SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 fuse_conn *fc, u64 nodeid, struct file *file,
  19                           int opcode, struct fuse_open_out *outargp)
  20 {
  21         struct fuse_open_in inarg;
  22         struct fuse_req *req;
  23         int err;
  24
  25         req = fuse_get_req(fc);
  26         if (IS_ERR(req))
  27                 return PTR_ERR(req);
  28
  29         memset(&inarg, 0, sizeof(inarg));
  30         inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  31         if (!fc->atomic_o_trunc)
  32                 inarg.flags &= ~O_TRUNC;
  33         req->in.h.opcode = opcode;
  34         req->in.h.nodeid = nodeid;
  35         req->in.numargs = 1;
  36         req->in.args[0].size = sizeof(inarg);
  37         req->in.args[0].value = &inarg;
  38         req->out.numargs = 1;
  39         req->out.args[0].size = sizeof(*outargp);
  40         req->out.args[0].value = outargp;
  41         fuse_request_send(fc, req);
  42         err = req->out.h.error;
  43         fuse_put_request(fc, req);
  44
  45         return err;
  46 }
  47
  48 struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  49 {
  50         struct fuse_file *ff;
  51
  52         ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  53         if (unlikely(!ff))
  54                 return NULL;
  55
  56         ff->fc = fc;
  57         ff->reserved_req = fuse_request_alloc();
  58         if (unlikely(!ff->reserved_req)) {
  59                 kfree(ff);
  60                 return NULL;
  61         }
  62
  63         INIT_LIST_HEAD(&ff->write_entry);
  64         atomic_set(&ff->count, 0);
  65         RB_CLEAR_NODE(&ff->polled_node);
  66         init_waitqueue_head(&ff->poll_wait);
  67
  68         spin_lock(&fc->lock);
  69         ff->kh = ++fc->khctr;
  70         spin_unlock(&fc->lock);
  71
  72         return ff;
  73 }
  74
  75 void fuse_file_free(struct fuse_file *ff)
  76 {
  77         fuse_request_free(ff->reserved_req);
  78         kfree(ff);
  79 }
  80
  81 struct fuse_file *fuse_file_get(struct fuse_file *ff)
  82 {
  83         atomic_inc(&ff->count);
  84         return ff;
  85 }
  86
  87 static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
  88 {
  89         path_put(&req->misc.release.path);
  90 }
  91
  92 static void fuse_file_put(struct fuse_file *ff)
  93 {
  94         if (atomic_dec_and_test(&ff->count)) {
  95                 struct fuse_req *req = ff->reserved_req;
  96
  97                 req->end = fuse_release_end;
  98                 fuse_request_send_background(ff->fc, req);
  99                 kfree(ff);
 100         }
 101 }
 102
 103 static int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 104                         bool isdir)
 105 {
 106         struct fuse_open_out outarg;
 107         struct fuse_file *ff;
 108         int err;
 109         int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 110
 111         ff = fuse_file_alloc(fc);
 112         if (!ff)
 113                 return -ENOMEM;
 114
 115         err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 116         if (err) {
 117                 fuse_file_free(ff);
 118                 return err;
 119         }
 120
 121         if (isdir)
 122                 outarg.open_flags &= ~FOPEN_DIRECT_IO;
 123
 124         ff->fh = outarg.fh;
 125         ff->nodeid = nodeid;
 126         ff->open_flags = outarg.open_flags;
 127         file->private_data = fuse_file_get(ff);
 128
 129         return 0;
 130 }
 131
 132 void fuse_finish_open(struct inode *inode, struct file *file)
 133 {
 134         struct fuse_file *ff = file->private_data;
 135
 136         if (ff->open_flags & FOPEN_DIRECT_IO)
 137                 file->f_op = &fuse_direct_io_file_operations;
 138         if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 139                 invalidate_inode_pages2(inode->i_mapping);
 140         if (ff->open_flags & FOPEN_NONSEEKABLE)
 141                 nonseekable_open(inode, file);
 142 }
 143
 144 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 145 {
 146         struct fuse_conn *fc = get_fuse_conn(inode);
 147         int err;
 148
 149         /* VFS checks this, but only _after_ ->open() */
 150         if (file->f_flags & O_DIRECT)
 151                 return -EINVAL;
 152
 153         err = generic_file_open(inode, file);
 154         if (err)
 155                 return err;
 156
 157         err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 158         if (err)
 159                 return err;
 160
 161         fuse_finish_open(inode, file);
 162
 163         return 0;
 164 }
 165
 166 static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 167 {
 168         struct fuse_conn *fc = ff->fc;
 169         struct fuse_req *req = ff->reserved_req;
 170         struct fuse_release_in *inarg = &req->misc.release.in;
 171
 172         spin_lock(&fc->lock);
 173         list_del(&ff->write_entry);
 174         if (!RB_EMPTY_NODE(&ff->polled_node))
 175                 rb_erase(&ff->polled_node, &fc->polled_files);
 176         spin_unlock(&fc->lock);
 177
 178         wake_up_interruptible_sync(&ff->poll_wait);
 179
 180         inarg->fh = ff->fh;
 181         inarg->flags = flags;
 182         req->in.h.opcode = opcode;
 183         req->in.h.nodeid = ff->nodeid;
 184         req->in.numargs = 1;
 185         req->in.args[0].size = sizeof(struct fuse_release_in);
 186         req->in.args[0].value = inarg;
 187 }
 188
 189 void fuse_release_common(struct file *file, int opcode)
 190 {
 191         struct fuse_file *ff;
 192         struct fuse_req *req;
 193
 194         ff = file->private_data;
 195         if (unlikely(!ff))
 196                 return;
 197
 198         req = ff->reserved_req;
 199         fuse_prepare_release(ff, file->f_flags, opcode);
 200
 201         /* Hold vfsmount and dentry until release is finished */
 202         path_get(&file->f_path);
 203         req->misc.release.path = file->f_path;
 204
 205         /*
 206          * Normally this will send the RELEASE request, however if
 207          * some asynchronous READ or WRITE requests are outstanding,
 208          * the sending will be delayed.
 209          */
 210         fuse_file_put(ff);
 211 }
 212
 213 static int fuse_open(struct inode *inode, struct file *file)
 214 {
 215         return fuse_open_common(inode, file, false);
 216 }
 217
 218 static int fuse_release(struct inode *inode, struct file *file)
 219 {
 220         fuse_release_common(file, FUSE_RELEASE);
 221
 222         /* return value is ignored by VFS */
 223         return 0;
 224 }
 225
 226 void fuse_sync_release(struct fuse_file *ff, int flags)
 227 {
 228         WARN_ON(atomic_read(&ff->count) > 1);
 229         fuse_prepare_release(ff, flags, FUSE_RELEASE);
 230         ff->reserved_req->force = 1;
 231         fuse_request_send(ff->fc, ff->reserved_req);
 232         fuse_put_request(ff->fc, ff->reserved_req);
 233         kfree(ff);
 234 }
 235
 236 /*
 237  * Scramble the ID space with XTEA, so that the value of the files_struct
 238  * pointer is not exposed to userspace.
 239  */
 240 u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 241 {
 242         u32 *k = fc->scramble_key;
 243         u64 v = (unsigned long) id;
 244         u32 v0 = v;
 245         u32 v1 = v >> 32;
 246         u32 sum = 0;
 247         int i;
 248
 249         for (i = 0; i < 32; i++) {
 250                 v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 251                 sum += 0x9E3779B9;
 252                 v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 253         }
 254
 255         return (u64) v0 + ((u64) v1 << 32);
 256 }
 257
 258 /*
 259  * Check if page is under writeback
 260  *
 261  * This is currently done by walking the list of writepage requests
 262  * for the inode, which can be pretty inefficient.
 263  */
 264 static bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 265 {
 266         struct fuse_conn *fc = get_fuse_conn(inode);
 267         struct fuse_inode *fi = get_fuse_inode(inode);
 268         struct fuse_req *req;
 269         bool found = false;
 270
 271         spin_lock(&fc->lock);
 272         list_for_each_entry(req, &fi->writepages, writepages_entry) {
 273                 pgoff_t curr_index;
 274
 275                 BUG_ON(req->inode != inode);
 276                 curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 277                 if (curr_index == index) {
 278                         found = true;
 279                         break;
 280                 }
 281         }
 282         spin_unlock(&fc->lock);
 283
 284         return found;
 285 }
 286
 287 /*
 288  * Wait for page writeback to be completed.
