NFS: Add helper functions for allocating filehandles and fattr structs
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/memcontrol.h>
25 #include <linux/mm_inline.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
33 #include <linux/gfp.h>
34
35 /*
36  * Attempt to steal a page from a pipe buffer. This should perhaps go into
37  * a vm helper function, it's already simplified quite a bit by the
38  * addition of remove_mapping(). If success is returned, the caller may
39  * attempt to reuse this page for another destination.
40  */
41 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
42                                      struct pipe_buffer *buf)
43 {
44         struct page *page = buf->page;
45         struct address_space *mapping;
46
47         lock_page(page);
48
49         mapping = page_mapping(page);
50         if (mapping) {
51                 WARN_ON(!PageUptodate(page));
52
53                 /*
54                  * At least for ext2 with nobh option, we need to wait on
55                  * writeback completing on this page, since we'll remove it
56                  * from the pagecache.  Otherwise truncate wont wait on the
57                  * page, allowing the disk blocks to be reused by someone else
58                  * before we actually wrote our data to them. fs corruption
59                  * ensues.
60                  */
61                 wait_on_page_writeback(page);
62
63                 if (page_has_private(page) &&
64                     !try_to_release_page(page, GFP_KERNEL))
65                         goto out_unlock;
66
67                 /*
68                  * If we succeeded in removing the mapping, set LRU flag
69                  * and return good.
70                  */
71                 if (remove_mapping(mapping, page)) {
72                         buf->flags |= PIPE_BUF_FLAG_LRU;
73                         return 0;
74                 }
75         }
76
77         /*
78          * Raced with truncate or failed to remove page from current
79          * address space, unlock and return failure.
80          */
81 out_unlock:
82         unlock_page(page);
83         return 1;
84 }
85
86 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
87                                         struct pipe_buffer *buf)
88 {
89         page_cache_release(buf->page);
90         buf->flags &= ~PIPE_BUF_FLAG_LRU;
91 }
92
93 /*
94  * Check whether the contents of buf is OK to access. Since the content
95  * is a page cache page, IO may be in flight.
96  */
97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
98                                        struct pipe_buffer *buf)
99 {
100         struct page *page = buf->page;
101         int err;
102
103         if (!PageUptodate(page)) {
104                 lock_page(page);
105
106                 /*
107                  * Page got truncated/unhashed. This will cause a 0-byte
108                  * splice, if this is the first page.
109                  */
110                 if (!page->mapping) {
111                         err = -ENODATA;
112                         goto error;
113                 }
114
115                 /*
116                  * Uh oh, read-error from disk.
117                  */
118                 if (!PageUptodate(page)) {
119                         err = -EIO;
120                         goto error;
121                 }
122
123                 /*
124                  * Page is ok afterall, we are done.
125                  */
126                 unlock_page(page);
127         }
128
129         return 0;
130 error:
131         unlock_page(page);
132         return err;
133 }
134
135 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
136         .can_merge = 0,
137         .map = generic_pipe_buf_map,
138         .unmap = generic_pipe_buf_unmap,
139         .confirm = page_cache_pipe_buf_confirm,
140         .release = page_cache_pipe_buf_release,
141         .steal = page_cache_pipe_buf_steal,
142         .get = generic_pipe_buf_get,
143 };
144
145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
146                                     struct pipe_buffer *buf)
147 {
148         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
149                 return 1;
150
151         buf->flags |= PIPE_BUF_FLAG_LRU;
152         return generic_pipe_buf_steal(pipe, buf);
153 }
154
155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
156         .can_merge = 0,
157         .map = generic_pipe_buf_map,
158         .unmap = generic_pipe_buf_unmap,
159         .confirm = generic_pipe_buf_confirm,
160         .release = page_cache_pipe_buf_release,
161         .steal = user_page_pipe_buf_steal,
162         .get = generic_pipe_buf_get,
163 };
164
165 /**
166  * splice_to_pipe - fill passed data into a pipe
167  * @pipe:       pipe to fill
168  * @spd:        data to fill
169  *
170  * Description:
171  *    @spd contains a map of pages and len/offset tuples, along with
172  *    the struct pipe_buf_operations associated with these pages. This
173  *    function will link that data to the pipe.
174  *
175  */
176 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
177                        struct splice_pipe_desc *spd)
178 {
179         unsigned int spd_pages = spd->nr_pages;
180         int ret, do_wakeup, page_nr;
181
182         ret = 0;
183         do_wakeup = 0;
184         page_nr = 0;
185
186         pipe_lock(pipe);
187
188         for (;;) {
189                 if (!pipe->readers) {
190                         send_sig(SIGPIPE, current, 0);
191                         if (!ret)
192                                 ret = -EPIPE;
193                         break;
194                 }
195
196                 if (pipe->nrbufs < PIPE_BUFFERS) {
197                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
198                         struct pipe_buffer *buf = pipe->bufs + newbuf;
199
200                         buf->page = spd->pages[page_nr];
201                         buf->offset = spd->partial[page_nr].offset;
202                         buf->len = spd->partial[page_nr].len;
203                         buf->private = spd->partial[page_nr].private;
204                         buf->ops = spd->ops;
205                         if (spd->flags & SPLICE_F_GIFT)
206                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
207
208                         pipe->nrbufs++;
209                         page_nr++;
210                         ret += buf->len;
211
212                         if (pipe->inode)
213                                 do_wakeup = 1;
214
215                         if (!--spd->nr_pages)
216                                 break;
217                         if (pipe->nrbufs < PIPE_BUFFERS)
218                                 continue;
219
220                         break;
221                 }
222
223                 if (spd->flags & SPLICE_F_NONBLOCK) {
224                         if (!ret)
225                                 ret = -EAGAIN;
226                         break;
227                 }
228
229                 if (signal_pending(current)) {
230                         if (!ret)
231                                 ret = -ERESTARTSYS;
232                         break;
233                 }
234
235                 if (do_wakeup) {
236                         smp_mb();
237                         if (waitqueue_active(&pipe->wait))
238                                 wake_up_interruptible_sync(&pipe->wait);
239                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
240                         do_wakeup = 0;
241                 }
242
243                 pipe->waiting_writers++;
244                 pipe_wait(pipe);
245                 pipe->waiting_writers--;
246         }
247
248         pipe_unlock(pipe);
249
250         if (do_wakeup) {
251                 smp_mb();
252                 if (waitqueue_active(&pipe->wait))
253                         wake_up_interruptible(&pipe->wait);
254                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
255         }
256
257         while (page_nr < spd_pages)
258                 spd->spd_release(spd, page_nr++);
259
260         return ret;
261 }
262
263 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
264 {
265         page_cache_release(spd->pages[i]);
266 }
267
268 static int
269 __generic_file_splice_read(struct file *in, loff_t *ppos,
270                            struct pipe_inode_info *pipe, size_t len,
271                            unsigned int flags)
272 {
273         struct address_space *mapping = in->f_mapping;
274         unsigned int loff, nr_pages, req_pages;
275         struct page *pages[PIPE_BUFFERS];
276         struct partial_page partial[PIPE_BUFFERS];
277         struct page *page;
278         pgoff_t index, end_index;
279         loff_t isize;
280         int error, page_nr;
281         struct splice_pipe_desc spd = {
282                 .pages = pages,
283                 .partial = partial,
284                 .flags = flags,
285                 .ops = &page_cache_pipe_buf_ops,
286                 .spd_release = spd_release_page,
287         };
288
289         index = *ppos >> PAGE_CACHE_SHIFT;
290         loff = *ppos & ~PAGE_CACHE_MASK;
291         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
292         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
293
294         /*
295          * Lookup the (hopefully) full range of pages we need.
