]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/page_alloc.c | |
3 | * | |
4 | * Manages the free list, the system allocates free pages here. | |
5 | * Note that kmalloc() lives in slab.c | |
6 | * | |
7 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
8 | * Swap reorganised 29.12.95, Stephen Tweedie | |
9 | * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 | |
10 | * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 | |
11 | * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 | |
12 | * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 | |
13 | * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 | |
14 | * (lots of bits borrowed from Ingo Molnar & Andrew Morton) | |
15 | */ | |
16 | ||
17 | #include <linux/stddef.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/swap.h> | |
20 | #include <linux/interrupt.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/jiffies.h> | |
23 | #include <linux/bootmem.h> | |
24 | #include <linux/memblock.h> | |
25 | #include <linux/compiler.h> | |
26 | #include <linux/kernel.h> | |
27 | #include <linux/kmemcheck.h> | |
28 | #include <linux/module.h> | |
29 | #include <linux/suspend.h> | |
30 | #include <linux/pagevec.h> | |
31 | #include <linux/blkdev.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/ratelimit.h> | |
34 | #include <linux/oom.h> | |
35 | #include <linux/notifier.h> | |
36 | #include <linux/topology.h> | |
37 | #include <linux/sysctl.h> | |
38 | #include <linux/cpu.h> | |
39 | #include <linux/cpuset.h> | |
40 | #include <linux/memory_hotplug.h> | |
41 | #include <linux/nodemask.h> | |
42 | #include <linux/vmalloc.h> | |
43 | #include <linux/vmstat.h> | |
44 | #include <linux/mempolicy.h> | |
45 | #include <linux/stop_machine.h> | |
46 | #include <linux/sort.h> | |
47 | #include <linux/pfn.h> | |
48 | #include <linux/backing-dev.h> | |
49 | #include <linux/fault-inject.h> | |
50 | #include <linux/page-isolation.h> | |
51 | #include <linux/page_cgroup.h> | |
52 | #include <linux/debugobjects.h> | |
53 | #include <linux/kmemleak.h> | |
54 | #include <linux/memory.h> | |
55 | #include <linux/compaction.h> | |
56 | #include <trace/events/kmem.h> | |
57 | #include <linux/ftrace_event.h> | |
58 | #include <linux/memcontrol.h> | |
59 | #include <linux/prefetch.h> | |
60 | #include <linux/migrate.h> | |
61 | #include <linux/page-debug-flags.h> | |
62 | ||
63 | #include <asm/tlbflush.h> | |
64 | #include <asm/div64.h> | |
65 | #include "internal.h" | |
66 | ||
67 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID | |
68 | DEFINE_PER_CPU(int, numa_node); | |
69 | EXPORT_PER_CPU_SYMBOL(numa_node); | |
70 | #endif | |
71 | ||
72 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
73 | /* | |
74 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. | |
75 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. | |
76 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() | |
77 | * defined in <linux/topology.h>. | |
78 | */ | |
79 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ | |
80 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); | |
81 | #endif | |
82 | ||
83 | /* | |
84 | * Array of node states. | |
85 | */ | |
86 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { | |
87 | [N_POSSIBLE] = NODE_MASK_ALL, | |
88 | [N_ONLINE] = { { [0] = 1UL } }, | |
89 | #ifndef CONFIG_NUMA | |
90 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
91 | #ifdef CONFIG_HIGHMEM | |
92 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
93 | #endif | |
94 | [N_CPU] = { { [0] = 1UL } }, | |
95 | #endif /* NUMA */ | |
96 | }; | |
97 | EXPORT_SYMBOL(node_states); | |
98 | ||
99 | unsigned long totalram_pages __read_mostly; | |
100 | unsigned long totalreserve_pages __read_mostly; | |
101 | /* | |
102 | * When calculating the number of globally allowed dirty pages, there | |
103 | * is a certain number of per-zone reserves that should not be | |
104 | * considered dirtyable memory. This is the sum of those reserves | |
105 | * over all existing zones that contribute dirtyable memory. | |
106 | */ | |
107 | unsigned long dirty_balance_reserve __read_mostly; | |
108 | ||
109 | int percpu_pagelist_fraction; | |
110 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; | |
111 | ||
112 | #ifdef CONFIG_PM_SLEEP | |
113 | /* | |
114 | * The following functions are used by the suspend/hibernate code to temporarily | |
115 | * change gfp_allowed_mask in order to avoid using I/O during memory allocations | |
116 | * while devices are suspended. To avoid races with the suspend/hibernate code, | |
117 | * they should always be called with pm_mutex held (gfp_allowed_mask also should | |
118 | * only be modified with pm_mutex held, unless the suspend/hibernate code is | |
119 | * guaranteed not to run in parallel with that modification). | |
120 | */ | |
121 | ||
122 | static gfp_t saved_gfp_mask; | |
123 | ||
124 | void pm_restore_gfp_mask(void) | |
125 | { | |
126 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
127 | if (saved_gfp_mask) { | |
128 | gfp_allowed_mask = saved_gfp_mask; | |
129 | saved_gfp_mask = 0; | |
130 | } | |
131 | } | |
132 | ||
133 | void pm_restrict_gfp_mask(void) | |
134 | { | |
135 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
136 | WARN_ON(saved_gfp_mask); | |
137 | saved_gfp_mask = gfp_allowed_mask; | |
138 | gfp_allowed_mask &= ~GFP_IOFS; | |
139 | } | |
140 | ||
141 | bool pm_suspended_storage(void) | |
142 | { | |
143 | if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) | |
144 | return false; | |
145 | return true; | |
146 | } | |
147 | #endif /* CONFIG_PM_SLEEP */ | |
148 | ||
149 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
150 | int pageblock_order __read_mostly; | |
151 | #endif | |
152 | ||
153 | static void __free_pages_ok(struct page *page, unsigned int order); | |
154 | ||
155 | /* | |
156 | * results with 256, 32 in the lowmem_reserve sysctl: | |
157 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
158 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
159 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
160 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
161 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA | |
162 | * | |
163 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
164 | * don't need any ZONE_NORMAL reservation | |
165 | */ | |
166 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { | |
167 | #ifdef CONFIG_ZONE_DMA | |
168 | 256, | |
169 | #endif | |
170 | #ifdef CONFIG_ZONE_DMA32 | |
171 | 256, | |
172 | #endif | |
173 | #ifdef CONFIG_HIGHMEM | |
174 | 32, | |
175 | #endif | |
176 | 32, | |
177 | }; | |
178 | ||
179 | EXPORT_SYMBOL(totalram_pages); | |
180 | ||
181 | static char * const zone_names[MAX_NR_ZONES] = { | |
182 | #ifdef CONFIG_ZONE_DMA | |
183 | "DMA", | |
184 | #endif | |
185 | #ifdef CONFIG_ZONE_DMA32 | |
186 | "DMA32", | |
187 | #endif | |
188 | "Normal", | |
189 | #ifdef CONFIG_HIGHMEM | |
190 | "HighMem", | |
191 | #endif | |
192 | "Movable", | |
193 | }; | |
194 | ||
195 | int min_free_kbytes = 1024; | |
196 | ||
197 | static unsigned long __meminitdata nr_kernel_pages; | |
198 | static unsigned long __meminitdata nr_all_pages; | |
199 | static unsigned long __meminitdata dma_reserve; | |
200 | ||
201 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
202 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; | |
203 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; | |
204 | static unsigned long __initdata required_kernelcore; | |
205 | static unsigned long __initdata required_movablecore; | |
206 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; | |
207 | ||
208 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
209 | int movable_zone; | |
210 | EXPORT_SYMBOL(movable_zone); | |
211 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
212 | ||
213 | #if MAX_NUMNODES > 1 | |
214 | int nr_node_ids __read_mostly = MAX_NUMNODES; | |
215 | int nr_online_nodes __read_mostly = 1; | |
216 | EXPORT_SYMBOL(nr_node_ids); | |
217 | EXPORT_SYMBOL(nr_online_nodes); | |
218 | #endif | |
219 | ||
220 | int page_group_by_mobility_disabled __read_mostly; | |
221 | ||
222 | static void set_pageblock_migratetype(struct page *page, int migratetype) | |
223 | { | |
224 | ||
225 | if (unlikely(page_group_by_mobility_disabled)) | |
226 | migratetype = MIGRATE_UNMOVABLE; | |
227 | ||
228 | set_pageblock_flags_group(page, (unsigned long)migratetype, | |
229 | PB_migrate, PB_migrate_end); | |
230 | } | |
231 | ||
232 | bool oom_killer_disabled __read_mostly; | |
233 | ||
234 | #ifdef CONFIG_DEBUG_VM | |
235 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) | |
236 | { | |
237 | int ret = 0; | |
238 | unsigned seq; | |
239 | unsigned long pfn = page_to_pfn(page); | |
240 | ||
241 | do { | |
242 | seq = zone_span_seqbegin(zone); | |
243 | if (pfn >= zone->zone_start_pfn + zone->spanned_pages) | |
244 | ret = 1; | |
245 | else if (pfn < zone->zone_start_pfn) | |
246 | ret = 1; | |
247 | } while (zone_span_seqretry(zone, seq)); | |
248 | ||
249 | return ret; | |
250 | } | |
251 | ||
252 | static int page_is_consistent(struct zone *zone, struct page *page) | |
253 | { | |
254 | if (!pfn_valid_within(page_to_pfn(page))) | |
255 | return 0; | |
256 | if (zone != page_zone(page)) | |
257 | return 0; | |
258 | ||
259 | return 1; | |
260 | } | |
261 | /* | |
262 | * Temporary debugging check for pages not lying within a given zone. | |
263 | */ | |
264 | static int bad_range(struct zone *zone, struct page *page) | |
265 | { | |
266 | if (page_outside_zone_boundaries(zone, page)) | |
267 | return 1; | |
268 | if (!page_is_consistent(zone, page)) | |
269 | return 1; | |
270 | ||
271 | return 0; | |
272 | } | |
273 | #else | |
274 | static inline int bad_range(struct zone *zone, struct page *page) | |
275 | { | |
276 | return 0; | |
277 | } | |
278 | #endif | |
279 | ||
280 | static void bad_page(struct page *page) | |
281 | { | |
282 | static unsigned long resume; | |
283 | static unsigned long nr_shown; | |
284 | static unsigned long nr_unshown; | |
285 | ||
286 | /* Don't complain about poisoned pages */ | |
287 | if (PageHWPoison(page)) { | |
288 | reset_page_mapcount(page); /* remove PageBuddy */ | |
289 | return; | |
290 | } | |
291 | ||
292 | /* | |
293 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
294 | * or allow a steady drip of one report per second. | |
295 | */ | |
296 | if (nr_shown == 60) { | |
297 | if (time_before(jiffies, resume)) { | |
298 | nr_unshown++; | |
299 | goto out; | |
300 | } | |
301 | if (nr_unshown) { | |
302 | printk(KERN_ALERT | |
303 | "BUG: Bad page state: %lu messages suppressed\n", | |
304 | nr_unshown); | |
305 | nr_unshown = 0; | |
306 | } | |
307 | nr_shown = 0; | |
308 | } | |
309 | if (nr_shown++ == 0) | |
310 | resume = jiffies + 60 * HZ; | |
311 | ||
312 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", | |
313 | current->comm, page_to_pfn(page)); | |
314 | dump_page(page); | |
315 | ||
316 | print_modules(); | |
317 | dump_stack(); | |
318 | out: | |
319 | /* Leave bad fields for debug, except PageBuddy could make trouble */ | |
320 | reset_page_mapcount(page); /* remove PageBuddy */ | |
321 | add_taint(TAINT_BAD_PAGE); | |
322 | } | |
323 | ||
324 | /* | |
325 | * Higher-order pages are called "compound pages". They are structured thusly: | |
326 | * | |
327 | * The first PAGE_SIZE page is called the "head page". | |
328 | * | |
329 | * The remaining PAGE_SIZE pages are called "tail pages". | |
330 | * | |
331 | * All pages have PG_compound set. All tail pages have their ->first_page | |
332 | * pointing at the head page. | |
333 | * | |
334 | * The first tail page's ->lru.next holds the address of the compound page's | |
335 | * put_page() function. Its ->lru.prev holds the order of allocation. | |
336 | * This usage means that zero-order pages may not be compound. | |
337 | */ | |
338 | ||
339 | static void free_compound_page(struct page *page) | |
340 | { | |
341 | __free_pages_ok(page, compound_order(page)); | |
342 | } | |
343 | ||
344 | void prep_compound_page(struct page *page, unsigned long order) | |
345 | { | |
346 | int i; | |
347 | int nr_pages = 1 << order; | |
348 | ||
349 | set_compound_page_dtor(page, free_compound_page); | |
350 | set_compound_order(page, order); | |
351 | __SetPageHead(page); | |
352 | for (i = 1; i < nr_pages; i++) { | |
353 | struct page *p = page + i; | |
354 | __SetPageTail(p); | |
355 | set_page_count(p, 0); | |
356 | p->first_page = page; | |
357 | } | |
358 | } | |
359 | ||
360 | /* update __split_huge_page_refcount if you change this function */ | |
361 | static int destroy_compound_page(struct page *page, unsigned long order) | |
362 | { | |
363 | int i; | |
364 | int nr_pages = 1 << order; | |
365 | int bad = 0; | |
366 | ||
367 | if (unlikely(compound_order(page) != order) || | |
368 | unlikely(!PageHead(page))) { | |
369 | bad_page(page); | |
370 | bad++; | |
371 | } | |
372 | ||
373 | __ClearPageHead(page); | |
374 | ||
375 | for (i = 1; i < nr_pages; i++) { | |
376 | struct page *p = page + i; | |
377 | ||
378 | if (unlikely(!PageTail(p) || (p->first_page != page))) { | |
379 | bad_page(page); | |
380 | bad++; | |
381 | } | |
382 | __ClearPageTail(p); | |
383 | } | |
384 | ||
385 | return bad; | |
386 | } | |
387 | ||
388 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) | |
389 | { | |
390 | int i; | |
391 | ||
392 | /* | |
393 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO | |
394 | * and __GFP_HIGHMEM from hard or soft interrupt context. | |
395 | */ | |
396 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); | |
397 | for (i = 0; i < (1 << order); i++) | |
398 | clear_highpage(page + i); | |
399 | } | |
400 | ||
401 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
402 | unsigned int _debug_guardpage_minorder; | |
403 | ||
404 | static int __init debug_guardpage_minorder_setup(char *buf) | |
405 | { | |
406 | unsigned long res; | |
407 | ||
408 | if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { | |
409 | printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); | |
410 | return 0; | |
411 | } | |
412 | _debug_guardpage_minorder = res; | |
413 | printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); | |
414 | return 0; | |
415 | } | |
416 | __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); | |
417 | ||
418 | static inline void set_page_guard_flag(struct page *page) | |
419 | { | |
420 | __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); | |
421 | } | |
422 | ||
423 | static inline void clear_page_guard_flag(struct page *page) | |
424 | { | |
425 | __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); | |
426 | } | |
427 | #else | |
428 | static inline void set_page_guard_flag(struct page *page) { } | |
429 | static inline void clear_page_guard_flag(struct page *page) { } | |
430 | #endif | |
431 | ||
432 | static inline void set_page_order(struct page *page, int order) | |
433 | { | |
434 | set_page_private(page, order); | |
435 | __SetPageBuddy(page); | |
436 | } | |
437 | ||
438 | static inline void rmv_page_order(struct page *page) | |
439 | { | |
440 | __ClearPageBuddy(page); | |
441 | set_page_private(page, 0); | |
442 | } | |
443 | ||
444 | /* | |
445 | * Locate the struct page for both the matching buddy in our | |
446 | * pair (buddy1) and the combined O(n+1) page they form (page). | |
447 | * | |
448 | * 1) Any buddy B1 will have an order O twin B2 which satisfies | |
449 | * the following equation: | |
450 | * B2 = B1 ^ (1 << O) | |
451 | * For example, if the starting buddy (buddy2) is #8 its order | |
452 | * 1 buddy is #10: | |
453 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 | |
454 | * | |
455 | * 2) Any buddy B will have an order O+1 parent P which | |
456 | * satisfies the following equation: | |
457 | * P = B & ~(1 << O) | |
458 | * | |
459 | * Assumption: *_mem_map is contiguous at least up to MAX_ORDER | |
460 | */ | |
461 | static inline unsigned long | |
462 | __find_buddy_index(unsigned long page_idx, unsigned int order) | |
463 | { | |
464 | return page_idx ^ (1 << order); | |
465 | } | |
466 | ||
467 | /* | |
468 | * This function checks whether a page is free && is the buddy | |
469 | * we can do coalesce a page and its buddy if | |
470 | * (a) the buddy is not in a hole && | |
471 | * (b) the buddy is in the buddy system && | |
472 | * (c) a page and its buddy have the same order && | |
473 | * (d) a page and its buddy are in the same zone. | |
474 | * | |
475 | * For recording whether a page is in the buddy system, we set ->_mapcount -2. | |
476 | * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. | |
477 | * | |
478 | * For recording page's order, we use page_private(page). | |
479 | */ | |
480 | static inline int page_is_buddy(struct page *page, struct page *buddy, | |
481 | int order) | |
482 | { | |
483 | if (!pfn_valid_within(page_to_pfn(buddy))) | |
484 | return 0; | |
485 | ||
486 | if (page_zone_id(page) != page_zone_id(buddy)) | |
487 | return 0; | |
488 | ||
489 | if (page_is_guard(buddy) && page_order(buddy) == order) { | |
490 | VM_BUG_ON(page_count(buddy) != 0); | |
491 | return 1; | |
492 | } | |
493 | ||
494 | if (PageBuddy(buddy) && page_order(buddy) == order) { | |
495 | VM_BUG_ON(page_count(buddy) != 0); | |
496 | return 1; | |
497 | } | |
498 | return 0; | |
499 | } | |
500 | ||
501 | /* | |
502 | * Freeing function for a buddy system allocator. | |
503 | * | |
504 | * The concept of a buddy system is to maintain direct-mapped table | |
505 | * (containing bit values) for memory blocks of various "orders". | |
506 | * The bottom level table contains the map for the smallest allocatable | |
507 | * units of memory (here, pages), and each level above it describes | |
508 | * pairs of units from the levels below, hence, "buddies". | |
509 | * At a high level, all that happens here is marking the table entry | |
510 | * at the bottom level available, and propagating the changes upward | |
511 | * as necessary, plus some accounting needed to play nicely with other | |
512 | * parts of the VM system. | |
513 | * At each level, we keep a list of pages, which are heads of continuous | |
514 | * free pages of length of (1 << order) and marked with _mapcount -2. Page's | |
515 | * order is recorded in page_private(page) field. | |
516 | * So when we are allocating or freeing one, we can derive the state of the | |
517 | * other. That is, if we allocate a small block, and both were | |
518 | * free, the remainder of the region must be split into blocks. | |
519 | * If a block is freed, and its buddy is also free, then this | |
520 | * triggers coalescing into a block of larger size. | |
521 | * | |
522 | * -- wli | |
523 | */ | |
524 | ||
525 | static inline void __free_one_page(struct page *page, | |
526 | struct zone *zone, unsigned int order, | |
527 | int migratetype) | |
528 | { | |
529 | unsigned long page_idx; | |
530 | unsigned long combined_idx; | |
531 | unsigned long uninitialized_var(buddy_idx); | |
532 | struct page *buddy; | |
533 | ||
534 | if (unlikely(PageCompound(page))) | |
535 | if (unlikely(destroy_compound_page(page, order))) | |
536 | return; | |
537 | ||
538 | VM_BUG_ON(migratetype == -1); | |
539 | ||
540 | page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); | |
541 | ||
542 | VM_BUG_ON(page_idx & ((1 << order) - 1)); | |
543 | VM_BUG_ON(bad_range(zone, page)); | |
544 | ||
545 | while (order < MAX_ORDER-1) { | |
546 | buddy_idx = __find_buddy_index(page_idx, order); | |
547 | buddy = page + (buddy_idx - page_idx); | |
548 | if (!page_is_buddy(page, buddy, order)) | |
549 | break; | |
550 | /* | |
551 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, | |
552 | * merge with it and move up one order. | |
553 | */ | |
554 | if (page_is_guard(buddy)) { | |
555 | clear_page_guard_flag(buddy); | |
556 | set_page_private(page, 0); | |
557 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); | |
558 | } else { | |
559 | list_del(&buddy->lru); | |
560 | zone->free_area[order].nr_free--; | |
561 | rmv_page_order(buddy); | |
562 | } | |
563 | combined_idx = buddy_idx & page_idx; | |
564 | page = page + (combined_idx - page_idx); | |
565 | page_idx = combined_idx; | |
566 | order++; | |
567 | } | |
568 | set_page_order(page, order); | |
569 | ||
570 | /* | |
571 | * If this is not the largest possible page, check if the buddy | |
572 | * of the next-highest order is free. If it is, it's possible | |
573 | * that pages are being freed that will coalesce soon. In case, | |
574 | * that is happening, add the free page to the tail of the list | |
575 | * so it's less likely to be used soon and more likely to be merged | |
576 | * as a higher order page | |
577 | */ | |
578 | if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { | |
579 | struct page *higher_page, *higher_buddy; | |
580 | combined_idx = buddy_idx & page_idx; | |
581 | higher_page = page + (combined_idx - page_idx); | |
582 | buddy_idx = __find_buddy_index(combined_idx, order + 1); | |
583 | higher_buddy = page + (buddy_idx - combined_idx); | |
584 | if (page_is_buddy(higher_page, higher_buddy, order + 1)) { | |
585 | list_add_tail(&page->lru, | |
586 | &zone->free_area[order].free_list[migratetype]); | |
587 | goto out; | |
588 | } | |
589 | } | |
590 | ||
591 | list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); | |
592 | out: | |
593 | zone->free_area[order].nr_free++; | |
594 | } | |
595 | ||
596 | /* | |
597 | * free_page_mlock() -- clean up attempts to free and mlocked() page. | |
598 | * Page should not be on lru, so no need to fix that up. | |
599 | * free_pages_check() will verify... | |
600 | */ | |
601 | static inline void free_page_mlock(struct page *page) | |
602 | { | |
603 | __dec_zone_page_state(page, NR_MLOCK); | |
604 | __count_vm_event(UNEVICTABLE_MLOCKFREED); | |
605 | } | |
606 | ||
607 | static inline int free_pages_check(struct page *page) | |
608 | { | |
609 | if (unlikely(page_mapcount(page) | | |
610 | (page->mapping != NULL) | | |
611 | (atomic_read(&page->_count) != 0) | | |
612 | (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | | |
613 | (mem_cgroup_bad_page_check(page)))) { | |
614 | bad_page(page); | |
615 | return 1; | |
616 | } | |
617 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | |
618 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
619 | return 0; | |
620 | } | |
621 | ||
622 | /* | |
623 | * Frees a number of pages from the PCP lists | |
624 | * Assumes all pages on list are in same zone, and of same order. | |
625 | * count is the number of pages to free. | |
626 | * | |
627 | * If the zone was previously in an "all pages pinned" state then look to | |
628 | * see if this freeing clears that state. | |
629 | * | |
630 | * And clear the zone's pages_scanned counter, to hold off the "all pages are | |
631 | * pinned" detection logic. | |
632 | */ | |
633 | static void free_pcppages_bulk(struct zone *zone, int count, | |
634 | struct per_cpu_pages *pcp) | |
635 | { | |
636 | int migratetype = 0; | |
637 | int batch_free = 0; | |
638 | int to_free = count; | |
639 | ||
640 | spin_lock(&zone->lock); | |
641 | zone->all_unreclaimable = 0; | |
642 | zone->pages_scanned = 0; | |
643 | ||
644 | while (to_free) { | |
645 | struct page *page; | |
646 | struct list_head *list; | |
647 | ||
648 | /* | |
649 | * Remove pages from lists in a round-robin fashion. A | |
650 | * batch_free count is maintained that is incremented when an | |
651 | * empty list is encountered. This is so more pages are freed | |
652 | * off fuller lists instead of spinning excessively around empty | |
653 | * lists | |
654 | */ | |
655 | do { | |
656 | batch_free++; | |
657 | if (++migratetype == MIGRATE_PCPTYPES) | |
658 | migratetype = 0; | |
659 | list = &pcp->lists[migratetype]; | |
660 | } while (list_empty(list)); | |
661 | ||
662 | /* This is the only non-empty list. Free them all. */ | |
663 | if (batch_free == MIGRATE_PCPTYPES) | |
664 | batch_free = to_free; | |
665 | ||
666 | do { | |
667 | page = list_entry(list->prev, struct page, lru); | |
668 | /* must delete as __free_one_page list manipulates */ | |
669 | list_del(&page->lru); | |
670 | /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ | |
671 | __free_one_page(page, zone, 0, page_private(page)); | |
672 | trace_mm_page_pcpu_drain(page, 0, page_private(page)); | |
673 | } while (--to_free && --batch_free && !list_empty(list)); | |
674 | } | |
675 | __mod_zone_page_state(zone, NR_FREE_PAGES, count); | |
676 | spin_unlock(&zone->lock); | |
677 | } | |
678 | ||
679 | static void free_one_page(struct zone *zone, struct page *page, int order, | |
680 | int migratetype) | |
681 | { | |
682 | spin_lock(&zone->lock); | |
683 | zone->all_unreclaimable = 0; | |
684 | zone->pages_scanned = 0; | |
685 | ||
686 | __free_one_page(page, zone, order, migratetype); | |
687 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); | |
688 | spin_unlock(&zone->lock); | |
689 | } | |
690 | ||
691 | static bool free_pages_prepare(struct page *page, unsigned int order) | |
692 | { | |
693 | int i; | |
694 | int bad = 0; | |
695 | ||
696 | trace_mm_page_free(page, order); | |
697 | kmemcheck_free_shadow(page, order); | |
698 | ||
699 | if (PageAnon(page)) | |
700 | page->mapping = NULL; | |
701 | for (i = 0; i < (1 << order); i++) | |
702 | bad += free_pages_check(page + i); | |
703 | if (bad) | |
704 | return false; | |
705 | ||
706 | if (!PageHighMem(page)) { | |
707 | debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); | |
708 | debug_check_no_obj_freed(page_address(page), | |
709 | PAGE_SIZE << order); | |
710 | } | |
711 | arch_free_page(page, order); | |
712 | kernel_map_pages(page, 1 << order, 0); | |
713 | ||
714 | return true; | |
715 | } | |
716 | ||
717 | static void __free_pages_ok(struct page *page, unsigned int order) | |
718 | { | |
719 | unsigned long flags; | |
720 | int wasMlocked = __TestClearPageMlocked(page); | |
721 | ||
722 | if (!free_pages_prepare(page, order)) | |
723 | return; | |
724 | ||
725 | local_irq_save(flags); | |
726 | if (unlikely(wasMlocked)) | |
727 | free_page_mlock(page); | |
728 | __count_vm_events(PGFREE, 1 << order); | |
729 | free_one_page(page_zone(page), page, order, | |
730 | get_pageblock_migratetype(page)); | |
731 | local_irq_restore(flags); | |
732 | } | |
733 | ||
734 | void __meminit __free_pages_bootmem(struct page *page, unsigned int order) | |
735 | { | |
736 | unsigned int nr_pages = 1 << order; | |
737 | unsigned int loop; | |
738 | ||
739 | prefetchw(page); | |
740 | for (loop = 0; loop < nr_pages; loop++) { | |
741 | struct page *p = &page[loop]; | |
742 | ||
743 | if (loop + 1 < nr_pages) | |
744 | prefetchw(p + 1); | |
745 | __ClearPageReserved(p); | |
746 | set_page_count(p, 0); | |
747 | } | |
748 | ||
749 | set_page_refcounted(page); | |
750 | __free_pages(page, order); | |
751 | } | |
752 | ||
753 | #ifdef CONFIG_CMA | |
754 | /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */ | |
755 | void __init init_cma_reserved_pageblock(struct page *page) | |
756 | { | |
757 | unsigned i = pageblock_nr_pages; | |
758 | struct page *p = page; | |
759 | ||
760 | do { | |
761 | __ClearPageReserved(p); | |
762 | set_page_count(p, 0); | |
763 | } while (++p, --i); | |
764 | ||
765 | set_page_refcounted(page); | |
