Root/
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/kmemcheck.h> |
27 | #include <linux/module.h> |
28 | #include <linux/suspend.h> |
29 | #include <linux/pagevec.h> |
30 | #include <linux/blkdev.h> |
31 | #include <linux/slab.h> |
32 | #include <linux/oom.h> |
33 | #include <linux/notifier.h> |
34 | #include <linux/topology.h> |
35 | #include <linux/sysctl.h> |
36 | #include <linux/cpu.h> |
37 | #include <linux/cpuset.h> |
38 | #include <linux/memory_hotplug.h> |
39 | #include <linux/nodemask.h> |
40 | #include <linux/vmalloc.h> |
41 | #include <linux/mempolicy.h> |
42 | #include <linux/stop_machine.h> |
43 | #include <linux/sort.h> |
44 | #include <linux/pfn.h> |
45 | #include <linux/backing-dev.h> |
46 | #include <linux/fault-inject.h> |
47 | #include <linux/page-isolation.h> |
48 | #include <linux/page_cgroup.h> |
49 | #include <linux/debugobjects.h> |
50 | #include <linux/kmemleak.h> |
51 | |
52 | #include <asm/tlbflush.h> |
53 | #include <asm/div64.h> |
54 | #include "internal.h" |
55 | |
56 | /* |
57 | * Array of node states. |
58 | */ |
59 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { |
60 | [N_POSSIBLE] = NODE_MASK_ALL, |
61 | [N_ONLINE] = { { [0] = 1UL } }, |
62 | #ifndef CONFIG_NUMA |
63 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, |
64 | #ifdef CONFIG_HIGHMEM |
65 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, |
66 | #endif |
67 | [N_CPU] = { { [0] = 1UL } }, |
68 | #endif /* NUMA */ |
69 | }; |
70 | EXPORT_SYMBOL(node_states); |
71 | |
72 | unsigned long totalram_pages __read_mostly; |
73 | unsigned long totalreserve_pages __read_mostly; |
74 | unsigned long highest_memmap_pfn __read_mostly; |
75 | int percpu_pagelist_fraction; |
76 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; |
77 | |
78 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
79 | int pageblock_order __read_mostly; |
80 | #endif |
81 | |
82 | static void __free_pages_ok(struct page *page, unsigned int order); |
83 | |
84 | /* |
85 | * results with 256, 32 in the lowmem_reserve sysctl: |
86 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) |
87 | * 1G machine -> (16M dma, 784M normal, 224M high) |
88 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA |
89 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL |
90 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA |
91 | * |
92 | * TBD: should special case ZONE_DMA32 machines here - in those we normally |
93 | * don't need any ZONE_NORMAL reservation |
94 | */ |
95 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { |
96 | #ifdef CONFIG_ZONE_DMA |
97 | 256, |
98 | #endif |
99 | #ifdef CONFIG_ZONE_DMA32 |
100 | 256, |
101 | #endif |
102 | #ifdef CONFIG_HIGHMEM |
103 | 32, |
104 | #endif |
105 | 32, |
106 | }; |
107 | |
108 | EXPORT_SYMBOL(totalram_pages); |
109 | |
110 | static char * const zone_names[MAX_NR_ZONES] = { |
111 | #ifdef CONFIG_ZONE_DMA |
112 | "DMA", |
113 | #endif |
114 | #ifdef CONFIG_ZONE_DMA32 |
115 | "DMA32", |
116 | #endif |
117 | "Normal", |
118 | #ifdef CONFIG_HIGHMEM |
119 | "HighMem", |
120 | #endif |
121 | "Movable", |
122 | }; |
123 | |
124 | int min_free_kbytes = 1024; |
125 | |
126 | unsigned long __meminitdata nr_kernel_pages; |
127 | unsigned long __meminitdata nr_all_pages; |
128 | static unsigned long __meminitdata dma_reserve; |
129 | |
130 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP |
131 | /* |
132 | * MAX_ACTIVE_REGIONS determines the maximum number of distinct |
133 | * ranges of memory (RAM) that may be registered with add_active_range(). |
134 | * Ranges passed to add_active_range() will be merged if possible |
135 | * so the number of times add_active_range() can be called is |
136 | * related to the number of nodes and the number of holes |
137 | */ |
138 | #ifdef CONFIG_MAX_ACTIVE_REGIONS |
139 | /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ |
140 | #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS |
141 | #else |
142 | #if MAX_NUMNODES >= 32 |
143 | /* If there can be many nodes, allow up to 50 holes per node */ |
144 | #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) |
145 | #else |
146 | /* By default, allow up to 256 distinct regions */ |
147 | #define MAX_ACTIVE_REGIONS 256 |
148 | #endif |
149 | #endif |
150 | |
151 | static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; |
152 | static int __meminitdata nr_nodemap_entries; |
153 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; |
154 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; |
155 | static unsigned long __initdata required_kernelcore; |
156 | static unsigned long __initdata required_movablecore; |
157 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; |
158 | |
159 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ |
160 | int movable_zone; |
161 | EXPORT_SYMBOL(movable_zone); |
162 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ |
163 | |
164 | #if MAX_NUMNODES > 1 |
165 | int nr_node_ids __read_mostly = MAX_NUMNODES; |
166 | int nr_online_nodes __read_mostly = 1; |
167 | EXPORT_SYMBOL(nr_node_ids); |
168 | EXPORT_SYMBOL(nr_online_nodes); |
169 | #endif |
170 | |
171 | int page_group_by_mobility_disabled __read_mostly; |
172 | |
173 | static void set_pageblock_migratetype(struct page *page, int migratetype) |
174 | { |
175 | |
176 | if (unlikely(page_group_by_mobility_disabled)) |
177 | migratetype = MIGRATE_UNMOVABLE; |
178 | |
179 | set_pageblock_flags_group(page, (unsigned long)migratetype, |
180 | PB_migrate, PB_migrate_end); |
181 | } |
182 | |
183 | bool oom_killer_disabled __read_mostly; |
184 | |
185 | #ifdef CONFIG_DEBUG_VM |
186 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) |
187 | { |
188 | int ret = 0; |
189 | unsigned seq; |
190 | unsigned long pfn = page_to_pfn(page); |
191 | |
192 | do { |
193 | seq = zone_span_seqbegin(zone); |
194 | if (pfn >= zone->zone_start_pfn + zone->spanned_pages) |
195 | ret = 1; |
196 | else if (pfn < zone->zone_start_pfn) |
197 | ret = 1; |
198 | } while (zone_span_seqretry(zone, seq)); |
199 | |
200 | return ret; |
201 | } |
202 | |
203 | static int page_is_consistent(struct zone *zone, struct page *page) |
204 | { |
205 | if (!pfn_valid_within(page_to_pfn(page))) |
206 | return 0; |
207 | if (zone != page_zone(page)) |
208 | return 0; |
209 | |
210 | return 1; |
211 | } |
212 | /* |
213 | * Temporary debugging check for pages not lying within a given zone. |
214 | */ |
215 | static int bad_range(struct zone *zone, struct page *page) |
216 | { |
217 | if (page_outside_zone_boundaries(zone, page)) |
218 | return 1; |
219 | if (!page_is_consistent(zone, page)) |
220 | return 1; |
221 | |
222 | return 0; |
223 | } |
224 | #else |
225 | static inline int bad_range(struct zone *zone, struct page *page) |
226 | { |
227 | return 0; |
228 | } |
229 | #endif |
230 | |
231 | static void bad_page(struct page *page) |
232 | { |
233 | static unsigned long resume; |
234 | static unsigned long nr_shown; |
235 | static unsigned long nr_unshown; |
236 | |
237 | /* |
238 | * Allow a burst of 60 reports, then keep quiet for that minute; |
239 | * or allow a steady drip of one report per second. |
240 | */ |
241 | if (nr_shown == 60) { |
242 | if (time_before(jiffies, resume)) { |
243 | nr_unshown++; |
244 | goto out; |
245 | } |
246 | if (nr_unshown) { |
247 | printk(KERN_ALERT |
248 | "BUG: Bad page state: %lu messages suppressed\n", |
249 | nr_unshown); |
250 | nr_unshown = 0; |
251 | } |
252 | nr_shown = 0; |
253 | } |
254 | if (nr_shown++ == 0) |
255 | resume = jiffies + 60 * HZ; |
256 | |
257 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", |
258 | current->comm, page_to_pfn(page)); |
259 | printk(KERN_ALERT |
260 | "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", |
261 | page, (void *)page->flags, page_count(page), |
262 | page_mapcount(page), page->mapping, page->index); |
263 | |
264 | dump_stack(); |
265 | out: |
266 | /* Leave bad fields for debug, except PageBuddy could make trouble */ |
267 | __ClearPageBuddy(page); |
268 | add_taint(TAINT_BAD_PAGE); |
269 | } |
270 | |
271 | /* |
272 | * Higher-order pages are called "compound pages". They are structured thusly: |
273 | * |
274 | * The first PAGE_SIZE page is called the "head page". |
275 | * |
276 | * The remaining PAGE_SIZE pages are called "tail pages". |
277 | * |
278 | * All pages have PG_compound set. All pages have their ->private pointing at |
279 | * the head page (even the head page has this). |
280 | * |
281 | * The first tail page's ->lru.next holds the address of the compound page's |
282 | * put_page() function. Its ->lru.prev holds the order of allocation. |
283 | * This usage means that zero-order pages may not be compound. |
284 | */ |
285 | |
286 | static void free_compound_page(struct page *page) |
287 | { |
288 | __free_pages_ok(page, compound_order(page)); |
289 | } |
290 | |
291 | void prep_compound_page(struct page *page, unsigned long order) |
292 | { |
293 | int i; |
294 | int nr_pages = 1 << order; |
295 | |
296 | set_compound_page_dtor(page, free_compound_page); |
297 | set_compound_order(page, order); |
298 | __SetPageHead(page); |
299 | for (i = 1; i < nr_pages; i++) { |
300 | struct page *p = page + i; |
301 | |
302 | __SetPageTail(p); |
303 | p->first_page = page; |
304 | } |
305 | } |
306 | |
307 | static int destroy_compound_page(struct page *page, unsigned long order) |
308 | { |
309 | int i; |
310 | int nr_pages = 1 << order; |
311 | int bad = 0; |
312 | |
313 | if (unlikely(compound_order(page) != order) || |
314 | unlikely(!PageHead(page))) { |
315 | bad_page(page); |
316 | bad++; |
317 | } |
318 | |
319 | __ClearPageHead(page); |
320 | |
321 | for (i = 1; i < nr_pages; i++) { |
322 | struct page *p = page + i; |
323 | |
324 | if (unlikely(!PageTail(p) || (p->first_page != page))) { |
325 | bad_page(page); |
326 | bad++; |
327 | } |
328 | __ClearPageTail(p); |
329 | } |
330 | |
331 | return bad; |
332 | } |
333 | |
334 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) |
335 | { |
336 | int i; |
337 | |
338 | /* |
339 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO |
340 | * and __GFP_HIGHMEM from hard or soft interrupt context. |
341 | */ |
342 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); |
343 | for (i = 0; i < (1 << order); i++) |
344 | clear_highpage(page + i); |
345 | } |
346 | |
347 | static inline void set_page_order(struct page *page, int order) |
348 | { |
349 | set_page_private(page, order); |
350 | __SetPageBuddy(page); |
351 | } |
352 | |
353 | static inline void rmv_page_order(struct page *page) |
354 | { |
355 | __ClearPageBuddy(page); |
356 | set_page_private(page, 0); |
357 | } |
358 | |
359 | /* |
360 | * Locate the struct page for both the matching buddy in our |
361 | * pair (buddy1) and the combined O(n+1) page they form (page). |
362 | * |
363 | * 1) Any buddy B1 will have an order O twin B2 which satisfies |
364 | * the following equation: |
365 | * B2 = B1 ^ (1 << O) |
366 | * For example, if the starting buddy (buddy2) is #8 its order |
367 | * 1 buddy is #10: |
368 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 |
369 | * |
370 | * 2) Any buddy B will have an order O+1 parent P which |
371 | * satisfies the following equation: |
372 | * P = B & ~(1 << O) |
373 | * |
374 | * Assumption: *_mem_map is contiguous at least up to MAX_ORDER |
375 | */ |
376 | static inline struct page * |
377 | __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) |
378 | { |
379 | unsigned long buddy_idx = page_idx ^ (1 << order); |
380 | |
381 | return page + (buddy_idx - page_idx); |
382 | } |
383 | |
384 | static inline unsigned long |
385 | __find_combined_index(unsigned long page_idx, unsigned int order) |
386 | { |
387 | return (page_idx & ~(1 << order)); |
388 | } |
389 | |
390 | /* |
391 | * This function checks whether a page is free && is the buddy |
392 | * we can do coalesce a page and its buddy if |
393 | * (a) the buddy is not in a hole && |
394 | * (b) the buddy is in the buddy system && |
395 | * (c) a page and its buddy have the same order && |
396 | * (d) a page and its buddy are in the same zone. |
397 | * |
398 | * For recording whether a page is in the buddy system, we use PG_buddy. |
399 | * Setting, clearing, and testing PG_buddy is serialized by zone->lock. |
400 | * |
401 | * For recording page's order, we use page_private(page). |
402 | */ |
403 | static inline int page_is_buddy(struct page *page, struct page *buddy, |
404 | int order) |
405 | { |
406 | if (!pfn_valid_within(page_to_pfn(buddy))) |
407 | return 0; |
408 | |
409 | if (page_zone_id(page) != page_zone_id(buddy)) |
410 | return 0; |
411 | |
412 | if (PageBuddy(buddy) && page_order(buddy) == order) { |
413 | VM_BUG_ON(page_count(buddy) != 0); |
414 | return 1; |
415 | } |
416 | return 0; |
417 | } |
418 | |
419 | /* |
420 | * Freeing function for a buddy system allocator. |
421 | * |
422 | * The concept of a buddy system is to maintain direct-mapped table |
423 | * (containing bit values) for memory blocks of various "orders". |
424 | * The bottom level table contains the map for the smallest allocatable |
425 | * units of memory (here, pages), and each level above it describes |
426 | * pairs of units from the levels below, hence, "buddies". |
427 | * At a high level, all that happens here is marking the table entry |
428 | * at the bottom level available, and propagating the changes upward |
429 | * as necessary, plus some accounting needed to play nicely with other |
430 | * parts of the VM system. |
431 | * At each level, we keep a list of pages, which are heads of continuous |
432 | * free pages of length of (1 << order) and marked with PG_buddy. Page's |
433 | * order is recorded in page_private(page) field. |
434 | * So when we are allocating or freeing one, we can derive the state of the |
435 | * other. That is, if we allocate a small block, and both were |
436 | * free, the remainder of the region must be split into blocks. |
437 | * If a block is freed, and its buddy is also free, then this |
438 | * triggers coalescing into a block of larger size. |
439 | * |
440 | * -- wli |
441 | */ |
442 | |
443 | static inline void __free_one_page(struct page *page, |
444 | struct zone *zone, unsigned int order, |
445 | int migratetype) |
446 | { |
447 | unsigned long page_idx; |
448 | |
449 | if (unlikely(PageCompound(page))) |
450 | if (unlikely(destroy_compound_page(page, order))) |
451 | return; |
452 | |
453 | VM_BUG_ON(migratetype == -1); |
454 | |
455 | page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); |
456 | |
457 | VM_BUG_ON(page_idx & ((1 << order) - 1)); |
458 | VM_BUG_ON(bad_range(zone, page)); |
459 | |
460 | while (order < MAX_ORDER-1) { |
461 | unsigned long combined_idx; |
462 | struct page *buddy; |
463 | |
464 | buddy = __page_find_buddy(page, page_idx, order); |
465 | if (!page_is_buddy(page, buddy, order)) |
466 | break; |
467 | |
468 | /* Our buddy is free, merge with it and move up one order. */ |
469 | list_del(&buddy->lru); |
470 | zone->free_area[order].nr_free--; |
471 | rmv_page_order(buddy); |
472 | combined_idx = __find_combined_index(page_idx, order); |
473 | page = page + (combined_idx - page_idx); |
474 | page_idx = combined_idx; |
475 | order++; |
476 | } |
477 | set_page_order(page, order); |
478 | list_add(&page->lru, |
479 | &zone->free_area[order].free_list[migratetype]); |
480 | zone->free_area[order].nr_free++; |
481 | } |
482 | |
483 | #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT |
484 | /* |
485 | * free_page_mlock() -- clean up attempts to free and mlocked() page. |
486 | * Page should not be on lru, so no need to fix that up. |
487 | * free_pages_check() will verify... |
488 | */ |
489 | static inline void free_page_mlock(struct page *page) |
490 | { |
491 | __dec_zone_page_state(page, NR_MLOCK); |
492 | __count_vm_event(UNEVICTABLE_MLOCKFREED); |
493 | } |
494 | #else |
495 | static void free_page_mlock(struct page *page) { } |
496 | #endif |
497 | |
498 | static inline int free_pages_check(struct page *page) |
499 | { |
500 | if (unlikely(page_mapcount(page) | |
501 | (page->mapping != NULL) | |
502 | (atomic_read(&page->_count) != 0) | |
503 | (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { |
504 | bad_page(page); |
505 | return 1; |
506 | } |
507 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
508 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
509 | return 0; |
510 | } |
511 | |
512 | /* |
513 | * Frees a list of pages. |
514 | * Assumes all pages on list are in same zone, and of same order. |
515 | * count is the number of pages to free. |
516 | * |
517 | * If the zone was previously in an "all pages pinned" state then look to |
518 | * see if this freeing clears that state. |
519 | * |
520 | * And clear the zone's pages_scanned counter, to hold off the "all pages are |
521 | * pinned" detection logic. |
522 | */ |
523 | static void free_pages_bulk(struct zone *zone, int count, |
524 | struct list_head *list, int order) |
525 | { |
526 | spin_lock(&zone->lock); |
527 | zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); |
528 | zone->pages_scanned = 0; |
529 | |
530 | __mod_zone_page_state(zone, NR_FREE_PAGES, count << order); |
531 | while (count--) { |
532 | struct page *page; |
533 | |
534 | VM_BUG_ON(list_empty(list)); |
535 | page = list_entry(list->prev, struct page, lru); |
536 | /* have to delete it as __free_one_page list manipulates */ |
537 | list_del(&page->lru); |
538 | __free_one_page(page, zone, order, page_private(page)); |
539 | } |
540 | spin_unlock(&zone->lock); |
541 | } |
542 | |
543 | static void free_one_page(struct zone *zone, struct page *page, int order, |
544 | int migratetype) |
545 | { |
546 | spin_lock(&zone->lock); |
547 | zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); |
548 | zone->pages_scanned = 0; |
549 | |
550 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); |
551 | __free_one_page(page, zone, order, migratetype); |
552 | spin_unlock(&zone->lock); |
553 | } |
554 | |
555 | static void __free_pages_ok(struct page *page, unsigned int order) |
556 | { |
557 | unsigned long flags; |
558 | int i; |
559 | int bad = 0; |
560 | int wasMlocked = TestClearPageMlocked(page); |
561 | |
562 | kmemcheck_free_shadow(page, order); |
563 | |
564 | for (i = 0 ; i < (1 << order) ; ++i) |
565 | bad += free_pages_check(page + i); |
566 | if (bad) |
567 | return; |
568 | |
569 | if (!PageHighMem(page)) { |
570 | debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); |
571 | debug_check_no_obj_freed(page_address(page), |
572 | PAGE_SIZE << order); |
573 | } |
574 | arch_free_page(page, order); |
575 | kernel_map_pages(page, 1 << order, 0); |
576 | |
577 | local_irq_save(flags); |
578 | if (unlikely(wasMlocked)) |
579 | free_page_mlock(page); |
580 | __count_vm_events(PGFREE, 1 << order); |
581 | free_one_page(page_zone(page), page, order, |
582 | get_pageblock_migratetype(page)); |
583 | local_irq_restore(flags); |
584 | } |
585 | |
586 | /* |
587 | * permit the bootmem allocator to evade page validation on high-order frees |
588 | */ |
589 | void __meminit __free_pages_bootmem(struct page *page, unsigned int order) |
590 | { |
591 | if (order == 0) { |
592 | __ClearPageReserved(page); |
593 | set_page_count(page, 0); |
594 | set_page_refcounted(page); |
595 | __free_page(page); |
596 | } else { |
597 | int loop; |
598 | |
599 | prefetchw(page); |
600 | for (loop = 0; loop < BITS_PER_LONG; loop++) { |
601 | struct page *p = &page[loop]; |
602 | |
603 | if (loop + 1 < BITS_PER_LONG) |
604 | prefetchw(p + 1); |
605 | __ClearPageReserved(p); |
606 | set_page_count(p, 0); |
607 | } |
608 | |
609 | set_page_refcounted(page); |
610 | __free_pages(page, order); |
611 | } |
612 | } |
613 | |
614 | |
615 | /* |
616 | * The order of subdivision here is critical for the IO subsystem. |
617 | * Please do not alter this order without good reasons and regression |
618 | * testing. Specifically, as large blocks of memory are subdivided, |
619 | * the order in which smaller blocks are delivered depends on the order |
620 | * they're subdivided in this function. This is the primary factor |
621 | * influencing the order in which pages are delivered to the IO |
622 | * subsystem according to empirical testing, and this is also justified |
623 | * by considering the behavior of a buddy system containing a single |
624 | * large block of memory acted on by a series of small allocations. |
625 | * This behavior is a critical factor in sglist merging's success. |
626 | * |
627 | * -- wli |
628 | */ |
629 | static inline void expand(struct zone *zone, struct page *page, |
630 | int low, int high, struct free_area *area, |
631 | int migratetype) |
632 | { |
633 | unsigned long size = 1 << high; |
634 | |
635 | while (high > low) { |
636 | area--; |
637 | high--; |
638 | size >>= 1; |
639 | VM_BUG_ON(bad_range(zone, &page[size])); |
640 | list_add(&page[size].lru, &area->free_list[migratetype]); |
641 | area->nr_free++; |
642 | set_page_order(&page[size], high); |
643 | } |
644 | } |
645 | |
646 | /* |
647 | * This page is about to be returned from the page allocator |
648 | */ |
649 | static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) |
650 | { |
651 | if (unlikely(page_mapcount(page) | |
652 | (page->mapping != NULL) | |
653 | (atomic_read(&page->_count) != 0) | |
