]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/page_alloc.c | |
3 | * | |
4 | * Manages the free list, the system allocates free pages here. | |
5 | * Note that kmalloc() lives in slab.c | |
6 | * | |
7 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
8 | * Swap reorganised 29.12.95, Stephen Tweedie | |
9 | * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 | |
10 | * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 | |
11 | * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 | |
12 | * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 | |
13 | * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 | |
14 | * (lots of bits borrowed from Ingo Molnar & Andrew Morton) | |
15 | */ | |
16 | ||
17 | #include <linux/stddef.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/swap.h> | |
20 | #include <linux/interrupt.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/jiffies.h> | |
23 | #include <linux/bootmem.h> | |
24 | #include <linux/memblock.h> | |
25 | #include <linux/compiler.h> | |
26 | #include <linux/kernel.h> | |
27 | #include <linux/kmemcheck.h> | |
28 | #include <linux/kasan.h> | |
29 | #include <linux/module.h> | |
30 | #include <linux/suspend.h> | |
31 | #include <linux/pagevec.h> | |
32 | #include <linux/blkdev.h> | |
33 | #include <linux/slab.h> | |
34 | #include <linux/ratelimit.h> | |
35 | #include <linux/oom.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/topology.h> | |
38 | #include <linux/sysctl.h> | |
39 | #include <linux/cpu.h> | |
40 | #include <linux/cpuset.h> | |
41 | #include <linux/memory_hotplug.h> | |
42 | #include <linux/nodemask.h> | |
43 | #include <linux/vmalloc.h> | |
44 | #include <linux/vmstat.h> | |
45 | #include <linux/mempolicy.h> | |
46 | #include <linux/stop_machine.h> | |
47 | #include <linux/sort.h> | |
48 | #include <linux/pfn.h> | |
49 | #include <linux/backing-dev.h> | |
50 | #include <linux/fault-inject.h> | |
51 | #include <linux/page-isolation.h> | |
52 | #include <linux/page_ext.h> | |
53 | #include <linux/debugobjects.h> | |
54 | #include <linux/kmemleak.h> | |
55 | #include <linux/compaction.h> | |
56 | #include <trace/events/kmem.h> | |
57 | #include <linux/prefetch.h> | |
58 | #include <linux/mm_inline.h> | |
59 | #include <linux/migrate.h> | |
60 | #include <linux/page_ext.h> | |
61 | #include <linux/hugetlb.h> | |
62 | #include <linux/sched/rt.h> | |
63 | #include <linux/page_owner.h> | |
64 | ||
65 | #include <asm/sections.h> | |
66 | #include <asm/tlbflush.h> | |
67 | #include <asm/div64.h> | |
68 | #include "internal.h" | |
69 | ||
70 | /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ | |
71 | static DEFINE_MUTEX(pcp_batch_high_lock); | |
72 | #define MIN_PERCPU_PAGELIST_FRACTION (8) | |
73 | ||
74 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID | |
75 | DEFINE_PER_CPU(int, numa_node); | |
76 | EXPORT_PER_CPU_SYMBOL(numa_node); | |
77 | #endif | |
78 | ||
79 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
80 | /* | |
81 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. | |
82 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. | |
83 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() | |
84 | * defined in <linux/topology.h>. | |
85 | */ | |
86 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ | |
87 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); | |
88 | int _node_numa_mem_[MAX_NUMNODES]; | |
89 | #endif | |
90 | ||
91 | /* | |
92 | * Array of node states. | |
93 | */ | |
94 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { | |
95 | [N_POSSIBLE] = NODE_MASK_ALL, | |
96 | [N_ONLINE] = { { [0] = 1UL } }, | |
97 | #ifndef CONFIG_NUMA | |
98 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
99 | #ifdef CONFIG_HIGHMEM | |
100 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
101 | #endif | |
102 | #ifdef CONFIG_MOVABLE_NODE | |
103 | [N_MEMORY] = { { [0] = 1UL } }, | |
104 | #endif | |
105 | [N_CPU] = { { [0] = 1UL } }, | |
106 | #endif /* NUMA */ | |
107 | }; | |
108 | EXPORT_SYMBOL(node_states); | |
109 | ||
110 | /* Protect totalram_pages and zone->managed_pages */ | |
111 | static DEFINE_SPINLOCK(managed_page_count_lock); | |
112 | ||
113 | unsigned long totalram_pages __read_mostly; | |
114 | unsigned long totalreserve_pages __read_mostly; | |
115 | unsigned long totalcma_pages __read_mostly; | |
116 | /* | |
117 | * When calculating the number of globally allowed dirty pages, there | |
118 | * is a certain number of per-zone reserves that should not be | |
119 | * considered dirtyable memory. This is the sum of those reserves | |
120 | * over all existing zones that contribute dirtyable memory. | |
121 | */ | |
122 | unsigned long dirty_balance_reserve __read_mostly; | |
123 | ||
124 | int percpu_pagelist_fraction; | |
125 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; | |
126 | ||
127 | #ifdef CONFIG_PM_SLEEP | |
128 | /* | |
129 | * The following functions are used by the suspend/hibernate code to temporarily | |
130 | * change gfp_allowed_mask in order to avoid using I/O during memory allocations | |
131 | * while devices are suspended. To avoid races with the suspend/hibernate code, | |
132 | * they should always be called with pm_mutex held (gfp_allowed_mask also should | |
133 | * only be modified with pm_mutex held, unless the suspend/hibernate code is | |
134 | * guaranteed not to run in parallel with that modification). | |
135 | */ | |
136 | ||
137 | static gfp_t saved_gfp_mask; | |
138 | ||
139 | void pm_restore_gfp_mask(void) | |
140 | { | |
141 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
142 | if (saved_gfp_mask) { | |
143 | gfp_allowed_mask = saved_gfp_mask; | |
144 | saved_gfp_mask = 0; | |
145 | } | |
146 | } | |
147 | ||
148 | void pm_restrict_gfp_mask(void) | |
149 | { | |
150 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
151 | WARN_ON(saved_gfp_mask); | |
152 | saved_gfp_mask = gfp_allowed_mask; | |
153 | gfp_allowed_mask &= ~GFP_IOFS; | |
154 | } | |
155 | ||
156 | bool pm_suspended_storage(void) | |
157 | { | |
158 | if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) | |
159 | return false; | |
160 | return true; | |
161 | } | |
162 | #endif /* CONFIG_PM_SLEEP */ | |
163 | ||
164 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
165 | int pageblock_order __read_mostly; | |
166 | #endif | |
167 | ||
168 | static void __free_pages_ok(struct page *page, unsigned int order); | |
169 | ||
170 | /* | |
171 | * results with 256, 32 in the lowmem_reserve sysctl: | |
172 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
173 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
174 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
175 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
176 | * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA | |
177 | * | |
178 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
179 | * don't need any ZONE_NORMAL reservation | |
180 | */ | |
181 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { | |
182 | #ifdef CONFIG_ZONE_DMA | |
183 | 256, | |
184 | #endif | |
185 | #ifdef CONFIG_ZONE_DMA32 | |
186 | 256, | |
187 | #endif | |
188 | #ifdef CONFIG_HIGHMEM | |
189 | 32, | |
190 | #endif | |
191 | 32, | |
192 | }; | |
193 | ||
194 | EXPORT_SYMBOL(totalram_pages); | |
195 | ||
196 | static char * const zone_names[MAX_NR_ZONES] = { | |
197 | #ifdef CONFIG_ZONE_DMA | |
198 | "DMA", | |
199 | #endif | |
200 | #ifdef CONFIG_ZONE_DMA32 | |
201 | "DMA32", | |
202 | #endif | |
203 | "Normal", | |
204 | #ifdef CONFIG_HIGHMEM | |
205 | "HighMem", | |
206 | #endif | |
207 | "Movable", | |
208 | }; | |
209 | ||
210 | int min_free_kbytes = 1024; | |
211 | int user_min_free_kbytes = -1; | |
212 | ||
213 | static unsigned long __meminitdata nr_kernel_pages; | |
214 | static unsigned long __meminitdata nr_all_pages; | |
215 | static unsigned long __meminitdata dma_reserve; | |
216 | ||
217 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
218 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; | |
219 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; | |
220 | static unsigned long __initdata required_kernelcore; | |
221 | static unsigned long __initdata required_movablecore; | |
222 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; | |
223 | ||
224 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
225 | int movable_zone; | |
226 | EXPORT_SYMBOL(movable_zone); | |
227 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
228 | ||
229 | #if MAX_NUMNODES > 1 | |
230 | int nr_node_ids __read_mostly = MAX_NUMNODES; | |
231 | int nr_online_nodes __read_mostly = 1; | |
232 | EXPORT_SYMBOL(nr_node_ids); | |
233 | EXPORT_SYMBOL(nr_online_nodes); | |
234 | #endif | |
235 | ||
236 | int page_group_by_mobility_disabled __read_mostly; | |
237 | ||
238 | void set_pageblock_migratetype(struct page *page, int migratetype) | |
239 | { | |
240 | if (unlikely(page_group_by_mobility_disabled && | |
241 | migratetype < MIGRATE_PCPTYPES)) | |
242 | migratetype = MIGRATE_UNMOVABLE; | |
243 | ||
244 | set_pageblock_flags_group(page, (unsigned long)migratetype, | |
245 | PB_migrate, PB_migrate_end); | |
246 | } | |
247 | ||
248 | #ifdef CONFIG_DEBUG_VM | |
249 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) | |
250 | { | |
251 | int ret = 0; | |
252 | unsigned seq; | |
253 | unsigned long pfn = page_to_pfn(page); | |
254 | unsigned long sp, start_pfn; | |
255 | ||
256 | do { | |
257 | seq = zone_span_seqbegin(zone); | |
258 | start_pfn = zone->zone_start_pfn; | |
259 | sp = zone->spanned_pages; | |
260 | if (!zone_spans_pfn(zone, pfn)) | |
261 | ret = 1; | |
262 | } while (zone_span_seqretry(zone, seq)); | |
263 | ||
264 | if (ret) | |
265 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", | |
266 | pfn, zone_to_nid(zone), zone->name, | |
267 | start_pfn, start_pfn + sp); | |
268 | ||
269 | return ret; | |
270 | } | |
271 | ||
272 | static int page_is_consistent(struct zone *zone, struct page *page) | |
273 | { | |
274 | if (!pfn_valid_within(page_to_pfn(page))) | |
275 | return 0; | |
276 | if (zone != page_zone(page)) | |
277 | return 0; | |
278 | ||
279 | return 1; | |
280 | } | |
281 | /* | |
282 | * Temporary debugging check for pages not lying within a given zone. | |
283 | */ | |
284 | static int bad_range(struct zone *zone, struct page *page) | |
285 | { | |
286 | if (page_outside_zone_boundaries(zone, page)) | |
287 | return 1; | |
288 | if (!page_is_consistent(zone, page)) | |
289 | return 1; | |
290 | ||
291 | return 0; | |
292 | } | |
293 | #else | |
294 | static inline int bad_range(struct zone *zone, struct page *page) | |
295 | { | |
296 | return 0; | |
297 | } | |
298 | #endif | |
299 | ||
300 | static void bad_page(struct page *page, const char *reason, | |
301 | unsigned long bad_flags) | |
302 | { | |
303 | static unsigned long resume; | |
304 | static unsigned long nr_shown; | |
305 | static unsigned long nr_unshown; | |
306 | ||
307 | /* Don't complain about poisoned pages */ | |
308 | if (PageHWPoison(page)) { | |
309 | page_mapcount_reset(page); /* remove PageBuddy */ | |
310 | return; | |
311 | } | |
312 | ||
313 | /* | |
314 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
315 | * or allow a steady drip of one report per second. | |
316 | */ | |
317 | if (nr_shown == 60) { | |
318 | if (time_before(jiffies, resume)) { | |
319 | nr_unshown++; | |
320 | goto out; | |
321 | } | |
322 | if (nr_unshown) { | |
323 | printk(KERN_ALERT | |
324 | "BUG: Bad page state: %lu messages suppressed\n", | |
325 | nr_unshown); | |
326 | nr_unshown = 0; | |
327 | } | |
328 | nr_shown = 0; | |
329 | } | |
330 | if (nr_shown++ == 0) | |
331 | resume = jiffies + 60 * HZ; | |
332 | ||
333 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", | |
334 | current->comm, page_to_pfn(page)); | |
335 | dump_page_badflags(page, reason, bad_flags); | |
336 | ||
337 | print_modules(); | |
338 | dump_stack(); | |
339 | out: | |
340 | /* Leave bad fields for debug, except PageBuddy could make trouble */ | |
341 | page_mapcount_reset(page); /* remove PageBuddy */ | |
342 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); | |
343 | } | |
344 | ||
345 | /* | |
346 | * Higher-order pages are called "compound pages". They are structured thusly: | |
347 | * | |
348 | * The first PAGE_SIZE page is called the "head page". | |
349 | * | |
350 | * The remaining PAGE_SIZE pages are called "tail pages". | |
351 | * | |
352 | * All pages have PG_compound set. All tail pages have their ->first_page | |
353 | * pointing at the head page. | |
354 | * | |
355 | * The first tail page's ->lru.next holds the address of the compound page's | |
356 | * put_page() function. Its ->lru.prev holds the order of allocation. | |
357 | * This usage means that zero-order pages may not be compound. | |
358 | */ | |
359 | ||
360 | static void free_compound_page(struct page *page) | |
361 | { | |
362 | __free_pages_ok(page, compound_order(page)); | |
363 | } | |
364 | ||
365 | void prep_compound_page(struct page *page, unsigned long order) | |
366 | { | |
367 | int i; | |
368 | int nr_pages = 1 << order; | |
369 | ||
370 | set_compound_page_dtor(page, free_compound_page); | |
371 | set_compound_order(page, order); | |
372 | __SetPageHead(page); | |
373 | for (i = 1; i < nr_pages; i++) { | |
374 | struct page *p = page + i; | |
375 | set_page_count(p, 0); | |
376 | p->first_page = page; | |
377 | /* Make sure p->first_page is always valid for PageTail() */ | |
378 | smp_wmb(); | |
379 | __SetPageTail(p); | |
380 | } | |
381 | } | |
382 | ||
383 | static inline void prep_zero_page(struct page *page, unsigned int order, | |
384 | gfp_t gfp_flags) | |
385 | { | |
386 | int i; | |
387 | ||
388 | /* | |
389 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO | |
390 | * and __GFP_HIGHMEM from hard or soft interrupt context. | |
391 | */ | |
392 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); | |
393 | for (i = 0; i < (1 << order); i++) | |
394 | clear_highpage(page + i); | |
395 | } | |
396 | ||
397 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
398 | unsigned int _debug_guardpage_minorder; | |
399 | bool _debug_pagealloc_enabled __read_mostly; | |
400 | bool _debug_guardpage_enabled __read_mostly; | |
401 | ||
402 | static int __init early_debug_pagealloc(char *buf) | |
403 | { | |
404 | if (!buf) | |
405 | return -EINVAL; | |
406 | ||
407 | if (strcmp(buf, "on") == 0) | |
408 | _debug_pagealloc_enabled = true; | |
409 | ||
410 | return 0; | |
411 | } | |
412 | early_param("debug_pagealloc", early_debug_pagealloc); | |
413 | ||
414 | static bool need_debug_guardpage(void) | |
415 | { | |
416 | /* If we don't use debug_pagealloc, we don't need guard page */ | |
417 | if (!debug_pagealloc_enabled()) | |
418 | return false; | |
419 | ||
420 | return true; | |
421 | } | |
422 | ||
423 | static void init_debug_guardpage(void) | |
424 | { | |
425 | if (!debug_pagealloc_enabled()) | |
426 | return; | |
427 | ||
428 | _debug_guardpage_enabled = true; | |
429 | } | |
430 | ||
431 | struct page_ext_operations debug_guardpage_ops = { | |
432 | .need = need_debug_guardpage, | |
433 | .init = init_debug_guardpage, | |
434 | }; | |
435 | ||
436 | static int __init debug_guardpage_minorder_setup(char *buf) | |
437 | { | |
438 | unsigned long res; | |
439 | ||
440 | if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { | |
441 | printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); | |
442 | return 0; | |
443 | } | |
444 | _debug_guardpage_minorder = res; | |
445 | printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); | |
446 | return 0; | |
447 | } | |
448 | __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); | |
449 | ||
450 | static inline void set_page_guard(struct zone *zone, struct page *page, | |
451 | unsigned int order, int migratetype) | |
452 | { | |
453 | struct page_ext *page_ext; | |
454 | ||
455 | if (!debug_guardpage_enabled()) | |
456 | return; | |
457 | ||
458 | page_ext = lookup_page_ext(page); | |
459 | __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); | |
460 | ||
461 | INIT_LIST_HEAD(&page->lru); | |
462 | set_page_private(page, order); | |
463 | /* Guard pages are not available for any usage */ | |
464 | __mod_zone_freepage_state(zone, -(1 << order), migratetype); | |
465 | } | |
466 | ||
467 | static inline void clear_page_guard(struct zone *zone, struct page *page, | |
468 | unsigned int order, int migratetype) | |
469 | { | |
470 | struct page_ext *page_ext; | |
471 | ||
472 | if (!debug_guardpage_enabled()) | |
473 | return; | |
474 | ||
475 | page_ext = lookup_page_ext(page); | |
476 | __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); | |
477 | ||
478 | set_page_private(page, 0); | |
479 | if (!is_migrate_isolate(migratetype)) | |
480 | __mod_zone_freepage_state(zone, (1 << order), migratetype); | |
481 | } | |
482 | #else | |
483 | struct page_ext_operations debug_guardpage_ops = { NULL, }; | |
484 | static inline void set_page_guard(struct zone *zone, struct page *page, | |
485 | unsigned int order, int migratetype) {} | |
486 | static inline void clear_page_guard(struct zone *zone, struct page *page, | |
487 | unsigned int order, int migratetype) {} | |
488 | #endif | |
489 | ||
490 | static inline void set_page_order(struct page *page, unsigned int order) | |
491 | { | |
492 | set_page_private(page, order); | |
493 | __SetPageBuddy(page); | |
494 | } | |
495 | ||
496 | static inline void rmv_page_order(struct page *page) | |
497 | { | |
498 | __ClearPageBuddy(page); | |
499 | set_page_private(page, 0); | |
500 | } | |
501 | ||
502 | /* | |
503 | * This function checks whether a page is free && is the buddy | |
504 | * we can do coalesce a page and its buddy if | |
505 | * (a) the buddy is not in a hole && | |
506 | * (b) the buddy is in the buddy system && | |
507 | * (c) a page and its buddy have the same order && | |
508 | * (d) a page and its buddy are in the same zone. | |
509 | * | |
510 | * For recording whether a page is in the buddy system, we set ->_mapcount | |
511 | * PAGE_BUDDY_MAPCOUNT_VALUE. | |
512 | * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is | |
513 | * serialized by zone->lock. | |
514 | * | |
515 | * For recording page's order, we use page_private(page). | |
516 | */ | |
517 | static inline int page_is_buddy(struct page *page, struct page *buddy, | |
518 | unsigned int order) | |
519 | { | |
520 | if (!pfn_valid_within(page_to_pfn(buddy))) | |
521 | return 0; | |
522 | ||
523 | if (page_is_guard(buddy) && page_order(buddy) == order) { | |
524 | if (page_zone_id(page) != page_zone_id(buddy)) | |
525 | return 0; | |
526 | ||
527 | VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); | |
528 | ||
529 | return 1; | |
530 | } | |
531 | ||
532 | if (PageBuddy(buddy) && page_order(buddy) == order) { | |
533 | /* | |
534 | * zone check is done late to avoid uselessly | |
535 | * calculating zone/node ids for pages that could | |
536 | * never merge. | |
537 | */ | |
538 | if (page_zone_id(page) != page_zone_id(buddy)) | |
539 | return 0; | |
540 | ||
541 | VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); | |
542 | ||
543 | return 1; | |
544 | } | |
545 | return 0; | |
546 | } | |
547 | ||
548 | /* | |
549 | * Freeing function for a buddy system allocator. | |
550 | * | |
551 | * The concept of a buddy system is to maintain direct-mapped table | |
552 | * (containing bit values) for memory blocks of various "orders". | |
553 | * The bottom level table contains the map for the smallest allocatable | |
554 | * units of memory (here, pages), and each level above it describes | |
555 | * pairs of units from the levels below, hence, "buddies". | |
556 | * At a high level, all that happens here is marking the table entry | |
557 | * at the bottom level available, and propagating the changes upward | |
558 | * as necessary, plus some accounting needed to play nicely with other | |
559 | * parts of the VM system. | |
560 | * At each level, we keep a list of pages, which are heads of continuous | |
561 | * free pages of length of (1 << order) and marked with _mapcount | |
562 | * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) | |
563 | * field. | |
564 | * So when we are allocating or freeing one, we can derive the state of the | |
565 | * other. That is, if we allocate a small block, and both were | |
566 | * free, the remainder of the region must be split into blocks. | |
567 | * If a block is freed, and its buddy is also free, then this | |
568 | * triggers coalescing into a block of larger size. | |
569 | * | |
570 | * -- nyc | |
571 | */ | |
572 | ||
573 | static inline void __free_one_page(struct page *page, | |
574 | unsigned long pfn, | |
575 | struct zone *zone, unsigned int order, | |
576 | int migratetype) | |
577 | { | |
578 | unsigned long page_idx; | |
579 | unsigned long combined_idx; | |
580 | unsigned long uninitialized_var(buddy_idx); | |
581 | struct page *buddy; | |
582 | int max_order = MAX_ORDER; | |
583 | ||
584 | VM_BUG_ON(!zone_is_initialized(zone)); | |
585 | VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); | |
586 | ||
587 | VM_BUG_ON(migratetype == -1); | |
588 | if (is_migrate_isolate(migratetype)) { | |
589 | /* | |
590 | * We restrict max order of merging to prevent merge | |
591 | * between freepages on isolate pageblock and normal | |
592 | * pageblock. Without this, pageblock isolation | |
593 | * could cause incorrect freepage accounting. | |
594 | */ | |
595 | max_order = min(MAX_ORDER, pageblock_order + 1); | |
596 | } else { | |
597 | __mod_zone_freepage_state(zone, 1 << order, migratetype); | |
598 | } | |
599 | ||
600 | page_idx = pfn & ((1 << max_order) - 1); | |
601 | ||
602 | VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page); | |
603 | VM_BUG_ON_PAGE(bad_range(zone, page), page); | |
604 | ||
605 | while (order < max_order - 1) { | |
606 | buddy_idx = __find_buddy_index(page_idx, order); | |
607 | buddy = page + (buddy_idx - page_idx); | |
608 | if (!page_is_buddy(page, buddy, order)) | |
609 | break; | |
610 | /* | |
611 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, | |
612 | * merge with it and move up one order. | |
613 | */ | |
614 | if (page_is_guard(buddy)) { | |
615 | clear_page_guard(zone, buddy, order, migratetype); | |
616 | } else { | |
617 | list_del(&buddy->lru); | |
618 | zone->free_area[order].nr_free--; | |
619 | rmv_page_order(buddy); | |
620 | } | |
621 | combined_idx = buddy_idx & page_idx; | |
622 | page = page + (combined_idx - page_idx); | |
623 | page_idx = combined_idx; | |
624 | order++; | |
625 | } | |
626 | set_page_order(page, order); | |
627 | ||
628 | /* | |
629 | * If this is not the largest possible page, check if the buddy | |
630 | * of the next-highest order is free. If it is, it's possible | |
631 | * that pages are being freed that will coalesce soon. In case, | |
632 | * that is happening, add the free page to the tail of the list | |
633 | * so it's less likely to be used soon and more likely to be merged | |
634 | * as a higher order page | |
635 | */ | |
636 | if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { | |
637 | struct page *higher_page, *higher_buddy; | |
638 | combined_idx = buddy_idx & page_idx; | |
639 | higher_page = page + (combined_idx - page_idx); | |
640 | buddy_idx = __find_buddy_index(combined_idx, order + 1); | |
641 | higher_buddy = higher_page + (buddy_idx - combined_idx); | |
642 | if (page_is_buddy(higher_page, higher_buddy, order + 1)) { | |
643 | list_add_tail(&page->lru, | |
644 | &zone->free_area[order].free_list[migratetype]); | |
645 | goto out; | |
646 | } | |
647 | } | |
648 | ||
649 | list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); | |
650 | out: | |
651 | zone->free_area[order].nr_free++; | |
652 | } | |
653 | ||
654 | static inline int free_pages_check(struct page *page) | |
655 | { | |
656 | const char *bad_reason = NULL; | |
657 | unsigned long bad_flags = 0; | |
658 | ||
659 | if (unlikely(page_mapcount(page))) | |
660 | bad_reason = "nonzero mapcount"; | |
661 | if (unlikely(page->mapping != NULL)) | |
662 | bad_reason = "non-NULL mapping"; | |
663 | if (unlikely(atomic_read(&page->_count) != 0)) | |
664 | bad_reason = "nonzero _count"; | |
665 | if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { | |
666 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; | |
667 | bad_flags = PAGE_FLAGS_CHECK_AT_FREE; | |
668 | } | |
669 | #ifdef CONFIG_MEMCG | |
670 | if (unlikely(page->mem_cgroup)) | |
671 | bad_reason = "page still charged to cgroup"; | |
672 | #endif | |
673 | if (unlikely(bad_reason)) { | |
674 | bad_page(page, bad_reason, bad_flags); | |
675 | return 1; | |
676 | } | |
677 | page_cpupid_reset_last(page); | |
678 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | |
679 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
680 | return 0; | |
681 | } | |
682 | ||
683 | /* | |
684 | * Frees a number of pages from the PCP lists | |
685 | * Assumes all pages on list are in same zone, and of same order. | |
686 | * count is the number of pages to free. | |
687 | * | |
688 | * If the zone was previously in an "all pages pinned" state then look to | |
689 | * see if this freeing clears that state. | |
690 | * | |
691 | * And clear the zone's pages_scanned counter, to hold off the "all pages are | |
692 | * pinned" detection logic. | |
693 | */ | |
694 | static void free_pcppages_bulk(struct zone *zone, int count, | |
695 | struct per_cpu_pages *pcp) | |
696 | { | |
697 | int migratetype = 0; | |
698 | int batch_free = 0; | |
699 | int to_free = count; | |
700 | unsigned long nr_scanned; | |
701 | ||
702 | spin_lock(&zone->lock); | |
703 | nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); | |
704 | if (nr_scanned) | |
705 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); | |
706 | ||
707 | while (to_free) { | |
708 | struct page *page; | |
709 | struct list_head *list; | |
710 | ||
711 | /* | |
712 | * Remove pages from lists in a round-robin fashion. A | |
713 | * batch_free count is maintained that is incremented when an | |
714 | * empty list is encountered. This is so more pages are freed | |
715 | * off fuller lists instead of spinning excessively around empty | |
716 | * lists | |
717 | */ | |
718 | do { | |
719 | batch_free++; | |
720 | if (++migratetype == MIGRATE_PCPTYPES) | |
721 | migratetype = 0; | |
722 | list = &pcp->lists[migratetype]; | |
723 | } while (list_empty(list)); | |
724 | ||
725 | /* This is the only non-empty list. Free them all. */ | |
726 | if (batch_free == MIGRATE_PCPTYPES) | |
727 | batch_free = to_free; | |
728 | ||
729 | do { | |
730 | int mt; /* migratetype of the to-be-freed page */ | |
731 | ||
732 | page = list_entry(list->prev, struct page, lru); | |
733 | /* must delete as __free_one_page list manipulates */ | |
734 | list_del(&page->lru); | |
735 | mt = get_freepage_migratetype(page); | |
736 | if (unlikely(has_isolate_pageblock(zone))) | |
737 | mt = get_pageblock_migratetype(page); | |
738 | ||
739 | /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ | |
740 | __free_one_page(page, page_to_pfn(page), zone, 0, mt); | |
741 | trace_mm_page_pcpu_drain(page, 0, mt); | |
742 | } while (--to_free && --batch_free && !list_empty(list)); | |
743 | } | |
744 | spin_unlock(&zone->lock); | |
745 | } | |
746 | ||
747 | static void free_one_page(struct zone *zone, | |
748 | struct page *page, unsigned long pfn, | |
749 | unsigned int order, | |
750 | int migratetype) | |
751 | { | |
752 | unsigned long nr_scanned; | |
753 | spin_lock(&zone->lock); | |
754 | nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); | |
755 | if (nr_scanned) | |
756 | __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); | |
757 | ||
758 | if (unlikely(has_isolate_pageblock(zone) || | |
759 | is_migrate_isolate(migratetype))) { | |
760 | migratetype = get_pfnblock_migratetype(page, pfn); | |
