Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/pablo/nf
[linux-2.6-microblaze.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/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 <linux/compat.h>
35 #include "internal.h"
36
37 /*
38  * Attempt to steal a page from a pipe buffer. This should perhaps go into
39  * a vm helper function, it's already simplified quite a bit by the
40  * addition of remove_mapping(). If success is returned, the caller may
41  * attempt to reuse this page for another destination.
42  */
43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
44                                      struct pipe_buffer *buf)
45 {
46         struct page *page = buf->page;
47         struct address_space *mapping;
48
49         lock_page(page);
50
51         mapping = page_mapping(page);
52         if (mapping) {
53                 WARN_ON(!PageUptodate(page));
54
55                 /*
56                  * At least for ext2 with nobh option, we need to wait on
57                  * writeback completing on this page, since we'll remove it
58                  * from the pagecache.  Otherwise truncate wont wait on the
59                  * page, allowing the disk blocks to be reused by someone else
60                  * before we actually wrote our data to them. fs corruption
61                  * ensues.
62                  */
63                 wait_on_page_writeback(page);
64
65                 if (page_has_private(page) &&
66                     !try_to_release_page(page, GFP_KERNEL))
67                         goto out_unlock;
68
69                 /*
70                  * If we succeeded in removing the mapping, set LRU flag
71                  * and return good.
72                  */
73                 if (remove_mapping(mapping, page)) {
74                         buf->flags |= PIPE_BUF_FLAG_LRU;
75                         return 0;
76                 }
77         }
78
79         /*
80          * Raced with truncate or failed to remove page from current
81          * address space, unlock and return failure.
82          */
83 out_unlock:
84         unlock_page(page);
85         return 1;
86 }
87
88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
89                                         struct pipe_buffer *buf)
90 {
91         page_cache_release(buf->page);
92         buf->flags &= ~PIPE_BUF_FLAG_LRU;
93 }
94
95 /*
96  * Check whether the contents of buf is OK to access. Since the content
97  * is a page cache page, IO may be in flight.
98  */
99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
100                                        struct pipe_buffer *buf)
101 {
102         struct page *page = buf->page;
103         int err;
104
105         if (!PageUptodate(page)) {
106                 lock_page(page);
107
108                 /*
109                  * Page got truncated/unhashed. This will cause a 0-byte
110                  * splice, if this is the first page.
111                  */
112                 if (!page->mapping) {
113                         err = -ENODATA;
114                         goto error;
115                 }
116
117                 /*
118                  * Uh oh, read-error from disk.
119                  */
120                 if (!PageUptodate(page)) {
121                         err = -EIO;
122                         goto error;
123                 }
124
125                 /*
126                  * Page is ok afterall, we are done.
127                  */
128                 unlock_page(page);
129         }
130
131         return 0;
132 error:
133         unlock_page(page);
134         return err;
135 }
136
137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
138         .can_merge = 0,
139         .map = generic_pipe_buf_map,
140         .unmap = generic_pipe_buf_unmap,
141         .confirm = page_cache_pipe_buf_confirm,
142         .release = page_cache_pipe_buf_release,
143         .steal = page_cache_pipe_buf_steal,
144         .get = generic_pipe_buf_get,
145 };
146
147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
148                                     struct pipe_buffer *buf)
149 {
150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
151                 return 1;
152
153         buf->flags |= PIPE_BUF_FLAG_LRU;
154         return generic_pipe_buf_steal(pipe, buf);
155 }
156
157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158         .can_merge = 0,
159         .map = generic_pipe_buf_map,
160         .unmap = generic_pipe_buf_unmap,
161         .confirm = generic_pipe_buf_confirm,
162         .release = page_cache_pipe_buf_release,
163         .steal = user_page_pipe_buf_steal,
164         .get = generic_pipe_buf_get,
165 };
166
167 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
168 {
169         smp_mb();
170         if (waitqueue_active(&pipe->wait))
171                 wake_up_interruptible(&pipe->wait);
172         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
173 }
174
175 /**
176  * splice_to_pipe - fill passed data into a pipe
177  * @pipe:       pipe to fill
178  * @spd:        data to fill
179  *
180  * Description:
181  *    @spd contains a map of pages and len/offset tuples, along with
182  *    the struct pipe_buf_operations associated with these pages. This
183  *    function will link that data to the pipe.
184  *
185  */
186 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
187                        struct splice_pipe_desc *spd)
188 {
189         unsigned int spd_pages = spd->nr_pages;
190         int ret, do_wakeup, page_nr;
191
192         ret = 0;
193         do_wakeup = 0;
194         page_nr = 0;
195
196         pipe_lock(pipe);
197
198         for (;;) {
199                 if (!pipe->readers) {
200                         send_sig(SIGPIPE, current, 0);
201                         if (!ret)
202                                 ret = -EPIPE;
203                         break;
204                 }
205
206                 if (pipe->nrbufs < pipe->buffers) {
207                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
208                         struct pipe_buffer *buf = pipe->bufs + newbuf;
209
210                         buf->page = spd->pages[page_nr];
211                         buf->offset = spd->partial[page_nr].offset;
212                         buf->len = spd->partial[page_nr].len;
213                         buf->private = spd->partial[page_nr].private;
214                         buf->ops = spd->ops;
215                         if (spd->flags & SPLICE_F_GIFT)
216                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
217
218                         pipe->nrbufs++;
219                         page_nr++;
220                         ret += buf->len;
221
222                         if (pipe->files)
223                                 do_wakeup = 1;
224
225                         if (!--spd->nr_pages)
226                                 break;
227                         if (pipe->nrbufs < pipe->buffers)
228                                 continue;
229
230                         break;
231                 }
232
233                 if (spd->flags & SPLICE_F_NONBLOCK) {
234                         if (!ret)
235                                 ret = -EAGAIN;
236                         break;
237                 }
238
239                 if (signal_pending(current)) {
240                         if (!ret)
241                                 ret = -ERESTARTSYS;
242                         break;
243                 }
244
245                 if (do_wakeup) {
246                         smp_mb();
247                         if (waitqueue_active(&pipe->wait))
248                                 wake_up_interruptible_sync(&pipe->wait);
249                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
250                         do_wakeup = 0;
251                 }
252
253                 pipe->waiting_writers++;
254                 pipe_wait(pipe);
255                 pipe->waiting_writers--;
256         }
257
258         pipe_unlock(pipe);
259
260         if (do_wakeup)
261                 wakeup_pipe_readers(pipe);
262
263         while (page_nr < spd_pages)
264                 spd->spd_release(spd, page_nr++);
265
266         return ret;
267 }
268
269 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
270 {
271         page_cache_release(spd->pages[i]);
272 }
273
274 /*
275  * Check if we need to grow the arrays holding pages and partial page
276  * descriptions.