 289  *
 290  * Since fuse doesn't rely on the VM writeback tracking, this has to
 291  * use some other means.
 292  */
 293 static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 294 {
 295         struct fuse_inode *fi = get_fuse_inode(inode);
 296
 297         wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 298         return 0;
 299 }
 300
 301 static int fuse_flush(struct file *file, fl_owner_t id)
 302 {
 303         struct inode *inode = file->f_path.dentry->d_inode;
 304         struct fuse_conn *fc = get_fuse_conn(inode);
 305         struct fuse_file *ff = file->private_data;
 306         struct fuse_req *req;
 307         struct fuse_flush_in inarg;
 308         int err;
 309
 310         if (is_bad_inode(inode))
 311                 return -EIO;
 312
 313         if (fc->no_flush)
 314                 return 0;
 315
 316         req = fuse_get_req_nofail(fc, file);
 317         memset(&inarg, 0, sizeof(inarg));
 318         inarg.fh = ff->fh;
 319         inarg.lock_owner = fuse_lock_owner_id(fc, id);
 320         req->in.h.opcode = FUSE_FLUSH;
 321         req->in.h.nodeid = get_node_id(inode);
 322         req->in.numargs = 1;
 323         req->in.args[0].size = sizeof(inarg);
 324         req->in.args[0].value = &inarg;
 325         req->force = 1;
 326         fuse_request_send(fc, req);
 327         err = req->out.h.error;
 328         fuse_put_request(fc, req);
 329         if (err == -ENOSYS) {
 330                 fc->no_flush = 1;
 331                 err = 0;
 332         }
 333         return err;
 334 }
 335
 336 /*
 337  * Wait for all pending writepages on the inode to finish.
 338  *
 339  * This is currently done by blocking further writes with FUSE_NOWRITE
 340  * and waiting for all sent writes to complete.
 341  *
 342  * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 343  * could conflict with truncation.
 344  */
 345 static void fuse_sync_writes(struct inode *inode)
 346 {
 347         fuse_set_nowrite(inode);
 348         fuse_release_nowrite(inode);
 349 }
 350
 351 int fuse_fsync_common(struct file *file, struct dentry *de, int datasync,
 352                       int isdir)
 353 {
 354         struct inode *inode = de->d_inode;
 355         struct fuse_conn *fc = get_fuse_conn(inode);
 356         struct fuse_file *ff = file->private_data;
 357         struct fuse_req *req;
 358         struct fuse_fsync_in inarg;
 359         int err;
 360
 361         if (is_bad_inode(inode))
 362                 return -EIO;
 363
 364         if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 365                 return 0;
 366
 367         /*
 368          * Start writeback against all dirty pages of the inode, then
 369          * wait for all outstanding writes, before sending the FSYNC
 370          * request.
 371          */
 372         err = write_inode_now(inode, 0);
 373         if (err)
 374                 return err;
 375
 376         fuse_sync_writes(inode);
 377
 378         req = fuse_get_req(fc);
 379         if (IS_ERR(req))
 380                 return PTR_ERR(req);
 381
 382         memset(&inarg, 0, sizeof(inarg));
 383         inarg.fh = ff->fh;
 384         inarg.fsync_flags = datasync ? 1 : 0;
 385         req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 386         req->in.h.nodeid = get_node_id(inode);
 387         req->in.numargs = 1;
 388         req->in.args[0].size = sizeof(inarg);
 389         req->in.args[0].value = &inarg;
 390         fuse_request_send(fc, req);
 391         err = req->out.h.error;
 392         fuse_put_request(fc, req);
 393         if (err == -ENOSYS) {
 394                 if (isdir)
 395                         fc->no_fsyncdir = 1;
 396                 else
 397                         fc->no_fsync = 1;
 398                 err = 0;
 399         }
 400         return err;
 401 }
 402
 403 static int fuse_fsync(struct file *file, struct dentry *de, int datasync)
 404 {
 405         return fuse_fsync_common(file, de, datasync, 0);
 406 }
 407
 408 void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 409                     size_t count, int opcode)
 410 {
 411         struct fuse_read_in *inarg = &req->misc.read.in;
 412         struct fuse_file *ff = file->private_data;
 413
 414         inarg->fh = ff->fh;
 415         inarg->offset = pos;
 416         inarg->size = count;
 417         inarg->flags = file->f_flags;
 418         req->in.h.opcode = opcode;
 419         req->in.h.nodeid = ff->nodeid;
 420         req->in.numargs = 1;
 421         req->in.args[0].size = sizeof(struct fuse_read_in);
 422         req->in.args[0].value = inarg;
 423         req->out.argvar = 1;
 424         req->out.numargs = 1;
 425         req->out.args[0].size = count;
 426 }
 427
 428 static size_t fuse_send_read(struct fuse_req *req, struct file *file,
 429                              loff_t pos, size_t count, fl_owner_t owner)
 430 {
 431         struct fuse_file *ff = file->private_data;
 432         struct fuse_conn *fc = ff->fc;
 433
 434         fuse_read_fill(req, file, pos, count, FUSE_READ);
 435         if (owner != NULL) {
 436                 struct fuse_read_in *inarg = &req->misc.read.in;
 437
 438                 inarg->read_flags |= FUSE_READ_LOCKOWNER;
 439                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 440         }
 441         fuse_request_send(fc, req);
 442         return req->out.args[0].size;
 443 }
 444
 445 static void fuse_read_update_size(struct inode *inode, loff_t size,
 446                                   u64 attr_ver)
 447 {
 448         struct fuse_conn *fc = get_fuse_conn(inode);
 449         struct fuse_inode *fi = get_fuse_inode(inode);
 450
 451         spin_lock(&fc->lock);
 452         if (attr_ver == fi->attr_version && size < inode->i_size) {
 453                 fi->attr_version = ++fc->attr_version;
 454                 i_size_write(inode, size);
 455         }
 456         spin_unlock(&fc->lock);
 457 }
 458
 459 static int fuse_readpage(struct file *file, struct page *page)
 460 {
 461         struct inode *inode = page->mapping->host;
 462         struct fuse_conn *fc = get_fuse_conn(inode);
 463         struct fuse_req *req;
 464         size_t num_read;
 465         loff_t pos = page_offset(page);
 466         size_t count = PAGE_CACHE_SIZE;
 467         u64 attr_ver;
 468         int err;
 469
 470         err = -EIO;
 471         if (is_bad_inode(inode))
 472                 goto out;
 473
 474         /*
 475          * Page writeback can extend beyond the liftime of the
 476          * page-cache page, so make sure we read a properly synced
 477          * page.
