Merge tag 'mm-compaction-5.1-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-block.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/bvec.h>
21 #include <linux/fs.h>
22 #include <linux/file.h>
23 #include <linux/pagemap.h>
24 #include <linux/splice.h>
25 #include <linux/memcontrol.h>
26 #include <linux/mm_inline.h>
27 #include <linux/swap.h>
28 #include <linux/writeback.h>
29 #include <linux/export.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
33 #include <linux/gfp.h>
34 #include <linux/socket.h>
35 #include <linux/compat.h>
36 #include <linux/sched/signal.h>
37
38 #include "internal.h"
39
40 /*
41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
42  * a vm helper function, it's already simplified quite a bit by the
43  * addition of remove_mapping(). If success is returned, the caller may
44  * attempt to reuse this page for another destination.
45  */
46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
47                                      struct pipe_buffer *buf)
48 {
49         struct page *page = buf->page;
50         struct address_space *mapping;
51
52         lock_page(page);
53
54         mapping = page_mapping(page);
55         if (mapping) {
56                 WARN_ON(!PageUptodate(page));
57
58                 /*
59                  * At least for ext2 with nobh option, we need to wait on
60                  * writeback completing on this page, since we'll remove it
61                  * from the pagecache.  Otherwise truncate wont wait on the
62                  * page, allowing the disk blocks to be reused by someone else
63                  * before we actually wrote our data to them. fs corruption
64                  * ensues.
65                  */
66                 wait_on_page_writeback(page);
67
68                 if (page_has_private(page) &&
69                     !try_to_release_page(page, GFP_KERNEL))
70                         goto out_unlock;
71
72                 /*
73                  * If we succeeded in removing the mapping, set LRU flag
74                  * and return good.
75                  */
76                 if (remove_mapping(mapping, page)) {
77                         buf->flags |= PIPE_BUF_FLAG_LRU;
78                         return 0;
79                 }
80         }
81
82         /*
83          * Raced with truncate or failed to remove page from current
84          * address space, unlock and return failure.
85          */
86 out_unlock:
87         unlock_page(page);
88         return 1;
89 }
90
91 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
92                                         struct pipe_buffer *buf)
93 {
94         put_page(buf->page);
95         buf->flags &= ~PIPE_BUF_FLAG_LRU;
96 }
97
98 /*
99  * Check whether the contents of buf is OK to access. Since the content
100  * is a page cache page, IO may be in flight.
101  */
102 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
103                                        struct pipe_buffer *buf)
104 {
105         struct page *page = buf->page;
106         int err;
107
108         if (!PageUptodate(page)) {
109                 lock_page(page);
110
111                 /*
112                  * Page got truncated/unhashed. This will cause a 0-byte
113                  * splice, if this is the first page.
114                  */
115                 if (!page->mapping) {
116                         err = -ENODATA;
117                         goto error;
118                 }
119
120                 /*
121                  * Uh oh, read-error from disk.
122                  */
123                 if (!PageUptodate(page)) {
124                         err = -EIO;
125                         goto error;
126                 }
127
128                 /*
129                  * Page is ok afterall, we are done.
130                  */
131                 unlock_page(page);
132         }
133
134         return 0;
135 error:
136         unlock_page(page);
137         return err;
138 }
139
140 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
141         .confirm = page_cache_pipe_buf_confirm,
142         .release = page_cache_pipe_buf_release,
143         .steal = page_cache_pipe_buf_steal,
144         .get = generic_pipe_buf_get,
145 };
146
147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
148                                     struct pipe_buffer *buf)
149 {
150         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
151                 return 1;
152
153         buf->flags |= PIPE_BUF_FLAG_LRU;
154         return generic_pipe_buf_steal(pipe, buf);
155 }
156
157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158         .confirm = generic_pipe_buf_confirm,
159         .release = page_cache_pipe_buf_release,
160         .steal = user_page_pipe_buf_steal,
161         .get = generic_pipe_buf_get,
162 };
163
164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
165 {
166         smp_mb();
167         if (waitqueue_active(&pipe->wait))
168                 wake_up_interruptible(&pipe->wait);
169         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
170 }
171
172 /**
173  * splice_to_pipe - fill passed data into a pipe
174  * @pipe:       pipe to fill
175  * @spd:        data to fill
176  *
177  * Description:
178  *    @spd contains a map of pages and len/offset tuples, along with
179  *    the struct pipe_buf_operations associated with these pages. This
180  *    function will link that data to the pipe.
181  *
182  */
183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
184                        struct splice_pipe_desc *spd)
185 {
186         unsigned int spd_pages = spd->nr_pages;
187         int ret = 0, page_nr = 0;
188
189         if (!spd_pages)
190                 return 0;
191
192         if (unlikely(!pipe->readers)) {
193                 send_sig(SIGPIPE, current, 0);
194                 ret = -EPIPE;
195                 goto out;
196         }
197
198         while (pipe->nrbufs < pipe->buffers) {
199                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
200                 struct pipe_buffer *buf = pipe->bufs + newbuf;
201
202                 buf->page = spd->pages[page_nr];
203                 buf->offset = spd->partial[page_nr].offset;
204                 buf->len = spd->partial[page_nr].len;
205                 buf->private = spd->partial[page_nr].private;
206                 buf->ops = spd->ops;
207                 buf->flags = 0;
208
209                 pipe->nrbufs++;
210                 page_nr++;
211                 ret += buf->len;
212
213                 if (!--spd->nr_pages)
214                         break;
215         }
216
217         if (!ret)
218                 ret = -EAGAIN;
219
220 out:
221         while (page_nr < spd_pages)
222                 spd->spd_release(spd, page_nr++);
223
224         return ret;
225 }
226 EXPORT_SYMBOL_GPL(splice_to_pipe);
227
228 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
229 {
230         int ret;
231
232         if (unlikely(!pipe->readers)) {
233                 send_sig(SIGPIPE, current, 0);
234                 ret = -EPIPE;
235         } else if (pipe->nrbufs == pipe->buffers) {
236                 ret = -EAGAIN;
237         } else {
238                 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
239                 pipe->bufs[newbuf] = *buf;
240                 pipe->nrbufs++;
241                 return buf->len;
242         }
243         pipe_buf_release(pipe, buf);
244         return ret;
245 }
246 EXPORT_SYMBOL(add_to_pipe);
247
248 /*
249  * Check if we need to grow the arrays holding pages and partial page
250  * descriptions.