654 | (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { |
655 | bad_page(page); |
656 | return 1; |
657 | } |
658 | |
659 | set_page_private(page, 0); |
660 | set_page_refcounted(page); |
661 | |
662 | arch_alloc_page(page, order); |
663 | kernel_map_pages(page, 1 << order, 1); |
664 | |
665 | if (gfp_flags & __GFP_ZERO) |
666 | prep_zero_page(page, order, gfp_flags); |
667 | |
668 | if (order && (gfp_flags & __GFP_COMP)) |
669 | prep_compound_page(page, order); |
670 | |
671 | return 0; |
672 | } |
673 | |
674 | /* |
675 | * Go through the free lists for the given migratetype and remove |
676 | * the smallest available page from the freelists |
677 | */ |
678 | static inline |
679 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, |
680 | int migratetype) |
681 | { |
682 | unsigned int current_order; |
683 | struct free_area * area; |
684 | struct page *page; |
685 | |
686 | /* Find a page of the appropriate size in the preferred list */ |
687 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { |
688 | area = &(zone->free_area[current_order]); |
689 | if (list_empty(&area->free_list[migratetype])) |
690 | continue; |
691 | |
692 | page = list_entry(area->free_list[migratetype].next, |
693 | struct page, lru); |
694 | list_del(&page->lru); |
695 | rmv_page_order(page); |
696 | area->nr_free--; |
697 | expand(zone, page, order, current_order, area, migratetype); |
698 | return page; |
699 | } |
700 | |
701 | return NULL; |
702 | } |
703 | |
704 | |
705 | /* |
706 | * This array describes the order lists are fallen back to when |
707 | * the free lists for the desirable migrate type are depleted |
708 | */ |
709 | static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { |
710 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, |
711 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, |
712 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, |
713 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ |
714 | }; |
715 | |
716 | /* |
717 | * Move the free pages in a range to the free lists of the requested type. |
718 | * Note that start_page and end_pages are not aligned on a pageblock |
719 | * boundary. If alignment is required, use move_freepages_block() |
720 | */ |
721 | static int move_freepages(struct zone *zone, |
722 | struct page *start_page, struct page *end_page, |
723 | int migratetype) |
724 | { |
725 | struct page *page; |
726 | unsigned long order; |
727 | int pages_moved = 0; |
728 | |
729 | #ifndef CONFIG_HOLES_IN_ZONE |
730 | /* |
731 | * page_zone is not safe to call in this context when |
732 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant |
733 | * anyway as we check zone boundaries in move_freepages_block(). |
734 | * Remove at a later date when no bug reports exist related to |
735 | * grouping pages by mobility |
736 | */ |
737 | BUG_ON(page_zone(start_page) != page_zone(end_page)); |
738 | #endif |
739 | |
740 | for (page = start_page; page <= end_page;) { |
741 | /* Make sure we are not inadvertently changing nodes */ |
742 | VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); |
743 | |
744 | if (!pfn_valid_within(page_to_pfn(page))) { |
745 | page++; |
746 | continue; |
747 | } |
748 | |
749 | if (!PageBuddy(page)) { |
750 | page++; |
751 | continue; |
752 | } |
753 | |
754 | order = page_order(page); |
755 | list_del(&page->lru); |
756 | list_add(&page->lru, |
757 | &zone->free_area[order].free_list[migratetype]); |
758 | page += 1 << order; |
759 | pages_moved += 1 << order; |
760 | } |
761 | |
762 | return pages_moved; |
763 | } |
764 | |
765 | static int move_freepages_block(struct zone *zone, struct page *page, |
766 | int migratetype) |
767 | { |
768 | unsigned long start_pfn, end_pfn; |
769 | struct page *start_page, *end_page; |
770 | |
771 | start_pfn = page_to_pfn(page); |
772 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); |
773 | start_page = pfn_to_page(start_pfn); |
774 | end_page = start_page + pageblock_nr_pages - 1; |
775 | end_pfn = start_pfn + pageblock_nr_pages - 1; |
776 | |
777 | /* Do not cross zone boundaries */ |
778 | if (start_pfn < zone->zone_start_pfn) |
779 | start_page = page; |
780 | if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) |
781 | return 0; |
782 | |
783 | return move_freepages(zone, start_page, end_page, migratetype); |
784 | } |
785 | |
786 | /* Remove an element from the buddy allocator from the fallback list */ |
787 | static inline struct page * |
788 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) |
789 | { |
790 | struct free_area * area; |
791 | int current_order; |
792 | struct page *page; |
793 | int migratetype, i; |
794 | |
795 | /* Find the largest possible block of pages in the other list */ |
796 | for (current_order = MAX_ORDER-1; current_order >= order; |
797 | --current_order) { |
798 | for (i = 0; i < MIGRATE_TYPES - 1; i++) { |
799 | migratetype = fallbacks[start_migratetype][i]; |
800 | |
801 | /* MIGRATE_RESERVE handled later if necessary */ |
802 | if (migratetype == MIGRATE_RESERVE) |
803 | continue; |
804 | |
805 | area = &(zone->free_area[current_order]); |
806 | if (list_empty(&area->free_list[migratetype])) |
807 | continue; |
808 | |
809 | page = list_entry(area->free_list[migratetype].next, |
810 | struct page, lru); |
811 | area->nr_free--; |
812 | |
813 | /* |
814 | * If breaking a large block of pages, move all free |
815 | * pages to the preferred allocation list. If falling |
816 | * back for a reclaimable kernel allocation, be more |
817 | * agressive about taking ownership of free pages |
818 | */ |
819 | if (unlikely(current_order >= (pageblock_order >> 1)) || |
820 | start_migratetype == MIGRATE_RECLAIMABLE || |
821 | page_group_by_mobility_disabled) { |
822 | unsigned long pages; |
823 | pages = move_freepages_block(zone, page, |
824 | start_migratetype); |
825 | |
826 | /* Claim the whole block if over half of it is free */ |
827 | if (pages >= (1 << (pageblock_order-1)) || |
828 | page_group_by_mobility_disabled) |
829 | set_pageblock_migratetype(page, |
830 | start_migratetype); |
831 | |
832 | migratetype = start_migratetype; |
833 | } |
834 | |
835 | /* Remove the page from the freelists */ |
836 | list_del(&page->lru); |
837 | rmv_page_order(page); |
838 | |
839 | if (current_order == pageblock_order) |
840 | set_pageblock_migratetype(page, |
841 | start_migratetype); |
842 | |
843 | expand(zone, page, order, current_order, area, migratetype); |
844 | return page; |
845 | } |
846 | } |
847 | |
848 | return NULL; |
849 | } |
850 | |
851 | /* |
852 | * Do the hard work of removing an element from the buddy allocator. |
853 | * Call me with the zone->lock already held. |
854 | */ |
855 | static struct page *__rmqueue(struct zone *zone, unsigned int order, |
856 | int migratetype) |
857 | { |
858 | struct page *page; |
859 | |
860 | retry_reserve: |
861 | page = __rmqueue_smallest(zone, order, migratetype); |
862 | |
863 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { |
864 | page = __rmqueue_fallback(zone, order, migratetype); |
865 | |
866 | /* |
867 | * Use MIGRATE_RESERVE rather than fail an allocation. goto |
868 | * is used because __rmqueue_smallest is an inline function |
869 | * and we want just one call site |
870 | */ |
871 | if (!page) { |
872 | migratetype = MIGRATE_RESERVE; |
873 | goto retry_reserve; |
874 | } |
875 | } |
876 | |
877 | return page; |
878 | } |
879 | |
880 | /* |
881 | * Obtain a specified number of elements from the buddy allocator, all under |
882 | * a single hold of the lock, for efficiency. Add them to the supplied list. |
883 | * Returns the number of new pages which were placed at *list. |
884 | */ |
885 | static int rmqueue_bulk(struct zone *zone, unsigned int order, |
886 | unsigned long count, struct list_head *list, |
887 | int migratetype, int cold) |
888 | { |
889 | int i; |
890 | |
891 | spin_lock(&zone->lock); |
892 | for (i = 0; i < count; ++i) { |
893 | struct page *page = __rmqueue(zone, order, migratetype); |
894 | if (unlikely(page == NULL)) |
895 | break; |
896 | |
897 | /* |
898 | * Split buddy pages returned by expand() are received here |
899 | * in physical page order. The page is added to the callers and |
900 | * list and the list head then moves forward. From the callers |
901 | * perspective, the linked list is ordered by page number in |
902 | * some conditions. This is useful for IO devices that can |
903 | * merge IO requests if the physical pages are ordered |
904 | * properly. |
905 | */ |
906 | if (likely(cold == 0)) |
907 | list_add(&page->lru, list); |
908 | else |
909 | list_add_tail(&page->lru, list); |
910 | set_page_private(page, migratetype); |
911 | list = &page->lru; |
912 | } |
913 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); |
914 | spin_unlock(&zone->lock); |
915 | return i; |
916 | } |
917 | |
918 | #ifdef CONFIG_NUMA |
919 | /* |
920 | * Called from the vmstat counter updater to drain pagesets of this |
921 | * currently executing processor on remote nodes after they have |
922 | * expired. |
923 | * |
924 | * Note that this function must be called with the thread pinned to |
925 | * a single processor. |
926 | */ |
927 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) |
928 | { |
929 | unsigned long flags; |
930 | int to_drain; |
931 | |
932 | local_irq_save(flags); |
933 | if (pcp->count >= pcp->batch) |
934 | to_drain = pcp->batch; |
935 | else |
936 | to_drain = pcp->count; |
937 | free_pages_bulk(zone, to_drain, &pcp->list, 0); |
938 | pcp->count -= to_drain; |
939 | local_irq_restore(flags); |
940 | } |
941 | #endif |
942 | |
943 | /* |
944 | * Drain pages of the indicated processor. |
945 | * |
946 | * The processor must either be the current processor and the |
947 | * thread pinned to the current processor or a processor that |
948 | * is not online. |
949 | */ |
950 | static void drain_pages(unsigned int cpu) |
951 | { |
952 | unsigned long flags; |
953 | struct zone *zone; |
954 | |
955 | for_each_populated_zone(zone) { |
956 | struct per_cpu_pageset *pset; |
957 | struct per_cpu_pages *pcp; |
958 | |
959 | pset = zone_pcp(zone, cpu); |
960 | |
961 | pcp = &pset->pcp; |
962 | local_irq_save(flags); |
963 | free_pages_bulk(zone, pcp->count, &pcp->list, 0); |
964 | pcp->count = 0; |
965 | local_irq_restore(flags); |
966 | } |
967 | } |
968 | |
969 | /* |
970 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. |
971 | */ |
972 | void drain_local_pages(void *arg) |
973 | { |
974 | drain_pages(smp_processor_id()); |
975 | } |
976 | |
977 | /* |
978 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator |
979 | */ |
980 | void drain_all_pages(void) |
981 | { |
982 | on_each_cpu(drain_local_pages, NULL, 1); |
983 | } |
984 | |
985 | #ifdef CONFIG_HIBERNATION |
986 | |
987 | void mark_free_pages(struct zone *zone) |
988 | { |
989 | unsigned long pfn, max_zone_pfn; |
990 | unsigned long flags; |
991 | int order, t; |
992 | struct list_head *curr; |
993 | |
994 | if (!zone->spanned_pages) |
995 | return; |
996 | |
997 | spin_lock_irqsave(&zone->lock, flags); |
998 | |
999 | max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; |
1000 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) |
1001 | if (pfn_valid(pfn)) { |
1002 | struct page *page = pfn_to_page(pfn); |
1003 | |
1004 | if (!swsusp_page_is_forbidden(page)) |
1005 | swsusp_unset_page_free(page); |
1006 | } |
1007 | |
1008 | for_each_migratetype_order(order, t) { |
1009 | list_for_each(curr, &zone->free_area[order].free_list[t]) { |
1010 | unsigned long i; |
1011 | |
1012 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); |
1013 | for (i = 0; i < (1UL << order); i++) |
1014 | swsusp_set_page_free(pfn_to_page(pfn + i)); |
1015 | } |
1016 | } |
1017 | spin_unlock_irqrestore(&zone->lock, flags); |
1018 | } |
1019 | #endif /* CONFIG_PM */ |
1020 | |
1021 | /* |
1022 | * Free a 0-order page |
1023 | */ |
1024 | static void free_hot_cold_page(struct page *page, int cold) |
1025 | { |
1026 | struct zone *zone = page_zone(page); |
1027 | struct per_cpu_pages *pcp; |
1028 | unsigned long flags; |
1029 | int wasMlocked = TestClearPageMlocked(page); |
1030 | |
1031 | kmemcheck_free_shadow(page, 0); |
1032 | |
1033 | if (PageAnon(page)) |
1034 | page->mapping = NULL; |
1035 | if (free_pages_check(page)) |
1036 | return; |
1037 | |
1038 | if (!PageHighMem(page)) { |
1039 | debug_check_no_locks_freed(page_address(page), PAGE_SIZE); |
1040 | debug_check_no_obj_freed(page_address(page), PAGE_SIZE); |
1041 | } |
1042 | arch_free_page(page, 0); |
1043 | kernel_map_pages(page, 1, 0); |
1044 | |
1045 | pcp = &zone_pcp(zone, get_cpu())->pcp; |
1046 | set_page_private(page, get_pageblock_migratetype(page)); |
1047 | local_irq_save(flags); |
1048 | if (unlikely(wasMlocked)) |
1049 | free_page_mlock(page); |
1050 | __count_vm_event(PGFREE); |
1051 | |
1052 | if (cold) |
1053 | list_add_tail(&page->lru, &pcp->list); |
1054 | else |
1055 | list_add(&page->lru, &pcp->list); |
1056 | pcp->count++; |
1057 | if (pcp->count >= pcp->high) { |
1058 | free_pages_bulk(zone, pcp->batch, &pcp->list, 0); |
1059 | pcp->count -= pcp->batch; |
1060 | } |
1061 | local_irq_restore(flags); |
1062 | put_cpu(); |
1063 | } |
1064 | |
1065 | void free_hot_page(struct page *page) |
1066 | { |
1067 | free_hot_cold_page(page, 0); |
1068 | } |
1069 | |
1070 | void free_cold_page(struct page *page) |
1071 | { |
1072 | free_hot_cold_page(page, 1); |
1073 | } |
1074 | |
1075 | /* |
1076 | * split_page takes a non-compound higher-order page, and splits it into |
1077 | * n (1<<order) sub-pages: page[0..n] |
1078 | * Each sub-page must be freed individually. |
1079 | * |
1080 | * Note: this is probably too low level an operation for use in drivers. |
1081 | * Please consult with lkml before using this in your driver. |
1082 | */ |
1083 | void split_page(struct page *page, unsigned int order) |
1084 | { |
1085 | int i; |
1086 | |
1087 | VM_BUG_ON(PageCompound(page)); |
1088 | VM_BUG_ON(!page_count(page)); |
1089 | |
1090 | #ifdef CONFIG_KMEMCHECK |
1091 | /* |
1092 | * Split shadow pages too, because free(page[0]) would |
1093 | * otherwise free the whole shadow. |
1094 | */ |
1095 | if (kmemcheck_page_is_tracked(page)) |
1096 | split_page(virt_to_page(page[0].shadow), order); |
1097 | #endif |
1098 | |
1099 | for (i = 1; i < (1 << order); i++) |
1100 | set_page_refcounted(page + i); |
1101 | } |
1102 | |
1103 | /* |
1104 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But |
1105 | * we cheat by calling it from here, in the order > 0 path. Saves a branch |
1106 | * or two. |
1107 | */ |
1108 | static inline |
1109 | struct page *buffered_rmqueue(struct zone *preferred_zone, |
1110 | struct zone *zone, int order, gfp_t gfp_flags, |
1111 | int migratetype) |
1112 | { |
1113 | unsigned long flags; |
1114 | struct page *page; |
1115 | int cold = !!(gfp_flags & __GFP_COLD); |
1116 | int cpu; |
1117 | |
1118 | again: |
1119 | cpu = get_cpu(); |
1120 | if (likely(order == 0)) { |
1121 | struct per_cpu_pages *pcp; |
1122 | |
1123 | pcp = &zone_pcp(zone, cpu)->pcp; |
1124 | local_irq_save(flags); |
1125 | if (!pcp->count) { |
1126 | pcp->count = rmqueue_bulk(zone, 0, |
1127 | pcp->batch, &pcp->list, |
1128 | migratetype, cold); |
1129 | if (unlikely(!pcp->count)) |
1130 | goto failed; |
1131 | } |
1132 | |
1133 | /* Find a page of the appropriate migrate type */ |
1134 | if (cold) { |
1135 | list_for_each_entry_reverse(page, &pcp->list, lru) |
1136 | if (page_private(page) == migratetype) |
1137 | break; |
1138 | } else { |
1139 | list_for_each_entry(page, &pcp->list, lru) |
1140 | if (page_private(page) == migratetype) |
1141 | break; |
1142 | } |
1143 | |
1144 | /* Allocate more to the pcp list if necessary */ |
1145 | if (unlikely(&page->lru == &pcp->list)) { |
1146 | pcp->count += rmqueue_bulk(zone, 0, |
1147 | pcp->batch, &pcp->list, |
1148 | migratetype, cold); |
1149 | page = list_entry(pcp->list.next, struct page, lru); |
1150 | } |
1151 | |
1152 | list_del(&page->lru); |
1153 | pcp->count--; |
1154 | } else { |
1155 | if (unlikely(gfp_flags & __GFP_NOFAIL)) { |
1156 | /* |
1157 | * __GFP_NOFAIL is not to be used in new code. |
1158 | * |
1159 | * All __GFP_NOFAIL callers should be fixed so that they |
1160 | * properly detect and handle allocation failures. |
1161 | * |
1162 | * We most definitely don't want callers attempting to |
1163 | * allocate greater than order-1 page units with |
1164 | * __GFP_NOFAIL. |
1165 | */ |
1166 | WARN_ON_ONCE(order > 1); |
1167 | } |
1168 | spin_lock_irqsave(&zone->lock, flags); |
1169 | page = __rmqueue(zone, order, migratetype); |
1170 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); |
1171 | spin_unlock(&zone->lock); |
1172 | if (!page) |
1173 | goto failed; |
1174 | } |
1175 | |
1176 | __count_zone_vm_events(PGALLOC, zone, 1 << order); |
1177 | zone_statistics(preferred_zone, zone); |
1178 | local_irq_restore(flags); |
1179 | put_cpu(); |
1180 | |
1181 | VM_BUG_ON(bad_range(zone, page)); |
1182 | if (prep_new_page(page, order, gfp_flags)) |
1183 | goto again; |
1184 | return page; |
1185 | |
1186 | failed: |
1187 | local_irq_restore(flags); |
1188 | put_cpu(); |
1189 | return NULL; |
1190 | } |
1191 | |
1192 | /* The ALLOC_WMARK bits are used as an index to zone->watermark */ |
1193 | #define ALLOC_WMARK_MIN WMARK_MIN |
1194 | #define ALLOC_WMARK_LOW WMARK_LOW |
1195 | #define ALLOC_WMARK_HIGH WMARK_HIGH |
1196 | #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ |
1197 | |
1198 | /* Mask to get the watermark bits */ |
1199 | #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) |
1200 | |
1201 | #define ALLOC_HARDER 0x10 /* try to alloc harder */ |
1202 | #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ |
1203 | #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ |
1204 | |
1205 | #ifdef CONFIG_FAIL_PAGE_ALLOC |
1206 | |
1207 | static struct fail_page_alloc_attr { |
1208 | struct fault_attr attr; |
1209 | |
1210 | u32 ignore_gfp_highmem; |
1211 | u32 ignore_gfp_wait; |
1212 | u32 min_order; |
1213 | |
1214 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS |
1215 | |
1216 | struct dentry *ignore_gfp_highmem_file; |
1217 | struct dentry *ignore_gfp_wait_file; |
1218 | struct dentry *min_order_file; |
1219 | |
1220 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ |
1221 | |
1222 | } fail_page_alloc = { |
1223 | .attr = FAULT_ATTR_INITIALIZER, |
1224 | .ignore_gfp_wait = 1, |
1225 | .ignore_gfp_highmem = 1, |
1226 | .min_order = 1, |
1227 | }; |
1228 | |
1229 | static int __init setup_fail_page_alloc(char *str) |
1230 | { |
1231 | return setup_fault_attr(&fail_page_alloc.attr, str); |
1232 | } |
1233 | __setup("fail_page_alloc=", setup_fail_page_alloc); |
1234 | |
1235 | static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
1236 | { |
1237 | if (order < fail_page_alloc.min_order) |
1238 | return 0; |
1239 | if (gfp_mask & __GFP_NOFAIL) |
1240 | return 0; |
1241 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) |
1242 | return 0; |
1243 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) |
1244 | return 0; |
1245 | |
1246 | return should_fail(&fail_page_alloc.attr, 1 << order); |
1247 | } |
1248 | |
1249 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS |
1250 | |
1251 | static int __init fail_page_alloc_debugfs(void) |
1252 | { |
1253 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; |
1254 | struct dentry *dir; |
1255 | int err; |
1256 | |
1257 | err = init_fault_attr_dentries(&fail_page_alloc.attr, |
1258 | "fail_page_alloc"); |
1259 | if (err) |
1260 | return err; |
1261 | dir = fail_page_alloc.attr.dentries.dir; |
1262 | |
1263 | fail_page_alloc.ignore_gfp_wait_file = |
1264 | debugfs_create_bool("ignore-gfp-wait", mode, dir, |
1265 | &fail_page_alloc.ignore_gfp_wait); |
1266 | |
1267 | fail_page_alloc.ignore_gfp_highmem_file = |
1268 | debugfs_create_bool("ignore-gfp-highmem", mode, dir, |
1269 | &fail_page_alloc.ignore_gfp_highmem); |
1270 | fail_page_alloc.min_order_file = |
1271 | debugfs_create_u32("min-order", mode, dir, |
1272 | &fail_page_alloc.min_order); |
1273 | |
1274 | if (!fail_page_alloc.ignore_gfp_wait_file || |
1275 | !fail_page_alloc.ignore_gfp_highmem_file || |
1276 | !fail_page_alloc.min_order_file) { |
1277 | err = -ENOMEM; |
1278 | debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); |
1279 | debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); |
1280 | debugfs_remove(fail_page_alloc.min_order_file); |
1281 | cleanup_fault_attr_dentries(&fail_page_alloc.attr); |
1282 | } |
1283 | |
1284 | return err; |
1285 | } |
1286 | |
1287 | late_initcall(fail_page_alloc_debugfs); |
1288 | |
1289 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ |
1290 | |
1291 | #else /* CONFIG_FAIL_PAGE_ALLOC */ |
1292 | |
1293 | static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
1294 | { |
1295 | return 0; |
1296 | } |
1297 | |
1298 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ |
1299 | |
1300 | /* |
1301 | * Return 1 if free pages are above 'mark'. This takes into account the order |
1302 | * of the allocation. |
1303 | */ |
1304 | int zone_watermark_ok(struct zone *z, int order, unsigned long mark, |
1305 | int classzone_idx, int alloc_flags) |
1306 | { |
1307 | /* free_pages my go negative - that's OK */ |
1308 | long min = mark; |
1309 | long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; |