296          */
297         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
298         index += spd.nr_pages;
299
300         /*
301          * If find_get_pages_contig() returned fewer pages than we needed,
302          * readahead/allocate the rest and fill in the holes.
303          */
304         if (spd.nr_pages < nr_pages)
305                 page_cache_sync_readahead(mapping, &in->f_ra, in,
306                                 index, req_pages - spd.nr_pages);
307
308         error = 0;
309         while (spd.nr_pages < nr_pages) {
310                 /*
311                  * Page could be there, find_get_pages_contig() breaks on
312                  * the first hole.
313                  */
314                 page = find_get_page(mapping, index);
315                 if (!page) {
316                         /*
317                          * page didn't exist, allocate one.
318                          */
319                         page = page_cache_alloc_cold(mapping);
320                         if (!page)
321                                 break;
322
323                         error = add_to_page_cache_lru(page, mapping, index,
324                                                 mapping_gfp_mask(mapping));
325                         if (unlikely(error)) {
326                                 page_cache_release(page);
327                                 if (error == -EEXIST)
328                                         continue;
329                                 break;
330                         }
331                         /*
332                          * add_to_page_cache() locks the page, unlock it
333                          * to avoid convoluting the logic below even more.
334                          */
335                         unlock_page(page);
336                 }
337
338                 pages[spd.nr_pages++] = page;
339                 index++;
340         }
341
342         /*
343          * Now loop over the map and see if we need to start IO on any
344          * pages, fill in the partial map, etc.
345          */
346         index = *ppos >> PAGE_CACHE_SHIFT;
347         nr_pages = spd.nr_pages;
348         spd.nr_pages = 0;
349         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
350                 unsigned int this_len;
351
352                 if (!len)
353                         break;
354
355                 /*
356                  * this_len is the max we'll use from this page
357                  */
358                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
359                 page = pages[page_nr];
360
361                 if (PageReadahead(page))
362                         page_cache_async_readahead(mapping, &in->f_ra, in,
363                                         page, index, req_pages - page_nr);
364
365                 /*
366                  * If the page isn't uptodate, we may need to start io on it
367                  */
368                 if (!PageUptodate(page)) {
369                         /*
370                          * If in nonblock mode then dont block on waiting
371                          * for an in-flight io page
372                          */
373                         if (flags & SPLICE_F_NONBLOCK) {
374                                 if (!trylock_page(page)) {
375                                         error = -EAGAIN;
376                                         break;
377                                 }
378                         } else
379                                 lock_page(page);
380
381                         /*
382                          * Page was truncated, or invalidated by the
383                          * filesystem.  Redo the find/create, but this time the
384                          * page is kept locked, so there's no chance of another
385                          * race with truncate/invalidate.
386                          */
387                         if (!page->mapping) {
388                                 unlock_page(page);
389                                 page = find_or_create_page(mapping, index,
390                                                 mapping_gfp_mask(mapping));
391
392                                 if (!page) {
393                                         error = -ENOMEM;
394                                         break;
395                                 }
396                                 page_cache_release(pages[page_nr]);
397                                 pages[page_nr] = page;
398                         }
399                         /*
400                          * page was already under io and is now done, great
401                          */
402                         if (PageUptodate(page)) {
403                                 unlock_page(page);
404                                 goto fill_it;
405                         }
406
407                         /*
408                          * need to read in the page
409                          */
410                         error = mapping->a_ops->readpage(in, page);
411                         if (unlikely(error)) {
412                                 /*
413                                  * We really should re-lookup the page here,
414                                  * but it complicates things a lot. Instead
415                                  * lets just do what we already stored, and
416                                  * we'll get it the next time we are called.
417                                  */
418                                 if (error == AOP_TRUNCATED_PAGE)
419                                         error = 0;
420
421                                 break;
422                         }
423                 }
424 fill_it:
425                 /*
426                  * i_size must be checked after PageUptodate.
427                  */
428                 isize = i_size_read(mapping->host);
429                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
430                 if (unlikely(!isize || index > end_index))
431                         break;
432
433                 /*
434                  * if this is the last page, see if we need to shrink
435                  * the length and stop
436                  */
437                 if (end_index == index) {
438                         unsigned int plen;
439
440                         /*
441                          * max good bytes in this page
442                          */
443                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
444                         if (plen <= loff)
445                                 break;
446
447                         /*
448                          * force quit after adding this page
449                          */
450                         this_len = min(this_len, plen - loff);
451                         len = this_len;
452                 }
453
454                 partial[page_nr].offset = loff;
455                 partial[page_nr].len = this_len;
456                 len -= this_len;
457                 loff = 0;
458                 spd.nr_pages++;
459                 index++;
460         }
461
462         /*
463          * Release any pages at the end, if we quit early. 'page_nr' is how far
464          * we got, 'nr_pages' is how many pages are in the map.
465          */
466         while (page_nr < nr_pages)
467                 page_cache_release(pages[page_nr++]);
468         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
469
470         if (spd.nr_pages)
471                 return splice_to_pipe(pipe, &spd);
472
473         return error;
474 }
475
476 /**
477  * generic_file_splice_read - splice data from file to a pipe
478  * @in:         file to splice from
479  * @ppos:       position in @in
480  * @pipe:       pipe to splice to
481  * @len:        number of bytes to splice
482  * @flags:      splice modifier flags
483  *
484  * Description:
485  *    Will read pages from given file and fill them into a pipe. Can be
486  *    used as long as the address_space operations for the source implements
487  *    a readpage() hook.
