blob: 1633778f36528a9179de90222e7bf81a544d6e77 [file] [log] [blame]
Jens Axboe5274f052006-03-30 15:15:30 +02001/*
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
Jens Axboec2058e02006-04-11 13:56:34 +020012 * Jens to support splicing to files, network, direct splicing, etc and
13 * fixing lots of bugs.
Jens Axboe5274f052006-03-30 15:15:30 +020014 *
Jens Axboec2058e02006-04-11 13:56:34 +020015 * Copyright (C) 2005-2006 Jens Axboe <axboe@suse.de>
16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
Jens Axboe5274f052006-03-30 15:15:30 +020018 *
19 */
20#include <linux/fs.h>
21#include <linux/file.h>
22#include <linux/pagemap.h>
23#include <linux/pipe_fs_i.h>
24#include <linux/mm_inline.h>
Jens Axboe5abc97a2006-03-30 15:16:46 +020025#include <linux/swap.h>
Jens Axboe4f6f0bd2006-04-02 23:04:46 +020026#include <linux/writeback.h>
27#include <linux/buffer_head.h>
Jeff Garzika0f06782006-03-30 23:06:13 -050028#include <linux/module.h>
Jens Axboe4f6f0bd2006-04-02 23:04:46 +020029#include <linux/syscalls.h>
Jens Axboe912d35f2006-04-26 10:59:21 +020030#include <linux/uio.h>
Jens Axboe5274f052006-03-30 15:15:30 +020031
Jens Axboe912d35f2006-04-26 10:59:21 +020032struct partial_page {
33 unsigned int offset;
34 unsigned int len;
35};
36
37/*
Jens Axboe00522fb2006-04-26 14:39:29 +020038 * Passed to splice_to_pipe
Jens Axboe912d35f2006-04-26 10:59:21 +020039 */
40struct splice_pipe_desc {
41 struct page **pages; /* page map */
42 struct partial_page *partial; /* pages[] may not be contig */
43 int nr_pages; /* number of pages in map */
44 unsigned int flags; /* splice flags */
45 struct pipe_buf_operations *ops;/* ops associated with output pipe */
46};
47
Jens Axboe83f91352006-04-02 23:05:09 +020048/*
49 * Attempt to steal a page from a pipe buffer. This should perhaps go into
50 * a vm helper function, it's already simplified quite a bit by the
51 * addition of remove_mapping(). If success is returned, the caller may
52 * attempt to reuse this page for another destination.
53 */
Jens Axboe5abc97a2006-03-30 15:16:46 +020054static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
55 struct pipe_buffer *buf)
56{
57 struct page *page = buf->page;
Jens Axboe4f6f0bd2006-04-02 23:04:46 +020058 struct address_space *mapping = page_mapping(page);
Jens Axboe5abc97a2006-03-30 15:16:46 +020059
Jens Axboe9e0267c2006-04-19 15:57:31 +020060 lock_page(page);
61
Jens Axboe5abc97a2006-03-30 15:16:46 +020062 WARN_ON(!PageUptodate(page));
63
Jens Axboead8d6f02006-04-02 23:10:32 +020064 /*
65 * At least for ext2 with nobh option, we need to wait on writeback
66 * completing on this page, since we'll remove it from the pagecache.
67 * Otherwise truncate wont wait on the page, allowing the disk
68 * blocks to be reused by someone else before we actually wrote our
69 * data to them. fs corruption ensues.
70 */
71 wait_on_page_writeback(page);
72
Jens Axboe4f6f0bd2006-04-02 23:04:46 +020073 if (PagePrivate(page))
74 try_to_release_page(page, mapping_gfp_mask(mapping));
75
Jens Axboe9e0267c2006-04-19 15:57:31 +020076 if (!remove_mapping(mapping, page)) {
77 unlock_page(page);
Jens Axboe5abc97a2006-03-30 15:16:46 +020078 return 1;
Jens Axboe9e0267c2006-04-19 15:57:31 +020079 }
Jens Axboe5abc97a2006-03-30 15:16:46 +020080
Jens Axboe0568b402006-05-01 19:50:48 +020081 buf->flags |= PIPE_BUF_FLAG_LRU;
Jens Axboe5abc97a2006-03-30 15:16:46 +020082 return 0;
83}
84
Jens Axboe5274f052006-03-30 15:15:30 +020085static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
86 struct pipe_buffer *buf)
87{
88 page_cache_release(buf->page);
89 buf->page = NULL;
Jens Axboe0568b402006-05-01 19:50:48 +020090 buf->flags &= ~PIPE_BUF_FLAG_LRU;
Jens Axboe5274f052006-03-30 15:15:30 +020091}
92
93static void *page_cache_pipe_buf_map(struct file *file,
94 struct pipe_inode_info *info,
95 struct pipe_buffer *buf)
96{
97 struct page *page = buf->page;
Jens Axboe49d0b212006-04-10 09:04:41 +020098 int err;
Jens Axboe5274f052006-03-30 15:15:30 +020099
100 if (!PageUptodate(page)) {
Jens Axboe49d0b212006-04-10 09:04:41 +0200101 lock_page(page);
102
103 /*
104 * Page got truncated/unhashed. This will cause a 0-byte
Ingo Molnar73d62d82006-04-11 13:57:21 +0200105 * splice, if this is the first page.
Jens Axboe49d0b212006-04-10 09:04:41 +0200106 */
107 if (!page->mapping) {
108 err = -ENODATA;
109 goto error;
110 }
111
112 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200113 * Uh oh, read-error from disk.
Jens Axboe49d0b212006-04-10 09:04:41 +0200114 */
115 if (!PageUptodate(page)) {
116 err = -EIO;
117 goto error;
118 }
119
120 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200121 * Page is ok afterall, fall through to mapping.
Jens Axboe49d0b212006-04-10 09:04:41 +0200122 */
Jens Axboe5274f052006-03-30 15:15:30 +0200123 unlock_page(page);
Jens Axboe5274f052006-03-30 15:15:30 +0200124 }
125
Jens Axboe49d0b212006-04-10 09:04:41 +0200126 return kmap(page);
127error:
128 unlock_page(page);
129 return ERR_PTR(err);
Jens Axboe5274f052006-03-30 15:15:30 +0200130}
131
132static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
133 struct pipe_buffer *buf)
134{
Jens Axboe5274f052006-03-30 15:15:30 +0200135 kunmap(buf->page);
136}
137
Jens Axboe912d35f2006-04-26 10:59:21 +0200138static void *user_page_pipe_buf_map(struct file *file,
139 struct pipe_inode_info *pipe,
140 struct pipe_buffer *buf)
141{
142 return kmap(buf->page);
143}
144
145static void user_page_pipe_buf_unmap(struct pipe_inode_info *pipe,
146 struct pipe_buffer *buf)
147{
148 kunmap(buf->page);
149}
150
Jens Axboe70524492006-04-11 15:51:17 +0200151static void page_cache_pipe_buf_get(struct pipe_inode_info *info,
152 struct pipe_buffer *buf)
153{
154 page_cache_get(buf->page);
155}
156
Jens Axboe5274f052006-03-30 15:15:30 +0200157static struct pipe_buf_operations page_cache_pipe_buf_ops = {
158 .can_merge = 0,
159 .map = page_cache_pipe_buf_map,
160 .unmap = page_cache_pipe_buf_unmap,
161 .release = page_cache_pipe_buf_release,
Jens Axboe5abc97a2006-03-30 15:16:46 +0200162 .steal = page_cache_pipe_buf_steal,
Jens Axboe70524492006-04-11 15:51:17 +0200163 .get = page_cache_pipe_buf_get,
Jens Axboe5274f052006-03-30 15:15:30 +0200164};
165
Jens Axboe912d35f2006-04-26 10:59:21 +0200166static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
167 struct pipe_buffer *buf)
168{
169 return 1;
170}
171
172static struct pipe_buf_operations user_page_pipe_buf_ops = {
173 .can_merge = 0,
174 .map = user_page_pipe_buf_map,
