]>
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/compiler.h> | |
25 | #include <linux/kernel.h> | |
26 | #include <linux/module.h> | |
27 | #include <linux/suspend.h> | |
28 | #include <linux/pagevec.h> | |
29 | #include <linux/blkdev.h> | |
30 | #include <linux/slab.h> | |
31 | #include <linux/oom.h> | |
32 | #include <linux/notifier.h> | |
33 | #include <linux/topology.h> | |
34 | #include <linux/sysctl.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
37 | #include <linux/memory_hotplug.h> | |
38 | #include <linux/nodemask.h> | |
39 | #include <linux/vmalloc.h> | |
40 | #include <linux/mempolicy.h> | |
41 | #include <linux/stop_machine.h> | |
42 | #include <linux/sort.h> | |
43 | #include <linux/pfn.h> | |
44 | #include <linux/backing-dev.h> | |
45 | #include <linux/fault-inject.h> | |
46 | #include <linux/page-isolation.h> | |
47 | #include <linux/page_cgroup.h> | |
48 | #include <linux/debugobjects.h> | |
49 | #include <linux/kmemleak.h> | |
50 | ||
51 | #include <asm/tlbflush.h> | |
52 | #include <asm/div64.h> | |
53 | #include "internal.h" | |
54 | ||
55 | /* | |
56 | * Array of node states. | |
57 | */ | |
58 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { | |
59 | [N_POSSIBLE] = NODE_MASK_ALL, | |
60 | [N_ONLINE] = { { [0] = 1UL } }, | |
61 | #ifndef CONFIG_NUMA | |
62 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
63 | #ifdef CONFIG_HIGHMEM | |
64 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
65 | #endif | |
66 | [N_CPU] = { { [0] = 1UL } }, | |
67 | #endif /* NUMA */ | |
68 | }; | |
69 | EXPORT_SYMBOL(node_states); | |
70 | ||
71 | unsigned long totalram_pages __read_mostly; | |
72 | unsigned long totalreserve_pages __read_mostly; | |
73 | unsigned long highest_memmap_pfn __read_mostly; | |
74 | int percpu_pagelist_fraction; | |
75 | ||
76 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
77 | int pageblock_order __read_mostly; | |
78 | #endif | |
79 | ||
80 | static void __free_pages_ok(struct page *page, unsigned int order); | |
81 | ||
82 | /* | |
83 | * results with 256, 32 in the lowmem_reserve sysctl: | |
84 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
85 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
86 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
87 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
88 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA | |
89 | * | |
90 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
91 | * don't need any ZONE_NORMAL reservation | |
92 | */ | |
93 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { | |
94 | #ifdef CONFIG_ZONE_DMA | |
95 | 256, | |
96 | #endif | |
97 | #ifdef CONFIG_ZONE_DMA32 | |
98 | 256, | |
99 | #endif | |
100 | #ifdef CONFIG_HIGHMEM | |
101 | 32, | |
102 | #endif | |
103 | 32, | |
104 | }; | |
105 | ||
106 | EXPORT_SYMBOL(totalram_pages); | |
107 | ||
108 | static char * const zone_names[MAX_NR_ZONES] = { | |
109 | #ifdef CONFIG_ZONE_DMA | |
110 | "DMA", | |
111 | #endif | |
112 | #ifdef CONFIG_ZONE_DMA32 | |
113 | "DMA32", | |
114 | #endif | |
115 | "Normal", | |
116 | #ifdef CONFIG_HIGHMEM | |
117 | "HighMem", | |
118 | #endif | |
119 | "Movable", | |
120 | }; | |
121 | ||
122 | int min_free_kbytes = 1024; | |
123 | ||
124 | unsigned long __meminitdata nr_kernel_pages; | |
125 | unsigned long __meminitdata nr_all_pages; | |
126 | static unsigned long __meminitdata dma_reserve; | |
127 | ||
128 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP | |
129 | /* | |
130 | * MAX_ACTIVE_REGIONS determines the maximum number of distinct | |
131 | * ranges of memory (RAM) that may be registered with add_active_range(). | |
132 | * Ranges passed to add_active_range() will be merged if possible | |
133 | * so the number of times add_active_range() can be called is | |
134 | * related to the number of nodes and the number of holes | |
135 | */ | |
136 | #ifdef CONFIG_MAX_ACTIVE_REGIONS | |
137 | /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ | |
138 | #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS | |
139 | #else | |
140 | #if MAX_NUMNODES >= 32 | |
141 | /* If there can be many nodes, allow up to 50 holes per node */ | |
142 | #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) | |
143 | #else | |
144 | /* By default, allow up to 256 distinct regions */ | |
145 | #define MAX_ACTIVE_REGIONS 256 | |
146 | #endif | |
147 | #endif | |
148 | ||
149 | static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; | |
150 | static int __meminitdata nr_nodemap_entries; | |
151 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; | |
152 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; | |
153 | static unsigned long __initdata required_kernelcore; | |
154 | static unsigned long __initdata required_movablecore; | |
155 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; | |
156 | ||
157 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
158 | int movable_zone; | |
159 | EXPORT_SYMBOL(movable_zone); | |
160 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ | |
161 | ||
162 | #if MAX_NUMNODES > 1 | |
163 | int nr_node_ids __read_mostly = MAX_NUMNODES; | |
164 | int nr_online_nodes __read_mostly = 1; | |
165 | EXPORT_SYMBOL(nr_node_ids); | |
166 | EXPORT_SYMBOL(nr_online_nodes); | |
167 | #endif | |
168 | ||
169 | int page_group_by_mobility_disabled __read_mostly; | |
170 | ||
171 | static void set_pageblock_migratetype(struct page *page, int migratetype) | |
172 | { | |
173 | ||
174 | if (unlikely(page_group_by_mobility_disabled)) | |
175 | migratetype = MIGRATE_UNMOVABLE; | |
176 | ||
177 | set_pageblock_flags_group(page, (unsigned long)migratetype, | |
178 | PB_migrate, PB_migrate_end); | |
179 | } | |
180 | ||
181 | #ifdef CONFIG_DEBUG_VM | |
182 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) | |
183 | { | |
184 | int ret = 0; | |
185 | unsigned seq; | |
186 | unsigned long pfn = page_to_pfn(page); | |
187 | ||
188 | do { | |
189 | seq = zone_span_seqbegin(zone); | |
190 | if (pfn >= zone->zone_start_pfn + zone->spanned_pages) | |
191 | ret = 1; | |
192 | else if (pfn < zone->zone_start_pfn) | |
193 | ret = 1; | |
194 | } while (zone_span_seqretry(zone, seq)); | |
195 | ||
196 | return ret; | |
197 | } | |
198 | ||
199 | static int page_is_consistent(struct zone *zone, struct page *page) | |
200 | { | |
201 | if (!pfn_valid_within(page_to_pfn(page))) | |
202 | return 0; | |
203 | if (zone != page_zone(page)) | |
204 | return 0; | |
205 | ||
206 | return 1; | |
207 | } | |
208 | /* | |
209 | * Temporary debugging check for pages not lying within a given zone. | |
210 | */ | |
211 | static int bad_range(struct zone *zone, struct page *page) | |
212 | { | |
213 | if (page_outside_zone_boundaries(zone, page)) | |
214 | return 1; | |
215 | if (!page_is_consistent(zone, page)) | |
216 | return 1; | |
217 | ||
218 | return 0; | |
219 | } | |
220 | #else | |
221 | static inline int bad_range(struct zone *zone, struct page *page) | |
222 | { | |
223 | return 0; | |
224 | } | |
225 | #endif | |
226 | ||
227 | static void bad_page(struct page *page) | |
228 | { | |
229 | static unsigned long resume; | |
230 | static unsigned long nr_shown; | |
231 | static unsigned long nr_unshown; | |
232 | ||
233 | /* | |
234 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
235 | * or allow a steady drip of one report per second. | |
236 | */ | |
237 | if (nr_shown == 60) { | |
238 | if (time_before(jiffies, resume)) { | |
239 | nr_unshown++; | |
240 | goto out; | |
241 | } | |
242 | if (nr_unshown) { | |
243 | printk(KERN_ALERT | |
244 | "BUG: Bad page state: %lu messages suppressed\n", | |
245 | nr_unshown); | |
246 | nr_unshown = 0; | |
247 | } | |
248 | nr_shown = 0; | |
249 | } | |
250 | if (nr_shown++ == 0) | |
251 | resume = jiffies + 60 * HZ; | |
252 | ||
253 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", | |
254 | current->comm, page_to_pfn(page)); | |
255 | printk(KERN_ALERT | |
256 | "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", | |
257 | page, (void *)page->flags, page_count(page), | |
258 | page_mapcount(page), page->mapping, page->index); | |
259 | ||
260 | dump_stack(); | |
261 | out: | |
262 | /* Leave bad fields for debug, except PageBuddy could make trouble */ | |
263 | __ClearPageBuddy(page); | |
264 | add_taint(TAINT_BAD_PAGE); | |
265 | } | |
266 | ||
267 | /* | |
268 | * Higher-order pages are called "compound pages". They are structured thusly: | |
269 | * | |
270 | * The first PAGE_SIZE page is called the "head page". | |
271 | * | |
272 | * The remaining PAGE_SIZE pages are called "tail pages". | |
273 | * | |
274 | * All pages have PG_compound set. All pages have their ->private pointing at | |
275 | * the head page (even the head page has this). | |
276 | * | |
277 | * The first tail page's ->lru.next holds the address of the compound page's | |
278 | * put_page() function. Its ->lru.prev holds the order of allocation. | |
279 | * This usage means that zero-order pages may not be compound. | |
280 | */ | |
281 | ||
282 | static void free_compound_page(struct page *page) | |
283 | { | |
284 | __free_pages_ok(page, compound_order(page)); | |
285 | } | |
286 | ||
287 | void prep_compound_page(struct page *page, unsigned long order) | |
288 | { | |
289 | int i; | |
290 | int nr_pages = 1 << order; | |
291 | ||
292 | set_compound_page_dtor(page, free_compound_page); | |
293 | set_compound_order(page, order); | |
294 | __SetPageHead(page); | |
295 | for (i = 1; i < nr_pages; i++) { | |
296 | struct page *p = page + i; | |
297 | ||
298 | __SetPageTail(p); | |
299 | p->first_page = page; | |
300 | } | |
301 | } | |
302 | ||
303 | static int destroy_compound_page(struct page *page, unsigned long order) | |
304 | { | |
305 | int i; | |
306 | int nr_pages = 1 << order; | |
307 | int bad = 0; | |
308 | ||
309 | if (unlikely(compound_order(page) != order) || | |
310 | unlikely(!PageHead(page))) { | |
311 | bad_page(page); | |
312 | bad++; | |
313 | } | |
314 | ||
315 | __ClearPageHead(page); | |
316 | ||
317 | for (i = 1; i < nr_pages; i++) { | |
318 | struct page *p = page + i; | |
319 | ||
320 | if (unlikely(!PageTail(p) || (p->first_page != page))) { | |
321 | bad_page(page); | |
322 | bad++; | |
323 | } | |
324 | __ClearPageTail(p); | |
325 | } | |
326 | ||
327 | return bad; | |
328 | } | |
329 | ||
330 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) | |
331 | { | |
332 | int i; | |
333 | ||
334 | /* | |
335 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO | |
336 | * and __GFP_HIGHMEM from hard or soft interrupt context. | |
337 | */ | |
338 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); | |
339 | for (i = 0; i < (1 << order); i++) | |
340 | clear_highpage(page + i); | |
341 | } | |
342 | ||
343 | static inline void set_page_order(struct page *page, int order) | |
344 | { | |
345 | set_page_private(page, order); | |
346 | __SetPageBuddy(page); | |
347 | } | |
348 | ||
349 | static inline void rmv_page_order(struct page *page) | |
350 | { | |
351 | __ClearPageBuddy(page); | |
352 | set_page_private(page, 0); | |
353 | } | |
354 | ||
355 | /* | |
356 | * Locate the struct page for both the matching buddy in our | |
357 | * pair (buddy1) and the combined O(n+1) page they form (page). | |
358 | * | |
359 | * 1) Any buddy B1 will have an order O twin B2 which satisfies | |
360 | * the following equation: | |
361 | * B2 = B1 ^ (1 << O) | |
362 | * For example, if the starting buddy (buddy2) is #8 its order | |
363 | * 1 buddy is #10: | |
364 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 | |
365 | * | |
366 | * 2) Any buddy B will have an order O+1 parent P which | |
367 | * satisfies the following equation: | |
368 | * P = B & ~(1 << O) | |
369 | * | |
370 | * Assumption: *_mem_map is contiguous at least up to MAX_ORDER | |
371 | */ | |
372 | static inline struct page * | |
373 | __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) | |
374 | { | |
375 | unsigned long buddy_idx = page_idx ^ (1 << order); | |
376 | ||
377 | return page + (buddy_idx - page_idx); | |
378 | } | |
379 | ||
380 | static inline unsigned long | |
381 | __find_combined_index(unsigned long page_idx, unsigned int order) | |
382 | { | |
383 | return (page_idx & ~(1 << order)); | |
384 | } | |
385 | ||
386 | /* | |
387 | * This function checks whether a page is free && is the buddy | |
388 | * we can do coalesce a page and its buddy if | |
389 | * (a) the buddy is not in a hole && | |
390 | * (b) the buddy is in the buddy system && | |
391 | * (c) a page and its buddy have the same order && | |
392 | * (d) a page and its buddy are in the same zone. | |
393 | * | |
394 | * For recording whether a page is in the buddy system, we use PG_buddy. | |
395 | * Setting, clearing, and testing PG_buddy is serialized by zone->lock. | |
396 | * | |
397 | * For recording page's order, we use page_private(page). | |
398 | */ | |
399 | static inline int page_is_buddy(struct page *page, struct page *buddy, | |
400 | int order) | |
401 | { | |
402 | if (!pfn_valid_within(page_to_pfn(buddy))) | |
403 | return 0; | |
404 | ||
405 | if (page_zone_id(page) != page_zone_id(buddy)) | |
406 | return 0; | |
407 | ||
408 | if (PageBuddy(buddy) && page_order(buddy) == order) { | |
409 | VM_BUG_ON(page_count(buddy) != 0); | |
410 | return 1; | |
411 | } | |
412 | return 0; | |
413 | } | |
414 | ||
415 | /* | |
416 | * Freeing function for a buddy system allocator. | |
417 | * | |
418 | * The concept of a buddy system is to maintain direct-mapped table | |
419 | * (containing bit values) for memory blocks of various "orders". | |
420 | * The bottom level table contains the map for the smallest allocatable | |
421 | * units of memory (here, pages), and each level above it describes | |
422 | * pairs of units from the levels below, hence, "buddies". | |
423 | * At a high level, all that happens here is marking the table entry | |
424 | * at the bottom level available, and propagating the changes upward | |
425 | * as necessary, plus some accounting needed to play nicely with other | |
426 | * parts of the VM system. | |
427 | * At each level, we keep a list of pages, which are heads of continuous | |
428 | * free pages of length of (1 << order) and marked with PG_buddy. Page's | |
429 | * order is recorded in page_private(page) field. | |
430 | * So when we are allocating or freeing one, we can derive the state of the | |
431 | * other. That is, if we allocate a small block, and both were | |
432 | * free, the remainder of the region must be split into blocks. | |
433 | * If a block is freed, and its buddy is also free, then this | |
434 | * triggers coalescing into a block of larger size. | |
435 | * | |
436 | * -- wli | |
437 | */ | |
438 | ||
439 | static inline void __free_one_page(struct page *page, | |
440 | struct zone *zone, unsigned int order, | |
441 | int migratetype) | |
442 | { | |
443 | unsigned long page_idx; | |
444 | ||
445 | if (unlikely(PageCompound(page))) | |
446 | if (unlikely(destroy_compound_page(page, order))) | |
447 | return; | |
448 | ||
449 | VM_BUG_ON(migratetype == -1); | |
450 | ||
451 | page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); | |
452 | ||
453 | VM_BUG_ON(page_idx & ((1 << order) - 1)); | |
454 | VM_BUG_ON(bad_range(zone, page)); | |
455 | ||
456 | while (order < MAX_ORDER-1) { | |
457 | unsigned long combined_idx; | |
458 | struct page *buddy; | |
459 | ||
460 | buddy = __page_find_buddy(page, page_idx, order); | |
461 | if (!page_is_buddy(page, buddy, order)) | |
462 | break; | |
463 | ||
464 | /* Our buddy is free, merge with it and move up one order. */ | |
465 | list_del(&buddy->lru); | |
466 | zone->free_area[order].nr_free--; | |
467 | rmv_page_order(buddy); | |
468 | combined_idx = __find_combined_index(page_idx, order); | |
469 | page = page + (combined_idx - page_idx); | |
470 | page_idx = combined_idx; | |
471 | order++; | |
472 | } | |
473 | set_page_order(page, order); | |
474 | list_add(&page->lru, | |
475 | &zone->free_area[order].free_list[migratetype]); | |
476 | zone->free_area[order].nr_free++; | |
477 | } | |
478 | ||
479 | #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT | |
480 | /* | |
481 | * free_page_mlock() -- clean up attempts to free and mlocked() page. | |
482 | * Page should not be on lru, so no need to fix that up. | |
483 | * free_pages_check() will verify... | |
484 | */ | |
485 | static inline void free_page_mlock(struct page *page) | |
486 | { | |
487 | __ClearPageMlocked(page); | |
488 | __dec_zone_page_state(page, NR_MLOCK); | |
489 | __count_vm_event(UNEVICTABLE_MLOCKFREED); | |
490 | } | |
491 | #else | |
492 | static void free_page_mlock(struct page *page) { } | |
493 | #endif | |
494 | ||
495 | static inline int free_pages_check(struct page *page) | |
496 | { | |
497 | if (unlikely(page_mapcount(page) | | |
498 | (page->mapping != NULL) | | |
499 | (atomic_read(&page->_count) != 0) | | |
500 | (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { | |
501 | bad_page(page); | |
502 | return 1; | |
503 | } | |
504 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | |
505 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
506 | return 0; | |
507 | } | |
508 | ||
509 | /* | |
510 | * Frees a list of pages. | |
511 | * Assumes all pages on list are in same zone, and of same order. | |
512 | * count is the number of pages to free. | |
513 | * | |
514 | * If the zone was previously in an "all pages pinned" state then look to | |
515 | * see if this freeing clears that state. | |
516 | * | |
517 | * And clear the zone's pages_scanned counter, to hold off the "all pages are | |
518 | * pinned" detection logic. | |
519 | */ | |
520 | static void free_pages_bulk(struct zone *zone, int count, | |
521 | struct list_head *list, int order) | |
522 | { | |
523 | spin_lock(&zone->lock); | |
524 | zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); | |
525 | zone->pages_scanned = 0; | |
526 | ||
527 | __mod_zone_page_state(zone, NR_FREE_PAGES, count << order); | |
528 | while (count--) { | |
529 | struct page *page; | |
530 | ||
531 | VM_BUG_ON(list_empty(list)); | |
532 | page = list_entry(list->prev, struct page, lru); | |
533 | /* have to delete it as __free_one_page list manipulates */ | |
534 | list_del(&page->lru); | |
535 | __free_one_page(page, zone, order, page_private(page)); | |
536 | } | |
537 | spin_unlock(&zone->lock); | |
538 | } | |
539 | ||
540 | static void free_one_page(struct zone *zone, struct page *page, int order, | |
541 | int migratetype) | |
542 | { | |
543 | spin_lock(&zone->lock); | |
544 | zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); | |
545 | zone->pages_scanned = 0; | |
546 | ||
547 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); | |
548 | __free_one_page(page, zone, order, migratetype); | |
549 | spin_unlock(&zone->lock); | |
550 | } | |
551 | ||
552 | static void __free_pages_ok(struct page *page, unsigned int order) | |
553 | { | |
554 | unsigned long flags; | |
555 | int i; | |
556 | int bad = 0; | |
557 | int clearMlocked = PageMlocked(page); | |
558 | ||
559 | for (i = 0 ; i < (1 << order) ; ++i) | |
560 | bad += free_pages_check(page + i); | |
561 | if (bad) | |
562 | return; | |
563 | ||
564 | if (!PageHighMem(page)) { | |
565 | debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); | |
566 | debug_check_no_obj_freed(page_address(page), | |
567 | PAGE_SIZE << order); | |
568 | } | |
569 | arch_free_page(page, order); | |
570 | kernel_map_pages(page, 1 << order, 0); | |
571 | ||
572 | local_irq_save(flags); | |
573 | if (unlikely(clearMlocked)) | |
574 | free_page_mlock(page); | |
575 | __count_vm_events(PGFREE, 1 << order); | |
576 | free_one_page(page_zone(page), page, order, | |
577 | get_pageblock_migratetype(page)); | |
578 | local_irq_restore(flags); | |
579 | } | |
580 | ||
581 | /* | |
582 | * permit the bootmem allocator to evade page validation on high-order frees | |
583 | */ | |
584 | void __meminit __free_pages_bootmem(struct page *page, unsigned int order) | |
585 | { | |
586 | if (order == 0) { | |
587 | __ClearPageReserved(page); | |
588 | set_page_count(page, 0); | |
589 | set_page_refcounted(page); | |
590 | __free_page(page); | |
591 | } else { | |
592 | int loop; | |
593 | ||
594 | prefetchw(page); | |
595 | for (loop = 0; loop < BITS_PER_LONG; loop++) { | |
596 | struct page *p = &page[loop]; | |
597 | ||
598 | if (loop + 1 < BITS_PER_LONG) | |
599 | prefetchw(p + 1); | |
600 | __ClearPageReserved(p); | |
601 | set_page_count(p, 0); | |
602 | } | |
603 | ||
604 | set_page_refcounted(page); | |
605 | __free_pages(page, order); | |
606 | } | |
607 | } | |
608 | ||
609 | ||
610 | /* | |
611 | * The order of subdivision here is critical for the IO subsystem. | |
612 | * Please do not alter this order without good reasons and regression | |
613 | * testing. Specifically, as large blocks of memory are subdivided, | |
614 | * the order in which smaller blocks are delivered depends on the order | |
615 | * they're subdivided in this function. This is the primary factor | |
616 | * influencing the order in which pages are delivered to the IO | |
617 | * subsystem according to empirical testing, and this is also justified | |
618 | * by considering the behavior of a buddy system containing a single | |
619 | * large block of memory acted on by a series of small allocations. | |
620 | * This behavior is a critical factor in sglist merging's success. | |
621 | * | |
622 | * -- wli | |
623 | */ | |
624 | static inline void expand(struct zone *zone, struct page *page, | |
625 | int low, int high, struct free_area *area, | |
626 | int migratetype) | |
627 | { | |
628 | unsigned long size = 1 << high; | |
629 | ||
630 | while (high > low) { | |
631 | area--; | |
632 | high--; | |
633 | size >>= 1; | |
634 | VM_BUG_ON(bad_range(zone, &page[size])); | |
635 | list_add(&page[size].lru, &area->free_list[migratetype]); | |
636 | area->nr_free++; | |
637 | set_page_order(&page[size], high); | |
638 | } | |
639 | } | |
640 | ||
641 | /* | |
642 | * This page is about to be returned from the page allocator | |
643 | */ | |
644 | static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) | |
