Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | #ifndef _LINUX_MMZONE_H |
| 2 | #define _LINUX_MMZONE_H |
| 3 | |
| 4 | #ifdef __KERNEL__ |
| 5 | #ifndef __ASSEMBLY__ |
| 6 | |
| 7 | #include <linux/config.h> |
| 8 | #include <linux/spinlock.h> |
| 9 | #include <linux/list.h> |
| 10 | #include <linux/wait.h> |
| 11 | #include <linux/cache.h> |
| 12 | #include <linux/threads.h> |
| 13 | #include <linux/numa.h> |
| 14 | #include <linux/init.h> |
| 15 | #include <asm/atomic.h> |
| 16 | |
| 17 | /* Free memory management - zoned buddy allocator. */ |
| 18 | #ifndef CONFIG_FORCE_MAX_ZONEORDER |
| 19 | #define MAX_ORDER 11 |
| 20 | #else |
| 21 | #define MAX_ORDER CONFIG_FORCE_MAX_ZONEORDER |
| 22 | #endif |
| 23 | |
| 24 | struct free_area { |
| 25 | struct list_head free_list; |
| 26 | unsigned long nr_free; |
| 27 | }; |
| 28 | |
| 29 | struct pglist_data; |
| 30 | |
| 31 | /* |
| 32 | * zone->lock and zone->lru_lock are two of the hottest locks in the kernel. |
| 33 | * So add a wild amount of padding here to ensure that they fall into separate |
| 34 | * cachelines. There are very few zone structures in the machine, so space |
| 35 | * consumption is not a concern here. |
| 36 | */ |
| 37 | #if defined(CONFIG_SMP) |
| 38 | struct zone_padding { |
| 39 | char x[0]; |
| 40 | } ____cacheline_maxaligned_in_smp; |
| 41 | #define ZONE_PADDING(name) struct zone_padding name; |
| 42 | #else |
| 43 | #define ZONE_PADDING(name) |
| 44 | #endif |
| 45 | |
| 46 | struct per_cpu_pages { |
| 47 | int count; /* number of pages in the list */ |
| 48 | int low; /* low watermark, refill needed */ |
| 49 | int high; /* high watermark, emptying needed */ |
| 50 | int batch; /* chunk size for buddy add/remove */ |
| 51 | struct list_head list; /* the list of pages */ |
| 52 | }; |
| 53 | |
| 54 | struct per_cpu_pageset { |
| 55 | struct per_cpu_pages pcp[2]; /* 0: hot. 1: cold */ |
| 56 | #ifdef CONFIG_NUMA |
| 57 | unsigned long numa_hit; /* allocated in intended node */ |
| 58 | unsigned long numa_miss; /* allocated in non intended node */ |
| 59 | unsigned long numa_foreign; /* was intended here, hit elsewhere */ |
| 60 | unsigned long interleave_hit; /* interleaver prefered this zone */ |
| 61 | unsigned long local_node; /* allocation from local node */ |
| 62 | unsigned long other_node; /* allocation from other node */ |
| 63 | #endif |
| 64 | } ____cacheline_aligned_in_smp; |
| 65 | |
| 66 | #define ZONE_DMA 0 |
| 67 | #define ZONE_NORMAL 1 |
| 68 | #define ZONE_HIGHMEM 2 |
| 69 | |
| 70 | #define MAX_NR_ZONES 3 /* Sync this with ZONES_SHIFT */ |
| 71 | #define ZONES_SHIFT 2 /* ceil(log2(MAX_NR_ZONES)) */ |
| 72 | |
| 73 | |
| 74 | /* |
| 75 | * When a memory allocation must conform to specific limitations (such |
| 76 | * as being suitable for DMA) the caller will pass in hints to the |
| 77 | * allocator in the gfp_mask, in the zone modifier bits. These bits |
| 78 | * are used to select a priority ordered list of memory zones which |
| 79 | * match the requested limits. GFP_ZONEMASK defines which bits within |
| 80 | * the gfp_mask should be considered as zone modifiers. Each valid |
| 81 | * combination of the zone modifier bits has a corresponding list |
| 82 | * of zones (in node_zonelists). Thus for two zone modifiers there |
| 83 | * will be a maximum of 4 (2 ** 2) zonelists, for 3 modifiers there will |
| 84 | * be 8 (2 ** 3) zonelists. GFP_ZONETYPES defines the number of possible |
| 85 | * combinations of zone modifiers in "zone modifier space". |
