linux/fs/splice.c
<<
>>
Prefs
   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, network, direct splicing, etc and
  13 * fixing lots of bugs.
  14 *
  15 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
  16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
  17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
  18 *
  19 */
  20#include <linux/fs.h>
  21#include <linux/file.h>
  22#include <linux/pagemap.h>
  23#include <linux/splice.h>
  24#include <linux/memcontrol.h>
  25#include <linux/mm_inline.h>
  26#include <linux/swap.h>
  27#include <linux/writeback.h>
  28#include <linux/export.h>
  29#include <linux/syscalls.h>
  30#include <linux/uio.h>
  31#include <linux/security.h>
  32#include <linux/gfp.h>
  33#include <linux/socket.h>
  34
  35/*
  36 * Attempt to steal a page from a pipe buffer. This should perhaps go into
  37 * a vm helper function, it's already simplified quite a bit by the
  38 * addition of remove_mapping(). If success is returned, the caller may
  39 * attempt to reuse this page for another destination.
  40 */
  41static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
  42                                     struct pipe_buffer *buf)
  43{
  44        struct page *page = buf->page;
  45        struct address_space *mapping;
  46
  47        lock_page(page);
  48
  49        mapping = page_mapping(page);
  50        if (mapping) {
  51                WARN_ON(!PageUptodate(page));
  52
  53                /*
  54                 * At least for ext2 with nobh option, we need to wait on
  55                 * writeback completing on this page, since we'll remove it
  56                 * from the pagecache.  Otherwise truncate wont wait on the
  57                 * page, allowing the disk blocks to be reused by someone else
  58                 * before we actually wrote our data to them. fs corruption
  59                 * ensues.
  60                 */
  61                wait_on_page_writeback(page);
  62
  63                if (page_has_private(page) &&
  64                    !try_to_release_page(page, GFP_KERNEL))
  65                        goto out_unlock;
  66
  67                /*
  68                 * If we succeeded in removing the mapping, set LRU flag
  69                 * and return good.
  70                 */
  71                if (remove_mapping(mapping, page)) {
  72                        buf->flags |= PIPE_BUF_FLAG_LRU;
  73                        return 0;
  74                }
  75        }
  76
  77        /*
  78         * Raced with truncate or failed to remove page from current
  79         * address space, unlock and return failure.
  80         */
  81out_unlock:
  82        unlock_page(page);
  83        return 1;
  84}
  85
  86static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
  87                                        struct pipe_buffer *buf)
  88{
  89        page_cache_release(buf->page);
  90        buf->flags &= ~PIPE_BUF_FLAG_LRU;
  91}
  92
  93/*
  94 * Check whether the contents of buf is OK to access. Since the content
  95 * is a page cache page, IO may be in flight.
  96 */
  97static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
  98                                       struct pipe_buffer *buf)
  99{
 100        struct page *page = buf->page;
 101        int err;
 102
 103        if (!PageUptodate(page)) {
 104                lock_page(page);
 105
 106                /*
 107                 * Page got truncated/unhashed. This will cause a 0-byte
 108                 * splice, if this is the first page.
 109                 */
 110                if (!page->mapping) {
 111                        err = -ENODATA;
 112                        goto error;
 113                }
 114
 115                /*
 116                 * Uh oh, read-error from disk.
 117                 */
 118                if (!PageUptodate(page)) {
 119                        err = -EIO;
 120                        goto error;
 121                }
 122
 123                /*
 124                 * Page is ok afterall, we are done.
 125                 */
 126                unlock_page(page);
 127        }
 128
 129        return 0;
 130error:
 131        unlock_page(page);
 132        return err;
 133}
 134
 135const struct pipe_buf_operations page_cache_pipe_buf_ops = {
 136        .can_merge = 0,
 137        .map = generic_pipe_buf_map,
 138        .unmap = generic_pipe_buf_unmap,
 139        .confirm = page_cache_pipe_buf_confirm,
 140        .release = page_cache_pipe_buf_release,
 141        .steal = page_cache_pipe_buf_steal,
 142        .get = generic_pipe_buf_get,
 143};
 144
 145static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
 146                                    struct pipe_buffer *buf)
 147{
 148        if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
 149                return 1;
 150
 151        buf->flags |= PIPE_BUF_FLAG_LRU;
 152        return generic_pipe_buf_steal(pipe, buf);
 153}
 154
 155static const struct pipe_buf_operations user_page_pipe_buf_ops = {
 156        .can_merge = 0,
 157        .map = generic_pipe_buf_map,
 158        .unmap = generic_pipe_buf_unmap,
 159        .confirm = generic_pipe_buf_confirm,
 160        .release = page_cache_pipe_buf_release,
 161        .steal = user_page_pipe_buf_steal,
 162        .get = generic_pipe_buf_get,
 163};
 164
 165static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
 166{
 167        smp_mb();
 168        if (waitqueue_active(&pipe->wait))
 169                wake_up_interruptible(&pipe->wait);
 170        kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 171}
 172
 173/**
 174 * splice_to_pipe - fill passed data into a pipe
 175 * @pipe:       pipe to fill
 176 * @spd:        data to fill
 177 *
 178 * Description:
 179 *    @spd contains a map of pages and len/offset tuples, along with
 180 *    the struct pipe_buf_operations associated with these pages. This
 181 *    function will link that data to the pipe.
 182 *
 183 */
 184ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 185                       struct splice_pipe_desc *spd)
 186{
 187        unsigned int spd_pages = spd->nr_pages;
 188        int ret, do_wakeup, page_nr;
 189
 190        ret = 0;
 191        do_wakeup = 0;
 192        page_nr = 0;
 193
 194        pipe_lock(pipe);
 195
 196        for (;;) {
 197                if (!pipe->readers) {
 198                        send_sig(SIGPIPE, current, 0);
 199                        if (!ret)
 200                                ret = -EPIPE;
 201                        break;
 202                }
 203
 204                if (pipe->nrbufs < pipe->buffers) {
 205                        int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
 206                        struct pipe_buffer *buf = pipe->bufs + newbuf;
 207
 208                        buf->page = spd->pages[page_nr];
 209                        buf->offset = spd->partial[page_nr].offset;
 210                        buf->len = spd->partial[page_nr].len;
 211                        buf->private = spd->partial[page_nr].private;
 212                        buf->ops = spd->ops;
 213                        if (spd->flags & SPLICE_F_GIFT)
 214                                buf->flags |= PIPE_BUF_FLAG_GIFT;
 215
 216                        pipe->nrbufs++;
 217                        page_nr++;
 218                        ret += buf->len;
 219
 220                        if (pipe->inode)
 221                                do_wakeup = 1;
 222
 223                        if (!--spd->nr_pages)
 224                                break;
 225                        if (pipe->nrbufs < pipe->buffers)
 226                                continue;
 227
 228                        break;
 229                }
 230
 231                if (spd->flags & SPLICE_F_NONBLOCK) {
 232                        if (!ret)
 233                                ret = -EAGAIN;
 234                        break;
 235                }
 236
 237                if (signal_pending(current)) {
 238                        if (!ret)
 239                                ret = -ERESTARTSYS;
 240                        break;
 241                }
 242
 243                if (do_wakeup) {
 244                        smp_mb();
 245                        if (waitqueue_active(&pipe->wait))
 246                                wake_up_interruptible_sync(&pipe->wait);
 247                        kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 248                        do_wakeup = 0;
 249                }
 250
 251                pipe->waiting_writers++;
 252                pipe_wait(pipe);
 253                pipe->waiting_writers--;
 254        }
 255
 256        pipe_unlock(pipe);
 257
 258        if (do_wakeup)
 259                wakeup_pipe_readers(pipe);
 260
 261        while (page_nr < spd_pages)
 262                spd->spd_release(spd, page_nr++);
 263
 264        return ret;
 265}
 266
 267void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
 268{
 269        page_cache_release(spd->pages[i]);
 270}
 271
 272/*
 273 * Check if we need to grow the arrays holding pages and partial page
 274 * descriptions.
