SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/splice.c

   1 /*
   2  * "splice": joining two ropes together by interweaving their strands.
   3  *
   4  * This is the "extended pipe" functionality, where a pipe is used as
   5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
   6  * buffer that you can use to transfer data from one end to the other.
   7  *
   8  * The traditional unix read/write is extended with a "splice()" operation
   9  * that transfers data buffers to or from a pipe buffer.
  10  *
  11  * Named by Larry McVoy, original implementation from Linus, extended by
  12  * Jens to support splicing to files and fixing the initial implementation
  13  * bugs.
  14  *
  15  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
  16  * Copyright (C) 2005 Linus Torvalds <torvalds@osdl.org>
  17  *
  18  */
  19 #include <linux/fs.h>
  20 #include <linux/file.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/pipe_fs_i.h>
  23 #include <linux/mm_inline.h>
  24 #include <linux/swap.h>
  25
  26 /*
  27  * Passed to the actors
  28  */
  29 struct splice_desc {
  30         unsigned int len, total_len;    /* current and remaining length */
  31         unsigned int flags;             /* splice flags */
  32         struct file *file;              /* file to read/write */
  33         loff_t pos;                     /* file position */
  34 };
  35
  36 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  37                                      struct pipe_buffer *buf)
  38 {
  39         struct page *page = buf->page;
  40
  41         WARN_ON(!PageLocked(page));
  42         WARN_ON(!PageUptodate(page));
  43
  44         if (!remove_mapping(page_mapping(page), page))
  45                 return 1;
  46
  47         if (PageLRU(page)) {
  48                 struct zone *zone = page_zone(page);
  49
  50                 spin_lock_irq(&zone->lru_lock);
  51                 BUG_ON(!PageLRU(page));
  52                 __ClearPageLRU(page);
  53                 del_page_from_lru(zone, page);
  54                 spin_unlock_irq(&zone->lru_lock);
  55         }
  56
  57         buf->stolen = 1;
  58         return 0;
  59 }
  60
  61 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  62                                         struct pipe_buffer *buf)
  63 {
  64         page_cache_release(buf->page);
  65         buf->page = NULL;
  66         buf->stolen = 0;
  67 }
  68
  69 static void *page_cache_pipe_buf_map(struct file *file,
  70                                      struct pipe_inode_info *info,
  71                                      struct pipe_buffer *buf)
  72 {
  73         struct page *page = buf->page;
  74
  75         lock_page(page);
  76
  77         if (!PageUptodate(page)) {
  78                 unlock_page(page);
  79                 return ERR_PTR(-EIO);
  80         }
  81
  82         if (!page->mapping) {
  83                 unlock_page(page);
  84                 return ERR_PTR(-ENODATA);
  85         }
  86
  87         return kmap(buf->page);
  88 }
  89
  90 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
  91                                       struct pipe_buffer *buf)
  92 {
  93         if (!buf->stolen)
  94                 unlock_page(buf->page);
  95         kunmap(buf->page);
  96 }
  97
  98 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
  99         .can_merge = 0,
 100         .map = page_cache_pipe_buf_map,
 101         .unmap = page_cache_pipe_buf_unmap,
 102         .release = page_cache_pipe_buf_release,
 103         .steal = page_cache_pipe_buf_steal,
 104 };
 105
 106 static ssize_t move_to_pipe(struct inode *inode, struct page **pages,
 107                             int nr_pages, unsigned long offset,
 108                             unsigned long len)
 109 {
 110         struct pipe_inode_info *info;
 111         int ret, do_wakeup, i;
 112
 113         ret = 0;
 114         do_wakeup = 0;
 115         i = 0;
 116
 117         mutex_lock(PIPE_MUTEX(*inode));
 118
 119         info = inode->i_pipe;
 120         for (;;) {
 121                 int bufs;
 122
 123                 if (!PIPE_READERS(*inode)) {
 124                         send_sig(SIGPIPE, current, 0);
 125                         if (!ret)
 126                                 ret = -EPIPE;
 127                         break;
 128                 }
 129
 130                 bufs = info->nrbufs;
 131                 if (bufs < PIPE_BUFFERS) {
 132                         int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS - 1);
 133                         struct pipe_buffer *buf = info->bufs + newbuf;
 134                         struct page *page = pages[i++];
 135                         unsigned long this_len;
 136
 137                         this_len = PAGE_CACHE_SIZE - offset;
 138                         if (this_len > len)
 139                                 this_len = len;
 140
 141                         buf->page = page;