251  */
252 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
253 {
254         unsigned int buffers = READ_ONCE(pipe->buffers);
255
256         spd->nr_pages_max = buffers;
257         if (buffers <= PIPE_DEF_BUFFERS)
258                 return 0;
259
260         spd->pages = kmalloc_array(buffers, sizeof(struct page *), GFP_KERNEL);
261         spd->partial = kmalloc_array(buffers, sizeof(struct partial_page),
262                                      GFP_KERNEL);
263
264         if (spd->pages && spd->partial)
265                 return 0;
266
267         kfree(spd->pages);
268         kfree(spd->partial);
269         return -ENOMEM;
270 }
271
272 void splice_shrink_spd(struct splice_pipe_desc *spd)
273 {
274         if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
275                 return;
276
277         kfree(spd->pages);
278         kfree(spd->partial);
279 }
280
281 /**
282  * generic_file_splice_read - splice data from file to a pipe
283  * @in:         file to splice from
284  * @ppos:       position in @in
285  * @pipe:       pipe to splice to
286  * @len:        number of bytes to splice
287  * @flags:      splice modifier flags
288  *
289  * Description:
290  *    Will read pages from given file and fill them into a pipe. Can be
291  *    used as long as it has more or less sane ->read_iter().
292  *
293  */
294 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
295                                  struct pipe_inode_info *pipe, size_t len,
296                                  unsigned int flags)
297 {
298         struct iov_iter to;
299         struct kiocb kiocb;
300         int idx, ret;
301
302         iov_iter_pipe(&to, READ, pipe, len);
303         idx = to.idx;
304         init_sync_kiocb(&kiocb, in);
305         kiocb.ki_pos = *ppos;
306         ret = call_read_iter(in, &kiocb, &to);
307         if (ret > 0) {
308                 *ppos = kiocb.ki_pos;
309                 file_accessed(in);
310         } else if (ret < 0) {
311                 to.idx = idx;
312                 to.iov_offset = 0;
313                 iov_iter_advance(&to, 0); /* to free what was emitted */
314                 /*
315                  * callers of ->splice_read() expect -EAGAIN on
316                  * "can't put anything in there", rather than -EFAULT.
317                  */
318                 if (ret == -EFAULT)
319                         ret = -EAGAIN;
320         }
321
322         return ret;
323 }
324 EXPORT_SYMBOL(generic_file_splice_read);
325
326 const struct pipe_buf_operations default_pipe_buf_ops = {
327         .confirm = generic_pipe_buf_confirm,
328         .release = generic_pipe_buf_release,
329         .steal = generic_pipe_buf_steal,
330         .get = generic_pipe_buf_get,
331 };
332
333 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
334                                     struct pipe_buffer *buf)
335 {
336         return 1;
337 }
338
339 /* Pipe buffer operations for a socket and similar. */
340 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
341         .confirm = generic_pipe_buf_confirm,
342         .release = generic_pipe_buf_release,
343         .steal = generic_pipe_buf_nosteal,
344         .get = generic_pipe_buf_get,
345 };
346 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
347
348 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
349                             unsigned long vlen, loff_t offset)
350 {
351         mm_segment_t old_fs;
352         loff_t pos = offset;
353         ssize_t res;
354
355         old_fs = get_fs();
356         set_fs(KERNEL_DS);
357         /* The cast to a user pointer is valid due to the set_fs() */
358         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
359         set_fs(old_fs);
360
361         return res;
362 }
363
364 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
365                                  struct pipe_inode_info *pipe, size_t len,
366                                  unsigned int flags)
367 {
368         struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
369         struct iov_iter to;
370         struct page **pages;
371         unsigned int nr_pages;
372         size_t offset, base, copied = 0;
373         ssize_t res;
374         int i;
375
376         if (pipe->nrbufs == pipe->buffers)
377                 return -EAGAIN;
378
379         /*
380          * Try to keep page boundaries matching to source pagecache ones -
381          * it probably won't be much help, but...
382          */
383         offset = *ppos & ~PAGE_MASK;
384
385         iov_iter_pipe(&to, READ, pipe, len + offset);
386
387         res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
388         if (res <= 0)
389                 return -ENOMEM;
390
391         nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
392
393         vec = __vec;
394         if (nr_pages > PIPE_DEF_BUFFERS) {
395                 vec = kmalloc_array(nr_pages, sizeof(struct kvec), GFP_KERNEL);
396                 if (unlikely(!vec)) {
397                         res = -ENOMEM;
398                         goto out;
399                 }
400         }
401
402         pipe->bufs[to.idx].offset = offset;
403         pipe->bufs[to.idx].len -= offset;
404
405         for (i = 0; i < nr_pages; i++) {
406                 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
407                 vec[i].iov_base = page_address(pages[i]) + offset;
408                 vec[i].iov_len = this_len;
409                 len -= this_len;
410                 offset = 0;
411         }
412
413         res = kernel_readv(in, vec, nr_pages, *ppos);
414         if (res > 0) {
415                 copied = res;
416                 *ppos += res;
417         }
418
419         if (vec != __vec)
420                 kfree(vec);
421 out:
422         for (i = 0; i < nr_pages; i++)
423                 put_page(pages[i]);
424         kvfree(pages);
425         iov_iter_advance(&to, copied);  /* truncates and discards */
426         return res;
427 }
428
429 /*
430  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
431  * using sendpage(). Return the number of bytes sent.
432  */
433 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
434                             struct pipe_buffer *buf, struct splice_desc *sd)
435 {
436         struct file *file = sd->u.file;
437         loff_t pos = sd->pos;
438         int more;
439
440         if (!likely(file->f_op->sendpage))
441                 return -EINVAL;
442
443         more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
444
445         if (sd->len < sd->total_len && pipe->nrbufs > 1)
446                 more |= MSG_SENDPAGE_NOTLAST;
447
448         return file->f_op->sendpage(file, buf->page, buf->offset,
449                                     sd->len, &pos, more);
450 }
451
452 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
453 {
454         smp_mb();
455         if (waitqueue_active(&pipe->wait))
456                 wake_up_interruptible(&pipe->wait);
457         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
458 }
459
460 /**
461  * splice_from_pipe_feed - feed available data from a pipe to a file
462  * @pipe:       pipe to splice from
463  * @sd:         information to @actor
464  * @actor:      handler that splices the data
465  *
466  * Description:
467  *    This function loops over the pipe and calls @actor to do the
468  *    actual moving of a single struct pipe_buffer to the desired
469  *    destination.  It returns when there's no more buffers left in
470  *    the pipe or if the requested number of bytes (@sd->total_len)
471  *    have been copied.  It returns a positive number (one) if the
472  *    pipe needs to be filled with more data, zero if the required
473  *    number of bytes have been copied and -errno on error.