175 .unmap = user_page_pipe_buf_unmap,
176 .release = page_cache_pipe_buf_release,
177 .steal = user_page_pipe_buf_steal,
178 .get = page_cache_pipe_buf_get,
179};
180
Jens Axboe83f91352006-04-02 23:05:09 +0200181/*
182 * Pipe output worker. This sets up our pipe format with the page cache
183 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
184 */
Jens Axboe00522fb2006-04-26 14:39:29 +0200185static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
186 struct splice_pipe_desc *spd)
Jens Axboe5274f052006-03-30 15:15:30 +0200187{
Jens Axboe912d35f2006-04-26 10:59:21 +0200188 int ret, do_wakeup, page_nr;
Jens Axboe5274f052006-03-30 15:15:30 +0200189
190 ret = 0;
191 do_wakeup = 0;
Jens Axboe912d35f2006-04-26 10:59:21 +0200192 page_nr = 0;
Jens Axboe5274f052006-03-30 15:15:30 +0200193
Ingo Molnar3a326a22006-04-10 15:18:35 +0200194 if (pipe->inode)
195 mutex_lock(&pipe->inode->i_mutex);
Jens Axboe5274f052006-03-30 15:15:30 +0200196
Jens Axboe5274f052006-03-30 15:15:30 +0200197 for (;;) {
Ingo Molnar3a326a22006-04-10 15:18:35 +0200198 if (!pipe->readers) {
Jens Axboe5274f052006-03-30 15:15:30 +0200199 send_sig(SIGPIPE, current, 0);
200 if (!ret)
201 ret = -EPIPE;
202 break;
203 }
204
Jens Axboe6f767b02006-04-11 13:53:56 +0200205 if (pipe->nrbufs < PIPE_BUFFERS) {
206 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
Ingo Molnar3a326a22006-04-10 15:18:35 +0200207 struct pipe_buffer *buf = pipe->bufs + newbuf;
Jens Axboe5274f052006-03-30 15:15:30 +0200208
Jens Axboe912d35f2006-04-26 10:59:21 +0200209 buf->page = spd->pages[page_nr];
210 buf->offset = spd->partial[page_nr].offset;
211 buf->len = spd->partial[page_nr].len;
212 buf->ops = spd->ops;
Jens Axboe6f767b02006-04-11 13:53:56 +0200213 pipe->nrbufs++;
Jens Axboe912d35f2006-04-26 10:59:21 +0200214 page_nr++;
215 ret += buf->len;
216
Jens Axboe6f767b02006-04-11 13:53:56 +0200217 if (pipe->inode)
218 do_wakeup = 1;
Jens Axboe5274f052006-03-30 15:15:30 +0200219
Jens Axboe912d35f2006-04-26 10:59:21 +0200220 if (!--spd->nr_pages)
Jens Axboe5274f052006-03-30 15:15:30 +0200221 break;
Jens Axboe6f767b02006-04-11 13:53:56 +0200222 if (pipe->nrbufs < PIPE_BUFFERS)
Jens Axboe5274f052006-03-30 15:15:30 +0200223 continue;
224
225 break;
226 }
227
Jens Axboe912d35f2006-04-26 10:59:21 +0200228 if (spd->flags & SPLICE_F_NONBLOCK) {
Linus Torvalds29e35092006-04-02 12:46:35 -0700229 if (!ret)
230 ret = -EAGAIN;
231 break;
232 }
233
Jens Axboe5274f052006-03-30 15:15:30 +0200234 if (signal_pending(current)) {
235 if (!ret)
236 ret = -ERESTARTSYS;
237 break;
238 }
239
240 if (do_wakeup) {
Jens Axboec0bd1f62006-04-10 09:03:32 +0200241 smp_mb();
Ingo Molnar3a326a22006-04-10 15:18:35 +0200242 if (waitqueue_active(&pipe->wait))
243 wake_up_interruptible_sync(&pipe->wait);
244 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
Jens Axboe5274f052006-03-30 15:15:30 +0200245 do_wakeup = 0;
246 }
247
Ingo Molnar3a326a22006-04-10 15:18:35 +0200248 pipe->waiting_writers++;
249 pipe_wait(pipe);
250 pipe->waiting_writers--;
Jens Axboe5274f052006-03-30 15:15:30 +0200251 }
252
Ingo Molnar3a326a22006-04-10 15:18:35 +0200253 if (pipe->inode)
254 mutex_unlock(&pipe->inode->i_mutex);
Jens Axboe5274f052006-03-30 15:15:30 +0200255
256 if (do_wakeup) {
Jens Axboec0bd1f62006-04-10 09:03:32 +0200257 smp_mb();
Ingo Molnar3a326a22006-04-10 15:18:35 +0200258 if (waitqueue_active(&pipe->wait))
259 wake_up_interruptible(&pipe->wait);
260 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
Jens Axboe5274f052006-03-30 15:15:30 +0200261 }
262
Jens Axboe912d35f2006-04-26 10:59:21 +0200263 while (page_nr < spd->nr_pages)
264 page_cache_release(spd->pages[page_nr++]);
Jens Axboe5274f052006-03-30 15:15:30 +0200265
266 return ret;
267}
268
Ingo Molnar3a326a22006-04-10 15:18:35 +0200269static int
Jens Axboecbb7e572006-04-11 14:57:50 +0200270__generic_file_splice_read(struct file *in, loff_t *ppos,
271 struct pipe_inode_info *pipe, size_t len,
272 unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200273{
274 struct address_space *mapping = in->f_mapping;
Jens Axboe912d35f2006-04-26 10:59:21 +0200275 unsigned int loff, nr_pages;
Jens Axboe16c523d2006-04-10 09:03:58 +0200276 struct page *pages[PIPE_BUFFERS];
Jens Axboe912d35f2006-04-26 10:59:21 +0200277 struct partial_page partial[PIPE_BUFFERS];
Jens Axboe5274f052006-03-30 15:15:30 +0200278 struct page *page;
Jens Axboe91ad66e2006-04-19 15:55:10 +0200279 pgoff_t index, end_index;
280 loff_t isize;
Jens Axboe912d35f2006-04-26 10:59:21 +0200281 size_t total_len;
Jens Axboeeb207962006-04-27 11:05:22 +0200282 int error, page_nr;
Jens Axboe912d35f2006-04-26 10:59:21 +0200283 struct splice_pipe_desc spd = {
284 .pages = pages,
285 .partial = partial,
286 .flags = flags,
287 .ops = &page_cache_pipe_buf_ops,
288 };
Jens Axboe5274f052006-03-30 15:15:30 +0200289
Jens Axboecbb7e572006-04-11 14:57:50 +0200290 index = *ppos >> PAGE_CACHE_SHIFT;
Jens Axboe912d35f2006-04-26 10:59:21 +0200291 loff = *ppos & ~PAGE_CACHE_MASK;
292 nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
Jens Axboe5274f052006-03-30 15:15:30 +0200293
294 if (nr_pages > PIPE_BUFFERS)
295 nr_pages = PIPE_BUFFERS;
296
297 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200298 * Initiate read-ahead on this page range. however, don't call into
Jens Axboe0b749ce2006-04-10 09:05:04 +0200299 * read-ahead if this is a non-zero offset (we are likely doing small
300 * chunk splice and the page is already there) for a single page.
Jens Axboe5274f052006-03-30 15:15:30 +0200301 */
Jens Axboeeb645a22006-04-27 08:44:27 +0200302 if (!loff || nr_pages > 1)
303 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
Jens Axboe5274f052006-03-30 15:15:30 +0200304
305 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200306 * Now fill in the holes:
Jens Axboe5274f052006-03-30 15:15:30 +0200307 */
Jens Axboe7480a902006-04-11 13:52:47 +0200308 error = 0;
Jens Axboe912d35f2006-04-26 10:59:21 +0200309 total_len = 0;
Jens Axboeeb207962006-04-27 11:05:22 +0200310
311 /*
312 * Lookup the (hopefully) full range of pages we need.
313 */
314 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
315
316 /*
317 * If find_get_pages_contig() returned fewer pages than we needed,
318 * allocate the rest.
319 */
320 index += spd.nr_pages;
321 while (spd.nr_pages < nr_pages) {
322 /*
323 * Page could be there, find_get_pages_contig() breaks on
324 * the first hole.
325 */
326 page = find_get_page(mapping, index);
327 if (!page) {
328 /*
329 * page didn't exist, allocate one.
