct_write(file, &iov, 1, ppos);
1269         mutex_unlock(&inode->i_mutex);
1270
1271         return res;
1272 }
1273
1274 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1275 {
1276         __free_page(req->pages[0]);
1277         fuse_file_put(req->ff, false);
1278 }
1279
1280 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1281 {
1282         struct inode *inode = req->inode;
1283         struct fuse_inode *fi = get_fuse_inode(inode);
1284         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1285
1286         list_del(&req->writepages_entry);
1287         dec_bdi_stat(bdi, BDI_WRITEBACK);
1288         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1289         bdi_writeout_inc(bdi);
1290         wake_up(&fi->page_waitq);
1291 }
1292
1293 /* Called under fc->lock, may release and reacquire it */
1294 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1295 __releases(fc->lock)
1296 __acquires(fc->lock)
1297 {
1298         struct fuse_inode *fi = get_fuse_inode(req->inode);
1299         loff_t size = i_size_read(req->inode);
1300         struct fuse_write_in *inarg = &req->misc.write.in;
1301
1302         if (!fc->connected)
1303                 goto out_free;
1304
1305         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1306                 inarg->size = PAGE_CACHE_SIZE;
1307         } else if (inarg->offset < size) {
1308                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1309         } else {
1310                 /* Got truncated off completely */
1311                 goto out_free;
1312         }
1313
1314         req->in.args[1].size = inarg->size;
1315         fi->writectr++;
1316         fuse_request_send_background_locked(fc, req);
1317         return;
1318
1319  out_free:
1320         fuse_writepage_finish(fc, req);
1321         spin_unlock(&fc->lock);
1322         fuse_writepage_free(fc, req);
1323         fuse_put_request(fc, req);
1324         spin_lock(&fc->lock);
1325 }
1326
1327 /*
1328  * If fi->writectr is positive (no truncate or fsync going on) send
1329  * all queued writepage requests.
1330  *
1331  * Called with fc->lock
1332  */
1333 void fuse_flush_writepages(struct inode *inode)
1334 __releases(fc->lock)
1335 __acquires(fc->lock)
1336 {
1337         struct fuse_conn *fc = get_fuse_conn(inode);
1338         struct fuse_inode *fi = get_fuse_inode(inode);
1339         struct fuse_req *req;
1340
1341         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1342                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1343                 list_del_init(&req->list);
1344                 fuse_send_writepage(fc, req);
1345         }
1346 }
1347
1348 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1349 {
1350         struct inode *inode = req->inode;
1351         struct fuse_inode *fi = get_fuse_inode(inode);
1352
1353         mapping_set_error(inode->i_mapping, req->out.h.error);
1354         spin_lock(&fc->lock);
1355         fi->writectr--;
1356         fuse_writepage_finish(fc, req);
1357         spin_unlock(&fc->lock);
1358         fuse_writepage_free(fc, req);
1359 }
1360
1361 static int fuse_writepage_locked(struct page *page)
1362 {
1363         struct address_space *mapping = page->mapping;
1364         struct inode *inode = mapping->host;
1365         struct fuse_conn *fc = get_fuse_conn(inode);
1366         struct fuse_inode *fi = get_fuse_inode(inode);
1367         struct fuse_req *req;
1368         struct fuse_file *ff;
1369         struct page *tmp_page;
1370
1371         set_page_writeback(page);
1372
1373         req = fuse_request_alloc_nofs(1);
1374         if (!req)
1375                 goto err;
1376
1377         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1378         if (!tmp_page)
1379                 goto err_free;
1380
1381         spin_lock(&fc->lock);
1382         BUG_ON(list_empty(&fi->write_files));
1383         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1384         req->ff = fuse_file_get(ff);
1385         spin_unlock(&fc->lock);
1386
1387         fuse_write_fill(req, ff, page_offset(page), 0);
1388
1389         copy_highpage(tmp_page, page);
1390         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1391         req->in.argpages = 1;
1392         req->num_pages = 1;
1393         req->pages[0] = tmp_page;
1394         req->page_descs[0].offset = 0;
1395         req->page_descs[0].length = PAGE_SIZE;
1396         req->end = fuse_writepage_end;
1397         req->inode = inode;
1398
1399         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1400         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1401         end_page_writeback(page);
1402
1403         spin_lock(&fc->lock);
1404         list_add(&req->writepages_entry, &fi->writepages);
