fs/fuse/file.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15
  16 static const struct file_operations fuse_direct_io_file_operations;
  17
  18 static int fuse_send_open(struct inode *inode, struct file *file, int isdir,
  19                           struct fuse_open_out *outargp)
  20 {
  21         struct fuse_conn *fc = get_fuse_conn(inode);
  22         struct fuse_open_in inarg;
  23         struct fuse_req *req;
  24         int err;
  25
  26         req = fuse_get_req(fc);
  27         if (IS_ERR(req))
  28                 return PTR_ERR(req);
  29
  30         memset(&inarg, 0, sizeof(inarg));
  31         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  32         if (!fc->atomic_o_trunc)
  33                 inarg.flags &= ~O_TRUNC;
  34         req->in.h.opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
  35         req->in.h.nodeid = get_node_id(inode);
  36         req->in.numargs = 1;
  37         req->in.args[0].size = sizeof(inarg);
  38         req->in.args[0].value = &inarg;
  39         req->out.numargs = 1;
  40         req->out.args[0].size = sizeof(*outargp);
  41         req->out.args[0].value = outargp;
  42         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 }
  83
  84 static void fuse_file_put(struct fuse_file *ff)
  85 {
  86         if (atomic_dec_and_test(&ff->count)) {
  87                 struct fuse_req *req = ff->reserved_req;
  88                 struct inode *inode = req->misc.release.dentry->d_inode;
  89                 struct fuse_conn *fc = get_fuse_conn(inode);
  90                 req->end = fuse_release_end;
  91                 request_send_background(fc, req);
  92                 kfree(ff);
  93         }
  94 }
  95
  96 void fuse_finish_open(struct inode *inode, struct file *file,
  97                       struct fuse_file *ff, struct fuse_open_out *outarg)
  98 {
  99         if (outarg->open_flags & FOPEN_DIRECT_IO)
 100                 file->f_op = &fuse_direct_io_file_operations;
 101         if (!(outarg->open_flags & FOPEN_KEEP_CACHE))
 102                 invalidate_inode_pages2(inode->i_mapping);
 103         if (outarg->open_flags & FOPEN_NONSEEKABLE)
 104                 nonseekable_open(inode, file);
 105         ff->fh = outarg->fh;
 106         file->private_data = fuse_file_get(ff);
 107 }
 108
 109 int fuse_open_common(struct inode *inode, struct file *file, int isdir)
 110 {
 111         struct fuse_open_out outarg;
 112         struct fuse_file *ff;
 113         int err;
 114
 115         /* VFS checks this, but only _after_ ->open() */
 116         if (file->f_flags & O_DIRECT)
 117                 return -EINVAL;
 118
 119         err = generic_file_open(inode, file);
 120         if (err)
 121                 return err;
 122
 123         ff = fuse_file_alloc();
 124         if (!ff)
 125                 return -ENOMEM;
 126
 127         err = fuse_send_open(inode, file, isdir, &outarg);
 128         if (err)
 129                 fuse_file_free(ff);
 130         else {
 131                 if (isdir)
 132                         outarg.open_flags &= ~FOPEN_DIRECT_IO;
 133                 fuse_finish_open(inode, file, ff, &outarg);
 134         }
 135
 136         return err;
 137 }
 138
 139 void fuse_release_fill(struct fuse_file *ff, u64 nodeid, int flags, int opcode)
 140 {
 141         struct fuse_req *req = ff->reserved_req;
 142         struct fuse_release_in *inarg = &req->misc.release.in;
 143
 144         inarg->fh = ff->fh;
 145         inarg->flags = flags;
 146         req->in.h.opcode = opcode;
 147         req->in.h.nodeid = nodeid;
 148         req->in.numargs = 1;
 149         req->in.args[0].size = sizeof(struct fuse_release_in);
 150         req->in.args[0].value = inarg;
 151 }
 152
 153 int fuse_release_common(struct inode *inode, struct file *file, int isdir)
 154 {
 155         struct fuse_file *ff = file->private_data;
 156         if (ff) {
 157                 struct fuse_conn *fc = get_fuse_conn(inode);
 158                 struct fuse_req *req = ff->reserved_req;
 159
 160                 fuse_release_fill(ff, get_node_id(inode), file->f_flags,
 161                                   isdir ? FUSE_RELEASEDIR : FUSE_RELEASE);
 162
 163                 /* Hold vfsmount and dentry until release is finished */
 164                 req->misc.release.vfsmount = mntget(file->f_path.mnt);
 165                 req->misc.release.dentry = dget(file->f_path.dentry);
 166
 167                 spin_lock(&fc->lock);
 168                 list_del(&ff->write_entry);
 169                 spin_unlock(&fc->lock);
 170                 /*
 171                  * Normally this will send the RELEASE request,
 172                  * however if some asynchronous READ or WRITE requests
 173                  * are outstanding, the sending will be delayed
 174                  */
 175                 fuse_file_put(ff);
 176         }
 177
 178         /* Return value is ignored by VFS */
 179         return 0;
 180 }
 181
 182 static int fuse_open(struct inode *inode, struct file *file)
 183 {
 184         return fuse_open_common(inode, file, 0);
 185 }
 186
 187 static int fuse_release(struct inode *inode, struct file *file)
 188 {
 189         return fuse_release_common(inode, file, 0);
 190 }
 191
 192 /*
 193  * Scramble the ID space with XTEA, so that the value of the files_struct
 194  * pointer is not exposed to userspace.
 195  */
 196 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 197 {
 198         u32 *k = fc->scramble_key;
 199         u64 v = (unsigned long) id;
 200         u32 v0 = v;
 201         u32 v1 = v >> 32;
 202         u32 sum = 0;
 203         int i;
 204
 205         for (i = 0; i < 32; i++) {
 206                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 207                 sum += 0x9E3779B9;
 208                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 209         }
 210
 211         return (u64) v0 + ((u64) v1 << 32);
 212 }
 213
 214 /*
 215  * Check if page is under writeback
 216  *
 217  * This is currently done by walking the list of writepage requests
 218  * for the inode, which can be pretty inefficient.
 219  */
 220 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 221 {
 222         struct fuse_conn *fc = get_fuse_conn(inode);
 223         struct fuse_inode *fi = get_fuse_inode(inode);
 224         struct fuse_req *req;
 225         bool found = false;
 226
 227         spin_lock(&fc->lock);
 228         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 229                 pgoff_t curr_index;
 230
 231                 BUG_ON(req->inode != inode);
 232                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 233                 if (curr_index == index) {
 234                         found = true;
 235                         break;
 236                 }
 237         }
 238         spin_unlock(&fc->lock);
 239
 240         return found;
 241 }
 242
 243 /*
 244  * Wait for page writeback to be completed.
 245  *
 246  * Since fuse doesn't rely on the VM writeback tracking, this has to
 247  * use some other means.