761 | } | |
762 | __free_one_page(page, pfn, zone, order, migratetype); | |
763 | spin_unlock(&zone->lock); | |
764 | } | |
765 | ||
766 | static int free_tail_pages_check(struct page *head_page, struct page *page) | |
767 | { | |
768 | if (!IS_ENABLED(CONFIG_DEBUG_VM)) | |
769 | return 0; | |
770 | if (unlikely(!PageTail(page))) { | |
771 | bad_page(page, "PageTail not set", 0); | |
772 | return 1; | |
773 | } | |
774 | if (unlikely(page->first_page != head_page)) { | |
775 | bad_page(page, "first_page not consistent", 0); | |
776 | return 1; | |
777 | } | |
778 | return 0; | |
779 | } | |
780 | ||
781 | static bool free_pages_prepare(struct page *page, unsigned int order) | |
782 | { | |
783 | bool compound = PageCompound(page); | |
784 | int i, bad = 0; | |
785 | ||
786 | VM_BUG_ON_PAGE(PageTail(page), page); | |
787 | VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); | |
788 | ||
789 | trace_mm_page_free(page, order); | |
790 | kmemcheck_free_shadow(page, order); | |
791 | kasan_free_pages(page, order); | |
792 | ||
793 | if (PageAnon(page)) | |
794 | page->mapping = NULL; | |
795 | bad += free_pages_check(page); | |
796 | for (i = 1; i < (1 << order); i++) { | |
797 | if (compound) | |
798 | bad += free_tail_pages_check(page, page + i); | |
799 | bad += free_pages_check(page + i); | |
800 | } | |
801 | if (bad) | |
802 | return false; | |
803 | ||
804 | reset_page_owner(page, order); | |
805 | ||
806 | if (!PageHighMem(page)) { | |
807 | debug_check_no_locks_freed(page_address(page), | |
808 | PAGE_SIZE << order); | |
809 | debug_check_no_obj_freed(page_address(page), | |
810 | PAGE_SIZE << order); | |
811 | } | |
812 | arch_free_page(page, order); | |
813 | kernel_map_pages(page, 1 << order, 0); | |
814 | ||
815 | return true; | |
816 | } | |
817 | ||
818 | static void __free_pages_ok(struct page *page, unsigned int order) | |
819 | { | |
820 | unsigned long flags; | |
821 | int migratetype; | |
822 | unsigned long pfn = page_to_pfn(page); | |
823 | ||
824 | if (!free_pages_prepare(page, order)) | |
825 | return; | |
826 | ||
827 | migratetype = get_pfnblock_migratetype(page, pfn); | |
828 | local_irq_save(flags); | |
829 | __count_vm_events(PGFREE, 1 << order); | |
830 | set_freepage_migratetype(page, migratetype); | |
831 | free_one_page(page_zone(page), page, pfn, order, migratetype); | |
832 | local_irq_restore(flags); | |
833 | } | |
834 | ||
835 | void __init __free_pages_bootmem(struct page *page, unsigned int order) | |
836 | { | |
837 | unsigned int nr_pages = 1 << order; | |
838 | struct page *p = page; | |
839 | unsigned int loop; | |
840 | ||
841 | prefetchw(p); | |
842 | for (loop = 0; loop < (nr_pages - 1); loop++, p++) { | |
843 | prefetchw(p + 1); | |
844 | __ClearPageReserved(p); | |
845 | set_page_count(p, 0); | |
846 | } | |
847 | __ClearPageReserved(p); | |
848 | set_page_count(p, 0); | |
849 | ||
850 | page_zone(page)->managed_pages += nr_pages; | |
851 | set_page_refcounted(page); | |
852 | __free_pages(page, order); | |
853 | } | |
854 | ||
855 | #ifdef CONFIG_CMA | |
856 | /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ | |
857 | void __init init_cma_reserved_pageblock(struct page *page) | |
858 | { | |
859 | unsigned i = pageblock_nr_pages; | |
860 | struct page *p = page; | |
861 | ||
862 | do { | |
863 | __ClearPageReserved(p); | |
864 | set_page_count(p, 0); | |
865 | } while (++p, --i); | |
866 | ||
867 | set_pageblock_migratetype(page, MIGRATE_CMA); | |
868 | ||
869 | if (pageblock_order >= MAX_ORDER) { | |
870 | i = pageblock_nr_pages; | |
871 | p = page; | |
872 | do { | |
873 | set_page_refcounted(p); | |
874 | __free_pages(p, MAX_ORDER - 1); | |
875 | p += MAX_ORDER_NR_PAGES; | |
876 | } while (i -= MAX_ORDER_NR_PAGES); | |
877 | } else { | |
878 | set_page_refcounted(page); | |
879 | __free_pages(page, pageblock_order); | |
880 | } | |
881 | ||
882 | adjust_managed_page_count(page, pageblock_nr_pages); | |
883 | } | |
884 | #endif | |
885 | ||
886 | /* | |
887 | * The order of subdivision here is critical for the IO subsystem. | |
888 | * Please do not alter this order without good reasons and regression | |
889 | * testing. Specifically, as large blocks of memory are subdivided, | |
890 | * the order in which smaller blocks are delivered depends on the order | |
891 | * they're subdivided in this function. This is the primary factor | |
892 | * influencing the order in which pages are delivered to the IO | |
893 | * subsystem according to empirical testing, and this is also justified | |
894 | * by considering the behavior of a buddy system containing a single | |
895 | * large block of memory acted on by a series of small allocations. | |
896 | * This behavior is a critical factor in sglist merging's success. | |
897 | * | |
898 | * -- nyc | |
899 | */ | |
900 | static inline void expand(struct zone *zone, struct page *page, | |
901 | int low, int high, struct free_area *area, | |
902 | int migratetype) | |
903 | { | |
904 | unsigned long size = 1 << high; | |
905 | ||
906 | while (high > low) { | |
907 | area--; | |
908 | high--; | |
909 | size >>= 1; | |
910 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); | |
911 | ||
912 | if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) && | |
913 | debug_guardpage_enabled() && | |
914 | high < debug_guardpage_minorder()) { | |
915 | /* | |
916 | * Mark as guard pages (or page), that will allow to | |
917 | * merge back to allocator when buddy will be freed. | |
918 | * Corresponding page table entries will not be touched, | |
919 | * pages will stay not present in virtual address space | |
920 | */ | |
921 | set_page_guard(zone, &page[size], high, migratetype); | |
922 | continue; | |
923 | } | |
924 | list_add(&page[size].lru, &area->free_list[migratetype]); | |
925 | area->nr_free++; | |
926 | set_page_order(&page[size], high); | |
927 | } | |
928 | } | |
929 | ||
930 | /* | |
931 | * This page is about to be returned from the page allocator | |
932 | */ | |
933 | static inline int check_new_page(struct page *page) | |
934 | { | |
935 | const char *bad_reason = NULL; | |
936 | unsigned long bad_flags = 0; | |
937 | ||
938 | if (unlikely(page_mapcount(page))) | |
939 | bad_reason = "nonzero mapcount"; | |
940 | if (unlikely(page->mapping != NULL)) | |
941 | bad_reason = "non-NULL mapping"; | |
942 | if (unlikely(atomic_read(&page->_count) != 0)) | |
943 | bad_reason = "nonzero _count"; | |
944 | if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { | |
945 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; | |
946 | bad_flags = PAGE_FLAGS_CHECK_AT_PREP; | |
947 | } | |
948 | #ifdef CONFIG_MEMCG | |
949 | if (unlikely(page->mem_cgroup)) | |
950 | bad_reason = "page still charged to cgroup"; | |
951 | #endif | |
952 | if (unlikely(bad_reason)) { | |
953 | bad_page(page, bad_reason, bad_flags); | |
954 | return 1; | |
955 | } | |
956 | return 0; | |
957 | } | |
958 | ||
959 | static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, | |
960 | int alloc_flags) | |
961 | { | |
962 | int i; | |
963 | ||
964 | for (i = 0; i < (1 << order); i++) { | |
965 | struct page *p = page + i; | |
966 | if (unlikely(check_new_page(p))) | |
967 | return 1; | |
968 | } | |
969 | ||
970 | set_page_private(page, 0); | |
971 | set_page_refcounted(page); | |
972 | ||
973 | arch_alloc_page(page, order); | |
974 | kernel_map_pages(page, 1 << order, 1); | |
975 | kasan_alloc_pages(page, order); | |
976 | ||
977 | if (gfp_flags & __GFP_ZERO) | |
978 | prep_zero_page(page, order, gfp_flags); | |
979 | ||
980 | if (order && (gfp_flags & __GFP_COMP)) | |
981 | prep_compound_page(page, order); | |
982 | ||
983 | set_page_owner(page, order, gfp_flags); | |
984 | ||
985 | /* | |
986 | * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to | |
987 | * allocate the page. The expectation is that the caller is taking | |
988 | * steps that will free more memory. The caller should avoid the page | |
989 | * being used for !PFMEMALLOC purposes. | |
990 | */ | |
991 | page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); | |
992 | ||
993 | return 0; | |
994 | } | |
995 | ||
996 | /* | |
997 | * Go through the free lists for the given migratetype and remove | |
998 | * the smallest available page from the freelists | |
999 | */ | |
1000 | static inline | |
1001 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, | |
1002 | int migratetype) | |
1003 | { | |
1004 | unsigned int current_order; | |
1005 | struct free_area *area; | |
1006 | struct page *page; | |
1007 | ||
1008 | /* Find a page of the appropriate size in the preferred list */ | |
1009 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { | |
1010 | area = &(zone->free_area[current_order]); | |
1011 | if (list_empty(&area->free_list[migratetype])) | |
1012 | continue; | |
1013 | ||
1014 | page = list_entry(area->free_list[migratetype].next, | |
1015 | struct page, lru); | |
1016 | list_del(&page->lru); | |
1017 | rmv_page_order(page); | |
1018 | area->nr_free--; | |
1019 | expand(zone, page, order, current_order, area, migratetype); | |
1020 | set_freepage_migratetype(page, migratetype); | |
1021 | return page; | |
1022 | } | |
1023 | ||
1024 | return NULL; | |
1025 | } | |
1026 | ||
1027 | ||
1028 | /* | |
1029 | * This array describes the order lists are fallen back to when | |
1030 | * the free lists for the desirable migrate type are depleted | |
1031 | */ | |
1032 | static int fallbacks[MIGRATE_TYPES][4] = { | |
1033 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
1034 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
1035 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
1036 | #ifdef CONFIG_CMA | |
1037 | [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ | |
1038 | #endif | |
1039 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ | |
1040 | #ifdef CONFIG_MEMORY_ISOLATION | |
1041 | [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ | |
1042 | #endif | |
1043 | }; | |
1044 | ||
1045 | #ifdef CONFIG_CMA | |
1046 | static struct page *__rmqueue_cma_fallback(struct zone *zone, | |
1047 | unsigned int order) | |
1048 | { | |
1049 | return __rmqueue_smallest(zone, order, MIGRATE_CMA); | |
1050 | } | |
1051 | #else | |
1052 | static inline struct page *__rmqueue_cma_fallback(struct zone *zone, | |
1053 | unsigned int order) { return NULL; } | |
1054 | #endif | |
1055 | ||
1056 | /* | |
1057 | * Move the free pages in a range to the free lists of the requested type. | |
1058 | * Note that start_page and end_pages are not aligned on a pageblock | |
1059 | * boundary. If alignment is required, use move_freepages_block() | |
1060 | */ | |
1061 | int move_freepages(struct zone *zone, | |
1062 | struct page *start_page, struct page *end_page, | |
1063 | int migratetype) | |
1064 | { | |
1065 | struct page *page; | |
1066 | unsigned long order; | |
1067 | int pages_moved = 0; | |
1068 | ||
1069 | #ifndef CONFIG_HOLES_IN_ZONE | |
1070 | /* | |
1071 | * page_zone is not safe to call in this context when | |
1072 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant | |
1073 | * anyway as we check zone boundaries in move_freepages_block(). | |
1074 | * Remove at a later date when no bug reports exist related to | |
1075 | * grouping pages by mobility | |
1076 | */ | |
1077 | VM_BUG_ON(page_zone(start_page) != page_zone(end_page)); | |
1078 | #endif | |
1079 | ||
1080 | for (page = start_page; page <= end_page;) { | |
1081 | /* Make sure we are not inadvertently changing nodes */ | |
1082 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); | |
1083 | ||
1084 | if (!pfn_valid_within(page_to_pfn(page))) { | |
1085 | page++; | |
1086 | continue; | |
1087 | } | |
1088 | ||
1089 | if (!PageBuddy(page)) { | |
1090 | page++; | |
1091 | continue; | |
1092 | } | |
1093 | ||
1094 | order = page_order(page); | |
1095 | list_move(&page->lru, | |
1096 | &zone->free_area[order].free_list[migratetype]); | |
1097 | set_freepage_migratetype(page, migratetype); | |
1098 | page += 1 << order; | |
1099 | pages_moved += 1 << order; | |
1100 | } | |
1101 | ||
1102 | return pages_moved; | |
1103 | } | |
1104 | ||
1105 | int move_freepages_block(struct zone *zone, struct page *page, | |
1106 | int migratetype) | |
1107 | { | |
1108 | unsigned long start_pfn, end_pfn; | |
1109 | struct page *start_page, *end_page; | |
1110 | ||
1111 | start_pfn = page_to_pfn(page); | |
1112 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); | |
1113 | start_page = pfn_to_page(start_pfn); | |
1114 | end_page = start_page + pageblock_nr_pages - 1; | |
1115 | end_pfn = start_pfn + pageblock_nr_pages - 1; | |
1116 | ||
1117 | /* Do not cross zone boundaries */ | |
1118 | if (!zone_spans_pfn(zone, start_pfn)) | |
1119 | start_page = page; | |
1120 | if (!zone_spans_pfn(zone, end_pfn)) | |
1121 | return 0; | |
1122 | ||
1123 | return move_freepages(zone, start_page, end_page, migratetype); | |
1124 | } | |
1125 | ||
1126 | static void change_pageblock_range(struct page *pageblock_page, | |
1127 | int start_order, int migratetype) | |
1128 | { | |
1129 | int nr_pageblocks = 1 << (start_order - pageblock_order); | |
1130 | ||
1131 | while (nr_pageblocks--) { | |
1132 | set_pageblock_migratetype(pageblock_page, migratetype); | |
1133 | pageblock_page += pageblock_nr_pages; | |
1134 | } | |
1135 | } | |
1136 | ||
1137 | /* | |
1138 | * When we are falling back to another migratetype during allocation, try to | |
1139 | * steal extra free pages from the same pageblocks to satisfy further | |
1140 | * allocations, instead of polluting multiple pageblocks. | |
1141 | * | |
1142 | * If we are stealing a relatively large buddy page, it is likely there will | |
1143 | * be more free pages in the pageblock, so try to steal them all. For | |
1144 | * reclaimable and unmovable allocations, we steal regardless of page size, | |
1145 | * as fragmentation caused by those allocations polluting movable pageblocks | |
1146 | * is worse than movable allocations stealing from unmovable and reclaimable | |
1147 | * pageblocks. | |
1148 | * | |
1149 | * If we claim more than half of the pageblock, change pageblock's migratetype | |
1150 | * as well. | |
1151 | */ | |
1152 | static void try_to_steal_freepages(struct zone *zone, struct page *page, | |
1153 | int start_type, int fallback_type) | |
1154 | { | |
1155 | int current_order = page_order(page); | |
1156 | ||
1157 | /* Take ownership for orders >= pageblock_order */ | |
1158 | if (current_order >= pageblock_order) { | |
1159 | change_pageblock_range(page, current_order, start_type); | |
1160 | return; | |
1161 | } | |
1162 | ||
1163 | if (current_order >= pageblock_order / 2 || | |
1164 | start_type == MIGRATE_RECLAIMABLE || | |
1165 | start_type == MIGRATE_UNMOVABLE || | |
1166 | page_group_by_mobility_disabled) { | |
1167 | int pages; | |
1168 | ||
1169 | pages = move_freepages_block(zone, page, start_type); | |
1170 | ||
1171 | /* Claim the whole block if over half of it is free */ | |
1172 | if (pages >= (1 << (pageblock_order-1)) || | |
1173 | page_group_by_mobility_disabled) | |
1174 | set_pageblock_migratetype(page, start_type); | |
1175 | } | |
1176 | } | |
1177 | ||
1178 | /* Remove an element from the buddy allocator from the fallback list */ | |
1179 | static inline struct page * | |
1180 | __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) | |
1181 | { | |
1182 | struct free_area *area; | |
1183 | unsigned int current_order; | |
1184 | struct page *page; | |
1185 | ||
1186 | /* Find the largest possible block of pages in the other list */ | |
1187 | for (current_order = MAX_ORDER-1; | |
1188 | current_order >= order && current_order <= MAX_ORDER-1; | |
1189 | --current_order) { | |
1190 | int i; | |
1191 | for (i = 0;; i++) { | |
1192 | int migratetype = fallbacks[start_migratetype][i]; | |
1193 | int buddy_type = start_migratetype; | |
1194 | ||
1195 | /* MIGRATE_RESERVE handled later if necessary */ | |
1196 | if (migratetype == MIGRATE_RESERVE) | |
1197 | break; | |
1198 | ||
1199 | area = &(zone->free_area[current_order]); | |
1200 | if (list_empty(&area->free_list[migratetype])) | |
1201 | continue; | |
1202 | ||
1203 | page = list_entry(area->free_list[migratetype].next, | |
1204 | struct page, lru); | |
1205 | area->nr_free--; | |
1206 | ||
1207 | try_to_steal_freepages(zone, page, start_migratetype, | |
1208 | migratetype); | |
1209 | ||
1210 | /* Remove the page from the freelists */ | |
1211 | list_del(&page->lru); | |
1212 | rmv_page_order(page); | |
1213 | ||
1214 | expand(zone, page, order, current_order, area, | |
1215 | buddy_type); | |
1216 | ||
1217 | /* | |
1218 | * The freepage_migratetype may differ from pageblock's | |
1219 | * migratetype depending on the decisions in | |
1220 | * try_to_steal_freepages(). This is OK as long as it | |
1221 | * does not differ for MIGRATE_CMA pageblocks. For CMA | |
1222 | * we need to make sure unallocated pages flushed from | |
1223 | * pcp lists are returned to the correct freelist. | |
1224 | */ | |
1225 | set_freepage_migratetype(page, buddy_type); | |
1226 | ||
1227 | trace_mm_page_alloc_extfrag(page, order, current_order, | |
1228 | start_migratetype, migratetype); | |
1229 | ||
1230 | return page; | |
1231 | } | |
1232 | } | |
1233 | ||
1234 | return NULL; | |
1235 | } | |
1236 | ||
1237 | /* | |
1238 | * Do the hard work of removing an element from the buddy allocator. | |
1239 | * Call me with the zone->lock already held. | |
1240 | */ | |
1241 | static struct page *__rmqueue(struct zone *zone, unsigned int order, | |
1242 | int migratetype) | |
1243 | { | |
1244 | struct page *page; | |
1245 | ||
1246 | retry_reserve: | |
1247 | page = __rmqueue_smallest(zone, order, migratetype); | |
1248 | ||
1249 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { | |
1250 | if (migratetype == MIGRATE_MOVABLE) | |
1251 | page = __rmqueue_cma_fallback(zone, order); | |
1252 | ||
1253 | if (!page) | |
1254 | page = __rmqueue_fallback(zone, order, migratetype); | |
1255 | ||
1256 | /* | |
1257 | * Use MIGRATE_RESERVE rather than fail an allocation. goto | |
1258 | * is used because __rmqueue_smallest is an inline function | |
1259 | * and we want just one call site | |
1260 | */ | |
1261 | if (!page) { | |
1262 | migratetype = MIGRATE_RESERVE; | |
1263 | goto retry_reserve; | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | trace_mm_page_alloc_zone_locked(page, order, migratetype); | |
1268 | return page; | |
1269 | } | |
1270 | ||
1271 | /* | |
1272 | * Obtain a specified number of elements from the buddy allocator, all under | |
1273 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
1274 | * Returns the number of new pages which were placed at *list. | |
1275 | */ | |
1276 | static int rmqueue_bulk(struct zone *zone, unsigned int order, | |
1277 | unsigned long count, struct list_head *list, | |
1278 | int migratetype, bool cold) | |
1279 | { | |
1280 | int i; | |
1281 | ||
1282 | spin_lock(&zone->lock); | |
1283 | for (i = 0; i < count; ++i) { | |
1284 | struct page *page = __rmqueue(zone, order, migratetype); | |
1285 | if (unlikely(page == NULL)) | |
1286 | break; | |
1287 | ||
1288 | /* | |
1289 | * Split buddy pages returned by expand() are received here | |
1290 | * in physical page order. The page is added to the callers and | |
1291 | * list and the list head then moves forward. From the callers | |
1292 | * perspective, the linked list is ordered by page number in | |
1293 | * some conditions. This is useful for IO devices that can | |
1294 | * merge IO requests if the physical pages are ordered | |
1295 | * properly. | |
1296 | */ | |
1297 | if (likely(!cold)) | |
1298 | list_add(&page->lru, list); | |
1299 | else | |
1300 | list_add_tail(&page->lru, list); | |
1301 | list = &page->lru; | |
1302 | if (is_migrate_cma(get_freepage_migratetype(page))) | |
1303 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, | |
1304 | -(1 << order)); | |
1305 | } | |
1306 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); | |
1307 | spin_unlock(&zone->lock); | |
1308 | return i; | |
1309 | } | |
1310 | ||
1311 | #ifdef CONFIG_NUMA | |
1312 | /* | |
1313 | * Called from the vmstat counter updater to drain pagesets of this | |
1314 | * currently executing processor on remote nodes after they have | |
1315 | * expired. | |
1316 | * | |
1317 | * Note that this function must be called with the thread pinned to | |
1318 | * a single processor. | |
1319 | */ | |
1320 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) | |
1321 | { | |
1322 | unsigned long flags; | |
1323 | int to_drain, batch; | |
1324 | ||
1325 | local_irq_save(flags); | |
1326 | batch = ACCESS_ONCE(pcp->batch); | |
1327 | to_drain = min(pcp->count, batch); | |
1328 | if (to_drain > 0) { | |
1329 | free_pcppages_bulk(zone, to_drain, pcp); | |
1330 | pcp->count -= to_drain; | |
1331 | } | |
1332 | local_irq_restore(flags); | |
1333 | } | |
1334 | #endif | |
1335 | ||
1336 | /* | |
1337 | * Drain pcplists of the indicated processor and zone. | |
1338 | * | |
1339 | * The processor must either be the current processor and the | |
1340 | * thread pinned to the current processor or a processor that | |
1341 | * is not online. | |
1342 | */ | |
1343 | static void drain_pages_zone(unsigned int cpu, struct zone *zone) | |
1344 | { | |
1345 | unsigned long flags; | |
1346 | struct per_cpu_pageset *pset; | |
1347 | struct per_cpu_pages *pcp; | |
1348 | ||
1349 | local_irq_save(flags); | |
1350 | pset = per_cpu_ptr(zone->pageset, cpu); | |
1351 | ||
1352 | pcp = &pset->pcp; | |
1353 | if (pcp->count) { | |
1354 | free_pcppages_bulk(zone, pcp->count, pcp); | |
1355 | pcp->count = 0; | |
1356 | } | |
1357 | local_irq_restore(flags); | |
1358 | } | |
1359 | ||
1360 | /* | |
1361 | * Drain pcplists of all zones on the indicated processor. | |
1362 | * | |
1363 | * The processor must either be the current processor and the | |
1364 | * thread pinned to the current processor or a processor that | |
1365 | * is not online. | |
1366 | */ | |
1367 | static void drain_pages(unsigned int cpu) | |
1368 | { | |
1369 | struct zone *zone; | |
1370 | ||
1371 | for_each_populated_zone(zone) { | |
1372 | drain_pages_zone(cpu, zone); | |
1373 | } | |
1374 | } | |
1375 | ||
1376 | /* | |
1377 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
1378 | * | |
1379 | * The CPU has to be pinned. When zone parameter is non-NULL, spill just | |
1380 | * the single zone's pages. | |
1381 | */ | |
1382 | void drain_local_pages(struct zone *zone) | |
1383 | { | |
1384 | int cpu = smp_processor_id(); | |
1385 | ||
1386 | if (zone) | |
1387 | drain_pages_zone(cpu, zone); | |
1388 | else | |
1389 | drain_pages(cpu); | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. | |
1394 | * | |
1395 | * When zone parameter is non-NULL, spill just the single zone's pages. | |
1396 | * | |
1397 | * Note that this code is protected against sending an IPI to an offline | |
1398 | * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: | |
1399 | * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but | |
1400 | * nothing keeps CPUs from showing up after we populated the cpumask and | |
1401 | * before the call to on_each_cpu_mask(). | |
1402 | */ | |
1403 | void drain_all_pages(struct zone *zone) | |
1404 | { | |
1405 | int cpu; | |
1406 | ||
1407 | /* | |
1408 | * Allocate in the BSS so we wont require allocation in | |
1409 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y | |
1410 | */ | |
1411 | static cpumask_t cpus_with_pcps; | |
1412 | ||
1413 | /* | |
1414 | * We don't care about racing with CPU hotplug event | |
1415 | * as offline notification will cause the notified | |
1416 | * cpu to drain that CPU pcps and on_each_cpu_mask | |
1417 | * disables preemption as part of its processing | |
1418 | */ | |
1419 | for_each_online_cpu(cpu) { | |
1420 | struct per_cpu_pageset *pcp; | |
1421 | struct zone *z; | |
1422 | bool has_pcps = false; | |
1423 | ||
1424 | if (zone) { | |
1425 | pcp = per_cpu_ptr(zone->pageset, cpu); | |
1426 | if (pcp->pcp.count) | |
1427 | has_pcps = true; | |
1428 | } else { | |
1429 | for_each_populated_zone(z) { | |
1430 | pcp = per_cpu_ptr(z->pageset, cpu); | |
1431 | if (pcp->pcp.count) { | |
1432 | has_pcps = true; | |
1433 | break; | |
1434 | } | |
1435 | } | |
1436 | } | |
1437 | ||
1438 | if (has_pcps) | |
1439 | cpumask_set_cpu(cpu, &cpus_with_pcps); | |
1440 | else | |
1441 | cpumask_clear_cpu(cpu, &cpus_with_pcps); | |
1442 | } | |
1443 | on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages, | |
1444 | zone, 1); | |
1445 | } | |
1446 | ||
1447 | #ifdef CONFIG_HIBERNATION | |
1448 | ||
1449 | void mark_free_pages(struct zone *zone) | |
1450 | { | |
1451 | unsigned long pfn, max_zone_pfn; | |
1452 | unsigned long flags; | |
1453 | unsigned int order, t; | |
1454 | struct list_head *curr; | |
1455 | ||
1456 | if (zone_is_empty(zone)) | |
1457 | return; | |
1458 | ||
1459 | spin_lock_irqsave(&zone->lock, flags); | |
1460 | ||
1461 | max_zone_pfn = zone_end_pfn(zone); | |
1462 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) | |
1463 | if (pfn_valid(pfn)) { | |
1464 | struct page *page = pfn_to_page(pfn); | |
1465 | ||
1466 | if (!swsusp_page_is_forbidden(page)) | |
1467 | swsusp_unset_page_free(page); | |
1468 | } | |
1469 | ||
1470 | for_each_migratetype_order(order, t) { | |
1471 | list_for_each(curr, &zone->free_area[order].free_list[t]) { | |
1472 | unsigned long i; | |
1473 | ||
1474 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); | |
1475 | for (i = 0; i < (1UL << order); i++) | |
1476 | swsusp_set_page_free(pfn_to_page(pfn + i)); | |
1477 | } | |
1478 | } | |
1479 | spin_unlock_irqrestore(&zone->lock, flags); | |
1480 | } | |
1481 | #endif /* CONFIG_PM */ | |
1482 | ||
1483 | /* | |
1484 | * Free a 0-order page | |
1485 | * cold == true ? free a cold page : free a hot page | |
1486 | */ | |
1487 | void free_hot_cold_page(struct page *page, bool cold) | |
1488 | { | |
1489 | struct zone *zone = page_zone(page); | |
1490 | struct per_cpu_pages *pcp; | |
1491 | unsigned long flags; | |
1492 | unsigned long pfn = page_to_pfn(page); | |
1493 | int migratetype; | |
1494 | ||
1495 | if (!free_pages_prepare(page, 0)) | |
1496 | return; | |
1497 | ||
1498 | migratetype = get_pfnblock_migratetype(page, pfn); | |
1499 | set_freepage_migratetype(page, migratetype); | |
1500 | local_irq_save(flags); | |
1501 | __count_vm_event(PGFREE); | |
1502 | ||
1503 | /* | |
1504 | * We only track unmovable, reclaimable and movable on pcp lists. | |
1505 | * Free ISOLATE pages back to the allocator because they are being | |
1506 | * offlined but treat RESERVE as movable pages so we can get those | |
1507 | * areas back if necessary. Otherwise, we may have to free | |
1508 | * excessively into the page allocator | |
1509 | */ | |
1510 | if (migratetype >= MIGRATE_PCPTYPES) { | |
1511 | if (unlikely(is_migrate_isolate(migratetype))) { | |
1512 | free_one_page(zone, page, pfn, 0, migratetype); | |
1513 | goto out; | |
1514 | } | |
1515 | migratetype = MIGRATE_MOVABLE; | |
1516 | } | |
1517 | ||
1518 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1519 | if (!cold) | |
1520 | list_add(&page->lru, &pcp->lists[migratetype]); | |
1521 | else | |
1522 | list_add_tail(&page->lru, &pcp->lists[migratetype]); | |
1523 | pcp->count++; | |
1524 | if (pcp->count >= pcp->high) { | |
1525 | unsigned long batch = ACCESS_ONCE(pcp->batch); | |
1526 | free_pcppages_bulk(zone, batch, pcp); | |
1527 | pcp->count -= batch; | |
1528 | } | |
1529 | ||
1530 | out: | |
1531 | local_irq_restore(flags); | |
1532 | } | |
1533 | ||
1534 | /* | |
1535 | * Free a list of 0-order pages | |
1536 | */ | |
1537 | void free_hot_cold_page_list(struct list_head *list, bool cold) | |
1538 | { | |
1539 | struct page *page, *next; | |
1540 | ||
1541 | list_for_each_entry_safe(page, next, list, lru) { | |