645 | { | |
646 | if (unlikely(page_mapcount(page) | | |
647 | (page->mapping != NULL) | | |
648 | (atomic_read(&page->_count) != 0) | | |
649 | (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { | |
650 | bad_page(page); | |
651 | return 1; | |
652 | } | |
653 | ||
654 | set_page_private(page, 0); | |
655 | set_page_refcounted(page); | |
656 | ||
657 | arch_alloc_page(page, order); | |
658 | kernel_map_pages(page, 1 << order, 1); | |
659 | ||
660 | if (gfp_flags & __GFP_ZERO) | |
661 | prep_zero_page(page, order, gfp_flags); | |
662 | ||
663 | if (order && (gfp_flags & __GFP_COMP)) | |
664 | prep_compound_page(page, order); | |
665 | ||
666 | return 0; | |
667 | } | |
668 | ||
669 | /* | |
670 | * Go through the free lists for the given migratetype and remove | |
671 | * the smallest available page from the freelists | |
672 | */ | |
673 | static inline | |
674 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, | |
675 | int migratetype) | |
676 | { | |
677 | unsigned int current_order; | |
678 | struct free_area * area; | |
679 | struct page *page; | |
680 | ||
681 | /* Find a page of the appropriate size in the preferred list */ | |
682 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { | |
683 | area = &(zone->free_area[current_order]); | |
684 | if (list_empty(&area->free_list[migratetype])) | |
685 | continue; | |
686 | ||
687 | page = list_entry(area->free_list[migratetype].next, | |
688 | struct page, lru); | |
689 | list_del(&page->lru); | |
690 | rmv_page_order(page); | |
691 | area->nr_free--; | |
692 | expand(zone, page, order, current_order, area, migratetype); | |
693 | return page; | |
694 | } | |
695 | ||
696 | return NULL; | |
697 | } | |
698 | ||
699 | ||
700 | /* | |
701 | * This array describes the order lists are fallen back to when | |
702 | * the free lists for the desirable migrate type are depleted | |
703 | */ | |
704 | static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { | |
705 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
706 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
707 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
708 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ | |
709 | }; | |
710 | ||
711 | /* | |
712 | * Move the free pages in a range to the free lists of the requested type. | |
713 | * Note that start_page and end_pages are not aligned on a pageblock | |
714 | * boundary. If alignment is required, use move_freepages_block() | |
715 | */ | |
716 | static int move_freepages(struct zone *zone, | |
717 | struct page *start_page, struct page *end_page, | |
718 | int migratetype) | |
719 | { | |
720 | struct page *page; | |
721 | unsigned long order; | |
722 | int pages_moved = 0; | |
723 | ||
724 | #ifndef CONFIG_HOLES_IN_ZONE | |
725 | /* | |
726 | * page_zone is not safe to call in this context when | |
727 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant | |
728 | * anyway as we check zone boundaries in move_freepages_block(). | |
729 | * Remove at a later date when no bug reports exist related to | |
730 | * grouping pages by mobility | |
731 | */ | |
732 | BUG_ON(page_zone(start_page) != page_zone(end_page)); | |
733 | #endif | |
734 | ||
735 | for (page = start_page; page <= end_page;) { | |
736 | /* Make sure we are not inadvertently changing nodes */ | |
737 | VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); | |
738 | ||
739 | if (!pfn_valid_within(page_to_pfn(page))) { | |
740 | page++; | |
741 | continue; | |
742 | } | |
743 | ||
744 | if (!PageBuddy(page)) { | |
745 | page++; | |
746 | continue; | |
747 | } | |
748 | ||
749 | order = page_order(page); | |
750 | list_del(&page->lru); | |
751 | list_add(&page->lru, | |
752 | &zone->free_area[order].free_list[migratetype]); | |
753 | page += 1 << order; | |
754 | pages_moved += 1 << order; | |
755 | } | |
756 | ||
757 | return pages_moved; | |
758 | } | |
759 | ||
760 | static int move_freepages_block(struct zone *zone, struct page *page, | |
761 | int migratetype) | |
762 | { | |
763 | unsigned long start_pfn, end_pfn; | |
764 | struct page *start_page, *end_page; | |
765 | ||
766 | start_pfn = page_to_pfn(page); | |
767 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); | |
768 | start_page = pfn_to_page(start_pfn); | |
769 | end_page = start_page + pageblock_nr_pages - 1; | |
770 | end_pfn = start_pfn + pageblock_nr_pages - 1; | |
771 | ||
772 | /* Do not cross zone boundaries */ | |
773 | if (start_pfn < zone->zone_start_pfn) | |
774 | start_page = page; | |
775 | if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) | |
776 | return 0; | |
777 | ||
778 | return move_freepages(zone, start_page, end_page, migratetype); | |
779 | } | |
780 | ||
781 | /* Remove an element from the buddy allocator from the fallback list */ | |
782 | static inline struct page * | |
783 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) | |
784 | { | |
785 | struct free_area * area; | |
786 | int current_order; | |
787 | struct page *page; | |
788 | int migratetype, i; | |
789 | ||
790 | /* Find the largest possible block of pages in the other list */ | |
791 | for (current_order = MAX_ORDER-1; current_order >= order; | |
792 | --current_order) { | |
793 | for (i = 0; i < MIGRATE_TYPES - 1; i++) { | |
794 | migratetype = fallbacks[start_migratetype][i]; | |
795 | ||
796 | /* MIGRATE_RESERVE handled later if necessary */ | |
797 | if (migratetype == MIGRATE_RESERVE) | |
798 | continue; | |
799 | ||
800 | area = &(zone->free_area[current_order]); | |
801 | if (list_empty(&area->free_list[migratetype])) | |
802 | continue; | |
803 | ||
804 | page = list_entry(area->free_list[migratetype].next, | |
805 | struct page, lru); | |
806 | area->nr_free--; | |
807 | ||
808 | /* | |
809 | * If breaking a large block of pages, move all free | |
810 | * pages to the preferred allocation list. If falling | |
811 | * back for a reclaimable kernel allocation, be more | |
812 | * agressive about taking ownership of free pages | |
813 | */ | |
814 | if (unlikely(current_order >= (pageblock_order >> 1)) || | |
815 | start_migratetype == MIGRATE_RECLAIMABLE) { | |
816 | unsigned long pages; | |
817 | pages = move_freepages_block(zone, page, | |
818 | start_migratetype); | |
819 | ||
820 | /* Claim the whole block if over half of it is free */ | |
821 | if (pages >= (1 << (pageblock_order-1))) | |
822 | set_pageblock_migratetype(page, | |
823 | start_migratetype); | |
824 | ||
825 | migratetype = start_migratetype; | |
826 | } | |
827 | ||
828 | /* Remove the page from the freelists */ | |
829 | list_del(&page->lru); | |
830 | rmv_page_order(page); | |
831 | ||
832 | if (current_order == pageblock_order) | |
833 | set_pageblock_migratetype(page, | |
834 | start_migratetype); | |
835 | ||
836 | expand(zone, page, order, current_order, area, migratetype); | |
837 | return page; | |
838 | } | |
839 | } | |
840 | ||
841 | return NULL; | |
842 | } | |
843 | ||
844 | /* | |
845 | * Do the hard work of removing an element from the buddy allocator. | |
846 | * Call me with the zone->lock already held. | |
847 | */ | |
848 | static struct page *__rmqueue(struct zone *zone, unsigned int order, | |
849 | int migratetype) | |
850 | { | |
851 | struct page *page; | |
852 | ||
853 | retry_reserve: | |
854 | page = __rmqueue_smallest(zone, order, migratetype); | |
855 | ||
856 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { | |
857 | page = __rmqueue_fallback(zone, order, migratetype); | |
858 | ||
859 | /* | |
860 | * Use MIGRATE_RESERVE rather than fail an allocation. goto | |
861 | * is used because __rmqueue_smallest is an inline function | |
862 | * and we want just one call site | |
863 | */ | |
864 | if (!page) { | |
865 | migratetype = MIGRATE_RESERVE; | |
866 | goto retry_reserve; | |
867 | } | |
868 | } | |
869 | ||
870 | return page; | |
871 | } | |
872 | ||
873 | /* | |
874 | * Obtain a specified number of elements from the buddy allocator, all under | |
875 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
876 | * Returns the number of new pages which were placed at *list. | |
877 | */ | |
878 | static int rmqueue_bulk(struct zone *zone, unsigned int order, | |
879 | unsigned long count, struct list_head *list, | |
880 | int migratetype) | |
881 | { | |
882 | int i; | |
883 | ||
884 | spin_lock(&zone->lock); | |
885 | for (i = 0; i < count; ++i) { | |
886 | struct page *page = __rmqueue(zone, order, migratetype); | |
887 | if (unlikely(page == NULL)) | |
888 | break; | |
889 | ||
890 | /* | |
891 | * Split buddy pages returned by expand() are received here | |
892 | * in physical page order. The page is added to the callers and | |
893 | * list and the list head then moves forward. From the callers | |
894 | * perspective, the linked list is ordered by page number in | |
895 | * some conditions. This is useful for IO devices that can | |
896 | * merge IO requests if the physical pages are ordered | |
897 | * properly. | |
898 | */ | |
899 | list_add(&page->lru, list); | |
900 | set_page_private(page, migratetype); | |
901 | list = &page->lru; | |
902 | } | |
903 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); | |
904 | spin_unlock(&zone->lock); | |
905 | return i; | |
906 | } | |
907 | ||
908 | #ifdef CONFIG_NUMA | |
909 | /* | |
910 | * Called from the vmstat counter updater to drain pagesets of this | |
911 | * currently executing processor on remote nodes after they have | |
912 | * expired. | |
913 | * | |
914 | * Note that this function must be called with the thread pinned to | |
915 | * a single processor. | |
916 | */ | |
917 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) | |
918 | { | |
919 | unsigned long flags; | |
920 | int to_drain; | |
921 | ||
922 | local_irq_save(flags); | |
923 | if (pcp->count >= pcp->batch) | |
924 | to_drain = pcp->batch; | |
925 | else | |
926 | to_drain = pcp->count; | |
927 | free_pages_bulk(zone, to_drain, &pcp->list, 0); | |
928 | pcp->count -= to_drain; | |
929 | local_irq_restore(flags); | |
930 | } | |
931 | #endif | |
932 | ||
933 | /* | |
934 | * Drain pages of the indicated processor. | |
935 | * | |
936 | * The processor must either be the current processor and the | |
937 | * thread pinned to the current processor or a processor that | |
938 | * is not online. | |
939 | */ | |
940 | static void drain_pages(unsigned int cpu) | |
941 | { | |
942 | unsigned long flags; | |
943 | struct zone *zone; | |
944 | ||
945 | for_each_populated_zone(zone) { | |
946 | struct per_cpu_pageset *pset; | |
947 | struct per_cpu_pages *pcp; | |
948 | ||
949 | pset = zone_pcp(zone, cpu); | |
950 | ||
951 | pcp = &pset->pcp; | |
952 | local_irq_save(flags); | |
953 | free_pages_bulk(zone, pcp->count, &pcp->list, 0); | |
954 | pcp->count = 0; | |
955 | local_irq_restore(flags); | |
956 | } | |
957 | } | |
958 | ||
959 | /* | |
960 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
961 | */ | |
962 | void drain_local_pages(void *arg) | |
963 | { | |
964 | drain_pages(smp_processor_id()); | |
965 | } | |
966 | ||
967 | /* | |
968 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator | |
969 | */ | |
970 | void drain_all_pages(void) | |
971 | { | |
972 | on_each_cpu(drain_local_pages, NULL, 1); | |
973 | } | |
974 | ||
975 | #ifdef CONFIG_HIBERNATION | |
976 | ||
977 | void mark_free_pages(struct zone *zone) | |
978 | { | |
979 | unsigned long pfn, max_zone_pfn; | |
980 | unsigned long flags; | |
981 | int order, t; | |
982 | struct list_head *curr; | |
983 | ||
984 | if (!zone->spanned_pages) | |
985 | return; | |
986 | ||
987 | spin_lock_irqsave(&zone->lock, flags); | |
988 | ||
989 | max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; | |
990 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) | |
991 | if (pfn_valid(pfn)) { | |
992 | struct page *page = pfn_to_page(pfn); | |
993 | ||
994 | if (!swsusp_page_is_forbidden(page)) | |
995 | swsusp_unset_page_free(page); | |
996 | } | |
997 | ||
998 | for_each_migratetype_order(order, t) { | |
999 | list_for_each(curr, &zone->free_area[order].free_list[t]) { | |
1000 | unsigned long i; | |
1001 | ||
1002 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); | |
1003 | for (i = 0; i < (1UL << order); i++) | |
1004 | swsusp_set_page_free(pfn_to_page(pfn + i)); | |
1005 | } | |
1006 | } | |
1007 | spin_unlock_irqrestore(&zone->lock, flags); | |
1008 | } | |
1009 | #endif /* CONFIG_PM */ | |
1010 | ||
1011 | /* | |
1012 | * Free a 0-order page | |
1013 | */ | |
1014 | static void free_hot_cold_page(struct page *page, int cold) | |
1015 | { | |
1016 | struct zone *zone = page_zone(page); | |
1017 | struct per_cpu_pages *pcp; | |
1018 | unsigned long flags; | |
1019 | int clearMlocked = PageMlocked(page); | |
1020 | ||
1021 | if (PageAnon(page)) | |
1022 | page->mapping = NULL; | |
1023 | if (free_pages_check(page)) | |
1024 | return; | |
1025 | ||
1026 | if (!PageHighMem(page)) { | |
1027 | debug_check_no_locks_freed(page_address(page), PAGE_SIZE); | |
1028 | debug_check_no_obj_freed(page_address(page), PAGE_SIZE); | |
1029 | } | |
1030 | arch_free_page(page, 0); | |
1031 | kernel_map_pages(page, 1, 0); | |
1032 | ||
1033 | pcp = &zone_pcp(zone, get_cpu())->pcp; | |
1034 | set_page_private(page, get_pageblock_migratetype(page)); | |
1035 | local_irq_save(flags); | |
1036 | if (unlikely(clearMlocked)) | |
1037 | free_page_mlock(page); | |
1038 | __count_vm_event(PGFREE); | |
1039 | ||
1040 | if (cold) | |
1041 | list_add_tail(&page->lru, &pcp->list); | |
1042 | else | |
1043 | list_add(&page->lru, &pcp->list); | |
1044 | pcp->count++; | |
1045 | if (pcp->count >= pcp->high) { | |
1046 | free_pages_bulk(zone, pcp->batch, &pcp->list, 0); | |
1047 | pcp->count -= pcp->batch; | |
1048 | } | |
1049 | local_irq_restore(flags); | |
1050 | put_cpu(); | |
1051 | } | |
1052 | ||
1053 | void free_hot_page(struct page *page) | |
1054 | { | |
1055 | free_hot_cold_page(page, 0); | |
1056 | } | |
1057 | ||
1058 | void free_cold_page(struct page *page) | |
1059 | { | |
1060 | free_hot_cold_page(page, 1); | |
1061 | } | |
1062 | ||
1063 | /* | |
1064 | * split_page takes a non-compound higher-order page, and splits it into | |
1065 | * n (1<<order) sub-pages: page[0..n] | |
1066 | * Each sub-page must be freed individually. | |
1067 | * | |
1068 | * Note: this is probably too low level an operation for use in drivers. | |
1069 | * Please consult with lkml before using this in your driver. | |
1070 | */ | |
1071 | void split_page(struct page *page, unsigned int order) | |
1072 | { | |
1073 | int i; | |
1074 | ||
1075 | VM_BUG_ON(PageCompound(page)); | |
1076 | VM_BUG_ON(!page_count(page)); | |
1077 | for (i = 1; i < (1 << order); i++) | |
1078 | set_page_refcounted(page + i); | |
1079 | } | |
1080 | ||
1081 | /* | |
1082 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But | |
1083 | * we cheat by calling it from here, in the order > 0 path. Saves a branch | |
1084 | * or two. | |
1085 | */ | |
1086 | static inline | |
1087 | struct page *buffered_rmqueue(struct zone *preferred_zone, | |
1088 | struct zone *zone, int order, gfp_t gfp_flags, | |
1089 | int migratetype) | |
1090 | { | |
1091 | unsigned long flags; | |
1092 | struct page *page; | |
1093 | int cold = !!(gfp_flags & __GFP_COLD); | |
1094 | int cpu; | |
1095 | ||
1096 | again: | |
1097 | cpu = get_cpu(); | |
1098 | if (likely(order == 0)) { | |
1099 | struct per_cpu_pages *pcp; | |
1100 | ||
1101 | pcp = &zone_pcp(zone, cpu)->pcp; | |
1102 | local_irq_save(flags); | |
1103 | if (!pcp->count) { | |
1104 | pcp->count = rmqueue_bulk(zone, 0, | |
1105 | pcp->batch, &pcp->list, migratetype); | |
1106 | if (unlikely(!pcp->count)) | |
1107 | goto failed; | |
1108 | } | |
1109 | ||
1110 | /* Find a page of the appropriate migrate type */ | |
1111 | if (cold) { | |
1112 | list_for_each_entry_reverse(page, &pcp->list, lru) | |
1113 | if (page_private(page) == migratetype) | |
1114 | break; | |
1115 | } else { | |
1116 | list_for_each_entry(page, &pcp->list, lru) | |
1117 | if (page_private(page) == migratetype) | |
1118 | break; | |
1119 | } | |
1120 | ||
1121 | /* Allocate more to the pcp list if necessary */ | |
1122 | if (unlikely(&page->lru == &pcp->list)) { | |
1123 | pcp->count += rmqueue_bulk(zone, 0, | |
1124 | pcp->batch, &pcp->list, migratetype); | |
1125 | page = list_entry(pcp->list.next, struct page, lru); | |
1126 | } | |
1127 | ||
1128 | list_del(&page->lru); | |
1129 | pcp->count--; | |
1130 | } else { | |
1131 | spin_lock_irqsave(&zone->lock, flags); | |
1132 | page = __rmqueue(zone, order, migratetype); | |
1133 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); | |
1134 | spin_unlock(&zone->lock); | |
1135 | if (!page) | |
1136 | goto failed; | |
1137 | } | |
1138 | ||
1139 | __count_zone_vm_events(PGALLOC, zone, 1 << order); | |
1140 | zone_statistics(preferred_zone, zone); | |
1141 | local_irq_restore(flags); | |
1142 | put_cpu(); | |
1143 | ||
1144 | VM_BUG_ON(bad_range(zone, page)); | |
1145 | if (prep_new_page(page, order, gfp_flags)) | |
1146 | goto again; | |
1147 | return page; | |
1148 | ||
1149 | failed: | |
1150 | local_irq_restore(flags); | |
1151 | put_cpu(); | |
1152 | return NULL; | |
1153 | } | |
1154 | ||
1155 | /* The ALLOC_WMARK bits are used as an index to zone->watermark */ | |
1156 | #define ALLOC_WMARK_MIN WMARK_MIN | |
1157 | #define ALLOC_WMARK_LOW WMARK_LOW | |
1158 | #define ALLOC_WMARK_HIGH WMARK_HIGH | |
1159 | #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ | |
1160 | ||
1161 | /* Mask to get the watermark bits */ | |
1162 | #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) | |
1163 | ||
1164 | #define ALLOC_HARDER 0x10 /* try to alloc harder */ | |
1165 | #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ | |
1166 | #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ | |
1167 | ||
1168 | #ifdef CONFIG_FAIL_PAGE_ALLOC | |
1169 | ||
1170 | static struct fail_page_alloc_attr { | |
1171 | struct fault_attr attr; | |
1172 | ||
1173 | u32 ignore_gfp_highmem; | |
1174 | u32 ignore_gfp_wait; | |
1175 | u32 min_order; | |
1176 | ||
1177 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
1178 | ||
1179 | struct dentry *ignore_gfp_highmem_file; | |
1180 | struct dentry *ignore_gfp_wait_file; | |
1181 | struct dentry *min_order_file; | |
1182 | ||
1183 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
1184 | ||
1185 | } fail_page_alloc = { | |
1186 | .attr = FAULT_ATTR_INITIALIZER, | |
1187 | .ignore_gfp_wait = 1, | |
1188 | .ignore_gfp_highmem = 1, | |
1189 | .min_order = 1, | |
1190 | }; | |
1191 | ||
1192 | static int __init setup_fail_page_alloc(char *str) | |
1193 | { | |
1194 | return setup_fault_attr(&fail_page_alloc.attr, str); | |
1195 | } | |
1196 | __setup("fail_page_alloc=", setup_fail_page_alloc); | |
1197 | ||
1198 | static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1199 | { | |
1200 | if (order < fail_page_alloc.min_order) | |
1201 | return 0; | |
1202 | if (gfp_mask & __GFP_NOFAIL) | |
1203 | return 0; | |
1204 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) | |
1205 | return 0; | |
1206 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) | |
1207 | return 0; | |
1208 | ||
1209 | return should_fail(&fail_page_alloc.attr, 1 << order); | |
1210 | } | |
1211 | ||
1212 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
1213 | ||
1214 | static int __init fail_page_alloc_debugfs(void) | |
1215 | { | |
1216 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
1217 | struct dentry *dir; | |
1218 | int err; | |
1219 | ||
1220 | err = init_fault_attr_dentries(&fail_page_alloc.attr, | |
1221 | "fail_page_alloc"); | |
1222 | if (err) | |
1223 | return err; | |
1224 | dir = fail_page_alloc.attr.dentries.dir; | |
1225 | ||
1226 | fail_page_alloc.ignore_gfp_wait_file = | |
1227 | debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
1228 | &fail_page_alloc.ignore_gfp_wait); | |
1229 | ||
1230 | fail_page_alloc.ignore_gfp_highmem_file = | |
1231 | debugfs_create_bool("ignore-gfp-highmem", mode, dir, | |
1232 | &fail_page_alloc.ignore_gfp_highmem); | |
1233 | fail_page_alloc.min_order_file = | |
1234 | debugfs_create_u32("min-order", mode, dir, | |
1235 | &fail_page_alloc.min_order); | |
1236 | ||
1237 | if (!fail_page_alloc.ignore_gfp_wait_file || | |
1238 | !fail_page_alloc.ignore_gfp_highmem_file || | |
1239 | !fail_page_alloc.min_order_file) { | |
1240 | err = -ENOMEM; | |
1241 | debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); | |
1242 | debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); | |
1243 | debugfs_remove(fail_page_alloc.min_order_file); | |
1244 | cleanup_fault_attr_dentries(&fail_page_alloc.attr); | |
1245 | } | |
1246 | ||
1247 | return err; | |
1248 | } | |
1249 | ||
1250 | late_initcall(fail_page_alloc_debugfs); | |
1251 | ||
1252 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
1253 | ||
1254 | #else /* CONFIG_FAIL_PAGE_ALLOC */ | |
1255 | ||
1256 | static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1257 | { | |
1258 | return 0; | |
1259 | } | |
1260 | ||
1261 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ | |
1262 | ||
1263 | /* | |
1264 | * Return 1 if free pages are above 'mark'. This takes into account the order | |
1265 | * of the allocation. | |
1266 | */ | |
1267 | int zone_watermark_ok(struct zone *z, int order, unsigned long mark, | |
1268 | int classzone_idx, int alloc_flags) | |
1269 | { | |
1270 | /* free_pages my go negative - that's OK */ | |
1271 | long min = mark; | |
1272 | long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; | |
1273 | int o; | |
1274 | ||
1275 | if (alloc_flags & ALLOC_HIGH) | |
1276 | min -= min / 2; | |
1277 | if (alloc_flags & ALLOC_HARDER) | |
1278 | min -= min / 4; | |
1279 | ||
1280 | if (free_pages <= min + z->lowmem_reserve[classzone_idx]) | |
1281 | return 0; | |
1282 | for (o = 0; o < order; o++) { | |
1283 | /* At the next order, this order's pages become unavailable */ | |
1284 | free_pages -= z->free_area[o].nr_free << o; | |
1285 | ||
1286 | /* Require fewer higher order pages to be free */ | |
1287 | min >>= 1; | |
1288 | ||
1289 | if (free_pages <= min) | |
1290 | return 0; | |
1291 | } | |
1292 | return 1; | |
1293 | } | |
1294 | ||