1405         list_add_tail(&req->list, &fi->queued_writes);
1406         fuse_flush_writepages(inode);
1407         spin_unlock(&fc->lock);
1408
1409         return 0;
1410
1411 err_free:
1412         fuse_request_free(req);
1413 err:
1414         end_page_writeback(page);
1415         return -ENOMEM;
1416 }
1417
1418 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1419 {
1420         int err;
1421
1422         err = fuse_writepage_locked(page);
1423         unlock_page(page);
1424
1425         return err;
1426 }
1427
1428 static int fuse_launder_page(struct page *page)
1429 {
1430         int err = 0;
1431         if (clear_page_dirty_for_io(page)) {
1432                 struct inode *inode = page->mapping->host;
1433                 err = fuse_writepage_locked(page);
1434                 if (!err)
1435                         fuse_wait_on_page_writeback(inode, page->index);
1436         }
1437         return err;
1438 }
1439
1440 /*
1441  * Write back dirty pages now, because there may not be any suitable
1442  * open files later
1443  */
1444 static void fuse_vma_close(struct vm_area_struct *vma)
1445 {
1446         filemap_write_and_wait(vma->vm_file->f_mapping);
1447 }
1448
1449 /*
1450  * Wait for writeback against this page to complete before allowing it
1451  * to be marked dirty again, and hence written back again, possibly
1452  * before the previous writepage completed.
1453  *
1454  * Block here, instead of in ->writepage(), so that the userspace fs
1455  * can only block processes actually operating on the filesystem.
1456  *
1457  * Otherwise unprivileged userspace fs would be able to block
1458  * unrelated:
1459  *
1460  * - page migration
1461  * - sync(2)
1462  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1463  */
1464 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1465 {
1466         struct page *page = vmf->page;
1467         /*
1468          * Don't use page->mapping as it may become NULL from a
1469          * concurrent truncate.
1470          */
1471         struct inode *inode = vma->vm_file->f_mapping->host;
1472
1473         fuse_wait_on_page_writeback(inode, page->index);
1474         return 0;
1475 }
1476
1477 static const struct vm_operations_struct fuse_file_vm_ops = {
1478         .close          = fuse_vma_close,
1479         .fault          = filemap_fault,
1480         .page_mkwrite   = fuse_page_mkwrite,
1481         .remap_pages    = generic_file_remap_pages,
1482 };
1483
1484 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1485 {
1486         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1487                 struct inode *inode = file_inode(file);
1488                 struct fuse_conn *fc = get_fuse_conn(inode);
1489                 struct fuse_inode *fi = get_fuse_inode(inode);
1490                 struct fuse_file *ff = file->private_data;
1491                 /*
1492                  * file may be written through mmap, so chain it onto the
1493                  * inodes's write_file list
1494                  */
1495                 spin_lock(&fc->lock);
1496                 if (list_empty(&ff->write_entry))
1497                         list_add(&ff->write_entry, &fi->write_files);
1498                 spin_unlock(&fc->lock);
1499         }
1500         file_accessed(file);
1501         vma->vm_ops = &fuse_file_vm_ops;
1502         return 0;
1503 }
1504
1505 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1506 {
1507         /* Can't provide the coherency needed for MAP_SHARED */
1508         if (vma->vm_flags & VM_MAYSHARE)
1509                 return -ENODEV;
1510
1511         invalidate_inode_pages2(file->f_mapping);
1512
1513         return generic_file_mmap(file, vma);
1514 }
1515
1516 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1517                                   struct file_lock *fl)
1518 {
1519         switch (ffl->type) {
1520         case F_UNLCK:
1521                 break;
1522
1523         case F_RDLCK:
1524         case F_WRLCK:
1525                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1526                     ffl->end < ffl->start)
1527                         return -EIO;
1528
1529                 fl->fl_start = ffl->start;
1530                 fl->fl_end = ffl->end;
1531                 fl->fl_pid = ffl->pid;
1532                 break;
1533
1534         default:
1535                 return -EIO;
1536         }
1537         fl->fl_type = ffl->type;
1538         return 0;
1539 }
1540
1541 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1542                          const struct file_lock *fl, int opcode, pid_t pid,
1543                          int flock)
1544 {
1545         struct inode *inode = file_inode(file);
1546         struct fuse_conn *fc = get_fuse_conn(inode);
1547         struct fuse_file *ff = file->private_data;