330 */
331 page = page_cache_alloc_cold(mapping);
332 if (!page)
333 break;
334
335 error = add_to_page_cache_lru(page, mapping, index,
336 mapping_gfp_mask(mapping));
337 if (unlikely(error)) {
338 page_cache_release(page);
339 break;
340 }
341 /*
342 * add_to_page_cache() locks the page, unlock it
343 * to avoid convoluting the logic below even more.
344 */
345 unlock_page(page);
346 }
347
348 pages[spd.nr_pages++] = page;
349 index++;
350 }
351
352 /*
353 * Now loop over the map and see if we need to start IO on any
354 * pages, fill in the partial map, etc.
355 */
356 index = *ppos >> PAGE_CACHE_SHIFT;
357 nr_pages = spd.nr_pages;
358 spd.nr_pages = 0;
359 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
Jens Axboe82aa5d62006-04-20 13:05:48 +0200360 unsigned int this_len;
361
362 if (!len)
363 break;
364
365 /*
366 * this_len is the max we'll use from this page
367 */
Andrew Mortonba5f5d92006-04-25 15:33:34 +0200368 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
Jens Axboeeb207962006-04-27 11:05:22 +0200369 page = pages[page_nr];
Jens Axboe7480a902006-04-11 13:52:47 +0200370
371 /*
372 * If the page isn't uptodate, we may need to start io on it
373 */
374 if (!PageUptodate(page)) {
Jens Axboec4f895c2006-04-19 15:56:12 +0200375 /*
376 * If in nonblock mode then dont block on waiting
377 * for an in-flight io page
378 */
379 if (flags & SPLICE_F_NONBLOCK)
380 break;
381
Jens Axboe7480a902006-04-11 13:52:47 +0200382 lock_page(page);
383
384 /*
385 * page was truncated, stop here. if this isn't the
386 * first page, we'll just complete what we already
387 * added
388 */
389 if (!page->mapping) {
390 unlock_page(page);
Jens Axboe7480a902006-04-11 13:52:47 +0200391 break;
392 }
393 /*
394 * page was already under io and is now done, great
395 */
396 if (PageUptodate(page)) {
397 unlock_page(page);
398 goto fill_it;
399 }
400
Jens Axboe7480a902006-04-11 13:52:47 +0200401 /*
402 * need to read in the page
403 */
404 error = mapping->a_ops->readpage(in, page);
Jens Axboe7480a902006-04-11 13:52:47 +0200405 if (unlikely(error)) {
Jens Axboeeb207962006-04-27 11:05:22 +0200406 /*
407 * We really should re-lookup the page here,
408 * but it complicates things a lot. Instead
409 * lets just do what we already stored, and
410 * we'll get it the next time we are called.
411 */
Jens Axboe7480a902006-04-11 13:52:47 +0200412 if (error == AOP_TRUNCATED_PAGE)
Jens Axboeeb207962006-04-27 11:05:22 +0200413 error = 0;
414
Jens Axboe7480a902006-04-11 13:52:47 +0200415 break;
416 }
Jens Axboe91ad66e2006-04-19 15:55:10 +0200417
418 /*
419 * i_size must be checked after ->readpage().
420 */
421 isize = i_size_read(mapping->host);
422 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
Jens Axboeeb207962006-04-27 11:05:22 +0200423 if (unlikely(!isize || index > end_index))
Jens Axboe91ad66e2006-04-19 15:55:10 +0200424 break;
Jens Axboe91ad66e2006-04-19 15:55:10 +0200425
426 /*
427 * if this is the last page, see if we need to shrink
428 * the length and stop
429 */
430 if (end_index == index) {
431 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
Jens Axboeeb207962006-04-27 11:05:22 +0200432 if (total_len + loff > isize)
Jens Axboe91ad66e2006-04-19 15:55:10 +0200433 break;
Jens Axboe91ad66e2006-04-19 15:55:10 +0200434 /*
435 * force quit after adding this page
436 */
Jens Axboeeb207962006-04-27 11:05:22 +0200437 len = this_len;
Jens Axboe82aa5d62006-04-20 13:05:48 +0200438 this_len = min(this_len, loff);
Jens Axboe912d35f2006-04-26 10:59:21 +0200439 loff = 0;
Jens Axboe91ad66e2006-04-19 15:55:10 +0200440 }
Jens Axboe7480a902006-04-11 13:52:47 +0200441 }
442fill_it:
Jens Axboeeb207962006-04-27 11:05:22 +0200443 partial[page_nr].offset = loff;
444 partial[page_nr].len = this_len;
Jens Axboe82aa5d62006-04-20 13:05:48 +0200445 len -= this_len;
Jens Axboe912d35f2006-04-26 10:59:21 +0200446 total_len += this_len;
Jens Axboe91ad66e2006-04-19 15:55:10 +0200447 loff = 0;
Jens Axboeeb207962006-04-27 11:05:22 +0200448 spd.nr_pages++;
449 index++;
Jens Axboe5274f052006-03-30 15:15:30 +0200450 }
451
Jens Axboeeb207962006-04-27 11:05:22 +0200452 /*
453 * Release any pages at the end, if we quit early. 'i' is how far
454 * we got, 'nr_pages' is how many pages are in the map.
455 */
456 while (page_nr < nr_pages)
457 page_cache_release(pages[page_nr++]);
458
Jens Axboe912d35f2006-04-26 10:59:21 +0200459 if (spd.nr_pages)
Jens Axboe00522fb2006-04-26 14:39:29 +0200460 return splice_to_pipe(pipe, &spd);
Jens Axboe5274f052006-03-30 15:15:30 +0200461
Jens Axboe7480a902006-04-11 13:52:47 +0200462 return error;
Jens Axboe5274f052006-03-30 15:15:30 +0200463}
464
Jens Axboe83f91352006-04-02 23:05:09 +0200465/**
466 * generic_file_splice_read - splice data from file to a pipe
467 * @in: file to splice from
468 * @pipe: pipe to splice to
469 * @len: number of bytes to splice
470 * @flags: splice modifier flags
471 *
472 * Will read pages from given file and fill them into a pipe.
Jens Axboe83f91352006-04-02 23:05:09 +0200473 */
Jens Axboecbb7e572006-04-11 14:57:50 +0200474ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
475 struct pipe_inode_info *pipe, size_t len,
476 unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200477{
478 ssize_t spliced;
479 int ret;
480
481 ret = 0;
482 spliced = 0;
Ingo Molnar3a326a22006-04-10 15:18:35 +0200483
Jens Axboe5274f052006-03-30 15:15:30 +0200484 while (len) {
Jens Axboecbb7e572006-04-11 14:57:50 +0200485 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
Jens Axboe5274f052006-03-30 15:15:30 +0200486
Jens Axboec4f895c2006-04-19 15:56:12 +0200487 if (ret < 0)
Jens Axboe5274f052006-03-30 15:15:30 +0200488 break;
Jens Axboec4f895c2006-04-19 15:56:12 +0200489 else if (!ret) {
490 if (spliced)
491 break;
492 if (flags & SPLICE_F_NONBLOCK) {
493 ret = -EAGAIN;
494 break;
495 }
496 }
Jens Axboe5274f052006-03-30 15:15:30 +0200497
Jens Axboecbb7e572006-04-11 14:57:50 +0200498 *ppos += ret;
Jens Axboe5274f052006-03-30 15:15:30 +0200499 len -= ret;
500 spliced += ret;
501 }
502
503 if (spliced)
504 return spliced;
505
506 return ret;
507}
508
Jens Axboe059a8f32006-04-02 23:06:05 +0200509EXPORT_SYMBOL(generic_file_splice_read);
510
Jens Axboe5274f052006-03-30 15:15:30 +0200511/*
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200512 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
Jens Axboe016b6612006-04-25 15:42:00 +0200513 * using sendpage(). Return the number of bytes sent.
Jens Axboe5274f052006-03-30 15:15:30 +0200514 */
515static int pipe_to_sendpage(struct pipe_inode_info *info,
516 struct pipe_buffer *buf, struct splice_desc *sd)
517{
518 struct file *file = sd->file;
519 loff_t pos = sd->pos;
Jens Axboe5274f052006-03-30 15:15:30 +0200520 ssize_t ret;
521 void *ptr;
Jens Axboeb2b39fa2006-04-02 23:05:41 +0200522 int more;
Jens Axboe5274f052006-03-30 15:15:30 +0200523
524 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200525 * Sub-optimal, but we are limited by the pipe ->map. We don't
Jens Axboe5274f052006-03-30 15:15:30 +0200526 * need a kmap'ed buffer here, we just want to make sure we
527 * have the page pinned if the pipe page originates from the
Ingo Molnar73d62d82006-04-11 13:57:21 +0200528 * page cache.