766 | set_pageblock_migratetype(page, MIGRATE_CMA); | |
767 | __free_pages(page, pageblock_order); | |
768 | totalram_pages += pageblock_nr_pages; | |
769 | } | |
770 | #endif | |
771 | ||
772 | /* | |
773 | * The order of subdivision here is critical for the IO subsystem. | |
774 | * Please do not alter this order without good reasons and regression | |
775 | * testing. Specifically, as large blocks of memory are subdivided, | |
776 | * the order in which smaller blocks are delivered depends on the order | |
777 | * they're subdivided in this function. This is the primary factor | |
778 | * influencing the order in which pages are delivered to the IO | |
779 | * subsystem according to empirical testing, and this is also justified | |
780 | * by considering the behavior of a buddy system containing a single | |
781 | * large block of memory acted on by a series of small allocations. | |
782 | * This behavior is a critical factor in sglist merging's success. | |
783 | * | |
784 | * -- wli | |
785 | */ | |
786 | static inline void expand(struct zone *zone, struct page *page, | |
787 | int low, int high, struct free_area *area, | |
788 | int migratetype) | |
789 | { | |
790 | unsigned long size = 1 << high; | |
791 | ||
792 | while (high > low) { | |
793 | area--; | |
794 | high--; | |
795 | size >>= 1; | |
796 | VM_BUG_ON(bad_range(zone, &page[size])); | |
797 | ||
798 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
799 | if (high < debug_guardpage_minorder()) { | |
800 | /* | |
801 | * Mark as guard pages (or page), that will allow to | |
802 | * merge back to allocator when buddy will be freed. | |
803 | * Corresponding page table entries will not be touched, | |
804 | * pages will stay not present in virtual address space | |
805 | */ | |
806 | INIT_LIST_HEAD(&page[size].lru); | |
807 | set_page_guard_flag(&page[size]); | |
808 | set_page_private(&page[size], high); | |
809 | /* Guard pages are not available for any usage */ | |
810 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high)); | |
811 | continue; | |
812 | } | |
813 | #endif | |
814 | list_add(&page[size].lru, &area->free_list[migratetype]); | |
815 | area->nr_free++; | |
816 | set_page_order(&page[size], high); | |
817 | } | |
818 | } | |
819 | ||
820 | /* | |
821 | * This page is about to be returned from the page allocator | |
822 | */ | |
823 | static inline int check_new_page(struct page *page) | |
824 | { | |
825 | if (unlikely(page_mapcount(page) | | |
826 | (page->mapping != NULL) | | |
827 | (atomic_read(&page->_count) != 0) | | |
828 | (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | | |
829 | (mem_cgroup_bad_page_check(page)))) { | |
830 | bad_page(page); | |
831 | return 1; | |
832 | } | |
833 | return 0; | |
834 | } | |
835 | ||
836 | static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) | |
837 | { | |
838 | int i; | |
839 | ||
840 | for (i = 0; i < (1 << order); i++) { | |
841 | struct page *p = page + i; | |
842 | if (unlikely(check_new_page(p))) | |
843 | return 1; | |
844 | } | |
845 | ||
846 | set_page_private(page, 0); | |
847 | set_page_refcounted(page); | |
848 | ||
849 | arch_alloc_page(page, order); | |
850 | kernel_map_pages(page, 1 << order, 1); | |
851 | ||
852 | if (gfp_flags & __GFP_ZERO) | |
853 | prep_zero_page(page, order, gfp_flags); | |
854 | ||
855 | if (order && (gfp_flags & __GFP_COMP)) | |
856 | prep_compound_page(page, order); | |
857 | ||
858 | return 0; | |
859 | } | |
860 | ||
861 | /* | |
862 | * Go through the free lists for the given migratetype and remove | |
863 | * the smallest available page from the freelists | |
864 | */ | |
865 | static inline | |
866 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, | |
867 | int migratetype) | |
868 | { | |
869 | unsigned int current_order; | |
870 | struct free_area * area; | |
871 | struct page *page; | |
872 | ||
873 | /* Find a page of the appropriate size in the preferred list */ | |
874 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { | |
875 | area = &(zone->free_area[current_order]); | |
876 | if (list_empty(&area->free_list[migratetype])) | |
877 | continue; | |
878 | ||
879 | page = list_entry(area->free_list[migratetype].next, | |
880 | struct page, lru); | |
881 | list_del(&page->lru); | |
882 | rmv_page_order(page); | |
883 | area->nr_free--; | |
884 | expand(zone, page, order, current_order, area, migratetype); | |
885 | return page; | |
886 | } | |
887 | ||
888 | return NULL; | |
889 | } | |
890 | ||
891 | ||
892 | /* | |
893 | * This array describes the order lists are fallen back to when | |
894 | * the free lists for the desirable migrate type are depleted | |
895 | */ | |
896 | static int fallbacks[MIGRATE_TYPES][4] = { | |
897 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
898 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
899 | #ifdef CONFIG_CMA | |
900 | [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
901 | [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ | |
902 | #else | |
903 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
904 | #endif | |
905 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ | |
906 | [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ | |
907 | }; | |
908 | ||
909 | /* | |
910 | * Move the free pages in a range to the free lists of the requested type. | |
911 | * Note that start_page and end_pages are not aligned on a pageblock | |
912 | * boundary. If alignment is required, use move_freepages_block() | |
913 | */ | |
914 | static int move_freepages(struct zone *zone, | |
915 | struct page *start_page, struct page *end_page, | |
916 | int migratetype) | |
917 | { | |
918 | struct page *page; | |
919 | unsigned long order; | |
920 | int pages_moved = 0; | |
921 | ||
922 | #ifndef CONFIG_HOLES_IN_ZONE | |
923 | /* | |
924 | * page_zone is not safe to call in this context when | |
925 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant | |
926 | * anyway as we check zone boundaries in move_freepages_block(). | |
927 | * Remove at a later date when no bug reports exist related to | |
928 | * grouping pages by mobility | |
929 | */ | |
930 | BUG_ON(page_zone(start_page) != page_zone(end_page)); | |
931 | #endif | |
932 | ||
933 | for (page = start_page; page <= end_page;) { | |
934 | /* Make sure we are not inadvertently changing nodes */ | |
935 | VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); | |
936 | ||
937 | if (!pfn_valid_within(page_to_pfn(page))) { | |
938 | page++; | |
939 | continue; | |
940 | } | |
941 | ||
942 | if (!PageBuddy(page)) { | |
943 | page++; | |
944 | continue; | |
945 | } | |
946 | ||
947 | order = page_order(page); | |
948 | list_move(&page->lru, | |
949 | &zone->free_area[order].free_list[migratetype]); | |
950 | page += 1 << order; | |
951 | pages_moved += 1 << order; | |
952 | } | |
953 | ||
954 | return pages_moved; | |
955 | } | |
956 | ||
957 | static int move_freepages_block(struct zone *zone, struct page *page, | |
958 | int migratetype) | |
959 | { | |
960 | unsigned long start_pfn, end_pfn; | |
961 | struct page *start_page, *end_page; | |
962 | ||
963 | start_pfn = page_to_pfn(page); | |
964 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); | |
965 | start_page = pfn_to_page(start_pfn); | |
966 | end_page = start_page + pageblock_nr_pages - 1; | |
967 | end_pfn = start_pfn + pageblock_nr_pages - 1; | |
968 | ||
969 | /* Do not cross zone boundaries */ | |
970 | if (start_pfn < zone->zone_start_pfn) | |
971 | start_page = page; | |
972 | if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) | |
973 | return 0; | |
974 | ||
975 | return move_freepages(zone, start_page, end_page, migratetype); | |
976 | } | |
977 | ||
978 | static void change_pageblock_range(struct page *pageblock_page, | |
979 | int start_order, int migratetype) | |
980 | { | |
981 | int nr_pageblocks = 1 << (start_order - pageblock_order); | |
982 | ||
983 | while (nr_pageblocks--) { | |
984 | set_pageblock_migratetype(pageblock_page, migratetype); | |
985 | pageblock_page += pageblock_nr_pages; | |
986 | } | |
987 | } | |
988 | ||
989 | /* Remove an element from the buddy allocator from the fallback list */ | |
990 | static inline struct page * | |
991 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) | |
992 | { | |
993 | struct free_area * area; | |
994 | int current_order; | |
995 | struct page *page; | |
996 | int migratetype, i; | |
997 | ||
998 | /* Find the largest possible block of pages in the other list */ | |
999 | for (current_order = MAX_ORDER-1; current_order >= order; | |
1000 | --current_order) { | |
1001 | for (i = 0;; i++) { | |
1002 | migratetype = fallbacks[start_migratetype][i]; | |
1003 | ||
1004 | /* MIGRATE_RESERVE handled later if necessary */ | |
1005 | if (migratetype == MIGRATE_RESERVE) | |
1006 | break; | |
1007 | ||
1008 | area = &(zone->free_area[current_order]); | |
1009 | if (list_empty(&area->free_list[migratetype])) | |
1010 | continue; | |
1011 | ||
1012 | page = list_entry(area->free_list[migratetype].next, | |
1013 | struct page, lru); | |
1014 | area->nr_free--; | |
1015 | ||
1016 | /* | |
1017 | * If breaking a large block of pages, move all free | |
1018 | * pages to the preferred allocation list. If falling | |
1019 | * back for a reclaimable kernel allocation, be more | |
1020 | * aggressive about taking ownership of free pages | |
1021 | * | |
1022 | * On the other hand, never change migration | |
1023 | * type of MIGRATE_CMA pageblocks nor move CMA | |
1024 | * pages on different free lists. We don't | |
1025 | * want unmovable pages to be allocated from | |
1026 | * MIGRATE_CMA areas. | |
1027 | */ | |
1028 | if (!is_migrate_cma(migratetype) && | |
1029 | (unlikely(current_order >= pageblock_order / 2) || | |
1030 | start_migratetype == MIGRATE_RECLAIMABLE || | |
1031 | page_group_by_mobility_disabled)) { | |
1032 | int pages; | |
1033 | pages = move_freepages_block(zone, page, | |
1034 | start_migratetype); | |
1035 | ||
1036 | /* Claim the whole block if over half of it is free */ | |
1037 | if (pages >= (1 << (pageblock_order-1)) || | |
1038 | page_group_by_mobility_disabled) | |
1039 | set_pageblock_migratetype(page, | |
1040 | start_migratetype); | |
1041 | ||
1042 | migratetype = start_migratetype; | |
1043 | } | |
1044 | ||
1045 | /* Remove the page from the freelists */ | |
1046 | list_del(&page->lru); | |
1047 | rmv_page_order(page); | |
1048 | ||
1049 | /* Take ownership for orders >= pageblock_order */ | |
1050 | if (current_order >= pageblock_order && | |
1051 | !is_migrate_cma(migratetype)) | |
1052 | change_pageblock_range(page, current_order, | |
1053 | start_migratetype); | |
1054 | ||
1055 | expand(zone, page, order, current_order, area, | |
1056 | is_migrate_cma(migratetype) | |
1057 | ? migratetype : start_migratetype); | |
1058 | ||
1059 | trace_mm_page_alloc_extfrag(page, order, current_order, | |
1060 | start_migratetype, migratetype); | |
1061 | ||
1062 | return page; | |
1063 | } | |
1064 | } | |
1065 | ||
1066 | return NULL; | |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * Do the hard work of removing an element from the buddy allocator. | |
1071 | * Call me with the zone->lock already held. | |
1072 | */ | |
1073 | static struct page *__rmqueue(struct zone *zone, unsigned int order, | |
1074 | int migratetype) | |
1075 | { | |
1076 | struct page *page; | |
1077 | ||
1078 | retry_reserve: | |
1079 | page = __rmqueue_smallest(zone, order, migratetype); | |
1080 | ||
1081 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { | |
1082 | page = __rmqueue_fallback(zone, order, migratetype); | |
1083 | ||
1084 | /* | |
1085 | * Use MIGRATE_RESERVE rather than fail an allocation. goto | |
1086 | * is used because __rmqueue_smallest is an inline function | |
1087 | * and we want just one call site | |
1088 | */ | |
1089 | if (!page) { | |
1090 | migratetype = MIGRATE_RESERVE; | |
1091 | goto retry_reserve; | |
1092 | } | |
1093 | } | |
1094 | ||
1095 | trace_mm_page_alloc_zone_locked(page, order, migratetype); | |
1096 | return page; | |
1097 | } | |
1098 | ||
1099 | /* | |
1100 | * Obtain a specified number of elements from the buddy allocator, all under | |
1101 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
1102 | * Returns the number of new pages which were placed at *list. | |
1103 | */ | |
1104 | static int rmqueue_bulk(struct zone *zone, unsigned int order, | |
1105 | unsigned long count, struct list_head *list, | |
1106 | int migratetype, int cold) | |
1107 | { | |
1108 | int mt = migratetype, i; | |
1109 | ||
1110 | spin_lock(&zone->lock); | |
1111 | for (i = 0; i < count; ++i) { | |
1112 | struct page *page = __rmqueue(zone, order, migratetype); | |
1113 | if (unlikely(page == NULL)) | |
1114 | break; | |
1115 | ||
1116 | /* | |
1117 | * Split buddy pages returned by expand() are received here | |
1118 | * in physical page order. The page is added to the callers and | |
1119 | * list and the list head then moves forward. From the callers | |
1120 | * perspective, the linked list is ordered by page number in | |
1121 | * some conditions. This is useful for IO devices that can | |
1122 | * merge IO requests if the physical pages are ordered | |
1123 | * properly. | |
1124 | */ | |
1125 | if (likely(cold == 0)) | |
1126 | list_add(&page->lru, list); | |
1127 | else | |
1128 | list_add_tail(&page->lru, list); | |
1129 | if (IS_ENABLED(CONFIG_CMA)) { | |
1130 | mt = get_pageblock_migratetype(page); | |
1131 | if (!is_migrate_cma(mt) && mt != MIGRATE_ISOLATE) | |
1132 | mt = migratetype; | |
1133 | } | |
1134 | set_page_private(page, mt); | |
1135 | list = &page->lru; | |
1136 | } | |
1137 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); | |
1138 | spin_unlock(&zone->lock); | |
1139 | return i; | |
1140 | } | |
1141 | ||
1142 | #ifdef CONFIG_NUMA | |
1143 | /* | |
1144 | * Called from the vmstat counter updater to drain pagesets of this | |
1145 | * currently executing processor on remote nodes after they have | |
1146 | * expired. | |
1147 | * | |
1148 | * Note that this function must be called with the thread pinned to | |
1149 | * a single processor. | |
1150 | */ | |
1151 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) | |
1152 | { | |
1153 | unsigned long flags; | |
1154 | int to_drain; | |
1155 | ||
1156 | local_irq_save(flags); | |
1157 | if (pcp->count >= pcp->batch) | |
1158 | to_drain = pcp->batch; | |
1159 | else | |
1160 | to_drain = pcp->count; | |
1161 | free_pcppages_bulk(zone, to_drain, pcp); | |
1162 | pcp->count -= to_drain; | |
1163 | local_irq_restore(flags); | |
1164 | } | |
1165 | #endif | |
1166 | ||
1167 | /* | |
1168 | * Drain pages of the indicated processor. | |
1169 | * | |
1170 | * The processor must either be the current processor and the | |
1171 | * thread pinned to the current processor or a processor that | |
1172 | * is not online. | |
1173 | */ | |
1174 | static void drain_pages(unsigned int cpu) | |
1175 | { | |
1176 | unsigned long flags; | |
1177 | struct zone *zone; | |
1178 | ||
1179 | for_each_populated_zone(zone) { | |
1180 | struct per_cpu_pageset *pset; | |
1181 | struct per_cpu_pages *pcp; | |
1182 | ||
1183 | local_irq_save(flags); | |
1184 | pset = per_cpu_ptr(zone->pageset, cpu); | |
1185 | ||
1186 | pcp = &pset->pcp; | |
1187 | if (pcp->count) { | |
1188 | free_pcppages_bulk(zone, pcp->count, pcp); | |
1189 | pcp->count = 0; | |
1190 | } | |
1191 | local_irq_restore(flags); | |
1192 | } | |
1193 | } | |
1194 | ||
1195 | /* | |
1196 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
1197 | */ | |
1198 | void drain_local_pages(void *arg) | |
1199 | { | |
1200 | drain_pages(smp_processor_id()); | |
1201 | } | |
1202 | ||
1203 | /* | |
1204 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. | |
1205 | * | |
1206 | * Note that this code is protected against sending an IPI to an offline | |
1207 | * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: | |
1208 | * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but | |
1209 | * nothing keeps CPUs from showing up after we populated the cpumask and | |
1210 | * before the call to on_each_cpu_mask(). | |
1211 | */ | |
1212 | void drain_all_pages(void) | |
1213 | { | |
1214 | int cpu; | |
1215 | struct per_cpu_pageset *pcp; | |
1216 | struct zone *zone; | |
1217 | ||
1218 | /* | |
1219 | * Allocate in the BSS so we wont require allocation in | |
1220 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y | |
1221 | */ | |
1222 | static cpumask_t cpus_with_pcps; | |
1223 | ||
1224 | /* | |
1225 | * We don't care about racing with CPU hotplug event | |
1226 | * as offline notification will cause the notified | |
1227 | * cpu to drain that CPU pcps and on_each_cpu_mask | |
1228 | * disables preemption as part of its processing | |
1229 | */ | |
1230 | for_each_online_cpu(cpu) { | |
1231 | bool has_pcps = false; | |
1232 | for_each_populated_zone(zone) { | |
1233 | pcp = per_cpu_ptr(zone->pageset, cpu); | |
1234 | if (pcp->pcp.count) { | |
1235 | has_pcps = true; | |
1236 | break; | |
1237 | } | |
1238 | } | |
1239 | if (has_pcps) | |
1240 | cpumask_set_cpu(cpu, &cpus_with_pcps); | |
1241 | else | |
1242 | cpumask_clear_cpu(cpu, &cpus_with_pcps); | |
1243 | } | |
1244 | on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1); | |
1245 | } | |
1246 | ||
1247 | #ifdef CONFIG_HIBERNATION | |
1248 | ||
1249 | void mark_free_pages(struct zone *zone) | |
1250 | { | |
1251 | unsigned long pfn, max_zone_pfn; | |
1252 | unsigned long flags; | |
1253 | int order, t; | |
1254 | struct list_head *curr; | |
1255 | ||
1256 | if (!zone->spanned_pages) | |
1257 | return; | |
1258 | ||
1259 | spin_lock_irqsave(&zone->lock, flags); | |
1260 | ||
1261 | max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; | |
1262 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) | |
1263 | if (pfn_valid(pfn)) { | |
1264 | struct page *page = pfn_to_page(pfn); | |
1265 | ||
1266 | if (!swsusp_page_is_forbidden(page)) | |
1267 | swsusp_unset_page_free(page); | |
1268 | } | |
1269 | ||
1270 | for_each_migratetype_order(order, t) { | |
1271 | list_for_each(curr, &zone->free_area[order].free_list[t]) { | |
1272 | unsigned long i; | |
1273 | ||
1274 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); | |
1275 | for (i = 0; i < (1UL << order); i++) | |
1276 | swsusp_set_page_free(pfn_to_page(pfn + i)); | |
1277 | } | |
1278 | } | |
1279 | spin_unlock_irqrestore(&zone->lock, flags); | |
1280 | } | |
1281 | #endif /* CONFIG_PM */ | |
1282 | ||
1283 | /* | |
1284 | * Free a 0-order page | |
1285 | * cold == 1 ? free a cold page : free a hot page | |
1286 | */ | |
1287 | void free_hot_cold_page(struct page *page, int cold) | |
1288 | { | |
1289 | struct zone *zone = page_zone(page); | |
1290 | struct per_cpu_pages *pcp; | |
1291 | unsigned long flags; | |
1292 | int migratetype; | |
1293 | int wasMlocked = __TestClearPageMlocked(page); | |
1294 | ||
1295 | if (!free_pages_prepare(page, 0)) | |
1296 | return; | |
1297 | ||
1298 | migratetype = get_pageblock_migratetype(page); | |
1299 | set_page_private(page, migratetype); | |
1300 | local_irq_save(flags); | |
1301 | if (unlikely(wasMlocked)) | |
1302 | free_page_mlock(page); | |
1303 | __count_vm_event(PGFREE); | |
1304 | ||
1305 | /* | |
1306 | * We only track unmovable, reclaimable and movable on pcp lists. | |
1307 | * Free ISOLATE pages back to the allocator because they are being | |
1308 | * offlined but treat RESERVE as movable pages so we can get those | |
1309 | * areas back if necessary. Otherwise, we may have to free | |
1310 | * excessively into the page allocator | |
1311 | */ | |
1312 | if (migratetype >= MIGRATE_PCPTYPES) { | |
1313 | if (unlikely(migratetype == MIGRATE_ISOLATE)) { | |
1314 | free_one_page(zone, page, 0, migratetype); | |
1315 | goto out; | |
1316 | } | |
1317 | migratetype = MIGRATE_MOVABLE; | |
1318 | } | |
1319 | ||
1320 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1321 | if (cold) | |
1322 | list_add_tail(&page->lru, &pcp->lists[migratetype]); | |
1323 | else | |
1324 | list_add(&page->lru, &pcp->lists[migratetype]); | |
1325 | pcp->count++; | |
1326 | if (pcp->count >= pcp->high) { | |
1327 | free_pcppages_bulk(zone, pcp->batch, pcp); | |
1328 | pcp->count -= pcp->batch; | |
1329 | } | |
1330 | ||
1331 | out: | |
1332 | local_irq_restore(flags); | |
1333 | } | |
1334 | ||
1335 | /* | |
1336 | * Free a list of 0-order pages | |
1337 | */ | |
1338 | void free_hot_cold_page_list(struct list_head *list, int cold) | |
1339 | { | |
1340 | struct page *page, *next; | |
1341 | ||
1342 | list_for_each_entry_safe(page, next, list, lru) { | |
1343 | trace_mm_page_free_batched(page, cold); | |
1344 | free_hot_cold_page(page, cold); | |
1345 | } | |
1346 | } | |
1347 | ||
1348 | /* | |
1349 | * split_page takes a non-compound higher-order page, and splits it into | |
1350 | * n (1<<order) sub-pages: page[0..n] | |
1351 | * Each sub-page must be freed individually. | |
1352 | * | |
1353 | * Note: this is probably too low level an operation for use in drivers. | |
1354 | * Please consult with lkml before using this in your driver. | |
1355 | */ | |
1356 | void split_page(struct page *page, unsigned int order) | |
1357 | { | |
1358 | int i; | |
1359 | ||
1360 | VM_BUG_ON(PageCompound(page)); | |
1361 | VM_BUG_ON(!page_count(page)); | |
1362 | ||
1363 | #ifdef CONFIG_KMEMCHECK | |
1364 | /* | |
1365 | * Split shadow pages too, because free(page[0]) would | |
1366 | * otherwise free the whole shadow. | |
1367 | */ | |
1368 | if (kmemcheck_page_is_tracked(page)) | |
1369 | split_page(virt_to_page(page[0].shadow), order); | |
1370 | #endif | |
1371 | ||
1372 | for (i = 1; i < (1 << order); i++) | |
1373 | set_page_refcounted(page + i); | |
1374 | } | |
1375 | ||
1376 | /* | |
1377 | * Similar to split_page except the page is already free. As this is only | |
1378 | * being used for migration, the migratetype of the block also changes. | |
1379 | * As this is called with interrupts disabled, the caller is responsible | |
1380 | * for calling arch_alloc_page() and kernel_map_page() after interrupts | |
1381 | * are enabled. | |
1382 | * | |
1383 | * Note: this is probably too low level an operation for use in drivers. | |
1384 | * Please consult with lkml before using this in your driver. | |
1385 | */ | |
1386 | int split_free_page(struct page *page) | |
1387 | { | |
1388 | unsigned int order; | |
1389 | unsigned long watermark; | |
1390 | struct zone *zone; | |
1391 | ||
1392 | BUG_ON(!PageBuddy(page)); | |
1393 | ||
1394 | zone = page_zone(page); | |
1395 | order = page_order(page); | |
1396 | ||
1397 | /* Obey watermarks as if the page was being allocated */ | |
1398 | watermark = low_wmark_pages(zone) + (1 << order); | |
1399 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
1400 | return 0; | |
1401 | ||
1402 | /* Remove page from free list */ | |
1403 | list_del(&page->lru); | |
1404 | zone->free_area[order].nr_free--; | |
1405 | rmv_page_order(page); | |
1406 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); | |
1407 | ||
1408 | /* Split into individual pages */ | |
1409 | set_page_refcounted(page); | |
1410 | split_page(page, order); | |
1411 | ||
1412 | if (order >= pageblock_order - 1) { | |
1413 | struct page *endpage = page + (1 << order) - 1; | |
1414 | for (; page < endpage; page += pageblock_nr_pages) { | |
1415 | int mt = get_pageblock_migratetype(page); | |
1416 | if (mt != MIGRATE_ISOLATE && !is_migrate_cma(mt)) | |
1417 | set_pageblock_migratetype(page, | |
1418 | MIGRATE_MOVABLE); | |
1419 | } | |
1420 | } | |
1421 | ||
1422 | return 1 << order; | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But | |
1427 | * we cheat by calling it from here, in the order > 0 path. Saves a branch | |
1428 | * or two. | |
1429 | */ | |
1430 | static inline | |
1431 | struct page *buffered_rmqueue(struct zone *preferred_zone, | |
1432 | struct zone *zone, int order, gfp_t gfp_flags, | |
1433 | int migratetype) | |
1434 | { | |
1435 | unsigned long flags; | |
1436 | struct page *page; | |
1437 | int cold = !!(gfp_flags & __GFP_COLD); | |
1438 | ||
1439 | again: | |
1440 | if (likely(order == 0)) { | |
1441 | struct per_cpu_pages *pcp; | |
1442 | struct list_head *list; | |
1443 | ||
1444 | local_irq_save(flags); | |
1445 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1446 | list = &pcp->lists[migratetype]; | |
1447 | if (list_empty(list)) { | |
1448 | pcp->count += rmqueue_bulk(zone, 0, | |
1449 | pcp->batch, list, | |
1450 | migratetype, cold); | |
1451 | if (unlikely(list_empty(list))) | |
1452 | goto failed; | |
1453 | } | |
1454 | ||
1455 | if (cold) | |
1456 | page = list_entry(list->prev, struct page, lru); | |
1457 | else | |
1458 | page = list_entry(list->next, struct page, lru); | |
1459 | ||
1460 | list_del(&page->lru); | |
1461 | pcp->count--; | |
1462 | } else { | |
1463 | if (unlikely(gfp_flags & __GFP_NOFAIL)) { | |
1464 | /* | |
1465 | * __GFP_NOFAIL is not to be used in new code. | |
1466 | * | |
1467 | * All __GFP_NOFAIL callers should be fixed so that they | |
1468 | * properly detect and handle allocation failures. | |
1469 | * | |
1470 | * We most definitely don't want callers attempting to | |
1471 | * allocate greater than order-1 page units with | |
1472 | * __GFP_NOFAIL. | |
1473 | */ | |
1474 | WARN_ON_ONCE(order > 1); | |
1475 | } | |
1476 | spin_lock_irqsave(&zone->lock, flags); | |
1477 | page = __rmqueue(zone, order, migratetype); | |
1478 | spin_unlock(&zone->lock); | |
1479 | if (!page) | |
1480 | goto failed; | |
1481 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); | |
1482 | } | |
1483 | ||
1484 | __count_zone_vm_events(PGALLOC, zone, 1 << order); | |
1485 | zone_statistics(preferred_zone, zone, gfp_flags); | |
1486 | local_irq_restore(flags); | |
1487 | ||
1488 | VM_BUG_ON(bad_range(zone, page)); | |
1489 | if (prep_new_page(page, order, gfp_flags)) | |
1490 | goto again; | |
1491 | return page; | |
1492 | ||
1493 | failed: | |
1494 | local_irq_restore(flags); | |
1495 | return NULL; | |
1496 | } | |
1497 | ||
1498 | /* The ALLOC_WMARK bits are used as an index to zone->watermark */ | |
1499 | #define ALLOC_WMARK_MIN WMARK_MIN | |
1500 | #define ALLOC_WMARK_LOW WMARK_LOW | |
1501 | #define ALLOC_WMARK_HIGH WMARK_HIGH | |
1502 | #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ | |
1503 | ||
1504 | /* Mask to get the watermark bits */ | |
1505 | #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) | |
1506 | ||
1507 | #define ALLOC_HARDER 0x10 /* try to alloc harder */ | |
1508 | #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ | |
1509 | #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ | |
1510 | ||
1511 | #ifdef CONFIG_FAIL_PAGE_ALLOC | |
1512 | ||
1513 | static struct { | |
1514 | struct fault_attr attr; | |
1515 | ||
1516 | u32 ignore_gfp_highmem; | |
1517 | u32 ignore_gfp_wait; | |
1518 | u32 min_order; | |
1519 | } fail_page_alloc = { | |
1520 | .attr = FAULT_ATTR_INITIALIZER, | |
1521 | .ignore_gfp_wait = 1, | |
1522 | .ignore_gfp_highmem = 1, | |
1523 | .min_order = 1, | |
1524 | }; | |
1525 | ||
1526 | static int __init setup_fail_page_alloc(char *str) | |
1527 | { | |
1528 | return setup_fault_attr(&fail_page_alloc.attr, str); | |
1529 | } | |
1530 | __setup("fail_page_alloc=", setup_fail_page_alloc); | |
1531 | ||