1310 | int o; |
1311 | |
1312 | if (alloc_flags & ALLOC_HIGH) |
1313 | min -= min / 2; |
1314 | if (alloc_flags & ALLOC_HARDER) |
1315 | min -= min / 4; |
1316 | |
1317 | if (free_pages <= min + z->lowmem_reserve[classzone_idx]) |
1318 | return 0; |
1319 | for (o = 0; o < order; o++) { |
1320 | /* At the next order, this order's pages become unavailable */ |
1321 | free_pages -= z->free_area[o].nr_free << o; |
1322 | |
1323 | /* Require fewer higher order pages to be free */ |
1324 | min >>= 1; |
1325 | |
1326 | if (free_pages <= min) |
1327 | return 0; |
1328 | } |
1329 | return 1; |
1330 | } |
1331 | |
1332 | #ifdef CONFIG_NUMA |
1333 | /* |
1334 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to |
1335 | * skip over zones that are not allowed by the cpuset, or that have |
1336 | * been recently (in last second) found to be nearly full. See further |
1337 | * comments in mmzone.h. Reduces cache footprint of zonelist scans |
1338 | * that have to skip over a lot of full or unallowed zones. |
1339 | * |
1340 | * If the zonelist cache is present in the passed in zonelist, then |
1341 | * returns a pointer to the allowed node mask (either the current |
1342 | * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) |
1343 | * |
1344 | * If the zonelist cache is not available for this zonelist, does |
1345 | * nothing and returns NULL. |
1346 | * |
1347 | * If the fullzones BITMAP in the zonelist cache is stale (more than |
1348 | * a second since last zap'd) then we zap it out (clear its bits.) |
1349 | * |
1350 | * We hold off even calling zlc_setup, until after we've checked the |
1351 | * first zone in the zonelist, on the theory that most allocations will |
1352 | * be satisfied from that first zone, so best to examine that zone as |
1353 | * quickly as we can. |
1354 | */ |
1355 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) |
1356 | { |
1357 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
1358 | nodemask_t *allowednodes; /* zonelist_cache approximation */ |
1359 | |
1360 | zlc = zonelist->zlcache_ptr; |
1361 | if (!zlc) |
1362 | return NULL; |
1363 | |
1364 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { |
1365 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); |
1366 | zlc->last_full_zap = jiffies; |
1367 | } |
1368 | |
1369 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? |
1370 | &cpuset_current_mems_allowed : |
1371 | &node_states[N_HIGH_MEMORY]; |
1372 | return allowednodes; |
1373 | } |
1374 | |
1375 | /* |
1376 | * Given 'z' scanning a zonelist, run a couple of quick checks to see |
1377 | * if it is worth looking at further for free memory: |
1378 | * 1) Check that the zone isn't thought to be full (doesn't have its |
1379 | * bit set in the zonelist_cache fullzones BITMAP). |
1380 | * 2) Check that the zones node (obtained from the zonelist_cache |
1381 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. |
1382 | * Return true (non-zero) if zone is worth looking at further, or |
1383 | * else return false (zero) if it is not. |
1384 | * |
1385 | * This check -ignores- the distinction between various watermarks, |
1386 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is |
1387 | * found to be full for any variation of these watermarks, it will |
1388 | * be considered full for up to one second by all requests, unless |
1389 | * we are so low on memory on all allowed nodes that we are forced |
1390 | * into the second scan of the zonelist. |
1391 | * |
1392 | * In the second scan we ignore this zonelist cache and exactly |
1393 | * apply the watermarks to all zones, even it is slower to do so. |
1394 | * We are low on memory in the second scan, and should leave no stone |
1395 | * unturned looking for a free page. |
1396 | */ |
1397 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, |
1398 | nodemask_t *allowednodes) |
1399 | { |
1400 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
1401 | int i; /* index of *z in zonelist zones */ |
1402 | int n; /* node that zone *z is on */ |
1403 | |
1404 | zlc = zonelist->zlcache_ptr; |
1405 | if (!zlc) |
1406 | return 1; |
1407 | |
1408 | i = z - zonelist->_zonerefs; |
1409 | n = zlc->z_to_n[i]; |
1410 | |
1411 | /* This zone is worth trying if it is allowed but not full */ |
1412 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); |
1413 | } |
1414 | |
1415 | /* |
1416 | * Given 'z' scanning a zonelist, set the corresponding bit in |
1417 | * zlc->fullzones, so that subsequent attempts to allocate a page |
1418 | * from that zone don't waste time re-examining it. |
1419 | */ |
1420 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) |
1421 | { |
1422 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
1423 | int i; /* index of *z in zonelist zones */ |
1424 | |
1425 | zlc = zonelist->zlcache_ptr; |
1426 | if (!zlc) |
1427 | return; |
1428 | |
1429 | i = z - zonelist->_zonerefs; |
1430 | |
1431 | set_bit(i, zlc->fullzones); |
1432 | } |
1433 | |
1434 | #else /* CONFIG_NUMA */ |
1435 | |
1436 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) |
1437 | { |
1438 | return NULL; |
1439 | } |
1440 | |
1441 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, |
1442 | nodemask_t *allowednodes) |
1443 | { |
1444 | return 1; |
1445 | } |
1446 | |
1447 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) |
1448 | { |
1449 | } |
1450 | #endif /* CONFIG_NUMA */ |
1451 | |
1452 | /* |
1453 | * get_page_from_freelist goes through the zonelist trying to allocate |
1454 | * a page. |
1455 | */ |
1456 | static struct page * |
1457 | get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, |
1458 | struct zonelist *zonelist, int high_zoneidx, int alloc_flags, |
1459 | struct zone *preferred_zone, int migratetype) |
1460 | { |
1461 | struct zoneref *z; |
1462 | struct page *page = NULL; |
1463 | int classzone_idx; |
1464 | struct zone *zone; |
1465 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ |
1466 | int zlc_active = 0; /* set if using zonelist_cache */ |
1467 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ |
1468 | |
1469 | classzone_idx = zone_idx(preferred_zone); |
1470 | zonelist_scan: |
1471 | /* |
1472 | * Scan zonelist, looking for a zone with enough free. |
1473 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. |
1474 | */ |
1475 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
1476 | high_zoneidx, nodemask) { |
1477 | if (NUMA_BUILD && zlc_active && |
1478 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) |
1479 | continue; |
1480 | if ((alloc_flags & ALLOC_CPUSET) && |
1481 | !cpuset_zone_allowed_softwall(zone, gfp_mask)) |
1482 | goto try_next_zone; |
1483 | |
1484 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); |
1485 | if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { |
1486 | unsigned long mark; |
1487 | int ret; |
1488 | |
1489 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; |
1490 | if (zone_watermark_ok(zone, order, mark, |
1491 | classzone_idx, alloc_flags)) |
1492 | goto try_this_zone; |
1493 | |
1494 | if (zone_reclaim_mode == 0) |
1495 | goto this_zone_full; |
1496 | |
1497 | ret = zone_reclaim(zone, gfp_mask, order); |
1498 | switch (ret) { |
1499 | case ZONE_RECLAIM_NOSCAN: |
1500 | /* did not scan */ |
1501 | goto try_next_zone; |
1502 | case ZONE_RECLAIM_FULL: |
1503 | /* scanned but unreclaimable */ |
1504 | goto this_zone_full; |
1505 | default: |
1506 | /* did we reclaim enough */ |
1507 | if (!zone_watermark_ok(zone, order, mark, |
1508 | classzone_idx, alloc_flags)) |
1509 | goto this_zone_full; |
1510 | } |
1511 | } |
1512 | |
1513 | try_this_zone: |
1514 | page = buffered_rmqueue(preferred_zone, zone, order, |
1515 | gfp_mask, migratetype); |
1516 | if (page) |
1517 | break; |
1518 | this_zone_full: |
1519 | if (NUMA_BUILD) |
1520 | zlc_mark_zone_full(zonelist, z); |
1521 | try_next_zone: |
1522 | if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { |
1523 | /* |
1524 | * we do zlc_setup after the first zone is tried but only |
1525 | * if there are multiple nodes make it worthwhile |
1526 | */ |
1527 | allowednodes = zlc_setup(zonelist, alloc_flags); |
1528 | zlc_active = 1; |
1529 | did_zlc_setup = 1; |
1530 | } |
1531 | } |
1532 | |
1533 | if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { |
1534 | /* Disable zlc cache for second zonelist scan */ |
1535 | zlc_active = 0; |
1536 | goto zonelist_scan; |
1537 | } |
1538 | return page; |
1539 | } |
1540 | |
1541 | static inline int |
1542 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, |
1543 | unsigned long pages_reclaimed) |
1544 | { |
1545 | /* Do not loop if specifically requested */ |
1546 | if (gfp_mask & __GFP_NORETRY) |
1547 | return 0; |
1548 | |
1549 | /* |
1550 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER |
1551 | * means __GFP_NOFAIL, but that may not be true in other |
1552 | * implementations. |
1553 | */ |
1554 | if (order <= PAGE_ALLOC_COSTLY_ORDER) |
1555 | return 1; |
1556 | |
1557 | /* |
1558 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is |
1559 | * specified, then we retry until we no longer reclaim any pages |
1560 | * (above), or we've reclaimed an order of pages at least as |
1561 | * large as the allocation's order. In both cases, if the |
1562 | * allocation still fails, we stop retrying. |
1563 | */ |
1564 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) |
1565 | return 1; |
1566 | |
1567 | /* |
1568 | * Don't let big-order allocations loop unless the caller |
1569 | * explicitly requests that. |
1570 | */ |
1571 | if (gfp_mask & __GFP_NOFAIL) |
1572 | return 1; |
1573 | |
1574 | return 0; |
1575 | } |
1576 | |
1577 | static inline struct page * |
1578 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, |
1579 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
1580 | nodemask_t *nodemask, struct zone *preferred_zone, |
1581 | int migratetype) |
1582 | { |
1583 | struct page *page; |
1584 | |
1585 | /* Acquire the OOM killer lock for the zones in zonelist */ |
1586 | if (!try_set_zone_oom(zonelist, gfp_mask)) { |
1587 | schedule_timeout_uninterruptible(1); |
1588 | return NULL; |
1589 | } |
1590 | |
1591 | /* |
1592 | * Go through the zonelist yet one more time, keep very high watermark |
1593 | * here, this is only to catch a parallel oom killing, we must fail if |
1594 | * we're still under heavy pressure. |
1595 | */ |
1596 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, |
1597 | order, zonelist, high_zoneidx, |
1598 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, |
1599 | preferred_zone, migratetype); |
1600 | if (page) |
1601 | goto out; |
1602 | |
1603 | /* The OOM killer will not help higher order allocs */ |
1604 | if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL)) |
1605 | goto out; |
1606 | |
1607 | /* Exhausted what can be done so it's blamo time */ |
1608 | out_of_memory(zonelist, gfp_mask, order); |
1609 | |
1610 | out: |
1611 | clear_zonelist_oom(zonelist, gfp_mask); |
1612 | return page; |
1613 | } |
1614 | |
1615 | /* The really slow allocator path where we enter direct reclaim */ |
1616 | static inline struct page * |
1617 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, |
1618 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
1619 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, |
1620 | int migratetype, unsigned long *did_some_progress) |
1621 | { |
1622 | struct page *page = NULL; |
1623 | struct reclaim_state reclaim_state; |
1624 | struct task_struct *p = current; |
1625 | |
1626 | cond_resched(); |
1627 | |
1628 | /* We now go into synchronous reclaim */ |
1629 | cpuset_memory_pressure_bump(); |
1630 | |
1631 | /* |
1632 | * The task's cpuset might have expanded its set of allowable nodes |
1633 | */ |
1634 | p->flags |= PF_MEMALLOC; |
1635 | lockdep_set_current_reclaim_state(gfp_mask); |
1636 | reclaim_state.reclaimed_slab = 0; |
1637 | p->reclaim_state = &reclaim_state; |
1638 | |
1639 | *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); |
1640 | |
1641 | p->reclaim_state = NULL; |
1642 | lockdep_clear_current_reclaim_state(); |
1643 | p->flags &= ~PF_MEMALLOC; |
1644 | |
1645 | cond_resched(); |
1646 | |
1647 | if (order != 0) |
1648 | drain_all_pages(); |
1649 | |
1650 | if (likely(*did_some_progress)) |
1651 | page = get_page_from_freelist(gfp_mask, nodemask, order, |
1652 | zonelist, high_zoneidx, |
1653 | alloc_flags, preferred_zone, |
1654 | migratetype); |
1655 | return page; |
1656 | } |
1657 | |
1658 | /* |
1659 | * This is called in the allocator slow-path if the allocation request is of |
1660 | * sufficient urgency to ignore watermarks and take other desperate measures |
1661 | */ |
1662 | static inline struct page * |
1663 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, |
1664 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
1665 | nodemask_t *nodemask, struct zone *preferred_zone, |
1666 | int migratetype) |
1667 | { |
1668 | struct page *page; |
1669 | |
1670 | do { |
1671 | page = get_page_from_freelist(gfp_mask, nodemask, order, |
1672 | zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, |
1673 | preferred_zone, migratetype); |
1674 | |
1675 | if (!page && gfp_mask & __GFP_NOFAIL) |
1676 | congestion_wait(BLK_RW_ASYNC, HZ/50); |
1677 | } while (!page && (gfp_mask & __GFP_NOFAIL)); |
1678 | |
1679 | return page; |
1680 | } |
1681 | |
1682 | static inline |
1683 | void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, |
1684 | enum zone_type high_zoneidx) |
1685 | { |
1686 | struct zoneref *z; |
1687 | struct zone *zone; |
1688 | |
1689 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) |
1690 | wakeup_kswapd(zone, order); |
1691 | } |
1692 | |
1693 | static inline int |
1694 | gfp_to_alloc_flags(gfp_t gfp_mask) |
1695 | { |
1696 | struct task_struct *p = current; |
1697 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; |
1698 | const gfp_t wait = gfp_mask & __GFP_WAIT; |
1699 | |
1700 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ |
1701 | BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); |
1702 | |
1703 | /* |
1704 | * The caller may dip into page reserves a bit more if the caller |
1705 | * cannot run direct reclaim, or if the caller has realtime scheduling |
1706 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will |
1707 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). |
1708 | */ |
1709 | alloc_flags |= (gfp_mask & __GFP_HIGH); |
1710 | |
1711 | if (!wait) { |
1712 | alloc_flags |= ALLOC_HARDER; |
1713 | /* |
1714 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. |
1715 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. |
1716 | */ |
1717 | alloc_flags &= ~ALLOC_CPUSET; |
1718 | } else if (unlikely(rt_task(p))) |
1719 | alloc_flags |= ALLOC_HARDER; |
1720 | |
1721 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { |
1722 | if (!in_interrupt() && |
1723 | ((p->flags & PF_MEMALLOC) || |
1724 | unlikely(test_thread_flag(TIF_MEMDIE)))) |
1725 | alloc_flags |= ALLOC_NO_WATERMARKS; |
1726 | } |
1727 | |
1728 | return alloc_flags; |
1729 | } |
1730 | |
1731 | static inline struct page * |
1732 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, |
1733 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
1734 | nodemask_t *nodemask, struct zone *preferred_zone, |
1735 | int migratetype) |
1736 | { |
1737 | const gfp_t wait = gfp_mask & __GFP_WAIT; |
1738 | struct page *page = NULL; |
1739 | int alloc_flags; |
1740 | unsigned long pages_reclaimed = 0; |
1741 | unsigned long did_some_progress; |
1742 | struct task_struct *p = current; |
1743 | |
1744 | /* |
1745 | * In the slowpath, we sanity check order to avoid ever trying to |
1746 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may |
1747 | * be using allocators in order of preference for an area that is |
1748 | * too large. |
1749 | */ |
1750 | if (order >= MAX_ORDER) { |
1751 | WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); |
1752 | return NULL; |
1753 | } |
1754 | |
1755 | /* |
1756 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and |
1757 | * __GFP_NOWARN set) should not cause reclaim since the subsystem |
1758 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim |
1759 | * using a larger set of nodes after it has established that the |
1760 | * allowed per node queues are empty and that nodes are |
1761 | * over allocated. |
1762 | */ |
1763 | if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) |
1764 | goto nopage; |
1765 | |
1766 | wake_all_kswapd(order, zonelist, high_zoneidx); |
1767 | |
1768 | /* |
1769 | * OK, we're below the kswapd watermark and have kicked background |
1770 | * reclaim. Now things get more complex, so set up alloc_flags according |
1771 | * to how we want to proceed. |
1772 | */ |
1773 | alloc_flags = gfp_to_alloc_flags(gfp_mask); |
1774 | |
1775 | restart: |
1776 | /* This is the last chance, in general, before the goto nopage. */ |
1777 | page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, |
1778 | high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, |
1779 | preferred_zone, migratetype); |
1780 | if (page) |
1781 | goto got_pg; |
1782 | |
1783 | rebalance: |
1784 | /* Allocate without watermarks if the context allows */ |
1785 | if (alloc_flags & ALLOC_NO_WATERMARKS) { |
1786 | page = __alloc_pages_high_priority(gfp_mask, order, |
1787 | zonelist, high_zoneidx, nodemask, |
1788 | preferred_zone, migratetype); |
1789 | if (page) |
1790 | goto got_pg; |
1791 | } |
1792 | |
1793 | /* Atomic allocations - we can't balance anything */ |
1794 | if (!wait) |
1795 | goto nopage; |
1796 | |
1797 | /* Avoid recursion of direct reclaim */ |
1798 | if (p->flags & PF_MEMALLOC) |
1799 | goto nopage; |
1800 | |
1801 | /* Avoid allocations with no watermarks from looping endlessly */ |
1802 | if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) |
1803 | goto nopage; |
1804 | |
1805 | /* Try direct reclaim and then allocating */ |
1806 | page = __alloc_pages_direct_reclaim(gfp_mask, order, |
1807 | zonelist, high_zoneidx, |
1808 | nodemask, |
1809 | alloc_flags, preferred_zone, |
1810 | migratetype, &did_some_progress); |
1811 | if (page) |
1812 | goto got_pg; |
1813 | |
1814 | /* |
1815 | * If we failed to make any progress reclaiming, then we are |
1816 | * running out of options and have to consider going OOM |
1817 | */ |
1818 | if (!did_some_progress) { |
1819 | if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { |
1820 | if (oom_killer_disabled) |
1821 | goto nopage; |
1822 | page = __alloc_pages_may_oom(gfp_mask, order, |
1823 | zonelist, high_zoneidx, |
1824 | nodemask, preferred_zone, |
1825 | migratetype); |
1826 | if (page) |
1827 | goto got_pg; |
1828 | |
1829 | /* |
1830 | * The OOM killer does not trigger for high-order |
1831 | * ~__GFP_NOFAIL allocations so if no progress is being |
1832 | * made, there are no other options and retrying is |
1833 | * unlikely to help. |
1834 | */ |
1835 | if (order > PAGE_ALLOC_COSTLY_ORDER && |
1836 | !(gfp_mask & __GFP_NOFAIL)) |
1837 | goto nopage; |
1838 | |
1839 | goto restart; |
1840 | } |
1841 | } |
1842 | |
1843 | /* Check if we should retry the allocation */ |
1844 | pages_reclaimed += did_some_progress; |
1845 | if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { |
1846 | /* Wait for some write requests to complete then retry */ |
1847 | congestion_wait(BLK_RW_ASYNC, HZ/50); |
1848 | goto rebalance; |
1849 | } |
1850 | |
1851 | nopage: |
1852 | if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { |
1853 | printk(KERN_WARNING "%s: page allocation failure." |
1854 | " order:%d, mode:0x%x\n", |
1855 | p->comm, order, gfp_mask); |
1856 | dump_stack(); |
1857 | show_mem(); |
1858 | } |
1859 | return page; |
1860 | got_pg: |
1861 | if (kmemcheck_enabled) |
1862 | kmemcheck_pagealloc_alloc(page, order, gfp_mask); |
1863 | return page; |
1864 | |
1865 | } |
1866 | |
1867 | /* |
1868 | * This is the 'heart' of the zoned buddy allocator. |
1869 | */ |
1870 | struct page * |
1871 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, |
1872 | struct zonelist *zonelist, nodemask_t *nodemask) |
1873 | { |
1874 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); |
1875 | struct zone *preferred_zone; |
1876 | struct page *page; |
1877 | int migratetype = allocflags_to_migratetype(gfp_mask); |
1878 | |
1879 | gfp_mask &= gfp_allowed_mask; |
1880 | |
1881 | lockdep_trace_alloc(gfp_mask); |
1882 | |
1883 | might_sleep_if(gfp_mask & __GFP_WAIT); |
1884 | |
1885 | if (should_fail_alloc_page(gfp_mask, order)) |
1886 | return NULL; |
1887 | |
1888 | /* |
1889 | * Check the zones suitable for the gfp_mask contain at least one |
1890 | * valid zone. It's possible to have an empty zonelist as a result |
1891 | * of GFP_THISNODE and a memoryless node |
1892 | */ |
1893 | if (unlikely(!zonelist->_zonerefs->zone)) |
1894 | return NULL; |
1895 | |
1896 | /* The preferred zone is used for statistics later */ |
1897 | first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); |
1898 | if (!preferred_zone) |
1899 | return NULL; |
1900 | |
1901 | /* First allocation attempt */ |
1902 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, |
1903 | zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, |
1904 | preferred_zone, migratetype); |
1905 | if (unlikely(!page)) |
1906 | page = __alloc_pages_slowpath(gfp_mask, order, |
1907 | zonelist, high_zoneidx, nodemask, |