 275 */
 276int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
 277{
 278        unsigned int buffers = ACCESS_ONCE(pipe->buffers);
 279
 280        spd->nr_pages_max = buffers;
 281        if (buffers <= PIPE_DEF_BUFFERS)
 282                return 0;
 283
 284        spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
 285        spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
 286
 287        if (spd->pages && spd->partial)
 288                return 0;
 289
 290        kfree(spd->pages);
 291        kfree(spd->partial);
 292        return -ENOMEM;
 293}
 294
 295void splice_shrink_spd(struct splice_pipe_desc *spd)
 296{
 297        if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
 298                return;
 299
 300        kfree(spd->pages);
 301        kfree(spd->partial);
 302}
 303
 304static int
 305__generic_file_splice_read(struct file *in, loff_t *ppos,
 306                           struct pipe_inode_info *pipe, size_t len,
 307                           unsigned int flags)
 308{
 309        struct address_space *mapping = in->f_mapping;
 310        unsigned int loff, nr_pages, req_pages;
 311        struct page *pages[PIPE_DEF_BUFFERS];
 312        struct partial_page partial[PIPE_DEF_BUFFERS];
 313        struct page *page;
 314        pgoff_t index, end_index;
 315        loff_t isize;
 316        int error, page_nr;
 317        struct splice_pipe_desc spd = {
 318                .pages = pages,
 319                .partial = partial,
 320                .nr_pages_max = PIPE_DEF_BUFFERS,
 321                .flags = flags,
 322                .ops = &page_cache_pipe_buf_ops,
 323                .spd_release = spd_release_page,
 324        };
 325
 326        if (splice_grow_spd(pipe, &spd))
 327                return -ENOMEM;
 328
 329        index = *ppos >> PAGE_CACHE_SHIFT;
 330        loff = *ppos & ~PAGE_CACHE_MASK;
 331        req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 332        nr_pages = min(req_pages, spd.nr_pages_max);
 333
 334        /*
 335         * Lookup the (hopefully) full range of pages we need.
 336         */
 337        spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
 338        index += spd.nr_pages;
 339
 340        /*
 341         * If find_get_pages_contig() returned fewer pages than we needed,
 342         * readahead/allocate the rest and fill in the holes.
 343         */
 344        if (spd.nr_pages < nr_pages)
 345                page_cache_sync_readahead(mapping, &in->f_ra, in,
 346                                index, req_pages - spd.nr_pages);
 347
 348        error = 0;
 349        while (spd.nr_pages < nr_pages) {
 350                /*
 351                 * Page could be there, find_get_pages_contig() breaks on
 352                 * the first hole.
 353                 */
 354                page = find_get_page(mapping, index);
 355                if (!page) {
 356                        /*
 357                         * page didn't exist, allocate one.
 358                         */
 359                        page = page_cache_alloc_cold(mapping);
 360                        if (!page)
 361                                break;
 362
 363                        error = add_to_page_cache_lru(page, mapping, index,
 364                                                GFP_KERNEL);
 365                        if (unlikely(error)) {
 366                                page_cache_release(page);
 367                                if (error == -EEXIST)
 368                                        continue;
 369                                break;
 370                        }
 371                        /*
 372                         * add_to_page_cache() locks the page, unlock it
 373                         * to avoid convoluting the logic below even more.
 374                         */
 375                        unlock_page(page);
 376                }
 377
 378                spd.pages[spd.nr_pages++] = page;
 379                index++;
 380        }
 381
 382        /*
 383         * Now loop over the map and see if we need to start IO on any
 384         * pages, fill in the partial map, etc.
 385         */
 386        index = *ppos >> PAGE_CACHE_SHIFT;
 387        nr_pages = spd.nr_pages;
 388        spd.nr_pages = 0;
 389        for (page_nr = 0; page_nr < nr_pages; page_nr++) {
 390                unsigned int this_len;
 391
 392                if (!len)
 393                        break;
 394
 395                /*
 396                 * this_len is the max we'll use from this page
 397                 */
 398                this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
 399                page = spd.pages[page_nr];
 400
 401                if (PageReadahead(page))
 402                        page_cache_async_readahead(mapping, &in->f_ra, in,
 403                                        page, index, req_pages - page_nr);
 404
 405                /*
 406                 * If the page isn't uptodate, we may need to start io on it
 407                 */
 408                if (!PageUptodate(page)) {
 409                        lock_page(page);
 410
 411                        /*
 412                         * Page was truncated, or invalidated by the
 413                         * filesystem.  Redo the find/create, but this time the
 414                         * page is kept locked, so there's no chance of another
 415                         * race with truncate/invalidate.
 416                         */
 417                        if (!page->mapping) {
 418                                unlock_page(page);
 419                                page = find_or_create_page(mapping, index,
 420                                                mapping_gfp_mask(mapping));
 421
 422                                if (!page) {
 423                                        error = -ENOMEM;
 424                                        break;
 425                                }
 426                                page_cache_release(spd.pages[page_nr]);
 427                                spd.pages[page_nr] = page;
 428                        }
 429                        /*
 430                         * page was already under io and is now done, great
 431                         */
 432                        if (PageUptodate(page)) {
 433                                unlock_page(page);
 434                                goto fill_it;
 435                        }
 436
 437                        /*
 438                         * need to read in the page
 439                         */
 440                        error = mapping->a_ops->readpage(in, page);
 441                        if (unlikely(error)) {
 442                                /*
 443                                 * We really should re-lookup the page here,
 444                                 * but it complicates things a lot. Instead
 445                                 * lets just do what we already stored, and
 446                                 * we'll get it the next time we are called.
 447                                 */
 448                                if (error == AOP_TRUNCATED_PAGE)
 449                                        error = 0;
 450
 451                                break;
 452                        }
 453                }
 454fill_it:
 455                /*
 456                 * i_size must be checked after PageUptodate.
 457                 */
 458                isize = i_size_read(mapping->host);
 459                end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 460                if (unlikely(!isize || index > end_index))
 461                        break;
 462
 463                /*
 464                 * if this is the last page, see if we need to shrink
 465                 * the length and stop
 466                 */
 467                if (end_index == index) {
 468                        unsigned int plen;
 469
 470                        /*
 471                         * max good bytes in this page
 472                         */
 473                        plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 474                        if (plen <= loff)
 475                                break;
 476
 477                        /*
 478                         * force quit after adding this page
 479                         */
 480                        this_len = min(this_len, plen - loff);
 481                        len = this_len;
 482                }
 483
 484                spd.partial[page_nr].offset = loff;
 485                spd.partial[page_nr].len = this_len;
 486                len -= this_len;
 487                loff = 0;
 488                spd.nr_pages++;
 489                index++;
 490        }
 491
 492        /*
 493         * Release any pages at the end, if we quit early. 'page_nr' is how far
 494         * we got, 'nr_pages' is how many pages are in the map.
 495         */
 496        while (page_nr < nr_pages)
 497                page_cache_release(spd.pages[page_nr++]);
 498        in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 499
 500        if (spd.nr_pages)
 501                error = splice_to_pipe(pipe, &spd);
 502
 503        splice_shrink_spd(&spd);
 504        return error;
 505}
 506
 507/**
 508 * generic_file_splice_read - splice data from file to a pipe
 509 * @in:         file to splice from
 510 * @ppos:       position in @in
 511 * @pipe:       pipe to splice to
 512 * @len:        number of bytes to splice
 513 * @flags:      splice modifier flags
 514 *
 515 * Description:
 516 *    Will read pages from given file and fill them into a pipe. Can be
 517 *    used as long as the address_space operations for the source implements
 518 *    a readpage() hook.