474  *
475  *    This, together with splice_from_pipe_{begin,end,next}, may be
476  *    used to implement the functionality of __splice_from_pipe() when
477  *    locking is required around copying the pipe buffers to the
478  *    destination.
479  */
480 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
481                           splice_actor *actor)
482 {
483         int ret;
484
485         while (pipe->nrbufs) {
486                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
487
488                 sd->len = buf->len;
489                 if (sd->len > sd->total_len)
490                         sd->len = sd->total_len;
491
492                 ret = pipe_buf_confirm(pipe, buf);
493                 if (unlikely(ret)) {
494                         if (ret == -ENODATA)
495                                 ret = 0;
496                         return ret;
497                 }
498
499                 ret = actor(pipe, buf, sd);
500                 if (ret <= 0)
501                         return ret;
502
503                 buf->offset += ret;
504                 buf->len -= ret;
505
506                 sd->num_spliced += ret;
507                 sd->len -= ret;
508                 sd->pos += ret;
509                 sd->total_len -= ret;
510
511                 if (!buf->len) {
512                         pipe_buf_release(pipe, buf);
513                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
514                         pipe->nrbufs--;
515                         if (pipe->files)
516                                 sd->need_wakeup = true;
517                 }
518
519                 if (!sd->total_len)
520                         return 0;
521         }
522
523         return 1;
524 }
525
526 /**
527  * splice_from_pipe_next - wait for some data to splice from
528  * @pipe:       pipe to splice from
529  * @sd:         information about the splice operation
530  *
531  * Description:
532  *    This function will wait for some data and return a positive
533  *    value (one) if pipe buffers are available.  It will return zero
534  *    or -errno if no more data needs to be spliced.
535  */
536 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
537 {
538         /*
539          * Check for signal early to make process killable when there are
540          * always buffers available
541          */
542         if (signal_pending(current))
543                 return -ERESTARTSYS;
544
545         while (!pipe->nrbufs) {
546                 if (!pipe->writers)
547                         return 0;
548
549                 if (!pipe->waiting_writers && sd->num_spliced)
550                         return 0;
551
552                 if (sd->flags & SPLICE_F_NONBLOCK)
553                         return -EAGAIN;
554
555                 if (signal_pending(current))
556                         return -ERESTARTSYS;
557
558                 if (sd->need_wakeup) {
559                         wakeup_pipe_writers(pipe);
560                         sd->need_wakeup = false;
561                 }
562
563                 pipe_wait(pipe);
564         }
565
566         return 1;
567 }
568
569 /**
570  * splice_from_pipe_begin - start splicing from pipe
571  * @sd:         information about the splice operation
572  *
573  * Description:
574  *    This function should be called before a loop containing
575  *    splice_from_pipe_next() and splice_from_pipe_feed() to
576  *    initialize the necessary fields of @sd.
577  */
578 static void splice_from_pipe_begin(struct splice_desc *sd)
579 {
580         sd->num_spliced = 0;
581         sd->need_wakeup = false;
582 }
583
584 /**
585  * splice_from_pipe_end - finish splicing from pipe
586  * @pipe:       pipe to splice from
587  * @sd:         information about the splice operation
588  *
589  * Description:
590  *    This function will wake up pipe writers if necessary.  It should
591  *    be called after a loop containing splice_from_pipe_next() and
592  *    splice_from_pipe_feed().
593  */
594 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
595 {
596         if (sd->need_wakeup)
597                 wakeup_pipe_writers(pipe);
598 }
599
600 /**
601  * __splice_from_pipe - splice data from a pipe to given actor
602  * @pipe:       pipe to splice from
603  * @sd:         information to @actor
604  * @actor:      handler that splices the data
605  *
606  * Description:
607  *    This function does little more than loop over the pipe and call
608  *    @actor to do the actual moving of a single struct pipe_buffer to
609  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
610  *    pipe_to_user.
611  *
612  */
613 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
614                            splice_actor *actor)
615 {
616         int ret;
617
618         splice_from_pipe_begin(sd);
619         do {
620                 cond_resched();
621                 ret = splice_from_pipe_next(pipe, sd);
622                 if (ret > 0)
623                         ret = splice_from_pipe_feed(pipe, sd, actor);
624         } while (ret > 0);
625         splice_from_pipe_end(pipe, sd);
626
627         return sd->num_spliced ? sd->num_spliced : ret;
628 }
629 EXPORT_SYMBOL(__splice_from_pipe);
630
631 /**
632  * splice_from_pipe - splice data from a pipe to a file
633  * @pipe:       pipe to splice from
634  * @out:        file to splice to
635  * @ppos:       position in @out
636  * @len:        how many bytes to splice
637  * @flags:      splice modifier flags
638  * @actor:      handler that splices the data
639  *
640  * Description:
641  *    See __splice_from_pipe. This function locks the pipe inode,
642  *    otherwise it's identical to __splice_from_pipe().
643  *
644  */
645 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
646                          loff_t *ppos, size_t len, unsigned int flags,
647                          splice_actor *actor)
648 {
649         ssize_t ret;
650         struct splice_desc sd = {
651                 .total_len = len,
652                 .flags = flags,
653                 .pos = *ppos,
654                 .u.file = out,
655         };
656
657         pipe_lock(pipe);
658         ret = __splice_from_pipe(pipe, &sd, actor);
659         pipe_unlock(pipe);
660
661         return ret;
662 }
663
664 /**
665  * iter_file_splice_write - splice data from a pipe to a file
666  * @pipe:       pipe info
667  * @out:        file to write to
668  * @ppos:       position in @out
669  * @len:        number of bytes to splice
670  * @flags:      splice modifier flags
671  *
672  * Description:
673  *    Will either move or copy pages (determined by @flags options) from
674  *    the given pipe inode to the given file.
675  *    This one is ->write_iter-based.
