[PATCH] splice: cleanup __generic_file_splice_read()
[pandora-kernel.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 and fixing the initial implementation
13  * bugs.
14  *
15  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
16  * Copyright (C) 2005 Linus Torvalds <torvalds@osdl.org>
17  *
18  */
19 #include <linux/fs.h>
20 #include <linux/file.h>
21 #include <linux/pagemap.h>
22 #include <linux/pipe_fs_i.h>
23 #include <linux/mm_inline.h>
24 #include <linux/swap.h>
25 #include <linux/writeback.h>
26 #include <linux/buffer_head.h>
27 #include <linux/module.h>
28 #include <linux/syscalls.h>
29
30 /*
31  * Passed to the actors
32  */
33 struct splice_desc {
34         unsigned int len, total_len;    /* current and remaining length */
35         unsigned int flags;             /* splice flags */
36         struct file *file;              /* file to read/write */
37         loff_t pos;                     /* file position */
38 };
39
40 /*
41  * Attempt to steal a page from a pipe buffer. This should perhaps go into
42  * a vm helper function, it's already simplified quite a bit by the
43  * addition of remove_mapping(). If success is returned, the caller may
44  * attempt to reuse this page for another destination.
45  */
46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
47                                      struct pipe_buffer *buf)
48 {
49         struct page *page = buf->page;
50         struct address_space *mapping = page_mapping(page);
51
52         WARN_ON(!PageLocked(page));
53         WARN_ON(!PageUptodate(page));
54
55         /*
56          * At least for ext2 with nobh option, we need to wait on writeback
57          * completing on this page, since we'll remove it from the pagecache.
58          * Otherwise truncate wont wait on the page, allowing the disk
59          * blocks to be reused by someone else before we actually wrote our
60          * data to them. fs corruption ensues.
61          */
62         wait_on_page_writeback(page);
63
64         if (PagePrivate(page))
65                 try_to_release_page(page, mapping_gfp_mask(mapping));
66
67         if (!remove_mapping(mapping, page))
68                 return 1;
69
70         buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
71         return 0;
72 }
73
74 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
75                                         struct pipe_buffer *buf)
76 {
77         page_cache_release(buf->page);
78         buf->page = NULL;
79         buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
80 }
81
82 static void *page_cache_pipe_buf_map(struct file *file,
83                                      struct pipe_inode_info *info,
84                                      struct pipe_buffer *buf)
85 {
86         struct page *page = buf->page;
87
88         lock_page(page);
89
90         if (!PageUptodate(page)) {
91                 unlock_page(page);
92                 return ERR_PTR(-EIO);
93         }
94
95         if (!page->mapping) {
96                 unlock_page(page);
97                 return ERR_PTR(-ENODATA);
98         }
99
100         return kmap(buf->page);
101 }
102
103 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
104                                       struct pipe_buffer *buf)
105 {
106         unlock_page(buf->page);
107         kunmap(buf->page);
108 }
109
110 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
111         .can_merge = 0,
112         .map = page_cache_pipe_buf_map,
113         .unmap = page_cache_pipe_buf_unmap,
114         .release = page_cache_pipe_buf_release,
115         .steal = page_cache_pipe_buf_steal,
116 };
117
118 /*
119  * Pipe output worker. This sets up our pipe format with the page cache
120  * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
121  */
122 static ssize_t move_to_pipe(struct inode *inode, struct page **pages,
123                             int nr_pages, unsigned long offset,
124                             unsigned long len, unsigned int flags)
125 {
126         struct pipe_inode_info *info;
127         int ret, do_wakeup, i;
128
129         ret = 0;
130         do_wakeup = 0;
131         i = 0;
132
133         mutex_lock(PIPE_MUTEX(*inode));
134
135         info = inode->i_pipe;
136         for (;;) {
137                 int bufs;
138
139                 if (!PIPE_READERS(*inode)) {
140                         send_sig(SIGPIPE, current, 0);
141                         if (!ret)
142                                 ret = -EPIPE;
143                         break;
144                 }
145
146                 bufs = info->nrbufs;
147                 if (bufs < PIPE_BUFFERS) {
148                         int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS - 1);
149                         struct pipe_buffer *buf = info->bufs + newbuf;
150                         struct page *page = pages[i++];
151                         unsigned long this_len;
152
153                         this_len = PAGE_CACHE_SIZE - offset;
154                         if (this_len > len)
155                                 this_len = len;
156
157                         buf->page = page;
158                         buf->offset = offset;
