Loading...
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 */
46static 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 */
86out_unlock:
87 unlock_page(page);
88 return 1;
89}
90
91static 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 */
102static 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;
135error:
136 unlock_page(page);
137 return err;
138}
139
140const struct pipe_buf_operations page_cache_pipe_buf_ops = {
141 .can_merge = 0,
142 .confirm = page_cache_pipe_buf_confirm,
143 .release = page_cache_pipe_buf_release,
144 .steal = page_cache_pipe_buf_steal,
145 .get = generic_pipe_buf_get,
146};
147
148static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
149 struct pipe_buffer *buf)
150{
151 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
152 return 1;
153
154 buf->flags |= PIPE_BUF_FLAG_LRU;
155 return generic_pipe_buf_steal(pipe, buf);
156}
157
158static const struct pipe_buf_operations user_page_pipe_buf_ops = {
159 .can_merge = 0,
160 .confirm = generic_pipe_buf_confirm,
161 .release = page_cache_pipe_buf_release,
162 .steal = user_page_pipe_buf_steal,
163 .get = generic_pipe_buf_get,
164};
165
166static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
167{
168 smp_mb();
169 if (waitqueue_active(&pipe->wait))
170 wake_up_interruptible(&pipe->wait);
171 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
172}
173
174/**
175 * splice_to_pipe - fill passed data into a pipe
176 * @pipe: pipe to fill
177 * @spd: data to fill
178 *
179 * Description:
180 * @spd contains a map of pages and len/offset tuples, along with
181 * the struct pipe_buf_operations associated with these pages. This
182 * function will link that data to the pipe.
183 *
184 */
185ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
186 struct splice_pipe_desc *spd)
187{
188 unsigned int spd_pages = spd->nr_pages;
189 int ret = 0, page_nr = 0;
190
191 if (!spd_pages)
192 return 0;
193
194 if (unlikely(!pipe->readers)) {
195 send_sig(SIGPIPE, current, 0);
196 ret = -EPIPE;
197 goto out;
198 }
199
200 while (pipe->nrbufs < pipe->buffers) {
201 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
202 struct pipe_buffer *buf = pipe->bufs + newbuf;
203
204 buf->page = spd->pages[page_nr];
205 buf->offset = spd->partial[page_nr].offset;
206 buf->len = spd->partial[page_nr].len;
207 buf->private = spd->partial[page_nr].private;
208 buf->ops = spd->ops;
209 buf->flags = 0;
210
211 pipe->nrbufs++;
212 page_nr++;
213 ret += buf->len;
214
215 if (!--spd->nr_pages)
216 break;
217 }
218
219 if (!ret)
220 ret = -EAGAIN;
221
222out:
223 while (page_nr < spd_pages)
224 spd->spd_release(spd, page_nr++);
225
226 return ret;
227}
228EXPORT_SYMBOL_GPL(splice_to_pipe);
229
230ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
231{
232 int ret;
233
234 if (unlikely(!pipe->readers)) {
235 send_sig(SIGPIPE, current, 0);
236 ret = -EPIPE;
237 } else if (pipe->nrbufs == pipe->buffers) {
238 ret = -EAGAIN;
239 } else {
240 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
241 pipe->bufs[newbuf] = *buf;
242 pipe->nrbufs++;
243 return buf->len;
244 }
245 pipe_buf_release(pipe, buf);
246 return ret;
247}
248EXPORT_SYMBOL(add_to_pipe);
249
250/*
251 * Check if we need to grow the arrays holding pages and partial page
252 * descriptions.
253 */
254int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
255{
256 unsigned int buffers = READ_ONCE(pipe->buffers);
257
258 spd->nr_pages_max = buffers;
259 if (buffers <= PIPE_DEF_BUFFERS)
260 return 0;
261
262 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
263 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
264
265 if (spd->pages && spd->partial)
266 return 0;
267
268 kfree(spd->pages);
269 kfree(spd->partial);
270 return -ENOMEM;
271}
272
273void splice_shrink_spd(struct splice_pipe_desc *spd)
274{
275 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
276 return;
277
278 kfree(spd->pages);
279 kfree(spd->partial);
280}
281
282/**
283 * generic_file_splice_read - splice data from file to a pipe
284 * @in: file to splice from
285 * @ppos: position in @in
286 * @pipe: pipe to splice to
287 * @len: number of bytes to splice
288 * @flags: splice modifier flags
289 *
290 * Description:
291 * Will read pages from given file and fill them into a pipe. Can be
292 * used as long as it has more or less sane ->read_iter().
293 *
294 */
295ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
296 struct pipe_inode_info *pipe, size_t len,
297 unsigned int flags)
298{
299 struct iov_iter to;
300 struct kiocb kiocb;
301 int idx, ret;
302
303 iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
304 idx = to.idx;
305 init_sync_kiocb(&kiocb, in);
306 kiocb.ki_pos = *ppos;
307 ret = call_read_iter(in, &kiocb, &to);
308 if (ret > 0) {
309 *ppos = kiocb.ki_pos;
310 file_accessed(in);
311 } else if (ret < 0) {
312 to.idx = idx;
313 to.iov_offset = 0;
314 iov_iter_advance(&to, 0); /* to free what was emitted */
315 /*
316 * callers of ->splice_read() expect -EAGAIN on
317 * "can't put anything in there", rather than -EFAULT.
318 */
319 if (ret == -EFAULT)
320 ret = -EAGAIN;
321 }
322
323 return ret;
324}
325EXPORT_SYMBOL(generic_file_splice_read);
326
327const struct pipe_buf_operations default_pipe_buf_ops = {
328 .can_merge = 0,
329 .confirm = generic_pipe_buf_confirm,
330 .release = generic_pipe_buf_release,
331 .steal = generic_pipe_buf_steal,
332 .get = generic_pipe_buf_get,
333};
334
335static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
336 struct pipe_buffer *buf)
337{
338 return 1;
339}
340
341/* Pipe buffer operations for a socket and similar. */
342const struct pipe_buf_operations nosteal_pipe_buf_ops = {
343 .can_merge = 0,
344 .confirm = generic_pipe_buf_confirm,
345 .release = generic_pipe_buf_release,
346 .steal = generic_pipe_buf_nosteal,
347 .get = generic_pipe_buf_get,
348};
349EXPORT_SYMBOL(nosteal_pipe_buf_ops);
350
351static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
352 unsigned long vlen, loff_t offset)
353{
354 mm_segment_t old_fs;
355 loff_t pos = offset;
356 ssize_t res;
357
358 old_fs = get_fs();
359 set_fs(get_ds());
360 /* The cast to a user pointer is valid due to the set_fs() */
361 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
362 set_fs(old_fs);
363
364 return res;
365}
366
367static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
368 struct pipe_inode_info *pipe, size_t len,
369 unsigned int flags)
370{
371 struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
372 struct iov_iter to;
373 struct page **pages;
374 unsigned int nr_pages;
375 size_t offset, base, copied = 0;
376 ssize_t res;
377 int i;
378
379 if (pipe->nrbufs == pipe->buffers)
380 return -EAGAIN;
381
382 /*
383 * Try to keep page boundaries matching to source pagecache ones -
384 * it probably won't be much help, but...
385 */
386 offset = *ppos & ~PAGE_MASK;
387
388 iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
389
390 res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &base);
391 if (res <= 0)
392 return -ENOMEM;
393
394 nr_pages = DIV_ROUND_UP(res + base, PAGE_SIZE);
395
396 vec = __vec;
397 if (nr_pages > PIPE_DEF_BUFFERS) {
398 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
399 if (unlikely(!vec)) {
400 res = -ENOMEM;
401 goto out;
402 }
403 }
404
405 pipe->bufs[to.idx].offset = offset;
406 pipe->bufs[to.idx].len -= offset;
407
408 for (i = 0; i < nr_pages; i++) {
409 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
410 vec[i].iov_base = page_address(pages[i]) + offset;
411 vec[i].iov_len = this_len;
412 len -= this_len;
413 offset = 0;
414 }
415
416 res = kernel_readv(in, vec, nr_pages, *ppos);
417 if (res > 0) {
418 copied = res;
419 *ppos += res;
420 }
421
422 if (vec != __vec)
423 kfree(vec);
424out:
425 for (i = 0; i < nr_pages; i++)
426 put_page(pages[i]);
427 kvfree(pages);
428 iov_iter_advance(&to, copied); /* truncates and discards */
429 return res;
430}
431
432/*
433 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
434 * using sendpage(). Return the number of bytes sent.
435 */
436static int pipe_to_sendpage(struct pipe_inode_info *pipe,
437 struct pipe_buffer *buf, struct splice_desc *sd)
438{
439 struct file *file = sd->u.file;
440 loff_t pos = sd->pos;
441 int more;
442
443 if (!likely(file->f_op->sendpage))
444 return -EINVAL;
445
446 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
447
448 if (sd->len < sd->total_len && pipe->nrbufs > 1)
449 more |= MSG_SENDPAGE_NOTLAST;
450
451 return file->f_op->sendpage(file, buf->page, buf->offset,
452 sd->len, &pos, more);
453}
454
455static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
456{
457 smp_mb();
458 if (waitqueue_active(&pipe->wait))
459 wake_up_interruptible(&pipe->wait);
460 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
461}
462
463/**
464 * splice_from_pipe_feed - feed available data from a pipe to a file
465 * @pipe: pipe to splice from
466 * @sd: information to @actor
467 * @actor: handler that splices the data
468 *
469 * Description:
470 * This function loops over the pipe and calls @actor to do the
471 * actual moving of a single struct pipe_buffer to the desired
472 * destination. It returns when there's no more buffers left in
473 * the pipe or if the requested number of bytes (@sd->total_len)
474 * have been copied. It returns a positive number (one) if the
475 * pipe needs to be filled with more data, zero if the required
476 * number of bytes have been copied and -errno on error.
477 *
478 * This, together with splice_from_pipe_{begin,end,next}, may be
479 * used to implement the functionality of __splice_from_pipe() when
480 * locking is required around copying the pipe buffers to the
481 * destination.
482 */
483static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
484 splice_actor *actor)
485{
486 int ret;
487
488 while (pipe->nrbufs) {
489 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
490
491 sd->len = buf->len;
492 if (sd->len > sd->total_len)
493 sd->len = sd->total_len;
494
495 ret = pipe_buf_confirm(pipe, buf);
496 if (unlikely(ret)) {
497 if (ret == -ENODATA)
498 ret = 0;
499 return ret;
500 }
501
502 ret = actor(pipe, buf, sd);
503 if (ret <= 0)
504 return ret;
505
506 buf->offset += ret;
507 buf->len -= ret;
508
509 sd->num_spliced += ret;
510 sd->len -= ret;
511 sd->pos += ret;
512 sd->total_len -= ret;
513
514 if (!buf->len) {
515 pipe_buf_release(pipe, buf);
516 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
517 pipe->nrbufs--;
518 if (pipe->files)
519 sd->need_wakeup = true;
520 }
521
522 if (!sd->total_len)
523 return 0;
524 }
525
526 return 1;
527}
528
529/**
530 * splice_from_pipe_next - wait for some data to splice from
531 * @pipe: pipe to splice from
532 * @sd: information about the splice operation
533 *
534 * Description:
535 * This function will wait for some data and return a positive
536 * value (one) if pipe buffers are available. It will return zero
537 * or -errno if no more data needs to be spliced.
538 */
539static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
540{
541 /*
542 * Check for signal early to make process killable when there are
543 * always buffers available
544 */
545 if (signal_pending(current))
546 return -ERESTARTSYS;
547
548 while (!pipe->nrbufs) {
549 if (!pipe->writers)
550 return 0;
551
552 if (!pipe->waiting_writers && sd->num_spliced)
553 return 0;
554
555 if (sd->flags & SPLICE_F_NONBLOCK)
556 return -EAGAIN;
557
558 if (signal_pending(current))
559 return -ERESTARTSYS;
560
561 if (sd->need_wakeup) {
562 wakeup_pipe_writers(pipe);
563 sd->need_wakeup = false;
564 }
565
566 pipe_wait(pipe);
567 }
568
569 return 1;
570}
571
572/**
573 * splice_from_pipe_begin - start splicing from pipe
574 * @sd: information about the splice operation
575 *
576 * Description:
577 * This function should be called before a loop containing
578 * splice_from_pipe_next() and splice_from_pipe_feed() to
579 * initialize the necessary fields of @sd.
