Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * mm/readahead.c - address_space-level file readahead. |
| 3 | * |
| 4 | * Copyright (C) 2002, Linus Torvalds |
| 5 | * |
Francois Cami | e1f8e87 | 2008-10-15 22:01:59 -0700 | [diff] [blame] | 6 | * 09Apr2002 Andrew Morton |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 7 | * Initial version. |
| 8 | */ |
| 9 | |
| 10 | #include <linux/kernel.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 11 | #include <linux/gfp.h> |
Paul Gortmaker | b95f1b31 | 2011-10-16 02:01:52 -0400 | [diff] [blame] | 12 | #include <linux/export.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 13 | #include <linux/blkdev.h> |
| 14 | #include <linux/backing-dev.h> |
Andrew Morton | 8bde37f | 2006-12-10 02:19:40 -0800 | [diff] [blame] | 15 | #include <linux/task_io_accounting_ops.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 16 | #include <linux/pagevec.h> |
Jens Axboe | f5ff842 | 2007-09-21 09:19:54 +0200 | [diff] [blame] | 17 | #include <linux/pagemap.h> |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 18 | #include <linux/syscalls.h> |
| 19 | #include <linux/file.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 20 | |
Fabian Frederick | 29f175d | 2014-04-07 15:37:55 -0700 | [diff] [blame] | 21 | #include "internal.h" |
| 22 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 23 | /* |
| 24 | * Initialise a struct file's readahead state. Assumes that the caller has |
| 25 | * memset *ra to zero. |
| 26 | */ |
| 27 | void |
| 28 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) |
| 29 | { |
| 30 | ra->ra_pages = mapping->backing_dev_info->ra_pages; |
Fengguang Wu | f4e6b49 | 2007-10-16 01:24:33 -0700 | [diff] [blame] | 31 | ra->prev_pos = -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 32 | } |
Steven Whitehouse | d41cc70 | 2006-01-30 08:53:33 +0000 | [diff] [blame] | 33 | EXPORT_SYMBOL_GPL(file_ra_state_init); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 34 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 35 | #define list_to_page(head) (list_entry((head)->prev, struct page, lru)) |
| 36 | |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 37 | /* |
| 38 | * see if a page needs releasing upon read_cache_pages() failure |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 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 |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 43 | */ |
| 44 | static void read_cache_pages_invalidate_page(struct address_space *mapping, |
| 45 | struct page *page) |
| 46 | { |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 47 | if (page_has_private(page)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 48 | if (!trylock_page(page)) |
| 49 | BUG(); |
| 50 | page->mapping = mapping; |
Lukas Czerner | d47992f | 2013-05-21 23:17:23 -0400 | [diff] [blame] | 51 | do_invalidatepage(page, 0, PAGE_CACHE_SIZE); |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 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 | */ |
| 61 | static 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 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 73 | /** |
Randy Dunlap | bd40cdd | 2006-06-25 05:48:08 -0700 | [diff] [blame] | 74 | * read_cache_pages - populate an address space with some pages & start reads against them |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | */ |
| 83 | int read_cache_pages(struct address_space *mapping, struct list_head *pages, |
| 84 | int (*filler)(void *, struct page *), void *data) |
| 85 | { |
| 86 | struct page *page; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 87 | int ret = 0; |
| 88 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | while (!list_empty(pages)) { |
| 90 | page = list_to_page(pages); |
| 91 | list_del(&page->lru); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 92 | if (add_to_page_cache_lru(page, mapping, |
| 93 | page->index, GFP_KERNEL)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 94 | read_cache_pages_invalidate_page(mapping, page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 95 | continue; |
| 96 | } |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 97 | page_cache_release(page); |
| 98 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 99 | ret = filler(data, page); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 100 | if (unlikely(ret)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 101 | read_cache_pages_invalidate_pages(mapping, pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 102 | break; |
