splice: implement default splice_write method
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, 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/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
33
34 /*
35  * Attempt to steal a page from a pipe buffer. This should perhaps go into
36  * a vm helper function, it's already simplified quite a bit by the
37  * addition of remove_mapping(). If success is returned, the caller may
38  * attempt to reuse this page for another destination.
39  */
40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
41                                      struct pipe_buffer *buf)
42 {
43         struct page *page = buf->page;
44         struct address_space *mapping;
45
46         lock_page(page);
47
48         mapping = page_mapping(page);
49         if (mapping) {
50                 WARN_ON(!PageUptodate(page));
51
52                 /*
53                  * At least for ext2 with nobh option, we need to wait on
54                  * writeback completing on this page, since we'll remove it
55                  * from the pagecache.  Otherwise truncate wont wait on the
56                  * page, allowing the disk blocks to be reused by someone else
57                  * before we actually wrote our data to them. fs corruption
58                  * ensues.
59                  */
60                 wait_on_page_writeback(page);
61
62                 if (page_has_private(page) &&
63                     !try_to_release_page(page, GFP_KERNEL))
64                         goto out_unlock;
65
66                 /*
67                  * If we succeeded in removing the mapping, set LRU flag
68                  * and return good.
69                  */
70                 if (remove_mapping(mapping, page)) {
71                         buf->flags |= PIPE_BUF_FLAG_LRU;
72                         return 0;
73                 }
74         }
75
76         /*
77          * Raced with truncate or failed to remove page from current
78          * address space, unlock and return failure.
79          */
80 out_unlock:
81         unlock_page(page);
82         return 1;
83 }
84
85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
86                                         struct pipe_buffer *buf)
87 {
88         page_cache_release(buf->page);
89         buf->flags &= ~PIPE_BUF_FLAG_LRU;
90 }
91
92 /*
93  * Check whether the contents of buf is OK to access. Since the content
94  * is a page cache page, IO may be in flight.
95  */
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
97                                        struct pipe_buffer *buf)
98 {
99         struct page *page = buf->page;
100         int err;
101
102         if (!PageUptodate(page)) {
103                 lock_page(page);
104
105                 /*
106                  * Page got truncated/unhashed. This will cause a 0-byte
107                  * splice, if this is the first page.
108                  */
109                 if (!page->mapping) {
110                         err = -ENODATA;
111                         goto error;
112                 }
113
114                 /*
115                  * Uh oh, read-error from disk.
116                  */
117                 if (!PageUptodate(page)) {
118                         err = -EIO;
119                         goto error;
120                 }
121
122                 /*
123                  * Page is ok afterall, we are done.
124                  */
125                 unlock_page(page);
126         }
127
128         return 0;
129 error:
130         unlock_page(page);
131         return err;
132 }
133
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
135         .can_merge = 0,
136         .map = generic_pipe_buf_map,
137         .unmap = generic_pipe_buf_unmap,
138         .confirm = page_cache_pipe_buf_confirm,
139         .release = page_cache_pipe_buf_release,
140         .steal = page_cache_pipe_buf_steal,
141         .get = generic_pipe_buf_get,
142 };
143
144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
145                                     struct pipe_buffer *buf)
146 {
147         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
148                 return 1;
149
150         buf->flags |= PIPE_BUF_FLAG_LRU;
151         return generic_pipe_buf_steal(pipe, buf);
152 }
153
154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
155         .can_merge = 0,
156         .map = generic_pipe_buf_map,
157         .unmap = generic_pipe_buf_unmap,
158         .confirm = generic_pipe_buf_confirm,
159         .release = page_cache_pipe_buf_release,
160         .steal = user_page_pipe_buf_steal,
161         .get = generic_pipe_buf_get,
162 };
163
164 /**
165  * splice_to_pipe - fill passed data into a pipe
166  * @pipe:       pipe to fill
167  * @spd:        data to fill
168  *
169  * Description:
170  *    @spd contains a map of pages and len/offset tuples, along with
171  *    the struct pipe_buf_operations associated with these pages. This
172  *    function will link that data to the pipe.
173  *
174  */
175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
176                        struct splice_pipe_desc *spd)
177 {
178         unsigned int spd_pages = spd->nr_pages;
179         int ret, do_wakeup, page_nr;
180
181         ret = 0;
182         do_wakeup = 0;
183         page_nr = 0;
184
185         pipe_lock(pipe);
186
187         for (;;) {
188                 if (!pipe->readers) {
189                         send_sig(SIGPIPE, current, 0);
190                         if (!ret)
191                                 ret = -EPIPE;
192                         break;
193                 }
194
195                 if (pipe->nrbufs < PIPE_BUFFERS) {
196                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
197                         struct pipe_buffer *buf = pipe->bufs + newbuf;
198
199                         buf->page = spd->pages[page_nr];
200                         buf->offset = spd->partial[page_nr].offset;
201                         buf->len = spd->partial[page_nr].len;
202                         buf->private = spd->partial[page_nr].private;
203                         buf->ops = spd->ops;
204                         if (spd->flags & SPLICE_F_GIFT)
205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
206
207                         pipe->nrbufs++;
208                         page_nr++;
209                         ret += buf->len;
210
211                         if (pipe->inode)
212                                 do_wakeup = 1;
213
214                         if (!--spd->nr_pages)
215                                 break;
216                         if (pipe->nrbufs < PIPE_BUFFERS)
217                                 continue;
218
219                         break;
220                 }
221
222                 if (spd->flags & SPLICE_F_NONBLOCK) {
223                         if (!ret)
224                                 ret = -EAGAIN;
225                         break;
226                 }
227
228                 if (signal_pending(current)) {
229                         if (!ret)
230                                 ret = -ERESTARTSYS;
231                         break;
232                 }
233
234                 if (do_wakeup) {
235                         smp_mb();
236                         if (waitqueue_active(&pipe->wait))
237                                 wake_up_interruptible_sync(&pipe->wait);
238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
239                         do_wakeup = 0;
240                 }
241
242                 pipe->waiting_writers++;
243                 pipe_wait(pipe);
244                 pipe->waiting_writers--;
245         }
246
247         pipe_unlock(pipe);
248
249         if (do_wakeup) {
250                 smp_mb();
251                 if (waitqueue_active(&pipe->wait))
252                         wake_up_interruptible(&pipe->wait);
253                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
254         }
255
256         while (page_nr < spd_pages)
257                 spd->spd_release(spd, page_nr++);
258
259         return ret;
260 }
261
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
263 {
264         page_cache_release(spd->pages[i]);
265 }
266
267 static int
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269                            struct pipe_inode_info *pipe, size_t len,
270                            unsigned int flags)
271 {
272         struct address_space *mapping = in->f_mapping;
273         unsigned int loff, nr_pages, req_pages;
274         struct page *pages[PIPE_BUFFERS];
275         struct partial_page partial[PIPE_BUFFERS];
276         struct page *page;
277         pgoff_t index, end_index;
278         loff_t isize;
279         int error, page_nr;
280         struct splice_pipe_desc spd = {
281                 .pages = pages,
282                 .partial = partial,
283                 .flags = flags,
284                 .ops = &page_cache_pipe_buf_ops,
285                 .spd_release = spd_release_page,
286         };
287
288         index = *ppos >> PAGE_CACHE_SHIFT;
289         loff = *ppos & ~PAGE_CACHE_MASK;
290         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
292
293         /*
294          * Lookup the (hopefully) full range of pages we need.
295          */
296         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297         index += spd.nr_pages;
298
299         /*
300          * If find_get_pages_contig() returned fewer pages than we needed,
301          * readahead/allocate the rest and fill in the holes.