488  *
489  */
490 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
491                                  struct pipe_inode_info *pipe, size_t len,
492                                  unsigned int flags)
493 {
494         loff_t isize, left;
495         int ret;
496
497         isize = i_size_read(in->f_mapping->host);
498         if (unlikely(*ppos >= isize))
499                 return 0;
500
501         left = isize - *ppos;
502         if (unlikely(left < len))
503                 len = left;
504
505         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
506         if (ret > 0) {
507                 *ppos += ret;
508                 file_accessed(in);
509         }
510
511         return ret;
512 }
513 EXPORT_SYMBOL(generic_file_splice_read);
514
515 static const struct pipe_buf_operations default_pipe_buf_ops = {
516         .can_merge = 0,
517         .map = generic_pipe_buf_map,
518         .unmap = generic_pipe_buf_unmap,
519         .confirm = generic_pipe_buf_confirm,
520         .release = generic_pipe_buf_release,
521         .steal = generic_pipe_buf_steal,
522         .get = generic_pipe_buf_get,
523 };
524
525 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
526                             unsigned long vlen, loff_t offset)
527 {
528         mm_segment_t old_fs;
529         loff_t pos = offset;
530         ssize_t res;
531
532         old_fs = get_fs();
533         set_fs(get_ds());
534         /* The cast to a user pointer is valid due to the set_fs() */
535         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
536         set_fs(old_fs);
537
538         return res;
539 }
540
541 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
542                             loff_t pos)
543 {
544         mm_segment_t old_fs;
545         ssize_t res;
546
547         old_fs = get_fs();
548         set_fs(get_ds());
549         /* The cast to a user pointer is valid due to the set_fs() */
550         res = vfs_write(file, (const char __user *)buf, count, &pos);
551         set_fs(old_fs);
552
553         return res;
554 }
555
556 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
557                                  struct pipe_inode_info *pipe, size_t len,
558                                  unsigned int flags)
559 {
560         unsigned int nr_pages;
561         unsigned int nr_freed;
562         size_t offset;
563         struct page *pages[PIPE_BUFFERS];
564         struct partial_page partial[PIPE_BUFFERS];
565         struct iovec vec[PIPE_BUFFERS];
566         pgoff_t index;
567         ssize_t res;
568         size_t this_len;
569         int error;
570         int i;
571         struct splice_pipe_desc spd = {
572                 .pages = pages,
573                 .partial = partial,
574                 .flags = flags,
575                 .ops = &default_pipe_buf_ops,
576                 .spd_release = spd_release_page,
577         };
578
579         index = *ppos >> PAGE_CACHE_SHIFT;
580         offset = *ppos & ~PAGE_CACHE_MASK;
581         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
582
583         for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
584                 struct page *page;
585
586                 page = alloc_page(GFP_USER);
587                 error = -ENOMEM;
588                 if (!page)
589                         goto err;
590
591                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
592                 vec[i].iov_base = (void __user *) page_address(page);
593                 vec[i].iov_len = this_len;
594                 pages[i] = page;
595                 spd.nr_pages++;
596                 len -= this_len;
597                 offset = 0;
598         }
599
600         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
601         if (res < 0) {
602                 error = res;
603                 goto err;
604         }
605
606         error = 0;
607         if (!res)
608                 goto err;
609
610         nr_freed = 0;
611         for (i = 0; i < spd.nr_pages; i++) {
612                 this_len = min_t(size_t, vec[i].iov_len, res);
613                 partial[i].offset = 0;
614                 partial[i].len = this_len;
615                 if (!this_len) {
616                         __free_page(pages[i]);
617                         pages[i] = NULL;
618                         nr_freed++;
619                 }
620                 res -= this_len;
621         }
622         spd.nr_pages -= nr_freed;
623
624         res = splice_to_pipe(pipe, &spd);
625         if (res > 0)
626                 *ppos += res;
627
628         return res;
629
630 err:
631         for (i = 0; i < spd.nr_pages; i++)
632                 __free_page(pages[i]);
633
634         return error;
635 }
636 EXPORT_SYMBOL(default_file_splice_read);
637
638 /*
639  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
640  * using sendpage(). Return the number of bytes sent.
641  */
642 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
643                             struct pipe_buffer *buf, struct splice_desc *sd)
644 {
645         struct file *file = sd->u.file;
646         loff_t pos = sd->pos;
647         int ret, more;
648
649         ret = buf->ops->confirm(pipe, buf);
650         if (!ret) {
651                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
652                 if (file->f_op && file->f_op->sendpage)
653                         ret = file->f_op->sendpage(file, buf->page, buf->offset,
654                                                    sd->len, &pos, more);
655                 else
656                         ret = -EINVAL;
657         }
658
659         return ret;
660 }
661
662 /*
663  * This is a little more tricky than the file -> pipe splicing. There are
664  * basically three cases:
665  *
666  *      - Destination page already exists in the address space and there
667  *        are users of it. For that case we have no other option that
668  *        copying the data. Tough luck.
669  *      - Destination page already exists in the address space, but there
670  *        are no users of it. Make sure it's uptodate, then drop it. Fall
671  *        through to last case.
672  *      - Destination page does not exist, we can add the pipe page to
673  *        the page cache and avoid the copy.
674  *
675  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
676  * sd->flags), we attempt to migrate pages from the pipe to the output
677  * file address space page cache. This is possible if no one else has
678  * the pipe page referenced outside of the pipe and page cache. If
679  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
680  * a new page in the output file page cache and fill/dirty that.
681  */
682 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
683                  struct splice_desc *sd)
684 {
685         struct file *file = sd->u.file;
686         struct address_space *mapping = file->f_mapping;
687         unsigned int offset, this_len;
688         struct page *page;
689         void *fsdata;
690         int ret;
691
692         /*
693          * make sure the data in this buffer is uptodate
694          */
695         ret = buf->ops->confirm(pipe, buf);
696         if (unlikely(ret))
697                 return ret;
698
699         offset = sd->pos & ~PAGE_CACHE_MASK;
700
701         this_len = sd->len;
702         if (this_len + offset > PAGE_CACHE_SIZE)
703                 this_len = PAGE_CACHE_SIZE - offset;
704
705         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
706                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
707         if (unlikely(ret))
708                 goto out;
709
710         if (buf->page != page) {
711                 /*
712                  * Careful, ->map() uses KM_USER0!
713                  */
714                 char *src = buf->ops->map(pipe, buf, 1);
715                 char *dst = kmap_atomic(page, KM_USER1);
716
717                 memcpy(dst + offset, src + buf->offset, this_len);
718                 flush_dcache_page(page);
719                 kunmap_atomic(dst, KM_USER1);
720                 buf->ops->unmap(pipe, buf, src);
721         }
722         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
723                                 page, fsdata);
724 out:
725         return ret;
726 }
727 EXPORT_SYMBOL(pipe_to_file);
728
729 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
730 {
731         smp_mb();
732         if (waitqueue_active(&pipe->wait))
733                 wake_up_interruptible(&pipe->wait);
734         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
735 }
736
737 /**
738  * splice_from_pipe_feed - feed available data from a pipe to a file
739  * @pipe:       pipe to splice from
740  * @sd:         information to @actor
741  * @actor:      handler that splices the data
742  *
743  * Description:
744  *    This function loops over the pipe and calls @actor to do the
745  *    actual moving of a single struct pipe_buffer to the desired
746  *    destination.  It returns when there's no more buffers left in
747  *    the pipe or if the requested number of bytes (@sd->total_len)
748  *    have been copied.  It returns a positive number (one) if the
749  *    pipe needs to be filled with more data, zero if the required
750  *    number of bytes have been copied and -errno on error.
751  *
752  *    This, together with splice_from_pipe_{begin,end,next}, may be
753  *    used to implement the functionality of __splice_from_pipe() when
754  *    locking is required around copying the pipe buffers to the
755  *    destination.
756  */
757 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
758                           splice_actor *actor)
759 {
760         int ret;
761
762         while (pipe->nrbufs) {
763                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
764                 const struct pipe_buf_operations *ops = buf->ops;
765
766                 sd->len = buf->len;
767                 if (sd->len > sd->total_len)
768                         sd->len = sd->total_len;
769
770                 ret = actor(pipe, buf, sd);
771                 if (ret <= 0) {
772                         if (ret == -ENODATA)
773                                 ret = 0;
774                         return ret;
775                 }
776                 buf->offset += ret;
777                 buf->len -= ret;
778
779                 sd->num_spliced += ret;
780                 sd->len -= ret;
781                 sd->pos += ret;
782                 sd->total_len -= ret;
783
784                 if (!buf->len) {
785                         buf->ops = NULL;
786                         ops->release(pipe, buf);
787                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
788                         pipe->nrbufs--;
789                         if (pipe->inode)
790                                 sd->need_wakeup = true;
791                 }
792
793                 if (!sd->total_len)
794                         return 0;
795         }
796
797         return 1;
798 }
799 EXPORT_SYMBOL(splice_from_pipe_feed);
800
801 /**
802  * splice_from_pipe_next - wait for some data to splice from
803  * @pipe:       pipe to splice from
804  * @sd:         information about the splice operation
805  *
806  * Description:
807  *    This function will wait for some data and return a positive
808  *    value (one) if pipe buffers are available.  It will return zero
809  *    or -errno if no more data needs to be spliced.