 142                         buf->offset = offset;
 143                         buf->len = this_len;
 144                         buf->ops = &page_cache_pipe_buf_ops;
 145                         info->nrbufs = ++bufs;
 146                         do_wakeup = 1;
 147
 148                         ret += this_len;
 149                         len -= this_len;
 150                         offset = 0;
 151                         if (!--nr_pages)
 152                                 break;
 153                         if (!len)
 154                                 break;
 155                         if (bufs < PIPE_BUFFERS)
 156                                 continue;
 157
 158                         break;
 159                 }
 160
 161                 if (signal_pending(current)) {
 162                         if (!ret)
 163                                 ret = -ERESTARTSYS;
 164                         break;
 165                 }
 166
 167                 if (do_wakeup) {
 168                         wake_up_interruptible_sync(PIPE_WAIT(*inode));
 169                         kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO,
 170                                     POLL_IN);
 171                         do_wakeup = 0;
 172                 }
 173
 174                 PIPE_WAITING_WRITERS(*inode)++;
 175                 pipe_wait(inode);
 176                 PIPE_WAITING_WRITERS(*inode)--;
 177         }
 178
 179         mutex_unlock(PIPE_MUTEX(*inode));
 180
 181         if (do_wakeup) {
 182                 wake_up_interruptible(PIPE_WAIT(*inode));
 183                 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
 184         }
 185
 186         while (i < nr_pages)
 187                 page_cache_release(pages[i++]);
 188
 189         return ret;
 190 }
 191
 192 static int __generic_file_splice_read(struct file *in, struct inode *pipe,
 193                                       size_t len)
 194 {
 195         struct address_space *mapping = in->f_mapping;
 196         unsigned int offset, nr_pages;
 197         struct page *pages[PIPE_BUFFERS], *shadow[PIPE_BUFFERS];
 198         struct page *page;
 199         pgoff_t index, pidx;
 200         int i, j;
 201
 202         index = in->f_pos >> PAGE_CACHE_SHIFT;
 203         offset = in->f_pos & ~PAGE_CACHE_MASK;
 204         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 205
 206         if (nr_pages > PIPE_BUFFERS)
 207                 nr_pages = PIPE_BUFFERS;
 208
 209         /*
 210          * initiate read-ahead on this page range
 211          */
 212         do_page_cache_readahead(mapping, in, index, nr_pages);
 213
 214         /*
 215          * Get as many pages from the page cache as possible..
 216          * Start IO on the page cache entries we create (we
 217          * can assume that any pre-existing ones we find have
 218          * already had IO started on them).
 219          */
 220         i = find_get_pages(mapping, index, nr_pages, pages);
 221
 222         /*
 223          * common case - we found all pages and they are contiguous,
 224          * kick them off
 225          */
 226         if (i && (pages[i - 1]->index == index + i - 1))
 227                 goto splice_them;
 228
 229         /*
 230          * fill shadow[] with pages at the right locations, so we only
 231          * have to fill holes
 232          */
 233         memset(shadow, 0, i * sizeof(struct page *));
 234         for (j = 0, pidx = index; j < i; pidx++, j++)
 235                 shadow[pages[j]->index - pidx] = pages[j];
 236
 237         /*
 238          * now fill in the holes
 239          */
 240         for (i = 0, pidx = index; i < nr_pages; pidx++, i++) {
 241                 int error;
 242
 243                 if (shadow[i])
 244                         continue;
 245
 246                 /*
 247                  * no page there, look one up / create it
 248                  */
 249                 page = find_or_create_page(mapping, pidx,
 250                                                    mapping_gfp_mask(mapping));
 251                 if (!page)
 252                         break;
 253
 254                 if (PageUptodate(page))
 255                         unlock_page(page);
 256                 else {
 257                         error = mapping->a_ops->readpage(in, page);
 258
 259                         if (unlikely(error)) {
 260                                 page_cache_release(page);
 261                                 break;
 262                         }
 263                 }
 264                 shadow[i] = page;
 265         }
 266
 267         if (!i) {
 268                 for (i = 0; i < nr_pages; i++) {
 269                          if (shadow[i])
 270                                 page_cache_release(shadow[i]);
 271                 }
 272                 return 0;
 273         }
 274
 275         memcpy(pages, shadow, i * sizeof(struct page *));
 276
 277         /*
 278          * Now we splice them into the pipe..