| 86 | */ |
| 87 | #define GFP_ZONEMASK 0x03 |
| 88 | /* |
| 89 | * As an optimisation any zone modifier bits which are only valid when |
| 90 | * no other zone modifier bits are set (loners) should be placed in |
| 91 | * the highest order bits of this field. This allows us to reduce the |
| 92 | * extent of the zonelists thus saving space. For example in the case |
| 93 | * of three zone modifier bits, we could require up to eight zonelists. |
| 94 | * If the left most zone modifier is a "loner" then the highest valid |
| 95 | * zonelist would be four allowing us to allocate only five zonelists. |
| 96 | * Use the first form when the left most bit is not a "loner", otherwise |
| 97 | * use the second. |
| 98 | */ |
| 99 | /* #define GFP_ZONETYPES (GFP_ZONEMASK + 1) */ /* Non-loner */ |
| 100 | #define GFP_ZONETYPES ((GFP_ZONEMASK + 1) / 2 + 1) /* Loner */ |
| 101 | |
| 102 | /* |
| 103 | * On machines where it is needed (eg PCs) we divide physical memory |
| 104 | * into multiple physical zones. On a PC we have 3 zones: |
| 105 | * |
| 106 | * ZONE_DMA < 16 MB ISA DMA capable memory |
| 107 | * ZONE_NORMAL 16-896 MB direct mapped by the kernel |
| 108 | * ZONE_HIGHMEM > 896 MB only page cache and user processes |
| 109 | */ |
| 110 | |
| 111 | struct zone { |
| 112 | /* Fields commonly accessed by the page allocator */ |
| 113 | unsigned long free_pages; |
| 114 | unsigned long pages_min, pages_low, pages_high; |
| 115 | /* |
| 116 | * We don't know if the memory that we're going to allocate will be freeable |
| 117 | * or/and it will be released eventually, so to avoid totally wasting several |
| 118 | * GB of ram we must reserve some of the lower zone memory (otherwise we risk |
| 119 | * to run OOM on the lower zones despite there's tons of freeable ram |
| 120 | * on the higher zones). This array is recalculated at runtime if the |
| 121 | * sysctl_lowmem_reserve_ratio sysctl changes. |
| 122 | */ |
| 123 | unsigned long lowmem_reserve[MAX_NR_ZONES]; |
| 124 | |
| 125 | struct per_cpu_pageset pageset[NR_CPUS]; |
| 126 | |
| 127 | /* |
| 128 | * free areas of different sizes |
| 129 | */ |
| 130 | spinlock_t lock; |
| 131 | struct free_area free_area[MAX_ORDER]; |
| 132 | |
| 133 | |
| 134 | ZONE_PADDING(_pad1_) |
| 135 | |
| 136 | /* Fields commonly accessed by the page reclaim scanner */ |
| 137 | spinlock_t lru_lock; |
| 138 | struct list_head active_list; |
| 139 | struct list_head inactive_list; |
| 140 | unsigned long nr_scan_active; |
| 141 | unsigned long nr_scan_inactive; |
| 142 | unsigned long nr_active; |
| 143 | unsigned long nr_inactive; |
| 144 | unsigned long pages_scanned; /* since last reclaim */ |
| 145 | int all_unreclaimable; /* All pages pinned */ |
| 146 | |
| 147 | /* |
| 148 | * prev_priority holds the scanning priority for this zone. It is |
| 149 | * defined as the scanning priority at which we achieved our reclaim |
| 150 | * target at the previous try_to_free_pages() or balance_pgdat() |
| 151 | * invokation. |
| 152 | * |
| 153 | * We use prev_priority as a measure of how much stress page reclaim is |
| 154 | * under - it drives the swappiness decision: whether to unmap mapped |
| 155 | * pages. |
| 156 | * |
| 157 | * temp_priority is used to remember the scanning priority at which |
| 158 | * this zone was successfully refilled to free_pages == pages_high. |
| 159 | * |
| 160 | * Access to both these fields is quite racy even on uniprocessor. But |