| 103 | } |
Andrew Morton | 8bde37f | 2006-12-10 02:19:40 -0800 | [diff] [blame] | 104 | task_io_account_read(PAGE_CACHE_SIZE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 105 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 106 | return ret; |
| 107 | } |
| 108 | |
| 109 | EXPORT_SYMBOL(read_cache_pages); |
| 110 | |
| 111 | static int read_pages(struct address_space *mapping, struct file *filp, |
| 112 | struct list_head *pages, unsigned nr_pages) |
| 113 | { |
Jens Axboe | 5b417b1 | 2010-04-19 10:04:38 +0200 | [diff] [blame] | 114 | struct blk_plug plug; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 115 | unsigned page_idx; |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 116 | int ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 117 | |
Jens Axboe | 5b417b1 | 2010-04-19 10:04:38 +0200 | [diff] [blame] | 118 | blk_start_plug(&plug); |
| 119 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 120 | if (mapping->a_ops->readpages) { |
| 121 | ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages); |
OGAWA Hirofumi | 029e332 | 2006-11-02 22:07:06 -0800 | [diff] [blame] | 122 | /* Clean up the remaining pages */ |
| 123 | put_pages_list(pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 124 | goto out; |
| 125 | } |
| 126 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 127 | for (page_idx = 0; page_idx < nr_pages; page_idx++) { |
| 128 | struct page *page = list_to_page(pages); |
| 129 | list_del(&page->lru); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 130 | if (!add_to_page_cache_lru(page, mapping, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | page->index, GFP_KERNEL)) { |
Zach Brown | 9f1a3cf | 2006-06-25 05:46:46 -0700 | [diff] [blame] | 132 | mapping->a_ops->readpage(filp, page); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 133 | } |
| 134 | page_cache_release(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 135 | } |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 136 | ret = 0; |
Jens Axboe | 5b417b1 | 2010-04-19 10:04:38 +0200 | [diff] [blame] | 137 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 138 | out: |
Jens Axboe | 5b417b1 | 2010-04-19 10:04:38 +0200 | [diff] [blame] | 139 | blk_finish_plug(&plug); |
| 140 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 141 | return ret; |
| 142 | } |
| 143 | |
| 144 | /* |
Wu Fengguang | d30a110 | 2009-06-16 15:31:30 -0700 | [diff] [blame] | 145 | * __do_page_cache_readahead() actually reads a chunk of disk. It allocates all |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | */ |
Fabian Frederick | 29f175d | 2014-04-07 15:37:55 -0700 | [diff] [blame] | 152 | int __do_page_cache_readahead(struct address_space *mapping, struct file *filp, |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 153 | pgoff_t offset, unsigned long nr_to_read, |
| 154 | unsigned long lookahead_size) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 167 | end_index = ((isize - 1) >> PAGE_CACHE_SHIFT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 168 | |
| 169 | /* |
| 170 | * Preallocate as many pages as we will need. |
| 171 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 172 | for (page_idx = 0; page_idx < nr_to_read; page_idx++) { |
Andrew Morton | 7361f4d | 2005-11-07 00:59:28 -0800 | [diff] [blame] | 173 | pgoff_t page_offset = offset + page_idx; |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 174 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 175 | if (page_offset > end_index) |
| 176 | break; |
| 177 | |
Nick Piggin | 0012818 | 2007-10-16 01:24:40 -0700 | [diff] [blame] | 178 | rcu_read_lock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 179 | page = radix_tree_lookup(&mapping->page_tree, page_offset); |
Nick Piggin | 0012818 | 2007-10-16 01:24:40 -0700 | [diff] [blame] | 180 | rcu_read_unlock(); |
Johannes Weiner | 0cd6144 | 2014-04-03 14:47:46 -0700 | [diff] [blame] | 181 | if (page && !radix_tree_exceptional_entry(page)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 182 | continue; |
| 183 | |
Wu Fengguang | 7b1de58 | 2011-05-24 17:12:25 -0700 | [diff] [blame] | 184 | page = page_cache_alloc_readahead(mapping); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 185 | if (!page) |
| 186 | break; |
| 187 | page->index = page_offset; |
| 188 | list_add(&page->lru, &page_pool); |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 189 | if (page_idx == nr_to_read - lookahead_size) |