 478          */
 479         fuse_wait_on_page_writeback(inode, page->index);
 480
 481         req = fuse_get_req(fc);
 482         err = PTR_ERR(req);
 483         if (IS_ERR(req))
 484                 goto out;
 485
 486         attr_ver = fuse_get_attr_version(fc);
 487
 488         req->out.page_zeroing = 1;
 489         req->out.argpages = 1;
 490         req->num_pages = 1;
 491         req->pages[0] = page;
 492         num_read = fuse_send_read(req, file, pos, count, NULL);
 493         err = req->out.h.error;
 494         fuse_put_request(fc, req);
 495
 496         if (!err) {
 497                 /*
 498                  * Short read means EOF.  If file size is larger, truncate it
 499                  */
 500                 if (num_read < count)
 501                         fuse_read_update_size(inode, pos + num_read, attr_ver);
 502
 503                 SetPageUptodate(page);
 504         }
 505
 506         fuse_invalidate_attr(inode); /* atime changed */
 507  out:
 508         unlock_page(page);
 509         return err;
 510 }
 511
 512 static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 513 {
 514         int i;
 515         size_t count = req->misc.read.in.size;
 516         size_t num_read = req->out.args[0].size;
 517         struct inode *inode = req->pages[0]->mapping->host;
 518
 519         /*
 520          * Short read means EOF.  If file size is larger, truncate it
 521          */
 522         if (!req->out.h.error && num_read < count) {
 523                 loff_t pos = page_offset(req->pages[0]) + num_read;
 524                 fuse_read_update_size(inode, pos, req->misc.read.attr_ver);
 525         }
 526
 527         fuse_invalidate_attr(inode); /* atime changed */
 528
 529         for (i = 0; i < req->num_pages; i++) {
 530                 struct page *page = req->pages[i];
 531                 if (!req->out.h.error)
 532                         SetPageUptodate(page);
 533                 else
 534                         SetPageError(page);
 535                 unlock_page(page);
 536         }
 537         if (req->ff)
 538                 fuse_file_put(req->ff);
 539 }
 540
 541 static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 542 {
 543         struct fuse_file *ff = file->private_data;
 544         struct fuse_conn *fc = ff->fc;
 545         loff_t pos = page_offset(req->pages[0]);
 546         size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 547
 548         req->out.argpages = 1;
 549         req->out.page_zeroing = 1;
 550         fuse_read_fill(req, file, pos, count, FUSE_READ);
 551         req->misc.read.attr_ver = fuse_get_attr_version(fc);
 552         if (fc->async_read) {
 553                 req->ff = fuse_file_get(ff);
 554                 req->end = fuse_readpages_end;
 555                 fuse_request_send_background(fc, req);
 556         } else {
 557                 fuse_request_send(fc, req);
 558                 fuse_readpages_end(fc, req);
 559                 fuse_put_request(fc, req);
 560         }
 561 }
 562
 563 struct fuse_fill_data {
 564         struct fuse_req *req;
 565         struct file *file;
 566         struct inode *inode;
 567 };
 568
 569 static int fuse_readpages_fill(void *_data, struct page *page)
 570 {
 571         struct fuse_fill_data *data = _data;
 572         struct fuse_req *req = data->req;
 573         struct inode *inode = data->inode;
 574         struct fuse_conn *fc = get_fuse_conn(inode);
 575
 576         fuse_wait_on_page_writeback(inode, page->index);
 577
 578         if (req->num_pages &&
 579             (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 580              (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 581              req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 582                 fuse_send_readpages(req, data->file);
 583                 data->req = req = fuse_get_req(fc);
 584                 if (IS_ERR(req)) {
 585                         unlock_page(page);
 586                         return PTR_ERR(req);
 587                 }
 588         }
 589         req->pages[req->num_pages] = page;
 590         req->num_pages++;
 591         return 0;
 592 }
 593
 594 static int fuse_readpages(struct file *file, struct address_space *mapping,
 595                           struct list_head *pages, unsigned nr_pages)
 596 {
 597         struct inode *inode = mapping->host;
 598         struct fuse_conn *fc = get_fuse_conn(inode);
 599         struct fuse_fill_data data;
 600         int err;
 601
 602         err = -EIO;
 603         if (is_bad_inode(inode))
 604                 goto out;
 605
 606         data.file = file;
 607         data.inode = inode;
 608         data.req = fuse_get_req(fc);
 609         err = PTR_ERR(data.req);
 610         if (IS_ERR(data.req))
 611                 goto out;
 612
 613         err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 614         if (!err) {
 615                 if (data.req->num_pages)
 616                         fuse_send_readpages(data.req, file);
 617                 else
 618                         fuse_put_request(fc, data.req);
 619         }
 620 out:
 621         return err;
 622 }
 623
 624 static ssize_t fuse_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 625                                   unsigned long nr_segs, loff_t pos)
 626 {
 627         struct inode *inode = iocb->ki_filp->f_mapping->host;
 628
 629         if (pos + iov_length(iov, nr_segs) > i_size_read(inode)) {
 630                 int err;
 631                 /*
 632                  * If trying to read past EOF, make sure the i_size
 633                  * attribute is up-to-date.
 634                  */
 635                 err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 636                 if (err)
 637                         return err;
 638         }
 639
 640         return generic_file_aio_read(iocb, iov, nr_segs, pos);
 641 }
 642
 643 static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 644                             loff_t pos, size_t count)
 645 {
 646         struct fuse_write_in *inarg = &req->misc.write.in;
 647         struct fuse_write_out *outarg = &req->misc.write.out;
 648
 649         inarg->fh = ff->fh;
 650         inarg->offset = pos;
 651         inarg->size = count;
 652         req->in.h.opcode = FUSE_WRITE;
 653         req->in.h.nodeid = ff->nodeid;
 654         req->in.numargs = 2;
 655         if (ff->fc->minor < 9)
 656                 req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 657         else
 658                 req->in.args[0].size = sizeof(struct fuse_write_in);
 659         req->in.args[0].value = inarg;
 660         req->in.args[1].size = count;
 661         req->out.numargs = 1;
 662         req->out.args[0].size = sizeof(struct fuse_write_out);
 663         req->out.args[0].value = outarg;
 664 }
 665
 666 static size_t fuse_send_write(struct fuse_req *req, struct file *file,
 667                               loff_t pos, size_t count, fl_owner_t owner)
 668 {
 669         struct fuse_file *ff = file->private_data;
 670         struct fuse_conn *fc = ff->fc;
 671         struct fuse_write_in *inarg = &req->misc.write.in;
 672
 673         fuse_write_fill(req, ff, pos, count);
 674         inarg->flags = file->f_flags;
 675         if (owner != NULL) {
 676                 inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 677                 inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 678         }
 679         fuse_request_send(fc, req);
 680         return req->misc.write.out.size;
 681 }
 682
 683 static int fuse_write_begin(struct file *file, struct address_space *mapping,
 684                         loff_t pos, unsigned len, unsigned flags,
 685                         struct page **pagep, void **fsdata)
 686 {
 687         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 688
 689         *pagep = grab_cache_page_write_begin(mapping, index, flags);
 690         if (!*pagep)
 691                 return -ENOMEM;
 692         return 0;
 693 }
 694
 695 static void fuse_write_update_size(struct inode *inode, loff_t pos)
 696 {
 697         struct fuse_conn *fc = get_fuse_conn(inode);
 698         struct fuse_inode *fi = get_fuse_inode(inode);
 699
 700         spin_lock(&fc->lock);
 701         fi->attr_version = ++fc->attr_version;
 702         if (pos > inode->i_size)
 703                 i_size_write(inode, pos);
 704         spin_unlock(&fc->lock);
 705 }
 706
 707 static int fuse_buffered_write(struct file *file, struct inode *inode,
 708                                loff_t pos, unsigned count, struct page *page)
 709 {
 710         int err;
 711         size_t nres;
 712         struct fuse_conn *fc = get_fuse_conn(inode);
 713         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 714         struct fuse_req *req;
 715
 716         if (is_bad_inode(inode))
 717                 return -EIO;
 718
 719         /*
 720          * Make sure writepages on the same page are not mixed up with
 721          * plain writes.
