fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2006  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         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(void)
  50 {
  51         struct fuse_file *ff;
  52         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  53         if (ff) {
  54                 ff->reserved_req = fuse_request_alloc();
  55                 if (!ff->reserved_req) {
  56                         kfree(ff);
  57                         ff = NULL;
  58                 } else {
  59                         INIT_LIST_HEAD(&ff->write_entry);
  60                         atomic_set(&ff->count, 0);
  61                 }
  62         }
  63         return ff;
  64 }
  65
  66 void fuse_file_free(struct fuse_file *ff)
  67 {
  68         fuse_request_free(ff->reserved_req);
  69         kfree(ff);
  70 }
  71
  72 static struct fuse_file *fuse_file_get(struct fuse_file *ff)
  73 {
  74         atomic_inc(&ff->count);
  75         return ff;
  76 }
  77
  78 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  79 {
  80         dput(req->misc.release.dentry);
  81         mntput(req->misc.release.vfsmount);
  82         fuse_put_request(fc, req);
  83 }
  84
  85 static void fuse_file_put(struct fuse_file *ff)
  86 {
  87         if (atomic_dec_and_test(&ff->count)) {
  88                 struct fuse_req *req = ff->reserved_req;
  89                 struct inode *inode = req->misc.release.dentry->d_inode;
  90                 struct fuse_conn *fc = get_fuse_conn(inode);
  91                 req->end = fuse_release_end;
  92                 request_send_background(fc, req);
  93                 kfree(ff);
  94         }
  95 }
  96
  97 void fuse_finish_open(struct inode *inode, struct file *file,
  98                       struct fuse_file *ff, struct fuse_open_out *outarg)
  99 {
 100         if (outarg->open_flags & FOPEN_DIRECT_IO)
 101                 file->f_op = &fuse_direct_io_file_operations;
 102         if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
 103                 invalidate_inode_pages2(inode->i_mapping);
 104         ff->fh = outarg->fh;
 105         file->private_data = fuse_file_get(ff);
 106 }
 107
 108 int fuse_open_common(struct inode *inode, struct file *file, int isdir)
 109 {
 110         struct fuse_open_out outarg;
 111         struct fuse_file *ff;
 112         int err;
 113
 114         /* VFS checks this, but only _after_ ->open() */
 115         if (file->f_flags & O_DIRECT)
 116                 return -EINVAL;
 117
 118         err = generic_file_open(inode, file);
 119         if (err)
 120                 return err;
 121
 122         ff = fuse_file_alloc();
 123         if (!ff)
 124                 return -ENOMEM;
 125
 126         err = fuse_send_open(inode, file, isdir, &outarg);
 127         if (err)
 128                 fuse_file_free(ff);
 129         else {
 130                 if (isdir)
 131                         outarg.open_flags &= ~FOPEN_DIRECT_IO;
 132                 fuse_finish_open(inode, file, ff, &outarg);
 133         }
 134
 135         return err;
 136 }
 137
 138 void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
 139 {
 140         struct fuse_req *req = ff->reserved_req;
 141         struct fuse_release_in *inarg = &req->misc.release.in;
 142
 143         inarg->fh = ff->fh;
 144         inarg->flags = flags;
 145         req->in.h.opcode = opcode;
 146         req->in.h.nodeid = nodeid;
 147         req->in.numargs = 1;
 148         req->in.args[0].size = sizeof(struct fuse_release_in);
 149         req->in.args[0].value = inarg;
 150 }
 151
 152 int fuse_release_common(struct inode *inode, struct file *file, int isdir)
 153 {
 154         struct fuse_file *ff = file->private_data;
 155         if (ff) {
 156                 struct fuse_conn *fc = get_fuse_conn(inode);
 157                 struct fuse_req *req = ff->reserved_req;
 158
 159                 fuse_release_fill(ff, get_node_id(inode), file->f_flags,
 160                                   isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
 161
 162                 /* Hold vfsmount and dentry until release is finished */
 163                 req->misc.release.vfsmount = mntget(file->f_path.mnt);
 164                 req->misc.release.dentry = dget(file->f_path.dentry);
 165
 166                 spin_lock(&fc->lock);
 167                 list_del(&ff->write_entry);
 168                 spin_unlock(&fc->lock);
 169                 /*
 170                  * Normally this will send the RELEASE request,
 171                  * however if some asynchronous READ or WRITE requests
 172                  * are outstanding, the sending will be delayed
 173                  */
 174                 fuse_file_put(ff);
 175         }
 176
 177         /* Return value is ignored by VFS */
 178         return 0;
 179 }
 180
 181 static int fuse_open(struct inode *inode, struct file *file)
 182 {
 183         return fuse_open_common(inode, file, 0);
 184 }
 185
 186 static int fuse_release(struct inode *inode, struct file *file)
 187 {
 188         return fuse_release_common(inode, file, 0);
 189 }
 190
 191 /*
 192  * Scramble the ID space with XTEA, so that the value of the files_struct
 193  * pointer is not exposed to userspace.