302          */
303         if (spd.nr_pages < nr_pages)
304                 page_cache_sync_readahead(mapping, &in->f_ra, in,
305                                 index, req_pages - spd.nr_pages);
306
307         error = 0;
308         while (spd.nr_pages < nr_pages) {
309                 /*
310                  * Page could be there, find_get_pages_contig() breaks on
311                  * the first hole.
312                  */
313                 page = find_get_page(mapping, index);
314                 if (!page) {
315                         /*
316                          * page didn't exist, allocate one.
317                          */
318                         page = page_cache_alloc_cold(mapping);
319                         if (!page)
320                                 break;
321
322                         error = add_to_page_cache_lru(page, mapping, index,
323                                                 mapping_gfp_mask(mapping));
324                         if (unlikely(error)) {
325                                 page_cache_release(page);
326                                 if (error == -EEXIST)
327                                         continue;
328                                 break;
329                         }
330                         /*
331                          * add_to_page_cache() locks the page, unlock it
332                          * to avoid convoluting the logic below even more.
333                          */
334                         unlock_page(page);
335                 }
336
337                 pages[spd.nr_pages++] = page;
338                 index++;
339         }
340
341         /*
342          * Now loop over the map and see if we need to start IO on any
343          * pages, fill in the partial map, etc.
344          */
345         index = *ppos >> PAGE_CACHE_SHIFT;
346         nr_pages = spd.nr_pages;
347         spd.nr_pages = 0;
348         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349                 unsigned int this_len;
350
351                 if (!len)
352                         break;
353
354                 /*
355                  * this_len is the max we'll use from this page
356                  */
357                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358                 page = pages[page_nr];
359
360                 if (PageReadahead(page))
361                         page_cache_async_readahead(mapping, &in->f_ra, in,
362                                         page, index, req_pages - page_nr);
363
364                 /*
365                  * If the page isn't uptodate, we may need to start io on it
366                  */
367                 if (!PageUptodate(page)) {
368                         /*
369                          * If in nonblock mode then dont block on waiting
370                          * for an in-flight io page
371                          */
372                         if (flags & SPLICE_F_NONBLOCK) {
373                                 if (!trylock_page(page)) {
374                                         error = -EAGAIN;
375                                         break;
376                                 }
377                         } else
378                                 lock_page(page);
379
380                         /*
381                          * Page was truncated, or invalidated by the
382                          * filesystem.  Redo the find/create, but this time the
383                          * page is kept locked, so there's no chance of another
384                          * race with truncate/invalidate.
385                          */
386                         if (!page->mapping) {
387                                 unlock_page(page);
388                                 page = find_or_create_page(mapping, index,
389                                                 mapping_gfp_mask(mapping));
390
391                                 if (!page) {
392                                         error = -ENOMEM;
393                                         break;
394                                 }
395                                 page_cache_release(pages[page_nr]);
396                                 pages[page_nr] = page;
397                         }
398                         /*
399                          * page was already under io and is now done, great
400                          */
401                         if (PageUptodate(page)) {
402                                 unlock_page(page);
403                                 goto fill_it;
404                         }
405
406                         /*
407                          * need to read in the page
408                          */
409                         error = mapping->a_ops->readpage(in, page);
410                         if (unlikely(error)) {
411                                 /*
412                                  * We really should re-lookup the page here,
413                                  * but it complicates things a lot. Instead
414                                  * lets just do what we already stored, and
415                                  * we'll get it the next time we are called.
416                                  */
417                                 if (error == AOP_TRUNCATED_PAGE)
418                                         error = 0;
419
420                                 break;
421                         }
422                 }
423 fill_it:
424                 /*
425                  * i_size must be checked after PageUptodate.
426                  */
427                 isize = i_size_read(mapping->host);
428                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
429                 if (unlikely(!isize || index > end_index))
430                         break;
431
432                 /*
433                  * if this is the last page, see if we need to shrink
434                  * the length and stop
435                  */
436                 if (end_index == index) {
437                         unsigned int plen;
438
439                         /*
440                          * max good bytes in this page
441                          */
442                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
443                         if (plen <= loff)
444                                 break;
445
446                         /*
447                          * force quit after adding this page
448                          */
449                         this_len = min(this_len, plen - loff);
450                         len = this_len;
451                 }
452
453                 partial[page_nr].offset = loff;
454                 partial[page_nr].len = this_len;
455                 len -= this_len;
456                 loff = 0;
457                 spd.nr_pages++;
458                 index++;
459         }
460
461         /*
462          * Release any pages at the end, if we quit early. 'page_nr' is how far
463          * we got, 'nr_pages' is how many pages are in the map.
464          */
465         while (page_nr < nr_pages)
466                 page_cache_release(pages[page_nr++]);
467         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
468
469         if (spd.nr_pages)
470                 return splice_to_pipe(pipe, &spd);
471
472         return error;
473 }
474
475 /**
476  * generic_file_splice_read - splice data from file to a pipe
477  * @in:         file to splice from
478  * @ppos:       position in @in
479  * @pipe:       pipe to splice to
480  * @len:        number of bytes to splice
481  * @flags:      splice modifier flags
482  *
483  * Description:
484  *    Will read pages from given file and fill them into a pipe. Can be
485  *    used as long as the address_space operations for the source implements
486  *    a readpage() hook.
487  *
488  */
489 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
490                                  struct pipe_inode_info *pipe, size_t len,
491                                  unsigned int flags)
492 {
493         loff_t isize, left;
494         int ret;
495
496         isize = i_size_read(in->f_mapping->host);
497         if (unlikely(*ppos >= isize))
498                 return 0;
499
500         left = isize - *ppos;
501         if (unlikely(left < len))
502                 len = left;
503
504         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
505         if (ret > 0)
506                 *ppos += ret;
507
508         return ret;
509 }
510 EXPORT_SYMBOL(generic_file_splice_read);
511
512 static const struct pipe_buf_operations default_pipe_buf_ops = {
513         .can_merge = 0,
514         .map = generic_pipe_buf_map,
515         .unmap = generic_pipe_buf_unmap,
516         .confirm = generic_pipe_buf_confirm,
517         .release = generic_pipe_buf_release,
518         .steal = generic_pipe_buf_steal,
519         .get = generic_pipe_buf_get,
520 };
521
522 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
523                             unsigned long vlen, loff_t offset)
524 {
525         mm_segment_t old_fs;
526         loff_t pos = offset;
527         ssize_t res;
528
529         old_fs = get_fs();
530         set_fs(get_ds());
531         /* The cast to a user pointer is valid due to the set_fs() */
532         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
533         set_fs(old_fs);
534
535         return res;
536 }
537
538 static ssize_t kernel_writev(struct file *file, const struct iovec *vec,
539                             unsigned long vlen, loff_t *ppos)
540 {
541         mm_segment_t old_fs;
542         ssize_t res;
543
544         old_fs = get_fs();
545         set_fs(get_ds());
546         /* The cast to a user pointer is valid due to the set_fs() */
547         res = vfs_writev(file, (const struct iovec __user *)vec, vlen, ppos);
548         set_fs(old_fs);
549
550         return res;
551 }
552
553 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
554                                  struct pipe_inode_info *pipe, size_t len,
555                                  unsigned int flags)
556 {
557         unsigned int nr_pages;
558         unsigned int nr_freed;
559         size_t offset;
560         struct page *pages[PIPE_BUFFERS];
561         struct partial_page partial[PIPE_BUFFERS];
562         struct iovec vec[PIPE_BUFFERS];
563         pgoff_t index;
564         ssize_t res;
565         size_t this_len;
566         int error;
567         int i;
568         struct splice_pipe_desc spd = {
569                 .pages = pages,
570                 .partial = partial,
571                 .flags = flags,
572                 .ops = &default_pipe_buf_ops,
573                 .spd_release = spd_release_page,
574         };
575
576         index = *ppos >> PAGE_CACHE_SHIFT;
577         offset = *ppos & ~PAGE_CACHE_MASK;
578         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
579
580         for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
581                 struct page *page;
582
583                 page = alloc_page(GFP_HIGHUSER);
584                 error = -ENOMEM;
585                 if (!page)
586                         goto err;
587
588                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
589                 vec[i].iov_base = (void __user *) kmap(page);
590                 vec[i].iov_len = this_len;
591                 pages[i] = page;
592                 spd.nr_pages++;
593                 len -= this_len;
594                 offset = 0;
595         }
596
597         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
598         if (res < 0)
599                 goto err;
600
601         error = 0;
602         if (!res)
603                 goto err;
604
605         nr_freed = 0;
606         for (i = 0; i < spd.nr_pages; i++) {
607                 kunmap(pages[i]);
608                 this_len = min_t(size_t, vec[i].iov_len, res);
609                 partial[i].offset = 0;
610                 partial[i].len = this_len;
611                 if (!this_len) {
612                         __free_page(pages[i]);
613                         pages[i] = NULL;
614                         nr_freed++;
615                 }
616                 res -= this_len;
617         }
618         spd.nr_pages -= nr_freed;
619
620         res = splice_to_pipe(pipe, &spd);
621         if (res > 0)
622                 *ppos += res;
623
624         return res;
625
626 err:
627         for (i = 0; i < spd.nr_pages; i++) {
628                 kunmap(pages[i]);
629                 __free_page(pages[i]);
630         }
631         return error;
632 }
633 EXPORT_SYMBOL(default_file_splice_read);
634
635 /*
636  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
637  * using sendpage(). Return the number of bytes sent.