676  *
677  */
678 ssize_t
679 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
680                           loff_t *ppos, size_t len, unsigned int flags)
681 {
682         struct splice_desc sd = {
683                 .total_len = len,
684                 .flags = flags,
685                 .pos = *ppos,
686                 .u.file = out,
687         };
688         int nbufs = pipe->buffers;
689         struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
690                                         GFP_KERNEL);
691         ssize_t ret;
692
693         if (unlikely(!array))
694                 return -ENOMEM;
695
696         pipe_lock(pipe);
697
698         splice_from_pipe_begin(&sd);
699         while (sd.total_len) {
700                 struct iov_iter from;
701                 size_t left;
702                 int n, idx;
703
704                 ret = splice_from_pipe_next(pipe, &sd);
705                 if (ret <= 0)
706                         break;
707
708                 if (unlikely(nbufs < pipe->buffers)) {
709                         kfree(array);
710                         nbufs = pipe->buffers;
711                         array = kcalloc(nbufs, sizeof(struct bio_vec),
712                                         GFP_KERNEL);
713                         if (!array) {
714                                 ret = -ENOMEM;
715                                 break;
716                         }
717                 }
718
719                 /* build the vector */
720                 left = sd.total_len;
721                 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
722                         struct pipe_buffer *buf = pipe->bufs + idx;
723                         size_t this_len = buf->len;
724
725                         if (this_len > left)
726                                 this_len = left;
727
728                         if (idx == pipe->buffers - 1)
729                                 idx = -1;
730
731                         ret = pipe_buf_confirm(pipe, buf);
732                         if (unlikely(ret)) {
733                                 if (ret == -ENODATA)
734                                         ret = 0;
735                                 goto done;
736                         }
737
738                         array[n].bv_page = buf->page;
739                         array[n].bv_len = this_len;
740                         array[n].bv_offset = buf->offset;
741                         left -= this_len;
742                 }
743
744                 iov_iter_bvec(&from, WRITE, array, n, sd.total_len - left);
745                 ret = vfs_iter_write(out, &from, &sd.pos, 0);
746                 if (ret <= 0)
747                         break;
748
749                 sd.num_spliced += ret;
750                 sd.total_len -= ret;
751                 *ppos = sd.pos;
752
753                 /* dismiss the fully eaten buffers, adjust the partial one */
754                 while (ret) {
755                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
756                         if (ret >= buf->len) {
757                                 ret -= buf->len;
758                                 buf->len = 0;
759                                 pipe_buf_release(pipe, buf);
760                                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
761                                 pipe->nrbufs--;
762                                 if (pipe->files)
763                                         sd.need_wakeup = true;
764                         } else {
765                                 buf->offset += ret;
766                                 buf->len -= ret;
767                                 ret = 0;
768                         }
769                 }
770         }
771 done:
772         kfree(array);
773         splice_from_pipe_end(pipe, &sd);
774
775         pipe_unlock(pipe);
776
777         if (sd.num_spliced)
778                 ret = sd.num_spliced;
779
780         return ret;
781 }
782
783 EXPORT_SYMBOL(iter_file_splice_write);
784
785 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
786                           struct splice_desc *sd)
787 {
788         int ret;
789         void *data;
790         loff_t tmp = sd->pos;
791
792         data = kmap(buf->page);
793         ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
794         kunmap(buf->page);
795
796         return ret;
797 }
798
799 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
800                                          struct file *out, loff_t *ppos,
801                                          size_t len, unsigned int flags)
802 {
803         ssize_t ret;
804
805         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
806         if (ret > 0)
807                 *ppos += ret;
808
809         return ret;
810 }
811
812 /**
813  * generic_splice_sendpage - splice data from a pipe to a socket
814  * @pipe:       pipe to splice from
815  * @out:        socket to write to
816  * @ppos:       position in @out
817  * @len:        number of bytes to splice
818  * @flags:      splice modifier flags
819  *
820  * Description:
821  *    Will send @len bytes from the pipe to a network socket. No data copying
822  *    is involved.
823  *
824  */
825 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
826                                 loff_t *ppos, size_t len, unsigned int flags)
827 {
828         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
829 }
830
831 EXPORT_SYMBOL(generic_splice_sendpage);
832
833 /*
834  * Attempt to initiate a splice from pipe to file.
835  */
836 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
837                            loff_t *ppos, size_t len, unsigned int flags)
838 {
839         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
840                                 loff_t *, size_t, unsigned int);
841
842         if (out->f_op->splice_write)
843                 splice_write = out->f_op->splice_write;
844         else
845                 splice_write = default_file_splice_write;
846
847         return splice_write(pipe, out, ppos, len, flags);
848 }
849
850 /*
851  * Attempt to initiate a splice from a file to a pipe.
852  */
853 static long do_splice_to(struct file *in, loff_t *ppos,
854                          struct pipe_inode_info *pipe, size_t len,
855                          unsigned int flags)
856 {
857         ssize_t (*splice_read)(struct file *, loff_t *,
858                                struct pipe_inode_info *, size_t, unsigned int);
859         int ret;
860
861         if (unlikely(!(in->f_mode & FMODE_READ)))
862                 return -EBADF;
863
864         ret = rw_verify_area(READ, in, ppos, len);
865         if (unlikely(ret < 0))
866                 return ret;
867
868         if (unlikely(len > MAX_RW_COUNT))
869                 len = MAX_RW_COUNT;
870
871         if (in->f_op->splice_read)
872                 splice_read = in->f_op->splice_read;
873         else
874                 splice_read = default_file_splice_read;
875
876         return splice_read(in, ppos, pipe, len, flags);
877 }
878
879 /**
880  * splice_direct_to_actor - splices data directly between two non-pipes
881  * @in:         file to splice from
882  * @sd:         actor information on where to splice to
883  * @actor:      handles the data splicing
884  *
885  * Description:
886  *    This is a special case helper to splice directly between two
887  *    points, without requiring an explicit pipe. Internally an allocated
888  *    pipe is cached in the process, and reused during the lifetime of
889  *    that process.
890  *
891  */
892 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
893                                splice_direct_actor *actor)
894 {
895         struct pipe_inode_info *pipe;
896         long ret, bytes;
897         umode_t i_mode;
898         size_t len;
899         int i, flags, more;
900
901         /*
902          * We require the input being a regular file, as we don't want to
903          * randomly drop data for eg socket -> socket splicing. Use the
904          * piped splicing for that!
905          */
906         i_mode = file_inode(in)->i_mode;
907         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
908                 return -EINVAL;
909
910         /*
911          * neither in nor out is a pipe, setup an internal pipe attached to
912          * 'out' and transfer the wanted data from 'in' to 'out' through that
913          */
914         pipe = current->splice_pipe;
915         if (unlikely(!pipe)) {
916                 pipe = alloc_pipe_info();
917                 if (!pipe)
918                         return -ENOMEM;
919
920                 /*
921                  * We don't have an immediate reader, but we'll read the stuff
922                  * out of the pipe right after the splice_to_pipe(). So set
923                  * PIPE_READERS appropriately.