| 190 | SetPageReadahead(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 191 | ret++; |
| 192 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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)); |
| 202 | out: |
| 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 | */ |
| 210 | int force_page_cache_readahead(struct address_space *mapping, struct file *filp, |
Andrew Morton | 7361f4d | 2005-11-07 00:59:28 -0800 | [diff] [blame] | 211 | pgoff_t offset, unsigned long nr_to_read) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 212 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 213 | if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages)) |
| 214 | return -EINVAL; |
| 215 | |
Wu Fengguang | f7e839d | 2009-06-16 15:31:20 -0700 | [diff] [blame] | 216 | nr_to_read = max_sane_readahead(nr_to_read); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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, |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 225 | offset, this_chunk, 0); |
Mark Rutland | 58d5640 | 2014-01-29 14:05:51 -0800 | [diff] [blame] | 226 | if (err < 0) |
| 227 | return err; |
| 228 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 229 | offset += this_chunk; |
| 230 | nr_to_read -= this_chunk; |
| 231 | } |
Mark Rutland | 58d5640 | 2014-01-29 14:05:51 -0800 | [diff] [blame] | 232 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 233 | } |
| 234 | |
Raghavendra K T | 6d2be91 | 2014-04-03 14:48:23 -0700 | [diff] [blame] | 235 | #define MAX_READAHEAD ((512*4096)/PAGE_CACHE_SIZE) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 236 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 237 | * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a |
| 238 | * sensible upper limit. |
| 239 | */ |
| 240 | unsigned long max_sane_readahead(unsigned long nr) |
| 241 | { |
Raghavendra K T | 6d2be91 | 2014-04-03 14:48:23 -0700 | [diff] [blame] | 242 | return min(nr, MAX_READAHEAD); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 243 | } |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 244 | |
| 245 | /* |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 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 | */ |
| 251 | static 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 | /* |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 266 | * Get the previous window size, ramp it up, and |
| 267 | * return it as the new window size. |
| 268 | */ |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 269 | static unsigned long get_next_ra_size(struct file_ra_state *ra, |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 270 | unsigned long max) |
| 271 | { |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 272 | unsigned long cur = ra->size; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 273 | unsigned long newsize; |
| 274 | |
| 275 | if (cur < max / 16) |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 276 | newsize = 4 * cur; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 277 | else |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 278 | newsize = 2 * cur; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 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 | * |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 289 | * |<----- async_size ---------| |
| 290 | * |------------------- size -------------------->| |
| 291 | * |==================#===========================| |
| 292 | * ^start ^page marked with PG_readahead |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 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; |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 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. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 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 |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 303 | * page at (start+size-async_size) with PG_readahead, and use it as readahead |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 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 | * |
Fengguang Wu | f4e6b49 | 2007-10-16 01:24:33 -0700 | [diff] [blame] | 307 | * prev_pos tracks the last visited byte in the _previous_ read request. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 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 | /* |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 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 | */ |