1532 | static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1533 | { | |
1534 | if (order < fail_page_alloc.min_order) | |
1535 | return 0; | |
1536 | if (gfp_mask & __GFP_NOFAIL) | |
1537 | return 0; | |
1538 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) | |
1539 | return 0; | |
1540 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) | |
1541 | return 0; | |
1542 | ||
1543 | return should_fail(&fail_page_alloc.attr, 1 << order); | |
1544 | } | |
1545 | ||
1546 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
1547 | ||
1548 | static int __init fail_page_alloc_debugfs(void) | |
1549 | { | |
1550 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
1551 | struct dentry *dir; | |
1552 | ||
1553 | dir = fault_create_debugfs_attr("fail_page_alloc", NULL, | |
1554 | &fail_page_alloc.attr); | |
1555 | if (IS_ERR(dir)) | |
1556 | return PTR_ERR(dir); | |
1557 | ||
1558 | if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
1559 | &fail_page_alloc.ignore_gfp_wait)) | |
1560 | goto fail; | |
1561 | if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, | |
1562 | &fail_page_alloc.ignore_gfp_highmem)) | |
1563 | goto fail; | |
1564 | if (!debugfs_create_u32("min-order", mode, dir, | |
1565 | &fail_page_alloc.min_order)) | |
1566 | goto fail; | |
1567 | ||
1568 | return 0; | |
1569 | fail: | |
1570 | debugfs_remove_recursive(dir); | |
1571 | ||
1572 | return -ENOMEM; | |
1573 | } | |
1574 | ||
1575 | late_initcall(fail_page_alloc_debugfs); | |
1576 | ||
1577 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
1578 | ||
1579 | #else /* CONFIG_FAIL_PAGE_ALLOC */ | |
1580 | ||
1581 | static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1582 | { | |
1583 | return 0; | |
1584 | } | |
1585 | ||
1586 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ | |
1587 | ||
1588 | /* | |
1589 | * Return true if free pages are above 'mark'. This takes into account the order | |
1590 | * of the allocation. | |
1591 | */ | |
1592 | static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, | |
1593 | int classzone_idx, int alloc_flags, long free_pages) | |
1594 | { | |
1595 | /* free_pages my go negative - that's OK */ | |
1596 | long min = mark; | |
1597 | int o; | |
1598 | ||
1599 | free_pages -= (1 << order) - 1; | |
1600 | if (alloc_flags & ALLOC_HIGH) | |
1601 | min -= min / 2; | |
1602 | if (alloc_flags & ALLOC_HARDER) | |
1603 | min -= min / 4; | |
1604 | ||
1605 | if (free_pages <= min + z->lowmem_reserve[classzone_idx]) | |
1606 | return false; | |
1607 | for (o = 0; o < order; o++) { | |
1608 | /* At the next order, this order's pages become unavailable */ | |
1609 | free_pages -= z->free_area[o].nr_free << o; | |
1610 | ||
1611 | /* Require fewer higher order pages to be free */ | |
1612 | min >>= 1; | |
1613 | ||
1614 | if (free_pages <= min) | |
1615 | return false; | |
1616 | } | |
1617 | return true; | |
1618 | } | |
1619 | ||
1620 | bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, | |
1621 | int classzone_idx, int alloc_flags) | |
1622 | { | |
1623 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1624 | zone_page_state(z, NR_FREE_PAGES)); | |
1625 | } | |
1626 | ||
1627 | bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, | |
1628 | int classzone_idx, int alloc_flags) | |
1629 | { | |
1630 | long free_pages = zone_page_state(z, NR_FREE_PAGES); | |
1631 | ||
1632 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) | |
1633 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); | |
1634 | ||
1635 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1636 | free_pages); | |
1637 | } | |
1638 | ||
1639 | #ifdef CONFIG_NUMA | |
1640 | /* | |
1641 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to | |
1642 | * skip over zones that are not allowed by the cpuset, or that have | |
1643 | * been recently (in last second) found to be nearly full. See further | |
1644 | * comments in mmzone.h. Reduces cache footprint of zonelist scans | |
1645 | * that have to skip over a lot of full or unallowed zones. | |
1646 | * | |
1647 | * If the zonelist cache is present in the passed in zonelist, then | |
1648 | * returns a pointer to the allowed node mask (either the current | |
1649 | * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) | |
1650 | * | |
1651 | * If the zonelist cache is not available for this zonelist, does | |
1652 | * nothing and returns NULL. | |
1653 | * | |
1654 | * If the fullzones BITMAP in the zonelist cache is stale (more than | |
1655 | * a second since last zap'd) then we zap it out (clear its bits.) | |
1656 | * | |
1657 | * We hold off even calling zlc_setup, until after we've checked the | |
1658 | * first zone in the zonelist, on the theory that most allocations will | |
1659 | * be satisfied from that first zone, so best to examine that zone as | |
1660 | * quickly as we can. | |
1661 | */ | |
1662 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1663 | { | |
1664 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1665 | nodemask_t *allowednodes; /* zonelist_cache approximation */ | |
1666 | ||
1667 | zlc = zonelist->zlcache_ptr; | |
1668 | if (!zlc) | |
1669 | return NULL; | |
1670 | ||
1671 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { | |
1672 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1673 | zlc->last_full_zap = jiffies; | |
1674 | } | |
1675 | ||
1676 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? | |
1677 | &cpuset_current_mems_allowed : | |
1678 | &node_states[N_HIGH_MEMORY]; | |
1679 | return allowednodes; | |
1680 | } | |
1681 | ||
1682 | /* | |
1683 | * Given 'z' scanning a zonelist, run a couple of quick checks to see | |
1684 | * if it is worth looking at further for free memory: | |
1685 | * 1) Check that the zone isn't thought to be full (doesn't have its | |
1686 | * bit set in the zonelist_cache fullzones BITMAP). | |
1687 | * 2) Check that the zones node (obtained from the zonelist_cache | |
1688 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. | |
1689 | * Return true (non-zero) if zone is worth looking at further, or | |
1690 | * else return false (zero) if it is not. | |
1691 | * | |
1692 | * This check -ignores- the distinction between various watermarks, | |
1693 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is | |
1694 | * found to be full for any variation of these watermarks, it will | |
1695 | * be considered full for up to one second by all requests, unless | |
1696 | * we are so low on memory on all allowed nodes that we are forced | |
1697 | * into the second scan of the zonelist. | |
1698 | * | |
1699 | * In the second scan we ignore this zonelist cache and exactly | |
1700 | * apply the watermarks to all zones, even it is slower to do so. | |
1701 | * We are low on memory in the second scan, and should leave no stone | |
1702 | * unturned looking for a free page. | |
1703 | */ | |
1704 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1705 | nodemask_t *allowednodes) | |
1706 | { | |
1707 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1708 | int i; /* index of *z in zonelist zones */ | |
1709 | int n; /* node that zone *z is on */ | |
1710 | ||
1711 | zlc = zonelist->zlcache_ptr; | |
1712 | if (!zlc) | |
1713 | return 1; | |
1714 | ||
1715 | i = z - zonelist->_zonerefs; | |
1716 | n = zlc->z_to_n[i]; | |
1717 | ||
1718 | /* This zone is worth trying if it is allowed but not full */ | |
1719 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); | |
1720 | } | |
1721 | ||
1722 | /* | |
1723 | * Given 'z' scanning a zonelist, set the corresponding bit in | |
1724 | * zlc->fullzones, so that subsequent attempts to allocate a page | |
1725 | * from that zone don't waste time re-examining it. | |
1726 | */ | |
1727 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1728 | { | |
1729 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1730 | int i; /* index of *z in zonelist zones */ | |
1731 | ||
1732 | zlc = zonelist->zlcache_ptr; | |
1733 | if (!zlc) | |
1734 | return; | |
1735 | ||
1736 | i = z - zonelist->_zonerefs; | |
1737 | ||
1738 | set_bit(i, zlc->fullzones); | |
1739 | } | |
1740 | ||
1741 | /* | |
1742 | * clear all zones full, called after direct reclaim makes progress so that | |
1743 | * a zone that was recently full is not skipped over for up to a second | |
1744 | */ | |
1745 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
1746 | { | |
1747 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1748 | ||
1749 | zlc = zonelist->zlcache_ptr; | |
1750 | if (!zlc) | |
1751 | return; | |
1752 | ||
1753 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1754 | } | |
1755 | ||
1756 | #else /* CONFIG_NUMA */ | |
1757 | ||
1758 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1759 | { | |
1760 | return NULL; | |
1761 | } | |
1762 | ||
1763 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1764 | nodemask_t *allowednodes) | |
1765 | { | |
1766 | return 1; | |
1767 | } | |
1768 | ||
1769 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1770 | { | |
1771 | } | |
1772 | ||
1773 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
1774 | { | |
1775 | } | |
1776 | #endif /* CONFIG_NUMA */ | |
1777 | ||
1778 | /* | |
1779 | * get_page_from_freelist goes through the zonelist trying to allocate | |
1780 | * a page. | |
1781 | */ | |
1782 | static struct page * | |
1783 | get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, | |
1784 | struct zonelist *zonelist, int high_zoneidx, int alloc_flags, | |
1785 | struct zone *preferred_zone, int migratetype) | |
1786 | { | |
1787 | struct zoneref *z; | |
1788 | struct page *page = NULL; | |
1789 | int classzone_idx; | |
1790 | struct zone *zone; | |
1791 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ | |
1792 | int zlc_active = 0; /* set if using zonelist_cache */ | |
1793 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ | |
1794 | ||
1795 | classzone_idx = zone_idx(preferred_zone); | |
1796 | zonelist_scan: | |
1797 | /* | |
1798 | * Scan zonelist, looking for a zone with enough free. | |
1799 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
1800 | */ | |
1801 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
1802 | high_zoneidx, nodemask) { | |
1803 | if (NUMA_BUILD && zlc_active && | |
1804 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
1805 | continue; | |
1806 | if ((alloc_flags & ALLOC_CPUSET) && | |
1807 | !cpuset_zone_allowed_softwall(zone, gfp_mask)) | |
1808 | continue; | |
1809 | /* | |
1810 | * When allocating a page cache page for writing, we | |
1811 | * want to get it from a zone that is within its dirty | |
1812 | * limit, such that no single zone holds more than its | |
1813 | * proportional share of globally allowed dirty pages. | |
1814 | * The dirty limits take into account the zone's | |
1815 | * lowmem reserves and high watermark so that kswapd | |
1816 | * should be able to balance it without having to | |
1817 | * write pages from its LRU list. | |
1818 | * | |
1819 | * This may look like it could increase pressure on | |
1820 | * lower zones by failing allocations in higher zones | |
1821 | * before they are full. But the pages that do spill | |
1822 | * over are limited as the lower zones are protected | |
1823 | * by this very same mechanism. It should not become | |
1824 | * a practical burden to them. | |
1825 | * | |
1826 | * XXX: For now, allow allocations to potentially | |
1827 | * exceed the per-zone dirty limit in the slowpath | |
1828 | * (ALLOC_WMARK_LOW unset) before going into reclaim, | |
1829 | * which is important when on a NUMA setup the allowed | |
1830 | * zones are together not big enough to reach the | |
1831 | * global limit. The proper fix for these situations | |
1832 | * will require awareness of zones in the | |
1833 | * dirty-throttling and the flusher threads. | |
1834 | */ | |
1835 | if ((alloc_flags & ALLOC_WMARK_LOW) && | |
1836 | (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone)) | |
1837 | goto this_zone_full; | |
1838 | ||
1839 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
1840 | if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { | |
1841 | unsigned long mark; | |
1842 | int ret; | |
1843 | ||
1844 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; | |
1845 | if (zone_watermark_ok(zone, order, mark, | |
1846 | classzone_idx, alloc_flags)) | |
1847 | goto try_this_zone; | |
1848 | ||
1849 | if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { | |
1850 | /* | |
1851 | * we do zlc_setup if there are multiple nodes | |
1852 | * and before considering the first zone allowed | |
1853 | * by the cpuset. | |
1854 | */ | |
1855 | allowednodes = zlc_setup(zonelist, alloc_flags); | |
1856 | zlc_active = 1; | |
1857 | did_zlc_setup = 1; | |
1858 | } | |
1859 | ||
1860 | if (zone_reclaim_mode == 0) | |
1861 | goto this_zone_full; | |
1862 | ||
1863 | /* | |
1864 | * As we may have just activated ZLC, check if the first | |
1865 | * eligible zone has failed zone_reclaim recently. | |
1866 | */ | |
1867 | if (NUMA_BUILD && zlc_active && | |
1868 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
1869 | continue; | |
1870 | ||
1871 | ret = zone_reclaim(zone, gfp_mask, order); | |
1872 | switch (ret) { | |
1873 | case ZONE_RECLAIM_NOSCAN: | |
1874 | /* did not scan */ | |
1875 | continue; | |
1876 | case ZONE_RECLAIM_FULL: | |
1877 | /* scanned but unreclaimable */ | |
1878 | continue; | |
1879 | default: | |
1880 | /* did we reclaim enough */ | |
1881 | if (!zone_watermark_ok(zone, order, mark, | |
1882 | classzone_idx, alloc_flags)) | |
1883 | goto this_zone_full; | |
1884 | } | |
1885 | } | |
1886 | ||
1887 | try_this_zone: | |
1888 | page = buffered_rmqueue(preferred_zone, zone, order, | |
1889 | gfp_mask, migratetype); | |
1890 | if (page) | |
1891 | break; | |
1892 | this_zone_full: | |
1893 | if (NUMA_BUILD) | |
1894 | zlc_mark_zone_full(zonelist, z); | |
1895 | } | |
1896 | ||
1897 | if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { | |
1898 | /* Disable zlc cache for second zonelist scan */ | |
1899 | zlc_active = 0; | |
1900 | goto zonelist_scan; | |
1901 | } | |
1902 | return page; | |
1903 | } | |
1904 | ||
1905 | /* | |
1906 | * Large machines with many possible nodes should not always dump per-node | |
1907 | * meminfo in irq context. | |
1908 | */ | |
1909 | static inline bool should_suppress_show_mem(void) | |
1910 | { | |
1911 | bool ret = false; | |
1912 | ||
1913 | #if NODES_SHIFT > 8 | |
1914 | ret = in_interrupt(); | |
1915 | #endif | |
1916 | return ret; | |
1917 | } | |
1918 | ||
1919 | static DEFINE_RATELIMIT_STATE(nopage_rs, | |
1920 | DEFAULT_RATELIMIT_INTERVAL, | |
1921 | DEFAULT_RATELIMIT_BURST); | |
1922 | ||
1923 | void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) | |
1924 | { | |
1925 | unsigned int filter = SHOW_MEM_FILTER_NODES; | |
1926 | ||
1927 | if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || | |
1928 | debug_guardpage_minorder() > 0) | |
1929 | return; | |
1930 | ||
1931 | /* | |
1932 | * This documents exceptions given to allocations in certain | |
1933 | * contexts that are allowed to allocate outside current's set | |
1934 | * of allowed nodes. | |
1935 | */ | |
1936 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
1937 | if (test_thread_flag(TIF_MEMDIE) || | |
1938 | (current->flags & (PF_MEMALLOC | PF_EXITING))) | |
1939 | filter &= ~SHOW_MEM_FILTER_NODES; | |
1940 | if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) | |
1941 | filter &= ~SHOW_MEM_FILTER_NODES; | |
1942 | ||
1943 | if (fmt) { | |
1944 | struct va_format vaf; | |
1945 | va_list args; | |
1946 | ||
1947 | va_start(args, fmt); | |
1948 | ||
1949 | vaf.fmt = fmt; | |
1950 | vaf.va = &args; | |
1951 | ||
1952 | pr_warn("%pV", &vaf); | |
1953 | ||
1954 | va_end(args); | |
1955 | } | |
1956 | ||
1957 | pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", | |
1958 | current->comm, order, gfp_mask); | |
1959 | ||
1960 | dump_stack(); | |
1961 | if (!should_suppress_show_mem()) | |
1962 | show_mem(filter); | |
1963 | } | |
1964 | ||
1965 | static inline int | |
1966 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, | |
1967 | unsigned long did_some_progress, | |
1968 | unsigned long pages_reclaimed) | |
1969 | { | |
1970 | /* Do not loop if specifically requested */ | |
1971 | if (gfp_mask & __GFP_NORETRY) | |
1972 | return 0; | |
1973 | ||
1974 | /* Always retry if specifically requested */ | |
1975 | if (gfp_mask & __GFP_NOFAIL) | |
1976 | return 1; | |
1977 | ||
1978 | /* | |
1979 | * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim | |
1980 | * making forward progress without invoking OOM. Suspend also disables | |
1981 | * storage devices so kswapd will not help. Bail if we are suspending. | |
1982 | */ | |
1983 | if (!did_some_progress && pm_suspended_storage()) | |
1984 | return 0; | |
1985 | ||
1986 | /* | |
1987 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER | |
1988 | * means __GFP_NOFAIL, but that may not be true in other | |
1989 | * implementations. | |
1990 | */ | |
1991 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
1992 | return 1; | |
1993 | ||
1994 | /* | |
1995 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is | |
1996 | * specified, then we retry until we no longer reclaim any pages | |
1997 | * (above), or we've reclaimed an order of pages at least as | |
1998 | * large as the allocation's order. In both cases, if the | |
1999 | * allocation still fails, we stop retrying. | |
2000 | */ | |
2001 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) | |
2002 | return 1; | |
2003 | ||
2004 | return 0; | |
2005 | } | |
2006 | ||
2007 | static inline struct page * | |
2008 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
2009 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2010 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2011 | int migratetype) | |
2012 | { | |
2013 | struct page *page; | |
2014 | ||
2015 | /* Acquire the OOM killer lock for the zones in zonelist */ | |
2016 | if (!try_set_zonelist_oom(zonelist, gfp_mask)) { | |
2017 | schedule_timeout_uninterruptible(1); | |
2018 | return NULL; | |
2019 | } | |
2020 | ||
2021 | /* | |
2022 | * Go through the zonelist yet one more time, keep very high watermark | |
2023 | * here, this is only to catch a parallel oom killing, we must fail if | |
2024 | * we're still under heavy pressure. | |
2025 | */ | |
2026 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, | |
2027 | order, zonelist, high_zoneidx, | |
2028 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, | |
2029 | preferred_zone, migratetype); | |
2030 | if (page) | |
2031 | goto out; | |
2032 | ||
2033 | if (!(gfp_mask & __GFP_NOFAIL)) { | |
2034 | /* The OOM killer will not help higher order allocs */ | |
2035 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
2036 | goto out; | |
2037 | /* The OOM killer does not needlessly kill tasks for lowmem */ | |
2038 | if (high_zoneidx < ZONE_NORMAL) | |
2039 | goto out; | |
2040 | /* | |
2041 | * GFP_THISNODE contains __GFP_NORETRY and we never hit this. | |
2042 | * Sanity check for bare calls of __GFP_THISNODE, not real OOM. | |
2043 | * The caller should handle page allocation failure by itself if | |
2044 | * it specifies __GFP_THISNODE. | |
2045 | * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. | |
2046 | */ | |
2047 | if (gfp_mask & __GFP_THISNODE) | |
2048 | goto out; | |
2049 | } | |
2050 | /* Exhausted what can be done so it's blamo time */ | |
2051 | out_of_memory(zonelist, gfp_mask, order, nodemask, false); | |
2052 | ||
2053 | out: | |
2054 | clear_zonelist_oom(zonelist, gfp_mask); | |
2055 | return page; | |
2056 | } | |
2057 | ||
2058 | #ifdef CONFIG_COMPACTION | |
2059 | /* Try memory compaction for high-order allocations before reclaim */ | |
2060 | static struct page * | |
2061 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2062 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2063 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2064 | int migratetype, bool sync_migration, | |
2065 | bool *deferred_compaction, | |
2066 | unsigned long *did_some_progress) | |
2067 | { | |
2068 | struct page *page; | |
2069 | ||
2070 | if (!order) | |
2071 | return NULL; | |
2072 | ||
2073 | if (compaction_deferred(preferred_zone, order)) { | |
2074 | *deferred_compaction = true; | |
2075 | return NULL; | |
2076 | } | |
2077 | ||
2078 | current->flags |= PF_MEMALLOC; | |
2079 | *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, | |
2080 | nodemask, sync_migration); | |
2081 | current->flags &= ~PF_MEMALLOC; | |
2082 | if (*did_some_progress != COMPACT_SKIPPED) { | |
2083 | ||
2084 | /* Page migration frees to the PCP lists but we want merging */ | |
2085 | drain_pages(get_cpu()); | |
2086 | put_cpu(); | |
2087 | ||
2088 | page = get_page_from_freelist(gfp_mask, nodemask, | |
2089 | order, zonelist, high_zoneidx, | |
2090 | alloc_flags, preferred_zone, | |
2091 | migratetype); | |
2092 | if (page) { | |
2093 | preferred_zone->compact_considered = 0; | |
2094 | preferred_zone->compact_defer_shift = 0; | |
2095 | if (order >= preferred_zone->compact_order_failed) | |
2096 | preferred_zone->compact_order_failed = order + 1; | |
2097 | count_vm_event(COMPACTSUCCESS); | |
2098 | return page; | |
2099 | } | |
2100 | ||
2101 | /* | |
2102 | * It's bad if compaction run occurs and fails. | |
2103 | * The most likely reason is that pages exist, | |
2104 | * but not enough to satisfy watermarks. | |
2105 | */ | |
2106 | count_vm_event(COMPACTFAIL); | |
2107 | ||
2108 | /* | |
2109 | * As async compaction considers a subset of pageblocks, only | |
2110 | * defer if the failure was a sync compaction failure. | |
2111 | */ | |
2112 | if (sync_migration) | |
2113 | defer_compaction(preferred_zone, order); | |
2114 | ||
2115 | cond_resched(); | |
2116 | } | |
2117 | ||
2118 | return NULL; | |
2119 | } | |
2120 | #else | |
2121 | static inline struct page * | |
2122 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2123 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2124 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2125 | int migratetype, bool sync_migration, | |
2126 | bool *deferred_compaction, | |
2127 | unsigned long *did_some_progress) | |
2128 | { | |
2129 | return NULL; | |
2130 | } | |
2131 | #endif /* CONFIG_COMPACTION */ | |
2132 | ||
2133 | /* The really slow allocator path where we enter direct reclaim */ | |
2134 | static inline struct page * | |
2135 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
2136 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2137 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2138 | int migratetype, unsigned long *did_some_progress) | |
2139 | { | |
2140 | struct page *page = NULL; | |
2141 | struct reclaim_state reclaim_state; | |
2142 | bool drained = false; | |
2143 | ||
2144 | cond_resched(); | |
2145 | ||
2146 | /* We now go into synchronous reclaim */ | |
2147 | cpuset_memory_pressure_bump(); | |
2148 | current->flags |= PF_MEMALLOC; | |
2149 | lockdep_set_current_reclaim_state(gfp_mask); | |
2150 | reclaim_state.reclaimed_slab = 0; | |
2151 | current->reclaim_state = &reclaim_state; | |
2152 | ||
2153 | *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); | |
2154 | ||
2155 | current->reclaim_state = NULL; | |
2156 | lockdep_clear_current_reclaim_state(); | |
2157 | current->flags &= ~PF_MEMALLOC; | |
2158 | ||
2159 | cond_resched(); | |
2160 | ||
2161 | if (unlikely(!(*did_some_progress))) | |
2162 | return NULL; | |
2163 | ||
2164 | /* After successful reclaim, reconsider all zones for allocation */ | |
2165 | if (NUMA_BUILD) | |
2166 | zlc_clear_zones_full(zonelist); | |
2167 | ||
2168 | retry: | |
2169 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
2170 | zonelist, high_zoneidx, | |
2171 | alloc_flags, preferred_zone, | |
2172 | migratetype); | |
2173 | ||
2174 | /* | |
2175 | * If an allocation failed after direct reclaim, it could be because | |
2176 | * pages are pinned on the per-cpu lists. Drain them and try again | |
2177 | */ | |
2178 | if (!page && !drained) { | |
2179 | drain_all_pages(); | |
2180 | drained = true; | |
2181 | goto retry; | |
2182 | } | |
2183 | ||
2184 | return page; | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * This is called in the allocator slow-path if the allocation request is of | |
2189 | * sufficient urgency to ignore watermarks and take other desperate measures | |
2190 | */ | |
2191 | static inline struct page * | |
2192 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, | |
2193 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2194 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2195 | int migratetype) | |
2196 | { | |
2197 | struct page *page; | |
2198 | ||
2199 | do { | |
2200 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
2201 | zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, | |
2202 | preferred_zone, migratetype); | |
2203 | ||
2204 | if (!page && gfp_mask & __GFP_NOFAIL) | |
2205 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); | |
2206 | } while (!page && (gfp_mask & __GFP_NOFAIL)); | |
2207 | ||
2208 | return page; | |
2209 | } | |
2210 | ||
2211 | static inline | |
2212 | void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, | |
2213 | enum zone_type high_zoneidx, | |
2214 | enum zone_type classzone_idx) | |
2215 | { | |
2216 | struct zoneref *z; | |
2217 | struct zone *zone; | |
2218 | ||
2219 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) | |
2220 | wakeup_kswapd(zone, order, classzone_idx); | |
2221 | } | |
2222 | ||
2223 | static inline int | |
2224 | gfp_to_alloc_flags(gfp_t gfp_mask) | |
2225 | { | |
2226 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; | |
2227 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
2228 | ||
2229 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ | |
2230 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); | |
2231 | ||
2232 | /* | |
2233 | * The caller may dip into page reserves a bit more if the caller | |
2234 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
2235 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
2236 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). | |
2237 | */ | |
2238 | alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); | |
2239 | ||
2240 | if (!wait) { | |
2241 | /* | |
2242 | * Not worth trying to allocate harder for | |
2243 | * __GFP_NOMEMALLOC even if it can't schedule. | |
2244 | */ | |
2245 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
2246 | alloc_flags |= ALLOC_HARDER; | |
2247 | /* | |
2248 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. | |
2249 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
2250 | */ | |
2251 | alloc_flags &= ~ALLOC_CPUSET; | |
2252 | } else if (unlikely(rt_task(current)) && !in_interrupt()) | |
2253 | alloc_flags |= ALLOC_HARDER; | |
2254 | ||
2255 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { | |
2256 | if (!in_interrupt() && | |
2257 | ((current->flags & PF_MEMALLOC) || | |
2258 | unlikely(test_thread_flag(TIF_MEMDIE)))) | |
2259 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2260 | } | |
2261 | ||
2262 | return alloc_flags; | |
2263 | } | |
2264 | ||
2265 | static inline struct page * | |
2266 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