1295 | #ifdef CONFIG_NUMA | |
1296 | /* | |
1297 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to | |
1298 | * skip over zones that are not allowed by the cpuset, or that have | |
1299 | * been recently (in last second) found to be nearly full. See further | |
1300 | * comments in mmzone.h. Reduces cache footprint of zonelist scans | |
1301 | * that have to skip over a lot of full or unallowed zones. | |
1302 | * | |
1303 | * If the zonelist cache is present in the passed in zonelist, then | |
1304 | * returns a pointer to the allowed node mask (either the current | |
1305 | * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) | |
1306 | * | |
1307 | * If the zonelist cache is not available for this zonelist, does | |
1308 | * nothing and returns NULL. | |
1309 | * | |
1310 | * If the fullzones BITMAP in the zonelist cache is stale (more than | |
1311 | * a second since last zap'd) then we zap it out (clear its bits.) | |
1312 | * | |
1313 | * We hold off even calling zlc_setup, until after we've checked the | |
1314 | * first zone in the zonelist, on the theory that most allocations will | |
1315 | * be satisfied from that first zone, so best to examine that zone as | |
1316 | * quickly as we can. | |
1317 | */ | |
1318 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1319 | { | |
1320 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1321 | nodemask_t *allowednodes; /* zonelist_cache approximation */ | |
1322 | ||
1323 | zlc = zonelist->zlcache_ptr; | |
1324 | if (!zlc) | |
1325 | return NULL; | |
1326 | ||
1327 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { | |
1328 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1329 | zlc->last_full_zap = jiffies; | |
1330 | } | |
1331 | ||
1332 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? | |
1333 | &cpuset_current_mems_allowed : | |
1334 | &node_states[N_HIGH_MEMORY]; | |
1335 | return allowednodes; | |
1336 | } | |
1337 | ||
1338 | /* | |
1339 | * Given 'z' scanning a zonelist, run a couple of quick checks to see | |
1340 | * if it is worth looking at further for free memory: | |
1341 | * 1) Check that the zone isn't thought to be full (doesn't have its | |
1342 | * bit set in the zonelist_cache fullzones BITMAP). | |
1343 | * 2) Check that the zones node (obtained from the zonelist_cache | |
1344 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. | |
1345 | * Return true (non-zero) if zone is worth looking at further, or | |
1346 | * else return false (zero) if it is not. | |
1347 | * | |
1348 | * This check -ignores- the distinction between various watermarks, | |
1349 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is | |
1350 | * found to be full for any variation of these watermarks, it will | |
1351 | * be considered full for up to one second by all requests, unless | |
1352 | * we are so low on memory on all allowed nodes that we are forced | |
1353 | * into the second scan of the zonelist. | |
1354 | * | |
1355 | * In the second scan we ignore this zonelist cache and exactly | |
1356 | * apply the watermarks to all zones, even it is slower to do so. | |
1357 | * We are low on memory in the second scan, and should leave no stone | |
1358 | * unturned looking for a free page. | |
1359 | */ | |
1360 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1361 | nodemask_t *allowednodes) | |
1362 | { | |
1363 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1364 | int i; /* index of *z in zonelist zones */ | |
1365 | int n; /* node that zone *z is on */ | |
1366 | ||
1367 | zlc = zonelist->zlcache_ptr; | |
1368 | if (!zlc) | |
1369 | return 1; | |
1370 | ||
1371 | i = z - zonelist->_zonerefs; | |
1372 | n = zlc->z_to_n[i]; | |
1373 | ||
1374 | /* This zone is worth trying if it is allowed but not full */ | |
1375 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); | |
1376 | } | |
1377 | ||
1378 | /* | |
1379 | * Given 'z' scanning a zonelist, set the corresponding bit in | |
1380 | * zlc->fullzones, so that subsequent attempts to allocate a page | |
1381 | * from that zone don't waste time re-examining it. | |
1382 | */ | |
1383 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1384 | { | |
1385 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1386 | int i; /* index of *z in zonelist zones */ | |
1387 | ||
1388 | zlc = zonelist->zlcache_ptr; | |
1389 | if (!zlc) | |
1390 | return; | |
1391 | ||
1392 | i = z - zonelist->_zonerefs; | |
1393 | ||
1394 | set_bit(i, zlc->fullzones); | |
1395 | } | |
1396 | ||
1397 | #else /* CONFIG_NUMA */ | |
1398 | ||
1399 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1400 | { | |
1401 | return NULL; | |
1402 | } | |
1403 | ||
1404 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1405 | nodemask_t *allowednodes) | |
1406 | { | |
1407 | return 1; | |
1408 | } | |
1409 | ||
1410 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1411 | { | |
1412 | } | |
1413 | #endif /* CONFIG_NUMA */ | |
1414 | ||
1415 | /* | |
1416 | * get_page_from_freelist goes through the zonelist trying to allocate | |
1417 | * a page. | |
1418 | */ | |
1419 | static struct page * | |
1420 | get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, | |
1421 | struct zonelist *zonelist, int high_zoneidx, int alloc_flags, | |
1422 | struct zone *preferred_zone, int migratetype) | |
1423 | { | |
1424 | struct zoneref *z; | |
1425 | struct page *page = NULL; | |
1426 | int classzone_idx; | |
1427 | struct zone *zone; | |
1428 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ | |
1429 | int zlc_active = 0; /* set if using zonelist_cache */ | |
1430 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ | |
1431 | ||
1432 | classzone_idx = zone_idx(preferred_zone); | |
1433 | zonelist_scan: | |
1434 | /* | |
1435 | * Scan zonelist, looking for a zone with enough free. | |
1436 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
1437 | */ | |
1438 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
1439 | high_zoneidx, nodemask) { | |
1440 | if (NUMA_BUILD && zlc_active && | |
1441 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
1442 | continue; | |
1443 | if ((alloc_flags & ALLOC_CPUSET) && | |
1444 | !cpuset_zone_allowed_softwall(zone, gfp_mask)) | |
1445 | goto try_next_zone; | |
1446 | ||
1447 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
1448 | if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { | |
1449 | unsigned long mark; | |
1450 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; | |
1451 | if (!zone_watermark_ok(zone, order, mark, | |
1452 | classzone_idx, alloc_flags)) { | |
1453 | if (!zone_reclaim_mode || | |
1454 | !zone_reclaim(zone, gfp_mask, order)) | |
1455 | goto this_zone_full; | |
1456 | } | |
1457 | } | |
1458 | ||
1459 | page = buffered_rmqueue(preferred_zone, zone, order, | |
1460 | gfp_mask, migratetype); | |
1461 | if (page) | |
1462 | break; | |
1463 | this_zone_full: | |
1464 | if (NUMA_BUILD) | |
1465 | zlc_mark_zone_full(zonelist, z); | |
1466 | try_next_zone: | |
1467 | if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { | |
1468 | /* | |
1469 | * we do zlc_setup after the first zone is tried but only | |
1470 | * if there are multiple nodes make it worthwhile | |
1471 | */ | |
1472 | allowednodes = zlc_setup(zonelist, alloc_flags); | |
1473 | zlc_active = 1; | |
1474 | did_zlc_setup = 1; | |
1475 | } | |
1476 | } | |
1477 | ||
1478 | if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { | |
1479 | /* Disable zlc cache for second zonelist scan */ | |
1480 | zlc_active = 0; | |
1481 | goto zonelist_scan; | |
1482 | } | |
1483 | return page; | |
1484 | } | |
1485 | ||
1486 | static inline int | |
1487 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, | |
1488 | unsigned long pages_reclaimed) | |
1489 | { | |
1490 | /* Do not loop if specifically requested */ | |
1491 | if (gfp_mask & __GFP_NORETRY) | |
1492 | return 0; | |
1493 | ||
1494 | /* | |
1495 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER | |
1496 | * means __GFP_NOFAIL, but that may not be true in other | |
1497 | * implementations. | |
1498 | */ | |
1499 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
1500 | return 1; | |
1501 | ||
1502 | /* | |
1503 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is | |
1504 | * specified, then we retry until we no longer reclaim any pages | |
1505 | * (above), or we've reclaimed an order of pages at least as | |
1506 | * large as the allocation's order. In both cases, if the | |
1507 | * allocation still fails, we stop retrying. | |
1508 | */ | |
1509 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) | |
1510 | return 1; | |
1511 | ||
1512 | /* | |
1513 | * Don't let big-order allocations loop unless the caller | |
1514 | * explicitly requests that. | |
1515 | */ | |
1516 | if (gfp_mask & __GFP_NOFAIL) | |
1517 | return 1; | |
1518 | ||
1519 | return 0; | |
1520 | } | |
1521 | ||
1522 | static inline struct page * | |
1523 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
1524 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
1525 | nodemask_t *nodemask, struct zone *preferred_zone, | |
1526 | int migratetype) | |
1527 | { | |
1528 | struct page *page; | |
1529 | ||
1530 | /* Acquire the OOM killer lock for the zones in zonelist */ | |
1531 | if (!try_set_zone_oom(zonelist, gfp_mask)) { | |
1532 | schedule_timeout_uninterruptible(1); | |
1533 | return NULL; | |
1534 | } | |
1535 | ||
1536 | /* | |
1537 | * Go through the zonelist yet one more time, keep very high watermark | |
1538 | * here, this is only to catch a parallel oom killing, we must fail if | |
1539 | * we're still under heavy pressure. | |
1540 | */ | |
1541 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, | |
1542 | order, zonelist, high_zoneidx, | |
1543 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, | |
1544 | preferred_zone, migratetype); | |
1545 | if (page) | |
1546 | goto out; | |
1547 | ||
1548 | /* The OOM killer will not help higher order allocs */ | |
1549 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
1550 | goto out; | |
1551 | ||
1552 | /* Exhausted what can be done so it's blamo time */ | |
1553 | out_of_memory(zonelist, gfp_mask, order); | |
1554 | ||
1555 | out: | |
1556 | clear_zonelist_oom(zonelist, gfp_mask); | |
1557 | return page; | |
1558 | } | |
1559 | ||
1560 | /* The really slow allocator path where we enter direct reclaim */ | |
1561 | static inline struct page * | |
1562 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
1563 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
1564 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
1565 | int migratetype, unsigned long *did_some_progress) | |
1566 | { | |
1567 | struct page *page = NULL; | |
1568 | struct reclaim_state reclaim_state; | |
1569 | struct task_struct *p = current; | |
1570 | ||
1571 | cond_resched(); | |
1572 | ||
1573 | /* We now go into synchronous reclaim */ | |
1574 | cpuset_memory_pressure_bump(); | |
1575 | ||
1576 | /* | |
1577 | * The task's cpuset might have expanded its set of allowable nodes | |
1578 | */ | |
1579 | p->flags |= PF_MEMALLOC; | |
1580 | lockdep_set_current_reclaim_state(gfp_mask); | |
1581 | reclaim_state.reclaimed_slab = 0; | |
1582 | p->reclaim_state = &reclaim_state; | |
1583 | ||
1584 | *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); | |
1585 | ||
1586 | p->reclaim_state = NULL; | |
1587 | lockdep_clear_current_reclaim_state(); | |
1588 | p->flags &= ~PF_MEMALLOC; | |
1589 | ||
1590 | cond_resched(); | |
1591 | ||
1592 | if (order != 0) | |
1593 | drain_all_pages(); | |
1594 | ||
1595 | if (likely(*did_some_progress)) | |
1596 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
1597 | zonelist, high_zoneidx, | |
1598 | alloc_flags, preferred_zone, | |
1599 | migratetype); | |
1600 | return page; | |
1601 | } | |
1602 | ||
1603 | /* | |
1604 | * This is called in the allocator slow-path if the allocation request is of | |
1605 | * sufficient urgency to ignore watermarks and take other desperate measures | |
1606 | */ | |
1607 | static inline struct page * | |
1608 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, | |
1609 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
1610 | nodemask_t *nodemask, struct zone *preferred_zone, | |
1611 | int migratetype) | |
1612 | { | |
1613 | struct page *page; | |
1614 | ||
1615 | do { | |
1616 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
1617 | zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, | |
1618 | preferred_zone, migratetype); | |
1619 | ||
1620 | if (!page && gfp_mask & __GFP_NOFAIL) | |
1621 | congestion_wait(WRITE, HZ/50); | |
1622 | } while (!page && (gfp_mask & __GFP_NOFAIL)); | |
1623 | ||
1624 | return page; | |
1625 | } | |
1626 | ||
1627 | static inline | |
1628 | void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, | |
1629 | enum zone_type high_zoneidx) | |
1630 | { | |
1631 | struct zoneref *z; | |
1632 | struct zone *zone; | |
1633 | ||
1634 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) | |
1635 | wakeup_kswapd(zone, order); | |
1636 | } | |
1637 | ||
1638 | static inline int | |
1639 | gfp_to_alloc_flags(gfp_t gfp_mask) | |
1640 | { | |
1641 | struct task_struct *p = current; | |
1642 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; | |
1643 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
1644 | ||
1645 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ | |
1646 | BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); | |
1647 | ||
1648 | /* | |
1649 | * The caller may dip into page reserves a bit more if the caller | |
1650 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
1651 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
1652 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). | |
1653 | */ | |
1654 | alloc_flags |= (gfp_mask & __GFP_HIGH); | |
1655 | ||
1656 | if (!wait) { | |
1657 | alloc_flags |= ALLOC_HARDER; | |
1658 | /* | |
1659 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. | |
1660 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
1661 | */ | |
1662 | alloc_flags &= ~ALLOC_CPUSET; | |
1663 | } else if (unlikely(rt_task(p))) | |
1664 | alloc_flags |= ALLOC_HARDER; | |
1665 | ||
1666 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { | |
1667 | if (!in_interrupt() && | |
1668 | ((p->flags & PF_MEMALLOC) || | |
1669 | unlikely(test_thread_flag(TIF_MEMDIE)))) | |
1670 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
1671 | } | |
1672 | ||
1673 | return alloc_flags; | |
1674 | } | |
1675 | ||
1676 | static inline struct page * | |
1677 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
1678 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
1679 | nodemask_t *nodemask, struct zone *preferred_zone, | |
1680 | int migratetype) | |
1681 | { | |
1682 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
1683 | struct page *page = NULL; | |
1684 | int alloc_flags; | |
1685 | unsigned long pages_reclaimed = 0; | |
1686 | unsigned long did_some_progress; | |
1687 | struct task_struct *p = current; | |
1688 | ||
1689 | /* | |
1690 | * In the slowpath, we sanity check order to avoid ever trying to | |
1691 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may | |
1692 | * be using allocators in order of preference for an area that is | |
1693 | * too large. | |
1694 | */ | |
1695 | if (WARN_ON_ONCE(order >= MAX_ORDER)) | |
1696 | return NULL; | |
1697 | ||
1698 | /* | |
1699 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and | |
1700 | * __GFP_NOWARN set) should not cause reclaim since the subsystem | |
1701 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim | |
1702 | * using a larger set of nodes after it has established that the | |
1703 | * allowed per node queues are empty and that nodes are | |
1704 | * over allocated. | |
1705 | */ | |
1706 | if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) | |
1707 | goto nopage; | |
1708 | ||
1709 | wake_all_kswapd(order, zonelist, high_zoneidx); | |
1710 | ||
1711 | /* | |
1712 | * OK, we're below the kswapd watermark and have kicked background | |
1713 | * reclaim. Now things get more complex, so set up alloc_flags according | |
1714 | * to how we want to proceed. | |
1715 | */ | |
1716 | alloc_flags = gfp_to_alloc_flags(gfp_mask); | |
1717 | ||
1718 | restart: | |
1719 | /* This is the last chance, in general, before the goto nopage. */ | |
1720 | page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, | |
1721 | high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, | |
1722 | preferred_zone, migratetype); | |
1723 | if (page) | |
1724 | goto got_pg; | |
1725 | ||
1726 | rebalance: | |
1727 | /* Allocate without watermarks if the context allows */ | |
1728 | if (alloc_flags & ALLOC_NO_WATERMARKS) { | |
1729 | page = __alloc_pages_high_priority(gfp_mask, order, | |
1730 | zonelist, high_zoneidx, nodemask, | |
1731 | preferred_zone, migratetype); | |
1732 | if (page) | |
1733 | goto got_pg; | |
1734 | } | |
1735 | ||
1736 | /* Atomic allocations - we can't balance anything */ | |
1737 | if (!wait) | |
1738 | goto nopage; | |
1739 | ||
1740 | /* Avoid recursion of direct reclaim */ | |
1741 | if (p->flags & PF_MEMALLOC) | |
1742 | goto nopage; | |
1743 | ||
1744 | /* Try direct reclaim and then allocating */ | |
1745 | page = __alloc_pages_direct_reclaim(gfp_mask, order, | |
1746 | zonelist, high_zoneidx, | |
1747 | nodemask, | |
1748 | alloc_flags, preferred_zone, | |
1749 | migratetype, &did_some_progress); | |
1750 | if (page) | |
1751 | goto got_pg; | |
1752 | ||
1753 | /* | |
1754 | * If we failed to make any progress reclaiming, then we are | |
1755 | * running out of options and have to consider going OOM | |
1756 | */ | |
1757 | if (!did_some_progress) { | |
1758 | if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { | |
1759 | page = __alloc_pages_may_oom(gfp_mask, order, | |
1760 | zonelist, high_zoneidx, | |
1761 | nodemask, preferred_zone, | |
1762 | migratetype); | |
1763 | if (page) | |
1764 | goto got_pg; | |
1765 | ||
1766 | /* | |
1767 | * The OOM killer does not trigger for high-order allocations | |
1768 | * but if no progress is being made, there are no other | |
1769 | * options and retrying is unlikely to help | |
1770 | */ | |
1771 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
1772 | goto nopage; | |
1773 | ||
1774 | goto restart; | |
1775 | } | |
1776 | } | |
1777 | ||
1778 | /* Check if we should retry the allocation */ | |
1779 | pages_reclaimed += did_some_progress; | |
1780 | if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { | |
1781 | /* Wait for some write requests to complete then retry */ | |
1782 | congestion_wait(WRITE, HZ/50); | |
1783 | goto rebalance; | |
1784 | } | |
1785 | ||
1786 | nopage: | |
1787 | if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { | |
1788 | printk(KERN_WARNING "%s: page allocation failure." | |
1789 | " order:%d, mode:0x%x\n", | |
1790 | p->comm, order, gfp_mask); | |
1791 | dump_stack(); | |
1792 | show_mem(); | |
1793 | } | |
1794 | got_pg: | |
1795 | return page; | |
1796 | ||
1797 | } | |
1798 | ||
1799 | /* | |
1800 | * This is the 'heart' of the zoned buddy allocator. | |
1801 | */ | |
1802 | struct page * | |
1803 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, | |
1804 | struct zonelist *zonelist, nodemask_t *nodemask) | |
1805 | { | |
1806 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
1807 | struct zone *preferred_zone; | |
1808 | struct page *page; | |
1809 | int migratetype = allocflags_to_migratetype(gfp_mask); | |
1810 | ||
1811 | lockdep_trace_alloc(gfp_mask); | |
1812 | ||
1813 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
1814 | ||
1815 | if (should_fail_alloc_page(gfp_mask, order)) | |
1816 | return NULL; | |
1817 | ||
1818 | /* | |
1819 | * Check the zones suitable for the gfp_mask contain at least one | |
1820 | * valid zone. It's possible to have an empty zonelist as a result | |
1821 | * of GFP_THISNODE and a memoryless node | |
1822 | */ | |
1823 | if (unlikely(!zonelist->_zonerefs->zone)) | |
1824 | return NULL; | |
1825 | ||
1826 | /* The preferred zone is used for statistics later */ | |
1827 | first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); | |
1828 | if (!preferred_zone) | |
1829 | return NULL; | |
1830 | ||
1831 | /* First allocation attempt */ | |
1832 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, | |
1833 | zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, | |
1834 | preferred_zone, migratetype); | |
1835 | if (unlikely(!page)) | |
1836 | page = __alloc_pages_slowpath(gfp_mask, order, | |
1837 | zonelist, high_zoneidx, nodemask, | |
1838 | preferred_zone, migratetype); | |
1839 | ||
1840 | return page; | |
1841 | } | |
1842 | EXPORT_SYMBOL(__alloc_pages_nodemask); | |
1843 | ||
1844 | /* | |
1845 | * Common helper functions. | |
1846 | */ | |
1847 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) | |
1848 | { | |
1849 | struct page * page; | |
1850 | page = alloc_pages(gfp_mask, order); | |
1851 | if (!page) | |
1852 | return 0; | |
1853 | return (unsigned long) page_address(page); | |
1854 | } | |
1855 | ||
1856 | EXPORT_SYMBOL(__get_free_pages); | |
1857 | ||
1858 | unsigned long get_zeroed_page(gfp_t gfp_mask) | |
1859 | { | |
1860 | struct page * page; | |
1861 | ||
1862 | /* | |
1863 | * get_zeroed_page() returns a 32-bit address, which cannot represent | |
1864 | * a highmem page | |
1865 | */ | |
1866 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); | |
1867 | ||
1868 | page = alloc_pages(gfp_mask | __GFP_ZERO, 0); | |
1869 | if (page) | |
1870 | return (unsigned long) page_address(page); | |
1871 | return 0; | |
1872 | } | |
1873 | ||
1874 | EXPORT_SYMBOL(get_zeroed_page); | |
1875 | ||
1876 | void __pagevec_free(struct pagevec *pvec) | |
1877 | { | |
1878 | int i = pagevec_count(pvec); | |
1879 | ||
1880 | while (--i >= 0) | |
1881 | free_hot_cold_page(pvec->pages[i], pvec->cold); | |
1882 | } | |
1883 | ||
1884 | void __free_pages(struct page *page, unsigned int order) | |
1885 | { | |
1886 | if (put_page_testzero(page)) { | |
1887 | if (order == 0) | |
1888 | free_hot_page(page); | |
1889 | else | |
1890 | __free_pages_ok(page, order); | |
1891 | } | |