| 328 | static pgoff_t count_history_pages(struct address_space *mapping, |
| 329 | struct file_ra_state *ra, |
| 330 | pgoff_t offset, unsigned long max) |
| 331 | { |
| 332 | pgoff_t head; |
| 333 | |
| 334 | rcu_read_lock(); |
Johannes Weiner | e7b563b | 2014-04-03 14:47:44 -0700 | [diff] [blame] | 335 | head = page_cache_prev_hole(mapping, offset - 1, max); |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 336 | rcu_read_unlock(); |
| 337 | |
| 338 | return offset - 1 - head; |
| 339 | } |
| 340 | |
| 341 | /* |
| 342 | * page cache context based read-ahead |
| 343 | */ |
| 344 | static int try_context_readahead(struct address_space *mapping, |
| 345 | struct file_ra_state *ra, |
| 346 | pgoff_t offset, |
| 347 | unsigned long req_size, |
| 348 | unsigned long max) |
| 349 | { |
| 350 | pgoff_t size; |
| 351 | |
| 352 | size = count_history_pages(mapping, ra, offset, max); |
| 353 | |
| 354 | /* |
Fengguang Wu | 2cad401 | 2013-09-11 14:21:47 -0700 | [diff] [blame] | 355 | * not enough history pages: |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 356 | * it could be a random read |
| 357 | */ |
Fengguang Wu | 2cad401 | 2013-09-11 14:21:47 -0700 | [diff] [blame] | 358 | if (size <= req_size) |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 359 | return 0; |
| 360 | |
| 361 | /* |
| 362 | * starts from beginning of file: |
| 363 | * it is a strong indication of long-run stream (or whole-file-read) |
| 364 | */ |
| 365 | if (size >= offset) |
| 366 | size *= 2; |
| 367 | |
| 368 | ra->start = offset; |
Fengguang Wu | 2cad401 | 2013-09-11 14:21:47 -0700 | [diff] [blame] | 369 | ra->size = min(size + req_size, max); |
| 370 | ra->async_size = 1; |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 371 | |
| 372 | return 1; |
| 373 | } |
| 374 | |
| 375 | /* |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 376 | * A minimal readahead algorithm for trivial sequential/random reads. |
| 377 | */ |
| 378 | static unsigned long |
| 379 | ondemand_readahead(struct address_space *mapping, |
| 380 | struct file_ra_state *ra, struct file *filp, |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 381 | bool hit_readahead_marker, pgoff_t offset, |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 382 | unsigned long req_size) |
| 383 | { |
Wu Fengguang | fc31d16 | 2009-06-16 15:31:21 -0700 | [diff] [blame] | 384 | unsigned long max = max_sane_readahead(ra->ra_pages); |
Damien Ramonda | af248a0 | 2013-11-12 15:08:16 -0800 | [diff] [blame] | 385 | pgoff_t prev_offset; |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 386 | |
| 387 | /* |
| 388 | * start of file |
| 389 | */ |
| 390 | if (!offset) |
| 391 | goto initial_readahead; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 392 | |
| 393 | /* |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 394 | * It's the expected callback offset, assume sequential access. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 395 | * Ramp up sizes, and push forward the readahead window. |
| 396 | */ |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 397 | if ((offset == (ra->start + ra->size - ra->async_size) || |
| 398 | offset == (ra->start + ra->size))) { |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 399 | ra->start += ra->size; |
| 400 | ra->size = get_next_ra_size(ra, max); |
| 401 | ra->async_size = ra->size; |
| 402 | goto readit; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 403 | } |
| 404 | |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 405 | /* |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 406 | * Hit a marked page without valid readahead state. |
| 407 | * E.g. interleaved reads. |
| 408 | * Query the pagecache for async_size, which normally equals to |
| 409 | * readahead size. Ramp it up and use it as the new readahead size. |
| 410 | */ |
| 411 | if (hit_readahead_marker) { |
| 412 | pgoff_t start; |
| 413 | |
Nick Piggin | 30002ed | 2008-07-25 19:45:28 -0700 | [diff] [blame] | 414 | rcu_read_lock(); |
Johannes Weiner | e7b563b | 2014-04-03 14:47:44 -0700 | [diff] [blame] | 415 | start = page_cache_next_hole(mapping, offset + 1, max); |
Nick Piggin | 30002ed | 2008-07-25 19:45:28 -0700 | [diff] [blame] | 416 | rcu_read_unlock(); |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 417 | |
| 418 | if (!start || start - offset > max) |
| 419 | return 0; |
| 420 | |
| 421 | ra->start = start; |
| 422 | ra->size = start - offset; /* old async_size */ |
