git.oblomov.eu Git - 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 #include <linux/module.h>
  26
  27 /*
  28  * Passed to the actors
  29  */
  30 struct splice_desc {
  31         unsigned int len, total_len;    /* current and remaining length */
  32         unsigned int flags;             /* splice flags */
  33         struct file *file;              /* file to read/write */
  34         loff_t pos;                     /* file position */
  35 };
  36
  37 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
  38                                      struct pipe_buffer *buf)
  39 {
  40         struct page *page = buf->page;
  41
  42         WARN_ON(!PageLocked(page));
  43         WARN_ON(!PageUptodate(page));
  44
  45         if (!remove_mapping(page_mapping(page), page))
  46                 return 1;
  47
  48         if (PageLRU(page)) {
  49                 struct zone *zone = page_zone(page);
  50
  51                 spin_lock_irq(&zone->lru_lock);
  52                 BUG_ON(!PageLRU(page));
  53                 __ClearPageLRU(page);
  54                 del_page_from_lru(zone, page);
  55                 spin_unlock_irq(&zone->lru_lock);
  56         }
  57
  58         buf->stolen = 1;
  59         return 0;
  60 }
  61
  62 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
  63                                         struct pipe_buffer *buf)
  64 {
  65         page_cache_release(buf->page);
  66         buf->page = NULL;
  67         buf->stolen = 0;
  68 }
  69
  70 static void *page_cache_pipe_buf_map(struct file *file,
  71                                      struct pipe_inode_info *info,
  72                                      struct pipe_buffer *buf)
  73 {
  74         struct page *page = buf->page;
  75
  76         lock_page(page);
  77
  78         if (!PageUptodate(page)) {
  79                 unlock_page(page);
  80                 return ERR_PTR(-EIO);
  81         }
  82
  83         if (!page->mapping) {
  84                 unlock_page(page);
  85                 return ERR_PTR(-ENODATA);
  86         }
  87
  88         return kmap(buf->page);
  89 }
  90
  91 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
  92                                       struct pipe_buffer *buf)
  93 {
  94         if (!buf->stolen)
  95                 unlock_page(buf->page);
  96         kunmap(buf->page);
  97 }
  98
  99 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
 100         .can_merge = 0,
 101         .map = page_cache_pipe_buf_map,
 102         .unmap = page_cache_pipe_buf_unmap,
 103         .release = page_cache_pipe_buf_release,
 104         .steal = page_cache_pipe_buf_steal,
 105 };
 106
 107 static ssize_t move_to_pipe(struct inode *inode, struct page **pages,
 108                             int nr_pages, unsigned long offset,
 109                             unsigned long len)
 110 {
 111         struct pipe_inode_info *info;
 112         int ret, do_wakeup, i;
 113
 114         ret = 0;
 115         do_wakeup = 0;
 116         i = 0;
 117
 118         mutex_lock(PIPE_MUTEX(*inode));
 119
 120         info = inode->i_pipe;
 121         for (;;) {
 122                 int bufs;
 123
 124                 if (!PIPE_READERS(*inode)) {
 125                         send_sig(SIGPIPE, current, 0);
 126                         if (!ret)
 127                                 ret = -EPIPE;
 128                         break;
 129                 }
 130
 131                 bufs = info->nrbufs;
 132                 if (bufs < PIPE_BUFFERS) {
 133                         int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS - 1);
 134                         struct pipe_buffer *buf = info->bufs + newbuf;
 135                         struct page *page = pages[i++];
 136                         unsigned long this_len;
 137
 138                         this_len = PAGE_CACHE_SIZE - offset;
 139                         if (this_len > len)
 140                                 this_len = len;
 141
 142                         buf->page = page;
 143                         buf->offset = offset;
 144                         buf->len = this_len;
 145                         buf->ops = &page_cache_pipe_buf_ops;
 146                         info->nrbufs = ++bufs;
 147                         do_wakeup = 1;
 148
 149                         ret += this_len;
 150                         len -= this_len;
 151                         offset = 0;
 152                         if (!--nr_pages)
 153                                 break;