1892 | } | |
1893 | ||
1894 | EXPORT_SYMBOL(__free_pages); | |
1895 | ||
1896 | void free_pages(unsigned long addr, unsigned int order) | |
1897 | { | |
1898 | if (addr != 0) { | |
1899 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
1900 | __free_pages(virt_to_page((void *)addr), order); | |
1901 | } | |
1902 | } | |
1903 | ||
1904 | EXPORT_SYMBOL(free_pages); | |
1905 | ||
1906 | /** | |
1907 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
1908 | * @size: the number of bytes to allocate | |
1909 | * @gfp_mask: GFP flags for the allocation | |
1910 | * | |
1911 | * This function is similar to alloc_pages(), except that it allocates the | |
1912 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
1913 | * allocate memory in power-of-two pages. | |
1914 | * | |
1915 | * This function is also limited by MAX_ORDER. | |
1916 | * | |
1917 | * Memory allocated by this function must be released by free_pages_exact(). | |
1918 | */ | |
1919 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
1920 | { | |
1921 | unsigned int order = get_order(size); | |
1922 | unsigned long addr; | |
1923 | ||
1924 | addr = __get_free_pages(gfp_mask, order); | |
1925 | if (addr) { | |
1926 | unsigned long alloc_end = addr + (PAGE_SIZE << order); | |
1927 | unsigned long used = addr + PAGE_ALIGN(size); | |
1928 | ||
1929 | split_page(virt_to_page(addr), order); | |
1930 | while (used < alloc_end) { | |
1931 | free_page(used); | |
1932 | used += PAGE_SIZE; | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | return (void *)addr; | |
1937 | } | |
1938 | EXPORT_SYMBOL(alloc_pages_exact); | |
1939 | ||
1940 | /** | |
1941 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
1942 | * @virt: the value returned by alloc_pages_exact. | |
1943 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
1944 | * | |
1945 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
1946 | */ | |
1947 | void free_pages_exact(void *virt, size_t size) | |
1948 | { | |
1949 | unsigned long addr = (unsigned long)virt; | |
1950 | unsigned long end = addr + PAGE_ALIGN(size); | |
1951 | ||
1952 | while (addr < end) { | |
1953 | free_page(addr); | |
1954 | addr += PAGE_SIZE; | |
1955 | } | |
1956 | } | |
1957 | EXPORT_SYMBOL(free_pages_exact); | |
1958 | ||
1959 | static unsigned int nr_free_zone_pages(int offset) | |
1960 | { | |
1961 | struct zoneref *z; | |
1962 | struct zone *zone; | |
1963 | ||
1964 | /* Just pick one node, since fallback list is circular */ | |
1965 | unsigned int sum = 0; | |
1966 | ||
1967 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); | |
1968 | ||
1969 | for_each_zone_zonelist(zone, z, zonelist, offset) { | |
1970 | unsigned long size = zone->present_pages; | |
1971 | unsigned long high = high_wmark_pages(zone); | |
1972 | if (size > high) | |
1973 | sum += size - high; | |
1974 | } | |
1975 | ||
1976 | return sum; | |
1977 | } | |
1978 | ||
1979 | /* | |
1980 | * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL | |
1981 | */ | |
1982 | unsigned int nr_free_buffer_pages(void) | |
1983 | { | |
1984 | return nr_free_zone_pages(gfp_zone(GFP_USER)); | |
1985 | } | |
1986 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); | |
1987 | ||
1988 | /* | |
1989 | * Amount of free RAM allocatable within all zones | |
1990 | */ | |
1991 | unsigned int nr_free_pagecache_pages(void) | |
1992 | { | |
1993 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
1994 | } | |
1995 | ||
1996 | static inline void show_node(struct zone *zone) | |
1997 | { | |
1998 | if (NUMA_BUILD) | |
1999 | printk("Node %d ", zone_to_nid(zone)); | |
2000 | } | |
2001 | ||
2002 | void si_meminfo(struct sysinfo *val) | |
2003 | { | |
2004 | val->totalram = totalram_pages; | |
2005 | val->sharedram = 0; | |
2006 | val->freeram = global_page_state(NR_FREE_PAGES); | |
2007 | val->bufferram = nr_blockdev_pages(); | |
2008 | val->totalhigh = totalhigh_pages; | |
2009 | val->freehigh = nr_free_highpages(); | |
2010 | val->mem_unit = PAGE_SIZE; | |
2011 | } | |
2012 | ||
2013 | EXPORT_SYMBOL(si_meminfo); | |
2014 | ||
2015 | #ifdef CONFIG_NUMA | |
2016 | void si_meminfo_node(struct sysinfo *val, int nid) | |
2017 | { | |
2018 | pg_data_t *pgdat = NODE_DATA(nid); | |
2019 | ||
2020 | val->totalram = pgdat->node_present_pages; | |
2021 | val->freeram = node_page_state(nid, NR_FREE_PAGES); | |
2022 | #ifdef CONFIG_HIGHMEM | |
2023 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; | |
2024 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], | |
2025 | NR_FREE_PAGES); | |
2026 | #else | |
2027 | val->totalhigh = 0; | |
2028 | val->freehigh = 0; | |
2029 | #endif | |
2030 | val->mem_unit = PAGE_SIZE; | |
2031 | } | |
2032 | #endif | |
2033 | ||
2034 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
2035 | ||
2036 | /* | |
2037 | * Show free area list (used inside shift_scroll-lock stuff) | |
2038 | * We also calculate the percentage fragmentation. We do this by counting the | |
2039 | * memory on each free list with the exception of the first item on the list. | |
2040 | */ | |
2041 | void show_free_areas(void) | |
2042 | { | |
2043 | int cpu; | |
2044 | struct zone *zone; | |
2045 | ||
2046 | for_each_populated_zone(zone) { | |
2047 | show_node(zone); | |
2048 | printk("%s per-cpu:\n", zone->name); | |
2049 | ||
2050 | for_each_online_cpu(cpu) { | |
2051 | struct per_cpu_pageset *pageset; | |
2052 | ||
2053 | pageset = zone_pcp(zone, cpu); | |
2054 | ||
2055 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", | |
2056 | cpu, pageset->pcp.high, | |
2057 | pageset->pcp.batch, pageset->pcp.count); | |
2058 | } | |
2059 | } | |
2060 | ||
2061 | printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n" | |
2062 | " inactive_file:%lu" | |
2063 | //TODO: check/adjust line lengths | |
2064 | #ifdef CONFIG_UNEVICTABLE_LRU | |
2065 | " unevictable:%lu" | |
2066 | #endif | |
2067 | " dirty:%lu writeback:%lu unstable:%lu\n" | |
2068 | " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n", | |
2069 | global_page_state(NR_ACTIVE_ANON), | |
2070 | global_page_state(NR_ACTIVE_FILE), | |
2071 | global_page_state(NR_INACTIVE_ANON), | |
2072 | global_page_state(NR_INACTIVE_FILE), | |
2073 | #ifdef CONFIG_UNEVICTABLE_LRU | |
2074 | global_page_state(NR_UNEVICTABLE), | |
2075 | #endif | |
2076 | global_page_state(NR_FILE_DIRTY), | |
2077 | global_page_state(NR_WRITEBACK), | |
2078 | global_page_state(NR_UNSTABLE_NFS), | |
2079 | global_page_state(NR_FREE_PAGES), | |
2080 | global_page_state(NR_SLAB_RECLAIMABLE) + | |
2081 | global_page_state(NR_SLAB_UNRECLAIMABLE), | |
2082 | global_page_state(NR_FILE_MAPPED), | |
2083 | global_page_state(NR_PAGETABLE), | |
2084 | global_page_state(NR_BOUNCE)); | |
2085 | ||
2086 | for_each_populated_zone(zone) { | |
2087 | int i; | |
2088 | ||
2089 | show_node(zone); | |
2090 | printk("%s" | |
2091 | " free:%lukB" | |
2092 | " min:%lukB" | |
2093 | " low:%lukB" | |
2094 | " high:%lukB" | |
2095 | " active_anon:%lukB" | |
2096 | " inactive_anon:%lukB" | |
2097 | " active_file:%lukB" | |
2098 | " inactive_file:%lukB" | |
2099 | #ifdef CONFIG_UNEVICTABLE_LRU | |
2100 | " unevictable:%lukB" | |
2101 | #endif | |
2102 | " present:%lukB" | |
2103 | " pages_scanned:%lu" | |
2104 | " all_unreclaimable? %s" | |
2105 | "\n", | |
2106 | zone->name, | |
2107 | K(zone_page_state(zone, NR_FREE_PAGES)), | |
2108 | K(min_wmark_pages(zone)), | |
2109 | K(low_wmark_pages(zone)), | |
2110 | K(high_wmark_pages(zone)), | |
2111 | K(zone_page_state(zone, NR_ACTIVE_ANON)), | |
2112 | K(zone_page_state(zone, NR_INACTIVE_ANON)), | |
2113 | K(zone_page_state(zone, NR_ACTIVE_FILE)), | |
2114 | K(zone_page_state(zone, NR_INACTIVE_FILE)), | |
2115 | #ifdef CONFIG_UNEVICTABLE_LRU | |
2116 | K(zone_page_state(zone, NR_UNEVICTABLE)), | |
2117 | #endif | |
2118 | K(zone->present_pages), | |
2119 | zone->pages_scanned, | |
2120 | (zone_is_all_unreclaimable(zone) ? "yes" : "no") | |
2121 | ); | |
2122 | printk("lowmem_reserve[]:"); | |
2123 | for (i = 0; i < MAX_NR_ZONES; i++) | |
2124 | printk(" %lu", zone->lowmem_reserve[i]); | |
2125 | printk("\n"); | |
2126 | } | |
2127 | ||
2128 | for_each_populated_zone(zone) { | |
2129 | unsigned long nr[MAX_ORDER], flags, order, total = 0; | |
2130 | ||
2131 | show_node(zone); | |
2132 | printk("%s: ", zone->name); | |
2133 | ||
2134 | spin_lock_irqsave(&zone->lock, flags); | |
2135 | for (order = 0; order < MAX_ORDER; order++) { | |
2136 | nr[order] = zone->free_area[order].nr_free; | |
2137 | total += nr[order] << order; | |
2138 | } | |
2139 | spin_unlock_irqrestore(&zone->lock, flags); | |
2140 | for (order = 0; order < MAX_ORDER; order++) | |
2141 | printk("%lu*%lukB ", nr[order], K(1UL) << order); | |
2142 | printk("= %lukB\n", K(total)); | |
2143 | } | |
2144 | ||
2145 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); | |
2146 | ||
2147 | show_swap_cache_info(); | |
2148 | } | |
2149 | ||
2150 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) | |
2151 | { | |
2152 | zoneref->zone = zone; | |
2153 | zoneref->zone_idx = zone_idx(zone); | |
2154 | } | |
2155 | ||
2156 | /* | |
2157 | * Builds allocation fallback zone lists. | |
2158 | * | |
2159 | * Add all populated zones of a node to the zonelist. | |
2160 | */ | |
2161 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, | |
2162 | int nr_zones, enum zone_type zone_type) | |
2163 | { | |
2164 | struct zone *zone; | |
2165 | ||
2166 | BUG_ON(zone_type >= MAX_NR_ZONES); | |
2167 | zone_type++; | |
2168 | ||
2169 | do { | |
2170 | zone_type--; | |
2171 | zone = pgdat->node_zones + zone_type; | |
2172 | if (populated_zone(zone)) { | |
2173 | zoneref_set_zone(zone, | |
2174 | &zonelist->_zonerefs[nr_zones++]); | |
2175 | check_highest_zone(zone_type); | |
2176 | } | |
2177 | ||
2178 | } while (zone_type); | |
2179 | return nr_zones; | |
2180 | } | |
2181 | ||
2182 | ||
2183 | /* | |
2184 | * zonelist_order: | |
2185 | * 0 = automatic detection of better ordering. | |
2186 | * 1 = order by ([node] distance, -zonetype) | |
2187 | * 2 = order by (-zonetype, [node] distance) | |
2188 | * | |
2189 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create | |
2190 | * the same zonelist. So only NUMA can configure this param. | |
2191 | */ | |
2192 | #define ZONELIST_ORDER_DEFAULT 0 | |
2193 | #define ZONELIST_ORDER_NODE 1 | |
2194 | #define ZONELIST_ORDER_ZONE 2 | |
2195 | ||
2196 | /* zonelist order in the kernel. | |
2197 | * set_zonelist_order() will set this to NODE or ZONE. | |
2198 | */ | |
2199 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2200 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; | |
2201 | ||
2202 | ||
2203 | #ifdef CONFIG_NUMA | |
2204 | /* The value user specified ....changed by config */ | |
2205 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2206 | /* string for sysctl */ | |
2207 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
2208 | char numa_zonelist_order[16] = "default"; | |
2209 | ||
2210 | /* | |
2211 | * interface for configure zonelist ordering. | |
2212 | * command line option "numa_zonelist_order" | |
2213 | * = "[dD]efault - default, automatic configuration. | |
2214 | * = "[nN]ode - order by node locality, then by zone within node | |
2215 | * = "[zZ]one - order by zone, then by locality within zone | |
2216 | */ | |
2217 | ||
2218 | static int __parse_numa_zonelist_order(char *s) | |
2219 | { | |
2220 | if (*s == 'd' || *s == 'D') { | |
2221 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
2222 | } else if (*s == 'n' || *s == 'N') { | |
2223 | user_zonelist_order = ZONELIST_ORDER_NODE; | |
2224 | } else if (*s == 'z' || *s == 'Z') { | |
2225 | user_zonelist_order = ZONELIST_ORDER_ZONE; | |
2226 | } else { | |
2227 | printk(KERN_WARNING | |
2228 | "Ignoring invalid numa_zonelist_order value: " | |
2229 | "%s\n", s); | |
2230 | return -EINVAL; | |
2231 | } | |
2232 | return 0; | |
2233 | } | |
2234 | ||
2235 | static __init int setup_numa_zonelist_order(char *s) | |
2236 | { | |
2237 | if (s) | |
2238 | return __parse_numa_zonelist_order(s); | |
2239 | return 0; | |
2240 | } | |
2241 | early_param("numa_zonelist_order", setup_numa_zonelist_order); | |
2242 | ||
2243 | /* | |
2244 | * sysctl handler for numa_zonelist_order | |
2245 | */ | |
2246 | int numa_zonelist_order_handler(ctl_table *table, int write, | |
2247 | struct file *file, void __user *buffer, size_t *length, | |
2248 | loff_t *ppos) | |
2249 | { | |
2250 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; | |
2251 | int ret; | |
2252 | ||
2253 | if (write) | |
2254 | strncpy(saved_string, (char*)table->data, | |
2255 | NUMA_ZONELIST_ORDER_LEN); | |
2256 | ret = proc_dostring(table, write, file, buffer, length, ppos); | |
2257 | if (ret) | |
2258 | return ret; | |
2259 | if (write) { | |
2260 | int oldval = user_zonelist_order; | |
2261 | if (__parse_numa_zonelist_order((char*)table->data)) { | |
2262 | /* | |
2263 | * bogus value. restore saved string | |
2264 | */ | |
2265 | strncpy((char*)table->data, saved_string, | |
2266 | NUMA_ZONELIST_ORDER_LEN); | |
2267 | user_zonelist_order = oldval; | |
2268 | } else if (oldval != user_zonelist_order) | |
2269 | build_all_zonelists(); | |
2270 | } | |
2271 | return 0; | |
2272 | } | |
2273 | ||
2274 | ||
2275 | #define MAX_NODE_LOAD (nr_online_nodes) | |
2276 | static int node_load[MAX_NUMNODES]; | |
2277 | ||
2278 | /** | |
2279 | * find_next_best_node - find the next node that should appear in a given node's fallback list | |
2280 | * @node: node whose fallback list we're appending | |
2281 | * @used_node_mask: nodemask_t of already used nodes | |
2282 | * | |
2283 | * We use a number of factors to determine which is the next node that should | |
2284 | * appear on a given node's fallback list. The node should not have appeared | |
2285 | * already in @node's fallback list, and it should be the next closest node | |
2286 | * according to the distance array (which contains arbitrary distance values | |
2287 | * from each node to each node in the system), and should also prefer nodes | |
2288 | * with no CPUs, since presumably they'll have very little allocation pressure | |
2289 | * on them otherwise. | |
2290 | * It returns -1 if no node is found. | |
2291 | */ | |
2292 | static int find_next_best_node(int node, nodemask_t *used_node_mask) | |
2293 | { | |
2294 | int n, val; | |
2295 | int min_val = INT_MAX; | |
2296 | int best_node = -1; | |
2297 | const struct cpumask *tmp = cpumask_of_node(0); | |
2298 | ||
2299 | /* Use the local node if we haven't already */ | |
2300 | if (!node_isset(node, *used_node_mask)) { | |
2301 | node_set(node, *used_node_mask); | |
2302 | return node; | |
2303 | } | |
2304 | ||
2305 | for_each_node_state(n, N_HIGH_MEMORY) { | |
2306 | ||
2307 | /* Don't want a node to appear more than once */ | |
2308 | if (node_isset(n, *used_node_mask)) | |
2309 | continue; | |
2310 | ||
2311 | /* Use the distance array to find the distance */ | |
2312 | val = node_distance(node, n); | |
2313 | ||
2314 | /* Penalize nodes under us ("prefer the next node") */ | |
2315 | val += (n < node); | |
2316 | ||
2317 | /* Give preference to headless and unused nodes */ | |
2318 | tmp = cpumask_of_node(n); | |
2319 | if (!cpumask_empty(tmp)) | |
2320 | val += PENALTY_FOR_NODE_WITH_CPUS; | |
2321 | ||
2322 | /* Slight preference for less loaded node */ | |
2323 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); | |
2324 | val += node_load[n]; | |
2325 | ||
2326 | if (val < min_val) { | |
2327 | min_val = val; | |
2328 | best_node = n; | |
2329 | } | |
2330 | } | |
2331 | ||
2332 | if (best_node >= 0) | |
2333 | node_set(best_node, *used_node_mask); | |
2334 | ||
2335 | return best_node; | |
2336 | } | |
2337 | ||
2338 | ||
2339 | /* | |
2340 | * Build zonelists ordered by node and zones within node. | |
2341 | * This results in maximum locality--normal zone overflows into local | |
2342 | * DMA zone, if any--but risks exhausting DMA zone. | |
2343 | */ | |
2344 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) | |
2345 | { | |
2346 | int j; | |
2347 | struct zonelist *zonelist; | |
2348 | ||
2349 | zonelist = &pgdat->node_zonelists[0]; | |
2350 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) | |
2351 | ; | |
2352 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
2353 | MAX_NR_ZONES - 1); | |
2354 | zonelist->_zonerefs[j].zone = NULL; | |
2355 | zonelist->_zonerefs[j].zone_idx = 0; | |
2356 | } | |
2357 | ||
2358 | /* | |
2359 | * Build gfp_thisnode zonelists | |
2360 | */ | |
2361 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
2362 | { | |
2363 | int j; | |
2364 | struct zonelist *zonelist; | |
2365 | ||
2366 | zonelist = &pgdat->node_zonelists[1]; | |
2367 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
2368 | zonelist->_zonerefs[j].zone = NULL; | |
2369 | zonelist->_zonerefs[j].zone_idx = 0; | |
2370 | } | |
2371 | ||
2372 | /* | |
2373 | * Build zonelists ordered by zone and nodes within zones. | |
2374 | * This results in conserving DMA zone[s] until all Normal memory is | |
2375 | * exhausted, but results in overflowing to remote node while memory | |
2376 | * may still exist in local DMA zone. | |
2377 | */ | |
2378 | static int node_order[MAX_NUMNODES]; | |
2379 | ||
2380 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) | |
2381 | { | |
2382 | int pos, j, node; | |
2383 | int zone_type; /* needs to be signed */ | |
2384 | struct zone *z; | |
2385 | struct zonelist *zonelist; | |
2386 | ||
2387 | zonelist = &pgdat->node_zonelists[0]; | |
2388 | pos = 0; | |
2389 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { | |
2390 | for (j = 0; j < nr_nodes; j++) { | |
2391 | node = node_order[j]; | |
2392 | z = &NODE_DATA(node)->node_zones[zone_type]; | |
2393 | if (populated_zone(z)) { | |
2394 | zoneref_set_zone(z, | |
2395 | &zonelist->_zonerefs[pos++]); | |
2396 | check_highest_zone(zone_type); | |
2397 | } | |
2398 | } | |
2399 | } | |
2400 | zonelist->_zonerefs[pos].zone = NULL; | |
2401 | zonelist->_zonerefs[pos].zone_idx = 0; | |
2402 | } | |
2403 | ||
2404 | static int default_zonelist_order(void) | |
2405 | { | |
2406 | int nid, zone_type; | |
2407 | unsigned long low_kmem_size,total_size; | |
2408 | struct zone *z; | |
2409 | int average_size; | |
2410 | /* | |
2411 | * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. | |
2412 | * If they are really small and used heavily, the system can fall | |
2413 | * into OOM very easily. | |
2414 | * This function detect ZONE_DMA/DMA32 size and confgigures zone order. | |
2415 | */ | |
2416 | /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ | |
2417 | low_kmem_size = 0; | |
2418 | total_size = 0; | |
2419 | for_each_online_node(nid) { | |
2420 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
2421 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
2422 | if (populated_zone(z)) { | |
2423 | if (zone_type < ZONE_NORMAL) | |
2424 | low_kmem_size += z->present_pages; | |
2425 | total_size += z->present_pages; | |
2426 | } | |
2427 | } | |
2428 | } | |
2429 | if (!low_kmem_size || /* there are no DMA area. */ | |
2430 | low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ | |
2431 | return ZONELIST_ORDER_NODE; | |
2432 | /* | |
2433 | * look into each node's config. | |
2434 | * If there is a node whose DMA/DMA32 memory is very big area on | |
2435 | * local memory, NODE_ORDER may be suitable. | |
2436 | */ | |
2437 | average_size = total_size / | |
2438 | (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); | |
2439 | for_each_online_node(nid) { | |
2440 | low_kmem_size = 0; | |
2441 | total_size = 0; | |
2442 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
2443 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
2444 | if (populated_zone(z)) { | |
2445 | if (zone_type < ZONE_NORMAL) | |
2446 | low_kmem_size += z->present_pages; | |
2447 | total_size += z->present_pages; | |
2448 | } | |
2449 | } | |
2450 | if (low_kmem_size && | |
2451 | total_size > average_size && /* ignore small node */ | |
2452 | low_kmem_size > total_size * 70/100) | |
2453 | return ZONELIST_ORDER_NODE; | |
2454 | } | |
2455 | return ZONELIST_ORDER_ZONE; | |
2456 | } | |
2457 | ||
2458 | static void set_zonelist_order(void) | |
2459 | { | |
2460 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) | |
2461 | current_zonelist_order = default_zonelist_order(); | |
2462 | else | |
2463 | current_zonelist_order = user_zonelist_order; | |
2464 | } | |
2465 | ||
2466 | static void build_zonelists(pg_data_t *pgdat) | |
2467 | { | |
2468 | int j, node, load; | |
2469 | enum zone_type i; | |
2470 | nodemask_t used_mask; | |
2471 | int local_node, prev_node; | |
2472 | struct zonelist *zonelist; | |
2473 | int order = current_zonelist_order; | |
2474 | ||
2475 | /* initialize zonelists */ | |
2476 | for (i = 0; i < MAX_ZONELISTS; i++) { | |
2477 | zonelist = pgdat->node_zonelists + i; | |
2478 | zonelist->_zonerefs[0].zone = NULL; | |
2479 | zonelist->_zonerefs[0].zone_idx = 0; | |
2480 | } | |
2481 | ||
2482 | /* NUMA-aware ordering of nodes */ | |
2483 | local_node = pgdat->node_id; | |
2484 | load = nr_online_nodes; | |
2485 | prev_node = local_node; | |
2486 | nodes_clear(used_mask); | |
2487 | ||
2488 | memset(node_load, 0, sizeof(node_load)); | |
2489 | memset(node_order, 0, sizeof(node_order)); | |
2490 | j = 0; | |
2491 | ||
2492 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { | |
2493 | int distance = node_distance(local_node, node); | |
2494 | ||
2495 | /* | |
2496 | * If another node is sufficiently far away then it is better | |
2497 | * to reclaim pages in a zone before going off node. | |
2498 | */ | |
2499 | if (distance > RECLAIM_DISTANCE) | |
2500 | zone_reclaim_mode = 1; | |
2501 | ||
2502 | /* | |
2503 | * We don't want to pressure a particular node. | |
2504 | * So adding penalty to the first node in same | |
2505 | * distance group to make it round-robin. | |
2506 | */ | |
2507 | if (distance != node_distance(local_node, prev_node)) | |
2508 | node_load[node] = load; | |
2509 | ||
2510 | prev_node = node; | |
2511 | load--; | |
2512 | if (order == ZONELIST_ORDER_NODE) | |
2513 | build_zonelists_in_node_order(pgdat, node); | |
2514 | else | |