1548         struct fuse_lk_in *arg = &req->misc.lk_in;
1549
1550         arg->fh = ff->fh;
1551         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1552         arg->lk.start = fl->fl_start;
1553         arg->lk.end = fl->fl_end;
1554         arg->lk.type = fl->fl_type;
1555         arg->lk.pid = pid;
1556         if (flock)
1557                 arg->lk_flags |= FUSE_LK_FLOCK;
1558         req->in.h.opcode = opcode;
1559         req->in.h.nodeid = get_node_id(inode);
1560         req->in.numargs = 1;
1561         req->in.args[0].size = sizeof(*arg);
1562         req->in.args[0].value = arg;
1563 }
1564
1565 static int fuse_getlk(struct file *file, struct file_lock *fl)
1566 {
1567         struct inode *inode = file_inode(file);
1568         struct fuse_conn *fc = get_fuse_conn(inode);
1569         struct fuse_req *req;
1570         struct fuse_lk_out outarg;
1571         int err;
1572
1573         req = fuse_get_req_nopages(fc);
1574         if (IS_ERR(req))
1575                 return PTR_ERR(req);
1576
1577         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1578         req->out.numargs = 1;
1579         req->out.args[0].size = sizeof(outarg);
1580         req->out.args[0].value = &outarg;
1581         fuse_request_send(fc, req);
1582         err = req->out.h.error;
1583         fuse_put_request(fc, req);
1584         if (!err)
1585                 err = convert_fuse_file_lock(&outarg.lk, fl);
1586
1587         return err;
1588 }
1589
1590 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1591 {
1592         struct inode *inode = file_inode(file);
1593         struct fuse_conn *fc = get_fuse_conn(inode);
1594         struct fuse_req *req;
1595         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1596         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1597         int err;
1598
1599         if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
1600                 /* NLM needs asynchronous locks, which we don't support yet */
1601                 return -ENOLCK;
1602         }
1603
1604         /* Unlock on close is handled by the flush method */
1605         if (fl->fl_flags & FL_CLOSE)
1606                 return 0;
1607
1608         req = fuse_get_req_nopages(fc);
1609         if (IS_ERR(req))
1610                 return PTR_ERR(req);
1611
1612         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1613         fuse_request_send(fc, req);
1614         err = req->out.h.error;
1615         /* locking is restartable */
1616         if (err == -EINTR)
1617                 err = -ERESTARTSYS;
1618         fuse_put_request(fc, req);
1619         return err;
1620 }
1621
1622 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1623 {
1624         struct inode *inode = file_inode(file);
1625         struct fuse_conn *fc = get_fuse_conn(inode);
1626         int err;
1627
1628         if (cmd == F_CANCELLK) {
1629                 err = 0;
1630         } else if (cmd == F_GETLK) {
1631                 if (fc->no_lock) {
1632                         posix_test_lock(file, fl);
1633                         err = 0;
1634                 } else
1635                         err = fuse_getlk(file, fl);
1636         } else {
1637                 if (fc->no_lock)
1638                         err = posix_lock_file(file, fl, NULL);
1639                 else
1640                         err = fuse_setlk(file, fl, 0);
1641         }
1642         return err;
1643 }
1644
1645 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1646 {
1647         struct inode *inode = file_inode(file);
1648         struct fuse_conn *fc = get_fuse_conn(inode);
1649         int err;
1650
1651         if (fc->no_flock) {
1652                 err = flock_lock_file_wait(file, fl);
1653         } else {
1654                 struct fuse_file *ff = file->private_data;
1655
1656                 /* emulate flock with POSIX locks */
1657                 fl->fl_owner = (fl_owner_t) file;
1658                 ff->flock = true;
1659                 err = fuse_setlk(file, fl, 1);
1660         }
1661
1662         return err;
1663 }
1664
1665 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1666 {
1667         struct inode *inode = mapping->host;
1668         struct fuse_conn *fc = get_fuse_conn(inode);
1669         struct fuse_req *req;
1670         struct fuse_bmap_in inarg;
1671         struct fuse_bmap_out outarg;
1672         int err;
1673
1674         if (!inode->i_sb->s_bdev || fc->no_bmap)
1675                 return 0;
1676
1677         req = fuse_get_req_nopages(fc);
1678         if (IS_ERR(req))
1679                 return 0;
1680
1681         memset(&inarg, 0, sizeof(inarg));
1682         inarg.block = block;
1683         inarg.blocksize = inode->i_sb->s_blocksize;