 279          */
 280 splice_them:
 281         return move_to_pipe(pipe, pages, i, offset, len);
 282 }
 283
 284 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
 285                                  size_t len, unsigned int flags)
 286 {
 287         ssize_t spliced;
 288         int ret;
 289
 290         ret = 0;
 291         spliced = 0;
 292         while (len) {
 293                 ret = __generic_file_splice_read(in, pipe, len);
 294
 295                 if (ret <= 0)
 296                         break;
 297
 298                 in->f_pos += ret;
 299                 len -= ret;
 300                 spliced += ret;
 301         }
 302
 303         if (spliced)
 304                 return spliced;
 305
 306         return ret;
 307 }
 308
 309 /*
 310  * Send 'len' bytes to socket from 'file' at position 'pos' using sendpage().
 311  */
 312 static int pipe_to_sendpage(struct pipe_inode_info *info,
 313                             struct pipe_buffer *buf, struct splice_desc *sd)
 314 {
 315         struct file *file = sd->file;
 316         loff_t pos = sd->pos;
 317         unsigned int offset;
 318         ssize_t ret;
 319         void *ptr;
 320
 321         /*
 322          * sub-optimal, but we are limited by the pipe ->map. we don't
 323          * need a kmap'ed buffer here, we just want to make sure we
 324          * have the page pinned if the pipe page originates from the
 325          * page cache
 326          */
 327         ptr = buf->ops->map(file, info, buf);
 328         if (IS_ERR(ptr))
 329                 return PTR_ERR(ptr);
 330
 331         offset = pos & ~PAGE_CACHE_MASK;
 332
 333         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,
 334                                         sd->len < sd->total_len);
 335
 336         buf->ops->unmap(info, buf);
 337         if (ret == sd->len)
 338                 return 0;
 339
 340         return -EIO;
 341 }
 342
 343 /*
 344  * This is a little more tricky than the file -> pipe splicing. There are
 345  * basically three cases:
 346  *
 347  *      - Destination page already exists in the address space and there
 348  *        are users of it. For that case we have no other option that
 349  *        copying the data. Tough luck.
 350  *      - Destination page already exists in the address space, but there
 351  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 352  *        through to last case.
 353  *      - Destination page does not exist, we can add the pipe page to
 354  *        the page cache and avoid the copy.
 355  *
 356  * For now we just do the slower thing and always copy pages over, it's
 357  * easier than migrating pages from the pipe to the target file. For the
 358  * case of doing file | file splicing, the migrate approach had some LRU
 359  * nastiness...
 360  */
 361 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 362                         struct splice_desc *sd)
 363 {
 364         struct file *file = sd->file;
 365         struct address_space *mapping = file->f_mapping;
 366         unsigned int offset;
 367         struct page *page;
 368         pgoff_t index;
 369         char *src;
 370         int ret;
 371
 372         /*
 373          * after this, page will be locked and unmapped
 374          */
 375         src = buf->ops->map(file, info, buf);
 376         if (IS_ERR(src))
 377                 return PTR_ERR(src);
 378
 379         index = sd->pos >> PAGE_CACHE_SHIFT;
 380         offset = sd->pos & ~PAGE_CACHE_MASK;
 381
 382         /*
 383          * reuse buf page, if SPLICE_F_MOVE is set
 384          */
 385         if (sd->flags & SPLICE_F_MOVE) {
 386                 if (buf->ops->steal(info, buf))
 387                         goto find_page;
 388
 389                 page = buf->page;
 390                 if (add_to_page_cache_lru(page, mapping, index,
 391                                                 mapping_gfp_mask(mapping)))
 392                         goto find_page;
 393         } else {
 394 find_page:
 395                 ret = -ENOMEM;
 396                 page = find_or_create_page(mapping, index,
 397                                                 mapping_gfp_mask(mapping));
 398                 if (!page)
 399                         goto out;
 400
 401                 /*
 402                  * If the page is uptodate, it is also locked. If it isn't
 403                  * uptodate, we can mark it uptodate if we are filling the
 404                  * full page. Otherwise we need to read it in first...