638  */
639 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
640                             struct pipe_buffer *buf, struct splice_desc *sd)
641 {
642         struct file *file = sd->u.file;
643         loff_t pos = sd->pos;
644         int ret, more;
645
646         ret = buf->ops->confirm(pipe, buf);
647         if (!ret) {
648                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
649
650                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
651                                            sd->len, &pos, more);
652         }
653
654         return ret;
655 }
656
657 /*
658  * This is a little more tricky than the file -> pipe splicing. There are
659  * basically three cases:
660  *
661  *      - Destination page already exists in the address space and there
662  *        are users of it. For that case we have no other option that
663  *        copying the data. Tough luck.
664  *      - Destination page already exists in the address space, but there
665  *        are no users of it. Make sure it's uptodate, then drop it. Fall
666  *        through to last case.
667  *      - Destination page does not exist, we can add the pipe page to
668  *        the page cache and avoid the copy.
669  *
670  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
671  * sd->flags), we attempt to migrate pages from the pipe to the output
672  * file address space page cache. This is possible if no one else has
673  * the pipe page referenced outside of the pipe and page cache. If
674  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
675  * a new page in the output file page cache and fill/dirty that.
676  */
677 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
678                  struct splice_desc *sd)
679 {
680         struct file *file = sd->u.file;
681         struct address_space *mapping = file->f_mapping;
682         unsigned int offset, this_len;
683         struct page *page;
684         void *fsdata;
685         int ret;
686
687         /*
688          * make sure the data in this buffer is uptodate
689          */
690         ret = buf->ops->confirm(pipe, buf);
691         if (unlikely(ret))
692                 return ret;
693
694         offset = sd->pos & ~PAGE_CACHE_MASK;
695
696         this_len = sd->len;
697         if (this_len + offset > PAGE_CACHE_SIZE)
698                 this_len = PAGE_CACHE_SIZE - offset;
699
700         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
701                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
702         if (unlikely(ret))
703                 goto out;
704
705         if (buf->page != page) {
706                 /*
707                  * Careful, ->map() uses KM_USER0!
708                  */
709                 char *src = buf->ops->map(pipe, buf, 1);
710                 char *dst = kmap_atomic(page, KM_USER1);
711
712                 memcpy(dst + offset, src + buf->offset, this_len);
713                 flush_dcache_page(page);
714                 kunmap_atomic(dst, KM_USER1);
715                 buf->ops->unmap(pipe, buf, src);
716         }
717         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
718                                 page, fsdata);
719 out:
720         return ret;
721 }
722 EXPORT_SYMBOL(pipe_to_file);
723
724 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
725 {
726         smp_mb();
727         if (waitqueue_active(&pipe->wait))
728                 wake_up_interruptible(&pipe->wait);
729         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
730 }
731
732 /**
733  * splice_from_pipe_feed - feed available data from a pipe to a file
734  * @pipe:       pipe to splice from
735  * @sd:         information to @actor
736  * @actor:      handler that splices the data
737  *
738  * Description:
739  *    This function loops over the pipe and calls @actor to do the
740  *    actual moving of a single struct pipe_buffer to the desired
741  *    destination.  It returns when there's no more buffers left in
742  *    the pipe or if the requested number of bytes (@sd->total_len)
743  *    have been copied.  It returns a positive number (one) if the
744  *    pipe needs to be filled with more data, zero if the required
745  *    number of bytes have been copied and -errno on error.
746  *
747  *    This, together with splice_from_pipe_{begin,end,next}, may be
748  *    used to implement the functionality of __splice_from_pipe() when
749  *    locking is required around copying the pipe buffers to the
750  *    destination.
751  */
752 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
753                           splice_actor *actor)
754 {
755         int ret;
756
757         while (pipe->nrbufs) {
758                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
759                 const struct pipe_buf_operations *ops = buf->ops;
760
761                 sd->len = buf->len;
762                 if (sd->len > sd->total_len)
763                         sd->len = sd->total_len;
764
765                 ret = actor(pipe, buf, sd);
766                 if (ret <= 0) {
767                         if (ret == -ENODATA)
768                                 ret = 0;
769                         return ret;
770                 }
771                 buf->offset += ret;
772                 buf->len -= ret;
773
774                 sd->num_spliced += ret;
775                 sd->len -= ret;
776                 sd->pos += ret;
777                 sd->total_len -= ret;
778
779                 if (!buf->len) {
780                         buf->ops = NULL;
781                         ops->release(pipe, buf);
782                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
783                         pipe->nrbufs--;
784                         if (pipe->inode)
785                                 sd->need_wakeup = true;
786                 }
787
788                 if (!sd->total_len)
789                         return 0;
790         }
791
792         return 1;
793 }
794 EXPORT_SYMBOL(splice_from_pipe_feed);
795
796 /**
797  * splice_from_pipe_next - wait for some data to splice from
798  * @pipe:       pipe to splice from
799  * @sd:         information about the splice operation
800  *
801  * Description:
802  *    This function will wait for some data and return a positive
803  *    value (one) if pipe buffers are available.  It will return zero
804  *    or -errno if no more data needs to be spliced.
805  */
806 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
807 {
808         while (!pipe->nrbufs) {
809                 if (!pipe->writers)
810                         return 0;
811
812                 if (!pipe->waiting_writers && sd->num_spliced)
813                         return 0;
814
815                 if (sd->flags & SPLICE_F_NONBLOCK)
816                         return -EAGAIN;
817
818                 if (signal_pending(current))
819                         return -ERESTARTSYS;
820
821                 if (sd->need_wakeup) {
822                         wakeup_pipe_writers(pipe);
823                         sd->need_wakeup = false;
824                 }
825
826                 pipe_wait(pipe);
827         }
828
829         return 1;
830 }
831 EXPORT_SYMBOL(splice_from_pipe_next);
832
833 /**
834  * splice_from_pipe_begin - start splicing from pipe
835  * @sd:         information about the splice operation
836  *
837  * Description:
838  *    This function should be called before a loop containing
839  *    splice_from_pipe_next() and splice_from_pipe_feed() to
840  *    initialize the necessary fields of @sd.