 194  */
 195 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 196 {
 197         u32 *k = fc->scramble_key;
 198         u64 v = (unsigned long) id;
 199         u32 v0 = v;
 200         u32 v1 = v >> 32;
 201         u32 sum = 0;
 202         int i;
 203
 204         for (i = 0; i < 32; i++) {
 205                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 206                 sum += 0x9E3779B9;
 207                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 208         }
 209
 210         return (u64) v0 + ((u64) v1 << 32);
 211 }
 212
 213 /*
 214  * Check if page is under writeback
 215  *
 216  * This is currently done by walking the list of writepage requests
 217  * for the inode, which can be pretty inefficient.
 218  */
 219 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 220 {
 221         struct fuse_conn *fc = get_fuse_conn(inode);
 222         struct fuse_inode *fi = get_fuse_inode(inode);
 223         struct fuse_req *req;
 224         bool found = false;
 225
 226         spin_lock(&fc->lock);
 227         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 228                 pgoff_t curr_index;
 229
 230                 BUG_ON(req->inode != inode);
 231                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 232                 if (curr_index == index) {
 233                         found = true;
 234                         break;
 235                 }
 236         }
 237         spin_unlock(&fc->lock);
 238
 239         return found;
 240 }
 241
 242 /*
 243  * Wait for page writeback to be completed.
 244  *
 245  * Since fuse doesn't rely on the VM writeback tracking, this has to
 246  * use some other means.
 247  */
 248 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 249 {
 250         struct fuse_inode *fi = get_fuse_inode(inode);
 251
 252         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 253         return 0;
 254 }
 255
 256 static int fuse_flush(struct file *file, fl_owner_t id)
 257 {
 258         struct inode *inode = file->f_path.dentry->d_inode;
 259         struct fuse_conn *fc = get_fuse_conn(inode);
 260         struct fuse_file *ff = file->private_data;
 261         struct fuse_req *req;
 262         struct fuse_flush_in inarg;
 263         int err;
 264
 265         if (is_bad_inode(inode))
 266                 return -EIO;
 267
 268         if (fc->no_flush)
 269                 return 0;
 270
 271         req = fuse_get_req_nofail(fc, file);
 272         memset(&inarg, 0, sizeof(inarg));
 273         inarg.fh = ff->fh;
 274         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 275         req->in.h.opcode = FUSE_FLUSH;
 276         req->in.h.nodeid = get_node_id(inode);
 277         req->in.numargs = 1;
 278         req->in.args[0].size = sizeof(inarg);
 279         req->in.args[0].value = &inarg;
 280         req->force = 1;
 281         request_send(fc, req);
 282         err = req->out.h.error;
 283         fuse_put_request(fc, req);
 284         if (err == -ENOSYS) {
 285                 fc->no_flush = 1;
 286                 err = 0;
 287         }
 288         return err;
 289 }
 290
 291 /*
 292  * Wait for all pending writepages on the inode to finish.
 293  *
 294  * This is currently done by blocking further writes with FUSE_NOWRITE
 295  * and waiting for all sent writes to complete.
 296  *
 297  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 298  * could conflict with truncation.
 299  */
 300 static void fuse_sync_writes(struct inode *inode)
 301 {
 302         fuse_set_nowrite(inode);
 303         fuse_release_nowrite(inode);
 304 }
 305
 306 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
 307                       int isdir)
 308 {
 309         struct inode *inode = de->d_inode;
 310         struct fuse_conn *fc = get_fuse_conn(inode);
 311         struct fuse_file *ff = file->private_data;
 312         struct fuse_req *req;
 313         struct fuse_fsync_in inarg;
 314         int err;
 315
 316         if (is_bad_inode(inode))
 317                 return -EIO;
 318
 319         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 320                 return 0;
 321
 322         /*
 323          * Start writeback against all dirty pages of the inode, then
 324          * wait for all outstanding writes, before sending the FSYNC
 325          * request.