 519 *
 520 */
 521ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
 522                                 struct pipe_inode_info *pipe, size_t len,
 523                                 unsigned int flags)
 524{
 525        loff_t isize, left;
 526        int ret;
 527
 528        isize = i_size_read(in->f_mapping->host);
 529        if (unlikely(*ppos >= isize))
 530                return 0;
 531
 532        left = isize - *ppos;
 533        if (unlikely(left < len))
 534                len = left;
 535
 536        ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
 537        if (ret > 0) {
 538                *ppos += ret;
 539                file_accessed(in);
 540        }
 541
 542        return ret;
 543}
 544EXPORT_SYMBOL(generic_file_splice_read);
 545
 546static const struct pipe_buf_operations default_pipe_buf_ops = {
 547        .can_merge = 0,
 548        .map = generic_pipe_buf_map,
 549        .unmap = generic_pipe_buf_unmap,
 550        .confirm = generic_pipe_buf_confirm,
 551        .release = generic_pipe_buf_release,
 552        .steal = generic_pipe_buf_steal,
 553        .get = generic_pipe_buf_get,
 554};
 555
 556static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
 557                            unsigned long vlen, loff_t offset)
 558{
 559        mm_segment_t old_fs;
 560        loff_t pos = offset;
 561        ssize_t res;
 562
 563        old_fs = get_fs();
 564        set_fs(get_ds());
 565        /* The cast to a user pointer is valid due to the set_fs() */
 566        res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
 567        set_fs(old_fs);
 568
 569        return res;
 570}
 571
 572static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
 573                            loff_t pos)
 574{
 575        mm_segment_t old_fs;
 576        ssize_t res;
 577
 578        old_fs = get_fs();
 579        set_fs(get_ds());
 580        /* The cast to a user pointer is valid due to the set_fs() */
 581        res = vfs_write(file, (const char __user *)buf, count, &pos);
 582        set_fs(old_fs);
 583
 584        return res;
 585}
 586
 587ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
 588                                 struct pipe_inode_info *pipe, size_t len,
 589                                 unsigned int flags)
 590{
 591        unsigned int nr_pages;
 592        unsigned int nr_freed;
 593        size_t offset;
 594        struct page *pages[PIPE_DEF_BUFFERS];
 595        struct partial_page partial[PIPE_DEF_BUFFERS];
 596        struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
 597        ssize_t res;
 598        size_t this_len;
 599        int error;
 600        int i;
 601        struct splice_pipe_desc spd = {
 602                .pages = pages,
 603                .partial = partial,
 604                .nr_pages_max = PIPE_DEF_BUFFERS,
 605                .flags = flags,
 606                .ops = &default_pipe_buf_ops,
 607                .spd_release = spd_release_page,
 608        };
 609
 610        if (splice_grow_spd(pipe, &spd))
 611                return -ENOMEM;
 612
 613        res = -ENOMEM;
 614        vec = __vec;
 615        if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
 616                vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
 617                if (!vec)
 618                        goto shrink_ret;
 619        }
 620
 621        offset = *ppos & ~PAGE_CACHE_MASK;
 622        nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 623
 624        for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
 625                struct page *page;
 626
 627                page = alloc_page(GFP_USER);
 628                error = -ENOMEM;
 629                if (!page)
 630                        goto err;
 631
 632                this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
 633                vec[i].iov_base = (void __user *) page_address(page);
 634                vec[i].iov_len = this_len;
 635                spd.pages[i] = page;
 636                spd.nr_pages++;
 637                len -= this_len;
 638                offset = 0;
 639        }
 640
 641        res = kernel_readv(in, vec, spd.nr_pages, *ppos);
 642        if (res < 0) {
 643                error = res;
 644                goto err;
 645        }
 646
 647        error = 0;
 648        if (!res)
 649                goto err;
 650
 651        nr_freed = 0;
 652        for (i = 0; i < spd.nr_pages; i++) {
 653                this_len = min_t(size_t, vec[i].iov_len, res);
 654                spd.partial[i].offset = 0;
 655                spd.partial[i].len = this_len;
 656                if (!this_len) {
 657                        __free_page(spd.pages[i]);
 658                        spd.pages[i] = NULL;
 659                        nr_freed++;
 660                }
 661                res -= this_len;
 662        }
 663        spd.nr_pages -= nr_freed;
 664
 665        res = splice_to_pipe(pipe, &spd);
 666        if (res > 0)
 667                *ppos += res;
 668
 669shrink_ret:
 670        if (vec != __vec)
 671                kfree(vec);
 672        splice_shrink_spd(&spd);
 673        return res;
 674
 675err:
 676        for (i = 0; i < spd.nr_pages; i++)
 677                __free_page(spd.pages[i]);
 678
 679        res = error;
 680        goto shrink_ret;
 681}
 682EXPORT_SYMBOL(default_file_splice_read);
 683
 684/*
 685 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
 686 * using sendpage(). Return the number of bytes sent.
 687 */
 688static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 689                            struct pipe_buffer *buf, struct splice_desc *sd)
 690{
 691        struct file *file = sd->u.file;
 692        loff_t pos = sd->pos;
 693        int more;
 694
 695        if (!likely(file->f_op && file->f_op->sendpage))
 696                return -EINVAL;
 697
 698        more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 699        if (sd->len < sd->total_len)
 700                more |= MSG_SENDPAGE_NOTLAST;
 701        return file->f_op->sendpage(file, buf->page, buf->offset,
 702                                    sd->len, &pos, more);
 703}
 704
 705/*
 706 * This is a little more tricky than the file -> pipe splicing. There are
 707 * basically three cases:
 708 *
 709 *      - Destination page already exists in the address space and there
 710 *        are users of it. For that case we have no other option that
 711 *        copying the data. Tough luck.
 712 *      - Destination page already exists in the address space, but there
 713 *        are no users of it. Make sure it's uptodate, then drop it. Fall
 714 *        through to last case.
 715 *      - Destination page does not exist, we can add the pipe page to
 716 *        the page cache and avoid the copy.
 717 *
 718 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
 719 * sd->flags), we attempt to migrate pages from the pipe to the output
 720 * file address space page cache. This is possible if no one else has
 721 * the pipe page referenced outside of the pipe and page cache. If
 722 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
 723 * a new page in the output file page cache and fill/dirty that.
 724 */
 725int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
 726                 struct splice_desc *sd)
 727{
 728        struct file *file = sd->u.file;
 729        struct address_space *mapping = file->f_mapping;
 730        unsigned int offset, this_len;
 731        struct page *page;
 732        void *fsdata;
 733        int ret;
 734
 735        offset = sd->pos & ~PAGE_CACHE_MASK;
 736
 737        this_len = sd->len;
 738        if (this_len + offset > PAGE_CACHE_SIZE)
 739                this_len = PAGE_CACHE_SIZE - offset;
 740
 741        ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
 742                                AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
 743        if (unlikely(ret))
 744                goto out;
 745
 746        if (buf->page != page) {
 747                char *src = buf->ops->map(pipe, buf, 1);
 748                char *dst = kmap_atomic(page);
 749
 750                memcpy(dst + offset, src + buf->offset, this_len);
 751                flush_dcache_page(page);
 752                kunmap_atomic(dst);
 753                buf->ops->unmap(pipe, buf, src);
 754        }
 755        ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
 756                                page, fsdata);
 757out:
 758        return ret;
 759}
 760EXPORT_SYMBOL(pipe_to_file);
 761
 762static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 763{
 764        smp_mb();
 765        if (waitqueue_active(&pipe->wait))
 766                wake_up_interruptible(&pipe->wait);
 767        kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
 768}
 769
 770/**
 771 * splice_from_pipe_feed - feed available data from a pipe to a file
 772 * @pipe:       pipe to splice from
 773 * @sd:         information to @actor
 774 * @actor:      handler that splices the data
 775 *
 776 * Description:
 777 *    This function loops over the pipe and calls @actor to do the
 778 *    actual moving of a single struct pipe_buffer to the desired
 779 *    destination.  It returns when there's no more buffers left in
 780 *    the pipe or if the requested number of bytes (@sd->total_len)
 781 *    have been copied.  It returns a positive number (one) if the
 782 *    pipe needs to be filled with more data, zero if the required
 783 *    number of bytes have been copied and -errno on error.