1908 | preferred_zone, migratetype); |
1909 | |
1910 | return page; |
1911 | } |
1912 | EXPORT_SYMBOL(__alloc_pages_nodemask); |
1913 | |
1914 | /* |
1915 | * Common helper functions. |
1916 | */ |
1917 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) |
1918 | { |
1919 | struct page * page; |
1920 | page = alloc_pages(gfp_mask, order); |
1921 | if (!page) |
1922 | return 0; |
1923 | return (unsigned long) page_address(page); |
1924 | } |
1925 | |
1926 | EXPORT_SYMBOL(__get_free_pages); |
1927 | |
1928 | unsigned long get_zeroed_page(gfp_t gfp_mask) |
1929 | { |
1930 | struct page * page; |
1931 | |
1932 | /* |
1933 | * get_zeroed_page() returns a 32-bit address, which cannot represent |
1934 | * a highmem page |
1935 | */ |
1936 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); |
1937 | |
1938 | page = alloc_pages(gfp_mask | __GFP_ZERO, 0); |
1939 | if (page) |
1940 | return (unsigned long) page_address(page); |
1941 | return 0; |
1942 | } |
1943 | |
1944 | EXPORT_SYMBOL(get_zeroed_page); |
1945 | |
1946 | void __pagevec_free(struct pagevec *pvec) |
1947 | { |
1948 | int i = pagevec_count(pvec); |
1949 | |
1950 | while (--i >= 0) |
1951 | free_hot_cold_page(pvec->pages[i], pvec->cold); |
1952 | } |
1953 | |
1954 | void __free_pages(struct page *page, unsigned int order) |
1955 | { |
1956 | if (put_page_testzero(page)) { |
1957 | if (order == 0) |
1958 | free_hot_page(page); |
1959 | else |
1960 | __free_pages_ok(page, order); |
1961 | } |
1962 | } |
1963 | |
1964 | EXPORT_SYMBOL(__free_pages); |
1965 | |
1966 | void free_pages(unsigned long addr, unsigned int order) |
1967 | { |
1968 | if (addr != 0) { |
1969 | VM_BUG_ON(!virt_addr_valid((void *)addr)); |
1970 | __free_pages(virt_to_page((void *)addr), order); |
1971 | } |
1972 | } |
1973 | |
1974 | EXPORT_SYMBOL(free_pages); |
1975 | |
1976 | /** |
1977 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. |
1978 | * @size: the number of bytes to allocate |
1979 | * @gfp_mask: GFP flags for the allocation |
1980 | * |
1981 | * This function is similar to alloc_pages(), except that it allocates the |
1982 | * minimum number of pages to satisfy the request. alloc_pages() can only |
1983 | * allocate memory in power-of-two pages. |
1984 | * |
1985 | * This function is also limited by MAX_ORDER. |
1986 | * |
1987 | * Memory allocated by this function must be released by free_pages_exact(). |
1988 | */ |
1989 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) |
1990 | { |
1991 | unsigned int order = get_order(size); |
1992 | unsigned long addr; |
1993 | |
1994 | addr = __get_free_pages(gfp_mask, order); |
1995 | if (addr) { |
1996 | unsigned long alloc_end = addr + (PAGE_SIZE << order); |
1997 | unsigned long used = addr + PAGE_ALIGN(size); |
1998 | |
1999 | split_page(virt_to_page((void *)addr), order); |
2000 | while (used < alloc_end) { |
2001 | free_page(used); |
2002 | used += PAGE_SIZE; |
2003 | } |
2004 | } |
2005 | |
2006 | return (void *)addr; |
2007 | } |
2008 | EXPORT_SYMBOL(alloc_pages_exact); |
2009 | |
2010 | /** |
2011 | * free_pages_exact - release memory allocated via alloc_pages_exact() |
2012 | * @virt: the value returned by alloc_pages_exact. |
2013 | * @size: size of allocation, same value as passed to alloc_pages_exact(). |
2014 | * |
2015 | * Release the memory allocated by a previous call to alloc_pages_exact. |
2016 | */ |
2017 | void free_pages_exact(void *virt, size_t size) |
2018 | { |
2019 | unsigned long addr = (unsigned long)virt; |
2020 | unsigned long end = addr + PAGE_ALIGN(size); |
2021 | |
2022 | while (addr < end) { |
2023 | free_page(addr); |
2024 | addr += PAGE_SIZE; |
2025 | } |
2026 | } |
2027 | EXPORT_SYMBOL(free_pages_exact); |
2028 | |
2029 | static unsigned int nr_free_zone_pages(int offset) |
2030 | { |
2031 | struct zoneref *z; |
2032 | struct zone *zone; |
2033 | |
2034 | /* Just pick one node, since fallback list is circular */ |
2035 | unsigned int sum = 0; |
2036 | |
2037 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); |
2038 | |
2039 | for_each_zone_zonelist(zone, z, zonelist, offset) { |
2040 | unsigned long size = zone->present_pages; |
2041 | unsigned long high = high_wmark_pages(zone); |
2042 | if (size > high) |
2043 | sum += size - high; |
2044 | } |
2045 | |
2046 | return sum; |
2047 | } |
2048 | |
2049 | /* |
2050 | * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL |
2051 | */ |
2052 | unsigned int nr_free_buffer_pages(void) |
2053 | { |
2054 | return nr_free_zone_pages(gfp_zone(GFP_USER)); |
2055 | } |
2056 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); |
2057 | |
2058 | /* |
2059 | * Amount of free RAM allocatable within all zones |
2060 | */ |
2061 | unsigned int nr_free_pagecache_pages(void) |
2062 | { |
2063 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); |
2064 | } |
2065 | |
2066 | static inline void show_node(struct zone *zone) |
2067 | { |
2068 | if (NUMA_BUILD) |
2069 | printk("Node %d ", zone_to_nid(zone)); |
2070 | } |
2071 | |
2072 | void si_meminfo(struct sysinfo *val) |
2073 | { |
2074 | val->totalram = totalram_pages; |
2075 | val->sharedram = 0; |
2076 | val->freeram = global_page_state(NR_FREE_PAGES); |
2077 | val->bufferram = nr_blockdev_pages(); |
2078 | val->totalhigh = totalhigh_pages; |
2079 | val->freehigh = nr_free_highpages(); |
2080 | val->mem_unit = PAGE_SIZE; |
2081 | } |
2082 | |
2083 | EXPORT_SYMBOL(si_meminfo); |
2084 | |
2085 | #ifdef CONFIG_NUMA |
2086 | void si_meminfo_node(struct sysinfo *val, int nid) |
2087 | { |
2088 | pg_data_t *pgdat = NODE_DATA(nid); |
2089 | |
2090 | val->totalram = pgdat->node_present_pages; |
2091 | val->freeram = node_page_state(nid, NR_FREE_PAGES); |
2092 | #ifdef CONFIG_HIGHMEM |
2093 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; |
2094 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], |
2095 | NR_FREE_PAGES); |
2096 | #else |
2097 | val->totalhigh = 0; |
2098 | val->freehigh = 0; |
2099 | #endif |
2100 | val->mem_unit = PAGE_SIZE; |
2101 | } |
2102 | #endif |
2103 | |
2104 | #define K(x) ((x) << (PAGE_SHIFT-10)) |
2105 | |
2106 | /* |
2107 | * Show free area list (used inside shift_scroll-lock stuff) |
2108 | * We also calculate the percentage fragmentation. We do this by counting the |
2109 | * memory on each free list with the exception of the first item on the list. |
2110 | */ |
2111 | void show_free_areas(void) |
2112 | { |
2113 | int cpu; |
2114 | struct zone *zone; |
2115 | |
2116 | for_each_populated_zone(zone) { |
2117 | show_node(zone); |
2118 | printk("%s per-cpu:\n", zone->name); |
2119 | |
2120 | for_each_online_cpu(cpu) { |
2121 | struct per_cpu_pageset *pageset; |
2122 | |
2123 | pageset = zone_pcp(zone, cpu); |
2124 | |
2125 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", |
2126 | cpu, pageset->pcp.high, |
2127 | pageset->pcp.batch, pageset->pcp.count); |
2128 | } |
2129 | } |
2130 | |
2131 | printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n" |
2132 | " inactive_file:%lu" |
2133 | " unevictable:%lu" |
2134 | " dirty:%lu writeback:%lu unstable:%lu\n" |
2135 | " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n", |
2136 | global_page_state(NR_ACTIVE_ANON), |
2137 | global_page_state(NR_ACTIVE_FILE), |
2138 | global_page_state(NR_INACTIVE_ANON), |
2139 | global_page_state(NR_INACTIVE_FILE), |
2140 | global_page_state(NR_UNEVICTABLE), |
2141 | global_page_state(NR_FILE_DIRTY), |
2142 | global_page_state(NR_WRITEBACK), |
2143 | global_page_state(NR_UNSTABLE_NFS), |
2144 | global_page_state(NR_FREE_PAGES), |
2145 | global_page_state(NR_SLAB_RECLAIMABLE) + |
2146 | global_page_state(NR_SLAB_UNRECLAIMABLE), |
2147 | global_page_state(NR_FILE_MAPPED), |
2148 | global_page_state(NR_PAGETABLE), |
2149 | global_page_state(NR_BOUNCE)); |
2150 | |
2151 | for_each_populated_zone(zone) { |
2152 | int i; |
2153 | |
2154 | show_node(zone); |
2155 | printk("%s" |
2156 | " free:%lukB" |
2157 | " min:%lukB" |
2158 | " low:%lukB" |
2159 | " high:%lukB" |
2160 | " active_anon:%lukB" |
2161 | " inactive_anon:%lukB" |
2162 | " active_file:%lukB" |
2163 | " inactive_file:%lukB" |
2164 | " unevictable:%lukB" |
2165 | " present:%lukB" |
2166 | " pages_scanned:%lu" |
2167 | " all_unreclaimable? %s" |
2168 | "\n", |
2169 | zone->name, |
2170 | K(zone_page_state(zone, NR_FREE_PAGES)), |
2171 | K(min_wmark_pages(zone)), |
2172 | K(low_wmark_pages(zone)), |
2173 | K(high_wmark_pages(zone)), |
2174 | K(zone_page_state(zone, NR_ACTIVE_ANON)), |
2175 | K(zone_page_state(zone, NR_INACTIVE_ANON)), |
2176 | K(zone_page_state(zone, NR_ACTIVE_FILE)), |
2177 | K(zone_page_state(zone, NR_INACTIVE_FILE)), |
2178 | K(zone_page_state(zone, NR_UNEVICTABLE)), |
2179 | K(zone->present_pages), |
2180 | zone->pages_scanned, |
2181 | (zone_is_all_unreclaimable(zone) ? "yes" : "no") |
2182 | ); |
2183 | printk("lowmem_reserve[]:"); |
2184 | for (i = 0; i < MAX_NR_ZONES; i++) |
2185 | printk(" %lu", zone->lowmem_reserve[i]); |
2186 | printk("\n"); |
2187 | } |
2188 | |
2189 | for_each_populated_zone(zone) { |
2190 | unsigned long nr[MAX_ORDER], flags, order, total = 0; |
2191 | |
2192 | show_node(zone); |
2193 | printk("%s: ", zone->name); |
2194 | |
2195 | spin_lock_irqsave(&zone->lock, flags); |
2196 | for (order = 0; order < MAX_ORDER; order++) { |
2197 | nr[order] = zone->free_area[order].nr_free; |
2198 | total += nr[order] << order; |
2199 | } |
2200 | spin_unlock_irqrestore(&zone->lock, flags); |
2201 | for (order = 0; order < MAX_ORDER; order++) |
2202 | printk("%lu*%lukB ", nr[order], K(1UL) << order); |
2203 | printk("= %lukB\n", K(total)); |
2204 | } |
2205 | |
2206 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); |
2207 | |
2208 | show_swap_cache_info(); |
2209 | } |
2210 | |
2211 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) |
2212 | { |
2213 | zoneref->zone = zone; |
2214 | zoneref->zone_idx = zone_idx(zone); |
2215 | } |
2216 | |
2217 | /* |
2218 | * Builds allocation fallback zone lists. |
2219 | * |
2220 | * Add all populated zones of a node to the zonelist. |
2221 | */ |
2222 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, |
2223 | int nr_zones, enum zone_type zone_type) |
2224 | { |
2225 | struct zone *zone; |
2226 | |
2227 | BUG_ON(zone_type >= MAX_NR_ZONES); |
2228 | zone_type++; |
2229 | |
2230 | do { |
2231 | zone_type--; |
2232 | zone = pgdat->node_zones + zone_type; |
2233 | if (populated_zone(zone)) { |
2234 | zoneref_set_zone(zone, |
2235 | &zonelist->_zonerefs[nr_zones++]); |
2236 | check_highest_zone(zone_type); |
2237 | } |
2238 | |
2239 | } while (zone_type); |
2240 | return nr_zones; |
2241 | } |
2242 | |
2243 | |
2244 | /* |
2245 | * zonelist_order: |
2246 | * 0 = automatic detection of better ordering. |
2247 | * 1 = order by ([node] distance, -zonetype) |
2248 | * 2 = order by (-zonetype, [node] distance) |
2249 | * |
2250 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create |
2251 | * the same zonelist. So only NUMA can configure this param. |
2252 | */ |
2253 | #define ZONELIST_ORDER_DEFAULT 0 |
2254 | #define ZONELIST_ORDER_NODE 1 |
2255 | #define ZONELIST_ORDER_ZONE 2 |
2256 | |
2257 | /* zonelist order in the kernel. |
2258 | * set_zonelist_order() will set this to NODE or ZONE. |
2259 | */ |
2260 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; |
2261 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; |
2262 | |
2263 | |
2264 | #ifdef CONFIG_NUMA |
2265 | /* The value user specified ....changed by config */ |
2266 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; |
2267 | /* string for sysctl */ |
2268 | #define NUMA_ZONELIST_ORDER_LEN 16 |
2269 | char numa_zonelist_order[16] = "default"; |
2270 | |
2271 | /* |
2272 | * interface for configure zonelist ordering. |
2273 | * command line option "numa_zonelist_order" |
2274 | * = "[dD]efault - default, automatic configuration. |
2275 | * = "[nN]ode - order by node locality, then by zone within node |
2276 | * = "[zZ]one - order by zone, then by locality within zone |
2277 | */ |
2278 | |
2279 | static int __parse_numa_zonelist_order(char *s) |
2280 | { |
2281 | if (*s == 'd' || *s == 'D') { |
2282 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; |
2283 | } else if (*s == 'n' || *s == 'N') { |
2284 | user_zonelist_order = ZONELIST_ORDER_NODE; |
2285 | } else if (*s == 'z' || *s == 'Z') { |
2286 | user_zonelist_order = ZONELIST_ORDER_ZONE; |
2287 | } else { |
2288 | printk(KERN_WARNING |
2289 | "Ignoring invalid numa_zonelist_order value: " |
2290 | "%s\n", s); |
2291 | return -EINVAL; |
2292 | } |
2293 | return 0; |
2294 | } |
2295 | |
2296 | static __init int setup_numa_zonelist_order(char *s) |
2297 | { |
2298 | if (s) |
2299 | return __parse_numa_zonelist_order(s); |
2300 | return 0; |
2301 | } |
2302 | early_param("numa_zonelist_order", setup_numa_zonelist_order); |
2303 | |
2304 | /* |
2305 | * sysctl handler for numa_zonelist_order |
2306 | */ |
2307 | int numa_zonelist_order_handler(ctl_table *table, int write, |
2308 | struct file *file, void __user *buffer, size_t *length, |
2309 | loff_t *ppos) |
2310 | { |
2311 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; |
2312 | int ret; |
2313 | |
2314 | if (write) |
2315 | strncpy(saved_string, (char*)table->data, |
2316 | NUMA_ZONELIST_ORDER_LEN); |
2317 | ret = proc_dostring(table, write, file, buffer, length, ppos); |
2318 | if (ret) |
2319 | return ret; |
2320 | if (write) { |
2321 | int oldval = user_zonelist_order; |
2322 | if (__parse_numa_zonelist_order((char*)table->data)) { |
2323 | /* |
2324 | * bogus value. restore saved string |
2325 | */ |
2326 | strncpy((char*)table->data, saved_string, |
2327 | NUMA_ZONELIST_ORDER_LEN); |
2328 | user_zonelist_order = oldval; |
2329 | } else if (oldval != user_zonelist_order) |
2330 | build_all_zonelists(); |
2331 | } |
2332 | return 0; |
2333 | } |
2334 | |
2335 | |
2336 | #define MAX_NODE_LOAD (nr_online_nodes) |
2337 | static int node_load[MAX_NUMNODES]; |
2338 | |
2339 | /** |
2340 | * find_next_best_node - find the next node that should appear in a given node's fallback list |
2341 | * @node: node whose fallback list we're appending |
2342 | * @used_node_mask: nodemask_t of already used nodes |
2343 | * |
2344 | * We use a number of factors to determine which is the next node that should |
2345 | * appear on a given node's fallback list. The node should not have appeared |
2346 | * already in @node's fallback list, and it should be the next closest node |
2347 | * according to the distance array (which contains arbitrary distance values |
2348 | * from each node to each node in the system), and should also prefer nodes |
2349 | * with no CPUs, since presumably they'll have very little allocation pressure |
2350 | * on them otherwise. |
2351 | * It returns -1 if no node is found. |
2352 | */ |
2353 | static int find_next_best_node(int node, nodemask_t *used_node_mask) |
2354 | { |
2355 | int n, val; |
2356 | int min_val = INT_MAX; |
2357 | int best_node = -1; |
2358 | const struct cpumask *tmp = cpumask_of_node(0); |
2359 | |
2360 | /* Use the local node if we haven't already */ |
2361 | if (!node_isset(node, *used_node_mask)) { |
2362 | node_set(node, *used_node_mask); |
2363 | return node; |
2364 | } |
2365 | |
2366 | for_each_node_state(n, N_HIGH_MEMORY) { |
2367 | |
2368 | /* Don't want a node to appear more than once */ |
2369 | if (node_isset(n, *used_node_mask)) |
2370 | continue; |
2371 | |
2372 | /* Use the distance array to find the distance */ |
2373 | val = node_distance(node, n); |
2374 | |
2375 | /* Penalize nodes under us ("prefer the next node") */ |
2376 | val += (n < node); |
2377 | |
2378 | /* Give preference to headless and unused nodes */ |
2379 | tmp = cpumask_of_node(n); |
2380 | if (!cpumask_empty(tmp)) |
2381 | val += PENALTY_FOR_NODE_WITH_CPUS; |
2382 | |
2383 | /* Slight preference for less loaded node */ |
2384 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); |
2385 | val += node_load[n]; |
2386 | |
2387 | if (val < min_val) { |
2388 | min_val = val; |
2389 | best_node = n; |
2390 | } |
2391 | } |
2392 | |
2393 | if (best_node >= 0) |
2394 | node_set(best_node, *used_node_mask); |
2395 | |
2396 | return best_node; |
2397 | } |
2398 | |
2399 | |
2400 | /* |
2401 | * Build zonelists ordered by node and zones within node. |
2402 | * This results in maximum locality--normal zone overflows into local |
2403 | * DMA zone, if any--but risks exhausting DMA zone. |
2404 | */ |
2405 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) |
2406 | { |
2407 | int j; |
2408 | struct zonelist *zonelist; |
2409 | |
2410 | zonelist = &pgdat->node_zonelists[0]; |
2411 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) |
2412 | ; |
2413 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, |
2414 | MAX_NR_ZONES - 1); |
2415 | zonelist->_zonerefs[j].zone = NULL; |
2416 | zonelist->_zonerefs[j].zone_idx = 0; |
2417 | } |
2418 | |
2419 | /* |
2420 | * Build gfp_thisnode zonelists |
2421 | */ |
2422 | static void build_thisnode_zonelists(pg_data_t *pgdat) |
2423 | { |
2424 | int j; |
2425 | struct zonelist *zonelist; |
2426 | |
2427 | zonelist = &pgdat->node_zonelists[1]; |
2428 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); |
2429 | zonelist->_zonerefs[j].zone = NULL; |
2430 | zonelist->_zonerefs[j].zone_idx = 0; |
2431 | } |
2432 | |
2433 | /* |
2434 | * Build zonelists ordered by zone and nodes within zones. |
2435 | * This results in conserving DMA zone[s] until all Normal memory is |
2436 | * exhausted, but results in overflowing to remote node while memory |
2437 | * may still exist in local DMA zone. |
2438 | */ |
2439 | static int node_order[MAX_NUMNODES]; |
2440 | |
2441 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) |
2442 | { |
2443 | int pos, j, node; |
2444 | int zone_type; /* needs to be signed */ |
2445 | struct zone *z; |
2446 | struct zonelist *zonelist; |
2447 | |
2448 | zonelist = &pgdat->node_zonelists[0]; |
2449 | pos = 0; |
2450 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { |
2451 | for (j = 0; j < nr_nodes; j++) { |
2452 | node = node_order[j]; |
2453 | z = &NODE_DATA(node)->node_zones[zone_type]; |
2454 | if (populated_zone(z)) { |
2455 | zoneref_set_zone(z, |
2456 | &zonelist->_zonerefs[pos++]); |
2457 | check_highest_zone(zone_type); |
2458 | } |
2459 | } |
2460 | } |
2461 | zonelist->_zonerefs[pos].zone = NULL; |
2462 | zonelist->_zonerefs[pos].zone_idx = 0; |
2463 | } |
2464 | |
2465 | static int default_zonelist_order(void) |
2466 | { |
2467 | int nid, zone_type; |
2468 | unsigned long low_kmem_size,total_size; |
2469 | struct zone *z; |
2470 | int average_size; |
2471 | /* |
2472 | * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. |
2473 | * If they are really small and used heavily, the system can fall |
2474 | * into OOM very easily. |
2475 | * This function detect ZONE_DMA/DMA32 size and confgigures zone order. |
2476 | */ |
2477 | /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ |
2478 | low_kmem_size = 0; |
2479 | total_size = 0; |
2480 | for_each_online_node(nid) { |
2481 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { |
2482 | z = &NODE_DATA(nid)->node_zones[zone_type]; |
2483 | if (populated_zone(z)) { |
2484 | if (zone_type < ZONE_NORMAL) |
2485 | low_kmem_size += z->present_pages; |
2486 | total_size += z->present_pages; |
2487 | } |
2488 | } |
2489 | } |
2490 | if (!low_kmem_size || /* there are no DMA area. */ |
2491 | low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ |
2492 | return ZONELIST_ORDER_NODE; |
2493 | /* |
2494 | * look into each node's config. |
2495 | * If there is a node whose DMA/DMA32 memory is very big area on |
2496 | * local memory, NODE_ORDER may be suitable. |
2497 | */ |
2498 | average_size = total_size / |
2499 | (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); |
2500 | for_each_online_node(nid) { |
2501 | low_kmem_size = 0; |
2502 | total_size = 0; |
2503 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { |
2504 | z = &NODE_DATA(nid)->node_zones[zone_type]; |
2505 | if (populated_zone(z)) { |
2506 | if (zone_type < ZONE_NORMAL) |
2507 | low_kmem_size += z->present_pages; |
2508 | total_size += z->present_pages; |
2509 | } |
2510 | } |
2511 | if (low_kmem_size && |
2512 | total_size > average_size && /* ignore small node */ |
2513 | low_kmem_size > total_size * 70/100) |
2514 | return ZONELIST_ORDER_NODE; |
2515 | } |
2516 | return ZONELIST_ORDER_ZONE; |
2517 | } |
2518 | |
2519 | static void set_zonelist_order(void) |
2520 | { |
2521 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) |
2522 | current_zonelist_order = default_zonelist_order(); |
2523 | else |
2524 | current_zonelist_order = user_zonelist_order; |
2525 | } |
2526 | |
2527 | static void build_zonelists(pg_data_t *pgdat) |
2528 | { |
2529 | int j, node, load; |
2530 | enum zone_type i; |
2531 | nodemask_t used_mask; |
2532 | int local_node, prev_node; |
2533 | struct zonelist *zonelist; |
2534 | int order = current_zonelist_order; |
2535 | |
2536 | /* initialize zonelists */ |
2537 | for (i = 0; i < MAX_ZONELISTS; i++) { |
2538 | zonelist = pgdat->node_zonelists + i; |
2539 | zonelist->_zonerefs[0].zone = NULL; |
2540 | zonelist->_zonerefs[0].zone_idx = 0; |
2541 | } |
2542 | |
2543 | /* NUMA-aware ordering of nodes */ |
2544 | local_node = pgdat->node_id; |
2545 | load = nr_online_nodes; |
2546 | prev_node = local_node; |
