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