2515 | node_order[j++] = node; /* remember order */ | |
2516 | } | |
2517 | ||
2518 | if (order == ZONELIST_ORDER_ZONE) { | |
2519 | /* calculate node order -- i.e., DMA last! */ | |
2520 | build_zonelists_in_zone_order(pgdat, j); | |
2521 | } | |
2522 | ||
2523 | build_thisnode_zonelists(pgdat); | |
2524 | } | |
2525 | ||
2526 | /* Construct the zonelist performance cache - see further mmzone.h */ | |
2527 | static void build_zonelist_cache(pg_data_t *pgdat) | |
2528 | { | |
2529 | struct zonelist *zonelist; | |
2530 | struct zonelist_cache *zlc; | |
2531 | struct zoneref *z; | |
2532 | ||
2533 | zonelist = &pgdat->node_zonelists[0]; | |
2534 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; | |
2535 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
2536 | for (z = zonelist->_zonerefs; z->zone; z++) | |
2537 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); | |
2538 | } | |
2539 | ||
2540 | ||
2541 | #else /* CONFIG_NUMA */ | |
2542 | ||
2543 | static void set_zonelist_order(void) | |
2544 | { | |
2545 | current_zonelist_order = ZONELIST_ORDER_ZONE; | |
2546 | } | |
2547 | ||
2548 | static void build_zonelists(pg_data_t *pgdat) | |
2549 | { | |
2550 | int node, local_node; | |
2551 | enum zone_type j; | |
2552 | struct zonelist *zonelist; | |
2553 | ||
2554 | local_node = pgdat->node_id; | |
2555 | ||
2556 | zonelist = &pgdat->node_zonelists[0]; | |
2557 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
2558 | ||
2559 | /* | |
2560 | * Now we build the zonelist so that it contains the zones | |
2561 | * of all the other nodes. | |
2562 | * We don't want to pressure a particular node, so when | |
2563 | * building the zones for node N, we make sure that the | |
2564 | * zones coming right after the local ones are those from | |
2565 | * node N+1 (modulo N) | |
2566 | */ | |
2567 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
2568 | if (!node_online(node)) | |
2569 | continue; | |
2570 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
2571 | MAX_NR_ZONES - 1); | |
2572 | } | |
2573 | for (node = 0; node < local_node; node++) { | |
2574 | if (!node_online(node)) | |
2575 | continue; | |
2576 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
2577 | MAX_NR_ZONES - 1); | |
2578 | } | |
2579 | ||
2580 | zonelist->_zonerefs[j].zone = NULL; | |
2581 | zonelist->_zonerefs[j].zone_idx = 0; | |
2582 | } | |
2583 | ||
2584 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ | |
2585 | static void build_zonelist_cache(pg_data_t *pgdat) | |
2586 | { | |
2587 | pgdat->node_zonelists[0].zlcache_ptr = NULL; | |
2588 | } | |
2589 | ||
2590 | #endif /* CONFIG_NUMA */ | |
2591 | ||
2592 | /* return values int ....just for stop_machine() */ | |
2593 | static int __build_all_zonelists(void *dummy) | |
2594 | { | |
2595 | int nid; | |
2596 | ||
2597 | for_each_online_node(nid) { | |
2598 | pg_data_t *pgdat = NODE_DATA(nid); | |
2599 | ||
2600 | build_zonelists(pgdat); | |
2601 | build_zonelist_cache(pgdat); | |
2602 | } | |
2603 | return 0; | |
2604 | } | |
2605 | ||
2606 | void build_all_zonelists(void) | |
2607 | { | |
2608 | set_zonelist_order(); | |
2609 | ||
2610 | if (system_state == SYSTEM_BOOTING) { | |
2611 | __build_all_zonelists(NULL); | |
2612 | mminit_verify_zonelist(); | |
2613 | cpuset_init_current_mems_allowed(); | |
2614 | } else { | |
2615 | /* we have to stop all cpus to guarantee there is no user | |
2616 | of zonelist */ | |
2617 | stop_machine(__build_all_zonelists, NULL, NULL); | |
2618 | /* cpuset refresh routine should be here */ | |
2619 | } | |
2620 | vm_total_pages = nr_free_pagecache_pages(); | |
2621 | /* | |
2622 | * Disable grouping by mobility if the number of pages in the | |
2623 | * system is too low to allow the mechanism to work. It would be | |
2624 | * more accurate, but expensive to check per-zone. This check is | |
2625 | * made on memory-hotadd so a system can start with mobility | |
2626 | * disabled and enable it later | |
2627 | */ | |
2628 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) | |
2629 | page_group_by_mobility_disabled = 1; | |
2630 | else | |
2631 | page_group_by_mobility_disabled = 0; | |
2632 | ||
2633 | printk("Built %i zonelists in %s order, mobility grouping %s. " | |
2634 | "Total pages: %ld\n", | |
2635 | nr_online_nodes, | |
2636 | zonelist_order_name[current_zonelist_order], | |
2637 | page_group_by_mobility_disabled ? "off" : "on", | |
2638 | vm_total_pages); | |
2639 | #ifdef CONFIG_NUMA | |
2640 | printk("Policy zone: %s\n", zone_names[policy_zone]); | |
2641 | #endif | |
2642 | } | |
2643 | ||
2644 | /* | |
2645 | * Helper functions to size the waitqueue hash table. | |
2646 | * Essentially these want to choose hash table sizes sufficiently | |
2647 | * large so that collisions trying to wait on pages are rare. | |
2648 | * But in fact, the number of active page waitqueues on typical | |
2649 | * systems is ridiculously low, less than 200. So this is even | |
2650 | * conservative, even though it seems large. | |
2651 | * | |
2652 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to | |
2653 | * waitqueues, i.e. the size of the waitq table given the number of pages. | |
2654 | */ | |
2655 | #define PAGES_PER_WAITQUEUE 256 | |
2656 | ||
2657 | #ifndef CONFIG_MEMORY_HOTPLUG | |
2658 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
2659 | { | |
2660 | unsigned long size = 1; | |
2661 | ||
2662 | pages /= PAGES_PER_WAITQUEUE; | |
2663 | ||
2664 | while (size < pages) | |
2665 | size <<= 1; | |
2666 | ||
2667 | /* | |
2668 | * Once we have dozens or even hundreds of threads sleeping | |
2669 | * on IO we've got bigger problems than wait queue collision. | |
2670 | * Limit the size of the wait table to a reasonable size. | |
2671 | */ | |
2672 | size = min(size, 4096UL); | |
2673 | ||
2674 | return max(size, 4UL); | |
2675 | } | |
2676 | #else | |
2677 | /* | |
2678 | * A zone's size might be changed by hot-add, so it is not possible to determine | |
2679 | * a suitable size for its wait_table. So we use the maximum size now. | |
2680 | * | |
2681 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: | |
2682 | * | |
2683 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. | |
2684 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. | |
2685 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. | |
2686 | * | |
2687 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages | |
2688 | * or more by the traditional way. (See above). It equals: | |
2689 | * | |
2690 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. | |
2691 | * ia64(16K page size) : = ( 8G + 4M)byte. | |
2692 | * powerpc (64K page size) : = (32G +16M)byte. | |
2693 | */ | |
2694 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
2695 | { | |
2696 | return 4096UL; | |
2697 | } | |
2698 | #endif | |
2699 | ||
2700 | /* | |
2701 | * This is an integer logarithm so that shifts can be used later | |
2702 | * to extract the more random high bits from the multiplicative | |
2703 | * hash function before the remainder is taken. | |
2704 | */ | |
2705 | static inline unsigned long wait_table_bits(unsigned long size) | |
2706 | { | |
2707 | return ffz(~size); | |
2708 | } | |
2709 | ||
2710 | #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) | |
2711 | ||
2712 | /* | |
2713 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number | |
2714 | * of blocks reserved is based on min_wmark_pages(zone). The memory within | |
2715 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes | |
2716 | * higher will lead to a bigger reserve which will get freed as contiguous | |
2717 | * blocks as reclaim kicks in | |
2718 | */ | |
2719 | static void setup_zone_migrate_reserve(struct zone *zone) | |
2720 | { | |
2721 | unsigned long start_pfn, pfn, end_pfn; | |
2722 | struct page *page; | |
2723 | unsigned long reserve, block_migratetype; | |
2724 | ||
2725 | /* Get the start pfn, end pfn and the number of blocks to reserve */ | |
2726 | start_pfn = zone->zone_start_pfn; | |
2727 | end_pfn = start_pfn + zone->spanned_pages; | |
2728 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> | |
2729 | pageblock_order; | |
2730 | ||
2731 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
2732 | if (!pfn_valid(pfn)) | |
2733 | continue; | |
2734 | page = pfn_to_page(pfn); | |
2735 | ||
2736 | /* Watch out for overlapping nodes */ | |
2737 | if (page_to_nid(page) != zone_to_nid(zone)) | |
2738 | continue; | |
2739 | ||
2740 | /* Blocks with reserved pages will never free, skip them. */ | |
2741 | if (PageReserved(page)) | |
2742 | continue; | |
2743 | ||
2744 | block_migratetype = get_pageblock_migratetype(page); | |
2745 | ||
2746 | /* If this block is reserved, account for it */ | |
2747 | if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { | |
2748 | reserve--; | |
2749 | continue; | |
2750 | } | |
2751 | ||
2752 | /* Suitable for reserving if this block is movable */ | |
2753 | if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { | |
2754 | set_pageblock_migratetype(page, MIGRATE_RESERVE); | |
2755 | move_freepages_block(zone, page, MIGRATE_RESERVE); | |
2756 | reserve--; | |
2757 | continue; | |
2758 | } | |
2759 | ||
2760 | /* | |
2761 | * If the reserve is met and this is a previous reserved block, | |
2762 | * take it back | |
2763 | */ | |
2764 | if (block_migratetype == MIGRATE_RESERVE) { | |
2765 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
2766 | move_freepages_block(zone, page, MIGRATE_MOVABLE); | |
2767 | } | |
2768 | } | |
2769 | } | |
2770 | ||
2771 | /* | |
2772 | * Initially all pages are reserved - free ones are freed | |
2773 | * up by free_all_bootmem() once the early boot process is | |
2774 | * done. Non-atomic initialization, single-pass. | |
2775 | */ | |
2776 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, | |
2777 | unsigned long start_pfn, enum memmap_context context) | |
2778 | { | |
2779 | struct page *page; | |
2780 | unsigned long end_pfn = start_pfn + size; | |
2781 | unsigned long pfn; | |
2782 | struct zone *z; | |
2783 | ||
2784 | if (highest_memmap_pfn < end_pfn - 1) | |
2785 | highest_memmap_pfn = end_pfn - 1; | |
2786 | ||
2787 | z = &NODE_DATA(nid)->node_zones[zone]; | |
2788 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
2789 | /* | |
2790 | * There can be holes in boot-time mem_map[]s | |
2791 | * handed to this function. They do not | |
2792 | * exist on hotplugged memory. | |
2793 | */ | |
2794 | if (context == MEMMAP_EARLY) { | |
2795 | if (!early_pfn_valid(pfn)) | |
2796 | continue; | |
2797 | if (!early_pfn_in_nid(pfn, nid)) | |
2798 | continue; | |
2799 | } | |
2800 | page = pfn_to_page(pfn); | |
2801 | set_page_links(page, zone, nid, pfn); | |
2802 | mminit_verify_page_links(page, zone, nid, pfn); | |
2803 | init_page_count(page); | |
2804 | reset_page_mapcount(page); | |
2805 | SetPageReserved(page); | |
2806 | /* | |
2807 | * Mark the block movable so that blocks are reserved for | |
2808 | * movable at startup. This will force kernel allocations | |
2809 | * to reserve their blocks rather than leaking throughout | |
2810 | * the address space during boot when many long-lived | |
2811 | * kernel allocations are made. Later some blocks near | |
2812 | * the start are marked MIGRATE_RESERVE by | |
2813 | * setup_zone_migrate_reserve() | |
2814 | * | |
2815 | * bitmap is created for zone's valid pfn range. but memmap | |
2816 | * can be created for invalid pages (for alignment) | |
2817 | * check here not to call set_pageblock_migratetype() against | |
2818 | * pfn out of zone. | |
2819 | */ | |
2820 | if ((z->zone_start_pfn <= pfn) | |
2821 | && (pfn < z->zone_start_pfn + z->spanned_pages) | |
2822 | && !(pfn & (pageblock_nr_pages - 1))) | |
2823 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
2824 | ||
2825 | INIT_LIST_HEAD(&page->lru); | |
2826 | #ifdef WANT_PAGE_VIRTUAL | |
2827 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ | |
2828 | if (!is_highmem_idx(zone)) | |
2829 | set_page_address(page, __va(pfn << PAGE_SHIFT)); | |
2830 | #endif | |
2831 | } | |
2832 | } | |
2833 | ||
2834 | static void __meminit zone_init_free_lists(struct zone *zone) | |
2835 | { | |
2836 | int order, t; | |
2837 | for_each_migratetype_order(order, t) { | |
2838 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); | |
2839 | zone->free_area[order].nr_free = 0; | |
2840 | } | |
2841 | } | |
2842 | ||
2843 | #ifndef __HAVE_ARCH_MEMMAP_INIT | |
2844 | #define memmap_init(size, nid, zone, start_pfn) \ | |
2845 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) | |
2846 | #endif | |
2847 | ||
2848 | static int zone_batchsize(struct zone *zone) | |
2849 | { | |
2850 | #ifdef CONFIG_MMU | |
2851 | int batch; | |
2852 | ||
2853 | /* | |
2854 | * The per-cpu-pages pools are set to around 1000th of the | |
2855 | * size of the zone. But no more than 1/2 of a meg. | |
2856 | * | |
2857 | * OK, so we don't know how big the cache is. So guess. | |
2858 | */ | |
2859 | batch = zone->present_pages / 1024; | |
2860 | if (batch * PAGE_SIZE > 512 * 1024) | |
2861 | batch = (512 * 1024) / PAGE_SIZE; | |
2862 | batch /= 4; /* We effectively *= 4 below */ | |
2863 | if (batch < 1) | |
2864 | batch = 1; | |
2865 | ||
2866 | /* | |
2867 | * Clamp the batch to a 2^n - 1 value. Having a power | |
2868 | * of 2 value was found to be more likely to have | |
2869 | * suboptimal cache aliasing properties in some cases. | |
2870 | * | |
2871 | * For example if 2 tasks are alternately allocating | |
2872 | * batches of pages, one task can end up with a lot | |
2873 | * of pages of one half of the possible page colors | |
2874 | * and the other with pages of the other colors. | |
2875 | */ | |
2876 | batch = rounddown_pow_of_two(batch + batch/2) - 1; | |
2877 | ||
2878 | return batch; | |
2879 | ||
2880 | #else | |
2881 | /* The deferral and batching of frees should be suppressed under NOMMU | |
2882 | * conditions. | |
2883 | * | |
2884 | * The problem is that NOMMU needs to be able to allocate large chunks | |
2885 | * of contiguous memory as there's no hardware page translation to | |
2886 | * assemble apparent contiguous memory from discontiguous pages. | |
2887 | * | |
2888 | * Queueing large contiguous runs of pages for batching, however, | |
2889 | * causes the pages to actually be freed in smaller chunks. As there | |
2890 | * can be a significant delay between the individual batches being | |
2891 | * recycled, this leads to the once large chunks of space being | |
2892 | * fragmented and becoming unavailable for high-order allocations. | |
2893 | */ | |
2894 | return 0; | |
2895 | #endif | |
2896 | } | |
2897 | ||
2898 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) | |
2899 | { | |
2900 | struct per_cpu_pages *pcp; | |
2901 | ||
2902 | memset(p, 0, sizeof(*p)); | |
2903 | ||
2904 | pcp = &p->pcp; | |
2905 | pcp->count = 0; | |
2906 | pcp->high = 6 * batch; | |
2907 | pcp->batch = max(1UL, 1 * batch); | |
2908 | INIT_LIST_HEAD(&pcp->list); | |
2909 | } | |
2910 | ||
2911 | /* | |
2912 | * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist | |
2913 | * to the value high for the pageset p. | |
2914 | */ | |
2915 | ||
2916 | static void setup_pagelist_highmark(struct per_cpu_pageset *p, | |
2917 | unsigned long high) | |
2918 | { | |
2919 | struct per_cpu_pages *pcp; | |
2920 | ||
2921 | pcp = &p->pcp; | |
2922 | pcp->high = high; | |
2923 | pcp->batch = max(1UL, high/4); | |
2924 | if ((high/4) > (PAGE_SHIFT * 8)) | |
2925 | pcp->batch = PAGE_SHIFT * 8; | |
2926 | } | |
2927 | ||
2928 | ||
2929 | #ifdef CONFIG_NUMA | |
2930 | /* | |
2931 | * Boot pageset table. One per cpu which is going to be used for all | |
2932 | * zones and all nodes. The parameters will be set in such a way | |
2933 | * that an item put on a list will immediately be handed over to | |
2934 | * the buddy list. This is safe since pageset manipulation is done | |
2935 | * with interrupts disabled. | |
2936 | * | |
2937 | * Some NUMA counter updates may also be caught by the boot pagesets. | |
2938 | * | |
2939 | * The boot_pagesets must be kept even after bootup is complete for | |
2940 | * unused processors and/or zones. They do play a role for bootstrapping | |
2941 | * hotplugged processors. | |
2942 | * | |
2943 | * zoneinfo_show() and maybe other functions do | |
2944 | * not check if the processor is online before following the pageset pointer. | |
2945 | * Other parts of the kernel may not check if the zone is available. | |
2946 | */ | |
2947 | static struct per_cpu_pageset boot_pageset[NR_CPUS]; | |
2948 | ||
2949 | /* | |
2950 | * Dynamically allocate memory for the | |
2951 | * per cpu pageset array in struct zone. | |
2952 | */ | |
2953 | static int __cpuinit process_zones(int cpu) | |
2954 | { | |
2955 | struct zone *zone, *dzone; | |
2956 | int node = cpu_to_node(cpu); | |
2957 | ||
2958 | node_set_state(node, N_CPU); /* this node has a cpu */ | |
2959 | ||
2960 | for_each_populated_zone(zone) { | |
2961 | zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), | |
2962 | GFP_KERNEL, node); | |
2963 | if (!zone_pcp(zone, cpu)) | |
2964 | goto bad; | |
2965 | ||
2966 | setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); | |
2967 | ||
2968 | if (percpu_pagelist_fraction) | |
2969 | setup_pagelist_highmark(zone_pcp(zone, cpu), | |
2970 | (zone->present_pages / percpu_pagelist_fraction)); | |
2971 | } | |
2972 | ||
2973 | return 0; | |
2974 | bad: | |
2975 | for_each_zone(dzone) { | |
2976 | if (!populated_zone(dzone)) | |
2977 | continue; | |
2978 | if (dzone == zone) | |
2979 | break; | |
2980 | kfree(zone_pcp(dzone, cpu)); | |
2981 | zone_pcp(dzone, cpu) = NULL; | |
2982 | } | |
2983 | return -ENOMEM; | |
2984 | } | |
2985 | ||
2986 | static inline void free_zone_pagesets(int cpu) | |
2987 | { | |
2988 | struct zone *zone; | |
2989 | ||
2990 | for_each_zone(zone) { | |
2991 | struct per_cpu_pageset *pset = zone_pcp(zone, cpu); | |
2992 | ||
2993 | /* Free per_cpu_pageset if it is slab allocated */ | |
2994 | if (pset != &boot_pageset[cpu]) | |
2995 | kfree(pset); | |
2996 | zone_pcp(zone, cpu) = NULL; | |
2997 | } | |
2998 | } | |
2999 | ||
3000 | static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, | |
3001 | unsigned long action, | |
3002 | void *hcpu) | |
3003 | { | |
3004 | int cpu = (long)hcpu; | |
3005 | int ret = NOTIFY_OK; | |
3006 | ||
3007 | switch (action) { | |
3008 | case CPU_UP_PREPARE: | |
3009 | case CPU_UP_PREPARE_FROZEN: | |
3010 | if (process_zones(cpu)) | |
3011 | ret = NOTIFY_BAD; | |
3012 | break; | |
3013 | case CPU_UP_CANCELED: | |
3014 | case CPU_UP_CANCELED_FROZEN: | |
3015 | case CPU_DEAD: | |
3016 | case CPU_DEAD_FROZEN: | |
3017 | free_zone_pagesets(cpu); | |
3018 | break; | |
3019 | default: | |
3020 | break; | |
3021 | } | |
3022 | return ret; | |
3023 | } | |
3024 | ||
3025 | static struct notifier_block __cpuinitdata pageset_notifier = | |
3026 | { &pageset_cpuup_callback, NULL, 0 }; | |
3027 | ||
3028 | void __init setup_per_cpu_pageset(void) | |
3029 | { | |
3030 | int err; | |
3031 | ||
3032 | /* Initialize per_cpu_pageset for cpu 0. | |
3033 | * A cpuup callback will do this for every cpu | |
3034 | * as it comes online | |
3035 | */ | |
3036 | err = process_zones(smp_processor_id()); | |
3037 | BUG_ON(err); | |
3038 | register_cpu_notifier(&pageset_notifier); | |
3039 | } | |
3040 | ||
3041 | #endif | |
3042 | ||
3043 | static noinline __init_refok | |
3044 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) | |
3045 | { | |
3046 | int i; | |
3047 | struct pglist_data *pgdat = zone->zone_pgdat; | |
3048 | size_t alloc_size; | |
3049 | ||
3050 | /* | |
3051 | * The per-page waitqueue mechanism uses hashed waitqueues | |
3052 | * per zone. | |
3053 | */ | |
3054 | zone->wait_table_hash_nr_entries = | |
3055 | wait_table_hash_nr_entries(zone_size_pages); | |
3056 | zone->wait_table_bits = | |
3057 | wait_table_bits(zone->wait_table_hash_nr_entries); | |
3058 | alloc_size = zone->wait_table_hash_nr_entries | |
3059 | * sizeof(wait_queue_head_t); | |
3060 | ||
3061 | if (!slab_is_available()) { | |
3062 | zone->wait_table = (wait_queue_head_t *) | |
3063 | alloc_bootmem_node(pgdat, alloc_size); | |
3064 | } else { | |
3065 | /* | |
3066 | * This case means that a zone whose size was 0 gets new memory | |
3067 | * via memory hot-add. | |
3068 | * But it may be the case that a new node was hot-added. In | |
3069 | * this case vmalloc() will not be able to use this new node's | |
3070 | * memory - this wait_table must be initialized to use this new | |
3071 | * node itself as well. | |
3072 | * To use this new node's memory, further consideration will be | |
3073 | * necessary. | |
3074 | */ | |
3075 | zone->wait_table = vmalloc(alloc_size); | |
3076 | } | |
3077 | if (!zone->wait_table) | |
3078 | return -ENOMEM; | |
3079 | ||
3080 | for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) | |
3081 | init_waitqueue_head(zone->wait_table + i); | |
3082 | ||
3083 | return 0; | |
3084 | } | |
3085 | ||
3086 | static __meminit void zone_pcp_init(struct zone *zone) | |
3087 | { | |
3088 | int cpu; | |
3089 | unsigned long batch = zone_batchsize(zone); | |
3090 | ||
3091 | for (cpu = 0; cpu < NR_CPUS; cpu++) { | |
3092 | #ifdef CONFIG_NUMA | |
3093 | /* Early boot. Slab allocator not functional yet */ | |
3094 | zone_pcp(zone, cpu) = &boot_pageset[cpu]; | |
3095 | setup_pageset(&boot_pageset[cpu],0); | |
3096 | #else | |
3097 | setup_pageset(zone_pcp(zone,cpu), batch); | |
3098 | #endif | |
3099 | } | |
3100 | if (zone->present_pages) | |
3101 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", | |
3102 | zone->name, zone->present_pages, batch); | |
3103 | } | |
3104 | ||
3105 | __meminit int init_currently_empty_zone(struct zone *zone, | |
3106 | unsigned long zone_start_pfn, | |
3107 | unsigned long size, | |
3108 | enum memmap_context context) | |
3109 | { | |
3110 | struct pglist_data *pgdat = zone->zone_pgdat; | |
3111 | int ret; | |
3112 | ret = zone_wait_table_init(zone, size); | |
3113 | if (ret) | |
3114 | return ret; | |
3115 | pgdat->nr_zones = zone_idx(zone) + 1; | |
3116 | ||
3117 | zone->zone_start_pfn = zone_start_pfn; | |
3118 | ||
3119 | mminit_dprintk(MMINIT_TRACE, "memmap_init", | |
