1/*
2 * mm/readahead.c - address_space-level file readahead.
3 *
4 * Copyright (C) 2002, Linus Torvalds
5 *
6 * 09Apr2002	Andrew Morton
7 *		Initial version.
8 */
9
10#include <linux/kernel.h>
11#include <linux/gfp.h>
12#include <linux/export.h>
13#include <linux/blkdev.h>
14#include <linux/backing-dev.h>
15#include <linux/task_io_accounting_ops.h>
16#include <linux/pagevec.h>
17#include <linux/pagemap.h>
18#include <linux/syscalls.h>
19#include <linux/file.h>
20
21#include "internal.h"
22
23/*
24 * Initialise a struct file's readahead state.  Assumes that the caller has
25 * memset *ra to zero.
26 */
27void
28file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
29{
30	ra->ra_pages = inode_to_bdi(mapping->host)->ra_pages;
31	ra->prev_pos = -1;
32}
33EXPORT_SYMBOL_GPL(file_ra_state_init);
34
35#define list_to_page(head) (list_entry((head)->prev, struct page, lru))
36
37/*
38 * see if a page needs releasing upon read_cache_pages() failure
39 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
40 *   before calling, such as the NFS fs marking pages that are cached locally
41 *   on disk, thus we need to give the fs a chance to clean up in the event of
42 *   an error
43 */
44static void read_cache_pages_invalidate_page(struct address_space *mapping,
45					     struct page *page)
46{
47	if (page_has_private(page)) {
48		if (!trylock_page(page))
49			BUG();
50		page->mapping = mapping;
51		do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
52		page->mapping = NULL;
53		unlock_page(page);
54	}
55	page_cache_release(page);
56}
57
58/*
59 * release a list of pages, invalidating them first if need be
60 */
61static void read_cache_pages_invalidate_pages(struct address_space *mapping,
62					      struct list_head *pages)
63{
64	struct page *victim;
65
66	while (!list_empty(pages)) {
67		victim = list_to_page(pages);
68		list_del(&victim->lru);
69		read_cache_pages_invalidate_page(mapping, victim);
70	}
71}
72
73/**
74 * read_cache_pages - populate an address space with some pages & start reads against them
75 * @mapping: the address_space
76 * @pages: The address of a list_head which contains the target pages.  These
77 *   pages have their ->index populated and are otherwise uninitialised.
78 * @filler: callback routine for filling a single page.
79 * @data: private data for the callback routine.
80 *
81 * Hides the details of the LRU cache etc from the filesystems.
82 */
83int read_cache_pages(struct address_space *mapping, struct list_head *pages,
84			int (*filler)(void *, struct page *), void *data)
85{
86	struct page *page;
87	int ret = 0;
88
89	while (!list_empty(pages)) {
90		page = list_to_page(pages);
91		list_del(&page->lru);
92		if (add_to_page_cache_lru(page, mapping,
93					page->index, GFP_KERNEL)) {
94			read_cache_pages_invalidate_page(mapping, page);
95			continue;
96		}
97		page_cache_release(page);
98
99		ret = filler(data, page);
100		if (unlikely(ret)) {
101			read_cache_pages_invalidate_pages(mapping, pages);
102			break;
103		}
104		task_io_account_read(PAGE_CACHE_SIZE);
105	}
106	return ret;
107}
108
109EXPORT_SYMBOL(read_cache_pages);
110
111static int read_pages(struct address_space *mapping, struct file *filp,
112		struct list_head *pages, unsigned nr_pages)
113{
114	struct blk_plug plug;
115	unsigned page_idx;
116	int ret;
117
118	blk_start_plug(&plug);
119
120	if (mapping->a_ops->readpages) {
121		ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
122		/* Clean up the remaining pages */
123		put_pages_list(pages);
124		goto out;
125	}
126
127	for (page_idx = 0; page_idx < nr_pages; page_idx++) {
128		struct page *page = list_to_page(pages);
129		list_del(&page->lru);
130		if (!add_to_page_cache_lru(page, mapping,
131					page->index, GFP_KERNEL)) {
132			mapping->a_ops->readpage(filp, page);
133		}
134		page_cache_release(page);
135	}
136	ret = 0;
137
138out:
139	blk_finish_plug(&plug);
140
141	return ret;
142}
143
144/*
145 * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
146 * the pages first, then submits them all for I/O. This avoids the very bad
147 * behaviour which would occur if page allocations are causing VM writeback.
148 * We really don't want to intermingle reads and writes like that.