 154                         if (!len)
 155                                 break;
 156                         if (bufs < PIPE_BUFFERS)
 157                                 continue;
 158
 159                         break;
 160                 }
 161
 162                 if (signal_pending(current)) {
 163                         if (!ret)
 164                                 ret = -ERESTARTSYS;
 165                         break;
 166                 }
 167
 168                 if (do_wakeup) {
 169                         wake_up_interruptible_sync(PIPE_WAIT(*inode));
 170                         kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO,
 171                                     POLL_IN);
 172                         do_wakeup = 0;
 173                 }
 174
 175                 PIPE_WAITING_WRITERS(*inode)++;
 176                 pipe_wait(inode);
 177                 PIPE_WAITING_WRITERS(*inode)--;
 178         }
 179
 180         mutex_unlock(PIPE_MUTEX(*inode));
 181
 182         if (do_wakeup) {
 183                 wake_up_interruptible(PIPE_WAIT(*inode));
 184                 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
 185         }
 186
 187         while (i < nr_pages)
 188                 page_cache_release(pages[i++]);
 189
 190         return ret;
 191 }
 192
 193 static int __generic_file_splice_read(struct file *in, struct inode *pipe,
 194                                       size_t len)
 195 {
 196         struct address_space *mapping = in->f_mapping;
 197         unsigned int offset, nr_pages;
 198         struct page *pages[PIPE_BUFFERS], *shadow[PIPE_BUFFERS];
 199         struct page *page;
 200         pgoff_t index, pidx;
 201         int i, j;
 202
 203         index = in->f_pos >> PAGE_CACHE_SHIFT;
 204         offset = in->f_pos & ~PAGE_CACHE_MASK;
 205         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 206
 207         if (nr_pages > PIPE_BUFFERS)
 208                 nr_pages = PIPE_BUFFERS;
 209
 210         /*
 211          * initiate read-ahead on this page range
 212          */
 213         do_page_cache_readahead(mapping, in, index, nr_pages);
 214
 215         /*
 216          * Get as many pages from the page cache as possible..
 217          * Start IO on the page cache entries we create (we
 218          * can assume that any pre-existing ones we find have
 219          * already had IO started on them).
 220          */
 221         i = find_get_pages(mapping, index, nr_pages, pages);
 222
 223         /*
 224          * common case - we found all pages and they are contiguous,
 225          * kick them off
 226          */
 227         if (i && (pages[i - 1]->index == index + i - 1))
 228                 goto splice_them;
 229
 230         /*
 231          * fill shadow[] with pages at the right locations, so we only
 232          * have to fill holes
 233          */
 234         memset(shadow, 0, i * sizeof(struct page *));
 235         for (j = 0, pidx = index; j < i; pidx++, j++)
 236                 shadow[pages[j]->index - pidx] = pages[j];
 237
 238         /*
 239          * now fill in the holes
 240          */
 241         for (i = 0, pidx = index; i < nr_pages; pidx++, i++) {
 242                 int error;
 243
 244                 if (shadow[i])
 245                         continue;
 246
 247                 /*
 248                  * no page there, look one up / create it
 249                  */
 250                 page = find_or_create_page(mapping, pidx,
 251                                                    mapping_gfp_mask(mapping));
 252                 if (!page)
 253                         break;
 254
 255                 if (PageUptodate(page))
 256                         unlock_page(page);
 257                 else {
 258                         error = mapping->a_ops->readpage(in, page);
 259
 260                         if (unlikely(error)) {
 261                                 page_cache_release(page);
 262                                 break;
 263                         }
 264                 }
 265                 shadow[i] = page;
 266         }
 267
 268         if (!i) {
 269                 for (i = 0; i < nr_pages; i++) {
 270                          if (shadow[i])
 271                                 page_cache_release(shadow[i]);
 272                 }
 273                 return 0;
 274         }
 275
 276         memcpy(pages, shadow, i * sizeof(struct page *));
 277
 278         /*
 279          * Now we splice them into the pipe..
