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