 784 *
 785 *    This, together with splice_from_pipe_{begin,end,next}, may be
 786 *    used to implement the functionality of __splice_from_pipe() when
 787 *    locking is required around copying the pipe buffers to the
 788 *    destination.
 789 */
 790int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
 791                          splice_actor *actor)
 792{
 793        int ret;
 794
 795        while (pipe->nrbufs) {
 796                struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
 797                const struct pipe_buf_operations *ops = buf->ops;
 798
 799                sd->len = buf->len;
 800                if (sd->len > sd->total_len)
 801                        sd->len = sd->total_len;
 802
 803                ret = buf->ops->confirm(pipe, buf);
 804                if (unlikely(ret)) {
 805                        if (ret == -ENODATA)
 806                                ret = 0;
 807                        return ret;
 808                }
 809
 810                ret = actor(pipe, buf, sd);
 811                if (ret <= 0)
 812                        return ret;
 813
 814                buf->offset += ret;
 815                buf->len -= ret;
 816
 817                sd->num_spliced += ret;
 818                sd->len -= ret;
 819                sd->pos += ret;
 820                sd->total_len -= ret;
 821
 822                if (!buf->len) {
 823                        buf->ops = NULL;
 824                        ops->release(pipe, buf);
 825                        pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
 826                        pipe->nrbufs--;
 827                        if (pipe->inode)
 828                                sd->need_wakeup = true;
 829                }
 830
 831                if (!sd->total_len)
 832                        return 0;
 833        }
 834
 835        return 1;
 836}
 837EXPORT_SYMBOL(splice_from_pipe_feed);
 838
 839/**
 840 * splice_from_pipe_next - wait for some data to splice from
 841 * @pipe:       pipe to splice from
 842 * @sd:         information about the splice operation
 843 *
 844 * Description:
 845 *    This function will wait for some data and return a positive
 846 *    value (one) if pipe buffers are available.  It will return zero
 847 *    or -errno if no more data needs to be spliced.
 848 */
 849int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
 850{
 851        while (!pipe->nrbufs) {
 852                if (!pipe->writers)
 853                        return 0;
 854
 855                if (!pipe->waiting_writers && sd->num_spliced)
 856                        return 0;
 857
 858                if (sd->flags & SPLICE_F_NONBLOCK)
 859                        return -EAGAIN;
 860
 861                if (signal_pending(current))
 862                        return -ERESTARTSYS;
 863
 864                if (sd->need_wakeup) {
 865                        wakeup_pipe_writers(pipe);
 866                        sd->need_wakeup = false;
 867                }
 868
 869                pipe_wait(pipe);
 870        }
 871
 872        return 1;
 873}
 874EXPORT_SYMBOL(splice_from_pipe_next);
 875
 876/**
 877 * splice_from_pipe_begin - start splicing from pipe
 878 * @sd:         information about the splice operation
 879 *
 880 * Description:
 881 *    This function should be called before a loop containing
 882 *    splice_from_pipe_next() and splice_from_pipe_feed() to
 883 *    initialize the necessary fields of @sd.
 884 */
 885void splice_from_pipe_begin(struct splice_desc *sd)
 886{
 887        sd->num_spliced = 0;
 888        sd->need_wakeup = false;
 889}
 890EXPORT_SYMBOL(splice_from_pipe_begin);
 891
 892/**
 893 * splice_from_pipe_end - finish splicing from pipe
 894 * @pipe:       pipe to splice from
 895 * @sd:         information about the splice operation
 896 *
 897 * Description:
 898 *    This function will wake up pipe writers if necessary.  It should
 899 *    be called after a loop containing splice_from_pipe_next() and
 900 *    splice_from_pipe_feed().
 901 */
 902void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
 903{
 904        if (sd->need_wakeup)
 905                wakeup_pipe_writers(pipe);
 906}
 907EXPORT_SYMBOL(splice_from_pipe_end);
 908
 909/**
 910 * __splice_from_pipe - splice data from a pipe to given actor
 911 * @pipe:       pipe to splice from
 912 * @sd:         information to @actor
 913 * @actor:      handler that splices the data
 914 *
 915 * Description:
 916 *    This function does little more than loop over the pipe and call
 917 *    @actor to do the actual moving of a single struct pipe_buffer to
 918 *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
 919 *    pipe_to_user.
 920 *
 921 */
 922ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
 923                           splice_actor *actor)
 924{
 925        int ret;
 926
 927        splice_from_pipe_begin(sd);
 928        do {
 929                ret = splice_from_pipe_next(pipe, sd);
 930                if (ret > 0)
 931                        ret = splice_from_pipe_feed(pipe, sd, actor);
 932        } while (ret > 0);
 933        splice_from_pipe_end(pipe, sd);
 934
 935        return sd->num_spliced ? sd->num_spliced : ret;
 936}
 937EXPORT_SYMBOL(__splice_from_pipe);
 938
 939/**
 940 * splice_from_pipe - splice data from a pipe to a file
 941 * @pipe:       pipe to splice from
 942 * @out:        file to splice to
 943 * @ppos:       position in @out
 944 * @len:        how many bytes to splice
 945 * @flags:      splice modifier flags
 946 * @actor:      handler that splices the data
 947 *
 948 * Description:
 949 *    See __splice_from_pipe. This function locks the pipe inode,
 950 *    otherwise it's identical to __splice_from_pipe().
 951 *
 952 */
 953ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
 954                         loff_t *ppos, size_t len, unsigned int flags,
 955                         splice_actor *actor)
 956{
 957        ssize_t ret;
 958        struct splice_desc sd = {
 959                .total_len = len,
 960                .flags = flags,
 961                .pos = *ppos,
 962                .u.file = out,
 963        };
 964
 965        pipe_lock(pipe);
 966        ret = __splice_from_pipe(pipe, &sd, actor);
 967        pipe_unlock(pipe);
 968
 969        return ret;
 970}
 971
 972/**
 973 * generic_file_splice_write - splice data from a pipe to a file
 974 * @pipe:       pipe info
 975 * @out:        file to write to
 976 * @ppos:       position in @out
 977 * @len:        number of bytes to splice
 978 * @flags:      splice modifier flags
 979 *
 980 * Description:
 981 *    Will either move or copy pages (determined by @flags options) from
 982 *    the given pipe inode to the given file.