 326          */
 327         err = write_inode_now(inode, 0);
 328         if (err)
 329                 return err;
 330
 331         fuse_sync_writes(inode);
 332
 333         req = fuse_get_req(fc);
 334         if (IS_ERR(req))
 335                 return PTR_ERR(req);
 336
 337         memset(&inarg, 0, sizeof(inarg));
 338         inarg.fh = ff->fh;
 339         inarg.fsync_flags = datasync ? 1 : 0;
 340         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 341         req->in.h.nodeid = get_node_id(inode);
 342         req->in.numargs = 1;
 343         req->in.args[0].size = sizeof(inarg);
 344         req->in.args[0].value = &inarg;
 345         request_send(fc, req);
 346         err = req->out.h.error;
 347         fuse_put_request(fc, req);
 348         if (err == -ENOSYS) {
 349                 if (isdir)
 350                         fc->no_fsyncdir = 1;
 351                 else
 352                         fc->no_fsync = 1;
 353                 err = 0;
 354         }
 355         return err;
 356 }
 357
 358 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
 359 {
 360         return fuse_fsync_common(file, de, datasync, 0);
 361 }
 362
 363 void fuse_read_fill(struct fuse_req *req, struct file *file,
 364                     struct inode *inode, loff_t pos, size_t count, int opcode)
 365 {
 366         struct fuse_read_in *inarg = &req->misc.read_in;
 367         struct fuse_file *ff = file->private_data;
 368
 369         inarg->fh = ff->fh;
 370         inarg->offset = pos;
 371         inarg->size = count;
 372         inarg->flags = file->f_flags;
 373         req->in.h.opcode = opcode;
 374         req->in.h.nodeid = get_node_id(inode);
 375         req->in.numargs = 1;
 376         req->in.args[0].size = sizeof(struct fuse_read_in);
 377         req->in.args[0].value = inarg;
 378         req->out.argpages = 1;
 379         req->out.argvar = 1;
 380         req->out.numargs = 1;
 381         req->out.args[0].size = count;
 382 }
 383
 384 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 385                              struct inode *inode, loff_t pos, size_t count,
 386                              fl_owner_t owner)
 387 {
 388         struct fuse_conn *fc = get_fuse_conn(inode);
 389
 390         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
 391         if (owner != NULL) {
 392                 struct fuse_read_in *inarg = &req->misc.read_in;
 393
 394                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 395                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 396         }
 397         request_send(fc, req);
 398         return req->out.args[0].size;
 399 }
 400
 401 static int fuse_readpage(struct file *file, struct page *page)
 402 {
 403         struct inode *inode = page->mapping->host;
 404         struct fuse_conn *fc = get_fuse_conn(inode);
 405         struct fuse_req *req;
 406         int err;
 407
 408         err = -EIO;
 409         if (is_bad_inode(inode))
 410                 goto out;
 411
 412         /*
 413          * Page writeback can extend beyond the liftime of the
 414          * page-cache page, so make sure we read a properly synced
 415          * page.
 416          */
 417         fuse_wait_on_page_writeback(inode, page->index);
 418
 419         req = fuse_get_req(fc);
 420         err = PTR_ERR(req);
 421         if (IS_ERR(req))
 422                 goto out;
 423
 424         req->out.page_zeroing = 1;
 425         req->num_pages = 1;
 426         req->pages[0] = page;
 427         fuse_send_read(req, file, inode, page_offset(page), PAGE_CACHE_SIZE,
 428                        NULL);
 429         err = req->out.h.error;
 430         fuse_put_request(fc, req);
 431         if (!err)
 432                 SetPageUptodate(page);
 433         fuse_invalidate_attr(inode); /* atime changed */
 434  out:
 435         unlock_page(page);
 436         return err;
 437 }
 438
 439 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 440 {
 441         int i;
 442
 443         fuse_invalidate_attr(req->pages[0]->mapping->host); /* atime changed */
 444
 445         for (i = 0; i < req->num_pages; i++) {
 446                 struct page *page = req->pages[i];
 447                 if (!req->out.h.error)
 448                         SetPageUptodate(page);
 449                 else
 450                         SetPageError(page);
 451                 unlock_page(page);
 452         }
 453         if (req->ff)
 454                 fuse_file_put(req->ff);
 455         fuse_put_request(fc, req);
 456 }
 457
 458 static void fuse_send_readpages(struct fuse_req *req, struct file *file,
 459                                 struct inode *inode)
 460 {
 461         struct fuse_conn *fc = get_fuse_conn(inode);
 462         loff_t pos = page_offset(req->pages[0]);
 463         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 464         req->out.page_zeroing = 1;
 465         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
 466         if (fc->async_read) {
 467                 struct fuse_file *ff = file->private_data;
 468                 req->ff = fuse_file_get(ff);
 469                 req->end = fuse_readpages_end;
 470                 request_send_background(fc, req);
 471         } else {
 472                 request_send(fc, req);
 473                 fuse_readpages_end(fc, req);
 474         }
 475 }
 476
 477 struct fuse_fill_data {
 478         struct fuse_req *req;
 479         struct file *file;
 480         struct inode *inode;
 481 };
 482
 483 static int fuse_readpages_fill(void *_data, struct page *page)
 484 {
 485         struct fuse_fill_data *data = _data;
 486         struct fuse_req *req = data->req;
 487         struct inode *inode = data->inode;
 488         struct fuse_conn *fc = get_fuse_conn(inode);
 489
 490         fuse_wait_on_page_writeback(inode, page->index);
 491
 492         if (req->num_pages &&
 493             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 494              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 495              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 496                 fuse_send_readpages(req, data->file, inode);
 497                 data->req = req = fuse_get_req(fc);
 498                 if (IS_ERR(req)) {
 499                         unlock_page(page);
 500                         return PTR_ERR(req);
 501                 }
 502         }
 503         req->pages[req->num_pages] = page;
 504         req->num_pages ++;
 505         return 0;
 506 }
 507
 508 static int fuse_readpages(struct file *file, struct address_space *mapping,
 509                           struct list_head *pages, unsigned nr_pages)
 510 {
 511         struct inode *inode = mapping->host;
 512         struct fuse_conn *fc = get_fuse_conn(inode);
 513         struct fuse_fill_data data;
 514         int err;
 515
 516         err = -EIO;
 517         if (is_bad_inode(inode))
 518                 goto out;
 519
 520         data.file = file;
 521         data.inode = inode;
 522         data.req = fuse_get_req(fc);
 523         err = PTR_ERR(data.req);
 524         if (IS_ERR(data.req))
 525                 goto out;
 526
 527         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 528         if (!err) {
 529                 if (data.req->num_pages)
 530                         fuse_send_readpages(data.req, file, inode);
 531                 else
 532                         fuse_put_request(fc, data.req);
 533         }
 534 out:
 535         return err;
 536 }
 537
 538 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 539                                   unsigned long nr_segs, loff_t pos)
 540 {
 541         struct inode *inode = iocb->ki_filp->f_mapping->host;
 542
 543         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 544                 int err;
 545                 /*
 546                  * If trying to read past EOF, make sure the i_size
 547                  * attribute is up-to-date.