1542 | trace_mm_page_free_batched(page, cold); | |
1543 | free_hot_cold_page(page, cold); | |
1544 | } | |
1545 | } | |
1546 | ||
1547 | /* | |
1548 | * split_page takes a non-compound higher-order page, and splits it into | |
1549 | * n (1<<order) sub-pages: page[0..n] | |
1550 | * Each sub-page must be freed individually. | |
1551 | * | |
1552 | * Note: this is probably too low level an operation for use in drivers. | |
1553 | * Please consult with lkml before using this in your driver. | |
1554 | */ | |
1555 | void split_page(struct page *page, unsigned int order) | |
1556 | { | |
1557 | int i; | |
1558 | ||
1559 | VM_BUG_ON_PAGE(PageCompound(page), page); | |
1560 | VM_BUG_ON_PAGE(!page_count(page), page); | |
1561 | ||
1562 | #ifdef CONFIG_KMEMCHECK | |
1563 | /* | |
1564 | * Split shadow pages too, because free(page[0]) would | |
1565 | * otherwise free the whole shadow. | |
1566 | */ | |
1567 | if (kmemcheck_page_is_tracked(page)) | |
1568 | split_page(virt_to_page(page[0].shadow), order); | |
1569 | #endif | |
1570 | ||
1571 | set_page_owner(page, 0, 0); | |
1572 | for (i = 1; i < (1 << order); i++) { | |
1573 | set_page_refcounted(page + i); | |
1574 | set_page_owner(page + i, 0, 0); | |
1575 | } | |
1576 | } | |
1577 | EXPORT_SYMBOL_GPL(split_page); | |
1578 | ||
1579 | int __isolate_free_page(struct page *page, unsigned int order) | |
1580 | { | |
1581 | unsigned long watermark; | |
1582 | struct zone *zone; | |
1583 | int mt; | |
1584 | ||
1585 | BUG_ON(!PageBuddy(page)); | |
1586 | ||
1587 | zone = page_zone(page); | |
1588 | mt = get_pageblock_migratetype(page); | |
1589 | ||
1590 | if (!is_migrate_isolate(mt)) { | |
1591 | /* Obey watermarks as if the page was being allocated */ | |
1592 | watermark = low_wmark_pages(zone) + (1 << order); | |
1593 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
1594 | return 0; | |
1595 | ||
1596 | __mod_zone_freepage_state(zone, -(1UL << order), mt); | |
1597 | } | |
1598 | ||
1599 | /* Remove page from free list */ | |
1600 | list_del(&page->lru); | |
1601 | zone->free_area[order].nr_free--; | |
1602 | rmv_page_order(page); | |
1603 | ||
1604 | /* Set the pageblock if the isolated page is at least a pageblock */ | |
1605 | if (order >= pageblock_order - 1) { | |
1606 | struct page *endpage = page + (1 << order) - 1; | |
1607 | for (; page < endpage; page += pageblock_nr_pages) { | |
1608 | int mt = get_pageblock_migratetype(page); | |
1609 | if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) | |
1610 | set_pageblock_migratetype(page, | |
1611 | MIGRATE_MOVABLE); | |
1612 | } | |
1613 | } | |
1614 | ||
1615 | set_page_owner(page, order, 0); | |
1616 | return 1UL << order; | |
1617 | } | |
1618 | ||
1619 | /* | |
1620 | * Similar to split_page except the page is already free. As this is only | |
1621 | * being used for migration, the migratetype of the block also changes. | |
1622 | * As this is called with interrupts disabled, the caller is responsible | |
1623 | * for calling arch_alloc_page() and kernel_map_page() after interrupts | |
1624 | * are enabled. | |
1625 | * | |
1626 | * Note: this is probably too low level an operation for use in drivers. | |
1627 | * Please consult with lkml before using this in your driver. | |
1628 | */ | |
1629 | int split_free_page(struct page *page) | |
1630 | { | |
1631 | unsigned int order; | |
1632 | int nr_pages; | |
1633 | ||
1634 | order = page_order(page); | |
1635 | ||
1636 | nr_pages = __isolate_free_page(page, order); | |
1637 | if (!nr_pages) | |
1638 | return 0; | |
1639 | ||
1640 | /* Split into individual pages */ | |
1641 | set_page_refcounted(page); | |
1642 | split_page(page, order); | |
1643 | return nr_pages; | |
1644 | } | |
1645 | ||
1646 | /* | |
1647 | * Allocate a page from the given zone. Use pcplists for order-0 allocations. | |
1648 | */ | |
1649 | static inline | |
1650 | struct page *buffered_rmqueue(struct zone *preferred_zone, | |
1651 | struct zone *zone, unsigned int order, | |
1652 | gfp_t gfp_flags, int migratetype) | |
1653 | { | |
1654 | unsigned long flags; | |
1655 | struct page *page; | |
1656 | bool cold = ((gfp_flags & __GFP_COLD) != 0); | |
1657 | ||
1658 | if (likely(order == 0)) { | |
1659 | struct per_cpu_pages *pcp; | |
1660 | struct list_head *list; | |
1661 | ||
1662 | local_irq_save(flags); | |
1663 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1664 | list = &pcp->lists[migratetype]; | |
1665 | if (list_empty(list)) { | |
1666 | pcp->count += rmqueue_bulk(zone, 0, | |
1667 | pcp->batch, list, | |
1668 | migratetype, cold); | |
1669 | if (unlikely(list_empty(list))) | |
1670 | goto failed; | |
1671 | } | |
1672 | ||
1673 | if (cold) | |
1674 | page = list_entry(list->prev, struct page, lru); | |
1675 | else | |
1676 | page = list_entry(list->next, struct page, lru); | |
1677 | ||
1678 | list_del(&page->lru); | |
1679 | pcp->count--; | |
1680 | } else { | |
1681 | if (unlikely(gfp_flags & __GFP_NOFAIL)) { | |
1682 | /* | |
1683 | * __GFP_NOFAIL is not to be used in new code. | |
1684 | * | |
1685 | * All __GFP_NOFAIL callers should be fixed so that they | |
1686 | * properly detect and handle allocation failures. | |
1687 | * | |
1688 | * We most definitely don't want callers attempting to | |
1689 | * allocate greater than order-1 page units with | |
1690 | * __GFP_NOFAIL. | |
1691 | */ | |
1692 | WARN_ON_ONCE(order > 1); | |
1693 | } | |
1694 | spin_lock_irqsave(&zone->lock, flags); | |
1695 | page = __rmqueue(zone, order, migratetype); | |
1696 | spin_unlock(&zone->lock); | |
1697 | if (!page) | |
1698 | goto failed; | |
1699 | __mod_zone_freepage_state(zone, -(1 << order), | |
1700 | get_freepage_migratetype(page)); | |
1701 | } | |
1702 | ||
1703 | __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order)); | |
1704 | if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 && | |
1705 | !test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) | |
1706 | set_bit(ZONE_FAIR_DEPLETED, &zone->flags); | |
1707 | ||
1708 | __count_zone_vm_events(PGALLOC, zone, 1 << order); | |
1709 | zone_statistics(preferred_zone, zone, gfp_flags); | |
1710 | local_irq_restore(flags); | |
1711 | ||
1712 | VM_BUG_ON_PAGE(bad_range(zone, page), page); | |
1713 | return page; | |
1714 | ||
1715 | failed: | |
1716 | local_irq_restore(flags); | |
1717 | return NULL; | |
1718 | } | |
1719 | ||
1720 | #ifdef CONFIG_FAIL_PAGE_ALLOC | |
1721 | ||
1722 | static struct { | |
1723 | struct fault_attr attr; | |
1724 | ||
1725 | u32 ignore_gfp_highmem; | |
1726 | u32 ignore_gfp_wait; | |
1727 | u32 min_order; | |
1728 | } fail_page_alloc = { | |
1729 | .attr = FAULT_ATTR_INITIALIZER, | |
1730 | .ignore_gfp_wait = 1, | |
1731 | .ignore_gfp_highmem = 1, | |
1732 | .min_order = 1, | |
1733 | }; | |
1734 | ||
1735 | static int __init setup_fail_page_alloc(char *str) | |
1736 | { | |
1737 | return setup_fault_attr(&fail_page_alloc.attr, str); | |
1738 | } | |
1739 | __setup("fail_page_alloc=", setup_fail_page_alloc); | |
1740 | ||
1741 | static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1742 | { | |
1743 | if (order < fail_page_alloc.min_order) | |
1744 | return false; | |
1745 | if (gfp_mask & __GFP_NOFAIL) | |
1746 | return false; | |
1747 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) | |
1748 | return false; | |
1749 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) | |
1750 | return false; | |
1751 | ||
1752 | return should_fail(&fail_page_alloc.attr, 1 << order); | |
1753 | } | |
1754 | ||
1755 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
1756 | ||
1757 | static int __init fail_page_alloc_debugfs(void) | |
1758 | { | |
1759 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
1760 | struct dentry *dir; | |
1761 | ||
1762 | dir = fault_create_debugfs_attr("fail_page_alloc", NULL, | |
1763 | &fail_page_alloc.attr); | |
1764 | if (IS_ERR(dir)) | |
1765 | return PTR_ERR(dir); | |
1766 | ||
1767 | if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
1768 | &fail_page_alloc.ignore_gfp_wait)) | |
1769 | goto fail; | |
1770 | if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, | |
1771 | &fail_page_alloc.ignore_gfp_highmem)) | |
1772 | goto fail; | |
1773 | if (!debugfs_create_u32("min-order", mode, dir, | |
1774 | &fail_page_alloc.min_order)) | |
1775 | goto fail; | |
1776 | ||
1777 | return 0; | |
1778 | fail: | |
1779 | debugfs_remove_recursive(dir); | |
1780 | ||
1781 | return -ENOMEM; | |
1782 | } | |
1783 | ||
1784 | late_initcall(fail_page_alloc_debugfs); | |
1785 | ||
1786 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
1787 | ||
1788 | #else /* CONFIG_FAIL_PAGE_ALLOC */ | |
1789 | ||
1790 | static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1791 | { | |
1792 | return false; | |
1793 | } | |
1794 | ||
1795 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ | |
1796 | ||
1797 | /* | |
1798 | * Return true if free pages are above 'mark'. This takes into account the order | |
1799 | * of the allocation. | |
1800 | */ | |
1801 | static bool __zone_watermark_ok(struct zone *z, unsigned int order, | |
1802 | unsigned long mark, int classzone_idx, int alloc_flags, | |
1803 | long free_pages) | |
1804 | { | |
1805 | /* free_pages may go negative - that's OK */ | |
1806 | long min = mark; | |
1807 | int o; | |
1808 | long free_cma = 0; | |
1809 | ||
1810 | free_pages -= (1 << order) - 1; | |
1811 | if (alloc_flags & ALLOC_HIGH) | |
1812 | min -= min / 2; | |
1813 | if (alloc_flags & ALLOC_HARDER) | |
1814 | min -= min / 4; | |
1815 | #ifdef CONFIG_CMA | |
1816 | /* If allocation can't use CMA areas don't use free CMA pages */ | |
1817 | if (!(alloc_flags & ALLOC_CMA)) | |
1818 | free_cma = zone_page_state(z, NR_FREE_CMA_PAGES); | |
1819 | #endif | |
1820 | ||
1821 | if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx]) | |
1822 | return false; | |
1823 | for (o = 0; o < order; o++) { | |
1824 | /* At the next order, this order's pages become unavailable */ | |
1825 | free_pages -= z->free_area[o].nr_free << o; | |
1826 | ||
1827 | /* Require fewer higher order pages to be free */ | |
1828 | min >>= 1; | |
1829 | ||
1830 | if (free_pages <= min) | |
1831 | return false; | |
1832 | } | |
1833 | return true; | |
1834 | } | |
1835 | ||
1836 | bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, | |
1837 | int classzone_idx, int alloc_flags) | |
1838 | { | |
1839 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1840 | zone_page_state(z, NR_FREE_PAGES)); | |
1841 | } | |
1842 | ||
1843 | bool zone_watermark_ok_safe(struct zone *z, unsigned int order, | |
1844 | unsigned long mark, int classzone_idx, int alloc_flags) | |
1845 | { | |
1846 | long free_pages = zone_page_state(z, NR_FREE_PAGES); | |
1847 | ||
1848 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) | |
1849 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); | |
1850 | ||
1851 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1852 | free_pages); | |
1853 | } | |
1854 | ||
1855 | #ifdef CONFIG_NUMA | |
1856 | /* | |
1857 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to | |
1858 | * skip over zones that are not allowed by the cpuset, or that have | |
1859 | * been recently (in last second) found to be nearly full. See further | |
1860 | * comments in mmzone.h. Reduces cache footprint of zonelist scans | |
1861 | * that have to skip over a lot of full or unallowed zones. | |
1862 | * | |
1863 | * If the zonelist cache is present in the passed zonelist, then | |
1864 | * returns a pointer to the allowed node mask (either the current | |
1865 | * tasks mems_allowed, or node_states[N_MEMORY].) | |
1866 | * | |
1867 | * If the zonelist cache is not available for this zonelist, does | |
1868 | * nothing and returns NULL. | |
1869 | * | |
1870 | * If the fullzones BITMAP in the zonelist cache is stale (more than | |
1871 | * a second since last zap'd) then we zap it out (clear its bits.) | |
1872 | * | |
1873 | * We hold off even calling zlc_setup, until after we've checked the | |
1874 | * first zone in the zonelist, on the theory that most allocations will | |
1875 | * be satisfied from that first zone, so best to examine that zone as | |
1876 | * quickly as we can. | |
1877 | */ | |
1878 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1879 | { | |
1880 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1881 | nodemask_t *allowednodes; /* zonelist_cache approximation */ | |
1882 | ||
1883 | zlc = zonelist->zlcache_ptr; | |
1884 | if (!zlc) | |
1885 | return NULL; | |
1886 | ||
1887 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { | |
1888 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1889 | zlc->last_full_zap = jiffies; | |
1890 | } | |
1891 | ||
1892 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? | |
1893 | &cpuset_current_mems_allowed : | |
1894 | &node_states[N_MEMORY]; | |
1895 | return allowednodes; | |
1896 | } | |
1897 | ||
1898 | /* | |
1899 | * Given 'z' scanning a zonelist, run a couple of quick checks to see | |
1900 | * if it is worth looking at further for free memory: | |
1901 | * 1) Check that the zone isn't thought to be full (doesn't have its | |
1902 | * bit set in the zonelist_cache fullzones BITMAP). | |
1903 | * 2) Check that the zones node (obtained from the zonelist_cache | |
1904 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. | |
1905 | * Return true (non-zero) if zone is worth looking at further, or | |
1906 | * else return false (zero) if it is not. | |
1907 | * | |
1908 | * This check -ignores- the distinction between various watermarks, | |
1909 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is | |
1910 | * found to be full for any variation of these watermarks, it will | |
1911 | * be considered full for up to one second by all requests, unless | |
1912 | * we are so low on memory on all allowed nodes that we are forced | |
1913 | * into the second scan of the zonelist. | |
1914 | * | |
1915 | * In the second scan we ignore this zonelist cache and exactly | |
1916 | * apply the watermarks to all zones, even it is slower to do so. | |
1917 | * We are low on memory in the second scan, and should leave no stone | |
1918 | * unturned looking for a free page. | |
1919 | */ | |
1920 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1921 | nodemask_t *allowednodes) | |
1922 | { | |
1923 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1924 | int i; /* index of *z in zonelist zones */ | |
1925 | int n; /* node that zone *z is on */ | |
1926 | ||
1927 | zlc = zonelist->zlcache_ptr; | |
1928 | if (!zlc) | |
1929 | return 1; | |
1930 | ||
1931 | i = z - zonelist->_zonerefs; | |
1932 | n = zlc->z_to_n[i]; | |
1933 | ||
1934 | /* This zone is worth trying if it is allowed but not full */ | |
1935 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); | |
1936 | } | |
1937 | ||
1938 | /* | |
1939 | * Given 'z' scanning a zonelist, set the corresponding bit in | |
1940 | * zlc->fullzones, so that subsequent attempts to allocate a page | |
1941 | * from that zone don't waste time re-examining it. | |
1942 | */ | |
1943 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1944 | { | |
1945 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1946 | int i; /* index of *z in zonelist zones */ | |
1947 | ||
1948 | zlc = zonelist->zlcache_ptr; | |
1949 | if (!zlc) | |
1950 | return; | |
1951 | ||
1952 | i = z - zonelist->_zonerefs; | |
1953 | ||
1954 | set_bit(i, zlc->fullzones); | |
1955 | } | |
1956 | ||
1957 | /* | |
1958 | * clear all zones full, called after direct reclaim makes progress so that | |
1959 | * a zone that was recently full is not skipped over for up to a second | |
1960 | */ | |
1961 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
1962 | { | |
1963 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1964 | ||
1965 | zlc = zonelist->zlcache_ptr; | |
1966 | if (!zlc) | |
1967 | return; | |
1968 | ||
1969 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1970 | } | |
1971 | ||
1972 | static bool zone_local(struct zone *local_zone, struct zone *zone) | |
1973 | { | |
1974 | return local_zone->node == zone->node; | |
1975 | } | |
1976 | ||
1977 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) | |
1978 | { | |
1979 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) < | |
1980 | RECLAIM_DISTANCE; | |
1981 | } | |
1982 | ||
1983 | #else /* CONFIG_NUMA */ | |
1984 | ||
1985 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1986 | { | |
1987 | return NULL; | |
1988 | } | |
1989 | ||
1990 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1991 | nodemask_t *allowednodes) | |
1992 | { | |
1993 | return 1; | |
1994 | } | |
1995 | ||
1996 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1997 | { | |
1998 | } | |
1999 | ||
2000 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
2001 | { | |
2002 | } | |
2003 | ||
2004 | static bool zone_local(struct zone *local_zone, struct zone *zone) | |
2005 | { | |
2006 | return true; | |
2007 | } | |
2008 | ||
2009 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) | |
2010 | { | |
2011 | return true; | |
2012 | } | |
2013 | ||
2014 | #endif /* CONFIG_NUMA */ | |
2015 | ||
2016 | static void reset_alloc_batches(struct zone *preferred_zone) | |
2017 | { | |
2018 | struct zone *zone = preferred_zone->zone_pgdat->node_zones; | |
2019 | ||
2020 | do { | |
2021 | mod_zone_page_state(zone, NR_ALLOC_BATCH, | |
2022 | high_wmark_pages(zone) - low_wmark_pages(zone) - | |
2023 | atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); | |
2024 | clear_bit(ZONE_FAIR_DEPLETED, &zone->flags); | |
2025 | } while (zone++ != preferred_zone); | |
2026 | } | |
2027 | ||
2028 | /* | |
2029 | * get_page_from_freelist goes through the zonelist trying to allocate | |
2030 | * a page. | |
2031 | */ | |
2032 | static struct page * | |
2033 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, | |
2034 | const struct alloc_context *ac) | |
2035 | { | |
2036 | struct zonelist *zonelist = ac->zonelist; | |
2037 | struct zoneref *z; | |
2038 | struct page *page = NULL; | |
2039 | struct zone *zone; | |
2040 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ | |
2041 | int zlc_active = 0; /* set if using zonelist_cache */ | |
2042 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ | |
2043 | bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) && | |
2044 | (gfp_mask & __GFP_WRITE); | |
2045 | int nr_fair_skipped = 0; | |
2046 | bool zonelist_rescan; | |
2047 | ||
2048 | zonelist_scan: | |
2049 | zonelist_rescan = false; | |
2050 | ||
2051 | /* | |
2052 | * Scan zonelist, looking for a zone with enough free. | |
2053 | * See also __cpuset_node_allowed() comment in kernel/cpuset.c. | |
2054 | */ | |
2055 | for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx, | |
2056 | ac->nodemask) { | |
2057 | unsigned long mark; | |
2058 | ||
2059 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && | |
2060 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
2061 | continue; | |
2062 | if (cpusets_enabled() && | |
2063 | (alloc_flags & ALLOC_CPUSET) && | |
2064 | !cpuset_zone_allowed(zone, gfp_mask)) | |
2065 | continue; | |
2066 | /* | |
2067 | * Distribute pages in proportion to the individual | |
2068 | * zone size to ensure fair page aging. The zone a | |
2069 | * page was allocated in should have no effect on the | |
2070 | * time the page has in memory before being reclaimed. | |
2071 | */ | |
2072 | if (alloc_flags & ALLOC_FAIR) { | |
2073 | if (!zone_local(ac->preferred_zone, zone)) | |
2074 | break; | |
2075 | if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) { | |
2076 | nr_fair_skipped++; | |
2077 | continue; | |
2078 | } | |
2079 | } | |
2080 | /* | |
2081 | * When allocating a page cache page for writing, we | |
2082 | * want to get it from a zone that is within its dirty | |
2083 | * limit, such that no single zone holds more than its | |
2084 | * proportional share of globally allowed dirty pages. | |
2085 | * The dirty limits take into account the zone's | |
2086 | * lowmem reserves and high watermark so that kswapd | |
2087 | * should be able to balance it without having to | |
2088 | * write pages from its LRU list. | |
2089 | * | |
2090 | * This may look like it could increase pressure on | |
2091 | * lower zones by failing allocations in higher zones | |
2092 | * before they are full. But the pages that do spill | |
2093 | * over are limited as the lower zones are protected | |
2094 | * by this very same mechanism. It should not become | |
2095 | * a practical burden to them. | |
2096 | * | |
2097 | * XXX: For now, allow allocations to potentially | |
2098 | * exceed the per-zone dirty limit in the slowpath | |
2099 | * (ALLOC_WMARK_LOW unset) before going into reclaim, | |
2100 | * which is important when on a NUMA setup the allowed | |
2101 | * zones are together not big enough to reach the | |
2102 | * global limit. The proper fix for these situations | |
2103 | * will require awareness of zones in the | |
2104 | * dirty-throttling and the flusher threads. | |
2105 | */ | |
2106 | if (consider_zone_dirty && !zone_dirty_ok(zone)) | |
2107 | continue; | |
2108 | ||
2109 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; | |
2110 | if (!zone_watermark_ok(zone, order, mark, | |
2111 | ac->classzone_idx, alloc_flags)) { | |
2112 | int ret; | |
2113 | ||
2114 | /* Checked here to keep the fast path fast */ | |
2115 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
2116 | if (alloc_flags & ALLOC_NO_WATERMARKS) | |
2117 | goto try_this_zone; | |
2118 | ||
2119 | if (IS_ENABLED(CONFIG_NUMA) && | |
2120 | !did_zlc_setup && nr_online_nodes > 1) { | |
2121 | /* | |
2122 | * we do zlc_setup if there are multiple nodes | |
2123 | * and before considering the first zone allowed | |
2124 | * by the cpuset. | |
2125 | */ | |
2126 | allowednodes = zlc_setup(zonelist, alloc_flags); | |
2127 | zlc_active = 1; | |
2128 | did_zlc_setup = 1; | |
2129 | } | |
2130 | ||
2131 | if (zone_reclaim_mode == 0 || | |
2132 | !zone_allows_reclaim(ac->preferred_zone, zone)) | |
2133 | goto this_zone_full; | |
2134 | ||
2135 | /* | |
2136 | * As we may have just activated ZLC, check if the first | |
2137 | * eligible zone has failed zone_reclaim recently. | |
2138 | */ | |
2139 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && | |
2140 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
2141 | continue; | |
2142 | ||
2143 | ret = zone_reclaim(zone, gfp_mask, order); | |
2144 | switch (ret) { | |
2145 | case ZONE_RECLAIM_NOSCAN: | |
2146 | /* did not scan */ | |
2147 | continue; | |
2148 | case ZONE_RECLAIM_FULL: | |
2149 | /* scanned but unreclaimable */ | |
2150 | continue; | |
2151 | default: | |
2152 | /* did we reclaim enough */ | |
2153 | if (zone_watermark_ok(zone, order, mark, | |
2154 | ac->classzone_idx, alloc_flags)) | |
2155 | goto try_this_zone; | |
2156 | ||
2157 | /* | |
2158 | * Failed to reclaim enough to meet watermark. | |
2159 | * Only mark the zone full if checking the min | |
2160 | * watermark or if we failed to reclaim just | |
2161 | * 1<<order pages or else the page allocator | |
2162 | * fastpath will prematurely mark zones full | |
2163 | * when the watermark is between the low and | |
2164 | * min watermarks. | |
2165 | */ | |
2166 | if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) || | |
2167 | ret == ZONE_RECLAIM_SOME) | |
2168 | goto this_zone_full; | |
2169 | ||
2170 | continue; | |
2171 | } | |
2172 | } | |
2173 | ||
2174 | try_this_zone: | |
2175 | page = buffered_rmqueue(ac->preferred_zone, zone, order, | |
2176 | gfp_mask, ac->migratetype); | |
2177 | if (page) { | |
2178 | if (prep_new_page(page, order, gfp_mask, alloc_flags)) | |
2179 | goto try_this_zone; | |
2180 | return page; | |
2181 | } | |
2182 | this_zone_full: | |
2183 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active) | |
2184 | zlc_mark_zone_full(zonelist, z); | |
2185 | } | |
2186 | ||
2187 | /* | |
2188 | * The first pass makes sure allocations are spread fairly within the | |
2189 | * local node. However, the local node might have free pages left | |
2190 | * after the fairness batches are exhausted, and remote zones haven't | |
2191 | * even been considered yet. Try once more without fairness, and | |
2192 | * include remote zones now, before entering the slowpath and waking | |
2193 | * kswapd: prefer spilling to a remote zone over swapping locally. | |
2194 | */ | |
2195 | if (alloc_flags & ALLOC_FAIR) { | |
2196 | alloc_flags &= ~ALLOC_FAIR; | |
2197 | if (nr_fair_skipped) { | |
2198 | zonelist_rescan = true; | |
2199 | reset_alloc_batches(ac->preferred_zone); | |
2200 | } | |
2201 | if (nr_online_nodes > 1) | |
2202 | zonelist_rescan = true; | |
2203 | } | |
2204 | ||
2205 | if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) { | |
2206 | /* Disable zlc cache for second zonelist scan */ | |
2207 | zlc_active = 0; | |
2208 | zonelist_rescan = true; | |
2209 | } | |
2210 | ||
2211 | if (zonelist_rescan) | |
2212 | goto zonelist_scan; | |
2213 | ||
2214 | return NULL; | |
2215 | } | |
2216 | ||
2217 | /* | |
2218 | * Large machines with many possible nodes should not always dump per-node | |
2219 | * meminfo in irq context. | |
2220 | */ | |
2221 | static inline bool should_suppress_show_mem(void) | |
2222 | { | |
2223 | bool ret = false; | |
2224 | ||
2225 | #if NODES_SHIFT > 8 | |
2226 | ret = in_interrupt(); | |
2227 | #endif | |
2228 | return ret; | |
2229 | } | |
2230 | ||
2231 | static DEFINE_RATELIMIT_STATE(nopage_rs, | |
2232 | DEFAULT_RATELIMIT_INTERVAL, | |
2233 | DEFAULT_RATELIMIT_BURST); | |
2234 | ||
2235 | void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) | |
2236 | { | |
2237 | unsigned int filter = SHOW_MEM_FILTER_NODES; | |
2238 | ||
2239 | if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || | |
2240 | debug_guardpage_minorder() > 0) | |
2241 | return; | |
2242 | ||
2243 | /* | |
2244 | * This documents exceptions given to allocations in certain | |
2245 | * contexts that are allowed to allocate outside current's set | |
2246 | * of allowed nodes. | |
2247 | */ | |
2248 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
2249 | if (test_thread_flag(TIF_MEMDIE) || | |
2250 | (current->flags & (PF_MEMALLOC | PF_EXITING))) | |
2251 | filter &= ~SHOW_MEM_FILTER_NODES; | |
2252 | if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) | |
2253 | filter &= ~SHOW_MEM_FILTER_NODES; | |
2254 | ||
2255 | if (fmt) { | |
2256 | struct va_format vaf; | |
2257 | va_list args; | |
2258 | ||
2259 | va_start(args, fmt); | |
2260 | ||
2261 | vaf.fmt = fmt; | |
2262 | vaf.va = &args; | |
2263 | ||
2264 | pr_warn("%pV", &vaf); | |
2265 | ||
2266 | va_end(args); | |
2267 | } | |
2268 | ||
2269 | pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", | |
2270 | current->comm, order, gfp_mask); | |
2271 | ||
2272 | dump_stack(); | |
2273 | if (!should_suppress_show_mem()) | |
2274 | show_mem(filter); | |
2275 | } | |
2276 | ||
2277 | static inline int | |
2278 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, | |
2279 | unsigned long did_some_progress, | |
2280 | unsigned long pages_reclaimed) | |
2281 | { | |
2282 | /* Do not loop if specifically requested */ | |
2283 | if (gfp_mask & __GFP_NORETRY) | |
2284 | return 0; | |
2285 | ||
2286 | /* Always retry if specifically requested */ | |
2287 | if (gfp_mask & __GFP_NOFAIL) | |
2288 | return 1; | |
2289 | ||
2290 | /* | |
2291 | * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim | |
2292 | * making forward progress without invoking OOM. Suspend also disables | |
2293 | * storage devices so kswapd will not help. Bail if we are suspending. | |
2294 | */ | |
2295 | if (!did_some_progress && pm_suspended_storage()) | |
2296 | return 0; | |
2297 | ||
2298 | /* | |
2299 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER | |
2300 | * means __GFP_NOFAIL, but that may not be true in other | |