2547 | nodes_clear(used_mask); |
2548 | |
2549 | memset(node_order, 0, sizeof(node_order)); |
2550 | j = 0; |
2551 | |
2552 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { |
2553 | int distance = node_distance(local_node, node); |
2554 | |
2555 | /* |
2556 | * If another node is sufficiently far away then it is better |
2557 | * to reclaim pages in a zone before going off node. |
2558 | */ |
2559 | if (distance > RECLAIM_DISTANCE) |
2560 | zone_reclaim_mode = 1; |
2561 | |
2562 | /* |
2563 | * We don't want to pressure a particular node. |
2564 | * So adding penalty to the first node in same |
2565 | * distance group to make it round-robin. |
2566 | */ |
2567 | if (distance != node_distance(local_node, prev_node)) |
2568 | node_load[node] = load; |
2569 | |
2570 | prev_node = node; |
2571 | load--; |
2572 | if (order == ZONELIST_ORDER_NODE) |
2573 | build_zonelists_in_node_order(pgdat, node); |
2574 | else |
2575 | node_order[j++] = node; /* remember order */ |
2576 | } |
2577 | |
2578 | if (order == ZONELIST_ORDER_ZONE) { |
2579 | /* calculate node order -- i.e., DMA last! */ |
2580 | build_zonelists_in_zone_order(pgdat, j); |
2581 | } |
2582 | |
2583 | build_thisnode_zonelists(pgdat); |
2584 | } |
2585 | |
2586 | /* Construct the zonelist performance cache - see further mmzone.h */ |
2587 | static void build_zonelist_cache(pg_data_t *pgdat) |
2588 | { |
2589 | struct zonelist *zonelist; |
2590 | struct zonelist_cache *zlc; |
2591 | struct zoneref *z; |
2592 | |
2593 | zonelist = &pgdat->node_zonelists[0]; |
2594 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; |
2595 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); |
2596 | for (z = zonelist->_zonerefs; z->zone; z++) |
2597 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); |
2598 | } |
2599 | |
2600 | |
2601 | #else /* CONFIG_NUMA */ |
2602 | |
2603 | static void set_zonelist_order(void) |
2604 | { |
2605 | current_zonelist_order = ZONELIST_ORDER_ZONE; |
2606 | } |
2607 | |
2608 | static void build_zonelists(pg_data_t *pgdat) |
2609 | { |
2610 | int node, local_node; |
2611 | enum zone_type j; |
2612 | struct zonelist *zonelist; |
2613 | |
2614 | local_node = pgdat->node_id; |
2615 | |
2616 | zonelist = &pgdat->node_zonelists[0]; |
2617 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); |
2618 | |
2619 | /* |
2620 | * Now we build the zonelist so that it contains the zones |
2621 | * of all the other nodes. |
2622 | * We don't want to pressure a particular node, so when |
2623 | * building the zones for node N, we make sure that the |
2624 | * zones coming right after the local ones are those from |
2625 | * node N+1 (modulo N) |
2626 | */ |
2627 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { |
2628 | if (!node_online(node)) |
2629 | continue; |
2630 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, |
2631 | MAX_NR_ZONES - 1); |
2632 | } |
2633 | for (node = 0; node < local_node; node++) { |
2634 | if (!node_online(node)) |
2635 | continue; |
2636 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, |
2637 | MAX_NR_ZONES - 1); |
2638 | } |
2639 | |
2640 | zonelist->_zonerefs[j].zone = NULL; |
2641 | zonelist->_zonerefs[j].zone_idx = 0; |
2642 | } |
2643 | |
2644 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ |
2645 | static void build_zonelist_cache(pg_data_t *pgdat) |
2646 | { |
2647 | pgdat->node_zonelists[0].zlcache_ptr = NULL; |
2648 | } |
2649 | |
2650 | #endif /* CONFIG_NUMA */ |
2651 | |
2652 | /* return values int ....just for stop_machine() */ |
2653 | static int __build_all_zonelists(void *dummy) |
2654 | { |
2655 | int nid; |
2656 | |
2657 | #ifdef CONFIG_NUMA |
2658 | memset(node_load, 0, sizeof(node_load)); |
2659 | #endif |
2660 | for_each_online_node(nid) { |
2661 | pg_data_t *pgdat = NODE_DATA(nid); |
2662 | |
2663 | build_zonelists(pgdat); |
2664 | build_zonelist_cache(pgdat); |
2665 | } |
2666 | return 0; |
2667 | } |
2668 | |
2669 | void build_all_zonelists(void) |
2670 | { |
2671 | set_zonelist_order(); |
2672 | |
2673 | if (system_state == SYSTEM_BOOTING) { |
2674 | __build_all_zonelists(NULL); |
2675 | mminit_verify_zonelist(); |
2676 | cpuset_init_current_mems_allowed(); |
2677 | } else { |
2678 | /* we have to stop all cpus to guarantee there is no user |
2679 | of zonelist */ |
2680 | stop_machine(__build_all_zonelists, NULL, NULL); |
2681 | /* cpuset refresh routine should be here */ |
2682 | } |
2683 | vm_total_pages = nr_free_pagecache_pages(); |
2684 | /* |
2685 | * Disable grouping by mobility if the number of pages in the |
2686 | * system is too low to allow the mechanism to work. It would be |
2687 | * more accurate, but expensive to check per-zone. This check is |
2688 | * made on memory-hotadd so a system can start with mobility |
2689 | * disabled and enable it later |
2690 | */ |
2691 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) |
2692 | page_group_by_mobility_disabled = 1; |
2693 | else |
2694 | page_group_by_mobility_disabled = 0; |
2695 | |
2696 | printk("Built %i zonelists in %s order, mobility grouping %s. " |
2697 | "Total pages: %ld\n", |
2698 | nr_online_nodes, |
2699 | zonelist_order_name[current_zonelist_order], |
2700 | page_group_by_mobility_disabled ? "off" : "on", |
2701 | vm_total_pages); |
2702 | #ifdef CONFIG_NUMA |
2703 | printk("Policy zone: %s\n", zone_names[policy_zone]); |
2704 | #endif |
2705 | } |
2706 | |
2707 | /* |
2708 | * Helper functions to size the waitqueue hash table. |
2709 | * Essentially these want to choose hash table sizes sufficiently |
2710 | * large so that collisions trying to wait on pages are rare. |
2711 | * But in fact, the number of active page waitqueues on typical |
2712 | * systems is ridiculously low, less than 200. So this is even |
2713 | * conservative, even though it seems large. |
2714 | * |
2715 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to |
2716 | * waitqueues, i.e. the size of the waitq table given the number of pages. |
2717 | */ |
2718 | #define PAGES_PER_WAITQUEUE 256 |
2719 | |
2720 | #ifndef CONFIG_MEMORY_HOTPLUG |
2721 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) |
2722 | { |
2723 | unsigned long size = 1; |
2724 | |
2725 | pages /= PAGES_PER_WAITQUEUE; |
2726 | |
2727 | while (size < pages) |
2728 | size <<= 1; |
2729 | |
2730 | /* |
2731 | * Once we have dozens or even hundreds of threads sleeping |
2732 | * on IO we've got bigger problems than wait queue collision. |
2733 | * Limit the size of the wait table to a reasonable size. |
2734 | */ |
2735 | size = min(size, 4096UL); |
2736 | |
2737 | return max(size, 4UL); |
2738 | } |
2739 | #else |
2740 | /* |
2741 | * A zone's size might be changed by hot-add, so it is not possible to determine |
2742 | * a suitable size for its wait_table. So we use the maximum size now. |
2743 | * |
2744 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: |
2745 | * |
2746 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. |
2747 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. |
2748 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. |
2749 | * |
2750 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages |
2751 | * or more by the traditional way. (See above). It equals: |
2752 | * |
2753 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. |
2754 | * ia64(16K page size) : = ( 8G + 4M)byte. |
2755 | * powerpc (64K page size) : = (32G +16M)byte. |
2756 | */ |
2757 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) |
2758 | { |
2759 | return 4096UL; |
2760 | } |
2761 | #endif |
2762 | |
2763 | /* |
2764 | * This is an integer logarithm so that shifts can be used later |
2765 | * to extract the more random high bits from the multiplicative |
2766 | * hash function before the remainder is taken. |
2767 | */ |
2768 | static inline unsigned long wait_table_bits(unsigned long size) |
2769 | { |
2770 | return ffz(~size); |
2771 | } |
2772 | |
2773 | #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) |
2774 | |
2775 | /* |
2776 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number |
2777 | * of blocks reserved is based on min_wmark_pages(zone). The memory within |
2778 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes |
2779 | * higher will lead to a bigger reserve which will get freed as contiguous |
2780 | * blocks as reclaim kicks in |
2781 | */ |
2782 | static void setup_zone_migrate_reserve(struct zone *zone) |
2783 | { |
2784 | unsigned long start_pfn, pfn, end_pfn; |
2785 | struct page *page; |
2786 | unsigned long reserve, block_migratetype; |
2787 | |
2788 | /* Get the start pfn, end pfn and the number of blocks to reserve */ |
2789 | start_pfn = zone->zone_start_pfn; |
2790 | end_pfn = start_pfn + zone->spanned_pages; |
2791 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> |
2792 | pageblock_order; |
2793 | |
2794 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { |
2795 | if (!pfn_valid(pfn)) |
2796 | continue; |
2797 | page = pfn_to_page(pfn); |
2798 | |
2799 | /* Watch out for overlapping nodes */ |
2800 | if (page_to_nid(page) != zone_to_nid(zone)) |
2801 | continue; |
2802 | |
2803 | /* Blocks with reserved pages will never free, skip them. */ |
2804 | if (PageReserved(page)) |
2805 | continue; |
2806 | |
2807 | block_migratetype = get_pageblock_migratetype(page); |
2808 | |
2809 | /* If this block is reserved, account for it */ |
2810 | if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { |
2811 | reserve--; |
2812 | continue; |
2813 | } |
2814 | |
2815 | /* Suitable for reserving if this block is movable */ |
2816 | if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { |
2817 | set_pageblock_migratetype(page, MIGRATE_RESERVE); |
2818 | move_freepages_block(zone, page, MIGRATE_RESERVE); |
2819 | reserve--; |
2820 | continue; |
2821 | } |
2822 | |
2823 | /* |
2824 | * If the reserve is met and this is a previous reserved block, |
2825 | * take it back |
2826 | */ |
2827 | if (block_migratetype == MIGRATE_RESERVE) { |
2828 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); |
2829 | move_freepages_block(zone, page, MIGRATE_MOVABLE); |
2830 | } |
2831 | } |
2832 | } |
2833 | |
2834 | /* |
2835 | * Initially all pages are reserved - free ones are freed |
2836 | * up by free_all_bootmem() once the early boot process is |
2837 | * done. Non-atomic initialization, single-pass. |
2838 | */ |
2839 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, |
2840 | unsigned long start_pfn, enum memmap_context context) |
2841 | { |
2842 | struct page *page; |
2843 | unsigned long end_pfn = start_pfn + size; |
2844 | unsigned long pfn; |
2845 | struct zone *z; |
2846 | |
2847 | if (highest_memmap_pfn < end_pfn - 1) |
2848 | highest_memmap_pfn = end_pfn - 1; |
2849 | |
2850 | z = &NODE_DATA(nid)->node_zones[zone]; |
2851 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { |
2852 | /* |
2853 | * There can be holes in boot-time mem_map[]s |
2854 | * handed to this function. They do not |
2855 | * exist on hotplugged memory. |
2856 | */ |
2857 | if (context == MEMMAP_EARLY) { |
2858 | if (!early_pfn_valid(pfn)) |
2859 | continue; |
2860 | if (!early_pfn_in_nid(pfn, nid)) |
2861 | continue; |
2862 | } |
2863 | page = pfn_to_page(pfn); |
2864 | set_page_links(page, zone, nid, pfn); |
2865 | mminit_verify_page_links(page, zone, nid, pfn); |
2866 | init_page_count(page); |
2867 | reset_page_mapcount(page); |
2868 | SetPageReserved(page); |
2869 | /* |
2870 | * Mark the block movable so that blocks are reserved for |
2871 | * movable at startup. This will force kernel allocations |
2872 | * to reserve their blocks rather than leaking throughout |
2873 | * the address space during boot when many long-lived |
2874 | * kernel allocations are made. Later some blocks near |
2875 | * the start are marked MIGRATE_RESERVE by |
2876 | * setup_zone_migrate_reserve() |
2877 | * |
2878 | * bitmap is created for zone's valid pfn range. but memmap |
2879 | * can be created for invalid pages (for alignment) |
2880 | * check here not to call set_pageblock_migratetype() against |
2881 | * pfn out of zone. |
2882 | */ |
2883 | if ((z->zone_start_pfn <= pfn) |
2884 | && (pfn < z->zone_start_pfn + z->spanned_pages) |
2885 | && !(pfn & (pageblock_nr_pages - 1))) |
2886 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); |
2887 | |
2888 | INIT_LIST_HEAD(&page->lru); |
2889 | #ifdef WANT_PAGE_VIRTUAL |
2890 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ |
2891 | if (!is_highmem_idx(zone)) |
2892 | set_page_address(page, __va(pfn << PAGE_SHIFT)); |
2893 | #endif |
2894 | } |
2895 | } |
2896 | |
2897 | static void __meminit zone_init_free_lists(struct zone *zone) |
2898 | { |
2899 | int order, t; |
2900 | for_each_migratetype_order(order, t) { |
2901 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); |
2902 | zone->free_area[order].nr_free = 0; |
2903 | } |
2904 | } |
2905 | |
2906 | #ifndef __HAVE_ARCH_MEMMAP_INIT |
2907 | #define memmap_init(size, nid, zone, start_pfn) \ |
2908 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) |
2909 | #endif |
2910 | |
2911 | static int zone_batchsize(struct zone *zone) |
2912 | { |
2913 | #ifdef CONFIG_MMU |
2914 | int batch; |
2915 | |
2916 | /* |
2917 | * The per-cpu-pages pools are set to around 1000th of the |
2918 | * size of the zone. But no more than 1/2 of a meg. |
2919 | * |
2920 | * OK, so we don't know how big the cache is. So guess. |
2921 | */ |
2922 | batch = zone->present_pages / 1024; |
2923 | if (batch * PAGE_SIZE > 512 * 1024) |
2924 | batch = (512 * 1024) / PAGE_SIZE; |
2925 | batch /= 4; /* We effectively *= 4 below */ |
2926 | if (batch < 1) |
2927 | batch = 1; |
2928 | |
2929 | /* |
2930 | * Clamp the batch to a 2^n - 1 value. Having a power |
2931 | * of 2 value was found to be more likely to have |
2932 | * suboptimal cache aliasing properties in some cases. |
2933 | * |
2934 | * For example if 2 tasks are alternately allocating |
2935 | * batches of pages, one task can end up with a lot |
2936 | * of pages of one half of the possible page colors |
2937 | * and the other with pages of the other colors. |
2938 | */ |
2939 | batch = rounddown_pow_of_two(batch + batch/2) - 1; |
2940 | |
2941 | return batch; |
2942 | |
2943 | #else |
2944 | /* The deferral and batching of frees should be suppressed under NOMMU |
2945 | * conditions. |
2946 | * |
2947 | * The problem is that NOMMU needs to be able to allocate large chunks |
2948 | * of contiguous memory as there's no hardware page translation to |
2949 | * assemble apparent contiguous memory from discontiguous pages. |
2950 | * |
2951 | * Queueing large contiguous runs of pages for batching, however, |
2952 | * causes the pages to actually be freed in smaller chunks. As there |
2953 | * can be a significant delay between the individual batches being |
2954 | * recycled, this leads to the once large chunks of space being |
2955 | * fragmented and becoming unavailable for high-order allocations. |
2956 | */ |
2957 | return 0; |
2958 | #endif |
2959 | } |
2960 | |
2961 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) |
2962 | { |
2963 | struct per_cpu_pages *pcp; |
2964 | |
2965 | memset(p, 0, sizeof(*p)); |
2966 | |
2967 | pcp = &p->pcp; |
2968 | pcp->count = 0; |
2969 | pcp->high = 6 * batch; |
2970 | pcp->batch = max(1UL, 1 * batch); |
2971 | INIT_LIST_HEAD(&pcp->list); |
2972 | } |
2973 | |
2974 | /* |
2975 | * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist |
2976 | * to the value high for the pageset p. |
2977 | */ |
2978 | |
2979 | static void setup_pagelist_highmark(struct per_cpu_pageset *p, |
2980 | unsigned long high) |
2981 | { |
2982 | struct per_cpu_pages *pcp; |
2983 | |
2984 | pcp = &p->pcp; |
2985 | pcp->high = high; |
2986 | pcp->batch = max(1UL, high/4); |
2987 | if ((high/4) > (PAGE_SHIFT * 8)) |
2988 | pcp->batch = PAGE_SHIFT * 8; |
2989 | } |
2990 | |
2991 | |
2992 | #ifdef CONFIG_NUMA |
2993 | /* |
2994 | * Boot pageset table. One per cpu which is going to be used for all |
2995 | * zones and all nodes. The parameters will be set in such a way |
2996 | * that an item put on a list will immediately be handed over to |
2997 | * the buddy list. This is safe since pageset manipulation is done |
2998 | * with interrupts disabled. |
2999 | * |
3000 | * Some NUMA counter updates may also be caught by the boot pagesets. |
3001 | * |
3002 | * The boot_pagesets must be kept even after bootup is complete for |
3003 | * unused processors and/or zones. They do play a role for bootstrapping |
3004 | * hotplugged processors. |
3005 | * |
3006 | * zoneinfo_show() and maybe other functions do |
3007 | * not check if the processor is online before following the pageset pointer. |
3008 | * Other parts of the kernel may not check if the zone is available. |
3009 | */ |
3010 | static struct per_cpu_pageset boot_pageset[NR_CPUS]; |
3011 | |
3012 | /* |
3013 | * Dynamically allocate memory for the |
3014 | * per cpu pageset array in struct zone. |
3015 | */ |
3016 | static int __cpuinit process_zones(int cpu) |
3017 | { |
3018 | struct zone *zone, *dzone; |
3019 | int node = cpu_to_node(cpu); |
3020 | |
3021 | node_set_state(node, N_CPU); /* this node has a cpu */ |
3022 | |
3023 | for_each_populated_zone(zone) { |
3024 | zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), |
3025 | GFP_KERNEL, node); |
3026 | if (!zone_pcp(zone, cpu)) |
3027 | goto bad; |
3028 | |
3029 | setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); |
3030 | |
3031 | if (percpu_pagelist_fraction) |
3032 | setup_pagelist_highmark(zone_pcp(zone, cpu), |
3033 | (zone->present_pages / percpu_pagelist_fraction)); |
3034 | } |
3035 | |
3036 | return 0; |
3037 | bad: |
3038 | for_each_zone(dzone) { |
3039 | if (!populated_zone(dzone)) |
3040 | continue; |
3041 | if (dzone == zone) |
3042 | break; |
3043 | kfree(zone_pcp(dzone, cpu)); |
3044 | zone_pcp(dzone, cpu) = &boot_pageset[cpu]; |
3045 | } |
3046 | return -ENOMEM; |
3047 | } |
3048 | |
3049 | static inline void free_zone_pagesets(int cpu) |
3050 | { |
3051 | struct zone *zone; |
3052 | |
3053 | for_each_zone(zone) { |
3054 | struct per_cpu_pageset *pset = zone_pcp(zone, cpu); |
3055 | |
3056 | /* Free per_cpu_pageset if it is slab allocated */ |
3057 | if (pset != &boot_pageset[cpu]) |
3058 | kfree(pset); |
3059 | zone_pcp(zone, cpu) = &boot_pageset[cpu]; |
3060 | } |
3061 | } |
3062 | |
3063 | static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, |
3064 | unsigned long action, |
3065 | void *hcpu) |
3066 | { |
3067 | int cpu = (long)hcpu; |
3068 | int ret = NOTIFY_OK; |
3069 | |
3070 | switch (action) { |
3071 | case CPU_UP_PREPARE: |
3072 | case CPU_UP_PREPARE_FROZEN: |
3073 | if (process_zones(cpu)) |
3074 | ret = NOTIFY_BAD; |
3075 | break; |
3076 | case CPU_UP_CANCELED: |
3077 | case CPU_UP_CANCELED_FROZEN: |
3078 | case CPU_DEAD: |
3079 | case CPU_DEAD_FROZEN: |
3080 | free_zone_pagesets(cpu); |
3081 | break; |
3082 | default: |
3083 | break; |
3084 | } |
3085 | return ret; |
3086 | } |
3087 | |
3088 | static struct notifier_block __cpuinitdata pageset_notifier = |
3089 | { &pageset_cpuup_callback, NULL, 0 }; |
3090 | |
3091 | void __init setup_per_cpu_pageset(void) |
3092 | { |
3093 | int err; |
3094 | |
3095 | /* Initialize per_cpu_pageset for cpu 0. |
3096 | * A cpuup callback will do this for every cpu |
3097 | * as it comes online |
3098 | */ |
3099 | err = process_zones(smp_processor_id()); |
3100 | BUG_ON(err); |
3101 | register_cpu_notifier(&pageset_notifier); |
3102 | } |
3103 | |
3104 | #endif |
3105 | |
3106 | static noinline __init_refok |
3107 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) |
3108 | { |
3109 | int i; |
3110 | struct pglist_data *pgdat = zone->zone_pgdat; |
3111 | size_t alloc_size; |
3112 | |
3113 | /* |
3114 | * The per-page waitqueue mechanism uses hashed waitqueues |
3115 | * per zone. |
3116 | */ |
3117 | zone->wait_table_hash_nr_entries = |
3118 | wait_table_hash_nr_entries(zone_size_pages); |
3119 | zone->wait_table_bits = |
3120 | wait_table_bits(zone->wait_table_hash_nr_entries); |
3121 | alloc_size = zone->wait_table_hash_nr_entries |
3122 | * sizeof(wait_queue_head_t); |
3123 | |
3124 | if (!slab_is_available()) { |
3125 | zone->wait_table = (wait_queue_head_t *) |
3126 | alloc_bootmem_node(pgdat, alloc_size); |
3127 | } else { |
3128 | /* |
3129 | * This case means that a zone whose size was 0 gets new memory |
3130 | * via memory hot-add. |
3131 | * But it may be the case that a new node was hot-added. In |
3132 | * this case vmalloc() will not be able to use this new node's |
3133 | * memory - this wait_table must be initialized to use this new |
3134 | * node itself as well. |
3135 | * To use this new node's memory, further consideration will be |
3136 | * necessary. |
3137 | */ |
3138 | zone->wait_table = vmalloc(alloc_size); |
3139 | } |
3140 | if (!zone->wait_table) |
3141 | return -ENOMEM; |
3142 | |
3143 | for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) |
3144 | init_waitqueue_head(zone->wait_table + i); |
3145 | |
3146 | return 0; |
3147 | } |
3148 | |
3149 | static __meminit void zone_pcp_init(struct zone *zone) |
3150 | { |
3151 | int cpu; |
3152 | unsigned long batch = zone_batchsize(zone); |
3153 | |
3154 | for (cpu = 0; cpu < NR_CPUS; cpu++) { |
3155 | #ifdef CONFIG_NUMA |
3156 | /* Early boot. Slab allocator not functional yet */ |
3157 | zone_pcp(zone, cpu) = &boot_pageset[cpu]; |