159                         buf->len = this_len;
160                         buf->ops = &page_cache_pipe_buf_ops;
161                         info->nrbufs = ++bufs;
162                         do_wakeup = 1;
163
164                         ret += this_len;
165                         len -= this_len;
166                         offset = 0;
167                         if (!--nr_pages)
168                                 break;
169                         if (!len)
170                                 break;
171                         if (bufs < PIPE_BUFFERS)
172                                 continue;
173
174                         break;
175                 }
176
177                 if (flags & SPLICE_F_NONBLOCK) {
178                         if (!ret)
179                                 ret = -EAGAIN;
180                         break;
181                 }
182
183                 if (signal_pending(current)) {
184                         if (!ret)
185                                 ret = -ERESTARTSYS;
186                         break;
187                 }
188
189                 if (do_wakeup) {
190                         smp_mb();
191                         if (waitqueue_active(PIPE_WAIT(*inode)))
192                                 wake_up_interruptible_sync(PIPE_WAIT(*inode));
193                         kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO,
194                                     POLL_IN);
195                         do_wakeup = 0;
196                 }
197
198                 PIPE_WAITING_WRITERS(*inode)++;
199                 pipe_wait(inode);
200                 PIPE_WAITING_WRITERS(*inode)--;
201         }
202
203         mutex_unlock(PIPE_MUTEX(*inode));
204
205         if (do_wakeup) {
206                 smp_mb();
207                 if (waitqueue_active(PIPE_WAIT(*inode)))
208                         wake_up_interruptible(PIPE_WAIT(*inode));
209                 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
210         }
211
212         while (i < nr_pages)
213                 page_cache_release(pages[i++]);
214
215         return ret;
216 }
217
218 static int __generic_file_splice_read(struct file *in, struct inode *pipe,
219                                       size_t len, unsigned int flags)
220 {
221         struct address_space *mapping = in->f_mapping;
222         unsigned int offset, nr_pages;
223         struct page *pages[PIPE_BUFFERS];
224         struct page *page;
225         pgoff_t index;
226         int i;
227
228         index = in->f_pos >> PAGE_CACHE_SHIFT;
229         offset = in->f_pos & ~PAGE_CACHE_MASK;
230         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
231
232         if (nr_pages > PIPE_BUFFERS)
233                 nr_pages = PIPE_BUFFERS;
234
235         /*
236          * initiate read-ahead on this page range
237          */
238         do_page_cache_readahead(mapping, in, index, nr_pages);
239
240         /*
241          * now fill in the holes
242          */
243         for (i = 0; i < nr_pages; i++, index++) {
244                 /*
245                  * no page there, look one up / create it
246                  */
247                 page = find_or_create_page(mapping, index,
248                                                    mapping_gfp_mask(mapping));
249                 if (!page)
250                         break;
251
252                 if (PageUptodate(page))
253                         unlock_page(page);
254                 else {
255                         int error = mapping->a_ops->readpage(in, page);
256
257                         if (unlikely(error)) {
258                                 page_cache_release(page);
259                                 break;
260                         }
261                 }
262                 pages[i] = page;
263         }
264
265         if (i)
266                 return move_to_pipe(pipe, pages, i, offset, len, flags);
267
268         return 0;
269 }
270
271 /**
272  * generic_file_splice_read - splice data from file to a pipe
273  * @in:         file to splice from
274  * @pipe:       pipe to splice to
275  * @len:        number of bytes to splice
276  * @flags:      splice modifier flags
277  *
278  * Will read pages from given file and fill them into a pipe.
279  *
280  */
281 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
282                                  size_t len, unsigned int flags)
283 {
284         ssize_t spliced;
285         int ret;
286
287         ret = 0;
288         spliced = 0;
289         while (len) {
290                 ret = __generic_file_splice_read(in, pipe, len, flags);
291
292                 if (ret <= 0)
293                         break;
294
295                 in->f_pos += ret;
296                 len -= ret;
297                 spliced += ret;
298
299                 if (!(flags & SPLICE_F_NONBLOCK))
300                         continue;
301                 ret = -EAGAIN;
302                 break;
303         }
304
305         if (spliced)
306                 return spliced;
307
308         return ret;
309 }
310
311 EXPORT_SYMBOL(generic_file_splice_read);
312
313 /*
314  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
315  * using sendpage().