149 *
150 * Returns the number of pages requested, or the maximum amount of I/O allowed.
151 */
152int __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
153			pgoff_t offset, unsigned long nr_to_read,
154			unsigned long lookahead_size)
155{
156	struct inode *inode = mapping->host;
157	struct page *page;
158	unsigned long end_index;	/* The last page we want to read */
159	LIST_HEAD(page_pool);
160	int page_idx;
161	int ret = 0;
162	loff_t isize = i_size_read(inode);
163
164	if (isize == 0)
165		goto out;
166
167	end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
168
169	/*
170	 * Preallocate as many pages as we will need.
171	 */
172	for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
173		pgoff_t page_offset = offset + page_idx;
174
175		if (page_offset > end_index)
176			break;
177
178		rcu_read_lock();
179		page = radix_tree_lookup(&mapping->page_tree, page_offset);
180		rcu_read_unlock();
181		if (page && !radix_tree_exceptional_entry(page))
182			continue;
183
184		page = page_cache_alloc_readahead(mapping);
185		if (!page)
186			break;
187		page->index = page_offset;
188		list_add(&page->lru, &page_pool);
189		if (page_idx == nr_to_read - lookahead_size)
190			SetPageReadahead(page);
191		ret++;
192	}
193
194	/*
195	 * Now start the IO.  We ignore I/O errors - if the page is not
196	 * uptodate then the caller will launch readpage again, and
197	 * will then handle the error.
198	 */
199	if (ret)
200		read_pages(mapping, filp, &page_pool, ret);
201	BUG_ON(!list_empty(&page_pool));
202out:
203	return ret;
204}
205
206/*
207 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
208 * memory at once.
209 */
210int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
211		pgoff_t offset, unsigned long nr_to_read)
212{
213	if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
214		return -EINVAL;
215
216	nr_to_read = max_sane_readahead(nr_to_read);
217	while (nr_to_read) {
218		int err;
219
220		unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
221
222		if (this_chunk > nr_to_read)
223			this_chunk = nr_to_read;
224		err = __do_page_cache_readahead(mapping, filp,
225						offset, this_chunk, 0);
226		if (err < 0)
227			return err;
228
229		offset += this_chunk;
230		nr_to_read -= this_chunk;
231	}
232	return 0;
233}
234
235#define MAX_READAHEAD   ((512*4096)/PAGE_CACHE_SIZE)
236/*
237 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
238 * sensible upper limit.
239 */
240unsigned long max_sane_readahead(unsigned long nr)
241{
242	return min(nr, MAX_READAHEAD);
243}
244
245/*
246 * Set the initial window size, round to next power of 2 and square
247 * for small size, x 4 for medium, and x 2 for large
248 * for 128k (32 page) max ra
249 * 1-8 page = 32k initial, > 8 page = 128k initial
250 */
251static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
252{
253	unsigned long newsize = roundup_pow_of_two(size);
254
255	if (newsize <= max / 32)
256		newsize = newsize * 4;
257	else if (newsize <= max / 4)
258		newsize = newsize * 2;
259	else
260		newsize = max;
261
262	return newsize;
263}
264
265/*
266 *  Get the previous window size, ramp it up, and
267 *  return it as the new window size.
268 */
269static unsigned long get_next_ra_size(struct file_ra_state *ra,
270						unsigned long max)
271{
272	unsigned long cur = ra->size;
273	unsigned long newsize;
274
275	if (cur < max / 16)
276		newsize = 4 * cur;
277	else
278		newsize = 2 * cur;
279
280	return min(newsize, max);
281}
282
283/*
284 * On-demand readahead design.
285 *
286 * The fields in struct file_ra_state represent the most-recently-executed
287 * readahead attempt:
288 *
289 *                        |<----- async_size ---------|
290 *     |------------------- size -------------------->|
291 *     |==================#===========================|
292 *     ^start             ^page marked with PG_readahead
293 *
294 * To overlap application thinking time and disk I/O time, we do
295 * `readahead pipelining': Do not wait until the application consumed all
296 * readahead pages and stalled on the missing page at readahead_index;
297 * Instead, submit an asynchronous readahead I/O as soon as there are
298 * only async_size pages left in the readahead window. Normally async_size
299 * will be equal to size, for maximum pipelining.
300 *
301 * In interleaved sequential reads, concurrent streams on the same fd can
302 * be invalidating each other's readahead state. So we flag the new readahead
303 * page at (start+size-async_size) with PG_readahead, and use it as readahead
304 * indicator. The flag won't be set on already cached pages, to avoid the
305 * readahead-for-nothing fuss, saving pointless page cache lookups.