 983 *
 984 */
 985ssize_t
 986generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 987                          loff_t *ppos, size_t len, unsigned int flags)
 988{
 989        struct address_space *mapping = out->f_mapping;
 990        struct inode *inode = mapping->host;
 991        struct splice_desc sd = {
 992                .total_len = len,
 993                .flags = flags,
 994                .pos = *ppos,
 995                .u.file = out,
 996        };
 997        ssize_t ret;
 998
 999        sb_start_write(inode->i_sb);
1000
1001        pipe_lock(pipe);
1002
1003        splice_from_pipe_begin(&sd);
1004        do {
1005                ret = splice_from_pipe_next(pipe, &sd);
1006                if (ret <= 0)
1007                        break;
1008
1009                mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1010                ret = file_remove_suid(out);
1011                if (!ret) {
1012                        ret = file_update_time(out);
1013                        if (!ret)
1014                                ret = splice_from_pipe_feed(pipe, &sd,
1015                                                            pipe_to_file);
1016                }
1017                mutex_unlock(&inode->i_mutex);
1018        } while (ret > 0);
1019        splice_from_pipe_end(pipe, &sd);
1020
1021        pipe_unlock(pipe);
1022
1023        if (sd.num_spliced)
1024                ret = sd.num_spliced;
1025
1026        if (ret > 0) {
1027                unsigned long nr_pages;
1028                int err;
1029
1030                nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1031
1032                err = generic_write_sync(out, *ppos, ret);
1033                if (err)
1034                        ret = err;
1035                else
1036                        *ppos += ret;
1037                balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1038        }
1039        sb_end_write(inode->i_sb);
1040
1041        return ret;
1042}
1043
1044EXPORT_SYMBOL(generic_file_splice_write);
1045
1046static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1047                          struct splice_desc *sd)
1048{
1049        int ret;
1050        void *data;
1051
1052        data = buf->ops->map(pipe, buf, 0);
1053        ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1054        buf->ops->unmap(pipe, buf, data);
1055
1056        return ret;
1057}
1058
1059static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1060                                         struct file *out, loff_t *ppos,
1061                                         size_t len, unsigned int flags)
1062{
1063        ssize_t ret;
1064
1065        ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1066        if (ret > 0)
1067                *ppos += ret;
1068
1069        return ret;
1070}
1071
1072/**
1073 * generic_splice_sendpage - splice data from a pipe to a socket
1074 * @pipe:       pipe to splice from
1075 * @out:        socket to write to
1076 * @ppos:       position in @out
1077 * @len:        number of bytes to splice
1078 * @flags:      splice modifier flags
1079 *
1080 * Description:
1081 *    Will send @len bytes from the pipe to a network socket. No data copying
1082 *    is involved.
1083 *
1084 */
1085ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1086                                loff_t *ppos, size_t len, unsigned int flags)
1087{
1088        return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1089}
1090
1091EXPORT_SYMBOL(generic_splice_sendpage);
1092
1093/*
1094 * Attempt to initiate a splice from pipe to file.
1095 */
1096static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1097                           loff_t *ppos, size_t len, unsigned int flags)
1098{
1099        ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1100                                loff_t *, size_t, unsigned int);
1101        int ret;
1102
1103        if (unlikely(!(out->f_mode & FMODE_WRITE)))
1104                return -EBADF;
1105
1106        if (unlikely(out->f_flags & O_APPEND))
1107                return -EINVAL;
1108
1109        ret = rw_verify_area(WRITE, out, ppos, len);
1110        if (unlikely(ret < 0))
1111                return ret;
1112
1113        if (out->f_op && out->f_op->splice_write)
1114                splice_write = out->f_op->splice_write;
1115        else
1116                splice_write = default_file_splice_write;
1117
1118        return splice_write(pipe, out, ppos, len, flags);
1119}
1120
1121/*
1122 * Attempt to initiate a splice from a file to a pipe.
1123 */
1124static long do_splice_to(struct file *in, loff_t *ppos,
1125                         struct pipe_inode_info *pipe, size_t len,
1126                         unsigned int flags)
1127{
1128        ssize_t (*splice_read)(struct file *, loff_t *,
1129                               struct pipe_inode_info *, size_t, unsigned int);
1130        int ret;
1131
1132        if (unlikely(!(in->f_mode & FMODE_READ)))
1133                return -EBADF;
1134
1135        ret = rw_verify_area(READ, in, ppos, len);
1136        if (unlikely(ret < 0))
1137                return ret;
1138
1139        if (in->f_op && in->f_op->splice_read)
1140                splice_read = in->f_op->splice_read;
1141        else
1142                splice_read = default_file_splice_read;
1143
1144        return splice_read(in, ppos, pipe, len, flags);
1145}
1146
1147/**
1148 * splice_direct_to_actor - splices data directly between two non-pipes
1149 * @in:         file to splice from
1150 * @sd:         actor information on where to splice to
1151 * @actor:      handles the data splicing
1152 *
1153 * Description:
1154 *    This is a special case helper to splice directly between two
1155 *    points, without requiring an explicit pipe. Internally an allocated
1156 *    pipe is cached in the process, and reused during the lifetime of
1157 *    that process.
1158 *
1159 */
1160ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1161                               splice_direct_actor *actor)
1162{
1163        struct pipe_inode_info *pipe;
1164        long ret, bytes;
1165        umode_t i_mode;
1166        size_t len;
1167        int i, flags;
1168
1169        /*
1170         * We require the input being a regular file, as we don't want to
1171         * randomly drop data for eg socket -> socket splicing. Use the
1172         * piped splicing for that!
1173         */
1174        i_mode = in->f_path.dentry->d_inode->i_mode;
1175        if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1176                return -EINVAL;
1177
1178        /*
1179         * neither in nor out is a pipe, setup an internal pipe attached to
1180         * 'out' and transfer the wanted data from 'in' to 'out' through that
1181         */
1182        pipe = current->splice_pipe;
1183        if (unlikely(!pipe)) {
1184                pipe = alloc_pipe_info(NULL);
1185                if (!pipe)
1186                        return -ENOMEM;
1187
1188                /*
1189                 * We don't have an immediate reader, but we'll read the stuff
1190                 * out of the pipe right after the splice_to_pipe(). So set
1191                 * PIPE_READERS appropriately.
1192                 */
1193                pipe->readers = 1;
1194
1195                current->splice_pipe = pipe;
1196        }
1197
1198        /*
1199         * Do the splice.
1200         */
1201        ret = 0;
1202        bytes = 0;
1203        len = sd->total_len;
1204        flags = sd->flags;
1205
1206        /*
1207         * Don't block on output, we have to drain the direct pipe.
1208         */
1209        sd->flags &= ~SPLICE_F_NONBLOCK;
1210
1211        while (len) {
1212                size_t read_len;
1213                loff_t pos = sd->pos, prev_pos = pos;
1214
1215                ret = do_splice_to(in, &pos, pipe, len, flags);
1216                if (unlikely(ret <= 0))
1217                        goto out_release;
1218
1219                read_len = ret;
1220                sd->total_len = read_len;
1221
1222                /*
1223                 * NOTE: nonblocking mode only applies to the input. We
1224                 * must not do the output in nonblocking mode as then we
1225                 * could get stuck data in the internal pipe:
1226                 */
1227                ret = actor(pipe, sd);
1228                if (unlikely(ret <= 0)) {
1229                        sd->pos = prev_pos;
1230                        goto out_release;
1231                }
1232
1233                bytes += ret;
1234                len -= ret;
1235                sd->pos = pos;
1236
1237                if (ret < read_len) {
1238                        sd->pos = prev_pos + ret;
1239                        goto out_release;
1240                }
1241        }
1242
1243done:
1244        pipe->nrbufs = pipe->curbuf = 0;
1245        file_accessed(in);
1246        return bytes;
1247
1248out_release:
1249        /*
1250         * If we did an incomplete transfer we must release
1251         * the pipe buffers in question:
1252         */
1253        for (i = 0; i < pipe->buffers; i++) {
1254                struct pipe_buffer *buf = pipe->bufs + i;
1255
1256                if (buf->ops) {
1257                        buf->ops->release(pipe, buf);
1258                        buf->ops = NULL;
1259                }
1260        }
1261
1262        if (!bytes)
1263                bytes = ret;
1264
1265        goto done;
1266}
1267EXPORT_SYMBOL(splice_direct_to_actor);
1268
1269static int direct_splice_actor(struct pipe_inode_info *pipe,
1270                               struct splice_desc *sd)
1271{
1272        struct file *file = sd->u.file;
1273
1274        return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1275                              sd->flags);
1276}
1277
1278/**
1279 * do_splice_direct - splices data directly between two files
1280 * @in:         file to splice from
1281 * @ppos:       input file offset
1282 * @out:        file to splice to
1283 * @len:        number of bytes to splice
1284 * @flags:      splice modifier flags
1285 *
1286 * Description:
1287 *    For use by do_sendfile(). splice can easily emulate sendfile, but
1288 *    doing it in the application would incur an extra system call
1289 *    (splice in + splice out, as compared to just sendfile()). So this helper
1290 *    can splice directly through a process-private pipe.