2267 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2268 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2269 | int migratetype) | |
2270 | { | |
2271 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
2272 | struct page *page = NULL; | |
2273 | int alloc_flags; | |
2274 | unsigned long pages_reclaimed = 0; | |
2275 | unsigned long did_some_progress; | |
2276 | bool sync_migration = false; | |
2277 | bool deferred_compaction = false; | |
2278 | ||
2279 | /* | |
2280 | * In the slowpath, we sanity check order to avoid ever trying to | |
2281 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may | |
2282 | * be using allocators in order of preference for an area that is | |
2283 | * too large. | |
2284 | */ | |
2285 | if (order >= MAX_ORDER) { | |
2286 | WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); | |
2287 | return NULL; | |
2288 | } | |
2289 | ||
2290 | /* | |
2291 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and | |
2292 | * __GFP_NOWARN set) should not cause reclaim since the subsystem | |
2293 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim | |
2294 | * using a larger set of nodes after it has established that the | |
2295 | * allowed per node queues are empty and that nodes are | |
2296 | * over allocated. | |
2297 | */ | |
2298 | if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) | |
2299 | goto nopage; | |
2300 | ||
2301 | restart: | |
2302 | if (!(gfp_mask & __GFP_NO_KSWAPD)) | |
2303 | wake_all_kswapd(order, zonelist, high_zoneidx, | |
2304 | zone_idx(preferred_zone)); | |
2305 | ||
2306 | /* | |
2307 | * OK, we're below the kswapd watermark and have kicked background | |
2308 | * reclaim. Now things get more complex, so set up alloc_flags according | |
2309 | * to how we want to proceed. | |
2310 | */ | |
2311 | alloc_flags = gfp_to_alloc_flags(gfp_mask); | |
2312 | ||
2313 | /* | |
2314 | * Find the true preferred zone if the allocation is unconstrained by | |
2315 | * cpusets. | |
2316 | */ | |
2317 | if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) | |
2318 | first_zones_zonelist(zonelist, high_zoneidx, NULL, | |
2319 | &preferred_zone); | |
2320 | ||
2321 | rebalance: | |
2322 | /* This is the last chance, in general, before the goto nopage. */ | |
2323 | page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, | |
2324 | high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, | |
2325 | preferred_zone, migratetype); | |
2326 | if (page) | |
2327 | goto got_pg; | |
2328 | ||
2329 | /* Allocate without watermarks if the context allows */ | |
2330 | if (alloc_flags & ALLOC_NO_WATERMARKS) { | |
2331 | page = __alloc_pages_high_priority(gfp_mask, order, | |
2332 | zonelist, high_zoneidx, nodemask, | |
2333 | preferred_zone, migratetype); | |
2334 | if (page) | |
2335 | goto got_pg; | |
2336 | } | |
2337 | ||
2338 | /* Atomic allocations - we can't balance anything */ | |
2339 | if (!wait) | |
2340 | goto nopage; | |
2341 | ||
2342 | /* Avoid recursion of direct reclaim */ | |
2343 | if (current->flags & PF_MEMALLOC) | |
2344 | goto nopage; | |
2345 | ||
2346 | /* Avoid allocations with no watermarks from looping endlessly */ | |
2347 | if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) | |
2348 | goto nopage; | |
2349 | ||
2350 | /* | |
2351 | * Try direct compaction. The first pass is asynchronous. Subsequent | |
2352 | * attempts after direct reclaim are synchronous | |
2353 | */ | |
2354 | page = __alloc_pages_direct_compact(gfp_mask, order, | |
2355 | zonelist, high_zoneidx, | |
2356 | nodemask, | |
2357 | alloc_flags, preferred_zone, | |
2358 | migratetype, sync_migration, | |
2359 | &deferred_compaction, | |
2360 | &did_some_progress); | |
2361 | if (page) | |
2362 | goto got_pg; | |
2363 | sync_migration = true; | |
2364 | ||
2365 | /* | |
2366 | * If compaction is deferred for high-order allocations, it is because | |
2367 | * sync compaction recently failed. In this is the case and the caller | |
2368 | * has requested the system not be heavily disrupted, fail the | |
2369 | * allocation now instead of entering direct reclaim | |
2370 | */ | |
2371 | if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD)) | |
2372 | goto nopage; | |
2373 | ||
2374 | /* Try direct reclaim and then allocating */ | |
2375 | page = __alloc_pages_direct_reclaim(gfp_mask, order, | |
2376 | zonelist, high_zoneidx, | |
2377 | nodemask, | |
2378 | alloc_flags, preferred_zone, | |
2379 | migratetype, &did_some_progress); | |
2380 | if (page) | |
2381 | goto got_pg; | |
2382 | ||
2383 | /* | |
2384 | * If we failed to make any progress reclaiming, then we are | |
2385 | * running out of options and have to consider going OOM | |
2386 | */ | |
2387 | if (!did_some_progress) { | |
2388 | if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { | |
2389 | if (oom_killer_disabled) | |
2390 | goto nopage; | |
2391 | /* Coredumps can quickly deplete all memory reserves */ | |
2392 | if ((current->flags & PF_DUMPCORE) && | |
2393 | !(gfp_mask & __GFP_NOFAIL)) | |
2394 | goto nopage; | |
2395 | page = __alloc_pages_may_oom(gfp_mask, order, | |
2396 | zonelist, high_zoneidx, | |
2397 | nodemask, preferred_zone, | |
2398 | migratetype); | |
2399 | if (page) | |
2400 | goto got_pg; | |
2401 | ||
2402 | if (!(gfp_mask & __GFP_NOFAIL)) { | |
2403 | /* | |
2404 | * The oom killer is not called for high-order | |
2405 | * allocations that may fail, so if no progress | |
2406 | * is being made, there are no other options and | |
2407 | * retrying is unlikely to help. | |
2408 | */ | |
2409 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
2410 | goto nopage; | |
2411 | /* | |
2412 | * The oom killer is not called for lowmem | |
2413 | * allocations to prevent needlessly killing | |
2414 | * innocent tasks. | |
2415 | */ | |
2416 | if (high_zoneidx < ZONE_NORMAL) | |
2417 | goto nopage; | |
2418 | } | |
2419 | ||
2420 | goto restart; | |
2421 | } | |
2422 | } | |
2423 | ||
2424 | /* Check if we should retry the allocation */ | |
2425 | pages_reclaimed += did_some_progress; | |
2426 | if (should_alloc_retry(gfp_mask, order, did_some_progress, | |
2427 | pages_reclaimed)) { | |
2428 | /* Wait for some write requests to complete then retry */ | |
2429 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); | |
2430 | goto rebalance; | |
2431 | } else { | |
2432 | /* | |
2433 | * High-order allocations do not necessarily loop after | |
2434 | * direct reclaim and reclaim/compaction depends on compaction | |
2435 | * being called after reclaim so call directly if necessary | |
2436 | */ | |
2437 | page = __alloc_pages_direct_compact(gfp_mask, order, | |
2438 | zonelist, high_zoneidx, | |
2439 | nodemask, | |
2440 | alloc_flags, preferred_zone, | |
2441 | migratetype, sync_migration, | |
2442 | &deferred_compaction, | |
2443 | &did_some_progress); | |
2444 | if (page) | |
2445 | goto got_pg; | |
2446 | } | |
2447 | ||
2448 | nopage: | |
2449 | warn_alloc_failed(gfp_mask, order, NULL); | |
2450 | return page; | |
2451 | got_pg: | |
2452 | if (kmemcheck_enabled) | |
2453 | kmemcheck_pagealloc_alloc(page, order, gfp_mask); | |
2454 | return page; | |
2455 | ||
2456 | } | |
2457 | ||
2458 | /* | |
2459 | * This is the 'heart' of the zoned buddy allocator. | |
2460 | */ | |
2461 | struct page * | |
2462 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, | |
2463 | struct zonelist *zonelist, nodemask_t *nodemask) | |
2464 | { | |
2465 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
2466 | struct zone *preferred_zone; | |
2467 | struct page *page = NULL; | |
2468 | int migratetype = allocflags_to_migratetype(gfp_mask); | |
2469 | unsigned int cpuset_mems_cookie; | |
2470 | ||
2471 | gfp_mask &= gfp_allowed_mask; | |
2472 | ||
2473 | lockdep_trace_alloc(gfp_mask); | |
2474 | ||
2475 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2476 | ||
2477 | if (should_fail_alloc_page(gfp_mask, order)) | |
2478 | return NULL; | |
2479 | ||
2480 | /* | |
2481 | * Check the zones suitable for the gfp_mask contain at least one | |
2482 | * valid zone. It's possible to have an empty zonelist as a result | |
2483 | * of GFP_THISNODE and a memoryless node | |
2484 | */ | |
2485 | if (unlikely(!zonelist->_zonerefs->zone)) | |
2486 | return NULL; | |
2487 | ||
2488 | retry_cpuset: | |
2489 | cpuset_mems_cookie = get_mems_allowed(); | |
2490 | ||
2491 | /* The preferred zone is used for statistics later */ | |
2492 | first_zones_zonelist(zonelist, high_zoneidx, | |
2493 | nodemask ? : &cpuset_current_mems_allowed, | |
2494 | &preferred_zone); | |
2495 | if (!preferred_zone) | |
2496 | goto out; | |
2497 | ||
2498 | /* First allocation attempt */ | |
2499 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, | |
2500 | zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, | |
2501 | preferred_zone, migratetype); | |
2502 | if (unlikely(!page)) | |
2503 | page = __alloc_pages_slowpath(gfp_mask, order, | |
2504 | zonelist, high_zoneidx, nodemask, | |
2505 | preferred_zone, migratetype); | |
2506 | ||
2507 | trace_mm_page_alloc(page, order, gfp_mask, migratetype); | |
2508 | ||
2509 | out: | |
2510 | /* | |
2511 | * When updating a task's mems_allowed, it is possible to race with | |
2512 | * parallel threads in such a way that an allocation can fail while | |
2513 | * the mask is being updated. If a page allocation is about to fail, | |
2514 | * check if the cpuset changed during allocation and if so, retry. | |
2515 | */ | |
2516 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) | |
2517 | goto retry_cpuset; | |
2518 | ||
2519 | return page; | |
2520 | } | |
2521 | EXPORT_SYMBOL(__alloc_pages_nodemask); | |
2522 | ||
2523 | /* | |
2524 | * Common helper functions. | |
2525 | */ | |
2526 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) | |
2527 | { | |
2528 | struct page *page; | |
2529 | ||
2530 | /* | |
2531 | * __get_free_pages() returns a 32-bit address, which cannot represent | |
2532 | * a highmem page | |
2533 | */ | |
2534 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); | |
2535 | ||
2536 | page = alloc_pages(gfp_mask, order); | |
2537 | if (!page) | |
2538 | return 0; | |
2539 | return (unsigned long) page_address(page); | |
2540 | } | |
2541 | EXPORT_SYMBOL(__get_free_pages); | |
2542 | ||
2543 | unsigned long get_zeroed_page(gfp_t gfp_mask) | |
2544 | { | |
2545 | return __get_free_pages(gfp_mask | __GFP_ZERO, 0); | |
2546 | } | |
2547 | EXPORT_SYMBOL(get_zeroed_page); | |
2548 | ||
2549 | void __free_pages(struct page *page, unsigned int order) | |
2550 | { | |
2551 | if (put_page_testzero(page)) { | |
2552 | if (order == 0) | |
2553 | free_hot_cold_page(page, 0); | |
2554 | else | |
2555 | __free_pages_ok(page, order); | |
2556 | } | |
2557 | } | |
2558 | ||
2559 | EXPORT_SYMBOL(__free_pages); | |
2560 | ||
2561 | void free_pages(unsigned long addr, unsigned int order) | |
2562 | { | |
2563 | if (addr != 0) { | |
2564 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
2565 | __free_pages(virt_to_page((void *)addr), order); | |
2566 | } | |
2567 | } | |
2568 | ||
2569 | EXPORT_SYMBOL(free_pages); | |
2570 | ||
2571 | static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) | |
2572 | { | |
2573 | if (addr) { | |
2574 | unsigned long alloc_end = addr + (PAGE_SIZE << order); | |
2575 | unsigned long used = addr + PAGE_ALIGN(size); | |
2576 | ||
2577 | split_page(virt_to_page((void *)addr), order); | |
2578 | while (used < alloc_end) { | |
2579 | free_page(used); | |
2580 | used += PAGE_SIZE; | |
2581 | } | |
2582 | } | |
2583 | return (void *)addr; | |
2584 | } | |
2585 | ||
2586 | /** | |
2587 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
2588 | * @size: the number of bytes to allocate | |
2589 | * @gfp_mask: GFP flags for the allocation | |
2590 | * | |
2591 | * This function is similar to alloc_pages(), except that it allocates the | |
2592 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
2593 | * allocate memory in power-of-two pages. | |
2594 | * | |
2595 | * This function is also limited by MAX_ORDER. | |
2596 | * | |
2597 | * Memory allocated by this function must be released by free_pages_exact(). | |
2598 | */ | |
2599 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
2600 | { | |
2601 | unsigned int order = get_order(size); | |
2602 | unsigned long addr; | |
2603 | ||
2604 | addr = __get_free_pages(gfp_mask, order); | |
2605 | return make_alloc_exact(addr, order, size); | |
2606 | } | |
2607 | EXPORT_SYMBOL(alloc_pages_exact); | |
2608 | ||
2609 | /** | |
2610 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous | |
2611 | * pages on a node. | |
2612 | * @nid: the preferred node ID where memory should be allocated | |
2613 | * @size: the number of bytes to allocate | |
2614 | * @gfp_mask: GFP flags for the allocation | |
2615 | * | |
2616 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling | |
2617 | * back. | |
2618 | * Note this is not alloc_pages_exact_node() which allocates on a specific node, | |
2619 | * but is not exact. | |
2620 | */ | |
2621 | void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) | |
2622 | { | |
2623 | unsigned order = get_order(size); | |
2624 | struct page *p = alloc_pages_node(nid, gfp_mask, order); | |
2625 | if (!p) | |
2626 | return NULL; | |
2627 | return make_alloc_exact((unsigned long)page_address(p), order, size); | |
2628 | } | |
2629 | EXPORT_SYMBOL(alloc_pages_exact_nid); | |
2630 | ||
2631 | /** | |
2632 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
2633 | * @virt: the value returned by alloc_pages_exact. | |
2634 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
2635 | * | |
2636 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
2637 | */ | |
2638 | void free_pages_exact(void *virt, size_t size) | |
2639 | { | |
2640 | unsigned long addr = (unsigned long)virt; | |
2641 | unsigned long end = addr + PAGE_ALIGN(size); | |
2642 | ||
2643 | while (addr < end) { | |
2644 | free_page(addr); | |
2645 | addr += PAGE_SIZE; | |
2646 | } | |
2647 | } | |
2648 | EXPORT_SYMBOL(free_pages_exact); | |
2649 | ||
2650 | static unsigned int nr_free_zone_pages(int offset) | |
2651 | { | |
2652 | struct zoneref *z; | |
2653 | struct zone *zone; | |
2654 | ||
2655 | /* Just pick one node, since fallback list is circular */ | |
2656 | unsigned int sum = 0; | |
2657 | ||
2658 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); | |
2659 | ||
2660 | for_each_zone_zonelist(zone, z, zonelist, offset) { | |
2661 | unsigned long size = zone->present_pages; | |
2662 | unsigned long high = high_wmark_pages(zone); | |
2663 | if (size > high) | |
2664 | sum += size - high; | |
2665 | } | |
2666 | ||
2667 | return sum; | |
2668 | } | |
2669 | ||
2670 | /* | |
2671 | * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL | |
2672 | */ | |
2673 | unsigned int nr_free_buffer_pages(void) | |
2674 | { | |
2675 | return nr_free_zone_pages(gfp_zone(GFP_USER)); | |
2676 | } | |
2677 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); | |
2678 | ||
2679 | /* | |
2680 | * Amount of free RAM allocatable within all zones | |
2681 | */ | |
2682 | unsigned int nr_free_pagecache_pages(void) | |
2683 | { | |
2684 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
2685 | } | |
2686 | ||
2687 | static inline void show_node(struct zone *zone) | |
2688 | { | |
2689 | if (NUMA_BUILD) | |
2690 | printk("Node %d ", zone_to_nid(zone)); | |
2691 | } | |
2692 | ||
2693 | void si_meminfo(struct sysinfo *val) | |
2694 | { | |
2695 | val->totalram = totalram_pages; | |
2696 | val->sharedram = 0; | |
2697 | val->freeram = global_page_state(NR_FREE_PAGES); | |
2698 | val->bufferram = nr_blockdev_pages(); | |
2699 | val->totalhigh = totalhigh_pages; | |
2700 | val->freehigh = nr_free_highpages(); | |
2701 | val->mem_unit = PAGE_SIZE; | |
2702 | } | |
2703 | ||
2704 | EXPORT_SYMBOL(si_meminfo); | |
2705 | ||
2706 | #ifdef CONFIG_NUMA | |
2707 | void si_meminfo_node(struct sysinfo *val, int nid) | |
2708 | { | |
2709 | pg_data_t *pgdat = NODE_DATA(nid); | |
2710 | ||
2711 | val->totalram = pgdat->node_present_pages; | |
2712 | val->freeram = node_page_state(nid, NR_FREE_PAGES); | |
2713 | #ifdef CONFIG_HIGHMEM | |
2714 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; | |
2715 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], | |
2716 | NR_FREE_PAGES); | |
2717 | #else | |
2718 | val->totalhigh = 0; | |
2719 | val->freehigh = 0; | |
2720 | #endif | |
2721 | val->mem_unit = PAGE_SIZE; | |
2722 | } | |
2723 | #endif | |
2724 | ||
2725 | /* | |
2726 | * Determine whether the node should be displayed or not, depending on whether | |
2727 | * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). | |
2728 | */ | |
2729 | bool skip_free_areas_node(unsigned int flags, int nid) | |
2730 | { | |
2731 | bool ret = false; | |
2732 | unsigned int cpuset_mems_cookie; | |
2733 | ||
2734 | if (!(flags & SHOW_MEM_FILTER_NODES)) | |
2735 | goto out; | |
2736 | ||
2737 | do { | |
2738 | cpuset_mems_cookie = get_mems_allowed(); | |
2739 | ret = !node_isset(nid, cpuset_current_mems_allowed); | |
2740 | } while (!put_mems_allowed(cpuset_mems_cookie)); | |
2741 | out: | |
2742 | return ret; | |
2743 | } | |
2744 | ||
2745 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
2746 | ||
2747 | /* | |
2748 | * Show free area list (used inside shift_scroll-lock stuff) | |
2749 | * We also calculate the percentage fragmentation. We do this by counting the | |
2750 | * memory on each free list with the exception of the first item on the list. | |
2751 | * Suppresses nodes that are not allowed by current's cpuset if | |
2752 | * SHOW_MEM_FILTER_NODES is passed. | |
2753 | */ | |
2754 | void show_free_areas(unsigned int filter) | |
2755 | { | |
2756 | int cpu; | |
2757 | struct zone *zone; | |
2758 | ||
2759 | for_each_populated_zone(zone) { | |
2760 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
2761 | continue; | |
2762 | show_node(zone); | |
2763 | printk("%s per-cpu:\n", zone->name); | |
2764 | ||
2765 | for_each_online_cpu(cpu) { | |
2766 | struct per_cpu_pageset *pageset; | |
2767 | ||
2768 | pageset = per_cpu_ptr(zone->pageset, cpu); | |
2769 | ||
2770 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", | |
2771 | cpu, pageset->pcp.high, | |
2772 | pageset->pcp.batch, pageset->pcp.count); | |
2773 | } | |
2774 | } | |
2775 | ||
2776 | printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" | |
2777 | " active_file:%lu inactive_file:%lu isolated_file:%lu\n" | |
2778 | " unevictable:%lu" | |
2779 | " dirty:%lu writeback:%lu unstable:%lu\n" | |
2780 | " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" | |
2781 | " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", | |
2782 | global_page_state(NR_ACTIVE_ANON), | |
2783 | global_page_state(NR_INACTIVE_ANON), | |
2784 | global_page_state(NR_ISOLATED_ANON), | |
2785 | global_page_state(NR_ACTIVE_FILE), | |
2786 | global_page_state(NR_INACTIVE_FILE), | |
2787 | global_page_state(NR_ISOLATED_FILE), | |
2788 | global_page_state(NR_UNEVICTABLE), | |
2789 | global_page_state(NR_FILE_DIRTY), | |
2790 | global_page_state(NR_WRITEBACK), | |
2791 | global_page_state(NR_UNSTABLE_NFS), | |
2792 | global_page_state(NR_FREE_PAGES), | |
2793 | global_page_state(NR_SLAB_RECLAIMABLE), | |
2794 | global_page_state(NR_SLAB_UNRECLAIMABLE), | |
2795 | global_page_state(NR_FILE_MAPPED), | |
2796 | global_page_state(NR_SHMEM), | |
2797 | global_page_state(NR_PAGETABLE), | |
2798 | global_page_state(NR_BOUNCE)); | |
2799 | ||
2800 | for_each_populated_zone(zone) { | |
2801 | int i; | |
2802 | ||
2803 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
2804 | continue; | |
2805 | show_node(zone); | |
2806 | printk("%s" | |
2807 | " free:%lukB" | |
2808 | " min:%lukB" | |
2809 | " low:%lukB" | |
2810 | " high:%lukB" | |
2811 | " active_anon:%lukB" | |
2812 | " inactive_anon:%lukB" | |
2813 | " active_file:%lukB" | |
2814 | " inactive_file:%lukB" | |
2815 | " unevictable:%lukB" | |
2816 | " isolated(anon):%lukB" | |
2817 | " isolated(file):%lukB" | |
2818 | " present:%lukB" | |
2819 | " mlocked:%lukB" | |
2820 | " dirty:%lukB" | |
2821 | " writeback:%lukB" | |
2822 | " mapped:%lukB" | |
2823 | " shmem:%lukB" | |
2824 | " slab_reclaimable:%lukB" | |
2825 | " slab_unreclaimable:%lukB" | |
2826 | " kernel_stack:%lukB" | |
2827 | " pagetables:%lukB" | |
2828 | " unstable:%lukB" | |
2829 | " bounce:%lukB" | |
2830 | " writeback_tmp:%lukB" | |
2831 | " pages_scanned:%lu" | |
2832 | " all_unreclaimable? %s" | |
2833 | "\n", | |
2834 | zone->name, | |
2835 | K(zone_page_state(zone, NR_FREE_PAGES)), | |
2836 | K(min_wmark_pages(zone)), | |
2837 | K(low_wmark_pages(zone)), | |
2838 | K(high_wmark_pages(zone)), | |
2839 | K(zone_page_state(zone, NR_ACTIVE_ANON)), | |
2840 | K(zone_page_state(zone, NR_INACTIVE_ANON)), | |
2841 | K(zone_page_state(zone, NR_ACTIVE_FILE)), | |
2842 | K(zone_page_state(zone, NR_INACTIVE_FILE)), | |
2843 | K(zone_page_state(zone, NR_UNEVICTABLE)), | |
2844 | K(zone_page_state(zone, NR_ISOLATED_ANON)), | |
2845 | K(zone_page_state(zone, NR_ISOLATED_FILE)), | |
2846 | K(zone->present_pages), | |
2847 | K(zone_page_state(zone, NR_MLOCK)), | |
2848 | K(zone_page_state(zone, NR_FILE_DIRTY)), | |
2849 | K(zone_page_state(zone, NR_WRITEBACK)), | |
2850 | K(zone_page_state(zone, NR_FILE_MAPPED)), | |
2851 | K(zone_page_state(zone, NR_SHMEM)), | |
2852 | K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), | |
2853 | K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), | |
2854 | zone_page_state(zone, NR_KERNEL_STACK) * | |
2855 | THREAD_SIZE / 1024, | |
2856 | K(zone_page_state(zone, NR_PAGETABLE)), | |
2857 | K(zone_page_state(zone, NR_UNSTABLE_NFS)), | |
2858 | K(zone_page_state(zone, NR_BOUNCE)), | |
2859 | K(zone_page_state(zone, NR_WRITEBACK_TEMP)), | |
2860 | zone->pages_scanned, | |
2861 | (zone->all_unreclaimable ? "yes" : "no") | |
2862 | ); | |
2863 | printk("lowmem_reserve[]:"); | |
2864 | for (i = 0; i < MAX_NR_ZONES; i++) | |
2865 | printk(" %lu", zone->lowmem_reserve[i]); | |
2866 | printk("\n"); | |
2867 | } | |
2868 | ||
2869 | for_each_populated_zone(zone) { | |
2870 | unsigned long nr[MAX_ORDER], flags, order, total = 0; | |
2871 | ||
2872 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
2873 | continue; | |
2874 | show_node(zone); | |
2875 | printk("%s: ", zone->name); | |
2876 | ||
2877 | spin_lock_irqsave(&zone->lock, flags); | |
2878 | for (order = 0; order < MAX_ORDER; order++) { | |
2879 | nr[order] = zone->free_area[order].nr_free; | |
2880 | total += nr[order] << order; | |
2881 | } | |
2882 | spin_unlock_irqrestore(&zone->lock, flags); | |
2883 | for (order = 0; order < MAX_ORDER; order++) | |
2884 | printk("%lu*%lukB ", nr[order], K(1UL) << order); | |
2885 | printk("= %lukB\n", K(total)); | |
2886 | } | |
2887 | ||
2888 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); | |
2889 | ||
2890 | show_swap_cache_info(); | |
2891 | } | |
2892 | ||
2893 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) | |
2894 | { | |
2895 | zoneref->zone = zone; | |
2896 | zoneref->zone_idx = zone_idx(zone); | |
2897 | } | |
2898 | ||
2899 | /* | |
2900 | * Builds allocation fallback zone lists. | |
2901 | * | |
2902 | * Add all populated zones of a node to the zonelist. | |
2903 | */ | |
2904 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, | |
2905 | int nr_zones, enum zone_type zone_type) | |
2906 | { | |
2907 | struct zone *zone; | |
2908 | ||
2909 | BUG_ON(zone_type >= MAX_NR_ZONES); | |
2910 | zone_type++; | |
2911 | ||
2912 | do { | |
2913 | zone_type--; | |
2914 | zone = pgdat->node_zones + zone_type; | |
2915 | if (populated_zone(zone)) { | |
2916 | zoneref_set_zone(zone, | |
2917 | &zonelist->_zonerefs[nr_zones++]); | |
2918 | check_highest_zone(zone_type); | |
2919 | } | |
2920 | ||
2921 | } while (zone_type); | |
2922 | return nr_zones; | |
2923 | } | |
2924 | ||
2925 | ||
2926 | /* | |
2927 | * zonelist_order: | |
2928 | * 0 = automatic detection of better ordering. | |
2929 | * 1 = order by ([node] distance, -zonetype) | |
2930 | * 2 = order by (-zonetype, [node] distance) | |
2931 | * | |
2932 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create | |
2933 | * the same zonelist. So only NUMA can configure this param. | |
2934 | */ | |
2935 | #define ZONELIST_ORDER_DEFAULT 0 | |
2936 | #define ZONELIST_ORDER_NODE 1 | |
2937 | #define ZONELIST_ORDER_ZONE 2 | |
2938 | ||
2939 | /* zonelist order in the kernel. | |
2940 | * set_zonelist_order() will set this to NODE or ZONE. | |
2941 | */ | |
2942 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2943 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; | |
2944 | ||
2945 | ||
2946 | #ifdef CONFIG_NUMA | |
2947 | /* The value user specified ....changed by config */ | |
2948 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2949 | /* string for sysctl */ | |
2950 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
2951 | char numa_zonelist_order[16] = "default"; | |
2952 | ||
2953 | /* | |
2954 | * interface for configure zonelist ordering. | |
2955 | * command line option "numa_zonelist_order" | |
2956 | * = "[dD]efault - default, automatic configuration. | |
2957 | * = "[nN]ode - order by node locality, then by zone within node | |
2958 | * = "[zZ]one - order by zone, then by locality within zone | |
2959 | */ | |
2960 | ||
2961 | static int __parse_numa_zonelist_order(char *s) | |
2962 | { | |
2963 | if (*s == 'd' || *s == 'D') { | |
2964 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2965 | } else if (*s == 'n' || *s == 'N') { | |
2966 | user_zonelist_order = ZONELIST_ORDER_NODE; | |
2967 | } else if (*s == 'z' || *s == 'Z') { | |
2968 | user_zonelist_order = ZONELIST_ORDER_ZONE; | |
2969 | } else { | |
2970 | printk(KERN_WARNING | |
2971 | "Ignoring invalid numa_zonelist_order value: " | |
2972 | "%s\n", s); | |
2973 | return -EINVAL; | |
2974 | } | |
2975 | return 0; | |
2976 | } | |
2977 | ||
2978 | static __init int setup_numa_zonelist_order(char *s) | |
2979 | { | |
2980 | int ret; | |
2981 | ||
2982 | if (!s) | |
2983 | return 0; | |
2984 | ||
2985 | ret = __parse_numa_zonelist_order(s); | |
2986 | if (ret == 0) | |
2987 | strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); | |
2988 | ||
2989 | return ret; | |
2990 | } | |
2991 | early_param("numa_zonelist_order", setup_numa_zonelist_order); | |
2992 | ||
2993 | /* | |
2994 | * sysctl handler for numa_zonelist_order | |
2995 | */ | |
2996 | int numa_zonelist_order_handler(ctl_table *table, int write, | |
2997 | void __user *buffer, size_t *length, | |
2998 | loff_t *ppos) | |
2999 | { | |
3000 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; | |
3001 | int ret; | |
3002 | static DEFINE_MUTEX(zl_order_mutex); | |
3003 | ||
3004 | mutex_lock(&zl_order_mutex); | |
3005 | if (write) | |
3006 | strcpy(saved_string, (char*)table->data); | |
3007 | ret = proc_dostring(table, write, buffer, length, ppos); | |
3008 | if (ret) | |
3009 | goto out; | |
3010 | if (write) { | |
3011 | int oldval = user_zonelist_order; | |
3012 | if (__parse_numa_zonelist_order((char*)table->data)) { | |
3013 | /* | |
3014 | * bogus value. restore saved string | |
3015 | */ | |
3016 | strncpy((char*)table->data, saved_string, | |
3017 | NUMA_ZONELIST_ORDER_LEN); | |
3018 | user_zonelist_order = oldval; | |
3019 | } else if (oldval != user_zonelist_order) { | |