 280          */
 281 splice_them:
 282         return move_to_pipe(pipe, pages, i, offset, len);
 283 }
 284
 285 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
 286                                  size_t len, unsigned int flags)
 287 {
 288         ssize_t spliced;
 289         int ret;
 290
 291         ret = 0;
 292         spliced = 0;
 293         while (len) {
 294                 ret = __generic_file_splice_read(in, pipe, len);
 295
 296                 if (ret <= 0)
 297                         break;
 298
 299                 in->f_pos += ret;
 300                 len -= ret;
 301                 spliced += ret;
 302         }
 303
 304         if (spliced)
 305                 return spliced;
 306
 307         return ret;
 308 }
 309
 310 /*
 311  * Send 'len' bytes to socket from 'file' at position 'pos' using sendpage().
 312  */
 313 static int pipe_to_sendpage(struct pipe_inode_info *info,
 314                             struct pipe_buffer *buf, struct splice_desc *sd)
 315 {
 316         struct file *file = sd->file;
 317         loff_t pos = sd->pos;
 318         unsigned int offset;
 319         ssize_t ret;
 320         void *ptr;
 321
 322         /*
 323          * sub-optimal, but we are limited by the pipe ->map. we don't
 324          * need a kmap'ed buffer here, we just want to make sure we
 325          * have the page pinned if the pipe page originates from the
 326          * page cache
 327          */
 328         ptr = buf->ops->map(file, info, buf);
 329         if (IS_ERR(ptr))
 330                 return PTR_ERR(ptr);
 331
 332         offset = pos & ~PAGE_CACHE_MASK;
 333
 334         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,
 335                                         sd->len < sd->total_len);
 336
 337         buf->ops->unmap(info, buf);
 338         if (ret == sd->len)
 339                 return 0;
 340
 341         return -EIO;
 342 }
 343
 344 /*
 345  * This is a little more tricky than the file -> pipe splicing. There are
 346  * basically three cases:
 347  *
 348  *      - Destination page already exists in the address space and there
 349  *        are users of it. For that case we have no other option that
 350  *        copying the data. Tough luck.
 351  *      - Destination page already exists in the address space, but there
 352  *        are no users of it. Make sure it's uptodate, then drop it. Fall
 353  *        through to last case.
 354  *      - Destination page does not exist, we can add the pipe page to
 355  *        the page cache and avoid the copy.
 356  *
 357  * For now we just do the slower thing and always copy pages over, it's
 358  * easier than migrating pages from the pipe to the target file. For the
 359  * case of doing file | file splicing, the migrate approach had some LRU
 360  * nastiness...
 361  */
 362 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
 363                         struct splice_desc *sd)
 364 {
 365         struct file *file = sd->file;
 366         struct address_space *mapping = file->f_mapping;
 367         unsigned int offset;
 368         struct page *page;
 369         pgoff_t index;
 370         char *src;
 371         int ret;
 372
 373         /*
 374          * after this, page will be locked and unmapped
 375          */
 376         src = buf->ops->map(file, info, buf);
 377         if (IS_ERR(src))
 378                 return PTR_ERR(src);
 379
 380         index = sd->pos >> PAGE_CACHE_SHIFT;
 381         offset = sd->pos & ~PAGE_CACHE_MASK;
 382
 383         /*
 384          * reuse buf page, if SPLICE_F_MOVE is set
 385          */
 386         if (sd->flags & SPLICE_F_MOVE) {
 387                 if (buf->ops->steal(info, buf))
 388                         goto find_page;
 389
 390                 page = buf->page;
 391                 if (add_to_page_cache_lru(page, mapping, index,
 392                                                 mapping_gfp_mask(mapping)))
 393                         goto find_page;
 394         } else {
 395 find_page:
 396                 ret = -ENOMEM;
 397                 page = find_or_create_page(mapping, index,
 398                                                 mapping_gfp_mask(mapping));
 399                 if (!page)
 400                         goto out;
 401
 402                 /*
 403                  * If the page is uptodate, it is also locked. If it isn't
 404                  * uptodate, we can mark it uptodate if we are filling the
 405                  * full page. Otherwise we need to read it in first...