 248  */
 249 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 250 {
 251         struct fuse_inode *fi = get_fuse_inode(inode);
 252
 253         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 254         return 0;
 255 }
 256
 257 static int fuse_flush(struct file *file, fl_owner_t id)
 258 {
 259         struct inode *inode = file->f_path.dentry->d_inode;
 260         struct fuse_conn *fc = get_fuse_conn(inode);
 261         struct fuse_file *ff = file->private_data;
 262         struct fuse_req *req;
 263         struct fuse_flush_in inarg;
 264         int err;
 265
 266         if (is_bad_inode(inode))
 267                 return -EIO;
 268
 269         if (fc->no_flush)
 270                 return 0;
 271
 272         req = fuse_get_req_nofail(fc, file);
 273         memset(&inarg, 0, sizeof(inarg));
 274         inarg.fh = ff->fh;
 275         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 276         req->in.h.opcode = FUSE_FLUSH;
 277         req->in.h.nodeid = get_node_id(inode);
 278         req->in.numargs = 1;
 279         req->in.args[0].size = sizeof(inarg);
 280         req->in.args[0].value = &inarg;
 281         req->force = 1;
 282         request_send(fc, req);
 283         err = req->out.h.error;
 284         fuse_put_request(fc, req);
 285         if (err == -ENOSYS) {
 286                 fc->no_flush = 1;
 287                 err = 0;
 288         }
 289         return err;
 290 }
 291
 292 /*
 293  * Wait for all pending writepages on the inode to finish.
 294  *
 295  * This is currently done by blocking further writes with FUSE_NOWRITE
 296  * and waiting for all sent writes to complete.
 297  *
 298  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 299  * could conflict with truncation.
 300  */
 301 static void fuse_sync_writes(struct inode *inode)
 302 {
 303         fuse_set_nowrite(inode);
 304         fuse_release_nowrite(inode);
 305 }
 306
 307 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
 308                       int isdir)
 309 {
 310         struct inode *inode = de->d_inode;
 311         struct fuse_conn *fc = get_fuse_conn(inode);
 312         struct fuse_file *ff = file->private_data;
 313         struct fuse_req *req;
 314         struct fuse_fsync_in inarg;
 315         int err;
 316
 317         if (is_bad_inode(inode))
 318                 return -EIO;
 319
 320         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 321                 return 0;
 322
 323         /*
 324          * Start writeback against all dirty pages of the inode, then
 325          * wait for all outstanding writes, before sending the FSYNC
 326          * request.
 327          */
 328         err = write_inode_now(inode, 0);
 329         if (err)
 330                 return err;
 331
 332         fuse_sync_writes(inode);
 333
 334         req = fuse_get_req(fc);
 335         if (IS_ERR(req))
 336                 return PTR_ERR(req);
 337
 338         memset(&inarg, 0, sizeof(inarg));
 339         inarg.fh = ff->fh;
 340         inarg.fsync_flags = datasync ? 1 : 0;
 341         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 342         req->in.h.nodeid = get_node_id(inode);
 343         req->in.numargs = 1;
 344         req->in.args[0].size = sizeof(inarg);
 345         req->in.args[0].value = &inarg;
 346         request_send(fc, req);
 347         err = req->out.h.error;
 348         fuse_put_request(fc, req);
 349         if (err == -ENOSYS) {
 350                 if (isdir)
 351                         fc->no_fsyncdir = 1;
 352                 else
 353                         fc->no_fsync = 1;
 354                 err = 0;
 355         }
 356         return err;
 357 }
 358
 359 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
 360 {
 361         return fuse_fsync_common(file, de, datasync, 0);
 362 }
 363
 364 void fuse_read_fill(struct fuse_req *req, struct file *file,
 365                     struct inode *inode, loff_t pos, size_t count, int opcode)
 366 {
 367         struct fuse_read_in *inarg = &req->misc.read.in;
 368         struct fuse_file *ff = file->private_data;
 369
 370         inarg->fh = ff->fh;
 371         inarg->offset = pos;
 372         inarg->size = count;
 373         inarg->flags = file->f_flags;
 374         req->in.h.opcode = opcode;
 375         req->in.h.nodeid = get_node_id(inode);
 376         req->in.numargs = 1;
 377         req->in.args[0].size = sizeof(struct fuse_read_in);
 378         req->in.args[0].value = inarg;
 379         req->out.argpages = 1;
 380         req->out.argvar = 1;
 381         req->out.numargs = 1;
 382         req->out.args[0].size = count;
 383 }
 384
 385 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 386                              struct inode *inode, loff_t pos, size_t count,
 387                              fl_owner_t owner)
 388 {
 389         struct fuse_conn *fc = get_fuse_conn(inode);
 390
 391         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
 392         if (owner != NULL) {
 393                 struct fuse_read_in *inarg = &req->misc.read.in;
 394
 395                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 396                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 397         }
 398         request_send(fc, req);
 399         return req->out.args[0].size;
 400 }
 401
 402 static void fuse_read_update_size(struct inode *inode, loff_t size,
 403                                   u64 attr_ver)
 404 {
 405         struct fuse_conn *fc = get_fuse_conn(inode);
 406         struct fuse_inode *fi = get_fuse_inode(inode);
 407
 408         spin_lock(&fc->lock);
 409         if (attr_ver == fi->attr_version && size < inode->i_size) {
 410                 fi->attr_version = ++fc->attr_version;
 411                 i_size_write(inode, size);
 412         }
 413         spin_unlock(&fc->lock);
 414 }
 415
 416 static int fuse_readpage(struct file *file, struct page *page)
 417 {
 418         struct inode *inode = page->mapping->host;
 419         struct fuse_conn *fc = get_fuse_conn(inode);
 420         struct fuse_req *req;
 421         size_t num_read;
 422         loff_t pos = page_offset(page);
 423         size_t count = PAGE_CACHE_SIZE;
 424         u64 attr_ver;
 425         int err;
 426
 427         err = -EIO;
 428         if (is_bad_inode(inode))
 429                 goto out;
 430
 431         /*
 432          * Page writeback can extend beyond the liftime of the
 433          * page-cache page, so make sure we read a properly synced
 434          * page.