580 */
581static void splice_from_pipe_begin(struct splice_desc *sd)
582{
583 sd->num_spliced = 0;
584 sd->need_wakeup = false;
585}
586
587/**
588 * splice_from_pipe_end - finish splicing from pipe
589 * @pipe: pipe to splice from
590 * @sd: information about the splice operation
591 *
592 * Description:
593 * This function will wake up pipe writers if necessary. It should
594 * be called after a loop containing splice_from_pipe_next() and
595 * splice_from_pipe_feed().
596 */
597static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
598{
599 if (sd->need_wakeup)
600 wakeup_pipe_writers(pipe);
601}
602
603/**
604 * __splice_from_pipe - splice data from a pipe to given actor
605 * @pipe: pipe to splice from
606 * @sd: information to @actor
607 * @actor: handler that splices the data
608 *
609 * Description:
610 * This function does little more than loop over the pipe and call
611 * @actor to do the actual moving of a single struct pipe_buffer to
612 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
613 * pipe_to_user.
614 *
615 */
616ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
617 splice_actor *actor)
618{
619 int ret;
620
621 splice_from_pipe_begin(sd);
622 do {
623 cond_resched();
624 ret = splice_from_pipe_next(pipe, sd);
625 if (ret > 0)
626 ret = splice_from_pipe_feed(pipe, sd, actor);
627 } while (ret > 0);
628 splice_from_pipe_end(pipe, sd);
629
630 return sd->num_spliced ? sd->num_spliced : ret;
631}
632EXPORT_SYMBOL(__splice_from_pipe);
633
634/**
635 * splice_from_pipe - splice data from a pipe to a file
636 * @pipe: pipe to splice from
637 * @out: file to splice to
638 * @ppos: position in @out
639 * @len: how many bytes to splice
640 * @flags: splice modifier flags
641 * @actor: handler that splices the data
642 *
643 * Description:
644 * See __splice_from_pipe. This function locks the pipe inode,
645 * otherwise it's identical to __splice_from_pipe().
646 *
647 */
648ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
649 loff_t *ppos, size_t len, unsigned int flags,
650 splice_actor *actor)
651{
652 ssize_t ret;
653 struct splice_desc sd = {
654 .total_len = len,
655 .flags = flags,
656 .pos = *ppos,
657 .u.file = out,
658 };
659
660 pipe_lock(pipe);
661 ret = __splice_from_pipe(pipe, &sd, actor);
662 pipe_unlock(pipe);
663
664 return ret;
665}
666
667/**
668 * iter_file_splice_write - splice data from a pipe to a file
669 * @pipe: pipe info
670 * @out: file to write to
671 * @ppos: position in @out
672 * @len: number of bytes to splice
673 * @flags: splice modifier flags
674 *
675 * Description:
676 * Will either move or copy pages (determined by @flags options) from
677 * the given pipe inode to the given file.
678 * This one is ->write_iter-based.
679 *
680 */
681ssize_t
682iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
683 loff_t *ppos, size_t len, unsigned int flags)
684{
685 struct splice_desc sd = {
686 .total_len = len,
687 .flags = flags,
688 .pos = *ppos,
689 .u.file = out,
690 };
691 int nbufs = pipe->buffers;
692 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
693 GFP_KERNEL);
694 ssize_t ret;
695
696 if (unlikely(!array))
697 return -ENOMEM;
698
699 pipe_lock(pipe);
700
701 splice_from_pipe_begin(&sd);
702 while (sd.total_len) {
703 struct iov_iter from;
704 size_t left;
705 int n, idx;
706
707 ret = splice_from_pipe_next(pipe, &sd);
708 if (ret <= 0)
709 break;
710
711 if (unlikely(nbufs < pipe->buffers)) {
712 kfree(array);
713 nbufs = pipe->buffers;
714 array = kcalloc(nbufs, sizeof(struct bio_vec),
715 GFP_KERNEL);
716 if (!array) {
717 ret = -ENOMEM;
718 break;
719 }
720 }
721
722 /* build the vector */
723 left = sd.total_len;
724 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
725 struct pipe_buffer *buf = pipe->bufs + idx;
726 size_t this_len = buf->len;
727
728 if (this_len > left)
729 this_len = left;
730
731 if (idx == pipe->buffers - 1)
732 idx = -1;
733
734 ret = pipe_buf_confirm(pipe, buf);
735 if (unlikely(ret)) {
736 if (ret == -ENODATA)
737 ret = 0;
738 goto done;
739 }
740
741 array[n].bv_page = buf->page;
742 array[n].bv_len = this_len;
743 array[n].bv_offset = buf->offset;
744 left -= this_len;
745 }
746
747 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
748 sd.total_len - left);
749 ret = vfs_iter_write(out, &from, &sd.pos, 0);
750 if (ret <= 0)
751 break;
752
753 sd.num_spliced += ret;
754 sd.total_len -= ret;
755 *ppos = sd.pos;
756
757 /* dismiss the fully eaten buffers, adjust the partial one */
758 while (ret) {
759 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
760 if (ret >= buf->len) {
761 ret -= buf->len;
762 buf->len = 0;
763 pipe_buf_release(pipe, buf);
764 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
765 pipe->nrbufs--;
766 if (pipe->files)
767 sd.need_wakeup = true;
768 } else {
769 buf->offset += ret;
770 buf->len -= ret;
771 ret = 0;
772 }
773 }
774 }
775done:
776 kfree(array);
777 splice_from_pipe_end(pipe, &sd);
778
779 pipe_unlock(pipe);
780
781 if (sd.num_spliced)
782 ret = sd.num_spliced;
783
784 return ret;
785}
786
787EXPORT_SYMBOL(iter_file_splice_write);
788
789static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
790 struct splice_desc *sd)
791{
792 int ret;
793 void *data;
794 loff_t tmp = sd->pos;
795
796 data = kmap(buf->page);
797 ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
798 kunmap(buf->page);
799
800 return ret;
801}
802
803static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
804 struct file *out, loff_t *ppos,
805 size_t len, unsigned int flags)
806{
807 ssize_t ret;
808
809 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
810 if (ret > 0)
811 *ppos += ret;
812
813 return ret;
814}
815
816/**
817 * generic_splice_sendpage - splice data from a pipe to a socket
818 * @pipe: pipe to splice from
819 * @out: socket to write to
820 * @ppos: position in @out
821 * @len: number of bytes to splice
822 * @flags: splice modifier flags
823 *
824 * Description:
825 * Will send @len bytes from the pipe to a network socket. No data copying
826 * is involved.
827 *
828 */
829ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
830 loff_t *ppos, size_t len, unsigned int flags)
831{
832 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
833}
834
835EXPORT_SYMBOL(generic_splice_sendpage);
836
837/*
838 * Attempt to initiate a splice from pipe to file.
839 */
840static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
841 loff_t *ppos, size_t len, unsigned int flags)
842{
843 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
844 loff_t *, size_t, unsigned int);
845
846 if (out->f_op->splice_write)
847 splice_write = out->f_op->splice_write;
848 else
849 splice_write = default_file_splice_write;
850
851 return splice_write(pipe, out, ppos, len, flags);
852}
853
854/*
855 * Attempt to initiate a splice from a file to a pipe.
856 */
857static long do_splice_to(struct file *in, loff_t *ppos,
858 struct pipe_inode_info *pipe, size_t len,
859 unsigned int flags)
860{
861 ssize_t (*splice_read)(struct file *, loff_t *,
862 struct pipe_inode_info *, size_t, unsigned int);
863 int ret;
864
865 if (unlikely(!(in->f_mode & FMODE_READ)))
866 return -EBADF;
867
868 ret = rw_verify_area(READ, in, ppos, len);
869 if (unlikely(ret < 0))
870 return ret;
871
872 if (unlikely(len > MAX_RW_COUNT))
873 len = MAX_RW_COUNT;
874
875 if (in->f_op->splice_read)
876 splice_read = in->f_op->splice_read;
877 else
878 splice_read = default_file_splice_read;
879
880 return splice_read(in, ppos, pipe, len, flags);
881}
882
883/**
884 * splice_direct_to_actor - splices data directly between two non-pipes
885 * @in: file to splice from
886 * @sd: actor information on where to splice to
887 * @actor: handles the data splicing
888 *
889 * Description:
890 * This is a special case helper to splice directly between two
891 * points, without requiring an explicit pipe. Internally an allocated
892 * pipe is cached in the process, and reused during the lifetime of
893 * that process.
894 *
895 */
896ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
897 splice_direct_actor *actor)
898{
899 struct pipe_inode_info *pipe;
900 long ret, bytes;
901 umode_t i_mode;
902 size_t len;
903 int i, flags, more;
904
905 /*
906 * We require the input being a regular file, as we don't want to
907 * randomly drop data for eg socket -> socket splicing. Use the
908 * piped splicing for that!
909 */
910 i_mode = file_inode(in)->i_mode;
911 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
912 return -EINVAL;
913
914 /*
915 * neither in nor out is a pipe, setup an internal pipe attached to
916 * 'out' and transfer the wanted data from 'in' to 'out' through that
917 */
918 pipe = current->splice_pipe;
919 if (unlikely(!pipe)) {
920 pipe = alloc_pipe_info();
921 if (!pipe)
922 return -ENOMEM;
923
924 /*
925 * We don't have an immediate reader, but we'll read the stuff
926 * out of the pipe right after the splice_to_pipe(). So set
927 * PIPE_READERS appropriately.
928 */
929 pipe->readers = 1;
930
931 current->splice_pipe = pipe;
932 }
933
934 /*
935 * Do the splice.
936 */
937 ret = 0;
938 bytes = 0;
939 len = sd->total_len;
940 flags = sd->flags;
941
942 /*
943 * Don't block on output, we have to drain the direct pipe.
944 */
945 sd->flags &= ~SPLICE_F_NONBLOCK;
946 more = sd->flags & SPLICE_F_MORE;
947
948 while (len) {
949 size_t read_len;
950 loff_t pos = sd->pos, prev_pos = pos;
951
952 ret = do_splice_to(in, &pos, pipe, len, flags);
953 if (unlikely(ret <= 0))
954 goto out_release;
955
956 read_len = ret;
957 sd->total_len = read_len;
958
959 /*
960 * If more data is pending, set SPLICE_F_MORE
961 * If this is the last data and SPLICE_F_MORE was not set
962 * initially, clears it.
963 */
964 if (read_len < len)
965 sd->flags |= SPLICE_F_MORE;
966 else if (!more)
967 sd->flags &= ~SPLICE_F_MORE;
968 /*
969 * NOTE: nonblocking mode only applies to the input. We
970 * must not do the output in nonblocking mode as then we
971 * could get stuck data in the internal pipe:
972 */
973 ret = actor(pipe, sd);
974 if (unlikely(ret <= 0)) {
975 sd->pos = prev_pos;
976 goto out_release;
977 }
978
979 bytes += ret;
980 len -= ret;
981 sd->pos = pos;
982
983 if (ret < read_len) {
984 sd->pos = prev_pos + ret;
985 goto out_release;
986 }
987 }
988
989done:
990 pipe->nrbufs = pipe->curbuf = 0;
991 file_accessed(in);
992 return bytes;
993
994out_release:
995 /*
996 * If we did an incomplete transfer we must release
997 * the pipe buffers in question:
998 */
999 for (i = 0; i < pipe->buffers; i++) {
1000 struct pipe_buffer *buf = pipe->bufs + i;
1001
1002 if (buf->ops)
1003 pipe_buf_release(pipe, buf);
1004 }
1005
1006 if (!bytes)
1007 bytes = ret;
1008
1009 goto done;
1010}
1011EXPORT_SYMBOL(splice_direct_to_actor);
1012
1013static int direct_splice_actor(struct pipe_inode_info *pipe,
1014 struct splice_desc *sd)
1015{
1016 struct file *file = sd->u.file;
1017
1018 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1019 sd->flags);
1020}
1021
1022/**
1023 * do_splice_direct - splices data directly between two files
1024 * @in: file to splice from
1025 * @ppos: input file offset
1026 * @out: file to splice to
1027 * @opos: output file offset
1028 * @len: number of bytes to splice
1029 * @flags: splice modifier flags
1030 *
1031 * Description:
1032 * For use by do_sendfile(). splice can easily emulate sendfile, but
1033 * doing it in the application would incur an extra system call
1034 * (splice in + splice out, as compared to just sendfile()). So this helper
1035 * can splice directly through a process-private pipe.