Jens Axboe5274f052006-03-30 15:15:30 +0200529 */
530 ptr = buf->ops->map(file, info, buf);
531 if (IS_ERR(ptr))
532 return PTR_ERR(ptr);
533
Jens Axboeb2b39fa2006-04-02 23:05:41 +0200534 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
Jens Axboe5274f052006-03-30 15:15:30 +0200535
Jens Axboe016b6612006-04-25 15:42:00 +0200536 ret = file->f_op->sendpage(file, buf->page, buf->offset, sd->len,
537 &pos, more);
Jens Axboe5274f052006-03-30 15:15:30 +0200538
539 buf->ops->unmap(info, buf);
Jens Axboe016b6612006-04-25 15:42:00 +0200540 return ret;
Jens Axboe5274f052006-03-30 15:15:30 +0200541}
542
543/*
544 * This is a little more tricky than the file -> pipe splicing. There are
545 * basically three cases:
546 *
547 * - Destination page already exists in the address space and there
548 * are users of it. For that case we have no other option that
549 * copying the data. Tough luck.
550 * - Destination page already exists in the address space, but there
551 * are no users of it. Make sure it's uptodate, then drop it. Fall
552 * through to last case.
553 * - Destination page does not exist, we can add the pipe page to
554 * the page cache and avoid the copy.
555 *
Jens Axboe83f91352006-04-02 23:05:09 +0200556 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
557 * sd->flags), we attempt to migrate pages from the pipe to the output
558 * file address space page cache. This is possible if no one else has
559 * the pipe page referenced outside of the pipe and page cache. If
560 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
561 * a new page in the output file page cache and fill/dirty that.
Jens Axboe5274f052006-03-30 15:15:30 +0200562 */
563static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
564 struct splice_desc *sd)
565{
566 struct file *file = sd->file;
567 struct address_space *mapping = file->f_mapping;
Jens Axboe3e7ee3e2006-04-02 23:11:04 +0200568 gfp_t gfp_mask = mapping_gfp_mask(mapping);
Jens Axboe016b6612006-04-25 15:42:00 +0200569 unsigned int offset, this_len;
Jens Axboe5274f052006-03-30 15:15:30 +0200570 struct page *page;
Jens Axboe5274f052006-03-30 15:15:30 +0200571 pgoff_t index;
Jens Axboe5abc97a2006-03-30 15:16:46 +0200572 char *src;
Jens Axboe3e7ee3e2006-04-02 23:11:04 +0200573 int ret;
Jens Axboe5274f052006-03-30 15:15:30 +0200574
575 /*
Jens Axboe49d0b212006-04-10 09:04:41 +0200576 * make sure the data in this buffer is uptodate
Jens Axboe5274f052006-03-30 15:15:30 +0200577 */
578 src = buf->ops->map(file, info, buf);
579 if (IS_ERR(src))
580 return PTR_ERR(src);
581
582 index = sd->pos >> PAGE_CACHE_SHIFT;
583 offset = sd->pos & ~PAGE_CACHE_MASK;
584
Jens Axboe016b6612006-04-25 15:42:00 +0200585 this_len = sd->len;
586 if (this_len + offset > PAGE_CACHE_SIZE)
587 this_len = PAGE_CACHE_SIZE - offset;
588
Jens Axboe5274f052006-03-30 15:15:30 +0200589 /*
Jens Axboe0568b402006-05-01 19:50:48 +0200590 * Reuse buf page, if SPLICE_F_MOVE is set and we are doing a full
591 * page.
Jens Axboe5274f052006-03-30 15:15:30 +0200592 */
Jens Axboe0568b402006-05-01 19:50:48 +0200593 if ((sd->flags & SPLICE_F_MOVE) && this_len == PAGE_CACHE_SIZE) {
Jens Axboe83f91352006-04-02 23:05:09 +0200594 /*
595 * If steal succeeds, buf->page is now pruned from the vm
Jens Axboe9e0267c2006-04-19 15:57:31 +0200596 * side (LRU and page cache) and we can reuse it. The page
597 * will also be looked on successful return.
Jens Axboe83f91352006-04-02 23:05:09 +0200598 */
Jens Axboe5abc97a2006-03-30 15:16:46 +0200599 if (buf->ops->steal(info, buf))
600 goto find_page;
Jens Axboe5274f052006-03-30 15:15:30 +0200601
Jens Axboe5abc97a2006-03-30 15:16:46 +0200602 page = buf->page;
Jens Axboe46e678c2006-04-30 16:36:32 +0200603 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
604 unlock_page(page);
Jens Axboe5abc97a2006-03-30 15:16:46 +0200605 goto find_page;
Jens Axboe46e678c2006-04-30 16:36:32 +0200606 }
Jens Axboe3e7ee3e2006-04-02 23:11:04 +0200607
Jens Axboe0568b402006-05-01 19:50:48 +0200608 page_cache_get(page);
609
Jens Axboe3e7ee3e2006-04-02 23:11:04 +0200610 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
611 lru_cache_add(page);
Jens Axboe5abc97a2006-03-30 15:16:46 +0200612 } else {
613find_page:
Jens Axboe9e0267c2006-04-19 15:57:31 +0200614 page = find_lock_page(mapping, index);
615 if (!page) {
616 ret = -ENOMEM;
617 page = page_cache_alloc_cold(mapping);
618 if (unlikely(!page))
619 goto out_nomem;
620
621 /*
622 * This will also lock the page
623 */
624 ret = add_to_page_cache_lru(page, mapping, index,
625 gfp_mask);
626 if (unlikely(ret))
627 goto out;
628 }
Jens Axboe5274f052006-03-30 15:15:30 +0200629
Jens Axboe5abc97a2006-03-30 15:16:46 +0200630 /*
Jens Axboe9e0267c2006-04-19 15:57:31 +0200631 * We get here with the page locked. If the page is also
632 * uptodate, we don't need to do more. If it isn't, we
633 * may need to bring it in if we are not going to overwrite
634 * the full page.
Jens Axboe5abc97a2006-03-30 15:16:46 +0200635 */
636 if (!PageUptodate(page)) {
Jens Axboe016b6612006-04-25 15:42:00 +0200637 if (this_len < PAGE_CACHE_SIZE) {
Jens Axboe5abc97a2006-03-30 15:16:46 +0200638 ret = mapping->a_ops->readpage(file, page);
639 if (unlikely(ret))
640 goto out;
641
642 lock_page(page);
643
644 if (!PageUptodate(page)) {
645 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200646 * Page got invalidated, repeat.
Jens Axboe5abc97a2006-03-30 15:16:46 +0200647 */
648 if (!page->mapping) {
649 unlock_page(page);
650 page_cache_release(page);
651 goto find_page;
652 }
653 ret = -EIO;
654 goto out;
Jens Axboe5274f052006-03-30 15:15:30 +0200655 }
Jens Axboe9e0267c2006-04-19 15:57:31 +0200656 } else
Jens Axboe5abc97a2006-03-30 15:16:46 +0200657 SetPageUptodate(page);
Jens Axboe5274f052006-03-30 15:15:30 +0200658 }
659 }
660
Jens Axboe016b6612006-04-25 15:42:00 +0200661 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200662 if (ret == AOP_TRUNCATED_PAGE) {
663 page_cache_release(page);
664 goto find_page;
665 } else if (ret)
Jens Axboe5274f052006-03-30 15:15:30 +0200666 goto out;
667
Jens Axboe0568b402006-05-01 19:50:48 +0200668 if (buf->page != page) {
Jens Axboe5abc97a2006-03-30 15:16:46 +0200669 char *dst = kmap_atomic(page, KM_USER0);
670
Jens Axboe016b6612006-04-25 15:42:00 +0200671 memcpy(dst + offset, src + buf->offset, this_len);
Jens Axboe5abc97a2006-03-30 15:16:46 +0200672 flush_dcache_page(page);
673 kunmap_atomic(dst, KM_USER0);
674 }
Jens Axboe5274f052006-03-30 15:15:30 +0200675
Jens Axboe016b6612006-04-25 15:42:00 +0200676 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
Jens Axboe0568b402006-05-01 19:50:48 +0200677 if (!ret) {
678 /*
679 * Return the number of bytes written and mark page as
680 * accessed, we are now done!
