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> |
| 11 | #include <linux/fs.h> |
| 12 | #include <linux/mm.h> |
| 13 | #include <linux/module.h> |
| 14 | #include <linux/blkdev.h> |
| 15 | #include <linux/backing-dev.h> |
Andrew Morton | 8bde37f | 2006-12-10 02:19:40 -0800 | [diff] [blame] | 16 | #include <linux/task_io_accounting_ops.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | #include <linux/pagevec.h> |
Jens Axboe | f5ff842 | 2007-09-21 09:19:54 +0200 | [diff] [blame] | 18 | #include <linux/pagemap.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 19 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 20 | /* |
| 21 | * Initialise a struct file's readahead state. Assumes that the caller has |
| 22 | * memset *ra to zero. |
| 23 | */ |
| 24 | void |
| 25 | file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping) |
| 26 | { |
| 27 | ra->ra_pages = mapping->backing_dev_info->ra_pages; |
Fengguang Wu | f4e6b49 | 2007-10-16 01:24:33 -0700 | [diff] [blame] | 28 | ra->prev_pos = -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 29 | } |
Steven Whitehouse | d41cc70 | 2006-01-30 08:53:33 +0000 | [diff] [blame] | 30 | EXPORT_SYMBOL_GPL(file_ra_state_init); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 31 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 32 | #define list_to_page(head) (list_entry((head)->prev, struct page, lru)) |
| 33 | |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 34 | /* |
| 35 | * see if a page needs releasing upon read_cache_pages() failure |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 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 |
| 39 | * an error |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 40 | */ |
| 41 | static void read_cache_pages_invalidate_page(struct address_space *mapping, |
| 42 | struct page *page) |
| 43 | { |
David Howells | 266cf65 | 2009-04-03 16:42:36 +0100 | [diff] [blame] | 44 | if (page_has_private(page)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 45 | if (!trylock_page(page)) |
| 46 | BUG(); |
| 47 | page->mapping = mapping; |
| 48 | do_invalidatepage(page, 0); |
| 49 | page->mapping = NULL; |
| 50 | unlock_page(page); |
| 51 | } |
| 52 | page_cache_release(page); |
| 53 | } |
| 54 | |
| 55 | /* |
| 56 | * release a list of pages, invalidating them first if need be |
| 57 | */ |
| 58 | static void read_cache_pages_invalidate_pages(struct address_space *mapping, |
| 59 | struct list_head *pages) |
| 60 | { |
| 61 | struct page *victim; |
| 62 | |
| 63 | while (!list_empty(pages)) { |
| 64 | victim = list_to_page(pages); |
| 65 | list_del(&victim->lru); |
| 66 | read_cache_pages_invalidate_page(mapping, victim); |
| 67 | } |
| 68 | } |
| 69 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 70 | /** |
Randy Dunlap | bd40cdd | 2006-06-25 05:48:08 -0700 | [diff] [blame] | 71 | * 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] | 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. |
| 77 | * |
| 78 | * Hides the details of the LRU cache etc from the filesystems. |
| 79 | */ |
| 80 | int read_cache_pages(struct address_space *mapping, struct list_head *pages, |
| 81 | int (*filler)(void *, struct page *), void *data) |
| 82 | { |
| 83 | struct page *page; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 84 | int ret = 0; |
| 85 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 86 | while (!list_empty(pages)) { |
| 87 | page = list_to_page(pages); |
| 88 | list_del(&page->lru); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 89 | if (add_to_page_cache_lru(page, mapping, |
| 90 | page->index, GFP_KERNEL)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 91 | read_cache_pages_invalidate_page(mapping, page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 92 | continue; |
| 93 | } |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 94 | page_cache_release(page); |
| 95 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 96 | ret = filler(data, page); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 97 | if (unlikely(ret)) { |
David Howells | 03fb3d2 | 2009-04-03 16:42:35 +0100 | [diff] [blame] | 98 | read_cache_pages_invalidate_pages(mapping, pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 99 | break; |
| 100 | } |
Andrew Morton | 8bde37f | 2006-12-10 02:19:40 -0800 | [diff] [blame] | 101 | task_io_account_read(PAGE_CACHE_SIZE); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 102 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | return ret; |
| 104 | } |
| 105 | |
| 106 | EXPORT_SYMBOL(read_cache_pages); |
| 107 | |
| 108 | static int read_pages(struct address_space *mapping, struct file *filp, |
| 109 | struct list_head *pages, unsigned nr_pages) |
| 110 | { |
| 111 | unsigned page_idx; |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 112 | int ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 113 | |