810  */
811 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
812 {
813         while (!pipe->nrbufs) {
814                 if (!pipe->writers)
815                         return 0;
816
817                 if (!pipe->waiting_writers && sd->num_spliced)
818                         return 0;
819
820                 if (sd->flags & SPLICE_F_NONBLOCK)
821                         return -EAGAIN;
822
823                 if (signal_pending(current))
824                         return -ERESTARTSYS;
825
826                 if (sd->need_wakeup) {
827                         wakeup_pipe_writers(pipe);
828                         sd->need_wakeup = false;
829                 }
830
831                 pipe_wait(pipe);
832         }
833
834         return 1;
835 }
836 EXPORT_SYMBOL(splice_from_pipe_next);
837
838 /**
839  * splice_from_pipe_begin - start splicing from pipe
840  * @sd:         information about the splice operation
841  *
842  * Description:
843  *    This function should be called before a loop containing
844  *    splice_from_pipe_next() and splice_from_pipe_feed() to
845  *    initialize the necessary fields of @sd.
846  */
847 void splice_from_pipe_begin(struct splice_desc *sd)
848 {
849         sd->num_spliced = 0;
850         sd->need_wakeup = false;
851 }
852 EXPORT_SYMBOL(splice_from_pipe_begin);
853
854 /**
855  * splice_from_pipe_end - finish splicing from pipe
856  * @pipe:       pipe to splice from
857  * @sd:         information about the splice operation
858  *
859  * Description:
860  *    This function will wake up pipe writers if necessary.  It should
861  *    be called after a loop containing splice_from_pipe_next() and
862  *    splice_from_pipe_feed().
863  */
864 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
865 {
866         if (sd->need_wakeup)
867                 wakeup_pipe_writers(pipe);
868 }
869 EXPORT_SYMBOL(splice_from_pipe_end);
870
871 /**
872  * __splice_from_pipe - splice data from a pipe to given actor
873  * @pipe:       pipe to splice from
874  * @sd:         information to @actor
875  * @actor:      handler that splices the data
876  *
877  * Description:
878  *    This function does little more than loop over the pipe and call
879  *    @actor to do the actual moving of a single struct pipe_buffer to
880  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
881  *    pipe_to_user.
882  *
883  */
884 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
885                            splice_actor *actor)
886 {
887         int ret;
888
889         splice_from_pipe_begin(sd);
890         do {
891                 ret = splice_from_pipe_next(pipe, sd);
892                 if (ret > 0)
893                         ret = splice_from_pipe_feed(pipe, sd, actor);
894         } while (ret > 0);
895         splice_from_pipe_end(pipe, sd);
896
897         return sd->num_spliced ? sd->num_spliced : ret;
898 }
899 EXPORT_SYMBOL(__splice_from_pipe);
900
901 /**
902  * splice_from_pipe - splice data from a pipe to a file
903  * @pipe:       pipe to splice from
904  * @out:        file to splice to
905  * @ppos:       position in @out
906  * @len:        how many bytes to splice
907  * @flags:      splice modifier flags
908  * @actor:      handler that splices the data
909  *
910  * Description:
911  *    See __splice_from_pipe. This function locks the pipe inode,
912  *    otherwise it's identical to __splice_from_pipe().
913  *
914  */
915 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
916                          loff_t *ppos, size_t len, unsigned int flags,
917                          splice_actor *actor)
918 {
919         ssize_t ret;
920         struct splice_desc sd = {
921                 .total_len = len,
922                 .flags = flags,
923                 .pos = *ppos,
924                 .u.file = out,
925         };
926
927         pipe_lock(pipe);
928         ret = __splice_from_pipe(pipe, &sd, actor);
929         pipe_unlock(pipe);
930
931         return ret;
932 }
933
934 /**
935  * generic_file_splice_write - splice data from a pipe to a file
936  * @pipe:       pipe info
937  * @out:        file to write to
938  * @ppos:       position in @out
939  * @len:        number of bytes to splice
940  * @flags:      splice modifier flags
941  *
942  * Description:
943  *    Will either move or copy pages (determined by @flags options) from
944  *    the given pipe inode to the given file.
945  *
946  */
947 ssize_t
948 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
949                           loff_t *ppos, size_t len, unsigned int flags)
950 {
951         struct address_space *mapping = out->f_mapping;
952         struct inode *inode = mapping->host;
953         struct splice_desc sd = {
954                 .total_len = len,
955                 .flags = flags,
956                 .pos = *ppos,
957                 .u.file = out,
958         };
959         ssize_t ret;
960
961         pipe_lock(pipe);
962
963         splice_from_pipe_begin(&sd);
964         do {
965                 ret = splice_from_pipe_next(pipe, &sd);
966                 if (ret <= 0)
967                         break;
968
969                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
970                 ret = file_remove_suid(out);
971                 if (!ret) {
972                         file_update_time(out);
973                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
974                 }
975                 mutex_unlock(&inode->i_mutex);
976         } while (ret > 0);
977         splice_from_pipe_end(pipe, &sd);
978
979         pipe_unlock(pipe);
980
981         if (sd.num_spliced)
982                 ret = sd.num_spliced;
983
984         if (ret > 0) {
985                 unsigned long nr_pages;
986                 int err;
987
988                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
989
990                 err = generic_write_sync(out, *ppos, ret);
991                 if (err)
992                         ret = err;
993                 else
994                         *ppos += ret;
995                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
996         }
997
998         return ret;
999 }
1000
1001 EXPORT_SYMBOL(generic_file_splice_write);
1002
1003 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1004                           struct splice_desc *sd)
1005 {
1006         int ret;
1007         void *data;
1008
1009         ret = buf->ops->confirm(pipe, buf);
1010         if (ret)
1011                 return ret;
1012
1013         data = buf->ops->map(pipe, buf, 0);
1014         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1015         buf->ops->unmap(pipe, buf, data);
1016
1017         return ret;
1018 }
1019
1020 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1021                                          struct file *out, loff_t *ppos,
1022                                          size_t len, unsigned int flags)
1023 {
1024         ssize_t ret;
1025
1026         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1027         if (ret > 0)
1028                 *ppos += ret;
1029
1030         return ret;
1031 }
1032
1033 /**
1034  * generic_splice_sendpage - splice data from a pipe to a socket
1035  * @pipe:       pipe to splice from
1036  * @out:        socket to write to
1037  * @ppos:       position in @out
1038  * @len:        number of bytes to splice
1039  * @flags:      splice modifier flags
1040  *
1041  * Description:
1042  *    Will send @len bytes from the pipe to a network socket. No data copying
1043  *    is involved.