3120 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", | |
3121 | pgdat->node_id, | |
3122 | (unsigned long)zone_idx(zone), | |
3123 | zone_start_pfn, (zone_start_pfn + size)); | |
3124 | ||
3125 | zone_init_free_lists(zone); | |
3126 | ||
3127 | return 0; | |
3128 | } | |
3129 | ||
3130 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP | |
3131 | /* | |
3132 | * Basic iterator support. Return the first range of PFNs for a node | |
3133 | * Note: nid == MAX_NUMNODES returns first region regardless of node | |
3134 | */ | |
3135 | static int __meminit first_active_region_index_in_nid(int nid) | |
3136 | { | |
3137 | int i; | |
3138 | ||
3139 | for (i = 0; i < nr_nodemap_entries; i++) | |
3140 | if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) | |
3141 | return i; | |
3142 | ||
3143 | return -1; | |
3144 | } | |
3145 | ||
3146 | /* | |
3147 | * Basic iterator support. Return the next active range of PFNs for a node | |
3148 | * Note: nid == MAX_NUMNODES returns next region regardless of node | |
3149 | */ | |
3150 | static int __meminit next_active_region_index_in_nid(int index, int nid) | |
3151 | { | |
3152 | for (index = index + 1; index < nr_nodemap_entries; index++) | |
3153 | if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) | |
3154 | return index; | |
3155 | ||
3156 | return -1; | |
3157 | } | |
3158 | ||
3159 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID | |
3160 | /* | |
3161 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. | |
3162 | * Architectures may implement their own version but if add_active_range() | |
3163 | * was used and there are no special requirements, this is a convenient | |
3164 | * alternative | |
3165 | */ | |
3166 | int __meminit __early_pfn_to_nid(unsigned long pfn) | |
3167 | { | |
3168 | int i; | |
3169 | ||
3170 | for (i = 0; i < nr_nodemap_entries; i++) { | |
3171 | unsigned long start_pfn = early_node_map[i].start_pfn; | |
3172 | unsigned long end_pfn = early_node_map[i].end_pfn; | |
3173 | ||
3174 | if (start_pfn <= pfn && pfn < end_pfn) | |
3175 | return early_node_map[i].nid; | |
3176 | } | |
3177 | /* This is a memory hole */ | |
3178 | return -1; | |
3179 | } | |
3180 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ | |
3181 | ||
3182 | int __meminit early_pfn_to_nid(unsigned long pfn) | |
3183 | { | |
3184 | int nid; | |
3185 | ||
3186 | nid = __early_pfn_to_nid(pfn); | |
3187 | if (nid >= 0) | |
3188 | return nid; | |
3189 | /* just returns 0 */ | |
3190 | return 0; | |
3191 | } | |
3192 | ||
3193 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES | |
3194 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) | |
3195 | { | |
3196 | int nid; | |
3197 | ||
3198 | nid = __early_pfn_to_nid(pfn); | |
3199 | if (nid >= 0 && nid != node) | |
3200 | return false; | |
3201 | return true; | |
3202 | } | |
3203 | #endif | |
3204 | ||
3205 | /* Basic iterator support to walk early_node_map[] */ | |
3206 | #define for_each_active_range_index_in_nid(i, nid) \ | |
3207 | for (i = first_active_region_index_in_nid(nid); i != -1; \ | |
3208 | i = next_active_region_index_in_nid(i, nid)) | |
3209 | ||
3210 | /** | |
3211 | * free_bootmem_with_active_regions - Call free_bootmem_node for each active range | |
3212 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. | |
3213 | * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node | |
3214 | * | |
3215 | * If an architecture guarantees that all ranges registered with | |
3216 | * add_active_ranges() contain no holes and may be freed, this | |
3217 | * this function may be used instead of calling free_bootmem() manually. | |
3218 | */ | |
3219 | void __init free_bootmem_with_active_regions(int nid, | |
3220 | unsigned long max_low_pfn) | |
3221 | { | |
3222 | int i; | |
3223 | ||
3224 | for_each_active_range_index_in_nid(i, nid) { | |
3225 | unsigned long size_pages = 0; | |
3226 | unsigned long end_pfn = early_node_map[i].end_pfn; | |
3227 | ||
3228 | if (early_node_map[i].start_pfn >= max_low_pfn) | |
3229 | continue; | |
3230 | ||
3231 | if (end_pfn > max_low_pfn) | |
3232 | end_pfn = max_low_pfn; | |
3233 | ||
3234 | size_pages = end_pfn - early_node_map[i].start_pfn; | |
3235 | free_bootmem_node(NODE_DATA(early_node_map[i].nid), | |
3236 | PFN_PHYS(early_node_map[i].start_pfn), | |
3237 | size_pages << PAGE_SHIFT); | |
3238 | } | |
3239 | } | |
3240 | ||
3241 | void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) | |
3242 | { | |
3243 | int i; | |
3244 | int ret; | |
3245 | ||
3246 | for_each_active_range_index_in_nid(i, nid) { | |
3247 | ret = work_fn(early_node_map[i].start_pfn, | |
3248 | early_node_map[i].end_pfn, data); | |
3249 | if (ret) | |
3250 | break; | |
3251 | } | |
3252 | } | |
3253 | /** | |
3254 | * sparse_memory_present_with_active_regions - Call memory_present for each active range | |
3255 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. | |
3256 | * | |
3257 | * If an architecture guarantees that all ranges registered with | |
3258 | * add_active_ranges() contain no holes and may be freed, this | |
3259 | * function may be used instead of calling memory_present() manually. | |
3260 | */ | |
3261 | void __init sparse_memory_present_with_active_regions(int nid) | |
3262 | { | |
3263 | int i; | |
3264 | ||
3265 | for_each_active_range_index_in_nid(i, nid) | |
3266 | memory_present(early_node_map[i].nid, | |
3267 | early_node_map[i].start_pfn, | |
3268 | early_node_map[i].end_pfn); | |
3269 | } | |
3270 | ||
3271 | /** | |
3272 | * get_pfn_range_for_nid - Return the start and end page frames for a node | |
3273 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. | |
3274 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. | |
3275 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. | |
3276 | * | |
3277 | * It returns the start and end page frame of a node based on information | |
3278 | * provided by an arch calling add_active_range(). If called for a node | |
3279 | * with no available memory, a warning is printed and the start and end | |
3280 | * PFNs will be 0. | |
3281 | */ | |
3282 | void __meminit get_pfn_range_for_nid(unsigned int nid, | |
3283 | unsigned long *start_pfn, unsigned long *end_pfn) | |
3284 | { | |
3285 | int i; | |
3286 | *start_pfn = -1UL; | |
3287 | *end_pfn = 0; | |
3288 | ||
3289 | for_each_active_range_index_in_nid(i, nid) { | |
3290 | *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); | |
3291 | *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); | |
3292 | } | |
3293 | ||
3294 | if (*start_pfn == -1UL) | |
3295 | *start_pfn = 0; | |
3296 | } | |
3297 | ||
3298 | /* | |
3299 | * This finds a zone that can be used for ZONE_MOVABLE pages. The | |
3300 | * assumption is made that zones within a node are ordered in monotonic | |
3301 | * increasing memory addresses so that the "highest" populated zone is used | |
3302 | */ | |
3303 | static void __init find_usable_zone_for_movable(void) | |
3304 | { | |
3305 | int zone_index; | |
3306 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { | |
3307 | if (zone_index == ZONE_MOVABLE) | |
3308 | continue; | |
3309 | ||
3310 | if (arch_zone_highest_possible_pfn[zone_index] > | |
3311 | arch_zone_lowest_possible_pfn[zone_index]) | |
3312 | break; | |
3313 | } | |
3314 | ||
3315 | VM_BUG_ON(zone_index == -1); | |
3316 | movable_zone = zone_index; | |
3317 | } | |
3318 | ||
3319 | /* | |
3320 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE | |
3321 | * because it is sized independant of architecture. Unlike the other zones, | |
3322 | * the starting point for ZONE_MOVABLE is not fixed. It may be different | |
3323 | * in each node depending on the size of each node and how evenly kernelcore | |
3324 | * is distributed. This helper function adjusts the zone ranges | |
3325 | * provided by the architecture for a given node by using the end of the | |
3326 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that | |
3327 | * zones within a node are in order of monotonic increases memory addresses | |
3328 | */ | |
3329 | static void __meminit adjust_zone_range_for_zone_movable(int nid, | |
3330 | unsigned long zone_type, | |
3331 | unsigned long node_start_pfn, | |
3332 | unsigned long node_end_pfn, | |
3333 | unsigned long *zone_start_pfn, | |
3334 | unsigned long *zone_end_pfn) | |
3335 | { | |
3336 | /* Only adjust if ZONE_MOVABLE is on this node */ | |
3337 | if (zone_movable_pfn[nid]) { | |
3338 | /* Size ZONE_MOVABLE */ | |
3339 | if (zone_type == ZONE_MOVABLE) { | |
3340 | *zone_start_pfn = zone_movable_pfn[nid]; | |
3341 | *zone_end_pfn = min(node_end_pfn, | |
3342 | arch_zone_highest_possible_pfn[movable_zone]); | |
3343 | ||
3344 | /* Adjust for ZONE_MOVABLE starting within this range */ | |
3345 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && | |
3346 | *zone_end_pfn > zone_movable_pfn[nid]) { | |
3347 | *zone_end_pfn = zone_movable_pfn[nid]; | |
3348 | ||
3349 | /* Check if this whole range is within ZONE_MOVABLE */ | |
3350 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) | |
3351 | *zone_start_pfn = *zone_end_pfn; | |
3352 | } | |
3353 | } | |
3354 | ||
3355 | /* | |
3356 | * Return the number of pages a zone spans in a node, including holes | |
3357 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() | |
3358 | */ | |
3359 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
3360 | unsigned long zone_type, | |
3361 | unsigned long *ignored) | |
3362 | { | |
3363 | unsigned long node_start_pfn, node_end_pfn; | |
3364 | unsigned long zone_start_pfn, zone_end_pfn; | |
3365 | ||
3366 | /* Get the start and end of the node and zone */ | |
3367 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
3368 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; | |
3369 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; | |
3370 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
3371 | node_start_pfn, node_end_pfn, | |
3372 | &zone_start_pfn, &zone_end_pfn); | |
3373 | ||
3374 | /* Check that this node has pages within the zone's required range */ | |
3375 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) | |
3376 | return 0; | |
3377 | ||
3378 | /* Move the zone boundaries inside the node if necessary */ | |
3379 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); | |
3380 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); | |
3381 | ||
3382 | /* Return the spanned pages */ | |
3383 | return zone_end_pfn - zone_start_pfn; | |
3384 | } | |
3385 | ||
3386 | /* | |
3387 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, | |
3388 | * then all holes in the requested range will be accounted for. | |
3389 | */ | |
3390 | static unsigned long __meminit __absent_pages_in_range(int nid, | |
3391 | unsigned long range_start_pfn, | |
3392 | unsigned long range_end_pfn) | |
3393 | { | |
3394 | int i = 0; | |
3395 | unsigned long prev_end_pfn = 0, hole_pages = 0; | |
3396 | unsigned long start_pfn; | |
3397 | ||
3398 | /* Find the end_pfn of the first active range of pfns in the node */ | |
3399 | i = first_active_region_index_in_nid(nid); | |
3400 | if (i == -1) | |
3401 | return 0; | |
3402 | ||
3403 | prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); | |
3404 | ||
3405 | /* Account for ranges before physical memory on this node */ | |
3406 | if (early_node_map[i].start_pfn > range_start_pfn) | |
3407 | hole_pages = prev_end_pfn - range_start_pfn; | |
3408 | ||
3409 | /* Find all holes for the zone within the node */ | |
3410 | for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { | |
3411 | ||
3412 | /* No need to continue if prev_end_pfn is outside the zone */ | |
3413 | if (prev_end_pfn >= range_end_pfn) | |
3414 | break; | |
3415 | ||
3416 | /* Make sure the end of the zone is not within the hole */ | |
3417 | start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); | |
3418 | prev_end_pfn = max(prev_end_pfn, range_start_pfn); | |
3419 | ||
3420 | /* Update the hole size cound and move on */ | |
3421 | if (start_pfn > range_start_pfn) { | |
3422 | BUG_ON(prev_end_pfn > start_pfn); | |
3423 | hole_pages += start_pfn - prev_end_pfn; | |
3424 | } | |
3425 | prev_end_pfn = early_node_map[i].end_pfn; | |
3426 | } | |
3427 | ||
3428 | /* Account for ranges past physical memory on this node */ | |
3429 | if (range_end_pfn > prev_end_pfn) | |
3430 | hole_pages += range_end_pfn - | |
3431 | max(range_start_pfn, prev_end_pfn); | |
3432 | ||
3433 | return hole_pages; | |
3434 | } | |
3435 | ||
3436 | /** | |
3437 | * absent_pages_in_range - Return number of page frames in holes within a range | |
3438 | * @start_pfn: The start PFN to start searching for holes | |
3439 | * @end_pfn: The end PFN to stop searching for holes | |
3440 | * | |
3441 | * It returns the number of pages frames in memory holes within a range. | |
3442 | */ | |
3443 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, | |
3444 | unsigned long end_pfn) | |
3445 | { | |
3446 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); | |
3447 | } | |
3448 | ||
3449 | /* Return the number of page frames in holes in a zone on a node */ | |
3450 | static unsigned long __meminit zone_absent_pages_in_node(int nid, | |
3451 | unsigned long zone_type, | |
3452 | unsigned long *ignored) | |
3453 | { | |
3454 | unsigned long node_start_pfn, node_end_pfn; | |
3455 | unsigned long zone_start_pfn, zone_end_pfn; | |
3456 | ||
3457 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
3458 | zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], | |
3459 | node_start_pfn); | |
3460 | zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], | |
3461 | node_end_pfn); | |
3462 | ||
3463 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
3464 | node_start_pfn, node_end_pfn, | |
3465 | &zone_start_pfn, &zone_end_pfn); | |
3466 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); | |
3467 | } | |
3468 | ||
3469 | #else | |
3470 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
3471 | unsigned long zone_type, | |
3472 | unsigned long *zones_size) | |
3473 | { | |
3474 | return zones_size[zone_type]; | |
3475 | } | |
3476 | ||
3477 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, | |
3478 | unsigned long zone_type, | |
3479 | unsigned long *zholes_size) | |
3480 | { | |
3481 | if (!zholes_size) | |
3482 | return 0; | |
3483 | ||
3484 | return zholes_size[zone_type]; | |
3485 | } | |
3486 | ||
3487 | #endif | |
3488 | ||
3489 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, | |
3490 | unsigned long *zones_size, unsigned long *zholes_size) | |
3491 | { | |
3492 | unsigned long realtotalpages, totalpages = 0; | |
3493 | enum zone_type i; | |
3494 | ||
3495 | for (i = 0; i < MAX_NR_ZONES; i++) | |
3496 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, | |
3497 | zones_size); | |
3498 | pgdat->node_spanned_pages = totalpages; | |
3499 | ||
3500 | realtotalpages = totalpages; | |
3501 | for (i = 0; i < MAX_NR_ZONES; i++) | |
3502 | realtotalpages -= | |
3503 | zone_absent_pages_in_node(pgdat->node_id, i, | |
3504 | zholes_size); | |
3505 | pgdat->node_present_pages = realtotalpages; | |
3506 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, | |
3507 | realtotalpages); | |
3508 | } | |
3509 | ||
3510 | #ifndef CONFIG_SPARSEMEM | |
3511 | /* | |
3512 | * Calculate the size of the zone->blockflags rounded to an unsigned long | |
3513 | * Start by making sure zonesize is a multiple of pageblock_order by rounding | |
3514 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally | |
3515 | * round what is now in bits to nearest long in bits, then return it in | |
3516 | * bytes. | |
3517 | */ | |
3518 | static unsigned long __init usemap_size(unsigned long zonesize) | |
3519 | { | |
3520 | unsigned long usemapsize; | |
3521 | ||
3522 | usemapsize = roundup(zonesize, pageblock_nr_pages); | |
3523 | usemapsize = usemapsize >> pageblock_order; | |
3524 | usemapsize *= NR_PAGEBLOCK_BITS; | |
3525 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); | |
3526 | ||
3527 | return usemapsize / 8; | |
3528 | } | |
3529 | ||
3530 | static void __init setup_usemap(struct pglist_data *pgdat, | |
3531 | struct zone *zone, unsigned long zonesize) | |
3532 | { | |
3533 | unsigned long usemapsize = usemap_size(zonesize); | |
3534 | zone->pageblock_flags = NULL; | |
3535 | if (usemapsize) | |
3536 | zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); | |
3537 | } | |
3538 | #else | |
3539 | static void inline setup_usemap(struct pglist_data *pgdat, | |
3540 | struct zone *zone, unsigned long zonesize) {} | |
3541 | #endif /* CONFIG_SPARSEMEM */ | |
3542 | ||
3543 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
3544 | ||
3545 | /* Return a sensible default order for the pageblock size. */ | |
3546 | static inline int pageblock_default_order(void) | |
3547 | { | |
3548 | if (HPAGE_SHIFT > PAGE_SHIFT) | |
3549 | return HUGETLB_PAGE_ORDER; | |
3550 | ||
3551 | return MAX_ORDER-1; | |
3552 | } | |
3553 | ||
3554 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ | |
3555 | static inline void __init set_pageblock_order(unsigned int order) | |
3556 | { | |
3557 | /* Check that pageblock_nr_pages has not already been setup */ | |
3558 | if (pageblock_order) | |
3559 | return; | |
3560 | ||
3561 | /* | |
3562 | * Assume the largest contiguous order of interest is a huge page. | |
3563 | * This value may be variable depending on boot parameters on IA64 | |
3564 | */ | |
3565 | pageblock_order = order; | |
3566 | } | |
3567 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
3568 | ||
3569 | /* | |
3570 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() | |
3571 | * and pageblock_default_order() are unused as pageblock_order is set | |
3572 | * at compile-time. See include/linux/pageblock-flags.h for the values of | |
3573 | * pageblock_order based on the kernel config | |
3574 | */ | |
3575 | static inline int pageblock_default_order(unsigned int order) | |
3576 | { | |
3577 | return MAX_ORDER-1; | |
3578 | } | |
3579 | #define set_pageblock_order(x) do {} while (0) | |
3580 | ||
3581 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
3582 | ||
3583 | /* | |
3584 | * Set up the zone data structures: | |
3585 | * - mark all pages reserved | |
3586 | * - mark all memory queues empty | |
3587 | * - clear the memory bitmaps | |
3588 | */ | |
3589 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, | |
3590 | unsigned long *zones_size, unsigned long *zholes_size) | |
3591 | { | |
3592 | enum zone_type j; | |
3593 | int nid = pgdat->node_id; | |
3594 | unsigned long zone_start_pfn = pgdat->node_start_pfn; | |
3595 | int ret; | |
3596 | ||
3597 | pgdat_resize_init(pgdat); | |
3598 | pgdat->nr_zones = 0; | |
3599 | init_waitqueue_head(&pgdat->kswapd_wait); | |
3600 | pgdat->kswapd_max_order = 0; | |
3601 | pgdat_page_cgroup_init(pgdat); | |
3602 | ||
3603 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
3604 | struct zone *zone = pgdat->node_zones + j; | |
3605 | unsigned long size, realsize, memmap_pages; | |
3606 | enum lru_list l; | |
3607 | ||
3608 | size = zone_spanned_pages_in_node(nid, j, zones_size); | |
3609 | realsize = size - zone_absent_pages_in_node(nid, j, | |
3610 | zholes_size); | |
3611 | ||
3612 | /* | |
3613 | * Adjust realsize so that it accounts for how much memory | |
3614 | * is used by this zone for memmap. This affects the watermark | |
3615 | * and per-cpu initialisations | |
3616 | */ | |
3617 | memmap_pages = | |
3618 | PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; | |
3619 | if (realsize >= memmap_pages) { | |
3620 | realsize -= memmap_pages; | |
3621 | if (memmap_pages) | |
3622 | printk(KERN_DEBUG | |
3623 | " %s zone: %lu pages used for memmap\n", | |
3624 | zone_names[j], memmap_pages); | |
3625 | } else | |
3626 | printk(KERN_WARNING | |
3627 | " %s zone: %lu pages exceeds realsize %lu\n", | |
3628 | zone_names[j], memmap_pages, realsize); | |
3629 | ||
3630 | /* Account for reserved pages */ | |
3631 | if (j == 0 && realsize > dma_reserve) { | |
3632 | realsize -= dma_reserve; | |
3633 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", | |
3634 | zone_names[0], dma_reserve); | |
3635 | } | |
3636 | ||
3637 | if (!is_highmem_idx(j)) | |
3638 | nr_kernel_pages += realsize; | |
3639 | nr_all_pages += realsize; | |
3640 | ||
3641 | zone->spanned_pages = size; | |
3642 | zone->present_pages = realsize; | |
3643 | #ifdef CONFIG_NUMA | |
3644 | zone->node = nid; | |
3645 | zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) | |
3646 | / 100; | |
3647 | zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; | |
3648 | #endif | |
3649 | zone->name = zone_names[j]; | |
3650 | spin_lock_init(&zone->lock); | |
3651 | spin_lock_init(&zone->lru_lock); | |
3652 | zone_seqlock_init(zone); | |
3653 | zone->zone_pgdat = pgdat; | |
3654 | ||
3655 | zone->prev_priority = DEF_PRIORITY; | |
3656 | ||
3657 | zone_pcp_init(zone); | |
3658 | for_each_lru(l) { | |
3659 | INIT_LIST_HEAD(&zone->lru[l].list); | |
3660 | zone->lru[l].nr_saved_scan = 0; | |
3661 | } | |
3662 | zone->reclaim_stat.recent_rotated[0] = 0; | |
3663 | zone->reclaim_stat.recent_rotated[1] = 0; | |
3664 | zone->reclaim_stat.recent_scanned[0] = 0; | |
3665 | zone->reclaim_stat.recent_scanned[1] = 0; | |
3666 | zap_zone_vm_stats(zone); | |
3667 | zone->flags = 0; | |
3668 | if (!size) | |
3669 | continue; | |
3670 | ||
3671 | set_pageblock_order(pageblock_default_order()); | |
3672 | setup_usemap(pgdat, zone, size); | |
3673 | ret = init_currently_empty_zone(zone, zone_start_pfn, | |
3674 | size, MEMMAP_EARLY); | |
3675 | BUG_ON(ret); | |
3676 | memmap_init(size, nid, j, zone_start_pfn); | |
3677 | zone_start_pfn += size; | |
3678 | } | |
3679 | } | |
3680 | ||
3681 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) | |
3682 | { | |
3683 | /* Skip empty nodes */ | |
3684 | if (!pgdat->node_spanned_pages) | |
3685 | return; | |
3686 | ||
3687 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
3688 | /* ia64 gets its own node_mem_map, before this, without bootmem */ | |
3689 | if (!pgdat->node_mem_map) { | |
3690 | unsigned long size, start, end; | |
3691 | struct page *map; | |
3692 | ||
3693 | /* | |