1684         req->in.h.opcode = FUSE_BMAP;
1685         req->in.h.nodeid = get_node_id(inode);
1686         req->in.numargs = 1;
1687         req->in.args[0].size = sizeof(inarg);
1688         req->in.args[0].value = &inarg;
1689         req->out.numargs = 1;
1690         req->out.args[0].size = sizeof(outarg);
1691         req->out.args[0].value = &outarg;
1692         fuse_request_send(fc, req);
1693         err = req->out.h.error;
1694         fuse_put_request(fc, req);
1695         if (err == -ENOSYS)
1696                 fc->no_bmap = 1;
1697
1698         return err ? 0 : outarg.block;
1699 }
1700
1701 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
1702 {
1703         loff_t retval;
1704         struct inode *inode = file_inode(file);
1705
1706         /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
1707         if (whence == SEEK_CUR || whence == SEEK_SET)
1708                 return generic_file_llseek(file, offset, whence);
1709
1710         mutex_lock(&inode->i_mutex);
1711         retval = fuse_update_attributes(inode, NULL, file, NULL);
1712         if (!retval)
1713                 retval = generic_file_llseek(file, offset, whence);
1714         mutex_unlock(&inode->i_mutex);
1715
1716         return retval;
1717 }
1718
1719 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1720                         unsigned int nr_segs, size_t bytes, bool to_user)
1721 {
1722         struct iov_iter ii;
1723         int page_idx = 0;
1724
1725         if (!bytes)
1726                 return 0;
1727
1728         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1729
1730         while (iov_iter_count(&ii)) {
1731                 struct page *page = pages[page_idx++];
1732                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1733                 void *kaddr;
1734
1735                 kaddr = kmap(page);
1736
1737                 while (todo) {
1738                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1739                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1740                         size_t copy = min(todo, iov_len);
1741                         size_t left;
1742
1743                         if (!to_user)
1744                                 left = copy_from_user(kaddr, uaddr, copy);
1745                         else
1746                                 left = copy_to_user(uaddr, kaddr, copy);
1747
1748                         if (unlikely(left))
1749                                 return -EFAULT;
1750
1751                         iov_iter_advance(&ii, copy);
1752                         todo -= copy;
1753                         kaddr += copy;
1754                 }
1755
1756                 kunmap(page);
1757         }
1758
1759         return 0;
1760 }
1761
1762 /*
1763  * CUSE servers compiled on 32bit broke on 64bit kernels because the
1764  * ABI was defined to be 'struct iovec' which is different on 32bit
1765  * and 64bit.  Fortunately we can determine which structure the server
1766  * used from the size of the reply.
1767  */
1768 static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
1769                                      size_t transferred, unsigned count,
1770                                      bool is_compat)
1771 {
1772 #ifdef CONFIG_COMPAT
1773         if (count * sizeof(struct compat_iovec) == transferred) {
1774                 struct compat_iovec *ciov = src;
1775                 unsigned i;
1776
1777                 /*
1778                  * With this interface a 32bit server cannot support
1779                  * non-compat (i.e. ones coming from 64bit apps) ioctl
1780                  * requests
1781                  */
1782                 if (!is_compat)
1783                         return -EINVAL;
1784
1785                 for (i = 0; i < count; i++) {
1786                         dst[i].iov_base = compat_ptr(ciov[i].iov_base);
1787                         dst[i].iov_len = ciov[i].iov_len;
1788                 }
1789                 return 0;
1790         }
1791 #endif
1792
1793         if (count * sizeof(struct iovec) != transferred)
1794                 return -EIO;
1795
1796         memcpy(dst, src, transferred);
1797         return 0;
1798 }
1799
1800 /* Make sure iov_length() won't overflow */
1801 static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
1802 {
1803         size_t n;
1804         u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
1805
1806         for (n = 0; n < count; n++, iov++) {
1807                 if (iov->iov_len > (size_t) max)
1808                         return -ENOMEM;
1809                 max -= iov->iov_len;
1810         }
1811         return 0;
1812 }
1813
1814 static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
1815                                  void *src, size_t transferred, unsigned count,