Wu Fengguang | 160334a | 2009-06-16 15:31:23 -0700 | [diff] [blame] | 423 | ra->size += req_size; |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 424 | ra->size = get_next_ra_size(ra, max); |
| 425 | ra->async_size = ra->size; |
| 426 | goto readit; |
| 427 | } |
| 428 | |
| 429 | /* |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 430 | * oversize read |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 431 | */ |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 432 | if (req_size > max) |
| 433 | goto initial_readahead; |
| 434 | |
| 435 | /* |
| 436 | * sequential cache miss |
Damien Ramonda | af248a0 | 2013-11-12 15:08:16 -0800 | [diff] [blame] | 437 | * trivial case: (offset - prev_offset) == 1 |
| 438 | * unaligned reads: (offset - prev_offset) == 0 |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 439 | */ |
Damien Ramonda | af248a0 | 2013-11-12 15:08:16 -0800 | [diff] [blame] | 440 | prev_offset = (unsigned long long)ra->prev_pos >> PAGE_CACHE_SHIFT; |
| 441 | if (offset - prev_offset <= 1UL) |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 442 | goto initial_readahead; |
| 443 | |
| 444 | /* |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 445 | * Query the page cache and look for the traces(cached history pages) |
| 446 | * that a sequential stream would leave behind. |
| 447 | */ |
| 448 | if (try_context_readahead(mapping, ra, offset, req_size, max)) |
| 449 | goto readit; |
| 450 | |
| 451 | /* |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 452 | * standalone, small random read |
| 453 | * Read as is, and do not pollute the readahead state. |
| 454 | */ |
| 455 | return __do_page_cache_readahead(mapping, filp, offset, req_size, 0); |
| 456 | |
| 457 | initial_readahead: |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 458 | ra->start = offset; |
| 459 | ra->size = get_init_ra_size(req_size, max); |
| 460 | ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 461 | |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 462 | readit: |
Wu Fengguang | 51daa88 | 2009-06-16 15:31:24 -0700 | [diff] [blame] | 463 | /* |
| 464 | * Will this read hit the readahead marker made by itself? |
| 465 | * If so, trigger the readahead marker hit now, and merge |
| 466 | * the resulted next readahead window into the current one. |
| 467 | */ |
| 468 | if (offset == ra->start && ra->size == ra->async_size) { |
| 469 | ra->async_size = get_next_ra_size(ra, max); |
| 470 | ra->size += ra->async_size; |
| 471 | } |
| 472 | |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 473 | return ra_submit(ra, mapping, filp); |
| 474 | } |
| 475 | |
| 476 | /** |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 477 | * page_cache_sync_readahead - generic file readahead |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 478 | * @mapping: address_space which holds the pagecache and I/O vectors |
| 479 | * @ra: file_ra_state which holds the readahead state |
| 480 | * @filp: passed on to ->readpage() and ->readpages() |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 481 | * @offset: start offset into @mapping, in pagecache page-sized units |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 482 | * @req_size: hint: total size of the read which the caller is performing in |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 483 | * pagecache pages |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 484 | * |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 485 | * page_cache_sync_readahead() should be called when a cache miss happened: |
| 486 | * it will submit the read. The readahead logic may decide to piggyback more |
| 487 | * pages onto the read request if access patterns suggest it will improve |
| 488 | * performance. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 489 | */ |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 490 | void page_cache_sync_readahead(struct address_space *mapping, |
| 491 | struct file_ra_state *ra, struct file *filp, |
| 492 | pgoff_t offset, unsigned long req_size) |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 493 | { |
| 494 | /* no read-ahead */ |
| 495 | if (!ra->ra_pages) |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 496 | return; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 497 | |
Wu Fengguang | 0141450 | 2010-03-05 13:42:03 -0800 | [diff] [blame] | 498 | /* be dumb */ |
Wu Fengguang | 70655c0 | 2010-04-06 14:34:53 -0700 | [diff] [blame] | 499 | if (filp && (filp->f_mode & FMODE_RANDOM)) { |
