2 * mm/readahead.c - address_space-level file readahead.
4 * Copyright (C) 2002, Linus Torvalds
6 * 09Apr2002 Andrew Morton
10 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/blkdev.h>
15 #include <linux/backing-dev.h>
16 #include <linux/task_io_accounting_ops.h>
17 #include <linux/pagevec.h>
18 #include <linux/pagemap.h>
21 * Initialise a struct file's readahead state. Assumes that the caller has
25 file_ra_state_init(struct file_ra_state
*ra
, struct address_space
*mapping
)
27 ra
->ra_pages
= mapping
->backing_dev_info
->ra_pages
;
30 EXPORT_SYMBOL_GPL(file_ra_state_init
);
32 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
35 * see if a page needs releasing upon read_cache_pages() failure
36 * - the caller of read_cache_pages() may have set PG_private or PG_fscache
37 * before calling, such as the NFS fs marking pages that are cached locally
38 * on disk, thus we need to give the fs a chance to clean up in the event of
41 static void read_cache_pages_invalidate_page(struct address_space
*mapping
,
44 if (page_has_private(page
)) {
45 if (!trylock_page(page
))
47 page
->mapping
= mapping
;
48 do_invalidatepage(page
, 0);
52 page_cache_release(page
);
56 * release a list of pages, invalidating them first if need be
58 static void read_cache_pages_invalidate_pages(struct address_space
*mapping
,
59 struct list_head
*pages
)
63 while (!list_empty(pages
)) {
64 victim
= list_to_page(pages
);
65 list_del(&victim
->lru
);
66 read_cache_pages_invalidate_page(mapping
, victim
);
71 * read_cache_pages - populate an address space with some pages & start reads against them
72 * @mapping: the address_space
73 * @pages: The address of a list_head which contains the target pages. These
74 * pages have their ->index populated and are otherwise uninitialised.
75 * @filler: callback routine for filling a single page.
76 * @data: private data for the callback routine.
78 * Hides the details of the LRU cache etc from the filesystems.
80 int read_cache_pages(struct address_space
*mapping
, struct list_head
*pages
,
81 int (*filler
)(void *, struct page
*), void *data
)
86 while (!list_empty(pages
)) {
87 page
= list_to_page(pages
);
89 if (add_to_page_cache_lru(page
, mapping
,
90 page
->index
, GFP_KERNEL
)) {
91 read_cache_pages_invalidate_page(mapping
, page
);
94 page_cache_release(page
);
96 ret
= filler(data
, page
);
98 read_cache_pages_invalidate_pages(mapping
, pages
);
101 task_io_account_read(PAGE_CACHE_SIZE
);
106 EXPORT_SYMBOL(read_cache_pages
);
108 static int read_pages(struct address_space
*mapping
, struct file
*filp
,
109 struct list_head
*pages
, unsigned nr_pages
)
114 if (mapping
->a_ops
->readpages
) {
115 ret
= mapping
->a_ops
->readpages(filp
, mapping
, pages
, nr_pages
);
116 /* Clean up the remaining pages */
117 put_pages_list(pages
);
121 for (page_idx
= 0; page_idx
< nr_pages
; page_idx
++) {
122 struct page
*page
= list_to_page(pages
);
123 list_del(&page
->lru
);
124 if (!add_to_page_cache_lru(page
, mapping
,
125 page
->index
, GFP_KERNEL
)) {
126 mapping
->a_ops
->readpage(filp
, page
);
128 page_cache_release(page
);
136 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
137 * the pages first, then submits them all for I/O. This avoids the very bad
138 * behaviour which would occur if page allocations are causing VM writeback.
139 * We really don't want to intermingle reads and writes like that.
141 * Returns the number of pages requested, or the maximum amount of I/O allowed.
143 * do_page_cache_readahead() returns -1 if it encountered request queue
147 __do_page_cache_readahead(struct address_space
*mapping
, struct file
*filp
,
148 pgoff_t offset
, unsigned long nr_to_read
,
149 unsigned long lookahead_size
)
151 struct inode
*inode
= mapping
->host
;
153 unsigned long end_index
; /* The last page we want to read */
154 LIST_HEAD(page_pool
);
157 loff_t isize
= i_size_read(inode
);
162 end_index
= ((isize
- 1) >> PAGE_CACHE_SHIFT
);
165 * Preallocate as many pages as we will need.
167 for (page_idx
= 0; page_idx
< nr_to_read
; page_idx
++) {
168 pgoff_t page_offset
= offset
+ page_idx
;
170 if (page_offset
> end_index
)
174 page
= radix_tree_lookup(&mapping
->page_tree
, page_offset
);
179 page
= page_cache_alloc_cold(mapping
);
182 page
->index
= page_offset
;
183 list_add(&page
->lru
, &page_pool
);
184 if (page_idx
== nr_to_read
- lookahead_size
)
185 SetPageReadahead(page
);
190 * Now start the IO. We ignore I/O errors - if the page is not
191 * uptodate then the caller will launch readpage again, and
192 * will then handle the error.