2301 | * implementations. | |
2302 | */ | |
2303 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
2304 | return 1; | |
2305 | ||
2306 | /* | |
2307 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is | |
2308 | * specified, then we retry until we no longer reclaim any pages | |
2309 | * (above), or we've reclaimed an order of pages at least as | |
2310 | * large as the allocation's order. In both cases, if the | |
2311 | * allocation still fails, we stop retrying. | |
2312 | */ | |
2313 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) | |
2314 | return 1; | |
2315 | ||
2316 | return 0; | |
2317 | } | |
2318 | ||
2319 | static inline struct page * | |
2320 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
2321 | const struct alloc_context *ac, unsigned long *did_some_progress) | |
2322 | { | |
2323 | struct page *page; | |
2324 | ||
2325 | *did_some_progress = 0; | |
2326 | ||
2327 | /* | |
2328 | * Acquire the per-zone oom lock for each zone. If that | |
2329 | * fails, somebody else is making progress for us. | |
2330 | */ | |
2331 | if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) { | |
2332 | *did_some_progress = 1; | |
2333 | schedule_timeout_uninterruptible(1); | |
2334 | return NULL; | |
2335 | } | |
2336 | ||
2337 | /* | |
2338 | * Go through the zonelist yet one more time, keep very high watermark | |
2339 | * here, this is only to catch a parallel oom killing, we must fail if | |
2340 | * we're still under heavy pressure. | |
2341 | */ | |
2342 | page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order, | |
2343 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); | |
2344 | if (page) | |
2345 | goto out; | |
2346 | ||
2347 | if (!(gfp_mask & __GFP_NOFAIL)) { | |
2348 | /* Coredumps can quickly deplete all memory reserves */ | |
2349 | if (current->flags & PF_DUMPCORE) | |
2350 | goto out; | |
2351 | /* The OOM killer will not help higher order allocs */ | |
2352 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
2353 | goto out; | |
2354 | /* The OOM killer does not needlessly kill tasks for lowmem */ | |
2355 | if (ac->high_zoneidx < ZONE_NORMAL) | |
2356 | goto out; | |
2357 | /* The OOM killer does not compensate for light reclaim */ | |
2358 | if (!(gfp_mask & __GFP_FS)) { | |
2359 | /* | |
2360 | * XXX: Page reclaim didn't yield anything, | |
2361 | * and the OOM killer can't be invoked, but | |
2362 | * keep looping as per should_alloc_retry(). | |
2363 | */ | |
2364 | *did_some_progress = 1; | |
2365 | goto out; | |
2366 | } | |
2367 | /* | |
2368 | * GFP_THISNODE contains __GFP_NORETRY and we never hit this. | |
2369 | * Sanity check for bare calls of __GFP_THISNODE, not real OOM. | |
2370 | * The caller should handle page allocation failure by itself if | |
2371 | * it specifies __GFP_THISNODE. | |
2372 | * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. | |
2373 | */ | |
2374 | if (gfp_mask & __GFP_THISNODE) | |
2375 | goto out; | |
2376 | } | |
2377 | /* Exhausted what can be done so it's blamo time */ | |
2378 | if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false) | |
2379 | || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) | |
2380 | *did_some_progress = 1; | |
2381 | out: | |
2382 | oom_zonelist_unlock(ac->zonelist, gfp_mask); | |
2383 | return page; | |
2384 | } | |
2385 | ||
2386 | #ifdef CONFIG_COMPACTION | |
2387 | /* Try memory compaction for high-order allocations before reclaim */ | |
2388 | static struct page * | |
2389 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2390 | int alloc_flags, const struct alloc_context *ac, | |
2391 | enum migrate_mode mode, int *contended_compaction, | |
2392 | bool *deferred_compaction) | |
2393 | { | |
2394 | unsigned long compact_result; | |
2395 | struct page *page; | |
2396 | ||
2397 | if (!order) | |
2398 | return NULL; | |
2399 | ||
2400 | current->flags |= PF_MEMALLOC; | |
2401 | compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, | |
2402 | mode, contended_compaction); | |
2403 | current->flags &= ~PF_MEMALLOC; | |
2404 | ||
2405 | switch (compact_result) { | |
2406 | case COMPACT_DEFERRED: | |
2407 | *deferred_compaction = true; | |
2408 | /* fall-through */ | |
2409 | case COMPACT_SKIPPED: | |
2410 | return NULL; | |
2411 | default: | |
2412 | break; | |
2413 | } | |
2414 | ||
2415 | /* | |
2416 | * At least in one zone compaction wasn't deferred or skipped, so let's | |
2417 | * count a compaction stall | |
2418 | */ | |
2419 | count_vm_event(COMPACTSTALL); | |
2420 | ||
2421 | page = get_page_from_freelist(gfp_mask, order, | |
2422 | alloc_flags & ~ALLOC_NO_WATERMARKS, ac); | |
2423 | ||
2424 | if (page) { | |
2425 | struct zone *zone = page_zone(page); | |
2426 | ||
2427 | zone->compact_blockskip_flush = false; | |
2428 | compaction_defer_reset(zone, order, true); | |
2429 | count_vm_event(COMPACTSUCCESS); | |
2430 | return page; | |
2431 | } | |
2432 | ||
2433 | /* | |
2434 | * It's bad if compaction run occurs and fails. The most likely reason | |
2435 | * is that pages exist, but not enough to satisfy watermarks. | |
2436 | */ | |
2437 | count_vm_event(COMPACTFAIL); | |
2438 | ||
2439 | cond_resched(); | |
2440 | ||
2441 | return NULL; | |
2442 | } | |
2443 | #else | |
2444 | static inline struct page * | |
2445 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2446 | int alloc_flags, const struct alloc_context *ac, | |
2447 | enum migrate_mode mode, int *contended_compaction, | |
2448 | bool *deferred_compaction) | |
2449 | { | |
2450 | return NULL; | |
2451 | } | |
2452 | #endif /* CONFIG_COMPACTION */ | |
2453 | ||
2454 | /* Perform direct synchronous page reclaim */ | |
2455 | static int | |
2456 | __perform_reclaim(gfp_t gfp_mask, unsigned int order, | |
2457 | const struct alloc_context *ac) | |
2458 | { | |
2459 | struct reclaim_state reclaim_state; | |
2460 | int progress; | |
2461 | ||
2462 | cond_resched(); | |
2463 | ||
2464 | /* We now go into synchronous reclaim */ | |
2465 | cpuset_memory_pressure_bump(); | |
2466 | current->flags |= PF_MEMALLOC; | |
2467 | lockdep_set_current_reclaim_state(gfp_mask); | |
2468 | reclaim_state.reclaimed_slab = 0; | |
2469 | current->reclaim_state = &reclaim_state; | |
2470 | ||
2471 | progress = try_to_free_pages(ac->zonelist, order, gfp_mask, | |
2472 | ac->nodemask); | |
2473 | ||
2474 | current->reclaim_state = NULL; | |
2475 | lockdep_clear_current_reclaim_state(); | |
2476 | current->flags &= ~PF_MEMALLOC; | |
2477 | ||
2478 | cond_resched(); | |
2479 | ||
2480 | return progress; | |
2481 | } | |
2482 | ||
2483 | /* The really slow allocator path where we enter direct reclaim */ | |
2484 | static inline struct page * | |
2485 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
2486 | int alloc_flags, const struct alloc_context *ac, | |
2487 | unsigned long *did_some_progress) | |
2488 | { | |
2489 | struct page *page = NULL; | |
2490 | bool drained = false; | |
2491 | ||
2492 | *did_some_progress = __perform_reclaim(gfp_mask, order, ac); | |
2493 | if (unlikely(!(*did_some_progress))) | |
2494 | return NULL; | |
2495 | ||
2496 | /* After successful reclaim, reconsider all zones for allocation */ | |
2497 | if (IS_ENABLED(CONFIG_NUMA)) | |
2498 | zlc_clear_zones_full(ac->zonelist); | |
2499 | ||
2500 | retry: | |
2501 | page = get_page_from_freelist(gfp_mask, order, | |
2502 | alloc_flags & ~ALLOC_NO_WATERMARKS, ac); | |
2503 | ||
2504 | /* | |
2505 | * If an allocation failed after direct reclaim, it could be because | |
2506 | * pages are pinned on the per-cpu lists. Drain them and try again | |
2507 | */ | |
2508 | if (!page && !drained) { | |
2509 | drain_all_pages(NULL); | |
2510 | drained = true; | |
2511 | goto retry; | |
2512 | } | |
2513 | ||
2514 | return page; | |
2515 | } | |
2516 | ||
2517 | /* | |
2518 | * This is called in the allocator slow-path if the allocation request is of | |
2519 | * sufficient urgency to ignore watermarks and take other desperate measures | |
2520 | */ | |
2521 | static inline struct page * | |
2522 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, | |
2523 | const struct alloc_context *ac) | |
2524 | { | |
2525 | struct page *page; | |
2526 | ||
2527 | do { | |
2528 | page = get_page_from_freelist(gfp_mask, order, | |
2529 | ALLOC_NO_WATERMARKS, ac); | |
2530 | ||
2531 | if (!page && gfp_mask & __GFP_NOFAIL) | |
2532 | wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, | |
2533 | HZ/50); | |
2534 | } while (!page && (gfp_mask & __GFP_NOFAIL)); | |
2535 | ||
2536 | return page; | |
2537 | } | |
2538 | ||
2539 | static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac) | |
2540 | { | |
2541 | struct zoneref *z; | |
2542 | struct zone *zone; | |
2543 | ||
2544 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, | |
2545 | ac->high_zoneidx, ac->nodemask) | |
2546 | wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone)); | |
2547 | } | |
2548 | ||
2549 | static inline int | |
2550 | gfp_to_alloc_flags(gfp_t gfp_mask) | |
2551 | { | |
2552 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; | |
2553 | const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD)); | |
2554 | ||
2555 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ | |
2556 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); | |
2557 | ||
2558 | /* | |
2559 | * The caller may dip into page reserves a bit more if the caller | |
2560 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
2561 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
2562 | * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH). | |
2563 | */ | |
2564 | alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); | |
2565 | ||
2566 | if (atomic) { | |
2567 | /* | |
2568 | * Not worth trying to allocate harder for __GFP_NOMEMALLOC even | |
2569 | * if it can't schedule. | |
2570 | */ | |
2571 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
2572 | alloc_flags |= ALLOC_HARDER; | |
2573 | /* | |
2574 | * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the | |
2575 | * comment for __cpuset_node_allowed(). | |
2576 | */ | |
2577 | alloc_flags &= ~ALLOC_CPUSET; | |
2578 | } else if (unlikely(rt_task(current)) && !in_interrupt()) | |
2579 | alloc_flags |= ALLOC_HARDER; | |
2580 | ||
2581 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { | |
2582 | if (gfp_mask & __GFP_MEMALLOC) | |
2583 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2584 | else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) | |
2585 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2586 | else if (!in_interrupt() && | |
2587 | ((current->flags & PF_MEMALLOC) || | |
2588 | unlikely(test_thread_flag(TIF_MEMDIE)))) | |
2589 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2590 | } | |
2591 | #ifdef CONFIG_CMA | |
2592 | if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) | |
2593 | alloc_flags |= ALLOC_CMA; | |
2594 | #endif | |
2595 | return alloc_flags; | |
2596 | } | |
2597 | ||
2598 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) | |
2599 | { | |
2600 | return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); | |
2601 | } | |
2602 | ||
2603 | static inline struct page * | |
2604 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
2605 | struct alloc_context *ac) | |
2606 | { | |
2607 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
2608 | struct page *page = NULL; | |
2609 | int alloc_flags; | |
2610 | unsigned long pages_reclaimed = 0; | |
2611 | unsigned long did_some_progress; | |
2612 | enum migrate_mode migration_mode = MIGRATE_ASYNC; | |
2613 | bool deferred_compaction = false; | |
2614 | int contended_compaction = COMPACT_CONTENDED_NONE; | |
2615 | ||
2616 | /* | |
2617 | * In the slowpath, we sanity check order to avoid ever trying to | |
2618 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may | |
2619 | * be using allocators in order of preference for an area that is | |
2620 | * too large. | |
2621 | */ | |
2622 | if (order >= MAX_ORDER) { | |
2623 | WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); | |
2624 | return NULL; | |
2625 | } | |
2626 | ||
2627 | /* | |
2628 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and | |
2629 | * __GFP_NOWARN set) should not cause reclaim since the subsystem | |
2630 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim | |
2631 | * using a larger set of nodes after it has established that the | |
2632 | * allowed per node queues are empty and that nodes are | |
2633 | * over allocated. | |
2634 | */ | |
2635 | if (IS_ENABLED(CONFIG_NUMA) && | |
2636 | (gfp_mask & GFP_THISNODE) == GFP_THISNODE) | |
2637 | goto nopage; | |
2638 | ||
2639 | retry: | |
2640 | if (!(gfp_mask & __GFP_NO_KSWAPD)) | |
2641 | wake_all_kswapds(order, ac); | |
2642 | ||
2643 | /* | |
2644 | * OK, we're below the kswapd watermark and have kicked background | |
2645 | * reclaim. Now things get more complex, so set up alloc_flags according | |
2646 | * to how we want to proceed. | |
2647 | */ | |
2648 | alloc_flags = gfp_to_alloc_flags(gfp_mask); | |
2649 | ||
2650 | /* | |
2651 | * Find the true preferred zone if the allocation is unconstrained by | |
2652 | * cpusets. | |
2653 | */ | |
2654 | if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) { | |
2655 | struct zoneref *preferred_zoneref; | |
2656 | preferred_zoneref = first_zones_zonelist(ac->zonelist, | |
2657 | ac->high_zoneidx, NULL, &ac->preferred_zone); | |
2658 | ac->classzone_idx = zonelist_zone_idx(preferred_zoneref); | |
2659 | } | |
2660 | ||
2661 | /* This is the last chance, in general, before the goto nopage. */ | |
2662 | page = get_page_from_freelist(gfp_mask, order, | |
2663 | alloc_flags & ~ALLOC_NO_WATERMARKS, ac); | |
2664 | if (page) | |
2665 | goto got_pg; | |
2666 | ||
2667 | /* Allocate without watermarks if the context allows */ | |
2668 | if (alloc_flags & ALLOC_NO_WATERMARKS) { | |
2669 | /* | |
2670 | * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds | |
2671 | * the allocation is high priority and these type of | |
2672 | * allocations are system rather than user orientated | |
2673 | */ | |
2674 | ac->zonelist = node_zonelist(numa_node_id(), gfp_mask); | |
2675 | ||
2676 | page = __alloc_pages_high_priority(gfp_mask, order, ac); | |
2677 | ||
2678 | if (page) { | |
2679 | goto got_pg; | |
2680 | } | |
2681 | } | |
2682 | ||
2683 | /* Atomic allocations - we can't balance anything */ | |
2684 | if (!wait) { | |
2685 | /* | |
2686 | * All existing users of the deprecated __GFP_NOFAIL are | |
2687 | * blockable, so warn of any new users that actually allow this | |
2688 | * type of allocation to fail. | |
2689 | */ | |
2690 | WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL); | |
2691 | goto nopage; | |
2692 | } | |
2693 | ||
2694 | /* Avoid recursion of direct reclaim */ | |
2695 | if (current->flags & PF_MEMALLOC) | |
2696 | goto nopage; | |
2697 | ||
2698 | /* Avoid allocations with no watermarks from looping endlessly */ | |
2699 | if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) | |
2700 | goto nopage; | |
2701 | ||
2702 | /* | |
2703 | * Try direct compaction. The first pass is asynchronous. Subsequent | |
2704 | * attempts after direct reclaim are synchronous | |
2705 | */ | |
2706 | page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, | |
2707 | migration_mode, | |
2708 | &contended_compaction, | |
2709 | &deferred_compaction); | |
2710 | if (page) | |
2711 | goto got_pg; | |
2712 | ||
2713 | /* Checks for THP-specific high-order allocations */ | |
2714 | if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) { | |
2715 | /* | |
2716 | * If compaction is deferred for high-order allocations, it is | |
2717 | * because sync compaction recently failed. If this is the case | |
2718 | * and the caller requested a THP allocation, we do not want | |
2719 | * to heavily disrupt the system, so we fail the allocation | |
2720 | * instead of entering direct reclaim. | |
2721 | */ | |
2722 | if (deferred_compaction) | |
2723 | goto nopage; | |
2724 | ||
2725 | /* | |
2726 | * In all zones where compaction was attempted (and not | |
2727 | * deferred or skipped), lock contention has been detected. | |
2728 | * For THP allocation we do not want to disrupt the others | |
2729 | * so we fallback to base pages instead. | |
2730 | */ | |
2731 | if (contended_compaction == COMPACT_CONTENDED_LOCK) | |
2732 | goto nopage; | |
2733 | ||
2734 | /* | |
2735 | * If compaction was aborted due to need_resched(), we do not | |
2736 | * want to further increase allocation latency, unless it is | |
2737 | * khugepaged trying to collapse. | |
2738 | */ | |
2739 | if (contended_compaction == COMPACT_CONTENDED_SCHED | |
2740 | && !(current->flags & PF_KTHREAD)) | |
2741 | goto nopage; | |
2742 | } | |
2743 | ||
2744 | /* | |
2745 | * It can become very expensive to allocate transparent hugepages at | |
2746 | * fault, so use asynchronous memory compaction for THP unless it is | |
2747 | * khugepaged trying to collapse. | |
2748 | */ | |
2749 | if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE || | |
2750 | (current->flags & PF_KTHREAD)) | |
2751 | migration_mode = MIGRATE_SYNC_LIGHT; | |
2752 | ||
2753 | /* Try direct reclaim and then allocating */ | |
2754 | page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, | |
2755 | &did_some_progress); | |
2756 | if (page) | |
2757 | goto got_pg; | |
2758 | ||
2759 | /* Check if we should retry the allocation */ | |
2760 | pages_reclaimed += did_some_progress; | |
2761 | if (should_alloc_retry(gfp_mask, order, did_some_progress, | |
2762 | pages_reclaimed)) { | |
2763 | /* | |
2764 | * If we fail to make progress by freeing individual | |
2765 | * pages, but the allocation wants us to keep going, | |
2766 | * start OOM killing tasks. | |
2767 | */ | |
2768 | if (!did_some_progress) { | |
2769 | page = __alloc_pages_may_oom(gfp_mask, order, ac, | |
2770 | &did_some_progress); | |
2771 | if (page) | |
2772 | goto got_pg; | |
2773 | if (!did_some_progress) | |
2774 | goto nopage; | |
2775 | } | |
2776 | /* Wait for some write requests to complete then retry */ | |
2777 | wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50); | |
2778 | goto retry; | |
2779 | } else { | |
2780 | /* | |
2781 | * High-order allocations do not necessarily loop after | |
2782 | * direct reclaim and reclaim/compaction depends on compaction | |
2783 | * being called after reclaim so call directly if necessary | |
2784 | */ | |
2785 | page = __alloc_pages_direct_compact(gfp_mask, order, | |
2786 | alloc_flags, ac, migration_mode, | |
2787 | &contended_compaction, | |
2788 | &deferred_compaction); | |
2789 | if (page) | |
2790 | goto got_pg; | |
2791 | } | |
2792 | ||
2793 | nopage: | |
2794 | warn_alloc_failed(gfp_mask, order, NULL); | |
2795 | got_pg: | |
2796 | return page; | |
2797 | } | |
2798 | ||
2799 | /* | |
2800 | * This is the 'heart' of the zoned buddy allocator. | |
2801 | */ | |
2802 | struct page * | |
2803 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, | |
2804 | struct zonelist *zonelist, nodemask_t *nodemask) | |
2805 | { | |
2806 | struct zoneref *preferred_zoneref; | |
2807 | struct page *page = NULL; | |
2808 | unsigned int cpuset_mems_cookie; | |
2809 | int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR; | |
2810 | gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */ | |
2811 | struct alloc_context ac = { | |
2812 | .high_zoneidx = gfp_zone(gfp_mask), | |
2813 | .nodemask = nodemask, | |
2814 | .migratetype = gfpflags_to_migratetype(gfp_mask), | |
2815 | }; | |
2816 | ||
2817 | gfp_mask &= gfp_allowed_mask; | |
2818 | ||
2819 | lockdep_trace_alloc(gfp_mask); | |
2820 | ||
2821 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2822 | ||
2823 | if (should_fail_alloc_page(gfp_mask, order)) | |
2824 | return NULL; | |
2825 | ||
2826 | /* | |
2827 | * Check the zones suitable for the gfp_mask contain at least one | |
2828 | * valid zone. It's possible to have an empty zonelist as a result | |
2829 | * of GFP_THISNODE and a memoryless node | |
2830 | */ | |
2831 | if (unlikely(!zonelist->_zonerefs->zone)) | |
2832 | return NULL; | |
2833 | ||
2834 | if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE) | |
2835 | alloc_flags |= ALLOC_CMA; | |
2836 | ||
2837 | retry_cpuset: | |
2838 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
2839 | ||
2840 | /* We set it here, as __alloc_pages_slowpath might have changed it */ | |
2841 | ac.zonelist = zonelist; | |
2842 | /* The preferred zone is used for statistics later */ | |
2843 | preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx, | |
2844 | ac.nodemask ? : &cpuset_current_mems_allowed, | |
2845 | &ac.preferred_zone); | |
2846 | if (!ac.preferred_zone) | |
2847 | goto out; | |
2848 | ac.classzone_idx = zonelist_zone_idx(preferred_zoneref); | |
2849 | ||
2850 | /* First allocation attempt */ | |
2851 | alloc_mask = gfp_mask|__GFP_HARDWALL; | |
2852 | page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac); | |
2853 | if (unlikely(!page)) { | |
2854 | /* | |
2855 | * Runtime PM, block IO and its error handling path | |
2856 | * can deadlock because I/O on the device might not | |
2857 | * complete. | |
2858 | */ | |
2859 | alloc_mask = memalloc_noio_flags(gfp_mask); | |
2860 | ||
2861 | page = __alloc_pages_slowpath(alloc_mask, order, &ac); | |
2862 | } | |
2863 | ||
2864 | if (kmemcheck_enabled && page) | |
2865 | kmemcheck_pagealloc_alloc(page, order, gfp_mask); | |
2866 | ||
2867 | trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype); | |
2868 | ||
2869 | out: | |
2870 | /* | |
2871 | * When updating a task's mems_allowed, it is possible to race with | |
2872 | * parallel threads in such a way that an allocation can fail while | |
2873 | * the mask is being updated. If a page allocation is about to fail, | |
2874 | * check if the cpuset changed during allocation and if so, retry. | |
2875 | */ | |
2876 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) | |
2877 | goto retry_cpuset; | |
2878 | ||
2879 | return page; | |
2880 | } | |
2881 | EXPORT_SYMBOL(__alloc_pages_nodemask); | |
2882 | ||
2883 | /* | |
2884 | * Common helper functions. | |
2885 | */ | |
2886 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) | |
2887 | { | |
2888 | struct page *page; | |
2889 | ||
2890 | /* | |
2891 | * __get_free_pages() returns a 32-bit address, which cannot represent | |
2892 | * a highmem page | |
2893 | */ | |
2894 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); | |
2895 | ||
2896 | page = alloc_pages(gfp_mask, order); | |
2897 | if (!page) | |
2898 | return 0; | |
2899 | return (unsigned long) page_address(page); | |
2900 | } | |
2901 | EXPORT_SYMBOL(__get_free_pages); | |
2902 | ||
2903 | unsigned long get_zeroed_page(gfp_t gfp_mask) | |
2904 | { | |
2905 | return __get_free_pages(gfp_mask | __GFP_ZERO, 0); | |
2906 | } | |
2907 | EXPORT_SYMBOL(get_zeroed_page); | |
2908 | ||
2909 | void __free_pages(struct page *page, unsigned int order) | |
2910 | { | |
2911 | if (put_page_testzero(page)) { | |
2912 | if (order == 0) | |
2913 | free_hot_cold_page(page, false); | |
2914 | else | |
2915 | __free_pages_ok(page, order); | |
2916 | } | |
2917 | } | |
2918 | ||
2919 | EXPORT_SYMBOL(__free_pages); | |
2920 | ||
2921 | void free_pages(unsigned long addr, unsigned int order) | |
2922 | { | |
2923 | if (addr != 0) { | |
2924 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
2925 | __free_pages(virt_to_page((void *)addr), order); | |
2926 | } | |
2927 | } | |
2928 | ||
2929 | EXPORT_SYMBOL(free_pages); | |
2930 | ||
2931 | /* | |
2932 | * alloc_kmem_pages charges newly allocated pages to the kmem resource counter | |
2933 | * of the current memory cgroup. | |
2934 | * | |
2935 | * It should be used when the caller would like to use kmalloc, but since the | |
2936 | * allocation is large, it has to fall back to the page allocator. | |
2937 | */ | |
2938 | struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order) | |
2939 | { | |
2940 | struct page *page; | |
2941 | struct mem_cgroup *memcg = NULL; | |
2942 | ||
2943 | if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) | |
2944 | return NULL; | |
2945 | page = alloc_pages(gfp_mask, order); | |
2946 | memcg_kmem_commit_charge(page, memcg, order); | |
2947 | return page; | |
2948 | } | |
2949 | ||
2950 | struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order) | |
2951 | { | |
2952 | struct page *page; | |
2953 | struct mem_cgroup *memcg = NULL; | |
2954 | ||
2955 | if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) | |
2956 | return NULL; | |
2957 | page = alloc_pages_node(nid, gfp_mask, order); | |
2958 | memcg_kmem_commit_charge(page, memcg, order); | |
2959 | return page; | |
2960 | } | |
2961 | ||
2962 | /* | |
2963 | * __free_kmem_pages and free_kmem_pages will free pages allocated with | |
2964 | * alloc_kmem_pages. | |
2965 | */ | |
2966 | void __free_kmem_pages(struct page *page, unsigned int order) | |
2967 | { | |
2968 | memcg_kmem_uncharge_pages(page, order); | |
2969 | __free_pages(page, order); | |
2970 | } | |
2971 | ||
2972 | void free_kmem_pages(unsigned long addr, unsigned int order) | |
2973 | { | |
2974 | if (addr != 0) { | |
2975 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
2976 | __free_kmem_pages(virt_to_page((void *)addr), order); | |
2977 | } | |
2978 | } | |
2979 | ||
2980 | static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) | |
2981 | { | |
2982 | if (addr) { | |
2983 | unsigned long alloc_end = addr + (PAGE_SIZE << order); | |
2984 | unsigned long used = addr + PAGE_ALIGN(size); | |
2985 | ||
2986 | split_page(virt_to_page((void *)addr), order); | |
2987 | while (used < alloc_end) { | |
2988 | free_page(used); | |
2989 | used += PAGE_SIZE; | |
2990 | } | |
2991 | } | |
2992 | return (void *)addr; | |
2993 | } | |
2994 | ||
2995 | /** | |
2996 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
2997 | * @size: the number of bytes to allocate | |
2998 | * @gfp_mask: GFP flags for the allocation | |
2999 | * | |
3000 | * This function is similar to alloc_pages(), except that it allocates the | |
3001 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
3002 | * allocate memory in power-of-two pages. | |
3003 | * | |
3004 | * This function is also limited by MAX_ORDER. | |
3005 | * | |
3006 | * Memory allocated by this function must be released by free_pages_exact(). | |
3007 | */ | |
3008 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
3009 | { | |
3010 | unsigned int order = get_order(size); | |
3011 | unsigned long addr; | |
3012 | ||
3013 | addr = __get_free_pages(gfp_mask, order); | |
3014 | return make_alloc_exact(addr, order, size); | |
3015 | } | |
3016 | EXPORT_SYMBOL(alloc_pages_exact); | |
3017 | ||
3018 | /** | |
3019 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous | |
3020 | * pages on a node. | |
3021 | * @nid: the preferred node ID where memory should be allocated | |
3022 | * @size: the number of bytes to allocate | |
3023 | * @gfp_mask: GFP flags for the allocation | |
3024 | * | |
3025 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling | |
3026 | * back. | |
3027 | * Note this is not alloc_pages_exact_node() which allocates on a specific node, | |
3028 | * but is not exact. | |
3029 | */ | |
3030 | void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) | |
3031 | { | |
3032 | unsigned order = get_order(size); | |
3033 | struct page *p = alloc_pages_node(nid, gfp_mask, order); | |
3034 | if (!p) | |
3035 | return NULL; | |
3036 | return make_alloc_exact((unsigned long)page_address(p), order, size); | |
3037 | } | |
3038 | ||
3039 | /** | |
3040 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
3041 | * @virt: the value returned by alloc_pages_exact. | |
3042 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
3043 | * | |
3044 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
3045 | */ | |
3046 | void free_pages_exact(void *virt, size_t size) | |