| 161 | * it is expected to average out OK. |
| 162 | */ |
| 163 | int temp_priority; |
| 164 | int prev_priority; |
| 165 | |
| 166 | |
| 167 | ZONE_PADDING(_pad2_) |
| 168 | /* Rarely used or read-mostly fields */ |
| 169 | |
| 170 | /* |
| 171 | * wait_table -- the array holding the hash table |
| 172 | * wait_table_size -- the size of the hash table array |
| 173 | * wait_table_bits -- wait_table_size == (1 << wait_table_bits) |
| 174 | * |
| 175 | * The purpose of all these is to keep track of the people |
| 176 | * waiting for a page to become available and make them |
| 177 | * runnable again when possible. The trouble is that this |
| 178 | * consumes a lot of space, especially when so few things |
| 179 | * wait on pages at a given time. So instead of using |
| 180 | * per-page waitqueues, we use a waitqueue hash table. |
| 181 | * |
| 182 | * The bucket discipline is to sleep on the same queue when |
| 183 | * colliding and wake all in that wait queue when removing. |
| 184 | * When something wakes, it must check to be sure its page is |
| 185 | * truly available, a la thundering herd. The cost of a |
| 186 | * collision is great, but given the expected load of the |
| 187 | * table, they should be so rare as to be outweighed by the |
| 188 | * benefits from the saved space. |
| 189 | * |
| 190 | * __wait_on_page_locked() and unlock_page() in mm/filemap.c, are the |
| 191 | * primary users of these fields, and in mm/page_alloc.c |
| 192 | * free_area_init_core() performs the initialization of them. |
| 193 | */ |
| 194 | wait_queue_head_t * wait_table; |
| 195 | unsigned long wait_table_size; |
| 196 | unsigned long wait_table_bits; |
| 197 | |
| 198 | /* |
| 199 | * Discontig memory support fields. |
| 200 | */ |
| 201 | struct pglist_data *zone_pgdat; |
| 202 | struct page *zone_mem_map; |
| 203 | /* zone_start_pfn == zone_start_paddr >> PAGE_SHIFT */ |
| 204 | unsigned long zone_start_pfn; |
| 205 | |
| 206 | unsigned long spanned_pages; /* total size, including holes */ |
| 207 | unsigned long present_pages; /* amount of memory (excluding holes) */ |
| 208 | |
| 209 | /* |
| 210 | * rarely used fields: |
| 211 | */ |
| 212 | char *name; |
| 213 | } ____cacheline_maxaligned_in_smp; |
| 214 | |
| 215 | |
| 216 | /* |
| 217 | * The "priority" of VM scanning is how much of the queues we will scan in one |
| 218 | * go. A value of 12 for DEF_PRIORITY implies that we will scan 1/4096th of the |
| 219 | * queues ("queue_length >> 12") during an aging round. |
| 220 | */ |
| 221 | #define DEF_PRIORITY 12 |
| 222 | |
| 223 | /* |
| 224 | * One allocation request operates on a zonelist. A zonelist |
| 225 | * is a list of zones, the first one is the 'goal' of the |
| 226 | * allocation, the other zones are fallback zones, in decreasing |
| 227 | * priority. |
| 228 | * |
| 229 | * Right now a zonelist takes up less than a cacheline. We never |
| 230 | * modify it apart from boot-up, and only a few indices are used, |
| 231 | * so despite the zonelist table being relatively big, the cache |
| 232 | * footprint of this construct is very small. |
| 233 | */ |
| 234 | struct zonelist { |
| 235 | struct zone *zones[MAX_NUMNODES * MAX_NR_ZONES + 1]; // NULL delimited |
| 236 | }; |
| 237 | |
| 238 | |
| 239 | /* |
| 240 | * The pg_data_t structure is used in machines with CONFIG_DISCONTIGMEM |
| 241 | * (mostly NUMA machines?) to denote a higher-level memory zone than the |
| 242 | * zone denotes. |
| 243 | * |