681 */
682 ret = this_len;
683 mark_page_accessed(page);
684 balance_dirty_pages_ratelimited(mapping);
685 } else if (ret == AOP_TRUNCATED_PAGE) {
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200686 page_cache_release(page);
687 goto find_page;
Jens Axboe0568b402006-05-01 19:50:48 +0200688 }
Jens Axboe5274f052006-03-30 15:15:30 +0200689out:
Jens Axboe0568b402006-05-01 19:50:48 +0200690 page_cache_release(page);
Jens Axboe9e0267c2006-04-19 15:57:31 +0200691 unlock_page(page);
Dave Jones9aefe432006-04-10 09:02:40 +0200692out_nomem:
Jens Axboe5274f052006-03-30 15:15:30 +0200693 buf->ops->unmap(info, buf);
694 return ret;
695}
696
Jens Axboe83f91352006-04-02 23:05:09 +0200697/*
698 * Pipe input worker. Most of this logic works like a regular pipe, the
699 * key here is the 'actor' worker passed in that actually moves the data
700 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
701 */
Jens Axboe00522fb2006-04-26 14:39:29 +0200702ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
703 loff_t *ppos, size_t len, unsigned int flags,
704 splice_actor *actor)
Jens Axboe5274f052006-03-30 15:15:30 +0200705{
Jens Axboe5274f052006-03-30 15:15:30 +0200706 int ret, do_wakeup, err;
707 struct splice_desc sd;
708
709 ret = 0;
710 do_wakeup = 0;
711
712 sd.total_len = len;
713 sd.flags = flags;
714 sd.file = out;
Jens Axboecbb7e572006-04-11 14:57:50 +0200715 sd.pos = *ppos;
Jens Axboe5274f052006-03-30 15:15:30 +0200716
Ingo Molnar3a326a22006-04-10 15:18:35 +0200717 if (pipe->inode)
718 mutex_lock(&pipe->inode->i_mutex);
Jens Axboe5274f052006-03-30 15:15:30 +0200719
Jens Axboe5274f052006-03-30 15:15:30 +0200720 for (;;) {
Jens Axboe6f767b02006-04-11 13:53:56 +0200721 if (pipe->nrbufs) {
722 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
Jens Axboe5274f052006-03-30 15:15:30 +0200723 struct pipe_buf_operations *ops = buf->ops;
724
725 sd.len = buf->len;
726 if (sd.len > sd.total_len)
727 sd.len = sd.total_len;
728
Ingo Molnar3a326a22006-04-10 15:18:35 +0200729 err = actor(pipe, buf, &sd);
Jens Axboe016b6612006-04-25 15:42:00 +0200730 if (err <= 0) {
Jens Axboe5274f052006-03-30 15:15:30 +0200731 if (!ret && err != -ENODATA)
732 ret = err;
733
734 break;
735 }
736
Jens Axboe016b6612006-04-25 15:42:00 +0200737 ret += err;
738 buf->offset += err;
739 buf->len -= err;
740
741 sd.len -= err;
742 sd.pos += err;
743 sd.total_len -= err;
744 if (sd.len)
745 continue;
Ingo Molnar73d62d82006-04-11 13:57:21 +0200746
Jens Axboe5274f052006-03-30 15:15:30 +0200747 if (!buf->len) {
748 buf->ops = NULL;
Ingo Molnar3a326a22006-04-10 15:18:35 +0200749 ops->release(pipe, buf);
Jens Axboe6f767b02006-04-11 13:53:56 +0200750 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
751 pipe->nrbufs--;
752 if (pipe->inode)
753 do_wakeup = 1;
Jens Axboe5274f052006-03-30 15:15:30 +0200754 }
755
Jens Axboe5274f052006-03-30 15:15:30 +0200756 if (!sd.total_len)
757 break;
758 }
759
Jens Axboe6f767b02006-04-11 13:53:56 +0200760 if (pipe->nrbufs)
Jens Axboe5274f052006-03-30 15:15:30 +0200761 continue;
Ingo Molnar3a326a22006-04-10 15:18:35 +0200762 if (!pipe->writers)
Jens Axboe5274f052006-03-30 15:15:30 +0200763 break;
Ingo Molnar3a326a22006-04-10 15:18:35 +0200764 if (!pipe->waiting_writers) {
Jens Axboe5274f052006-03-30 15:15:30 +0200765 if (ret)
766 break;
767 }
768
Linus Torvalds29e35092006-04-02 12:46:35 -0700769 if (flags & SPLICE_F_NONBLOCK) {
770 if (!ret)
771 ret = -EAGAIN;
772 break;
773 }
774
Jens Axboe5274f052006-03-30 15:15:30 +0200775 if (signal_pending(current)) {
776 if (!ret)
777 ret = -ERESTARTSYS;
778 break;
779 }
780
781 if (do_wakeup) {
Jens Axboec0bd1f62006-04-10 09:03:32 +0200782 smp_mb();
Ingo Molnar3a326a22006-04-10 15:18:35 +0200783 if (waitqueue_active(&pipe->wait))
784 wake_up_interruptible_sync(&pipe->wait);
785 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
Jens Axboe5274f052006-03-30 15:15:30 +0200786 do_wakeup = 0;
787 }
788
Ingo Molnar3a326a22006-04-10 15:18:35 +0200789 pipe_wait(pipe);
Jens Axboe5274f052006-03-30 15:15:30 +0200790 }
791
Ingo Molnar3a326a22006-04-10 15:18:35 +0200792 if (pipe->inode)
793 mutex_unlock(&pipe->inode->i_mutex);
Jens Axboe5274f052006-03-30 15:15:30 +0200794
795 if (do_wakeup) {
Jens Axboec0bd1f62006-04-10 09:03:32 +0200796 smp_mb();
Ingo Molnar3a326a22006-04-10 15:18:35 +0200797 if (waitqueue_active(&pipe->wait))
798 wake_up_interruptible(&pipe->wait);
799 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
Jens Axboe5274f052006-03-30 15:15:30 +0200800 }
801
Jens Axboe5274f052006-03-30 15:15:30 +0200802 return ret;
Jens Axboe5274f052006-03-30 15:15:30 +0200803}
804
Jens Axboe83f91352006-04-02 23:05:09 +0200805/**
806 * generic_file_splice_write - splice data from a pipe to a file
Ingo Molnar3a326a22006-04-10 15:18:35 +0200807 * @pipe: pipe info
Jens Axboe83f91352006-04-02 23:05:09 +0200808 * @out: file to write to
809 * @len: number of bytes to splice
810 * @flags: splice modifier flags
811 *
812 * Will either move or copy pages (determined by @flags options) from
813 * the given pipe inode to the given file.
814 *
815 */
Ingo Molnar3a326a22006-04-10 15:18:35 +0200816ssize_t
817generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
Jens Axboecbb7e572006-04-11 14:57:50 +0200818 loff_t *ppos, size_t len, unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200819{
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200820 struct address_space *mapping = out->f_mapping;
Ingo Molnar3a326a22006-04-10 15:18:35 +0200821 ssize_t ret;
822
Jens Axboe00522fb2006-04-26 14:39:29 +0200823 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
Jens Axboea4514eb2006-04-19 15:57:05 +0200824 if (ret > 0) {
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200825 struct inode *inode = mapping->host;
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200826
Jens Axboea4514eb2006-04-19 15:57:05 +0200827 *ppos += ret;
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200828
Jens Axboea4514eb2006-04-19 15:57:05 +0200829 /*
830 * If file or inode is SYNC and we actually wrote some data,
831 * sync it.