1291 *
1292 */
1293long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1294                      size_t len, unsigned int flags)
1295{
1296        struct splice_desc sd = {
1297                .len            = len,
1298                .total_len      = len,
1299                .flags          = flags,
1300                .pos            = *ppos,
1301                .u.file         = out,
1302        };
1303        long ret;
1304
1305        ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1306        if (ret > 0)
1307                *ppos = sd.pos;
1308
1309        return ret;
1310}
1311
1312static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1313                               struct pipe_inode_info *opipe,
1314                               size_t len, unsigned int flags);
1315
1316/*
1317 * Determine where to splice to/from.
1318 */
1319static long do_splice(struct file *in, loff_t __user *off_in,
1320                      struct file *out, loff_t __user *off_out,
1321                      size_t len, unsigned int flags)
1322{
1323        struct pipe_inode_info *ipipe;
1324        struct pipe_inode_info *opipe;
1325        loff_t offset, *off;
1326        long ret;
1327
1328        ipipe = get_pipe_info(in);
1329        opipe = get_pipe_info(out);
1330
1331        if (ipipe && opipe) {
1332                if (off_in || off_out)
1333                        return -ESPIPE;
1334
1335                if (!(in->f_mode & FMODE_READ))
1336                        return -EBADF;
1337
1338                if (!(out->f_mode & FMODE_WRITE))
1339                        return -EBADF;
1340
1341                /* Splicing to self would be fun, but... */
1342                if (ipipe == opipe)
1343                        return -EINVAL;
1344
1345                return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1346        }
1347
1348        if (ipipe) {
1349                if (off_in)
1350                        return -ESPIPE;
1351                if (off_out) {
1352                        if (!(out->f_mode & FMODE_PWRITE))
1353                                return -EINVAL;
1354                        if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1355                                return -EFAULT;
1356                        off = &offset;
1357                } else
1358                        off = &out->f_pos;
1359
1360                ret = do_splice_from(ipipe, out, off, len, flags);
1361
1362                if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1363                        ret = -EFAULT;
1364
1365                return ret;
1366        }
1367
1368        if (opipe) {
1369                if (off_out)
1370                        return -ESPIPE;
1371                if (off_in) {
1372                        if (!(in->f_mode & FMODE_PREAD))
1373                                return -EINVAL;
1374                        if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1375                                return -EFAULT;
1376                        off = &offset;
1377                } else
1378                        off = &in->f_pos;
1379
1380                ret = do_splice_to(in, off, opipe, len, flags);
1381
1382                if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1383                        ret = -EFAULT;
1384
1385                return ret;
1386        }
1387
1388        return -EINVAL;
1389}
1390
1391/*
1392 * Map an iov into an array of pages and offset/length tupples. With the
1393 * partial_page structure, we can map several non-contiguous ranges into
1394 * our ones pages[] map instead of splitting that operation into pieces.
1395 * Could easily be exported as a generic helper for other users, in which
1396 * case one would probably want to add a 'max_nr_pages' parameter as well.
1397 */
1398static int get_iovec_page_array(const struct iovec __user *iov,
1399                                unsigned int nr_vecs, struct page **pages,
1400                                struct partial_page *partial, bool aligned,
1401                                unsigned int pipe_buffers)
1402{
1403        int buffers = 0, error = 0;
1404
1405        while (nr_vecs) {
1406                unsigned long off, npages;
1407                struct iovec entry;
1408                void __user *base;
1409                size_t len;
1410                int i;
1411
1412                error = -EFAULT;
1413                if (copy_from_user(&entry, iov, sizeof(entry)))
1414                        break;
1415
1416                base = entry.iov_base;
1417                len = entry.iov_len;
1418
1419                /*
1420                 * Sanity check this iovec. 0 read succeeds.
1421                 */
1422                error = 0;
1423                if (unlikely(!len))
1424                        break;
1425                error = -EFAULT;
1426                if (!access_ok(VERIFY_READ, base, len))
1427                        break;
1428
1429                /*
1430                 * Get this base offset and number of pages, then map
1431                 * in the user pages.
1432                 */
1433                off = (unsigned long) base & ~PAGE_MASK;
1434
1435                /*
1436                 * If asked for alignment, the offset must be zero and the
1437                 * length a multiple of the PAGE_SIZE.
1438                 */
1439                error = -EINVAL;
1440                if (aligned && (off || len & ~PAGE_MASK))
1441                        break;
1442
1443                npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1444                if (npages > pipe_buffers - buffers)
1445                        npages = pipe_buffers - buffers;
1446
1447                error = get_user_pages_fast((unsigned long)base, npages,
1448                                        0, &pages[buffers]);
1449
1450                if (unlikely(error <= 0))
1451                        break;
1452
1453                /*
1454                 * Fill this contiguous range into the partial page map.
1455                 */
1456                for (i = 0; i < error; i++) {
1457                        const int plen = min_t(size_t, len, PAGE_SIZE - off);
1458
1459                        partial[buffers].offset = off;
1460                        partial[buffers].len = plen;
1461
1462                        off = 0;
1463                        len -= plen;
1464                        buffers++;
1465                }
1466
1467                /*
1468                 * We didn't complete this iov, stop here since it probably
1469                 * means we have to move some of this into a pipe to
1470                 * be able to continue.
1471                 */
1472                if (len)
1473                        break;
1474
1475                /*
1476                 * Don't continue if we mapped fewer pages than we asked for,
1477                 * or if we mapped the max number of pages that we have
1478                 * room for.
1479                 */
1480                if (error < npages || buffers == pipe_buffers)
1481                        break;
1482
1483                nr_vecs--;
1484                iov++;
1485        }
1486
1487        if (buffers)
1488                return buffers;
1489
1490        return error;
1491}
1492
1493static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1494                        struct splice_desc *sd)
1495{
1496        char *src;
1497        int ret;
1498
1499        /*
1500         * See if we can use the atomic maps, by prefaulting in the
1501         * pages and doing an atomic copy
1502         */
1503        if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1504                src = buf->ops->map(pipe, buf, 1);
1505                ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1506                                                        sd->len);
1507                buf->ops->unmap(pipe, buf, src);
1508                if (!ret) {
1509                        ret = sd->len;
1510                        goto out;
1511                }
1512        }
1513
1514        /*
1515         * No dice, use slow non-atomic map and copy
1516         */
1517        src = buf->ops->map(pipe, buf, 0);
1518
1519        ret = sd->len;
1520        if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1521                ret = -EFAULT;
1522
1523        buf->ops->unmap(pipe, buf, src);
1524out:
1525        if (ret > 0)
1526                sd->u.userptr += ret;
1527        return ret;
1528}
1529
1530/*
1531 * For lack of a better implementation, implement vmsplice() to userspace
1532 * as a simple copy of the pipes pages to the user iov.