 722          */
 723         fuse_wait_on_page_writeback(inode, page->index);
 724
 725         req = fuse_get_req(fc);
 726         if (IS_ERR(req))
 727                 return PTR_ERR(req);
 728
 729         req->in.argpages = 1;
 730         req->num_pages = 1;
 731         req->pages[0] = page;
 732         req->page_offset = offset;
 733         nres = fuse_send_write(req, file, pos, count, NULL);
 734         err = req->out.h.error;
 735         fuse_put_request(fc, req);
 736         if (!err && !nres)
 737                 err = -EIO;
 738         if (!err) {
 739                 pos += nres;
 740                 fuse_write_update_size(inode, pos);
 741                 if (count == PAGE_CACHE_SIZE)
 742                         SetPageUptodate(page);
 743         }
 744         fuse_invalidate_attr(inode);
 745         return err ? err : nres;
 746 }
 747
 748 static int fuse_write_end(struct file *file, struct address_space *mapping,
 749                         loff_t pos, unsigned len, unsigned copied,
 750                         struct page *page, void *fsdata)
 751 {
 752         struct inode *inode = mapping->host;
 753         int res = 0;
 754
 755         if (copied)
 756                 res = fuse_buffered_write(file, inode, pos, copied, page);
 757
 758         unlock_page(page);
 759         page_cache_release(page);
 760         return res;
 761 }
 762
 763 static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
 764                                     struct inode *inode, loff_t pos,
 765                                     size_t count)
 766 {
 767         size_t res;
 768         unsigned offset;
 769         unsigned i;
 770
 771         for (i = 0; i < req->num_pages; i++)
 772                 fuse_wait_on_page_writeback(inode, req->pages[i]->index);
 773
 774         res = fuse_send_write(req, file, pos, count, NULL);
 775
 776         offset = req->page_offset;
 777         count = res;
 778         for (i = 0; i < req->num_pages; i++) {
 779                 struct page *page = req->pages[i];
 780
 781                 if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
 782                         SetPageUptodate(page);
 783
 784                 if (count > PAGE_CACHE_SIZE - offset)
 785                         count -= PAGE_CACHE_SIZE - offset;
 786                 else
 787                         count = 0;
 788                 offset = 0;
 789
 790                 unlock_page(page);
 791                 page_cache_release(page);
 792         }
 793
 794         return res;
 795 }
 796
 797 static ssize_t fuse_fill_write_pages(struct fuse_req *req,
 798                                struct address_space *mapping,
 799                                struct iov_iter *ii, loff_t pos)
 800 {
 801         struct fuse_conn *fc = get_fuse_conn(mapping->host);
 802         unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 803         size_t count = 0;
 804         int err;
 805
 806         req->in.argpages = 1;
 807         req->page_offset = offset;
 808
 809         do {
 810                 size_t tmp;
 811                 struct page *page;
 812                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
 813                 size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
 814                                      iov_iter_count(ii));
 815
 816                 bytes = min_t(size_t, bytes, fc->max_write - count);
 817
 818  again:
 819                 err = -EFAULT;
 820                 if (iov_iter_fault_in_readable(ii, bytes))
 821                         break;
 822
 823                 err = -ENOMEM;
 824                 page = grab_cache_page_write_begin(mapping, index, 0);
 825                 if (!page)
 826                         break;
 827
 828                 pagefault_disable();
 829                 tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
 830                 pagefault_enable();
 831                 flush_dcache_page(page);
 832
 833                 if (!tmp) {
 834                         unlock_page(page);
 835                         page_cache_release(page);
 836                         bytes = min(bytes, iov_iter_single_seg_count(ii));
 837                         goto again;
 838                 }
 839
 840                 err = 0;
 841                 req->pages[req->num_pages] = page;
 842                 req->num_pages++;
 843
 844                 iov_iter_advance(ii, tmp);
 845                 count += tmp;
 846                 pos += tmp;
 847                 offset += tmp;
 848                 if (offset == PAGE_CACHE_SIZE)
 849                         offset = 0;
 850
 851                 if (!fc->big_writes)
 852                         break;
 853         } while (iov_iter_count(ii) && count < fc->max_write &&
 854                  req->num_pages < FUSE_MAX_PAGES_PER_REQ && offset == 0);
 855
 856         return count > 0 ? count : err;
 857 }
 858
 859 static ssize_t fuse_perform_write(struct file *file,
 860                                   struct address_space *mapping,
 861                                   struct iov_iter *ii, loff_t pos)
 862 {
 863         struct inode *inode = mapping->host;
 864         struct fuse_conn *fc = get_fuse_conn(inode);
 865         int err = 0;
 866         ssize_t res = 0;
 867
 868         if (is_bad_inode(inode))
 869                 return -EIO;
 870
 871         do {
 872                 struct fuse_req *req;
 873                 ssize_t count;
 874
 875                 req = fuse_get_req(fc);
 876                 if (IS_ERR(req)) {
 877                         err = PTR_ERR(req);
 878                         break;
 879                 }
 880
 881                 count = fuse_fill_write_pages(req, mapping, ii, pos);
 882                 if (count <= 0) {
 883                         err = count;
 884                 } else {
 885                         size_t num_written;
 886
 887                         num_written = fuse_send_write_pages(req, file, inode,
 888                                                             pos, count);
 889                         err = req->out.h.error;
 890                         if (!err) {
 891                                 res += num_written;
 892                                 pos += num_written;
 893
 894                                 /* break out of the loop on short write */
 895                                 if (num_written != count)
 896                                         err = -EIO;
 897                         }
 898                 }
 899                 fuse_put_request(fc, req);
 900         } while (!err && iov_iter_count(ii));
 901
 902         if (res > 0)
 903                 fuse_write_update_size(inode, pos);
 904
 905         fuse_invalidate_attr(inode);
 906
 907         return res > 0 ? res : err;
 908 }
 909
 910 static ssize_t fuse_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 911                                    unsigned long nr_segs, loff_t pos)
 912 {
 913         struct file *file = iocb->ki_filp;
 914         struct address_space *mapping = file->f_mapping;
 915         size_t count = 0;
 916         ssize_t written = 0;
 917         struct inode *inode = mapping->host;
 918         ssize_t err;
 919         struct iov_iter i;
 920
 921         WARN_ON(iocb->ki_pos != pos);
 922
 923         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
 924         if (err)
 925                 return err;
 926
 927         mutex_lock(&inode->i_mutex);
 928         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
 929
 930         /* We can write back this queue in page reclaim */
 931         current->backing_dev_info = mapping->backing_dev_info;
 932
 933         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
 934         if (err)
 935                 goto out;
 936
 937         if (count == 0)
 938                 goto out;
 939
 940         err = file_remove_suid(file);
 941         if (err)
 942                 goto out;
 943
 944         file_update_time(file);
 945
 946         iov_iter_init(&i, iov, nr_segs, count, 0);
 947         written = fuse_perform_write(file, mapping, &i, pos);
 948         if (written >= 0)
 949                 iocb->ki_pos = pos + written;
 950
 951 out:
 952         current->backing_dev_info = NULL;
 953         mutex_unlock(&inode->i_mutex);
 954
 955         return written ? written : err;
 956 }
 957
 958 static void fuse_release_user_pages(struct fuse_req *req, int write)
 959 {
 960         unsigned i;
 961
 962         for (i = 0; i < req->num_pages; i++) {
 963                 struct page *page = req->pages[i];
 964                 if (write)
 965                         set_page_dirty_lock(page);
 966                 put_page(page);
 967         }
 968 }
 969
 970 static int fuse_get_user_pages(struct fuse_req *req, const char __user *buf,
 971                                size_t *nbytesp, int write)
 972 {
 973         size_t nbytes = *nbytesp;
 974         unsigned long user_addr = (unsigned long) buf;
 975         unsigned offset = user_addr & ~PAGE_MASK;
 976         int npages;
 977
 978         /* Special case for kernel I/O: can copy directly into the buffer */
 979         if (segment_eq(get_fs(), KERNEL_DS)) {
 980                 if (write)
 981                         req->in.args[1].value = (void *) user_addr;
 982                 else
 983                         req->out.args[0].value = (void *) user_addr;
 984
 985                 return 0;
 986         }
 987
 988         nbytes = min_t(size_t, nbytes, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
 989         npages = (nbytes + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 990         npages = clamp(npages, 1, FUSE_MAX_PAGES_PER_REQ);
 991         down_read(&current->mm->mmap_sem);
 992         npages = get_user_pages(current, current->mm, user_addr, npages, !write,
 993                                 0, req->pages, NULL);
 994         up_read(&current->mm->mmap_sem);
 995         if (npages < 0)
 996                 return npages;
 997
 998         req->num_pages = npages;
 999         req->page_offset = offset;
1000
1001         if (write)
1002                 req->in.argpages = 1;
1003         else
1004                 req->out.argpages = 1;