195 read_pages(mapping
, filp
, &page_pool
, ret
);
196 BUG_ON(!list_empty(&page_pool
));
202 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
205 int force_page_cache_readahead(struct address_space
*mapping
, struct file
*filp
,
206 pgoff_t offset
, unsigned long nr_to_read
)
210 if (unlikely(!mapping
->a_ops
->readpage
&& !mapping
->a_ops
->readpages
))
213 nr_to_read
= max_sane_readahead(nr_to_read
);
217 unsigned long this_chunk
= (2 * 1024 * 1024) / PAGE_CACHE_SIZE
;
219 if (this_chunk
> nr_to_read
)
220 this_chunk
= nr_to_read
;
221 err
= __do_page_cache_readahead(mapping
, filp
,
222 offset
, this_chunk
, 0);
228 offset
+= this_chunk
;
229 nr_to_read
-= this_chunk
;
235 * This version skips the IO if the queue is read-congested, and will tell the
236 * block layer to abandon the readahead if request allocation would block.
238 * force_page_cache_readahead() will ignore queue congestion and will block on
241 int do_page_cache_readahead(struct address_space
*mapping
, struct file
*filp
,
242 pgoff_t offset
, unsigned long nr_to_read
)
244 if (bdi_read_congested(mapping
->backing_dev_info
))
247 return __do_page_cache_readahead(mapping
, filp
, offset
, nr_to_read
, 0);
251 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
252 * sensible upper limit.
254 unsigned long max_sane_readahead(unsigned long nr
)
256 return min(nr
, (node_page_state(numa_node_id(), NR_INACTIVE_FILE
)
257 + node_page_state(numa_node_id(), NR_FREE_PAGES
)) / 2);
261 * Submit IO for the read-ahead request in file_ra_state.
263 static unsigned long ra_submit(struct file_ra_state
*ra
,
264 struct address_space
*mapping
, struct file
*filp
)
268 actual
= __do_page_cache_readahead(mapping
, filp
,
269 ra
->start
, ra
->size
, ra
->async_size
);
275 * Set the initial window size, round to next power of 2 and square
276 * for small size, x 4 for medium, and x 2 for large
277 * for 128k (32 page) max ra
278 * 1-8 page = 32k initial, > 8 page = 128k initial
280 static unsigned long get_init_ra_size(unsigned long size
, unsigned long max
)
282 unsigned long newsize
= roundup_pow_of_two(size
);
284 if (newsize
<= max
/ 32)
285 newsize
= newsize
* 4;
286 else if (newsize
<= max
/ 4)
287 newsize
= newsize
* 2;
295 * Get the previous window size, ramp it up, and
296 * return it as the new window size.
298 static unsigned long get_next_ra_size(struct file_ra_state
*ra
,
301 unsigned long cur
= ra
->size
;
302 unsigned long newsize
;
309 return min(newsize
, max
);
313 * On-demand readahead design.
315 * The fields in struct file_ra_state represent the most-recently-executed
318 * |<----- async_size ---------|
319 * |------------------- size -------------------->|
320 * |==================#===========================|
321 * ^start ^page marked with PG_readahead
323 * To overlap application thinking time and disk I/O time, we do
324 * `readahead pipelining': Do not wait until the application consumed all
325 * readahead pages and stalled on the missing page at readahead_index;
326 * Instead, submit an asynchronous readahead I/O as soon as there are
327 * only async_size pages left in the readahead window. Normally async_size
328 * will be equal to size, for maximum pipelining.
330 * In interleaved sequential reads, concurrent streams on the same fd can
331 * be invalidating each other's readahead state. So we flag the new readahead
332 * page at (start+size-async_size) with PG_readahead, and use it as readahead
333 * indicator. The flag won't be set on already cached pages, to avoid the
334 * readahead-for-nothing fuss, saving pointless page cache lookups.
336 * prev_pos tracks the last visited byte in the _previous_ read request.
337 * It should be maintained by the caller, and will be used for detecting
338 * small random reads. Note that the readahead algorithm checks loosely
339 * for sequential patterns. Hence interleaved reads might be served as
342 * There is a special-case: if the first page which the application tries to
343 * read happens to be the first page of the file, it is assumed that a linear
344 * read is about to happen and the window is immediately set to the initial size
345 * based on I/O request size and the max_readahead.
347 * The code ramps up the readahead size aggressively at first, but slow down as
348 * it approaches max_readhead.
352 * A minimal readahead algorithm for trivial sequential/random reads.