316  */
317 static int pipe_to_sendpage(struct pipe_inode_info *info,
318                             struct pipe_buffer *buf, struct splice_desc *sd)
319 {
320         struct file *file = sd->file;
321         loff_t pos = sd->pos;
322         unsigned int offset;
323         ssize_t ret;
324         void *ptr;
325         int more;
326
327         /*
328          * sub-optimal, but we are limited by the pipe ->map. we don't
329          * need a kmap'ed buffer here, we just want to make sure we
330          * have the page pinned if the pipe page originates from the
331          * page cache
332          */
333         ptr = buf->ops->map(file, info, buf);
334         if (IS_ERR(ptr))
335                 return PTR_ERR(ptr);
336
337         offset = pos & ~PAGE_CACHE_MASK;
338         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
339
340         ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
341
342         buf->ops->unmap(info, buf);
343         if (ret == sd->len)
344                 return 0;
345
346         return -EIO;
347 }
348
349 /*
350  * This is a little more tricky than the file -> pipe splicing. There are
351  * basically three cases:
352  *
353  *      - Destination page already exists in the address space and there
354  *        are users of it. For that case we have no other option that
355  *        copying the data. Tough luck.
356  *      - Destination page already exists in the address space, but there
357  *        are no users of it. Make sure it's uptodate, then drop it. Fall
358  *        through to last case.
359  *      - Destination page does not exist, we can add the pipe page to
360  *        the page cache and avoid the copy.
361  *
362  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
363  * sd->flags), we attempt to migrate pages from the pipe to the output
364  * file address space page cache. This is possible if no one else has
365  * the pipe page referenced outside of the pipe and page cache. If
366  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
367  * a new page in the output file page cache and fill/dirty that.
368  */
369 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
370                         struct splice_desc *sd)
371 {
372         struct file *file = sd->file;
373         struct address_space *mapping = file->f_mapping;
374         gfp_t gfp_mask = mapping_gfp_mask(mapping);
375         unsigned int offset;
376         struct page *page;
377         pgoff_t index;
378         char *src;
379         int ret;
380
381         /*
382          * after this, page will be locked and unmapped
383          */
384         src = buf->ops->map(file, info, buf);
385         if (IS_ERR(src))
386                 return PTR_ERR(src);
387
388         index = sd->pos >> PAGE_CACHE_SHIFT;
389         offset = sd->pos & ~PAGE_CACHE_MASK;
390
391         /*
392          * reuse buf page, if SPLICE_F_MOVE is set
393          */
394         if (sd->flags & SPLICE_F_MOVE) {
395                 /*
396                  * If steal succeeds, buf->page is now pruned from the vm
397                  * side (LRU and page cache) and we can reuse it.
398                  */
399                 if (buf->ops->steal(info, buf))
400                         goto find_page;
401
402                 page = buf->page;
403                 if (add_to_page_cache(page, mapping, index, gfp_mask))
404                         goto find_page;
405
406                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
407                         lru_cache_add(page);
408         } else {
409 find_page:
410                 ret = -ENOMEM;
411                 page = find_or_create_page(mapping, index, gfp_mask);
412                 if (!page)
413                         goto out_nomem;
414
415                 /*
416                  * If the page is uptodate, it is also locked. If it isn't
417                  * uptodate, we can mark it uptodate if we are filling the
418                  * full page. Otherwise we need to read it in first...