841  */
842 void splice_from_pipe_begin(struct splice_desc *sd)
843 {
844         sd->num_spliced = 0;
845         sd->need_wakeup = false;
846 }
847 EXPORT_SYMBOL(splice_from_pipe_begin);
848
849 /**
850  * splice_from_pipe_end - finish splicing from pipe
851  * @pipe:       pipe to splice from
852  * @sd:         information about the splice operation
853  *
854  * Description:
855  *    This function will wake up pipe writers if necessary.  It should
856  *    be called after a loop containing splice_from_pipe_next() and
857  *    splice_from_pipe_feed().
858  */
859 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
860 {
861         if (sd->need_wakeup)
862                 wakeup_pipe_writers(pipe);
863 }
864 EXPORT_SYMBOL(splice_from_pipe_end);
865
866 /**
867  * __splice_from_pipe - splice data from a pipe to given actor
868  * @pipe:       pipe to splice from
869  * @sd:         information to @actor
870  * @actor:      handler that splices the data
871  *
872  * Description:
873  *    This function does little more than loop over the pipe and call
874  *    @actor to do the actual moving of a single struct pipe_buffer to
875  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
876  *    pipe_to_user.
877  *
878  */
879 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
880                            splice_actor *actor)
881 {
882         int ret;
883
884         splice_from_pipe_begin(sd);
885         do {
886                 ret = splice_from_pipe_next(pipe, sd);
887                 if (ret > 0)
888                         ret = splice_from_pipe_feed(pipe, sd, actor);
889         } while (ret > 0);
890         splice_from_pipe_end(pipe, sd);
891
892         return sd->num_spliced ? sd->num_spliced : ret;
893 }
894 EXPORT_SYMBOL(__splice_from_pipe);
895
896 /**
897  * splice_from_pipe - splice data from a pipe to a file
898  * @pipe:       pipe to splice from
899  * @out:        file to splice to
900  * @ppos:       position in @out
901  * @len:        how many bytes to splice
902  * @flags:      splice modifier flags
903  * @actor:      handler that splices the data
904  *
905  * Description:
906  *    See __splice_from_pipe. This function locks the pipe inode,
907  *    otherwise it's identical to __splice_from_pipe().
908  *
909  */
910 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
911                          loff_t *ppos, size_t len, unsigned int flags,
912                          splice_actor *actor)
913 {
914         ssize_t ret;
915         struct splice_desc sd = {
916                 .total_len = len,
917                 .flags = flags,
918                 .pos = *ppos,
919                 .u.file = out,
920         };
921
922         pipe_lock(pipe);
923         ret = __splice_from_pipe(pipe, &sd, actor);
924         pipe_unlock(pipe);
925
926         return ret;
927 }
928
929 /**
930  * generic_file_splice_write - splice data from a pipe to a file
931  * @pipe:       pipe info
932  * @out:        file to write to
933  * @ppos:       position in @out
934  * @len:        number of bytes to splice
935  * @flags:      splice modifier flags
936  *
937  * Description:
938  *    Will either move or copy pages (determined by @flags options) from
939  *    the given pipe inode to the given file.
940  *
941  */
942 ssize_t
943 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
944                           loff_t *ppos, size_t len, unsigned int flags)
945 {
946         struct address_space *mapping = out->f_mapping;
947         struct inode *inode = mapping->host;
948         struct splice_desc sd = {
949                 .total_len = len,
950                 .flags = flags,
951                 .pos = *ppos,
952                 .u.file = out,
953         };
954         ssize_t ret;
955
956         pipe_lock(pipe);
957
958         splice_from_pipe_begin(&sd);
959         do {
960                 ret = splice_from_pipe_next(pipe, &sd);
961                 if (ret <= 0)
962                         break;
963
964                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
965                 ret = file_remove_suid(out);
966                 if (!ret)
967                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
968                 mutex_unlock(&inode->i_mutex);
969         } while (ret > 0);
970         splice_from_pipe_end(pipe, &sd);
971
972         pipe_unlock(pipe);
973
974         if (sd.num_spliced)
975                 ret = sd.num_spliced;
976
977         if (ret > 0) {
978                 unsigned long nr_pages;
979
980                 *ppos += ret;
981                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
982
983                 /*
984                  * If file or inode is SYNC and we actually wrote some data,
985                  * sync it.
986                  */
987                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
988                         int err;
989
990                         mutex_lock(&inode->i_mutex);
991                         err = generic_osync_inode(inode, mapping,
992                                                   OSYNC_METADATA|OSYNC_DATA);
993                         mutex_unlock(&inode->i_mutex);
994
995                         if (err)
996                                 ret = err;
997                 }
998                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
999         }
1000
1001         return ret;
1002 }
1003
1004 EXPORT_SYMBOL(generic_file_splice_write);
1005
1006 static struct pipe_buffer *nth_pipe_buf(struct pipe_inode_info *pipe, int n)
1007 {
1008         return &pipe->bufs[(pipe->curbuf + n) % PIPE_BUFFERS];
1009 }
1010
1011 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1012                                          struct file *out, loff_t *ppos,
1013                                          size_t len, unsigned int flags)
1014 {
1015         ssize_t ret = 0;
1016         ssize_t total_len = 0;
1017         int do_wakeup = 0;
1018
1019         pipe_lock(pipe);
1020         while (len) {
1021                 struct pipe_buffer *buf;
1022                 void *data[PIPE_BUFFERS];
1023                 struct iovec vec[PIPE_BUFFERS];
1024                 unsigned int nr_pages = 0;
1025                 unsigned int write_len = 0;
1026                 unsigned int now_len = len;
1027                 unsigned int this_len;
1028                 int i;
1029
1030                 BUG_ON(pipe->nrbufs > PIPE_BUFFERS);
1031                 for (i = 0; i < pipe->nrbufs && now_len; i++) {
1032                         buf = nth_pipe_buf(pipe, i);
1033
1034                         ret = buf->ops->confirm(pipe, buf);
1035                         if (ret)
1036                                 break;
1037
1038                         data[i] = buf->ops->map(pipe, buf, 0);
1039                         this_len = min(buf->len, now_len);
1040                         vec[i].iov_base = (void __user *) data[i] + buf->offset;
1041                         vec[i].iov_len = this_len;
1042                         now_len -= this_len;
1043                         write_len += this_len;
1044                         nr_pages++;
1045                 }
1046
1047                 if (nr_pages) {
1048                         ret = kernel_writev(out, vec, nr_pages, ppos);
1049                         if (ret == 0)
1050                                 ret = -EIO;
1051                         if (ret > 0) {
1052                                 len -= ret;
1053                                 total_len += ret;
1054                         }
1055                 }
1056
1057                 for (i = 0; i < nr_pages; i++) {
1058                         buf = nth_pipe_buf(pipe, i);
1059                         buf->ops->unmap(pipe, buf, data[i]);
1060
1061                         if (ret > 0) {
1062                                 this_len = min_t(unsigned, vec[i].iov_len, ret);
1063                                 buf->offset += this_len;
1064                                 buf->len -= this_len;
1065                                 ret -= this_len;
1066                         }
1067                 }
1068
1069                 if (ret < 0)
1070                         break;
1071
1072                 while (pipe->nrbufs) {
1073                         const struct pipe_buf_operations *ops;
1074
1075                         buf = nth_pipe_buf(pipe, 0);
1076                         if (buf->len)
1077                                 break;
1078
1079                         ops = buf->ops;
1080                         buf->ops = NULL;
1081                         ops->release(pipe, buf);
1082                         pipe->curbuf = (pipe->curbuf + 1) % PIPE_BUFFERS;
1083                         pipe->nrbufs--;
1084                         if (pipe->inode)
1085                                 do_wakeup = 1;
1086                 }
1087
1088                 if (pipe->nrbufs)
1089                         continue;
1090                 if (!pipe->writers)
1091                         break;
1092                 if (!pipe->waiting_writers) {
1093                         if (total_len)
1094                                 break;
1095                 }
1096
1097                 if (flags & SPLICE_F_NONBLOCK) {
1098                         ret = -EAGAIN;
1099                         break;
1100                 }
1101
1102                 if (signal_pending(current)) {
1103                         ret = -ERESTARTSYS;
1104                         break;
1105                 }
1106
1107                 if (do_wakeup) {
1108                         wakeup_pipe_writers(pipe);
1109                         do_wakeup = 0;
1110                 }
1111
1112                 pipe_wait(pipe);
1113         }
1114         pipe_unlock(pipe);
1115
1116         if (do_wakeup)
1117                 wakeup_pipe_writers(pipe);
1118
1119         return total_len ? total_len : ret;
1120 }
1121
1122 /**
1123  * generic_splice_sendpage - splice data from a pipe to a socket
1124  * @pipe:       pipe to splice from
1125  * @out:        socket to write to
1126  * @ppos:       position in @out
1127  * @len:        number of bytes to splice
1128  * @flags:      splice modifier flags
1129  *
1130  * Description:
1131  *    Will send @len bytes from the pipe to a network socket. No data copying
1132  *    is involved.
