sysfs: fix race between readdir and lseek
[linux-3.10.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
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 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 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
166 {
167         smp_mb();
168         if (waitqueue_active(&pipe->wait))
169                 wake_up_interruptible(&pipe->wait);
170         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
171 }
172
173 /**
174  * splice_to_pipe - fill passed data into a pipe
175  * @pipe:       pipe to fill
176  * @spd:        data to fill
177  *
178  * Description:
179  *    @spd contains a map of pages and len/offset tuples, along with
180  *    the struct pipe_buf_operations associated with these pages. This
181  *    function will link that data to the pipe.
182  *
183  */
184 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
185                        struct splice_pipe_desc *spd)
186 {
187         unsigned int spd_pages = spd->nr_pages;
188         int ret, do_wakeup, page_nr;
189
190         ret = 0;
191         do_wakeup = 0;
192         page_nr = 0;
193
194         pipe_lock(pipe);
195
196         for (;;) {
197                 if (!pipe->readers) {
198                         send_sig(SIGPIPE, current, 0);
199                         if (!ret)
200                                 ret = -EPIPE;
201                         break;
202                 }
203
204                 if (pipe->nrbufs < pipe->buffers) {
205                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
206                         struct pipe_buffer *buf = pipe->bufs + newbuf;
207
208                         buf->page = spd->pages[page_nr];
209                         buf->offset = spd->partial[page_nr].offset;
210                         buf->len = spd->partial[page_nr].len;
211                         buf->private = spd->partial[page_nr].private;
212                         buf->ops = spd->ops;
213                         if (spd->flags & SPLICE_F_GIFT)
214                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
215
216                         pipe->nrbufs++;
217                         page_nr++;
218                         ret += buf->len;
219
220                         if (pipe->inode)
221                                 do_wakeup = 1;
222
223                         if (!--spd->nr_pages)
224                                 break;
225                         if (pipe->nrbufs < pipe->buffers)
226                                 continue;
227
228                         break;
229                 }
230
231                 if (spd->flags & SPLICE_F_NONBLOCK) {
232                         if (!ret)
233                                 ret = -EAGAIN;
234                         break;
235                 }
236
237                 if (signal_pending(current)) {
238                         if (!ret)
239                                 ret = -ERESTARTSYS;
240                         break;
241                 }
242
243                 if (do_wakeup) {
244                         smp_mb();
245                         if (waitqueue_active(&pipe->wait))
246                                 wake_up_interruptible_sync(&pipe->wait);
247                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
248                         do_wakeup = 0;
249                 }
250
251                 pipe->waiting_writers++;
252                 pipe_wait(pipe);
253                 pipe->waiting_writers--;
254         }
255
256         pipe_unlock(pipe);
257
258         if (do_wakeup)
259                 wakeup_pipe_readers(pipe);
260
261         while (page_nr < spd_pages)
262                 spd->spd_release(spd, page_nr++);
263
264         return ret;
265 }
266
267 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
268 {
269         page_cache_release(spd->pages[i]);
270 }
271
272 /*
273  * Check if we need to grow the arrays holding pages and partial page
274  * descriptions.
275  */
276 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
277 {
278         unsigned int buffers = ACCESS_ONCE(pipe->buffers);
279
280         spd->nr_pages_max = buffers;
281         if (buffers <= PIPE_DEF_BUFFERS)
282                 return 0;
283
284         spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
285         spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
286
287         if (spd->pages && spd->partial)
288                 return 0;
289
290         kfree(spd->pages);
291         kfree(spd->partial);
292         return -ENOMEM;
293 }
294
295 void splice_shrink_spd(struct splice_pipe_desc *spd)
296 {
297         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
298                 return;
299
300         kfree(spd->pages);
301         kfree(spd->partial);
302 }
303
304 static int
305 __generic_file_splice_read(struct file *in, loff_t *ppos,
306                            struct pipe_inode_info *pipe, size_t len,
307                            unsigned int flags)
308 {
309         struct address_space *mapping = in->f_mapping;
310         unsigned int loff, nr_pages, req_pages;
311         struct page *pages[PIPE_DEF_BUFFERS];
312         struct partial_page partial[PIPE_DEF_BUFFERS];
313         struct page *page;
314         pgoff_t index, end_index;
315         loff_t isize;
316         int error, page_nr;
317         struct splice_pipe_desc spd = {
318                 .pages = pages,
319                 .partial = partial,
320                 .nr_pages_max = PIPE_DEF_BUFFERS,
321                 .flags = flags,
322                 .ops = &page_cache_pipe_buf_ops,
323                 .spd_release = spd_release_page,
324         };
325
326         if (splice_grow_spd(pipe, &spd))
327                 return -ENOMEM;
328
329         index = *ppos >> PAGE_CACHE_SHIFT;
330         loff = *ppos & ~PAGE_CACHE_MASK;
331         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
332         nr_pages = min(req_pages, spd.nr_pages_max);
333
334         /*
335          * Lookup the (hopefully) full range of pages we need.
336          */
337         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
338         index += spd.nr_pages;
339
340         /*
341          * If find_get_pages_contig() returned fewer pages than we needed,
342          * readahead/allocate the rest and fill in the holes.
343          */
344         if (spd.nr_pages < nr_pages)
345                 page_cache_sync_readahead(mapping, &in->f_ra, in,
346                                 index, req_pages - spd.nr_pages);
347
348         error = 0;
349         while (spd.nr_pages < nr_pages) {
350                 /*
351                  * Page could be there, find_get_pages_contig() breaks on
352                  * the first hole.
353                  */
354                 page = find_get_page(mapping, index);
355                 if (!page) {
356                         /*
357                          * page didn't exist, allocate one.
358                          */
359                         page = page_cache_alloc_cold(mapping);
360                         if (!page)
361                                 break;
362
363                         error = add_to_page_cache_lru(page, mapping, index,
364                                                 GFP_KERNEL);
365                         if (unlikely(error)) {
366                                 page_cache_release(page);
367                                 if (error == -EEXIST)
368                                         continue;
369                                 break;
370                         }
371                         /*
372                          * add_to_page_cache() locks the page, unlock it
373                          * to avoid convoluting the logic below even more.