| 244 | * On NUMA machines, each NUMA node would have a pg_data_t to describe |
| 245 | * it's memory layout. |
| 246 | * |
| 247 | * Memory statistics and page replacement data structures are maintained on a |
| 248 | * per-zone basis. |
| 249 | */ |
| 250 | struct bootmem_data; |
| 251 | typedef struct pglist_data { |
| 252 | struct zone node_zones[MAX_NR_ZONES]; |
| 253 | struct zonelist node_zonelists[GFP_ZONETYPES]; |
| 254 | int nr_zones; |
| 255 | struct page *node_mem_map; |
| 256 | struct bootmem_data *bdata; |
| 257 | unsigned long node_start_pfn; |
| 258 | unsigned long node_present_pages; /* total number of physical pages */ |
| 259 | unsigned long node_spanned_pages; /* total size of physical page |
| 260 | range, including holes */ |
| 261 | int node_id; |
| 262 | struct pglist_data *pgdat_next; |
| 263 | wait_queue_head_t kswapd_wait; |
| 264 | struct task_struct *kswapd; |
| 265 | int kswapd_max_order; |
| 266 | } pg_data_t; |
| 267 | |
| 268 | #define node_present_pages(nid) (NODE_DATA(nid)->node_present_pages) |
| 269 | #define node_spanned_pages(nid) (NODE_DATA(nid)->node_spanned_pages) |
| 270 | |
| 271 | extern struct pglist_data *pgdat_list; |
| 272 | |
| 273 | void __get_zone_counts(unsigned long *active, unsigned long *inactive, |
| 274 | unsigned long *free, struct pglist_data *pgdat); |
| 275 | void get_zone_counts(unsigned long *active, unsigned long *inactive, |
| 276 | unsigned long *free); |
| 277 | void build_all_zonelists(void); |
| 278 | void wakeup_kswapd(struct zone *zone, int order); |
| 279 | int zone_watermark_ok(struct zone *z, int order, unsigned long mark, |
| 280 | int alloc_type, int can_try_harder, int gfp_high); |
| 281 | |
| 282 | #ifdef CONFIG_HAVE_MEMORY_PRESENT |
| 283 | void memory_present(int nid, unsigned long start, unsigned long end); |
| 284 | #else |
| 285 | static inline void memory_present(int nid, unsigned long start, unsigned long end) {} |
| 286 | #endif |
| 287 | |
| 288 | #ifdef CONFIG_NEED_NODE_MEMMAP_SIZE |
| 289 | unsigned long __init node_memmap_size_bytes(int, unsigned long, unsigned long); |
| 290 | #endif |
| 291 | |
| 292 | /* |
| 293 | * zone_idx() returns 0 for the ZONE_DMA zone, 1 for the ZONE_NORMAL zone, etc. |
| 294 | */ |
| 295 | #define zone_idx(zone) ((zone) - (zone)->zone_pgdat->node_zones) |
| 296 | |
| 297 | /** |
| 298 | * for_each_pgdat - helper macro to iterate over all nodes |
| 299 | * @pgdat - pointer to a pg_data_t variable |
| 300 | * |
| 301 | * Meant to help with common loops of the form |
| 302 | * pgdat = pgdat_list; |
| 303 | * while(pgdat) { |
| 304 | * ... |
| 305 | * pgdat = pgdat->pgdat_next; |
| 306 | * } |
| 307 | */ |
| 308 | #define for_each_pgdat(pgdat) \ |
| 309 | for (pgdat = pgdat_list; pgdat; pgdat = pgdat->pgdat_next) |
| 310 | |
| 311 | /* |
| 312 | * next_zone - helper magic for for_each_zone() |
| 313 | * Thanks to William Lee Irwin III for this piece of ingenuity. |
| 314 | */ |
| 315 | static inline struct zone *next_zone(struct zone *zone) |
| 316 | { |
| 317 | pg_data_t *pgdat = zone->zone_pgdat; |
| 318 | |
| 319 | if (zone < pgdat->node_zones + MAX_NR_ZONES - 1) |
| 320 | zone++; |
| 321 | else if (pgdat->pgdat_next) { |
| 322 | pgdat = pgdat->pgdat_next; |
| 323 | zone = pgdat->node_zones; |
| 324 | } else |
| 325 | zone = NULL; |
| 326 | |
| 327 | return zone; |
| 328 | } |
| 329 | |
| 330 | /** |
| 331 | * for_each_zone - helper macro to iterate over all memory zones |
| 332 | * @zone - pointer to struct zone variable |
| 333 | * |