1816                                  bool is_compat)
1817 {
1818         unsigned i;
1819         struct fuse_ioctl_iovec *fiov = src;
1820
1821         if (fc->minor < 16) {
1822                 return fuse_copy_ioctl_iovec_old(dst, src, transferred,
1823                                                  count, is_compat);
1824         }
1825
1826         if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
1827                 return -EIO;
1828
1829         for (i = 0; i < count; i++) {
1830                 /* Did the server supply an inappropriate value? */
1831                 if (fiov[i].base != (unsigned long) fiov[i].base ||
1832                     fiov[i].len != (unsigned long) fiov[i].len)
1833                         return -EIO;
1834
1835                 dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
1836                 dst[i].iov_len = (size_t) fiov[i].len;
1837
1838 #ifdef CONFIG_COMPAT
1839                 if (is_compat &&
1840                     (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
1841                      (compat_size_t) dst[i].iov_len != fiov[i].len))
1842                         return -EIO;
1843 #endif
1844         }
1845
1846         return 0;
1847 }
1848
1849
1850 /*
1851  * For ioctls, there is no generic way to determine how much memory
1852  * needs to be read and/or written.  Furthermore, ioctls are allowed
1853  * to dereference the passed pointer, so the parameter requires deep
1854  * copying but FUSE has no idea whatsoever about what to copy in or
1855  * out.
1856  *
1857  * This is solved by allowing FUSE server to retry ioctl with
1858  * necessary in/out iovecs.  Let's assume the ioctl implementation
1859  * needs to read in the following structure.
1860  *
1861  * struct a {
1862  *      char    *buf;
1863  *      size_t  buflen;
1864  * }
1865  *
1866  * On the first callout to FUSE server, inarg->in_size and
1867  * inarg->out_size will be NULL; then, the server completes the ioctl
1868  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1869  * the actual iov array to
1870  *
1871  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1872  *
1873  * which tells FUSE to copy in the requested area and retry the ioctl.
1874  * On the second round, the server has access to the structure and
1875  * from that it can tell what to look for next, so on the invocation,
1876  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1877  *
1878  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1879  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1880  *
1881  * FUSE will copy both struct a and the pointed buffer from the
1882  * process doing the ioctl and retry ioctl with both struct a and the
1883  * buffer.
1884  *
1885  * This time, FUSE server has everything it needs and completes ioctl
1886  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1887  *
1888  * Copying data out works the same way.
1889  *
1890  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1891  * automatically initializes in and out iovs by decoding @cmd with
1892  * _IOC_* macros and the server is not allowed to request RETRY.  This
1893  * limits ioctl data transfers to well-formed ioctls and is the forced
1894  * behavior for all FUSE servers.
1895  */
1896 long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
1897                    unsigned int flags)
1898 {
1899         struct fuse_file *ff = file->private_data;
1900         struct fuse_conn *fc = ff->fc;
1901         struct fuse_ioctl_in inarg = {
1902                 .fh = ff->fh,
1903                 .cmd = cmd,
1904                 .arg = arg,
1905                 .flags = flags
1906         };
1907         struct fuse_ioctl_out outarg;
1908         struct fuse_req *req = NULL;
1909         struct page **pages = NULL;
1910         struct iovec *iov_page = NULL;
1911         struct iovec *in_iov = NULL, *out_iov = NULL;
1912         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1913         size_t in_size, out_size, transferred;
1914         int err;
1915
1916 #if BITS_PER_LONG == 32
1917         inarg.flags |= FUSE_IOCTL_32BIT;
1918 #else
1919         if (flags & FUSE_IOCTL_COMPAT)
1920                 inarg.flags |= FUSE_IOCTL_32BIT;
1921 #endif
1922
1923         /* assume all the iovs returned by client always fits in a page */
1924         BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1925
1926         err = -ENOMEM;
1927         pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
1928         iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
1929         if (!pages || !iov_page)
1930                 goto out;
1931
1932         /*
1933          * If restricted, initialize IO parameters as encoded in @cmd.
1934          * RETRY from server is not allowed.