374                          */
375                         unlock_page(page);
376                 }
377
378                 spd.pages[spd.nr_pages++] = page;
379                 index++;
380         }
381
382         /*
383          * Now loop over the map and see if we need to start IO on any
384          * pages, fill in the partial map, etc.
385          */
386         index = *ppos >> PAGE_CACHE_SHIFT;
387         nr_pages = spd.nr_pages;
388         spd.nr_pages = 0;
389         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
390                 unsigned int this_len;
391
392                 if (!len)
393                         break;
394
395                 /*
396                  * this_len is the max we'll use from this page
397                  */
398                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
399                 page = spd.pages[page_nr];
400
401                 if (PageReadahead(page))
402                         page_cache_async_readahead(mapping, &in->f_ra, in,
403                                         page, index, req_pages - page_nr);
404
405                 /*
406                  * If the page isn't uptodate, we may need to start io on it
407                  */
408                 if (!PageUptodate(page)) {
409                         lock_page(page);
410
411                         /*
412                          * Page was truncated, or invalidated by the
413                          * filesystem.  Redo the find/create, but this time the
414                          * page is kept locked, so there's no chance of another
415                          * race with truncate/invalidate.
416                          */
417                         if (!page->mapping) {
418                                 unlock_page(page);
419                                 page = find_or_create_page(mapping, index,
420                                                 mapping_gfp_mask(mapping));
421
422                                 if (!page) {
423                                         error = -ENOMEM;
424                                         break;
425                                 }
426                                 page_cache_release(spd.pages[page_nr]);
427                                 spd.pages[page_nr] = page;
428                         }
429                         /*
430                          * page was already under io and is now done, great
431                          */
432                         if (PageUptodate(page)) {
433                                 unlock_page(page);
434                                 goto fill_it;
435                         }
436
437                         /*
438                          * need to read in the page
439                          */
440                         error = mapping->a_ops->readpage(in, page);
441                         if (unlikely(error)) {
442                                 /*
443                                  * We really should re-lookup the page here,
444                                  * but it complicates things a lot. Instead
445                                  * lets just do what we already stored, and
446                                  * we'll get it the next time we are called.
447                                  */
448                                 if (error == AOP_TRUNCATED_PAGE)
449                                         error = 0;
450
451                                 break;
452                         }
453                 }
454 fill_it:
455                 /*
456                  * i_size must be checked after PageUptodate.
457                  */
458                 isize = i_size_read(mapping->host);
459                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
460                 if (unlikely(!isize || index > end_index))
461                         break;
462
463                 /*
464                  * if this is the last page, see if we need to shrink
465                  * the length and stop
466                  */
467                 if (end_index == index) {
468                         unsigned int plen;
469
470                         /*
471                          * max good bytes in this page
472                          */
473                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
474                         if (plen <= loff)
475                                 break;
476
477                         /*
478                          * force quit after adding this page
479                          */
480                         this_len = min(this_len, plen - loff);
481                         len = this_len;
482                 }
483
484                 spd.partial[page_nr].offset = loff;
485                 spd.partial[page_nr].len = this_len;
486                 len -= this_len;
487                 loff = 0;
488                 spd.nr_pages++;
489                 index++;
490         }
491
492         /*
493          * Release any pages at the end, if we quit early. 'page_nr' is how far
494          * we got, 'nr_pages' is how many pages are in the map.
495          */
496         while (page_nr < nr_pages)
497                 page_cache_release(spd.pages[page_nr++]);
498         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
499
500         if (spd.nr_pages)
501                 error = splice_to_pipe(pipe, &spd);
502
503         splice_shrink_spd(&spd);
504         return error;
505 }
506
507 /**
508  * generic_file_splice_read - splice data from file to a pipe
509  * @in:         file to splice from
510  * @ppos:       position in @in
511  * @pipe:       pipe to splice to
512  * @len:        number of bytes to splice
513  * @flags:      splice modifier flags
514  *
515  * Description:
516  *    Will read pages from given file and fill them into a pipe. Can be
517  *    used as long as the address_space operations for the source implements
518  *    a readpage() hook.
519  *
520  */
521 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
522                                  struct pipe_inode_info *pipe, size_t len,
523                                  unsigned int flags)
524 {
525         loff_t isize, left;
526         int ret;
527
528         isize = i_size_read(in->f_mapping->host);
529         if (unlikely(*ppos >= isize))
530                 return 0;
531
532         left = isize - *ppos;
533         if (unlikely(left < len))
534                 len = left;
535
536         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
537         if (ret > 0) {
538                 *ppos += ret;
539                 file_accessed(in);
540         }
541
542         return ret;
543 }
544 EXPORT_SYMBOL(generic_file_splice_read);
545
546 static const struct pipe_buf_operations default_pipe_buf_ops = {
547         .can_merge = 0,
548         .map = generic_pipe_buf_map,
549         .unmap = generic_pipe_buf_unmap,
550         .confirm = generic_pipe_buf_confirm,
551         .release = generic_pipe_buf_release,
552         .steal = generic_pipe_buf_steal,
553         .get = generic_pipe_buf_get,
554 };
555
556 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
557                             unsigned long vlen, loff_t offset)
558 {
559         mm_segment_t old_fs;
560         loff_t pos = offset;
561         ssize_t res;
562
563         old_fs = get_fs();
564         set_fs(get_ds());
565         /* The cast to a user pointer is valid due to the set_fs() */
566         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
567         set_fs(old_fs);
568
569         return res;
570 }
571
572 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
573                             loff_t pos)
574 {
575         mm_segment_t old_fs;
576         ssize_t res;
577
578         old_fs = get_fs();
579         set_fs(get_ds());
580         /* The cast to a user pointer is valid due to the set_fs() */