3020 | mutex_lock(&zonelists_mutex); | |
3021 | build_all_zonelists(NULL); | |
3022 | mutex_unlock(&zonelists_mutex); | |
3023 | } | |
3024 | } | |
3025 | out: | |
3026 | mutex_unlock(&zl_order_mutex); | |
3027 | return ret; | |
3028 | } | |
3029 | ||
3030 | ||
3031 | #define MAX_NODE_LOAD (nr_online_nodes) | |
3032 | static int node_load[MAX_NUMNODES]; | |
3033 | ||
3034 | /** | |
3035 | * find_next_best_node - find the next node that should appear in a given node's fallback list | |
3036 | * @node: node whose fallback list we're appending | |
3037 | * @used_node_mask: nodemask_t of already used nodes | |
3038 | * | |
3039 | * We use a number of factors to determine which is the next node that should | |
3040 | * appear on a given node's fallback list. The node should not have appeared | |
3041 | * already in @node's fallback list, and it should be the next closest node | |
3042 | * according to the distance array (which contains arbitrary distance values | |
3043 | * from each node to each node in the system), and should also prefer nodes | |
3044 | * with no CPUs, since presumably they'll have very little allocation pressure | |
3045 | * on them otherwise. | |
3046 | * It returns -1 if no node is found. | |
3047 | */ | |
3048 | static int find_next_best_node(int node, nodemask_t *used_node_mask) | |
3049 | { | |
3050 | int n, val; | |
3051 | int min_val = INT_MAX; | |
3052 | int best_node = -1; | |
3053 | const struct cpumask *tmp = cpumask_of_node(0); | |
3054 | ||
3055 | /* Use the local node if we haven't already */ | |
3056 | if (!node_isset(node, *used_node_mask)) { | |
3057 | node_set(node, *used_node_mask); | |
3058 | return node; | |
3059 | } | |
3060 | ||
3061 | for_each_node_state(n, N_HIGH_MEMORY) { | |
3062 | ||
3063 | /* Don't want a node to appear more than once */ | |
3064 | if (node_isset(n, *used_node_mask)) | |
3065 | continue; | |
3066 | ||
3067 | /* Use the distance array to find the distance */ | |
3068 | val = node_distance(node, n); | |
3069 | ||
3070 | /* Penalize nodes under us ("prefer the next node") */ | |
3071 | val += (n < node); | |
3072 | ||
3073 | /* Give preference to headless and unused nodes */ | |
3074 | tmp = cpumask_of_node(n); | |
3075 | if (!cpumask_empty(tmp)) | |
3076 | val += PENALTY_FOR_NODE_WITH_CPUS; | |
3077 | ||
3078 | /* Slight preference for less loaded node */ | |
3079 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); | |
3080 | val += node_load[n]; | |
3081 | ||
3082 | if (val < min_val) { | |
3083 | min_val = val; | |
3084 | best_node = n; | |
3085 | } | |
3086 | } | |
3087 | ||
3088 | if (best_node >= 0) | |
3089 | node_set(best_node, *used_node_mask); | |
3090 | ||
3091 | return best_node; | |
3092 | } | |
3093 | ||
3094 | ||
3095 | /* | |
3096 | * Build zonelists ordered by node and zones within node. | |
3097 | * This results in maximum locality--normal zone overflows into local | |
3098 | * DMA zone, if any--but risks exhausting DMA zone. | |
3099 | */ | |
3100 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) | |
3101 | { | |
3102 | int j; | |
3103 | struct zonelist *zonelist; | |
3104 | ||
3105 | zonelist = &pgdat->node_zonelists[0]; | |
3106 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) | |
3107 | ; | |
3108 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3109 | MAX_NR_ZONES - 1); | |
3110 | zonelist->_zonerefs[j].zone = NULL; | |
3111 | zonelist->_zonerefs[j].zone_idx = 0; | |
3112 | } | |
3113 | ||
3114 | /* | |
3115 | * Build gfp_thisnode zonelists | |
3116 | */ | |
3117 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
3118 | { | |
3119 | int j; | |
3120 | struct zonelist *zonelist; | |
3121 | ||
3122 | zonelist = &pgdat->node_zonelists[1]; | |
3123 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
3124 | zonelist->_zonerefs[j].zone = NULL; | |
3125 | zonelist->_zonerefs[j].zone_idx = 0; | |
3126 | } | |
3127 | ||
3128 | /* | |
3129 | * Build zonelists ordered by zone and nodes within zones. | |
3130 | * This results in conserving DMA zone[s] until all Normal memory is | |
3131 | * exhausted, but results in overflowing to remote node while memory | |
3132 | * may still exist in local DMA zone. | |
3133 | */ | |
3134 | static int node_order[MAX_NUMNODES]; | |
3135 | ||
3136 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) | |
3137 | { | |
3138 | int pos, j, node; | |
3139 | int zone_type; /* needs to be signed */ | |
3140 | struct zone *z; | |
3141 | struct zonelist *zonelist; | |
3142 | ||
3143 | zonelist = &pgdat->node_zonelists[0]; | |
3144 | pos = 0; | |
3145 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { | |
3146 | for (j = 0; j < nr_nodes; j++) { | |
3147 | node = node_order[j]; | |
3148 | z = &NODE_DATA(node)->node_zones[zone_type]; | |
3149 | if (populated_zone(z)) { | |
3150 | zoneref_set_zone(z, | |
3151 | &zonelist->_zonerefs[pos++]); | |
3152 | check_highest_zone(zone_type); | |
3153 | } | |
3154 | } | |
3155 | } | |
3156 | zonelist->_zonerefs[pos].zone = NULL; | |
3157 | zonelist->_zonerefs[pos].zone_idx = 0; | |
3158 | } | |
3159 | ||
3160 | static int default_zonelist_order(void) | |
3161 | { | |
3162 | int nid, zone_type; | |
3163 | unsigned long low_kmem_size,total_size; | |
3164 | struct zone *z; | |
3165 | int average_size; | |
3166 | /* | |
3167 | * ZONE_DMA and ZONE_DMA32 can be very small area in the system. | |
3168 | * If they are really small and used heavily, the system can fall | |
3169 | * into OOM very easily. | |
3170 | * This function detect ZONE_DMA/DMA32 size and configures zone order. | |
3171 | */ | |
3172 | /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ | |
3173 | low_kmem_size = 0; | |
3174 | total_size = 0; | |
3175 | for_each_online_node(nid) { | |
3176 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
3177 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
3178 | if (populated_zone(z)) { | |
3179 | if (zone_type < ZONE_NORMAL) | |
3180 | low_kmem_size += z->present_pages; | |
3181 | total_size += z->present_pages; | |
3182 | } else if (zone_type == ZONE_NORMAL) { | |
3183 | /* | |
3184 | * If any node has only lowmem, then node order | |
3185 | * is preferred to allow kernel allocations | |
3186 | * locally; otherwise, they can easily infringe | |
3187 | * on other nodes when there is an abundance of | |
3188 | * lowmem available to allocate from. | |
3189 | */ | |
3190 | return ZONELIST_ORDER_NODE; | |
3191 | } | |
3192 | } | |
3193 | } | |
3194 | if (!low_kmem_size || /* there are no DMA area. */ | |
3195 | low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ | |
3196 | return ZONELIST_ORDER_NODE; | |
3197 | /* | |
3198 | * look into each node's config. | |
3199 | * If there is a node whose DMA/DMA32 memory is very big area on | |
3200 | * local memory, NODE_ORDER may be suitable. | |
3201 | */ | |
3202 | average_size = total_size / | |
3203 | (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); | |
3204 | for_each_online_node(nid) { | |
3205 | low_kmem_size = 0; | |
3206 | total_size = 0; | |
3207 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
3208 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
3209 | if (populated_zone(z)) { | |
3210 | if (zone_type < ZONE_NORMAL) | |
3211 | low_kmem_size += z->present_pages; | |
3212 | total_size += z->present_pages; | |
3213 | } | |
3214 | } | |
3215 | if (low_kmem_size && | |
3216 | total_size > average_size && /* ignore small node */ | |
3217 | low_kmem_size > total_size * 70/100) | |
3218 | return ZONELIST_ORDER_NODE; | |
3219 | } | |
3220 | return ZONELIST_ORDER_ZONE; | |
3221 | } | |
3222 | ||
3223 | static void set_zonelist_order(void) | |
3224 | { | |
3225 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) | |
3226 | current_zonelist_order = default_zonelist_order(); | |
3227 | else | |
3228 | current_zonelist_order = user_zonelist_order; | |
3229 | } | |
3230 | ||
3231 | static void build_zonelists(pg_data_t *pgdat) | |
3232 | { | |
3233 | int j, node, load; | |
3234 | enum zone_type i; | |
3235 | nodemask_t used_mask; | |
3236 | int local_node, prev_node; | |
3237 | struct zonelist *zonelist; | |
3238 | int order = current_zonelist_order; | |
3239 | ||
3240 | /* initialize zonelists */ | |
3241 | for (i = 0; i < MAX_ZONELISTS; i++) { | |
3242 | zonelist = pgdat->node_zonelists + i; | |
3243 | zonelist->_zonerefs[0].zone = NULL; | |
3244 | zonelist->_zonerefs[0].zone_idx = 0; | |
3245 | } | |
3246 | ||
3247 | /* NUMA-aware ordering of nodes */ | |
3248 | local_node = pgdat->node_id; | |
3249 | load = nr_online_nodes; | |
3250 | prev_node = local_node; | |
3251 | nodes_clear(used_mask); | |
3252 | ||
3253 | memset(node_order, 0, sizeof(node_order)); | |
3254 | j = 0; | |
3255 | ||
3256 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { | |
3257 | int distance = node_distance(local_node, node); | |
3258 | ||
3259 | /* | |
3260 | * If another node is sufficiently far away then it is better | |
3261 | * to reclaim pages in a zone before going off node. | |
3262 | */ | |
3263 | if (distance > RECLAIM_DISTANCE) | |
3264 | zone_reclaim_mode = 1; | |
3265 | ||
3266 | /* | |
3267 | * We don't want to pressure a particular node. | |
3268 | * So adding penalty to the first node in same | |
3269 | * distance group to make it round-robin. | |
3270 | */ | |
3271 | if (distance != node_distance(local_node, prev_node)) | |
3272 | node_load[node] = load; | |
3273 | ||
3274 | prev_node = node; | |
3275 | load--; | |
3276 | if (order == ZONELIST_ORDER_NODE) | |
3277 | build_zonelists_in_node_order(pgdat, node); | |
3278 | else | |
3279 | node_order[j++] = node; /* remember order */ | |
3280 | } | |
3281 | ||
3282 | if (order == ZONELIST_ORDER_ZONE) { | |
3283 | /* calculate node order -- i.e., DMA last! */ | |
3284 | build_zonelists_in_zone_order(pgdat, j); | |
3285 | } | |
3286 | ||
3287 | build_thisnode_zonelists(pgdat); | |
3288 | } | |
3289 | ||
3290 | /* Construct the zonelist performance cache - see further mmzone.h */ | |
3291 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3292 | { | |
3293 | struct zonelist *zonelist; | |
3294 | struct zonelist_cache *zlc; | |
3295 | struct zoneref *z; | |
3296 | ||
3297 | zonelist = &pgdat->node_zonelists[0]; | |
3298 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; | |
3299 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
3300 | for (z = zonelist->_zonerefs; z->zone; z++) | |
3301 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); | |
3302 | } | |
3303 | ||
3304 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3305 | /* | |
3306 | * Return node id of node used for "local" allocations. | |
3307 | * I.e., first node id of first zone in arg node's generic zonelist. | |
3308 | * Used for initializing percpu 'numa_mem', which is used primarily | |
3309 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. | |
3310 | */ | |
3311 | int local_memory_node(int node) | |
3312 | { | |
3313 | struct zone *zone; | |
3314 | ||
3315 | (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), | |
3316 | gfp_zone(GFP_KERNEL), | |
3317 | NULL, | |
3318 | &zone); | |
3319 | return zone->node; | |
3320 | } | |
3321 | #endif | |
3322 | ||
3323 | #else /* CONFIG_NUMA */ | |
3324 | ||
3325 | static void set_zonelist_order(void) | |
3326 | { | |
3327 | current_zonelist_order = ZONELIST_ORDER_ZONE; | |
3328 | } | |
3329 | ||
3330 | static void build_zonelists(pg_data_t *pgdat) | |
3331 | { | |
3332 | int node, local_node; | |
3333 | enum zone_type j; | |
3334 | struct zonelist *zonelist; | |
3335 | ||
3336 | local_node = pgdat->node_id; | |
3337 | ||
3338 | zonelist = &pgdat->node_zonelists[0]; | |
3339 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
3340 | ||
3341 | /* | |
3342 | * Now we build the zonelist so that it contains the zones | |
3343 | * of all the other nodes. | |
3344 | * We don't want to pressure a particular node, so when | |
3345 | * building the zones for node N, we make sure that the | |
3346 | * zones coming right after the local ones are those from | |
3347 | * node N+1 (modulo N) | |
3348 | */ | |
3349 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
3350 | if (!node_online(node)) | |
3351 | continue; | |
3352 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3353 | MAX_NR_ZONES - 1); | |
3354 | } | |
3355 | for (node = 0; node < local_node; node++) { | |
3356 | if (!node_online(node)) | |
3357 | continue; | |
3358 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3359 | MAX_NR_ZONES - 1); | |
3360 | } | |
3361 | ||
3362 | zonelist->_zonerefs[j].zone = NULL; | |
3363 | zonelist->_zonerefs[j].zone_idx = 0; | |
3364 | } | |
3365 | ||
3366 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ | |
3367 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3368 | { | |
3369 | pgdat->node_zonelists[0].zlcache_ptr = NULL; | |
3370 | } | |
3371 | ||
3372 | #endif /* CONFIG_NUMA */ | |
3373 | ||
3374 | /* | |
3375 | * Boot pageset table. One per cpu which is going to be used for all | |
3376 | * zones and all nodes. The parameters will be set in such a way | |
3377 | * that an item put on a list will immediately be handed over to | |
3378 | * the buddy list. This is safe since pageset manipulation is done | |
3379 | * with interrupts disabled. | |
3380 | * | |
3381 | * The boot_pagesets must be kept even after bootup is complete for | |
3382 | * unused processors and/or zones. They do play a role for bootstrapping | |
3383 | * hotplugged processors. | |
3384 | * | |
3385 | * zoneinfo_show() and maybe other functions do | |
3386 | * not check if the processor is online before following the pageset pointer. | |
3387 | * Other parts of the kernel may not check if the zone is available. | |
3388 | */ | |
3389 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); | |
3390 | static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); | |
3391 | static void setup_zone_pageset(struct zone *zone); | |
3392 | ||
3393 | /* | |
3394 | * Global mutex to protect against size modification of zonelists | |
3395 | * as well as to serialize pageset setup for the new populated zone. | |
3396 | */ | |
3397 | DEFINE_MUTEX(zonelists_mutex); | |
3398 | ||
3399 | /* return values int ....just for stop_machine() */ | |
3400 | static __init_refok int __build_all_zonelists(void *data) | |
3401 | { | |
3402 | int nid; | |
3403 | int cpu; | |
3404 | ||
3405 | #ifdef CONFIG_NUMA | |
3406 | memset(node_load, 0, sizeof(node_load)); | |
3407 | #endif | |
3408 | for_each_online_node(nid) { | |
3409 | pg_data_t *pgdat = NODE_DATA(nid); | |
3410 | ||
3411 | build_zonelists(pgdat); | |
3412 | build_zonelist_cache(pgdat); | |
3413 | } | |
3414 | ||
3415 | /* | |
3416 | * Initialize the boot_pagesets that are going to be used | |
3417 | * for bootstrapping processors. The real pagesets for | |
3418 | * each zone will be allocated later when the per cpu | |
3419 | * allocator is available. | |
3420 | * | |
3421 | * boot_pagesets are used also for bootstrapping offline | |
3422 | * cpus if the system is already booted because the pagesets | |
3423 | * are needed to initialize allocators on a specific cpu too. | |
3424 | * F.e. the percpu allocator needs the page allocator which | |
3425 | * needs the percpu allocator in order to allocate its pagesets | |
3426 | * (a chicken-egg dilemma). | |
3427 | */ | |
3428 | for_each_possible_cpu(cpu) { | |
3429 | setup_pageset(&per_cpu(boot_pageset, cpu), 0); | |
3430 | ||
3431 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3432 | /* | |
3433 | * We now know the "local memory node" for each node-- | |
3434 | * i.e., the node of the first zone in the generic zonelist. | |
3435 | * Set up numa_mem percpu variable for on-line cpus. During | |
3436 | * boot, only the boot cpu should be on-line; we'll init the | |
3437 | * secondary cpus' numa_mem as they come on-line. During | |
3438 | * node/memory hotplug, we'll fixup all on-line cpus. | |
3439 | */ | |
3440 | if (cpu_online(cpu)) | |
3441 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); | |
3442 | #endif | |
3443 | } | |
3444 | ||
3445 | return 0; | |
3446 | } | |
3447 | ||
3448 | /* | |
3449 | * Called with zonelists_mutex held always | |
3450 | * unless system_state == SYSTEM_BOOTING. | |
3451 | */ | |
3452 | void __ref build_all_zonelists(void *data) | |
3453 | { | |
3454 | set_zonelist_order(); | |
3455 | ||
3456 | if (system_state == SYSTEM_BOOTING) { | |
3457 | __build_all_zonelists(NULL); | |
3458 | mminit_verify_zonelist(); | |
3459 | cpuset_init_current_mems_allowed(); | |
3460 | } else { | |
3461 | /* we have to stop all cpus to guarantee there is no user | |
3462 | of zonelist */ | |
3463 | #ifdef CONFIG_MEMORY_HOTPLUG | |
3464 | if (data) | |
3465 | setup_zone_pageset((struct zone *)data); | |
3466 | #endif | |
3467 | stop_machine(__build_all_zonelists, NULL, NULL); | |
3468 | /* cpuset refresh routine should be here */ | |
3469 | } | |
3470 | vm_total_pages = nr_free_pagecache_pages(); | |
3471 | /* | |
3472 | * Disable grouping by mobility if the number of pages in the | |
3473 | * system is too low to allow the mechanism to work. It would be | |
3474 | * more accurate, but expensive to check per-zone. This check is | |
3475 | * made on memory-hotadd so a system can start with mobility | |
3476 | * disabled and enable it later | |
3477 | */ | |
3478 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) | |
3479 | page_group_by_mobility_disabled = 1; | |
3480 | else | |
3481 | page_group_by_mobility_disabled = 0; | |
3482 | ||
3483 | printk("Built %i zonelists in %s order, mobility grouping %s. " | |
3484 | "Total pages: %ld\n", | |
3485 | nr_online_nodes, | |
3486 | zonelist_order_name[current_zonelist_order], | |
3487 | page_group_by_mobility_disabled ? "off" : "on", | |
3488 | vm_total_pages); | |
3489 | #ifdef CONFIG_NUMA | |
3490 | printk("Policy zone: %s\n", zone_names[policy_zone]); | |
3491 | #endif | |
3492 | } | |
3493 | ||
3494 | /* | |
3495 | * Helper functions to size the waitqueue hash table. | |
3496 | * Essentially these want to choose hash table sizes sufficiently | |
3497 | * large so that collisions trying to wait on pages are rare. | |
3498 | * But in fact, the number of active page waitqueues on typical | |
3499 | * systems is ridiculously low, less than 200. So this is even | |
3500 | * conservative, even though it seems large. | |
3501 | * | |
3502 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to | |
3503 | * waitqueues, i.e. the size of the waitq table given the number of pages. | |
3504 | */ | |
3505 | #define PAGES_PER_WAITQUEUE 256 | |
3506 | ||
3507 | #ifndef CONFIG_MEMORY_HOTPLUG | |
3508 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3509 | { | |
3510 | unsigned long size = 1; | |
3511 | ||
3512 | pages /= PAGES_PER_WAITQUEUE; | |
3513 | ||
3514 | while (size < pages) | |
3515 | size <<= 1; | |
3516 | ||
3517 | /* | |
3518 | * Once we have dozens or even hundreds of threads sleeping | |
3519 | * on IO we've got bigger problems than wait queue collision. | |
3520 | * Limit the size of the wait table to a reasonable size. | |
3521 | */ | |
3522 | size = min(size, 4096UL); | |
3523 | ||
3524 | return max(size, 4UL); | |
3525 | } | |
3526 | #else | |
3527 | /* | |
3528 | * A zone's size might be changed by hot-add, so it is not possible to determine | |
3529 | * a suitable size for its wait_table. So we use the maximum size now. | |
3530 | * | |
3531 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: | |
3532 | * | |
3533 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. | |
3534 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. | |
3535 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. | |
3536 | * | |
3537 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages | |
3538 | * or more by the traditional way. (See above). It equals: | |
3539 | * | |
3540 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. | |
3541 | * ia64(16K page size) : = ( 8G + 4M)byte. | |
3542 | * powerpc (64K page size) : = (32G +16M)byte. | |
3543 | */ | |
3544 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3545 | { | |
3546 | return 4096UL; | |
3547 | } | |
3548 | #endif | |
3549 | ||
3550 | /* | |
3551 | * This is an integer logarithm so that shifts can be used later | |
3552 | * to extract the more random high bits from the multiplicative | |
3553 | * hash function before the remainder is taken. | |
3554 | */ | |
3555 | static inline unsigned long wait_table_bits(unsigned long size) | |
3556 | { | |
3557 | return ffz(~size); | |
3558 | } | |
3559 | ||
3560 | #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) | |
3561 | ||
3562 | /* | |
3563 | * Check if a pageblock contains reserved pages | |
3564 | */ | |
3565 | static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) | |
3566 | { | |
3567 | unsigned long pfn; | |
3568 | ||
3569 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
3570 | if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) | |
3571 | return 1; | |
3572 | } | |
3573 | return 0; | |
3574 | } | |
3575 | ||
3576 | /* | |
3577 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number | |
3578 | * of blocks reserved is based on min_wmark_pages(zone). The memory within | |
3579 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes | |
3580 | * higher will lead to a bigger reserve which will get freed as contiguous | |
3581 | * blocks as reclaim kicks in | |
3582 | */ | |
3583 | static void setup_zone_migrate_reserve(struct zone *zone) | |
3584 | { | |
3585 | unsigned long start_pfn, pfn, end_pfn, block_end_pfn; | |
3586 | struct page *page; | |
3587 | unsigned long block_migratetype; | |
3588 | int reserve; | |
3589 | ||
3590 | /* | |
3591 | * Get the start pfn, end pfn and the number of blocks to reserve | |
3592 | * We have to be careful to be aligned to pageblock_nr_pages to | |
3593 | * make sure that we always check pfn_valid for the first page in | |
3594 | * the block. | |
3595 | */ | |
3596 | start_pfn = zone->zone_start_pfn; | |
3597 | end_pfn = start_pfn + zone->spanned_pages; | |
3598 | start_pfn = roundup(start_pfn, pageblock_nr_pages); | |
3599 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> | |
3600 | pageblock_order; | |
3601 | ||
3602 | /* | |
3603 | * Reserve blocks are generally in place to help high-order atomic | |
3604 | * allocations that are short-lived. A min_free_kbytes value that | |
3605 | * would result in more than 2 reserve blocks for atomic allocations | |
3606 | * is assumed to be in place to help anti-fragmentation for the | |
3607 | * future allocation of hugepages at runtime. | |
3608 | */ | |
3609 | reserve = min(2, reserve); | |
3610 | ||
3611 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
3612 | if (!pfn_valid(pfn)) | |
3613 | continue; | |
3614 | page = pfn_to_page(pfn); | |
3615 | ||
3616 | /* Watch out for overlapping nodes */ | |
3617 | if (page_to_nid(page) != zone_to_nid(zone)) | |
3618 | continue; | |
3619 | ||
3620 | block_migratetype = get_pageblock_migratetype(page); | |
3621 | ||
3622 | /* Only test what is necessary when the reserves are not met */ | |
3623 | if (reserve > 0) { | |
3624 | /* | |
3625 | * Blocks with reserved pages will never free, skip | |
3626 | * them. | |
3627 | */ | |
3628 | block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); | |
3629 | if (pageblock_is_reserved(pfn, block_end_pfn)) | |
3630 | continue; | |
3631 | ||
3632 | /* If this block is reserved, account for it */ | |
3633 | if (block_migratetype == MIGRATE_RESERVE) { | |
3634 | reserve--; | |
3635 | continue; | |
3636 | } | |
3637 | ||
3638 | /* Suitable for reserving if this block is movable */ | |
3639 | if (block_migratetype == MIGRATE_MOVABLE) { | |
3640 | set_pageblock_migratetype(page, | |
3641 | MIGRATE_RESERVE); | |
3642 | move_freepages_block(zone, page, | |
3643 | MIGRATE_RESERVE); | |
3644 | reserve--; | |
3645 | continue; | |
3646 | } | |
3647 | } | |
3648 | ||
3649 | /* | |
3650 | * If the reserve is met and this is a previous reserved block, | |
3651 | * take it back | |
3652 | */ | |
3653 | if (block_migratetype == MIGRATE_RESERVE) { | |
3654 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
3655 | move_freepages_block(zone, page, MIGRATE_MOVABLE); | |
3656 | } | |
3657 | } | |
3658 | } | |
3659 | ||
3660 | /* | |
3661 | * Initially all pages are reserved - free ones are freed | |
3662 | * up by free_all_bootmem() once the early boot process is | |
3663 | * done. Non-atomic initialization, single-pass. | |
3664 | */ | |
3665 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, | |
3666 | unsigned long start_pfn, enum memmap_context context) | |
3667 | { | |
3668 | struct page *page; | |
3669 | unsigned long end_pfn = start_pfn + size; | |
3670 | unsigned long pfn; | |
3671 | struct zone *z; | |
3672 | ||
3673 | if (highest_memmap_pfn < end_pfn - 1) | |
3674 | highest_memmap_pfn = end_pfn - 1; | |
3675 | ||
3676 | z = &NODE_DATA(nid)->node_zones[zone]; | |
3677 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
3678 | /* | |
3679 | * There can be holes in boot-time mem_map[]s | |
3680 | * handed to this function. They do not | |
3681 | * exist on hotplugged memory. | |
3682 | */ | |
3683 | if (context == MEMMAP_EARLY) { | |
3684 | if (!early_pfn_valid(pfn)) | |
3685 | continue; | |
3686 | if (!early_pfn_in_nid(pfn, nid)) | |
3687 | continue; | |
3688 | } | |
3689 | page = pfn_to_page(pfn); | |
3690 | set_page_links(page, zone, nid, pfn); | |
3691 | mminit_verify_page_links(page, zone, nid, pfn); | |
3692 | init_page_count(page); | |
3693 | reset_page_mapcount(page); | |
3694 | SetPageReserved(page); | |
3695 | /* | |
3696 | * Mark the block movable so that blocks are reserved for | |
3697 | * movable at startup. This will force kernel allocations | |
3698 | * to reserve their blocks rather than leaking throughout | |
3699 | * the address space during boot when many long-lived | |
3700 | * kernel allocations are made. Later some blocks near | |
3701 | * the start are marked MIGRATE_RESERVE by | |
3702 | * setup_zone_migrate_reserve() | |
3703 | * | |
3704 | * bitmap is created for zone's valid pfn range. but memmap | |
3705 | * can be created for invalid pages (for alignment) | |
3706 | * check here not to call set_pageblock_migratetype() against | |
3707 | * pfn out of zone. | |
3708 | */ | |
3709 | if ((z->zone_start_pfn <= pfn) | |
3710 | && (pfn < z->zone_start_pfn + z->spanned_pages) | |
3711 | && !(pfn & (pageblock_nr_pages - 1))) | |
3712 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
3713 | ||
3714 | INIT_LIST_HEAD(&page->lru); | |
3715 | #ifdef WANT_PAGE_VIRTUAL | |
3716 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ | |
3717 | if (!is_highmem_idx(zone)) | |
3718 | set_page_address(page, __va(pfn << PAGE_SHIFT)); | |
3719 | #endif | |
3720 | } | |
3721 | } | |
3722 | ||
3723 | static void __meminit zone_init_free_lists(struct zone *zone) | |
3724 | { | |
3725 | int order, t; | |
3726 | for_each_migratetype_order(order, t) { | |
3727 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); | |
3728 | zone->free_area[order].nr_free = 0; | |
3729 | } | |
3730 | } | |
3731 | ||
3732 | #ifndef __HAVE_ARCH_MEMMAP_INIT | |
3733 | #define memmap_init(size, nid, zone, start_pfn) \ | |
3734 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) | |
3735 | #endif | |
3736 | ||
3737 | static int zone_batchsize(struct zone *zone) | |