 548                  */
 549                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 550                 if (err)
 551                         return err;
 552         }
 553
 554         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 555 }
 556
 557 static void fuse_write_fill(struct fuse_req *req, struct file *file,
 558                             struct fuse_file *ff, struct inode *inode,
 559                             loff_t pos, size_t count, int writepage)
 560 {
 561         struct fuse_conn *fc = get_fuse_conn(inode);
 562         struct fuse_write_in *inarg = &req->misc.write.in;
 563         struct fuse_write_out *outarg = &req->misc.write.out;
 564
 565         memset(inarg, 0, sizeof(struct fuse_write_in));
 566         inarg->fh = ff->fh;
 567         inarg->offset = pos;
 568         inarg->size = count;
 569         inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
 570         inarg->flags = file ? file->f_flags : 0;
 571         req->in.h.opcode = FUSE_WRITE;
 572         req->in.h.nodeid = get_node_id(inode);
 573         req->in.argpages = 1;
 574         req->in.numargs = 2;
 575         if (fc->minor < 9)
 576                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 577         else
 578                 req->in.args[0].size = sizeof(struct fuse_write_in);
 579         req->in.args[0].value = inarg;
 580         req->in.args[1].size = count;
 581         req->out.numargs = 1;
 582         req->out.args[0].size = sizeof(struct fuse_write_out);
 583         req->out.args[0].value = outarg;
 584 }
 585
 586 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 587                               struct inode *inode, loff_t pos, size_t count,
 588                               fl_owner_t owner)
 589 {
 590         struct fuse_conn *fc = get_fuse_conn(inode);
 591         fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
 592         if (owner != NULL) {
 593                 struct fuse_write_in *inarg = &req->misc.write.in;
 594                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 595                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 596         }
 597         request_send(fc, req);
 598         return req->misc.write.out.size;
 599 }
 600
 601 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 602                         loff_t pos, unsigned len, unsigned flags,
 603                         struct page **pagep, void **fsdata)
 604 {
 605         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 606
 607         *pagep = __grab_cache_page(mapping, index);
 608         if (!*pagep)
 609                 return -ENOMEM;
 610         return 0;
 611 }
 612
 613 static int fuse_buffered_write(struct file *file, struct inode *inode,
 614                                loff_t pos, unsigned count, struct page *page)
 615 {
 616         int err;
 617         size_t nres;
 618         struct fuse_conn *fc = get_fuse_conn(inode);
 619         struct fuse_inode *fi = get_fuse_inode(inode);
 620         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 621         struct fuse_req *req;
 622
 623         if (is_bad_inode(inode))
 624                 return -EIO;
 625
 626         /*
 627          * Make sure writepages on the same page are not mixed up with
 628          * plain writes.