| 114 | if (mapping->a_ops->readpages) { |
| 115 | ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages); |
OGAWA Hirofumi | 029e332 | 2006-11-02 22:07:06 -0800 | [diff] [blame] | 116 | /* Clean up the remaining pages */ |
| 117 | put_pages_list(pages); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 118 | goto out; |
| 119 | } |
| 120 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 121 | for (page_idx = 0; page_idx < nr_pages; page_idx++) { |
| 122 | struct page *page = list_to_page(pages); |
| 123 | list_del(&page->lru); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 124 | if (!add_to_page_cache_lru(page, mapping, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 125 | page->index, GFP_KERNEL)) { |
Zach Brown | 9f1a3cf | 2006-06-25 05:46:46 -0700 | [diff] [blame] | 126 | mapping->a_ops->readpage(filp, page); |
Nick Piggin | eb2be18 | 2007-10-16 01:24:57 -0700 | [diff] [blame] | 127 | } |
| 128 | page_cache_release(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | } |
Zach Brown | 994fc28c | 2005-12-15 14:28:17 -0800 | [diff] [blame] | 130 | ret = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 131 | out: |
| 132 | return ret; |
| 133 | } |
| 134 | |
| 135 | /* |
Wu Fengguang | d30a110 | 2009-06-16 15:31:30 -0700 | [diff] [blame] | 136 | * __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] | 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. |
| 140 | * |
| 141 | * 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] | 142 | */ |
| 143 | static int |
| 144 | __do_page_cache_readahead(struct address_space *mapping, struct file *filp, |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 145 | pgoff_t offset, unsigned long nr_to_read, |
| 146 | unsigned long lookahead_size) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 147 | { |
| 148 | struct inode *inode = mapping->host; |
| 149 | struct page *page; |
| 150 | unsigned long end_index; /* The last page we want to read */ |
| 151 | LIST_HEAD(page_pool); |
| 152 | int page_idx; |
| 153 | int ret = 0; |
| 154 | loff_t isize = i_size_read(inode); |
| 155 | |
| 156 | if (isize == 0) |
| 157 | goto out; |
| 158 | |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 159 | end_index = ((isize - 1) >> PAGE_CACHE_SHIFT); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 160 | |
| 161 | /* |
| 162 | * Preallocate as many pages as we will need. |
| 163 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 164 | for (page_idx = 0; page_idx < nr_to_read; page_idx++) { |
Andrew Morton | 7361f4d | 2005-11-07 00:59:28 -0800 | [diff] [blame] | 165 | pgoff_t page_offset = offset + page_idx; |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 166 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | if (page_offset > end_index) |
| 168 | break; |
| 169 | |
Nick Piggin | 0012818 | 2007-10-16 01:24:40 -0700 | [diff] [blame] | 170 | rcu_read_lock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | page = radix_tree_lookup(&mapping->page_tree, page_offset); |
Nick Piggin | 0012818 | 2007-10-16 01:24:40 -0700 | [diff] [blame] | 172 | rcu_read_unlock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 173 | if (page) |
| 174 | continue; |
| 175 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 176 | page = page_cache_alloc_cold(mapping); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 177 | if (!page) |
| 178 | break; |
| 179 | page->index = page_offset; |
| 180 | list_add(&page->lru, &page_pool); |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 181 | if (page_idx == nr_to_read - lookahead_size) |
| 182 | SetPageReadahead(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 183 | ret++; |
| 184 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 185 | |
| 186 | /* |
| 187 | * Now start the IO. We ignore I/O errors - if the page is not |
| 188 | * uptodate then the caller will launch readpage again, and |
| 189 | * will then handle the error. |
| 190 | */ |
| 191 | if (ret) |
| 192 | read_pages(mapping, filp, &page_pool, ret); |
| 193 | BUG_ON(!list_empty(&page_pool)); |
| 194 | out: |
| 195 | return ret; |
| 196 | } |
| 197 | |
| 198 | /* |
| 199 | * Chunk the readahead into 2 megabyte units, so that we don't pin too much |
| 200 | * memory at once. |
| 201 | */ |
| 202 | int force_page_cache_readahead(struct address_space *mapping, struct file *filp, |
Andrew Morton | 7361f4d | 2005-11-07 00:59:28 -0800 | [diff] [blame] | 203 | pgoff_t offset, unsigned long nr_to_read) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 204 | { |
| 205 | int ret = 0; |
| 206 | |
| 207 | if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages)) |
| 208 | return -EINVAL; |
| 209 | |
Wu Fengguang | f7e839d | 2009-06-16 15:31:20 -0700 | [diff] [blame] | 210 | nr_to_read = max_sane_readahead(nr_to_read); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 211 | while (nr_to_read) { |