924                  */
925                 pipe->readers = 1;
926
927                 current->splice_pipe = pipe;
928         }
929
930         /*
931          * Do the splice.
932          */
933         ret = 0;
934         bytes = 0;
935         len = sd->total_len;
936         flags = sd->flags;
937
938         /*
939          * Don't block on output, we have to drain the direct pipe.
940          */
941         sd->flags &= ~SPLICE_F_NONBLOCK;
942         more = sd->flags & SPLICE_F_MORE;
943
944         WARN_ON_ONCE(pipe->nrbufs != 0);
945
946         while (len) {
947                 size_t read_len;
948                 loff_t pos = sd->pos, prev_pos = pos;
949
950                 /* Don't try to read more the pipe has space for. */
951                 read_len = min_t(size_t, len,
952                                  (pipe->buffers - pipe->nrbufs) << PAGE_SHIFT);
953                 ret = do_splice_to(in, &pos, pipe, read_len, flags);
954                 if (unlikely(ret <= 0))
955                         goto out_release;
956
957                 read_len = ret;
958                 sd->total_len = read_len;
959
960                 /*
961                  * If more data is pending, set SPLICE_F_MORE
962                  * If this is the last data and SPLICE_F_MORE was not set
963                  * initially, clears it.
964                  */
965                 if (read_len < len)
966                         sd->flags |= SPLICE_F_MORE;
967                 else if (!more)
968                         sd->flags &= ~SPLICE_F_MORE;
969                 /*
970                  * NOTE: nonblocking mode only applies to the input. We
971                  * must not do the output in nonblocking mode as then we
972                  * could get stuck data in the internal pipe:
973                  */
974                 ret = actor(pipe, sd);
975                 if (unlikely(ret <= 0)) {
976                         sd->pos = prev_pos;
977                         goto out_release;
978                 }
979
980                 bytes += ret;
981                 len -= ret;
982                 sd->pos = pos;
983
984                 if (ret < read_len) {
985                         sd->pos = prev_pos + ret;
986                         goto out_release;
987                 }
988         }
989
990 done:
991         pipe->nrbufs = pipe->curbuf = 0;
992         file_accessed(in);
993         return bytes;
994
995 out_release:
996         /*
997          * If we did an incomplete transfer we must release
998          * the pipe buffers in question:
999          */
1000         for (i = 0; i < pipe->buffers; i++) {
1001                 struct pipe_buffer *buf = pipe->bufs + i;
1002
1003                 if (buf->ops)
1004                         pipe_buf_release(pipe, buf);
1005         }
1006
1007         if (!bytes)
1008                 bytes = ret;
1009
1010         goto done;
1011 }
1012 EXPORT_SYMBOL(splice_direct_to_actor);
1013
1014 static int direct_splice_actor(struct pipe_inode_info *pipe,
1015                                struct splice_desc *sd)
1016 {
1017         struct file *file = sd->u.file;
1018
1019         return do_splice_from(pipe, file, sd->opos, sd->total_len,
1020                               sd->flags);
1021 }
1022
1023 /**
1024  * do_splice_direct - splices data directly between two files
1025  * @in:         file to splice from
1026  * @ppos:       input file offset
1027  * @out:        file to splice to
1028  * @opos:       output file offset
1029  * @len:        number of bytes to splice
1030  * @flags:      splice modifier flags
1031  *
1032  * Description:
1033  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1034  *    doing it in the application would incur an extra system call
1035  *    (splice in + splice out, as compared to just sendfile()). So this helper
1036  *    can splice directly through a process-private pipe.
1037  *
1038  */
1039 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1040                       loff_t *opos, size_t len, unsigned int flags)
1041 {
1042         struct splice_desc sd = {
1043                 .len            = len,
1044                 .total_len      = len,
1045                 .flags          = flags,
1046                 .pos            = *ppos,
1047                 .u.file         = out,
1048                 .opos           = opos,
1049         };
1050         long ret;
1051
1052         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1053                 return -EBADF;
1054
1055         if (unlikely(out->f_flags & O_APPEND))
1056                 return -EINVAL;
1057
1058         ret = rw_verify_area(WRITE, out, opos, len);
1059         if (unlikely(ret < 0))
1060                 return ret;
1061
1062         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1063         if (ret > 0)
1064                 *ppos = sd.pos;
1065
1066         return ret;
1067 }
1068 EXPORT_SYMBOL(do_splice_direct);
1069
1070 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1071 {
1072         for (;;) {
1073                 if (unlikely(!pipe->readers)) {
1074                         send_sig(SIGPIPE, current, 0);
1075                         return -EPIPE;
1076                 }
1077                 if (pipe->nrbufs != pipe->buffers)
1078                         return 0;
1079                 if (flags & SPLICE_F_NONBLOCK)
1080                         return -EAGAIN;
1081                 if (signal_pending(current))
1082                         return -ERESTARTSYS;
1083                 pipe->waiting_writers++;
1084                 pipe_wait(pipe);
1085                 pipe->waiting_writers--;
1086         }
1087 }
1088
1089 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1090                                struct pipe_inode_info *opipe,
1091                                size_t len, unsigned int flags);
1092
1093 /*
1094  * Determine where to splice to/from.