 406                  */
 407                 if (!PageUptodate(page)) {
 408                         if (sd->len < PAGE_CACHE_SIZE) {
 409                                 ret = mapping->a_ops->readpage(file, page);
 410                                 if (unlikely(ret))
 411                                         goto out;
 412
 413                                 lock_page(page);
 414
 415                                 if (!PageUptodate(page)) {
 416                                         /*
 417                                          * page got invalidated, repeat
 418                                          */
 419                                         if (!page->mapping) {
 420                                                 unlock_page(page);
 421                                                 page_cache_release(page);
 422                                                 goto find_page;
 423                                         }
 424                                         ret = -EIO;
 425                                         goto out;
 426                                 }
 427                         } else {
 428                                 WARN_ON(!PageLocked(page));
 429                                 SetPageUptodate(page);
 430                         }
 431                 }
 432         }
 433
 434         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
 435         if (ret)
 436                 goto out;
 437
 438         if (!buf->stolen) {
 439                 char *dst = kmap_atomic(page, KM_USER0);
 440
 441                 memcpy(dst + offset, src + buf->offset, sd->len);
 442                 flush_dcache_page(page);
 443                 kunmap_atomic(dst, KM_USER0);
 444         }
 445
 446         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
 447         if (ret < 0)
 448                 goto out;
 449
 450         set_page_dirty(page);
 451         ret = write_one_page(page, 0);
 452 out:
 453         if (ret < 0)
 454                 unlock_page(page);
 455         if (!buf->stolen)
 456                 page_cache_release(page);
 457         buf->ops->unmap(info, buf);
 458         return ret;
 459 }
 460
 461 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
 462                            struct splice_desc *);
 463
 464 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
 465                               size_t len, unsigned int flags,
 466                               splice_actor *actor)
 467 {
 468         struct pipe_inode_info *info;
 469         int ret, do_wakeup, err;
 470         struct splice_desc sd;
 471
 472         ret = 0;
 473         do_wakeup = 0;
 474
 475         sd.total_len = len;
 476         sd.flags = flags;
 477         sd.file = out;
 478         sd.pos = out->f_pos;
 479
 480         mutex_lock(PIPE_MUTEX(*inode));
 481
 482         info = inode->i_pipe;
 483         for (;;) {
 484                 int bufs = info->nrbufs;
 485
 486                 if (bufs) {
 487                         int curbuf = info->curbuf;
 488                         struct pipe_buffer *buf = info->bufs + curbuf;
 489                         struct pipe_buf_operations *ops = buf->ops;
 490
 491                         sd.len = buf->len;
 492                         if (sd.len > sd.total_len)
 493                                 sd.len = sd.total_len;
 494
 495                         err = actor(info, buf, &sd);
 496                         if (err) {
 497                                 if (!ret && err != -ENODATA)
 498                                         ret = err;
 499
 500                                 break;
 501                         }
 502
 503                         ret += sd.len;
 504                         buf->offset += sd.len;
 505                         buf->len -= sd.len;
 506                         if (!buf->len) {
 507                                 buf->ops = NULL;
 508                                 ops->release(info, buf);
 509                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
 510                                 info->curbuf = curbuf;
 511                                 info->nrbufs = --bufs;
 512                                 do_wakeup = 1;
 513                         }
 514
 515                         sd.pos += sd.len;
 516                         sd.total_len -= sd.len;
 517                         if (!sd.total_len)
 518                                 break;
 519                 }
 520
 521                 if (bufs)
 522                         continue;