 629          */
 630         fuse_wait_on_page_writeback(inode, page->index);
 631
 632         req = fuse_get_req(fc);
 633         if (IS_ERR(req))
 634                 return PTR_ERR(req);
 635
 636         req->num_pages = 1;
 637         req->pages[0] = page;
 638         req->page_offset = offset;
 639         nres = fuse_send_write(req, file, inode, pos, count, NULL);
 640         err = req->out.h.error;
 641         fuse_put_request(fc, req);
 642         if (!err && !nres)
 643                 err = -EIO;
 644         if (!err) {
 645                 pos += nres;
 646                 spin_lock(&fc->lock);
 647                 fi->attr_version = ++fc->attr_version;
 648                 if (pos > inode->i_size)
 649                         i_size_write(inode, pos);
 650                 spin_unlock(&fc->lock);
 651
 652                 if (count == PAGE_CACHE_SIZE)
 653                         SetPageUptodate(page);
 654         }
 655         fuse_invalidate_attr(inode);
 656         return err ? err : nres;
 657 }
 658
 659 static int fuse_write_end(struct file *file, struct address_space *mapping,
 660                         loff_t pos, unsigned len, unsigned copied,
 661                         struct page *page, void *fsdata)
 662 {
 663         struct inode *inode = mapping->host;
 664         int res = 0;
 665
 666         if (copied)
 667                 res = fuse_buffered_write(file, inode, pos, copied, page);
 668
 669         unlock_page(page);
 670         page_cache_release(page);
 671         return res;
 672 }
 673
 674 static void fuse_release_user_pages(struct fuse_req *req, int write)
 675 {
 676         unsigned i;
 677
 678         for (i = 0; i < req->num_pages; i++) {
 679                 struct page *page = req->pages[i];
 680                 if (write)
 681                         set_page_dirty_lock(page);
 682                 put_page(page);
 683         }
 684 }
 685
 686 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 687                                unsigned nbytes, int write)
 688 {
 689         unsigned long user_addr = (unsigned long) buf;
 690         unsigned offset = user_addr & ~PAGE_MASK;
 691         int npages;
 692
 693         /* This doesn't work with nfsd */
 694         if (!current->mm)
 695                 return -EPERM;
 696
 697         nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
 698         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 699         npages = min(max(npages, 1), FUSE_MAX_PAGES_PER_REQ);
 700         down_read(&current->mm->mmap_sem);
 701         npages = get_user_pages(current, current->mm, user_addr, npages, write,
 702                                 0, req->pages, NULL);
 703         up_read(&current->mm->mmap_sem);
 704         if (npages < 0)
 705                 return npages;
 706
 707         req->num_pages = npages;
 708         req->page_offset = offset;
 709         return 0;
 710 }
 711
 712 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
 713                               size_t count, loff_t *ppos, int write)
 714 {
 715         struct inode *inode = file->f_path.dentry->d_inode;
 716         struct fuse_conn *fc = get_fuse_conn(inode);
 717         size_t nmax = write ? fc->max_write : fc->max_read;
 718         loff_t pos = *ppos;
 719         ssize_t res = 0;
 720         struct fuse_req *req;
 721
 722         if (is_bad_inode(inode))
 723                 return -EIO;
 724
 725         req = fuse_get_req(fc);
 726         if (IS_ERR(req))
 727                 return PTR_ERR(req);
 728
 729         while (count) {
 730                 size_t nres;
 731                 size_t nbytes = min(count, nmax);
 732                 int err = fuse_get_user_pages(req, buf, nbytes, !write);
 733                 if (err) {
 734                         res = err;
 735                         break;
 736                 }
 737                 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
 738                 nbytes = min(count, nbytes);
 739                 if (write)
 740                         nres = fuse_send_write(req, file, inode, pos, nbytes,
 741                                                current->files);
 742                 else
 743                         nres = fuse_send_read(req, file, inode, pos, nbytes,
 744                                               current->files);
 745                 fuse_release_user_pages(req, !write);
 746                 if (req->out.h.error) {
 747                         if (!res)
 748                                 res = req->out.h.error;
 749                         break;
 750                 } else if (nres > nbytes) {
 751                         res = -EIO;
 752                         break;
 753                 }
 754                 count -= nres;
 755                 res += nres;
 756                 pos += nres;
 757                 buf += nres;
 758                 if (nres != nbytes)
 759                         break;
 760                 if (count) {
 761                         fuse_put_request(fc, req);
 762                         req = fuse_get_req(fc);
 763                         if (IS_ERR(req))
 764                                 break;
 765                 }
 766         }
 767         fuse_put_request(fc, req);
 768         if (res > 0) {
 769                 if (write) {
 770                         spin_lock(&fc->lock);
 771                         if (pos > inode->i_size)
 772                                 i_size_write(inode, pos);
 773                         spin_unlock(&fc->lock);
 774                 }
 775                 *ppos = pos;
 776         }
 777         fuse_invalidate_attr(inode);
 778
 779         return res;
 780 }
 781
 782 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
 783                                      size_t count, loff_t *ppos)
 784 {
 785         return fuse_direct_io(file, buf, count, ppos, 0);
 786 }
 787
 788 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
 789                                  size_t count, loff_t *ppos)
 790 {
 791         struct inode *inode = file->f_path.dentry->d_inode;
 792         ssize_t res;