277  */
278 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
279 {
280         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
281
282         spd->nr_pages_max = buffers;
283         if (buffers <= PIPE_DEF_BUFFERS)
284                 return 0;
285
286         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
287         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
288
289         if (spd->pages && spd->partial)
290                 return 0;
291
292         kfree(spd->pages);
293         kfree(spd->partial);
294         return -ENOMEM;
295 }
296
297 void splice_shrink_spd(struct splice_pipe_desc *spd)
298 {
299         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
300                 return;
301
302         kfree(spd->pages);
303         kfree(spd->partial);
304 }
305
306 static int
307 __generic_file_splice_read(struct file *in, loff_t *ppos,
308                            struct pipe_inode_info *pipe, size_t len,
309                            unsigned int flags)
310 {
311         struct address_space *mapping = in->f_mapping;
312         unsigned int loff, nr_pages, req_pages;
313         struct page *pages[PIPE_DEF_BUFFERS];
314         struct partial_page partial[PIPE_DEF_BUFFERS];
315         struct page *page;
316         pgoff_t index, end_index;
317         loff_t isize;
318         int error, page_nr;
319         struct splice_pipe_desc spd = {
320                 .pages = pages,
321                 .partial = partial,
322                 .nr_pages_max = PIPE_DEF_BUFFERS,
323                 .flags = flags,
324                 .ops = &page_cache_pipe_buf_ops,
325                 .spd_release = spd_release_page,
326         };
327
328         if (splice_grow_spd(pipe, &spd))
329                 return -ENOMEM;
330
331         index = *ppos >> PAGE_CACHE_SHIFT;
332         loff = *ppos & ~PAGE_CACHE_MASK;
333         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
334         nr_pages = min(req_pages, spd.nr_pages_max);
335
336         /*
337          * Lookup the (hopefully) full range of pages we need.
338          */
339         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
340         index += spd.nr_pages;
341
342         /*
343          * If find_get_pages_contig() returned fewer pages than we needed,
344          * readahead/allocate the rest and fill in the holes.
345          */
346         if (spd.nr_pages < nr_pages)
347                 page_cache_sync_readahead(mapping, &in->f_ra, in,
348                                 index, req_pages - spd.nr_pages);
349
350         error = 0;
351         while (spd.nr_pages < nr_pages) {
352                 /*
353                  * Page could be there, find_get_pages_contig() breaks on
354                  * the first hole.
355                  */
356                 page = find_get_page(mapping, index);
357                 if (!page) {
358                         /*
359                          * page didn't exist, allocate one.
360                          */
361                         page = page_cache_alloc_cold(mapping);
362                         if (!page)
363                                 break;
364
365                         error = add_to_page_cache_lru(page, mapping, index,
366                                                 GFP_KERNEL);
367                         if (unlikely(error)) {
368                                 page_cache_release(page);
369                                 if (error == -EEXIST)
370                                         continue;
371                                 break;
372                         }
373                         /*
374                          * add_to_page_cache() locks the page, unlock it
375                          * to avoid convoluting the logic below even more.
376                          */
377                         unlock_page(page);
378                 }
379
380                 spd.pages[spd.nr_pages++] = page;
381                 index++;
382         }
383
384         /*
385          * Now loop over the map and see if we need to start IO on any
386          * pages, fill in the partial map, etc.
387          */
388         index = *ppos >> PAGE_CACHE_SHIFT;
389         nr_pages = spd.nr_pages;
390         spd.nr_pages = 0;
391         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
392                 unsigned int this_len;
393
394                 if (!len)
395                         break;
396
397                 /*
398                  * this_len is the max we'll use from this page
399                  */
400                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
401                 page = spd.pages[page_nr];
402
403                 if (PageReadahead(page))
404                         page_cache_async_readahead(mapping, &in->f_ra, in,
405                                         page, index, req_pages - page_nr);
406
407                 /*
408                  * If the page isn't uptodate, we may need to start io on it
409                  */
410                 if (!PageUptodate(page)) {
411                         lock_page(page);
412
413                         /*
414                          * Page was truncated, or invalidated by the
415                          * filesystem.  Redo the find/create, but this time the
416                          * page is kept locked, so there's no chance of another
417                          * race with truncate/invalidate.
418                          */
419                         if (!page->mapping) {
420                                 unlock_page(page);
421                                 page = find_or_create_page(mapping, index,
422                                                 mapping_gfp_mask(mapping));
423
424                                 if (!page) {
425                                         error = -ENOMEM;
426                                         break;
427                                 }
428                                 page_cache_release(spd.pages[page_nr]);
429                                 spd.pages[page_nr] = page;
430                         }
431                         /*
432                          * page was already under io and is now done, great
433                          */
434                         if (PageUptodate(page)) {
435                                 unlock_page(page);
436                                 goto fill_it;
437                         }
438
439                         /*
440                          * need to read in the page
441                          */
442                         error = mapping->a_ops->readpage(in, page);
443                         if (unlikely(error)) {
444                                 /*
445                                  * We really should re-lookup the page here,
446                                  * but it complicates things a lot. Instead
447                                  * lets just do what we already stored, and
448                                  * we'll get it the next time we are called.
449                                  */
450                                 if (error == AOP_TRUNCATED_PAGE)
451                                         error = 0;
452
453                                 break;
454                         }
455                 }
456 fill_it:
457                 /*
458                  * i_size must be checked after PageUptodate.
459                  */
460                 isize = i_size_read(mapping->host);
461                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
462                 if (unlikely(!isize || index > end_index))
463                         break;
464
465                 /*
466                  * if this is the last page, see if we need to shrink
467                  * the length and stop
468                  */
469                 if (end_index == index) {
470                         unsigned int plen;
471
472                         /*
473                          * max good bytes in this page
474                          */
475                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
476                         if (plen <= loff)
477                                 break;
478
479                         /*
480                          * force quit after adding this page
481                          */
482                         this_len = min(this_len, plen - loff);
483                         len = this_len;
484                 }
485
486                 spd.partial[page_nr].offset = loff;
487                 spd.partial[page_nr].len = this_len;
488                 len -= this_len;
489                 loff = 0;
490                 spd.nr_pages++;
491                 index++;
492         }
493
494         /*
495          * Release any pages at the end, if we quit early. 'page_nr' is how far
496          * we got, 'nr_pages' is how many pages are in the map.
497          */
498         while (page_nr < nr_pages)
499                 page_cache_release(spd.pages[page_nr++]);
500         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
501
502         if (spd.nr_pages)
503                 error = splice_to_pipe(pipe, &spd);
504
505         splice_shrink_spd(&spd);
506         return error;
507 }
508
509 /**
510  * generic_file_splice_read - splice data from file to a pipe
511  * @in:         file to splice from
512  * @ppos:       position in @in
513  * @pipe:       pipe to splice to
514  * @len:        number of bytes to splice
515  * @flags:      splice modifier flags
516  *
517  * Description:
518  *    Will read pages from given file and fill them into a pipe. Can be
519  *    used as long as the address_space operations for the source implements
520  *    a readpage() hook.