1036 *
1037 */
1038long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1039 loff_t *opos, size_t len, unsigned int flags)
1040{
1041 struct splice_desc sd = {
1042 .len = len,
1043 .total_len = len,
1044 .flags = flags,
1045 .pos = *ppos,
1046 .u.file = out,
1047 .opos = opos,
1048 };
1049 long ret;
1050
1051 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1052 return -EBADF;
1053
1054 if (unlikely(out->f_flags & O_APPEND))
1055 return -EINVAL;
1056
1057 ret = rw_verify_area(WRITE, out, opos, len);
1058 if (unlikely(ret < 0))
1059 return ret;
1060
1061 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1062 if (ret > 0)
1063 *ppos = sd.pos;
1064
1065 return ret;
1066}
1067EXPORT_SYMBOL(do_splice_direct);
1068
1069static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1070{
1071 for (;;) {
1072 if (unlikely(!pipe->readers)) {
1073 send_sig(SIGPIPE, current, 0);
1074 return -EPIPE;
1075 }
1076 if (pipe->nrbufs != pipe->buffers)
1077 return 0;
1078 if (flags & SPLICE_F_NONBLOCK)
1079 return -EAGAIN;
1080 if (signal_pending(current))
1081 return -ERESTARTSYS;
1082 pipe->waiting_writers++;
1083 pipe_wait(pipe);
1084 pipe->waiting_writers--;
1085 }
1086}
1087
1088static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1089 struct pipe_inode_info *opipe,
1090 size_t len, unsigned int flags);
1091
1092/*
1093 * Determine where to splice to/from.
1094 */
1095static long do_splice(struct file *in, loff_t __user *off_in,
1096 struct file *out, loff_t __user *off_out,
1097 size_t len, unsigned int flags)
1098{
1099 struct pipe_inode_info *ipipe;
1100 struct pipe_inode_info *opipe;
1101 loff_t offset;
1102 long ret;
1103
1104 ipipe = get_pipe_info(in);
1105 opipe = get_pipe_info(out);
1106
1107 if (ipipe && opipe) {
1108 if (off_in || off_out)
1109 return -ESPIPE;
1110
1111 if (!(in->f_mode & FMODE_READ))
1112 return -EBADF;
1113
1114 if (!(out->f_mode & FMODE_WRITE))
1115 return -EBADF;
1116
1117 /* Splicing to self would be fun, but... */
1118 if (ipipe == opipe)
1119 return -EINVAL;
1120
1121 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1122 }
1123
1124 if (ipipe) {
1125 if (off_in)
1126 return -ESPIPE;
1127 if (off_out) {
1128 if (!(out->f_mode & FMODE_PWRITE))
1129 return -EINVAL;
1130 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1131 return -EFAULT;
1132 } else {
1133 offset = out->f_pos;
1134 }
1135
1136 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1137 return -EBADF;
1138
1139 if (unlikely(out->f_flags & O_APPEND))
1140 return -EINVAL;
1141
1142 ret = rw_verify_area(WRITE, out, &offset, len);
1143 if (unlikely(ret < 0))
1144 return ret;
1145
1146 file_start_write(out);
1147 ret = do_splice_from(ipipe, out, &offset, len, flags);
1148 file_end_write(out);
1149
1150 if (!off_out)
1151 out->f_pos = offset;
1152 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1153 ret = -EFAULT;
1154
1155 return ret;
1156 }
1157
1158 if (opipe) {
1159 if (off_out)
1160 return -ESPIPE;
1161 if (off_in) {
1162 if (!(in->f_mode & FMODE_PREAD))
1163 return -EINVAL;
1164 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1165 return -EFAULT;
1166 } else {
1167 offset = in->f_pos;
1168 }
1169
1170 pipe_lock(opipe);
1171 ret = wait_for_space(opipe, flags);
1172 if (!ret)
1173 ret = do_splice_to(in, &offset, opipe, len, flags);
1174 pipe_unlock(opipe);
1175 if (ret > 0)
1176 wakeup_pipe_readers(opipe);
1177 if (!off_in)
1178 in->f_pos = offset;
1179 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1180 ret = -EFAULT;
1181
1182 return ret;
1183 }
1184
1185 return -EINVAL;
1186}
1187
1188static int iter_to_pipe(struct iov_iter *from,
1189 struct pipe_inode_info *pipe,
1190 unsigned flags)
1191{
1192 struct pipe_buffer buf = {
1193 .ops = &user_page_pipe_buf_ops,
1194 .flags = flags
1195 };
1196 size_t total = 0;
1197 int ret = 0;
1198 bool failed = false;
1199
1200 while (iov_iter_count(from) && !failed) {
1201 struct page *pages[16];
1202 ssize_t copied;
1203 size_t start;
1204 int n;
1205
1206 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1207 if (copied <= 0) {
1208 ret = copied;
1209 break;
1210 }
1211
1212 for (n = 0; copied; n++, start = 0) {
1213 int size = min_t(int, copied, PAGE_SIZE - start);
1214 if (!failed) {
1215 buf.page = pages[n];
1216 buf.offset = start;
1217 buf.len = size;
1218 ret = add_to_pipe(pipe, &buf);
1219 if (unlikely(ret < 0)) {
1220 failed = true;
1221 } else {
1222 iov_iter_advance(from, ret);
1223 total += ret;
1224 }
1225 } else {
1226 put_page(pages[n]);
1227 }
1228 copied -= size;
1229 }
1230 }
1231 return total ? total : ret;
1232}
1233
1234static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1235 struct splice_desc *sd)
1236{
1237 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1238 return n == sd->len ? n : -EFAULT;
1239}
1240
1241/*
1242 * For lack of a better implementation, implement vmsplice() to userspace
1243 * as a simple copy of the pipes pages to the user iov.
1244 */
1245static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1246 unsigned long nr_segs, unsigned int flags)
1247{
1248 struct pipe_inode_info *pipe;
1249 struct splice_desc sd;
1250 long ret;
1251 struct iovec iovstack[UIO_FASTIOV];
1252 struct iovec *iov = iovstack;
1253 struct iov_iter iter;
1254
1255 pipe = get_pipe_info(file);
1256 if (!pipe)
1257 return -EBADF;
1258
1259 ret = import_iovec(READ, uiov, nr_segs,
1260 ARRAY_SIZE(iovstack), &iov, &iter);
1261 if (ret < 0)
1262 return ret;
1263
1264 sd.total_len = iov_iter_count(&iter);
1265 sd.len = 0;
1266 sd.flags = flags;
1267 sd.u.data = &iter;
1268 sd.pos = 0;
1269
1270 if (sd.total_len) {
1271 pipe_lock(pipe);
1272 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1273 pipe_unlock(pipe);
1274 }
1275
1276 kfree(iov);
1277 return ret;
1278}
1279
1280/*
1281 * vmsplice splices a user address range into a pipe. It can be thought of
1282 * as splice-from-memory, where the regular splice is splice-from-file (or
1283 * to file). In both cases the output is a pipe, naturally.
1284 */
1285static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1286 unsigned long nr_segs, unsigned int flags)
1287{
1288 struct pipe_inode_info *pipe;
1289 struct iovec iovstack[UIO_FASTIOV];
1290 struct iovec *iov = iovstack;
1291 struct iov_iter from;
1292 long ret;
1293 unsigned buf_flag = 0;
1294
1295 if (flags & SPLICE_F_GIFT)
1296 buf_flag = PIPE_BUF_FLAG_GIFT;
1297
1298 pipe = get_pipe_info(file);
1299 if (!pipe)
1300 return -EBADF;
1301
1302 ret = import_iovec(WRITE, uiov, nr_segs,
1303 ARRAY_SIZE(iovstack), &iov, &from);
1304 if (ret < 0)
1305 return ret;
1306
1307 pipe_lock(pipe);
1308 ret = wait_for_space(pipe, flags);
1309 if (!ret)
1310 ret = iter_to_pipe(&from, pipe, buf_flag);
1311 pipe_unlock(pipe);
1312 if (ret > 0)
1313 wakeup_pipe_readers(pipe);
1314 kfree(iov);
1315 return ret;
1316}
1317
1318/*
1319 * Note that vmsplice only really supports true splicing _from_ user memory
1320 * to a pipe, not the other way around. Splicing from user memory is a simple
1321 * operation that can be supported without any funky alignment restrictions
1322 * or nasty vm tricks. We simply map in the user memory and fill them into
1323 * a pipe. The reverse isn't quite as easy, though. There are two possible
1324 * solutions for that:
1325 *
1326 * - memcpy() the data internally, at which point we might as well just
1327 * do a regular read() on the buffer anyway.
1328 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1329 * has restriction limitations on both ends of the pipe).
1330 *
1331 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1332 *
1333 */
1334static long do_vmsplice(int fd, const struct iovec __user *iov,
1335 unsigned long nr_segs, unsigned int flags)
1336{
1337 struct fd f;
1338 long error;
1339
1340 if (unlikely(flags & ~SPLICE_F_ALL))
1341 return -EINVAL;
1342 if (unlikely(nr_segs > UIO_MAXIOV))
1343 return -EINVAL;
1344 else if (unlikely(!nr_segs))
1345 return 0;
1346
1347 error = -EBADF;
1348 f = fdget(fd);
1349 if (f.file) {
1350 if (f.file->f_mode & FMODE_WRITE)
1351 error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1352 else if (f.file->f_mode & FMODE_READ)
1353 error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1354
1355 fdput(f);
1356 }
1357
1358 return error;
1359}
1360
1361SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1362 unsigned long, nr_segs, unsigned int, flags)
1363{
1364 return do_vmsplice(fd, iov, nr_segs, flags);
1365}
1366
1367#ifdef CONFIG_COMPAT
1368COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1369 unsigned int, nr_segs, unsigned int, flags)
1370{
1371 unsigned i;
1372 struct iovec __user *iov;
1373 if (nr_segs > UIO_MAXIOV)
1374 return -EINVAL;
1375 iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1376 for (i = 0; i < nr_segs; i++) {
1377 struct compat_iovec v;
1378 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1379 get_user(v.iov_len, &iov32[i].iov_len) ||
1380 put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1381 put_user(v.iov_len, &iov[i].iov_len))
1382 return -EFAULT;
1383 }
1384 return do_vmsplice(fd, iov, nr_segs, flags);
1385}
1386#endif
1387
1388SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1389 int, fd_out, loff_t __user *, off_out,
1390 size_t, len, unsigned int, flags)
1391{
1392 struct fd in, out;
1393 long error;
1394
1395 if (unlikely(!len))
1396 return 0;
1397
1398 if (unlikely(flags & ~SPLICE_F_ALL))
1399 return -EINVAL;
1400
1401 error = -EBADF;
1402 in = fdget(fd_in);
1403 if (in.file) {
1404 if (in.file->f_mode & FMODE_READ) {
1405 out = fdget(fd_out);
1406 if (out.file) {
1407 if (out.file->f_mode & FMODE_WRITE)
1408 error = do_splice(in.file, off_in,
1409 out.file, off_out,
1410 len, flags);
1411 fdput(out);
1412 }
1413 }
1414 fdput(in);
1415 }
1416 return error;
1417}
1418
1419/*
1420 * Make sure there's data to read. Wait for input if we can, otherwise
1421 * return an appropriate error.