355 ondemand_readahead(struct address_space
*mapping
,
356 struct file_ra_state
*ra
, struct file
*filp
,
357 bool hit_readahead_marker
, pgoff_t offset
,
358 unsigned long req_size
)
360 unsigned long max
= max_sane_readahead(ra
->ra_pages
);
365 * It's the expected callback offset, assume sequential access.
366 * Ramp up sizes, and push forward the readahead window.
368 if (offset
&& (offset
== (ra
->start
+ ra
->size
- ra
->async_size
) ||
369 offset
== (ra
->start
+ ra
->size
))) {
370 ra
->start
+= ra
->size
;
371 ra
->size
= get_next_ra_size(ra
, max
);
372 ra
->async_size
= ra
->size
;
376 prev_offset
= ra
->prev_pos
>> PAGE_CACHE_SHIFT
;
377 sequential
= offset
- prev_offset
<= 1UL || req_size
> max
;
380 * Standalone, small read.
381 * Read as is, and do not pollute the readahead state.
383 if (!hit_readahead_marker
&& !sequential
) {
384 return __do_page_cache_readahead(mapping
, filp
,
385 offset
, req_size
, 0);
389 * Hit a marked page without valid readahead state.
390 * E.g. interleaved reads.
391 * Query the pagecache for async_size, which normally equals to
392 * readahead size. Ramp it up and use it as the new readahead size.
394 if (hit_readahead_marker
) {
398 start
= radix_tree_next_hole(&mapping
->page_tree
, offset
+1,max
);
401 if (!start
|| start
- offset
> max
)
405 ra
->size
= start
- offset
; /* old async_size */
406 ra
->size
+= req_size
;
407 ra
->size
= get_next_ra_size(ra
, max
);
408 ra
->async_size
= ra
->size
;
414 * - first read on start of file
415 * - sequential cache miss
416 * - oversize random read
417 * Start readahead for it.
420 ra
->size
= get_init_ra_size(req_size
, max
);
421 ra
->async_size
= ra
->size
> req_size
? ra
->size
- req_size
: ra
->size
;
425 * Will this read hit the readahead marker made by itself?
426 * If so, trigger the readahead marker hit now, and merge
427 * the resulted next readahead window into the current one.
429 if (offset
== ra
->start
&& ra
->size
== ra
->async_size
) {
430 ra
->async_size
= get_next_ra_size(ra
, max
);
431 ra
->size
+= ra
->async_size
;
434 return ra_submit(ra
, mapping
, filp
);
438 * page_cache_sync_readahead - generic file readahead
439 * @mapping: address_space which holds the pagecache and I/O vectors
440 * @ra: file_ra_state which holds the readahead state
441 * @filp: passed on to ->readpage() and ->readpages()
442 * @offset: start offset into @mapping, in pagecache page-sized units
443 * @req_size: hint: total size of the read which the caller is performing in
446 * page_cache_sync_readahead() should be called when a cache miss happened:
447 * it will submit the read. The readahead logic may decide to piggyback more
448 * pages onto the read request if access patterns suggest it will improve
451 void page_cache_sync_readahead(struct address_space
*mapping
,
452 struct file_ra_state
*ra
, struct file
*filp
,
453 pgoff_t offset
, unsigned long req_size
)
460 ondemand_readahead(mapping
, ra
, filp
, false, offset
, req_size
);
462 EXPORT_SYMBOL_GPL(page_cache_sync_readahead
);
465 * page_cache_async_readahead - file readahead for marked pages
466 * @mapping: address_space which holds the pagecache and I/O vectors
467 * @ra: file_ra_state which holds the readahead state
468 * @filp: passed on to ->readpage() and ->readpages()
469 * @page: the page at @offset which has the PG_readahead flag set
470 * @offset: start offset into @mapping, in pagecache page-sized units
471 * @req_size: hint: total size of the read which the caller is performing in
474 * page_cache_async_ondemand() should be called when a page is used which
475 * has the PG_readahead flag; this is a marker to suggest that the application
476 * has used up enough of the readahead window that we should start pulling in
480 page_cache_async_readahead(struct address_space
*mapping
,
481 struct file_ra_state
*ra
, struct file
*filp
,
482 struct page
*page
, pgoff_t offset
,
483 unsigned long req_size
)
490 * Same bit is used for PG_readahead and PG_reclaim.
492 if (PageWriteback(page
))
495 ClearPageReadahead(page
);
498 * Defer asynchronous read-ahead on IO congestion.
500 if (bdi_read_congested(mapping
->backing_dev_info
))
504 ondemand_readahead(mapping
, ra
, filp
, true, offset
, req_size
);
506 EXPORT_SYMBOL_GPL(page_cache_async_readahead
);