306 *
307 * prev_pos tracks the last visited byte in the _previous_ read request.
308 * It should be maintained by the caller, and will be used for detecting
309 * small random reads. Note that the readahead algorithm checks loosely
310 * for sequential patterns. Hence interleaved reads might be served as
311 * sequential ones.
312 *
313 * There is a special-case: if the first page which the application tries to
314 * read happens to be the first page of the file, it is assumed that a linear
315 * read is about to happen and the window is immediately set to the initial size
316 * based on I/O request size and the max_readahead.
317 *
318 * The code ramps up the readahead size aggressively at first, but slow down as
319 * it approaches max_readhead.
320 */
321
322/*
323 * Count contiguously cached pages from @offset-1 to @offset-@max,
324 * this count is a conservative estimation of
325 * 	- length of the sequential read sequence, or
326 * 	- thrashing threshold in memory tight systems
327 */
328static pgoff_t count_history_pages(struct address_space *mapping,
329				   pgoff_t offset, unsigned long max)
330{
331	pgoff_t head;
332
333	rcu_read_lock();
334	head = page_cache_prev_hole(mapping, offset - 1, max);
335	rcu_read_unlock();
336
337	return offset - 1 - head;
338}
339
340/*
341 * page cache context based read-ahead
342 */
343static int try_context_readahead(struct address_space *mapping,
344				 struct file_ra_state *ra,
345				 pgoff_t offset,
346				 unsigned long req_size,
347				 unsigned long max)
348{
349	pgoff_t size;
350
351	size = count_history_pages(mapping, offset, max);
352
353	/*
354	 * not enough history pages:
355	 * it could be a random read
356	 */
357	if (size <= req_size)
358		return 0;
359
360	/*
361	 * starts from beginning of file:
362	 * it is a strong indication of long-run stream (or whole-file-read)
363	 */
364	if (size >= offset)
365		size *= 2;
366
367	ra->start = offset;
368	ra->size = min(size + req_size, max);
369	ra->async_size = 1;
370
371	return 1;
372}
373
374/*
375 * A minimal readahead algorithm for trivial sequential/random reads.
376 */
377static unsigned long
378ondemand_readahead(struct address_space *mapping,
379		   struct file_ra_state *ra, struct file *filp,
380		   bool hit_readahead_marker, pgoff_t offset,
381		   unsigned long req_size)
382{
383	unsigned long max = max_sane_readahead(ra->ra_pages);
384	pgoff_t prev_offset;
385
386	/*
387	 * start of file
388	 */
389	if (!offset)
390		goto initial_readahead;
391
392	/*
393	 * It's the expected callback offset, assume sequential access.
394	 * Ramp up sizes, and push forward the readahead window.
395	 */
396	if ((offset == (ra->start + ra->size - ra->async_size) ||
397	     offset == (ra->start + ra->size))) {
398		ra->start += ra->size;
399		ra->size = get_next_ra_size(ra, max);
400		ra->async_size = ra->size;
401		goto readit;
402	}
403
404	/*
405	 * Hit a marked page without valid readahead state.
406	 * E.g. interleaved reads.
407	 * Query the pagecache for async_size, which normally equals to
408	 * readahead size. Ramp it up and use it as the new readahead size.
409	 */
410	if (hit_readahead_marker) {
411		pgoff_t start;
412
413		rcu_read_lock();
414		start = page_cache_next_hole(mapping, offset + 1, max);
415		rcu_read_unlock();
416
417		if (!start || start - offset > max)
418			return 0;
419
420		ra->start = start;
421		ra->size = start - offset;	/* old async_size */
422		ra->size += req_size;
423		ra->size = get_next_ra_size(ra, max);
424		ra->async_size = ra->size;
425		goto readit;
426	}
427
428	/*
429	 * oversize read
430	 */
431	if (req_size > max)
432		goto initial_readahead;
433
434	/*
435	 * sequential cache miss
436	 * trivial case: (offset - prev_offset) == 1
437	 * unaligned reads: (offset - prev_offset) == 0
438	 */
439	prev_offset = (unsigned long long)ra->prev_pos >> PAGE_CACHE_SHIFT;
440	if (offset - prev_offset <= 1UL)
441		goto initial_readahead;
442
443	/*
444	 * Query the page cache and look for the traces(cached history pages)
445	 * that a sequential stream would leave behind.