1422 */
1423static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1424{
1425 int ret;
1426
1427 /*
1428 * Check ->nrbufs without the inode lock first. This function
1429 * is speculative anyways, so missing one is ok.
1430 */
1431 if (pipe->nrbufs)
1432 return 0;
1433
1434 ret = 0;
1435 pipe_lock(pipe);
1436
1437 while (!pipe->nrbufs) {
1438 if (signal_pending(current)) {
1439 ret = -ERESTARTSYS;
1440 break;
1441 }
1442 if (!pipe->writers)
1443 break;
1444 if (!pipe->waiting_writers) {
1445 if (flags & SPLICE_F_NONBLOCK) {
1446 ret = -EAGAIN;
1447 break;
1448 }
1449 }
1450 pipe_wait(pipe);
1451 }
1452
1453 pipe_unlock(pipe);
1454 return ret;
1455}
1456
1457/*
1458 * Make sure there's writeable room. Wait for room if we can, otherwise
1459 * return an appropriate error.
1460 */
1461static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1462{
1463 int ret;
1464
1465 /*
1466 * Check ->nrbufs without the inode lock first. This function
1467 * is speculative anyways, so missing one is ok.
1468 */
1469 if (pipe->nrbufs < pipe->buffers)
1470 return 0;
1471
1472 ret = 0;
1473 pipe_lock(pipe);
1474
1475 while (pipe->nrbufs >= pipe->buffers) {
1476 if (!pipe->readers) {
1477 send_sig(SIGPIPE, current, 0);
1478 ret = -EPIPE;
1479 break;
1480 }
1481 if (flags & SPLICE_F_NONBLOCK) {
1482 ret = -EAGAIN;
1483 break;
1484 }
1485 if (signal_pending(current)) {
1486 ret = -ERESTARTSYS;
1487 break;
1488 }
1489 pipe->waiting_writers++;
1490 pipe_wait(pipe);
1491 pipe->waiting_writers--;
1492 }
1493
1494 pipe_unlock(pipe);
1495 return ret;
1496}
1497
1498/*
1499 * Splice contents of ipipe to opipe.
1500 */
1501static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1502 struct pipe_inode_info *opipe,
1503 size_t len, unsigned int flags)
1504{
1505 struct pipe_buffer *ibuf, *obuf;
1506 int ret = 0, nbuf;
1507 bool input_wakeup = false;
1508
1509
1510retry:
1511 ret = ipipe_prep(ipipe, flags);
1512 if (ret)
1513 return ret;
1514
1515 ret = opipe_prep(opipe, flags);
1516 if (ret)
1517 return ret;
1518
1519 /*
1520 * Potential ABBA deadlock, work around it by ordering lock
1521 * grabbing by pipe info address. Otherwise two different processes
1522 * could deadlock (one doing tee from A -> B, the other from B -> A).
1523 */
1524 pipe_double_lock(ipipe, opipe);
1525
1526 do {
1527 if (!opipe->readers) {
1528 send_sig(SIGPIPE, current, 0);
1529 if (!ret)
1530 ret = -EPIPE;
1531 break;
1532 }
1533
1534 if (!ipipe->nrbufs && !ipipe->writers)
1535 break;
1536
1537 /*
1538 * Cannot make any progress, because either the input
1539 * pipe is empty or the output pipe is full.
1540 */
1541 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1542 /* Already processed some buffers, break */
1543 if (ret)
1544 break;
1545
1546 if (flags & SPLICE_F_NONBLOCK) {
1547 ret = -EAGAIN;
1548 break;
1549 }
1550
1551 /*
1552 * We raced with another reader/writer and haven't
1553 * managed to process any buffers. A zero return
1554 * value means EOF, so retry instead.
1555 */
1556 pipe_unlock(ipipe);
1557 pipe_unlock(opipe);
1558 goto retry;
1559 }
1560
1561 ibuf = ipipe->bufs + ipipe->curbuf;
1562 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1563 obuf = opipe->bufs + nbuf;
1564
1565 if (len >= ibuf->len) {
1566 /*
1567 * Simply move the whole buffer from ipipe to opipe
1568 */
1569 *obuf = *ibuf;
1570 ibuf->ops = NULL;
1571 opipe->nrbufs++;
1572 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1573 ipipe->nrbufs--;
1574 input_wakeup = true;
1575 } else {
1576 /*
1577 * Get a reference to this pipe buffer,
1578 * so we can copy the contents over.
1579 */
1580 pipe_buf_get(ipipe, ibuf);
1581 *obuf = *ibuf;
1582
1583 /*
1584 * Don't inherit the gift flag, we need to
1585 * prevent multiple steals of this page.
1586 */
1587 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1588
1589 obuf->len = len;
1590 opipe->nrbufs++;
1591 ibuf->offset += obuf->len;
1592 ibuf->len -= obuf->len;
1593 }
1594 ret += obuf->len;
1595 len -= obuf->len;
1596 } while (len);
1597
1598 pipe_unlock(ipipe);
1599 pipe_unlock(opipe);
1600
1601 /*
1602 * If we put data in the output pipe, wakeup any potential readers.
1603 */
1604 if (ret > 0)
1605 wakeup_pipe_readers(opipe);
1606
1607 if (input_wakeup)
1608 wakeup_pipe_writers(ipipe);
1609
1610 return ret;
1611}
1612
1613/*
1614 * Link contents of ipipe to opipe.
1615 */
1616static int link_pipe(struct pipe_inode_info *ipipe,
1617 struct pipe_inode_info *opipe,
1618 size_t len, unsigned int flags)
1619{
1620 struct pipe_buffer *ibuf, *obuf;
1621 int ret = 0, i = 0, nbuf;
1622
1623 /*
1624 * Potential ABBA deadlock, work around it by ordering lock
1625 * grabbing by pipe info address. Otherwise two different processes
1626 * could deadlock (one doing tee from A -> B, the other from B -> A).
1627 */
1628 pipe_double_lock(ipipe, opipe);
1629
1630 do {
1631 if (!opipe->readers) {
1632 send_sig(SIGPIPE, current, 0);
1633 if (!ret)
1634 ret = -EPIPE;
1635 break;
1636 }
1637
1638 /*
1639 * If we have iterated all input buffers or ran out of
1640 * output room, break.
1641 */
1642 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1643 break;
1644
1645 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1646 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1647
1648 /*
1649 * Get a reference to this pipe buffer,
1650 * so we can copy the contents over.
1651 */
1652 pipe_buf_get(ipipe, ibuf);
1653
1654 obuf = opipe->bufs + nbuf;
1655 *obuf = *ibuf;
1656
1657 /*
1658 * Don't inherit the gift flag, we need to
1659 * prevent multiple steals of this page.
1660 */
1661 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1662
1663 if (obuf->len > len)
1664 obuf->len = len;
1665
1666 opipe->nrbufs++;
1667 ret += obuf->len;
1668 len -= obuf->len;
1669 i++;
1670 } while (len);
1671
1672 /*
1673 * return EAGAIN if we have the potential of some data in the
1674 * future, otherwise just return 0
1675 */
1676 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1677 ret = -EAGAIN;
1678
1679 pipe_unlock(ipipe);
1680 pipe_unlock(opipe);
1681
1682 /*
1683 * If we put data in the output pipe, wakeup any potential readers.
1684 */
1685 if (ret > 0)
1686 wakeup_pipe_readers(opipe);
1687
1688 return ret;
1689}
1690
1691/*
1692 * This is a tee(1) implementation that works on pipes. It doesn't copy
1693 * any data, it simply references the 'in' pages on the 'out' pipe.
1694 * The 'flags' used are the SPLICE_F_* variants, currently the only
1695 * applicable one is SPLICE_F_NONBLOCK.
1696 */
1697static long do_tee(struct file *in, struct file *out, size_t len,
1698 unsigned int flags)
1699{
1700 struct pipe_inode_info *ipipe = get_pipe_info(in);
1701 struct pipe_inode_info *opipe = get_pipe_info(out);
1702 int ret = -EINVAL;
1703
1704 /*
1705 * Duplicate the contents of ipipe to opipe without actually
1706 * copying the data.
1707 */
1708 if (ipipe && opipe && ipipe != opipe) {
1709 /*
1710 * Keep going, unless we encounter an error. The ipipe/opipe
1711 * ordering doesn't really matter.
1712 */
1713 ret = ipipe_prep(ipipe, flags);
1714 if (!ret) {
1715 ret = opipe_prep(opipe, flags);
1716 if (!ret)
1717 ret = link_pipe(ipipe, opipe, len, flags);
1718 }
1719 }
1720
1721 return ret;
1722}
1723
1724SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1725{
1726 struct fd in;
1727 int error;
1728
1729 if (unlikely(flags & ~SPLICE_F_ALL))
1730 return -EINVAL;
1731
1732 if (unlikely(!len))
1733 return 0;
1734
1735 error = -EBADF;
1736 in = fdget(fdin);
1737 if (in.file) {
1738 if (in.file->f_mode & FMODE_READ) {
1739 struct fd out = fdget(fdout);
1740 if (out.file) {
1741 if (out.file->f_mode & FMODE_WRITE)
1742 error = do_tee(in.file, out.file,
1743 len, flags);
1744 fdput(out);
1745 }
1746 }
1747 fdput(in);
1748 }
1749
1750 return error;
1751}
1/*
2 * "splice": joining two ropes together by interweaving their strands.
3 *
4 * This is the "extended pipe" functionality, where a pipe is used as
5 * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6 * buffer that you can use to transfer data from one end to the other.
7 *
8 * The traditional unix read/write is extended with a "splice()" operation
9 * that transfers data buffers to or from a pipe buffer.
10 *
11 * Named by Larry McVoy, original implementation from Linus, extended by
12 * Jens to support splicing to files, network, direct splicing, etc and
13 * fixing lots of bugs.
14 *
15 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18 *
19 */
20#include <linux/fs.h>
21#include <linux/file.h>
22#include <linux/pagemap.h>
23#include <linux/splice.h>
24#include <linux/memcontrol.h>
25#include <linux/mm_inline.h>
26#include <linux/swap.h>
27#include <linux/writeback.h>
28#include <linux/buffer_head.h>
29#include <linux/module.h>
30#include <linux/syscalls.h>
31#include <linux/uio.h>
32#include <linux/security.h>
33#include <linux/gfp.h>
34
35/*
36 * Attempt to steal a page from a pipe buffer. This should perhaps go into
37 * a vm helper function, it's already simplified quite a bit by the
38 * addition of remove_mapping(). If success is returned, the caller may
39 * attempt to reuse this page for another destination.
40 */
41static 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 */
81out_unlock:
82 unlock_page(page);
83 return 1;
84}
85
86static 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 */
97static 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;
130error:
131 unlock_page(page);
132 return err;
133}
134
135const 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
145static 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
155static 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
165static 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 */
184ssize_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
267void 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 */
276int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
277{
278 if (pipe->buffers <= PIPE_DEF_BUFFERS)
279 return 0;
280
281 spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
282 spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
283
284 if (spd->pages && spd->partial)
285 return 0;
286
287 kfree(spd->pages);
288 kfree(spd->partial);
289 return -ENOMEM;
290}
291
292void splice_shrink_spd(struct pipe_inode_info *pipe,
293 struct splice_pipe_desc *spd)
294{
295 if (pipe->buffers <= PIPE_DEF_BUFFERS)
296 return;
297
298 kfree(spd->pages);
299 kfree(spd->partial);
300}
301
302static int
303__generic_file_splice_read(struct file *in, loff_t *ppos,
304 struct pipe_inode_info *pipe, size_t len,
305 unsigned int flags)
306{
307 struct address_space *mapping = in->f_mapping;
308 unsigned int loff, nr_pages, req_pages;
309 struct page *pages[PIPE_DEF_BUFFERS];
310 struct partial_page partial[PIPE_DEF_BUFFERS];
311 struct page *page;
312 pgoff_t index, end_index;
313 loff_t isize;
314 int error, page_nr;
315 struct splice_pipe_desc spd = {
316 .pages = pages,
317 .partial = partial,
318 .flags = flags,
319 .ops = &page_cache_pipe_buf_ops,
320 .spd_release = spd_release_page,
321 };
322
323 if (splice_grow_spd(pipe, &spd))
324 return -ENOMEM;
325
326 index = *ppos >> PAGE_CACHE_SHIFT;
327 loff = *ppos & ~PAGE_CACHE_MASK;
328 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
329 nr_pages = min(req_pages, pipe->buffers);
330
331 /*
332 * Lookup the (hopefully) full range of pages we need.