1133  *
1134  */
1135 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1136                                 loff_t *ppos, size_t len, unsigned int flags)
1137 {
1138         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1139 }
1140
1141 EXPORT_SYMBOL(generic_splice_sendpage);
1142
1143 /*
1144  * Attempt to initiate a splice from pipe to file.
1145  */
1146 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1147                            loff_t *ppos, size_t len, unsigned int flags)
1148 {
1149         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1150                                 loff_t *, size_t, unsigned int);
1151         int ret;
1152
1153         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1154                 return -EBADF;
1155
1156         if (unlikely(out->f_flags & O_APPEND))
1157                 return -EINVAL;
1158
1159         ret = rw_verify_area(WRITE, out, ppos, len);
1160         if (unlikely(ret < 0))
1161                 return ret;
1162
1163         splice_write = out->f_op->splice_write;
1164         if (!splice_write)
1165                 splice_write = default_file_splice_write;
1166
1167         return splice_write(pipe, out, ppos, len, flags);
1168 }
1169
1170 /*
1171  * Attempt to initiate a splice from a file to a pipe.
1172  */
1173 static long do_splice_to(struct file *in, loff_t *ppos,
1174                          struct pipe_inode_info *pipe, size_t len,
1175                          unsigned int flags)
1176 {
1177         ssize_t (*splice_read)(struct file *, loff_t *,
1178                                struct pipe_inode_info *, size_t, unsigned int);
1179         int ret;
1180
1181         if (unlikely(!(in->f_mode & FMODE_READ)))
1182                 return -EBADF;
1183
1184         ret = rw_verify_area(READ, in, ppos, len);
1185         if (unlikely(ret < 0))
1186                 return ret;
1187
1188         splice_read = in->f_op->splice_read;
1189         if (!splice_read)
1190                 splice_read = default_file_splice_read;
1191
1192         return splice_read(in, ppos, pipe, len, flags);
1193 }
1194
1195 /**
1196  * splice_direct_to_actor - splices data directly between two non-pipes
1197  * @in:         file to splice from
1198  * @sd:         actor information on where to splice to
1199  * @actor:      handles the data splicing
1200  *
1201  * Description:
1202  *    This is a special case helper to splice directly between two
1203  *    points, without requiring an explicit pipe. Internally an allocated
1204  *    pipe is cached in the process, and reused during the lifetime of
1205  *    that process.
1206  *
1207  */
1208 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1209                                splice_direct_actor *actor)
1210 {
1211         struct pipe_inode_info *pipe;
1212         long ret, bytes;
1213         umode_t i_mode;
1214         size_t len;
1215         int i, flags;
1216
1217         /*
1218          * We require the input being a regular file, as we don't want to
1219          * randomly drop data for eg socket -> socket splicing. Use the
1220          * piped splicing for that!
1221          */
1222         i_mode = in->f_path.dentry->d_inode->i_mode;
1223         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1224                 return -EINVAL;
1225
1226         /*
1227          * neither in nor out is a pipe, setup an internal pipe attached to
1228          * 'out' and transfer the wanted data from 'in' to 'out' through that
1229          */
1230         pipe = current->splice_pipe;
1231         if (unlikely(!pipe)) {
1232                 pipe = alloc_pipe_info(NULL);
1233                 if (!pipe)
1234                         return -ENOMEM;
1235
1236                 /*
1237                  * We don't have an immediate reader, but we'll read the stuff
1238                  * out of the pipe right after the splice_to_pipe(). So set
1239                  * PIPE_READERS appropriately.
1240                  */
1241                 pipe->readers = 1;
1242
1243                 current->splice_pipe = pipe;
1244         }
1245
1246         /*
1247          * Do the splice.
1248          */
1249         ret = 0;
1250         bytes = 0;
1251         len = sd->total_len;
1252         flags = sd->flags;
1253
1254         /*
1255          * Don't block on output, we have to drain the direct pipe.
1256          */
1257         sd->flags &= ~SPLICE_F_NONBLOCK;
1258
1259         while (len) {
1260                 size_t read_len;
1261                 loff_t pos = sd->pos, prev_pos = pos;
1262
1263                 ret = do_splice_to(in, &pos, pipe, len, flags);
1264                 if (unlikely(ret <= 0))
1265                         goto out_release;
1266
1267                 read_len = ret;
1268                 sd->total_len = read_len;
1269
1270                 /*
1271                  * NOTE: nonblocking mode only applies to the input. We
1272                  * must not do the output in nonblocking mode as then we
1273                  * could get stuck data in the internal pipe:
1274                  */
1275                 ret = actor(pipe, sd);
1276                 if (unlikely(ret <= 0)) {
1277                         sd->pos = prev_pos;
1278                         goto out_release;
1279                 }
1280
1281                 bytes += ret;
1282                 len -= ret;
1283                 sd->pos = pos;
1284
1285                 if (ret < read_len) {
1286                         sd->pos = prev_pos + ret;
1287                         goto out_release;
1288                 }
1289         }
1290
1291 done:
1292         pipe->nrbufs = pipe->curbuf = 0;
1293         file_accessed(in);
1294         return bytes;
1295
1296 out_release:
1297         /*
1298          * If we did an incomplete transfer we must release
1299          * the pipe buffers in question:
1300          */
1301         for (i = 0; i < PIPE_BUFFERS; i++) {
1302                 struct pipe_buffer *buf = pipe->bufs + i;
1303
1304                 if (buf->ops) {
1305                         buf->ops->release(pipe, buf);
1306                         buf->ops = NULL;
1307                 }
1308         }
1309
1310         if (!bytes)
1311                 bytes = ret;
1312
1313         goto done;
1314 }
1315 EXPORT_SYMBOL(splice_direct_to_actor);
1316
1317 static int direct_splice_actor(struct pipe_inode_info *pipe,
1318                                struct splice_desc *sd)
1319 {
1320         struct file *file = sd->u.file;
1321
1322         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1323 }
1324
1325 /**
1326  * do_splice_direct - splices data directly between two files
1327  * @in:         file to splice from
1328  * @ppos:       input file offset
1329  * @out:        file to splice to
1330  * @len:        number of bytes to splice
1331  * @flags:      splice modifier flags
1332  *
1333  * Description:
1334  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1335  *    doing it in the application would incur an extra system call
1336  *    (splice in + splice out, as compared to just sendfile()). So this helper
1337  *    can splice directly through a process-private pipe.