581         res = vfs_write(file, (__force const char __user *)buf, count, &pos);
582         set_fs(old_fs);
583
584         return res;
585 }
586 EXPORT_SYMBOL(kernel_write);
587
588 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
589                                  struct pipe_inode_info *pipe, size_t len,
590                                  unsigned int flags)
591 {
592         unsigned int nr_pages;
593         unsigned int nr_freed;
594         size_t offset;
595         struct page *pages[PIPE_DEF_BUFFERS];
596         struct partial_page partial[PIPE_DEF_BUFFERS];
597         struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
598         ssize_t res;
599         size_t this_len;
600         int error;
601         int i;
602         struct splice_pipe_desc spd = {
603                 .pages = pages,
604                 .partial = partial,
605                 .nr_pages_max = PIPE_DEF_BUFFERS,
606                 .flags = flags,
607                 .ops = &default_pipe_buf_ops,
608                 .spd_release = spd_release_page,
609         };
610
611         if (splice_grow_spd(pipe, &spd))
612                 return -ENOMEM;
613
614         res = -ENOMEM;
615         vec = __vec;
616         if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
617                 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
618                 if (!vec)
619                         goto shrink_ret;
620         }
621
622         offset = *ppos & ~PAGE_CACHE_MASK;
623         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
624
625         for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
626                 struct page *page;
627
628                 page = alloc_page(GFP_USER);
629                 error = -ENOMEM;
630                 if (!page)
631                         goto err;
632
633                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
634                 vec[i].iov_base = (void __user *) page_address(page);
635                 vec[i].iov_len = this_len;
636                 spd.pages[i] = page;
637                 spd.nr_pages++;
638                 len -= this_len;
639                 offset = 0;
640         }
641
642         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
643         if (res < 0) {
644                 error = res;
645                 goto err;
646         }
647
648         error = 0;
649         if (!res)
650                 goto err;
651
652         nr_freed = 0;
653         for (i = 0; i < spd.nr_pages; i++) {
654                 this_len = min_t(size_t, vec[i].iov_len, res);
655                 spd.partial[i].offset = 0;
656                 spd.partial[i].len = this_len;
657                 if (!this_len) {
658                         __free_page(spd.pages[i]);
659                         spd.pages[i] = NULL;
660                         nr_freed++;
661                 }
662                 res -= this_len;
663         }
664         spd.nr_pages -= nr_freed;
665
666         res = splice_to_pipe(pipe, &spd);
667         if (res > 0)
668                 *ppos += res;
669
670 shrink_ret:
671         if (vec != __vec)
672                 kfree(vec);
673         splice_shrink_spd(&spd);
674         return res;
675
676 err:
677         for (i = 0; i < spd.nr_pages; i++)
678                 __free_page(spd.pages[i]);
679
680         res = error;
681         goto shrink_ret;
682 }
683 EXPORT_SYMBOL(default_file_splice_read);
684
685 /*
686  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
687  * using sendpage(). Return the number of bytes sent.
688  */
689 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
690                             struct pipe_buffer *buf, struct splice_desc *sd)
691 {
692         struct file *file = sd->u.file;
693         loff_t pos = sd->pos;
694         int more;
695
696         if (!likely(file->f_op && file->f_op->sendpage))
697                 return -EINVAL;
698
699         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
700
701         if (sd->len < sd->total_len && pipe->nrbufs > 1)
702                 more |= MSG_SENDPAGE_NOTLAST;
703
704         return file->f_op->sendpage(file, buf->page, buf->offset,
705                                     sd->len, &pos, more);
706 }
707
708 /*
709  * This is a little more tricky than the file -> pipe splicing. There are
710  * basically three cases:
711  *
712  *      - Destination page already exists in the address space and there
713  *        are users of it. For that case we have no other option that
714  *        copying the data. Tough luck.
715  *      - Destination page already exists in the address space, but there
716  *        are no users of it. Make sure it's uptodate, then drop it. Fall
717  *        through to last case.
718  *      - Destination page does not exist, we can add the pipe page to
719  *        the page cache and avoid the copy.
720  *
721  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
722  * sd->flags), we attempt to migrate pages from the pipe to the output
723  * file address space page cache. This is possible if no one else has
724  * the pipe page referenced outside of the pipe and page cache. If
725  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
726  * a new page in the output file page cache and fill/dirty that.
727  */
728 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
729                  struct splice_desc *sd)
730 {
731         struct file *file = sd->u.file;
732         struct address_space *mapping = file->f_mapping;
733         unsigned int offset, this_len;
734         struct page *page;
735         void *fsdata;
736         int ret;
737
738         offset = sd->pos & ~PAGE_CACHE_MASK;
739
740         this_len = sd->len;
741         if (this_len + offset > PAGE_CACHE_SIZE)
742                 this_len = PAGE_CACHE_SIZE - offset;
743
744         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
745                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
746         if (unlikely(ret))
747                 goto out;
748
749         if (buf->page != page) {
750                 char *src = buf->ops->map(pipe, buf, 1);
751                 char *dst = kmap_atomic(page);
752
753                 memcpy(dst + offset, src + buf->offset, this_len);
754                 flush_dcache_page(page);
755                 kunmap_atomic(dst);
756                 buf->ops->unmap(pipe, buf, src);
757         }
758         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
759                                 page, fsdata);
760 out:
761         return ret;
762 }
763 EXPORT_SYMBOL(pipe_to_file);
764
765 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
766 {
767         smp_mb();
768         if (waitqueue_active(&pipe->wait))
769                 wake_up_interruptible(&pipe->wait);
770         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
771 }
772
773 /**
774  * splice_from_pipe_feed - feed available data from a pipe to a file
775  * @pipe:       pipe to splice from
776  * @sd:         information to @actor
777  * @actor:      handler that splices the data
778  *
779  * Description:
780  *    This function loops over the pipe and calls @actor to do the
781  *    actual moving of a single struct pipe_buffer to the desired
782  *    destination.  It returns when there's no more buffers left in
783  *    the pipe or if the requested number of bytes (@sd->total_len)
784  *    have been copied.  It returns a positive number (one) if the
785  *    pipe needs to be filled with more data, zero if the required
786  *    number of bytes have been copied and -errno on error.
787  *
788  *    This, together with splice_from_pipe_{begin,end,next}, may be
789  *    used to implement the functionality of __splice_from_pipe() when
790  *    locking is required around copying the pipe buffers to the
791  *    destination.