1005
1006         nbytes = (req->num_pages << PAGE_SHIFT) - req->page_offset;
1007         *nbytesp = min(*nbytesp, nbytes);
1008
1009         return 0;
1010 }
1011
1012 static ssize_t fuse_direct_io(struct file *file, const char __user *buf,
1013                               size_t count, loff_t *ppos, int write)
1014 {
1015         struct fuse_file *ff = file->private_data;
1016         struct fuse_conn *fc = ff->fc;
1017         size_t nmax = write ? fc->max_write : fc->max_read;
1018         loff_t pos = *ppos;
1019         ssize_t res = 0;
1020         struct fuse_req *req;
1021
1022         req = fuse_get_req(fc);
1023         if (IS_ERR(req))
1024                 return PTR_ERR(req);
1025
1026         while (count) {
1027                 size_t nres;
1028                 fl_owner_t owner = current->files;
1029                 size_t nbytes = min(count, nmax);
1030                 int err = fuse_get_user_pages(req, buf, &nbytes, write);
1031                 if (err) {
1032                         res = err;
1033                         break;
1034                 }
1035
1036                 if (write)
1037                         nres = fuse_send_write(req, file, pos, nbytes, owner);
1038                 else
1039                         nres = fuse_send_read(req, file, pos, nbytes, owner);
1040
1041                 fuse_release_user_pages(req, !write);
1042                 if (req->out.h.error) {
1043                         if (!res)
1044                                 res = req->out.h.error;
1045                         break;
1046                 } else if (nres > nbytes) {
1047                         res = -EIO;
1048                         break;
1049                 }
1050                 count -= nres;
1051                 res += nres;
1052                 pos += nres;
1053                 buf += nres;
1054                 if (nres != nbytes)
1055                         break;
1056                 if (count) {
1057                         fuse_put_request(fc, req);
1058                         req = fuse_get_req(fc);
1059                         if (IS_ERR(req))
1060                                 break;
1061                 }
1062         }
1063         fuse_put_request(fc, req);
1064         if (res > 0)
1065                 *ppos = pos;
1066
1067         return res;
1068 }
1069
1070 static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1071                                      size_t count, loff_t *ppos)
1072 {
1073         ssize_t res;
1074         struct inode *inode = file->f_path.dentry->d_inode;
1075
1076         if (is_bad_inode(inode))
1077                 return -EIO;
1078
1079         res = fuse_direct_io(file, buf, count, ppos, 0);
1080
1081         fuse_invalidate_attr(inode);
1082
1083         return res;
1084 }
1085
1086 static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1087                                  size_t count, loff_t *ppos)
1088 {
1089         struct inode *inode = file->f_path.dentry->d_inode;
1090         ssize_t res;
1091
1092         if (is_bad_inode(inode))
1093                 return -EIO;
1094
1095         /* Don't allow parallel writes to the same file */
1096         mutex_lock(&inode->i_mutex);
1097         res = generic_write_checks(file, ppos, &count, 0);
1098         if (!res) {
1099                 res = fuse_direct_io(file, buf, count, ppos, 1);
1100                 if (res > 0)
1101                         fuse_write_update_size(inode, *ppos);
1102         }
1103         mutex_unlock(&inode->i_mutex);
1104
1105         fuse_invalidate_attr(inode);
1106
1107         return res;
1108 }
1109
1110 static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1111 {
1112         __free_page(req->pages[0]);
1113         fuse_file_put(req->ff);
1114 }
1115
1116 static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1117 {
1118         struct inode *inode = req->inode;
1119         struct fuse_inode *fi = get_fuse_inode(inode);
1120         struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1121
1122         list_del(&req->writepages_entry);
1123         dec_bdi_stat(bdi, BDI_WRITEBACK);
1124         dec_zone_page_state(req->pages[0], NR_WRITEBACK_TEMP);
1125         bdi_writeout_inc(bdi);
1126         wake_up(&fi->page_waitq);
1127 }
1128
1129 /* Called under fc->lock, may release and reacquire it */
1130 static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req)
1131 __releases(&fc->lock)
1132 __acquires(&fc->lock)
1133 {
1134         struct fuse_inode *fi = get_fuse_inode(req->inode);
1135         loff_t size = i_size_read(req->inode);
1136         struct fuse_write_in *inarg = &req->misc.write.in;
1137
1138         if (!fc->connected)
1139                 goto out_free;
1140
1141         if (inarg->offset + PAGE_CACHE_SIZE <= size) {
1142                 inarg->size = PAGE_CACHE_SIZE;
1143         } else if (inarg->offset < size) {
1144                 inarg->size = size & (PAGE_CACHE_SIZE - 1);
1145         } else {
1146                 /* Got truncated off completely */
1147                 goto out_free;
1148         }
1149
1150         req->in.args[1].size = inarg->size;
1151         fi->writectr++;
1152         fuse_request_send_background_locked(fc, req);
1153         return;
1154
1155  out_free:
1156         fuse_writepage_finish(fc, req);
1157         spin_unlock(&fc->lock);
1158         fuse_writepage_free(fc, req);
1159         fuse_put_request(fc, req);
1160         spin_lock(&fc->lock);
1161 }
1162
1163 /*
1164  * If fi->writectr is positive (no truncate or fsync going on) send
1165  * all queued writepage requests.
1166  *
1167  * Called with fc->lock
1168  */
1169 void fuse_flush_writepages(struct inode *inode)
1170 __releases(&fc->lock)
1171 __acquires(&fc->lock)
1172 {
1173         struct fuse_conn *fc = get_fuse_conn(inode);
1174         struct fuse_inode *fi = get_fuse_inode(inode);
1175         struct fuse_req *req;
1176
1177         while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1178                 req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1179                 list_del_init(&req->list);
1180                 fuse_send_writepage(fc, req);
1181         }
1182 }
1183
1184 static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1185 {
1186         struct inode *inode = req->inode;
1187         struct fuse_inode *fi = get_fuse_inode(inode);
1188
1189         mapping_set_error(inode->i_mapping, req->out.h.error);
1190         spin_lock(&fc->lock);
1191         fi->writectr--;
1192         fuse_writepage_finish(fc, req);
1193         spin_unlock(&fc->lock);
1194         fuse_writepage_free(fc, req);
1195 }
1196
1197 static int fuse_writepage_locked(struct page *page)
1198 {
1199         struct address_space *mapping = page->mapping;
1200         struct inode *inode = mapping->host;
1201         struct fuse_conn *fc = get_fuse_conn(inode);
1202         struct fuse_inode *fi = get_fuse_inode(inode);
1203         struct fuse_req *req;
1204         struct fuse_file *ff;
1205         struct page *tmp_page;
1206
1207         set_page_writeback(page);
1208
1209         req = fuse_request_alloc_nofs();
1210         if (!req)
1211                 goto err;
1212
1213         tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1214         if (!tmp_page)
1215                 goto err_free;
1216
1217         spin_lock(&fc->lock);
1218         BUG_ON(list_empty(&fi->write_files));
1219         ff = list_entry(fi->write_files.next, struct fuse_file, write_entry);
1220         req->ff = fuse_file_get(ff);
1221         spin_unlock(&fc->lock);
1222
1223         fuse_write_fill(req, ff, page_offset(page), 0);
1224
1225         copy_highpage(tmp_page, page);
1226         req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1227         req->in.argpages = 1;
1228         req->num_pages = 1;
1229         req->pages[0] = tmp_page;
1230         req->page_offset = 0;
1231         req->end = fuse_writepage_end;
1232         req->inode = inode;
1233
1234         inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1235         inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1236         end_page_writeback(page);
1237
1238         spin_lock(&fc->lock);
1239         list_add(&req->writepages_entry, &fi->writepages);
1240         list_add_tail(&req->list, &fi->queued_writes);
1241         fuse_flush_writepages(inode);
1242         spin_unlock(&fc->lock);
1243
1244         return 0;
1245
1246 err_free:
1247         fuse_request_free(req);
1248 err:
1249         end_page_writeback(page);
1250         return -ENOMEM;
1251 }
1252
1253 static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1254 {
1255         int err;
1256
1257         err = fuse_writepage_locked(page);
1258         unlock_page(page);
1259
1260         return err;
1261 }
1262
1263 static int fuse_launder_page(struct page *page)
1264 {
1265         int err = 0;
1266         if (clear_page_dirty_for_io(page)) {
1267                 struct inode *inode = page->mapping->host;
1268                 err = fuse_writepage_locked(page);
1269                 if (!err)
1270                         fuse_wait_on_page_writeback(inode, page->index);
1271         }
1272         return err;
1273 }
1274
1275 /*
1276  * Write back dirty pages now, because there may not be any suitable
1277  * open files later
1278  */
1279 static void fuse_vma_close(struct vm_area_struct *vma)
1280 {
1281         filemap_write_and_wait(vma->vm_file->f_mapping);
1282 }
1283
1284 /*
1285  * Wait for writeback against this page to complete before allowing it
1286  * to be marked dirty again, and hence written back again, possibly
1287  * before the previous writepage completed.
1288  *
1289  * Block here, instead of in ->writepage(), so that the userspace fs
1290  * can only block processes actually operating on the filesystem.
1291  *
1292  * Otherwise unprivileged userspace fs would be able to block
1293  * unrelated:
1294  *
1295  * - page migration
1296  * - sync(2)
1297  * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
1298  */
1299 static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1300 {
1301         struct page *page = vmf->page;
1302         /*
1303          * Don't use page->mapping as it may become NULL from a
1304          * concurrent truncate.