3738 | { | |
3739 | #ifdef CONFIG_MMU | |
3740 | int batch; | |
3741 | ||
3742 | /* | |
3743 | * The per-cpu-pages pools are set to around 1000th of the | |
3744 | * size of the zone. But no more than 1/2 of a meg. | |
3745 | * | |
3746 | * OK, so we don't know how big the cache is. So guess. | |
3747 | */ | |
3748 | batch = zone->present_pages / 1024; | |
3749 | if (batch * PAGE_SIZE > 512 * 1024) | |
3750 | batch = (512 * 1024) / PAGE_SIZE; | |
3751 | batch /= 4; /* We effectively *= 4 below */ | |
3752 | if (batch < 1) | |
3753 | batch = 1; | |
3754 | ||
3755 | /* | |
3756 | * Clamp the batch to a 2^n - 1 value. Having a power | |
3757 | * of 2 value was found to be more likely to have | |
3758 | * suboptimal cache aliasing properties in some cases. | |
3759 | * | |
3760 | * For example if 2 tasks are alternately allocating | |
3761 | * batches of pages, one task can end up with a lot | |
3762 | * of pages of one half of the possible page colors | |
3763 | * and the other with pages of the other colors. | |
3764 | */ | |
3765 | batch = rounddown_pow_of_two(batch + batch/2) - 1; | |
3766 | ||
3767 | return batch; | |
3768 | ||
3769 | #else | |
3770 | /* The deferral and batching of frees should be suppressed under NOMMU | |
3771 | * conditions. | |
3772 | * | |
3773 | * The problem is that NOMMU needs to be able to allocate large chunks | |
3774 | * of contiguous memory as there's no hardware page translation to | |
3775 | * assemble apparent contiguous memory from discontiguous pages. | |
3776 | * | |
3777 | * Queueing large contiguous runs of pages for batching, however, | |
3778 | * causes the pages to actually be freed in smaller chunks. As there | |
3779 | * can be a significant delay between the individual batches being | |
3780 | * recycled, this leads to the once large chunks of space being | |
3781 | * fragmented and becoming unavailable for high-order allocations. | |
3782 | */ | |
3783 | return 0; | |
3784 | #endif | |
3785 | } | |
3786 | ||
3787 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) | |
3788 | { | |
3789 | struct per_cpu_pages *pcp; | |
3790 | int migratetype; | |
3791 | ||
3792 | memset(p, 0, sizeof(*p)); | |
3793 | ||
3794 | pcp = &p->pcp; | |
3795 | pcp->count = 0; | |
3796 | pcp->high = 6 * batch; | |
3797 | pcp->batch = max(1UL, 1 * batch); | |
3798 | for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) | |
3799 | INIT_LIST_HEAD(&pcp->lists[migratetype]); | |
3800 | } | |
3801 | ||
3802 | /* | |
3803 | * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist | |
3804 | * to the value high for the pageset p. | |
3805 | */ | |
3806 | ||
3807 | static void setup_pagelist_highmark(struct per_cpu_pageset *p, | |
3808 | unsigned long high) | |
3809 | { | |
3810 | struct per_cpu_pages *pcp; | |
3811 | ||
3812 | pcp = &p->pcp; | |
3813 | pcp->high = high; | |
3814 | pcp->batch = max(1UL, high/4); | |
3815 | if ((high/4) > (PAGE_SHIFT * 8)) | |
3816 | pcp->batch = PAGE_SHIFT * 8; | |
3817 | } | |
3818 | ||
3819 | static void setup_zone_pageset(struct zone *zone) | |
3820 | { | |
3821 | int cpu; | |
3822 | ||
3823 | zone->pageset = alloc_percpu(struct per_cpu_pageset); | |
3824 | ||
3825 | for_each_possible_cpu(cpu) { | |
3826 | struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); | |
3827 | ||
3828 | setup_pageset(pcp, zone_batchsize(zone)); | |
3829 | ||
3830 | if (percpu_pagelist_fraction) | |
3831 | setup_pagelist_highmark(pcp, | |
3832 | (zone->present_pages / | |
3833 | percpu_pagelist_fraction)); | |
3834 | } | |
3835 | } | |
3836 | ||
3837 | /* | |
3838 | * Allocate per cpu pagesets and initialize them. | |
3839 | * Before this call only boot pagesets were available. | |
3840 | */ | |
3841 | void __init setup_per_cpu_pageset(void) | |
3842 | { | |
3843 | struct zone *zone; | |
3844 | ||
3845 | for_each_populated_zone(zone) | |
3846 | setup_zone_pageset(zone); | |
3847 | } | |
3848 | ||
3849 | static noinline __init_refok | |
3850 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) | |
3851 | { | |
3852 | int i; | |
3853 | struct pglist_data *pgdat = zone->zone_pgdat; | |
3854 | size_t alloc_size; | |
3855 | ||
3856 | /* | |
3857 | * The per-page waitqueue mechanism uses hashed waitqueues | |
3858 | * per zone. | |
3859 | */ | |
3860 | zone->wait_table_hash_nr_entries = | |
3861 | wait_table_hash_nr_entries(zone_size_pages); | |
3862 | zone->wait_table_bits = | |
3863 | wait_table_bits(zone->wait_table_hash_nr_entries); | |
3864 | alloc_size = zone->wait_table_hash_nr_entries | |
3865 | * sizeof(wait_queue_head_t); | |
3866 | ||
3867 | if (!slab_is_available()) { | |
3868 | zone->wait_table = (wait_queue_head_t *) | |
3869 | alloc_bootmem_node_nopanic(pgdat, alloc_size); | |
3870 | } else { | |
3871 | /* | |
3872 | * This case means that a zone whose size was 0 gets new memory | |
3873 | * via memory hot-add. | |
3874 | * But it may be the case that a new node was hot-added. In | |
3875 | * this case vmalloc() will not be able to use this new node's | |
3876 | * memory - this wait_table must be initialized to use this new | |
3877 | * node itself as well. | |
3878 | * To use this new node's memory, further consideration will be | |
3879 | * necessary. | |
3880 | */ | |
3881 | zone->wait_table = vmalloc(alloc_size); | |
3882 | } | |
3883 | if (!zone->wait_table) | |
3884 | return -ENOMEM; | |
3885 | ||
3886 | for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) | |
3887 | init_waitqueue_head(zone->wait_table + i); | |
3888 | ||
3889 | return 0; | |
3890 | } | |
3891 | ||
3892 | static int __zone_pcp_update(void *data) | |
3893 | { | |
3894 | struct zone *zone = data; | |
3895 | int cpu; | |
3896 | unsigned long batch = zone_batchsize(zone), flags; | |
3897 | ||
3898 | for_each_possible_cpu(cpu) { | |
3899 | struct per_cpu_pageset *pset; | |
3900 | struct per_cpu_pages *pcp; | |
3901 | ||
3902 | pset = per_cpu_ptr(zone->pageset, cpu); | |
3903 | pcp = &pset->pcp; | |
3904 | ||
3905 | local_irq_save(flags); | |
3906 | free_pcppages_bulk(zone, pcp->count, pcp); | |
3907 | setup_pageset(pset, batch); | |
3908 | local_irq_restore(flags); | |
3909 | } | |
3910 | return 0; | |
3911 | } | |
3912 | ||
3913 | void zone_pcp_update(struct zone *zone) | |
3914 | { | |
3915 | stop_machine(__zone_pcp_update, zone, NULL); | |
3916 | } | |
3917 | ||
3918 | static __meminit void zone_pcp_init(struct zone *zone) | |
3919 | { | |
3920 | /* | |
3921 | * per cpu subsystem is not up at this point. The following code | |
3922 | * relies on the ability of the linker to provide the | |
3923 | * offset of a (static) per cpu variable into the per cpu area. | |
3924 | */ | |
3925 | zone->pageset = &boot_pageset; | |
3926 | ||
3927 | if (zone->present_pages) | |
3928 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", | |
3929 | zone->name, zone->present_pages, | |
3930 | zone_batchsize(zone)); | |
3931 | } | |
3932 | ||
3933 | __meminit int init_currently_empty_zone(struct zone *zone, | |
3934 | unsigned long zone_start_pfn, | |
3935 | unsigned long size, | |
3936 | enum memmap_context context) | |
3937 | { | |
3938 | struct pglist_data *pgdat = zone->zone_pgdat; | |
3939 | int ret; | |
3940 | ret = zone_wait_table_init(zone, size); | |
3941 | if (ret) | |
3942 | return ret; | |
3943 | pgdat->nr_zones = zone_idx(zone) + 1; | |
3944 | ||
3945 | zone->zone_start_pfn = zone_start_pfn; | |
3946 | ||
3947 | mminit_dprintk(MMINIT_TRACE, "memmap_init", | |
3948 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", | |
3949 | pgdat->node_id, | |
3950 | (unsigned long)zone_idx(zone), | |
3951 | zone_start_pfn, (zone_start_pfn + size)); | |
3952 | ||
3953 | zone_init_free_lists(zone); | |
3954 | ||
3955 | return 0; | |
3956 | } | |
3957 | ||
3958 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
3959 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID | |
3960 | /* | |
3961 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. | |
3962 | * Architectures may implement their own version but if add_active_range() | |
3963 | * was used and there are no special requirements, this is a convenient | |
3964 | * alternative | |
3965 | */ | |
3966 | int __meminit __early_pfn_to_nid(unsigned long pfn) | |
3967 | { | |
3968 | unsigned long start_pfn, end_pfn; | |
3969 | int i, nid; | |
3970 | ||
3971 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) | |
3972 | if (start_pfn <= pfn && pfn < end_pfn) | |
3973 | return nid; | |
3974 | /* This is a memory hole */ | |
3975 | return -1; | |
3976 | } | |
3977 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ | |
3978 | ||
3979 | int __meminit early_pfn_to_nid(unsigned long pfn) | |
3980 | { | |
3981 | int nid; | |
3982 | ||
3983 | nid = __early_pfn_to_nid(pfn); | |
3984 | if (nid >= 0) | |
3985 | return nid; | |
3986 | /* just returns 0 */ | |
3987 | return 0; | |
3988 | } | |
3989 | ||
3990 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES | |
3991 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) | |
3992 | { | |
3993 | int nid; | |
3994 | ||
3995 | nid = __early_pfn_to_nid(pfn); | |
3996 | if (nid >= 0 && nid != node) | |
3997 | return false; | |
3998 | return true; | |
3999 | } | |
4000 | #endif | |
4001 | ||
4002 | /** | |
4003 | * free_bootmem_with_active_regions - Call free_bootmem_node for each active range | |
4004 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. | |
4005 | * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node | |
4006 | * | |
4007 | * If an architecture guarantees that all ranges registered with | |
4008 | * add_active_ranges() contain no holes and may be freed, this | |
4009 | * this function may be used instead of calling free_bootmem() manually. | |
4010 | */ | |
4011 | void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) | |
4012 | { | |
4013 | unsigned long start_pfn, end_pfn; | |
4014 | int i, this_nid; | |
4015 | ||
4016 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { | |
4017 | start_pfn = min(start_pfn, max_low_pfn); | |
4018 | end_pfn = min(end_pfn, max_low_pfn); | |
4019 | ||
4020 | if (start_pfn < end_pfn) | |
4021 | free_bootmem_node(NODE_DATA(this_nid), | |
4022 | PFN_PHYS(start_pfn), | |
4023 | (end_pfn - start_pfn) << PAGE_SHIFT); | |
4024 | } | |
4025 | } | |
4026 | ||
4027 | /** | |
4028 | * sparse_memory_present_with_active_regions - Call memory_present for each active range | |
4029 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. | |
4030 | * | |
4031 | * If an architecture guarantees that all ranges registered with | |
4032 | * add_active_ranges() contain no holes and may be freed, this | |
4033 | * function may be used instead of calling memory_present() manually. | |
4034 | */ | |
4035 | void __init sparse_memory_present_with_active_regions(int nid) | |
4036 | { | |
4037 | unsigned long start_pfn, end_pfn; | |
4038 | int i, this_nid; | |
4039 | ||
4040 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) | |
4041 | memory_present(this_nid, start_pfn, end_pfn); | |
4042 | } | |
4043 | ||
4044 | /** | |
4045 | * get_pfn_range_for_nid - Return the start and end page frames for a node | |
4046 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. | |
4047 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. | |
4048 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. | |
4049 | * | |
4050 | * It returns the start and end page frame of a node based on information | |
4051 | * provided by an arch calling add_active_range(). If called for a node | |
4052 | * with no available memory, a warning is printed and the start and end | |
4053 | * PFNs will be 0. | |
4054 | */ | |
4055 | void __meminit get_pfn_range_for_nid(unsigned int nid, | |
4056 | unsigned long *start_pfn, unsigned long *end_pfn) | |
4057 | { | |
4058 | unsigned long this_start_pfn, this_end_pfn; | |
4059 | int i; | |
4060 | ||
4061 | *start_pfn = -1UL; | |
4062 | *end_pfn = 0; | |
4063 | ||
4064 | for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { | |
4065 | *start_pfn = min(*start_pfn, this_start_pfn); | |
4066 | *end_pfn = max(*end_pfn, this_end_pfn); | |
4067 | } | |
4068 | ||
4069 | if (*start_pfn == -1UL) | |
4070 | *start_pfn = 0; | |
4071 | } | |
4072 | ||
4073 | /* | |
4074 | * This finds a zone that can be used for ZONE_MOVABLE pages. The | |
4075 | * assumption is made that zones within a node are ordered in monotonic | |
4076 | * increasing memory addresses so that the "highest" populated zone is used | |
4077 | */ | |
4078 | static void __init find_usable_zone_for_movable(void) | |
4079 | { | |
4080 | int zone_index; | |
4081 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { | |
4082 | if (zone_index == ZONE_MOVABLE) | |
4083 | continue; | |
4084 | ||
4085 | if (arch_zone_highest_possible_pfn[zone_index] > | |
4086 | arch_zone_lowest_possible_pfn[zone_index]) | |
4087 | break; | |
4088 | } | |
4089 | ||
4090 | VM_BUG_ON(zone_index == -1); | |
4091 | movable_zone = zone_index; | |
4092 | } | |
4093 | ||
4094 | /* | |
4095 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE | |
4096 | * because it is sized independent of architecture. Unlike the other zones, | |
4097 | * the starting point for ZONE_MOVABLE is not fixed. It may be different | |
4098 | * in each node depending on the size of each node and how evenly kernelcore | |
4099 | * is distributed. This helper function adjusts the zone ranges | |
4100 | * provided by the architecture for a given node by using the end of the | |
4101 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that | |
4102 | * zones within a node are in order of monotonic increases memory addresses | |
4103 | */ | |
4104 | static void __meminit adjust_zone_range_for_zone_movable(int nid, | |
4105 | unsigned long zone_type, | |
4106 | unsigned long node_start_pfn, | |
4107 | unsigned long node_end_pfn, | |
4108 | unsigned long *zone_start_pfn, | |
4109 | unsigned long *zone_end_pfn) | |
4110 | { | |
4111 | /* Only adjust if ZONE_MOVABLE is on this node */ | |
4112 | if (zone_movable_pfn[nid]) { | |
4113 | /* Size ZONE_MOVABLE */ | |
4114 | if (zone_type == ZONE_MOVABLE) { | |
4115 | *zone_start_pfn = zone_movable_pfn[nid]; | |
4116 | *zone_end_pfn = min(node_end_pfn, | |
4117 | arch_zone_highest_possible_pfn[movable_zone]); | |
4118 | ||
4119 | /* Adjust for ZONE_MOVABLE starting within this range */ | |
4120 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && | |
4121 | *zone_end_pfn > zone_movable_pfn[nid]) { | |
4122 | *zone_end_pfn = zone_movable_pfn[nid]; | |
4123 | ||
4124 | /* Check if this whole range is within ZONE_MOVABLE */ | |
4125 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) | |
4126 | *zone_start_pfn = *zone_end_pfn; | |
4127 | } | |
4128 | } | |
4129 | ||
4130 | /* | |
4131 | * Return the number of pages a zone spans in a node, including holes | |
4132 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() | |
4133 | */ | |
4134 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4135 | unsigned long zone_type, | |
4136 | unsigned long *ignored) | |
4137 | { | |
4138 | unsigned long node_start_pfn, node_end_pfn; | |
4139 | unsigned long zone_start_pfn, zone_end_pfn; | |
4140 | ||
4141 | /* Get the start and end of the node and zone */ | |
4142 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
4143 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; | |
4144 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; | |
4145 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4146 | node_start_pfn, node_end_pfn, | |
4147 | &zone_start_pfn, &zone_end_pfn); | |
4148 | ||
4149 | /* Check that this node has pages within the zone's required range */ | |
4150 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) | |
4151 | return 0; | |
4152 | ||
4153 | /* Move the zone boundaries inside the node if necessary */ | |
4154 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); | |
4155 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); | |
4156 | ||
4157 | /* Return the spanned pages */ | |
4158 | return zone_end_pfn - zone_start_pfn; | |
4159 | } | |
4160 | ||
4161 | /* | |
4162 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, | |
4163 | * then all holes in the requested range will be accounted for. | |
4164 | */ | |
4165 | unsigned long __meminit __absent_pages_in_range(int nid, | |
4166 | unsigned long range_start_pfn, | |
4167 | unsigned long range_end_pfn) | |
4168 | { | |
4169 | unsigned long nr_absent = range_end_pfn - range_start_pfn; | |
4170 | unsigned long start_pfn, end_pfn; | |
4171 | int i; | |
4172 | ||
4173 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
4174 | start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); | |
4175 | end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); | |
4176 | nr_absent -= end_pfn - start_pfn; | |
4177 | } | |
4178 | return nr_absent; | |
4179 | } | |
4180 | ||
4181 | /** | |
4182 | * absent_pages_in_range - Return number of page frames in holes within a range | |
4183 | * @start_pfn: The start PFN to start searching for holes | |
4184 | * @end_pfn: The end PFN to stop searching for holes | |
4185 | * | |
4186 | * It returns the number of pages frames in memory holes within a range. | |
4187 | */ | |
4188 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, | |
4189 | unsigned long end_pfn) | |
4190 | { | |
4191 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); | |
4192 | } | |
4193 | ||
4194 | /* Return the number of page frames in holes in a zone on a node */ | |
4195 | static unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4196 | unsigned long zone_type, | |
4197 | unsigned long *ignored) | |
4198 | { | |
4199 | unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; | |
4200 | unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; | |
4201 | unsigned long node_start_pfn, node_end_pfn; | |
4202 | unsigned long zone_start_pfn, zone_end_pfn; | |
4203 | ||
4204 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
4205 | zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); | |
4206 | zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); | |
4207 | ||
4208 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4209 | node_start_pfn, node_end_pfn, | |
4210 | &zone_start_pfn, &zone_end_pfn); | |
4211 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); | |
4212 | } | |
4213 | ||
4214 | #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4215 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4216 | unsigned long zone_type, | |
4217 | unsigned long *zones_size) | |
4218 | { | |
4219 | return zones_size[zone_type]; | |
4220 | } | |
4221 | ||
4222 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4223 | unsigned long zone_type, | |
4224 | unsigned long *zholes_size) | |
4225 | { | |
4226 | if (!zholes_size) | |
4227 | return 0; | |
4228 | ||
4229 | return zholes_size[zone_type]; | |
4230 | } | |
4231 | ||
4232 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4233 | ||
4234 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, | |
4235 | unsigned long *zones_size, unsigned long *zholes_size) | |
4236 | { | |
4237 | unsigned long realtotalpages, totalpages = 0; | |
4238 | enum zone_type i; | |
4239 | ||
4240 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4241 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, | |
4242 | zones_size); | |
4243 | pgdat->node_spanned_pages = totalpages; | |
4244 | ||
4245 | realtotalpages = totalpages; | |
4246 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4247 | realtotalpages -= | |
4248 | zone_absent_pages_in_node(pgdat->node_id, i, | |
4249 | zholes_size); | |
4250 | pgdat->node_present_pages = realtotalpages; | |
4251 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, | |
4252 | realtotalpages); | |
4253 | } | |
4254 | ||
4255 | #ifndef CONFIG_SPARSEMEM | |
4256 | /* | |
4257 | * Calculate the size of the zone->blockflags rounded to an unsigned long | |
4258 | * Start by making sure zonesize is a multiple of pageblock_order by rounding | |
4259 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally | |
4260 | * round what is now in bits to nearest long in bits, then return it in | |
4261 | * bytes. | |
4262 | */ | |
4263 | static unsigned long __init usemap_size(unsigned long zonesize) | |
4264 | { | |
4265 | unsigned long usemapsize; | |
4266 | ||
4267 | usemapsize = roundup(zonesize, pageblock_nr_pages); | |
4268 | usemapsize = usemapsize >> pageblock_order; | |
4269 | usemapsize *= NR_PAGEBLOCK_BITS; | |
4270 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); | |
4271 | ||
4272 | return usemapsize / 8; | |
4273 | } | |
4274 | ||
4275 | static void __init setup_usemap(struct pglist_data *pgdat, | |
4276 | struct zone *zone, unsigned long zonesize) | |
4277 | { | |
4278 | unsigned long usemapsize = usemap_size(zonesize); | |
4279 | zone->pageblock_flags = NULL; | |
4280 | if (usemapsize) | |
4281 | zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, | |
4282 | usemapsize); | |
4283 | } | |
4284 | #else | |
4285 | static inline void setup_usemap(struct pglist_data *pgdat, | |
4286 | struct zone *zone, unsigned long zonesize) {} | |
4287 | #endif /* CONFIG_SPARSEMEM */ | |
4288 | ||
4289 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
4290 | ||
4291 | /* Return a sensible default order for the pageblock size. */ | |
4292 | static inline int pageblock_default_order(void) | |
4293 | { | |
4294 | if (HPAGE_SHIFT > PAGE_SHIFT) | |
4295 | return HUGETLB_PAGE_ORDER; | |
4296 | ||
4297 | return MAX_ORDER-1; | |
4298 | } | |
4299 | ||
4300 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ | |
4301 | static inline void __init set_pageblock_order(unsigned int order) | |
4302 | { | |
4303 | /* Check that pageblock_nr_pages has not already been setup */ | |
4304 | if (pageblock_order) | |
4305 | return; | |
4306 | ||
4307 | /* | |
4308 | * Assume the largest contiguous order of interest is a huge page. | |
4309 | * This value may be variable depending on boot parameters on IA64 | |
4310 | */ | |
4311 | pageblock_order = order; | |
4312 | } | |
4313 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4314 | ||
4315 | /* | |
4316 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() | |
4317 | * and pageblock_default_order() are unused as pageblock_order is set | |
4318 | * at compile-time. See include/linux/pageblock-flags.h for the values of | |
4319 | * pageblock_order based on the kernel config | |
4320 | */ | |
4321 | static inline int pageblock_default_order(unsigned int order) | |
4322 | { | |
4323 | return MAX_ORDER-1; | |
4324 | } | |
4325 | #define set_pageblock_order(x) do {} while (0) | |
4326 | ||
4327 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4328 | ||
4329 | /* | |
4330 | * Set up the zone data structures: | |
4331 | * - mark all pages reserved | |
4332 | * - mark all memory queues empty | |
4333 | * - clear the memory bitmaps | |
4334 | */ | |
4335 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, | |
4336 | unsigned long *zones_size, unsigned long *zholes_size) | |
4337 | { | |
4338 | enum zone_type j; | |
4339 | int nid = pgdat->node_id; | |
4340 | unsigned long zone_start_pfn = pgdat->node_start_pfn; | |
4341 | int ret; | |
4342 | ||
4343 | pgdat_resize_init(pgdat); | |
4344 | pgdat->nr_zones = 0; | |
4345 | init_waitqueue_head(&pgdat->kswapd_wait); | |
4346 | pgdat->kswapd_max_order = 0; | |
4347 | pgdat_page_cgroup_init(pgdat); | |
4348 | ||
4349 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
4350 | struct zone *zone = pgdat->node_zones + j; | |
4351 | unsigned long size, realsize, memmap_pages; | |
4352 | enum lru_list lru; | |
4353 | ||
4354 | size = zone_spanned_pages_in_node(nid, j, zones_size); | |
4355 | realsize = size - zone_absent_pages_in_node(nid, j, | |
4356 | zholes_size); | |
4357 | ||
4358 | /* | |
4359 | * Adjust realsize so that it accounts for how much memory | |
4360 | * is used by this zone for memmap. This affects the watermark | |
4361 | * and per-cpu initialisations | |
4362 | */ | |
4363 | memmap_pages = | |
4364 | PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; | |
4365 | if (realsize >= memmap_pages) { | |
4366 | realsize -= memmap_pages; | |
4367 | if (memmap_pages) | |
4368 | printk(KERN_DEBUG | |
4369 | " %s zone: %lu pages used for memmap\n", | |
4370 | zone_names[j], memmap_pages); | |
4371 | } else | |
4372 | printk(KERN_WARNING | |
4373 | " %s zone: %lu pages exceeds realsize %lu\n", | |
4374 | zone_names[j], memmap_pages, realsize); | |
4375 | ||
4376 | /* Account for reserved pages */ | |
4377 | if (j == 0 && realsize > dma_reserve) { | |
4378 | realsize -= dma_reserve; | |
4379 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", | |
4380 | zone_names[0], dma_reserve); | |
4381 | } | |
4382 | ||
4383 | if (!is_highmem_idx(j)) | |
4384 | nr_kernel_pages += realsize; | |
4385 | nr_all_pages += realsize; | |
4386 | ||
4387 | zone->spanned_pages = size; | |
4388 | zone->present_pages = realsize; | |
4389 | #ifdef CONFIG_NUMA | |
4390 | zone->node = nid; | |
4391 | zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) | |
4392 | / 100; | |
4393 | zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; | |
4394 | #endif | |
4395 | zone->name = zone_names[j]; | |
4396 | spin_lock_init(&zone->lock); | |
4397 | spin_lock_init(&zone->lru_lock); | |
4398 | zone_seqlock_init(zone); | |
4399 | zone->zone_pgdat = pgdat; | |
4400 | ||
4401 | zone_pcp_init(zone); | |
4402 | for_each_lru(lru) | |
4403 | INIT_LIST_HEAD(&zone->lruvec.lists[lru]); | |
4404 | zone->reclaim_stat.recent_rotated[0] = 0; | |
4405 | zone->reclaim_stat.recent_rotated[1] = 0; | |
4406 | zone->reclaim_stat.recent_scanned[0] = 0; | |
4407 | zone->reclaim_stat.recent_scanned[1] = 0; | |
4408 | zap_zone_vm_stats(zone); | |
4409 | zone->flags = 0; | |
4410 | if (!size) | |
4411 | continue; | |
4412 | ||
4413 | set_pageblock_order(pageblock_default_order()); | |
4414 | setup_usemap(pgdat, zone, size); | |
4415 | ret = init_currently_empty_zone(zone, zone_start_pfn, | |
4416 | size, MEMMAP_EARLY); | |
4417 | BUG_ON(ret); | |
4418 | memmap_init(size, nid, j, zone_start_pfn); | |
4419 | zone_start_pfn += size; | |
4420 | } | |
4421 | } | |
4422 | ||
4423 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) | |
4424 | { | |
4425 | /* Skip empty nodes */ | |
4426 | if (!pgdat->node_spanned_pages) | |
4427 | return; | |
4428 | ||
4429 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
4430 | /* ia64 gets its own node_mem_map, before this, without bootmem */ | |
4431 | if (!pgdat->node_mem_map) { | |
4432 | unsigned long size, start, end; | |
4433 | struct page *map; | |
4434 | ||
4435 | /* | |
4436 | * The zone's endpoints aren't required to be MAX_ORDER | |
4437 | * aligned but the node_mem_map endpoints must be in order | |
4438 | * for the buddy allocator to function correctly. | |
4439 | */ | |
4440 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); | |
4441 | end = pgdat->node_start_pfn + pgdat->node_spanned_pages; | |
4442 | end = ALIGN(end, MAX_ORDER_NR_PAGES); | |
4443 | size = (end - start) * sizeof(struct page); | |
4444 | map = alloc_remap(pgdat->node_id, size); | |
4445 | if (!map) | |
4446 | map = alloc_bootmem_node_nopanic(pgdat, size); | |
4447 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); | |
4448 | } | |
4449 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
4450 | /* | |