333 */
334 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
335 index += spd.nr_pages;
336
337 /*
338 * If find_get_pages_contig() returned fewer pages than we needed,
339 * readahead/allocate the rest and fill in the holes.
340 */
341 if (spd.nr_pages < nr_pages)
342 page_cache_sync_readahead(mapping, &in->f_ra, in,
343 index, req_pages - spd.nr_pages);
344
345 error = 0;
346 while (spd.nr_pages < nr_pages) {
347 /*
348 * Page could be there, find_get_pages_contig() breaks on
349 * the first hole.
350 */
351 page = find_get_page(mapping, index);
352 if (!page) {
353 /*
354 * page didn't exist, allocate one.
355 */
356 page = page_cache_alloc_cold(mapping);
357 if (!page)
358 break;
359
360 error = add_to_page_cache_lru(page, mapping, index,
361 GFP_KERNEL);
362 if (unlikely(error)) {
363 page_cache_release(page);
364 if (error == -EEXIST)
365 continue;
366 break;
367 }
368 /*
369 * add_to_page_cache() locks the page, unlock it
370 * to avoid convoluting the logic below even more.
371 */
372 unlock_page(page);
373 }
374
375 spd.pages[spd.nr_pages++] = page;
376 index++;
377 }
378
379 /*
380 * Now loop over the map and see if we need to start IO on any
381 * pages, fill in the partial map, etc.
382 */
383 index = *ppos >> PAGE_CACHE_SHIFT;
384 nr_pages = spd.nr_pages;
385 spd.nr_pages = 0;
386 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
387 unsigned int this_len;
388
389 if (!len)
390 break;
391
392 /*
393 * this_len is the max we'll use from this page
394 */
395 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
396 page = spd.pages[page_nr];
397
398 if (PageReadahead(page))
399 page_cache_async_readahead(mapping, &in->f_ra, in,
400 page, index, req_pages - page_nr);
401
402 /*
403 * If the page isn't uptodate, we may need to start io on it
404 */
405 if (!PageUptodate(page)) {
406 lock_page(page);
407
408 /*
409 * Page was truncated, or invalidated by the
410 * filesystem. Redo the find/create, but this time the
411 * page is kept locked, so there's no chance of another
412 * race with truncate/invalidate.
413 */
414 if (!page->mapping) {
415 unlock_page(page);
416 page = find_or_create_page(mapping, index,
417 mapping_gfp_mask(mapping));
418
419 if (!page) {
420 error = -ENOMEM;
421 break;
422 }
423 page_cache_release(spd.pages[page_nr]);
424 spd.pages[page_nr] = page;
425 }
426 /*
427 * page was already under io and is now done, great
428 */
429 if (PageUptodate(page)) {
430 unlock_page(page);
431 goto fill_it;
432 }
433
434 /*
435 * need to read in the page
436 */
437 error = mapping->a_ops->readpage(in, page);
438 if (unlikely(error)) {
439 /*
440 * We really should re-lookup the page here,
441 * but it complicates things a lot. Instead
442 * lets just do what we already stored, and
443 * we'll get it the next time we are called.
444 */
445 if (error == AOP_TRUNCATED_PAGE)
446 error = 0;
447
448 break;
449 }
450 }
451fill_it:
452 /*
453 * i_size must be checked after PageUptodate.
454 */
455 isize = i_size_read(mapping->host);
456 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
457 if (unlikely(!isize || index > end_index))
458 break;
459
460 /*
461 * if this is the last page, see if we need to shrink
462 * the length and stop
463 */
464 if (end_index == index) {
465 unsigned int plen;
466
467 /*
468 * max good bytes in this page
469 */
470 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
471 if (plen <= loff)
472 break;
473
474 /*
475 * force quit after adding this page
476 */
477 this_len = min(this_len, plen - loff);
478 len = this_len;
479 }
480
481 spd.partial[page_nr].offset = loff;
482 spd.partial[page_nr].len = this_len;
483 len -= this_len;
484 loff = 0;
485 spd.nr_pages++;
486 index++;
487 }
488
489 /*
490 * Release any pages at the end, if we quit early. 'page_nr' is how far
491 * we got, 'nr_pages' is how many pages are in the map.
492 */
493 while (page_nr < nr_pages)
494 page_cache_release(spd.pages[page_nr++]);
495 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
496
497 if (spd.nr_pages)
498 error = splice_to_pipe(pipe, &spd);
499
500 splice_shrink_spd(pipe, &spd);
501 return error;
502}
503
504/**
505 * generic_file_splice_read - splice data from file to a pipe
506 * @in: file to splice from
507 * @ppos: position in @in
508 * @pipe: pipe to splice to
509 * @len: number of bytes to splice
510 * @flags: splice modifier flags
511 *
512 * Description:
513 * Will read pages from given file and fill them into a pipe. Can be
514 * used as long as the address_space operations for the source implements
515 * a readpage() hook.
516 *
517 */
518ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
519 struct pipe_inode_info *pipe, size_t len,
520 unsigned int flags)
521{
522 loff_t isize, left;
523 int ret;
524
525 isize = i_size_read(in->f_mapping->host);
526 if (unlikely(*ppos >= isize))
527 return 0;
528
529 left = isize - *ppos;
530 if (unlikely(left < len))
531 len = left;
532
533 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
534 if (ret > 0) {
535 *ppos += ret;
536 file_accessed(in);
537 }
538
539 return ret;
540}
541EXPORT_SYMBOL(generic_file_splice_read);
542
543static const struct pipe_buf_operations default_pipe_buf_ops = {
544 .can_merge = 0,
545 .map = generic_pipe_buf_map,
546 .unmap = generic_pipe_buf_unmap,
547 .confirm = generic_pipe_buf_confirm,
548 .release = generic_pipe_buf_release,
549 .steal = generic_pipe_buf_steal,
550 .get = generic_pipe_buf_get,
551};
552
553static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
554 unsigned long vlen, loff_t offset)
555{
556 mm_segment_t old_fs;
557 loff_t pos = offset;
558 ssize_t res;
559
560 old_fs = get_fs();
561 set_fs(get_ds());
562 /* The cast to a user pointer is valid due to the set_fs() */
563 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
564 set_fs(old_fs);
565
566 return res;
567}
568
569static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
570 loff_t pos)
571{
572 mm_segment_t old_fs;
573 ssize_t res;
574
575 old_fs = get_fs();
576 set_fs(get_ds());
577 /* The cast to a user pointer is valid due to the set_fs() */
578 res = vfs_write(file, (const char __user *)buf, count, &pos);
579 set_fs(old_fs);
580
581 return res;
582}
583
584ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
585 struct pipe_inode_info *pipe, size_t len,
586 unsigned int flags)
587{
588 unsigned int nr_pages;
589 unsigned int nr_freed;
590 size_t offset;
591 struct page *pages[PIPE_DEF_BUFFERS];
592 struct partial_page partial[PIPE_DEF_BUFFERS];
593 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
594 ssize_t res;
595 size_t this_len;
596 int error;
597 int i;
598 struct splice_pipe_desc spd = {
599 .pages = pages,
600 .partial = partial,
601 .flags = flags,
602 .ops = &default_pipe_buf_ops,
603 .spd_release = spd_release_page,
604 };
605
606 if (splice_grow_spd(pipe, &spd))
607 return -ENOMEM;
608
609 res = -ENOMEM;
610 vec = __vec;
611 if (pipe->buffers > PIPE_DEF_BUFFERS) {
612 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
613 if (!vec)
614 goto shrink_ret;
615 }
616
617 offset = *ppos & ~PAGE_CACHE_MASK;
618 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
619
620 for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
621 struct page *page;
622
623 page = alloc_page(GFP_USER);
624 error = -ENOMEM;
625 if (!page)
626 goto err;
627
628 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
629 vec[i].iov_base = (void __user *) page_address(page);
630 vec[i].iov_len = this_len;
631 spd.pages[i] = page;
632 spd.nr_pages++;
633 len -= this_len;
634 offset = 0;
635 }
636
637 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
638 if (res < 0) {
639 error = res;
640 goto err;
641 }
642
643 error = 0;
644 if (!res)
645 goto err;
646
647 nr_freed = 0;
648 for (i = 0; i < spd.nr_pages; i++) {
649 this_len = min_t(size_t, vec[i].iov_len, res);
650 spd.partial[i].offset = 0;
651 spd.partial[i].len = this_len;
652 if (!this_len) {
653 __free_page(spd.pages[i]);
654 spd.pages[i] = NULL;
655 nr_freed++;
656 }
657 res -= this_len;
658 }
659 spd.nr_pages -= nr_freed;
660
661 res = splice_to_pipe(pipe, &spd);
662 if (res > 0)
663 *ppos += res;
664
665shrink_ret:
666 if (vec != __vec)
667 kfree(vec);
668 splice_shrink_spd(pipe, &spd);
669 return res;
670
671err:
672 for (i = 0; i < spd.nr_pages; i++)
673 __free_page(spd.pages[i]);
674
675 res = error;
676 goto shrink_ret;
677}
678EXPORT_SYMBOL(default_file_splice_read);
679
680/*
681 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
682 * using sendpage(). Return the number of bytes sent.
683 */
684static int pipe_to_sendpage(struct pipe_inode_info *pipe,
685 struct pipe_buffer *buf, struct splice_desc *sd)
686{
687 struct file *file = sd->u.file;
688 loff_t pos = sd->pos;
689 int more;
690
691 if (!likely(file->f_op && file->f_op->sendpage))
692 return -EINVAL;
693
694 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
695 return file->f_op->sendpage(file, buf->page, buf->offset,
696 sd->len, &pos, more);
697}
698
699/*
700 * This is a little more tricky than the file -> pipe splicing. There are
701 * basically three cases:
702 *
703 * - Destination page already exists in the address space and there
704 * are users of it. For that case we have no other option that
705 * copying the data. Tough luck.
706 * - Destination page already exists in the address space, but there
707 * are no users of it. Make sure it's uptodate, then drop it. Fall
708 * through to last case.
709 * - Destination page does not exist, we can add the pipe page to
710 * the page cache and avoid the copy.
711 *
712 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
713 * sd->flags), we attempt to migrate pages from the pipe to the output
714 * file address space page cache. This is possible if no one else has
715 * the pipe page referenced outside of the pipe and page cache. If
716 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
717 * a new page in the output file page cache and fill/dirty that.
718 */
719int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
720 struct splice_desc *sd)
721{
722 struct file *file = sd->u.file;
723 struct address_space *mapping = file->f_mapping;
724 unsigned int offset, this_len;
725 struct page *page;
726 void *fsdata;
727 int ret;
728
729 offset = sd->pos & ~PAGE_CACHE_MASK;
730
731 this_len = sd->len;
732 if (this_len + offset > PAGE_CACHE_SIZE)
733 this_len = PAGE_CACHE_SIZE - offset;
734
735 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
736 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
737 if (unlikely(ret))
738 goto out;
739
740 if (buf->page != page) {
741 /*
742 * Careful, ->map() uses KM_USER0!