792  */
793 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
794                           splice_actor *actor)
795 {
796         int ret;
797
798         while (pipe->nrbufs) {
799                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
800                 const struct pipe_buf_operations *ops = buf->ops;
801
802                 sd->len = buf->len;
803                 if (sd->len > sd->total_len)
804                         sd->len = sd->total_len;
805
806                 ret = buf->ops->confirm(pipe, buf);
807                 if (unlikely(ret)) {
808                         if (ret == -ENODATA)
809                                 ret = 0;
810                         return ret;
811                 }
812
813                 ret = actor(pipe, buf, sd);
814                 if (ret <= 0)
815                         return ret;
816
817                 buf->offset += ret;
818                 buf->len -= ret;
819
820                 sd->num_spliced += ret;
821                 sd->len -= ret;
822                 sd->pos += ret;
823                 sd->total_len -= ret;
824
825                 if (!buf->len) {
826                         buf->ops = NULL;
827                         ops->release(pipe, buf);
828                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
829                         pipe->nrbufs--;
830                         if (pipe->inode)
831                                 sd->need_wakeup = true;
832                 }
833
834                 if (!sd->total_len)
835                         return 0;
836         }
837
838         return 1;
839 }
840 EXPORT_SYMBOL(splice_from_pipe_feed);
841
842 /**
843  * splice_from_pipe_next - wait for some data to splice from
844  * @pipe:       pipe to splice from
845  * @sd:         information about the splice operation
846  *
847  * Description:
848  *    This function will wait for some data and return a positive
849  *    value (one) if pipe buffers are available.  It will return zero
850  *    or -errno if no more data needs to be spliced.
851  */
852 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
853 {
854         while (!pipe->nrbufs) {
855                 if (!pipe->writers)
856                         return 0;
857
858                 if (!pipe->waiting_writers && sd->num_spliced)
859                         return 0;
860
861                 if (sd->flags & SPLICE_F_NONBLOCK)
862                         return -EAGAIN;
863
864                 if (signal_pending(current))
865                         return -ERESTARTSYS;
866
867                 if (sd->need_wakeup) {
868                         wakeup_pipe_writers(pipe);
869                         sd->need_wakeup = false;
870                 }
871
872                 pipe_wait(pipe);
873         }
874
875         return 1;
876 }
877 EXPORT_SYMBOL(splice_from_pipe_next);
878
879 /**
880  * splice_from_pipe_begin - start splicing from pipe
881  * @sd:         information about the splice operation
882  *
883  * Description:
884  *    This function should be called before a loop containing
885  *    splice_from_pipe_next() and splice_from_pipe_feed() to
886  *    initialize the necessary fields of @sd.
887  */
888 void splice_from_pipe_begin(struct splice_desc *sd)
889 {
890         sd->num_spliced = 0;
891         sd->need_wakeup = false;
892 }
893 EXPORT_SYMBOL(splice_from_pipe_begin);
894
895 /**
896  * splice_from_pipe_end - finish splicing from pipe
897  * @pipe:       pipe to splice from
898  * @sd:         information about the splice operation
899  *
900  * Description:
901  *    This function will wake up pipe writers if necessary.  It should
902  *    be called after a loop containing splice_from_pipe_next() and
903  *    splice_from_pipe_feed().
904  */
905 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
906 {
907         if (sd->need_wakeup)
908                 wakeup_pipe_writers(pipe);
909 }
910 EXPORT_SYMBOL(splice_from_pipe_end);
911
912 /**
913  * __splice_from_pipe - splice data from a pipe to given actor
914  * @pipe:       pipe to splice from
915  * @sd:         information to @actor
916  * @actor:      handler that splices the data
917  *
918  * Description:
919  *    This function does little more than loop over the pipe and call
920  *    @actor to do the actual moving of a single struct pipe_buffer to
921  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
922  *    pipe_to_user.
923  *
924  */
925 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
926                            splice_actor *actor)
927 {
928         int ret;
929
930         splice_from_pipe_begin(sd);
931         do {
932                 ret = splice_from_pipe_next(pipe, sd);
933                 if (ret > 0)
934                         ret = splice_from_pipe_feed(pipe, sd, actor);
935         } while (ret > 0);
936         splice_from_pipe_end(pipe, sd);
937
938         return sd->num_spliced ? sd->num_spliced : ret;
939 }
940 EXPORT_SYMBOL(__splice_from_pipe);
941
942 /**
943  * splice_from_pipe - splice data from a pipe to a file
944  * @pipe:       pipe to splice from
945  * @out:        file to splice to
946  * @ppos:       position in @out
947  * @len:        how many bytes to splice
948  * @flags:      splice modifier flags
949  * @actor:      handler that splices the data
950  *
951  * Description:
952  *    See __splice_from_pipe. This function locks the pipe inode,
953  *    otherwise it's identical to __splice_from_pipe().
954  *
955  */
956 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
957                          loff_t *ppos, size_t len, unsigned int flags,
958                          splice_actor *actor)
959 {
960         ssize_t ret;
961         struct splice_desc sd = {
962                 .total_len = len,
963                 .flags = flags,
964                 .pos = *ppos,
965                 .u.file = out,
966         };
967
968         pipe_lock(pipe);
969         ret = __splice_from_pipe(pipe, &sd, actor);
970         pipe_unlock(pipe);
971
972         return ret;
973 }
974
975 /**
976  * generic_file_splice_write - splice data from a pipe to a file
977  * @pipe:       pipe info
978  * @out:        file to write to
979  * @ppos:       position in @out
980  * @len:        number of bytes to splice
981  * @flags:      splice modifier flags
982  *
983  * Description:
984  *    Will either move or copy pages (determined by @flags options) from
985  *    the given pipe inode to the given file.
986  *
987  */
988 ssize_t
989 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
990                           loff_t *ppos, size_t len, unsigned int flags)
991 {
992         struct address_space *mapping = out->f_mapping;
993         struct inode *inode = mapping->host;
994         struct splice_desc sd = {
995                 .total_len = len,
996                 .flags = flags,
997                 .pos = *ppos,
998                 .u.file = out,
999         };
1000         ssize_t ret;
1001
1002         sb_start_write(inode->i_sb);
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         sb_end_write(inode->i_sb);
1040
1041         return ret;
1042 }
1043
1044 EXPORT_SYMBOL(generic_file_splice_write);
1045
1046 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1047                           struct splice_desc *sd)
1048 {
1049         int ret;
1050         void *data;
1051
1052         data = buf->ops->map(pipe, buf, 0);
1053         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1054         buf->ops->unmap(pipe, buf, data);
1055
1056         return ret;
1057 }
1058
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         return splice_write(pipe, out, ppos, len, flags);
1119 }
1120
1121 /*
1122  * Attempt to initiate a splice from a file to a pipe.