 405                  */
 406                 if (!PageUptodate(page)) {
 407                         if (sd->len < PAGE_CACHE_SIZE) {
 408                                 ret = mapping->a_ops->readpage(file, page);
 409                                 if (unlikely(ret))
 410                                         goto out;
 411
 412                                 lock_page(page);
 413
 414                                 if (!PageUptodate(page)) {
 415                                         /*
 416                                          * page got invalidated, repeat
 417                                          */
 418                                         if (!page->mapping) {
 419                                                 unlock_page(page);
 420                                                 page_cache_release(page);
 421                                                 goto find_page;
 422                                         }
 423                                         ret = -EIO;
 424                                         goto out;
 425                                 }
 426                         } else {
 427                                 WARN_ON(!PageLocked(page));
 428                                 SetPageUptodate(page);
 429                         }
 430                 }
 431         }
 432
 433         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 434         if (ret)
 435                 goto out;
 436
 437         if (!buf->stolen) {
 438                 char *dst = kmap_atomic(page, KM_USER0);
 439
 440                 memcpy(dst + offset, src + buf->offset, sd->len);
 441                 flush_dcache_page(page);
 442                 kunmap_atomic(dst, KM_USER0);
 443         }
 444
 445         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 446         if (ret < 0)
 447                 goto out;
 448
 449         set_page_dirty(page);
 450         ret = write_one_page(page, 0);
 451 out:
 452         if (ret < 0)
 453                 unlock_page(page);
 454         if (!buf->stolen)
 455                 page_cache_release(page);
 456         buf->ops->unmap(info, buf);
 457         return ret;
 458 }
 459
 460 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 461                            struct splice_desc *);
 462
 463 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
 464                               size_t len, unsigned int flags,
 465                               splice_actor *actor)
 466 {
 467         struct pipe_inode_info *info;
 468         int ret, do_wakeup, err;
 469         struct splice_desc sd;
 470
 471         ret = 0;
 472         do_wakeup = 0;
 473
 474         sd.total_len = len;
 475         sd.flags = flags;
 476         sd.file = out;
 477         sd.pos = out->f_pos;
 478
 479         mutex_lock(PIPE_MUTEX(*inode));
 480
 481         info = inode->i_pipe;
 482         for (;;) {
 483                 int bufs = info->nrbufs;
 484
 485                 if (bufs) {
 486                         int curbuf = info->curbuf;
 487                         struct pipe_buffer *buf = info->bufs + curbuf;
 488                         struct pipe_buf_operations *ops = buf->ops;
 489
 490                         sd.len = buf->len;
 491                         if (sd.len > sd.total_len)
 492                                 sd.len = sd.total_len;
 493
 494                         err = actor(info, buf, &sd);
 495                         if (err) {
 496                                 if (!ret && err != -ENODATA)
 497                                         ret = err;
 498
 499                                 break;
 500                         }
 501
 502                         ret += sd.len;
 503                         buf->offset += sd.len;
 504                         buf->len -= sd.len;
 505                         if (!buf->len) {
 506                                 buf->ops = NULL;
 507                                 ops->release(info, buf);
 508                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
 509                                 info->curbuf = curbuf;
 510                                 info->nrbufs = --bufs;
 511                                 do_wakeup = 1;
 512                         }
 513
 514                         sd.pos += sd.len;
 515                         sd.total_len -= sd.len;
 516                         if (!sd.total_len)
 517                                 break;
 518                 }
 519
 520                 if (bufs)