743 */
744 char *src = buf->ops->map(pipe, buf, 1);
745 char *dst = kmap_atomic(page, KM_USER1);
746
747 memcpy(dst + offset, src + buf->offset, this_len);
748 flush_dcache_page(page);
749 kunmap_atomic(dst, KM_USER1);
750 buf->ops->unmap(pipe, buf, src);
751 }
752 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
753 page, fsdata);
754out:
755 return ret;
756}
757EXPORT_SYMBOL(pipe_to_file);
758
759static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
760{
761 smp_mb();
762 if (waitqueue_active(&pipe->wait))
763 wake_up_interruptible(&pipe->wait);
764 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
765}
766
767/**
768 * splice_from_pipe_feed - feed available data from a pipe to a file
769 * @pipe: pipe to splice from
770 * @sd: information to @actor
771 * @actor: handler that splices the data
772 *
773 * Description:
774 * This function loops over the pipe and calls @actor to do the
775 * actual moving of a single struct pipe_buffer to the desired
776 * destination. It returns when there's no more buffers left in
777 * the pipe or if the requested number of bytes (@sd->total_len)
778 * have been copied. It returns a positive number (one) if the
779 * pipe needs to be filled with more data, zero if the required
780 * number of bytes have been copied and -errno on error.
781 *
782 * This, together with splice_from_pipe_{begin,end,next}, may be
783 * used to implement the functionality of __splice_from_pipe() when
784 * locking is required around copying the pipe buffers to the
785 * destination.
786 */
787int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
788 splice_actor *actor)
789{
790 int ret;
791
792 while (pipe->nrbufs) {
793 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
794 const struct pipe_buf_operations *ops = buf->ops;
795
796 sd->len = buf->len;
797 if (sd->len > sd->total_len)
798 sd->len = sd->total_len;
799
800 ret = buf->ops->confirm(pipe, buf);
801 if (unlikely(ret)) {
802 if (ret == -ENODATA)
803 ret = 0;
804 return ret;
805 }
806
807 ret = actor(pipe, buf, sd);
808 if (ret <= 0)
809 return ret;
810
811 buf->offset += ret;
812 buf->len -= ret;
813
814 sd->num_spliced += ret;
815 sd->len -= ret;
816 sd->pos += ret;
817 sd->total_len -= ret;
818
819 if (!buf->len) {
820 buf->ops = NULL;
821 ops->release(pipe, buf);
822 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
823 pipe->nrbufs--;
824 if (pipe->inode)
825 sd->need_wakeup = true;
826 }
827
828 if (!sd->total_len)
829 return 0;
830 }
831
832 return 1;
833}
834EXPORT_SYMBOL(splice_from_pipe_feed);
835
836/**
837 * splice_from_pipe_next - wait for some data to splice from
838 * @pipe: pipe to splice from
839 * @sd: information about the splice operation
840 *
841 * Description:
842 * This function will wait for some data and return a positive
843 * value (one) if pipe buffers are available. It will return zero
844 * or -errno if no more data needs to be spliced.
845 */
846int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
847{
848 while (!pipe->nrbufs) {
849 if (!pipe->writers)
850 return 0;
851
852 if (!pipe->waiting_writers && sd->num_spliced)
853 return 0;
854
855 if (sd->flags & SPLICE_F_NONBLOCK)
856 return -EAGAIN;
857
858 if (signal_pending(current))
859 return -ERESTARTSYS;
860
861 if (sd->need_wakeup) {
862 wakeup_pipe_writers(pipe);
863 sd->need_wakeup = false;
864 }
865
866 pipe_wait(pipe);
867 }
868
869 return 1;
870}
871EXPORT_SYMBOL(splice_from_pipe_next);
872
873/**
874 * splice_from_pipe_begin - start splicing from pipe
875 * @sd: information about the splice operation
876 *
877 * Description:
878 * This function should be called before a loop containing
879 * splice_from_pipe_next() and splice_from_pipe_feed() to
880 * initialize the necessary fields of @sd.
881 */
882void splice_from_pipe_begin(struct splice_desc *sd)
883{
884 sd->num_spliced = 0;
885 sd->need_wakeup = false;
886}
887EXPORT_SYMBOL(splice_from_pipe_begin);
888
889/**
890 * splice_from_pipe_end - finish splicing from pipe
891 * @pipe: pipe to splice from
892 * @sd: information about the splice operation
893 *
894 * Description:
895 * This function will wake up pipe writers if necessary. It should
896 * be called after a loop containing splice_from_pipe_next() and
897 * splice_from_pipe_feed().
898 */
899void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
900{
901 if (sd->need_wakeup)
902 wakeup_pipe_writers(pipe);
903}
904EXPORT_SYMBOL(splice_from_pipe_end);
905
906/**
907 * __splice_from_pipe - splice data from a pipe to given actor
908 * @pipe: pipe to splice from
909 * @sd: information to @actor
910 * @actor: handler that splices the data
911 *
912 * Description:
913 * This function does little more than loop over the pipe and call
914 * @actor to do the actual moving of a single struct pipe_buffer to
915 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
916 * pipe_to_user.
917 *
918 */
919ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
920 splice_actor *actor)
921{
922 int ret;
923
924 splice_from_pipe_begin(sd);
925 do {
926 ret = splice_from_pipe_next(pipe, sd);
927 if (ret > 0)
928 ret = splice_from_pipe_feed(pipe, sd, actor);
929 } while (ret > 0);
930 splice_from_pipe_end(pipe, sd);
931
932 return sd->num_spliced ? sd->num_spliced : ret;
933}
934EXPORT_SYMBOL(__splice_from_pipe);
935
936/**
937 * splice_from_pipe - splice data from a pipe to a file
938 * @pipe: pipe to splice from
939 * @out: file to splice to
940 * @ppos: position in @out
941 * @len: how many bytes to splice
942 * @flags: splice modifier flags
943 * @actor: handler that splices the data
944 *
945 * Description:
946 * See __splice_from_pipe. This function locks the pipe inode,
947 * otherwise it's identical to __splice_from_pipe().
948 *
949 */
950ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
951 loff_t *ppos, size_t len, unsigned int flags,
952 splice_actor *actor)
953{
954 ssize_t ret;
955 struct splice_desc sd = {
956 .total_len = len,
957 .flags = flags,
958 .pos = *ppos,
959 .u.file = out,
960 };
961
962 pipe_lock(pipe);
963 ret = __splice_from_pipe(pipe, &sd, actor);
964 pipe_unlock(pipe);
965
966 return ret;
967}
968
969/**
970 * generic_file_splice_write - splice data from a pipe to a file
971 * @pipe: pipe info
972 * @out: file to write to
973 * @ppos: position in @out
974 * @len: number of bytes to splice
975 * @flags: splice modifier flags
976 *
977 * Description:
978 * Will either move or copy pages (determined by @flags options) from
979 * the given pipe inode to the given file.
980 *
981 */
982ssize_t
983generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
984 loff_t *ppos, size_t len, unsigned int flags)
985{
986 struct address_space *mapping = out->f_mapping;
987 struct inode *inode = mapping->host;
988 struct splice_desc sd = {
989 .total_len = len,
990 .flags = flags,
991 .pos = *ppos,
992 .u.file = out,
993 };
994 ssize_t ret;
995
996 pipe_lock(pipe);
997
998 splice_from_pipe_begin(&sd);
999 do {
1000 ret = splice_from_pipe_next(pipe, &sd);
1001 if (ret <= 0)
1002 break;
1003
1004 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1005 ret = file_remove_suid(out);
1006 if (!ret) {
1007 file_update_time(out);
1008 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1009 }
1010 mutex_unlock(&inode->i_mutex);
1011 } while (ret > 0);
1012 splice_from_pipe_end(pipe, &sd);
1013
1014 pipe_unlock(pipe);
1015
1016 if (sd.num_spliced)
1017 ret = sd.num_spliced;
1018
1019 if (ret > 0) {
1020 unsigned long nr_pages;
1021 int err;
1022
1023 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1024
1025 err = generic_write_sync(out, *ppos, ret);
1026 if (err)
1027 ret = err;
1028 else
1029 *ppos += ret;
1030 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1031 }
1032
1033 return ret;
1034}
1035
1036EXPORT_SYMBOL(generic_file_splice_write);
1037
1038static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1039 struct splice_desc *sd)
1040{
1041 int ret;
1042 void *data;
1043
1044 data = buf->ops->map(pipe, buf, 0);
1045 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1046 buf->ops->unmap(pipe, buf, data);
1047
1048 return ret;
1049}
1050
1051static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1052 struct file *out, loff_t *ppos,
1053 size_t len, unsigned int flags)
1054{
1055 ssize_t ret;
1056
1057 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1058 if (ret > 0)
1059 *ppos += ret;
1060
1061 return ret;
1062}
1063
1064/**
1065 * generic_splice_sendpage - splice data from a pipe to a socket
1066 * @pipe: pipe to splice from
1067 * @out: socket to write to
1068 * @ppos: position in @out
1069 * @len: number of bytes to splice
1070 * @flags: splice modifier flags
1071 *
1072 * Description:
1073 * Will send @len bytes from the pipe to a network socket. No data copying
1074 * is involved.
1075 *
1076 */
1077ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1078 loff_t *ppos, size_t len, unsigned int flags)
1079{
1080 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1081}
1082
1083EXPORT_SYMBOL(generic_splice_sendpage);
1084
1085/*
1086 * Attempt to initiate a splice from pipe to file.
1087 */
1088static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1089 loff_t *ppos, size_t len, unsigned int flags)
1090{
1091 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1092 loff_t *, size_t, unsigned int);
1093 int ret;
1094
1095 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1096 return -EBADF;
1097
1098 if (unlikely(out->f_flags & O_APPEND))
1099 return -EINVAL;
1100
1101 ret = rw_verify_area(WRITE, out, ppos, len);
1102 if (unlikely(ret < 0))
1103 return ret;
1104
1105 if (out->f_op && out->f_op->splice_write)
1106 splice_write = out->f_op->splice_write;
1107 else
1108 splice_write = default_file_splice_write;
1109
1110 return splice_write(pipe, out, ppos, len, flags);
1111}
1112
1113/*
1114 * Attempt to initiate a splice from a file to a pipe.
1115 */
1116static long do_splice_to(struct file *in, loff_t *ppos,
1117 struct pipe_inode_info *pipe, size_t len,
1118 unsigned int flags)
1119{
1120 ssize_t (*splice_read)(struct file *, loff_t *,
1121 struct pipe_inode_info *, size_t, unsigned int);
1122 int ret;
1123
1124 if (unlikely(!(in->f_mode & FMODE_READ)))
1125 return -EBADF;
1126
1127 ret = rw_verify_area(READ, in, ppos, len);
1128 if (unlikely(ret < 0))
1129 return ret;
1130
1131 if (in->f_op && in->f_op->splice_read)
1132 splice_read = in->f_op->splice_read;
1133 else
1134 splice_read = default_file_splice_read;
1135
1136 return splice_read(in, ppos, pipe, len, flags);
1137}
1138
1139/**
1140 * splice_direct_to_actor - splices data directly between two non-pipes
1141 * @in: file to splice from
1142 * @sd: actor information on where to splice to
1143 * @actor: handles the data splicing
1144 *
1145 * Description:
1146 * This is a special case helper to splice directly between two
1147 * points, without requiring an explicit pipe. Internally an allocated
1148 * pipe is cached in the process, and reused during the lifetime of
1149 * that process.
1150 *
1151 */
1152ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1153 splice_direct_actor *actor)
1154{
1155 struct pipe_inode_info *pipe;
1156 long ret, bytes;
1157 umode_t i_mode;
1158 size_t len;
1159 int i, flags;
1160
1161 /*
1162 * We require the input being a regular file, as we don't want to
1163 * randomly drop data for eg socket -> socket splicing. Use the
1164 * piped splicing for that!
1165 */
1166 i_mode = in->f_path.dentry->d_inode->i_mode;
1167 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1168 return -EINVAL;
1169
1170 /*
1171 * neither in nor out is a pipe, setup an internal pipe attached to
1172 * 'out' and transfer the wanted data from 'in' to 'out' through that
1173 */
1174 pipe = current->splice_pipe;
1175 if (unlikely(!pipe)) {
1176 pipe = alloc_pipe_info(NULL);
1177 if (!pipe)
1178 return -ENOMEM;
1179
1180 /*
1181 * We don't have an immediate reader, but we'll read the stuff
1182 * out of the pipe right after the splice_to_pipe(). So set
1183 * PIPE_READERS appropriately.