 523                 if (!PIPE_WRITERS(*inode))
 524                         break;
 525                 if (!PIPE_WAITING_WRITERS(*inode)) {
 526                         if (ret)
 527                                 break;
 528                 }
 529
 530                 if (signal_pending(current)) {
 531                         if (!ret)
 532                                 ret = -ERESTARTSYS;
 533                         break;
 534                 }
 535
 536                 if (do_wakeup) {
 537                         wake_up_interruptible_sync(PIPE_WAIT(*inode));
 538                         kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
 539                         do_wakeup = 0;
 540                 }
 541
 542                 pipe_wait(inode);
 543         }
 544
 545         mutex_unlock(PIPE_MUTEX(*inode));
 546
 547         if (do_wakeup) {
 548                 wake_up_interruptible(PIPE_WAIT(*inode));
 549                 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
 550         }
 551
 552         mutex_lock(&out->f_mapping->host->i_mutex);
 553         out->f_pos = sd.pos;
 554         mutex_unlock(&out->f_mapping->host->i_mutex);
 555         return ret;
 556
 557 }
 558
 559 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
 560                                   size_t len, unsigned int flags)
 561 {
 562         return move_from_pipe(inode, out, len, flags, pipe_to_file);
 563 }
 564
 565 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
 566                                 size_t len, unsigned int flags)
 567 {
 568         return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
 569 }
 570
 571 EXPORT_SYMBOL(generic_file_splice_write);
 572 EXPORT_SYMBOL(generic_file_splice_read);
 573
 574 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
 575                            unsigned int flags)
 576 {
 577         loff_t pos;
 578         int ret;
 579
 580         if (!out->f_op || !out->f_op->splice_write)
 581                 return -EINVAL;
 582
 583         if (!(out->f_mode & FMODE_WRITE))
 584                 return -EBADF;
 585
 586         pos = out->f_pos;
 587         ret = rw_verify_area(WRITE, out, &pos, len);
 588         if (unlikely(ret < 0))
 589                 return ret;
 590
 591         return out->f_op->splice_write(pipe, out, len, flags);
 592 }
 593
 594 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
 595                          unsigned int flags)
 596 {
 597         loff_t pos, isize, left;
 598         int ret;
 599
 600         if (!in->f_op || !in->f_op->splice_read)
 601                 return -EINVAL;
 602
 603         if (!(in->f_mode & FMODE_READ))
 604                 return -EBADF;
 605
 606         pos = in->f_pos;
 607         ret = rw_verify_area(READ, in, &pos, len);
 608         if (unlikely(ret < 0))
 609                 return ret;
 610
 611         isize = i_size_read(in->f_mapping->host);
 612         if (unlikely(in->f_pos >= isize))
 613                 return 0;
 614
 615         left = isize - in->f_pos;
 616         if (left < len)
 617                 len = left;
 618
 619         return in->f_op->splice_read(in, pipe, len, flags);
 620 }
 621
 622 static long do_splice(struct file *in, struct file *out, size_t len,
 623                       unsigned int flags)
 624 {
 625         struct inode *pipe;
 626
 627         pipe = in->f_dentry->d_inode;
 628         if (pipe->i_pipe)
 629                 return do_splice_from(pipe, out, len, flags);
 630
 631         pipe = out->f_dentry->d_inode;
 632         if (pipe->i_pipe)
 633                 return do_splice_to(in, pipe, len, flags);
 634
 635         return -EINVAL;
 636 }
 637
 638 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
 639 {
 640         long error;
 641         struct file *in, *out;
 642         int fput_in, fput_out;
 643
 644         if (unlikely(!len))
 645                 return 0;
 646
 647         error = -EBADF;
 648         in = fget_light(fdin, &fput_in);
 649         if (in) {
 650                 if (in->f_mode & FMODE_READ) {
 651                         out = fget_light(fdout, &fput_out);
 652                         if (out) {
 653                                 if (out->f_mode & FMODE_WRITE)
 654                                         error = do_splice(in, out, len, flags);
 655                                 fput_light(out, fput_out);
 656                         }
 657                 }
 658
 659                 fput_light(in, fput_in);
 660         }
 661
 662         return error;
 663 }