| 212 | int err; |
| 213 | |
| 214 | unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE; |
| 215 | |
| 216 | if (this_chunk > nr_to_read) |
| 217 | this_chunk = nr_to_read; |
| 218 | err = __do_page_cache_readahead(mapping, filp, |
Fengguang Wu | 46fc3e7 | 2007-07-19 01:47:57 -0700 | [diff] [blame] | 219 | offset, this_chunk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 220 | if (err < 0) { |
| 221 | ret = err; |
| 222 | break; |
| 223 | } |
| 224 | ret += err; |
| 225 | offset += this_chunk; |
| 226 | nr_to_read -= this_chunk; |
| 227 | } |
| 228 | return ret; |
| 229 | } |
| 230 | |
| 231 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 232 | * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a |
| 233 | * sensible upper limit. |
| 234 | */ |
| 235 | unsigned long max_sane_readahead(unsigned long nr) |
| 236 | { |
Rik van Riel | 4f98a2f | 2008-10-18 20:26:32 -0700 | [diff] [blame] | 237 | return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE) |
Christoph Lameter | 05a0416 | 2007-02-10 01:43:05 -0800 | [diff] [blame] | 238 | + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | } |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 240 | |
| 241 | /* |
| 242 | * Submit IO for the read-ahead request in file_ra_state. |
| 243 | */ |
Wu Fengguang | d30a110 | 2009-06-16 15:31:30 -0700 | [diff] [blame] | 244 | unsigned long ra_submit(struct file_ra_state *ra, |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 245 | struct address_space *mapping, struct file *filp) |
| 246 | { |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 247 | int actual; |
| 248 | |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 249 | actual = __do_page_cache_readahead(mapping, filp, |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 250 | ra->start, ra->size, ra->async_size); |
Fengguang Wu | 5ce1110 | 2007-07-19 01:47:59 -0700 | [diff] [blame] | 251 | |
| 252 | return actual; |
| 253 | } |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 254 | |
| 255 | /* |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 256 | * Set the initial window size, round to next power of 2 and square |
| 257 | * for small size, x 4 for medium, and x 2 for large |
| 258 | * for 128k (32 page) max ra |
| 259 | * 1-8 page = 32k initial, > 8 page = 128k initial |
| 260 | */ |
| 261 | static unsigned long get_init_ra_size(unsigned long size, unsigned long max) |
| 262 | { |
| 263 | unsigned long newsize = roundup_pow_of_two(size); |
| 264 | |
| 265 | if (newsize <= max / 32) |
| 266 | newsize = newsize * 4; |
| 267 | else if (newsize <= max / 4) |
| 268 | newsize = newsize * 2; |
| 269 | else |
| 270 | newsize = max; |
| 271 | |
| 272 | return newsize; |
| 273 | } |
| 274 | |
| 275 | /* |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 276 | * Get the previous window size, ramp it up, and |
| 277 | * return it as the new window size. |
| 278 | */ |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 279 | static unsigned long get_next_ra_size(struct file_ra_state *ra, |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 280 | unsigned long max) |
| 281 | { |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 282 | unsigned long cur = ra->size; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 283 | unsigned long newsize; |
| 284 | |
| 285 | if (cur < max / 16) |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 286 | newsize = 4 * cur; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 287 | else |
Fengguang Wu | c743d96 | 2007-07-19 01:48:04 -0700 | [diff] [blame] | 288 | newsize = 2 * cur; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 289 | |
| 290 | return min(newsize, max); |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * On-demand readahead design. |
| 295 | * |
| 296 | * The fields in struct file_ra_state represent the most-recently-executed |
| 297 | * readahead attempt: |
| 298 | * |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 299 | * |<----- async_size ---------| |
| 300 | * |------------------- size -------------------->| |
| 301 | * |==================#===========================| |
| 302 | * ^start ^page marked with PG_readahead |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 303 | * |
| 304 | * To overlap application thinking time and disk I/O time, we do |
| 305 | * `readahead pipelining': Do not wait until the application consumed all |
| 306 | * readahead pages and stalled on the missing page at readahead_index; |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 307 | * Instead, submit an asynchronous readahead I/O as soon as there are |
| 308 | * only async_size pages left in the readahead window. Normally async_size |
| 309 | * will be equal to size, for maximum pipelining. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 310 | * |