 435          */
 436         fuse_wait_on_page_writeback(inode, page->index);
 437
 438         req = fuse_get_req(fc);
 439         err = PTR_ERR(req);
 440         if (IS_ERR(req))
 441                 goto out;
 442
 443         attr_ver = fuse_get_attr_version(fc);
 444
 445         req->out.page_zeroing = 1;
 446         req->num_pages = 1;
 447         req->pages[0] = page;
 448         num_read = fuse_send_read(req, file, inode, pos, count, NULL);
 449         err = req->out.h.error;
 450         fuse_put_request(fc, req);
 451
 452         if (!err) {
 453                 /*
 454                  * Short read means EOF.  If file size is larger, truncate it
 455                  */
 456                 if (num_read < count)
 457                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 458
 459                 SetPageUptodate(page);
 460         }
 461
 462         fuse_invalidate_attr(inode); /* atime changed */
 463  out:
 464         unlock_page(page);
 465         return err;
 466 }
 467
 468 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 469 {
 470         int i;
 471         size_t count = req->misc.read.in.size;
 472         size_t num_read = req->out.args[0].size;
 473         struct inode *inode = req->pages[0]->mapping->host;
 474
 475         /*
 476          * Short read means EOF.  If file size is larger, truncate it
 477          */
 478         if (!req->out.h.error && num_read < count) {
 479                 loff_t pos = page_offset(req->pages[0]) + num_read;
 480                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
 481         }
 482
 483         fuse_invalidate_attr(inode); /* atime changed */
 484
 485         for (i = 0; i < req->num_pages; i++) {
 486                 struct page *page = req->pages[i];
 487                 if (!req->out.h.error)
 488                         SetPageUptodate(page);
 489                 else
 490                         SetPageError(page);
 491                 unlock_page(page);
 492         }
 493         if (req->ff)
 494                 fuse_file_put(req->ff);
 495 }
 496
 497 static void fuse_send_readpages(struct fuse_req *req, struct file *file,
 498                                 struct inode *inode)
 499 {
 500         struct fuse_conn *fc = get_fuse_conn(inode);
 501         loff_t pos = page_offset(req->pages[0]);
 502         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 503         req->out.page_zeroing = 1;
 504         fuse_read_fill(req, file, inode, pos, count, FUSE_READ);
 505         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 506         if (fc->async_read) {
 507                 struct fuse_file *ff = file->private_data;
 508                 req->ff = fuse_file_get(ff);
 509                 req->end = fuse_readpages_end;
 510                 request_send_background(fc, req);
 511         } else {
 512                 request_send(fc, req);
 513                 fuse_readpages_end(fc, req);
 514                 fuse_put_request(fc, req);
 515         }
 516 }
 517
 518 struct fuse_fill_data {
 519         struct fuse_req *req;
 520         struct file *file;
 521         struct inode *inode;
 522 };
 523
 524 static int fuse_readpages_fill(void *_data, struct page *page)
 525 {
 526         struct fuse_fill_data *data = _data;
 527         struct fuse_req *req = data->req;
 528         struct inode *inode = data->inode;
 529         struct fuse_conn *fc = get_fuse_conn(inode);
 530
 531         fuse_wait_on_page_writeback(inode, page->index);
 532
 533         if (req->num_pages &&
 534             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 535              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 536              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 537                 fuse_send_readpages(req, data->file, inode);
 538                 data->req = req = fuse_get_req(fc);
 539                 if (IS_ERR(req)) {
 540                         unlock_page(page);
 541                         return PTR_ERR(req);
 542                 }
 543         }
 544         req->pages[req->num_pages] = page;
 545         req->num_pages++;
 546         return 0;
 547 }
 548
 549 static int fuse_readpages(struct file *file, struct address_space *mapping,
 550                           struct list_head *pages, unsigned nr_pages)
 551 {
 552         struct inode *inode = mapping->host;
 553         struct fuse_conn *fc = get_fuse_conn(inode);
 554         struct fuse_fill_data data;
 555         int err;
 556
 557         err = -EIO;
 558         if (is_bad_inode(inode))
 559                 goto out;
 560
 561         data.file = file;
 562         data.inode = inode;
 563         data.req = fuse_get_req(fc);
 564         err = PTR_ERR(data.req);
 565         if (IS_ERR(data.req))
 566                 goto out;
 567
 568         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 569         if (!err) {
 570                 if (data.req->num_pages)
 571                         fuse_send_readpages(data.req, file, inode);
 572                 else
 573                         fuse_put_request(fc, data.req);
 574         }
 575 out:
 576         return err;
 577 }
 578
 579 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 580                                   unsigned long nr_segs, loff_t pos)
 581 {
 582         struct inode *inode = iocb->ki_filp->f_mapping->host;
 583
 584         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 585                 int err;
 586                 /*
 587                  * If trying to read past EOF, make sure the i_size
 588                  * attribute is up-to-date.
 589                  */
 590                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 591                 if (err)
 592                         return err;
 593         }
 594
 595         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 596 }
 597
 598 static void fuse_write_fill(struct fuse_req *req, struct file *file,
 599                             struct fuse_file *ff, struct inode *inode,
 600                             loff_t pos, size_t count, int writepage)
 601 {
 602         struct fuse_conn *fc = get_fuse_conn(inode);
 603         struct fuse_write_in *inarg = &req->misc.write.in;
 604         struct fuse_write_out *outarg = &req->misc.write.out;
 605
 606         memset(inarg, 0, sizeof(struct fuse_write_in));
 607         inarg->fh = ff->fh;
 608         inarg->offset = pos;
 609         inarg->size = count;
 610         inarg->write_flags = writepage ? FUSE_WRITE_CACHE : 0;
 611         inarg->flags = file ? file->f_flags : 0;
 612         req->in.h.opcode = FUSE_WRITE;
 613         req->in.h.nodeid = get_node_id(inode);
 614         req->in.argpages = 1;
 615         req->in.numargs = 2;
 616         if (fc->minor < 9)
 617                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 618         else
 619                 req->in.args[0].size = sizeof(struct fuse_write_in);
 620         req->in.args[0].value = inarg;
 621         req->in.args[1].size = count;
 622         req->out.numargs = 1;
 623         req->out.args[0].size = sizeof(struct fuse_write_out);
 624         req->out.args[0].value = outarg;
 625 }
 626
 627 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 628                               struct inode *inode, loff_t pos, size_t count,
 629                               fl_owner_t owner)
 630 {
 631         struct fuse_conn *fc = get_fuse_conn(inode);
 632         fuse_write_fill(req, file, file->private_data, inode, pos, count, 0);
 633         if (owner != NULL) {
 634                 struct fuse_write_in *inarg = &req->misc.write.in;
 635                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 636                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 637         }
 638         request_send(fc, req);
 639         return req->misc.write.out.size;
 640 }
 641
 642 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 643                         loff_t pos, unsigned len, unsigned flags,
 644                         struct page **pagep, void **fsdata)
 645 {
 646         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 647
 648         *pagep = __grab_cache_page(mapping, index);
 649         if (!*pagep)
 650                 return -ENOMEM;
 651         return 0;
 652 }
 653
 654 static void fuse_write_update_size(struct inode *inode, loff_t pos)
 655 {
 656         struct fuse_conn *fc = get_fuse_conn(inode);
 657         struct fuse_inode *fi = get_fuse_inode(inode);
 658
 659         spin_lock(&fc->lock);
 660         fi->attr_version = ++fc->attr_version;
 661         if (pos > inode->i_size)
 662                 i_size_write(inode, pos);
 663         spin_unlock(&fc->lock);
 664 }
 665
 666 static int fuse_buffered_write(struct file *file, struct inode *inode,
 667                                loff_t pos, unsigned count, struct page *page)
 668 {
 669         int err;
 670         size_t nres;
 671         struct fuse_conn *fc = get_fuse_conn(inode);
 672         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 673         struct fuse_req *req;
 674
 675         if (is_bad_inode(inode))
 676                 return -EIO;
 677
 678         /*
 679          * Make sure writepages on the same page are not mixed up with
 680          * plain writes.