446	 */
447	if (try_context_readahead(mapping, ra, offset, req_size, max))
448		goto readit;
449
450	/*
451	 * standalone, small random read
452	 * Read as is, and do not pollute the readahead state.
453	 */
454	return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
455
456initial_readahead:
457	ra->start = offset;
458	ra->size = get_init_ra_size(req_size, max);
459	ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
460
461readit:
462	/*
463	 * Will this read hit the readahead marker made by itself?
464	 * If so, trigger the readahead marker hit now, and merge
465	 * the resulted next readahead window into the current one.
466	 */
467	if (offset == ra->start && ra->size == ra->async_size) {
468		ra->async_size = get_next_ra_size(ra, max);
469		ra->size += ra->async_size;
470	}
471
472	return ra_submit(ra, mapping, filp);
473}
474
475/**
476 * page_cache_sync_readahead - generic file readahead
477 * @mapping: address_space which holds the pagecache and I/O vectors
478 * @ra: file_ra_state which holds the readahead state
479 * @filp: passed on to ->readpage() and ->readpages()
480 * @offset: start offset into @mapping, in pagecache page-sized units
481 * @req_size: hint: total size of the read which the caller is performing in
482 *            pagecache pages
483 *
484 * page_cache_sync_readahead() should be called when a cache miss happened:
485 * it will submit the read.  The readahead logic may decide to piggyback more
486 * pages onto the read request if access patterns suggest it will improve
487 * performance.
488 */
489void page_cache_sync_readahead(struct address_space *mapping,
490			       struct file_ra_state *ra, struct file *filp,
491			       pgoff_t offset, unsigned long req_size)
492{
493	/* no read-ahead */
494	if (!ra->ra_pages)
495		return;
496
497	/* be dumb */
498	if (filp && (filp->f_mode & FMODE_RANDOM)) {
499		force_page_cache_readahead(mapping, filp, offset, req_size);
500		return;
501	}
502
503	/* do read-ahead */
504	ondemand_readahead(mapping, ra, filp, false, offset, req_size);
505}
506EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
507
508/**
509 * page_cache_async_readahead - file readahead for marked pages
510 * @mapping: address_space which holds the pagecache and I/O vectors
511 * @ra: file_ra_state which holds the readahead state
512 * @filp: passed on to ->readpage() and ->readpages()
513 * @page: the page at @offset which has the PG_readahead flag set
514 * @offset: start offset into @mapping, in pagecache page-sized units
515 * @req_size: hint: total size of the read which the caller is performing in
516 *            pagecache pages
517 *
518 * page_cache_async_readahead() should be called when a page is used which
519 * has the PG_readahead flag; this is a marker to suggest that the application
520 * has used up enough of the readahead window that we should start pulling in
521 * more pages.
522 */
523void
524page_cache_async_readahead(struct address_space *mapping,
525			   struct file_ra_state *ra, struct file *filp,
526			   struct page *page, pgoff_t offset,
527			   unsigned long req_size)
528{
529	/* no read-ahead */
530	if (!ra->ra_pages)
531		return;
532
533	/*
534	 * Same bit is used for PG_readahead and PG_reclaim.
535	 */
536	if (PageWriteback(page))
537		return;
538
539	ClearPageReadahead(page);
540
541	/*
542	 * Defer asynchronous read-ahead on IO congestion.
543	 */
544	if (bdi_read_congested(inode_to_bdi(mapping->host)))
545		return;
546
547	/* do read-ahead */
548	ondemand_readahead(mapping, ra, filp, true, offset, req_size);
549}
550EXPORT_SYMBOL_GPL(page_cache_async_readahead);
551
552static ssize_t
553do_readahead(struct address_space *mapping, struct file *filp,
554	     pgoff_t index, unsigned long nr)
555{
556	if (!mapping || !mapping->a_ops)
557		return -EINVAL;
558
559	return force_page_cache_readahead(mapping, filp, index, nr);
560}
561
562SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count)
563{
564	ssize_t ret;
565	struct fd f;
566
567	ret = -EBADF;
568	f = fdget(fd);
569	if (f.file) {
570		if (f.file->f_mode & FMODE_READ) {
571			struct address_space *mapping = f.file->f_mapping;
572			pgoff_t start = offset >> PAGE_CACHE_SHIFT;
573			pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
574			unsigned long len = end - start + 1;
575			ret = do_readahead(mapping, f.file, start, len);
576		}
577		fdput(f);
578	}
579	return ret;
580}
581