419                  */
420                 if (!PageUptodate(page)) {
421                         if (sd->len < PAGE_CACHE_SIZE) {
422                                 ret = mapping->a_ops->readpage(file, page);
423                                 if (unlikely(ret))
424                                         goto out;
425
426                                 lock_page(page);
427
428                                 if (!PageUptodate(page)) {
429                                         /*
430                                          * page got invalidated, repeat
431                                          */
432                                         if (!page->mapping) {
433                                                 unlock_page(page);
434                                                 page_cache_release(page);
435                                                 goto find_page;
436                                         }
437                                         ret = -EIO;
438                                         goto out;
439                                 }
440                         } else {
441                                 WARN_ON(!PageLocked(page));
442                                 SetPageUptodate(page);
443                         }
444                 }
445         }
446
447         ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
448         if (ret == AOP_TRUNCATED_PAGE) {
449                 page_cache_release(page);
450                 goto find_page;
451         } else if (ret)
452                 goto out;
453
454         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
455                 char *dst = kmap_atomic(page, KM_USER0);
456
457                 memcpy(dst + offset, src + buf->offset, sd->len);
458                 flush_dcache_page(page);
459                 kunmap_atomic(dst, KM_USER0);
460         }
461
462         ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
463         if (ret == AOP_TRUNCATED_PAGE) {
464                 page_cache_release(page);
465                 goto find_page;
466         } else if (ret)
467                 goto out;
468
469         mark_page_accessed(page);
470         balance_dirty_pages_ratelimited(mapping);
471 out:
472         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
473                 page_cache_release(page);
474                 unlock_page(page);
475         }
476 out_nomem:
477         buf->ops->unmap(info, buf);
478         return ret;
479 }
480
481 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
482                            struct splice_desc *);
483
484 /*
485  * Pipe input worker. Most of this logic works like a regular pipe, the
486  * key here is the 'actor' worker passed in that actually moves the data
487  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
488  */
489 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
490                               size_t len, unsigned int flags,
491                               splice_actor *actor)
492 {
493         struct pipe_inode_info *info;
494         int ret, do_wakeup, err;
495         struct splice_desc sd;
496
497         ret = 0;
498         do_wakeup = 0;
499
500         sd.total_len = len;
501         sd.flags = flags;
502         sd.file = out;
503         sd.pos = out->f_pos;
504
505         mutex_lock(PIPE_MUTEX(*inode));
506
507         info = inode->i_pipe;
508         for (;;) {
509                 int bufs = info->nrbufs;
510
511                 if (bufs) {
512                         int curbuf = info->curbuf;
513                         struct pipe_buffer *buf = info->bufs + curbuf;
514                         struct pipe_buf_operations *ops = buf->ops;
515
516                         sd.len = buf->len;
517                         if (sd.len > sd.total_len)
518                                 sd.len = sd.total_len;
519
520                         err = actor(info, buf, &sd);
521                         if (err) {
522                                 if (!ret && err != -ENODATA)
523                                         ret = err;
524
525                                 break;
526                         }
527
528                         ret += sd.len;
529                         buf->offset += sd.len;
530                         buf->len -= sd.len;
531                         if (!buf->len) {
532                                 buf->ops = NULL;
533                                 ops->release(info, buf);
534                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
535                                 info->curbuf = curbuf;
536                                 info->nrbufs = --bufs;
537                                 do_wakeup = 1;
538                         }
539
540                         sd.pos += sd.len;
541                         sd.total_len -= sd.len;
542                         if (!sd.total_len)
543                                 break;
544                 }
545
546                 if (bufs)
547                         continue;
548                 if (!PIPE_WRITERS(*inode))
549                         break;
550                 if (!PIPE_WAITING_WRITERS(*inode)) {
551                         if (ret)
552                                 break;
553                 }
554
555                 if (flags & SPLICE_F_NONBLOCK) {
556                         if (!ret)
557                                 ret = -EAGAIN;
558                         break;
559                 }
560
561                 if (signal_pending(current)) {
562                         if (!ret)
563                                 ret = -ERESTARTSYS;
564                         break;
565                 }
566
567                 if (do_wakeup) {
568                         smp_mb();
569                         if (waitqueue_active(PIPE_WAIT(*inode)))
570                                 wake_up_interruptible_sync(PIPE_WAIT(*inode));
571                         kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
572                         do_wakeup = 0;
573                 }
574
575                 pipe_wait(inode);
576         }
577
578         mutex_unlock(PIPE_MUTEX(*inode));
579
580         if (do_wakeup) {
581                 smp_mb();
582                 if (waitqueue_active(PIPE_WAIT(*inode)))
583                         wake_up_interruptible(PIPE_WAIT(*inode));
584                 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
585         }
586
587         mutex_lock(&out->f_mapping->host->i_mutex);
588         out->f_pos = sd.pos;
589         mutex_unlock(&out->f_mapping->host->i_mutex);
590         return ret;
591
592 }
593
594 /**
595  * generic_file_splice_write - splice data from a pipe to a file
596  * @inode:      pipe inode
597  * @out:        file to write to
598  * @len:        number of bytes to splice
599  * @flags:      splice modifier flags
600  *
601  * Will either move or copy pages (determined by @flags options) from
602  * the given pipe inode to the given file.