 681          */
 682         fuse_wait_on_page_writeback(inode, page->index);
 683
 684         req = fuse_get_req(fc);
 685         if (IS_ERR(req))
 686                 return PTR_ERR(req);
 687
 688         req->num_pages = 1;
 689         req->pages[0] = page;
 690         req->page_offset = offset;
 691         nres = fuse_send_write(req, file, inode, pos, count, NULL);
 692         err = req->out.h.error;
 693         fuse_put_request(fc, req);
 694         if (!err && !nres)
 695                 err = -EIO;
 696         if (!err) {
 697                 pos += nres;
 698                 fuse_write_update_size(inode, pos);
 699                 if (count == PAGE_CACHE_SIZE)
 700                         SetPageUptodate(page);
 701         }
 702         fuse_invalidate_attr(inode);
 703         return err ? err : nres;
 704 }
 705
 706 static int fuse_write_end(struct file *file, struct address_space *mapping,
 707                         loff_t pos, unsigned len, unsigned copied,
 708                         struct page *page, void *fsdata)
 709 {
 710         struct inode *inode = mapping->host;
 711         int res = 0;
 712
 713         if (copied)
 714                 res = fuse_buffered_write(file, inode, pos, copied, page);
 715
 716         unlock_page(page);
 717         page_cache_release(page);
 718         return res;
 719 }
 720
 721 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 722                                     struct inode *inode, loff_t pos,
 723                                     size_t count)
 724 {
 725         size_t res;
 726         unsigned offset;
 727         unsigned i;
 728
 729         for (i = 0; i < req->num_pages; i++)
 730                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 731
 732         res = fuse_send_write(req, file, inode, pos, count, NULL);
 733
 734         offset = req->page_offset;
 735         count = res;
 736         for (i = 0; i < req->num_pages; i++) {
 737                 struct page *page = req->pages[i];
 738
 739                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 740                         SetPageUptodate(page);
 741
 742                 if (count > PAGE_CACHE_SIZE - offset)
 743                         count -= PAGE_CACHE_SIZE - offset;
 744                 else
 745                         count = 0;
 746                 offset = 0;
 747
 748                 unlock_page(page);
 749                 page_cache_release(page);
 750         }
 751
 752         return res;
 753 }
 754
 755 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 756                                struct address_space *mapping,
 757                                struct iov_iter *ii, loff_t pos)
 758 {
 759         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 760         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 761         size_t count = 0;
 762         int err;
 763
 764         req->page_offset = offset;
 765
 766         do {
 767                 size_t tmp;
 768                 struct page *page;
 769                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 770                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 771                                      iov_iter_count(ii));
 772
 773                 bytes = min_t(size_t, bytes, fc->max_write - count);
 774
 775  again:
 776                 err = -EFAULT;
 777                 if (iov_iter_fault_in_readable(ii, bytes))
 778                         break;
 779
 780                 err = -ENOMEM;
 781                 page = __grab_cache_page(mapping, index);
 782                 if (!page)
 783                         break;
 784
 785                 pagefault_disable();
 786                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 787                 pagefault_enable();
 788                 flush_dcache_page(page);
 789
 790                 if (!tmp) {
 791                         unlock_page(page);
 792                         page_cache_release(page);
 793                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 794                         goto again;
 795                 }
 796
 797                 err = 0;
 798                 req->pages[req->num_pages] = page;
 799                 req->num_pages++;
 800
 801                 iov_iter_advance(ii, tmp);
 802                 count += tmp;
 803                 pos += tmp;
 804                 offset += tmp;
 805                 if (offset == PAGE_CACHE_SIZE)
 806                         offset = 0;
 807
 808                 if (!fc->big_writes)
 809                         break;
 810         } while (iov_iter_count(ii) && count < fc->max_write &&
 811                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 812
 813         return count > 0 ? count : err;
 814 }
 815
 816 static ssize_t fuse_perform_write(struct file *file,
 817                                   struct address_space *mapping,
 818                                   struct iov_iter *ii, loff_t pos)
 819 {
 820         struct inode *inode = mapping->host;
 821         struct fuse_conn *fc = get_fuse_conn(inode);
 822         int err = 0;
 823         ssize_t res = 0;
 824
 825         if (is_bad_inode(inode))
 826                 return -EIO;
 827
 828         do {
 829                 struct fuse_req *req;
 830                 ssize_t count;
 831
 832                 req = fuse_get_req(fc);
 833                 if (IS_ERR(req)) {
 834                         err = PTR_ERR(req);
 835                         break;
 836                 }
 837
 838                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 839                 if (count <= 0) {
 840                         err = count;
 841                 } else {
 842                         size_t num_written;
 843
 844                         num_written = fuse_send_write_pages(req, file, inode,
 845                                                             pos, count);
 846                         err = req->out.h.error;
 847                         if (!err) {
 848                                 res += num_written;
 849                                 pos += num_written;
 850
 851                                 /* break out of the loop on short write */
 852                                 if (num_written != count)
 853                                         err = -EIO;
 854                         }
 855                 }
 856                 fuse_put_request(fc, req);
 857         } while (!err && iov_iter_count(ii));
 858
 859         if (res > 0)
 860                 fuse_write_update_size(inode, pos);
 861
 862         fuse_invalidate_attr(inode);
 863
 864         return res > 0 ? res : err;
 865 }
 866
 867 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 868                                    unsigned long nr_segs, loff_t pos)
 869 {
 870         struct file *file = iocb->ki_filp;
 871         struct address_space *mapping = file->f_mapping;
 872         size_t count = 0;
 873         ssize_t written = 0;
 874         struct inode *inode = mapping->host;
 875         ssize_t err;
 876         struct iov_iter i;
 877
 878         WARN_ON(iocb->ki_pos != pos);
 879
 880         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 881         if (err)
 882                 return err;
 883
 884         mutex_lock(&inode->i_mutex);
 885         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 886
 887         /* We can write back this queue in page reclaim */