1305          */
1306         struct inode *inode = vma->vm_file->f_mapping->host;
1307
1308         fuse_wait_on_page_writeback(inode, page->index);
1309         return 0;
1310 }
1311
1312 static struct vm_operations_struct fuse_file_vm_ops = {
1313         .close          = fuse_vma_close,
1314         .fault          = filemap_fault,
1315         .page_mkwrite   = fuse_page_mkwrite,
1316 };
1317
1318 static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
1319 {
1320         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) {
1321                 struct inode *inode = file->f_dentry->d_inode;
1322                 struct fuse_conn *fc = get_fuse_conn(inode);
1323                 struct fuse_inode *fi = get_fuse_inode(inode);
1324                 struct fuse_file *ff = file->private_data;
1325                 /*
1326                  * file may be written through mmap, so chain it onto the
1327                  * inodes's write_file list
1328                  */
1329                 spin_lock(&fc->lock);
1330                 if (list_empty(&ff->write_entry))
1331                         list_add(&ff->write_entry, &fi->write_files);
1332                 spin_unlock(&fc->lock);
1333         }
1334         file_accessed(file);
1335         vma->vm_ops = &fuse_file_vm_ops;
1336         return 0;
1337 }
1338
1339 static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
1340 {
1341         /* Can't provide the coherency needed for MAP_SHARED */
1342         if (vma->vm_flags & VM_MAYSHARE)
1343                 return -ENODEV;
1344
1345         invalidate_inode_pages2(file->f_mapping);
1346
1347         return generic_file_mmap(file, vma);
1348 }
1349
1350 static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
1351                                   struct file_lock *fl)
1352 {
1353         switch (ffl->type) {
1354         case F_UNLCK:
1355                 break;
1356
1357         case F_RDLCK:
1358         case F_WRLCK:
1359                 if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
1360                     ffl->end < ffl->start)
1361                         return -EIO;
1362
1363                 fl->fl_start = ffl->start;
1364                 fl->fl_end = ffl->end;
1365                 fl->fl_pid = ffl->pid;
1366                 break;
1367
1368         default:
1369                 return -EIO;
1370         }
1371         fl->fl_type = ffl->type;
1372         return 0;
1373 }
1374
1375 static void fuse_lk_fill(struct fuse_req *req, struct file *file,
1376                          const struct file_lock *fl, int opcode, pid_t pid,
1377                          int flock)
1378 {
1379         struct inode *inode = file->f_path.dentry->d_inode;
1380         struct fuse_conn *fc = get_fuse_conn(inode);
1381         struct fuse_file *ff = file->private_data;
1382         struct fuse_lk_in *arg = &req->misc.lk_in;
1383
1384         arg->fh = ff->fh;
1385         arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
1386         arg->lk.start = fl->fl_start;
1387         arg->lk.end = fl->fl_end;
1388         arg->lk.type = fl->fl_type;
1389         arg->lk.pid = pid;
1390         if (flock)
1391                 arg->lk_flags |= FUSE_LK_FLOCK;
1392         req->in.h.opcode = opcode;
1393         req->in.h.nodeid = get_node_id(inode);
1394         req->in.numargs = 1;
1395         req->in.args[0].size = sizeof(*arg);
1396         req->in.args[0].value = arg;
1397 }
1398
1399 static int fuse_getlk(struct file *file, struct file_lock *fl)
1400 {
1401         struct inode *inode = file->f_path.dentry->d_inode;
1402         struct fuse_conn *fc = get_fuse_conn(inode);
1403         struct fuse_req *req;
1404         struct fuse_lk_out outarg;
1405         int err;
1406
1407         req = fuse_get_req(fc);
1408         if (IS_ERR(req))
1409                 return PTR_ERR(req);
1410
1411         fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
1412         req->out.numargs = 1;
1413         req->out.args[0].size = sizeof(outarg);
1414         req->out.args[0].value = &outarg;
1415         fuse_request_send(fc, req);
1416         err = req->out.h.error;
1417         fuse_put_request(fc, req);
1418         if (!err)
1419                 err = convert_fuse_file_lock(&outarg.lk, fl);
1420
1421         return err;
1422 }
1423
1424 static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
1425 {
1426         struct inode *inode = file->f_path.dentry->d_inode;
1427         struct fuse_conn *fc = get_fuse_conn(inode);
1428         struct fuse_req *req;
1429         int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
1430         pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
1431         int err;
1432
1433         if (fl->fl_lmops && fl->fl_lmops->fl_grant) {
1434                 /* NLM needs asynchronous locks, which we don't support yet */
1435                 return -ENOLCK;
1436         }
1437
1438         /* Unlock on close is handled by the flush method */
1439         if (fl->fl_flags & FL_CLOSE)
1440                 return 0;
1441
1442         req = fuse_get_req(fc);
1443         if (IS_ERR(req))
1444                 return PTR_ERR(req);
1445
1446         fuse_lk_fill(req, file, fl, opcode, pid, flock);
1447         fuse_request_send(fc, req);
1448         err = req->out.h.error;
1449         /* locking is restartable */
1450         if (err == -EINTR)
1451                 err = -ERESTARTSYS;
1452         fuse_put_request(fc, req);
1453         return err;
1454 }
1455
1456 static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
1457 {
1458         struct inode *inode = file->f_path.dentry->d_inode;
1459         struct fuse_conn *fc = get_fuse_conn(inode);
1460         int err;
1461
1462         if (cmd == F_CANCELLK) {
1463                 err = 0;
1464         } else if (cmd == F_GETLK) {
1465                 if (fc->no_lock) {
1466                         posix_test_lock(file, fl);
1467                         err = 0;
1468                 } else
1469                         err = fuse_getlk(file, fl);
1470         } else {
1471                 if (fc->no_lock)
1472                         err = posix_lock_file(file, fl, NULL);
1473                 else
1474                         err = fuse_setlk(file, fl, 0);
1475         }
1476         return err;
1477 }
1478
1479 static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
1480 {
1481         struct inode *inode = file->f_path.dentry->d_inode;
1482         struct fuse_conn *fc = get_fuse_conn(inode);
1483         int err;
1484
1485         if (fc->no_lock) {
1486                 err = flock_lock_file_wait(file, fl);
1487         } else {
1488                 /* emulate flock with POSIX locks */
1489                 fl->fl_owner = (fl_owner_t) file;
1490                 err = fuse_setlk(file, fl, 1);
1491         }
1492
1493         return err;
1494 }
1495
1496 static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
1497 {
1498         struct inode *inode = mapping->host;
1499         struct fuse_conn *fc = get_fuse_conn(inode);
1500         struct fuse_req *req;
1501         struct fuse_bmap_in inarg;
1502         struct fuse_bmap_out outarg;
1503         int err;
1504
1505         if (!inode->i_sb->s_bdev || fc->no_bmap)
1506                 return 0;
1507
1508         req = fuse_get_req(fc);
1509         if (IS_ERR(req))
1510                 return 0;
1511
1512         memset(&inarg, 0, sizeof(inarg));
1513         inarg.block = block;
1514         inarg.blocksize = inode->i_sb->s_blocksize;
1515         req->in.h.opcode = FUSE_BMAP;
1516         req->in.h.nodeid = get_node_id(inode);
1517         req->in.numargs = 1;
1518         req->in.args[0].size = sizeof(inarg);
1519         req->in.args[0].value = &inarg;
1520         req->out.numargs = 1;
1521         req->out.args[0].size = sizeof(outarg);
1522         req->out.args[0].value = &outarg;
1523         fuse_request_send(fc, req);
1524         err = req->out.h.error;
1525         fuse_put_request(fc, req);
1526         if (err == -ENOSYS)
1527                 fc->no_bmap = 1;
1528
1529         return err ? 0 : outarg.block;
1530 }
1531
1532 static loff_t fuse_file_llseek(struct file *file, loff_t offset, int origin)
1533 {
1534         loff_t retval;
1535         struct inode *inode = file->f_path.dentry->d_inode;
1536
1537         mutex_lock(&inode->i_mutex);
1538         switch (origin) {
1539         case SEEK_END:
1540                 retval = fuse_update_attributes(inode, NULL, file, NULL);
1541                 if (retval)
1542                         goto exit;
1543                 offset += i_size_read(inode);
1544                 break;
1545         case SEEK_CUR:
1546                 offset += file->f_pos;
1547         }
1548         retval = -EINVAL;
1549         if (offset >= 0 && offset <= inode->i_sb->s_maxbytes) {
1550                 if (offset != file->f_pos) {
1551                         file->f_pos = offset;
1552                         file->f_version = 0;
1553                 }
1554                 retval = offset;
1555         }
1556 exit:
1557         mutex_unlock(&inode->i_mutex);
1558         return retval;
1559 }
1560
1561 static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
1562                         unsigned int nr_segs, size_t bytes, bool to_user)
1563 {
1564         struct iov_iter ii;
1565         int page_idx = 0;
1566
1567         if (!bytes)
1568                 return 0;
1569
1570         iov_iter_init(&ii, iov, nr_segs, bytes, 0);
1571
1572         while (iov_iter_count(&ii)) {
1573                 struct page *page = pages[page_idx++];
1574                 size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
1575                 void *kaddr, *map;
1576
1577                 kaddr = map = kmap(page);
1578
1579                 while (todo) {
1580                         char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
1581                         size_t iov_len = ii.iov->iov_len - ii.iov_offset;
1582                         size_t copy = min(todo, iov_len);
1583                         size_t left;
1584
1585                         if (!to_user)
1586                                 left = copy_from_user(kaddr, uaddr, copy);
1587                         else
1588                                 left = copy_to_user(uaddr, kaddr, copy);
1589
1590                         if (unlikely(left))
1591                                 return -EFAULT;
1592
1593                         iov_iter_advance(&ii, copy);
1594                         todo -= copy;
1595                         kaddr += copy;
1596                 }
1597
1598                 kunmap(map);
1599         }
1600
1601         return 0;
1602 }
1603
1604 /*
1605  * For ioctls, there is no generic way to determine how much memory
1606  * needs to be read and/or written.  Furthermore, ioctls are allowed
1607  * to dereference the passed pointer, so the parameter requires deep
1608  * copying but FUSE has no idea whatsoever about what to copy in or
1609  * out.