3047 | { | |
3048 | unsigned long addr = (unsigned long)virt; | |
3049 | unsigned long end = addr + PAGE_ALIGN(size); | |
3050 | ||
3051 | while (addr < end) { | |
3052 | free_page(addr); | |
3053 | addr += PAGE_SIZE; | |
3054 | } | |
3055 | } | |
3056 | EXPORT_SYMBOL(free_pages_exact); | |
3057 | ||
3058 | /** | |
3059 | * nr_free_zone_pages - count number of pages beyond high watermark | |
3060 | * @offset: The zone index of the highest zone | |
3061 | * | |
3062 | * nr_free_zone_pages() counts the number of counts pages which are beyond the | |
3063 | * high watermark within all zones at or below a given zone index. For each | |
3064 | * zone, the number of pages is calculated as: | |
3065 | * managed_pages - high_pages | |
3066 | */ | |
3067 | static unsigned long nr_free_zone_pages(int offset) | |
3068 | { | |
3069 | struct zoneref *z; | |
3070 | struct zone *zone; | |
3071 | ||
3072 | /* Just pick one node, since fallback list is circular */ | |
3073 | unsigned long sum = 0; | |
3074 | ||
3075 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); | |
3076 | ||
3077 | for_each_zone_zonelist(zone, z, zonelist, offset) { | |
3078 | unsigned long size = zone->managed_pages; | |
3079 | unsigned long high = high_wmark_pages(zone); | |
3080 | if (size > high) | |
3081 | sum += size - high; | |
3082 | } | |
3083 | ||
3084 | return sum; | |
3085 | } | |
3086 | ||
3087 | /** | |
3088 | * nr_free_buffer_pages - count number of pages beyond high watermark | |
3089 | * | |
3090 | * nr_free_buffer_pages() counts the number of pages which are beyond the high | |
3091 | * watermark within ZONE_DMA and ZONE_NORMAL. | |
3092 | */ | |
3093 | unsigned long nr_free_buffer_pages(void) | |
3094 | { | |
3095 | return nr_free_zone_pages(gfp_zone(GFP_USER)); | |
3096 | } | |
3097 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); | |
3098 | ||
3099 | /** | |
3100 | * nr_free_pagecache_pages - count number of pages beyond high watermark | |
3101 | * | |
3102 | * nr_free_pagecache_pages() counts the number of pages which are beyond the | |
3103 | * high watermark within all zones. | |
3104 | */ | |
3105 | unsigned long nr_free_pagecache_pages(void) | |
3106 | { | |
3107 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
3108 | } | |
3109 | ||
3110 | static inline void show_node(struct zone *zone) | |
3111 | { | |
3112 | if (IS_ENABLED(CONFIG_NUMA)) | |
3113 | printk("Node %d ", zone_to_nid(zone)); | |
3114 | } | |
3115 | ||
3116 | void si_meminfo(struct sysinfo *val) | |
3117 | { | |
3118 | val->totalram = totalram_pages; | |
3119 | val->sharedram = global_page_state(NR_SHMEM); | |
3120 | val->freeram = global_page_state(NR_FREE_PAGES); | |
3121 | val->bufferram = nr_blockdev_pages(); | |
3122 | val->totalhigh = totalhigh_pages; | |
3123 | val->freehigh = nr_free_highpages(); | |
3124 | val->mem_unit = PAGE_SIZE; | |
3125 | } | |
3126 | ||
3127 | EXPORT_SYMBOL(si_meminfo); | |
3128 | ||
3129 | #ifdef CONFIG_NUMA | |
3130 | void si_meminfo_node(struct sysinfo *val, int nid) | |
3131 | { | |
3132 | int zone_type; /* needs to be signed */ | |
3133 | unsigned long managed_pages = 0; | |
3134 | pg_data_t *pgdat = NODE_DATA(nid); | |
3135 | ||
3136 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) | |
3137 | managed_pages += pgdat->node_zones[zone_type].managed_pages; | |
3138 | val->totalram = managed_pages; | |
3139 | val->sharedram = node_page_state(nid, NR_SHMEM); | |
3140 | val->freeram = node_page_state(nid, NR_FREE_PAGES); | |
3141 | #ifdef CONFIG_HIGHMEM | |
3142 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; | |
3143 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], | |
3144 | NR_FREE_PAGES); | |
3145 | #else | |
3146 | val->totalhigh = 0; | |
3147 | val->freehigh = 0; | |
3148 | #endif | |
3149 | val->mem_unit = PAGE_SIZE; | |
3150 | } | |
3151 | #endif | |
3152 | ||
3153 | /* | |
3154 | * Determine whether the node should be displayed or not, depending on whether | |
3155 | * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). | |
3156 | */ | |
3157 | bool skip_free_areas_node(unsigned int flags, int nid) | |
3158 | { | |
3159 | bool ret = false; | |
3160 | unsigned int cpuset_mems_cookie; | |
3161 | ||
3162 | if (!(flags & SHOW_MEM_FILTER_NODES)) | |
3163 | goto out; | |
3164 | ||
3165 | do { | |
3166 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
3167 | ret = !node_isset(nid, cpuset_current_mems_allowed); | |
3168 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
3169 | out: | |
3170 | return ret; | |
3171 | } | |
3172 | ||
3173 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
3174 | ||
3175 | static void show_migration_types(unsigned char type) | |
3176 | { | |
3177 | static const char types[MIGRATE_TYPES] = { | |
3178 | [MIGRATE_UNMOVABLE] = 'U', | |
3179 | [MIGRATE_RECLAIMABLE] = 'E', | |
3180 | [MIGRATE_MOVABLE] = 'M', | |
3181 | [MIGRATE_RESERVE] = 'R', | |
3182 | #ifdef CONFIG_CMA | |
3183 | [MIGRATE_CMA] = 'C', | |
3184 | #endif | |
3185 | #ifdef CONFIG_MEMORY_ISOLATION | |
3186 | [MIGRATE_ISOLATE] = 'I', | |
3187 | #endif | |
3188 | }; | |
3189 | char tmp[MIGRATE_TYPES + 1]; | |
3190 | char *p = tmp; | |
3191 | int i; | |
3192 | ||
3193 | for (i = 0; i < MIGRATE_TYPES; i++) { | |
3194 | if (type & (1 << i)) | |
3195 | *p++ = types[i]; | |
3196 | } | |
3197 | ||
3198 | *p = '\0'; | |
3199 | printk("(%s) ", tmp); | |
3200 | } | |
3201 | ||
3202 | /* | |
3203 | * Show free area list (used inside shift_scroll-lock stuff) | |
3204 | * We also calculate the percentage fragmentation. We do this by counting the | |
3205 | * memory on each free list with the exception of the first item on the list. | |
3206 | * Suppresses nodes that are not allowed by current's cpuset if | |
3207 | * SHOW_MEM_FILTER_NODES is passed. | |
3208 | */ | |
3209 | void show_free_areas(unsigned int filter) | |
3210 | { | |
3211 | int cpu; | |
3212 | struct zone *zone; | |
3213 | ||
3214 | for_each_populated_zone(zone) { | |
3215 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
3216 | continue; | |
3217 | show_node(zone); | |
3218 | printk("%s per-cpu:\n", zone->name); | |
3219 | ||
3220 | for_each_online_cpu(cpu) { | |
3221 | struct per_cpu_pageset *pageset; | |
3222 | ||
3223 | pageset = per_cpu_ptr(zone->pageset, cpu); | |
3224 | ||
3225 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", | |
3226 | cpu, pageset->pcp.high, | |
3227 | pageset->pcp.batch, pageset->pcp.count); | |
3228 | } | |
3229 | } | |
3230 | ||
3231 | printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" | |
3232 | " active_file:%lu inactive_file:%lu isolated_file:%lu\n" | |
3233 | " unevictable:%lu" | |
3234 | " dirty:%lu writeback:%lu unstable:%lu\n" | |
3235 | " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" | |
3236 | " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" | |
3237 | " free_cma:%lu\n", | |
3238 | global_page_state(NR_ACTIVE_ANON), | |
3239 | global_page_state(NR_INACTIVE_ANON), | |
3240 | global_page_state(NR_ISOLATED_ANON), | |
3241 | global_page_state(NR_ACTIVE_FILE), | |
3242 | global_page_state(NR_INACTIVE_FILE), | |
3243 | global_page_state(NR_ISOLATED_FILE), | |
3244 | global_page_state(NR_UNEVICTABLE), | |
3245 | global_page_state(NR_FILE_DIRTY), | |
3246 | global_page_state(NR_WRITEBACK), | |
3247 | global_page_state(NR_UNSTABLE_NFS), | |
3248 | global_page_state(NR_FREE_PAGES), | |
3249 | global_page_state(NR_SLAB_RECLAIMABLE), | |
3250 | global_page_state(NR_SLAB_UNRECLAIMABLE), | |
3251 | global_page_state(NR_FILE_MAPPED), | |
3252 | global_page_state(NR_SHMEM), | |
3253 | global_page_state(NR_PAGETABLE), | |
3254 | global_page_state(NR_BOUNCE), | |
3255 | global_page_state(NR_FREE_CMA_PAGES)); | |
3256 | ||
3257 | for_each_populated_zone(zone) { | |
3258 | int i; | |
3259 | ||
3260 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
3261 | continue; | |
3262 | show_node(zone); | |
3263 | printk("%s" | |
3264 | " free:%lukB" | |
3265 | " min:%lukB" | |
3266 | " low:%lukB" | |
3267 | " high:%lukB" | |
3268 | " active_anon:%lukB" | |
3269 | " inactive_anon:%lukB" | |
3270 | " active_file:%lukB" | |
3271 | " inactive_file:%lukB" | |
3272 | " unevictable:%lukB" | |
3273 | " isolated(anon):%lukB" | |
3274 | " isolated(file):%lukB" | |
3275 | " present:%lukB" | |
3276 | " managed:%lukB" | |
3277 | " mlocked:%lukB" | |
3278 | " dirty:%lukB" | |
3279 | " writeback:%lukB" | |
3280 | " mapped:%lukB" | |
3281 | " shmem:%lukB" | |
3282 | " slab_reclaimable:%lukB" | |
3283 | " slab_unreclaimable:%lukB" | |
3284 | " kernel_stack:%lukB" | |
3285 | " pagetables:%lukB" | |
3286 | " unstable:%lukB" | |
3287 | " bounce:%lukB" | |
3288 | " free_cma:%lukB" | |
3289 | " writeback_tmp:%lukB" | |
3290 | " pages_scanned:%lu" | |
3291 | " all_unreclaimable? %s" | |
3292 | "\n", | |
3293 | zone->name, | |
3294 | K(zone_page_state(zone, NR_FREE_PAGES)), | |
3295 | K(min_wmark_pages(zone)), | |
3296 | K(low_wmark_pages(zone)), | |
3297 | K(high_wmark_pages(zone)), | |
3298 | K(zone_page_state(zone, NR_ACTIVE_ANON)), | |
3299 | K(zone_page_state(zone, NR_INACTIVE_ANON)), | |
3300 | K(zone_page_state(zone, NR_ACTIVE_FILE)), | |
3301 | K(zone_page_state(zone, NR_INACTIVE_FILE)), | |
3302 | K(zone_page_state(zone, NR_UNEVICTABLE)), | |
3303 | K(zone_page_state(zone, NR_ISOLATED_ANON)), | |
3304 | K(zone_page_state(zone, NR_ISOLATED_FILE)), | |
3305 | K(zone->present_pages), | |
3306 | K(zone->managed_pages), | |
3307 | K(zone_page_state(zone, NR_MLOCK)), | |
3308 | K(zone_page_state(zone, NR_FILE_DIRTY)), | |
3309 | K(zone_page_state(zone, NR_WRITEBACK)), | |
3310 | K(zone_page_state(zone, NR_FILE_MAPPED)), | |
3311 | K(zone_page_state(zone, NR_SHMEM)), | |
3312 | K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), | |
3313 | K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), | |
3314 | zone_page_state(zone, NR_KERNEL_STACK) * | |
3315 | THREAD_SIZE / 1024, | |
3316 | K(zone_page_state(zone, NR_PAGETABLE)), | |
3317 | K(zone_page_state(zone, NR_UNSTABLE_NFS)), | |
3318 | K(zone_page_state(zone, NR_BOUNCE)), | |
3319 | K(zone_page_state(zone, NR_FREE_CMA_PAGES)), | |
3320 | K(zone_page_state(zone, NR_WRITEBACK_TEMP)), | |
3321 | K(zone_page_state(zone, NR_PAGES_SCANNED)), | |
3322 | (!zone_reclaimable(zone) ? "yes" : "no") | |
3323 | ); | |
3324 | printk("lowmem_reserve[]:"); | |
3325 | for (i = 0; i < MAX_NR_ZONES; i++) | |
3326 | printk(" %ld", zone->lowmem_reserve[i]); | |
3327 | printk("\n"); | |
3328 | } | |
3329 | ||
3330 | for_each_populated_zone(zone) { | |
3331 | unsigned long nr[MAX_ORDER], flags, order, total = 0; | |
3332 | unsigned char types[MAX_ORDER]; | |
3333 | ||
3334 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
3335 | continue; | |
3336 | show_node(zone); | |
3337 | printk("%s: ", zone->name); | |
3338 | ||
3339 | spin_lock_irqsave(&zone->lock, flags); | |
3340 | for (order = 0; order < MAX_ORDER; order++) { | |
3341 | struct free_area *area = &zone->free_area[order]; | |
3342 | int type; | |
3343 | ||
3344 | nr[order] = area->nr_free; | |
3345 | total += nr[order] << order; | |
3346 | ||
3347 | types[order] = 0; | |
3348 | for (type = 0; type < MIGRATE_TYPES; type++) { | |
3349 | if (!list_empty(&area->free_list[type])) | |
3350 | types[order] |= 1 << type; | |
3351 | } | |
3352 | } | |
3353 | spin_unlock_irqrestore(&zone->lock, flags); | |
3354 | for (order = 0; order < MAX_ORDER; order++) { | |
3355 | printk("%lu*%lukB ", nr[order], K(1UL) << order); | |
3356 | if (nr[order]) | |
3357 | show_migration_types(types[order]); | |
3358 | } | |
3359 | printk("= %lukB\n", K(total)); | |
3360 | } | |
3361 | ||
3362 | hugetlb_show_meminfo(); | |
3363 | ||
3364 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); | |
3365 | ||
3366 | show_swap_cache_info(); | |
3367 | } | |
3368 | ||
3369 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) | |
3370 | { | |
3371 | zoneref->zone = zone; | |
3372 | zoneref->zone_idx = zone_idx(zone); | |
3373 | } | |
3374 | ||
3375 | /* | |
3376 | * Builds allocation fallback zone lists. | |
3377 | * | |
3378 | * Add all populated zones of a node to the zonelist. | |
3379 | */ | |
3380 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, | |
3381 | int nr_zones) | |
3382 | { | |
3383 | struct zone *zone; | |
3384 | enum zone_type zone_type = MAX_NR_ZONES; | |
3385 | ||
3386 | do { | |
3387 | zone_type--; | |
3388 | zone = pgdat->node_zones + zone_type; | |
3389 | if (populated_zone(zone)) { | |
3390 | zoneref_set_zone(zone, | |
3391 | &zonelist->_zonerefs[nr_zones++]); | |
3392 | check_highest_zone(zone_type); | |
3393 | } | |
3394 | } while (zone_type); | |
3395 | ||
3396 | return nr_zones; | |
3397 | } | |
3398 | ||
3399 | ||
3400 | /* | |
3401 | * zonelist_order: | |
3402 | * 0 = automatic detection of better ordering. | |
3403 | * 1 = order by ([node] distance, -zonetype) | |
3404 | * 2 = order by (-zonetype, [node] distance) | |
3405 | * | |
3406 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create | |
3407 | * the same zonelist. So only NUMA can configure this param. | |
3408 | */ | |
3409 | #define ZONELIST_ORDER_DEFAULT 0 | |
3410 | #define ZONELIST_ORDER_NODE 1 | |
3411 | #define ZONELIST_ORDER_ZONE 2 | |
3412 | ||
3413 | /* zonelist order in the kernel. | |
3414 | * set_zonelist_order() will set this to NODE or ZONE. | |
3415 | */ | |
3416 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3417 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; | |
3418 | ||
3419 | ||
3420 | #ifdef CONFIG_NUMA | |
3421 | /* The value user specified ....changed by config */ | |
3422 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3423 | /* string for sysctl */ | |
3424 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
3425 | char numa_zonelist_order[16] = "default"; | |
3426 | ||
3427 | /* | |
3428 | * interface for configure zonelist ordering. | |
3429 | * command line option "numa_zonelist_order" | |
3430 | * = "[dD]efault - default, automatic configuration. | |
3431 | * = "[nN]ode - order by node locality, then by zone within node | |
3432 | * = "[zZ]one - order by zone, then by locality within zone | |
3433 | */ | |
3434 | ||
3435 | static int __parse_numa_zonelist_order(char *s) | |
3436 | { | |
3437 | if (*s == 'd' || *s == 'D') { | |
3438 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3439 | } else if (*s == 'n' || *s == 'N') { | |
3440 | user_zonelist_order = ZONELIST_ORDER_NODE; | |
3441 | } else if (*s == 'z' || *s == 'Z') { | |
3442 | user_zonelist_order = ZONELIST_ORDER_ZONE; | |
3443 | } else { | |
3444 | printk(KERN_WARNING | |
3445 | "Ignoring invalid numa_zonelist_order value: " | |
3446 | "%s\n", s); | |
3447 | return -EINVAL; | |
3448 | } | |
3449 | return 0; | |
3450 | } | |
3451 | ||
3452 | static __init int setup_numa_zonelist_order(char *s) | |
3453 | { | |
3454 | int ret; | |
3455 | ||
3456 | if (!s) | |
3457 | return 0; | |
3458 | ||
3459 | ret = __parse_numa_zonelist_order(s); | |
3460 | if (ret == 0) | |
3461 | strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); | |
3462 | ||
3463 | return ret; | |
3464 | } | |
3465 | early_param("numa_zonelist_order", setup_numa_zonelist_order); | |
3466 | ||
3467 | /* | |
3468 | * sysctl handler for numa_zonelist_order | |
3469 | */ | |
3470 | int numa_zonelist_order_handler(struct ctl_table *table, int write, | |
3471 | void __user *buffer, size_t *length, | |
3472 | loff_t *ppos) | |
3473 | { | |
3474 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; | |
3475 | int ret; | |
3476 | static DEFINE_MUTEX(zl_order_mutex); | |
3477 | ||
3478 | mutex_lock(&zl_order_mutex); | |
3479 | if (write) { | |
3480 | if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { | |
3481 | ret = -EINVAL; | |
3482 | goto out; | |
3483 | } | |
3484 | strcpy(saved_string, (char *)table->data); | |
3485 | } | |
3486 | ret = proc_dostring(table, write, buffer, length, ppos); | |
3487 | if (ret) | |
3488 | goto out; | |
3489 | if (write) { | |
3490 | int oldval = user_zonelist_order; | |
3491 | ||
3492 | ret = __parse_numa_zonelist_order((char *)table->data); | |
3493 | if (ret) { | |
3494 | /* | |
3495 | * bogus value. restore saved string | |
3496 | */ | |
3497 | strncpy((char *)table->data, saved_string, | |
3498 | NUMA_ZONELIST_ORDER_LEN); | |
3499 | user_zonelist_order = oldval; | |
3500 | } else if (oldval != user_zonelist_order) { | |
3501 | mutex_lock(&zonelists_mutex); | |
3502 | build_all_zonelists(NULL, NULL); | |
3503 | mutex_unlock(&zonelists_mutex); | |
3504 | } | |
3505 | } | |
3506 | out: | |
3507 | mutex_unlock(&zl_order_mutex); | |
3508 | return ret; | |
3509 | } | |
3510 | ||
3511 | ||
3512 | #define MAX_NODE_LOAD (nr_online_nodes) | |
3513 | static int node_load[MAX_NUMNODES]; | |
3514 | ||
3515 | /** | |
3516 | * find_next_best_node - find the next node that should appear in a given node's fallback list | |
3517 | * @node: node whose fallback list we're appending | |
3518 | * @used_node_mask: nodemask_t of already used nodes | |
3519 | * | |
3520 | * We use a number of factors to determine which is the next node that should | |
3521 | * appear on a given node's fallback list. The node should not have appeared | |
3522 | * already in @node's fallback list, and it should be the next closest node | |
3523 | * according to the distance array (which contains arbitrary distance values | |
3524 | * from each node to each node in the system), and should also prefer nodes | |
3525 | * with no CPUs, since presumably they'll have very little allocation pressure | |
3526 | * on them otherwise. | |
3527 | * It returns -1 if no node is found. | |
3528 | */ | |
3529 | static int find_next_best_node(int node, nodemask_t *used_node_mask) | |
3530 | { | |
3531 | int n, val; | |
3532 | int min_val = INT_MAX; | |
3533 | int best_node = NUMA_NO_NODE; | |
3534 | const struct cpumask *tmp = cpumask_of_node(0); | |
3535 | ||
3536 | /* Use the local node if we haven't already */ | |
3537 | if (!node_isset(node, *used_node_mask)) { | |
3538 | node_set(node, *used_node_mask); | |
3539 | return node; | |
3540 | } | |
3541 | ||
3542 | for_each_node_state(n, N_MEMORY) { | |
3543 | ||
3544 | /* Don't want a node to appear more than once */ | |
3545 | if (node_isset(n, *used_node_mask)) | |
3546 | continue; | |
3547 | ||
3548 | /* Use the distance array to find the distance */ | |
3549 | val = node_distance(node, n); | |
3550 | ||
3551 | /* Penalize nodes under us ("prefer the next node") */ | |
3552 | val += (n < node); | |
3553 | ||
3554 | /* Give preference to headless and unused nodes */ | |
3555 | tmp = cpumask_of_node(n); | |
3556 | if (!cpumask_empty(tmp)) | |
3557 | val += PENALTY_FOR_NODE_WITH_CPUS; | |
3558 | ||
3559 | /* Slight preference for less loaded node */ | |
3560 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); | |
3561 | val += node_load[n]; | |
3562 | ||
3563 | if (val < min_val) { | |
3564 | min_val = val; | |
3565 | best_node = n; | |
3566 | } | |
3567 | } | |
3568 | ||
3569 | if (best_node >= 0) | |
3570 | node_set(best_node, *used_node_mask); | |
3571 | ||
3572 | return best_node; | |
3573 | } | |
3574 | ||
3575 | ||
3576 | /* | |
3577 | * Build zonelists ordered by node and zones within node. | |
3578 | * This results in maximum locality--normal zone overflows into local | |
3579 | * DMA zone, if any--but risks exhausting DMA zone. | |
3580 | */ | |
3581 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) | |
3582 | { | |
3583 | int j; | |
3584 | struct zonelist *zonelist; | |
3585 | ||
3586 | zonelist = &pgdat->node_zonelists[0]; | |
3587 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) | |
3588 | ; | |
3589 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); | |
3590 | zonelist->_zonerefs[j].zone = NULL; | |
3591 | zonelist->_zonerefs[j].zone_idx = 0; | |
3592 | } | |
3593 | ||
3594 | /* | |
3595 | * Build gfp_thisnode zonelists | |
3596 | */ | |
3597 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
3598 | { | |
3599 | int j; | |
3600 | struct zonelist *zonelist; | |
3601 | ||
3602 | zonelist = &pgdat->node_zonelists[1]; | |
3603 | j = build_zonelists_node(pgdat, zonelist, 0); | |
3604 | zonelist->_zonerefs[j].zone = NULL; | |
3605 | zonelist->_zonerefs[j].zone_idx = 0; | |
3606 | } | |
3607 | ||
3608 | /* | |
3609 | * Build zonelists ordered by zone and nodes within zones. | |
3610 | * This results in conserving DMA zone[s] until all Normal memory is | |
3611 | * exhausted, but results in overflowing to remote node while memory | |
3612 | * may still exist in local DMA zone. | |
3613 | */ | |
3614 | static int node_order[MAX_NUMNODES]; | |
3615 | ||
3616 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) | |
3617 | { | |
3618 | int pos, j, node; | |
3619 | int zone_type; /* needs to be signed */ | |
3620 | struct zone *z; | |
3621 | struct zonelist *zonelist; | |
3622 | ||
3623 | zonelist = &pgdat->node_zonelists[0]; | |
3624 | pos = 0; | |
3625 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { | |
3626 | for (j = 0; j < nr_nodes; j++) { | |
3627 | node = node_order[j]; | |
3628 | z = &NODE_DATA(node)->node_zones[zone_type]; | |
3629 | if (populated_zone(z)) { | |
3630 | zoneref_set_zone(z, | |
3631 | &zonelist->_zonerefs[pos++]); | |
3632 | check_highest_zone(zone_type); | |
3633 | } | |
3634 | } | |
3635 | } | |
3636 | zonelist->_zonerefs[pos].zone = NULL; | |
3637 | zonelist->_zonerefs[pos].zone_idx = 0; | |
3638 | } | |
3639 | ||
3640 | #if defined(CONFIG_64BIT) | |
3641 | /* | |
3642 | * Devices that require DMA32/DMA are relatively rare and do not justify a | |
3643 | * penalty to every machine in case the specialised case applies. Default | |
3644 | * to Node-ordering on 64-bit NUMA machines | |
3645 | */ | |
3646 | static int default_zonelist_order(void) | |
3647 | { | |
3648 | return ZONELIST_ORDER_NODE; | |
3649 | } | |
3650 | #else | |
3651 | /* | |
3652 | * On 32-bit, the Normal zone needs to be preserved for allocations accessible | |
3653 | * by the kernel. If processes running on node 0 deplete the low memory zone | |
3654 | * then reclaim will occur more frequency increasing stalls and potentially | |
3655 | * be easier to OOM if a large percentage of the zone is under writeback or | |
3656 | * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set. | |
3657 | * Hence, default to zone ordering on 32-bit. | |
3658 | */ | |
3659 | static int default_zonelist_order(void) | |
3660 | { | |
3661 | return ZONELIST_ORDER_ZONE; | |
3662 | } | |
3663 | #endif /* CONFIG_64BIT */ | |
3664 | ||
3665 | static void set_zonelist_order(void) | |
3666 | { | |
3667 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) | |
3668 | current_zonelist_order = default_zonelist_order(); | |
3669 | else | |
3670 | current_zonelist_order = user_zonelist_order; | |
3671 | } | |
3672 | ||
3673 | static void build_zonelists(pg_data_t *pgdat) | |
3674 | { | |
3675 | int j, node, load; | |
3676 | enum zone_type i; | |
3677 | nodemask_t used_mask; | |
3678 | int local_node, prev_node; | |
3679 | struct zonelist *zonelist; | |
3680 | int order = current_zonelist_order; | |
3681 | ||
3682 | /* initialize zonelists */ | |
3683 | for (i = 0; i < MAX_ZONELISTS; i++) { | |
3684 | zonelist = pgdat->node_zonelists + i; | |
3685 | zonelist->_zonerefs[0].zone = NULL; | |
3686 | zonelist->_zonerefs[0].zone_idx = 0; | |
3687 | } | |
3688 | ||
3689 | /* NUMA-aware ordering of nodes */ | |
3690 | local_node = pgdat->node_id; | |
3691 | load = nr_online_nodes; | |
3692 | prev_node = local_node; | |
3693 | nodes_clear(used_mask); | |
3694 | ||
3695 | memset(node_order, 0, sizeof(node_order)); | |
3696 | j = 0; | |
3697 | ||
3698 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { | |
3699 | /* | |
3700 | * We don't want to pressure a particular node. | |
3701 | * So adding penalty to the first node in same | |
3702 | * distance group to make it round-robin. | |
3703 | */ | |
3704 | if (node_distance(local_node, node) != | |
3705 | node_distance(local_node, prev_node)) | |
3706 | node_load[node] = load; | |
3707 | ||
3708 | prev_node = node; | |
3709 | load--; | |
3710 | if (order == ZONELIST_ORDER_NODE) | |
3711 | build_zonelists_in_node_order(pgdat, node); | |
3712 | else | |
3713 | node_order[j++] = node; /* remember order */ | |
3714 | } | |
3715 | ||
3716 | if (order == ZONELIST_ORDER_ZONE) { | |
3717 | /* calculate node order -- i.e., DMA last! */ | |
3718 | build_zonelists_in_zone_order(pgdat, j); | |
3719 | } | |
3720 | ||
3721 | build_thisnode_zonelists(pgdat); | |
3722 | } | |
3723 | ||
3724 | /* Construct the zonelist performance cache - see further mmzone.h */ | |
3725 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3726 | { | |
3727 | struct zonelist *zonelist; | |
3728 | struct zonelist_cache *zlc; | |
3729 | struct zoneref *z; | |
3730 | ||
3731 | zonelist = &pgdat->node_zonelists[0]; | |
3732 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; | |
3733 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
3734 | for (z = zonelist->_zonerefs; z->zone; z++) | |
3735 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); | |
3736 | } | |
3737 | ||
3738 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3739 | /* | |
3740 | * Return node id of node used for "local" allocations. | |
3741 | * I.e., first node id of first zone in arg node's generic zonelist. | |
3742 | * Used for initializing percpu 'numa_mem', which is used primarily | |
3743 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. | |
3744 | */ | |
3745 | int local_memory_node(int node) | |
3746 | { | |
3747 | struct zone *zone; | |
3748 | ||
3749 | (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), | |
3750 | gfp_zone(GFP_KERNEL), | |
3751 | NULL, | |
3752 | &zone); | |
3753 | return zone->node; | |
3754 | } | |
3755 | #endif | |
3756 | ||
3757 | #else /* CONFIG_NUMA */ | |
3758 | ||
3759 | static void set_zonelist_order(void) | |
3760 | { | |
3761 | current_zonelist_order = ZONELIST_ORDER_ZONE; | |
3762 | } | |
3763 | ||
3764 | static void build_zonelists(pg_data_t *pgdat) | |
3765 | { | |
3766 | int node, local_node; | |
3767 | enum zone_type j; | |
3768 | struct zonelist *zonelist; | |
3769 | ||
3770 | local_node = pgdat->node_id; | |
3771 | ||
3772 | zonelist = &pgdat->node_zonelists[0]; | |
3773 | j = build_zonelists_node(pgdat, zonelist, 0); | |
3774 | ||
3775 | /* | |
3776 | * Now we build the zonelist so that it contains the zones | |
3777 | * of all the other nodes. | |
3778 | * We don't want to pressure a particular node, so when | |
3779 | * building the zones for node N, we make sure that the | |
3780 | * zones coming right after the local ones are those from | |
3781 | * node N+1 (modulo N) | |
3782 | */ | |
3783 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
3784 | if (!node_online(node)) | |
3785 | continue; | |
3786 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); | |
3787 | } | |
3788 | for (node = 0; node < local_node; node++) { | |
3789 | if (!node_online(node)) | |
3790 | continue; | |
3791 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); | |
3792 | } | |
3793 | ||
3794 | zonelist->_zonerefs[j].zone = NULL; | |
3795 | zonelist->_zonerefs[j].zone_idx = 0; | |
3796 | } | |
3797 | ||
3798 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ | |
3799 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3800 | { | |
3801 | pgdat->node_zonelists[0].zlcache_ptr = NULL; | |
3802 | } | |
3803 | ||
3804 | #endif /* CONFIG_NUMA */ | |
3805 | ||
3806 | /* | |
3807 | * Boot pageset table. One per cpu which is going to be used for all | |
3808 | * zones and all nodes. The parameters will be set in such a way | |
3809 | * that an item put on a list will immediately be handed over to | |
3810 | * the buddy list. This is safe since pageset manipulation is done | |
3811 | * with interrupts disabled. | |
3812 | * | |
3813 | * The boot_pagesets must be kept even after bootup is complete for | |