521  *
522  */
523 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
524                                  struct pipe_inode_info *pipe, size_t len,
525                                  unsigned int flags)
526 {
527         loff_t isize, left;
528         int ret;
529
530         isize = i_size_read(in->f_mapping->host);
531         if (unlikely(*ppos >= isize))
532                 return 0;
533
534         left = isize - *ppos;
535         if (unlikely(left < len))
536                 len = left;
537
538         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
539         if (ret > 0) {
540                 *ppos += ret;
541                 file_accessed(in);
542         }
543
544         return ret;
545 }
546 EXPORT_SYMBOL(generic_file_splice_read);
547
548 static const struct pipe_buf_operations default_pipe_buf_ops = {
549         .can_merge = 0,
550         .map = generic_pipe_buf_map,
551         .unmap = generic_pipe_buf_unmap,
552         .confirm = generic_pipe_buf_confirm,
553         .release = generic_pipe_buf_release,
554         .steal = generic_pipe_buf_steal,
555         .get = generic_pipe_buf_get,
556 };
557
558 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
559                             unsigned long vlen, loff_t offset)
560 {
561         mm_segment_t old_fs;
562         loff_t pos = offset;
563         ssize_t res;
564
565         old_fs = get_fs();
566         set_fs(get_ds());
567         /* The cast to a user pointer is valid due to the set_fs() */
568         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
569         set_fs(old_fs);
570
571         return res;
572 }
573
574 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
575                             loff_t pos)
576 {
577         mm_segment_t old_fs;
578         ssize_t res;
579
580         old_fs = get_fs();
581         set_fs(get_ds());
582         /* The cast to a user pointer is valid due to the set_fs() */
583         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
584         set_fs(old_fs);
585
586         return res;
587 }
588 EXPORT_SYMBOL(kernel_write);
589
590 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
591                                  struct pipe_inode_info *pipe, size_t len,
592                                  unsigned int flags)
593 {
594         unsigned int nr_pages;
595         unsigned int nr_freed;
596         size_t offset;
597         struct page *pages[PIPE_DEF_BUFFERS];
598         struct partial_page partial[PIPE_DEF_BUFFERS];
599         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
600         ssize_t res;
601         size_t this_len;
602         int error;
603         int i;
604         struct splice_pipe_desc spd = {
605                 .pages = pages,
606                 .partial = partial,
607                 .nr_pages_max = PIPE_DEF_BUFFERS,
608                 .flags = flags,
609                 .ops = &default_pipe_buf_ops,
610                 .spd_release = spd_release_page,
611         };
612
613         if (splice_grow_spd(pipe, &spd))
614                 return -ENOMEM;
615
616         res = -ENOMEM;
617         vec = __vec;
618         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
619                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
620                 if (!vec)
621                         goto shrink_ret;
622         }
623
624         offset = *ppos & ~PAGE_CACHE_MASK;
625         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
626
627         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
628                 struct page *page;
629
630                 page = alloc_page(GFP_USER);
631                 error = -ENOMEM;
632                 if (!page)
633                         goto err;
634
635                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
636                 vec[i].iov_base = (void __user *) page_address(page);
637                 vec[i].iov_len = this_len;
638                 spd.pages[i] = page;
639                 spd.nr_pages++;
640                 len -= this_len;
641                 offset = 0;
642         }
643
644         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
645         if (res < 0) {
646                 error = res;
647                 goto err;
648         }
649
650         error = 0;
651         if (!res)
652                 goto err;
653
654         nr_freed = 0;
655         for (i = 0; i < spd.nr_pages; i++) {
656                 this_len = min_t(size_t, vec[i].iov_len, res);
657                 spd.partial[i].offset = 0;
658                 spd.partial[i].len = this_len;
659                 if (!this_len) {
660                         __free_page(spd.pages[i]);
661                         spd.pages[i] = NULL;
662                         nr_freed++;
663                 }
664                 res -= this_len;
665         }
666         spd.nr_pages -= nr_freed;
667
668         res = splice_to_pipe(pipe, &spd);
669         if (res > 0)
670                 *ppos += res;
671
672 shrink_ret:
673         if (vec != __vec)
674                 kfree(vec);
675         splice_shrink_spd(&spd);
676         return res;
677
678 err:
679         for (i = 0; i < spd.nr_pages; i++)
680                 __free_page(spd.pages[i]);
681
682         res = error;
683         goto shrink_ret;
684 }
685 EXPORT_SYMBOL(default_file_splice_read);
686
687 /*
688  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
689  * using sendpage(). Return the number of bytes sent.
690  */
691 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
692                             struct pipe_buffer *buf, struct splice_desc *sd)
693 {
694         struct file *file = sd->u.file;
695         loff_t pos = sd->pos;
696         int more;
697
698         if (!likely(file->f_op && file->f_op->sendpage))
699                 return -EINVAL;
700
701         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
702
703         if (sd->len < sd->total_len && pipe->nrbufs > 1)
704                 more |= MSG_SENDPAGE_NOTLAST;
705
706         return file->f_op->sendpage(file, buf->page, buf->offset,
707                                     sd->len, &pos, more);
708 }
709
710 /*
711  * This is a little more tricky than the file -> pipe splicing. There are
712  * basically three cases:
713  *
714  *      - Destination page already exists in the address space and there
715  *        are users of it. For that case we have no other option that
716  *        copying the data. Tough luck.
717  *      - Destination page already exists in the address space, but there
718  *        are no users of it. Make sure it's uptodate, then drop it. Fall
719  *        through to last case.
720  *      - Destination page does not exist, we can add the pipe page to
721  *        the page cache and avoid the copy.
722  *
723  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
724  * sd->flags), we attempt to migrate pages from the pipe to the output
725  * file address space page cache. This is possible if no one else has
726  * the pipe page referenced outside of the pipe and page cache. If
727  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
728  * a new page in the output file page cache and fill/dirty that.
729  */
730 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
731                  struct splice_desc *sd)
732 {
733         struct file *file = sd->u.file;
734         struct address_space *mapping = file->f_mapping;
735         unsigned int offset, this_len;
736         struct page *page;
737         void *fsdata;
738         int ret;
739
740         offset = sd->pos & ~PAGE_CACHE_MASK;
741
742         this_len = sd->len;
743         if (this_len + offset > PAGE_CACHE_SIZE)
744                 this_len = PAGE_CACHE_SIZE - offset;
745
746         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
747                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
748         if (unlikely(ret))
749                 goto out;
750
751         if (buf->page != page) {
752                 char *src = buf->ops->map(pipe, buf, 1);
753                 char *dst = kmap_atomic(page);
754
755                 memcpy(dst + offset, src + buf->offset, this_len);
756                 flush_dcache_page(page);
757                 kunmap_atomic(dst);
758                 buf->ops->unmap(pipe, buf, src);
759         }
760         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
761                                 page, fsdata);
762 out:
763         return ret;
764 }
765 EXPORT_SYMBOL(pipe_to_file);
766
767 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
768 {
769         smp_mb();
770         if (waitqueue_active(&pipe->wait))
771                 wake_up_interruptible(&pipe->wait);
772         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
773 }
774
775 /**
776  * splice_from_pipe_feed - feed available data from a pipe to a file
777  * @pipe:       pipe to splice from
778  * @sd:         information to @actor
779  * @actor:      handler that splices the data
780  *
781  * Description:
782  *    This function loops over the pipe and calls @actor to do the
783  *    actual moving of a single struct pipe_buffer to the desired
784  *    destination.  It returns when there's no more buffers left in
785  *    the pipe or if the requested number of bytes (@sd->total_len)
786  *    have been copied.  It returns a positive number (one) if the
787  *    pipe needs to be filled with more data, zero if the required
788  *    number of bytes have been copied and -errno on error.
789  *
790  *    This, together with splice_from_pipe_{begin,end,next}, may be
791  *    used to implement the functionality of __splice_from_pipe() when
792  *    locking is required around copying the pipe buffers to the
793  *    destination.