1044  *
1045  */
1046 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1047                                 loff_t *ppos, size_t len, unsigned int flags)
1048 {
1049         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1050 }
1051
1052 EXPORT_SYMBOL(generic_splice_sendpage);
1053
1054 /*
1055  * Attempt to initiate a splice from pipe to file.
1056  */
1057 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1058                            loff_t *ppos, size_t len, unsigned int flags)
1059 {
1060         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1061                                 loff_t *, size_t, unsigned int);
1062         int ret;
1063
1064         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1065                 return -EBADF;
1066
1067         if (unlikely(out->f_flags & O_APPEND))
1068                 return -EINVAL;
1069
1070         ret = rw_verify_area(WRITE, out, ppos, len);
1071         if (unlikely(ret < 0))
1072                 return ret;
1073
1074         if (out->f_op && out->f_op->splice_write)
1075                 splice_write = out->f_op->splice_write;
1076         else
1077                 splice_write = default_file_splice_write;
1078
1079         return splice_write(pipe, out, ppos, len, flags);
1080 }
1081
1082 /*
1083  * Attempt to initiate a splice from a file to a pipe.
1084  */
1085 static long do_splice_to(struct file *in, loff_t *ppos,
1086                          struct pipe_inode_info *pipe, size_t len,
1087                          unsigned int flags)
1088 {
1089         ssize_t (*splice_read)(struct file *, loff_t *,
1090                                struct pipe_inode_info *, size_t, unsigned int);
1091         int ret;
1092
1093         if (unlikely(!(in->f_mode & FMODE_READ)))
1094                 return -EBADF;
1095
1096         ret = rw_verify_area(READ, in, ppos, len);
1097         if (unlikely(ret < 0))
1098                 return ret;
1099
1100         if (in->f_op && in->f_op->splice_read)
1101                 splice_read = in->f_op->splice_read;
1102         else
1103                 splice_read = default_file_splice_read;
1104
1105         return splice_read(in, ppos, pipe, len, flags);
1106 }
1107
1108 /**
1109  * splice_direct_to_actor - splices data directly between two non-pipes
1110  * @in:         file to splice from
1111  * @sd:         actor information on where to splice to
1112  * @actor:      handles the data splicing
1113  *
1114  * Description:
1115  *    This is a special case helper to splice directly between two
1116  *    points, without requiring an explicit pipe. Internally an allocated
1117  *    pipe is cached in the process, and reused during the lifetime of
1118  *    that process.
1119  *
1120  */
1121 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1122                                splice_direct_actor *actor)
1123 {
1124         struct pipe_inode_info *pipe;
1125         long ret, bytes;
1126         umode_t i_mode;
1127         size_t len;
1128         int i, flags;
1129
1130         /*
1131          * We require the input being a regular file, as we don't want to
1132          * randomly drop data for eg socket -> socket splicing. Use the
1133          * piped splicing for that!
1134          */
1135         i_mode = in->f_path.dentry->d_inode->i_mode;
1136         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1137                 return -EINVAL;
1138
1139         /*
1140          * neither in nor out is a pipe, setup an internal pipe attached to
1141          * 'out' and transfer the wanted data from 'in' to 'out' through that
1142          */
1143         pipe = current->splice_pipe;
1144         if (unlikely(!pipe)) {
1145                 pipe = alloc_pipe_info(NULL);
1146                 if (!pipe)
1147                         return -ENOMEM;
1148
1149                 /*
1150                  * We don't have an immediate reader, but we'll read the stuff
1151                  * out of the pipe right after the splice_to_pipe(). So set
1152                  * PIPE_READERS appropriately.
1153                  */
1154                 pipe->readers = 1;
1155
1156                 current->splice_pipe = pipe;
1157         }
1158
1159         /*
1160          * Do the splice.
1161          */
1162         ret = 0;
1163         bytes = 0;
1164         len = sd->total_len;
1165         flags = sd->flags;
1166
1167         /*
1168          * Don't block on output, we have to drain the direct pipe.
1169          */
1170         sd->flags &= ~SPLICE_F_NONBLOCK;
1171
1172         while (len) {
1173                 size_t read_len;
1174                 loff_t pos = sd->pos, prev_pos = pos;
1175
1176                 ret = do_splice_to(in, &pos, pipe, len, flags);
1177                 if (unlikely(ret <= 0))
1178                         goto out_release;
1179
1180                 read_len = ret;
1181                 sd->total_len = read_len;
1182
1183                 /*
1184                  * NOTE: nonblocking mode only applies to the input. We
1185                  * must not do the output in nonblocking mode as then we
1186                  * could get stuck data in the internal pipe:
1187                  */
1188                 ret = actor(pipe, sd);
1189                 if (unlikely(ret <= 0)) {
1190                         sd->pos = prev_pos;
1191                         goto out_release;
1192                 }
1193
1194                 bytes += ret;
1195                 len -= ret;
1196                 sd->pos = pos;
1197
1198                 if (ret < read_len) {
1199                         sd->pos = prev_pos + ret;
1200                         goto out_release;
1201                 }
1202         }
1203
1204 done:
1205         pipe->nrbufs = pipe->curbuf = 0;
1206         file_accessed(in);
1207         return bytes;
1208
1209 out_release:
1210         /*
1211          * If we did an incomplete transfer we must release
1212          * the pipe buffers in question:
1213          */
1214         for (i = 0; i < PIPE_BUFFERS; i++) {
1215                 struct pipe_buffer *buf = pipe->bufs + i;
1216
1217                 if (buf->ops) {
1218                         buf->ops->release(pipe, buf);
1219                         buf->ops = NULL;
1220                 }
1221         }
1222
1223         if (!bytes)
1224                 bytes = ret;
1225
1226         goto done;
1227 }
1228 EXPORT_SYMBOL(splice_direct_to_actor);
1229
1230 static int direct_splice_actor(struct pipe_inode_info *pipe,
1231                                struct splice_desc *sd)
1232 {
1233         struct file *file = sd->u.file;
1234
1235         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1236 }
1237
1238 /**
1239  * do_splice_direct - splices data directly between two files
1240  * @in:         file to splice from
1241  * @ppos:       input file offset
1242  * @out:        file to splice to
1243  * @len:        number of bytes to splice
1244  * @flags:      splice modifier flags
1245  *
1246  * Description:
1247  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1248  *    doing it in the application would incur an extra system call
1249  *    (splice in + splice out, as compared to just sendfile()). So this helper
1250  *    can splice directly through a process-private pipe.
1251  *
1252  */
1253 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1254                       size_t len, unsigned int flags)
1255 {
1256         struct splice_desc sd = {
1257                 .len            = len,
1258                 .total_len      = len,
1259                 .flags          = flags,
1260                 .pos            = *ppos,
1261                 .u.file         = out,
1262         };
1263         long ret;
1264
1265         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1266         if (ret > 0)
1267                 *ppos = sd.pos;
1268
1269         return ret;
1270 }
1271
1272 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1273                                struct pipe_inode_info *opipe,
1274                                size_t len, unsigned int flags);
1275 /*
1276  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1277  * location, so checking ->i_pipe is not enough to verify that this is a
1278  * pipe.