1338  *
1339  */
1340 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1341                       size_t len, unsigned int flags)
1342 {
1343         struct splice_desc sd = {
1344                 .len            = len,
1345                 .total_len      = len,
1346                 .flags          = flags,
1347                 .pos            = *ppos,
1348                 .u.file         = out,
1349         };
1350         long ret;
1351
1352         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1353         if (ret > 0)
1354                 *ppos = sd.pos;
1355
1356         return ret;
1357 }
1358
1359 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1360                                struct pipe_inode_info *opipe,
1361                                size_t len, unsigned int flags);
1362 /*
1363  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1364  * location, so checking ->i_pipe is not enough to verify that this is a
1365  * pipe.
1366  */
1367 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1368 {
1369         if (S_ISFIFO(inode->i_mode))
1370                 return inode->i_pipe;
1371
1372         return NULL;
1373 }
1374
1375 /*
1376  * Determine where to splice to/from.
1377  */
1378 static long do_splice(struct file *in, loff_t __user *off_in,
1379                       struct file *out, loff_t __user *off_out,
1380                       size_t len, unsigned int flags)
1381 {
1382         struct pipe_inode_info *ipipe;
1383         struct pipe_inode_info *opipe;
1384         loff_t offset, *off;
1385         long ret;
1386
1387         ipipe = pipe_info(in->f_path.dentry->d_inode);
1388         opipe = pipe_info(out->f_path.dentry->d_inode);
1389
1390         if (ipipe && opipe) {
1391                 if (off_in || off_out)
1392                         return -ESPIPE;
1393
1394                 if (!(in->f_mode & FMODE_READ))
1395                         return -EBADF;
1396
1397                 if (!(out->f_mode & FMODE_WRITE))
1398                         return -EBADF;
1399
1400                 /* Splicing to self would be fun, but... */
1401                 if (ipipe == opipe)
1402                         return -EINVAL;
1403
1404                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1405         }
1406
1407         if (ipipe) {
1408                 if (off_in)
1409                         return -ESPIPE;
1410                 if (off_out) {
1411                         if (out->f_op->llseek == no_llseek)
1412                                 return -EINVAL;
1413                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1414                                 return -EFAULT;
1415                         off = &offset;
1416                 } else
1417                         off = &out->f_pos;
1418
1419                 ret = do_splice_from(ipipe, out, off, len, flags);
1420
1421                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1422                         ret = -EFAULT;
1423
1424                 return ret;
1425         }
1426
1427         if (opipe) {
1428                 if (off_out)
1429                         return -ESPIPE;
1430                 if (off_in) {
1431                         if (in->f_op->llseek == no_llseek)
1432                                 return -EINVAL;
1433                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1434                                 return -EFAULT;
1435                         off = &offset;
1436                 } else
1437                         off = &in->f_pos;
1438
1439                 ret = do_splice_to(in, off, opipe, len, flags);
1440
1441                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1442                         ret = -EFAULT;
1443
1444                 return ret;
1445         }
1446
1447         return -EINVAL;
1448 }
1449
1450 /*
1451  * Map an iov into an array of pages and offset/length tupples. With the
1452  * partial_page structure, we can map several non-contiguous ranges into
1453  * our ones pages[] map instead of splitting that operation into pieces.
1454  * Could easily be exported as a generic helper for other users, in which
1455  * case one would probably want to add a 'max_nr_pages' parameter as well.
1456  */
1457 static int get_iovec_page_array(const struct iovec __user *iov,
1458                                 unsigned int nr_vecs, struct page **pages,
1459                                 struct partial_page *partial, int aligned)
1460 {
1461         int buffers = 0, error = 0;
1462
1463         while (nr_vecs) {
1464                 unsigned long off, npages;
1465                 struct iovec entry;
1466                 void __user *base;
1467                 size_t len;
1468                 int i;
1469
1470                 error = -EFAULT;
1471                 if (copy_from_user(&entry, iov, sizeof(entry)))
1472                         break;
1473
1474                 base = entry.iov_base;
1475                 len = entry.iov_len;
1476
1477                 /*
1478                  * Sanity check this iovec. 0 read succeeds.
1479                  */
1480                 error = 0;
1481                 if (unlikely(!len))
1482                         break;
1483                 error = -EFAULT;
1484                 if (!access_ok(VERIFY_READ, base, len))
1485                         break;
1486
1487                 /*
1488                  * Get this base offset and number of pages, then map
1489                  * in the user pages.
1490                  */
1491                 off = (unsigned long) base & ~PAGE_MASK;
1492
1493                 /*
1494                  * If asked for alignment, the offset must be zero and the
1495                  * length a multiple of the PAGE_SIZE.
1496                  */
1497                 error = -EINVAL;
1498                 if (aligned && (off || len & ~PAGE_MASK))
1499                         break;
1500
1501                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1502                 if (npages > PIPE_BUFFERS - buffers)
1503                         npages = PIPE_BUFFERS - buffers;
1504
1505                 error = get_user_pages_fast((unsigned long)base, npages,
1506                                         0, &pages[buffers]);
1507
1508                 if (unlikely(error <= 0))
1509                         break;
1510
1511                 /*
1512                  * Fill this contiguous range into the partial page map.
1513                  */
1514                 for (i = 0; i < error; i++) {
1515                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1516
1517                         partial[buffers].offset = off;
1518                         partial[buffers].len = plen;
1519
1520                         off = 0;
1521                         len -= plen;
1522                         buffers++;
1523                 }
1524
1525                 /*
1526                  * We didn't complete this iov, stop here since it probably
1527                  * means we have to move some of this into a pipe to
1528                  * be able to continue.
1529                  */
1530                 if (len)
1531                         break;
1532
1533                 /*
1534                  * Don't continue if we mapped fewer pages than we asked for,
1535                  * or if we mapped the max number of pages that we have
1536                  * room for.
1537                  */
1538                 if (error < npages || buffers == PIPE_BUFFERS)
1539                         break;
1540
1541                 nr_vecs--;
1542                 iov++;
1543         }
1544
1545         if (buffers)
1546                 return buffers;
1547
1548         return error;
1549 }
1550
1551 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1552                         struct splice_desc *sd)
1553 {
1554         char *src;
1555         int ret;
1556
1557         ret = buf->ops->confirm(pipe, buf);
1558         if (unlikely(ret))
1559                 return ret;
1560
1561         /*
1562          * See if we can use the atomic maps, by prefaulting in the
1563          * pages and doing an atomic copy
1564          */
1565         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1566                 src = buf->ops->map(pipe, buf, 1);
1567                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1568                                                         sd->len);
1569                 buf->ops->unmap(pipe, buf, src);
1570                 if (!ret) {
1571                         ret = sd->len;
1572                         goto out;
1573                 }
1574         }
1575
1576         /*
1577          * No dice, use slow non-atomic map and copy
1578          */
1579         src = buf->ops->map(pipe, buf, 0);
1580
1581         ret = sd->len;
1582         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1583                 ret = -EFAULT;
1584
1585         buf->ops->unmap(pipe, buf, src);
1586 out:
1587         if (ret > 0)
1588                 sd->u.userptr += ret;
1589         return ret;
1590 }
1591
1592 /*
1593  * For lack of a better implementation, implement vmsplice() to userspace
1594  * as a simple copy of the pipes pages to the user iov.