3158 | setup_pageset(&boot_pageset[cpu],0); |
3159 | #else |
3160 | setup_pageset(zone_pcp(zone,cpu), batch); |
3161 | #endif |
3162 | } |
3163 | if (zone->present_pages) |
3164 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", |
3165 | zone->name, zone->present_pages, batch); |
3166 | } |
3167 | |
3168 | __meminit int init_currently_empty_zone(struct zone *zone, |
3169 | unsigned long zone_start_pfn, |
3170 | unsigned long size, |
3171 | enum memmap_context context) |
3172 | { |
3173 | struct pglist_data *pgdat = zone->zone_pgdat; |
3174 | int ret; |
3175 | ret = zone_wait_table_init(zone, size); |
3176 | if (ret) |
3177 | return ret; |
3178 | pgdat->nr_zones = zone_idx(zone) + 1; |
3179 | |
3180 | zone->zone_start_pfn = zone_start_pfn; |
3181 | |
3182 | mminit_dprintk(MMINIT_TRACE, "memmap_init", |
3183 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", |
3184 | pgdat->node_id, |
3185 | (unsigned long)zone_idx(zone), |
3186 | zone_start_pfn, (zone_start_pfn + size)); |
3187 | |
3188 | zone_init_free_lists(zone); |
3189 | |
3190 | return 0; |
3191 | } |
3192 | |
3193 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP |
3194 | /* |
3195 | * Basic iterator support. Return the first range of PFNs for a node |
3196 | * Note: nid == MAX_NUMNODES returns first region regardless of node |
3197 | */ |
3198 | static int __meminit first_active_region_index_in_nid(int nid) |
3199 | { |
3200 | int i; |
3201 | |
3202 | for (i = 0; i < nr_nodemap_entries; i++) |
3203 | if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) |
3204 | return i; |
3205 | |
3206 | return -1; |
3207 | } |
3208 | |
3209 | /* |
3210 | * Basic iterator support. Return the next active range of PFNs for a node |
3211 | * Note: nid == MAX_NUMNODES returns next region regardless of node |
3212 | */ |
3213 | static int __meminit next_active_region_index_in_nid(int index, int nid) |
3214 | { |
3215 | for (index = index + 1; index < nr_nodemap_entries; index++) |
3216 | if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) |
3217 | return index; |
3218 | |
3219 | return -1; |
3220 | } |
3221 | |
3222 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID |
3223 | /* |
3224 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. |
3225 | * Architectures may implement their own version but if add_active_range() |
3226 | * was used and there are no special requirements, this is a convenient |
3227 | * alternative |
3228 | */ |
3229 | int __meminit __early_pfn_to_nid(unsigned long pfn) |
3230 | { |
3231 | int i; |
3232 | |
3233 | for (i = 0; i < nr_nodemap_entries; i++) { |
3234 | unsigned long start_pfn = early_node_map[i].start_pfn; |
3235 | unsigned long end_pfn = early_node_map[i].end_pfn; |
3236 | |
3237 | if (start_pfn <= pfn && pfn < end_pfn) |
3238 | return early_node_map[i].nid; |
3239 | } |
3240 | /* This is a memory hole */ |
3241 | return -1; |
3242 | } |
3243 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ |
3244 | |
3245 | int __meminit early_pfn_to_nid(unsigned long pfn) |
3246 | { |
3247 | int nid; |
3248 | |
3249 | nid = __early_pfn_to_nid(pfn); |
3250 | if (nid >= 0) |
3251 | return nid; |
3252 | /* just returns 0 */ |
3253 | return 0; |
3254 | } |
3255 | |
3256 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES |
3257 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) |
3258 | { |
3259 | int nid; |
3260 | |
3261 | nid = __early_pfn_to_nid(pfn); |
3262 | if (nid >= 0 && nid != node) |
3263 | return false; |
3264 | return true; |
3265 | } |
3266 | #endif |
3267 | |
3268 | /* Basic iterator support to walk early_node_map[] */ |
3269 | #define for_each_active_range_index_in_nid(i, nid) \ |
3270 | for (i = first_active_region_index_in_nid(nid); i != -1; \ |
3271 | i = next_active_region_index_in_nid(i, nid)) |
3272 | |
3273 | /** |
3274 | * free_bootmem_with_active_regions - Call free_bootmem_node for each active range |
3275 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. |
3276 | * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node |
3277 | * |
3278 | * If an architecture guarantees that all ranges registered with |
3279 | * add_active_ranges() contain no holes and may be freed, this |
3280 | * this function may be used instead of calling free_bootmem() manually. |
3281 | */ |
3282 | void __init free_bootmem_with_active_regions(int nid, |
3283 | unsigned long max_low_pfn) |
3284 | { |
3285 | int i; |
3286 | |
3287 | for_each_active_range_index_in_nid(i, nid) { |
3288 | unsigned long size_pages = 0; |
3289 | unsigned long end_pfn = early_node_map[i].end_pfn; |
3290 | |
3291 | if (early_node_map[i].start_pfn >= max_low_pfn) |
3292 | continue; |
3293 | |
3294 | if (end_pfn > max_low_pfn) |
3295 | end_pfn = max_low_pfn; |
3296 | |
3297 | size_pages = end_pfn - early_node_map[i].start_pfn; |
3298 | free_bootmem_node(NODE_DATA(early_node_map[i].nid), |
3299 | PFN_PHYS(early_node_map[i].start_pfn), |
3300 | size_pages << PAGE_SHIFT); |
3301 | } |
3302 | } |
3303 | |
3304 | void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) |
3305 | { |
3306 | int i; |
3307 | int ret; |
3308 | |
3309 | for_each_active_range_index_in_nid(i, nid) { |
3310 | ret = work_fn(early_node_map[i].start_pfn, |
3311 | early_node_map[i].end_pfn, data); |
3312 | if (ret) |
3313 | break; |
3314 | } |
3315 | } |
3316 | /** |
3317 | * sparse_memory_present_with_active_regions - Call memory_present for each active range |
3318 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. |
3319 | * |
3320 | * If an architecture guarantees that all ranges registered with |
3321 | * add_active_ranges() contain no holes and may be freed, this |
3322 | * function may be used instead of calling memory_present() manually. |
3323 | */ |
3324 | void __init sparse_memory_present_with_active_regions(int nid) |
3325 | { |
3326 | int i; |
3327 | |
3328 | for_each_active_range_index_in_nid(i, nid) |
3329 | memory_present(early_node_map[i].nid, |
3330 | early_node_map[i].start_pfn, |
3331 | early_node_map[i].end_pfn); |
3332 | } |
3333 | |
3334 | /** |
3335 | * get_pfn_range_for_nid - Return the start and end page frames for a node |
3336 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. |
3337 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. |
3338 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. |
3339 | * |
3340 | * It returns the start and end page frame of a node based on information |
3341 | * provided by an arch calling add_active_range(). If called for a node |
3342 | * with no available memory, a warning is printed and the start and end |
3343 | * PFNs will be 0. |
3344 | */ |
3345 | void __meminit get_pfn_range_for_nid(unsigned int nid, |
3346 | unsigned long *start_pfn, unsigned long *end_pfn) |
3347 | { |
3348 | int i; |
3349 | *start_pfn = -1UL; |
3350 | *end_pfn = 0; |
3351 | |
3352 | for_each_active_range_index_in_nid(i, nid) { |
3353 | *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); |
3354 | *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); |
3355 | } |
3356 | |
3357 | if (*start_pfn == -1UL) |
3358 | *start_pfn = 0; |
3359 | } |
3360 | |
3361 | /* |
3362 | * This finds a zone that can be used for ZONE_MOVABLE pages. The |
3363 | * assumption is made that zones within a node are ordered in monotonic |
3364 | * increasing memory addresses so that the "highest" populated zone is used |
3365 | */ |
3366 | static void __init find_usable_zone_for_movable(void) |
3367 | { |
3368 | int zone_index; |
3369 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { |
3370 | if (zone_index == ZONE_MOVABLE) |
3371 | continue; |
3372 | |
3373 | if (arch_zone_highest_possible_pfn[zone_index] > |
3374 | arch_zone_lowest_possible_pfn[zone_index]) |
3375 | break; |
3376 | } |
3377 | |
3378 | VM_BUG_ON(zone_index == -1); |
3379 | movable_zone = zone_index; |
3380 | } |
3381 | |
3382 | /* |
3383 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE |
3384 | * because it is sized independant of architecture. Unlike the other zones, |
3385 | * the starting point for ZONE_MOVABLE is not fixed. It may be different |
3386 | * in each node depending on the size of each node and how evenly kernelcore |
3387 | * is distributed. This helper function adjusts the zone ranges |
3388 | * provided by the architecture for a given node by using the end of the |
3389 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that |
3390 | * zones within a node are in order of monotonic increases memory addresses |
3391 | */ |
3392 | static void __meminit adjust_zone_range_for_zone_movable(int nid, |
3393 | unsigned long zone_type, |
3394 | unsigned long node_start_pfn, |
3395 | unsigned long node_end_pfn, |
3396 | unsigned long *zone_start_pfn, |
3397 | unsigned long *zone_end_pfn) |
3398 | { |
3399 | /* Only adjust if ZONE_MOVABLE is on this node */ |
3400 | if (zone_movable_pfn[nid]) { |
3401 | /* Size ZONE_MOVABLE */ |
3402 | if (zone_type == ZONE_MOVABLE) { |
3403 | *zone_start_pfn = zone_movable_pfn[nid]; |
3404 | *zone_end_pfn = min(node_end_pfn, |
3405 | arch_zone_highest_possible_pfn[movable_zone]); |
3406 | |
3407 | /* Adjust for ZONE_MOVABLE starting within this range */ |
3408 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && |
3409 | *zone_end_pfn > zone_movable_pfn[nid]) { |
3410 | *zone_end_pfn = zone_movable_pfn[nid]; |
3411 | |
3412 | /* Check if this whole range is within ZONE_MOVABLE */ |
3413 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) |
3414 | *zone_start_pfn = *zone_end_pfn; |
3415 | } |
3416 | } |
3417 | |
3418 | /* |
3419 | * Return the number of pages a zone spans in a node, including holes |
3420 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() |
3421 | */ |
3422 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, |
3423 | unsigned long zone_type, |
3424 | unsigned long *ignored) |
3425 | { |
3426 | unsigned long node_start_pfn, node_end_pfn; |
3427 | unsigned long zone_start_pfn, zone_end_pfn; |
3428 | |
3429 | /* Get the start and end of the node and zone */ |
3430 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); |
3431 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; |
3432 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; |
3433 | adjust_zone_range_for_zone_movable(nid, zone_type, |
3434 | node_start_pfn, node_end_pfn, |
3435 | &zone_start_pfn, &zone_end_pfn); |
3436 | |
3437 | /* Check that this node has pages within the zone's required range */ |
3438 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) |
3439 | return 0; |
3440 | |
3441 | /* Move the zone boundaries inside the node if necessary */ |
3442 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); |
3443 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); |
3444 | |
3445 | /* Return the spanned pages */ |
3446 | return zone_end_pfn - zone_start_pfn; |
3447 | } |
3448 | |
3449 | /* |
3450 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, |
3451 | * then all holes in the requested range will be accounted for. |
3452 | */ |
3453 | static unsigned long __meminit __absent_pages_in_range(int nid, |
3454 | unsigned long range_start_pfn, |
3455 | unsigned long range_end_pfn) |
3456 | { |
3457 | int i = 0; |
3458 | unsigned long prev_end_pfn = 0, hole_pages = 0; |
3459 | unsigned long start_pfn; |
3460 | |
3461 | /* Find the end_pfn of the first active range of pfns in the node */ |
3462 | i = first_active_region_index_in_nid(nid); |
3463 | if (i == -1) |
3464 | return 0; |
3465 | |
3466 | prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); |
3467 | |
3468 | /* Account for ranges before physical memory on this node */ |
3469 | if (early_node_map[i].start_pfn > range_start_pfn) |
3470 | hole_pages = prev_end_pfn - range_start_pfn; |
3471 | |
3472 | /* Find all holes for the zone within the node */ |
3473 | for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { |
3474 | |
3475 | /* No need to continue if prev_end_pfn is outside the zone */ |
3476 | if (prev_end_pfn >= range_end_pfn) |
3477 | break; |
3478 | |
3479 | /* Make sure the end of the zone is not within the hole */ |
3480 | start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); |
3481 | prev_end_pfn = max(prev_end_pfn, range_start_pfn); |
3482 | |
3483 | /* Update the hole size cound and move on */ |
3484 | if (start_pfn > range_start_pfn) { |
3485 | BUG_ON(prev_end_pfn > start_pfn); |
3486 | hole_pages += start_pfn - prev_end_pfn; |
3487 | } |
3488 | prev_end_pfn = early_node_map[i].end_pfn; |
3489 | } |
3490 | |
3491 | /* Account for ranges past physical memory on this node */ |
3492 | if (range_end_pfn > prev_end_pfn) |
3493 | hole_pages += range_end_pfn - |
3494 | max(range_start_pfn, prev_end_pfn); |
3495 | |
3496 | return hole_pages; |
3497 | } |
3498 | |
3499 | /** |
3500 | * absent_pages_in_range - Return number of page frames in holes within a range |
3501 | * @start_pfn: The start PFN to start searching for holes |
3502 | * @end_pfn: The end PFN to stop searching for holes |
3503 | * |
3504 | * It returns the number of pages frames in memory holes within a range. |
3505 | */ |
3506 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, |
3507 | unsigned long end_pfn) |
3508 | { |
3509 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); |
3510 | } |
3511 | |
3512 | /* Return the number of page frames in holes in a zone on a node */ |
3513 | static unsigned long __meminit zone_absent_pages_in_node(int nid, |
3514 | unsigned long zone_type, |
3515 | unsigned long *ignored) |
3516 | { |
3517 | unsigned long node_start_pfn, node_end_pfn; |
3518 | unsigned long zone_start_pfn, zone_end_pfn; |
3519 | |
3520 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); |
3521 | zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], |
3522 | node_start_pfn); |
3523 | zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], |
3524 | node_end_pfn); |
3525 | |
3526 | adjust_zone_range_for_zone_movable(nid, zone_type, |
3527 | node_start_pfn, node_end_pfn, |
3528 | &zone_start_pfn, &zone_end_pfn); |
3529 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); |
3530 | } |
3531 | |
3532 | #else |
3533 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, |
3534 | unsigned long zone_type, |
3535 | unsigned long *zones_size) |
3536 | { |
3537 | return zones_size[zone_type]; |
3538 | } |
3539 | |
3540 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, |
3541 | unsigned long zone_type, |
3542 | unsigned long *zholes_size) |
3543 | { |
3544 | if (!zholes_size) |
3545 | return 0; |
3546 | |
3547 | return zholes_size[zone_type]; |
3548 | } |
3549 | |
3550 | #endif |
3551 | |
3552 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, |
3553 | unsigned long *zones_size, unsigned long *zholes_size) |
3554 | { |
3555 | unsigned long realtotalpages, totalpages = 0; |
3556 | enum zone_type i; |
3557 | |
3558 | for (i = 0; i < MAX_NR_ZONES; i++) |
3559 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, |
3560 | zones_size); |
3561 | pgdat->node_spanned_pages = totalpages; |
3562 | |
3563 | realtotalpages = totalpages; |
3564 | for (i = 0; i < MAX_NR_ZONES; i++) |
3565 | realtotalpages -= |
3566 | zone_absent_pages_in_node(pgdat->node_id, i, |
3567 | zholes_size); |
3568 | pgdat->node_present_pages = realtotalpages; |
3569 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, |
3570 | realtotalpages); |
3571 | } |
3572 | |
3573 | #ifndef CONFIG_SPARSEMEM |
3574 | /* |
3575 | * Calculate the size of the zone->blockflags rounded to an unsigned long |
3576 | * Start by making sure zonesize is a multiple of pageblock_order by rounding |
3577 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally |
3578 | * round what is now in bits to nearest long in bits, then return it in |
3579 | * bytes. |
3580 | */ |
3581 | static unsigned long __init usemap_size(unsigned long zonesize) |
3582 | { |
3583 | unsigned long usemapsize; |
3584 | |
3585 | usemapsize = roundup(zonesize, pageblock_nr_pages); |
3586 | usemapsize = usemapsize >> pageblock_order; |
3587 | usemapsize *= NR_PAGEBLOCK_BITS; |
3588 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); |
3589 | |
3590 | return usemapsize / 8; |
3591 | } |
3592 | |
3593 | static void __init setup_usemap(struct pglist_data *pgdat, |
3594 | struct zone *zone, unsigned long zonesize) |
3595 | { |
3596 | unsigned long usemapsize = usemap_size(zonesize); |
3597 | zone->pageblock_flags = NULL; |
3598 | if (usemapsize) |
3599 | zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); |
3600 | } |
3601 | #else |
3602 | static void inline setup_usemap(struct pglist_data *pgdat, |
3603 | struct zone *zone, unsigned long zonesize) {} |
3604 | #endif /* CONFIG_SPARSEMEM */ |
3605 | |
3606 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
3607 | |
3608 | /* Return a sensible default order for the pageblock size. */ |
3609 | static inline int pageblock_default_order(void) |
3610 | { |
3611 | if (HPAGE_SHIFT > PAGE_SHIFT) |
3612 | return HUGETLB_PAGE_ORDER; |
3613 | |
3614 | return MAX_ORDER-1; |
3615 | } |
3616 | |
3617 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ |
3618 | static inline void __init set_pageblock_order(unsigned int order) |
3619 | { |
3620 | /* Check that pageblock_nr_pages has not already been setup */ |
3621 | if (pageblock_order) |
3622 | return; |
3623 | |
3624 | /* |
3625 | * Assume the largest contiguous order of interest is a huge page. |
3626 | * This value may be variable depending on boot parameters on IA64 |
3627 | */ |
3628 | pageblock_order = order; |
3629 | } |
3630 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ |
3631 | |
3632 | /* |
3633 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() |
3634 | * and pageblock_default_order() are unused as pageblock_order is set |
3635 | * at compile-time. See include/linux/pageblock-flags.h for the values of |
3636 | * pageblock_order based on the kernel config |
3637 | */ |
3638 | static inline int pageblock_default_order(unsigned int order) |
3639 | { |
3640 | return MAX_ORDER-1; |
3641 | } |
3642 | #define set_pageblock_order(x) do {} while (0) |
3643 | |
3644 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ |
3645 | |
3646 | /* |
3647 | * Set up the zone data structures: |
3648 | * - mark all pages reserved |
3649 | * - mark all memory queues empty |
3650 | * - clear the memory bitmaps |
3651 | */ |
3652 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, |
3653 | unsigned long *zones_size, unsigned long *zholes_size) |
3654 | { |
3655 | enum zone_type j; |
3656 | int nid = pgdat->node_id; |
3657 | unsigned long zone_start_pfn = pgdat->node_start_pfn; |
3658 | int ret; |
3659 | |
3660 | pgdat_resize_init(pgdat); |
3661 | pgdat->nr_zones = 0; |
3662 | init_waitqueue_head(&pgdat->kswapd_wait); |
3663 | pgdat->kswapd_max_order = 0; |
3664 | pgdat_page_cgroup_init(pgdat); |
3665 | |
3666 | for (j = 0; j < MAX_NR_ZONES; j++) { |
3667 | struct zone *zone = pgdat->node_zones + j; |
3668 | unsigned long size, realsize, memmap_pages; |
3669 | enum lru_list l; |
3670 | |
3671 | size = zone_spanned_pages_in_node(nid, j, zones_size); |
3672 | realsize = size - zone_absent_pages_in_node(nid, j, |
3673 | zholes_size); |
3674 | |
3675 | /* |
3676 | * Adjust realsize so that it accounts for how much memory |
3677 | * is used by this zone for memmap. This affects the watermark |
3678 | * and per-cpu initialisations |
3679 | */ |
3680 | memmap_pages = |
3681 | PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; |
3682 | if (realsize >= memmap_pages) { |
3683 | realsize -= memmap_pages; |
3684 | if (memmap_pages) |
3685 | printk(KERN_DEBUG |
3686 | " %s zone: %lu pages used for memmap\n", |
3687 | zone_names[j], memmap_pages); |
3688 | } else |
3689 | printk(KERN_WARNING |
3690 | " %s zone: %lu pages exceeds realsize %lu\n", |
3691 | zone_names[j], memmap_pages, realsize); |
3692 | |
3693 | /* Account for reserved pages */ |
3694 | if (j == 0 && realsize > dma_reserve) { |
3695 | realsize -= dma_reserve; |
3696 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", |
3697 | zone_names[0], dma_reserve); |
3698 | } |
3699 | |
3700 | if (!is_highmem_idx(j)) |
3701 | nr_kernel_pages += realsize; |
3702 | nr_all_pages += realsize; |
3703 | |
3704 | zone->spanned_pages = size; |
3705 | zone->present_pages = realsize; |
3706 | #ifdef CONFIG_NUMA |
3707 | zone->node = nid; |
3708 | zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) |
3709 | / 100; |
3710 | zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; |
3711 | #endif |
3712 | zone->name = zone_names[j]; |
3713 | spin_lock_init(&zone->lock); |
3714 | spin_lock_init(&zone->lru_lock); |
3715 | zone_seqlock_init(zone); |
3716 | zone->zone_pgdat = pgdat; |
3717 | |
3718 | zone->prev_priority = DEF_PRIORITY; |
3719 | |
3720 | zone_pcp_init(zone); |
3721 | for_each_lru(l) { |
3722 | INIT_LIST_HEAD(&zone->lru[l].list); |
3723 | zone->lru[l].nr_saved_scan = 0; |
3724 | } |
3725 | zone->reclaim_stat.recent_rotated[0] = 0; |
3726 | zone->reclaim_stat.recent_rotated[1] = 0; |
3727 | zone->reclaim_stat.recent_scanned[0] = 0; |
3728 | zone->reclaim_stat.recent_scanned[1] = 0; |
3729 | zap_zone_vm_stats(zone); |
3730 | zone->flags = 0; |
3731 | if (!size) |
3732 | continue; |
3733 | |
3734 | set_pageblock_order(pageblock_default_order()); |
3735 | setup_usemap(pgdat, zone, size); |
3736 | ret = init_currently_empty_zone(zone, zone_start_pfn, |