1123  */
1124 static long do_splice_to(struct file *in, loff_t *ppos,
1125                          struct pipe_inode_info *pipe, size_t len,
1126                          unsigned int flags)
1127 {
1128         ssize_t (*splice_read)(struct file *, loff_t *,
1129                                struct pipe_inode_info *, size_t, unsigned int);
1130         int ret;
1131
1132         if (unlikely(!(in->f_mode & FMODE_READ)))
1133                 return -EBADF;
1134
1135         ret = rw_verify_area(READ, in, ppos, len);
1136         if (unlikely(ret < 0))
1137                 return ret;
1138
1139         if (in->f_op && in->f_op->splice_read)
1140                 splice_read = in->f_op->splice_read;
1141         else
1142                 splice_read = default_file_splice_read;
1143
1144         return splice_read(in, ppos, pipe, len, flags);
1145 }
1146
1147 /**
1148  * splice_direct_to_actor - splices data directly between two non-pipes
1149  * @in:         file to splice from
1150  * @sd:         actor information on where to splice to
1151  * @actor:      handles the data splicing
1152  *
1153  * Description:
1154  *    This is a special case helper to splice directly between two
1155  *    points, without requiring an explicit pipe. Internally an allocated
1156  *    pipe is cached in the process, and reused during the lifetime of
1157  *    that process.
1158  *
1159  */
1160 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1161                                splice_direct_actor *actor)
1162 {
1163         struct pipe_inode_info *pipe;
1164         long ret, bytes;
1165         umode_t i_mode;
1166         size_t len;
1167         int i, flags;
1168
1169         /*
1170          * We require the input being a regular file, as we don't want to
1171          * randomly drop data for eg socket -> socket splicing. Use the
1172          * piped splicing for that!
1173          */
1174         i_mode = file_inode(in)->i_mode;
1175         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1176                 return -EINVAL;
1177
1178         /*
1179          * neither in nor out is a pipe, setup an internal pipe attached to
1180          * 'out' and transfer the wanted data from 'in' to 'out' through that
1181          */
1182         pipe = current->splice_pipe;
1183         if (unlikely(!pipe)) {
1184                 pipe = alloc_pipe_info(NULL);
1185                 if (!pipe)
1186                         return -ENOMEM;
1187
1188                 /*
1189                  * We don't have an immediate reader, but we'll read the stuff
1190                  * out of the pipe right after the splice_to_pipe(). So set
1191                  * PIPE_READERS appropriately.
1192                  */
1193                 pipe->readers = 1;
1194
1195                 current->splice_pipe = pipe;
1196         }
1197
1198         /*
1199          * Do the splice.
1200          */
1201         ret = 0;
1202         bytes = 0;
1203         len = sd->total_len;
1204         flags = sd->flags;
1205
1206         /*
1207          * Don't block on output, we have to drain the direct pipe.
1208          */
1209         sd->flags &= ~SPLICE_F_NONBLOCK;
1210
1211         while (len) {
1212                 size_t read_len;
1213                 loff_t pos = sd->pos, prev_pos = pos;
1214
1215                 ret = do_splice_to(in, &pos, pipe, len, flags);
1216                 if (unlikely(ret <= 0))
1217                         goto out_release;
1218
1219                 read_len = ret;
1220                 sd->total_len = read_len;
1221
1222                 /*
1223                  * NOTE: nonblocking mode only applies to the input. We
1224                  * must not do the output in nonblocking mode as then we
1225                  * could get stuck data in the internal pipe:
1226                  */
1227                 ret = actor(pipe, sd);
1228                 if (unlikely(ret <= 0)) {
1229                         sd->pos = prev_pos;
1230                         goto out_release;
1231                 }
1232
1233                 bytes += ret;
1234                 len -= ret;
1235                 sd->pos = pos;
1236
1237                 if (ret < read_len) {
1238                         sd->pos = prev_pos + ret;
1239                         goto out_release;
1240                 }
1241         }
1242
1243 done:
1244         pipe->nrbufs = pipe->curbuf = 0;
1245         file_accessed(in);
1246         return bytes;
1247
1248 out_release:
1249         /*
1250          * If we did an incomplete transfer we must release
1251          * the pipe buffers in question:
1252          */
1253         for (i = 0; i < pipe->buffers; i++) {
1254                 struct pipe_buffer *buf = pipe->bufs + i;
1255
1256                 if (buf->ops) {
1257                         buf->ops->release(pipe, buf);
1258                         buf->ops = NULL;
1259                 }
1260         }
1261
1262         if (!bytes)
1263                 bytes = ret;
1264
1265         goto done;
1266 }
1267 EXPORT_SYMBOL(splice_direct_to_actor);
1268
1269 static int direct_splice_actor(struct pipe_inode_info *pipe,
1270                                struct splice_desc *sd)
1271 {
1272         struct file *file = sd->u.file;
1273
1274         return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1275                               sd->flags);
1276 }
1277
1278 /**
1279  * do_splice_direct - splices data directly between two files
1280  * @in:         file to splice from
1281  * @ppos:       input file offset
1282  * @out:        file to splice to
1283  * @len:        number of bytes to splice
1284  * @flags:      splice modifier flags
1285  *
1286  * Description:
1287  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1288  *    doing it in the application would incur an extra system call
1289  *    (splice in + splice out, as compared to just sendfile()). So this helper
1290  *    can splice directly through a process-private pipe.
1291  *
1292  */
1293 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1294                       size_t len, unsigned int flags)
1295 {
1296         struct splice_desc sd = {
1297                 .len            = len,
1298                 .total_len      = len,
1299                 .flags          = flags,
1300                 .pos            = *ppos,
1301                 .u.file         = out,
1302         };
1303         long ret;
1304
1305         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1306         if (ret > 0)
1307                 *ppos = sd.pos;
1308
1309         return ret;
1310 }
1311
1312 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1313                                struct pipe_inode_info *opipe,
1314                                size_t len, unsigned int flags);
1315
1316 /*
1317  * Determine where to splice to/from.