1610  *
1611  * This is solved by allowing FUSE server to retry ioctl with
1612  * necessary in/out iovecs.  Let's assume the ioctl implementation
1613  * needs to read in the following structure.
1614  *
1615  * struct a {
1616  *      char    *buf;
1617  *      size_t  buflen;
1618  * }
1619  *
1620  * On the first callout to FUSE server, inarg->in_size and
1621  * inarg->out_size will be NULL; then, the server completes the ioctl
1622  * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
1623  * the actual iov array to
1624  *
1625  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
1626  *
1627  * which tells FUSE to copy in the requested area and retry the ioctl.
1628  * On the second round, the server has access to the structure and
1629  * from that it can tell what to look for next, so on the invocation,
1630  * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
1631  *
1632  * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
1633  *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
1634  *
1635  * FUSE will copy both struct a and the pointed buffer from the
1636  * process doing the ioctl and retry ioctl with both struct a and the
1637  * buffer.
1638  *
1639  * This time, FUSE server has everything it needs and completes ioctl
1640  * without FUSE_IOCTL_RETRY which finishes the ioctl call.
1641  *
1642  * Copying data out works the same way.
1643  *
1644  * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
1645  * automatically initializes in and out iovs by decoding @cmd with
1646  * _IOC_* macros and the server is not allowed to request RETRY.  This
1647  * limits ioctl data transfers to well-formed ioctls and is the forced
1648  * behavior for all FUSE servers.
1649  */
1650 static long fuse_do_ioctl(struct file *file, unsigned int cmd,
1651                           unsigned long arg, unsigned int flags)
1652 {
1653         struct fuse_file *ff = file->private_data;
1654         struct fuse_conn *fc = ff->fc;
1655         struct fuse_ioctl_in inarg = {
1656                 .fh = ff->fh,
1657                 .cmd = cmd,
1658                 .arg = arg,
1659                 .flags = flags
1660         };
1661         struct fuse_ioctl_out outarg;
1662         struct fuse_req *req = NULL;
1663         struct page **pages = NULL;
1664         struct page *iov_page = NULL;
1665         struct iovec *in_iov = NULL, *out_iov = NULL;
1666         unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
1667         size_t in_size, out_size, transferred;
1668         int err;
1669
1670         /* assume all the iovs returned by client always fits in a page */
1671         BUILD_BUG_ON(sizeof(struct iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
1672
1673         err = -ENOMEM;
1674         pages = kzalloc(sizeof(pages[0]) * FUSE_MAX_PAGES_PER_REQ, GFP_KERNEL);
1675         iov_page = alloc_page(GFP_KERNEL);
1676         if (!pages || !iov_page)
1677                 goto out;
1678
1679         /*
1680          * If restricted, initialize IO parameters as encoded in @cmd.
1681          * RETRY from server is not allowed.
1682          */
1683         if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
1684                 struct iovec *iov = page_address(iov_page);
1685
1686                 iov->iov_base = (void __user *)arg;
1687                 iov->iov_len = _IOC_SIZE(cmd);
1688
1689                 if (_IOC_DIR(cmd) & _IOC_WRITE) {
1690                         in_iov = iov;
1691                         in_iovs = 1;
1692                 }
1693
1694                 if (_IOC_DIR(cmd) & _IOC_READ) {
1695                         out_iov = iov;
1696                         out_iovs = 1;
1697                 }
1698         }
1699
1700  retry:
1701         inarg.in_size = in_size = iov_length(in_iov, in_iovs);
1702         inarg.out_size = out_size = iov_length(out_iov, out_iovs);
1703
1704         /*
1705          * Out data can be used either for actual out data or iovs,
1706          * make sure there always is at least one page.
1707          */
1708         out_size = max_t(size_t, out_size, PAGE_SIZE);
1709         max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
1710
1711         /* make sure there are enough buffer pages and init request with them */
1712         err = -ENOMEM;
1713         if (max_pages > FUSE_MAX_PAGES_PER_REQ)
1714                 goto out;
1715         while (num_pages < max_pages) {
1716                 pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
1717                 if (!pages[num_pages])
1718                         goto out;
1719                 num_pages++;
1720         }
1721
1722         req = fuse_get_req(fc);
1723         if (IS_ERR(req)) {
1724                 err = PTR_ERR(req);
1725                 req = NULL;
1726                 goto out;
1727         }
1728         memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
1729         req->num_pages = num_pages;
1730
1731         /* okay, let's send it to the client */
1732         req->in.h.opcode = FUSE_IOCTL;
1733         req->in.h.nodeid = ff->nodeid;
1734         req->in.numargs = 1;
1735         req->in.args[0].size = sizeof(inarg);
1736         req->in.args[0].value = &inarg;
1737         if (in_size) {
1738                 req->in.numargs++;
1739                 req->in.args[1].size = in_size;
1740                 req->in.argpages = 1;
1741
1742                 err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
1743                                            false);
1744                 if (err)
1745                         goto out;
1746         }
1747
1748         req->out.numargs = 2;
1749         req->out.args[0].size = sizeof(outarg);
1750         req->out.args[0].value = &outarg;
1751         req->out.args[1].size = out_size;
1752         req->out.argpages = 1;
1753         req->out.argvar = 1;
1754
1755         fuse_request_send(fc, req);
1756         err = req->out.h.error;
1757         transferred = req->out.args[1].size;
1758         fuse_put_request(fc, req);
1759         req = NULL;
1760         if (err)
1761                 goto out;
1762
1763         /* did it ask for retry? */
1764         if (outarg.flags & FUSE_IOCTL_RETRY) {
1765                 char *vaddr;
1766
1767                 /* no retry if in restricted mode */
1768                 err = -EIO;
1769                 if (!(flags & FUSE_IOCTL_UNRESTRICTED))
1770                         goto out;
1771
1772                 in_iovs = outarg.in_iovs;
1773                 out_iovs = outarg.out_iovs;
1774
1775                 /*
1776                  * Make sure things are in boundary, separate checks
1777                  * are to protect against overflow.