1533 */
1534static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1535                             unsigned long nr_segs, unsigned int flags)
1536{
1537        struct pipe_inode_info *pipe;
1538        struct splice_desc sd;
1539        ssize_t size;
1540        int error;
1541        long ret;
1542
1543        pipe = get_pipe_info(file);
1544        if (!pipe)
1545                return -EBADF;
1546
1547        pipe_lock(pipe);
1548
1549        error = ret = 0;
1550        while (nr_segs) {
1551                void __user *base;
1552                size_t len;
1553
1554                /*
1555                 * Get user address base and length for this iovec.
1556                 */
1557                error = get_user(base, &iov->iov_base);
1558                if (unlikely(error))
1559                        break;
1560                error = get_user(len, &iov->iov_len);
1561                if (unlikely(error))
1562                        break;
1563
1564                /*
1565                 * Sanity check this iovec. 0 read succeeds.
1566                 */
1567                if (unlikely(!len))
1568                        break;
1569                if (unlikely(!base)) {
1570                        error = -EFAULT;
1571                        break;
1572                }
1573
1574                if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1575                        error = -EFAULT;
1576                        break;
1577                }
1578
1579                sd.len = 0;
1580                sd.total_len = len;
1581                sd.flags = flags;
1582                sd.u.userptr = base;
1583                sd.pos = 0;
1584
1585                size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1586                if (size < 0) {
1587                        if (!ret)
1588                                ret = size;
1589
1590                        break;
1591                }
1592
1593                ret += size;
1594
1595                if (size < len)
1596                        break;
1597
1598                nr_segs--;
1599                iov++;
1600        }
1601
1602        pipe_unlock(pipe);
1603
1604        if (!ret)
1605                ret = error;
1606
1607        return ret;
1608}
1609
1610/*
1611 * vmsplice splices a user address range into a pipe. It can be thought of
1612 * as splice-from-memory, where the regular splice is splice-from-file (or
1613 * to file). In both cases the output is a pipe, naturally.
1614 */
1615static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1616                             unsigned long nr_segs, unsigned int flags)
1617{
1618        struct pipe_inode_info *pipe;
1619        struct page *pages[PIPE_DEF_BUFFERS];
1620        struct partial_page partial[PIPE_DEF_BUFFERS];
1621        struct splice_pipe_desc spd = {
1622                .pages = pages,
1623                .partial = partial,
1624                .nr_pages_max = PIPE_DEF_BUFFERS,
1625                .flags = flags,
1626                .ops = &user_page_pipe_buf_ops,
1627                .spd_release = spd_release_page,
1628        };
1629        long ret;
1630
1631        pipe = get_pipe_info(file);
1632        if (!pipe)
1633                return -EBADF;
1634
1635        if (splice_grow_spd(pipe, &spd))
1636                return -ENOMEM;
1637
1638        spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1639                                            spd.partial, false,
1640                                            spd.nr_pages_max);
1641        if (spd.nr_pages <= 0)
1642                ret = spd.nr_pages;
1643        else
1644                ret = splice_to_pipe(pipe, &spd);
1645
1646        splice_shrink_spd(&spd);
1647        return ret;
1648}
1649
1650/*
1651 * Note that vmsplice only really supports true splicing _from_ user memory
1652 * to a pipe, not the other way around. Splicing from user memory is a simple
1653 * operation that can be supported without any funky alignment restrictions
1654 * or nasty vm tricks. We simply map in the user memory and fill them into
1655 * a pipe. The reverse isn't quite as easy, though. There are two possible
1656 * solutions for that:
1657 *
1658 *      - memcpy() the data internally, at which point we might as well just
1659 *        do a regular read() on the buffer anyway.
1660 *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1661 *        has restriction limitations on both ends of the pipe).
1662 *
1663 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1664 *
1665 */
1666SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1667                unsigned long, nr_segs, unsigned int, flags)
1668{
1669        struct file *file;
1670        long error;
1671        int fput;
1672
1673        if (unlikely(nr_segs > UIO_MAXIOV))
1674                return -EINVAL;
1675        else if (unlikely(!nr_segs))
1676                return 0;
1677
1678        error = -EBADF;
1679        file = fget_light(fd, &fput);
1680        if (file) {
1681                if (file->f_mode & FMODE_WRITE)
1682                        error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1683                else if (file->f_mode & FMODE_READ)
1684                        error = vmsplice_to_user(file, iov, nr_segs, flags);
1685
1686                fput_light(file, fput);
1687        }
1688
1689        return error;
1690}
1691
1692SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1693                int, fd_out, loff_t __user *, off_out,
1694                size_t, len, unsigned int, flags)
1695{
1696        long error;
1697        struct file *in, *out;
1698        int fput_in, fput_out;
1699
1700        if (unlikely(!len))
1701                return 0;
1702
1703        error = -EBADF;
1704        in = fget_light(fd_in, &fput_in);
1705        if (in) {
1706                if (in->f_mode & FMODE_READ) {
1707                        out = fget_light(fd_out, &fput_out);
1708                        if (out) {
1709                                if (out->f_mode & FMODE_WRITE)
1710                                        error = do_splice(in, off_in,
1711                                                          out, off_out,
1712                                                          len, flags);
1713                                fput_light(out, fput_out);
1714                        }
1715                }
1716
1717                fput_light(in, fput_in);
1718        }
1719
1720        return error;
1721}
1722
1723/*
1724 * Make sure there's data to read. Wait for input if we can, otherwise
1725 * return an appropriate error.
1726 */
1727static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1728{
1729        int ret;
1730
1731        /*
1732         * Check ->nrbufs without the inode lock first. This function
1733         * is speculative anyways, so missing one is ok.
1734         */
1735        if (pipe->nrbufs)
1736                return 0;
1737
1738        ret = 0;
1739        pipe_lock(pipe);
1740
1741        while (!pipe->nrbufs) {
1742                if (signal_pending(current)) {
1743                        ret = -ERESTARTSYS;
1744                        break;
1745                }
1746                if (!pipe->writers)
1747                        break;
1748                if (!pipe->waiting_writers) {
1749                        if (flags & SPLICE_F_NONBLOCK) {
1750                                ret = -EAGAIN;
1751                                break;
1752                        }
1753                }
1754                pipe_wait(pipe);
1755        }
1756
1757        pipe_unlock(pipe);
1758        return ret;
1759}
1760
1761/*
1762 * Make sure there's writeable room. Wait for room if we can, otherwise
1763 * return an appropriate error.
1764 */
1765static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1766{
1767        int ret;
1768
1769        /*
1770         * Check ->nrbufs without the inode lock first. This function
1771         * is speculative anyways, so missing one is ok.
1772         */
1773        if (pipe->nrbufs < pipe->buffers)
1774                return 0;
1775
1776        ret = 0;
1777        pipe_lock(pipe);
1778
1779        while (pipe->nrbufs >= pipe->buffers) {
1780                if (!pipe->readers) {
1781                        send_sig(SIGPIPE, current, 0);
1782                        ret = -EPIPE;
1783                        break;
1784                }
1785                if (flags & SPLICE_F_NONBLOCK) {
1786                        ret = -EAGAIN;
1787                        break;
1788                }
1789                if (signal_pending(current)) {
1790                        ret = -ERESTARTSYS;
1791                        break;
1792                }
1793                pipe->waiting_writers++;
1794                pipe_wait(pipe);
1795                pipe->waiting_writers--;
1796        }
1797
1798        pipe_unlock(pipe);
1799        return ret;
1800}
1801
1802/*
1803 * Splice contents of ipipe to opipe.