1935          */
1936         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1937                 struct iovec *iov = iov_page;
1938
1939                 iov->iov_base = (void __user *)arg;
1940                 iov->iov_len = _IOC_SIZE(cmd);
1941
1942                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1943                         in_iov = iov;
1944                         in_iovs = 1;
1945                 }
1946
1947                 if (_IOC_DIR(cmd) & _IOC_READ) {
1948                         out_iov = iov;
1949                         out_iovs = 1;
1950                 }
1951         }
1952
1953  retry:
1954         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1955         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1956
1957         /*
1958          * Out data can be used either for actual out data or iovs,
1959          * make sure there always is at least one page.
1960          */
1961         out_size = max_t(size_t, out_size, PAGE_SIZE);
1962         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1963
1964         /* make sure there are enough buffer pages and init request with them */
1965         err = -ENOMEM;
1966         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1967                 goto out;
1968         while (num_pages < max_pages) {
1969                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1970                 if (!pages[num_pages])
1971                         goto out;
1972                 num_pages++;
1973         }
1974
1975         req = fuse_get_req(fc, num_pages);
1976         if (IS_ERR(req)) {
1977                 err = PTR_ERR(req);
1978                 req = NULL;
1979                 goto out;
1980         }
1981         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1982         req->num_pages = num_pages;
1983         fuse_page_descs_length_init(req, 0, req->num_pages);
1984
1985         /* okay, let's send it to the client */
1986         req->in.h.opcode = FUSE_IOCTL;
1987         req->in.h.nodeid = ff->nodeid;
1988         req->in.numargs = 1;
1989         req->in.args[0].size = sizeof(inarg);
1990         req->in.args[0].value = &inarg;
1991         if (in_size) {
1992                 req->in.numargs++;
1993                 req->in.args[1].size = in_size;
1994                 req->in.argpages = 1;
1995
1996                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1997                                            false);
1998                 if (err)
1999                         goto out;
2000         }
2001
2002         req->out.numargs = 2;
2003         req->out.args[0].size = sizeof(outarg);
2004         req->out.args[0].value = &outarg;
2005         req->out.args[1].size = out_size;
2006         req->out.argpages = 1;
2007         req->out.argvar = 1;
2008
2009         fuse_request_send(fc, req);
2010         err = req->out.h.error;
2011         transferred = req->out.args[1].size;
2012         fuse_put_request(fc, req);
2013         req = NULL;
2014         if (err)
2015                 goto out;
2016
2017         /* did it ask for retry? */
2018         if (outarg.flags & FUSE_IOCTL_RETRY) {
2019                 void *vaddr;
2020
2021                 /* no retry if in restricted mode */
2022                 err = -EIO;
2023                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2024                         goto out;
2025
2026                 in_iovs = outarg.in_iovs;
2027                 out_iovs = outarg.out_iovs;
2028
2029                 /*
2030                  * Make sure things are in boundary, separate checks
2031                  * are to protect against overflow.
2032                  */
2033                 err = -ENOMEM;
2034                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2035                     out_iovs > FUSE_IOCTL_MAX_IOV ||
2036                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2037                         goto out;
2038
2039                 vaddr = kmap_atomic(pages[0]);
2040                 err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2041                                             transferred, in_iovs + out_iovs,
2042                                             (flags & FUSE_IOCTL_COMPAT) != 0);
2043                 kunmap_atomic(vaddr);
2044                 if (err)
2045                         goto out;
2046
2047                 in_iov = iov_page;
2048                 out_iov = in_iov + in_iovs;
2049
2050                 err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2051                 if (err)
2052                         goto out;
2053
2054                 err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2055                 if (err)
2056                         goto out;
2057
2058                 goto retry;
2059         }
2060
2061         err = -EIO;
2062         if (transferred > inarg.out_size)
2063                 goto out;
2064
2065         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2066  out:
2067         if (req)
2068                 fuse_put_request(fc, req);
2069         free_page((unsigned long) iov_page);
2070         while (num_pages)
2071                 __free_page(pages[--num_pages]);
2072         kfree(pages);
2073
2074         return err ? err : outarg.result;
2075 }
2076 EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2077
2078 long fuse_ioctl_common(struct file *file, unsigned int cmd,
2079                        unsigned long arg, unsigned int flags)
2080 {
2081         struct inode *inode = file_inode(file);
2082         struct fuse_conn *fc = get_fuse_conn(inode);
2083
2084         if (!fuse_allow_current_process(fc))
2085                 return -EACCES;
2086
2087         if (is_bad_inode(inode))
2088                 return -EIO;
2089
2090         return fuse_do_ioctl(file, cmd, arg, flags);
2091 }
2092
2093 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2094                             unsigned long arg)
2095 {
2096         return fuse_ioctl_common(file, cmd, arg, 0);
2097 }
2098
2099 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2100                                    unsigned long arg)
2101 {
2102         return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2103 }
2104
2105 /*
2106  * All files which have been polled are linked to RB tree
2107  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2108  * find the matching one.