 793         /* Don't allow parallel writes to the same file */
 794         mutex_lock(&inode->i_mutex);
 795         res = generic_write_checks(file, ppos, &count, 0);
 796         if (!res)
 797                 res = fuse_direct_io(file, buf, count, ppos, 1);
 798         mutex_unlock(&inode->i_mutex);
 799         return res;
 800 }
 801
 802 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
 803 {
 804         __free_page(req->pages[0]);
 805         fuse_file_put(req->ff);
 806         fuse_put_request(fc, req);
 807 }
 808
 809 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
 810 {
 811         struct inode *inode = req->inode;
 812         struct fuse_inode *fi = get_fuse_inode(inode);
 813         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
 814
 815         list_del(&req->writepages_entry);
 816         dec_bdi_stat(bdi, BDI_WRITEBACK);
 817         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
 818         bdi_writeout_inc(bdi);
 819         wake_up(&fi->page_waitq);
 820 }
 821
 822 /* Called under fc->lock, may release and reacquire it */
 823 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
 824 {
 825         struct fuse_inode *fi = get_fuse_inode(req->inode);
 826         loff_t size = i_size_read(req->inode);
 827         struct fuse_write_in *inarg = &req->misc.write.in;
 828
 829         if (!fc->connected)
 830                 goto out_free;
 831
 832         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
 833                 inarg->size = PAGE_CACHE_SIZE;
 834         } else if (inarg->offset < size) {
 835                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
 836         } else {
 837                 /* Got truncated off completely */
 838                 goto out_free;
 839         }
 840
 841         req->in.args[1].size = inarg->size;
 842         fi->writectr++;
 843         request_send_background_locked(fc, req);
 844         return;
 845
 846  out_free:
 847         fuse_writepage_finish(fc, req);
 848         spin_unlock(&fc->lock);
 849         fuse_writepage_free(fc, req);
 850         spin_lock(&fc->lock);
 851 }
 852
 853 /*
 854  * If fi->writectr is positive (no truncate or fsync going on) send
 855  * all queued writepage requests.
 856  *
 857  * Called with fc->lock
 858  */
 859 void fuse_flush_writepages(struct inode *inode)
 860 {
 861         struct fuse_conn *fc = get_fuse_conn(inode);
 862         struct fuse_inode *fi = get_fuse_inode(inode);
 863         struct fuse_req *req;
 864
 865         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
 866                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
 867                 list_del_init(&req->list);
 868                 fuse_send_writepage(fc, req);
 869         }
 870 }
 871
 872 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
 873 {
 874         struct inode *inode = req->inode;
 875         struct fuse_inode *fi = get_fuse_inode(inode);
 876
 877         mapping_set_error(inode->i_mapping, req->out.h.error);
 878         spin_lock(&fc->lock);
 879         fi->writectr--;
 880         fuse_writepage_finish(fc, req);
 881         spin_unlock(&fc->lock);
 882         fuse_writepage_free(fc, req);
 883 }
 884
 885 static int fuse_writepage_locked(struct page *page)
 886 {
 887         struct address_space *mapping = page->mapping;
 888         struct inode *inode = mapping->host;
 889         struct fuse_conn *fc = get_fuse_conn(inode);
 890         struct fuse_inode *fi = get_fuse_inode(inode);
 891         struct fuse_req *req;
 892         struct fuse_file *ff;
 893         struct page *tmp_page;
 894
 895         set_page_writeback(page);
 896
 897         req = fuse_request_alloc_nofs();
 898         if (!req)
 899                 goto err;
 900
 901         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
 902         if (!tmp_page)
 903                 goto err_free;
 904
 905         spin_lock(&fc->lock);
 906         BUG_ON(list_empty(&fi->write_files));
 907         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
 908         req->ff = fuse_file_get(ff);
 909         spin_unlock(&fc->lock);
 910
 911         fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
 912
 913         copy_highpage(tmp_page, page);
 914         req->num_pages = 1;
 915         req->pages[0] = tmp_page;
 916         req->page_offset = 0;
 917         req->end = fuse_writepage_end;
 918         req->inode = inode;
 919
 920         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
 921         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
 922         end_page_writeback(page);
 923
 924         spin_lock(&fc->lock);
 925         list_add(&req->writepages_entry, &fi->writepages);
 926         list_add_tail(&req->list, &fi->queued_writes);
 927         fuse_flush_writepages(inode);
 928         spin_unlock(&fc->lock);
 929
 930         return 0;
 931
 932 err_free:
 933         fuse_request_free(req);
 934 err:
 935         end_page_writeback(page);
 936         return -ENOMEM;
 937 }
 938
 939 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
 940 {
 941         int err;
 942
 943         err = fuse_writepage_locked(page);
 944         unlock_page(page);
 945
 946         return err;
 947 }
 948
 949 static int fuse_launder_page(struct page *page)
 950 {
 951         int err = 0;
 952         if (clear_page_dirty_for_io(page)) {
 953                 struct inode *inode = page->mapping->host;
 954                 err = fuse_writepage_locked(page);
 955                 if (!err)
 956                         fuse_wait_on_page_writeback(inode, page->index);
 957         }
 958         return err;
 959 }
 960
 961 /*
 962  * Write back dirty pages now, because there may not be any suitable
 963  * open files later
 964  */
 965 static void fuse_vma_close(struct vm_area_struct *vma)
 966 {
 967         filemap_write_and_wait(vma->vm_file->f_mapping);
 968 }
 969
 970 /*
 971  * Wait for writeback against this page to complete before allowing it
 972  * to be marked dirty again, and hence written back again, possibly
 973  * before the previous writepage completed.
