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