3694 | * The zone's endpoints aren't required to be MAX_ORDER | |
3695 | * aligned but the node_mem_map endpoints must be in order | |
3696 | * for the buddy allocator to function correctly. | |
3697 | */ | |
3698 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); | |
3699 | end = pgdat->node_start_pfn + pgdat->node_spanned_pages; | |
3700 | end = ALIGN(end, MAX_ORDER_NR_PAGES); | |
3701 | size = (end - start) * sizeof(struct page); | |
3702 | map = alloc_remap(pgdat->node_id, size); | |
3703 | if (!map) | |
3704 | map = alloc_bootmem_node(pgdat, size); | |
3705 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); | |
3706 | } | |
3707 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
3708 | /* | |
3709 | * With no DISCONTIG, the global mem_map is just set as node 0's | |
3710 | */ | |
3711 | if (pgdat == NODE_DATA(0)) { | |
3712 | mem_map = NODE_DATA(0)->node_mem_map; | |
3713 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP | |
3714 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) | |
3715 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); | |
3716 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ | |
3717 | } | |
3718 | #endif | |
3719 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ | |
3720 | } | |
3721 | ||
3722 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, | |
3723 | unsigned long node_start_pfn, unsigned long *zholes_size) | |
3724 | { | |
3725 | pg_data_t *pgdat = NODE_DATA(nid); | |
3726 | ||
3727 | pgdat->node_id = nid; | |
3728 | pgdat->node_start_pfn = node_start_pfn; | |
3729 | calculate_node_totalpages(pgdat, zones_size, zholes_size); | |
3730 | ||
3731 | alloc_node_mem_map(pgdat); | |
3732 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
3733 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", | |
3734 | nid, (unsigned long)pgdat, | |
3735 | (unsigned long)pgdat->node_mem_map); | |
3736 | #endif | |
3737 | ||
3738 | free_area_init_core(pgdat, zones_size, zholes_size); | |
3739 | } | |
3740 | ||
3741 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP | |
3742 | ||
3743 | #if MAX_NUMNODES > 1 | |
3744 | /* | |
3745 | * Figure out the number of possible node ids. | |
3746 | */ | |
3747 | static void __init setup_nr_node_ids(void) | |
3748 | { | |
3749 | unsigned int node; | |
3750 | unsigned int highest = 0; | |
3751 | ||
3752 | for_each_node_mask(node, node_possible_map) | |
3753 | highest = node; | |
3754 | nr_node_ids = highest + 1; | |
3755 | } | |
3756 | #else | |
3757 | static inline void setup_nr_node_ids(void) | |
3758 | { | |
3759 | } | |
3760 | #endif | |
3761 | ||
3762 | /** | |
3763 | * add_active_range - Register a range of PFNs backed by physical memory | |
3764 | * @nid: The node ID the range resides on | |
3765 | * @start_pfn: The start PFN of the available physical memory | |
3766 | * @end_pfn: The end PFN of the available physical memory | |
3767 | * | |
3768 | * These ranges are stored in an early_node_map[] and later used by | |
3769 | * free_area_init_nodes() to calculate zone sizes and holes. If the | |
3770 | * range spans a memory hole, it is up to the architecture to ensure | |
3771 | * the memory is not freed by the bootmem allocator. If possible | |
3772 | * the range being registered will be merged with existing ranges. | |
3773 | */ | |
3774 | void __init add_active_range(unsigned int nid, unsigned long start_pfn, | |
3775 | unsigned long end_pfn) | |
3776 | { | |
3777 | int i; | |
3778 | ||
3779 | mminit_dprintk(MMINIT_TRACE, "memory_register", | |
3780 | "Entering add_active_range(%d, %#lx, %#lx) " | |
3781 | "%d entries of %d used\n", | |
3782 | nid, start_pfn, end_pfn, | |
3783 | nr_nodemap_entries, MAX_ACTIVE_REGIONS); | |
3784 | ||
3785 | mminit_validate_memmodel_limits(&start_pfn, &end_pfn); | |
3786 | ||
3787 | /* Merge with existing active regions if possible */ | |
3788 | for (i = 0; i < nr_nodemap_entries; i++) { | |
3789 | if (early_node_map[i].nid != nid) | |
3790 | continue; | |
3791 | ||
3792 | /* Skip if an existing region covers this new one */ | |
3793 | if (start_pfn >= early_node_map[i].start_pfn && | |
3794 | end_pfn <= early_node_map[i].end_pfn) | |
3795 | return; | |
3796 | ||
3797 | /* Merge forward if suitable */ | |
3798 | if (start_pfn <= early_node_map[i].end_pfn && | |
3799 | end_pfn > early_node_map[i].end_pfn) { | |
3800 | early_node_map[i].end_pfn = end_pfn; | |
3801 | return; | |
3802 | } | |
3803 | ||
3804 | /* Merge backward if suitable */ | |
3805 | if (start_pfn < early_node_map[i].end_pfn && | |
3806 | end_pfn >= early_node_map[i].start_pfn) { | |
3807 | early_node_map[i].start_pfn = start_pfn; | |
3808 | return; | |
3809 | } | |
3810 | } | |
3811 | ||
3812 | /* Check that early_node_map is large enough */ | |
3813 | if (i >= MAX_ACTIVE_REGIONS) { | |
3814 | printk(KERN_CRIT "More than %d memory regions, truncating\n", | |
3815 | MAX_ACTIVE_REGIONS); | |
3816 | return; | |
3817 | } | |
3818 | ||
3819 | early_node_map[i].nid = nid; | |
3820 | early_node_map[i].start_pfn = start_pfn; | |
3821 | early_node_map[i].end_pfn = end_pfn; | |
3822 | nr_nodemap_entries = i + 1; | |
3823 | } | |
3824 | ||
3825 | /** | |
3826 | * remove_active_range - Shrink an existing registered range of PFNs | |
3827 | * @nid: The node id the range is on that should be shrunk | |
3828 | * @start_pfn: The new PFN of the range | |
3829 | * @end_pfn: The new PFN of the range | |
3830 | * | |
3831 | * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. | |
3832 | * The map is kept near the end physical page range that has already been | |
3833 | * registered. This function allows an arch to shrink an existing registered | |
3834 | * range. | |
3835 | */ | |
3836 | void __init remove_active_range(unsigned int nid, unsigned long start_pfn, | |
3837 | unsigned long end_pfn) | |
3838 | { | |
3839 | int i, j; | |
3840 | int removed = 0; | |
3841 | ||
3842 | printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", | |
3843 | nid, start_pfn, end_pfn); | |
3844 | ||
3845 | /* Find the old active region end and shrink */ | |
3846 | for_each_active_range_index_in_nid(i, nid) { | |
3847 | if (early_node_map[i].start_pfn >= start_pfn && | |
3848 | early_node_map[i].end_pfn <= end_pfn) { | |
3849 | /* clear it */ | |
3850 | early_node_map[i].start_pfn = 0; | |
3851 | early_node_map[i].end_pfn = 0; | |
3852 | removed = 1; | |
3853 | continue; | |
3854 | } | |
3855 | if (early_node_map[i].start_pfn < start_pfn && | |
3856 | early_node_map[i].end_pfn > start_pfn) { | |
3857 | unsigned long temp_end_pfn = early_node_map[i].end_pfn; | |
3858 | early_node_map[i].end_pfn = start_pfn; | |
3859 | if (temp_end_pfn > end_pfn) | |
3860 | add_active_range(nid, end_pfn, temp_end_pfn); | |
3861 | continue; | |
3862 | } | |
3863 | if (early_node_map[i].start_pfn >= start_pfn && | |
3864 | early_node_map[i].end_pfn > end_pfn && | |
3865 | early_node_map[i].start_pfn < end_pfn) { | |
3866 | early_node_map[i].start_pfn = end_pfn; | |
3867 | continue; | |
3868 | } | |
3869 | } | |
3870 | ||
3871 | if (!removed) | |
3872 | return; | |
3873 | ||
3874 | /* remove the blank ones */ | |
3875 | for (i = nr_nodemap_entries - 1; i > 0; i--) { | |
3876 | if (early_node_map[i].nid != nid) | |
3877 | continue; | |
3878 | if (early_node_map[i].end_pfn) | |
3879 | continue; | |
3880 | /* we found it, get rid of it */ | |
3881 | for (j = i; j < nr_nodemap_entries - 1; j++) | |
3882 | memcpy(&early_node_map[j], &early_node_map[j+1], | |
3883 | sizeof(early_node_map[j])); | |
3884 | j = nr_nodemap_entries - 1; | |
3885 | memset(&early_node_map[j], 0, sizeof(early_node_map[j])); | |
3886 | nr_nodemap_entries--; | |
3887 | } | |
3888 | } | |
3889 | ||
3890 | /** | |
3891 | * remove_all_active_ranges - Remove all currently registered regions | |
3892 | * | |
3893 | * During discovery, it may be found that a table like SRAT is invalid | |
3894 | * and an alternative discovery method must be used. This function removes | |
3895 | * all currently registered regions. | |
3896 | */ | |
3897 | void __init remove_all_active_ranges(void) | |
3898 | { | |
3899 | memset(early_node_map, 0, sizeof(early_node_map)); | |
3900 | nr_nodemap_entries = 0; | |
3901 | } | |
3902 | ||
3903 | /* Compare two active node_active_regions */ | |
3904 | static int __init cmp_node_active_region(const void *a, const void *b) | |
3905 | { | |
3906 | struct node_active_region *arange = (struct node_active_region *)a; | |
3907 | struct node_active_region *brange = (struct node_active_region *)b; | |
3908 | ||
3909 | /* Done this way to avoid overflows */ | |
3910 | if (arange->start_pfn > brange->start_pfn) | |
3911 | return 1; | |
3912 | if (arange->start_pfn < brange->start_pfn) | |
3913 | return -1; | |
3914 | ||
3915 | return 0; | |
3916 | } | |
3917 | ||
3918 | /* sort the node_map by start_pfn */ | |
3919 | static void __init sort_node_map(void) | |
3920 | { | |
3921 | sort(early_node_map, (size_t)nr_nodemap_entries, | |
3922 | sizeof(struct node_active_region), | |
3923 | cmp_node_active_region, NULL); | |
3924 | } | |
3925 | ||
3926 | /* Find the lowest pfn for a node */ | |
3927 | static unsigned long __init find_min_pfn_for_node(int nid) | |
3928 | { | |
3929 | int i; | |
3930 | unsigned long min_pfn = ULONG_MAX; | |
3931 | ||
3932 | /* Assuming a sorted map, the first range found has the starting pfn */ | |
3933 | for_each_active_range_index_in_nid(i, nid) | |
3934 | min_pfn = min(min_pfn, early_node_map[i].start_pfn); | |
3935 | ||
3936 | if (min_pfn == ULONG_MAX) { | |
3937 | printk(KERN_WARNING | |
3938 | "Could not find start_pfn for node %d\n", nid); | |
3939 | return 0; | |
3940 | } | |
3941 | ||
3942 | return min_pfn; | |
3943 | } | |
3944 | ||
3945 | /** | |
3946 | * find_min_pfn_with_active_regions - Find the minimum PFN registered | |
3947 | * | |
3948 | * It returns the minimum PFN based on information provided via | |
3949 | * add_active_range(). | |
3950 | */ | |
3951 | unsigned long __init find_min_pfn_with_active_regions(void) | |
3952 | { | |
3953 | return find_min_pfn_for_node(MAX_NUMNODES); | |
3954 | } | |
3955 | ||
3956 | /* | |
3957 | * early_calculate_totalpages() | |
3958 | * Sum pages in active regions for movable zone. | |
3959 | * Populate N_HIGH_MEMORY for calculating usable_nodes. | |
3960 | */ | |
3961 | static unsigned long __init early_calculate_totalpages(void) | |
3962 | { | |
3963 | int i; | |
3964 | unsigned long totalpages = 0; | |
3965 | ||
3966 | for (i = 0; i < nr_nodemap_entries; i++) { | |
3967 | unsigned long pages = early_node_map[i].end_pfn - | |
3968 | early_node_map[i].start_pfn; | |
3969 | totalpages += pages; | |
3970 | if (pages) | |
3971 | node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); | |
3972 | } | |
3973 | return totalpages; | |
3974 | } | |
3975 | ||
3976 | /* | |
3977 | * Find the PFN the Movable zone begins in each node. Kernel memory | |
3978 | * is spread evenly between nodes as long as the nodes have enough | |
3979 | * memory. When they don't, some nodes will have more kernelcore than | |
3980 | * others | |
3981 | */ | |
3982 | static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) | |
3983 | { | |
3984 | int i, nid; | |
3985 | unsigned long usable_startpfn; | |
3986 | unsigned long kernelcore_node, kernelcore_remaining; | |
3987 | unsigned long totalpages = early_calculate_totalpages(); | |
3988 | int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); | |
3989 | ||
3990 | /* | |
3991 | * If movablecore was specified, calculate what size of | |
3992 | * kernelcore that corresponds so that memory usable for | |
3993 | * any allocation type is evenly spread. If both kernelcore | |
3994 | * and movablecore are specified, then the value of kernelcore | |
3995 | * will be used for required_kernelcore if it's greater than | |
3996 | * what movablecore would have allowed. | |
3997 | */ | |
3998 | if (required_movablecore) { | |
3999 | unsigned long corepages; | |
4000 | ||
4001 | /* | |
4002 | * Round-up so that ZONE_MOVABLE is at least as large as what | |
4003 | * was requested by the user | |
4004 | */ | |
4005 | required_movablecore = | |
4006 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); | |
4007 | corepages = totalpages - required_movablecore; | |
4008 | ||
4009 | required_kernelcore = max(required_kernelcore, corepages); | |
4010 | } | |
4011 | ||
4012 | /* If kernelcore was not specified, there is no ZONE_MOVABLE */ | |
4013 | if (!required_kernelcore) | |
4014 | return; | |
4015 | ||
4016 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ | |
4017 | find_usable_zone_for_movable(); | |
4018 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; | |
4019 | ||
4020 | restart: | |
4021 | /* Spread kernelcore memory as evenly as possible throughout nodes */ | |
4022 | kernelcore_node = required_kernelcore / usable_nodes; | |
4023 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4024 | /* | |
4025 | * Recalculate kernelcore_node if the division per node | |
4026 | * now exceeds what is necessary to satisfy the requested | |
4027 | * amount of memory for the kernel | |
4028 | */ | |
4029 | if (required_kernelcore < kernelcore_node) | |
4030 | kernelcore_node = required_kernelcore / usable_nodes; | |
4031 | ||
4032 | /* | |
4033 | * As the map is walked, we track how much memory is usable | |
4034 | * by the kernel using kernelcore_remaining. When it is | |
4035 | * 0, the rest of the node is usable by ZONE_MOVABLE | |
4036 | */ | |
4037 | kernelcore_remaining = kernelcore_node; | |
4038 | ||
4039 | /* Go through each range of PFNs within this node */ | |
4040 | for_each_active_range_index_in_nid(i, nid) { | |
4041 | unsigned long start_pfn, end_pfn; | |
4042 | unsigned long size_pages; | |
4043 | ||
4044 | start_pfn = max(early_node_map[i].start_pfn, | |
4045 | zone_movable_pfn[nid]); | |
4046 | end_pfn = early_node_map[i].end_pfn; | |
4047 | if (start_pfn >= end_pfn) | |
4048 | continue; | |
4049 | ||
4050 | /* Account for what is only usable for kernelcore */ | |
4051 | if (start_pfn < usable_startpfn) { | |
4052 | unsigned long kernel_pages; | |
4053 | kernel_pages = min(end_pfn, usable_startpfn) | |
4054 | - start_pfn; | |
4055 | ||
4056 | kernelcore_remaining -= min(kernel_pages, | |
4057 | kernelcore_remaining); | |
4058 | required_kernelcore -= min(kernel_pages, | |
4059 | required_kernelcore); | |
4060 | ||
4061 | /* Continue if range is now fully accounted */ | |
4062 | if (end_pfn <= usable_startpfn) { | |
4063 | ||
4064 | /* | |
4065 | * Push zone_movable_pfn to the end so | |
4066 | * that if we have to rebalance | |
4067 | * kernelcore across nodes, we will | |
4068 | * not double account here | |
4069 | */ | |
4070 | zone_movable_pfn[nid] = end_pfn; | |
4071 | continue; | |
4072 | } | |
4073 | start_pfn = usable_startpfn; | |
4074 | } | |
4075 | ||
4076 | /* | |
4077 | * The usable PFN range for ZONE_MOVABLE is from | |
4078 | * start_pfn->end_pfn. Calculate size_pages as the | |
4079 | * number of pages used as kernelcore | |
4080 | */ | |
4081 | size_pages = end_pfn - start_pfn; | |
4082 | if (size_pages > kernelcore_remaining) | |
4083 | size_pages = kernelcore_remaining; | |
4084 | zone_movable_pfn[nid] = start_pfn + size_pages; | |
4085 | ||
4086 | /* | |
4087 | * Some kernelcore has been met, update counts and | |
4088 | * break if the kernelcore for this node has been | |
4089 | * satisified | |
4090 | */ | |
4091 | required_kernelcore -= min(required_kernelcore, | |
4092 | size_pages); | |
4093 | kernelcore_remaining -= size_pages; | |
4094 | if (!kernelcore_remaining) | |
4095 | break; | |
4096 | } | |
4097 | } | |
4098 | ||
4099 | /* | |
4100 | * If there is still required_kernelcore, we do another pass with one | |
4101 | * less node in the count. This will push zone_movable_pfn[nid] further | |
4102 | * along on the nodes that still have memory until kernelcore is | |
4103 | * satisified | |
4104 | */ | |
4105 | usable_nodes--; | |
4106 | if (usable_nodes && required_kernelcore > usable_nodes) | |
4107 | goto restart; | |
4108 | ||
4109 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ | |
4110 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
4111 | zone_movable_pfn[nid] = | |
4112 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); | |
4113 | } | |
4114 | ||
4115 | /* Any regular memory on that node ? */ | |
4116 | static void check_for_regular_memory(pg_data_t *pgdat) | |
4117 | { | |
4118 | #ifdef CONFIG_HIGHMEM | |
4119 | enum zone_type zone_type; | |
4120 | ||
4121 | for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { | |
4122 | struct zone *zone = &pgdat->node_zones[zone_type]; | |
4123 | if (zone->present_pages) | |
4124 | node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); | |
4125 | } | |
4126 | #endif | |
4127 | } | |
4128 | ||
4129 | /** | |
4130 | * free_area_init_nodes - Initialise all pg_data_t and zone data | |
4131 | * @max_zone_pfn: an array of max PFNs for each zone | |
4132 | * | |
4133 | * This will call free_area_init_node() for each active node in the system. | |
4134 | * Using the page ranges provided by add_active_range(), the size of each | |
4135 | * zone in each node and their holes is calculated. If the maximum PFN | |
4136 | * between two adjacent zones match, it is assumed that the zone is empty. | |
4137 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed | |
4138 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone | |
4139 | * starts where the previous one ended. For example, ZONE_DMA32 starts | |
4140 | * at arch_max_dma_pfn. | |
4141 | */ | |
4142 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) | |
4143 | { | |
4144 | unsigned long nid; | |
4145 | int i; | |
4146 | ||
4147 | /* Sort early_node_map as initialisation assumes it is sorted */ | |
4148 | sort_node_map(); | |
4149 | ||
4150 | /* Record where the zone boundaries are */ | |
4151 | memset(arch_zone_lowest_possible_pfn, 0, | |
4152 | sizeof(arch_zone_lowest_possible_pfn)); | |
4153 | memset(arch_zone_highest_possible_pfn, 0, | |
4154 | sizeof(arch_zone_highest_possible_pfn)); | |
4155 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); | |
4156 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; | |
4157 | for (i = 1; i < MAX_NR_ZONES; i++) { | |
4158 | if (i == ZONE_MOVABLE) | |
4159 | continue; | |
4160 | arch_zone_lowest_possible_pfn[i] = | |
4161 | arch_zone_highest_possible_pfn[i-1]; | |
4162 | arch_zone_highest_possible_pfn[i] = | |
4163 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); | |
4164 | } | |
4165 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; | |
4166 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; | |
4167 | ||
4168 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ | |
4169 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); | |
4170 | find_zone_movable_pfns_for_nodes(zone_movable_pfn); | |
4171 | ||
4172 | /* Print out the zone ranges */ | |
4173 | printk("Zone PFN ranges:\n"); | |
4174 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
4175 | if (i == ZONE_MOVABLE) | |
4176 | continue; | |
4177 | printk(" %-8s %0#10lx -> %0#10lx\n", | |
4178 | zone_names[i], | |
4179 | arch_zone_lowest_possible_pfn[i], | |
4180 | arch_zone_highest_possible_pfn[i]); | |
4181 | } | |
4182 | ||
4183 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ | |
4184 | printk("Movable zone start PFN for each node\n"); | |
4185 | for (i = 0; i < MAX_NUMNODES; i++) { | |
4186 | if (zone_movable_pfn[i]) | |
4187 | printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); | |
4188 | } | |
4189 | ||
4190 | /* Print out the early_node_map[] */ | |
4191 | printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); | |
4192 | for (i = 0; i < nr_nodemap_entries; i++) | |
4193 | printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, | |
4194 | early_node_map[i].start_pfn, | |
4195 | early_node_map[i].end_pfn); | |
4196 | ||
4197 | /* Initialise every node */ | |
4198 | mminit_verify_pageflags_layout(); | |
4199 | setup_nr_node_ids(); | |
4200 | for_each_online_node(nid) { | |
4201 | pg_data_t *pgdat = NODE_DATA(nid); | |
4202 | free_area_init_node(nid, NULL, | |
4203 | find_min_pfn_for_node(nid), NULL); | |
4204 | ||
4205 | /* Any memory on that node */ | |
4206 | if (pgdat->node_present_pages) | |
4207 | node_set_state(nid, N_HIGH_MEMORY); | |
4208 | check_for_regular_memory(pgdat); | |
4209 | } | |
4210 | } | |
4211 | ||
4212 | static int __init cmdline_parse_core(char *p, unsigned long *core) | |
4213 | { | |
4214 | unsigned long long coremem; | |
4215 | if (!p) | |
4216 | return -EINVAL; | |
4217 | ||
4218 | coremem = memparse(p, &p); | |
4219 | *core = coremem >> PAGE_SHIFT; | |
4220 | ||
4221 | /* Paranoid check that UL is enough for the coremem value */ | |
4222 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); | |
4223 | ||
4224 | return 0; | |
4225 | } | |
4226 | ||
4227 | /* | |
4228 | * kernelcore=size sets the amount of memory for use for allocations that | |
4229 | * cannot be reclaimed or migrated. | |
4230 | */ | |
4231 | static int __init cmdline_parse_kernelcore(char *p) | |
4232 | { | |
4233 | return cmdline_parse_core(p, &required_kernelcore); | |
4234 | } | |
4235 | ||
4236 | /* | |
4237 | * movablecore=size sets the amount of memory for use for allocations that | |
4238 | * can be reclaimed or migrated. | |
4239 | */ | |
4240 | static int __init cmdline_parse_movablecore(char *p) | |
4241 | { | |
4242 | return cmdline_parse_core(p, &required_movablecore); | |
4243 | } | |
4244 | ||
4245 | early_param("kernelcore", cmdline_parse_kernelcore); | |
4246 | early_param("movablecore", cmdline_parse_movablecore); | |
4247 | ||
4248 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ | |
4249 | ||
4250 | /** | |
4251 | * set_dma_reserve - set the specified number of pages reserved in the first zone | |
4252 | * @new_dma_reserve: The number of pages to mark reserved | |
4253 | * | |
4254 | * The per-cpu batchsize and zone watermarks are determined by present_pages. | |
4255 | * In the DMA zone, a significant percentage may be consumed by kernel image | |
4256 | * and other unfreeable allocations which can skew the watermarks badly. This | |
4257 | * function may optionally be used to account for unfreeable pages in the | |