1184 */
1185 pipe->readers = 1;
1186
1187 current->splice_pipe = pipe;
1188 }
1189
1190 /*
1191 * Do the splice.
1192 */
1193 ret = 0;
1194 bytes = 0;
1195 len = sd->total_len;
1196 flags = sd->flags;
1197
1198 /*
1199 * Don't block on output, we have to drain the direct pipe.
1200 */
1201 sd->flags &= ~SPLICE_F_NONBLOCK;
1202
1203 while (len) {
1204 size_t read_len;
1205 loff_t pos = sd->pos, prev_pos = pos;
1206
1207 ret = do_splice_to(in, &pos, pipe, len, flags);
1208 if (unlikely(ret <= 0))
1209 goto out_release;
1210
1211 read_len = ret;
1212 sd->total_len = read_len;
1213
1214 /*
1215 * NOTE: nonblocking mode only applies to the input. We
1216 * must not do the output in nonblocking mode as then we
1217 * could get stuck data in the internal pipe:
1218 */
1219 ret = actor(pipe, sd);
1220 if (unlikely(ret <= 0)) {
1221 sd->pos = prev_pos;
1222 goto out_release;
1223 }
1224
1225 bytes += ret;
1226 len -= ret;
1227 sd->pos = pos;
1228
1229 if (ret < read_len) {
1230 sd->pos = prev_pos + ret;
1231 goto out_release;
1232 }
1233 }
1234
1235done:
1236 pipe->nrbufs = pipe->curbuf = 0;
1237 file_accessed(in);
1238 return bytes;
1239
1240out_release:
1241 /*
1242 * If we did an incomplete transfer we must release
1243 * the pipe buffers in question:
1244 */
1245 for (i = 0; i < pipe->buffers; i++) {
1246 struct pipe_buffer *buf = pipe->bufs + i;
1247
1248 if (buf->ops) {
1249 buf->ops->release(pipe, buf);
1250 buf->ops = NULL;
1251 }
1252 }
1253
1254 if (!bytes)
1255 bytes = ret;
1256
1257 goto done;
1258}
1259EXPORT_SYMBOL(splice_direct_to_actor);
1260
1261static int direct_splice_actor(struct pipe_inode_info *pipe,
1262 struct splice_desc *sd)
1263{
1264 struct file *file = sd->u.file;
1265
1266 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1267 sd->flags);
1268}
1269
1270/**
1271 * do_splice_direct - splices data directly between two files
1272 * @in: file to splice from
1273 * @ppos: input file offset
1274 * @out: file to splice to
1275 * @len: number of bytes to splice
1276 * @flags: splice modifier flags
1277 *
1278 * Description:
1279 * For use by do_sendfile(). splice can easily emulate sendfile, but
1280 * doing it in the application would incur an extra system call
1281 * (splice in + splice out, as compared to just sendfile()). So this helper
1282 * can splice directly through a process-private pipe.
1283 *
1284 */
1285long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1286 size_t len, unsigned int flags)
1287{
1288 struct splice_desc sd = {
1289 .len = len,
1290 .total_len = len,
1291 .flags = flags,
1292 .pos = *ppos,
1293 .u.file = out,
1294 };
1295 long ret;
1296
1297 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1298 if (ret > 0)
1299 *ppos = sd.pos;
1300
1301 return ret;
1302}
1303
1304static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1305 struct pipe_inode_info *opipe,
1306 size_t len, unsigned int flags);
1307
1308/*
1309 * Determine where to splice to/from.
1310 */
1311static long do_splice(struct file *in, loff_t __user *off_in,
1312 struct file *out, loff_t __user *off_out,
1313 size_t len, unsigned int flags)
1314{
1315 struct pipe_inode_info *ipipe;
1316 struct pipe_inode_info *opipe;
1317 loff_t offset, *off;
1318 long ret;
1319
1320 ipipe = get_pipe_info(in);
1321 opipe = get_pipe_info(out);
1322
1323 if (ipipe && opipe) {
1324 if (off_in || off_out)
1325 return -ESPIPE;
1326
1327 if (!(in->f_mode & FMODE_READ))
1328 return -EBADF;
1329
1330 if (!(out->f_mode & FMODE_WRITE))
1331 return -EBADF;
1332
1333 /* Splicing to self would be fun, but... */
1334 if (ipipe == opipe)
1335 return -EINVAL;
1336
1337 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1338 }
1339
1340 if (ipipe) {
1341 if (off_in)
1342 return -ESPIPE;
1343 if (off_out) {
1344 if (!(out->f_mode & FMODE_PWRITE))
1345 return -EINVAL;
1346 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1347 return -EFAULT;
1348 off = &offset;
1349 } else
1350 off = &out->f_pos;
1351
1352 ret = do_splice_from(ipipe, out, off, len, flags);
1353
1354 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1355 ret = -EFAULT;
1356
1357 return ret;
1358 }
1359
1360 if (opipe) {
1361 if (off_out)
1362 return -ESPIPE;
1363 if (off_in) {
1364 if (!(in->f_mode & FMODE_PREAD))
1365 return -EINVAL;
1366 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1367 return -EFAULT;
1368 off = &offset;
1369 } else
1370 off = &in->f_pos;
1371
1372 ret = do_splice_to(in, off, opipe, len, flags);
1373
1374 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1375 ret = -EFAULT;
1376
1377 return ret;
1378 }
1379
1380 return -EINVAL;
1381}
1382
1383/*
1384 * Map an iov into an array of pages and offset/length tupples. With the
1385 * partial_page structure, we can map several non-contiguous ranges into
1386 * our ones pages[] map instead of splitting that operation into pieces.
1387 * Could easily be exported as a generic helper for other users, in which
1388 * case one would probably want to add a 'max_nr_pages' parameter as well.
1389 */
1390static int get_iovec_page_array(const struct iovec __user *iov,
1391 unsigned int nr_vecs, struct page **pages,
1392 struct partial_page *partial, int aligned,
1393 unsigned int pipe_buffers)
1394{
1395 int buffers = 0, error = 0;
1396
1397 while (nr_vecs) {
1398 unsigned long off, npages;
1399 struct iovec entry;
1400 void __user *base;
1401 size_t len;
1402 int i;
1403
1404 error = -EFAULT;
1405 if (copy_from_user(&entry, iov, sizeof(entry)))
1406 break;
1407
1408 base = entry.iov_base;
1409 len = entry.iov_len;
1410
1411 /*
1412 * Sanity check this iovec. 0 read succeeds.
1413 */
1414 error = 0;
1415 if (unlikely(!len))
1416 break;
1417 error = -EFAULT;
1418 if (!access_ok(VERIFY_READ, base, len))
1419 break;
1420
1421 /*
1422 * Get this base offset and number of pages, then map
1423 * in the user pages.
1424 */
1425 off = (unsigned long) base & ~PAGE_MASK;
1426
1427 /*
1428 * If asked for alignment, the offset must be zero and the
1429 * length a multiple of the PAGE_SIZE.
1430 */
1431 error = -EINVAL;
1432 if (aligned && (off || len & ~PAGE_MASK))
1433 break;
1434
1435 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1436 if (npages > pipe_buffers - buffers)
1437 npages = pipe_buffers - buffers;
1438
1439 error = get_user_pages_fast((unsigned long)base, npages,
1440 0, &pages[buffers]);
1441
1442 if (unlikely(error <= 0))
1443 break;
1444
1445 /*
1446 * Fill this contiguous range into the partial page map.
1447 */
1448 for (i = 0; i < error; i++) {
1449 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1450
1451 partial[buffers].offset = off;
1452 partial[buffers].len = plen;
1453
1454 off = 0;
1455 len -= plen;
1456 buffers++;
1457 }
1458
1459 /*
1460 * We didn't complete this iov, stop here since it probably
1461 * means we have to move some of this into a pipe to
1462 * be able to continue.
1463 */
1464 if (len)
1465 break;
1466
1467 /*
1468 * Don't continue if we mapped fewer pages than we asked for,
1469 * or if we mapped the max number of pages that we have
1470 * room for.
1471 */
1472 if (error < npages || buffers == pipe_buffers)
1473 break;
1474
1475 nr_vecs--;
1476 iov++;
1477 }
1478
1479 if (buffers)
1480 return buffers;
1481
1482 return error;
1483}
1484
1485static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1486 struct splice_desc *sd)
1487{
1488 char *src;
1489 int ret;
1490
1491 /*
1492 * See if we can use the atomic maps, by prefaulting in the
1493 * pages and doing an atomic copy
1494 */
1495 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1496 src = buf->ops->map(pipe, buf, 1);
1497 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1498 sd->len);
1499 buf->ops->unmap(pipe, buf, src);
1500 if (!ret) {
1501 ret = sd->len;
1502 goto out;
1503 }
1504 }
1505
1506 /*
1507 * No dice, use slow non-atomic map and copy
1508 */
1509 src = buf->ops->map(pipe, buf, 0);
1510
1511 ret = sd->len;
1512 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1513 ret = -EFAULT;
1514
1515 buf->ops->unmap(pipe, buf, src);
1516out:
1517 if (ret > 0)
1518 sd->u.userptr += ret;
1519 return ret;
1520}
1521
1522/*
1523 * For lack of a better implementation, implement vmsplice() to userspace
1524 * as a simple copy of the pipes pages to the user iov.
1525 */
1526static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1527 unsigned long nr_segs, unsigned int flags)
1528{
1529 struct pipe_inode_info *pipe;
1530 struct splice_desc sd;
1531 ssize_t size;
1532 int error;
1533 long ret;
1534
1535 pipe = get_pipe_info(file);
1536 if (!pipe)
1537 return -EBADF;
1538
1539 pipe_lock(pipe);
1540
1541 error = ret = 0;
1542 while (nr_segs) {
1543 void __user *base;
1544 size_t len;
1545
1546 /*
1547 * Get user address base and length for this iovec.
1548 */
1549 error = get_user(base, &iov->iov_base);
1550 if (unlikely(error))
1551 break;
1552 error = get_user(len, &iov->iov_len);
1553 if (unlikely(error))
1554 break;
1555
1556 /*
1557 * Sanity check this iovec. 0 read succeeds.
1558 */
1559 if (unlikely(!len))
1560 break;
1561 if (unlikely(!base)) {
1562 error = -EFAULT;
1563 break;
1564 }
1565
1566 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1567 error = -EFAULT;
1568 break;
1569 }
1570
1571 sd.len = 0;
1572 sd.total_len = len;
1573 sd.flags = flags;
1574 sd.u.userptr = base;
1575 sd.pos = 0;
1576
1577 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1578 if (size < 0) {
1579 if (!ret)
1580 ret = size;
1581
1582 break;
1583 }
1584
1585 ret += size;
1586
1587 if (size < len)
1588 break;
1589
1590 nr_segs--;
1591 iov++;
1592 }
1593
1594 pipe_unlock(pipe);
1595
1596 if (!ret)
1597 ret = error;
1598
1599 return ret;
1600}
1601
1602/*
1603 * vmsplice splices a user address range into a pipe. It can be thought of
1604 * as splice-from-memory, where the regular splice is splice-from-file (or
1605 * to file). In both cases the output is a pipe, naturally.