Wu Fengguang | 0141450 | 2010-03-05 13:42:03 -0800 | [diff] [blame] | 500 | force_page_cache_readahead(mapping, filp, offset, req_size); |
| 501 | return; |
| 502 | } |
| 503 | |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 504 | /* do read-ahead */ |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 505 | ondemand_readahead(mapping, ra, filp, false, offset, req_size); |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 506 | } |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 507 | EXPORT_SYMBOL_GPL(page_cache_sync_readahead); |
| 508 | |
| 509 | /** |
| 510 | * page_cache_async_readahead - file readahead for marked pages |
| 511 | * @mapping: address_space which holds the pagecache and I/O vectors |
| 512 | * @ra: file_ra_state which holds the readahead state |
| 513 | * @filp: passed on to ->readpage() and ->readpages() |
| 514 | * @page: the page at @offset which has the PG_readahead flag set |
| 515 | * @offset: start offset into @mapping, in pagecache page-sized units |
| 516 | * @req_size: hint: total size of the read which the caller is performing in |
| 517 | * pagecache pages |
| 518 | * |
Huang Shijie | bf8abe8 | 2010-05-24 14:32:36 -0700 | [diff] [blame] | 519 | * page_cache_async_readahead() should be called when a page is used which |
Randy Dunlap | f7850d9 | 2008-03-19 17:01:02 -0700 | [diff] [blame] | 520 | * has the PG_readahead flag; this is a marker to suggest that the application |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 521 | * has used up enough of the readahead window that we should start pulling in |
Randy Dunlap | f7850d9 | 2008-03-19 17:01:02 -0700 | [diff] [blame] | 522 | * more pages. |
| 523 | */ |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 524 | void |
| 525 | page_cache_async_readahead(struct address_space *mapping, |
| 526 | struct file_ra_state *ra, struct file *filp, |
| 527 | struct page *page, pgoff_t offset, |
| 528 | unsigned long req_size) |
| 529 | { |
| 530 | /* no read-ahead */ |
| 531 | if (!ra->ra_pages) |
| 532 | return; |
| 533 | |
| 534 | /* |
| 535 | * Same bit is used for PG_readahead and PG_reclaim. |
| 536 | */ |
| 537 | if (PageWriteback(page)) |
| 538 | return; |
| 539 | |
| 540 | ClearPageReadahead(page); |
| 541 | |
| 542 | /* |
| 543 | * Defer asynchronous read-ahead on IO congestion. |
| 544 | */ |
| 545 | if (bdi_read_congested(mapping->backing_dev_info)) |
| 546 | return; |
| 547 | |
| 548 | /* do read-ahead */ |
| 549 | ondemand_readahead(mapping, ra, filp, true, offset, req_size); |
| 550 | } |
| 551 | EXPORT_SYMBOL_GPL(page_cache_async_readahead); |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 552 | |
| 553 | static ssize_t |
| 554 | do_readahead(struct address_space *mapping, struct file *filp, |
| 555 | pgoff_t index, unsigned long nr) |
| 556 | { |
Andrew Morton | 63d0f0a | 2013-11-12 15:07:09 -0800 | [diff] [blame] | 557 | if (!mapping || !mapping->a_ops) |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 558 | return -EINVAL; |
| 559 | |
Mark Rutland | 58d5640 | 2014-01-29 14:05:51 -0800 | [diff] [blame] | 560 | return force_page_cache_readahead(mapping, filp, index, nr); |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 561 | } |
| 562 | |
Al Viro | 4a0fd5b | 2013-01-21 15:16:58 -0500 | [diff] [blame] | 563 | SYSCALL_DEFINE3(readahead, int, fd, loff_t, offset, size_t, count) |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 564 | { |
| 565 | ssize_t ret; |
Al Viro | 2903ff0 | 2012-08-28 12:52:22 -0400 | [diff] [blame] | 566 | struct fd f; |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 567 | |
| 568 | ret = -EBADF; |
Al Viro | 2903ff0 | 2012-08-28 12:52:22 -0400 | [diff] [blame] | 569 | f = fdget(fd); |
| 570 | if (f.file) { |
| 571 | if (f.file->f_mode & FMODE_READ) { |
| 572 | struct address_space *mapping = f.file->f_mapping; |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 573 | pgoff_t start = offset >> PAGE_CACHE_SHIFT; |
| 574 | pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT; |
| 575 | unsigned long len = end - start + 1; |
Al Viro | 2903ff0 | 2012-08-28 12:52:22 -0400 | [diff] [blame] | 576 | ret = do_readahead(mapping, f.file, start, len); |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 577 | } |
Al Viro | 2903ff0 | 2012-08-28 12:52:22 -0400 | [diff] [blame] | 578 | fdput(f); |
Cong Wang | 782182e | 2012-05-29 15:06:43 -0700 | [diff] [blame] | 579 | } |
| 580 | return ret; |
| 581 | } |