1595  */
1596 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1597                              unsigned long nr_segs, unsigned int flags)
1598 {
1599         struct pipe_inode_info *pipe;
1600         struct splice_desc sd;
1601         ssize_t size;
1602         int error;
1603         long ret;
1604
1605         pipe = pipe_info(file->f_path.dentry->d_inode);
1606         if (!pipe)
1607                 return -EBADF;
1608
1609         pipe_lock(pipe);
1610
1611         error = ret = 0;
1612         while (nr_segs) {
1613                 void __user *base;
1614                 size_t len;
1615
1616                 /*
1617                  * Get user address base and length for this iovec.
1618                  */
1619                 error = get_user(base, &iov->iov_base);
1620                 if (unlikely(error))
1621                         break;
1622                 error = get_user(len, &iov->iov_len);
1623                 if (unlikely(error))
1624                         break;
1625
1626                 /*
1627                  * Sanity check this iovec. 0 read succeeds.
1628                  */
1629                 if (unlikely(!len))
1630                         break;
1631                 if (unlikely(!base)) {
1632                         error = -EFAULT;
1633                         break;
1634                 }
1635
1636                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1637                         error = -EFAULT;
1638                         break;
1639                 }
1640
1641                 sd.len = 0;
1642                 sd.total_len = len;
1643                 sd.flags = flags;
1644                 sd.u.userptr = base;
1645                 sd.pos = 0;
1646
1647                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1648                 if (size < 0) {
1649                         if (!ret)
1650                                 ret = size;
1651
1652                         break;
1653                 }
1654
1655                 ret += size;
1656
1657                 if (size < len)
1658                         break;
1659
1660                 nr_segs--;
1661                 iov++;
1662         }
1663
1664         pipe_unlock(pipe);
1665
1666         if (!ret)
1667                 ret = error;
1668
1669         return ret;
1670 }
1671
1672 /*
1673  * vmsplice splices a user address range into a pipe. It can be thought of
1674  * as splice-from-memory, where the regular splice is splice-from-file (or
1675  * to file). In both cases the output is a pipe, naturally.
1676  */
1677 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1678                              unsigned long nr_segs, unsigned int flags)
1679 {
1680         struct pipe_inode_info *pipe;
1681         struct page *pages[PIPE_BUFFERS];
1682         struct partial_page partial[PIPE_BUFFERS];
1683         struct splice_pipe_desc spd = {
1684                 .pages = pages,
1685                 .partial = partial,
1686                 .flags = flags,
1687                 .ops = &user_page_pipe_buf_ops,
1688                 .spd_release = spd_release_page,
1689         };
1690
1691         pipe = pipe_info(file->f_path.dentry->d_inode);
1692         if (!pipe)
1693                 return -EBADF;
1694
1695         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1696                                             flags & SPLICE_F_GIFT);
1697         if (spd.nr_pages <= 0)
1698                 return spd.nr_pages;
1699
1700         return splice_to_pipe(pipe, &spd);
1701 }
1702
1703 /*
1704  * Note that vmsplice only really supports true splicing _from_ user memory
1705  * to a pipe, not the other way around. Splicing from user memory is a simple
1706  * operation that can be supported without any funky alignment restrictions
1707  * or nasty vm tricks. We simply map in the user memory and fill them into
1708  * a pipe. The reverse isn't quite as easy, though. There are two possible
1709  * solutions for that:
1710  *
1711  *      - memcpy() the data internally, at which point we might as well just
1712  *        do a regular read() on the buffer anyway.
1713  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1714  *        has restriction limitations on both ends of the pipe).
1715  *
1716  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1717  *
1718  */
1719 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1720                 unsigned long, nr_segs, unsigned int, flags)
1721 {
1722         struct file *file;
1723         long error;
1724         int fput;
1725
1726         if (unlikely(nr_segs > UIO_MAXIOV))
1727                 return -EINVAL;
1728         else if (unlikely(!nr_segs))
1729                 return 0;
1730
1731         error = -EBADF;
1732         file = fget_light(fd, &fput);
1733         if (file) {
1734                 if (file->f_mode & FMODE_WRITE)
1735                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1736                 else if (file->f_mode & FMODE_READ)
1737                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1738
1739                 fput_light(file, fput);
1740         }
1741
1742         return error;
1743 }
1744
1745 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1746                 int, fd_out, loff_t __user *, off_out,
1747                 size_t, len, unsigned int, flags)
1748 {
1749         long error;
1750         struct file *in, *out;
1751         int fput_in, fput_out;
1752
1753         if (unlikely(!len))
1754                 return 0;
1755
1756         error = -EBADF;
1757         in = fget_light(fd_in, &fput_in);
1758         if (in) {
1759                 if (in->f_mode & FMODE_READ) {
1760                         out = fget_light(fd_out, &fput_out);
1761                         if (out) {
1762                                 if (out->f_mode & FMODE_WRITE)
1763                                         error = do_splice(in, off_in,
1764                                                           out, off_out,
1765                                                           len, flags);
1766                                 fput_light(out, fput_out);
1767                         }
1768                 }
1769
1770                 fput_light(in, fput_in);
1771         }
1772
1773         return error;
1774 }
1775
1776 /*
1777  * Make sure there's data to read. Wait for input if we can, otherwise
1778  * return an appropriate error.
1779  */
1780 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1781 {
1782         int ret;
1783
1784         /*
1785          * Check ->nrbufs without the inode lock first. This function
1786          * is speculative anyways, so missing one is ok.
1787          */
1788         if (pipe->nrbufs)
1789                 return 0;
1790
1791         ret = 0;
1792         pipe_lock(pipe);
1793
1794         while (!pipe->nrbufs) {
1795                 if (signal_pending(current)) {
1796                         ret = -ERESTARTSYS;
1797                         break;
1798                 }
1799                 if (!pipe->writers)
1800                         break;
1801                 if (!pipe->waiting_writers) {
1802                         if (flags & SPLICE_F_NONBLOCK) {
1803                                 ret = -EAGAIN;
1804                                 break;
1805                         }
1806                 }
1807                 pipe_wait(pipe);
1808         }
1809
1810         pipe_unlock(pipe);
1811         return ret;
1812 }
1813
1814 /*
1815  * Make sure there's writeable room. Wait for room if we can, otherwise
1816  * return an appropriate error.
1817  */
1818 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1819 {
1820         int ret;
1821
1822         /*
1823          * Check ->nrbufs without the inode lock first. This function
1824          * is speculative anyways, so missing one is ok.
1825          */
1826         if (pipe->nrbufs < PIPE_BUFFERS)
1827                 return 0;
1828
1829         ret = 0;
1830         pipe_lock(pipe);
1831
1832         while (pipe->nrbufs >= PIPE_BUFFERS) {
1833                 if (!pipe->readers) {
1834                         send_sig(SIGPIPE, current, 0);
1835                         ret = -EPIPE;
1836                         break;
1837                 }
1838                 if (flags & SPLICE_F_NONBLOCK) {
1839                         ret = -EAGAIN;
1840                         break;
1841                 }
1842                 if (signal_pending(current)) {
1843                         ret = -ERESTARTSYS;
1844                         break;
1845                 }
1846                 pipe->waiting_writers++;
1847                 pipe_wait(pipe);
1848                 pipe->waiting_writers--;
1849         }
1850
1851         pipe_unlock(pipe);
1852         return ret;
1853 }
1854
1855 /*
1856  * Splice contents of ipipe to opipe.