1095  */
1096 static long do_splice(struct file *in, loff_t __user *off_in,
1097                       struct file *out, loff_t __user *off_out,
1098                       size_t len, unsigned int flags)
1099 {
1100         struct pipe_inode_info *ipipe;
1101         struct pipe_inode_info *opipe;
1102         loff_t offset;
1103         long ret;
1104
1105         ipipe = get_pipe_info(in);
1106         opipe = get_pipe_info(out);
1107
1108         if (ipipe && opipe) {
1109                 if (off_in || off_out)
1110                         return -ESPIPE;
1111
1112                 if (!(in->f_mode & FMODE_READ))
1113                         return -EBADF;
1114
1115                 if (!(out->f_mode & FMODE_WRITE))
1116                         return -EBADF;
1117
1118                 /* Splicing to self would be fun, but... */
1119                 if (ipipe == opipe)
1120                         return -EINVAL;
1121
1122                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1123                         flags |= SPLICE_F_NONBLOCK;
1124
1125                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1126         }
1127
1128         if (ipipe) {
1129                 if (off_in)
1130                         return -ESPIPE;
1131                 if (off_out) {
1132                         if (!(out->f_mode & FMODE_PWRITE))
1133                                 return -EINVAL;
1134                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1135                                 return -EFAULT;
1136                 } else {
1137                         offset = out->f_pos;
1138                 }
1139
1140                 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1141                         return -EBADF;
1142
1143                 if (unlikely(out->f_flags & O_APPEND))
1144                         return -EINVAL;
1145
1146                 ret = rw_verify_area(WRITE, out, &offset, len);
1147                 if (unlikely(ret < 0))
1148                         return ret;
1149
1150                 if (in->f_flags & O_NONBLOCK)
1151                         flags |= SPLICE_F_NONBLOCK;
1152
1153                 file_start_write(out);
1154                 ret = do_splice_from(ipipe, out, &offset, len, flags);
1155                 file_end_write(out);
1156
1157                 if (!off_out)
1158                         out->f_pos = offset;
1159                 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1160                         ret = -EFAULT;
1161
1162                 return ret;
1163         }
1164
1165         if (opipe) {
1166                 if (off_out)
1167                         return -ESPIPE;
1168                 if (off_in) {
1169                         if (!(in->f_mode & FMODE_PREAD))
1170                                 return -EINVAL;
1171                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1172                                 return -EFAULT;
1173                 } else {
1174                         offset = in->f_pos;
1175                 }
1176
1177                 if (out->f_flags & O_NONBLOCK)
1178                         flags |= SPLICE_F_NONBLOCK;
1179
1180                 pipe_lock(opipe);
1181                 ret = wait_for_space(opipe, flags);
1182                 if (!ret)
1183                         ret = do_splice_to(in, &offset, opipe, len, flags);
1184                 pipe_unlock(opipe);
1185                 if (ret > 0)
1186                         wakeup_pipe_readers(opipe);
1187                 if (!off_in)
1188                         in->f_pos = offset;
1189                 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1190                         ret = -EFAULT;
1191
1192                 return ret;
1193         }
1194
1195         return -EINVAL;
1196 }
1197
1198 static int iter_to_pipe(struct iov_iter *from,
1199                         struct pipe_inode_info *pipe,
1200                         unsigned flags)
1201 {
1202         struct pipe_buffer buf = {
1203                 .ops = &user_page_pipe_buf_ops,
1204                 .flags = flags
1205         };
1206         size_t total = 0;
1207         int ret = 0;
1208         bool failed = false;
1209
1210         while (iov_iter_count(from) && !failed) {
1211                 struct page *pages[16];
1212                 ssize_t copied;
1213                 size_t start;
1214                 int n;
1215
1216                 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1217                 if (copied <= 0) {
1218                         ret = copied;
1219                         break;
1220                 }
1221
1222                 for (n = 0; copied; n++, start = 0) {
1223                         int size = min_t(int, copied, PAGE_SIZE - start);
1224                         if (!failed) {
1225                                 buf.page = pages[n];
1226                                 buf.offset = start;
1227                                 buf.len = size;
1228                                 ret = add_to_pipe(pipe, &buf);
1229                                 if (unlikely(ret < 0)) {
1230                                         failed = true;
1231                                 } else {
1232                                         iov_iter_advance(from, ret);
1233                                         total += ret;
1234                                 }
1235                         } else {
1236                                 put_page(pages[n]);
1237                         }
1238                         copied -= size;
1239                 }
1240         }
1241         return total ? total : ret;
1242 }
1243
1244 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1245                         struct splice_desc *sd)
1246 {
1247         int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1248         return n == sd->len ? n : -EFAULT;
1249 }
1250
1251 /*
1252  * For lack of a better implementation, implement vmsplice() to userspace
1253  * as a simple copy of the pipes pages to the user iov.
1254  */
1255 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1256                              unsigned int flags)
1257 {
1258         struct pipe_inode_info *pipe = get_pipe_info(file);
1259         struct splice_desc sd = {
1260                 .total_len = iov_iter_count(iter),
1261                 .flags = flags,
1262                 .u.data = iter
1263         };
1264         long ret = 0;
1265
1266         if (!pipe)
1267                 return -EBADF;
1268
1269         if (sd.total_len) {
1270                 pipe_lock(pipe);
1271                 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1272                 pipe_unlock(pipe);
1273         }
1274
1275         return ret;
1276 }
1277
1278 /*
1279  * vmsplice splices a user address range into a pipe. It can be thought of
1280  * as splice-from-memory, where the regular splice is splice-from-file (or
1281  * to file). In both cases the output is a pipe, naturally.
1282  */
1283 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1284                              unsigned int flags)
1285 {
1286         struct pipe_inode_info *pipe;
1287         long ret = 0;
1288         unsigned buf_flag = 0;
1289
1290         if (flags & SPLICE_F_GIFT)
1291                 buf_flag = PIPE_BUF_FLAG_GIFT;
1292
1293         pipe = get_pipe_info(file);
1294         if (!pipe)
1295                 return -EBADF;
1296
1297         pipe_lock(pipe);
1298         ret = wait_for_space(pipe, flags);
1299         if (!ret)
1300                 ret = iter_to_pipe(iter, pipe, buf_flag);
1301         pipe_unlock(pipe);
1302         if (ret > 0)
1303                 wakeup_pipe_readers(pipe);
1304         return ret;
1305 }
1306
1307 static int vmsplice_type(struct fd f, int *type)
1308 {
1309         if (!f.file)
1310                 return -EBADF;
1311         if (f.file->f_mode & FMODE_WRITE) {
1312                 *type = WRITE;
1313         } else if (f.file->f_mode & FMODE_READ) {
1314                 *type = READ;
1315         } else {
1316                 fdput(f);
1317                 return -EBADF;
1318         }
1319         return 0;
1320 }
1321
1322 /*
1323  * Note that vmsplice only really supports true splicing _from_ user memory
1324  * to a pipe, not the other way around. Splicing from user memory is a simple
1325  * operation that can be supported without any funky alignment restrictions
1326  * or nasty vm tricks. We simply map in the user memory and fill them into
1327  * a pipe. The reverse isn't quite as easy, though. There are two possible
1328  * solutions for that:
1329  *
1330  *      - memcpy() the data internally, at which point we might as well just
1331  *        do a regular read() on the buffer anyway.
1332  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1333  *        has restriction limitations on both ends of the pipe).