 521                         continue;
 522                 if (!PIPE_WRITERS(*inode))
 523                         break;
 524                 if (!PIPE_WAITING_WRITERS(*inode)) {
 525                         if (ret)
 526                                 break;
 527                 }
 528
 529                 if (signal_pending(current)) {
 530                         if (!ret)
 531                                 ret = -ERESTARTSYS;
 532                         break;
 533                 }
 534
 535                 if (do_wakeup) {
 536                         wake_up_interruptible_sync(PIPE_WAIT(*inode));
 537                         kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
 538                         do_wakeup = 0;
 539                 }
 540
 541                 pipe_wait(inode);
 542         }
 543
 544         mutex_unlock(PIPE_MUTEX(*inode));
 545
 546         if (do_wakeup) {
 547                 wake_up_interruptible(PIPE_WAIT(*inode));
 548                 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
 549         }
 550
 551         mutex_lock(&out->f_mapping->host->i_mutex);
 552         out->f_pos = sd.pos;
 553         mutex_unlock(&out->f_mapping->host->i_mutex);
 554         return ret;
 555
 556 }
 557
 558 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
 559                                   size_t len, unsigned int flags)
 560 {
 561         return move_from_pipe(inode, out, len, flags, pipe_to_file);
 562 }
 563
 564 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
 565                                 size_t len, unsigned int flags)
 566 {
 567         return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
 568 }
 569
 570 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
 571                            unsigned int flags)
 572 {
 573         loff_t pos;
 574         int ret;
 575
 576         if (!out->f_op || !out->f_op->splice_write)
 577                 return -EINVAL;
 578
 579         if (!(out->f_mode & FMODE_WRITE))
 580                 return -EBADF;
 581
 582         pos = out->f_pos;
 583         ret = rw_verify_area(WRITE, out, &pos, len);
 584         if (unlikely(ret < 0))
 585                 return ret;
 586
 587         return out->f_op->splice_write(pipe, out, len, flags);
 588 }
 589
 590 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
 591                          unsigned int flags)
 592 {
 593         loff_t pos, isize, left;
 594         int ret;
 595
 596         if (!in->f_op || !in->f_op->splice_read)
 597                 return -EINVAL;
 598
 599         if (!(in->f_mode & FMODE_READ))
 600                 return -EBADF;
 601
 602         pos = in->f_pos;
 603         ret = rw_verify_area(READ, in, &pos, len);
 604         if (unlikely(ret < 0))
 605                 return ret;
 606
 607         isize = i_size_read(in->f_mapping->host);
 608         if (unlikely(in->f_pos >= isize))
 609                 return 0;
 610
 611         left = isize - in->f_pos;
 612         if (left < len)
 613                 len = left;
 614
 615         return in->f_op->splice_read(in, pipe, len, flags);
 616 }
 617
 618 static long do_splice(struct file *in, struct file *out, size_t len,
 619                       unsigned int flags)
 620 {
 621         struct inode *pipe;
 622
 623         pipe = in->f_dentry->d_inode;
 624         if (pipe->i_pipe)
 625                 return do_splice_from(pipe, out, len, flags);
 626
 627         pipe = out->f_dentry->d_inode;
 628         if (pipe->i_pipe)
 629                 return do_splice_to(in, pipe, len, flags);
 630
 631         return -EINVAL;
 632 }
 633
 634 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
 635 {
 636         long error;
 637         struct file *in, *out;
 638         int fput_in, fput_out;
 639
 640         if (unlikely(!len))
 641                 return 0;
 642
 643         error = -EBADF;
 644         in = fget_light(fdin, &fput_in);
 645         if (in) {
 646                 if (in->f_mode & FMODE_READ) {
 647                         out = fget_light(fdout, &fput_out);
 648                         if (out) {
 649                                 if (out->f_mode & FMODE_WRITE)
 650                                         error = do_splice(in, out, len, flags);
 651                                 fput_light(out, fput_out);
 652                         }
 653                 }
 654
 655                 fput_light(in, fput_in);
 656         }
 657
 658         return error;
 659 }