3737 | size, MEMMAP_EARLY); |
3738 | BUG_ON(ret); |
3739 | memmap_init(size, nid, j, zone_start_pfn); |
3740 | zone_start_pfn += size; |
3741 | } |
3742 | } |
3743 | |
3744 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) |
3745 | { |
3746 | /* Skip empty nodes */ |
3747 | if (!pgdat->node_spanned_pages) |
3748 | return; |
3749 | |
3750 | #ifdef CONFIG_FLAT_NODE_MEM_MAP |
3751 | /* ia64 gets its own node_mem_map, before this, without bootmem */ |
3752 | if (!pgdat->node_mem_map) { |
3753 | unsigned long size, start, end; |
3754 | struct page *map; |
3755 | |
3756 | /* |
3757 | * The zone's endpoints aren't required to be MAX_ORDER |
3758 | * aligned but the node_mem_map endpoints must be in order |
3759 | * for the buddy allocator to function correctly. |
3760 | */ |
3761 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); |
3762 | end = pgdat->node_start_pfn + pgdat->node_spanned_pages; |
3763 | end = ALIGN(end, MAX_ORDER_NR_PAGES); |
3764 | size = (end - start) * sizeof(struct page); |
3765 | map = alloc_remap(pgdat->node_id, size); |
3766 | if (!map) |
3767 | map = alloc_bootmem_node(pgdat, size); |
3768 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); |
3769 | } |
3770 | #ifndef CONFIG_NEED_MULTIPLE_NODES |
3771 | /* |
3772 | * With no DISCONTIG, the global mem_map is just set as node 0's |
3773 | */ |
3774 | if (pgdat == NODE_DATA(0)) { |
3775 | mem_map = NODE_DATA(0)->node_mem_map; |
3776 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP |
3777 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) |
3778 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); |
3779 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ |
3780 | } |
3781 | #endif |
3782 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ |
3783 | } |
3784 | |
3785 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, |
3786 | unsigned long node_start_pfn, unsigned long *zholes_size) |
3787 | { |
3788 | pg_data_t *pgdat = NODE_DATA(nid); |
3789 | |
3790 | pgdat->node_id = nid; |
3791 | pgdat->node_start_pfn = node_start_pfn; |
3792 | calculate_node_totalpages(pgdat, zones_size, zholes_size); |
3793 | |
3794 | alloc_node_mem_map(pgdat); |
3795 | #ifdef CONFIG_FLAT_NODE_MEM_MAP |
3796 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", |
3797 | nid, (unsigned long)pgdat, |
3798 | (unsigned long)pgdat->node_mem_map); |
3799 | #endif |
3800 | |
3801 | free_area_init_core(pgdat, zones_size, zholes_size); |
3802 | } |
3803 | |
3804 | #ifdef CONFIG_ARCH_POPULATES_NODE_MAP |
3805 | |
3806 | #if MAX_NUMNODES > 1 |
3807 | /* |
3808 | * Figure out the number of possible node ids. |
3809 | */ |
3810 | static void __init setup_nr_node_ids(void) |
3811 | { |
3812 | unsigned int node; |
3813 | unsigned int highest = 0; |
3814 | |
3815 | for_each_node_mask(node, node_possible_map) |
3816 | highest = node; |
3817 | nr_node_ids = highest + 1; |
3818 | } |
3819 | #else |
3820 | static inline void setup_nr_node_ids(void) |
3821 | { |
3822 | } |
3823 | #endif |
3824 | |
3825 | /** |
3826 | * add_active_range - Register a range of PFNs backed by physical memory |
3827 | * @nid: The node ID the range resides on |
3828 | * @start_pfn: The start PFN of the available physical memory |
3829 | * @end_pfn: The end PFN of the available physical memory |
3830 | * |
3831 | * These ranges are stored in an early_node_map[] and later used by |
3832 | * free_area_init_nodes() to calculate zone sizes and holes. If the |
3833 | * range spans a memory hole, it is up to the architecture to ensure |
3834 | * the memory is not freed by the bootmem allocator. If possible |
3835 | * the range being registered will be merged with existing ranges. |
3836 | */ |
3837 | void __init add_active_range(unsigned int nid, unsigned long start_pfn, |
3838 | unsigned long end_pfn) |
3839 | { |
3840 | int i; |
3841 | |
3842 | mminit_dprintk(MMINIT_TRACE, "memory_register", |
3843 | "Entering add_active_range(%d, %#lx, %#lx) " |
3844 | "%d entries of %d used\n", |
3845 | nid, start_pfn, end_pfn, |
3846 | nr_nodemap_entries, MAX_ACTIVE_REGIONS); |
3847 | |
3848 | mminit_validate_memmodel_limits(&start_pfn, &end_pfn); |
3849 | |
3850 | /* Merge with existing active regions if possible */ |
3851 | for (i = 0; i < nr_nodemap_entries; i++) { |
3852 | if (early_node_map[i].nid != nid) |
3853 | continue; |
3854 | |
3855 | /* Skip if an existing region covers this new one */ |
3856 | if (start_pfn >= early_node_map[i].start_pfn && |
3857 | end_pfn <= early_node_map[i].end_pfn) |
3858 | return; |
3859 | |
3860 | /* Merge forward if suitable */ |
3861 | if (start_pfn <= early_node_map[i].end_pfn && |
3862 | end_pfn > early_node_map[i].end_pfn) { |
3863 | early_node_map[i].end_pfn = end_pfn; |
3864 | return; |
3865 | } |
3866 | |
3867 | /* Merge backward if suitable */ |
3868 | if (start_pfn < early_node_map[i].end_pfn && |
3869 | end_pfn >= early_node_map[i].start_pfn) { |
3870 | early_node_map[i].start_pfn = start_pfn; |
3871 | return; |
3872 | } |
3873 | } |
3874 | |
3875 | /* Check that early_node_map is large enough */ |
3876 | if (i >= MAX_ACTIVE_REGIONS) { |
3877 | printk(KERN_CRIT "More than %d memory regions, truncating\n", |
3878 | MAX_ACTIVE_REGIONS); |
3879 | return; |
3880 | } |
3881 | |
3882 | early_node_map[i].nid = nid; |
3883 | early_node_map[i].start_pfn = start_pfn; |
3884 | early_node_map[i].end_pfn = end_pfn; |
3885 | nr_nodemap_entries = i + 1; |
3886 | } |
3887 | |
3888 | /** |
3889 | * remove_active_range - Shrink an existing registered range of PFNs |
3890 | * @nid: The node id the range is on that should be shrunk |
3891 | * @start_pfn: The new PFN of the range |
3892 | * @end_pfn: The new PFN of the range |
3893 | * |
3894 | * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. |
3895 | * The map is kept near the end physical page range that has already been |
3896 | * registered. This function allows an arch to shrink an existing registered |
3897 | * range. |
3898 | */ |
3899 | void __init remove_active_range(unsigned int nid, unsigned long start_pfn, |
3900 | unsigned long end_pfn) |
3901 | { |
3902 | int i, j; |
3903 | int removed = 0; |
3904 | |
3905 | printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", |
3906 | nid, start_pfn, end_pfn); |
3907 | |
3908 | /* Find the old active region end and shrink */ |
3909 | for_each_active_range_index_in_nid(i, nid) { |
3910 | if (early_node_map[i].start_pfn >= start_pfn && |
3911 | early_node_map[i].end_pfn <= end_pfn) { |
3912 | /* clear it */ |
3913 | early_node_map[i].start_pfn = 0; |
3914 | early_node_map[i].end_pfn = 0; |
3915 | removed = 1; |
3916 | continue; |
3917 | } |
3918 | if (early_node_map[i].start_pfn < start_pfn && |
3919 | early_node_map[i].end_pfn > start_pfn) { |
3920 | unsigned long temp_end_pfn = early_node_map[i].end_pfn; |
3921 | early_node_map[i].end_pfn = start_pfn; |
3922 | if (temp_end_pfn > end_pfn) |
3923 | add_active_range(nid, end_pfn, temp_end_pfn); |
3924 | continue; |
3925 | } |
3926 | if (early_node_map[i].start_pfn >= start_pfn && |
3927 | early_node_map[i].end_pfn > end_pfn && |
3928 | early_node_map[i].start_pfn < end_pfn) { |
3929 | early_node_map[i].start_pfn = end_pfn; |
3930 | continue; |
3931 | } |
3932 | } |
3933 | |
3934 | if (!removed) |
3935 | return; |
3936 | |
3937 | /* remove the blank ones */ |
3938 | for (i = nr_nodemap_entries - 1; i > 0; i--) { |
3939 | if (early_node_map[i].nid != nid) |
3940 | continue; |
3941 | if (early_node_map[i].end_pfn) |
3942 | continue; |
3943 | /* we found it, get rid of it */ |
3944 | for (j = i; j < nr_nodemap_entries - 1; j++) |
3945 | memcpy(&early_node_map[j], &early_node_map[j+1], |
3946 | sizeof(early_node_map[j])); |
3947 | j = nr_nodemap_entries - 1; |
3948 | memset(&early_node_map[j], 0, sizeof(early_node_map[j])); |
3949 | nr_nodemap_entries--; |
3950 | } |
3951 | } |
3952 | |
3953 | /** |
3954 | * remove_all_active_ranges - Remove all currently registered regions |
3955 | * |
3956 | * During discovery, it may be found that a table like SRAT is invalid |
3957 | * and an alternative discovery method must be used. This function removes |
3958 | * all currently registered regions. |
3959 | */ |
3960 | void __init remove_all_active_ranges(void) |
3961 | { |
3962 | memset(early_node_map, 0, sizeof(early_node_map)); |
3963 | nr_nodemap_entries = 0; |
3964 | } |
3965 | |
3966 | /* Compare two active node_active_regions */ |
3967 | static int __init cmp_node_active_region(const void *a, const void *b) |
3968 | { |
3969 | struct node_active_region *arange = (struct node_active_region *)a; |
3970 | struct node_active_region *brange = (struct node_active_region *)b; |
3971 | |
3972 | /* Done this way to avoid overflows */ |
3973 | if (arange->start_pfn > brange->start_pfn) |
3974 | return 1; |
3975 | if (arange->start_pfn < brange->start_pfn) |
3976 | return -1; |
3977 | |
3978 | return 0; |
3979 | } |
3980 | |
3981 | /* sort the node_map by start_pfn */ |
3982 | static void __init sort_node_map(void) |
3983 | { |
3984 | sort(early_node_map, (size_t)nr_nodemap_entries, |
3985 | sizeof(struct node_active_region), |
3986 | cmp_node_active_region, NULL); |
3987 | } |
3988 | |
3989 | /* Find the lowest pfn for a node */ |
3990 | static unsigned long __init find_min_pfn_for_node(int nid) |
3991 | { |
3992 | int i; |
3993 | unsigned long min_pfn = ULONG_MAX; |
3994 | |
3995 | /* Assuming a sorted map, the first range found has the starting pfn */ |
3996 | for_each_active_range_index_in_nid(i, nid) |
3997 | min_pfn = min(min_pfn, early_node_map[i].start_pfn); |
3998 | |
3999 | if (min_pfn == ULONG_MAX) { |
4000 | printk(KERN_WARNING |
4001 | "Could not find start_pfn for node %d\n", nid); |
4002 | return 0; |
4003 | } |
4004 | |
4005 | return min_pfn; |
4006 | } |
4007 | |
4008 | /** |
4009 | * find_min_pfn_with_active_regions - Find the minimum PFN registered |
4010 | * |
4011 | * It returns the minimum PFN based on information provided via |
4012 | * add_active_range(). |
4013 | */ |
4014 | unsigned long __init find_min_pfn_with_active_regions(void) |
4015 | { |
4016 | return find_min_pfn_for_node(MAX_NUMNODES); |
4017 | } |
4018 | |
4019 | /* |
4020 | * early_calculate_totalpages() |
4021 | * Sum pages in active regions for movable zone. |
4022 | * Populate N_HIGH_MEMORY for calculating usable_nodes. |
4023 | */ |
4024 | static unsigned long __init early_calculate_totalpages(void) |
4025 | { |
4026 | int i; |
4027 | unsigned long totalpages = 0; |
4028 | |
4029 | for (i = 0; i < nr_nodemap_entries; i++) { |
4030 | unsigned long pages = early_node_map[i].end_pfn - |
4031 | early_node_map[i].start_pfn; |
4032 | totalpages += pages; |
4033 | if (pages) |
4034 | node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); |
4035 | } |
4036 | return totalpages; |
4037 | } |
4038 | |
4039 | /* |
4040 | * Find the PFN the Movable zone begins in each node. Kernel memory |
4041 | * is spread evenly between nodes as long as the nodes have enough |
4042 | * memory. When they don't, some nodes will have more kernelcore than |
4043 | * others |
4044 | */ |
4045 | static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) |
4046 | { |
4047 | int i, nid; |
4048 | unsigned long usable_startpfn; |
4049 | unsigned long kernelcore_node, kernelcore_remaining; |
4050 | /* save the state before borrow the nodemask */ |
4051 | nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; |
4052 | unsigned long totalpages = early_calculate_totalpages(); |
4053 | int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); |
4054 | |
4055 | /* |
4056 | * If movablecore was specified, calculate what size of |
4057 | * kernelcore that corresponds so that memory usable for |
4058 | * any allocation type is evenly spread. If both kernelcore |
4059 | * and movablecore are specified, then the value of kernelcore |
4060 | * will be used for required_kernelcore if it's greater than |
4061 | * what movablecore would have allowed. |
4062 | */ |
4063 | if (required_movablecore) { |
4064 | unsigned long corepages; |
4065 | |
4066 | /* |
4067 | * Round-up so that ZONE_MOVABLE is at least as large as what |
4068 | * was requested by the user |
4069 | */ |
4070 | required_movablecore = |
4071 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); |
4072 | corepages = totalpages - required_movablecore; |
4073 | |
4074 | required_kernelcore = max(required_kernelcore, corepages); |
4075 | } |
4076 | |
4077 | /* If kernelcore was not specified, there is no ZONE_MOVABLE */ |
4078 | if (!required_kernelcore) |
4079 | goto out; |
4080 | |
4081 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ |
4082 | find_usable_zone_for_movable(); |
4083 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; |
4084 | |
4085 | restart: |
4086 | /* Spread kernelcore memory as evenly as possible throughout nodes */ |
4087 | kernelcore_node = required_kernelcore / usable_nodes; |
4088 | for_each_node_state(nid, N_HIGH_MEMORY) { |
4089 | /* |
4090 | * Recalculate kernelcore_node if the division per node |
4091 | * now exceeds what is necessary to satisfy the requested |
4092 | * amount of memory for the kernel |
4093 | */ |
4094 | if (required_kernelcore < kernelcore_node) |
4095 | kernelcore_node = required_kernelcore / usable_nodes; |
4096 | |
4097 | /* |
4098 | * As the map is walked, we track how much memory is usable |
4099 | * by the kernel using kernelcore_remaining. When it is |
4100 | * 0, the rest of the node is usable by ZONE_MOVABLE |
4101 | */ |
4102 | kernelcore_remaining = kernelcore_node; |
4103 | |
4104 | /* Go through each range of PFNs within this node */ |
4105 | for_each_active_range_index_in_nid(i, nid) { |
4106 | unsigned long start_pfn, end_pfn; |
4107 | unsigned long size_pages; |
4108 | |
4109 | start_pfn = max(early_node_map[i].start_pfn, |
4110 | zone_movable_pfn[nid]); |
4111 | end_pfn = early_node_map[i].end_pfn; |
4112 | if (start_pfn >= end_pfn) |
4113 | continue; |
4114 | |
4115 | /* Account for what is only usable for kernelcore */ |
4116 | if (start_pfn < usable_startpfn) { |
4117 | unsigned long kernel_pages; |
4118 | kernel_pages = min(end_pfn, usable_startpfn) |
4119 | - start_pfn; |
4120 | |
4121 | kernelcore_remaining -= min(kernel_pages, |
4122 | kernelcore_remaining); |
4123 | required_kernelcore -= min(kernel_pages, |
4124 | required_kernelcore); |
4125 | |
4126 | /* Continue if range is now fully accounted */ |
4127 | if (end_pfn <= usable_startpfn) { |
4128 | |
4129 | /* |
4130 | * Push zone_movable_pfn to the end so |
4131 | * that if we have to rebalance |
4132 | * kernelcore across nodes, we will |
4133 | * not double account here |
4134 | */ |
4135 | zone_movable_pfn[nid] = end_pfn; |
4136 | continue; |
4137 | } |
4138 | start_pfn = usable_startpfn; |
4139 | } |
4140 | |
4141 | /* |
4142 | * The usable PFN range for ZONE_MOVABLE is from |
4143 | * start_pfn->end_pfn. Calculate size_pages as the |
4144 | * number of pages used as kernelcore |
4145 | */ |
4146 | size_pages = end_pfn - start_pfn; |
4147 | if (size_pages > kernelcore_remaining) |
4148 | size_pages = kernelcore_remaining; |
4149 | zone_movable_pfn[nid] = start_pfn + size_pages; |
4150 | |
4151 | /* |
4152 | * Some kernelcore has been met, update counts and |
4153 | * break if the kernelcore for this node has been |
4154 | * satisified |
4155 | */ |
4156 | required_kernelcore -= min(required_kernelcore, |
4157 | size_pages); |
4158 | kernelcore_remaining -= size_pages; |
4159 | if (!kernelcore_remaining) |
4160 | break; |
4161 | } |
4162 | } |
4163 | |
4164 | /* |
4165 | * If there is still required_kernelcore, we do another pass with one |
4166 | * less node in the count. This will push zone_movable_pfn[nid] further |
4167 | * along on the nodes that still have memory until kernelcore is |
4168 | * satisified |
4169 | */ |
4170 | usable_nodes--; |
4171 | if (usable_nodes && required_kernelcore > usable_nodes) |
4172 | goto restart; |
4173 | |
4174 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ |
4175 | for (nid = 0; nid < MAX_NUMNODES; nid++) |
4176 | zone_movable_pfn[nid] = |
4177 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); |
4178 | |
4179 | out: |
4180 | /* restore the node_state */ |
4181 | node_states[N_HIGH_MEMORY] = saved_node_state; |
4182 | } |
4183 | |
4184 | /* Any regular memory on that node ? */ |
4185 | static void check_for_regular_memory(pg_data_t *pgdat) |
4186 | { |
4187 | #ifdef CONFIG_HIGHMEM |
4188 | enum zone_type zone_type; |
4189 | |
4190 | for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { |
4191 | struct zone *zone = &pgdat->node_zones[zone_type]; |
4192 | if (zone->present_pages) |
4193 | node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); |
4194 | } |
4195 | #endif |
4196 | } |
4197 | |
4198 | /** |
4199 | * free_area_init_nodes - Initialise all pg_data_t and zone data |
4200 | * @max_zone_pfn: an array of max PFNs for each zone |
4201 | * |
4202 | * This will call free_area_init_node() for each active node in the system. |
4203 | * Using the page ranges provided by add_active_range(), the size of each |
4204 | * zone in each node and their holes is calculated. If the maximum PFN |
4205 | * between two adjacent zones match, it is assumed that the zone is empty. |
4206 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed |
4207 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone |
4208 | * starts where the previous one ended. For example, ZONE_DMA32 starts |
4209 | * at arch_max_dma_pfn. |
4210 | */ |
4211 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) |
4212 | { |
4213 | unsigned long nid; |
4214 | int i; |
4215 | |
4216 | /* Sort early_node_map as initialisation assumes it is sorted */ |
4217 | sort_node_map(); |
4218 | |
4219 | /* Record where the zone boundaries are */ |
4220 | memset(arch_zone_lowest_possible_pfn, 0, |
4221 | sizeof(arch_zone_lowest_possible_pfn)); |
4222 | memset(arch_zone_highest_possible_pfn, 0, |
4223 | sizeof(arch_zone_highest_possible_pfn)); |
4224 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); |
4225 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; |
4226 | for (i = 1; i < MAX_NR_ZONES; i++) { |
4227 | if (i == ZONE_MOVABLE) |
4228 | continue; |
4229 | arch_zone_lowest_possible_pfn[i] = |
4230 | arch_zone_highest_possible_pfn[i-1]; |
4231 | arch_zone_highest_possible_pfn[i] = |
4232 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); |
4233 | } |
4234 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; |
4235 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; |
4236 | |
4237 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ |
4238 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); |
4239 | find_zone_movable_pfns_for_nodes(zone_movable_pfn); |
4240 | |
4241 | /* Print out the zone ranges */ |
4242 | printk("Zone PFN ranges:\n"); |
4243 | for (i = 0; i < MAX_NR_ZONES; i++) { |
4244 | if (i == ZONE_MOVABLE) |
4245 | continue; |
4246 | printk(" %-8s %0#10lx -> %0#10lx\n", |
4247 | zone_names[i], |
4248 | arch_zone_lowest_possible_pfn[i], |
4249 | arch_zone_highest_possible_pfn[i]); |
4250 | } |
4251 | |
4252 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ |
4253 | printk("Movable zone start PFN for each node\n"); |
4254 | for (i = 0; i < MAX_NUMNODES; i++) { |
4255 | if (zone_movable_pfn[i]) |
4256 | printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); |
4257 | } |
4258 | |
4259 | /* Print out the early_node_map[] */ |
4260 | printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); |
4261 | for (i = 0; i < nr_nodemap_entries; i++) |
4262 | printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, |
4263 | early_node_map[i].start_pfn, |
4264 | early_node_map[i].end_pfn); |
4265 | |
4266 | /* Initialise every node */ |
4267 | mminit_verify_pageflags_layout(); |
4268 | setup_nr_node_ids(); |
4269 | for_each_online_node(nid) { |
4270 | pg_data_t *pgdat = NODE_DATA(nid); |
4271 | free_area_init_node(nid, NULL, |
4272 | find_min_pfn_for_node(nid), NULL); |
4273 | |
4274 | /* Any memory on that node */ |
4275 | if (pgdat->node_present_pages) |
4276 | node_set_state(nid, N_HIGH_MEMORY); |
4277 | check_for_regular_memory(pgdat); |
4278 | } |
4279 | } |
4280 | |
4281 | static int __init cmdline_parse_core(char *p, unsigned long *core) |
4282 | { |
4283 | unsigned long long coremem; |
4284 | if (!p) |
4285 | return -EINVAL; |
4286 | |
4287 | coremem = memparse(p, &p); |
4288 | *core = coremem >> PAGE_SHIFT; |
4289 | |
4290 | /* Paranoid check that UL is enough for the coremem value */ |
4291 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); |
4292 | |
4293 | return 0; |
4294 | } |
4295 | |
4296 | /* |
4297 | * kernelcore=size sets the amount of memory for use for allocations that |
4298 | * cannot be reclaimed or migrated. |
4299 | */ |
4300 | static int __init cmdline_parse_kernelcore(char *p) |
4301 | { |
4302 | return cmdline_parse_core(p, &required_kernelcore); |
4303 | } |
4304 | |
4305 | /* |
4306 | * movablecore=size sets the amount of memory for use for allocations that |
4307 | * can be reclaimed or migrated. |
4308 | */ |
4309 | static int __init cmdline_parse_movablecore(char *p) |
4310 | { |
4311 | return cmdline_parse_core(p, &required_movablecore); |
4312 | } |
4313 | |
4314 | early_param("kernelcore", cmdline_parse_kernelcore); |
4315 | early_param("movablecore", cmdline_parse_movablecore); |
4316 | |