794  */
795 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
796                           splice_actor *actor)
797 {
798         int ret;
799
800         while (pipe->nrbufs) {
801                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
802                 const struct pipe_buf_operations *ops = buf->ops;
803
804                 sd->len = buf->len;
805                 if (sd->len > sd->total_len)
806                         sd->len = sd->total_len;
807
808                 ret = buf->ops->confirm(pipe, buf);
809                 if (unlikely(ret)) {
810                         if (ret == -ENODATA)
811                                 ret = 0;
812                         return ret;
813                 }
814
815                 ret = actor(pipe, buf, sd);
816                 if (ret <= 0)
817                         return ret;
818
819                 buf->offset += ret;
820                 buf->len -= ret;
821
822                 sd->num_spliced += ret;
823                 sd->len -= ret;
824                 sd->pos += ret;
825                 sd->total_len -= ret;
826
827                 if (!buf->len) {
828                         buf->ops = NULL;
829                         ops->release(pipe, buf);
830                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
831                         pipe->nrbufs--;
832                         if (pipe->files)
833                                 sd->need_wakeup = true;
834                 }
835
836                 if (!sd->total_len)
837                         return 0;
838         }
839
840         return 1;
841 }
842 EXPORT_SYMBOL(splice_from_pipe_feed);
843
844 /**
845  * splice_from_pipe_next - wait for some data to splice from
846  * @pipe:       pipe to splice from
847  * @sd:         information about the splice operation
848  *
849  * Description:
850  *    This function will wait for some data and return a positive
851  *    value (one) if pipe buffers are available.  It will return zero
852  *    or -errno if no more data needs to be spliced.
853  */
854 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
855 {
856         while (!pipe->nrbufs) {
857                 if (!pipe->writers)
858                         return 0;
859
860                 if (!pipe->waiting_writers && sd->num_spliced)
861                         return 0;
862
863                 if (sd->flags & SPLICE_F_NONBLOCK)
864                         return -EAGAIN;
865
866                 if (signal_pending(current))
867                         return -ERESTARTSYS;
868
869                 if (sd->need_wakeup) {
870                         wakeup_pipe_writers(pipe);
871                         sd->need_wakeup = false;
872                 }
873
874                 pipe_wait(pipe);
875         }
876
877         return 1;
878 }
879 EXPORT_SYMBOL(splice_from_pipe_next);
880
881 /**
882  * splice_from_pipe_begin - start splicing from pipe
883  * @sd:         information about the splice operation
884  *
885  * Description:
886  *    This function should be called before a loop containing
887  *    splice_from_pipe_next() and splice_from_pipe_feed() to
888  *    initialize the necessary fields of @sd.
889  */
890 void splice_from_pipe_begin(struct splice_desc *sd)
891 {
892         sd->num_spliced = 0;
893         sd->need_wakeup = false;
894 }
895 EXPORT_SYMBOL(splice_from_pipe_begin);
896
897 /**
898  * splice_from_pipe_end - finish splicing from pipe
899  * @pipe:       pipe to splice from
900  * @sd:         information about the splice operation
901  *
902  * Description:
903  *    This function will wake up pipe writers if necessary.  It should
904  *    be called after a loop containing splice_from_pipe_next() and
905  *    splice_from_pipe_feed().
906  */
907 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
908 {
909         if (sd->need_wakeup)
910                 wakeup_pipe_writers(pipe);
911 }
912 EXPORT_SYMBOL(splice_from_pipe_end);
913
914 /**
915  * __splice_from_pipe - splice data from a pipe to given actor
916  * @pipe:       pipe to splice from
917  * @sd:         information to @actor
918  * @actor:      handler that splices the data
919  *
920  * Description:
921  *    This function does little more than loop over the pipe and call
922  *    @actor to do the actual moving of a single struct pipe_buffer to
923  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
924  *    pipe_to_user.
925  *
926  */
927 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
928                            splice_actor *actor)
929 {
930         int ret;
931
932         splice_from_pipe_begin(sd);
933         do {
934                 ret = splice_from_pipe_next(pipe, sd);
935                 if (ret > 0)
936                         ret = splice_from_pipe_feed(pipe, sd, actor);
937         } while (ret > 0);
938         splice_from_pipe_end(pipe, sd);
939
940         return sd->num_spliced ? sd->num_spliced : ret;
941 }
942 EXPORT_SYMBOL(__splice_from_pipe);
943
944 /**
945  * splice_from_pipe - splice data from a pipe to a file
946  * @pipe:       pipe to splice from
947  * @out:        file to splice to
948  * @ppos:       position in @out
949  * @len:        how many bytes to splice
950  * @flags:      splice modifier flags
951  * @actor:      handler that splices the data
952  *
953  * Description:
954  *    See __splice_from_pipe. This function locks the pipe inode,
955  *    otherwise it's identical to __splice_from_pipe().
956  *
957  */
958 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
959                          loff_t *ppos, size_t len, unsigned int flags,
960                          splice_actor *actor)
961 {
962         ssize_t ret;
963         struct splice_desc sd = {
964                 .total_len = len,
965                 .flags = flags,
966                 .pos = *ppos,
967                 .u.file = out,
968         };
969
970         pipe_lock(pipe);
971         ret = __splice_from_pipe(pipe, &sd, actor);
972         pipe_unlock(pipe);
973
974         return ret;
975 }
976
977 /**
978  * generic_file_splice_write - splice data from a pipe to a file
979  * @pipe:       pipe info
980  * @out:        file to write to
981  * @ppos:       position in @out
982  * @len:        number of bytes to splice
983  * @flags:      splice modifier flags
984  *
985  * Description:
986  *    Will either move or copy pages (determined by @flags options) from
987  *    the given pipe inode to the given file.
988  *
989  */
990 ssize_t
991 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
992                           loff_t *ppos, size_t len, unsigned int flags)
993 {
994         struct address_space *mapping = out->f_mapping;
995         struct inode *inode = mapping->host;
996         struct splice_desc sd = {
997                 .total_len = len,
998                 .flags = flags,
999                 .pos = *ppos,
1000                 .u.file = out,
1001         };
1002         ssize_t ret;
1003
1004         pipe_lock(pipe);
1005
1006         splice_from_pipe_begin(&sd);
1007         do {
1008                 ret = splice_from_pipe_next(pipe, &sd);
1009                 if (ret <= 0)
1010                         break;
1011
1012                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1013                 ret = file_remove_suid(out);
1014                 if (!ret) {
1015                         ret = file_update_time(out);
1016                         if (!ret)
1017                                 ret = splice_from_pipe_feed(pipe, &sd,
1018                                                             pipe_to_file);
1019                 }
1020                 mutex_unlock(&inode->i_mutex);
1021         } while (ret > 0);
1022         splice_from_pipe_end(pipe, &sd);
1023
1024         pipe_unlock(pipe);
1025
1026         if (sd.num_spliced)
1027                 ret = sd.num_spliced;
1028
1029         if (ret > 0) {
1030                 int err;
1031
1032                 err = generic_write_sync(out, *ppos, ret);
1033                 if (err)
1034                         ret = err;
1035                 else
1036                         *ppos += ret;
1037                 balance_dirty_pages_ratelimited(mapping);
1038         }
1039
1040         return ret;
1041 }
1042
1043 EXPORT_SYMBOL(generic_file_splice_write);
1044
1045 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1046                           struct splice_desc *sd)
1047 {
1048         int ret;
1049         void *data;
1050         loff_t tmp = sd->pos;
1051
1052         data = buf->ops->map(pipe, buf, 0);
1053         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1054         buf->ops->unmap(pipe, buf, data);
1055
1056         return ret;
1057 }
1058
1059 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1060                                          struct file *out, loff_t *ppos,
1061                                          size_t len, unsigned int flags)
1062 {
1063         ssize_t ret;
1064
1065         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1066         if (ret > 0)
1067                 *ppos += ret;
1068
1069         return ret;
1070 }
1071
1072 /**
1073  * generic_splice_sendpage - splice data from a pipe to a socket
1074  * @pipe:       pipe to splice from
1075  * @out:        socket to write to
1076  * @ppos:       position in @out
1077  * @len:        number of bytes to splice
1078  * @flags:      splice modifier flags
1079  *
1080  * Description:
1081  *    Will send @len bytes from the pipe to a network socket. No data copying
1082  *    is involved.