4258 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and | |
4259 | * smaller per-cpu batchsize. | |
4260 | */ | |
4261 | void __init set_dma_reserve(unsigned long new_dma_reserve) | |
4262 | { | |
4263 | dma_reserve = new_dma_reserve; | |
4264 | } | |
4265 | ||
4266 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
4267 | struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; | |
4268 | EXPORT_SYMBOL(contig_page_data); | |
4269 | #endif | |
4270 | ||
4271 | void __init free_area_init(unsigned long *zones_size) | |
4272 | { | |
4273 | free_area_init_node(0, zones_size, | |
4274 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); | |
4275 | } | |
4276 | ||
4277 | static int page_alloc_cpu_notify(struct notifier_block *self, | |
4278 | unsigned long action, void *hcpu) | |
4279 | { | |
4280 | int cpu = (unsigned long)hcpu; | |
4281 | ||
4282 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { | |
4283 | drain_pages(cpu); | |
4284 | ||
4285 | /* | |
4286 | * Spill the event counters of the dead processor | |
4287 | * into the current processors event counters. | |
4288 | * This artificially elevates the count of the current | |
4289 | * processor. | |
4290 | */ | |
4291 | vm_events_fold_cpu(cpu); | |
4292 | ||
4293 | /* | |
4294 | * Zero the differential counters of the dead processor | |
4295 | * so that the vm statistics are consistent. | |
4296 | * | |
4297 | * This is only okay since the processor is dead and cannot | |
4298 | * race with what we are doing. | |
4299 | */ | |
4300 | refresh_cpu_vm_stats(cpu); | |
4301 | } | |
4302 | return NOTIFY_OK; | |
4303 | } | |
4304 | ||
4305 | void __init page_alloc_init(void) | |
4306 | { | |
4307 | hotcpu_notifier(page_alloc_cpu_notify, 0); | |
4308 | } | |
4309 | ||
4310 | /* | |
4311 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio | |
4312 | * or min_free_kbytes changes. | |
4313 | */ | |
4314 | static void calculate_totalreserve_pages(void) | |
4315 | { | |
4316 | struct pglist_data *pgdat; | |
4317 | unsigned long reserve_pages = 0; | |
4318 | enum zone_type i, j; | |
4319 | ||
4320 | for_each_online_pgdat(pgdat) { | |
4321 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
4322 | struct zone *zone = pgdat->node_zones + i; | |
4323 | unsigned long max = 0; | |
4324 | ||
4325 | /* Find valid and maximum lowmem_reserve in the zone */ | |
4326 | for (j = i; j < MAX_NR_ZONES; j++) { | |
4327 | if (zone->lowmem_reserve[j] > max) | |
4328 | max = zone->lowmem_reserve[j]; | |
4329 | } | |
4330 | ||
4331 | /* we treat the high watermark as reserved pages. */ | |
4332 | max += high_wmark_pages(zone); | |
4333 | ||
4334 | if (max > zone->present_pages) | |
4335 | max = zone->present_pages; | |
4336 | reserve_pages += max; | |
4337 | } | |
4338 | } | |
4339 | totalreserve_pages = reserve_pages; | |
4340 | } | |
4341 | ||
4342 | /* | |
4343 | * setup_per_zone_lowmem_reserve - called whenever | |
4344 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone | |
4345 | * has a correct pages reserved value, so an adequate number of | |
4346 | * pages are left in the zone after a successful __alloc_pages(). | |
4347 | */ | |
4348 | static void setup_per_zone_lowmem_reserve(void) | |
4349 | { | |
4350 | struct pglist_data *pgdat; | |
4351 | enum zone_type j, idx; | |
4352 | ||
4353 | for_each_online_pgdat(pgdat) { | |
4354 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
4355 | struct zone *zone = pgdat->node_zones + j; | |
4356 | unsigned long present_pages = zone->present_pages; | |
4357 | ||
4358 | zone->lowmem_reserve[j] = 0; | |
4359 | ||
4360 | idx = j; | |
4361 | while (idx) { | |
4362 | struct zone *lower_zone; | |
4363 | ||
4364 | idx--; | |
4365 | ||
4366 | if (sysctl_lowmem_reserve_ratio[idx] < 1) | |
4367 | sysctl_lowmem_reserve_ratio[idx] = 1; | |
4368 | ||
4369 | lower_zone = pgdat->node_zones + idx; | |
4370 | lower_zone->lowmem_reserve[j] = present_pages / | |
4371 | sysctl_lowmem_reserve_ratio[idx]; | |
4372 | present_pages += lower_zone->present_pages; | |
4373 | } | |
4374 | } | |
4375 | } | |
4376 | ||
4377 | /* update totalreserve_pages */ | |
4378 | calculate_totalreserve_pages(); | |
4379 | } | |
4380 | ||
4381 | /** | |
4382 | * setup_per_zone_pages_min - called when min_free_kbytes changes. | |
4383 | * | |
4384 | * Ensures that the pages_{min,low,high} values for each zone are set correctly | |
4385 | * with respect to min_free_kbytes. | |
4386 | */ | |
4387 | void setup_per_zone_pages_min(void) | |
4388 | { | |
4389 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
4390 | unsigned long lowmem_pages = 0; | |
4391 | struct zone *zone; | |
4392 | unsigned long flags; | |
4393 | ||
4394 | /* Calculate total number of !ZONE_HIGHMEM pages */ | |
4395 | for_each_zone(zone) { | |
4396 | if (!is_highmem(zone)) | |
4397 | lowmem_pages += zone->present_pages; | |
4398 | } | |
4399 | ||
4400 | for_each_zone(zone) { | |
4401 | u64 tmp; | |
4402 | ||
4403 | spin_lock_irqsave(&zone->lock, flags); | |
4404 | tmp = (u64)pages_min * zone->present_pages; | |
4405 | do_div(tmp, lowmem_pages); | |
4406 | if (is_highmem(zone)) { | |
4407 | /* | |
4408 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't | |
4409 | * need highmem pages, so cap pages_min to a small | |
4410 | * value here. | |
4411 | * | |
4412 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) | |
4413 | * deltas controls asynch page reclaim, and so should | |
4414 | * not be capped for highmem. | |
4415 | */ | |
4416 | int min_pages; | |
4417 | ||
4418 | min_pages = zone->present_pages / 1024; | |
4419 | if (min_pages < SWAP_CLUSTER_MAX) | |
4420 | min_pages = SWAP_CLUSTER_MAX; | |
4421 | if (min_pages > 128) | |
4422 | min_pages = 128; | |
4423 | zone->watermark[WMARK_MIN] = min_pages; | |
4424 | } else { | |
4425 | /* | |
4426 | * If it's a lowmem zone, reserve a number of pages | |
4427 | * proportionate to the zone's size. | |
4428 | */ | |
4429 | zone->watermark[WMARK_MIN] = tmp; | |
4430 | } | |
4431 | ||
4432 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); | |
4433 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); | |
4434 | setup_zone_migrate_reserve(zone); | |
4435 | spin_unlock_irqrestore(&zone->lock, flags); | |
4436 | } | |
4437 | ||
4438 | /* update totalreserve_pages */ | |
4439 | calculate_totalreserve_pages(); | |
4440 | } | |
4441 | ||
4442 | /** | |
4443 | * The inactive anon list should be small enough that the VM never has to | |
4444 | * do too much work, but large enough that each inactive page has a chance | |
4445 | * to be referenced again before it is swapped out. | |
4446 | * | |
4447 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to | |
4448 | * INACTIVE_ANON pages on this zone's LRU, maintained by the | |
4449 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of | |
4450 | * the anonymous pages are kept on the inactive list. | |
4451 | * | |
4452 | * total target max | |
4453 | * memory ratio inactive anon | |
4454 | * ------------------------------------- | |
4455 | * 10MB 1 5MB | |
4456 | * 100MB 1 50MB | |
4457 | * 1GB 3 250MB | |
4458 | * 10GB 10 0.9GB | |
4459 | * 100GB 31 3GB | |
4460 | * 1TB 101 10GB | |
4461 | * 10TB 320 32GB | |
4462 | */ | |
4463 | static void __init setup_per_zone_inactive_ratio(void) | |
4464 | { | |
4465 | struct zone *zone; | |
4466 | ||
4467 | for_each_zone(zone) { | |
4468 | unsigned int gb, ratio; | |
4469 | ||
4470 | /* Zone size in gigabytes */ | |
4471 | gb = zone->present_pages >> (30 - PAGE_SHIFT); | |
4472 | if (gb) | |
4473 | ratio = int_sqrt(10 * gb); | |
4474 | else | |
4475 | ratio = 1; | |
4476 | ||
4477 | zone->inactive_ratio = ratio; | |
4478 | } | |
4479 | } | |
4480 | ||
4481 | /* | |
4482 | * Initialise min_free_kbytes. | |
4483 | * | |
4484 | * For small machines we want it small (128k min). For large machines | |
4485 | * we want it large (64MB max). But it is not linear, because network | |
4486 | * bandwidth does not increase linearly with machine size. We use | |
4487 | * | |
4488 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: | |
4489 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) | |
4490 | * | |
4491 | * which yields | |
4492 | * | |
4493 | * 16MB: 512k | |
4494 | * 32MB: 724k | |
4495 | * 64MB: 1024k | |
4496 | * 128MB: 1448k | |
4497 | * 256MB: 2048k | |
4498 | * 512MB: 2896k | |
4499 | * 1024MB: 4096k | |
4500 | * 2048MB: 5792k | |
4501 | * 4096MB: 8192k | |
4502 | * 8192MB: 11584k | |
4503 | * 16384MB: 16384k | |
4504 | */ | |
4505 | static int __init init_per_zone_pages_min(void) | |
4506 | { | |
4507 | unsigned long lowmem_kbytes; | |
4508 | ||
4509 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
4510 | ||
4511 | min_free_kbytes = int_sqrt(lowmem_kbytes * 16); | |
4512 | if (min_free_kbytes < 128) | |
4513 | min_free_kbytes = 128; | |
4514 | if (min_free_kbytes > 65536) | |
4515 | min_free_kbytes = 65536; | |
4516 | setup_per_zone_pages_min(); | |
4517 | setup_per_zone_lowmem_reserve(); | |
4518 | setup_per_zone_inactive_ratio(); | |
4519 | return 0; | |
4520 | } | |
4521 | module_init(init_per_zone_pages_min) | |
4522 | ||
4523 | /* | |
4524 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so | |
4525 | * that we can call two helper functions whenever min_free_kbytes | |
4526 | * changes. | |
4527 | */ | |
4528 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, | |
4529 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
4530 | { | |
4531 | proc_dointvec(table, write, file, buffer, length, ppos); | |
4532 | if (write) | |
4533 | setup_per_zone_pages_min(); | |
4534 | return 0; | |
4535 | } | |
4536 | ||
4537 | #ifdef CONFIG_NUMA | |
4538 | int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, | |
4539 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
4540 | { | |
4541 | struct zone *zone; | |
4542 | int rc; | |
4543 | ||
4544 | rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); | |
4545 | if (rc) | |
4546 | return rc; | |
4547 | ||
4548 | for_each_zone(zone) | |
4549 | zone->min_unmapped_pages = (zone->present_pages * | |
4550 | sysctl_min_unmapped_ratio) / 100; | |
4551 | return 0; | |
4552 | } | |
4553 | ||
4554 | int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, | |
4555 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
4556 | { | |
4557 | struct zone *zone; | |
4558 | int rc; | |
4559 | ||
4560 | rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); | |
4561 | if (rc) | |
4562 | return rc; | |
4563 | ||
4564 | for_each_zone(zone) | |
4565 | zone->min_slab_pages = (zone->present_pages * | |
4566 | sysctl_min_slab_ratio) / 100; | |
4567 | return 0; | |
4568 | } | |
4569 | #endif | |
4570 | ||
4571 | /* | |
4572 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
4573 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
4574 | * whenever sysctl_lowmem_reserve_ratio changes. | |
4575 | * | |
4576 | * The reserve ratio obviously has absolutely no relation with the | |
4577 | * minimum watermarks. The lowmem reserve ratio can only make sense | |
4578 | * if in function of the boot time zone sizes. | |
4579 | */ | |
4580 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, | |
4581 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
4582 | { | |
4583 | proc_dointvec_minmax(table, write, file, buffer, length, ppos); | |
4584 | setup_per_zone_lowmem_reserve(); | |
4585 | return 0; | |
4586 | } | |
4587 | ||
4588 | /* | |
4589 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each | |
4590 | * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist | |
4591 | * can have before it gets flushed back to buddy allocator. | |
4592 | */ | |
4593 | ||
4594 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, | |
4595 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) | |
4596 | { | |
4597 | struct zone *zone; | |
4598 | unsigned int cpu; | |
4599 | int ret; | |
4600 | ||
4601 | ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); | |
4602 | if (!write || (ret == -EINVAL)) | |
4603 | return ret; | |
4604 | for_each_zone(zone) { | |
4605 | for_each_online_cpu(cpu) { | |
4606 | unsigned long high; | |
4607 | high = zone->present_pages / percpu_pagelist_fraction; | |
4608 | setup_pagelist_highmark(zone_pcp(zone, cpu), high); | |
4609 | } | |
4610 | } | |
4611 | return 0; | |
4612 | } | |
4613 | ||
4614 | int hashdist = HASHDIST_DEFAULT; | |
4615 | ||
4616 | #ifdef CONFIG_NUMA | |
4617 | static int __init set_hashdist(char *str) | |
4618 | { | |
4619 | if (!str) | |
4620 | return 0; | |
4621 | hashdist = simple_strtoul(str, &str, 0); | |
4622 | return 1; | |
4623 | } | |
4624 | __setup("hashdist=", set_hashdist); | |
4625 | #endif | |
4626 | ||
4627 | /* | |
4628 | * allocate a large system hash table from bootmem | |
4629 | * - it is assumed that the hash table must contain an exact power-of-2 | |
4630 | * quantity of entries | |
4631 | * - limit is the number of hash buckets, not the total allocation size | |
4632 | */ | |
4633 | void *__init alloc_large_system_hash(const char *tablename, | |
4634 | unsigned long bucketsize, | |
4635 | unsigned long numentries, | |
4636 | int scale, | |
4637 | int flags, | |
4638 | unsigned int *_hash_shift, | |
4639 | unsigned int *_hash_mask, | |
4640 | unsigned long limit) | |
4641 | { | |
4642 | unsigned long long max = limit; | |
4643 | unsigned long log2qty, size; | |
4644 | void *table = NULL; | |
4645 | ||
4646 | /* allow the kernel cmdline to have a say */ | |
4647 | if (!numentries) { | |
4648 | /* round applicable memory size up to nearest megabyte */ | |
4649 | numentries = nr_kernel_pages; | |
4650 | numentries += (1UL << (20 - PAGE_SHIFT)) - 1; | |
4651 | numentries >>= 20 - PAGE_SHIFT; | |
4652 | numentries <<= 20 - PAGE_SHIFT; | |
4653 | ||
4654 | /* limit to 1 bucket per 2^scale bytes of low memory */ | |
4655 | if (scale > PAGE_SHIFT) | |
4656 | numentries >>= (scale - PAGE_SHIFT); | |
4657 | else | |
4658 | numentries <<= (PAGE_SHIFT - scale); | |
4659 | ||
4660 | /* Make sure we've got at least a 0-order allocation.. */ | |
4661 | if (unlikely((numentries * bucketsize) < PAGE_SIZE)) | |
4662 | numentries = PAGE_SIZE / bucketsize; | |
4663 | } | |
4664 | numentries = roundup_pow_of_two(numentries); | |
4665 | ||
4666 | /* limit allocation size to 1/16 total memory by default */ | |
4667 | if (max == 0) { | |
4668 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; | |
4669 | do_div(max, bucketsize); | |
4670 | } | |
4671 | ||
4672 | if (numentries > max) | |
4673 | numentries = max; | |
4674 | ||
4675 | log2qty = ilog2(numentries); | |
4676 | ||
4677 | do { | |
4678 | size = bucketsize << log2qty; | |
4679 | if (flags & HASH_EARLY) | |
4680 | table = alloc_bootmem_nopanic(size); | |
4681 | else if (hashdist) | |
4682 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); | |
4683 | else { | |
4684 | /* | |
4685 | * If bucketsize is not a power-of-two, we may free | |
4686 | * some pages at the end of hash table which | |
4687 | * alloc_pages_exact() automatically does | |
4688 | */ | |
4689 | if (get_order(size) < MAX_ORDER) | |
4690 | table = alloc_pages_exact(size, GFP_ATOMIC); | |
4691 | } | |
4692 | } while (!table && size > PAGE_SIZE && --log2qty); | |
4693 | ||
4694 | if (!table) | |
4695 | panic("Failed to allocate %s hash table\n", tablename); | |
4696 | ||
4697 | printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", | |
4698 | tablename, | |
4699 | (1U << log2qty), | |
4700 | ilog2(size) - PAGE_SHIFT, | |
4701 | size); | |
4702 | ||
4703 | if (_hash_shift) | |
4704 | *_hash_shift = log2qty; | |
4705 | if (_hash_mask) | |
4706 | *_hash_mask = (1 << log2qty) - 1; | |
4707 | ||
4708 | /* | |
4709 | * If hashdist is set, the table allocation is done with __vmalloc() | |
4710 | * which invokes the kmemleak_alloc() callback. This function may also | |
4711 | * be called before the slab and kmemleak are initialised when | |
4712 | * kmemleak simply buffers the request to be executed later | |
4713 | * (GFP_ATOMIC flag ignored in this case). | |
4714 | */ | |
4715 | if (!hashdist) | |
4716 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); | |
4717 | ||
4718 | return table; | |
4719 | } | |
4720 | ||
4721 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ | |
4722 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, | |
4723 | unsigned long pfn) | |
4724 | { | |
4725 | #ifdef CONFIG_SPARSEMEM | |
4726 | return __pfn_to_section(pfn)->pageblock_flags; | |
4727 | #else | |
4728 | return zone->pageblock_flags; | |
4729 | #endif /* CONFIG_SPARSEMEM */ | |
4730 | } | |
4731 | ||
4732 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) | |
4733 | { | |
4734 | #ifdef CONFIG_SPARSEMEM | |
4735 | pfn &= (PAGES_PER_SECTION-1); | |
4736 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
4737 | #else | |
4738 | pfn = pfn - zone->zone_start_pfn; | |
4739 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
4740 | #endif /* CONFIG_SPARSEMEM */ | |
4741 | } | |
4742 | ||
4743 | /** | |
4744 | * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages | |
4745 | * @page: The page within the block of interest | |
4746 | * @start_bitidx: The first bit of interest to retrieve | |
4747 | * @end_bitidx: The last bit of interest | |
4748 | * returns pageblock_bits flags | |
4749 | */ | |
4750 | unsigned long get_pageblock_flags_group(struct page *page, | |
4751 | int start_bitidx, int end_bitidx) | |
4752 | { | |
4753 | struct zone *zone; | |
4754 | unsigned long *bitmap; | |
4755 | unsigned long pfn, bitidx; | |
4756 | unsigned long flags = 0; | |
4757 | unsigned long value = 1; | |
4758 | ||
4759 | zone = page_zone(page); | |
4760 | pfn = page_to_pfn(page); | |
4761 | bitmap = get_pageblock_bitmap(zone, pfn); | |
4762 | bitidx = pfn_to_bitidx(zone, pfn); | |
4763 | ||
4764 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
4765 | if (test_bit(bitidx + start_bitidx, bitmap)) | |
4766 | flags |= value; | |
4767 | ||
4768 | return flags; | |
4769 | } | |
4770 | ||
4771 | /** | |
4772 | * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages | |
4773 | * @page: The page within the block of interest | |
4774 | * @start_bitidx: The first bit of interest | |
4775 | * @end_bitidx: The last bit of interest | |
4776 | * @flags: The flags to set | |
4777 | */ | |
4778 | void set_pageblock_flags_group(struct page *page, unsigned long flags, | |
4779 | int start_bitidx, int end_bitidx) | |
4780 | { | |
4781 | struct zone *zone; | |
4782 | unsigned long *bitmap; | |
4783 | unsigned long pfn, bitidx; | |
4784 | unsigned long value = 1; | |
4785 | ||
4786 | zone = page_zone(page); | |
4787 | pfn = page_to_pfn(page); | |
4788 | bitmap = get_pageblock_bitmap(zone, pfn); | |
4789 | bitidx = pfn_to_bitidx(zone, pfn); | |
4790 | VM_BUG_ON(pfn < zone->zone_start_pfn); | |
4791 | VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); | |
4792 | ||
4793 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
4794 | if (flags & value) | |
4795 | __set_bit(bitidx + start_bitidx, bitmap); | |
4796 | else | |
4797 | __clear_bit(bitidx + start_bitidx, bitmap); | |
4798 | } | |
4799 | ||
4800 | /* | |
4801 | * This is designed as sub function...plz see page_isolation.c also. | |
4802 | * set/clear page block's type to be ISOLATE. | |
4803 | * page allocater never alloc memory from ISOLATE block. | |
4804 | */ | |
4805 | ||
4806 | int set_migratetype_isolate(struct page *page) | |
4807 | { | |
4808 | struct zone *zone; | |
4809 | unsigned long flags; | |
4810 | int ret = -EBUSY; | |
4811 | ||
4812 | zone = page_zone(page); | |
4813 | spin_lock_irqsave(&zone->lock, flags); | |
4814 | /* | |
4815 | * In future, more migrate types will be able to be isolation target. | |
4816 | */ | |
4817 | if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE) | |
4818 | goto out; | |
4819 | set_pageblock_migratetype(page, MIGRATE_ISOLATE); | |
4820 | move_freepages_block(zone, page, MIGRATE_ISOLATE); | |
4821 | ret = 0; | |
4822 | out: | |
4823 | spin_unlock_irqrestore(&zone->lock, flags); | |
4824 | if (!ret) | |
4825 | drain_all_pages(); | |
4826 | return ret; | |
4827 | } | |
4828 | ||
4829 | void unset_migratetype_isolate(struct page *page) | |
4830 | { | |
4831 | struct zone *zone; | |
4832 | unsigned long flags; | |
4833 | zone = page_zone(page); | |
4834 | spin_lock_irqsave(&zone->lock, flags); | |
4835 | if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) | |
4836 | goto out; | |
4837 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
4838 | move_freepages_block(zone, page, MIGRATE_MOVABLE); | |
4839 | out: | |
4840 | spin_unlock_irqrestore(&zone->lock, flags); | |
4841 | } | |
4842 | ||
4843 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
4844 | /* | |
4845 | * All pages in the range must be isolated before calling this. | |
4846 | */ | |
4847 | void | |
4848 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) | |
4849 | { | |
4850 | struct page *page; | |
4851 | struct zone *zone; | |
4852 | int order, i; | |
4853 | unsigned long pfn; | |
4854 | unsigned long flags; | |
4855 | /* find the first valid pfn */ | |
4856 | for (pfn = start_pfn; pfn < end_pfn; pfn++) | |
4857 | if (pfn_valid(pfn)) | |
4858 | break; | |
4859 | if (pfn == end_pfn) | |
4860 | return; | |
4861 | zone = page_zone(pfn_to_page(pfn)); | |
4862 | spin_lock_irqsave(&zone->lock, flags); | |
4863 | pfn = start_pfn; | |
4864 | while (pfn < end_pfn) { | |
4865 | if (!pfn_valid(pfn)) { | |
4866 | pfn++; | |
4867 | continue; | |
4868 | } | |
4869 | page = pfn_to_page(pfn); | |
4870 | BUG_ON(page_count(page)); | |
4871 | BUG_ON(!PageBuddy(page)); | |
4872 | order = page_order(page); | |
4873 | #ifdef CONFIG_DEBUG_VM | |
4874 | printk(KERN_INFO "remove from free list %lx %d %lx\n", | |
4875 | pfn, 1 << order, end_pfn); | |
4876 | #endif | |
4877 | list_del(&page->lru); | |
4878 | rmv_page_order(page); | |
4879 | zone->free_area[order].nr_free--; | |
4880 | __mod_zone_page_state(zone, NR_FREE_PAGES, | |
4881 | - (1UL << order)); | |
4882 | for (i = 0; i < (1 << order); i++) | |
4883 | SetPageReserved((page+i)); | |
4884 | pfn += (1 << order); | |
4885 | } | |
4886 | spin_unlock_irqrestore(&zone->lock, flags); | |
4887 | } | |
4888 | #endif |