832 */
833 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
834 int err;
835
836 mutex_lock(&inode->i_mutex);
837 err = generic_osync_inode(inode, mapping,
838 OSYNC_METADATA|OSYNC_DATA);
839 mutex_unlock(&inode->i_mutex);
840
841 if (err)
842 ret = err;
843 }
Jens Axboe4f6f0bd2006-04-02 23:04:46 +0200844 }
845
846 return ret;
Jens Axboe5274f052006-03-30 15:15:30 +0200847}
848
Jens Axboe059a8f32006-04-02 23:06:05 +0200849EXPORT_SYMBOL(generic_file_splice_write);
850
Jens Axboe83f91352006-04-02 23:05:09 +0200851/**
852 * generic_splice_sendpage - splice data from a pipe to a socket
853 * @inode: pipe inode
854 * @out: socket to write to
855 * @len: number of bytes to splice
856 * @flags: splice modifier flags
857 *
858 * Will send @len bytes from the pipe to a network socket. No data copying
859 * is involved.
860 *
861 */
Ingo Molnar3a326a22006-04-10 15:18:35 +0200862ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
Jens Axboecbb7e572006-04-11 14:57:50 +0200863 loff_t *ppos, size_t len, unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200864{
Jens Axboe00522fb2006-04-26 14:39:29 +0200865 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
Jens Axboe5274f052006-03-30 15:15:30 +0200866}
867
Jens Axboe059a8f32006-04-02 23:06:05 +0200868EXPORT_SYMBOL(generic_splice_sendpage);
Jeff Garzika0f06782006-03-30 23:06:13 -0500869
Jens Axboe83f91352006-04-02 23:05:09 +0200870/*
871 * Attempt to initiate a splice from pipe to file.
872 */
Ingo Molnar3a326a22006-04-10 15:18:35 +0200873static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
Jens Axboecbb7e572006-04-11 14:57:50 +0200874 loff_t *ppos, size_t len, unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200875{
Jens Axboe5274f052006-03-30 15:15:30 +0200876 int ret;
877
Jens Axboe49570e92006-04-11 13:56:09 +0200878 if (unlikely(!out->f_op || !out->f_op->splice_write))
Jens Axboe5274f052006-03-30 15:15:30 +0200879 return -EINVAL;
880
Jens Axboe49570e92006-04-11 13:56:09 +0200881 if (unlikely(!(out->f_mode & FMODE_WRITE)))
Jens Axboe5274f052006-03-30 15:15:30 +0200882 return -EBADF;
883
Jens Axboecbb7e572006-04-11 14:57:50 +0200884 ret = rw_verify_area(WRITE, out, ppos, len);
Jens Axboe5274f052006-03-30 15:15:30 +0200885 if (unlikely(ret < 0))
886 return ret;
887
Jens Axboecbb7e572006-04-11 14:57:50 +0200888 return out->f_op->splice_write(pipe, out, ppos, len, flags);
Jens Axboe5274f052006-03-30 15:15:30 +0200889}
890
Jens Axboe83f91352006-04-02 23:05:09 +0200891/*
892 * Attempt to initiate a splice from a file to a pipe.
893 */
Jens Axboecbb7e572006-04-11 14:57:50 +0200894static long do_splice_to(struct file *in, loff_t *ppos,
895 struct pipe_inode_info *pipe, size_t len,
896 unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +0200897{
Jens Axboecbb7e572006-04-11 14:57:50 +0200898 loff_t isize, left;
Jens Axboe5274f052006-03-30 15:15:30 +0200899 int ret;
900
Jens Axboe49570e92006-04-11 13:56:09 +0200901 if (unlikely(!in->f_op || !in->f_op->splice_read))
Jens Axboe5274f052006-03-30 15:15:30 +0200902 return -EINVAL;
903
Jens Axboe49570e92006-04-11 13:56:09 +0200904 if (unlikely(!(in->f_mode & FMODE_READ)))
Jens Axboe5274f052006-03-30 15:15:30 +0200905 return -EBADF;
906
Jens Axboecbb7e572006-04-11 14:57:50 +0200907 ret = rw_verify_area(READ, in, ppos, len);
Jens Axboe5274f052006-03-30 15:15:30 +0200908 if (unlikely(ret < 0))
909 return ret;
910
911 isize = i_size_read(in->f_mapping->host);
Jens Axboecbb7e572006-04-11 14:57:50 +0200912 if (unlikely(*ppos >= isize))
Jens Axboe5274f052006-03-30 15:15:30 +0200913 return 0;
914
Jens Axboecbb7e572006-04-11 14:57:50 +0200915 left = isize - *ppos;
Jens Axboe49570e92006-04-11 13:56:09 +0200916 if (unlikely(left < len))
Jens Axboe5274f052006-03-30 15:15:30 +0200917 len = left;
918
Jens Axboecbb7e572006-04-11 14:57:50 +0200919 return in->f_op->splice_read(in, ppos, pipe, len, flags);
Jens Axboe5274f052006-03-30 15:15:30 +0200920}
921
Jens Axboecbb7e572006-04-11 14:57:50 +0200922long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
923 size_t len, unsigned int flags)
Jens Axboeb92ce552006-04-11 13:52:07 +0200924{
925 struct pipe_inode_info *pipe;
926 long ret, bytes;
Jens Axboecbb7e572006-04-11 14:57:50 +0200927 loff_t out_off;
Jens Axboeb92ce552006-04-11 13:52:07 +0200928 umode_t i_mode;
929 int i;
930
931 /*
932 * We require the input being a regular file, as we don't want to
933 * randomly drop data for eg socket -> socket splicing. Use the
934 * piped splicing for that!
935 */
936 i_mode = in->f_dentry->d_inode->i_mode;
937 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
938 return -EINVAL;
939
940 /*
941 * neither in nor out is a pipe, setup an internal pipe attached to
942 * 'out' and transfer the wanted data from 'in' to 'out' through that
943 */
944 pipe = current->splice_pipe;
Jens Axboe49570e92006-04-11 13:56:09 +0200945 if (unlikely(!pipe)) {
Jens Axboeb92ce552006-04-11 13:52:07 +0200946 pipe = alloc_pipe_info(NULL);
947 if (!pipe)
948 return -ENOMEM;
949
950 /*
951 * We don't have an immediate reader, but we'll read the stuff
Jens Axboe00522fb2006-04-26 14:39:29 +0200952 * out of the pipe right after the splice_to_pipe(). So set
Jens Axboeb92ce552006-04-11 13:52:07 +0200953 * PIPE_READERS appropriately.
954 */
955 pipe->readers = 1;
956
957 current->splice_pipe = pipe;
958 }
959
960 /*
Ingo Molnar73d62d82006-04-11 13:57:21 +0200961 * Do the splice.
Jens Axboeb92ce552006-04-11 13:52:07 +0200962 */
963 ret = 0;
964 bytes = 0;
Jens Axboecbb7e572006-04-11 14:57:50 +0200965 out_off = 0;
Jens Axboeb92ce552006-04-11 13:52:07 +0200966
967 while (len) {
968 size_t read_len, max_read_len;
969
970 /*
971 * Do at most PIPE_BUFFERS pages worth of transfer:
972 */
973 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
974
Jens Axboecbb7e572006-04-11 14:57:50 +0200975 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
Jens Axboeb92ce552006-04-11 13:52:07 +0200976 if (unlikely(ret < 0))
977 goto out_release;
978
979 read_len = ret;
980
981 /*
982 * NOTE: nonblocking mode only applies to the input. We
983 * must not do the output in nonblocking mode as then we
984 * could get stuck data in the internal pipe:
985 */
Jens Axboecbb7e572006-04-11 14:57:50 +0200986 ret = do_splice_from(pipe, out, &out_off, read_len,
Jens Axboeb92ce552006-04-11 13:52:07 +0200987 flags & ~SPLICE_F_NONBLOCK);
988 if (unlikely(ret < 0))
989 goto out_release;
990
991 bytes += ret;
992 len -= ret;
993
994 /*
995 * In nonblocking mode, if we got back a short read then
996 * that was due to either an IO error or due to the
997 * pagecache entry not being there. In the IO error case
998 * the _next_ splice attempt will produce a clean IO error
999 * return value (not a short read), so in both cases it's
1000 * correct to break out of the loop here:
1001 */
1002 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1003 break;
1004 }
1005
1006 pipe->nrbufs = pipe->curbuf = 0;
1007
1008 return bytes;
1009
1010out_release:
1011 /*
1012 * If we did an incomplete transfer we must release
1013 * the pipe buffers in question:
1014 */
1015 for (i = 0; i < PIPE_BUFFERS; i++) {
1016 struct pipe_buffer *buf = pipe->bufs + i;
1017
1018 if (buf->ops) {
1019 buf->ops->release(pipe, buf);
1020 buf->ops = NULL;
1021 }
1022 }
1023 pipe->nrbufs = pipe->curbuf = 0;
1024
1025 /*
1026 * If we transferred some data, return the number of bytes:
1027 */
1028 if (bytes > 0)
1029 return bytes;
1030
1031 return ret;
1032}
1033
1034EXPORT_SYMBOL(do_splice_direct);
1035
Jens Axboe83f91352006-04-02 23:05:09 +02001036/*
1037 * Determine where to splice to/from.