1083  *
1084  */
1085 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1086                                 loff_t *ppos, size_t len, unsigned int flags)
1087 {
1088         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1089 }
1090
1091 EXPORT_SYMBOL(generic_splice_sendpage);
1092
1093 /*
1094  * Attempt to initiate a splice from pipe to file.
1095  */
1096 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1097                            loff_t *ppos, size_t len, unsigned int flags)
1098 {
1099         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1100                                 loff_t *, size_t, unsigned int);
1101         int ret;
1102
1103         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1104                 return -EBADF;
1105
1106         if (unlikely(out->f_flags & O_APPEND))
1107                 return -EINVAL;
1108
1109         ret = rw_verify_area(WRITE, out, ppos, len);
1110         if (unlikely(ret < 0))
1111                 return ret;
1112
1113         if (out->f_op && out->f_op->splice_write)
1114                 splice_write = out->f_op->splice_write;
1115         else
1116                 splice_write = default_file_splice_write;
1117
1118         file_start_write(out);
1119         ret = splice_write(pipe, out, ppos, len, flags);
1120         file_end_write(out);
1121         return ret;
1122 }
1123
1124 /*
1125  * Attempt to initiate a splice from a file to a pipe.
1126  */
1127 static long do_splice_to(struct file *in, loff_t *ppos,
1128                          struct pipe_inode_info *pipe, size_t len,
1129                          unsigned int flags)
1130 {
1131         ssize_t (*splice_read)(struct file *, loff_t *,
1132                                struct pipe_inode_info *, size_t, unsigned int);
1133         int ret;
1134
1135         if (unlikely(!(in->f_mode & FMODE_READ)))
1136                 return -EBADF;
1137
1138         ret = rw_verify_area(READ, in, ppos, len);
1139         if (unlikely(ret < 0))
1140                 return ret;
1141
1142         if (in->f_op && in->f_op->splice_read)
1143                 splice_read = in->f_op->splice_read;
1144         else
1145                 splice_read = default_file_splice_read;
1146
1147         return splice_read(in, ppos, pipe, len, flags);
1148 }
1149
1150 /**
1151  * splice_direct_to_actor - splices data directly between two non-pipes
1152  * @in:         file to splice from
1153  * @sd:         actor information on where to splice to
1154  * @actor:      handles the data splicing
1155  *
1156  * Description:
1157  *    This is a special case helper to splice directly between two
1158  *    points, without requiring an explicit pipe. Internally an allocated
1159  *    pipe is cached in the process, and reused during the lifetime of
1160  *    that process.
1161  *
1162  */
1163 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1164                                splice_direct_actor *actor)
1165 {
1166         struct pipe_inode_info *pipe;
1167         long ret, bytes;
1168         umode_t i_mode;
1169         size_t len;
1170         int i, flags;
1171
1172         /*
1173          * We require the input being a regular file, as we don't want to
1174          * randomly drop data for eg socket -> socket splicing. Use the
1175          * piped splicing for that!
1176          */
1177         i_mode = file_inode(in)->i_mode;
1178         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1179                 return -EINVAL;
1180
1181         /*
1182          * neither in nor out is a pipe, setup an internal pipe attached to
1183          * 'out' and transfer the wanted data from 'in' to 'out' through that
1184          */
1185         pipe = current->splice_pipe;
1186         if (unlikely(!pipe)) {
1187                 pipe = alloc_pipe_info();
1188                 if (!pipe)
1189                         return -ENOMEM;
1190
1191                 /*
1192                  * We don't have an immediate reader, but we'll read the stuff
1193                  * out of the pipe right after the splice_to_pipe(). So set
1194                  * PIPE_READERS appropriately.
1195                  */
1196                 pipe->readers = 1;
1197
1198                 current->splice_pipe = pipe;
1199         }
1200
1201         /*
1202          * Do the splice.
1203          */
1204         ret = 0;
1205         bytes = 0;
1206         len = sd->total_len;
1207         flags = sd->flags;
1208
1209         /*
1210          * Don't block on output, we have to drain the direct pipe.
1211          */
1212         sd->flags &= ~SPLICE_F_NONBLOCK;
1213
1214         while (len) {
1215                 size_t read_len;
1216                 loff_t pos = sd->pos, prev_pos = pos;
1217
1218                 ret = do_splice_to(in, &pos, pipe, len, flags);
1219                 if (unlikely(ret <= 0))
1220                         goto out_release;
1221
1222                 read_len = ret;
1223                 sd->total_len = read_len;
1224
1225                 /*
1226                  * NOTE: nonblocking mode only applies to the input. We
1227                  * must not do the output in nonblocking mode as then we
1228                  * could get stuck data in the internal pipe:
1229                  */
1230                 ret = actor(pipe, sd);
1231                 if (unlikely(ret <= 0)) {
1232                         sd->pos = prev_pos;
1233                         goto out_release;
1234                 }
1235
1236                 bytes += ret;
1237                 len -= ret;
1238                 sd->pos = pos;
1239
1240                 if (ret < read_len) {
1241                         sd->pos = prev_pos + ret;
1242                         goto out_release;
1243                 }
1244         }
1245
1246 done:
1247         pipe->nrbufs = pipe->curbuf = 0;
1248         file_accessed(in);
1249         return bytes;
1250
1251 out_release:
1252         /*
1253          * If we did an incomplete transfer we must release
1254          * the pipe buffers in question:
1255          */
1256         for (i = 0; i < pipe->buffers; i++) {
1257                 struct pipe_buffer *buf = pipe->bufs + i;
1258
1259                 if (buf->ops) {
1260                         buf->ops->release(pipe, buf);
1261                         buf->ops = NULL;
1262                 }
1263         }
1264
1265         if (!bytes)
1266                 bytes = ret;
1267
1268         goto done;
1269 }
1270 EXPORT_SYMBOL(splice_direct_to_actor);
1271
1272 static int direct_splice_actor(struct pipe_inode_info *pipe,
1273                                struct splice_desc *sd)
1274 {
1275         struct file *file = sd->u.file;
1276
1277         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1278                               sd->flags);
1279 }
1280
1281 /**
1282  * do_splice_direct - splices data directly between two files
1283  * @in:         file to splice from
1284  * @ppos:       input file offset
1285  * @out:        file to splice to
1286  * @len:        number of bytes to splice
1287  * @flags:      splice modifier flags
1288  *
1289  * Description:
1290  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1291  *    doing it in the application would incur an extra system call
1292  *    (splice in + splice out, as compared to just sendfile()). So this helper
1293  *    can splice directly through a process-private pipe.
1294  *
1295  */
1296 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1297                       size_t len, unsigned int flags)
1298 {
1299         struct splice_desc sd = {
1300                 .len            = len,
1301                 .total_len      = len,
1302                 .flags          = flags,
1303                 .pos            = *ppos,
1304                 .u.file         = out,
1305         };
1306         long ret;
1307
1308         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1309         if (ret > 0)
1310                 *ppos = sd.pos;
1311
1312         return ret;
1313 }
1314
1315 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1316                                struct pipe_inode_info *opipe,
1317                                size_t len, unsigned int flags);
1318
1319 /*
1320  * Determine where to splice to/from.