603  *
604  */
605 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
606                                   size_t len, unsigned int flags)
607 {
608         struct address_space *mapping = out->f_mapping;
609         ssize_t ret = move_from_pipe(inode, out, len, flags, pipe_to_file);
610
611         /*
612          * if file or inode is SYNC and we actually wrote some data, sync it
613          */
614         if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
615             && ret > 0) {
616                 struct inode *inode = mapping->host;
617                 int err;
618
619                 mutex_lock(&inode->i_mutex);
620                 err = generic_osync_inode(mapping->host, mapping,
621                                                 OSYNC_METADATA|OSYNC_DATA);
622                 mutex_unlock(&inode->i_mutex);
623
624                 if (err)
625                         ret = err;
626         }
627
628         return ret;
629 }
630
631 EXPORT_SYMBOL(generic_file_splice_write);
632
633 /**
634  * generic_splice_sendpage - splice data from a pipe to a socket
635  * @inode:      pipe inode
636  * @out:        socket to write to
637  * @len:        number of bytes to splice
638  * @flags:      splice modifier flags
639  *
640  * Will send @len bytes from the pipe to a network socket. No data copying
641  * is involved.
642  *
643  */
644 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
645                                 size_t len, unsigned int flags)
646 {
647         return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
648 }
649
650 EXPORT_SYMBOL(generic_splice_sendpage);
651
652 /*
653  * Attempt to initiate a splice from pipe to file.
654  */
655 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
656                            unsigned int flags)
657 {
658         loff_t pos;
659         int ret;
660
661         if (!out->f_op || !out->f_op->splice_write)
662                 return -EINVAL;
663
664         if (!(out->f_mode & FMODE_WRITE))
665                 return -EBADF;
666
667         pos = out->f_pos;
668         ret = rw_verify_area(WRITE, out, &pos, len);
669         if (unlikely(ret < 0))
670                 return ret;
671
672         return out->f_op->splice_write(pipe, out, len, flags);
673 }
674
675 /*
676  * Attempt to initiate a splice from a file to a pipe.
677  */
678 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
679                          unsigned int flags)
680 {
681         loff_t pos, isize, left;
682         int ret;
683
684         if (!in->f_op || !in->f_op->splice_read)
685                 return -EINVAL;
686
687         if (!(in->f_mode & FMODE_READ))
688                 return -EBADF;
689
690         pos = in->f_pos;
691         ret = rw_verify_area(READ, in, &pos, len);
692         if (unlikely(ret < 0))
693                 return ret;
694
695         isize = i_size_read(in->f_mapping->host);
696         if (unlikely(in->f_pos >= isize))
697                 return 0;
698         
699         left = isize - in->f_pos;
700         if (left < len)
701                 len = left;
702
703         return in->f_op->splice_read(in, pipe, len, flags);
704 }
705
706 /*
707  * Determine where to splice to/from.
708  */
709 static long do_splice(struct file *in, struct file *out, size_t len,
710                       unsigned int flags)
711 {
712         struct inode *pipe;
713
714         pipe = in->f_dentry->d_inode;
715         if (pipe->i_pipe)
716                 return do_splice_from(pipe, out, len, flags);
717
718         pipe = out->f_dentry->d_inode;
719         if (pipe->i_pipe)
720                 return do_splice_to(in, pipe, len, flags);
721
722         return -EINVAL;
723 }
724
725 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
726 {
727         long error;
728         struct file *in, *out;
729         int fput_in, fput_out;
730
731         if (unlikely(!len))
732                 return 0;
733
734         error = -EBADF;
735         in = fget_light(fdin, &fput_in);
736         if (in) {
737                 if (in->f_mode & FMODE_READ) {
738                         out = fget_light(fdout, &fput_out);
739                         if (out) {
740                                 if (out->f_mode & FMODE_WRITE)
741                                         error = do_splice(in, out, len, flags);
742                                 fput_light(out, fput_out);
743                         }
744                 }
745
746                 fput_light(in, fput_in);
747         }
748
749         return error;
750 }