1334  *
1335  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1336  *
1337  */
1338 static long do_vmsplice(struct file *f, struct iov_iter *iter, unsigned int flags)
1339 {
1340         if (unlikely(flags & ~SPLICE_F_ALL))
1341                 return -EINVAL;
1342
1343         if (!iov_iter_count(iter))
1344                 return 0;
1345
1346         if (iov_iter_rw(iter) == WRITE)
1347                 return vmsplice_to_pipe(f, iter, flags);
1348         else
1349                 return vmsplice_to_user(f, iter, flags);
1350 }
1351
1352 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1353                 unsigned long, nr_segs, unsigned int, flags)
1354 {
1355         struct iovec iovstack[UIO_FASTIOV];
1356         struct iovec *iov = iovstack;
1357         struct iov_iter iter;
1358         long error;
1359         struct fd f;
1360         int type;
1361
1362         f = fdget(fd);
1363         error = vmsplice_type(f, &type);
1364         if (error)
1365                 return error;
1366
1367         error = import_iovec(type, uiov, nr_segs,
1368                              ARRAY_SIZE(iovstack), &iov, &iter);
1369         if (!error) {
1370                 error = do_vmsplice(f.file, &iter, flags);
1371                 kfree(iov);
1372         }
1373         fdput(f);
1374         return error;
1375 }
1376
1377 #ifdef CONFIG_COMPAT
1378 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1379                     unsigned int, nr_segs, unsigned int, flags)
1380 {
1381         struct iovec iovstack[UIO_FASTIOV];
1382         struct iovec *iov = iovstack;
1383         struct iov_iter iter;
1384         long error;
1385         struct fd f;
1386         int type;
1387
1388         f = fdget(fd);
1389         error = vmsplice_type(f, &type);
1390         if (error)
1391                 return error;
1392
1393         error = compat_import_iovec(type, iov32, nr_segs,
1394                              ARRAY_SIZE(iovstack), &iov, &iter);
1395         if (!error) {
1396                 error = do_vmsplice(f.file, &iter, flags);
1397                 kfree(iov);
1398         }
1399         fdput(f);
1400         return error;
1401 }
1402 #endif
1403
1404 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1405                 int, fd_out, loff_t __user *, off_out,
1406                 size_t, len, unsigned int, flags)
1407 {
1408         struct fd in, out;
1409         long error;
1410
1411         if (unlikely(!len))
1412                 return 0;
1413
1414         if (unlikely(flags & ~SPLICE_F_ALL))
1415                 return -EINVAL;
1416
1417         error = -EBADF;
1418         in = fdget(fd_in);
1419         if (in.file) {
1420                 if (in.file->f_mode & FMODE_READ) {
1421                         out = fdget(fd_out);
1422                         if (out.file) {
1423                                 if (out.file->f_mode & FMODE_WRITE)
1424                                         error = do_splice(in.file, off_in,
1425                                                           out.file, off_out,
1426                                                           len, flags);
1427                                 fdput(out);
1428                         }
1429                 }
1430                 fdput(in);
1431         }
1432         return error;
1433 }
1434
1435 /*
1436  * Make sure there's data to read. Wait for input if we can, otherwise
1437  * return an appropriate error.
1438  */
1439 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1440 {
1441         int ret;
1442
1443         /*
1444          * Check ->nrbufs without the inode lock first. This function
1445          * is speculative anyways, so missing one is ok.
1446          */
1447         if (pipe->nrbufs)
1448                 return 0;
1449
1450         ret = 0;
1451         pipe_lock(pipe);
1452
1453         while (!pipe->nrbufs) {
1454                 if (signal_pending(current)) {
1455                         ret = -ERESTARTSYS;
1456                         break;
1457                 }
1458                 if (!pipe->writers)
1459                         break;
1460                 if (!pipe->waiting_writers) {
1461                         if (flags & SPLICE_F_NONBLOCK) {
1462                                 ret = -EAGAIN;
1463                                 break;
1464                         }
1465                 }
1466                 pipe_wait(pipe);
1467         }
1468
1469         pipe_unlock(pipe);
1470         return ret;
1471 }
1472
1473 /*
1474  * Make sure there's writeable room. Wait for room if we can, otherwise
1475  * return an appropriate error.
1476  */
1477 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1478 {
1479         int ret;
1480
1481         /*
1482          * Check ->nrbufs without the inode lock first. This function
1483          * is speculative anyways, so missing one is ok.
1484          */
1485         if (pipe->nrbufs < pipe->buffers)
1486                 return 0;
1487
1488         ret = 0;
1489         pipe_lock(pipe);
1490
1491         while (pipe->nrbufs >= pipe->buffers) {
1492                 if (!pipe->readers) {
1493                         send_sig(SIGPIPE, current, 0);
1494                         ret = -EPIPE;
1495                         break;
1496                 }
1497                 if (flags & SPLICE_F_NONBLOCK) {
1498                         ret = -EAGAIN;
1499                         break;
1500                 }
1501                 if (signal_pending(current)) {
1502                         ret = -ERESTARTSYS;
1503                         break;
1504                 }
1505                 pipe->waiting_writers++;
1506                 pipe_wait(pipe);
1507                 pipe->waiting_writers--;
1508         }
1509
1510         pipe_unlock(pipe);
1511         return ret;
1512 }
1513
1514 /*
1515  * Splice contents of ipipe to opipe.
1516  */
1517 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1518                                struct pipe_inode_info *opipe,
1519                                size_t len, unsigned int flags)
1520 {
1521         struct pipe_buffer *ibuf, *obuf;
1522         int ret = 0, nbuf;
1523         bool input_wakeup = false;
1524
1525
1526 retry:
1527         ret = ipipe_prep(ipipe, flags);
1528         if (ret)
1529                 return ret;
1530
1531         ret = opipe_prep(opipe, flags);
1532         if (ret)
1533                 return ret;
1534
1535         /*
1536          * Potential ABBA deadlock, work around it by ordering lock
1537          * grabbing by pipe info address. Otherwise two different processes
1538          * could deadlock (one doing tee from A -> B, the other from B -> A).
1539          */
1540         pipe_double_lock(ipipe, opipe);
1541
1542         do {
1543                 if (!opipe->readers) {
1544                         send_sig(SIGPIPE, current, 0);
1545                         if (!ret)
1546                                 ret = -EPIPE;
1547                         break;
1548                 }
1549
1550                 if (!ipipe->nrbufs && !ipipe->writers)
1551                         break;
1552
1553                 /*
1554                  * Cannot make any progress, because either the input
1555                  * pipe is empty or the output pipe is full.