 888         current->backing_dev_info = mapping->backing_dev_info;
 889
 890         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 891         if (err)
 892                 goto out;
 893
 894         if (count == 0)
 895                 goto out;
 896
 897         err = file_remove_suid(file);
 898         if (err)
 899                 goto out;
 900
 901         file_update_time(file);
 902
 903         iov_iter_init(&i, iov, nr_segs, count, 0);
 904         written = fuse_perform_write(file, mapping, &i, pos);
 905         if (written >= 0)
 906                 iocb->ki_pos = pos + written;
 907
 908 out:
 909         current->backing_dev_info = NULL;
 910         mutex_unlock(&inode->i_mutex);
 911
 912         return written ? written : err;
 913 }
 914
 915 static void fuse_release_user_pages(struct fuse_req *req, int write)
 916 {
 917         unsigned i;
 918
 919         for (i = 0; i < req->num_pages; i++) {
 920                 struct page *page = req->pages[i];
 921                 if (write)
 922                         set_page_dirty_lock(page);
 923                 put_page(page);
 924         }
 925 }
 926
 927 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 928                                unsigned nbytes, int write)
 929 {
 930         unsigned long user_addr = (unsigned long) buf;
 931         unsigned offset = user_addr & ~PAGE_MASK;
 932         int npages;
 933
 934         /* This doesn't work with nfsd */
 935         if (!current->mm)
 936                 return -EPERM;
 937
 938         nbytes = min(nbytes, (unsigned) FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
 939         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 940         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
 941         down_read(&current->mm->mmap_sem);
 942         npages = get_user_pages(current, current->mm, user_addr, npages, write,
 943                                 0, req->pages, NULL);
 944         up_read(&current->mm->mmap_sem);
 945         if (npages < 0)
 946                 return npages;
 947
 948         req->num_pages = npages;
 949         req->page_offset = offset;
 950         return 0;
 951 }
 952
 953 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
 954                               size_t count, loff_t *ppos, int write)
 955 {
 956         struct inode *inode = file->f_path.dentry->d_inode;
 957         struct fuse_conn *fc = get_fuse_conn(inode);
 958         size_t nmax = write ? fc->max_write : fc->max_read;
 959         loff_t pos = *ppos;
 960         ssize_t res = 0;
 961         struct fuse_req *req;
 962
 963         if (is_bad_inode(inode))
 964                 return -EIO;
 965
 966         req = fuse_get_req(fc);
 967         if (IS_ERR(req))
 968                 return PTR_ERR(req);
 969
 970         while (count) {
 971                 size_t nres;
 972                 size_t nbytes_limit = min(count, nmax);
 973                 size_t nbytes;
 974                 int err = fuse_get_user_pages(req, buf, nbytes_limit, !write);
 975                 if (err) {
 976                         res = err;
 977                         break;
 978                 }
 979                 nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
 980                 nbytes = min(nbytes_limit, nbytes);
 981                 if (write)
 982                         nres = fuse_send_write(req, file, inode, pos, nbytes,
 983                                                current->files);
 984                 else
 985                         nres = fuse_send_read(req, file, inode, pos, nbytes,
 986                                               current->files);
 987                 fuse_release_user_pages(req, !write);
 988                 if (req->out.h.error) {
 989                         if (!res)
 990                                 res = req->out.h.error;
 991                         break;
 992                 } else if (nres > nbytes) {
 993                         res = -EIO;
 994                         break;
 995                 }
 996                 count -= nres;
 997                 res += nres;
 998                 pos += nres;
 999                 buf += nres;
1000                 if (nres != nbytes)
1001                         break;
1002                 if (count) {
1003                         fuse_put_request(fc, req);
1004                         req = fuse_get_req(fc);
1005                         if (IS_ERR(req))
1006                                 break;
1007                 }
1008         }
1009         fuse_put_request(fc, req);
1010         if (res > 0) {
1011                 if (write)
1012                         fuse_write_update_size(inode, pos);
1013                 *ppos = pos;
1014         }
1015         fuse_invalidate_attr(inode);
1016
1017         return res;
1018 }
1019
1020 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1021                                      size_t count, loff_t *ppos)
1022 {
1023         return fuse_direct_io(file, buf, count, ppos, 0);
1024 }
1025
1026 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1027                                  size_t count, loff_t *ppos)
1028 {
1029         struct inode *inode = file->f_path.dentry->d_inode;
1030         ssize_t res;
1031         /* Don't allow parallel writes to the same file */
1032         mutex_lock(&inode->i_mutex);
1033         res = generic_write_checks(file, ppos, &count, 0);
1034         if (!res)
1035                 res = fuse_direct_io(file, buf, count, ppos, 1);
1036         mutex_unlock(&inode->i_mutex);
1037         return res;
1038 }
1039
1040 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1041 {
1042         __free_page(req->pages[0]);
1043         fuse_file_put(req->ff);
1044 }
1045
1046 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1047 {
1048         struct inode *inode = req->inode;
1049         struct fuse_inode *fi = get_fuse_inode(inode);
1050         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1051
1052         list_del(&req->writepages_entry);
1053         dec_bdi_stat(bdi, BDI_WRITEBACK);
1054         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1055         bdi_writeout_inc(bdi);
1056         wake_up(&fi->page_waitq);
1057 }
1058
1059 /* Called under fc->lock, may release and reacquire it */
1060 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1061 {
1062         struct fuse_inode *fi = get_fuse_inode(req->inode);
1063         loff_t size = i_size_read(req->inode);
1064         struct fuse_write_in *inarg = &req->misc.write.in;
1065
1066         if (!fc->connected)
1067                 goto out_free;
1068
1069         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1070                 inarg->size = PAGE_CACHE_SIZE;
1071         } else if (inarg->offset < size) {
1072                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1073         } else {
1074                 /* Got truncated off completely */
1075                 goto out_free;
1076         }
1077
1078         req->in.args[1].size = inarg->size;
1079         fi->writectr++;
1080         request_send_background_locked(fc, req);
1081         return;
1082
1083  out_free:
1084         fuse_writepage_finish(fc, req);
1085         spin_unlock(&fc->lock);
1086         fuse_writepage_free(fc, req);
1087         fuse_put_request(fc, req);
1088         spin_lock(&fc->lock);
1089 }
1090
1091 /*
1092  * If fi->writectr is positive (no truncate or fsync going on) send
1093  * all queued writepage requests.