4451 | * With no DISCONTIG, the global mem_map is just set as node 0's | |
4452 | */ | |
4453 | if (pgdat == NODE_DATA(0)) { | |
4454 | mem_map = NODE_DATA(0)->node_mem_map; | |
4455 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4456 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) | |
4457 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); | |
4458 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4459 | } | |
4460 | #endif | |
4461 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ | |
4462 | } | |
4463 | ||
4464 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, | |
4465 | unsigned long node_start_pfn, unsigned long *zholes_size) | |
4466 | { | |
4467 | pg_data_t *pgdat = NODE_DATA(nid); | |
4468 | ||
4469 | pgdat->node_id = nid; | |
4470 | pgdat->node_start_pfn = node_start_pfn; | |
4471 | calculate_node_totalpages(pgdat, zones_size, zholes_size); | |
4472 | ||
4473 | alloc_node_mem_map(pgdat); | |
4474 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
4475 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", | |
4476 | nid, (unsigned long)pgdat, | |
4477 | (unsigned long)pgdat->node_mem_map); | |
4478 | #endif | |
4479 | ||
4480 | free_area_init_core(pgdat, zones_size, zholes_size); | |
4481 | } | |
4482 | ||
4483 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4484 | ||
4485 | #if MAX_NUMNODES > 1 | |
4486 | /* | |
4487 | * Figure out the number of possible node ids. | |
4488 | */ | |
4489 | static void __init setup_nr_node_ids(void) | |
4490 | { | |
4491 | unsigned int node; | |
4492 | unsigned int highest = 0; | |
4493 | ||
4494 | for_each_node_mask(node, node_possible_map) | |
4495 | highest = node; | |
4496 | nr_node_ids = highest + 1; | |
4497 | } | |
4498 | #else | |
4499 | static inline void setup_nr_node_ids(void) | |
4500 | { | |
4501 | } | |
4502 | #endif | |
4503 | ||
4504 | /** | |
4505 | * node_map_pfn_alignment - determine the maximum internode alignment | |
4506 | * | |
4507 | * This function should be called after node map is populated and sorted. | |
4508 | * It calculates the maximum power of two alignment which can distinguish | |
4509 | * all the nodes. | |
4510 | * | |
4511 | * For example, if all nodes are 1GiB and aligned to 1GiB, the return value | |
4512 | * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the | |
4513 | * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is | |
4514 | * shifted, 1GiB is enough and this function will indicate so. | |
4515 | * | |
4516 | * This is used to test whether pfn -> nid mapping of the chosen memory | |
4517 | * model has fine enough granularity to avoid incorrect mapping for the | |
4518 | * populated node map. | |
4519 | * | |
4520 | * Returns the determined alignment in pfn's. 0 if there is no alignment | |
4521 | * requirement (single node). | |
4522 | */ | |
4523 | unsigned long __init node_map_pfn_alignment(void) | |
4524 | { | |
4525 | unsigned long accl_mask = 0, last_end = 0; | |
4526 | unsigned long start, end, mask; | |
4527 | int last_nid = -1; | |
4528 | int i, nid; | |
4529 | ||
4530 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { | |
4531 | if (!start || last_nid < 0 || last_nid == nid) { | |
4532 | last_nid = nid; | |
4533 | last_end = end; | |
4534 | continue; | |
4535 | } | |
4536 | ||
4537 | /* | |
4538 | * Start with a mask granular enough to pin-point to the | |
4539 | * start pfn and tick off bits one-by-one until it becomes | |
4540 | * too coarse to separate the current node from the last. | |
4541 | */ | |
4542 | mask = ~((1 << __ffs(start)) - 1); | |
4543 | while (mask && last_end <= (start & (mask << 1))) | |
4544 | mask <<= 1; | |
4545 | ||
4546 | /* accumulate all internode masks */ | |
4547 | accl_mask |= mask; | |
4548 | } | |
4549 | ||
4550 | /* convert mask to number of pages */ | |
4551 | return ~accl_mask + 1; | |
4552 | } | |
4553 | ||
4554 | /* Find the lowest pfn for a node */ | |
4555 | static unsigned long __init find_min_pfn_for_node(int nid) | |
4556 | { | |
4557 | unsigned long min_pfn = ULONG_MAX; | |
4558 | unsigned long start_pfn; | |
4559 | int i; | |
4560 | ||
4561 | for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) | |
4562 | min_pfn = min(min_pfn, start_pfn); | |
4563 | ||
4564 | if (min_pfn == ULONG_MAX) { | |
4565 | printk(KERN_WARNING | |
4566 | "Could not find start_pfn for node %d\n", nid); | |
4567 | return 0; | |
4568 | } | |
4569 | ||
4570 | return min_pfn; | |
4571 | } | |
4572 | ||
4573 | /** | |
4574 | * find_min_pfn_with_active_regions - Find the minimum PFN registered | |
4575 | * | |
4576 | * It returns the minimum PFN based on information provided via | |
4577 | * add_active_range(). | |
4578 | */ | |
4579 | unsigned long __init find_min_pfn_with_active_regions(void) | |
4580 | { | |
4581 | return find_min_pfn_for_node(MAX_NUMNODES); | |
4582 | } | |
4583 | ||
4584 | /* | |
4585 | * early_calculate_totalpages() | |
4586 | * Sum pages in active regions for movable zone. | |
4587 | * Populate N_HIGH_MEMORY for calculating usable_nodes. | |
4588 | */ | |
4589 | static unsigned long __init early_calculate_totalpages(void) | |
4590 | { | |
4591 | unsigned long totalpages = 0; | |
4592 | unsigned long start_pfn, end_pfn; | |
4593 | int i, nid; | |
4594 | ||
4595 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { | |
4596 | unsigned long pages = end_pfn - start_pfn; | |
4597 | ||
4598 | totalpages += pages; | |
4599 | if (pages) | |
4600 | node_set_state(nid, N_HIGH_MEMORY); | |
4601 | } | |
4602 | return totalpages; | |
4603 | } | |
4604 | ||
4605 | /* | |
4606 | * Find the PFN the Movable zone begins in each node. Kernel memory | |
4607 | * is spread evenly between nodes as long as the nodes have enough | |
4608 | * memory. When they don't, some nodes will have more kernelcore than | |
4609 | * others | |
4610 | */ | |
4611 | static void __init find_zone_movable_pfns_for_nodes(void) | |
4612 | { | |
4613 | int i, nid; | |
4614 | unsigned long usable_startpfn; | |
4615 | unsigned long kernelcore_node, kernelcore_remaining; | |
4616 | /* save the state before borrow the nodemask */ | |
4617 | nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; | |
4618 | unsigned long totalpages = early_calculate_totalpages(); | |
4619 | int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); | |
4620 | ||
4621 | /* | |
4622 | * If movablecore was specified, calculate what size of | |
4623 | * kernelcore that corresponds so that memory usable for | |
4624 | * any allocation type is evenly spread. If both kernelcore | |
4625 | * and movablecore are specified, then the value of kernelcore | |
4626 | * will be used for required_kernelcore if it's greater than | |
4627 | * what movablecore would have allowed. | |
4628 | */ | |
4629 | if (required_movablecore) { | |
4630 | unsigned long corepages; | |
4631 | ||
4632 | /* | |
4633 | * Round-up so that ZONE_MOVABLE is at least as large as what | |
4634 | * was requested by the user | |
4635 | */ | |
4636 | required_movablecore = | |
4637 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); | |
4638 | corepages = totalpages - required_movablecore; | |
4639 | ||
4640 | required_kernelcore = max(required_kernelcore, corepages); | |
4641 | } | |
4642 | ||
4643 | /* If kernelcore was not specified, there is no ZONE_MOVABLE */ | |
4644 | if (!required_kernelcore) | |
4645 | goto out; | |
4646 | ||
4647 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ | |
4648 | find_usable_zone_for_movable(); | |
4649 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; | |
4650 | ||
4651 | restart: | |
4652 | /* Spread kernelcore memory as evenly as possible throughout nodes */ | |
4653 | kernelcore_node = required_kernelcore / usable_nodes; | |
4654 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4655 | unsigned long start_pfn, end_pfn; | |
4656 | ||
4657 | /* | |
4658 | * Recalculate kernelcore_node if the division per node | |
4659 | * now exceeds what is necessary to satisfy the requested | |
4660 | * amount of memory for the kernel | |
4661 | */ | |
4662 | if (required_kernelcore < kernelcore_node) | |
4663 | kernelcore_node = required_kernelcore / usable_nodes; | |
4664 | ||
4665 | /* | |
4666 | * As the map is walked, we track how much memory is usable | |
4667 | * by the kernel using kernelcore_remaining. When it is | |
4668 | * 0, the rest of the node is usable by ZONE_MOVABLE | |
4669 | */ | |
4670 | kernelcore_remaining = kernelcore_node; | |
4671 | ||
4672 | /* Go through each range of PFNs within this node */ | |
4673 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
4674 | unsigned long size_pages; | |
4675 | ||
4676 | start_pfn = max(start_pfn, zone_movable_pfn[nid]); | |
4677 | if (start_pfn >= end_pfn) | |
4678 | continue; | |
4679 | ||
4680 | /* Account for what is only usable for kernelcore */ | |
4681 | if (start_pfn < usable_startpfn) { | |
4682 | unsigned long kernel_pages; | |
4683 | kernel_pages = min(end_pfn, usable_startpfn) | |
4684 | - start_pfn; | |
4685 | ||
4686 | kernelcore_remaining -= min(kernel_pages, | |
4687 | kernelcore_remaining); | |
4688 | required_kernelcore -= min(kernel_pages, | |
4689 | required_kernelcore); | |
4690 | ||
4691 | /* Continue if range is now fully accounted */ | |
4692 | if (end_pfn <= usable_startpfn) { | |
4693 | ||
4694 | /* | |
4695 | * Push zone_movable_pfn to the end so | |
4696 | * that if we have to rebalance | |
4697 | * kernelcore across nodes, we will | |
4698 | * not double account here | |
4699 | */ | |
4700 | zone_movable_pfn[nid] = end_pfn; | |
4701 | continue; | |
4702 | } | |
4703 | start_pfn = usable_startpfn; | |
4704 | } | |
4705 | ||
4706 | /* | |
4707 | * The usable PFN range for ZONE_MOVABLE is from | |
4708 | * start_pfn->end_pfn. Calculate size_pages as the | |
4709 | * number of pages used as kernelcore | |
4710 | */ | |
4711 | size_pages = end_pfn - start_pfn; | |
4712 | if (size_pages > kernelcore_remaining) | |
4713 | size_pages = kernelcore_remaining; | |
4714 | zone_movable_pfn[nid] = start_pfn + size_pages; | |
4715 | ||
4716 | /* | |
4717 | * Some kernelcore has been met, update counts and | |
4718 | * break if the kernelcore for this node has been | |
4719 | * satisified | |
4720 | */ | |
4721 | required_kernelcore -= min(required_kernelcore, | |
4722 | size_pages); | |
4723 | kernelcore_remaining -= size_pages; | |
4724 | if (!kernelcore_remaining) | |
4725 | break; | |
4726 | } | |
4727 | } | |
4728 | ||
4729 | /* | |
4730 | * If there is still required_kernelcore, we do another pass with one | |
4731 | * less node in the count. This will push zone_movable_pfn[nid] further | |
4732 | * along on the nodes that still have memory until kernelcore is | |
4733 | * satisified | |
4734 | */ | |
4735 | usable_nodes--; | |
4736 | if (usable_nodes && required_kernelcore > usable_nodes) | |
4737 | goto restart; | |
4738 | ||
4739 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ | |
4740 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
4741 | zone_movable_pfn[nid] = | |
4742 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); | |
4743 | ||
4744 | out: | |
4745 | /* restore the node_state */ | |
4746 | node_states[N_HIGH_MEMORY] = saved_node_state; | |
4747 | } | |
4748 | ||
4749 | /* Any regular memory on that node ? */ | |
4750 | static void check_for_regular_memory(pg_data_t *pgdat) | |
4751 | { | |
4752 | #ifdef CONFIG_HIGHMEM | |
4753 | enum zone_type zone_type; | |
4754 | ||
4755 | for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { | |
4756 | struct zone *zone = &pgdat->node_zones[zone_type]; | |
4757 | if (zone->present_pages) { | |
4758 | node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); | |
4759 | break; | |
4760 | } | |
4761 | } | |
4762 | #endif | |
4763 | } | |
4764 | ||
4765 | /** | |
4766 | * free_area_init_nodes - Initialise all pg_data_t and zone data | |
4767 | * @max_zone_pfn: an array of max PFNs for each zone | |
4768 | * | |
4769 | * This will call free_area_init_node() for each active node in the system. | |
4770 | * Using the page ranges provided by add_active_range(), the size of each | |
4771 | * zone in each node and their holes is calculated. If the maximum PFN | |
4772 | * between two adjacent zones match, it is assumed that the zone is empty. | |
4773 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed | |
4774 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone | |
4775 | * starts where the previous one ended. For example, ZONE_DMA32 starts | |
4776 | * at arch_max_dma_pfn. | |
4777 | */ | |
4778 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) | |
4779 | { | |
4780 | unsigned long start_pfn, end_pfn; | |
4781 | int i, nid; | |
4782 | ||
4783 | /* Record where the zone boundaries are */ | |
4784 | memset(arch_zone_lowest_possible_pfn, 0, | |
4785 | sizeof(arch_zone_lowest_possible_pfn)); | |
4786 | memset(arch_zone_highest_possible_pfn, 0, | |
4787 | sizeof(arch_zone_highest_possible_pfn)); | |
4788 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); | |
4789 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; | |
4790 | for (i = 1; i < MAX_NR_ZONES; i++) { | |
4791 | if (i == ZONE_MOVABLE) | |
4792 | continue; | |
4793 | arch_zone_lowest_possible_pfn[i] = | |
4794 | arch_zone_highest_possible_pfn[i-1]; | |
4795 | arch_zone_highest_possible_pfn[i] = | |
4796 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); | |
4797 | } | |
4798 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; | |
4799 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; | |
4800 | ||
4801 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ | |
4802 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); | |
4803 | find_zone_movable_pfns_for_nodes(); | |
4804 | ||
4805 | /* Print out the zone ranges */ | |
4806 | printk("Zone PFN ranges:\n"); | |
4807 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
4808 | if (i == ZONE_MOVABLE) | |
4809 | continue; | |
4810 | printk(" %-8s ", zone_names[i]); | |
4811 | if (arch_zone_lowest_possible_pfn[i] == | |
4812 | arch_zone_highest_possible_pfn[i]) | |
4813 | printk("empty\n"); | |
4814 | else | |
4815 | printk("%0#10lx -> %0#10lx\n", | |
4816 | arch_zone_lowest_possible_pfn[i], | |
4817 | arch_zone_highest_possible_pfn[i]); | |
4818 | } | |
4819 | ||
4820 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ | |
4821 | printk("Movable zone start PFN for each node\n"); | |
4822 | for (i = 0; i < MAX_NUMNODES; i++) { | |
4823 | if (zone_movable_pfn[i]) | |
4824 | printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); | |
4825 | } | |
4826 | ||
4827 | /* Print out the early_node_map[] */ | |
4828 | printk("Early memory PFN ranges\n"); | |
4829 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) | |
4830 | printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn); | |
4831 | ||
4832 | /* Initialise every node */ | |
4833 | mminit_verify_pageflags_layout(); | |
4834 | setup_nr_node_ids(); | |
4835 | for_each_online_node(nid) { | |
4836 | pg_data_t *pgdat = NODE_DATA(nid); | |
4837 | free_area_init_node(nid, NULL, | |
4838 | find_min_pfn_for_node(nid), NULL); | |
4839 | ||
4840 | /* Any memory on that node */ | |
4841 | if (pgdat->node_present_pages) | |
4842 | node_set_state(nid, N_HIGH_MEMORY); | |
4843 | check_for_regular_memory(pgdat); | |
4844 | } | |
4845 | } | |
4846 | ||
4847 | static int __init cmdline_parse_core(char *p, unsigned long *core) | |
4848 | { | |
4849 | unsigned long long coremem; | |
4850 | if (!p) | |
4851 | return -EINVAL; | |
4852 | ||
4853 | coremem = memparse(p, &p); | |
4854 | *core = coremem >> PAGE_SHIFT; | |
4855 | ||
4856 | /* Paranoid check that UL is enough for the coremem value */ | |
4857 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); | |
4858 | ||
4859 | return 0; | |
4860 | } | |
4861 | ||
4862 | /* | |
4863 | * kernelcore=size sets the amount of memory for use for allocations that | |
4864 | * cannot be reclaimed or migrated. | |
4865 | */ | |
4866 | static int __init cmdline_parse_kernelcore(char *p) | |
4867 | { | |
4868 | return cmdline_parse_core(p, &required_kernelcore); | |
4869 | } | |
4870 | ||
4871 | /* | |
4872 | * movablecore=size sets the amount of memory for use for allocations that | |
4873 | * can be reclaimed or migrated. | |
4874 | */ | |
4875 | static int __init cmdline_parse_movablecore(char *p) | |
4876 | { | |
4877 | return cmdline_parse_core(p, &required_movablecore); | |
4878 | } | |
4879 | ||
4880 | early_param("kernelcore", cmdline_parse_kernelcore); | |
4881 | early_param("movablecore", cmdline_parse_movablecore); | |
4882 | ||
4883 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4884 | ||
4885 | /** | |
4886 | * set_dma_reserve - set the specified number of pages reserved in the first zone | |
4887 | * @new_dma_reserve: The number of pages to mark reserved | |
4888 | * | |
4889 | * The per-cpu batchsize and zone watermarks are determined by present_pages. | |
4890 | * In the DMA zone, a significant percentage may be consumed by kernel image | |
4891 | * and other unfreeable allocations which can skew the watermarks badly. This | |
4892 | * function may optionally be used to account for unfreeable pages in the | |
4893 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and | |
4894 | * smaller per-cpu batchsize. | |
4895 | */ | |
4896 | void __init set_dma_reserve(unsigned long new_dma_reserve) | |
4897 | { | |
4898 | dma_reserve = new_dma_reserve; | |
4899 | } | |
4900 | ||
4901 | void __init free_area_init(unsigned long *zones_size) | |
4902 | { | |
4903 | free_area_init_node(0, zones_size, | |
4904 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); | |
4905 | } | |
4906 | ||
4907 | static int page_alloc_cpu_notify(struct notifier_block *self, | |
4908 | unsigned long action, void *hcpu) | |
4909 | { | |
4910 | int cpu = (unsigned long)hcpu; | |
4911 | ||
4912 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { | |
4913 | lru_add_drain_cpu(cpu); | |
4914 | drain_pages(cpu); | |
4915 | ||
4916 | /* | |
4917 | * Spill the event counters of the dead processor | |
4918 | * into the current processors event counters. | |
4919 | * This artificially elevates the count of the current | |
4920 | * processor. | |
4921 | */ | |
4922 | vm_events_fold_cpu(cpu); | |
4923 | ||
4924 | /* | |
4925 | * Zero the differential counters of the dead processor | |
4926 | * so that the vm statistics are consistent. | |
4927 | * | |
4928 | * This is only okay since the processor is dead and cannot | |
4929 | * race with what we are doing. | |
4930 | */ | |
4931 | refresh_cpu_vm_stats(cpu); | |
4932 | } | |
4933 | return NOTIFY_OK; | |
4934 | } | |
4935 | ||
4936 | void __init page_alloc_init(void) | |
4937 | { | |
4938 | hotcpu_notifier(page_alloc_cpu_notify, 0); | |
4939 | } | |
4940 | ||
4941 | /* | |
4942 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio | |
4943 | * or min_free_kbytes changes. | |
4944 | */ | |
4945 | static void calculate_totalreserve_pages(void) | |
4946 | { | |
4947 | struct pglist_data *pgdat; | |
4948 | unsigned long reserve_pages = 0; | |
4949 | enum zone_type i, j; | |
4950 | ||
4951 | for_each_online_pgdat(pgdat) { | |
4952 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
4953 | struct zone *zone = pgdat->node_zones + i; | |
4954 | unsigned long max = 0; | |
4955 | ||
4956 | /* Find valid and maximum lowmem_reserve in the zone */ | |
4957 | for (j = i; j < MAX_NR_ZONES; j++) { | |
4958 | if (zone->lowmem_reserve[j] > max) | |
4959 | max = zone->lowmem_reserve[j]; | |
4960 | } | |
4961 | ||
4962 | /* we treat the high watermark as reserved pages. */ | |
4963 | max += high_wmark_pages(zone); | |
4964 | ||
4965 | if (max > zone->present_pages) | |
4966 | max = zone->present_pages; | |
4967 | reserve_pages += max; | |
4968 | /* | |
4969 | * Lowmem reserves are not available to | |
4970 | * GFP_HIGHUSER page cache allocations and | |
4971 | * kswapd tries to balance zones to their high | |
4972 | * watermark. As a result, neither should be | |
4973 | * regarded as dirtyable memory, to prevent a | |
4974 | * situation where reclaim has to clean pages | |
4975 | * in order to balance the zones. | |
4976 | */ | |
4977 | zone->dirty_balance_reserve = max; | |
4978 | } | |
4979 | } | |
4980 | dirty_balance_reserve = reserve_pages; | |
4981 | totalreserve_pages = reserve_pages; | |
4982 | } | |
4983 | ||
4984 | /* | |
4985 | * setup_per_zone_lowmem_reserve - called whenever | |
4986 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone | |
4987 | * has a correct pages reserved value, so an adequate number of | |
4988 | * pages are left in the zone after a successful __alloc_pages(). | |
4989 | */ | |
4990 | static void setup_per_zone_lowmem_reserve(void) | |
4991 | { | |
4992 | struct pglist_data *pgdat; | |
4993 | enum zone_type j, idx; | |
4994 | ||
4995 | for_each_online_pgdat(pgdat) { | |
4996 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
4997 | struct zone *zone = pgdat->node_zones + j; | |
4998 | unsigned long present_pages = zone->present_pages; | |
4999 | ||
5000 | zone->lowmem_reserve[j] = 0; | |
5001 | ||
5002 | idx = j; | |
5003 | while (idx) { | |
5004 | struct zone *lower_zone; | |
5005 | ||
5006 | idx--; | |
5007 | ||
5008 | if (sysctl_lowmem_reserve_ratio[idx] < 1) | |
5009 | sysctl_lowmem_reserve_ratio[idx] = 1; | |
5010 | ||
5011 | lower_zone = pgdat->node_zones + idx; | |
5012 | lower_zone->lowmem_reserve[j] = present_pages / | |
5013 | sysctl_lowmem_reserve_ratio[idx]; | |
5014 | present_pages += lower_zone->present_pages; | |
5015 | } | |
5016 | } | |
5017 | } | |
5018 | ||
5019 | /* update totalreserve_pages */ | |
5020 | calculate_totalreserve_pages(); | |
5021 | } | |
5022 | ||
5023 | static void __setup_per_zone_wmarks(void) | |
5024 | { | |
5025 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
5026 | unsigned long lowmem_pages = 0; | |
5027 | struct zone *zone; | |
5028 | unsigned long flags; | |
5029 | ||
5030 | /* Calculate total number of !ZONE_HIGHMEM pages */ | |
5031 | for_each_zone(zone) { | |
5032 | if (!is_highmem(zone)) | |
5033 | lowmem_pages += zone->present_pages; | |
5034 | } | |
5035 | ||
5036 | for_each_zone(zone) { | |
5037 | u64 tmp; | |
5038 | ||
5039 | spin_lock_irqsave(&zone->lock, flags); | |
5040 | tmp = (u64)pages_min * zone->present_pages; | |
5041 | do_div(tmp, lowmem_pages); | |
5042 | if (is_highmem(zone)) { | |
5043 | /* | |
5044 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't | |
5045 | * need highmem pages, so cap pages_min to a small | |
5046 | * value here. | |
5047 | * | |
5048 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) | |
5049 | * deltas controls asynch page reclaim, and so should | |
5050 | * not be capped for highmem. | |
5051 | */ | |
5052 | int min_pages; | |
5053 | ||
5054 | min_pages = zone->present_pages / 1024; | |
5055 | if (min_pages < SWAP_CLUSTER_MAX) | |
5056 | min_pages = SWAP_CLUSTER_MAX; | |
5057 | if (min_pages > 128) | |
5058 | min_pages = 128; | |
5059 | zone->watermark[WMARK_MIN] = min_pages; | |
5060 | } else { | |
5061 | /* | |
5062 | * If it's a lowmem zone, reserve a number of pages | |
5063 | * proportionate to the zone's size. | |
5064 | */ | |
5065 | zone->watermark[WMARK_MIN] = tmp; | |
5066 | } | |
5067 | ||
5068 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); | |
5069 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); | |
5070 | setup_zone_migrate_reserve(zone); | |
5071 | spin_unlock_irqrestore(&zone->lock, flags); | |
5072 | } | |
5073 | ||
5074 | /* update totalreserve_pages */ | |
5075 | calculate_totalreserve_pages(); | |
5076 | } | |
5077 | ||
5078 | /** | |
5079 | * setup_per_zone_wmarks - called when min_free_kbytes changes | |
5080 | * or when memory is hot-{added|removed} | |
5081 | * | |
5082 | * Ensures that the watermark[min,low,high] values for each zone are set | |
5083 | * correctly with respect to min_free_kbytes. | |
5084 | */ | |
5085 | void setup_per_zone_wmarks(void) | |
5086 | { | |
5087 | mutex_lock(&zonelists_mutex); | |
5088 | __setup_per_zone_wmarks(); | |
5089 | mutex_unlock(&zonelists_mutex); | |
5090 | } | |
5091 | ||
5092 | /* | |
5093 | * The inactive anon list should be small enough that the VM never has to | |
5094 | * do too much work, but large enough that each inactive page has a chance | |
5095 | * to be referenced again before it is swapped out. | |
5096 | * | |
5097 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to | |
5098 | * INACTIVE_ANON pages on this zone's LRU, maintained by the | |
5099 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of | |
5100 | * the anonymous pages are kept on the inactive list. | |
5101 | * | |
5102 | * total target max | |
5103 | * memory ratio inactive anon | |
5104 | * ------------------------------------- | |
5105 | * 10MB 1 5MB | |
5106 | * 100MB 1 50MB | |
5107 | * 1GB 3 250MB | |
5108 | * 10GB 10 0.9GB | |
5109 | * 100GB 31 3GB | |
5110 | * 1TB 101 10GB | |
5111 | * 10TB 320 32GB | |
5112 | */ | |
5113 | static void __meminit calculate_zone_inactive_ratio(struct zone *zone) | |
5114 | { | |
5115 | unsigned int gb, ratio; | |
5116 | ||
5117 | /* Zone size in gigabytes */ | |
5118 | gb = zone->present_pages >> (30 - PAGE_SHIFT); | |
5119 | if (gb) | |
5120 | ratio = int_sqrt(10 * gb); | |
5121 | else | |
5122 | ratio = 1; | |
5123 | ||
5124 | zone->inactive_ratio = ratio; | |
5125 | } | |
5126 | ||
5127 | static void __meminit setup_per_zone_inactive_ratio(void) | |
5128 | { | |
5129 | struct zone *zone; | |
5130 | ||
5131 | for_each_zone(zone) | |
5132 | calculate_zone_inactive_ratio(zone); | |
5133 | } | |
5134 | ||
5135 | /* | |
5136 | * Initialise min_free_kbytes. | |
5137 | * | |
5138 | * For small machines we want it small (128k min). For large machines | |
5139 | * we want it large (64MB max). But it is not linear, because network | |
5140 | * bandwidth does not increase linearly with machine size. We use | |
5141 | * | |
5142 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: | |
5143 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) | |
5144 | * | |
5145 | * which yields | |
5146 | * | |
5147 | * 16MB: 512k | |
5148 | * 32MB: 724k | |
5149 | * 64MB: 1024k | |
5150 | * 128MB: 1448k | |
5151 | * 256MB: 2048k | |
5152 | * 512MB: 2896k | |
5153 | * 1024MB: 4096k | |
5154 | * 2048MB: 5792k | |
5155 | * 4096MB: 8192k | |
5156 | * 8192MB: 11584k | |
5157 | * 16384MB: 16384k | |
5158 | */ | |
5159 | int __meminit init_per_zone_wmark_min(void) | |
5160 | { | |
5161 | unsigned long lowmem_kbytes; | |
5162 | ||
5163 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
5164 | ||
5165 | min_free_kbytes = int_sqrt(lowmem_kbytes * 16); | |
5166 | if (min_free_kbytes < 128) | |
5167 | min_free_kbytes = 128; | |
5168 | if (min_free_kbytes > 65536) | |
5169 | min_free_kbytes = 65536; | |
5170 | setup_per_zone_wmarks(); | |
5171 | refresh_zone_stat_thresholds(); | |
5172 | setup_per_zone_lowmem_reserve(); | |
5173 | setup_per_zone_inactive_ratio(); | |
5174 | return 0; | |
5175 | } | |
5176 | module_init(init_per_zone_wmark_min) | |
5177 | ||
5178 | /* | |
5179 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so | |
5180 | * that we can call two helper functions whenever min_free_kbytes | |
5181 | * changes. | |
5182 | */ | |
5183 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, | |