1038 */
Ingo Molnar529565d2006-04-10 15:18:58 +02001039static long do_splice(struct file *in, loff_t __user *off_in,
1040 struct file *out, loff_t __user *off_out,
1041 size_t len, unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +02001042{
Ingo Molnar3a326a22006-04-10 15:18:35 +02001043 struct pipe_inode_info *pipe;
Jens Axboecbb7e572006-04-11 14:57:50 +02001044 loff_t offset, *off;
Jens Axboea4514eb2006-04-19 15:57:05 +02001045 long ret;
Jens Axboe5274f052006-03-30 15:15:30 +02001046
Ingo Molnar3a326a22006-04-10 15:18:35 +02001047 pipe = in->f_dentry->d_inode->i_pipe;
Ingo Molnar529565d2006-04-10 15:18:58 +02001048 if (pipe) {
1049 if (off_in)
1050 return -ESPIPE;
Jens Axboeb92ce552006-04-11 13:52:07 +02001051 if (off_out) {
1052 if (out->f_op->llseek == no_llseek)
1053 return -EINVAL;
Jens Axboecbb7e572006-04-11 14:57:50 +02001054 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
Jens Axboeb92ce552006-04-11 13:52:07 +02001055 return -EFAULT;
Jens Axboecbb7e572006-04-11 14:57:50 +02001056 off = &offset;
1057 } else
1058 off = &out->f_pos;
Ingo Molnar529565d2006-04-10 15:18:58 +02001059
Jens Axboea4514eb2006-04-19 15:57:05 +02001060 ret = do_splice_from(pipe, out, off, len, flags);
1061
1062 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1063 ret = -EFAULT;
1064
1065 return ret;
Ingo Molnar529565d2006-04-10 15:18:58 +02001066 }
Jens Axboe5274f052006-03-30 15:15:30 +02001067
Ingo Molnar3a326a22006-04-10 15:18:35 +02001068 pipe = out->f_dentry->d_inode->i_pipe;
Ingo Molnar529565d2006-04-10 15:18:58 +02001069 if (pipe) {
1070 if (off_out)
1071 return -ESPIPE;
Jens Axboeb92ce552006-04-11 13:52:07 +02001072 if (off_in) {
1073 if (in->f_op->llseek == no_llseek)
1074 return -EINVAL;
Jens Axboecbb7e572006-04-11 14:57:50 +02001075 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
Jens Axboeb92ce552006-04-11 13:52:07 +02001076 return -EFAULT;
Jens Axboecbb7e572006-04-11 14:57:50 +02001077 off = &offset;
1078 } else
1079 off = &in->f_pos;
Ingo Molnar529565d2006-04-10 15:18:58 +02001080
Jens Axboea4514eb2006-04-19 15:57:05 +02001081 ret = do_splice_to(in, off, pipe, len, flags);
1082
1083 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1084 ret = -EFAULT;
1085
1086 return ret;
Ingo Molnar529565d2006-04-10 15:18:58 +02001087 }
Jens Axboe5274f052006-03-30 15:15:30 +02001088
1089 return -EINVAL;
1090}
1091
Jens Axboe912d35f2006-04-26 10:59:21 +02001092/*
1093 * Map an iov into an array of pages and offset/length tupples. With the
1094 * partial_page structure, we can map several non-contiguous ranges into
1095 * our ones pages[] map instead of splitting that operation into pieces.
1096 * Could easily be exported as a generic helper for other users, in which
1097 * case one would probably want to add a 'max_nr_pages' parameter as well.
1098 */
1099static int get_iovec_page_array(const struct iovec __user *iov,
1100 unsigned int nr_vecs, struct page **pages,
1101 struct partial_page *partial)
1102{
1103 int buffers = 0, error = 0;
1104
1105 /*
1106 * It's ok to take the mmap_sem for reading, even
1107 * across a "get_user()".
1108 */
1109 down_read(&current->mm->mmap_sem);
1110
1111 while (nr_vecs) {
1112 unsigned long off, npages;
1113 void __user *base;
1114 size_t len;
1115 int i;
1116
1117 /*
1118 * Get user address base and length for this iovec.
1119 */
1120 error = get_user(base, &iov->iov_base);
1121 if (unlikely(error))
1122 break;
1123 error = get_user(len, &iov->iov_len);
1124 if (unlikely(error))
1125 break;
1126
1127 /*
1128 * Sanity check this iovec. 0 read succeeds.
1129 */
1130 if (unlikely(!len))
1131 break;
1132 error = -EFAULT;
1133 if (unlikely(!base))
1134 break;
1135
1136 /*
1137 * Get this base offset and number of pages, then map
1138 * in the user pages.
1139 */
1140 off = (unsigned long) base & ~PAGE_MASK;
1141 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1142 if (npages > PIPE_BUFFERS - buffers)
1143 npages = PIPE_BUFFERS - buffers;
1144
1145 error = get_user_pages(current, current->mm,
1146 (unsigned long) base, npages, 0, 0,
1147 &pages[buffers], NULL);
1148
1149 if (unlikely(error <= 0))
1150 break;
1151
1152 /*
1153 * Fill this contiguous range into the partial page map.
1154 */
1155 for (i = 0; i < error; i++) {
1156 const int plen = min_t(size_t, len, PAGE_SIZE) - off;
1157
1158 partial[buffers].offset = off;
1159 partial[buffers].len = plen;
1160
1161 off = 0;
1162 len -= plen;
1163 buffers++;
1164 }
1165
1166 /*
1167 * We didn't complete this iov, stop here since it probably
1168 * means we have to move some of this into a pipe to
1169 * be able to continue.
1170 */
1171 if (len)
1172 break;
1173
1174 /*
1175 * Don't continue if we mapped fewer pages than we asked for,
1176 * or if we mapped the max number of pages that we have
1177 * room for.
1178 */
1179 if (error < npages || buffers == PIPE_BUFFERS)
1180 break;
1181
1182 nr_vecs--;
1183 iov++;
1184 }
1185
1186 up_read(&current->mm->mmap_sem);
1187
1188 if (buffers)
1189 return buffers;
1190
1191 return error;
1192}
1193
1194/*
1195 * vmsplice splices a user address range into a pipe. It can be thought of
1196 * as splice-from-memory, where the regular splice is splice-from-file (or
1197 * to file). In both cases the output is a pipe, naturally.
1198 *
1199 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1200 * not the other way around. Splicing from user memory is a simple operation
1201 * that can be supported without any funky alignment restrictions or nasty
1202 * vm tricks. We simply map in the user memory and fill them into a pipe.
1203 * The reverse isn't quite as easy, though. There are two possible solutions
1204 * for that:
1205 *
1206 * - memcpy() the data internally, at which point we might as well just
1207 * do a regular read() on the buffer anyway.
1208 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1209 * has restriction limitations on both ends of the pipe).
1210 *
1211 * Alas, it isn't here.