2109  */
2110 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2111                                               struct rb_node **parent_out)
2112 {
2113         struct rb_node **link = &fc->polled_files.rb_node;
2114         struct rb_node *last = NULL;
2115
2116         while (*link) {
2117                 struct fuse_file *ff;
2118
2119                 last = *link;
2120                 ff = rb_entry(last, struct fuse_file, polled_node);
2121
2122                 if (kh < ff->kh)
2123                         link = &last->rb_left;
2124                 else if (kh > ff->kh)
2125                         link = &last->rb_right;
2126                 else
2127                         return link;
2128         }
2129
2130         if (parent_out)
2131                 *parent_out = last;
2132         return link;
2133 }
2134
2135 /*
2136  * The file is about to be polled.  Make sure it's on the polled_files
2137  * RB tree.  Note that files once added to the polled_files tree are
2138  * not removed before the file is released.  This is because a file
2139  * polled once is likely to be polled again.
2140  */
2141 static void fuse_register_polled_file(struct fuse_conn *fc,
2142                                       struct fuse_file *ff)
2143 {
2144         spin_lock(&fc->lock);
2145         if (RB_EMPTY_NODE(&ff->polled_node)) {
2146                 struct rb_node **link, *parent;
2147
2148                 link = fuse_find_polled_node(fc, ff->kh, &parent);
2149                 BUG_ON(*link);
2150                 rb_link_node(&ff->polled_node, parent, link);
2151                 rb_insert_color(&ff->polled_node, &fc->polled_files);
2152         }
2153         spin_unlock(&fc->lock);
2154 }
2155
2156 unsigned fuse_file_poll(struct file *file, poll_table *wait)
2157 {
2158         struct fuse_file *ff = file->private_data;
2159         struct fuse_conn *fc = ff->fc;
2160         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2161         struct fuse_poll_out outarg;
2162         struct fuse_req *req;
2163         int err;
2164
2165         if (fc->no_poll)
2166                 return DEFAULT_POLLMASK;
2167
2168         poll_wait(file, &ff->poll_wait, wait);
2169         inarg.events = (__u32)poll_requested_events(wait);
2170
2171         /*
2172          * Ask for notification iff there's someone waiting for it.
2173          * The client may ignore the flag and always notify.
2174          */
2175         if (waitqueue_active(&ff->poll_wait)) {
2176                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2177                 fuse_register_polled_file(fc, ff);
2178         }
2179
2180         req = fuse_get_req_nopages(fc);
2181         if (IS_ERR(req))
2182                 return POLLERR;
2183
2184         req->in.h.opcode = FUSE_POLL;
2185         req->in.h.nodeid = ff->nodeid;
2186         req->in.numargs = 1;
2187         req->in.args[0].size = sizeof(inarg);
2188         req->in.args[0].value = &inarg;
2189         req->out.numargs = 1;
2190         req->out.args[0].size = sizeof(outarg);
2191         req->out.args[0].value = &outarg;
2192         fuse_request_send(fc, req);
2193         err = req->out.h.error;
2194         fuse_put_request(fc, req);
2195
2196         if (!err)
2197                 return outarg.revents;
2198         if (err == -ENOSYS) {
2199                 fc->no_poll = 1;
2200                 return DEFAULT_POLLMASK;
2201         }
2202         return POLLERR;
2203 }
2204 EXPORT_SYMBOL_GPL(fuse_file_poll);
2205
2206 /*
2207  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2208  * wakes up the poll waiters.