1279  */
1280 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1281 {
1282         if (S_ISFIFO(inode->i_mode))
1283                 return inode->i_pipe;
1284
1285         return NULL;
1286 }
1287
1288 /*
1289  * Determine where to splice to/from.
1290  */
1291 static long do_splice(struct file *in, loff_t __user *off_in,
1292                       struct file *out, loff_t __user *off_out,
1293                       size_t len, unsigned int flags)
1294 {
1295         struct pipe_inode_info *ipipe;
1296         struct pipe_inode_info *opipe;
1297         loff_t offset, *off;
1298         long ret;
1299
1300         ipipe = pipe_info(in->f_path.dentry->d_inode);
1301         opipe = pipe_info(out->f_path.dentry->d_inode);
1302
1303         if (ipipe && opipe) {
1304                 if (off_in || off_out)
1305                         return -ESPIPE;
1306
1307                 if (!(in->f_mode & FMODE_READ))
1308                         return -EBADF;
1309
1310                 if (!(out->f_mode & FMODE_WRITE))
1311                         return -EBADF;
1312
1313                 /* Splicing to self would be fun, but... */
1314                 if (ipipe == opipe)
1315                         return -EINVAL;
1316
1317                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1318         }
1319
1320         if (ipipe) {
1321                 if (off_in)
1322                         return -ESPIPE;
1323                 if (off_out) {
1324                         if (!out->f_op || !out->f_op->llseek ||
1325                             out->f_op->llseek == no_llseek)
1326                                 return -EINVAL;
1327                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1328                                 return -EFAULT;
1329                         off = &offset;
1330                 } else
1331                         off = &out->f_pos;
1332
1333                 ret = do_splice_from(ipipe, out, off, len, flags);
1334
1335                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1336                         ret = -EFAULT;
1337
1338                 return ret;
1339         }
1340
1341         if (opipe) {
1342                 if (off_out)
1343                         return -ESPIPE;
1344                 if (off_in) {
1345                         if (!in->f_op || !in->f_op->llseek ||
1346                             in->f_op->llseek == no_llseek)
1347                                 return -EINVAL;
1348                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1349                                 return -EFAULT;
1350                         off = &offset;
1351                 } else
1352                         off = &in->f_pos;
1353
1354                 ret = do_splice_to(in, off, opipe, len, flags);
1355
1356                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1357                         ret = -EFAULT;
1358
1359                 return ret;
1360         }
1361
1362         return -EINVAL;
1363 }
1364
1365 /*
1366  * Map an iov into an array of pages and offset/length tupples. With the
1367  * partial_page structure, we can map several non-contiguous ranges into
1368  * our ones pages[] map instead of splitting that operation into pieces.
1369  * Could easily be exported as a generic helper for other users, in which
1370  * case one would probably want to add a 'max_nr_pages' parameter as well.
1371  */
1372 static int get_iovec_page_array(const struct iovec __user *iov,
1373                                 unsigned int nr_vecs, struct page **pages,
1374                                 struct partial_page *partial, int aligned)
1375 {
1376         int buffers = 0, error = 0;
1377
1378         while (nr_vecs) {
1379                 unsigned long off, npages;
1380                 struct iovec entry;
1381                 void __user *base;
1382                 size_t len;
1383                 int i;
1384
1385                 error = -EFAULT;
1386                 if (copy_from_user(&entry, iov, sizeof(entry)))
1387                         break;
1388
1389                 base = entry.iov_base;
1390                 len = entry.iov_len;
1391
1392                 /*
1393                  * Sanity check this iovec. 0 read succeeds.
1394                  */
1395                 error = 0;
1396                 if (unlikely(!len))
1397                         break;
1398                 error = -EFAULT;
1399                 if (!access_ok(VERIFY_READ, base, len))
1400                         break;
1401
1402                 /*
1403                  * Get this base offset and number of pages, then map
1404                  * in the user pages.
1405                  */
1406                 off = (unsigned long) base & ~PAGE_MASK;
1407
1408                 /*
1409                  * If asked for alignment, the offset must be zero and the
1410                  * length a multiple of the PAGE_SIZE.
1411                  */
1412                 error = -EINVAL;
1413                 if (aligned && (off || len & ~PAGE_MASK))
1414                         break;
1415
1416                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1417                 if (npages > PIPE_BUFFERS - buffers)
1418                         npages = PIPE_BUFFERS - buffers;
1419
1420                 error = get_user_pages_fast((unsigned long)base, npages,
1421                                         0, &pages[buffers]);
1422
1423                 if (unlikely(error <= 0))
1424                         break;
1425
1426                 /*
1427                  * Fill this contiguous range into the partial page map.
1428                  */
1429                 for (i = 0; i < error; i++) {
1430                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1431
1432                         partial[buffers].offset = off;
1433                         partial[buffers].len = plen;
1434
1435                         off = 0;
1436                         len -= plen;
1437                         buffers++;
1438                 }
1439
1440                 /*
1441                  * We didn't complete this iov, stop here since it probably
1442                  * means we have to move some of this into a pipe to
1443                  * be able to continue.
1444                  */
1445                 if (len)
1446                         break;
1447
1448                 /*
1449                  * Don't continue if we mapped fewer pages than we asked for,
1450                  * or if we mapped the max number of pages that we have
1451                  * room for.
1452                  */
1453                 if (error < npages || buffers == PIPE_BUFFERS)
1454                         break;
1455
1456                 nr_vecs--;
1457                 iov++;
1458         }
1459
1460         if (buffers)
1461                 return buffers;
1462
1463         return error;
1464 }
1465
1466 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1467                         struct splice_desc *sd)
1468 {
1469         char *src;
1470         int ret;
1471
1472         ret = buf->ops->confirm(pipe, buf);
1473         if (unlikely(ret))
1474                 return ret;
1475
1476         /*
1477          * See if we can use the atomic maps, by prefaulting in the
1478          * pages and doing an atomic copy
1479          */
1480         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1481                 src = buf->ops->map(pipe, buf, 1);
1482                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1483                                                         sd->len);
1484                 buf->ops->unmap(pipe, buf, src);
1485                 if (!ret) {
1486                         ret = sd->len;
1487                         goto out;
1488                 }
1489         }
1490
1491         /*
1492          * No dice, use slow non-atomic map and copy
1493          */
1494         src = buf->ops->map(pipe, buf, 0);
1495
1496         ret = sd->len;
1497         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1498                 ret = -EFAULT;
1499
1500         buf->ops->unmap(pipe, buf, src);
1501 out:
1502         if (ret > 0)
1503                 sd->u.userptr += ret;
1504         return ret;
1505 }
1506
1507 /*
1508  * For lack of a better implementation, implement vmsplice() to userspace
1509  * as a simple copy of the pipes pages to the user iov.
1510  */
1511 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1512                              unsigned long nr_segs, unsigned int flags)
1513 {
1514         struct pipe_inode_info *pipe;
1515         struct splice_desc sd;
1516         ssize_t size;
1517         int error;
1518         long ret;
1519
1520         pipe = pipe_info(file->f_path.dentry->d_inode);
1521         if (!pipe)
1522                 return -EBADF;
1523
1524         pipe_lock(pipe);
1525
1526         error = ret = 0;
1527         while (nr_segs) {
1528                 void __user *base;
1529                 size_t len;
1530
1531                 /*
1532                  * Get user address base and length for this iovec.