1318  */
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 = get_pipe_info(in);
1329         opipe = get_pipe_info(out);
1330
1331         if (ipipe && opipe) {
1332                 if (off_in || off_out)
1333                         return -ESPIPE;
1334
1335                 if (!(in->f_mode & FMODE_READ))
1336                         return -EBADF;
1337
1338                 if (!(out->f_mode & FMODE_WRITE))
1339                         return -EBADF;
1340
1341                 /* Splicing to self would be fun, but... */
1342                 if (ipipe == opipe)
1343                         return -EINVAL;
1344
1345                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1346         }
1347
1348         if (ipipe) {
1349                 if (off_in)
1350                         return -ESPIPE;
1351                 if (off_out) {
1352                         if (!(out->f_mode & FMODE_PWRITE))
1353                                 return -EINVAL;
1354                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1355                                 return -EFAULT;
1356                         off = &offset;
1357                 } else
1358                         off = &out->f_pos;
1359
1360                 ret = do_splice_from(ipipe, out, off, len, flags);
1361
1362                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1363                         ret = -EFAULT;
1364
1365                 return ret;
1366         }
1367
1368         if (opipe) {
1369                 if (off_out)
1370                         return -ESPIPE;
1371                 if (off_in) {
1372                         if (!(in->f_mode & FMODE_PREAD))
1373                                 return -EINVAL;
1374                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1375                                 return -EFAULT;
1376                         off = &offset;
1377                 } else
1378                         off = &in->f_pos;
1379
1380                 ret = do_splice_to(in, off, opipe, len, flags);
1381
1382                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1383                         ret = -EFAULT;
1384
1385                 return ret;
1386         }
1387
1388         return -EINVAL;
1389 }
1390
1391 /*
1392  * Map an iov into an array of pages and offset/length tupples. With the
1393  * partial_page structure, we can map several non-contiguous ranges into
1394  * our ones pages[] map instead of splitting that operation into pieces.
1395  * Could easily be exported as a generic helper for other users, in which
1396  * case one would probably want to add a 'max_nr_pages' parameter as well.
1397  */
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, bool aligned,
1401                                 unsigned int pipe_buffers)
1402 {
1403         int buffers = 0, error = 0;
1404
1405         while (nr_vecs) {
1406                 unsigned long off, npages;
1407                 struct iovec entry;
1408                 void __user *base;
1409                 size_t len;
1410                 int i;
1411
1412                 error = -EFAULT;
1413                 if (copy_from_user(&entry, iov, sizeof(entry)))
1414                         break;
1415
1416                 base = entry.iov_base;
1417                 len = entry.iov_len;
1418
1419                 /*
1420                  * Sanity check this iovec. 0 read succeeds.
1421                  */
1422                 error = 0;
1423                 if (unlikely(!len))
1424                         break;
1425                 error = -EFAULT;
1426                 if (!access_ok(VERIFY_READ, base, len))
1427                         break;
1428
1429                 /*
1430                  * Get this base offset and number of pages, then map
1431                  * in the user pages.
1432                  */
1433                 off = (unsigned long) base & ~PAGE_MASK;
1434
1435                 /*
1436                  * If asked for alignment, the offset must be zero and the
1437                  * length a multiple of the PAGE_SIZE.
1438                  */
1439                 error = -EINVAL;
1440                 if (aligned && (off || len & ~PAGE_MASK))
1441                         break;
1442
1443                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1444                 if (npages > pipe_buffers - buffers)
1445                         npages = pipe_buffers - buffers;
1446
1447                 error = get_user_pages_fast((unsigned long)base, npages,
1448                                         0, &pages[buffers]);
1449
1450                 if (unlikely(error <= 0))
1451                         break;
1452
1453                 /*
1454                  * Fill this contiguous range into the partial page map.
1455                  */
1456                 for (i = 0; i < error; i++) {
1457                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1458
1459                         partial[buffers].offset = off;
1460                         partial[buffers].len = plen;
1461
1462                         off = 0;
1463                         len -= plen;
1464                         buffers++;
1465                 }
1466
1467                 /*
1468                  * We didn't complete this iov, stop here since it probably
1469                  * means we have to move some of this into a pipe to
1470                  * be able to continue.
1471                  */
1472                 if (len)
1473                         break;
1474
1475                 /*
1476                  * Don't continue if we mapped fewer pages than we asked for,
1477                  * or if we mapped the max number of pages that we have
1478                  * room for.
1479                  */
1480                 if (error < npages || buffers == pipe_buffers)
1481                         break;
1482
1483                 nr_vecs--;
1484                 iov++;
1485         }
1486
1487         if (buffers)
1488                 return buffers;
1489
1490         return error;
1491 }
1492
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 = get_pipe_info(file);
1544         if (!pipe)
1545                 return -EBADF;
1546
1547         pipe_lock(pipe);
1548
1549         error = ret = 0;
1550         while (nr_segs) {
1551                 void __user *base;
1552                 size_t len;
1553
1554                 /*
1555                  * Get user address base and length for this iovec.
1556                  */
1557                 error = get_user(base, &iov->iov_base);
1558                 if (unlikely(error))
1559                         break;
1560                 error = get_user(len, &iov->iov_len);
1561                 if (unlikely(error))
1562                         break;
1563
1564                 /*
1565                  * Sanity check this iovec. 0 read succeeds.
1566                  */
1567                 if (unlikely(!len))
1568                         break;
1569                 if (unlikely(!base)) {
1570                         error = -EFAULT;
1571                         break;
1572                 }
1573
1574                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1575                         error = -EFAULT;
1576                         break;
1577                 }
1578
1579                 sd.len = 0;
1580                 sd.total_len = len;
1581                 sd.flags = flags;
1582                 sd.u.userptr = base;
1583                 sd.pos = 0;
1584
1585                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1586                 if (size < 0) {
1587                         if (!ret)
1588                                 ret = size;
1589
1590                         break;
1591                 }
1592
1593                 ret += size;
1594
1595                 if (size < len)
1596                         break;
1597
1598                 nr_segs--;
1599                 iov++;
1600         }
1601
1602         pipe_unlock(pipe);
1603
1604         if (!ret)
1605                 ret = error;
1606
1607         return ret;
1608 }
1609
1610 /*
1611  * vmsplice splices a user address range into a pipe. It can be thought of
1612  * as splice-from-memory, where the regular splice is splice-from-file (or
1613  * to file). In both cases the output is a pipe, naturally.