1094  *
1095  * Called with fc->lock
1096  */
1097 void fuse_flush_writepages(struct inode *inode)
1098 {
1099         struct fuse_conn *fc = get_fuse_conn(inode);
1100         struct fuse_inode *fi = get_fuse_inode(inode);
1101         struct fuse_req *req;
1102
1103         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1104                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1105                 list_del_init(&req->list);
1106                 fuse_send_writepage(fc, req);
1107         }
1108 }
1109
1110 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1111 {
1112         struct inode *inode = req->inode;
1113         struct fuse_inode *fi = get_fuse_inode(inode);
1114
1115         mapping_set_error(inode->i_mapping, req->out.h.error);
1116         spin_lock(&fc->lock);
1117         fi->writectr--;
1118         fuse_writepage_finish(fc, req);
1119         spin_unlock(&fc->lock);
1120         fuse_writepage_free(fc, req);
1121 }
1122
1123 static int fuse_writepage_locked(struct page *page)
1124 {
1125         struct address_space *mapping = page->mapping;
1126         struct inode *inode = mapping->host;
1127         struct fuse_conn *fc = get_fuse_conn(inode);
1128         struct fuse_inode *fi = get_fuse_inode(inode);
1129         struct fuse_req *req;
1130         struct fuse_file *ff;
1131         struct page *tmp_page;
1132
1133         set_page_writeback(page);
1134
1135         req = fuse_request_alloc_nofs();
1136         if (!req)
1137                 goto err;
1138
1139         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1140         if (!tmp_page)
1141                 goto err_free;
1142
1143         spin_lock(&fc->lock);
1144         BUG_ON(list_empty(&fi->write_files));
1145         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1146         req->ff = fuse_file_get(ff);
1147         spin_unlock(&fc->lock);
1148
1149         fuse_write_fill(req, NULL, ff, inode, page_offset(page), 0, 1);
1150
1151         copy_highpage(tmp_page, page);
1152         req->num_pages = 1;
1153         req->pages[0] = tmp_page;
1154         req->page_offset = 0;
1155         req->end = fuse_writepage_end;
1156         req->inode = inode;
1157
1158         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1159         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1160         end_page_writeback(page);
1161
1162         spin_lock(&fc->lock);
1163         list_add(&req->writepages_entry, &fi->writepages);
1164         list_add_tail(&req->list, &fi->queued_writes);
1165         fuse_flush_writepages(inode);
1166         spin_unlock(&fc->lock);
1167
1168         return 0;
1169
1170 err_free:
1171         fuse_request_free(req);
1172 err:
1173         end_page_writeback(page);
1174         return -ENOMEM;
1175 }
1176
1177 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1178 {
1179         int err;
1180
1181         err = fuse_writepage_locked(page);
1182         unlock_page(page);
1183
1184         return err;
1185 }
1186
1187 static int fuse_launder_page(struct page *page)
1188 {
1189         int err = 0;
1190         if (clear_page_dirty_for_io(page)) {
1191                 struct inode *inode = page->mapping->host;
1192                 err = fuse_writepage_locked(page);
1193                 if (!err)
1194                         fuse_wait_on_page_writeback(inode, page->index);
1195         }
1196         return err;
1197 }
1198
1199 /*
1200  * Write back dirty pages now, because there may not be any suitable
1201  * open files later
1202  */
1203 static void fuse_vma_close(struct vm_area_struct *vma)
1204 {
1205         filemap_write_and_wait(vma->vm_file->f_mapping);
1206 }
1207
1208 /*
1209  * Wait for writeback against this page to complete before allowing it
1210  * to be marked dirty again, and hence written back again, possibly
1211  * before the previous writepage completed.
1212  *
1213  * Block here, instead of in ->writepage(), so that the userspace fs
1214  * can only block processes actually operating on the filesystem.
1215  *
1216  * Otherwise unprivileged userspace fs would be able to block
1217  * unrelated:
1218  *
1219  * - page migration
1220  * - sync(2)
1221  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1222  */
1223 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct page *page)
1224 {
1225         /*
1226          * Don't use page->mapping as it may become NULL from a
1227          * concurrent truncate.
1228          */
1229         struct inode *inode = vma->vm_file->f_mapping->host;
1230
1231         fuse_wait_on_page_writeback(inode, page->index);
1232         return 0;
1233 }
1234
1235 static struct vm_operations_struct fuse_file_vm_ops = {
1236         .close          = fuse_vma_close,
1237         .fault          = filemap_fault,
1238         .page_mkwrite   = fuse_page_mkwrite,
1239 };
1240
1241 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1242 {
1243         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1244                 struct inode *inode = file->f_dentry->d_inode;
1245                 struct fuse_conn *fc = get_fuse_conn(inode);
1246                 struct fuse_inode *fi = get_fuse_inode(inode);
1247                 struct fuse_file *ff = file->private_data;
1248                 /*
1249                  * file may be written through mmap, so chain it onto the
1250                  * inodes's write_file list
1251                  */
1252                 spin_lock(&fc->lock);
1253                 if (list_empty(&ff->write_entry))
1254                         list_add(&ff->write_entry, &fi->write_files);
1255                 spin_unlock(&fc->lock);
1256         }
1257         file_accessed(file);
1258         vma->vm_ops = &fuse_file_vm_ops;
1259         return 0;
1260 }
1261
1262 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1263                                   struct file_lock *fl)
1264 {
1265         switch (ffl->type) {
1266         case F_UNLCK:
1267                 break;
1268
1269         case F_RDLCK:
1270         case F_WRLCK:
1271                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1272                     ffl->end < ffl->start)
1273                         return -EIO;
1274
1275                 fl->fl_start = ffl->start;
1276                 fl->fl_end = ffl->end;
1277                 fl->fl_pid = ffl->pid;
1278                 break;
1279
1280         default:
1281                 return -EIO;
1282         }
1283         fl->fl_type = ffl->type;
1284         return 0;
1285 }
1286
1287 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1288                          const struct file_lock *fl, int opcode, pid_t pid,
1289                          int flock)
1290 {
1291         struct inode *inode = file->f_path.dentry->d_inode;
1292         struct fuse_conn *fc = get_fuse_conn(inode);
1293         struct fuse_file *ff = file->private_data;
1294         struct fuse_lk_in *arg = &req->misc.lk_in;
1295
1296         arg->fh = ff->fh;
1297         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1298         arg->lk.start = fl->fl_start;
1299         arg->lk.end = fl->fl_end;
1300         arg->lk.type = fl->fl_type;
1301         arg->lk.pid = pid;
1302         if (flock)
1303                 arg->lk_flags |= FUSE_LK_FLOCK;
1304         req->in.h.opcode = opcode;
1305         req->in.h.nodeid = get_node_id(inode);
1306         req->in.numargs = 1;
1307         req->in.args[0].size = sizeof(*arg);
1308         req->in.args[0].value = arg;
1309 }
1310
1311 static int fuse_getlk(struct file *file, struct file_lock *fl)
1312 {
1313         struct inode *inode = file->f_path.dentry->d_inode;
1314         struct fuse_conn *fc = get_fuse_conn(inode);
1315         struct fuse_req *req;
1316         struct fuse_lk_out outarg;
1317         int err;
1318
1319         req = fuse_get_req(fc);
1320         if (IS_ERR(req))
1321                 return PTR_ERR(req);
1322
1323         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1324         req->out.numargs = 1;
1325         req->out.args[0].size = sizeof(outarg);
1326         req->out.args[0].value = &outarg;
1327         request_send(fc, req);
1328         err = req->out.h.error;
1329         fuse_put_request(fc, req);
1330         if (!err)
1331                 err = convert_fuse_file_lock(&outarg.lk, fl);
1332
1333         return err;
1334 }
1335
1336 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1337 {
1338         struct inode *inode = file->f_path.dentry->d_inode;
1339         struct fuse_conn *fc = get_fuse_conn(inode);
1340         struct fuse_req *req;
1341         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1342         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1343         int err;
1344
1345         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1346                 /* NLM needs asynchronous locks, which we don't support yet */
1347                 return -ENOLCK;
1348         }
1349
1350         /* Unlock on close is handled by the flush method */
1351         if (fl->fl_flags & FL_CLOSE)
1352                 return 0;
1353