1857  */
1858 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1859                                struct pipe_inode_info *opipe,
1860                                size_t len, unsigned int flags)
1861 {
1862         struct pipe_buffer *ibuf, *obuf;
1863         int ret = 0, nbuf;
1864         bool input_wakeup = false;
1865
1866
1867 retry:
1868         ret = ipipe_prep(ipipe, flags);
1869         if (ret)
1870                 return ret;
1871
1872         ret = opipe_prep(opipe, flags);
1873         if (ret)
1874                 return ret;
1875
1876         /*
1877          * Potential ABBA deadlock, work around it by ordering lock
1878          * grabbing by pipe info address. Otherwise two different processes
1879          * could deadlock (one doing tee from A -> B, the other from B -> A).
1880          */
1881         pipe_double_lock(ipipe, opipe);
1882
1883         do {
1884                 if (!opipe->readers) {
1885                         send_sig(SIGPIPE, current, 0);
1886                         if (!ret)
1887                                 ret = -EPIPE;
1888                         break;
1889                 }
1890
1891                 if (!ipipe->nrbufs && !ipipe->writers)
1892                         break;
1893
1894                 /*
1895                  * Cannot make any progress, because either the input
1896                  * pipe is empty or the output pipe is full.
1897                  */
1898                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1899                         /* Already processed some buffers, break */
1900                         if (ret)
1901                                 break;
1902
1903                         if (flags & SPLICE_F_NONBLOCK) {
1904                                 ret = -EAGAIN;
1905                                 break;
1906                         }
1907
1908                         /*
1909                          * We raced with another reader/writer and haven't
1910                          * managed to process any buffers.  A zero return
1911                          * value means EOF, so retry instead.
1912                          */
1913                         pipe_unlock(ipipe);
1914                         pipe_unlock(opipe);
1915                         goto retry;
1916                 }
1917
1918                 ibuf = ipipe->bufs + ipipe->curbuf;
1919                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1920                 obuf = opipe->bufs + nbuf;
1921
1922                 if (len >= ibuf->len) {
1923                         /*
1924                          * Simply move the whole buffer from ipipe to opipe
1925                          */
1926                         *obuf = *ibuf;
1927                         ibuf->ops = NULL;
1928                         opipe->nrbufs++;
1929                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1930                         ipipe->nrbufs--;
1931                         input_wakeup = true;
1932                 } else {
1933                         /*
1934                          * Get a reference to this pipe buffer,
1935                          * so we can copy the contents over.
1936                          */
1937                         ibuf->ops->get(ipipe, ibuf);
1938                         *obuf = *ibuf;
1939
1940                         /*
1941                          * Don't inherit the gift flag, we need to
1942                          * prevent multiple steals of this page.
1943                          */
1944                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1945
1946                         obuf->len = len;
1947                         opipe->nrbufs++;
1948                         ibuf->offset += obuf->len;
1949                         ibuf->len -= obuf->len;
1950                 }
1951                 ret += obuf->len;
1952                 len -= obuf->len;
1953         } while (len);
1954
1955         pipe_unlock(ipipe);
1956         pipe_unlock(opipe);
1957
1958         /*
1959          * If we put data in the output pipe, wakeup any potential readers.
1960          */
1961         if (ret > 0) {
1962                 smp_mb();
1963                 if (waitqueue_active(&opipe->wait))
1964                         wake_up_interruptible(&opipe->wait);
1965                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1966         }
1967         if (input_wakeup)
1968                 wakeup_pipe_writers(ipipe);
1969
1970         return ret;
1971 }
1972
1973 /*
1974  * Link contents of ipipe to opipe.
1975  */
1976 static int link_pipe(struct pipe_inode_info *ipipe,
1977                      struct pipe_inode_info *opipe,
1978                      size_t len, unsigned int flags)
1979 {
1980         struct pipe_buffer *ibuf, *obuf;
1981         int ret = 0, i = 0, nbuf;
1982
1983         /*
1984          * Potential ABBA deadlock, work around it by ordering lock
1985          * grabbing by pipe info address. Otherwise two different processes
1986          * could deadlock (one doing tee from A -> B, the other from B -> A).
1987          */
1988         pipe_double_lock(ipipe, opipe);
1989
1990         do {
1991                 if (!opipe->readers) {
1992                         send_sig(SIGPIPE, current, 0);
1993                         if (!ret)
1994                                 ret = -EPIPE;
1995                         break;
1996                 }
1997
1998                 /*
1999                  * If we have iterated all input buffers or ran out of
2000                  * output room, break.
2001                  */
2002                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
2003                         break;
2004
2005                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
2006                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
2007
2008                 /*
2009                  * Get a reference to this pipe buffer,
2010                  * so we can copy the contents over.
2011                  */
2012                 ibuf->ops->get(ipipe, ibuf);
2013
2014                 obuf = opipe->bufs + nbuf;
2015                 *obuf = *ibuf;
2016
2017                 /*
2018                  * Don't inherit the gift flag, we need to
2019                  * prevent multiple steals of this page.
2020                  */
2021                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2022
2023                 if (obuf->len > len)
2024                         obuf->len = len;
2025
2026                 opipe->nrbufs++;
2027                 ret += obuf->len;
2028                 len -= obuf->len;
2029                 i++;
2030         } while (len);
2031
2032         /*
2033          * return EAGAIN if we have the potential of some data in the
2034          * future, otherwise just return 0
2035          */
2036         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2037                 ret = -EAGAIN;
2038
2039         pipe_unlock(ipipe);
2040         pipe_unlock(opipe);
2041
2042         /*
2043          * If we put data in the output pipe, wakeup any potential readers.
2044          */
2045         if (ret > 0) {
2046                 smp_mb();
2047                 if (waitqueue_active(&opipe->wait))
2048                         wake_up_interruptible(&opipe->wait);
2049                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2050         }
2051
2052         return ret;
2053 }
2054
2055 /*
2056  * This is a tee(1) implementation that works on pipes. It doesn't copy
2057  * any data, it simply references the 'in' pages on the 'out' pipe.
2058  * The 'flags' used are the SPLICE_F_* variants, currently the only
2059  * applicable one is SPLICE_F_NONBLOCK.
2060  */
2061 static long do_tee(struct file *in, struct file *out, size_t len,
2062                    unsigned int flags)
2063 {
2064         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2065         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2066         int ret = -EINVAL;
2067
2068         /*
2069          * Duplicate the contents of ipipe to opipe without actually
2070          * copying the data.
2071          */
2072         if (ipipe && opipe && ipipe != opipe) {
2073                 /*
2074                  * Keep going, unless we encounter an error. The ipipe/opipe
2075                  * ordering doesn't really matter.
2076                  */
2077                 ret = ipipe_prep(ipipe, flags);
2078                 if (!ret) {
2079                         ret = opipe_prep(opipe, flags);
2080                         if (!ret)
2081                                 ret = link_pipe(ipipe, opipe, len, flags);
2082                 }
2083         }
2084
2085         return ret;
2086 }
2087
2088 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2089 {
2090         struct file *in;
2091         int error, fput_in;
2092
2093         if (unlikely(!len))
2094                 return 0;
2095
2096         error = -EBADF;
2097         in = fget_light(fdin, &fput_in);
2098         if (in) {
2099                 if (in->f_mode & FMODE_READ) {
2100                         int fput_out;
2101                         struct file *out = fget_light(fdout, &fput_out);
2102
2103                         if (out) {
2104                                 if (out->f_mode & FMODE_WRITE)
2105                                         error = do_tee(in, out, len, flags);
2106                                 fput_light(out, fput_out);
2107                         }
2108                 }
2109                 fput_light(in, fput_in);
2110         }
2111
2112         return error;
2113 }