1614  */
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                 .nr_pages_max = PIPE_DEF_BUFFERS,
1625                 .flags = flags,
1626                 .ops = &user_page_pipe_buf_ops,
1627                 .spd_release = spd_release_page,
1628         };
1629         long ret;
1630
1631         pipe = get_pipe_info(file);
1632         if (!pipe)
1633                 return -EBADF;
1634
1635         if (splice_grow_spd(pipe, &spd))
1636                 return -ENOMEM;
1637
1638         spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1639                                             spd.partial, false,
1640                                             spd.nr_pages_max);
1641         if (spd.nr_pages <= 0)
1642                 ret = spd.nr_pages;
1643         else
1644                 ret = splice_to_pipe(pipe, &spd);
1645
1646         splice_shrink_spd(&spd);
1647         return ret;
1648 }
1649
1650 /*
1651  * Note that vmsplice only really supports true splicing _from_ user memory
1652  * to a pipe, not the other way around. Splicing from user memory is a simple
1653  * operation that can be supported without any funky alignment restrictions
1654  * or nasty vm tricks. We simply map in the user memory and fill them into
1655  * a pipe. The reverse isn't quite as easy, though. There are two possible
1656  * solutions for that:
1657  *
1658  *      - memcpy() the data internally, at which point we might as well just
1659  *        do a regular read() on the buffer anyway.
1660  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1661  *        has restriction limitations on both ends of the pipe).
1662  *
1663  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1664  *
1665  */
1666 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1667                 unsigned long, nr_segs, unsigned int, flags)
1668 {
1669         struct fd f;
1670         long error;
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         f = fdget(fd);
1679         if (f.file) {
1680                 if (f.file->f_mode & FMODE_WRITE)
1681                         error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1682                 else if (f.file->f_mode & FMODE_READ)
1683                         error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1684
1685                 fdput(f);
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         struct fd in, out;
1696         long error;
1697
1698         if (unlikely(!len))
1699                 return 0;
1700
1701         error = -EBADF;
1702         in = fdget(fd_in);
1703         if (in.file) {
1704                 if (in.file->f_mode & FMODE_READ) {
1705                         out = fdget(fd_out);
1706                         if (out.file) {
1707                                 if (out.file->f_mode & FMODE_WRITE)
1708                                         error = do_splice(in.file, off_in,
1709                                                           out.file, off_out,
1710                                                           len, flags);
1711                                 fdput(out);
1712                         }
1713                 }
1714                 fdput(in);
1715         }
1716         return error;
1717 }
1718
1719 /*
1720  * Make sure there's data to read. Wait for input if we can, otherwise
1721  * return an appropriate error.
1722  */
1723 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1724 {
1725         int ret;
1726
1727         /*
1728          * Check ->nrbufs without the inode lock first. This function
1729          * is speculative anyways, so missing one is ok.
1730          */
1731         if (pipe->nrbufs)
1732                 return 0;
1733
1734         ret = 0;
1735         pipe_lock(pipe);
1736
1737         while (!pipe->nrbufs) {
1738                 if (signal_pending(current)) {
1739                         ret = -ERESTARTSYS;
1740                         break;
1741                 }
1742                 if (!pipe->writers)
1743                         break;
1744                 if (!pipe->waiting_writers) {
1745                         if (flags & SPLICE_F_NONBLOCK) {
1746                                 ret = -EAGAIN;
1747                                 break;
1748                         }
1749                 }
1750                 pipe_wait(pipe);
1751         }
1752
1753         pipe_unlock(pipe);
1754         return ret;
1755 }
1756
1757 /*
1758  * Make sure there's writeable room. Wait for room if we can, otherwise
1759  * return an appropriate error.
1760  */
1761 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1762 {
1763         int ret;
1764
1765         /*
1766          * Check ->nrbufs without the inode lock first. This function
1767          * is speculative anyways, so missing one is ok.
1768          */
1769         if (pipe->nrbufs < pipe->buffers)
1770                 return 0;
1771
1772         ret = 0;
1773         pipe_lock(pipe);
1774
1775         while (pipe->nrbufs >= pipe->buffers) {
1776                 if (!pipe->readers) {
1777                         send_sig(SIGPIPE, current, 0);
1778                         ret = -EPIPE;
1779                         break;
1780                 }
1781                 if (flags & SPLICE_F_NONBLOCK) {
1782                         ret = -EAGAIN;
1783                         break;
1784                 }
1785                 if (signal_pending(current)) {
1786                         ret = -ERESTARTSYS;
1787                         break;
1788                 }
1789                 pipe->waiting_writers++;
1790                 pipe_wait(pipe);
1791                 pipe->waiting_writers--;
1792         }
1793
1794         pipe_unlock(pipe);
1795         return ret;
1796 }
1797
1798 /*
1799  * Splice contents of ipipe to opipe.
1800  */
1801 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1802                                struct pipe_inode_info *opipe,
1803                                size_t len, unsigned int flags)
1804 {
1805         struct pipe_buffer *ibuf, *obuf;
1806         int ret = 0, nbuf;
1807         bool input_wakeup = false;
1808
1809
1810 retry:
1811         ret = ipipe_prep(ipipe, flags);
1812         if (ret)
1813                 return ret;
1814
1815         ret = opipe_prep(opipe, flags);
1816         if (ret)
1817                 return ret;
1818
1819         /*
1820          * Potential ABBA deadlock, work around it by ordering lock
1821          * grabbing by pipe info address. Otherwise two different processes
1822          * could deadlock (one doing tee from A -> B, the other from B -> A).
1823          */
1824         pipe_double_lock(ipipe, opipe);
1825
1826         do {
1827                 if (!opipe->readers) {
1828                         send_sig(SIGPIPE, current, 0);
1829                         if (!ret)
1830                                 ret = -EPIPE;
1831                         break;
1832                 }
1833
1834                 if (!ipipe->nrbufs && !ipipe->writers)
1835                         break;
1836
1837                 /*
1838                  * Cannot make any progress, because either the input
1839                  * pipe is empty or the output pipe is full.