5184 | void __user *buffer, size_t *length, loff_t *ppos) | |
5185 | { | |
5186 | proc_dointvec(table, write, buffer, length, ppos); | |
5187 | if (write) | |
5188 | setup_per_zone_wmarks(); | |
5189 | return 0; | |
5190 | } | |
5191 | ||
5192 | #ifdef CONFIG_NUMA | |
5193 | int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, | |
5194 | void __user *buffer, size_t *length, loff_t *ppos) | |
5195 | { | |
5196 | struct zone *zone; | |
5197 | int rc; | |
5198 | ||
5199 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5200 | if (rc) | |
5201 | return rc; | |
5202 | ||
5203 | for_each_zone(zone) | |
5204 | zone->min_unmapped_pages = (zone->present_pages * | |
5205 | sysctl_min_unmapped_ratio) / 100; | |
5206 | return 0; | |
5207 | } | |
5208 | ||
5209 | int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, | |
5210 | void __user *buffer, size_t *length, loff_t *ppos) | |
5211 | { | |
5212 | struct zone *zone; | |
5213 | int rc; | |
5214 | ||
5215 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5216 | if (rc) | |
5217 | return rc; | |
5218 | ||
5219 | for_each_zone(zone) | |
5220 | zone->min_slab_pages = (zone->present_pages * | |
5221 | sysctl_min_slab_ratio) / 100; | |
5222 | return 0; | |
5223 | } | |
5224 | #endif | |
5225 | ||
5226 | /* | |
5227 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
5228 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
5229 | * whenever sysctl_lowmem_reserve_ratio changes. | |
5230 | * | |
5231 | * The reserve ratio obviously has absolutely no relation with the | |
5232 | * minimum watermarks. The lowmem reserve ratio can only make sense | |
5233 | * if in function of the boot time zone sizes. | |
5234 | */ | |
5235 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, | |
5236 | void __user *buffer, size_t *length, loff_t *ppos) | |
5237 | { | |
5238 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5239 | setup_per_zone_lowmem_reserve(); | |
5240 | return 0; | |
5241 | } | |
5242 | ||
5243 | /* | |
5244 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each | |
5245 | * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist | |
5246 | * can have before it gets flushed back to buddy allocator. | |
5247 | */ | |
5248 | ||
5249 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, | |
5250 | void __user *buffer, size_t *length, loff_t *ppos) | |
5251 | { | |
5252 | struct zone *zone; | |
5253 | unsigned int cpu; | |
5254 | int ret; | |
5255 | ||
5256 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5257 | if (!write || (ret < 0)) | |
5258 | return ret; | |
5259 | for_each_populated_zone(zone) { | |
5260 | for_each_possible_cpu(cpu) { | |
5261 | unsigned long high; | |
5262 | high = zone->present_pages / percpu_pagelist_fraction; | |
5263 | setup_pagelist_highmark( | |
5264 | per_cpu_ptr(zone->pageset, cpu), high); | |
5265 | } | |
5266 | } | |
5267 | return 0; | |
5268 | } | |
5269 | ||
5270 | int hashdist = HASHDIST_DEFAULT; | |
5271 | ||
5272 | #ifdef CONFIG_NUMA | |
5273 | static int __init set_hashdist(char *str) | |
5274 | { | |
5275 | if (!str) | |
5276 | return 0; | |
5277 | hashdist = simple_strtoul(str, &str, 0); | |
5278 | return 1; | |
5279 | } | |
5280 | __setup("hashdist=", set_hashdist); | |
5281 | #endif | |
5282 | ||
5283 | /* | |
5284 | * allocate a large system hash table from bootmem | |
5285 | * - it is assumed that the hash table must contain an exact power-of-2 | |
5286 | * quantity of entries | |
5287 | * - limit is the number of hash buckets, not the total allocation size | |
5288 | */ | |
5289 | void *__init alloc_large_system_hash(const char *tablename, | |
5290 | unsigned long bucketsize, | |
5291 | unsigned long numentries, | |
5292 | int scale, | |
5293 | int flags, | |
5294 | unsigned int *_hash_shift, | |
5295 | unsigned int *_hash_mask, | |
5296 | unsigned long limit) | |
5297 | { | |
5298 | unsigned long long max = limit; | |
5299 | unsigned long log2qty, size; | |
5300 | void *table = NULL; | |
5301 | ||
5302 | /* allow the kernel cmdline to have a say */ | |
5303 | if (!numentries) { | |
5304 | /* round applicable memory size up to nearest megabyte */ | |
5305 | numentries = nr_kernel_pages; | |
5306 | numentries += (1UL << (20 - PAGE_SHIFT)) - 1; | |
5307 | numentries >>= 20 - PAGE_SHIFT; | |
5308 | numentries <<= 20 - PAGE_SHIFT; | |
5309 | ||
5310 | /* limit to 1 bucket per 2^scale bytes of low memory */ | |
5311 | if (scale > PAGE_SHIFT) | |
5312 | numentries >>= (scale - PAGE_SHIFT); | |
5313 | else | |
5314 | numentries <<= (PAGE_SHIFT - scale); | |
5315 | ||
5316 | /* Make sure we've got at least a 0-order allocation.. */ | |
5317 | if (unlikely(flags & HASH_SMALL)) { | |
5318 | /* Makes no sense without HASH_EARLY */ | |
5319 | WARN_ON(!(flags & HASH_EARLY)); | |
5320 | if (!(numentries >> *_hash_shift)) { | |
5321 | numentries = 1UL << *_hash_shift; | |
5322 | BUG_ON(!numentries); | |
5323 | } | |
5324 | } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) | |
5325 | numentries = PAGE_SIZE / bucketsize; | |
5326 | } | |
5327 | numentries = roundup_pow_of_two(numentries); | |
5328 | ||
5329 | /* limit allocation size to 1/16 total memory by default */ | |
5330 | if (max == 0) { | |
5331 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; | |
5332 | do_div(max, bucketsize); | |
5333 | } | |
5334 | max = min(max, 0x80000000ULL); | |
5335 | ||
5336 | if (numentries > max) | |
5337 | numentries = max; | |
5338 | ||
5339 | log2qty = ilog2(numentries); | |
5340 | ||
5341 | do { | |
5342 | size = bucketsize << log2qty; | |
5343 | if (flags & HASH_EARLY) | |
5344 | table = alloc_bootmem_nopanic(size); | |
5345 | else if (hashdist) | |
5346 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); | |
5347 | else { | |
5348 | /* | |
5349 | * If bucketsize is not a power-of-two, we may free | |
5350 | * some pages at the end of hash table which | |
5351 | * alloc_pages_exact() automatically does | |
5352 | */ | |
5353 | if (get_order(size) < MAX_ORDER) { | |
5354 | table = alloc_pages_exact(size, GFP_ATOMIC); | |
5355 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); | |
5356 | } | |
5357 | } | |
5358 | } while (!table && size > PAGE_SIZE && --log2qty); | |
5359 | ||
5360 | if (!table) | |
5361 | panic("Failed to allocate %s hash table\n", tablename); | |
5362 | ||
5363 | printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", | |
5364 | tablename, | |
5365 | (1UL << log2qty), | |
5366 | ilog2(size) - PAGE_SHIFT, | |
5367 | size); | |
5368 | ||
5369 | if (_hash_shift) | |
5370 | *_hash_shift = log2qty; | |
5371 | if (_hash_mask) | |
5372 | *_hash_mask = (1 << log2qty) - 1; | |
5373 | ||
5374 | return table; | |
5375 | } | |
5376 | ||
5377 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ | |
5378 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, | |
5379 | unsigned long pfn) | |
5380 | { | |
5381 | #ifdef CONFIG_SPARSEMEM | |
5382 | return __pfn_to_section(pfn)->pageblock_flags; | |
5383 | #else | |
5384 | return zone->pageblock_flags; | |
5385 | #endif /* CONFIG_SPARSEMEM */ | |
5386 | } | |
5387 | ||
5388 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) | |
5389 | { | |
5390 | #ifdef CONFIG_SPARSEMEM | |
5391 | pfn &= (PAGES_PER_SECTION-1); | |
5392 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
5393 | #else | |
5394 | pfn = pfn - zone->zone_start_pfn; | |
5395 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
5396 | #endif /* CONFIG_SPARSEMEM */ | |
5397 | } | |
5398 | ||
5399 | /** | |
5400 | * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages | |
5401 | * @page: The page within the block of interest | |
5402 | * @start_bitidx: The first bit of interest to retrieve | |
5403 | * @end_bitidx: The last bit of interest | |
5404 | * returns pageblock_bits flags | |
5405 | */ | |
5406 | unsigned long get_pageblock_flags_group(struct page *page, | |
5407 | int start_bitidx, int end_bitidx) | |
5408 | { | |
5409 | struct zone *zone; | |
5410 | unsigned long *bitmap; | |
5411 | unsigned long pfn, bitidx; | |
5412 | unsigned long flags = 0; | |
5413 | unsigned long value = 1; | |
5414 | ||
5415 | zone = page_zone(page); | |
5416 | pfn = page_to_pfn(page); | |
5417 | bitmap = get_pageblock_bitmap(zone, pfn); | |
5418 | bitidx = pfn_to_bitidx(zone, pfn); | |
5419 | ||
5420 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
5421 | if (test_bit(bitidx + start_bitidx, bitmap)) | |
5422 | flags |= value; | |
5423 | ||
5424 | return flags; | |
5425 | } | |
5426 | ||
5427 | /** | |
5428 | * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages | |
5429 | * @page: The page within the block of interest | |
5430 | * @start_bitidx: The first bit of interest | |
5431 | * @end_bitidx: The last bit of interest | |
5432 | * @flags: The flags to set | |
5433 | */ | |
5434 | void set_pageblock_flags_group(struct page *page, unsigned long flags, | |
5435 | int start_bitidx, int end_bitidx) | |
5436 | { | |
5437 | struct zone *zone; | |
5438 | unsigned long *bitmap; | |
5439 | unsigned long pfn, bitidx; | |
5440 | unsigned long value = 1; | |
5441 | ||
5442 | zone = page_zone(page); | |
5443 | pfn = page_to_pfn(page); | |
5444 | bitmap = get_pageblock_bitmap(zone, pfn); | |
5445 | bitidx = pfn_to_bitidx(zone, pfn); | |
5446 | VM_BUG_ON(pfn < zone->zone_start_pfn); | |
5447 | VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); | |
5448 | ||
5449 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
5450 | if (flags & value) | |
5451 | __set_bit(bitidx + start_bitidx, bitmap); | |
5452 | else | |
5453 | __clear_bit(bitidx + start_bitidx, bitmap); | |
5454 | } | |
5455 | ||
5456 | /* | |
5457 | * This is designed as sub function...plz see page_isolation.c also. | |
5458 | * set/clear page block's type to be ISOLATE. | |
5459 | * page allocater never alloc memory from ISOLATE block. | |
5460 | */ | |
5461 | ||
5462 | static int | |
5463 | __count_immobile_pages(struct zone *zone, struct page *page, int count) | |
5464 | { | |
5465 | unsigned long pfn, iter, found; | |
5466 | int mt; | |
5467 | ||
5468 | /* | |
5469 | * For avoiding noise data, lru_add_drain_all() should be called | |
5470 | * If ZONE_MOVABLE, the zone never contains immobile pages | |
5471 | */ | |
5472 | if (zone_idx(zone) == ZONE_MOVABLE) | |
5473 | return true; | |
5474 | mt = get_pageblock_migratetype(page); | |
5475 | if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) | |
5476 | return true; | |
5477 | ||
5478 | pfn = page_to_pfn(page); | |
5479 | for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { | |
5480 | unsigned long check = pfn + iter; | |
5481 | ||
5482 | if (!pfn_valid_within(check)) | |
5483 | continue; | |
5484 | ||
5485 | page = pfn_to_page(check); | |
5486 | if (!page_count(page)) { | |
5487 | if (PageBuddy(page)) | |
5488 | iter += (1 << page_order(page)) - 1; | |
5489 | continue; | |
5490 | } | |
5491 | if (!PageLRU(page)) | |
5492 | found++; | |
5493 | /* | |
5494 | * If there are RECLAIMABLE pages, we need to check it. | |
5495 | * But now, memory offline itself doesn't call shrink_slab() | |
5496 | * and it still to be fixed. | |
5497 | */ | |
5498 | /* | |
5499 | * If the page is not RAM, page_count()should be 0. | |
5500 | * we don't need more check. This is an _used_ not-movable page. | |
5501 | * | |
5502 | * The problematic thing here is PG_reserved pages. PG_reserved | |
5503 | * is set to both of a memory hole page and a _used_ kernel | |
5504 | * page at boot. | |
5505 | */ | |
5506 | if (found > count) | |
5507 | return false; | |
5508 | } | |
5509 | return true; | |
5510 | } | |
5511 | ||
5512 | bool is_pageblock_removable_nolock(struct page *page) | |
5513 | { | |
5514 | struct zone *zone; | |
5515 | unsigned long pfn; | |
5516 | ||
5517 | /* | |
5518 | * We have to be careful here because we are iterating over memory | |
5519 | * sections which are not zone aware so we might end up outside of | |
5520 | * the zone but still within the section. | |
5521 | * We have to take care about the node as well. If the node is offline | |
5522 | * its NODE_DATA will be NULL - see page_zone. | |
5523 | */ | |
5524 | if (!node_online(page_to_nid(page))) | |
5525 | return false; | |
5526 | ||
5527 | zone = page_zone(page); | |
5528 | pfn = page_to_pfn(page); | |
5529 | if (zone->zone_start_pfn > pfn || | |
5530 | zone->zone_start_pfn + zone->spanned_pages <= pfn) | |
5531 | return false; | |
5532 | ||
5533 | return __count_immobile_pages(zone, page, 0); | |
5534 | } | |
5535 | ||
5536 | int set_migratetype_isolate(struct page *page) | |
5537 | { | |
5538 | struct zone *zone; | |
5539 | unsigned long flags, pfn; | |
5540 | struct memory_isolate_notify arg; | |
5541 | int notifier_ret; | |
5542 | int ret = -EBUSY; | |
5543 | ||
5544 | zone = page_zone(page); | |
5545 | ||
5546 | spin_lock_irqsave(&zone->lock, flags); | |
5547 | ||
5548 | pfn = page_to_pfn(page); | |
5549 | arg.start_pfn = pfn; | |
5550 | arg.nr_pages = pageblock_nr_pages; | |
5551 | arg.pages_found = 0; | |
5552 | ||
5553 | /* | |
5554 | * It may be possible to isolate a pageblock even if the | |
5555 | * migratetype is not MIGRATE_MOVABLE. The memory isolation | |
5556 | * notifier chain is used by balloon drivers to return the | |
5557 | * number of pages in a range that are held by the balloon | |
5558 | * driver to shrink memory. If all the pages are accounted for | |
5559 | * by balloons, are free, or on the LRU, isolation can continue. | |
5560 | * Later, for example, when memory hotplug notifier runs, these | |
5561 | * pages reported as "can be isolated" should be isolated(freed) | |
5562 | * by the balloon driver through the memory notifier chain. | |
5563 | */ | |
5564 | notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); | |
5565 | notifier_ret = notifier_to_errno(notifier_ret); | |
5566 | if (notifier_ret) | |
5567 | goto out; | |
5568 | /* | |
5569 | * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. | |
5570 | * We just check MOVABLE pages. | |
5571 | */ | |
5572 | if (__count_immobile_pages(zone, page, arg.pages_found)) | |
5573 | ret = 0; | |
5574 | ||
5575 | /* | |
5576 | * immobile means "not-on-lru" paes. If immobile is larger than | |
5577 | * removable-by-driver pages reported by notifier, we'll fail. | |
5578 | */ | |
5579 | ||
5580 | out: | |
5581 | if (!ret) { | |
5582 | set_pageblock_migratetype(page, MIGRATE_ISOLATE); | |
5583 | move_freepages_block(zone, page, MIGRATE_ISOLATE); | |
5584 | } | |
5585 | ||
5586 | spin_unlock_irqrestore(&zone->lock, flags); | |
5587 | if (!ret) | |
5588 | drain_all_pages(); | |
5589 | return ret; | |
5590 | } | |
5591 | ||
5592 | void unset_migratetype_isolate(struct page *page, unsigned migratetype) | |
5593 | { | |
5594 | struct zone *zone; | |
5595 | unsigned long flags; | |
5596 | zone = page_zone(page); | |
5597 | spin_lock_irqsave(&zone->lock, flags); | |
5598 | if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) | |
5599 | goto out; | |
5600 | set_pageblock_migratetype(page, migratetype); | |
5601 | move_freepages_block(zone, page, migratetype); | |
5602 | out: | |
5603 | spin_unlock_irqrestore(&zone->lock, flags); | |
5604 | } | |
5605 | ||
5606 | #ifdef CONFIG_CMA | |
5607 | ||
5608 | static unsigned long pfn_max_align_down(unsigned long pfn) | |
5609 | { | |
5610 | return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
5611 | pageblock_nr_pages) - 1); | |
5612 | } | |
5613 | ||
5614 | static unsigned long pfn_max_align_up(unsigned long pfn) | |
5615 | { | |
5616 | return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
5617 | pageblock_nr_pages)); | |
5618 | } | |
5619 | ||
5620 | static struct page * | |
5621 | __alloc_contig_migrate_alloc(struct page *page, unsigned long private, | |
5622 | int **resultp) | |
5623 | { | |
5624 | return alloc_page(GFP_HIGHUSER_MOVABLE); | |
5625 | } | |
5626 | ||
5627 | /* [start, end) must belong to a single zone. */ | |
5628 | static int __alloc_contig_migrate_range(unsigned long start, unsigned long end) | |
5629 | { | |
5630 | /* This function is based on compact_zone() from compaction.c. */ | |
5631 | ||
5632 | unsigned long pfn = start; | |
5633 | unsigned int tries = 0; | |
5634 | int ret = 0; | |
5635 | ||
5636 | struct compact_control cc = { | |
5637 | .nr_migratepages = 0, | |
5638 | .order = -1, | |
5639 | .zone = page_zone(pfn_to_page(start)), | |
5640 | .sync = true, | |
5641 | }; | |
5642 | INIT_LIST_HEAD(&cc.migratepages); | |
5643 | ||
5644 | migrate_prep_local(); | |
5645 | ||
5646 | while (pfn < end || !list_empty(&cc.migratepages)) { | |
5647 | if (fatal_signal_pending(current)) { | |
5648 | ret = -EINTR; | |
5649 | break; | |
5650 | } | |
5651 | ||
5652 | if (list_empty(&cc.migratepages)) { | |
5653 | cc.nr_migratepages = 0; | |
5654 | pfn = isolate_migratepages_range(cc.zone, &cc, | |
5655 | pfn, end); | |
5656 | if (!pfn) { | |
5657 | ret = -EINTR; | |
5658 | break; | |
5659 | } | |
5660 | tries = 0; | |
5661 | } else if (++tries == 5) { | |
5662 | ret = ret < 0 ? ret : -EBUSY; | |
5663 | break; | |
5664 | } | |
5665 | ||
5666 | ret = migrate_pages(&cc.migratepages, | |
5667 | __alloc_contig_migrate_alloc, | |
5668 | 0, false, true); | |
5669 | } | |
5670 | ||
5671 | putback_lru_pages(&cc.migratepages); | |
5672 | return ret > 0 ? 0 : ret; | |
5673 | } | |
5674 | ||
5675 | /** | |
5676 | * alloc_contig_range() -- tries to allocate given range of pages | |
5677 | * @start: start PFN to allocate | |
5678 | * @end: one-past-the-last PFN to allocate | |
5679 | * @migratetype: migratetype of the underlaying pageblocks (either | |
5680 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks | |
5681 | * in range must have the same migratetype and it must | |
5682 | * be either of the two. | |
5683 | * | |
5684 | * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES | |
5685 | * aligned, however it's the caller's responsibility to guarantee that | |
5686 | * we are the only thread that changes migrate type of pageblocks the | |
5687 | * pages fall in. | |
5688 | * | |
5689 | * The PFN range must belong to a single zone. | |
5690 | * | |
5691 | * Returns zero on success or negative error code. On success all | |
5692 | * pages which PFN is in [start, end) are allocated for the caller and | |
5693 | * need to be freed with free_contig_range(). | |
5694 | */ | |
5695 | int alloc_contig_range(unsigned long start, unsigned long end, | |
5696 | unsigned migratetype) | |
5697 | { | |
5698 | struct zone *zone = page_zone(pfn_to_page(start)); | |
5699 | unsigned long outer_start, outer_end; | |
5700 | int ret = 0, order; | |
5701 | ||
5702 | /* | |
5703 | * What we do here is we mark all pageblocks in range as | |
5704 | * MIGRATE_ISOLATE. Because pageblock and max order pages may | |
5705 | * have different sizes, and due to the way page allocator | |
5706 | * work, we align the range to biggest of the two pages so | |
5707 | * that page allocator won't try to merge buddies from | |
5708 | * different pageblocks and change MIGRATE_ISOLATE to some | |
5709 | * other migration type. | |
5710 | * | |
5711 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we | |
5712 | * migrate the pages from an unaligned range (ie. pages that | |
5713 | * we are interested in). This will put all the pages in | |
5714 | * range back to page allocator as MIGRATE_ISOLATE. | |
5715 | * | |
5716 | * When this is done, we take the pages in range from page | |
5717 | * allocator removing them from the buddy system. This way | |
5718 | * page allocator will never consider using them. | |
5719 | * | |
5720 | * This lets us mark the pageblocks back as | |
5721 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the | |
5722 | * aligned range but not in the unaligned, original range are | |
5723 | * put back to page allocator so that buddy can use them. | |
5724 | */ | |
5725 | ||
5726 | ret = start_isolate_page_range(pfn_max_align_down(start), | |
5727 | pfn_max_align_up(end), migratetype); | |
5728 | if (ret) | |
5729 | goto done; | |
5730 | ||
5731 | ret = __alloc_contig_migrate_range(start, end); | |
5732 | if (ret) | |
5733 | goto done; | |
5734 | ||
5735 | /* | |
5736 | * Pages from [start, end) are within a MAX_ORDER_NR_PAGES | |
5737 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's | |
5738 | * more, all pages in [start, end) are free in page allocator. | |
5739 | * What we are going to do is to allocate all pages from | |
5740 | * [start, end) (that is remove them from page allocator). | |
5741 | * | |
5742 | * The only problem is that pages at the beginning and at the | |
5743 | * end of interesting range may be not aligned with pages that | |
5744 | * page allocator holds, ie. they can be part of higher order | |
5745 | * pages. Because of this, we reserve the bigger range and | |
5746 | * once this is done free the pages we are not interested in. | |
5747 | * | |
5748 | * We don't have to hold zone->lock here because the pages are | |
5749 | * isolated thus they won't get removed from buddy. | |
5750 | */ | |
5751 | ||
5752 | lru_add_drain_all(); | |
5753 | drain_all_pages(); | |
5754 | ||
5755 | order = 0; | |
5756 | outer_start = start; | |
5757 | while (!PageBuddy(pfn_to_page(outer_start))) { | |
5758 | if (++order >= MAX_ORDER) { | |
5759 | ret = -EBUSY; | |
5760 | goto done; | |
5761 | } | |
5762 | outer_start &= ~0UL << order; | |
5763 | } | |
5764 | ||
5765 | /* Make sure the range is really isolated. */ | |
5766 | if (test_pages_isolated(outer_start, end)) { | |
5767 | pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n", | |
5768 | outer_start, end); | |
5769 | ret = -EBUSY; | |
5770 | goto done; | |
5771 | } | |
5772 | ||
5773 | outer_end = isolate_freepages_range(outer_start, end); | |
5774 | if (!outer_end) { | |
5775 | ret = -EBUSY; | |
5776 | goto done; | |
5777 | } | |
5778 | ||
5779 | /* Free head and tail (if any) */ | |
5780 | if (start != outer_start) | |
5781 | free_contig_range(outer_start, start - outer_start); | |
5782 | if (end != outer_end) | |
5783 | free_contig_range(end, outer_end - end); | |
5784 | ||
5785 | done: | |
5786 | undo_isolate_page_range(pfn_max_align_down(start), | |
5787 | pfn_max_align_up(end), migratetype); | |
5788 | return ret; | |
5789 | } | |
5790 | ||
5791 | void free_contig_range(unsigned long pfn, unsigned nr_pages) | |
5792 | { | |
5793 | for (; nr_pages--; ++pfn) | |
5794 | __free_page(pfn_to_page(pfn)); | |
5795 | } | |
5796 | #endif | |
5797 | ||
5798 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
5799 | /* | |
5800 | * All pages in the range must be isolated before calling this. | |
5801 | */ | |
5802 | void | |
5803 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) | |
5804 | { | |
5805 | struct page *page; | |
5806 | struct zone *zone; | |
5807 | int order, i; | |
5808 | unsigned long pfn; | |
5809 | unsigned long flags; | |
5810 | /* find the first valid pfn */ | |
5811 | for (pfn = start_pfn; pfn < end_pfn; pfn++) | |
5812 | if (pfn_valid(pfn)) | |
5813 | break; | |
5814 | if (pfn == end_pfn) | |
5815 | return; | |
5816 | zone = page_zone(pfn_to_page(pfn)); | |
5817 | spin_lock_irqsave(&zone->lock, flags); | |
5818 | pfn = start_pfn; | |
5819 | while (pfn < end_pfn) { | |
5820 | if (!pfn_valid(pfn)) { | |
5821 | pfn++; | |
5822 | continue; | |
5823 | } | |
5824 | page = pfn_to_page(pfn); | |
5825 | BUG_ON(page_count(page)); | |
5826 | BUG_ON(!PageBuddy(page)); | |
5827 | order = page_order(page); | |
5828 | #ifdef CONFIG_DEBUG_VM | |
5829 | printk(KERN_INFO "remove from free list %lx %d %lx\n", | |
5830 | pfn, 1 << order, end_pfn); | |
5831 | #endif | |
5832 | list_del(&page->lru); | |
5833 | rmv_page_order(page); | |
5834 | zone->free_area[order].nr_free--; | |
5835 | __mod_zone_page_state(zone, NR_FREE_PAGES, | |
5836 | - (1UL << order)); | |
5837 | for (i = 0; i < (1 << order); i++) | |
5838 | SetPageReserved((page+i)); | |
5839 | pfn += (1 << order); | |
5840 | } | |
5841 | spin_unlock_irqrestore(&zone->lock, flags); | |
5842 | } | |
5843 | #endif | |
5844 | ||
5845 | #ifdef CONFIG_MEMORY_FAILURE | |
5846 | bool is_free_buddy_page(struct page *page) | |
5847 | { | |
5848 | struct zone *zone = page_zone(page); | |
5849 | unsigned long pfn = page_to_pfn(page); | |
5850 | unsigned long flags; | |
5851 | int order; | |
5852 | ||
5853 | spin_lock_irqsave(&zone->lock, flags); | |
5854 | for (order = 0; order < MAX_ORDER; order++) { | |
5855 | struct page *page_head = page - (pfn & ((1 << order) - 1)); | |
5856 | ||
5857 | if (PageBuddy(page_head) && page_order(page_head) >= order) | |
5858 | break; | |
5859 | } | |
5860 | spin_unlock_irqrestore(&zone->lock, flags); | |
5861 | ||
5862 | return order < MAX_ORDER; | |
5863 | } | |
5864 | #endif | |
5865 | ||
5866 | static struct trace_print_flags pageflag_names[] = { | |
5867 | {1UL << PG_locked, "locked" }, | |
5868 | {1UL << PG_error, "error" }, | |
5869 | {1UL << PG_referenced, "referenced" }, | |
5870 | {1UL << PG_uptodate, "uptodate" }, | |
5871 | {1UL << PG_dirty, "dirty" }, | |
5872 | {1UL << PG_lru, "lru" }, | |
5873 | {1UL << PG_active, "active" }, | |
5874 | {1UL << PG_slab, "slab" }, | |
5875 | {1UL << PG_owner_priv_1, "owner_priv_1" }, | |
5876 | {1UL << PG_arch_1, "arch_1" }, | |
5877 | {1UL << PG_reserved, "reserved" }, | |
5878 | {1UL << PG_private, "private" }, | |
5879 | {1UL << PG_private_2, "private_2" }, | |
5880 | {1UL << PG_writeback, "writeback" }, | |
5881 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
5882 | {1UL << PG_head, "head" }, | |
5883 | {1UL << PG_tail, "tail" }, | |
5884 | #else | |
5885 | {1UL << PG_compound, "compound" }, | |
5886 | #endif | |
5887 | {1UL << PG_swapcache, "swapcache" }, | |
5888 | {1UL << PG_mappedtodisk, "mappedtodisk" }, | |
5889 | {1UL << PG_reclaim, "reclaim" }, | |
5890 | {1UL << PG_swapbacked, "swapbacked" }, | |
5891 | {1UL << PG_unevictable, "unevictable" }, | |
5892 | #ifdef CONFIG_MMU | |
5893 | {1UL << PG_mlocked, "mlocked" }, | |
5894 | #endif | |
5895 | #ifdef CONFIG_ARCH_USES_PG_UNCACHED | |
5896 | {1UL << PG_uncached, "uncached" }, | |
5897 | #endif | |
5898 | #ifdef CONFIG_MEMORY_FAILURE | |
5899 | {1UL << PG_hwpoison, "hwpoison" }, | |
5900 | #endif | |
5901 | {-1UL, NULL }, | |
5902 | }; | |
5903 | ||
5904 | static void dump_page_flags(unsigned long flags) | |
5905 | { | |
5906 | const char *delim = ""; | |
5907 | unsigned long mask; | |
5908 | int i; | |
5909 | ||
5910 | printk(KERN_ALERT "page flags: %#lx(", flags); | |
5911 | ||
5912 | /* remove zone id */ | |
5913 | flags &= (1UL << NR_PAGEFLAGS) - 1; | |
5914 | ||
5915 | for (i = 0; pageflag_names[i].name && flags; i++) { | |
5916 | ||
5917 | mask = pageflag_names[i].mask; | |
5918 | if ((flags & mask) != mask) | |
5919 | continue; | |
5920 | ||
5921 | flags &= ~mask; | |
5922 | printk("%s%s", delim, pageflag_names[i].name); | |
5923 | delim = "|"; | |
5924 | } | |
5925 | ||
5926 | /* check for left over flags */ | |
5927 | if (flags) | |
5928 | printk("%s%#lx", delim, flags); | |
5929 | ||
5930 | printk(")\n"); | |
5931 | } | |
5932 | ||
5933 | void dump_page(struct page *page) | |
5934 | { | |
5935 | printk(KERN_ALERT | |
5936 | "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", | |
5937 | page, atomic_read(&page->_count), page_mapcount(page), | |
5938 | page->mapping, page->index); | |
5939 | dump_page_flags(page->flags); | |
5940 | mem_cgroup_print_bad_page(page); | |
5941 | } |