3814 | * unused processors and/or zones. They do play a role for bootstrapping | |
3815 | * hotplugged processors. | |
3816 | * | |
3817 | * zoneinfo_show() and maybe other functions do | |
3818 | * not check if the processor is online before following the pageset pointer. | |
3819 | * Other parts of the kernel may not check if the zone is available. | |
3820 | */ | |
3821 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); | |
3822 | static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); | |
3823 | static void setup_zone_pageset(struct zone *zone); | |
3824 | ||
3825 | /* | |
3826 | * Global mutex to protect against size modification of zonelists | |
3827 | * as well as to serialize pageset setup for the new populated zone. | |
3828 | */ | |
3829 | DEFINE_MUTEX(zonelists_mutex); | |
3830 | ||
3831 | /* return values int ....just for stop_machine() */ | |
3832 | static int __build_all_zonelists(void *data) | |
3833 | { | |
3834 | int nid; | |
3835 | int cpu; | |
3836 | pg_data_t *self = data; | |
3837 | ||
3838 | #ifdef CONFIG_NUMA | |
3839 | memset(node_load, 0, sizeof(node_load)); | |
3840 | #endif | |
3841 | ||
3842 | if (self && !node_online(self->node_id)) { | |
3843 | build_zonelists(self); | |
3844 | build_zonelist_cache(self); | |
3845 | } | |
3846 | ||
3847 | for_each_online_node(nid) { | |
3848 | pg_data_t *pgdat = NODE_DATA(nid); | |
3849 | ||
3850 | build_zonelists(pgdat); | |
3851 | build_zonelist_cache(pgdat); | |
3852 | } | |
3853 | ||
3854 | /* | |
3855 | * Initialize the boot_pagesets that are going to be used | |
3856 | * for bootstrapping processors. The real pagesets for | |
3857 | * each zone will be allocated later when the per cpu | |
3858 | * allocator is available. | |
3859 | * | |
3860 | * boot_pagesets are used also for bootstrapping offline | |
3861 | * cpus if the system is already booted because the pagesets | |
3862 | * are needed to initialize allocators on a specific cpu too. | |
3863 | * F.e. the percpu allocator needs the page allocator which | |
3864 | * needs the percpu allocator in order to allocate its pagesets | |
3865 | * (a chicken-egg dilemma). | |
3866 | */ | |
3867 | for_each_possible_cpu(cpu) { | |
3868 | setup_pageset(&per_cpu(boot_pageset, cpu), 0); | |
3869 | ||
3870 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3871 | /* | |
3872 | * We now know the "local memory node" for each node-- | |
3873 | * i.e., the node of the first zone in the generic zonelist. | |
3874 | * Set up numa_mem percpu variable for on-line cpus. During | |
3875 | * boot, only the boot cpu should be on-line; we'll init the | |
3876 | * secondary cpus' numa_mem as they come on-line. During | |
3877 | * node/memory hotplug, we'll fixup all on-line cpus. | |
3878 | */ | |
3879 | if (cpu_online(cpu)) | |
3880 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); | |
3881 | #endif | |
3882 | } | |
3883 | ||
3884 | return 0; | |
3885 | } | |
3886 | ||
3887 | static noinline void __init | |
3888 | build_all_zonelists_init(void) | |
3889 | { | |
3890 | __build_all_zonelists(NULL); | |
3891 | mminit_verify_zonelist(); | |
3892 | cpuset_init_current_mems_allowed(); | |
3893 | } | |
3894 | ||
3895 | /* | |
3896 | * Called with zonelists_mutex held always | |
3897 | * unless system_state == SYSTEM_BOOTING. | |
3898 | * | |
3899 | * __ref due to (1) call of __meminit annotated setup_zone_pageset | |
3900 | * [we're only called with non-NULL zone through __meminit paths] and | |
3901 | * (2) call of __init annotated helper build_all_zonelists_init | |
3902 | * [protected by SYSTEM_BOOTING]. | |
3903 | */ | |
3904 | void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) | |
3905 | { | |
3906 | set_zonelist_order(); | |
3907 | ||
3908 | if (system_state == SYSTEM_BOOTING) { | |
3909 | build_all_zonelists_init(); | |
3910 | } else { | |
3911 | #ifdef CONFIG_MEMORY_HOTPLUG | |
3912 | if (zone) | |
3913 | setup_zone_pageset(zone); | |
3914 | #endif | |
3915 | /* we have to stop all cpus to guarantee there is no user | |
3916 | of zonelist */ | |
3917 | stop_machine(__build_all_zonelists, pgdat, NULL); | |
3918 | /* cpuset refresh routine should be here */ | |
3919 | } | |
3920 | vm_total_pages = nr_free_pagecache_pages(); | |
3921 | /* | |
3922 | * Disable grouping by mobility if the number of pages in the | |
3923 | * system is too low to allow the mechanism to work. It would be | |
3924 | * more accurate, but expensive to check per-zone. This check is | |
3925 | * made on memory-hotadd so a system can start with mobility | |
3926 | * disabled and enable it later | |
3927 | */ | |
3928 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) | |
3929 | page_group_by_mobility_disabled = 1; | |
3930 | else | |
3931 | page_group_by_mobility_disabled = 0; | |
3932 | ||
3933 | pr_info("Built %i zonelists in %s order, mobility grouping %s. " | |
3934 | "Total pages: %ld\n", | |
3935 | nr_online_nodes, | |
3936 | zonelist_order_name[current_zonelist_order], | |
3937 | page_group_by_mobility_disabled ? "off" : "on", | |
3938 | vm_total_pages); | |
3939 | #ifdef CONFIG_NUMA | |
3940 | pr_info("Policy zone: %s\n", zone_names[policy_zone]); | |
3941 | #endif | |
3942 | } | |
3943 | ||
3944 | /* | |
3945 | * Helper functions to size the waitqueue hash table. | |
3946 | * Essentially these want to choose hash table sizes sufficiently | |
3947 | * large so that collisions trying to wait on pages are rare. | |
3948 | * But in fact, the number of active page waitqueues on typical | |
3949 | * systems is ridiculously low, less than 200. So this is even | |
3950 | * conservative, even though it seems large. | |
3951 | * | |
3952 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to | |
3953 | * waitqueues, i.e. the size of the waitq table given the number of pages. | |
3954 | */ | |
3955 | #define PAGES_PER_WAITQUEUE 256 | |
3956 | ||
3957 | #ifndef CONFIG_MEMORY_HOTPLUG | |
3958 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3959 | { | |
3960 | unsigned long size = 1; | |
3961 | ||
3962 | pages /= PAGES_PER_WAITQUEUE; | |
3963 | ||
3964 | while (size < pages) | |
3965 | size <<= 1; | |
3966 | ||
3967 | /* | |
3968 | * Once we have dozens or even hundreds of threads sleeping | |
3969 | * on IO we've got bigger problems than wait queue collision. | |
3970 | * Limit the size of the wait table to a reasonable size. | |
3971 | */ | |
3972 | size = min(size, 4096UL); | |
3973 | ||
3974 | return max(size, 4UL); | |
3975 | } | |
3976 | #else | |
3977 | /* | |
3978 | * A zone's size might be changed by hot-add, so it is not possible to determine | |
3979 | * a suitable size for its wait_table. So we use the maximum size now. | |
3980 | * | |
3981 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: | |
3982 | * | |
3983 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. | |
3984 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. | |
3985 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. | |
3986 | * | |
3987 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages | |
3988 | * or more by the traditional way. (See above). It equals: | |
3989 | * | |
3990 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. | |
3991 | * ia64(16K page size) : = ( 8G + 4M)byte. | |
3992 | * powerpc (64K page size) : = (32G +16M)byte. | |
3993 | */ | |
3994 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3995 | { | |
3996 | return 4096UL; | |
3997 | } | |
3998 | #endif | |
3999 | ||
4000 | /* | |
4001 | * This is an integer logarithm so that shifts can be used later | |
4002 | * to extract the more random high bits from the multiplicative | |
4003 | * hash function before the remainder is taken. | |
4004 | */ | |
4005 | static inline unsigned long wait_table_bits(unsigned long size) | |
4006 | { | |
4007 | return ffz(~size); | |
4008 | } | |
4009 | ||
4010 | /* | |
4011 | * Check if a pageblock contains reserved pages | |
4012 | */ | |
4013 | static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) | |
4014 | { | |
4015 | unsigned long pfn; | |
4016 | ||
4017 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
4018 | if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) | |
4019 | return 1; | |
4020 | } | |
4021 | return 0; | |
4022 | } | |
4023 | ||
4024 | /* | |
4025 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number | |
4026 | * of blocks reserved is based on min_wmark_pages(zone). The memory within | |
4027 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes | |
4028 | * higher will lead to a bigger reserve which will get freed as contiguous | |
4029 | * blocks as reclaim kicks in | |
4030 | */ | |
4031 | static void setup_zone_migrate_reserve(struct zone *zone) | |
4032 | { | |
4033 | unsigned long start_pfn, pfn, end_pfn, block_end_pfn; | |
4034 | struct page *page; | |
4035 | unsigned long block_migratetype; | |
4036 | int reserve; | |
4037 | int old_reserve; | |
4038 | ||
4039 | /* | |
4040 | * Get the start pfn, end pfn and the number of blocks to reserve | |
4041 | * We have to be careful to be aligned to pageblock_nr_pages to | |
4042 | * make sure that we always check pfn_valid for the first page in | |
4043 | * the block. | |
4044 | */ | |
4045 | start_pfn = zone->zone_start_pfn; | |
4046 | end_pfn = zone_end_pfn(zone); | |
4047 | start_pfn = roundup(start_pfn, pageblock_nr_pages); | |
4048 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> | |
4049 | pageblock_order; | |
4050 | ||
4051 | /* | |
4052 | * Reserve blocks are generally in place to help high-order atomic | |
4053 | * allocations that are short-lived. A min_free_kbytes value that | |
4054 | * would result in more than 2 reserve blocks for atomic allocations | |
4055 | * is assumed to be in place to help anti-fragmentation for the | |
4056 | * future allocation of hugepages at runtime. | |
4057 | */ | |
4058 | reserve = min(2, reserve); | |
4059 | old_reserve = zone->nr_migrate_reserve_block; | |
4060 | ||
4061 | /* When memory hot-add, we almost always need to do nothing */ | |
4062 | if (reserve == old_reserve) | |
4063 | return; | |
4064 | zone->nr_migrate_reserve_block = reserve; | |
4065 | ||
4066 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
4067 | if (!pfn_valid(pfn)) | |
4068 | continue; | |
4069 | page = pfn_to_page(pfn); | |
4070 | ||
4071 | /* Watch out for overlapping nodes */ | |
4072 | if (page_to_nid(page) != zone_to_nid(zone)) | |
4073 | continue; | |
4074 | ||
4075 | block_migratetype = get_pageblock_migratetype(page); | |
4076 | ||
4077 | /* Only test what is necessary when the reserves are not met */ | |
4078 | if (reserve > 0) { | |
4079 | /* | |
4080 | * Blocks with reserved pages will never free, skip | |
4081 | * them. | |
4082 | */ | |
4083 | block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); | |
4084 | if (pageblock_is_reserved(pfn, block_end_pfn)) | |
4085 | continue; | |
4086 | ||
4087 | /* If this block is reserved, account for it */ | |
4088 | if (block_migratetype == MIGRATE_RESERVE) { | |
4089 | reserve--; | |
4090 | continue; | |
4091 | } | |
4092 | ||
4093 | /* Suitable for reserving if this block is movable */ | |
4094 | if (block_migratetype == MIGRATE_MOVABLE) { | |
4095 | set_pageblock_migratetype(page, | |
4096 | MIGRATE_RESERVE); | |
4097 | move_freepages_block(zone, page, | |
4098 | MIGRATE_RESERVE); | |
4099 | reserve--; | |
4100 | continue; | |
4101 | } | |
4102 | } else if (!old_reserve) { | |
4103 | /* | |
4104 | * At boot time we don't need to scan the whole zone | |
4105 | * for turning off MIGRATE_RESERVE. | |
4106 | */ | |
4107 | break; | |
4108 | } | |
4109 | ||
4110 | /* | |
4111 | * If the reserve is met and this is a previous reserved block, | |
4112 | * take it back | |
4113 | */ | |
4114 | if (block_migratetype == MIGRATE_RESERVE) { | |
4115 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
4116 | move_freepages_block(zone, page, MIGRATE_MOVABLE); | |
4117 | } | |
4118 | } | |
4119 | } | |
4120 | ||
4121 | /* | |
4122 | * Initially all pages are reserved - free ones are freed | |
4123 | * up by free_all_bootmem() once the early boot process is | |
4124 | * done. Non-atomic initialization, single-pass. | |
4125 | */ | |
4126 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, | |
4127 | unsigned long start_pfn, enum memmap_context context) | |
4128 | { | |
4129 | struct page *page; | |
4130 | unsigned long end_pfn = start_pfn + size; | |
4131 | unsigned long pfn; | |
4132 | struct zone *z; | |
4133 | ||
4134 | if (highest_memmap_pfn < end_pfn - 1) | |
4135 | highest_memmap_pfn = end_pfn - 1; | |
4136 | ||
4137 | z = &NODE_DATA(nid)->node_zones[zone]; | |
4138 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
4139 | /* | |
4140 | * There can be holes in boot-time mem_map[]s | |
4141 | * handed to this function. They do not | |
4142 | * exist on hotplugged memory. | |
4143 | */ | |
4144 | if (context == MEMMAP_EARLY) { | |
4145 | if (!early_pfn_valid(pfn)) | |
4146 | continue; | |
4147 | if (!early_pfn_in_nid(pfn, nid)) | |
4148 | continue; | |
4149 | } | |
4150 | page = pfn_to_page(pfn); | |
4151 | set_page_links(page, zone, nid, pfn); | |
4152 | mminit_verify_page_links(page, zone, nid, pfn); | |
4153 | init_page_count(page); | |
4154 | page_mapcount_reset(page); | |
4155 | page_cpupid_reset_last(page); | |
4156 | SetPageReserved(page); | |
4157 | /* | |
4158 | * Mark the block movable so that blocks are reserved for | |
4159 | * movable at startup. This will force kernel allocations | |
4160 | * to reserve their blocks rather than leaking throughout | |
4161 | * the address space during boot when many long-lived | |
4162 | * kernel allocations are made. Later some blocks near | |
4163 | * the start are marked MIGRATE_RESERVE by | |
4164 | * setup_zone_migrate_reserve() | |
4165 | * | |
4166 | * bitmap is created for zone's valid pfn range. but memmap | |
4167 | * can be created for invalid pages (for alignment) | |
4168 | * check here not to call set_pageblock_migratetype() against | |
4169 | * pfn out of zone. | |
4170 | */ | |
4171 | if ((z->zone_start_pfn <= pfn) | |
4172 | && (pfn < zone_end_pfn(z)) | |
4173 | && !(pfn & (pageblock_nr_pages - 1))) | |
4174 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
4175 | ||
4176 | INIT_LIST_HEAD(&page->lru); | |
4177 | #ifdef WANT_PAGE_VIRTUAL | |
4178 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ | |
4179 | if (!is_highmem_idx(zone)) | |
4180 | set_page_address(page, __va(pfn << PAGE_SHIFT)); | |
4181 | #endif | |
4182 | } | |
4183 | } | |
4184 | ||
4185 | static void __meminit zone_init_free_lists(struct zone *zone) | |
4186 | { | |
4187 | unsigned int order, t; | |
4188 | for_each_migratetype_order(order, t) { | |
4189 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); | |
4190 | zone->free_area[order].nr_free = 0; | |
4191 | } | |
4192 | } | |
4193 | ||
4194 | #ifndef __HAVE_ARCH_MEMMAP_INIT | |
4195 | #define memmap_init(size, nid, zone, start_pfn) \ | |
4196 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) | |
4197 | #endif | |
4198 | ||
4199 | static int zone_batchsize(struct zone *zone) | |
4200 | { | |
4201 | #ifdef CONFIG_MMU | |
4202 | int batch; | |
4203 | ||
4204 | /* | |
4205 | * The per-cpu-pages pools are set to around 1000th of the | |
4206 | * size of the zone. But no more than 1/2 of a meg. | |
4207 | * | |
4208 | * OK, so we don't know how big the cache is. So guess. | |
4209 | */ | |
4210 | batch = zone->managed_pages / 1024; | |
4211 | if (batch * PAGE_SIZE > 512 * 1024) | |
4212 | batch = (512 * 1024) / PAGE_SIZE; | |
4213 | batch /= 4; /* We effectively *= 4 below */ | |
4214 | if (batch < 1) | |
4215 | batch = 1; | |
4216 | ||
4217 | /* | |
4218 | * Clamp the batch to a 2^n - 1 value. Having a power | |
4219 | * of 2 value was found to be more likely to have | |
4220 | * suboptimal cache aliasing properties in some cases. | |
4221 | * | |
4222 | * For example if 2 tasks are alternately allocating | |
4223 | * batches of pages, one task can end up with a lot | |
4224 | * of pages of one half of the possible page colors | |
4225 | * and the other with pages of the other colors. | |
4226 | */ | |
4227 | batch = rounddown_pow_of_two(batch + batch/2) - 1; | |
4228 | ||
4229 | return batch; | |
4230 | ||
4231 | #else | |
4232 | /* The deferral and batching of frees should be suppressed under NOMMU | |
4233 | * conditions. | |
4234 | * | |
4235 | * The problem is that NOMMU needs to be able to allocate large chunks | |
4236 | * of contiguous memory as there's no hardware page translation to | |
4237 | * assemble apparent contiguous memory from discontiguous pages. | |
4238 | * | |
4239 | * Queueing large contiguous runs of pages for batching, however, | |
4240 | * causes the pages to actually be freed in smaller chunks. As there | |
4241 | * can be a significant delay between the individual batches being | |
4242 | * recycled, this leads to the once large chunks of space being | |
4243 | * fragmented and becoming unavailable for high-order allocations. | |
4244 | */ | |
4245 | return 0; | |
4246 | #endif | |
4247 | } | |
4248 | ||
4249 | /* | |
4250 | * pcp->high and pcp->batch values are related and dependent on one another: | |
4251 | * ->batch must never be higher then ->high. | |
4252 | * The following function updates them in a safe manner without read side | |
4253 | * locking. | |
4254 | * | |
4255 | * Any new users of pcp->batch and pcp->high should ensure they can cope with | |
4256 | * those fields changing asynchronously (acording the the above rule). | |
4257 | * | |
4258 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function | |
4259 | * outside of boot time (or some other assurance that no concurrent updaters | |
4260 | * exist). | |
4261 | */ | |
4262 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, | |
4263 | unsigned long batch) | |
4264 | { | |
4265 | /* start with a fail safe value for batch */ | |
4266 | pcp->batch = 1; | |
4267 | smp_wmb(); | |
4268 | ||
4269 | /* Update high, then batch, in order */ | |
4270 | pcp->high = high; | |
4271 | smp_wmb(); | |
4272 | ||
4273 | pcp->batch = batch; | |
4274 | } | |
4275 | ||
4276 | /* a companion to pageset_set_high() */ | |
4277 | static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) | |
4278 | { | |
4279 | pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); | |
4280 | } | |
4281 | ||
4282 | static void pageset_init(struct per_cpu_pageset *p) | |
4283 | { | |
4284 | struct per_cpu_pages *pcp; | |
4285 | int migratetype; | |
4286 | ||
4287 | memset(p, 0, sizeof(*p)); | |
4288 | ||
4289 | pcp = &p->pcp; | |
4290 | pcp->count = 0; | |
4291 | for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) | |
4292 | INIT_LIST_HEAD(&pcp->lists[migratetype]); | |
4293 | } | |
4294 | ||
4295 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) | |
4296 | { | |
4297 | pageset_init(p); | |
4298 | pageset_set_batch(p, batch); | |
4299 | } | |
4300 | ||
4301 | /* | |
4302 | * pageset_set_high() sets the high water mark for hot per_cpu_pagelist | |
4303 | * to the value high for the pageset p. | |
4304 | */ | |
4305 | static void pageset_set_high(struct per_cpu_pageset *p, | |
4306 | unsigned long high) | |
4307 | { | |
4308 | unsigned long batch = max(1UL, high / 4); | |
4309 | if ((high / 4) > (PAGE_SHIFT * 8)) | |
4310 | batch = PAGE_SHIFT * 8; | |
4311 | ||
4312 | pageset_update(&p->pcp, high, batch); | |
4313 | } | |
4314 | ||
4315 | static void pageset_set_high_and_batch(struct zone *zone, | |
4316 | struct per_cpu_pageset *pcp) | |
4317 | { | |
4318 | if (percpu_pagelist_fraction) | |
4319 | pageset_set_high(pcp, | |
4320 | (zone->managed_pages / | |
4321 | percpu_pagelist_fraction)); | |
4322 | else | |
4323 | pageset_set_batch(pcp, zone_batchsize(zone)); | |
4324 | } | |
4325 | ||
4326 | static void __meminit zone_pageset_init(struct zone *zone, int cpu) | |
4327 | { | |
4328 | struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); | |
4329 | ||
4330 | pageset_init(pcp); | |
4331 | pageset_set_high_and_batch(zone, pcp); | |
4332 | } | |
4333 | ||
4334 | static void __meminit setup_zone_pageset(struct zone *zone) | |
4335 | { | |
4336 | int cpu; | |
4337 | zone->pageset = alloc_percpu(struct per_cpu_pageset); | |
4338 | for_each_possible_cpu(cpu) | |
4339 | zone_pageset_init(zone, cpu); | |
4340 | } | |
4341 | ||
4342 | /* | |
4343 | * Allocate per cpu pagesets and initialize them. | |
4344 | * Before this call only boot pagesets were available. | |
4345 | */ | |
4346 | void __init setup_per_cpu_pageset(void) | |
4347 | { | |
4348 | struct zone *zone; | |
4349 | ||
4350 | for_each_populated_zone(zone) | |
4351 | setup_zone_pageset(zone); | |
4352 | } | |
4353 | ||
4354 | static noinline __init_refok | |
4355 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) | |
4356 | { | |
4357 | int i; | |
4358 | size_t alloc_size; | |
4359 | ||
4360 | /* | |
4361 | * The per-page waitqueue mechanism uses hashed waitqueues | |
4362 | * per zone. | |
4363 | */ | |
4364 | zone->wait_table_hash_nr_entries = | |
4365 | wait_table_hash_nr_entries(zone_size_pages); | |
4366 | zone->wait_table_bits = | |
4367 | wait_table_bits(zone->wait_table_hash_nr_entries); | |
4368 | alloc_size = zone->wait_table_hash_nr_entries | |
4369 | * sizeof(wait_queue_head_t); | |
4370 | ||
4371 | if (!slab_is_available()) { | |
4372 | zone->wait_table = (wait_queue_head_t *) | |
4373 | memblock_virt_alloc_node_nopanic( | |
4374 | alloc_size, zone->zone_pgdat->node_id); | |
4375 | } else { | |
4376 | /* | |
4377 | * This case means that a zone whose size was 0 gets new memory | |
4378 | * via memory hot-add. | |
4379 | * But it may be the case that a new node was hot-added. In | |
4380 | * this case vmalloc() will not be able to use this new node's | |
4381 | * memory - this wait_table must be initialized to use this new | |
4382 | * node itself as well. | |
4383 | * To use this new node's memory, further consideration will be | |
4384 | * necessary. | |
4385 | */ | |
4386 | zone->wait_table = vmalloc(alloc_size); | |
4387 | } | |
4388 | if (!zone->wait_table) | |
4389 | return -ENOMEM; | |
4390 | ||
4391 | for (i = 0; i < zone->wait_table_hash_nr_entries; ++i) | |
4392 | init_waitqueue_head(zone->wait_table + i); | |
4393 | ||
4394 | return 0; | |
4395 | } | |
4396 | ||
4397 | static __meminit void zone_pcp_init(struct zone *zone) | |
4398 | { | |
4399 | /* | |
4400 | * per cpu subsystem is not up at this point. The following code | |
4401 | * relies on the ability of the linker to provide the | |
4402 | * offset of a (static) per cpu variable into the per cpu area. | |
4403 | */ | |
4404 | zone->pageset = &boot_pageset; | |
4405 | ||
4406 | if (populated_zone(zone)) | |
4407 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", | |
4408 | zone->name, zone->present_pages, | |
4409 | zone_batchsize(zone)); | |
4410 | } | |
4411 | ||
4412 | int __meminit init_currently_empty_zone(struct zone *zone, | |
4413 | unsigned long zone_start_pfn, | |
4414 | unsigned long size, | |
4415 | enum memmap_context context) | |
4416 | { | |
4417 | struct pglist_data *pgdat = zone->zone_pgdat; | |
4418 | int ret; | |
4419 | ret = zone_wait_table_init(zone, size); | |
4420 | if (ret) | |
4421 | return ret; | |
4422 | pgdat->nr_zones = zone_idx(zone) + 1; | |
4423 | ||
4424 | zone->zone_start_pfn = zone_start_pfn; | |
4425 | ||
4426 | mminit_dprintk(MMINIT_TRACE, "memmap_init", | |
4427 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", | |
4428 | pgdat->node_id, | |
4429 | (unsigned long)zone_idx(zone), | |
4430 | zone_start_pfn, (zone_start_pfn + size)); | |
4431 | ||
4432 | zone_init_free_lists(zone); | |
4433 | ||
4434 | return 0; | |
4435 | } | |
4436 | ||
4437 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4438 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID | |
4439 | /* | |
4440 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. | |
4441 | */ | |
4442 | int __meminit __early_pfn_to_nid(unsigned long pfn) | |
4443 | { | |
4444 | unsigned long start_pfn, end_pfn; | |
4445 | int nid; | |
4446 | /* | |
4447 | * NOTE: The following SMP-unsafe globals are only used early in boot | |
4448 | * when the kernel is running single-threaded. | |
4449 | */ | |
4450 | static unsigned long __meminitdata last_start_pfn, last_end_pfn; | |
4451 | static int __meminitdata last_nid; | |
4452 | ||
4453 | if (last_start_pfn <= pfn && pfn < last_end_pfn) | |
4454 | return last_nid; | |
4455 | ||
4456 | nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); | |
4457 | if (nid != -1) { | |
4458 | last_start_pfn = start_pfn; | |
4459 | last_end_pfn = end_pfn; | |
4460 | last_nid = nid; | |
4461 | } | |
4462 | ||
4463 | return nid; | |
4464 | } | |
4465 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ | |
4466 | ||
4467 | int __meminit early_pfn_to_nid(unsigned long pfn) | |
4468 | { | |
4469 | int nid; | |
4470 | ||
4471 | nid = __early_pfn_to_nid(pfn); | |
4472 | if (nid >= 0) | |
4473 | return nid; | |
4474 | /* just returns 0 */ | |
4475 | return 0; | |
4476 | } | |
4477 | ||
4478 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES | |
4479 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) | |
4480 | { | |
4481 | int nid; | |
4482 | ||
4483 | nid = __early_pfn_to_nid(pfn); | |
4484 | if (nid >= 0 && nid != node) | |
4485 | return false; | |
4486 | return true; | |
4487 | } | |
4488 | #endif | |
4489 | ||
4490 | /** | |
4491 | * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range | |
4492 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. | |
4493 | * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid | |
4494 | * | |
4495 | * If an architecture guarantees that all ranges registered contain no holes | |
4496 | * and may be freed, this this function may be used instead of calling | |
4497 | * memblock_free_early_nid() manually. | |
4498 | */ | |
4499 | void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) | |
4500 | { | |
4501 | unsigned long start_pfn, end_pfn; | |
4502 | int i, this_nid; | |
4503 | ||
4504 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { | |
4505 | start_pfn = min(start_pfn, max_low_pfn); | |
4506 | end_pfn = min(end_pfn, max_low_pfn); | |
4507 | ||
4508 | if (start_pfn < end_pfn) | |
4509 | memblock_free_early_nid(PFN_PHYS(start_pfn), | |
4510 | (end_pfn - start_pfn) << PAGE_SHIFT, | |
4511 | this_nid); | |
4512 | } | |
4513 | } | |
4514 | ||
4515 | /** | |
4516 | * sparse_memory_present_with_active_regions - Call memory_present for each active range | |
4517 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. | |
4518 | * | |
4519 | * If an architecture guarantees that all ranges registered contain no holes and may | |
4520 | * be freed, this function may be used instead of calling memory_present() manually. | |
4521 | */ | |
4522 | void __init sparse_memory_present_with_active_regions(int nid) | |
4523 | { | |
4524 | unsigned long start_pfn, end_pfn; | |
4525 | int i, this_nid; | |
4526 | ||
4527 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) | |
4528 | memory_present(this_nid, start_pfn, end_pfn); | |
4529 | } | |
4530 | ||
4531 | /** | |
4532 | * get_pfn_range_for_nid - Return the start and end page frames for a node | |
4533 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. | |
4534 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. | |
4535 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. | |
4536 | * | |
4537 | * It returns the start and end page frame of a node based on information | |
4538 | * provided by memblock_set_node(). If called for a node | |
4539 | * with no available memory, a warning is printed and the start and end | |
4540 | * PFNs will be 0. | |
4541 | */ | |
4542 | void __meminit get_pfn_range_for_nid(unsigned int nid, | |
4543 | unsigned long *start_pfn, unsigned long *end_pfn) | |
4544 | { | |
4545 | unsigned long this_start_pfn, this_end_pfn; | |
4546 | int i; | |
4547 | ||