2209  */
2210 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2211                             struct fuse_notify_poll_wakeup_out *outarg)
2212 {
2213         u64 kh = outarg->kh;
2214         struct rb_node **link;
2215
2216         spin_lock(&fc->lock);
2217
2218         link = fuse_find_polled_node(fc, kh, NULL);
2219         if (*link) {
2220                 struct fuse_file *ff;
2221
2222                 ff = rb_entry(*link, struct fuse_file, polled_node);
2223                 wake_up_interruptible_sync(&ff->poll_wait);
2224         }
2225
2226         spin_unlock(&fc->lock);
2227         return 0;
2228 }
2229
2230 static ssize_t
2231 fuse_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2232                         loff_t offset, unsigned long nr_segs)
2233 {
2234         ssize_t ret = 0;
2235         struct file *file = NULL;
2236         loff_t pos = 0;
2237
2238         file = iocb->ki_filp;
2239         pos = offset;
2240
2241         if (rw == WRITE)
2242                 ret = __fuse_direct_write(file, iov, nr_segs, &pos);
2243         else
2244                 ret = __fuse_direct_read(file, iov, nr_segs, &pos);
2245
2246         return ret;
2247 }
2248
2249 static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2250                                 loff_t length)
2251 {
2252         struct fuse_file *ff = file->private_data;
2253         struct fuse_conn *fc = ff->fc;
2254         struct fuse_req *req;
2255         struct fuse_fallocate_in inarg = {
2256                 .fh = ff->fh,
2257                 .offset = offset,
2258                 .length = length,
2259                 .mode = mode
2260         };
2261         int err;
2262
2263         if (fc->no_fallocate)
2264                 return -EOPNOTSUPP;
2265
2266         req = fuse_get_req_nopages(fc);
2267         if (IS_ERR(req))
2268                 return PTR_ERR(req);
2269
2270         req->in.h.opcode = FUSE_FALLOCATE;
2271         req->in.h.nodeid = ff->nodeid;
2272         req->in.numargs = 1;
2273         req->in.args[0].size = sizeof(inarg);
2274         req->in.args[0].value = &inarg;
2275         fuse_request_send(fc, req);
2276         err = req->out.h.error;
2277         if (err == -ENOSYS) {
2278                 fc->no_fallocate = 1;
2279                 err = -EOPNOTSUPP;
2280         }
2281         fuse_put_request(fc, req);
2282
2283         return err;
2284 }
2285
2286 static const struct file_operations fuse_file_operations = {
2287         .llseek         = fuse_file_llseek,
2288         .read           = do_sync_read,
2289         .aio_read       = fuse_file_aio_read,
2290         .write          = do_sync_write,
2291         .aio_write      = fuse_file_aio_write,
2292         .mmap           = fuse_file_mmap,
2293         .open           = fuse_open,
2294         .flush          = fuse_flush,
2295         .release        = fuse_release,
2296         .fsync          = fuse_fsync,
2297         .lock           = fuse_file_lock,
2298         .flock          = fuse_file_flock,
2299         .splice_read    = generic_file_splice_read,
2300         .unlocked_ioctl = fuse_file_ioctl,
2301         .compat_ioctl   = fuse_file_compat_ioctl,
2302         .poll           = fuse_file_poll,
2303         .fallocate      = fuse_file_fallocate,
2304 };
2305
2306 static const struct file_operations fuse_direct_io_file_operations = {
2307         .llseek         = fuse_file_llseek,
2308         .read           = fuse_direct_read,
2309         .write          = fuse_direct_write,
2310         .mmap           = fuse_direct_mmap,
2311         .open           = fuse_open,
2312         .flush          = fuse_flush,
2313         .release        = fuse_release,
2314         .fsync          = fuse_fsync,
2315         .lock           = fuse_file_lock,
2316         .flock          = fuse_file_flock,
2317         .unlocked_ioctl = fuse_file_ioctl,
2318         .compat_ioctl   = fuse_file_compat_ioctl,
2319         .poll           = fuse_file_poll,
2320         .fallocate      = fuse_file_fallocate,
2321         /* no splice_read */
2322 };
2323
2324 static const struct address_space_operations fuse_file_aops  = {
2325         .readpage       = fuse_readpage,
2326         .writepage      = fuse_writepage,
2327         .launder_page   = fuse_launder_page,
2328         .readpages      = fuse_readpages,
2329         .set_page_dirty = __set_page_dirty_nobuffers,
2330         .bmap           = fuse_bmap,
2331         .direct_IO      = fuse_direct_IO,
2332 };
2333
2334 void fuse_init_file_inode(struct inode *inode)
2335 {
2336         inode->i_fop = &fuse_file_operations;
2337         inode->i_data.a_ops = &fuse_file_aops;
2338 }