1533                  */
1534                 error = get_user(base, &iov->iov_base);
1535                 if (unlikely(error))
1536                         break;
1537                 error = get_user(len, &iov->iov_len);
1538                 if (unlikely(error))
1539                         break;
1540
1541                 /*
1542                  * Sanity check this iovec. 0 read succeeds.
1543                  */
1544                 if (unlikely(!len))
1545                         break;
1546                 if (unlikely(!base)) {
1547                         error = -EFAULT;
1548                         break;
1549                 }
1550
1551                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1552                         error = -EFAULT;
1553                         break;
1554                 }
1555
1556                 sd.len = 0;
1557                 sd.total_len = len;
1558                 sd.flags = flags;
1559                 sd.u.userptr = base;
1560                 sd.pos = 0;
1561
1562                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1563                 if (size < 0) {
1564                         if (!ret)
1565                                 ret = size;
1566
1567                         break;
1568                 }
1569
1570                 ret += size;
1571
1572                 if (size < len)
1573                         break;
1574
1575                 nr_segs--;
1576                 iov++;
1577         }
1578
1579         pipe_unlock(pipe);
1580
1581         if (!ret)
1582                 ret = error;
1583
1584         return ret;
1585 }
1586
1587 /*
1588  * vmsplice splices a user address range into a pipe. It can be thought of
1589  * as splice-from-memory, where the regular splice is splice-from-file (or
1590  * to file). In both cases the output is a pipe, naturally.
1591  */
1592 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1593                              unsigned long nr_segs, unsigned int flags)
1594 {
1595         struct pipe_inode_info *pipe;
1596         struct page *pages[PIPE_BUFFERS];
1597         struct partial_page partial[PIPE_BUFFERS];
1598         struct splice_pipe_desc spd = {
1599                 .pages = pages,
1600                 .partial = partial,
1601                 .flags = flags,
1602                 .ops = &user_page_pipe_buf_ops,
1603                 .spd_release = spd_release_page,
1604         };
1605
1606         pipe = pipe_info(file->f_path.dentry->d_inode);
1607         if (!pipe)
1608                 return -EBADF;
1609
1610         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1611                                             flags & SPLICE_F_GIFT);
1612         if (spd.nr_pages <= 0)
1613                 return spd.nr_pages;
1614
1615         return splice_to_pipe(pipe, &spd);
1616 }
1617
1618 /*
1619  * Note that vmsplice only really supports true splicing _from_ user memory
1620  * to a pipe, not the other way around. Splicing from user memory is a simple
1621  * operation that can be supported without any funky alignment restrictions
1622  * or nasty vm tricks. We simply map in the user memory and fill them into
1623  * a pipe. The reverse isn't quite as easy, though. There are two possible
1624  * solutions for that:
1625  *
1626  *      - memcpy() the data internally, at which point we might as well just
1627  *        do a regular read() on the buffer anyway.
1628  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1629  *        has restriction limitations on both ends of the pipe).
1630  *
1631  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1632  *
1633  */
1634 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1635                 unsigned long, nr_segs, unsigned int, flags)
1636 {
1637         struct file *file;
1638         long error;
1639         int fput;
1640
1641         if (unlikely(nr_segs > UIO_MAXIOV))
1642                 return -EINVAL;
1643         else if (unlikely(!nr_segs))
1644                 return 0;
1645
1646         error = -EBADF;
1647         file = fget_light(fd, &fput);
1648         if (file) {
1649                 if (file->f_mode & FMODE_WRITE)
1650                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1651                 else if (file->f_mode & FMODE_READ)
1652                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1653
1654                 fput_light(file, fput);
1655         }
1656
1657         return error;
1658 }
1659
1660 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1661                 int, fd_out, loff_t __user *, off_out,
1662                 size_t, len, unsigned int, flags)
1663 {
1664         long error;
1665         struct file *in, *out;
1666         int fput_in, fput_out;
1667
1668         if (unlikely(!len))
1669                 return 0;
1670
1671         error = -EBADF;
1672         in = fget_light(fd_in, &fput_in);
1673         if (in) {
1674                 if (in->f_mode & FMODE_READ) {
1675                         out = fget_light(fd_out, &fput_out);
1676                         if (out) {
1677                                 if (out->f_mode & FMODE_WRITE)
1678                                         error = do_splice(in, off_in,
1679                                                           out, off_out,
1680                                                           len, flags);
1681                                 fput_light(out, fput_out);
1682                         }
1683                 }
1684
1685                 fput_light(in, fput_in);
1686         }
1687
1688         return error;
1689 }
1690
1691 /*
1692  * Make sure there's data to read. Wait for input if we can, otherwise
1693  * return an appropriate error.
1694  */
1695 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1696 {
1697         int ret;
1698
1699         /*
1700          * Check ->nrbufs without the inode lock first. This function
1701          * is speculative anyways, so missing one is ok.
1702          */
1703         if (pipe->nrbufs)
1704                 return 0;
1705
1706         ret = 0;
1707         pipe_lock(pipe);
1708
1709         while (!pipe->nrbufs) {
1710                 if (signal_pending(current)) {
1711                         ret = -ERESTARTSYS;
1712                         break;
1713                 }
1714                 if (!pipe->writers)
1715                         break;
1716                 if (!pipe->waiting_writers) {
1717                         if (flags & SPLICE_F_NONBLOCK) {
1718                                 ret = -EAGAIN;
1719                                 break;
1720                         }
1721                 }
1722                 pipe_wait(pipe);
1723         }
1724
1725         pipe_unlock(pipe);
1726         return ret;
1727 }
1728
1729 /*
1730  * Make sure there's writeable room. Wait for room if we can, otherwise
1731  * return an appropriate error.
1732  */
1733 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1734 {
1735         int ret;
1736
1737         /*
1738          * Check ->nrbufs without the inode lock first. This function
1739          * is speculative anyways, so missing one is ok.
1740          */
1741         if (pipe->nrbufs < PIPE_BUFFERS)
1742                 return 0;
1743
1744         ret = 0;
1745         pipe_lock(pipe);
1746
1747         while (pipe->nrbufs >= PIPE_BUFFERS) {
1748                 if (!pipe->readers) {
1749                         send_sig(SIGPIPE, current, 0);
1750                         ret = -EPIPE;
1751                         break;
1752                 }
1753                 if (flags & SPLICE_F_NONBLOCK) {
1754                         ret = -EAGAIN;
1755                         break;
1756                 }
1757                 if (signal_pending(current)) {
1758                         ret = -ERESTARTSYS;
1759                         break;
1760                 }
1761                 pipe->waiting_writers++;
1762                 pipe_wait(pipe);
1763                 pipe->waiting_writers--;
1764         }
1765
1766         pipe_unlock(pipe);
1767         return ret;
1768 }
1769
1770 /*
1771  * Splice contents of ipipe to opipe.
1772  */
1773 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1774                                struct pipe_inode_info *opipe,
1775                                size_t len, unsigned int flags)
1776 {
1777         struct pipe_buffer *ibuf, *obuf;
1778         int ret = 0, nbuf;
1779         bool input_wakeup = false;
1780
1781
1782 retry:
1783         ret = ipipe_prep(ipipe, flags);
1784         if (ret)
1785                 return ret;
1786
1787         ret = opipe_prep(opipe, flags);
1788         if (ret)
1789                 return ret;
1790
1791         /*
1792          * Potential ABBA deadlock, work around it by ordering lock
1793          * grabbing by pipe info address. Otherwise two different processes
1794          * could deadlock (one doing tee from A -> B, the other from B -> A).