1804 */
1805static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1806                               struct pipe_inode_info *opipe,
1807                               size_t len, unsigned int flags)
1808{
1809        struct pipe_buffer *ibuf, *obuf;
1810        int ret = 0, nbuf;
1811        bool input_wakeup = false;
1812
1813
1814retry:
1815        ret = ipipe_prep(ipipe, flags);
1816        if (ret)
1817                return ret;
1818
1819        ret = opipe_prep(opipe, flags);
1820        if (ret)
1821                return ret;
1822
1823        /*
1824         * Potential ABBA deadlock, work around it by ordering lock
1825         * grabbing by pipe info address. Otherwise two different processes
1826         * could deadlock (one doing tee from A -> B, the other from B -> A).
1827         */
1828        pipe_double_lock(ipipe, opipe);
1829
1830        do {
1831                if (!opipe->readers) {
1832                        send_sig(SIGPIPE, current, 0);
1833                        if (!ret)
1834                                ret = -EPIPE;
1835                        break;
1836                }
1837
1838                if (!ipipe->nrbufs && !ipipe->writers)
1839                        break;
1840
1841                /*
1842                 * Cannot make any progress, because either the input
1843                 * pipe is empty or the output pipe is full.
1844                 */
1845                if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1846                        /* Already processed some buffers, break */
1847                        if (ret)
1848                                break;
1849
1850                        if (flags & SPLICE_F_NONBLOCK) {
1851                                ret = -EAGAIN;
1852                                break;
1853                        }
1854
1855                        /*
1856                         * We raced with another reader/writer and haven't
1857                         * managed to process any buffers.  A zero return
1858                         * value means EOF, so retry instead.
1859                         */
1860                        pipe_unlock(ipipe);
1861                        pipe_unlock(opipe);
1862                        goto retry;
1863                }
1864
1865                ibuf = ipipe->bufs + ipipe->curbuf;
1866                nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1867                obuf = opipe->bufs + nbuf;
1868
1869                if (len >= ibuf->len) {
1870                        /*
1871                         * Simply move the whole buffer from ipipe to opipe
1872                         */
1873                        *obuf = *ibuf;
1874                        ibuf->ops = NULL;
1875                        opipe->nrbufs++;
1876                        ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1877                        ipipe->nrbufs--;
1878                        input_wakeup = true;
1879                } else {
1880                        /*
1881                         * Get a reference to this pipe buffer,
1882                         * so we can copy the contents over.
1883                         */
1884                        ibuf->ops->get(ipipe, ibuf);
1885                        *obuf = *ibuf;
1886
1887                        /*
1888                         * Don't inherit the gift flag, we need to
1889                         * prevent multiple steals of this page.
1890                         */
1891                        obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1892
1893                        obuf->len = len;
1894                        opipe->nrbufs++;
1895                        ibuf->offset += obuf->len;
1896                        ibuf->len -= obuf->len;
1897                }
1898                ret += obuf->len;
1899                len -= obuf->len;
1900        } while (len);
1901
1902        pipe_unlock(ipipe);
1903        pipe_unlock(opipe);
1904
1905        /*
1906         * If we put data in the output pipe, wakeup any potential readers.
1907         */
1908        if (ret > 0)
1909                wakeup_pipe_readers(opipe);
1910
1911        if (input_wakeup)
1912                wakeup_pipe_writers(ipipe);
1913
1914        return ret;
1915}
1916
1917/*
1918 * Link contents of ipipe to opipe.
1919 */
1920static int link_pipe(struct pipe_inode_info *ipipe,
1921                     struct pipe_inode_info *opipe,
1922                     size_t len, unsigned int flags)
1923{
1924        struct pipe_buffer *ibuf, *obuf;
1925        int ret = 0, i = 0, nbuf;
1926
1927        /*
1928         * Potential ABBA deadlock, work around it by ordering lock
1929         * grabbing by pipe info address. Otherwise two different processes
1930         * could deadlock (one doing tee from A -> B, the other from B -> A).
1931         */
1932        pipe_double_lock(ipipe, opipe);
1933
1934        do {
1935                if (!opipe->readers) {
1936                        send_sig(SIGPIPE, current, 0);
1937                        if (!ret)
1938                                ret = -EPIPE;
1939                        break;
1940                }
1941
1942                /*
1943                 * If we have iterated all input buffers or ran out of
1944                 * output room, break.
1945                 */
1946                if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1947                        break;
1948
1949                ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1950                nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1951
1952                /*
1953                 * Get a reference to this pipe buffer,
1954                 * so we can copy the contents over.
1955                 */
1956                ibuf->ops->get(ipipe, ibuf);
1957
1958                obuf = opipe->bufs + nbuf;
1959                *obuf = *ibuf;
1960
1961                /*
1962                 * Don't inherit the gift flag, we need to
1963                 * prevent multiple steals of this page.
1964                 */
1965                obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1966
1967                if (obuf->len > len)
1968                        obuf->len = len;
1969
1970                opipe->nrbufs++;
1971                ret += obuf->len;
1972                len -= obuf->len;
1973                i++;
1974        } while (len);
1975
1976        /*
1977         * return EAGAIN if we have the potential of some data in the
1978         * future, otherwise just return 0
1979         */
1980        if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1981                ret = -EAGAIN;
1982
1983        pipe_unlock(ipipe);
1984        pipe_unlock(opipe);
1985
1986        /*
1987         * If we put data in the output pipe, wakeup any potential readers.
1988         */
1989        if (ret > 0)
1990                wakeup_pipe_readers(opipe);
1991
1992        return ret;
1993}
1994
1995/*
1996 * This is a tee(1) implementation that works on pipes. It doesn't copy
1997 * any data, it simply references the 'in' pages on the 'out' pipe.
1998 * The 'flags' used are the SPLICE_F_* variants, currently the only
1999 * applicable one is SPLICE_F_NONBLOCK.
2000 */
2001static long do_tee(struct file *in, struct file *out, size_t len,
2002                   unsigned int flags)
2003{
2004        struct pipe_inode_info *ipipe = get_pipe_info(in);
2005        struct pipe_inode_info *opipe = get_pipe_info(out);
2006        int ret = -EINVAL;
2007
2008        /*
2009         * Duplicate the contents of ipipe to opipe without actually
2010         * copying the data.
2011         */
2012        if (ipipe && opipe && ipipe != opipe) {
2013                /*
2014                 * Keep going, unless we encounter an error. The ipipe/opipe
2015                 * ordering doesn't really matter.
2016                 */
2017                ret = ipipe_prep(ipipe, flags);
2018                if (!ret) {
2019                        ret = opipe_prep(opipe, flags);
2020                        if (!ret)
2021                                ret = link_pipe(ipipe, opipe, len, flags);
2022                }
2023        }
2024
2025        return ret;
2026}
2027
2028SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2029{
2030        struct file *in;
2031        int error, fput_in;
2032
2033        if (unlikely(!len))
2034                return 0;
2035
2036        error = -EBADF;
2037        in = fget_light(fdin, &fput_in);
2038        if (in) {
2039                if (in->f_mode & FMODE_READ) {
2040                        int fput_out;
2041                        struct file *out = fget_light(fdout, &fput_out);
2042
2043                        if (out) {
2044                                if (out->f_mode & FMODE_WRITE)
2045                                        error = do_tee(in, out, len, flags);
2046                                fput_light(out, fput_out);
2047                        }
2048                }
2049                fput_light(in, fput_in);
2050        }
2051
2052        return error;
2053}
2054
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.