1321  */
1322 static long do_splice(struct file *in, loff_t __user *off_in,
1323                       struct file *out, loff_t __user *off_out,
1324                       size_t len, unsigned int flags)
1325 {
1326         struct pipe_inode_info *ipipe;
1327         struct pipe_inode_info *opipe;
1328         loff_t offset, *off;
1329         long ret;
1330
1331         ipipe = get_pipe_info(in);
1332         opipe = get_pipe_info(out);
1333
1334         if (ipipe && opipe) {
1335                 if (off_in || off_out)
1336                         return -ESPIPE;
1337
1338                 if (!(in->f_mode & FMODE_READ))
1339                         return -EBADF;
1340
1341                 if (!(out->f_mode & FMODE_WRITE))
1342                         return -EBADF;
1343
1344                 /* Splicing to self would be fun, but... */
1345                 if (ipipe == opipe)
1346                         return -EINVAL;
1347
1348                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1349         }
1350
1351         if (ipipe) {
1352                 if (off_in)
1353                         return -ESPIPE;
1354                 if (off_out) {
1355                         if (!(out->f_mode & FMODE_PWRITE))
1356                                 return -EINVAL;
1357                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1358                                 return -EFAULT;
1359                         off = &offset;
1360                 } else
1361                         off = &out->f_pos;
1362
1363                 ret = do_splice_from(ipipe, out, off, len, flags);
1364
1365                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1366                         ret = -EFAULT;
1367
1368                 return ret;
1369         }
1370
1371         if (opipe) {
1372                 if (off_out)
1373                         return -ESPIPE;
1374                 if (off_in) {
1375                         if (!(in->f_mode & FMODE_PREAD))
1376                                 return -EINVAL;
1377                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1378                                 return -EFAULT;
1379                         off = &offset;
1380                 } else
1381                         off = &in->f_pos;
1382
1383                 ret = do_splice_to(in, off, opipe, len, flags);
1384
1385                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1386                         ret = -EFAULT;
1387
1388                 return ret;
1389         }
1390
1391         return -EINVAL;
1392 }
1393
1394 /*
1395  * Map an iov into an array of pages and offset/length tupples. With the
1396  * partial_page structure, we can map several non-contiguous ranges into
1397  * our ones pages[] map instead of splitting that operation into pieces.
1398  * Could easily be exported as a generic helper for other users, in which
1399  * case one would probably want to add a 'max_nr_pages' parameter as well.
1400  */
1401 static int get_iovec_page_array(const struct iovec __user *iov,
1402                                 unsigned int nr_vecs, struct page **pages,
1403                                 struct partial_page *partial, bool aligned,
1404                                 unsigned int pipe_buffers)
1405 {
1406         int buffers = 0, error = 0;
1407
1408         while (nr_vecs) {
1409                 unsigned long off, npages;
1410                 struct iovec entry;
1411                 void __user *base;
1412                 size_t len;
1413                 int i;
1414
1415                 error = -EFAULT;
1416                 if (copy_from_user(&entry, iov, sizeof(entry)))
1417                         break;
1418
1419                 base = entry.iov_base;
1420                 len = entry.iov_len;
1421
1422                 /*
1423                  * Sanity check this iovec. 0 read succeeds.
1424                  */
1425                 error = 0;
1426                 if (unlikely(!len))
1427                         break;
1428                 error = -EFAULT;
1429                 if (!access_ok(VERIFY_READ, base, len))
1430                         break;
1431
1432                 /*
1433                  * Get this base offset and number of pages, then map
1434                  * in the user pages.
1435                  */
1436                 off = (unsigned long) base & ~PAGE_MASK;
1437
1438                 /*
1439                  * If asked for alignment, the offset must be zero and the
1440                  * length a multiple of the PAGE_SIZE.
1441                  */
1442                 error = -EINVAL;
1443                 if (aligned && (off || len & ~PAGE_MASK))
1444                         break;
1445
1446                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1447                 if (npages > pipe_buffers - buffers)
1448                         npages = pipe_buffers - buffers;
1449
1450                 error = get_user_pages_fast((unsigned long)base, npages,
1451                                         0, &pages[buffers]);
1452
1453                 if (unlikely(error <= 0))
1454                         break;
1455
1456                 /*
1457                  * Fill this contiguous range into the partial page map.
1458                  */
1459                 for (i = 0; i < error; i++) {
1460                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1461
1462                         partial[buffers].offset = off;
1463                         partial[buffers].len = plen;
1464
1465                         off = 0;
1466                         len -= plen;
1467                         buffers++;
1468                 }
1469
1470                 /*
1471                  * We didn't complete this iov, stop here since it probably
1472                  * means we have to move some of this into a pipe to
1473                  * be able to continue.
1474                  */
1475                 if (len)
1476                         break;
1477
1478                 /*
1479                  * Don't continue if we mapped fewer pages than we asked for,
1480                  * or if we mapped the max number of pages that we have
1481                  * room for.
1482                  */
1483                 if (error < npages || buffers == pipe_buffers)
1484                         break;
1485
1486                 nr_vecs--;
1487                 iov++;
1488         }
1489
1490         if (buffers)
1491                 return buffers;
1492
1493         return error;
1494 }
1495
1496 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1497                         struct splice_desc *sd)
1498 {
1499         char *src;
1500         int ret;
1501
1502         /*
1503          * See if we can use the atomic maps, by prefaulting in the
1504          * pages and doing an atomic copy
1505          */
1506         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1507                 src = buf->ops->map(pipe, buf, 1);
1508                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1509                                                         sd->len);
1510                 buf->ops->unmap(pipe, buf, src);
1511                 if (!ret) {
1512                         ret = sd->len;
1513                         goto out;
1514                 }
1515         }
1516
1517         /*
1518          * No dice, use slow non-atomic map and copy
1519          */
1520         src = buf->ops->map(pipe, buf, 0);
1521
1522         ret = sd->len;
1523         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1524                 ret = -EFAULT;
1525
1526         buf->ops->unmap(pipe, buf, src);
1527 out:
1528         if (ret > 0)
1529                 sd->u.userptr += ret;
1530         return ret;
1531 }
1532
1533 /*
1534  * For lack of a better implementation, implement vmsplice() to userspace
1535  * as a simple copy of the pipes pages to the user iov.
1536  */
1537 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1538                              unsigned long nr_segs, unsigned int flags)
1539 {
1540         struct pipe_inode_info *pipe;
1541         struct splice_desc sd;
1542         ssize_t size;
1543         int error;
1544         long ret;
1545
1546         pipe = get_pipe_info(file);
1547         if (!pipe)
1548                 return -EBADF;
1549
1550         pipe_lock(pipe);
1551
1552         error = ret = 0;
1553         while (nr_segs) {
1554                 void __user *base;
1555                 size_t len;
1556
1557                 /*
1558                  * Get user address base and length for this iovec.
1559                  */
1560                 error = get_user(base, &iov->iov_base);
1561                 if (unlikely(error))
1562                         break;
1563                 error = get_user(len, &iov->iov_len);
1564                 if (unlikely(error))
1565                         break;
1566
1567                 /*
1568                  * Sanity check this iovec. 0 read succeeds.