4548 | *start_pfn = -1UL; | |
4549 | *end_pfn = 0; | |
4550 | ||
4551 | for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { | |
4552 | *start_pfn = min(*start_pfn, this_start_pfn); | |
4553 | *end_pfn = max(*end_pfn, this_end_pfn); | |
4554 | } | |
4555 | ||
4556 | if (*start_pfn == -1UL) | |
4557 | *start_pfn = 0; | |
4558 | } | |
4559 | ||
4560 | /* | |
4561 | * This finds a zone that can be used for ZONE_MOVABLE pages. The | |
4562 | * assumption is made that zones within a node are ordered in monotonic | |
4563 | * increasing memory addresses so that the "highest" populated zone is used | |
4564 | */ | |
4565 | static void __init find_usable_zone_for_movable(void) | |
4566 | { | |
4567 | int zone_index; | |
4568 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { | |
4569 | if (zone_index == ZONE_MOVABLE) | |
4570 | continue; | |
4571 | ||
4572 | if (arch_zone_highest_possible_pfn[zone_index] > | |
4573 | arch_zone_lowest_possible_pfn[zone_index]) | |
4574 | break; | |
4575 | } | |
4576 | ||
4577 | VM_BUG_ON(zone_index == -1); | |
4578 | movable_zone = zone_index; | |
4579 | } | |
4580 | ||
4581 | /* | |
4582 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE | |
4583 | * because it is sized independent of architecture. Unlike the other zones, | |
4584 | * the starting point for ZONE_MOVABLE is not fixed. It may be different | |
4585 | * in each node depending on the size of each node and how evenly kernelcore | |
4586 | * is distributed. This helper function adjusts the zone ranges | |
4587 | * provided by the architecture for a given node by using the end of the | |
4588 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that | |
4589 | * zones within a node are in order of monotonic increases memory addresses | |
4590 | */ | |
4591 | static void __meminit adjust_zone_range_for_zone_movable(int nid, | |
4592 | unsigned long zone_type, | |
4593 | unsigned long node_start_pfn, | |
4594 | unsigned long node_end_pfn, | |
4595 | unsigned long *zone_start_pfn, | |
4596 | unsigned long *zone_end_pfn) | |
4597 | { | |
4598 | /* Only adjust if ZONE_MOVABLE is on this node */ | |
4599 | if (zone_movable_pfn[nid]) { | |
4600 | /* Size ZONE_MOVABLE */ | |
4601 | if (zone_type == ZONE_MOVABLE) { | |
4602 | *zone_start_pfn = zone_movable_pfn[nid]; | |
4603 | *zone_end_pfn = min(node_end_pfn, | |
4604 | arch_zone_highest_possible_pfn[movable_zone]); | |
4605 | ||
4606 | /* Adjust for ZONE_MOVABLE starting within this range */ | |
4607 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && | |
4608 | *zone_end_pfn > zone_movable_pfn[nid]) { | |
4609 | *zone_end_pfn = zone_movable_pfn[nid]; | |
4610 | ||
4611 | /* Check if this whole range is within ZONE_MOVABLE */ | |
4612 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) | |
4613 | *zone_start_pfn = *zone_end_pfn; | |
4614 | } | |
4615 | } | |
4616 | ||
4617 | /* | |
4618 | * Return the number of pages a zone spans in a node, including holes | |
4619 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() | |
4620 | */ | |
4621 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4622 | unsigned long zone_type, | |
4623 | unsigned long node_start_pfn, | |
4624 | unsigned long node_end_pfn, | |
4625 | unsigned long *ignored) | |
4626 | { | |
4627 | unsigned long zone_start_pfn, zone_end_pfn; | |
4628 | ||
4629 | /* Get the start and end of the zone */ | |
4630 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; | |
4631 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; | |
4632 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4633 | node_start_pfn, node_end_pfn, | |
4634 | &zone_start_pfn, &zone_end_pfn); | |
4635 | ||
4636 | /* Check that this node has pages within the zone's required range */ | |
4637 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) | |
4638 | return 0; | |
4639 | ||
4640 | /* Move the zone boundaries inside the node if necessary */ | |
4641 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); | |
4642 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); | |
4643 | ||
4644 | /* Return the spanned pages */ | |
4645 | return zone_end_pfn - zone_start_pfn; | |
4646 | } | |
4647 | ||
4648 | /* | |
4649 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, | |
4650 | * then all holes in the requested range will be accounted for. | |
4651 | */ | |
4652 | unsigned long __meminit __absent_pages_in_range(int nid, | |
4653 | unsigned long range_start_pfn, | |
4654 | unsigned long range_end_pfn) | |
4655 | { | |
4656 | unsigned long nr_absent = range_end_pfn - range_start_pfn; | |
4657 | unsigned long start_pfn, end_pfn; | |
4658 | int i; | |
4659 | ||
4660 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
4661 | start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); | |
4662 | end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); | |
4663 | nr_absent -= end_pfn - start_pfn; | |
4664 | } | |
4665 | return nr_absent; | |
4666 | } | |
4667 | ||
4668 | /** | |
4669 | * absent_pages_in_range - Return number of page frames in holes within a range | |
4670 | * @start_pfn: The start PFN to start searching for holes | |
4671 | * @end_pfn: The end PFN to stop searching for holes | |
4672 | * | |
4673 | * It returns the number of pages frames in memory holes within a range. | |
4674 | */ | |
4675 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, | |
4676 | unsigned long end_pfn) | |
4677 | { | |
4678 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); | |
4679 | } | |
4680 | ||
4681 | /* Return the number of page frames in holes in a zone on a node */ | |
4682 | static unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4683 | unsigned long zone_type, | |
4684 | unsigned long node_start_pfn, | |
4685 | unsigned long node_end_pfn, | |
4686 | unsigned long *ignored) | |
4687 | { | |
4688 | unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; | |
4689 | unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; | |
4690 | unsigned long zone_start_pfn, zone_end_pfn; | |
4691 | ||
4692 | zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); | |
4693 | zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); | |
4694 | ||
4695 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4696 | node_start_pfn, node_end_pfn, | |
4697 | &zone_start_pfn, &zone_end_pfn); | |
4698 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); | |
4699 | } | |
4700 | ||
4701 | #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4702 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4703 | unsigned long zone_type, | |
4704 | unsigned long node_start_pfn, | |
4705 | unsigned long node_end_pfn, | |
4706 | unsigned long *zones_size) | |
4707 | { | |
4708 | return zones_size[zone_type]; | |
4709 | } | |
4710 | ||
4711 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4712 | unsigned long zone_type, | |
4713 | unsigned long node_start_pfn, | |
4714 | unsigned long node_end_pfn, | |
4715 | unsigned long *zholes_size) | |
4716 | { | |
4717 | if (!zholes_size) | |
4718 | return 0; | |
4719 | ||
4720 | return zholes_size[zone_type]; | |
4721 | } | |
4722 | ||
4723 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4724 | ||
4725 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, | |
4726 | unsigned long node_start_pfn, | |
4727 | unsigned long node_end_pfn, | |
4728 | unsigned long *zones_size, | |
4729 | unsigned long *zholes_size) | |
4730 | { | |
4731 | unsigned long realtotalpages, totalpages = 0; | |
4732 | enum zone_type i; | |
4733 | ||
4734 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4735 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, | |
4736 | node_start_pfn, | |
4737 | node_end_pfn, | |
4738 | zones_size); | |
4739 | pgdat->node_spanned_pages = totalpages; | |
4740 | ||
4741 | realtotalpages = totalpages; | |
4742 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4743 | realtotalpages -= | |
4744 | zone_absent_pages_in_node(pgdat->node_id, i, | |
4745 | node_start_pfn, node_end_pfn, | |
4746 | zholes_size); | |
4747 | pgdat->node_present_pages = realtotalpages; | |
4748 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, | |
4749 | realtotalpages); | |
4750 | } | |
4751 | ||
4752 | #ifndef CONFIG_SPARSEMEM | |
4753 | /* | |
4754 | * Calculate the size of the zone->blockflags rounded to an unsigned long | |
4755 | * Start by making sure zonesize is a multiple of pageblock_order by rounding | |
4756 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally | |
4757 | * round what is now in bits to nearest long in bits, then return it in | |
4758 | * bytes. | |
4759 | */ | |
4760 | static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) | |
4761 | { | |
4762 | unsigned long usemapsize; | |
4763 | ||
4764 | zonesize += zone_start_pfn & (pageblock_nr_pages-1); | |
4765 | usemapsize = roundup(zonesize, pageblock_nr_pages); | |
4766 | usemapsize = usemapsize >> pageblock_order; | |
4767 | usemapsize *= NR_PAGEBLOCK_BITS; | |
4768 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); | |
4769 | ||
4770 | return usemapsize / 8; | |
4771 | } | |
4772 | ||
4773 | static void __init setup_usemap(struct pglist_data *pgdat, | |
4774 | struct zone *zone, | |
4775 | unsigned long zone_start_pfn, | |
4776 | unsigned long zonesize) | |
4777 | { | |
4778 | unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); | |
4779 | zone->pageblock_flags = NULL; | |
4780 | if (usemapsize) | |
4781 | zone->pageblock_flags = | |
4782 | memblock_virt_alloc_node_nopanic(usemapsize, | |
4783 | pgdat->node_id); | |
4784 | } | |
4785 | #else | |
4786 | static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, | |
4787 | unsigned long zone_start_pfn, unsigned long zonesize) {} | |
4788 | #endif /* CONFIG_SPARSEMEM */ | |
4789 | ||
4790 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
4791 | ||
4792 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ | |
4793 | void __paginginit set_pageblock_order(void) | |
4794 | { | |
4795 | unsigned int order; | |
4796 | ||
4797 | /* Check that pageblock_nr_pages has not already been setup */ | |
4798 | if (pageblock_order) | |
4799 | return; | |
4800 | ||
4801 | if (HPAGE_SHIFT > PAGE_SHIFT) | |
4802 | order = HUGETLB_PAGE_ORDER; | |
4803 | else | |
4804 | order = MAX_ORDER - 1; | |
4805 | ||
4806 | /* | |
4807 | * Assume the largest contiguous order of interest is a huge page. | |
4808 | * This value may be variable depending on boot parameters on IA64 and | |
4809 | * powerpc. | |
4810 | */ | |
4811 | pageblock_order = order; | |
4812 | } | |
4813 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4814 | ||
4815 | /* | |
4816 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() | |
4817 | * is unused as pageblock_order is set at compile-time. See | |
4818 | * include/linux/pageblock-flags.h for the values of pageblock_order based on | |
4819 | * the kernel config | |
4820 | */ | |
4821 | void __paginginit set_pageblock_order(void) | |
4822 | { | |
4823 | } | |
4824 | ||
4825 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4826 | ||
4827 | static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, | |
4828 | unsigned long present_pages) | |
4829 | { | |
4830 | unsigned long pages = spanned_pages; | |
4831 | ||
4832 | /* | |
4833 | * Provide a more accurate estimation if there are holes within | |
4834 | * the zone and SPARSEMEM is in use. If there are holes within the | |
4835 | * zone, each populated memory region may cost us one or two extra | |
4836 | * memmap pages due to alignment because memmap pages for each | |
4837 | * populated regions may not naturally algined on page boundary. | |
4838 | * So the (present_pages >> 4) heuristic is a tradeoff for that. | |
4839 | */ | |
4840 | if (spanned_pages > present_pages + (present_pages >> 4) && | |
4841 | IS_ENABLED(CONFIG_SPARSEMEM)) | |
4842 | pages = present_pages; | |
4843 | ||
4844 | return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; | |
4845 | } | |
4846 | ||
4847 | /* | |
4848 | * Set up the zone data structures: | |
4849 | * - mark all pages reserved | |
4850 | * - mark all memory queues empty | |
4851 | * - clear the memory bitmaps | |
4852 | * | |
4853 | * NOTE: pgdat should get zeroed by caller. | |
4854 | */ | |
4855 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, | |
4856 | unsigned long node_start_pfn, unsigned long node_end_pfn, | |
4857 | unsigned long *zones_size, unsigned long *zholes_size) | |
4858 | { | |
4859 | enum zone_type j; | |
4860 | int nid = pgdat->node_id; | |
4861 | unsigned long zone_start_pfn = pgdat->node_start_pfn; | |
4862 | int ret; | |
4863 | ||
4864 | pgdat_resize_init(pgdat); | |
4865 | #ifdef CONFIG_NUMA_BALANCING | |
4866 | spin_lock_init(&pgdat->numabalancing_migrate_lock); | |
4867 | pgdat->numabalancing_migrate_nr_pages = 0; | |
4868 | pgdat->numabalancing_migrate_next_window = jiffies; | |
4869 | #endif | |
4870 | init_waitqueue_head(&pgdat->kswapd_wait); | |
4871 | init_waitqueue_head(&pgdat->pfmemalloc_wait); | |
4872 | pgdat_page_ext_init(pgdat); | |
4873 | ||
4874 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
4875 | struct zone *zone = pgdat->node_zones + j; | |
4876 | unsigned long size, realsize, freesize, memmap_pages; | |
4877 | ||
4878 | size = zone_spanned_pages_in_node(nid, j, node_start_pfn, | |
4879 | node_end_pfn, zones_size); | |
4880 | realsize = freesize = size - zone_absent_pages_in_node(nid, j, | |
4881 | node_start_pfn, | |
4882 | node_end_pfn, | |
4883 | zholes_size); | |
4884 | ||
4885 | /* | |
4886 | * Adjust freesize so that it accounts for how much memory | |
4887 | * is used by this zone for memmap. This affects the watermark | |
4888 | * and per-cpu initialisations | |
4889 | */ | |
4890 | memmap_pages = calc_memmap_size(size, realsize); | |
4891 | if (!is_highmem_idx(j)) { | |
4892 | if (freesize >= memmap_pages) { | |
4893 | freesize -= memmap_pages; | |
4894 | if (memmap_pages) | |
4895 | printk(KERN_DEBUG | |
4896 | " %s zone: %lu pages used for memmap\n", | |
4897 | zone_names[j], memmap_pages); | |
4898 | } else | |
4899 | printk(KERN_WARNING | |
4900 | " %s zone: %lu pages exceeds freesize %lu\n", | |
4901 | zone_names[j], memmap_pages, freesize); | |
4902 | } | |
4903 | ||
4904 | /* Account for reserved pages */ | |
4905 | if (j == 0 && freesize > dma_reserve) { | |
4906 | freesize -= dma_reserve; | |
4907 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", | |
4908 | zone_names[0], dma_reserve); | |
4909 | } | |
4910 | ||
4911 | if (!is_highmem_idx(j)) | |
4912 | nr_kernel_pages += freesize; | |
4913 | /* Charge for highmem memmap if there are enough kernel pages */ | |
4914 | else if (nr_kernel_pages > memmap_pages * 2) | |
4915 | nr_kernel_pages -= memmap_pages; | |
4916 | nr_all_pages += freesize; | |
4917 | ||
4918 | zone->spanned_pages = size; | |
4919 | zone->present_pages = realsize; | |
4920 | /* | |
4921 | * Set an approximate value for lowmem here, it will be adjusted | |
4922 | * when the bootmem allocator frees pages into the buddy system. | |
4923 | * And all highmem pages will be managed by the buddy system. | |
4924 | */ | |
4925 | zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; | |
4926 | #ifdef CONFIG_NUMA | |
4927 | zone->node = nid; | |
4928 | zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) | |
4929 | / 100; | |
4930 | zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; | |
4931 | #endif | |
4932 | zone->name = zone_names[j]; | |
4933 | spin_lock_init(&zone->lock); | |
4934 | spin_lock_init(&zone->lru_lock); | |
4935 | zone_seqlock_init(zone); | |
4936 | zone->zone_pgdat = pgdat; | |
4937 | zone_pcp_init(zone); | |
4938 | ||
4939 | /* For bootup, initialized properly in watermark setup */ | |
4940 | mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages); | |
4941 | ||
4942 | lruvec_init(&zone->lruvec); | |
4943 | if (!size) | |
4944 | continue; | |
4945 | ||
4946 | set_pageblock_order(); | |
4947 | setup_usemap(pgdat, zone, zone_start_pfn, size); | |
4948 | ret = init_currently_empty_zone(zone, zone_start_pfn, | |
4949 | size, MEMMAP_EARLY); | |
4950 | BUG_ON(ret); | |
4951 | memmap_init(size, nid, j, zone_start_pfn); | |
4952 | zone_start_pfn += size; | |
4953 | } | |
4954 | } | |
4955 | ||
4956 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) | |
4957 | { | |
4958 | /* Skip empty nodes */ | |
4959 | if (!pgdat->node_spanned_pages) | |
4960 | return; | |
4961 | ||
4962 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
4963 | /* ia64 gets its own node_mem_map, before this, without bootmem */ | |
4964 | if (!pgdat->node_mem_map) { | |
4965 | unsigned long size, start, end; | |
4966 | struct page *map; | |
4967 | ||
4968 | /* | |
4969 | * The zone's endpoints aren't required to be MAX_ORDER | |
4970 | * aligned but the node_mem_map endpoints must be in order | |
4971 | * for the buddy allocator to function correctly. | |
4972 | */ | |
4973 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); | |
4974 | end = pgdat_end_pfn(pgdat); | |
4975 | end = ALIGN(end, MAX_ORDER_NR_PAGES); | |
4976 | size = (end - start) * sizeof(struct page); | |
4977 | map = alloc_remap(pgdat->node_id, size); | |
4978 | if (!map) | |
4979 | map = memblock_virt_alloc_node_nopanic(size, | |
4980 | pgdat->node_id); | |
4981 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); | |
4982 | } | |
4983 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
4984 | /* | |
4985 | * With no DISCONTIG, the global mem_map is just set as node 0's | |
4986 | */ | |
4987 | if (pgdat == NODE_DATA(0)) { | |
4988 | mem_map = NODE_DATA(0)->node_mem_map; | |
4989 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4990 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) | |
4991 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); | |
4992 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4993 | } | |
4994 | #endif | |
4995 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ | |
4996 | } | |
4997 | ||
4998 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, | |
4999 | unsigned long node_start_pfn, unsigned long *zholes_size) | |
5000 | { | |
5001 | pg_data_t *pgdat = NODE_DATA(nid); | |
5002 | unsigned long start_pfn = 0; | |
5003 | unsigned long end_pfn = 0; | |
5004 | ||
5005 | /* pg_data_t should be reset to zero when it's allocated */ | |
5006 | WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); | |
5007 | ||
5008 | pgdat->node_id = nid; | |
5009 | pgdat->node_start_pfn = node_start_pfn; | |
5010 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
5011 | get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); | |
5012 | pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid, | |
5013 | (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1); | |
5014 | #endif | |
5015 | calculate_node_totalpages(pgdat, start_pfn, end_pfn, | |
5016 | zones_size, zholes_size); | |
5017 | ||
5018 | alloc_node_mem_map(pgdat); | |
5019 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
5020 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", | |
5021 | nid, (unsigned long)pgdat, | |
5022 | (unsigned long)pgdat->node_mem_map); | |
5023 | #endif | |
5024 | ||
5025 | free_area_init_core(pgdat, start_pfn, end_pfn, | |
5026 | zones_size, zholes_size); | |
5027 | } | |
5028 | ||
5029 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
5030 | ||
5031 | #if MAX_NUMNODES > 1 | |
5032 | /* | |
5033 | * Figure out the number of possible node ids. | |
5034 | */ | |
5035 | void __init setup_nr_node_ids(void) | |
5036 | { | |
5037 | unsigned int node; | |
5038 | unsigned int highest = 0; | |
5039 | ||
5040 | for_each_node_mask(node, node_possible_map) | |
5041 | highest = node; | |
5042 | nr_node_ids = highest + 1; | |
5043 | } | |
5044 | #endif | |
5045 | ||
5046 | /** | |
5047 | * node_map_pfn_alignment - determine the maximum internode alignment | |
5048 | * | |
5049 | * This function should be called after node map is populated and sorted. | |
5050 | * It calculates the maximum power of two alignment which can distinguish | |
5051 | * all the nodes. | |
5052 | * | |
5053 | * For example, if all nodes are 1GiB and aligned to 1GiB, the return value | |
5054 | * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the | |
5055 | * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is | |
5056 | * shifted, 1GiB is enough and this function will indicate so. | |
5057 | * | |
5058 | * This is used to test whether pfn -> nid mapping of the chosen memory | |
5059 | * model has fine enough granularity to avoid incorrect mapping for the | |
5060 | * populated node map. | |
5061 | * | |
5062 | * Returns the determined alignment in pfn's. 0 if there is no alignment | |
5063 | * requirement (single node). | |
5064 | */ | |
5065 | unsigned long __init node_map_pfn_alignment(void) | |
5066 | { | |
5067 | unsigned long accl_mask = 0, last_end = 0; | |
5068 | unsigned long start, end, mask; | |
5069 | int last_nid = -1; | |
5070 | int i, nid; | |
5071 | ||
5072 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { | |
5073 | if (!start || last_nid < 0 || last_nid == nid) { | |
5074 | last_nid = nid; | |
5075 | last_end = end; | |
5076 | continue; | |
5077 | } | |
5078 | ||
5079 | /* | |
5080 | * Start with a mask granular enough to pin-point to the | |
5081 | * start pfn and tick off bits one-by-one until it becomes | |
5082 | * too coarse to separate the current node from the last. | |
5083 | */ | |
5084 | mask = ~((1 << __ffs(start)) - 1); | |
5085 | while (mask && last_end <= (start & (mask << 1))) | |
5086 | mask <<= 1; | |
5087 | ||
5088 | /* accumulate all internode masks */ | |
5089 | accl_mask |= mask; | |
5090 | } | |
5091 | ||
5092 | /* convert mask to number of pages */ | |
5093 | return ~accl_mask + 1; | |
5094 | } | |
5095 | ||
5096 | /* Find the lowest pfn for a node */ | |
5097 | static unsigned long __init find_min_pfn_for_node(int nid) | |
5098 | { | |
5099 | unsigned long min_pfn = ULONG_MAX; | |
5100 | unsigned long start_pfn; | |
5101 | int i; | |
5102 | ||
5103 | for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) | |
5104 | min_pfn = min(min_pfn, start_pfn); | |
5105 | ||
5106 | if (min_pfn == ULONG_MAX) { | |
5107 | printk(KERN_WARNING | |
5108 | "Could not find start_pfn for node %d\n", nid); | |
5109 | return 0; | |
5110 | } | |
5111 | ||
5112 | return min_pfn; | |
5113 | } | |
5114 | ||
5115 | /** | |
5116 | * find_min_pfn_with_active_regions - Find the minimum PFN registered | |
5117 | * | |
5118 | * It returns the minimum PFN based on information provided via | |
5119 | * memblock_set_node(). | |
5120 | */ | |
5121 | unsigned long __init find_min_pfn_with_active_regions(void) | |
5122 | { | |
5123 | return find_min_pfn_for_node(MAX_NUMNODES); | |
5124 | } | |
5125 | ||
5126 | /* | |
5127 | * early_calculate_totalpages() | |
5128 | * Sum pages in active regions for movable zone. | |
5129 | * Populate N_MEMORY for calculating usable_nodes. | |
5130 | */ | |
5131 | static unsigned long __init early_calculate_totalpages(void) | |
5132 | { | |
5133 | unsigned long totalpages = 0; | |
5134 | unsigned long start_pfn, end_pfn; | |
5135 | int i, nid; | |
5136 | ||
5137 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { | |
5138 | unsigned long pages = end_pfn - start_pfn; | |
5139 | ||
5140 | totalpages += pages; | |
5141 | if (pages) | |
5142 | node_set_state(nid, N_MEMORY); | |
5143 | } | |
5144 | return totalpages; | |
5145 | } | |
5146 | ||
5147 | /* | |
5148 | * Find the PFN the Movable zone begins in each node. Kernel memory | |
5149 | * is spread evenly between nodes as long as the nodes have enough | |
5150 | * memory. When they don't, some nodes will have more kernelcore than | |
5151 | * others | |
5152 | */ | |
5153 | static void __init find_zone_movable_pfns_for_nodes(void) | |
5154 | { | |
5155 | int i, nid; | |
5156 | unsigned long usable_startpfn; | |
5157 | unsigned long kernelcore_node, kernelcore_remaining; | |
5158 | /* save the state before borrow the nodemask */ | |
5159 | nodemask_t saved_node_state = node_states[N_MEMORY]; | |
5160 | unsigned long totalpages = early_calculate_totalpages(); | |
5161 | int usable_nodes = nodes_weight(node_states[N_MEMORY]); | |
5162 | struct memblock_region *r; | |
5163 | ||
5164 | /* Need to find movable_zone earlier when movable_node is specified. */ | |
5165 | find_usable_zone_for_movable(); | |
5166 | ||
5167 | /* | |
5168 | * If movable_node is specified, ignore kernelcore and movablecore | |
5169 | * options. | |
5170 | */ | |
5171 | if (movable_node_is_enabled()) { | |
5172 | for_each_memblock(memory, r) { | |
5173 | if (!memblock_is_hotpluggable(r)) | |
5174 | continue; | |
5175 | ||
5176 | nid = r->nid; | |
5177 | ||
5178 | usable_startpfn = PFN_DOWN(r->base); | |
5179 | zone_movable_pfn[nid] = zone_movable_pfn[nid] ? | |
5180 | min(usable_startpfn, zone_movable_pfn[nid]) : | |
5181 | usable_startpfn; | |
5182 | } | |
5183 | ||
5184 | goto out2; | |
5185 | } | |
5186 | ||
5187 | /* | |
5188 | * If movablecore=nn[KMG] was specified, calculate what size of | |
5189 | * kernelcore that corresponds so that memory usable for | |
5190 | * any allocation type is evenly spread. If both kernelcore | |
5191 | * and movablecore are specified, then the value of kernelcore | |
5192 | * will be used for required_kernelcore if it's greater than | |
5193 | * what movablecore would have allowed. | |
5194 | */ | |
5195 | if (required_movablecore) { | |
5196 | unsigned long corepages; | |
5197 | ||
5198 | /* | |
5199 | * Round-up so that ZONE_MOVABLE is at least as large as what | |
5200 | * was requested by the user | |
5201 | */ | |
5202 | required_movablecore = | |
5203 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); | |
5204 | corepages = totalpages - required_movablecore; | |
5205 | ||
5206 | required_kernelcore = max(required_kernelcore, corepages); | |
5207 | } | |
5208 | ||
5209 | /* If kernelcore was not specified, there is no ZONE_MOVABLE */ | |
5210 | if (!required_kernelcore) | |
5211 | goto out; | |
5212 | ||
5213 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ | |
5214 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; | |
5215 | ||
5216 | restart: | |
5217 | /* Spread kernelcore memory as evenly as possible throughout nodes */ | |
5218 | kernelcore_node = required_kernelcore / usable_nodes; | |
5219 | for_each_node_state(nid, N_MEMORY) { | |
5220 | unsigned long start_pfn, end_pfn; | |
5221 | ||
5222 | /* | |
5223 | * Recalculate kernelcore_node if the division per node | |
5224 | * now exceeds what is necessary to satisfy the requested | |
5225 | * amount of memory for the kernel | |
5226 | */ | |
5227 | if (required_kernelcore < kernelcore_node) | |
5228 | kernelcore_node = required_kernelcore / usable_nodes; | |
5229 | ||
5230 | /* | |
5231 | * As the map is walked, we track how much memory is usable | |
5232 | * by the kernel using kernelcore_remaining. When it is | |
5233 | * 0, the rest of the node is usable by ZONE_MOVABLE | |
5234 | */ | |
5235 | kernelcore_remaining = kernelcore_node; | |
5236 | ||
5237 | /* Go through each range of PFNs within this node */ | |
5238 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
5239 | unsigned long size_pages; | |
5240 | ||
5241 | start_pfn = max(start_pfn, zone_movable_pfn[nid]); | |
5242 | if (start_pfn >= end_pfn) | |
5243 | continue; | |
5244 | ||
5245 | /* Account for what is only usable for kernelcore */ | |
5246 | if (start_pfn < usable_startpfn) { | |
5247 | unsigned long kernel_pages; | |
5248 | kernel_pages = min(end_pfn, usable_startpfn) | |
5249 | - start_pfn; | |
5250 | ||
5251 | kernelcore_remaining -= min(kernel_pages, | |
5252 | kernelcore_remaining); | |
5253 | required_kernelcore -= min(kernel_pages, | |
5254 | required_kernelcore); | |
5255 | ||
5256 | /* Continue if range is now fully accounted */ | |
5257 | if (end_pfn <= usable_startpfn) { | |
5258 | ||
5259 | /* | |
5260 | * Push zone_movable_pfn to the end so | |