1354         req = fuse_get_req(fc);
1355         if (IS_ERR(req))
1356                 return PTR_ERR(req);
1357
1358         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1359         request_send(fc, req);
1360         err = req->out.h.error;
1361         /* locking is restartable */
1362         if (err == -EINTR)
1363                 err = -ERESTARTSYS;
1364         fuse_put_request(fc, req);
1365         return err;
1366 }
1367
1368 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1369 {
1370         struct inode *inode = file->f_path.dentry->d_inode;
1371         struct fuse_conn *fc = get_fuse_conn(inode);
1372         int err;
1373
1374         if (cmd == F_CANCELLK) {
1375                 err = 0;
1376         } else if (cmd == F_GETLK) {
1377                 if (fc->no_lock) {
1378                         posix_test_lock(file, fl);
1379                         err = 0;
1380                 } else
1381                         err = fuse_getlk(file, fl);
1382         } else {
1383                 if (fc->no_lock)
1384                         err = posix_lock_file(file, fl, NULL);
1385                 else
1386                         err = fuse_setlk(file, fl, 0);
1387         }
1388         return err;
1389 }
1390
1391 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1392 {
1393         struct inode *inode = file->f_path.dentry->d_inode;
1394         struct fuse_conn *fc = get_fuse_conn(inode);
1395         int err;
1396
1397         if (fc->no_lock) {
1398                 err = flock_lock_file_wait(file, fl);
1399         } else {
1400                 /* emulate flock with POSIX locks */
1401                 fl->fl_owner = (fl_owner_t) file;
1402                 err = fuse_setlk(file, fl, 1);
1403         }
1404
1405         return err;
1406 }
1407
1408 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1409 {
1410         struct inode *inode = mapping->host;
1411         struct fuse_conn *fc = get_fuse_conn(inode);
1412         struct fuse_req *req;
1413         struct fuse_bmap_in inarg;
1414         struct fuse_bmap_out outarg;
1415         int err;
1416
1417         if (!inode->i_sb->s_bdev || fc->no_bmap)
1418                 return 0;
1419
1420         req = fuse_get_req(fc);
1421         if (IS_ERR(req))
1422                 return 0;
1423
1424         memset(&inarg, 0, sizeof(inarg));
1425         inarg.block = block;
1426         inarg.blocksize = inode->i_sb->s_blocksize;
1427         req->in.h.opcode = FUSE_BMAP;
1428         req->in.h.nodeid = get_node_id(inode);
1429         req->in.numargs = 1;
1430         req->in.args[0].size = sizeof(inarg);
1431         req->in.args[0].value = &inarg;
1432         req->out.numargs = 1;
1433         req->out.args[0].size = sizeof(outarg);
1434         req->out.args[0].value = &outarg;
1435         request_send(fc, req);
1436         err = req->out.h.error;
1437         fuse_put_request(fc, req);
1438         if (err == -ENOSYS)
1439                 fc->no_bmap = 1;
1440
1441         return err ? 0 : outarg.block;
1442 }
1443
1444 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1445 {
1446         loff_t retval;
1447         struct inode *inode = file->f_path.dentry->d_inode;
1448
1449         mutex_lock(&inode->i_mutex);
1450         switch (origin) {
1451         case SEEK_END:
1452                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1453                 if (retval)
1454                         return retval;
1455                 offset += i_size_read(inode);
1456                 break;
1457         case SEEK_CUR:
1458                 offset += file->f_pos;
1459         }
1460         retval = -EINVAL;
1461         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1462                 if (offset != file->f_pos) {
1463                         file->f_pos = offset;
1464                         file->f_version = 0;
1465                 }
1466                 retval = offset;
1467         }
1468         mutex_unlock(&inode->i_mutex);
1469         return retval;
1470 }
1471
1472 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1473                         unsigned int nr_segs, size_t bytes, bool to_user)
1474 {
1475         struct iov_iter ii;
1476         int page_idx = 0;
1477
1478         if (!bytes)
1479                 return 0;
1480
1481         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1482
1483         while (iov_iter_count(&ii)) {
1484                 struct page *page = pages[page_idx++];
1485                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1486                 void *kaddr, *map;
1487
1488                 kaddr = map = kmap(page);
1489
1490                 while (todo) {
1491                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1492                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1493                         size_t copy = min(todo, iov_len);
1494                         size_t left;
1495
1496                         if (!to_user)
1497                                 left = copy_from_user(kaddr, uaddr, copy);
1498                         else
1499                                 left = copy_to_user(uaddr, kaddr, copy);
1500
1501                         if (unlikely(left))
1502                                 return -EFAULT;
1503
1504                         iov_iter_advance(&ii, copy);
1505                         todo -= copy;
1506                         kaddr += copy;
1507                 }
1508
1509                 kunmap(map);
1510         }
1511
1512         return 0;
1513 }
1514
1515 /*
1516  * For ioctls, there is no generic way to determine how much memory
1517  * needs to be read and/or written.  Furthermore, ioctls are allowed
1518  * to dereference the passed pointer, so the parameter requires deep
1519  * copying but FUSE has no idea whatsoever about what to copy in or
1520  * out.
1521  *
1522  * This is solved by allowing FUSE server to retry ioctl with
1523  * necessary in/out iovecs.  Let's assume the ioctl implementation
1524  * needs to read in the following structure.
1525  *
1526  * struct a {
1527  *      char    *buf;
1528  *      size_t  buflen;
1529  * }
1530  *
1531  * On the first callout to FUSE server, inarg->in_size and
1532  * inarg->out_size will be NULL; then, the server completes the ioctl
1533  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1534  * the actual iov array to
1535  *
1536  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1537  *
1538  * which tells FUSE to copy in the requested area and retry the ioctl.
1539  * On the second round, the server has access to the structure and
1540  * from that it can tell what to look for next, so on the invocation,
1541  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1542  *
1543  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1544  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1545  *
1546  * FUSE will copy both struct a and the pointed buffer from the
1547  * process doing the ioctl and retry ioctl with both struct a and the
1548  * buffer.
1549  *
1550  * This time, FUSE server has everything it needs and completes ioctl
1551  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1552  *
1553  * Copying data out works the same way.
1554  *
1555  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1556  * automatically initializes in and out iovs by decoding @cmd with
1557  * _IOC_* macros and the server is not allowed to request RETRY.  This
1558  * limits ioctl data transfers to well-formed ioctls and is the forced
1559  * behavior for all FUSE servers.
1560  */
1561 static long fuse_file_do_ioctl(struct file *file, unsigned int cmd,
1562                                unsigned long arg, unsigned int flags)
1563 {
1564         struct inode *inode = file->f_dentry->d_inode;
1565         struct fuse_file *ff = file->private_data;
1566         struct fuse_conn *fc = get_fuse_conn(inode);
1567         struct fuse_ioctl_in inarg = {
1568                 .fh = ff->fh,
1569                 .cmd = cmd,
1570                 .arg = arg,
1571                 .flags = flags
1572         };
1573         struct fuse_ioctl_out outarg;
1574         struct fuse_req *req = NULL;
1575         struct page **pages = NULL;
1576         struct page *iov_page = NULL;
1577         struct iovec *in_iov = NULL, *out_iov = NULL;
1578         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1579         size_t in_size, out_size, transferred;
1580         int err;
1581
1582         /* assume all the iovs returned by client always fits in a page */
1583         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1584
1585         if (!fuse_allow_task(fc, current))
1586                 return -EACCES;
1587
1588         err = -EIO;
1589         if (is_bad_inode(inode))
1590                 goto out;
1591
1592         err = -ENOMEM;
1593         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1594         iov_page = alloc_page(GFP_KERNEL);
1595         if (!pages || !iov_page)
1596                 goto out;
1597
1598         /*
1599          * If restricted, initialize IO parameters as encoded in @cmd.
