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