4317 | #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ |
4318 | |
4319 | /** |
4320 | * set_dma_reserve - set the specified number of pages reserved in the first zone |
4321 | * @new_dma_reserve: The number of pages to mark reserved |
4322 | * |
4323 | * The per-cpu batchsize and zone watermarks are determined by present_pages. |
4324 | * In the DMA zone, a significant percentage may be consumed by kernel image |
4325 | * and other unfreeable allocations which can skew the watermarks badly. This |
4326 | * function may optionally be used to account for unfreeable pages in the |
4327 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and |
4328 | * smaller per-cpu batchsize. |
4329 | */ |
4330 | void __init set_dma_reserve(unsigned long new_dma_reserve) |
4331 | { |
4332 | dma_reserve = new_dma_reserve; |
4333 | } |
4334 | |
4335 | #ifndef CONFIG_NEED_MULTIPLE_NODES |
4336 | struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; |
4337 | EXPORT_SYMBOL(contig_page_data); |
4338 | #endif |
4339 | |
4340 | void __init free_area_init(unsigned long *zones_size) |
4341 | { |
4342 | free_area_init_node(0, zones_size, |
4343 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); |
4344 | } |
4345 | |
4346 | static int page_alloc_cpu_notify(struct notifier_block *self, |
4347 | unsigned long action, void *hcpu) |
4348 | { |
4349 | int cpu = (unsigned long)hcpu; |
4350 | |
4351 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { |
4352 | drain_pages(cpu); |
4353 | |
4354 | /* |
4355 | * Spill the event counters of the dead processor |
4356 | * into the current processors event counters. |
4357 | * This artificially elevates the count of the current |
4358 | * processor. |
4359 | */ |
4360 | vm_events_fold_cpu(cpu); |
4361 | |
4362 | /* |
4363 | * Zero the differential counters of the dead processor |
4364 | * so that the vm statistics are consistent. |
4365 | * |
4366 | * This is only okay since the processor is dead and cannot |
4367 | * race with what we are doing. |
4368 | */ |
4369 | refresh_cpu_vm_stats(cpu); |
4370 | } |
4371 | return NOTIFY_OK; |
4372 | } |
4373 | |
4374 | void __init page_alloc_init(void) |
4375 | { |
4376 | hotcpu_notifier(page_alloc_cpu_notify, 0); |
4377 | } |
4378 | |
4379 | /* |
4380 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio |
4381 | * or min_free_kbytes changes. |
4382 | */ |
4383 | static void calculate_totalreserve_pages(void) |
4384 | { |
4385 | struct pglist_data *pgdat; |
4386 | unsigned long reserve_pages = 0; |
4387 | enum zone_type i, j; |
4388 | |
4389 | for_each_online_pgdat(pgdat) { |
4390 | for (i = 0; i < MAX_NR_ZONES; i++) { |
4391 | struct zone *zone = pgdat->node_zones + i; |
4392 | unsigned long max = 0; |
4393 | |
4394 | /* Find valid and maximum lowmem_reserve in the zone */ |
4395 | for (j = i; j < MAX_NR_ZONES; j++) { |
4396 | if (zone->lowmem_reserve[j] > max) |
4397 | max = zone->lowmem_reserve[j]; |
4398 | } |
4399 | |
4400 | /* we treat the high watermark as reserved pages. */ |
4401 | max += high_wmark_pages(zone); |
4402 | |
4403 | if (max > zone->present_pages) |
4404 | max = zone->present_pages; |
4405 | reserve_pages += max; |
4406 | } |
4407 | } |
4408 | totalreserve_pages = reserve_pages; |
4409 | } |
4410 | |
4411 | /* |
4412 | * setup_per_zone_lowmem_reserve - called whenever |
4413 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone |
4414 | * has a correct pages reserved value, so an adequate number of |
4415 | * pages are left in the zone after a successful __alloc_pages(). |
4416 | */ |
4417 | static void setup_per_zone_lowmem_reserve(void) |
4418 | { |
4419 | struct pglist_data *pgdat; |
4420 | enum zone_type j, idx; |
4421 | |
4422 | for_each_online_pgdat(pgdat) { |
4423 | for (j = 0; j < MAX_NR_ZONES; j++) { |
4424 | struct zone *zone = pgdat->node_zones + j; |
4425 | unsigned long present_pages = zone->present_pages; |
4426 | |
4427 | zone->lowmem_reserve[j] = 0; |
4428 | |
4429 | idx = j; |
4430 | while (idx) { |
4431 | struct zone *lower_zone; |
4432 | |
4433 | idx--; |
4434 | |
4435 | if (sysctl_lowmem_reserve_ratio[idx] < 1) |
4436 | sysctl_lowmem_reserve_ratio[idx] = 1; |
4437 | |
4438 | lower_zone = pgdat->node_zones + idx; |
4439 | lower_zone->lowmem_reserve[j] = present_pages / |
4440 | sysctl_lowmem_reserve_ratio[idx]; |
4441 | present_pages += lower_zone->present_pages; |
4442 | } |
4443 | } |
4444 | } |
4445 | |
4446 | /* update totalreserve_pages */ |
4447 | calculate_totalreserve_pages(); |
4448 | } |
4449 | |
4450 | /** |
4451 | * setup_per_zone_wmarks - called when min_free_kbytes changes |
4452 | * or when memory is hot-{added|removed} |
4453 | * |
4454 | * Ensures that the watermark[min,low,high] values for each zone are set |
4455 | * correctly with respect to min_free_kbytes. |
4456 | */ |
4457 | void setup_per_zone_wmarks(void) |
4458 | { |
4459 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); |
4460 | unsigned long lowmem_pages = 0; |
4461 | struct zone *zone; |
4462 | unsigned long flags; |
4463 | |
4464 | /* Calculate total number of !ZONE_HIGHMEM pages */ |
4465 | for_each_zone(zone) { |
4466 | if (!is_highmem(zone)) |
4467 | lowmem_pages += zone->present_pages; |
4468 | } |
4469 | |
4470 | for_each_zone(zone) { |
4471 | u64 tmp; |
4472 | |
4473 | spin_lock_irqsave(&zone->lock, flags); |
4474 | tmp = (u64)pages_min * zone->present_pages; |
4475 | do_div(tmp, lowmem_pages); |
4476 | if (is_highmem(zone)) { |
4477 | /* |
4478 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't |
4479 | * need highmem pages, so cap pages_min to a small |
4480 | * value here. |
4481 | * |
4482 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) |
4483 | * deltas controls asynch page reclaim, and so should |
4484 | * not be capped for highmem. |
4485 | */ |
4486 | int min_pages; |
4487 | |
4488 | min_pages = zone->present_pages / 1024; |
4489 | if (min_pages < SWAP_CLUSTER_MAX) |
4490 | min_pages = SWAP_CLUSTER_MAX; |
4491 | if (min_pages > 128) |
4492 | min_pages = 128; |
4493 | zone->watermark[WMARK_MIN] = min_pages; |
4494 | } else { |
4495 | /* |
4496 | * If it's a lowmem zone, reserve a number of pages |
4497 | * proportionate to the zone's size. |
4498 | */ |
4499 | zone->watermark[WMARK_MIN] = tmp; |
4500 | } |
4501 | |
4502 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); |
4503 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); |
4504 | setup_zone_migrate_reserve(zone); |
4505 | spin_unlock_irqrestore(&zone->lock, flags); |
4506 | } |
4507 | |
4508 | /* update totalreserve_pages */ |
4509 | calculate_totalreserve_pages(); |
4510 | } |
4511 | |
4512 | /** |
4513 | * The inactive anon list should be small enough that the VM never has to |
4514 | * do too much work, but large enough that each inactive page has a chance |
4515 | * to be referenced again before it is swapped out. |
4516 | * |
4517 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to |
4518 | * INACTIVE_ANON pages on this zone's LRU, maintained by the |
4519 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of |
4520 | * the anonymous pages are kept on the inactive list. |
4521 | * |
4522 | * total target max |
4523 | * memory ratio inactive anon |
4524 | * ------------------------------------- |
4525 | * 10MB 1 5MB |
4526 | * 100MB 1 50MB |
4527 | * 1GB 3 250MB |
4528 | * 10GB 10 0.9GB |
4529 | * 100GB 31 3GB |
4530 | * 1TB 101 10GB |
4531 | * 10TB 320 32GB |
4532 | */ |
4533 | void calculate_zone_inactive_ratio(struct zone *zone) |
4534 | { |
4535 | unsigned int gb, ratio; |
4536 | |
4537 | /* Zone size in gigabytes */ |
4538 | gb = zone->present_pages >> (30 - PAGE_SHIFT); |
4539 | if (gb) |
4540 | ratio = int_sqrt(10 * gb); |
4541 | else |
4542 | ratio = 1; |
4543 | |
4544 | zone->inactive_ratio = ratio; |
4545 | } |
4546 | |
4547 | static void __init setup_per_zone_inactive_ratio(void) |
4548 | { |
4549 | struct zone *zone; |
4550 | |
4551 | for_each_zone(zone) |
4552 | calculate_zone_inactive_ratio(zone); |
4553 | } |
4554 | |
4555 | /* |
4556 | * Initialise min_free_kbytes. |
4557 | * |
4558 | * For small machines we want it small (128k min). For large machines |
4559 | * we want it large (64MB max). But it is not linear, because network |
4560 | * bandwidth does not increase linearly with machine size. We use |
4561 | * |
4562 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: |
4563 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) |
4564 | * |
4565 | * which yields |
4566 | * |
4567 | * 16MB: 512k |
4568 | * 32MB: 724k |
4569 | * 64MB: 1024k |
4570 | * 128MB: 1448k |
4571 | * 256MB: 2048k |
4572 | * 512MB: 2896k |
4573 | * 1024MB: 4096k |
4574 | * 2048MB: 5792k |
4575 | * 4096MB: 8192k |
4576 | * 8192MB: 11584k |
4577 | * 16384MB: 16384k |
4578 | */ |
4579 | static int __init init_per_zone_wmark_min(void) |
4580 | { |
4581 | unsigned long lowmem_kbytes; |
4582 | |
4583 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); |
4584 | |
4585 | min_free_kbytes = int_sqrt(lowmem_kbytes * 16); |
4586 | if (min_free_kbytes < 128) |
4587 | min_free_kbytes = 128; |
4588 | if (min_free_kbytes > 65536) |
4589 | min_free_kbytes = 65536; |
4590 | setup_per_zone_wmarks(); |
4591 | setup_per_zone_lowmem_reserve(); |
4592 | setup_per_zone_inactive_ratio(); |
4593 | return 0; |
4594 | } |
4595 | module_init(init_per_zone_wmark_min) |
4596 | |
4597 | /* |
4598 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so |
4599 | * that we can call two helper functions whenever min_free_kbytes |
4600 | * changes. |
4601 | */ |
4602 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, |
4603 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
4604 | { |
4605 | proc_dointvec(table, write, file, buffer, length, ppos); |
4606 | if (write) |
4607 | setup_per_zone_wmarks(); |
4608 | return 0; |
4609 | } |
4610 | |
4611 | #ifdef CONFIG_NUMA |
4612 | int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, |
4613 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
4614 | { |
4615 | struct zone *zone; |
4616 | int rc; |
4617 | |
4618 | rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
4619 | if (rc) |
4620 | return rc; |
4621 | |
4622 | for_each_zone(zone) |
4623 | zone->min_unmapped_pages = (zone->present_pages * |
4624 | sysctl_min_unmapped_ratio) / 100; |
4625 | return 0; |
4626 | } |
4627 | |
4628 | int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, |
4629 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
4630 | { |
4631 | struct zone *zone; |
4632 | int rc; |
4633 | |
4634 | rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
4635 | if (rc) |
4636 | return rc; |
4637 | |
4638 | for_each_zone(zone) |
4639 | zone->min_slab_pages = (zone->present_pages * |
4640 | sysctl_min_slab_ratio) / 100; |
4641 | return 0; |
4642 | } |
4643 | #endif |
4644 | |
4645 | /* |
4646 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around |
4647 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() |
4648 | * whenever sysctl_lowmem_reserve_ratio changes. |
4649 | * |
4650 | * The reserve ratio obviously has absolutely no relation with the |
4651 | * minimum watermarks. The lowmem reserve ratio can only make sense |
4652 | * if in function of the boot time zone sizes. |
4653 | */ |
4654 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, |
4655 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
4656 | { |
4657 | proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
4658 | setup_per_zone_lowmem_reserve(); |
4659 | return 0; |
4660 | } |
4661 | |
4662 | /* |
4663 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each |
4664 | * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist |
4665 | * can have before it gets flushed back to buddy allocator. |
4666 | */ |
4667 | |
4668 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, |
4669 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
4670 | { |
4671 | struct zone *zone; |
4672 | unsigned int cpu; |
4673 | int ret; |
4674 | |
4675 | ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
4676 | if (!write || (ret == -EINVAL)) |
4677 | return ret; |
4678 | for_each_populated_zone(zone) { |
4679 | for_each_online_cpu(cpu) { |
4680 | unsigned long high; |
4681 | high = zone->present_pages / percpu_pagelist_fraction; |
4682 | setup_pagelist_highmark(zone_pcp(zone, cpu), high); |
4683 | } |
4684 | } |
4685 | return 0; |
4686 | } |
4687 | |
4688 | int hashdist = HASHDIST_DEFAULT; |
4689 | |
4690 | #ifdef CONFIG_NUMA |
4691 | static int __init set_hashdist(char *str) |
4692 | { |
4693 | if (!str) |
4694 | return 0; |
4695 | hashdist = simple_strtoul(str, &str, 0); |
4696 | return 1; |
4697 | } |
4698 | __setup("hashdist=", set_hashdist); |
4699 | #endif |
4700 | |
4701 | /* |
4702 | * allocate a large system hash table from bootmem |
4703 | * - it is assumed that the hash table must contain an exact power-of-2 |
4704 | * quantity of entries |
4705 | * - limit is the number of hash buckets, not the total allocation size |
4706 | */ |
4707 | void *__init alloc_large_system_hash(const char *tablename, |
4708 | unsigned long bucketsize, |
4709 | unsigned long numentries, |
4710 | int scale, |
4711 | int flags, |
4712 | unsigned int *_hash_shift, |
4713 | unsigned int *_hash_mask, |
4714 | unsigned long limit) |
4715 | { |
4716 | unsigned long long max = limit; |
4717 | unsigned long log2qty, size; |
4718 | void *table = NULL; |
4719 | |
4720 | /* allow the kernel cmdline to have a say */ |
4721 | if (!numentries) { |
4722 | /* round applicable memory size up to nearest megabyte */ |
4723 | numentries = nr_kernel_pages; |
4724 | numentries += (1UL << (20 - PAGE_SHIFT)) - 1; |
4725 | numentries >>= 20 - PAGE_SHIFT; |
4726 | numentries <<= 20 - PAGE_SHIFT; |
4727 | |
4728 | /* limit to 1 bucket per 2^scale bytes of low memory */ |
4729 | if (scale > PAGE_SHIFT) |
4730 | numentries >>= (scale - PAGE_SHIFT); |
4731 | else |
4732 | numentries <<= (PAGE_SHIFT - scale); |
4733 | |
4734 | /* Make sure we've got at least a 0-order allocation.. */ |
4735 | if (unlikely((numentries * bucketsize) < PAGE_SIZE)) |
4736 | numentries = PAGE_SIZE / bucketsize; |
4737 | } |
4738 | numentries = roundup_pow_of_two(numentries); |
4739 | |
4740 | /* limit allocation size to 1/16 total memory by default */ |
4741 | if (max == 0) { |
4742 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; |
4743 | do_div(max, bucketsize); |
4744 | } |
4745 | |
4746 | if (numentries > max) |
4747 | numentries = max; |
4748 | |
4749 | log2qty = ilog2(numentries); |
4750 | |
4751 | do { |
4752 | size = bucketsize << log2qty; |
4753 | if (flags & HASH_EARLY) |
4754 | table = alloc_bootmem_nopanic(size); |
4755 | else if (hashdist) |
4756 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); |
4757 | else { |
4758 | /* |
4759 | * If bucketsize is not a power-of-two, we may free |
4760 | * some pages at the end of hash table which |
4761 | * alloc_pages_exact() automatically does |
4762 | */ |
4763 | if (get_order(size) < MAX_ORDER) { |
4764 | table = alloc_pages_exact(size, GFP_ATOMIC); |
4765 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); |
4766 | } |
4767 | } |
4768 | } while (!table && size > PAGE_SIZE && --log2qty); |
4769 | |
4770 | if (!table) |
4771 | panic("Failed to allocate %s hash table\n", tablename); |
4772 | |
4773 | printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", |
4774 | tablename, |
4775 | (1U << log2qty), |
4776 | ilog2(size) - PAGE_SHIFT, |
4777 | size); |
4778 | |
4779 | if (_hash_shift) |
4780 | *_hash_shift = log2qty; |
4781 | if (_hash_mask) |
4782 | *_hash_mask = (1 << log2qty) - 1; |
4783 | |
4784 | return table; |
4785 | } |
4786 | |
4787 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ |
4788 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, |
4789 | unsigned long pfn) |
4790 | { |
4791 | #ifdef CONFIG_SPARSEMEM |
4792 | return __pfn_to_section(pfn)->pageblock_flags; |
4793 | #else |
4794 | return zone->pageblock_flags; |
4795 | #endif /* CONFIG_SPARSEMEM */ |
4796 | } |
4797 | |
4798 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) |
4799 | { |
4800 | #ifdef CONFIG_SPARSEMEM |
4801 | pfn &= (PAGES_PER_SECTION-1); |
4802 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
4803 | #else |
4804 | pfn = pfn - zone->zone_start_pfn; |
4805 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
4806 | #endif /* CONFIG_SPARSEMEM */ |
4807 | } |
4808 | |
4809 | /** |
4810 | * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages |
4811 | * @page: The page within the block of interest |
4812 | * @start_bitidx: The first bit of interest to retrieve |
4813 | * @end_bitidx: The last bit of interest |
4814 | * returns pageblock_bits flags |
4815 | */ |
4816 | unsigned long get_pageblock_flags_group(struct page *page, |
4817 | int start_bitidx, int end_bitidx) |
4818 | { |
4819 | struct zone *zone; |
4820 | unsigned long *bitmap; |
4821 | unsigned long pfn, bitidx; |
4822 | unsigned long flags = 0; |
4823 | unsigned long value = 1; |
4824 | |
4825 | zone = page_zone(page); |
4826 | pfn = page_to_pfn(page); |
4827 | bitmap = get_pageblock_bitmap(zone, pfn); |
4828 | bitidx = pfn_to_bitidx(zone, pfn); |
4829 | |
4830 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) |
4831 | if (test_bit(bitidx + start_bitidx, bitmap)) |
4832 | flags |= value; |
4833 | |
4834 | return flags; |
4835 | } |
4836 | |
4837 | /** |
4838 | * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages |
4839 | * @page: The page within the block of interest |
4840 | * @start_bitidx: The first bit of interest |
4841 | * @end_bitidx: The last bit of interest |
4842 | * @flags: The flags to set |
4843 | */ |
4844 | void set_pageblock_flags_group(struct page *page, unsigned long flags, |
4845 | int start_bitidx, int end_bitidx) |
4846 | { |
4847 | struct zone *zone; |
4848 | unsigned long *bitmap; |
4849 | unsigned long pfn, bitidx; |
4850 | unsigned long value = 1; |
4851 | |
4852 | zone = page_zone(page); |
4853 | pfn = page_to_pfn(page); |
4854 | bitmap = get_pageblock_bitmap(zone, pfn); |
4855 | bitidx = pfn_to_bitidx(zone, pfn); |
4856 | VM_BUG_ON(pfn < zone->zone_start_pfn); |
4857 | VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); |
4858 | |
4859 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) |
4860 | if (flags & value) |
4861 | __set_bit(bitidx + start_bitidx, bitmap); |
4862 | else |
4863 | __clear_bit(bitidx + start_bitidx, bitmap); |
4864 | } |
4865 | |
4866 | /* |
4867 | * This is designed as sub function...plz see page_isolation.c also. |
4868 | * set/clear page block's type to be ISOLATE. |
4869 | * page allocater never alloc memory from ISOLATE block. |
4870 | */ |
4871 | |
4872 | int set_migratetype_isolate(struct page *page) |
4873 | { |
4874 | struct zone *zone; |
4875 | unsigned long flags; |
4876 | int ret = -EBUSY; |
4877 | |
4878 | zone = page_zone(page); |
4879 | spin_lock_irqsave(&zone->lock, flags); |
4880 | /* |
4881 | * In future, more migrate types will be able to be isolation target. |
4882 | */ |
4883 | if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE) |
4884 | goto out; |
4885 | set_pageblock_migratetype(page, MIGRATE_ISOLATE); |
4886 | move_freepages_block(zone, page, MIGRATE_ISOLATE); |
4887 | ret = 0; |
4888 | out: |
4889 | spin_unlock_irqrestore(&zone->lock, flags); |
4890 | if (!ret) |
4891 | drain_all_pages(); |
4892 | return ret; |
4893 | } |
4894 | |
4895 | void unset_migratetype_isolate(struct page *page) |
4896 | { |
4897 | struct zone *zone; |
4898 | unsigned long flags; |
4899 | zone = page_zone(page); |
4900 | spin_lock_irqsave(&zone->lock, flags); |
4901 | if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) |
4902 | goto out; |
4903 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); |
4904 | move_freepages_block(zone, page, MIGRATE_MOVABLE); |
4905 | out: |
4906 | spin_unlock_irqrestore(&zone->lock, flags); |
4907 | } |
4908 | |
4909 | #ifdef CONFIG_MEMORY_HOTREMOVE |
4910 | /* |
4911 | * All pages in the range must be isolated before calling this. |
4912 | */ |
4913 | void |
4914 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) |
4915 | { |
4916 | struct page *page; |
4917 | struct zone *zone; |
4918 | int order, i; |
4919 | unsigned long pfn; |
4920 | unsigned long flags; |
4921 | /* find the first valid pfn */ |
4922 | for (pfn = start_pfn; pfn < end_pfn; pfn++) |
4923 | if (pfn_valid(pfn)) |
4924 | break; |
4925 | if (pfn == end_pfn) |
4926 | return; |
4927 | zone = page_zone(pfn_to_page(pfn)); |
4928 | spin_lock_irqsave(&zone->lock, flags); |
4929 | pfn = start_pfn; |
4930 | while (pfn < end_pfn) { |
4931 | if (!pfn_valid(pfn)) { |
4932 | pfn++; |
4933 | continue; |
4934 | } |
4935 | page = pfn_to_page(pfn); |
4936 | BUG_ON(page_count(page)); |
4937 | BUG_ON(!PageBuddy(page)); |
4938 | order = page_order(page); |
4939 | #ifdef CONFIG_DEBUG_VM |
4940 | printk(KERN_INFO "remove from free list %lx %d %lx\n", |
4941 | pfn, 1 << order, end_pfn); |
4942 | #endif |
4943 | list_del(&page->lru); |
4944 | rmv_page_order(page); |
4945 | zone->free_area[order].nr_free--; |
4946 | __mod_zone_page_state(zone, NR_FREE_PAGES, |
4947 | - (1UL << order)); |
4948 | for (i = 0; i < (1 << order); i++) |
4949 | SetPageReserved((page+i)); |
4950 | pfn += (1 << order); |
4951 | } |
4952 | spin_unlock_irqrestore(&zone->lock, flags); |
4953 | } |
4954 | #endif |
4955 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9