1569                  */
1570                 if (unlikely(!len))
1571                         break;
1572                 if (unlikely(!base)) {
1573                         error = -EFAULT;
1574                         break;
1575                 }
1576
1577                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1578                         error = -EFAULT;
1579                         break;
1580                 }
1581
1582                 sd.len = 0;
1583                 sd.total_len = len;
1584                 sd.flags = flags;
1585                 sd.u.userptr = base;
1586                 sd.pos = 0;
1587
1588                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1589                 if (size < 0) {
1590                         if (!ret)
1591                                 ret = size;
1592
1593                         break;
1594                 }
1595
1596                 ret += size;
1597
1598                 if (size < len)
1599                         break;
1600
1601                 nr_segs--;
1602                 iov++;
1603         }
1604
1605         pipe_unlock(pipe);
1606
1607         if (!ret)
1608                 ret = error;
1609
1610         return ret;
1611 }
1612
1613 /*
1614  * vmsplice splices a user address range into a pipe. It can be thought of
1615  * as splice-from-memory, where the regular splice is splice-from-file (or
1616  * to file). In both cases the output is a pipe, naturally.
1617  */
1618 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1619                              unsigned long nr_segs, unsigned int flags)
1620 {
1621         struct pipe_inode_info *pipe;
1622         struct page *pages[PIPE_DEF_BUFFERS];
1623         struct partial_page partial[PIPE_DEF_BUFFERS];
1624         struct splice_pipe_desc spd = {
1625                 .pages = pages,
1626                 .partial = partial,
1627                 .nr_pages_max = PIPE_DEF_BUFFERS,
1628                 .flags = flags,
1629                 .ops = &user_page_pipe_buf_ops,
1630                 .spd_release = spd_release_page,
1631         };
1632         long ret;
1633
1634         pipe = get_pipe_info(file);
1635         if (!pipe)
1636                 return -EBADF;
1637
1638         if (splice_grow_spd(pipe, &spd))
1639                 return -ENOMEM;
1640
1641         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1642                                             spd.partial, false,
1643                                             spd.nr_pages_max);
1644         if (spd.nr_pages <= 0)
1645                 ret = spd.nr_pages;
1646         else
1647                 ret = splice_to_pipe(pipe, &spd);
1648
1649         splice_shrink_spd(&spd);
1650         return ret;
1651 }
1652
1653 /*
1654  * Note that vmsplice only really supports true splicing _from_ user memory
1655  * to a pipe, not the other way around. Splicing from user memory is a simple
1656  * operation that can be supported without any funky alignment restrictions
1657  * or nasty vm tricks. We simply map in the user memory and fill them into
1658  * a pipe. The reverse isn't quite as easy, though. There are two possible
1659  * solutions for that:
1660  *
1661  *      - memcpy() the data internally, at which point we might as well just
1662  *        do a regular read() on the buffer anyway.
1663  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1664  *        has restriction limitations on both ends of the pipe).
1665  *
1666  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1667  *
1668  */
1669 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1670                 unsigned long, nr_segs, unsigned int, flags)
1671 {
1672         struct fd f;
1673         long error;
1674
1675         if (unlikely(nr_segs > UIO_MAXIOV))
1676                 return -EINVAL;
1677         else if (unlikely(!nr_segs))
1678                 return 0;
1679
1680         error = -EBADF;
1681         f = fdget(fd);
1682         if (f.file) {
1683                 if (f.file->f_mode & FMODE_WRITE)
1684                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1685                 else if (f.file->f_mode & FMODE_READ)
1686                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1687
1688                 fdput(f);
1689         }
1690
1691         return error;
1692 }
1693
1694 #ifdef CONFIG_COMPAT
1695 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1696                     unsigned int, nr_segs, unsigned int, flags)
1697 {
1698         unsigned i;
1699         struct iovec __user *iov;
1700         if (nr_segs > UIO_MAXIOV)
1701                 return -EINVAL;
1702         iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1703         for (i = 0; i < nr_segs; i++) {
1704                 struct compat_iovec v;
1705                 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1706                     get_user(v.iov_len, &iov32[i].iov_len) ||
1707                     put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1708                     put_user(v.iov_len, &iov[i].iov_len))
1709                         return -EFAULT;
1710         }
1711         return sys_vmsplice(fd, iov, nr_segs, flags);
1712 }
1713 #endif
1714
1715 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1716                 int, fd_out, loff_t __user *, off_out,
1717                 size_t, len, unsigned int, flags)
1718 {
1719         struct fd in, out;
1720         long error;
1721
1722         if (unlikely(!len))
1723                 return 0;
1724
1725         error = -EBADF;
1726         in = fdget(fd_in);
1727         if (in.file) {
1728                 if (in.file->f_mode & FMODE_READ) {
1729                         out = fdget(fd_out);
1730                         if (out.file) {
1731                                 if (out.file->f_mode & FMODE_WRITE)
1732                                         error = do_splice(in.file, off_in,
1733                                                           out.file, off_out,
1734                                                           len, flags);
1735                                 fdput(out);
1736                         }
1737                 }
1738                 fdput(in);
1739         }
1740         return error;
1741 }
1742
1743 /*
1744  * Make sure there's data to read. Wait for input if we can, otherwise
1745  * return an appropriate error.
1746  */
1747 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1748 {
1749         int ret;
1750
1751         /*
1752          * Check ->nrbufs without the inode lock first. This function
1753          * is speculative anyways, so missing one is ok.
1754          */
1755         if (pipe->nrbufs)
1756                 return 0;
1757
1758         ret = 0;
1759         pipe_lock(pipe);
1760
1761         while (!pipe->nrbufs) {
1762                 if (signal_pending(current)) {
1763                         ret = -ERESTARTSYS;
1764                         break;
1765                 }
1766                 if (!pipe->writers)
1767                         break;
1768                 if (!pipe->waiting_writers) {
1769                         if (flags & SPLICE_F_NONBLOCK) {
1770                                 ret = -EAGAIN;
1771                                 break;
1772                         }
1773                 }
1774                 pipe_wait(pipe);
1775         }
1776
1777         pipe_unlock(pipe);
1778         return ret;
1779 }
1780
1781 /*
1782  * Make sure there's writeable room. Wait for room if we can, otherwise
1783  * return an appropriate error.
1784  */
1785 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1786 {
1787         int ret;
1788
1789         /*
1790          * Check ->nrbufs without the inode lock first. This function
1791          * is speculative anyways, so missing one is ok.
1792          */
1793         if (pipe->nrbufs < pipe->buffers)
1794                 return 0;
1795
1796         ret = 0;
1797         pipe_lock(pipe);
1798
1799         while (pipe->nrbufs >= pipe->buffers) {
1800                 if (!pipe->readers) {
1801                         send_sig(SIGPIPE, current, 0);
1802                         ret = -EPIPE;
1803                         break;
1804                 }
1805                 if (flags & SPLICE_F_NONBLOCK) {
1806                         ret = -EAGAIN;
1807                         break;
1808                 }
1809                 if (signal_pending(current)) {
1810                         ret = -ERESTARTSYS;
1811                         break;
1812                 }
1813                 pipe->waiting_writers++;
1814                 pipe_wait(pipe);
1815                 pipe->waiting_writers--;
1816         }
1817
1818         pipe_unlock(pipe);
1819         return ret;
1820 }
1821
1822 /*
1823  * Splice contents of ipipe to opipe.