1778                  */
1779                 err = -ENOMEM;
1780                 if (in_iovs > FUSE_IOCTL_MAX_IOV ||
1781                     out_iovs > FUSE_IOCTL_MAX_IOV ||
1782                     in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
1783                         goto out;
1784
1785                 err = -EIO;
1786                 if ((in_iovs + out_iovs) * sizeof(struct iovec) != transferred)
1787                         goto out;
1788
1789                 /* okay, copy in iovs and retry */
1790                 vaddr = kmap_atomic(pages[0], KM_USER0);
1791                 memcpy(page_address(iov_page), vaddr, transferred);
1792                 kunmap_atomic(vaddr, KM_USER0);
1793
1794                 in_iov = page_address(iov_page);
1795                 out_iov = in_iov + in_iovs;
1796
1797                 goto retry;
1798         }
1799
1800         err = -EIO;
1801         if (transferred > inarg.out_size)
1802                 goto out;
1803
1804         err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
1805  out:
1806         if (req)
1807                 fuse_put_request(fc, req);
1808         if (iov_page)
1809                 __free_page(iov_page);
1810         while (num_pages)
1811                 __free_page(pages[--num_pages]);
1812         kfree(pages);
1813
1814         return err ? err : outarg.result;
1815 }
1816
1817 static long fuse_file_ioctl_common(struct file *file, unsigned int cmd,
1818                                    unsigned long arg, unsigned int flags)
1819 {
1820         struct inode *inode = file->f_dentry->d_inode;
1821         struct fuse_conn *fc = get_fuse_conn(inode);
1822
1823         if (!fuse_allow_task(fc, current))
1824                 return -EACCES;
1825
1826         if (is_bad_inode(inode))
1827                 return -EIO;
1828
1829         return fuse_do_ioctl(file, cmd, arg, flags);
1830 }
1831
1832 static long fuse_file_ioctl(struct file *file, unsigned int cmd,
1833                             unsigned long arg)
1834 {
1835         return fuse_file_ioctl_common(file, cmd, arg, 0);
1836 }
1837
1838 static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
1839                                    unsigned long arg)
1840 {
1841         return fuse_file_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
1842 }
1843
1844 /*
1845  * All files which have been polled are linked to RB tree
1846  * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
1847  * find the matching one.
1848  */
1849 static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
1850                                               struct rb_node **parent_out)
1851 {
1852         struct rb_node **link = &fc->polled_files.rb_node;
1853         struct rb_node *last = NULL;
1854
1855         while (*link) {
1856                 struct fuse_file *ff;
1857
1858                 last = *link;
1859                 ff = rb_entry(last, struct fuse_file, polled_node);
1860
1861                 if (kh < ff->kh)
1862                         link = &last->rb_left;
1863                 else if (kh > ff->kh)
1864                         link = &last->rb_right;
1865                 else
1866                         return link;
1867         }
1868
1869         if (parent_out)
1870                 *parent_out = last;
1871         return link;
1872 }
1873
1874 /*
1875  * The file is about to be polled.  Make sure it's on the polled_files
1876  * RB tree.  Note that files once added to the polled_files tree are
1877  * not removed before the file is released.  This is because a file
1878  * polled once is likely to be polled again.
1879  */
1880 static void fuse_register_polled_file(struct fuse_conn *fc,
1881                                       struct fuse_file *ff)
1882 {
1883         spin_lock(&fc->lock);
1884         if (RB_EMPTY_NODE(&ff->polled_node)) {
1885                 struct rb_node **link, *parent;
1886
1887                 link = fuse_find_polled_node(fc, ff->kh, &parent);
1888                 BUG_ON(*link);
1889                 rb_link_node(&ff->polled_node, parent, link);
1890                 rb_insert_color(&ff->polled_node, &fc->polled_files);
1891         }
1892         spin_unlock(&fc->lock);
1893 }
1894
1895 static unsigned fuse_file_poll(struct file *file, poll_table *wait)
1896 {
1897         struct fuse_file *ff = file->private_data;
1898         struct fuse_conn *fc = ff->fc;
1899         struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
1900         struct fuse_poll_out outarg;
1901         struct fuse_req *req;
1902         int err;
1903
1904         if (fc->no_poll)
1905                 return DEFAULT_POLLMASK;
1906
1907         poll_wait(file, &ff->poll_wait, wait);
1908
1909         /*
1910          * Ask for notification iff there's someone waiting for it.
1911          * The client may ignore the flag and always notify.
1912          */
1913         if (waitqueue_active(&ff->poll_wait)) {
1914                 inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
1915                 fuse_register_polled_file(fc, ff);
1916         }
1917
1918         req = fuse_get_req(fc);
1919         if (IS_ERR(req))
1920                 return PTR_ERR(req);
1921
1922         req->in.h.opcode = FUSE_POLL;
1923         req->in.h.nodeid = ff->nodeid;
1924         req->in.numargs = 1;
1925         req->in.args[0].size = sizeof(inarg);
1926         req->in.args[0].value = &inarg;
1927         req->out.numargs = 1;
1928         req->out.args[0].size = sizeof(outarg);
1929         req->out.args[0].value = &outarg;
1930         fuse_request_send(fc, req);
1931         err = req->out.h.error;
1932         fuse_put_request(fc, req);
1933
1934         if (!err)
1935                 return outarg.revents;
1936         if (err == -ENOSYS) {
1937                 fc->no_poll = 1;
1938                 return DEFAULT_POLLMASK;
1939         }
1940         return POLLERR;
1941 }
1942
1943 /*
1944  * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
1945  * wakes up the poll waiters.
1946  */
1947 int fuse_notify_poll_wakeup(struct fuse_conn *fc,
1948                             struct fuse_notify_poll_wakeup_out *outarg)
1949 {
1950         u64 kh = outarg->kh;
1951         struct rb_node **link;
1952
1953         spin_lock(&fc->lock);
1954
1955         link = fuse_find_polled_node(fc, kh, NULL);
1956         if (*link) {
1957                 struct fuse_file *ff;
1958
1959                 ff = rb_entry(*link, struct fuse_file, polled_node);
1960                 wake_up_interruptible_sync(&ff->poll_wait);
1961         }
1962
1963         spin_unlock(&fc->lock);
1964         return 0;
1965 }
1966
1967 static const struct file_operations fuse_file_operations = {
1968         .llseek         = fuse_file_llseek,
1969         .read           = do_sync_read,
1970         .aio_read       = fuse_file_aio_read,
1971         .write          = do_sync_write,
1972         .aio_write      = fuse_file_aio_write,
1973         .mmap           = fuse_file_mmap,
1974         .open           = fuse_open,
1975         .flush          = fuse_flush,
1976         .release        = fuse_release,
1977         .fsync          = fuse_fsync,
1978         .lock           = fuse_file_lock,
1979         .flock          = fuse_file_flock,
1980         .splice_read    = generic_file_splice_read,
1981         .unlocked_ioctl = fuse_file_ioctl,
1982         .compat_ioctl   = fuse_file_compat_ioctl,
1983         .poll           = fuse_file_poll,
1984 };
1985
1986 static const struct file_operations fuse_direct_io_file_operations = {
1987         .llseek         = fuse_file_llseek,
1988         .read           = fuse_direct_read,
1989         .write          = fuse_direct_write,
1990         .mmap           = fuse_direct_mmap,
1991         .open           = fuse_open,
1992         .flush          = fuse_flush,
1993         .release        = fuse_release,
1994         .fsync          = fuse_fsync,
1995         .lock           = fuse_file_lock,
1996         .flock          = fuse_file_flock,
1997         .unlocked_ioctl = fuse_file_ioctl,
1998         .compat_ioctl   = fuse_file_compat_ioctl,
1999         .poll           = fuse_file_poll,
2000         /* no splice_read */
2001 };
2002
2003 static const struct address_space_operations fuse_file_aops  = {
2004         .readpage       = fuse_readpage,
2005         .writepage      = fuse_writepage,
2006         .launder_page   = fuse_launder_page,
2007         .write_begin    = fuse_write_begin,
2008         .write_end      = fuse_write_end,
2009         .readpages      = fuse_readpages,
2010         .set_page_dirty = __set_page_dirty_nobuffers,
2011         .bmap           = fuse_bmap,
2012 };
2013
2014 void fuse_init_file_inode(struct inode *inode)
2015 {
2016         inode->i_fop = &fuse_file_operations;
2017         inode->i_data.a_ops = &fuse_file_aops;
2018 }