| 311 | * In interleaved sequential reads, concurrent streams on the same fd can |
| 312 | * 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] | 313 | * 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] | 314 | * indicator. The flag won't be set on already cached pages, to avoid the |
| 315 | * readahead-for-nothing fuss, saving pointless page cache lookups. |
| 316 | * |
Fengguang Wu | f4e6b49 | 2007-10-16 01:24:33 -0700 | [diff] [blame] | 317 | * prev_pos tracks the last visited byte in the _previous_ read request. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 318 | * It should be maintained by the caller, and will be used for detecting |
| 319 | * small random reads. Note that the readahead algorithm checks loosely |
| 320 | * for sequential patterns. Hence interleaved reads might be served as |
| 321 | * sequential ones. |
| 322 | * |
| 323 | * There is a special-case: if the first page which the application tries to |
| 324 | * read happens to be the first page of the file, it is assumed that a linear |
| 325 | * read is about to happen and the window is immediately set to the initial size |
| 326 | * based on I/O request size and the max_readahead. |
| 327 | * |
| 328 | * The code ramps up the readahead size aggressively at first, but slow down as |
| 329 | * it approaches max_readhead. |
| 330 | */ |
| 331 | |
| 332 | /* |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 333 | * Count contiguously cached pages from @offset-1 to @offset-@max, |
| 334 | * this count is a conservative estimation of |
| 335 | * - length of the sequential read sequence, or |
| 336 | * - thrashing threshold in memory tight systems |
| 337 | */ |
| 338 | static pgoff_t count_history_pages(struct address_space *mapping, |
| 339 | struct file_ra_state *ra, |
| 340 | pgoff_t offset, unsigned long max) |
| 341 | { |
| 342 | pgoff_t head; |
| 343 | |
| 344 | rcu_read_lock(); |
| 345 | head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max); |
| 346 | rcu_read_unlock(); |
| 347 | |
| 348 | return offset - 1 - head; |
| 349 | } |
| 350 | |
| 351 | /* |
| 352 | * page cache context based read-ahead |
| 353 | */ |
| 354 | static int try_context_readahead(struct address_space *mapping, |
| 355 | struct file_ra_state *ra, |
| 356 | pgoff_t offset, |
| 357 | unsigned long req_size, |
| 358 | unsigned long max) |
| 359 | { |
| 360 | pgoff_t size; |
| 361 | |
| 362 | size = count_history_pages(mapping, ra, offset, max); |
| 363 | |
| 364 | /* |
| 365 | * no history pages: |
| 366 | * it could be a random read |
| 367 | */ |
| 368 | if (!size) |
| 369 | return 0; |
| 370 | |
| 371 | /* |
| 372 | * starts from beginning of file: |
| 373 | * it is a strong indication of long-run stream (or whole-file-read) |
| 374 | */ |
| 375 | if (size >= offset) |
| 376 | size *= 2; |
| 377 | |
| 378 | ra->start = offset; |
| 379 | ra->size = get_init_ra_size(size + req_size, max); |
| 380 | ra->async_size = ra->size; |
| 381 | |
| 382 | return 1; |
| 383 | } |
| 384 | |
| 385 | /* |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 386 | * A minimal readahead algorithm for trivial sequential/random reads. |
| 387 | */ |
| 388 | static unsigned long |
| 389 | ondemand_readahead(struct address_space *mapping, |
| 390 | struct file_ra_state *ra, struct file *filp, |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 391 | bool hit_readahead_marker, pgoff_t offset, |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 392 | unsigned long req_size) |
| 393 | { |
Wu Fengguang | fc31d16 | 2009-06-16 15:31:21 -0700 | [diff] [blame] | 394 | unsigned long max = max_sane_readahead(ra->ra_pages); |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 395 | |
| 396 | /* |
| 397 | * start of file |
| 398 | */ |
| 399 | if (!offset) |
| 400 | goto initial_readahead; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 401 | |
| 402 | /* |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 403 | * It's the expected callback offset, assume sequential access. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 404 | * Ramp up sizes, and push forward the readahead window. |
| 405 | */ |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 406 | if ((offset == (ra->start + ra->size - ra->async_size) || |
| 407 | offset == (ra->start + ra->size))) { |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 408 | ra->start += ra->size; |
| 409 | ra->size = get_next_ra_size(ra, max); |
| 410 | ra->async_size = ra->size; |
| 411 | goto readit; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 412 | } |
| 413 | |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 414 | /* |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 415 | * Hit a marked page without valid readahead state. |
| 416 | * E.g. interleaved reads. |
| 417 | * Query the pagecache for async_size, which normally equals to |