1556                  */
1557                 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1558                         /* Already processed some buffers, break */
1559                         if (ret)
1560                                 break;
1561
1562                         if (flags & SPLICE_F_NONBLOCK) {
1563                                 ret = -EAGAIN;
1564                                 break;
1565                         }
1566
1567                         /*
1568                          * We raced with another reader/writer and haven't
1569                          * managed to process any buffers.  A zero return
1570                          * value means EOF, so retry instead.
1571                          */
1572                         pipe_unlock(ipipe);
1573                         pipe_unlock(opipe);
1574                         goto retry;
1575                 }
1576
1577                 ibuf = ipipe->bufs + ipipe->curbuf;
1578                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1579                 obuf = opipe->bufs + nbuf;
1580
1581                 if (len >= ibuf->len) {
1582                         /*
1583                          * Simply move the whole buffer from ipipe to opipe
1584                          */
1585                         *obuf = *ibuf;
1586                         ibuf->ops = NULL;
1587                         opipe->nrbufs++;
1588                         ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1589                         ipipe->nrbufs--;
1590                         input_wakeup = true;
1591                 } else {
1592                         /*
1593                          * Get a reference to this pipe buffer,
1594                          * so we can copy the contents over.
1595                          */
1596                         pipe_buf_get(ipipe, ibuf);
1597                         *obuf = *ibuf;
1598
1599                         /*
1600                          * Don't inherit the gift flag, we need to
1601                          * prevent multiple steals of this page.
1602                          */
1603                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1604
1605                         pipe_buf_mark_unmergeable(obuf);
1606
1607                         obuf->len = len;
1608                         opipe->nrbufs++;
1609                         ibuf->offset += obuf->len;
1610                         ibuf->len -= obuf->len;
1611                 }
1612                 ret += obuf->len;
1613                 len -= obuf->len;
1614         } while (len);
1615
1616         pipe_unlock(ipipe);
1617         pipe_unlock(opipe);
1618
1619         /*
1620          * If we put data in the output pipe, wakeup any potential readers.
1621          */
1622         if (ret > 0)
1623                 wakeup_pipe_readers(opipe);
1624
1625         if (input_wakeup)
1626                 wakeup_pipe_writers(ipipe);
1627
1628         return ret;
1629 }
1630
1631 /*
1632  * Link contents of ipipe to opipe.
1633  */
1634 static int link_pipe(struct pipe_inode_info *ipipe,
1635                      struct pipe_inode_info *opipe,
1636                      size_t len, unsigned int flags)
1637 {
1638         struct pipe_buffer *ibuf, *obuf;
1639         int ret = 0, i = 0, nbuf;
1640
1641         /*
1642          * Potential ABBA deadlock, work around it by ordering lock
1643          * grabbing by pipe info address. Otherwise two different processes
1644          * could deadlock (one doing tee from A -> B, the other from B -> A).
1645          */
1646         pipe_double_lock(ipipe, opipe);
1647
1648         do {
1649                 if (!opipe->readers) {
1650                         send_sig(SIGPIPE, current, 0);
1651                         if (!ret)
1652                                 ret = -EPIPE;
1653                         break;
1654                 }
1655
1656                 /*
1657                  * If we have iterated all input buffers or ran out of
1658                  * output room, break.
1659                  */
1660                 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1661                         break;
1662
1663                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1664                 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1665
1666                 /*
1667                  * Get a reference to this pipe buffer,
1668                  * so we can copy the contents over.
1669                  */
1670                 pipe_buf_get(ipipe, ibuf);
1671
1672                 obuf = opipe->bufs + nbuf;
1673                 *obuf = *ibuf;
1674
1675                 /*
1676                  * Don't inherit the gift flag, we need to
1677                  * prevent multiple steals of this page.
1678                  */
1679                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1680
1681                 pipe_buf_mark_unmergeable(obuf);
1682
1683                 if (obuf->len > len)
1684                         obuf->len = len;
1685
1686                 opipe->nrbufs++;
1687                 ret += obuf->len;
1688                 len -= obuf->len;
1689                 i++;
1690         } while (len);
1691
1692         /*
1693          * return EAGAIN if we have the potential of some data in the
1694          * future, otherwise just return 0
1695          */
1696         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1697                 ret = -EAGAIN;
1698
1699         pipe_unlock(ipipe);
1700         pipe_unlock(opipe);
1701
1702         /*
1703          * If we put data in the output pipe, wakeup any potential readers.
1704          */
1705         if (ret > 0)
1706                 wakeup_pipe_readers(opipe);
1707
1708         return ret;
1709 }
1710
1711 /*
1712  * This is a tee(1) implementation that works on pipes. It doesn't copy
1713  * any data, it simply references the 'in' pages on the 'out' pipe.
1714  * The 'flags' used are the SPLICE_F_* variants, currently the only
1715  * applicable one is SPLICE_F_NONBLOCK.
1716  */
1717 static long do_tee(struct file *in, struct file *out, size_t len,
1718                    unsigned int flags)
1719 {
1720         struct pipe_inode_info *ipipe = get_pipe_info(in);
1721         struct pipe_inode_info *opipe = get_pipe_info(out);
1722         int ret = -EINVAL;
1723
1724         /*
1725          * Duplicate the contents of ipipe to opipe without actually
1726          * copying the data.
1727          */
1728         if (ipipe && opipe && ipipe != opipe) {
1729                 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1730                         flags |= SPLICE_F_NONBLOCK;
1731
1732                 /*
1733                  * Keep going, unless we encounter an error. The ipipe/opipe
1734                  * ordering doesn't really matter.
1735                  */
1736                 ret = ipipe_prep(ipipe, flags);
1737                 if (!ret) {
1738                         ret = opipe_prep(opipe, flags);
1739                         if (!ret)
1740                                 ret = link_pipe(ipipe, opipe, len, flags);
1741                 }
1742         }
1743
1744         return ret;
1745 }
1746
1747 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1748 {
1749         struct fd in;
1750         int error;
1751
1752         if (unlikely(flags & ~SPLICE_F_ALL))
1753                 return -EINVAL;
1754
1755         if (unlikely(!len))
1756                 return 0;
1757
1758         error = -EBADF;
1759         in = fdget(fdin);
1760         if (in.file) {
1761                 if (in.file->f_mode & FMODE_READ) {
1762                         struct fd out = fdget(fdout);
1763                         if (out.file) {
1764                                 if (out.file->f_mode & FMODE_WRITE)
1765                                         error = do_tee(in.file, out.file,
1766                                                         len, flags);
1767                                 fdput(out);
1768                         }
1769                 }
1770                 fdput(in);
1771         }
1772
1773         return error;
1774 }