fs/fuse/file.c

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