 974  *
 975  * Block here, instead of in ->writepage(), so that the userspace fs
 976  * can only block processes actually operating on the filesystem.
 977  *
 978  * Otherwise unprivileged userspace fs would be able to block
 979  * unrelated:
 980  *
 981  * - page migration
 982  * - sync(2)
 983  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
 984  */
 985 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct page *page)
 986 {
 987         /*
 988          * Don't use page->mapping as it may become NULL from a
 989          * concurrent truncate.
 990          */
 991         struct inode *inode = vma->vm_file->f_mapping->host;
 992
 993         fuse_wait_on_page_writeback(inode, page->index);
 994         return 0;
 995 }
 996
 997 static struct vm_operations_struct fuse_file_vm_ops = {
 998         .close          = fuse_vma_close,
 999         .fault          = filemap_fault,
1000         .page_mkwrite   = fuse_page_mkwrite,
1001 };
1002
1003 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1004 {
1005         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1006                 struct inode *inode = file->f_dentry->d_inode;
1007                 struct fuse_conn *fc = get_fuse_conn(inode);
1008                 struct fuse_inode *fi = get_fuse_inode(inode);
1009                 struct fuse_file *ff = file->private_data;
1010                 /*
1011                  * file may be written through mmap, so chain it onto the
1012                  * inodes's write_file list
1013                  */
1014                 spin_lock(&fc->lock);
1015                 if (list_empty(&ff->write_entry))
1016                         list_add(&ff->write_entry, &fi->write_files);
1017                 spin_unlock(&fc->lock);
1018         }
1019         file_accessed(file);
1020         vma->vm_ops = &fuse_file_vm_ops;
1021         return 0;
1022 }
1023
1024 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1025                                   struct file_lock *fl)
1026 {
1027         switch (ffl->type) {
1028         case F_UNLCK:
1029                 break;
1030
1031         case F_RDLCK:
1032         case F_WRLCK:
1033                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1034                     ffl->end < ffl->start)
1035                         return -EIO;
1036
1037                 fl->fl_start = ffl->start;
1038                 fl->fl_end = ffl->end;
1039                 fl->fl_pid = ffl->pid;
1040                 break;
1041
1042         default:
1043                 return -EIO;
1044         }
1045         fl->fl_type = ffl->type;
1046         return 0;
1047 }
1048
1049 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1050                          const struct file_lock *fl, int opcode, pid_t pid,
1051                          int flock)
1052 {
1053         struct inode *inode = file->f_path.dentry->d_inode;
1054         struct fuse_conn *fc = get_fuse_conn(inode);
1055         struct fuse_file *ff = file->private_data;
1056         struct fuse_lk_in *arg = &req->misc.lk_in;
1057
1058         arg->fh = ff->fh;
1059         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1060         arg->lk.start = fl->fl_start;
1061         arg->lk.end = fl->fl_end;
1062         arg->lk.type = fl->fl_type;
1063         arg->lk.pid = pid;
1064         if (flock)
1065                 arg->lk_flags |= FUSE_LK_FLOCK;
1066         req->in.h.opcode = opcode;
1067         req->in.h.nodeid = get_node_id(inode);
1068         req->in.numargs = 1;
1069         req->in.args[0].size = sizeof(*arg);
1070         req->in.args[0].value = arg;
1071 }
1072
1073 static int fuse_getlk(struct file *file, struct file_lock *fl)
1074 {
1075         struct inode *inode = file->f_path.dentry->d_inode;
1076         struct fuse_conn *fc = get_fuse_conn(inode);
1077         struct fuse_req *req;
1078         struct fuse_lk_out outarg;
1079         int err;
1080
1081         req = fuse_get_req(fc);
1082         if (IS_ERR(req))
1083                 return PTR_ERR(req);
1084
1085         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1086         req->out.numargs = 1;
1087         req->out.args[0].size = sizeof(outarg);
1088         req->out.args[0].value = &outarg;
1089         request_send(fc, req);
1090         err = req->out.h.error;
1091         fuse_put_request(fc, req);
1092         if (!err)
1093                 err = convert_fuse_file_lock(&outarg.lk, fl);
1094
1095         return err;
1096 }
1097
1098 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1099 {
1100         struct inode *inode = file->f_path.dentry->d_inode;
1101         struct fuse_conn *fc = get_fuse_conn(inode);
1102         struct fuse_req *req;
1103         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1104         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1105         int err;
1106
1107         /* Unlock on close is handled by the flush method */