1840                  */
1841                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1842                         /* Already processed some buffers, break */
1843                         if (ret)
1844                                 break;
1845
1846                         if (flags & SPLICE_F_NONBLOCK) {
1847                                 ret = -EAGAIN;
1848                                 break;
1849                         }
1850
1851                         /*
1852                          * We raced with another reader/writer and haven't
1853                          * managed to process any buffers.  A zero return
1854                          * value means EOF, so retry instead.
1855                          */
1856                         pipe_unlock(ipipe);
1857                         pipe_unlock(opipe);
1858                         goto retry;
1859                 }
1860
1861                 ibuf = ipipe->bufs + ipipe->curbuf;
1862                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1863                 obuf = opipe->bufs + nbuf;
1864
1865                 if (len >= ibuf->len) {
1866                         /*
1867                          * Simply move the whole buffer from ipipe to opipe
1868                          */
1869                         *obuf = *ibuf;
1870                         ibuf->ops = NULL;
1871                         opipe->nrbufs++;
1872                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1873                         ipipe->nrbufs--;
1874                         input_wakeup = true;
1875                 } else {
1876                         /*
1877                          * Get a reference to this pipe buffer,
1878                          * so we can copy the contents over.
1879                          */
1880                         ibuf->ops->get(ipipe, ibuf);
1881                         *obuf = *ibuf;
1882
1883                         /*
1884                          * Don't inherit the gift flag, we need to
1885                          * prevent multiple steals of this page.
1886                          */
1887                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1888
1889                         obuf->len = len;
1890                         opipe->nrbufs++;
1891                         ibuf->offset += obuf->len;
1892                         ibuf->len -= obuf->len;
1893                 }
1894                 ret += obuf->len;
1895                 len -= obuf->len;
1896         } while (len);
1897
1898         pipe_unlock(ipipe);
1899         pipe_unlock(opipe);
1900
1901         /*
1902          * If we put data in the output pipe, wakeup any potential readers.
1903          */
1904         if (ret > 0)
1905                 wakeup_pipe_readers(opipe);
1906
1907         if (input_wakeup)
1908                 wakeup_pipe_writers(ipipe);
1909
1910         return ret;
1911 }
1912
1913 /*
1914  * Link contents of ipipe to opipe.
1915  */
1916 static int link_pipe(struct pipe_inode_info *ipipe,
1917                      struct pipe_inode_info *opipe,
1918                      size_t len, unsigned int flags)
1919 {
1920         struct pipe_buffer *ibuf, *obuf;
1921         int ret = 0, i = 0, nbuf;
1922
1923         /*
1924          * Potential ABBA deadlock, work around it by ordering lock
1925          * grabbing by pipe info address. Otherwise two different processes
1926          * could deadlock (one doing tee from A -> B, the other from B -> A).
1927          */
1928         pipe_double_lock(ipipe, opipe);
1929
1930         do {
1931                 if (!opipe->readers) {
1932                         send_sig(SIGPIPE, current, 0);
1933                         if (!ret)
1934                                 ret = -EPIPE;
1935                         break;
1936                 }
1937
1938                 /*
1939                  * If we have iterated all input buffers or ran out of
1940                  * output room, break.
1941                  */
1942                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1943                         break;
1944
1945                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1946                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1947
1948                 /*
1949                  * Get a reference to this pipe buffer,
1950                  * so we can copy the contents over.
1951                  */
1952                 ibuf->ops->get(ipipe, ibuf);
1953
1954                 obuf = opipe->bufs + nbuf;
1955                 *obuf = *ibuf;
1956
1957                 /*
1958                  * Don't inherit the gift flag, we need to
1959                  * prevent multiple steals of this page.
1960                  */
1961                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1962
1963                 if (obuf->len > len)
1964                         obuf->len = len;
1965
1966                 opipe->nrbufs++;
1967                 ret += obuf->len;
1968                 len -= obuf->len;
1969                 i++;
1970         } while (len);
1971
1972         /*
1973          * return EAGAIN if we have the potential of some data in the
1974          * future, otherwise just return 0
1975          */
1976         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1977                 ret = -EAGAIN;
1978
1979         pipe_unlock(ipipe);
1980         pipe_unlock(opipe);
1981
1982         /*
1983          * If we put data in the output pipe, wakeup any potential readers.
1984          */
1985         if (ret > 0)
1986                 wakeup_pipe_readers(opipe);
1987
1988         return ret;
1989 }
1990
1991 /*
1992  * This is a tee(1) implementation that works on pipes. It doesn't copy
1993  * any data, it simply references the 'in' pages on the 'out' pipe.
1994  * The 'flags' used are the SPLICE_F_* variants, currently the only
1995  * applicable one is SPLICE_F_NONBLOCK.
1996  */
1997 static long do_tee(struct file *in, struct file *out, size_t len,
1998                    unsigned int flags)
1999 {
2000         struct pipe_inode_info *ipipe = get_pipe_info(in);
2001         struct pipe_inode_info *opipe = get_pipe_info(out);
2002         int ret = -EINVAL;
2003
2004         /*
2005          * Duplicate the contents of ipipe to opipe without actually
2006          * copying the data.
2007          */
2008         if (ipipe && opipe && ipipe != opipe) {
2009                 /*
2010                  * Keep going, unless we encounter an error. The ipipe/opipe
2011                  * ordering doesn't really matter.
2012                  */
2013                 ret = ipipe_prep(ipipe, flags);
2014                 if (!ret) {
2015                         ret = opipe_prep(opipe, flags);
2016                         if (!ret)
2017                                 ret = link_pipe(ipipe, opipe, len, flags);
2018                 }
2019         }
2020
2021         return ret;
2022 }
2023
2024 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2025 {
2026         struct fd in;
2027         int error;
2028
2029         if (unlikely(!len))
2030                 return 0;
2031
2032         error = -EBADF;
2033         in = fdget(fdin);
2034         if (in.file) {
2035                 if (in.file->f_mode & FMODE_READ) {
2036                         struct fd out = fdget(fdout);
2037                         if (out.file) {
2038                                 if (out.file->f_mode & FMODE_WRITE)
2039                                         error = do_tee(in.file, out.file,
2040                                                         len, flags);
2041                                 fdput(out);
2042                         }
2043                 }
2044                 fdput(in);
2045         }
2046
2047         return error;
2048 }