1824  */
1825 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1826                                struct pipe_inode_info *opipe,
1827                                size_t len, unsigned int flags)
1828 {
1829         struct pipe_buffer *ibuf, *obuf;
1830         int ret = 0, nbuf;
1831         bool input_wakeup = false;
1832
1833
1834 retry:
1835         ret = ipipe_prep(ipipe, flags);
1836         if (ret)
1837                 return ret;
1838
1839         ret = opipe_prep(opipe, flags);
1840         if (ret)
1841                 return ret;
1842
1843         /*
1844          * Potential ABBA deadlock, work around it by ordering lock
1845          * grabbing by pipe info address. Otherwise two different processes
1846          * could deadlock (one doing tee from A -> B, the other from B -> A).
1847          */
1848         pipe_double_lock(ipipe, opipe);
1849
1850         do {
1851                 if (!opipe->readers) {
1852                         send_sig(SIGPIPE, current, 0);
1853                         if (!ret)
1854                                 ret = -EPIPE;
1855                         break;
1856                 }
1857
1858                 if (!ipipe->nrbufs && !ipipe->writers)
1859                         break;
1860
1861                 /*
1862                  * Cannot make any progress, because either the input
1863                  * pipe is empty or the output pipe is full.
1864                  */
1865                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1866                         /* Already processed some buffers, break */
1867                         if (ret)
1868                                 break;
1869
1870                         if (flags & SPLICE_F_NONBLOCK) {
1871                                 ret = -EAGAIN;
1872                                 break;
1873                         }
1874
1875                         /*
1876                          * We raced with another reader/writer and haven't
1877                          * managed to process any buffers.  A zero return
1878                          * value means EOF, so retry instead.
1879                          */
1880                         pipe_unlock(ipipe);
1881                         pipe_unlock(opipe);
1882                         goto retry;
1883                 }
1884
1885                 ibuf = ipipe->bufs + ipipe->curbuf;
1886                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1887                 obuf = opipe->bufs + nbuf;
1888
1889                 if (len >= ibuf->len) {
1890                         /*
1891                          * Simply move the whole buffer from ipipe to opipe
1892                          */
1893                         *obuf = *ibuf;
1894                         ibuf->ops = NULL;
1895                         opipe->nrbufs++;
1896                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1897                         ipipe->nrbufs--;
1898                         input_wakeup = true;
1899                 } else {
1900                         /*
1901                          * Get a reference to this pipe buffer,
1902                          * so we can copy the contents over.
1903                          */
1904                         ibuf->ops->get(ipipe, ibuf);
1905                         *obuf = *ibuf;
1906
1907                         /*
1908                          * Don't inherit the gift flag, we need to
1909                          * prevent multiple steals of this page.
1910                          */
1911                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1912
1913                         obuf->len = len;
1914                         opipe->nrbufs++;
1915                         ibuf->offset += obuf->len;
1916                         ibuf->len -= obuf->len;
1917                 }
1918                 ret += obuf->len;
1919                 len -= obuf->len;
1920         } while (len);
1921
1922         pipe_unlock(ipipe);
1923         pipe_unlock(opipe);
1924
1925         /*
1926          * If we put data in the output pipe, wakeup any potential readers.
1927          */
1928         if (ret > 0)
1929                 wakeup_pipe_readers(opipe);
1930
1931         if (input_wakeup)
1932                 wakeup_pipe_writers(ipipe);
1933
1934         return ret;
1935 }
1936
1937 /*
1938  * Link contents of ipipe to opipe.
1939  */
1940 static int link_pipe(struct pipe_inode_info *ipipe,
1941                      struct pipe_inode_info *opipe,
1942                      size_t len, unsigned int flags)
1943 {
1944         struct pipe_buffer *ibuf, *obuf;
1945         int ret = 0, i = 0, nbuf;
1946
1947         /*
1948          * Potential ABBA deadlock, work around it by ordering lock
1949          * grabbing by pipe info address. Otherwise two different processes
1950          * could deadlock (one doing tee from A -> B, the other from B -> A).
1951          */
1952         pipe_double_lock(ipipe, opipe);
1953
1954         do {
1955                 if (!opipe->readers) {
1956                         send_sig(SIGPIPE, current, 0);
1957                         if (!ret)
1958                                 ret = -EPIPE;
1959                         break;
1960                 }
1961
1962                 /*
1963                  * If we have iterated all input buffers or ran out of
1964                  * output room, break.
1965                  */
1966                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1967                         break;
1968
1969                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1970                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1971
1972                 /*
1973                  * Get a reference to this pipe buffer,
1974                  * so we can copy the contents over.
1975                  */
1976                 ibuf->ops->get(ipipe, ibuf);
1977
1978                 obuf = opipe->bufs + nbuf;
1979                 *obuf = *ibuf;
1980
1981                 /*
1982                  * Don't inherit the gift flag, we need to
1983                  * prevent multiple steals of this page.
1984                  */
1985                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1986
1987                 if (obuf->len > len)
1988                         obuf->len = len;
1989
1990                 opipe->nrbufs++;
1991                 ret += obuf->len;
1992                 len -= obuf->len;
1993                 i++;
1994         } while (len);
1995
1996         /*
1997          * return EAGAIN if we have the potential of some data in the
1998          * future, otherwise just return 0
1999          */
2000         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2001                 ret = -EAGAIN;
2002
2003         pipe_unlock(ipipe);
2004         pipe_unlock(opipe);
2005
2006         /*
2007          * If we put data in the output pipe, wakeup any potential readers.
2008          */
2009         if (ret > 0)
2010                 wakeup_pipe_readers(opipe);
2011
2012         return ret;
2013 }
2014
2015 /*
2016  * This is a tee(1) implementation that works on pipes. It doesn't copy
2017  * any data, it simply references the 'in' pages on the 'out' pipe.
2018  * The 'flags' used are the SPLICE_F_* variants, currently the only
2019  * applicable one is SPLICE_F_NONBLOCK.
2020  */
2021 static long do_tee(struct file *in, struct file *out, size_t len,
2022                    unsigned int flags)
2023 {
2024         struct pipe_inode_info *ipipe = get_pipe_info(in);
2025         struct pipe_inode_info *opipe = get_pipe_info(out);
2026         int ret = -EINVAL;
2027
2028         /*
2029          * Duplicate the contents of ipipe to opipe without actually
2030          * copying the data.
2031          */
2032         if (ipipe && opipe && ipipe != opipe) {
2033                 /*
2034                  * Keep going, unless we encounter an error. The ipipe/opipe
2035                  * ordering doesn't really matter.
2036                  */
2037                 ret = ipipe_prep(ipipe, flags);
2038                 if (!ret) {
2039                         ret = opipe_prep(opipe, flags);
2040                         if (!ret)
2041                                 ret = link_pipe(ipipe, opipe, len, flags);
2042                 }
2043         }
2044
2045         return ret;
2046 }
2047
2048 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2049 {
2050         struct fd in;
2051         int error;
2052
2053         if (unlikely(!len))
2054                 return 0;
2055
2056         error = -EBADF;
2057         in = fdget(fdin);
2058         if (in.file) {
2059                 if (in.file->f_mode & FMODE_READ) {
2060                         struct fd out = fdget(fdout);
2061                         if (out.file) {
2062                                 if (out.file->f_mode & FMODE_WRITE)
2063                                         error = do_tee(in.file, out.file,
2064                                                         len, flags);
2065                                 fdput(out);
2066                         }
2067                 }
2068                 fdput(in);
2069         }
2070
2071         return error;
2072 }