1600          * RETRY from server is not allowed.
1601          */
1602         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1603                 struct iovec *iov = page_address(iov_page);
1604
1605                 iov->iov_base = (void *)arg;
1606                 iov->iov_len = _IOC_SIZE(cmd);
1607
1608                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1609                         in_iov = iov;
1610                         in_iovs = 1;
1611                 }
1612
1613                 if (_IOC_DIR(cmd) & _IOC_READ) {
1614                         out_iov = iov;
1615                         out_iovs = 1;
1616                 }
1617         }
1618
1619  retry:
1620         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1621         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1622
1623         /*
1624          * Out data can be used either for actual out data or iovs,
1625          * make sure there always is at least one page.
1626          */
1627         out_size = max_t(size_t, out_size, PAGE_SIZE);
1628         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1629
1630         /* make sure there are enough buffer pages and init request with them */
1631         err = -ENOMEM;
1632         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1633                 goto out;
1634         while (num_pages < max_pages) {
1635                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1636                 if (!pages[num_pages])
1637                         goto out;
1638                 num_pages++;
1639         }
1640
1641         req = fuse_get_req(fc);
1642         if (IS_ERR(req)) {
1643                 err = PTR_ERR(req);
1644                 req = NULL;
1645                 goto out;
1646         }
1647         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1648         req->num_pages = num_pages;
1649
1650         /* okay, let's send it to the client */
1651         req->in.h.opcode = FUSE_IOCTL;
1652         req->in.h.nodeid = get_node_id(inode);
1653         req->in.numargs = 1;
1654         req->in.args[0].size = sizeof(inarg);
1655         req->in.args[0].value = &inarg;
1656         if (in_size) {
1657                 req->in.numargs++;
1658                 req->in.args[1].size = in_size;
1659                 req->in.argpages = 1;
1660
1661                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1662                                            false);
1663                 if (err)
1664                         goto out;
1665         }
1666
1667         req->out.numargs = 2;
1668         req->out.args[0].size = sizeof(outarg);
1669         req->out.args[0].value = &outarg;
1670         req->out.args[1].size = out_size;
1671         req->out.argpages = 1;
1672         req->out.argvar = 1;
1673
1674         request_send(fc, req);
1675         err = req->out.h.error;
1676         transferred = req->out.args[1].size;
1677         fuse_put_request(fc, req);
1678         req = NULL;
1679         if (err)
1680                 goto out;
1681
1682         /* did it ask for retry? */
1683         if (outarg.flags & FUSE_IOCTL_RETRY) {
1684                 char *vaddr;
1685
1686                 /* no retry if in restricted mode */
1687                 err = -EIO;
1688                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1689                         goto out;
1690
1691                 in_iovs = outarg.in_iovs;
1692                 out_iovs = outarg.out_iovs;
1693
1694                 /*
1695                  * Make sure things are in boundary, separate checks
1696                  * are to protect against overflow.
1697                  */
1698                 err = -ENOMEM;
1699                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1700                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1701                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1702                         goto out;
1703
1704                 err = -EIO;
1705                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1706                         goto out;
1707
1708                 /* okay, copy in iovs and retry */
1709                 vaddr = kmap_atomic(pages[0], KM_USER0);
1710                 memcpy(page_address(iov_page), vaddr, transferred);
1711                 kunmap_atomic(vaddr, KM_USER0);
1712
1713                 in_iov = page_address(iov_page);
1714                 out_iov = in_iov + in_iovs;
1715
1716                 goto retry;
1717         }
1718
1719         err = -EIO;
1720         if (transferred > inarg.out_size)
1721                 goto out;
1722
1723         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1724  out:
1725         if (req)
1726                 fuse_put_request(fc, req);
1727         if (iov_page)
1728                 __free_page(iov_page);
1729         while (num_pages)
1730                 __free_page(pages[--num_pages]);
1731         kfree(pages);
1732
1733         return err ? err : outarg.result;
1734 }
1735
1736 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1737                             unsigned long arg)
1738 {
1739         return fuse_file_do_ioctl(file, cmd, arg, 0);
1740 }
1741
1742 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1743                                    unsigned long arg)
1744 {
1745         return fuse_file_do_ioctl(file, cmd, arg, FUSE_IOCTL_COMPAT);
1746 }
1747
1748 static const struct file_operations fuse_file_operations = {
1749         .llseek         = fuse_file_llseek,
1750         .read           = do_sync_read,
1751         .aio_read       = fuse_file_aio_read,
1752         .write          = do_sync_write,
1753         .aio_write      = fuse_file_aio_write,
1754         .mmap           = fuse_file_mmap,
1755         .open           = fuse_open,
1756         .flush          = fuse_flush,
1757         .release        = fuse_release,
1758         .fsync          = fuse_fsync,
1759         .lock           = fuse_file_lock,
1760         .flock          = fuse_file_flock,
1761         .splice_read    = generic_file_splice_read,
1762         .unlocked_ioctl = fuse_file_ioctl,
1763         .compat_ioctl   = fuse_file_compat_ioctl,
1764 };
1765
1766 static const struct file_operations fuse_direct_io_file_operations = {
1767         .llseek         = fuse_file_llseek,
1768         .read           = fuse_direct_read,
1769         .write          = fuse_direct_write,
1770         .open           = fuse_open,
1771         .flush          = fuse_flush,
1772         .release        = fuse_release,
1773         .fsync          = fuse_fsync,
1774         .lock           = fuse_file_lock,
1775         .flock          = fuse_file_flock,
1776         .unlocked_ioctl = fuse_file_ioctl,
1777         .compat_ioctl   = fuse_file_compat_ioctl,
1778         /* no mmap and splice_read */
1779 };
1780
1781 static const struct address_space_operations fuse_file_aops  = {
1782         .readpage       = fuse_readpage,
1783         .writepage      = fuse_writepage,
1784         .launder_page   = fuse_launder_page,
1785         .write_begin    = fuse_write_begin,
1786         .write_end      = fuse_write_end,
1787         .readpages      = fuse_readpages,
1788         .set_page_dirty = __set_page_dirty_nobuffers,
1789         .bmap           = fuse_bmap,
1790 };
1791
1792 void fuse_init_file_inode(struct inode *inode)
1793 {
1794         inode->i_fop = &fuse_file_operations;
1795         inode->i_data.a_ops = &fuse_file_aops;
1796 }