| 334 | * The user only needs to declare the zone variable, for_each_zone |
| 335 | * fills it in. This basically means for_each_zone() is an |
| 336 | * easier to read version of this piece of code: |
| 337 | * |
| 338 | * for (pgdat = pgdat_list; pgdat; pgdat = pgdat->node_next) |
| 339 | * for (i = 0; i < MAX_NR_ZONES; ++i) { |
| 340 | * struct zone * z = pgdat->node_zones + i; |
| 341 | * ... |
| 342 | * } |
| 343 | * } |
| 344 | */ |
| 345 | #define for_each_zone(zone) \ |
| 346 | for (zone = pgdat_list->node_zones; zone; zone = next_zone(zone)) |
| 347 | |
| 348 | static inline int is_highmem_idx(int idx) |
| 349 | { |
| 350 | return (idx == ZONE_HIGHMEM); |
| 351 | } |
| 352 | |
| 353 | static inline int is_normal_idx(int idx) |
| 354 | { |
| 355 | return (idx == ZONE_NORMAL); |
| 356 | } |
| 357 | /** |
| 358 | * is_highmem - helper function to quickly check if a struct zone is a |
| 359 | * highmem zone or not. This is an attempt to keep references |
| 360 | * to ZONE_{DMA/NORMAL/HIGHMEM/etc} in general code to a minimum. |
| 361 | * @zone - pointer to struct zone variable |
| 362 | */ |
| 363 | static inline int is_highmem(struct zone *zone) |
| 364 | { |
| 365 | return zone == zone->zone_pgdat->node_zones + ZONE_HIGHMEM; |
| 366 | } |
| 367 | |
| 368 | static inline int is_normal(struct zone *zone) |
| 369 | { |
| 370 | return zone == zone->zone_pgdat->node_zones + ZONE_NORMAL; |
| 371 | } |
| 372 | |
| 373 | /* These two functions are used to setup the per zone pages min values */ |
| 374 | struct ctl_table; |
| 375 | struct file; |
| 376 | int min_free_kbytes_sysctl_handler(struct ctl_table *, int, struct file *, |
| 377 | void __user *, size_t *, loff_t *); |
| 378 | extern int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1]; |
| 379 | int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *, int, struct file *, |
| 380 | void __user *, size_t *, loff_t *); |
| 381 | |
| 382 | #include <linux/topology.h> |
| 383 | /* Returns the number of the current Node. */ |
| 384 | #define numa_node_id() (cpu_to_node(_smp_processor_id())) |
| 385 | |
| 386 | #ifndef CONFIG_DISCONTIGMEM |
| 387 | |
| 388 | extern struct pglist_data contig_page_data; |
| 389 | #define NODE_DATA(nid) (&contig_page_data) |
| 390 | #define NODE_MEM_MAP(nid) mem_map |
| 391 | #define MAX_NODES_SHIFT 1 |
| 392 | #define pfn_to_nid(pfn) (0) |
| 393 | |
| 394 | #else /* CONFIG_DISCONTIGMEM */ |
| 395 | |
| 396 | #include <asm/mmzone.h> |
| 397 | |
| 398 | #if BITS_PER_LONG == 32 || defined(ARCH_HAS_ATOMIC_UNSIGNED) |
| 399 | /* |
| 400 | * with 32 bit page->flags field, we reserve 8 bits for node/zone info. |
| 401 | * there are 3 zones (2 bits) and this leaves 8-2=6 bits for nodes. |
| 402 | */ |
| 403 | #define MAX_NODES_SHIFT 6 |
| 404 | #elif BITS_PER_LONG == 64 |
| 405 | /* |
| 406 | * with 64 bit flags field, there's plenty of room. |
| 407 | */ |
| 408 | #define MAX_NODES_SHIFT 10 |
| 409 | #endif |
| 410 | |
| 411 | #endif /* !CONFIG_DISCONTIGMEM */ |
| 412 | |
| 413 | #if NODES_SHIFT > MAX_NODES_SHIFT |
| 414 | #error NODES_SHIFT > MAX_NODES_SHIFT |
| 415 | #endif |
| 416 | |
| 417 | /* There are currently 3 zones: DMA, Normal & Highmem, thus we need 2 bits */ |
| 418 | #define MAX_ZONES_SHIFT 2 |
| 419 | |
| 420 | #if ZONES_SHIFT > MAX_ZONES_SHIFT |
| 421 | #error ZONES_SHIFT > MAX_ZONES_SHIFT |
| 422 | #endif |
| 423 | |
| 424 | #endif /* !__ASSEMBLY__ */ |
| 425 | #endif /* __KERNEL__ */ |
| 426 | #endif /* _LINUX_MMZONE_H */ |