1606 */
1607static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1608 unsigned long nr_segs, unsigned int flags)
1609{
1610 struct pipe_inode_info *pipe;
1611 struct page *pages[PIPE_DEF_BUFFERS];
1612 struct partial_page partial[PIPE_DEF_BUFFERS];
1613 struct splice_pipe_desc spd = {
1614 .pages = pages,
1615 .partial = partial,
1616 .flags = flags,
1617 .ops = &user_page_pipe_buf_ops,
1618 .spd_release = spd_release_page,
1619 };
1620 long ret;
1621
1622 pipe = get_pipe_info(file);
1623 if (!pipe)
1624 return -EBADF;
1625
1626 if (splice_grow_spd(pipe, &spd))
1627 return -ENOMEM;
1628
1629 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1630 spd.partial, flags & SPLICE_F_GIFT,
1631 pipe->buffers);
1632 if (spd.nr_pages <= 0)
1633 ret = spd.nr_pages;
1634 else
1635 ret = splice_to_pipe(pipe, &spd);
1636
1637 splice_shrink_spd(pipe, &spd);
1638 return ret;
1639}
1640
1641/*
1642 * Note that vmsplice only really supports true splicing _from_ user memory
1643 * to a pipe, not the other way around. Splicing from user memory is a simple
1644 * operation that can be supported without any funky alignment restrictions
1645 * or nasty vm tricks. We simply map in the user memory and fill them into
1646 * a pipe. The reverse isn't quite as easy, though. There are two possible
1647 * solutions for that:
1648 *
1649 * - memcpy() the data internally, at which point we might as well just
1650 * do a regular read() on the buffer anyway.
1651 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1652 * has restriction limitations on both ends of the pipe).
1653 *
1654 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1655 *
1656 */
1657SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1658 unsigned long, nr_segs, unsigned int, flags)
1659{
1660 struct file *file;
1661 long error;
1662 int fput;
1663
1664 if (unlikely(nr_segs > UIO_MAXIOV))
1665 return -EINVAL;
1666 else if (unlikely(!nr_segs))
1667 return 0;
1668
1669 error = -EBADF;
1670 file = fget_light(fd, &fput);
1671 if (file) {
1672 if (file->f_mode & FMODE_WRITE)
1673 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1674 else if (file->f_mode & FMODE_READ)
1675 error = vmsplice_to_user(file, iov, nr_segs, flags);
1676
1677 fput_light(file, fput);
1678 }
1679
1680 return error;
1681}
1682
1683SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1684 int, fd_out, loff_t __user *, off_out,
1685 size_t, len, unsigned int, flags)
1686{
1687 long error;
1688 struct file *in, *out;
1689 int fput_in, fput_out;
1690
1691 if (unlikely(!len))
1692 return 0;
1693
1694 error = -EBADF;
1695 in = fget_light(fd_in, &fput_in);
1696 if (in) {
1697 if (in->f_mode & FMODE_READ) {
1698 out = fget_light(fd_out, &fput_out);
1699 if (out) {
1700 if (out->f_mode & FMODE_WRITE)
1701 error = do_splice(in, off_in,
1702 out, off_out,
1703 len, flags);
1704 fput_light(out, fput_out);
1705 }
1706 }
1707
1708 fput_light(in, fput_in);
1709 }
1710
1711 return error;
1712}
1713
1714/*
1715 * Make sure there's data to read. Wait for input if we can, otherwise
1716 * return an appropriate error.
1717 */
1718static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1719{
1720 int ret;
1721
1722 /*
1723 * Check ->nrbufs without the inode lock first. This function
1724 * is speculative anyways, so missing one is ok.
1725 */
1726 if (pipe->nrbufs)
1727 return 0;
1728
1729 ret = 0;
1730 pipe_lock(pipe);
1731
1732 while (!pipe->nrbufs) {
1733 if (signal_pending(current)) {
1734 ret = -ERESTARTSYS;
1735 break;
1736 }
1737 if (!pipe->writers)
1738 break;
1739 if (!pipe->waiting_writers) {
1740 if (flags & SPLICE_F_NONBLOCK) {
1741 ret = -EAGAIN;
1742 break;
1743 }
1744 }
1745 pipe_wait(pipe);
1746 }
1747
1748 pipe_unlock(pipe);
1749 return ret;
1750}
1751
1752/*
1753 * Make sure there's writeable room. Wait for room if we can, otherwise
1754 * return an appropriate error.
1755 */
1756static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1757{
1758 int ret;
1759
1760 /*
1761 * Check ->nrbufs without the inode lock first. This function
1762 * is speculative anyways, so missing one is ok.
1763 */
1764 if (pipe->nrbufs < pipe->buffers)
1765 return 0;
1766
1767 ret = 0;
1768 pipe_lock(pipe);
1769
1770 while (pipe->nrbufs >= pipe->buffers) {
1771 if (!pipe->readers) {
1772 send_sig(SIGPIPE, current, 0);
1773 ret = -EPIPE;
1774 break;
1775 }
1776 if (flags & SPLICE_F_NONBLOCK) {
1777 ret = -EAGAIN;
1778 break;
1779 }
1780 if (signal_pending(current)) {
1781 ret = -ERESTARTSYS;
1782 break;
1783 }
1784 pipe->waiting_writers++;
1785 pipe_wait(pipe);
1786 pipe->waiting_writers--;
1787 }
1788
1789 pipe_unlock(pipe);
1790 return ret;
1791}
1792
1793/*
1794 * Splice contents of ipipe to opipe.
1795 */
1796static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1797 struct pipe_inode_info *opipe,
1798 size_t len, unsigned int flags)
1799{
1800 struct pipe_buffer *ibuf, *obuf;
1801 int ret = 0, nbuf;
1802 bool input_wakeup = false;
1803
1804
1805retry:
1806 ret = ipipe_prep(ipipe, flags);
1807 if (ret)
1808 return ret;
1809
1810 ret = opipe_prep(opipe, flags);
1811 if (ret)
1812 return ret;
1813
1814 /*
1815 * Potential ABBA deadlock, work around it by ordering lock
1816 * grabbing by pipe info address. Otherwise two different processes
1817 * could deadlock (one doing tee from A -> B, the other from B -> A).
1818 */
1819 pipe_double_lock(ipipe, opipe);
1820
1821 do {
1822 if (!opipe->readers) {
1823 send_sig(SIGPIPE, current, 0);
1824 if (!ret)
1825 ret = -EPIPE;
1826 break;
1827 }
1828
1829 if (!ipipe->nrbufs && !ipipe->writers)
1830 break;
1831
1832 /*
1833 * Cannot make any progress, because either the input
1834 * pipe is empty or the output pipe is full.
1835 */
1836 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1837 /* Already processed some buffers, break */
1838 if (ret)
1839 break;
1840
1841 if (flags & SPLICE_F_NONBLOCK) {
1842 ret = -EAGAIN;
1843 break;
1844 }
1845
1846 /*
1847 * We raced with another reader/writer and haven't
1848 * managed to process any buffers. A zero return
1849 * value means EOF, so retry instead.
1850 */
1851 pipe_unlock(ipipe);
1852 pipe_unlock(opipe);
1853 goto retry;
1854 }
1855
1856 ibuf = ipipe->bufs + ipipe->curbuf;
1857 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1858 obuf = opipe->bufs + nbuf;
1859
1860 if (len >= ibuf->len) {
1861 /*
1862 * Simply move the whole buffer from ipipe to opipe
1863 */
1864 *obuf = *ibuf;
1865 ibuf->ops = NULL;
1866 opipe->nrbufs++;
1867 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1868 ipipe->nrbufs--;
1869 input_wakeup = true;
1870 } else {
1871 /*
1872 * Get a reference to this pipe buffer,
1873 * so we can copy the contents over.
1874 */
1875 ibuf->ops->get(ipipe, ibuf);
1876 *obuf = *ibuf;
1877
1878 /*
1879 * Don't inherit the gift flag, we need to
1880 * prevent multiple steals of this page.
1881 */
1882 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1883
1884 obuf->len = len;
1885 opipe->nrbufs++;
1886 ibuf->offset += obuf->len;
1887 ibuf->len -= obuf->len;
1888 }
1889 ret += obuf->len;
1890 len -= obuf->len;
1891 } while (len);
1892
1893 pipe_unlock(ipipe);
1894 pipe_unlock(opipe);
1895
1896 /*
1897 * If we put data in the output pipe, wakeup any potential readers.
1898 */
1899 if (ret > 0)
1900 wakeup_pipe_readers(opipe);
1901
1902 if (input_wakeup)
1903 wakeup_pipe_writers(ipipe);
1904
1905 return ret;
1906}
1907
1908/*
1909 * Link contents of ipipe to opipe.
1910 */
1911static int link_pipe(struct pipe_inode_info *ipipe,
1912 struct pipe_inode_info *opipe,
1913 size_t len, unsigned int flags)
1914{
1915 struct pipe_buffer *ibuf, *obuf;
1916 int ret = 0, i = 0, nbuf;
1917
1918 /*
1919 * Potential ABBA deadlock, work around it by ordering lock
1920 * grabbing by pipe info address. Otherwise two different processes
1921 * could deadlock (one doing tee from A -> B, the other from B -> A).
1922 */
1923 pipe_double_lock(ipipe, opipe);
1924
1925 do {
1926 if (!opipe->readers) {
1927 send_sig(SIGPIPE, current, 0);
1928 if (!ret)
1929 ret = -EPIPE;
1930 break;
1931 }
1932
1933 /*
1934 * If we have iterated all input buffers or ran out of
1935 * output room, break.
1936 */
1937 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1938 break;
1939
1940 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1941 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1942
1943 /*
1944 * Get a reference to this pipe buffer,
1945 * so we can copy the contents over.
1946 */
1947 ibuf->ops->get(ipipe, ibuf);
1948
1949 obuf = opipe->bufs + nbuf;
1950 *obuf = *ibuf;
1951
1952 /*
1953 * Don't inherit the gift flag, we need to
1954 * prevent multiple steals of this page.
1955 */
1956 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1957
1958 if (obuf->len > len)
1959 obuf->len = len;
1960
1961 opipe->nrbufs++;
1962 ret += obuf->len;
1963 len -= obuf->len;
1964 i++;
1965 } while (len);
1966
1967 /*
1968 * return EAGAIN if we have the potential of some data in the
1969 * future, otherwise just return 0
1970 */
1971 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1972 ret = -EAGAIN;
1973
1974 pipe_unlock(ipipe);
1975 pipe_unlock(opipe);
1976
1977 /*
1978 * If we put data in the output pipe, wakeup any potential readers.
1979 */
1980 if (ret > 0)
1981 wakeup_pipe_readers(opipe);
1982
1983 return ret;
1984}
1985
1986/*
1987 * This is a tee(1) implementation that works on pipes. It doesn't copy
1988 * any data, it simply references the 'in' pages on the 'out' pipe.
1989 * The 'flags' used are the SPLICE_F_* variants, currently the only
1990 * applicable one is SPLICE_F_NONBLOCK.
1991 */
1992static long do_tee(struct file *in, struct file *out, size_t len,
1993 unsigned int flags)
1994{
1995 struct pipe_inode_info *ipipe = get_pipe_info(in);
1996 struct pipe_inode_info *opipe = get_pipe_info(out);
1997 int ret = -EINVAL;
1998
1999 /*
2000 * Duplicate the contents of ipipe to opipe without actually
2001 * copying the data.
2002 */
2003 if (ipipe && opipe && ipipe != opipe) {
2004 /*
2005 * Keep going, unless we encounter an error. The ipipe/opipe
2006 * ordering doesn't really matter.
2007 */
2008 ret = ipipe_prep(ipipe, flags);
2009 if (!ret) {
2010 ret = opipe_prep(opipe, flags);
2011 if (!ret)
2012 ret = link_pipe(ipipe, opipe, len, flags);
2013 }
2014 }
2015
2016 return ret;
2017}
2018
2019SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2020{
2021 struct file *in;
2022 int error, fput_in;
2023
2024 if (unlikely(!len))
2025 return 0;
2026
2027 error = -EBADF;
2028 in = fget_light(fdin, &fput_in);
2029 if (in) {
2030 if (in->f_mode & FMODE_READ) {
2031 int fput_out;
2032 struct file *out = fget_light(fdout, &fput_out);
2033
2034 if (out) {
2035 if (out->f_mode & FMODE_WRITE)
2036 error = do_tee(in, out, len, flags);
2037 fput_light(out, fput_out);
2038 }
2039 }
2040 fput_light(in, fput_in);
2041 }
2042
2043 return error;
2044}