| 418 | * readahead size. Ramp it up and use it as the new readahead size. |
| 419 | */ |
| 420 | if (hit_readahead_marker) { |
| 421 | pgoff_t start; |
| 422 | |
Nick Piggin | 30002ed | 2008-07-25 19:45:28 -0700 | [diff] [blame] | 423 | rcu_read_lock(); |
Wu Fengguang | caca7cb | 2009-06-16 15:31:21 -0700 | [diff] [blame] | 424 | start = radix_tree_next_hole(&mapping->page_tree, offset+1,max); |
Nick Piggin | 30002ed | 2008-07-25 19:45:28 -0700 | [diff] [blame] | 425 | rcu_read_unlock(); |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 426 | |
| 427 | if (!start || start - offset > max) |
| 428 | return 0; |
| 429 | |
| 430 | ra->start = start; |
| 431 | ra->size = start - offset; /* old async_size */ |
Wu Fengguang | 160334a | 2009-06-16 15:31:23 -0700 | [diff] [blame] | 432 | ra->size += req_size; |
Fengguang Wu | 6b10c6c | 2007-10-16 01:24:34 -0700 | [diff] [blame] | 433 | ra->size = get_next_ra_size(ra, max); |
| 434 | ra->async_size = ra->size; |
| 435 | goto readit; |
| 436 | } |
| 437 | |
| 438 | /* |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 439 | * oversize read |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 440 | */ |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 441 | if (req_size > max) |
| 442 | goto initial_readahead; |
| 443 | |
| 444 | /* |
| 445 | * sequential cache miss |
| 446 | */ |
| 447 | if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL) |
| 448 | goto initial_readahead; |
| 449 | |
| 450 | /* |
Wu Fengguang | 10be0b3 | 2009-06-16 15:31:36 -0700 | [diff] [blame] | 451 | * Query the page cache and look for the traces(cached history pages) |
| 452 | * that a sequential stream would leave behind. |
| 453 | */ |
| 454 | if (try_context_readahead(mapping, ra, offset, req_size, max)) |
| 455 | goto readit; |
| 456 | |
| 457 | /* |
Wu Fengguang | 045a252 | 2009-06-16 15:31:33 -0700 | [diff] [blame] | 458 | * standalone, small random read |
| 459 | * Read as is, and do not pollute the readahead state. |
| 460 | */ |
| 461 | return __do_page_cache_readahead(mapping, filp, offset, req_size, 0); |
| 462 | |
| 463 | initial_readahead: |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 464 | ra->start = offset; |
| 465 | ra->size = get_init_ra_size(req_size, max); |
| 466 | 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] | 467 | |
Fengguang Wu | f9acc8c | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 468 | readit: |
Wu Fengguang | 51daa88 | 2009-06-16 15:31:24 -0700 | [diff] [blame] | 469 | /* |
| 470 | * Will this read hit the readahead marker made by itself? |
| 471 | * If so, trigger the readahead marker hit now, and merge |
| 472 | * the resulted next readahead window into the current one. |
| 473 | */ |
| 474 | if (offset == ra->start && ra->size == ra->async_size) { |
| 475 | ra->async_size = get_next_ra_size(ra, max); |
| 476 | ra->size += ra->async_size; |
| 477 | } |
| 478 | |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 479 | return ra_submit(ra, mapping, filp); |
| 480 | } |
| 481 | |
| 482 | /** |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 483 | * page_cache_sync_readahead - generic file readahead |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 484 | * @mapping: address_space which holds the pagecache and I/O vectors |
| 485 | * @ra: file_ra_state which holds the readahead state |
| 486 | * @filp: passed on to ->readpage() and ->readpages() |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 487 | * @offset: start offset into @mapping, in pagecache page-sized units |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 488 | * @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] | 489 | * pagecache pages |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 490 | * |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 491 | * page_cache_sync_readahead() should be called when a cache miss happened: |
| 492 | * it will submit the read. The readahead logic may decide to piggyback more |
| 493 | * pages onto the read request if access patterns suggest it will improve |
| 494 | * performance. |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 495 | */ |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 496 | void page_cache_sync_readahead(struct address_space *mapping, |
| 497 | struct file_ra_state *ra, struct file *filp, |
| 498 | pgoff_t offset, unsigned long req_size) |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 499 | { |
| 500 | /* no read-ahead */ |
| 501 | if (!ra->ra_pages) |
Rusty Russell | cf914a7 | 2007-07-19 01:48:08 -0700 | [diff] [blame] | 502 | return; |
Fengguang Wu | 122a21d | 2007-07-19 01:48:01 -0700 | [diff] [blame] | 503 | |
| 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 | * |
| 519 | * page_cache_async_ondemand() 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); |