1108         if (fl->fl_flags & FL_CLOSE)
1109                 return 0;
1110
1111         req = fuse_get_req(fc);
1112         if (IS_ERR(req))
1113                 return PTR_ERR(req);
1114
1115         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1116         request_send(fc, req);
1117         err = req->out.h.error;
1118         /* locking is restartable */
1119         if (err == -EINTR)
1120                 err = -ERESTARTSYS;
1121         fuse_put_request(fc, req);
1122         return err;
1123 }
1124
1125 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1126 {
1127         struct inode *inode = file->f_path.dentry->d_inode;
1128         struct fuse_conn *fc = get_fuse_conn(inode);
1129         int err;
1130
1131         if (cmd == F_GETLK) {
1132                 if (fc->no_lock) {
1133                         posix_test_lock(file, fl);
1134                         err = 0;
1135                 } else
1136                         err = fuse_getlk(file, fl);
1137         } else {
1138                 if (fc->no_lock)
1139                         err = posix_lock_file_wait(file, fl);
1140                 else
1141                         err = fuse_setlk(file, fl, 0);
1142         }
1143         return err;
1144 }
1145
1146 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1147 {
1148         struct inode *inode = file->f_path.dentry->d_inode;
1149         struct fuse_conn *fc = get_fuse_conn(inode);
1150         int err;
1151
1152         if (fc->no_lock) {
1153                 err = flock_lock_file_wait(file, fl);
1154         } else {
1155                 /* emulate flock with POSIX locks */
1156                 fl->fl_owner = (fl_owner_t) file;
1157                 err = fuse_setlk(file, fl, 1);
1158         }
1159
1160         return err;
1161 }
1162
1163 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1164 {
1165         struct inode *inode = mapping->host;
1166         struct fuse_conn *fc = get_fuse_conn(inode);
1167         struct fuse_req *req;
1168         struct fuse_bmap_in inarg;
1169         struct fuse_bmap_out outarg;
1170         int err;
1171
1172         if (!inode->i_sb->s_bdev || fc->no_bmap)
1173                 return 0;
1174
1175         req = fuse_get_req(fc);
1176         if (IS_ERR(req))
1177                 return 0;
1178
1179         memset(&inarg, 0, sizeof(inarg));
1180         inarg.block = block;
1181         inarg.blocksize = inode->i_sb->s_blocksize;
1182         req->in.h.opcode = FUSE_BMAP;
1183         req->in.h.nodeid = get_node_id(inode);
1184         req->in.numargs = 1;
1185         req->in.args[0].size = sizeof(inarg);
1186         req->in.args[0].value = &inarg;
1187         req->out.numargs = 1;
1188         req->out.args[0].size = sizeof(outarg);
1189         req->out.args[0].value = &outarg;
1190         request_send(fc, req);
1191         err = req->out.h.error;
1192         fuse_put_request(fc, req);
1193         if (err == -ENOSYS)
1194                 fc->no_bmap = 1;
1195
1196         return err ? 0 : outarg.block;
1197 }
1198
1199 static const struct file_operations fuse_file_operations = {
1200         .llseek         = generic_file_llseek,
1201         .read           = do_sync_read,
1202         .aio_read       = fuse_file_aio_read,
1203         .write          = do_sync_write,
1204         .aio_write      = generic_file_aio_write,
1205         .mmap           = fuse_file_mmap,
1206         .open           = fuse_open,
1207         .flush          = fuse_flush,
1208         .release        = fuse_release,
1209         .fsync          = fuse_fsync,
1210         .lock           = fuse_file_lock,
1211         .flock          = fuse_file_flock,
1212         .splice_read    = generic_file_splice_read,
1213 };
1214
1215 static const struct file_operations fuse_direct_io_file_operations = {
1216         .llseek         = generic_file_llseek,
1217         .read           = fuse_direct_read,
1218         .write          = fuse_direct_write,
1219         .open           = fuse_open,
1220         .flush          = fuse_flush,
1221         .release        = fuse_release,
1222         .fsync          = fuse_fsync,
1223         .lock           = fuse_file_lock,
1224         .flock          = fuse_file_flock,
1225         /* no mmap and splice_read */
1226 };
1227
1228 static const struct address_space_operations fuse_file_aops  = {
1229         .readpage       = fuse_readpage,
1230         .writepage      = fuse_writepage,
1231         .launder_page   = fuse_launder_page,
1232         .write_begin    = fuse_write_begin,
1233         .write_end      = fuse_write_end,
1234         .readpages      = fuse_readpages,
1235         .set_page_dirty = __set_page_dirty_nobuffers,
1236         .bmap           = fuse_bmap,
1237 };
1238
1239 void fuse_init_file_inode(struct inode *inode)
1240 {
1241         inode->i_fop = &fuse_file_operations;
1242         inode->i_data.a_ops = &fuse_file_aops;
1243 }