1212 *
1213 */
1214static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1215 unsigned long nr_segs, unsigned int flags)
1216{
1217 struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1218 struct page *pages[PIPE_BUFFERS];
1219 struct partial_page partial[PIPE_BUFFERS];
1220 struct splice_pipe_desc spd = {
1221 .pages = pages,
1222 .partial = partial,
1223 .flags = flags,
1224 .ops = &user_page_pipe_buf_ops,
1225 };
1226
1227 if (unlikely(!pipe))
1228 return -EBADF;
1229 if (unlikely(nr_segs > UIO_MAXIOV))
1230 return -EINVAL;
1231 else if (unlikely(!nr_segs))
1232 return 0;
1233
1234 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial);
1235 if (spd.nr_pages <= 0)
1236 return spd.nr_pages;
1237
Jens Axboe00522fb2006-04-26 14:39:29 +02001238 return splice_to_pipe(pipe, &spd);
Jens Axboe912d35f2006-04-26 10:59:21 +02001239}
1240
1241asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1242 unsigned long nr_segs, unsigned int flags)
1243{
1244 struct file *file;
1245 long error;
1246 int fput;
1247
1248 error = -EBADF;
1249 file = fget_light(fd, &fput);
1250 if (file) {
1251 if (file->f_mode & FMODE_WRITE)
1252 error = do_vmsplice(file, iov, nr_segs, flags);
1253
1254 fput_light(file, fput);
1255 }
1256
1257 return error;
1258}
1259
Ingo Molnar529565d2006-04-10 15:18:58 +02001260asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1261 int fd_out, loff_t __user *off_out,
1262 size_t len, unsigned int flags)
Jens Axboe5274f052006-03-30 15:15:30 +02001263{
1264 long error;
1265 struct file *in, *out;
1266 int fput_in, fput_out;
1267
1268 if (unlikely(!len))
1269 return 0;
1270
1271 error = -EBADF;
Ingo Molnar529565d2006-04-10 15:18:58 +02001272 in = fget_light(fd_in, &fput_in);
Jens Axboe5274f052006-03-30 15:15:30 +02001273 if (in) {
1274 if (in->f_mode & FMODE_READ) {
Ingo Molnar529565d2006-04-10 15:18:58 +02001275 out = fget_light(fd_out, &fput_out);
Jens Axboe5274f052006-03-30 15:15:30 +02001276 if (out) {
1277 if (out->f_mode & FMODE_WRITE)
Ingo Molnar529565d2006-04-10 15:18:58 +02001278 error = do_splice(in, off_in,
1279 out, off_out,
1280 len, flags);
Jens Axboe5274f052006-03-30 15:15:30 +02001281 fput_light(out, fput_out);
1282 }
1283 }
1284
1285 fput_light(in, fput_in);
1286 }
1287
1288 return error;
1289}
Jens Axboe70524492006-04-11 15:51:17 +02001290
1291/*
1292 * Link contents of ipipe to opipe.
1293 */
1294static int link_pipe(struct pipe_inode_info *ipipe,
1295 struct pipe_inode_info *opipe,
1296 size_t len, unsigned int flags)
1297{
1298 struct pipe_buffer *ibuf, *obuf;
Jens Axboe2a272502006-04-19 15:56:40 +02001299 int ret, do_wakeup, i, ipipe_first;
1300
1301 ret = do_wakeup = ipipe_first = 0;
Jens Axboe70524492006-04-11 15:51:17 +02001302
1303 /*
1304 * Potential ABBA deadlock, work around it by ordering lock
1305 * grabbing by inode address. Otherwise two different processes
1306 * could deadlock (one doing tee from A -> B, the other from B -> A).
1307 */
1308 if (ipipe->inode < opipe->inode) {
Jens Axboe2a272502006-04-19 15:56:40 +02001309 ipipe_first = 1;
Jens Axboe70524492006-04-11 15:51:17 +02001310 mutex_lock(&ipipe->inode->i_mutex);
1311 mutex_lock(&opipe->inode->i_mutex);
1312 } else {
1313 mutex_lock(&opipe->inode->i_mutex);
1314 mutex_lock(&ipipe->inode->i_mutex);
1315 }
1316
1317 for (i = 0;; i++) {
1318 if (!opipe->readers) {
1319 send_sig(SIGPIPE, current, 0);
1320 if (!ret)
1321 ret = -EPIPE;
1322 break;
1323 }
1324 if (ipipe->nrbufs - i) {
1325 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1326
1327 /*
1328 * If we have room, fill this buffer
1329 */
1330 if (opipe->nrbufs < PIPE_BUFFERS) {
1331 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1332
1333 /*
1334 * Get a reference to this pipe buffer,
1335 * so we can copy the contents over.
1336 */
1337 ibuf->ops->get(ipipe, ibuf);
1338
1339 obuf = opipe->bufs + nbuf;
1340 *obuf = *ibuf;
1341
1342 if (obuf->len > len)
1343 obuf->len = len;
1344
1345 opipe->nrbufs++;
1346 do_wakeup = 1;
1347 ret += obuf->len;
1348 len -= obuf->len;
1349
1350 if (!len)
1351 break;
1352 if (opipe->nrbufs < PIPE_BUFFERS)
1353 continue;
1354 }
1355
1356 /*
1357 * We have input available, but no output room.
Jens Axboe2a272502006-04-19 15:56:40 +02001358 * If we already copied data, return that. If we
1359 * need to drop the opipe lock, it must be ordered
1360 * last to avoid deadlocks.
Jens Axboe70524492006-04-11 15:51:17 +02001361 */
Jens Axboe2a272502006-04-19 15:56:40 +02001362 if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
Jens Axboe70524492006-04-11 15:51:17 +02001363 if (!ret)
1364 ret = -EAGAIN;
1365 break;
1366 }
1367 if (signal_pending(current)) {
1368 if (!ret)
1369 ret = -ERESTARTSYS;
1370 break;
1371 }
1372 if (do_wakeup) {
1373 smp_mb();
1374 if (waitqueue_active(&opipe->wait))
1375 wake_up_interruptible(&opipe->wait);
1376 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1377 do_wakeup = 0;
1378 }
1379
1380 opipe->waiting_writers++;
1381 pipe_wait(opipe);
1382 opipe->waiting_writers--;
1383 continue;
1384 }
1385
1386 /*
1387 * No input buffers, do the usual checks for available
1388 * writers and blocking and wait if necessary
1389 */
1390 if (!ipipe->writers)
1391 break;
1392 if (!ipipe->waiting_writers) {
1393 if (ret)
1394 break;
1395 }
Jens Axboe2a272502006-04-19 15:56:40 +02001396 /*
1397 * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1398 * with another process, we can only safely do that if
1399 * the ipipe lock is ordered last.
1400 */
1401 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
Jens Axboe70524492006-04-11 15:51:17 +02001402 if (!ret)
1403 ret = -EAGAIN;
1404 break;
1405 }
1406 if (signal_pending(current)) {
1407 if (!ret)
1408 ret = -ERESTARTSYS;
1409 break;
1410 }
1411
1412 if (waitqueue_active(&ipipe->wait))
1413 wake_up_interruptible_sync(&ipipe->wait);
1414 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1415
1416 pipe_wait(ipipe);
1417 }
1418
1419 mutex_unlock(&ipipe->inode->i_mutex);
1420 mutex_unlock(&opipe->inode->i_mutex);
1421
1422 if (do_wakeup) {
1423 smp_mb();
1424 if (waitqueue_active(&opipe->wait))
1425 wake_up_interruptible(&opipe->wait);
1426 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1427 }
1428
1429 return ret;
1430}
1431
1432/*
1433 * This is a tee(1) implementation that works on pipes. It doesn't copy
1434 * any data, it simply references the 'in' pages on the 'out' pipe.
1435 * The 'flags' used are the SPLICE_F_* variants, currently the only
1436 * applicable one is SPLICE_F_NONBLOCK.
1437 */
1438static long do_tee(struct file *in, struct file *out, size_t len,
1439 unsigned int flags)
1440{
1441 struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1442 struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1443
1444 /*
1445 * Link ipipe to the two output pipes, consuming as we go along.
1446 */
1447 if (ipipe && opipe)
1448 return link_pipe(ipipe, opipe, len, flags);
1449
1450 return -EINVAL;
1451}
1452
1453asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1454{
1455 struct file *in;
1456 int error, fput_in;
1457
1458 if (unlikely(!len))
1459 return 0;
1460
1461 error = -EBADF;
1462 in = fget_light(fdin, &fput_in);
1463 if (in) {
1464 if (in->f_mode & FMODE_READ) {
1465 int fput_out;
1466 struct file *out = fget_light(fdout, &fput_out);
1467
1468 if (out) {
1469 if (out->f_mode & FMODE_WRITE)
1470 error = do_tee(in, out, len, flags);
1471 fput_light(out, fput_out);
1472 }
1473 }
1474 fput_light(in, fput_in);
1475 }
1476
1477 return error;
1478}