1795          */
1796         pipe_double_lock(ipipe, opipe);
1797
1798         do {
1799                 if (!opipe->readers) {
1800                         send_sig(SIGPIPE, current, 0);
1801                         if (!ret)
1802                                 ret = -EPIPE;
1803                         break;
1804                 }
1805
1806                 if (!ipipe->nrbufs && !ipipe->writers)
1807                         break;
1808
1809                 /*
1810                  * Cannot make any progress, because either the input
1811                  * pipe is empty or the output pipe is full.
1812                  */
1813                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1814                         /* Already processed some buffers, break */
1815                         if (ret)
1816                                 break;
1817
1818                         if (flags & SPLICE_F_NONBLOCK) {
1819                                 ret = -EAGAIN;
1820                                 break;
1821                         }
1822
1823                         /*
1824                          * We raced with another reader/writer and haven't
1825                          * managed to process any buffers.  A zero return
1826                          * value means EOF, so retry instead.
1827                          */
1828                         pipe_unlock(ipipe);
1829                         pipe_unlock(opipe);
1830                         goto retry;
1831                 }
1832
1833                 ibuf = ipipe->bufs + ipipe->curbuf;
1834                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1835                 obuf = opipe->bufs + nbuf;
1836
1837                 if (len >= ibuf->len) {
1838                         /*
1839                          * Simply move the whole buffer from ipipe to opipe
1840                          */
1841                         *obuf = *ibuf;
1842                         ibuf->ops = NULL;
1843                         opipe->nrbufs++;
1844                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1845                         ipipe->nrbufs--;
1846                         input_wakeup = true;
1847                 } else {
1848                         /*
1849                          * Get a reference to this pipe buffer,
1850                          * so we can copy the contents over.
1851                          */
1852                         ibuf->ops->get(ipipe, ibuf);
1853                         *obuf = *ibuf;
1854
1855                         /*
1856                          * Don't inherit the gift flag, we need to
1857                          * prevent multiple steals of this page.
1858                          */
1859                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1860
1861                         obuf->len = len;
1862                         opipe->nrbufs++;
1863                         ibuf->offset += obuf->len;
1864                         ibuf->len -= obuf->len;
1865                 }
1866                 ret += obuf->len;
1867                 len -= obuf->len;
1868         } while (len);
1869
1870         pipe_unlock(ipipe);
1871         pipe_unlock(opipe);
1872
1873         /*
1874          * If we put data in the output pipe, wakeup any potential readers.
1875          */
1876         if (ret > 0) {
1877                 smp_mb();
1878                 if (waitqueue_active(&opipe->wait))
1879                         wake_up_interruptible(&opipe->wait);
1880                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1881         }
1882         if (input_wakeup)
1883                 wakeup_pipe_writers(ipipe);
1884
1885         return ret;
1886 }
1887
1888 /*
1889  * Link contents of ipipe to opipe.
1890  */
1891 static int link_pipe(struct pipe_inode_info *ipipe,
1892                      struct pipe_inode_info *opipe,
1893                      size_t len, unsigned int flags)
1894 {
1895         struct pipe_buffer *ibuf, *obuf;
1896         int ret = 0, i = 0, nbuf;
1897
1898         /*
1899          * Potential ABBA deadlock, work around it by ordering lock
1900          * grabbing by pipe info address. Otherwise two different processes
1901          * could deadlock (one doing tee from A -> B, the other from B -> A).
1902          */
1903         pipe_double_lock(ipipe, opipe);
1904
1905         do {
1906                 if (!opipe->readers) {
1907                         send_sig(SIGPIPE, current, 0);
1908                         if (!ret)
1909                                 ret = -EPIPE;
1910                         break;
1911                 }
1912
1913                 /*
1914                  * If we have iterated all input buffers or ran out of
1915                  * output room, break.
1916                  */
1917                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1918                         break;
1919
1920                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1921                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1922
1923                 /*
1924                  * Get a reference to this pipe buffer,
1925                  * so we can copy the contents over.
1926                  */
1927                 ibuf->ops->get(ipipe, ibuf);
1928
1929                 obuf = opipe->bufs + nbuf;
1930                 *obuf = *ibuf;
1931
1932                 /*
1933                  * Don't inherit the gift flag, we need to
1934                  * prevent multiple steals of this page.
1935                  */
1936                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1937
1938                 if (obuf->len > len)
1939                         obuf->len = len;
1940
1941                 opipe->nrbufs++;
1942                 ret += obuf->len;
1943                 len -= obuf->len;
1944                 i++;
1945         } while (len);
1946
1947         /*
1948          * return EAGAIN if we have the potential of some data in the
1949          * future, otherwise just return 0
1950          */
1951         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1952                 ret = -EAGAIN;
1953
1954         pipe_unlock(ipipe);
1955         pipe_unlock(opipe);
1956
1957         /*
1958          * If we put data in the output pipe, wakeup any potential readers.
1959          */
1960         if (ret > 0) {
1961                 smp_mb();
1962                 if (waitqueue_active(&opipe->wait))
1963                         wake_up_interruptible(&opipe->wait);
1964                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1965         }
1966
1967         return ret;
1968 }
1969
1970 /*
1971  * This is a tee(1) implementation that works on pipes. It doesn't copy
1972  * any data, it simply references the 'in' pages on the 'out' pipe.
1973  * The 'flags' used are the SPLICE_F_* variants, currently the only
1974  * applicable one is SPLICE_F_NONBLOCK.
1975  */
1976 static long do_tee(struct file *in, struct file *out, size_t len,
1977                    unsigned int flags)
1978 {
1979         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1980         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1981         int ret = -EINVAL;
1982
1983         /*
1984          * Duplicate the contents of ipipe to opipe without actually
1985          * copying the data.
1986          */
1987         if (ipipe && opipe && ipipe != opipe) {
1988                 /*
1989                  * Keep going, unless we encounter an error. The ipipe/opipe
1990                  * ordering doesn't really matter.
1991                  */
1992                 ret = ipipe_prep(ipipe, flags);
1993                 if (!ret) {
1994                         ret = opipe_prep(opipe, flags);
1995                         if (!ret)
1996                                 ret = link_pipe(ipipe, opipe, len, flags);
1997                 }
1998         }
1999
2000         return ret;
2001 }
2002
2003 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2004 {
2005         struct file *in;
2006         int error, fput_in;
2007
2008         if (unlikely(!len))
2009                 return 0;
2010
2011         error = -EBADF;
2012         in = fget_light(fdin, &fput_in);
2013         if (in) {
2014                 if (in->f_mode & FMODE_READ) {
2015                         int fput_out;
2016                         struct file *out = fget_light(fdout, &fput_out);
2017
2018                         if (out) {
2019                                 if (out->f_mode & FMODE_WRITE)
2020                                         error = do_tee(in, out, len, flags);
2021                                 fput_light(out, fput_out);
2022                         }
2023                 }
2024                 fput_light(in, fput_in);
2025         }
2026
2027         return error;
2028 }