5261 | * that if we have to rebalance | |
5262 | * kernelcore across nodes, we will | |
5263 | * not double account here | |
5264 | */ | |
5265 | zone_movable_pfn[nid] = end_pfn; | |
5266 | continue; | |
5267 | } | |
5268 | start_pfn = usable_startpfn; | |
5269 | } | |
5270 | ||
5271 | /* | |
5272 | * The usable PFN range for ZONE_MOVABLE is from | |
5273 | * start_pfn->end_pfn. Calculate size_pages as the | |
5274 | * number of pages used as kernelcore | |
5275 | */ | |
5276 | size_pages = end_pfn - start_pfn; | |
5277 | if (size_pages > kernelcore_remaining) | |
5278 | size_pages = kernelcore_remaining; | |
5279 | zone_movable_pfn[nid] = start_pfn + size_pages; | |
5280 | ||
5281 | /* | |
5282 | * Some kernelcore has been met, update counts and | |
5283 | * break if the kernelcore for this node has been | |
5284 | * satisfied | |
5285 | */ | |
5286 | required_kernelcore -= min(required_kernelcore, | |
5287 | size_pages); | |
5288 | kernelcore_remaining -= size_pages; | |
5289 | if (!kernelcore_remaining) | |
5290 | break; | |
5291 | } | |
5292 | } | |
5293 | ||
5294 | /* | |
5295 | * If there is still required_kernelcore, we do another pass with one | |
5296 | * less node in the count. This will push zone_movable_pfn[nid] further | |
5297 | * along on the nodes that still have memory until kernelcore is | |
5298 | * satisfied | |
5299 | */ | |
5300 | usable_nodes--; | |
5301 | if (usable_nodes && required_kernelcore > usable_nodes) | |
5302 | goto restart; | |
5303 | ||
5304 | out2: | |
5305 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ | |
5306 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
5307 | zone_movable_pfn[nid] = | |
5308 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); | |
5309 | ||
5310 | out: | |
5311 | /* restore the node_state */ | |
5312 | node_states[N_MEMORY] = saved_node_state; | |
5313 | } | |
5314 | ||
5315 | /* Any regular or high memory on that node ? */ | |
5316 | static void check_for_memory(pg_data_t *pgdat, int nid) | |
5317 | { | |
5318 | enum zone_type zone_type; | |
5319 | ||
5320 | if (N_MEMORY == N_NORMAL_MEMORY) | |
5321 | return; | |
5322 | ||
5323 | for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { | |
5324 | struct zone *zone = &pgdat->node_zones[zone_type]; | |
5325 | if (populated_zone(zone)) { | |
5326 | node_set_state(nid, N_HIGH_MEMORY); | |
5327 | if (N_NORMAL_MEMORY != N_HIGH_MEMORY && | |
5328 | zone_type <= ZONE_NORMAL) | |
5329 | node_set_state(nid, N_NORMAL_MEMORY); | |
5330 | break; | |
5331 | } | |
5332 | } | |
5333 | } | |
5334 | ||
5335 | /** | |
5336 | * free_area_init_nodes - Initialise all pg_data_t and zone data | |
5337 | * @max_zone_pfn: an array of max PFNs for each zone | |
5338 | * | |
5339 | * This will call free_area_init_node() for each active node in the system. | |
5340 | * Using the page ranges provided by memblock_set_node(), the size of each | |
5341 | * zone in each node and their holes is calculated. If the maximum PFN | |
5342 | * between two adjacent zones match, it is assumed that the zone is empty. | |
5343 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed | |
5344 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone | |
5345 | * starts where the previous one ended. For example, ZONE_DMA32 starts | |
5346 | * at arch_max_dma_pfn. | |
5347 | */ | |
5348 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) | |
5349 | { | |
5350 | unsigned long start_pfn, end_pfn; | |
5351 | int i, nid; | |
5352 | ||
5353 | /* Record where the zone boundaries are */ | |
5354 | memset(arch_zone_lowest_possible_pfn, 0, | |
5355 | sizeof(arch_zone_lowest_possible_pfn)); | |
5356 | memset(arch_zone_highest_possible_pfn, 0, | |
5357 | sizeof(arch_zone_highest_possible_pfn)); | |
5358 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); | |
5359 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; | |
5360 | for (i = 1; i < MAX_NR_ZONES; i++) { | |
5361 | if (i == ZONE_MOVABLE) | |
5362 | continue; | |
5363 | arch_zone_lowest_possible_pfn[i] = | |
5364 | arch_zone_highest_possible_pfn[i-1]; | |
5365 | arch_zone_highest_possible_pfn[i] = | |
5366 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); | |
5367 | } | |
5368 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; | |
5369 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; | |
5370 | ||
5371 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ | |
5372 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); | |
5373 | find_zone_movable_pfns_for_nodes(); | |
5374 | ||
5375 | /* Print out the zone ranges */ | |
5376 | pr_info("Zone ranges:\n"); | |
5377 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
5378 | if (i == ZONE_MOVABLE) | |
5379 | continue; | |
5380 | pr_info(" %-8s ", zone_names[i]); | |
5381 | if (arch_zone_lowest_possible_pfn[i] == | |
5382 | arch_zone_highest_possible_pfn[i]) | |
5383 | pr_cont("empty\n"); | |
5384 | else | |
5385 | pr_cont("[mem %#018Lx-%#018Lx]\n", | |
5386 | (u64)arch_zone_lowest_possible_pfn[i] | |
5387 | << PAGE_SHIFT, | |
5388 | ((u64)arch_zone_highest_possible_pfn[i] | |
5389 | << PAGE_SHIFT) - 1); | |
5390 | } | |
5391 | ||
5392 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ | |
5393 | pr_info("Movable zone start for each node\n"); | |
5394 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5395 | if (zone_movable_pfn[i]) | |
5396 | pr_info(" Node %d: %#018Lx\n", i, | |
5397 | (u64)zone_movable_pfn[i] << PAGE_SHIFT); | |
5398 | } | |
5399 | ||
5400 | /* Print out the early node map */ | |
5401 | pr_info("Early memory node ranges\n"); | |
5402 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) | |
5403 | pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid, | |
5404 | (u64)start_pfn << PAGE_SHIFT, | |
5405 | ((u64)end_pfn << PAGE_SHIFT) - 1); | |
5406 | ||
5407 | /* Initialise every node */ | |
5408 | mminit_verify_pageflags_layout(); | |
5409 | setup_nr_node_ids(); | |
5410 | for_each_online_node(nid) { | |
5411 | pg_data_t *pgdat = NODE_DATA(nid); | |
5412 | free_area_init_node(nid, NULL, | |
5413 | find_min_pfn_for_node(nid), NULL); | |
5414 | ||
5415 | /* Any memory on that node */ | |
5416 | if (pgdat->node_present_pages) | |
5417 | node_set_state(nid, N_MEMORY); | |
5418 | check_for_memory(pgdat, nid); | |
5419 | } | |
5420 | } | |
5421 | ||
5422 | static int __init cmdline_parse_core(char *p, unsigned long *core) | |
5423 | { | |
5424 | unsigned long long coremem; | |
5425 | if (!p) | |
5426 | return -EINVAL; | |
5427 | ||
5428 | coremem = memparse(p, &p); | |
5429 | *core = coremem >> PAGE_SHIFT; | |
5430 | ||
5431 | /* Paranoid check that UL is enough for the coremem value */ | |
5432 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); | |
5433 | ||
5434 | return 0; | |
5435 | } | |
5436 | ||
5437 | /* | |
5438 | * kernelcore=size sets the amount of memory for use for allocations that | |
5439 | * cannot be reclaimed or migrated. | |
5440 | */ | |
5441 | static int __init cmdline_parse_kernelcore(char *p) | |
5442 | { | |
5443 | return cmdline_parse_core(p, &required_kernelcore); | |
5444 | } | |
5445 | ||
5446 | /* | |
5447 | * movablecore=size sets the amount of memory for use for allocations that | |
5448 | * can be reclaimed or migrated. | |
5449 | */ | |
5450 | static int __init cmdline_parse_movablecore(char *p) | |
5451 | { | |
5452 | return cmdline_parse_core(p, &required_movablecore); | |
5453 | } | |
5454 | ||
5455 | early_param("kernelcore", cmdline_parse_kernelcore); | |
5456 | early_param("movablecore", cmdline_parse_movablecore); | |
5457 | ||
5458 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
5459 | ||
5460 | void adjust_managed_page_count(struct page *page, long count) | |
5461 | { | |
5462 | spin_lock(&managed_page_count_lock); | |
5463 | page_zone(page)->managed_pages += count; | |
5464 | totalram_pages += count; | |
5465 | #ifdef CONFIG_HIGHMEM | |
5466 | if (PageHighMem(page)) | |
5467 | totalhigh_pages += count; | |
5468 | #endif | |
5469 | spin_unlock(&managed_page_count_lock); | |
5470 | } | |
5471 | EXPORT_SYMBOL(adjust_managed_page_count); | |
5472 | ||
5473 | unsigned long free_reserved_area(void *start, void *end, int poison, char *s) | |
5474 | { | |
5475 | void *pos; | |
5476 | unsigned long pages = 0; | |
5477 | ||
5478 | start = (void *)PAGE_ALIGN((unsigned long)start); | |
5479 | end = (void *)((unsigned long)end & PAGE_MASK); | |
5480 | for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { | |
5481 | if ((unsigned int)poison <= 0xFF) | |
5482 | memset(pos, poison, PAGE_SIZE); | |
5483 | free_reserved_page(virt_to_page(pos)); | |
5484 | } | |
5485 | ||
5486 | if (pages && s) | |
5487 | pr_info("Freeing %s memory: %ldK (%p - %p)\n", | |
5488 | s, pages << (PAGE_SHIFT - 10), start, end); | |
5489 | ||
5490 | return pages; | |
5491 | } | |
5492 | EXPORT_SYMBOL(free_reserved_area); | |
5493 | ||
5494 | #ifdef CONFIG_HIGHMEM | |
5495 | void free_highmem_page(struct page *page) | |
5496 | { | |
5497 | __free_reserved_page(page); | |
5498 | totalram_pages++; | |
5499 | page_zone(page)->managed_pages++; | |
5500 | totalhigh_pages++; | |
5501 | } | |
5502 | #endif | |
5503 | ||
5504 | ||
5505 | void __init mem_init_print_info(const char *str) | |
5506 | { | |
5507 | unsigned long physpages, codesize, datasize, rosize, bss_size; | |
5508 | unsigned long init_code_size, init_data_size; | |
5509 | ||
5510 | physpages = get_num_physpages(); | |
5511 | codesize = _etext - _stext; | |
5512 | datasize = _edata - _sdata; | |
5513 | rosize = __end_rodata - __start_rodata; | |
5514 | bss_size = __bss_stop - __bss_start; | |
5515 | init_data_size = __init_end - __init_begin; | |
5516 | init_code_size = _einittext - _sinittext; | |
5517 | ||
5518 | /* | |
5519 | * Detect special cases and adjust section sizes accordingly: | |
5520 | * 1) .init.* may be embedded into .data sections | |
5521 | * 2) .init.text.* may be out of [__init_begin, __init_end], | |
5522 | * please refer to arch/tile/kernel/vmlinux.lds.S. | |
5523 | * 3) .rodata.* may be embedded into .text or .data sections. | |
5524 | */ | |
5525 | #define adj_init_size(start, end, size, pos, adj) \ | |
5526 | do { \ | |
5527 | if (start <= pos && pos < end && size > adj) \ | |
5528 | size -= adj; \ | |
5529 | } while (0) | |
5530 | ||
5531 | adj_init_size(__init_begin, __init_end, init_data_size, | |
5532 | _sinittext, init_code_size); | |
5533 | adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); | |
5534 | adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); | |
5535 | adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); | |
5536 | adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); | |
5537 | ||
5538 | #undef adj_init_size | |
5539 | ||
5540 | pr_info("Memory: %luK/%luK available " | |
5541 | "(%luK kernel code, %luK rwdata, %luK rodata, " | |
5542 | "%luK init, %luK bss, %luK reserved, %luK cma-reserved" | |
5543 | #ifdef CONFIG_HIGHMEM | |
5544 | ", %luK highmem" | |
5545 | #endif | |
5546 | "%s%s)\n", | |
5547 | nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10), | |
5548 | codesize >> 10, datasize >> 10, rosize >> 10, | |
5549 | (init_data_size + init_code_size) >> 10, bss_size >> 10, | |
5550 | (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10), | |
5551 | totalcma_pages << (PAGE_SHIFT-10), | |
5552 | #ifdef CONFIG_HIGHMEM | |
5553 | totalhigh_pages << (PAGE_SHIFT-10), | |
5554 | #endif | |
5555 | str ? ", " : "", str ? str : ""); | |
5556 | } | |
5557 | ||
5558 | /** | |
5559 | * set_dma_reserve - set the specified number of pages reserved in the first zone | |
5560 | * @new_dma_reserve: The number of pages to mark reserved | |
5561 | * | |
5562 | * The per-cpu batchsize and zone watermarks are determined by present_pages. | |
5563 | * In the DMA zone, a significant percentage may be consumed by kernel image | |
5564 | * and other unfreeable allocations which can skew the watermarks badly. This | |
5565 | * function may optionally be used to account for unfreeable pages in the | |
5566 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and | |
5567 | * smaller per-cpu batchsize. | |
5568 | */ | |
5569 | void __init set_dma_reserve(unsigned long new_dma_reserve) | |
5570 | { | |
5571 | dma_reserve = new_dma_reserve; | |
5572 | } | |
5573 | ||
5574 | void __init free_area_init(unsigned long *zones_size) | |
5575 | { | |
5576 | free_area_init_node(0, zones_size, | |
5577 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); | |
5578 | } | |
5579 | ||
5580 | static int page_alloc_cpu_notify(struct notifier_block *self, | |
5581 | unsigned long action, void *hcpu) | |
5582 | { | |
5583 | int cpu = (unsigned long)hcpu; | |
5584 | ||
5585 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { | |
5586 | lru_add_drain_cpu(cpu); | |
5587 | drain_pages(cpu); | |
5588 | ||
5589 | /* | |
5590 | * Spill the event counters of the dead processor | |
5591 | * into the current processors event counters. | |
5592 | * This artificially elevates the count of the current | |
5593 | * processor. | |
5594 | */ | |
5595 | vm_events_fold_cpu(cpu); | |
5596 | ||
5597 | /* | |
5598 | * Zero the differential counters of the dead processor | |
5599 | * so that the vm statistics are consistent. | |
5600 | * | |
5601 | * This is only okay since the processor is dead and cannot | |
5602 | * race with what we are doing. | |
5603 | */ | |
5604 | cpu_vm_stats_fold(cpu); | |
5605 | } | |
5606 | return NOTIFY_OK; | |
5607 | } | |
5608 | ||
5609 | void __init page_alloc_init(void) | |
5610 | { | |
5611 | hotcpu_notifier(page_alloc_cpu_notify, 0); | |
5612 | } | |
5613 | ||
5614 | /* | |
5615 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio | |
5616 | * or min_free_kbytes changes. | |
5617 | */ | |
5618 | static void calculate_totalreserve_pages(void) | |
5619 | { | |
5620 | struct pglist_data *pgdat; | |
5621 | unsigned long reserve_pages = 0; | |
5622 | enum zone_type i, j; | |
5623 | ||
5624 | for_each_online_pgdat(pgdat) { | |
5625 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
5626 | struct zone *zone = pgdat->node_zones + i; | |
5627 | long max = 0; | |
5628 | ||
5629 | /* Find valid and maximum lowmem_reserve in the zone */ | |
5630 | for (j = i; j < MAX_NR_ZONES; j++) { | |
5631 | if (zone->lowmem_reserve[j] > max) | |
5632 | max = zone->lowmem_reserve[j]; | |
5633 | } | |
5634 | ||
5635 | /* we treat the high watermark as reserved pages. */ | |
5636 | max += high_wmark_pages(zone); | |
5637 | ||
5638 | if (max > zone->managed_pages) | |
5639 | max = zone->managed_pages; | |
5640 | reserve_pages += max; | |
5641 | /* | |
5642 | * Lowmem reserves are not available to | |
5643 | * GFP_HIGHUSER page cache allocations and | |
5644 | * kswapd tries to balance zones to their high | |
5645 | * watermark. As a result, neither should be | |
5646 | * regarded as dirtyable memory, to prevent a | |
5647 | * situation where reclaim has to clean pages | |
5648 | * in order to balance the zones. | |
5649 | */ | |
5650 | zone->dirty_balance_reserve = max; | |
5651 | } | |
5652 | } | |
5653 | dirty_balance_reserve = reserve_pages; | |
5654 | totalreserve_pages = reserve_pages; | |
5655 | } | |
5656 | ||
5657 | /* | |
5658 | * setup_per_zone_lowmem_reserve - called whenever | |
5659 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone | |
5660 | * has a correct pages reserved value, so an adequate number of | |
5661 | * pages are left in the zone after a successful __alloc_pages(). | |
5662 | */ | |
5663 | static void setup_per_zone_lowmem_reserve(void) | |
5664 | { | |
5665 | struct pglist_data *pgdat; | |
5666 | enum zone_type j, idx; | |
5667 | ||
5668 | for_each_online_pgdat(pgdat) { | |
5669 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
5670 | struct zone *zone = pgdat->node_zones + j; | |
5671 | unsigned long managed_pages = zone->managed_pages; | |
5672 | ||
5673 | zone->lowmem_reserve[j] = 0; | |
5674 | ||
5675 | idx = j; | |
5676 | while (idx) { | |
5677 | struct zone *lower_zone; | |
5678 | ||
5679 | idx--; | |
5680 | ||
5681 | if (sysctl_lowmem_reserve_ratio[idx] < 1) | |
5682 | sysctl_lowmem_reserve_ratio[idx] = 1; | |
5683 | ||
5684 | lower_zone = pgdat->node_zones + idx; | |
5685 | lower_zone->lowmem_reserve[j] = managed_pages / | |
5686 | sysctl_lowmem_reserve_ratio[idx]; | |
5687 | managed_pages += lower_zone->managed_pages; | |
5688 | } | |
5689 | } | |
5690 | } | |
5691 | ||
5692 | /* update totalreserve_pages */ | |
5693 | calculate_totalreserve_pages(); | |
5694 | } | |
5695 | ||
5696 | static void __setup_per_zone_wmarks(void) | |
5697 | { | |
5698 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
5699 | unsigned long lowmem_pages = 0; | |
5700 | struct zone *zone; | |
5701 | unsigned long flags; | |
5702 | ||
5703 | /* Calculate total number of !ZONE_HIGHMEM pages */ | |
5704 | for_each_zone(zone) { | |
5705 | if (!is_highmem(zone)) | |
5706 | lowmem_pages += zone->managed_pages; | |
5707 | } | |
5708 | ||
5709 | for_each_zone(zone) { | |
5710 | u64 tmp; | |
5711 | ||
5712 | spin_lock_irqsave(&zone->lock, flags); | |
5713 | tmp = (u64)pages_min * zone->managed_pages; | |
5714 | do_div(tmp, lowmem_pages); | |
5715 | if (is_highmem(zone)) { | |
5716 | /* | |
5717 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't | |
5718 | * need highmem pages, so cap pages_min to a small | |
5719 | * value here. | |
5720 | * | |
5721 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) | |
5722 | * deltas controls asynch page reclaim, and so should | |
5723 | * not be capped for highmem. | |
5724 | */ | |
5725 | unsigned long min_pages; | |
5726 | ||
5727 | min_pages = zone->managed_pages / 1024; | |
5728 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); | |
5729 | zone->watermark[WMARK_MIN] = min_pages; | |
5730 | } else { | |
5731 | /* | |
5732 | * If it's a lowmem zone, reserve a number of pages | |
5733 | * proportionate to the zone's size. | |
5734 | */ | |
5735 | zone->watermark[WMARK_MIN] = tmp; | |
5736 | } | |
5737 | ||
5738 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); | |
5739 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); | |
5740 | ||
5741 | __mod_zone_page_state(zone, NR_ALLOC_BATCH, | |
5742 | high_wmark_pages(zone) - low_wmark_pages(zone) - | |
5743 | atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); | |
5744 | ||
5745 | setup_zone_migrate_reserve(zone); | |
5746 | spin_unlock_irqrestore(&zone->lock, flags); | |
5747 | } | |
5748 | ||
5749 | /* update totalreserve_pages */ | |
5750 | calculate_totalreserve_pages(); | |
5751 | } | |
5752 | ||
5753 | /** | |
5754 | * setup_per_zone_wmarks - called when min_free_kbytes changes | |
5755 | * or when memory is hot-{added|removed} | |
5756 | * | |
5757 | * Ensures that the watermark[min,low,high] values for each zone are set | |
5758 | * correctly with respect to min_free_kbytes. | |
5759 | */ | |
5760 | void setup_per_zone_wmarks(void) | |
5761 | { | |
5762 | mutex_lock(&zonelists_mutex); | |
5763 | __setup_per_zone_wmarks(); | |
5764 | mutex_unlock(&zonelists_mutex); | |
5765 | } | |
5766 | ||
5767 | /* | |
5768 | * The inactive anon list should be small enough that the VM never has to | |
5769 | * do too much work, but large enough that each inactive page has a chance | |
5770 | * to be referenced again before it is swapped out. | |
5771 | * | |
5772 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to | |
5773 | * INACTIVE_ANON pages on this zone's LRU, maintained by the | |
5774 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of | |
5775 | * the anonymous pages are kept on the inactive list. | |
5776 | * | |
5777 | * total target max | |
5778 | * memory ratio inactive anon | |
5779 | * ------------------------------------- | |
5780 | * 10MB 1 5MB | |
5781 | * 100MB 1 50MB | |
5782 | * 1GB 3 250MB | |
5783 | * 10GB 10 0.9GB | |
5784 | * 100GB 31 3GB | |
5785 | * 1TB 101 10GB | |
5786 | * 10TB 320 32GB | |
5787 | */ | |
5788 | static void __meminit calculate_zone_inactive_ratio(struct zone *zone) | |
5789 | { | |
5790 | unsigned int gb, ratio; | |
5791 | ||
5792 | /* Zone size in gigabytes */ | |
5793 | gb = zone->managed_pages >> (30 - PAGE_SHIFT); | |
5794 | if (gb) | |
5795 | ratio = int_sqrt(10 * gb); | |
5796 | else | |
5797 | ratio = 1; | |
5798 | ||
5799 | zone->inactive_ratio = ratio; | |
5800 | } | |
5801 | ||
5802 | static void __meminit setup_per_zone_inactive_ratio(void) | |
5803 | { | |
5804 | struct zone *zone; | |
5805 | ||
5806 | for_each_zone(zone) | |
5807 | calculate_zone_inactive_ratio(zone); | |
5808 | } | |
5809 | ||
5810 | /* | |
5811 | * Initialise min_free_kbytes. | |
5812 | * | |
5813 | * For small machines we want it small (128k min). For large machines | |
5814 | * we want it large (64MB max). But it is not linear, because network | |
5815 | * bandwidth does not increase linearly with machine size. We use | |
5816 | * | |
5817 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: | |
5818 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) | |
5819 | * | |
5820 | * which yields | |
5821 | * | |
5822 | * 16MB: 512k | |
5823 | * 32MB: 724k | |
5824 | * 64MB: 1024k | |
5825 | * 128MB: 1448k | |
5826 | * 256MB: 2048k | |
5827 | * 512MB: 2896k | |
5828 | * 1024MB: 4096k | |
5829 | * 2048MB: 5792k | |
5830 | * 4096MB: 8192k | |
5831 | * 8192MB: 11584k | |
5832 | * 16384MB: 16384k | |
5833 | */ | |
5834 | int __meminit init_per_zone_wmark_min(void) | |
5835 | { | |
5836 | unsigned long lowmem_kbytes; | |
5837 | int new_min_free_kbytes; | |
5838 | ||
5839 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
5840 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); | |
5841 | ||
5842 | if (new_min_free_kbytes > user_min_free_kbytes) { | |
5843 | min_free_kbytes = new_min_free_kbytes; | |
5844 | if (min_free_kbytes < 128) | |
5845 | min_free_kbytes = 128; | |
5846 | if (min_free_kbytes > 65536) | |
5847 | min_free_kbytes = 65536; | |
5848 | } else { | |
5849 | pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", | |
5850 | new_min_free_kbytes, user_min_free_kbytes); | |
5851 | } | |
5852 | setup_per_zone_wmarks(); | |
5853 | refresh_zone_stat_thresholds(); | |
5854 | setup_per_zone_lowmem_reserve(); | |
5855 | setup_per_zone_inactive_ratio(); | |
5856 | return 0; | |
5857 | } | |
5858 | module_init(init_per_zone_wmark_min) | |
5859 | ||
5860 | /* | |
5861 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so | |
5862 | * that we can call two helper functions whenever min_free_kbytes | |
5863 | * changes. | |
5864 | */ | |
5865 | int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, | |
5866 | void __user *buffer, size_t *length, loff_t *ppos) | |
5867 | { | |
5868 | int rc; | |
5869 | ||
5870 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5871 | if (rc) | |
5872 | return rc; | |
5873 | ||
5874 | if (write) { | |
5875 | user_min_free_kbytes = min_free_kbytes; | |
5876 | setup_per_zone_wmarks(); | |
5877 | } | |
5878 | return 0; | |
5879 | } | |
5880 | ||
5881 | #ifdef CONFIG_NUMA | |
5882 | int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, | |
5883 | void __user *buffer, size_t *length, loff_t *ppos) | |
5884 | { | |
5885 | struct zone *zone; | |
5886 | int rc; | |
5887 | ||
5888 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5889 | if (rc) | |
5890 | return rc; | |
5891 | ||
5892 | for_each_zone(zone) | |
5893 | zone->min_unmapped_pages = (zone->managed_pages * | |
5894 | sysctl_min_unmapped_ratio) / 100; | |
5895 | return 0; | |
5896 | } | |
5897 | ||
5898 | int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, | |
5899 | void __user *buffer, size_t *length, loff_t *ppos) | |
5900 | { | |
5901 | struct zone *zone; | |
5902 | int rc; | |
5903 | ||
5904 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5905 | if (rc) | |
5906 | return rc; | |
5907 | ||
5908 | for_each_zone(zone) | |
5909 | zone->min_slab_pages = (zone->managed_pages * | |
5910 | sysctl_min_slab_ratio) / 100; | |
5911 | return 0; | |
5912 | } | |
5913 | #endif | |
5914 | ||
5915 | /* | |
5916 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
5917 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
5918 | * whenever sysctl_lowmem_reserve_ratio changes. | |
5919 | * | |
5920 | * The reserve ratio obviously has absolutely no relation with the | |
5921 | * minimum watermarks. The lowmem reserve ratio can only make sense | |
5922 | * if in function of the boot time zone sizes. | |
5923 | */ | |
5924 | int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write, | |
5925 | void __user *buffer, size_t *length, loff_t *ppos) | |
5926 | { | |
5927 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5928 | setup_per_zone_lowmem_reserve(); | |
5929 | return 0; | |
5930 | } | |
5931 | ||
5932 | /* | |
5933 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each | |
5934 | * cpu. It is the fraction of total pages in each zone that a hot per cpu | |
5935 | * pagelist can have before it gets flushed back to buddy allocator. | |
5936 | */ | |
5937 | int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, | |
5938 | void __user *buffer, size_t *length, loff_t *ppos) | |
5939 | { | |
5940 | struct zone *zone; | |
5941 | int old_percpu_pagelist_fraction; | |
5942 | int ret; | |
5943 | ||
5944 | mutex_lock(&pcp_batch_high_lock); | |
5945 | old_percpu_pagelist_fraction = percpu_pagelist_fraction; | |
5946 | ||
5947 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5948 | if (!write || ret < 0) | |
5949 | goto out; | |
5950 | ||
5951 | /* Sanity checking to avoid pcp imbalance */ | |
5952 | if (percpu_pagelist_fraction && | |
5953 | percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) { | |
5954 | percpu_pagelist_fraction = old_percpu_pagelist_fraction; | |
5955 | ret = -EINVAL; | |
5956 | goto out; | |
5957 | } | |
5958 | ||
5959 | /* No change? */ | |
5960 | if (percpu_pagelist_fraction == old_percpu_pagelist_fraction) | |
5961 | goto out; | |
5962 | ||
5963 | for_each_populated_zone(zone) { | |
5964 | unsigned int cpu; | |
5965 | ||
5966 | for_each_possible_cpu(cpu) | |
5967 | pageset_set_high_and_batch(zone, | |
5968 | per_cpu_ptr(zone->pageset, cpu)); | |
5969 | } | |
5970 | out: | |
5971 | mutex_unlock(&pcp_batch_high_lock); | |
5972 | return ret; | |
5973 | } | |
5974 | ||
5975 | int hashdist = HASHDIST_DEFAULT; | |
5976 | ||
5977 | #ifdef CONFIG_NUMA | |
5978 | static int __init set_hashdist(char *str) | |
5979 | { | |
5980 | if (!str) | |
5981 | return 0; | |
5982 | hashdist = simple_strtoul(str, &str, 0); | |
5983 | return 1; | |
5984 | } | |
5985 | __setup("hashdist=", set_hashdist); | |
5986 | #endif | |
5987 | ||
5988 | /* | |
5989 | * allocate a large system hash table from bootmem | |
5990 | * - it is assumed that the hash table must contain an exact power-of-2 | |
5991 | * quantity of entries | |
5992 | * - limit is the number of hash buckets, not the total allocation size | |
5993 | */ | |
5994 | void *__init alloc_large_system_hash(const char *tablename, | |
5995 | unsigned long bucketsize, | |
5996 | unsigned long numentries, | |
5997 | int scale, | |
5998 | int flags, | |
5999 | unsigned int *_hash_shift, | |
6000 | unsigned int *_hash_mask, | |
6001 | unsigned long low_limit, | |
6002 | unsigned long high_limit) | |
6003 | { | |
6004 | unsigned long long max = high_limit; | |
6005 | unsigned long log2qty, size; | |
6006 | void *table = NULL; | |
6007 | ||
6008 | /* allow the kernel cmdline to have a say */ | |
6009 | if (!numentries) { | |
6010 | /* round applicable memory size up to nearest megabyte */ | |
6011 | numentries = nr_kernel_pages; | |
6012 | ||
6013 | /* It isn't necessary when PAGE_SIZE >= 1MB */ | |
6014 | if (PAGE_SHIFT < 20) | |
6015 | numentries = round_up(numentries, (1<<20)/PAGE_SIZE); | |
6016 | ||
6017 | /* limit to 1 bucket per 2^scale bytes of low memory */ | |
6018 | if (scale > PAGE_SHIFT) | |
6019 | numentries >>= (scale - PAGE_SHIFT); | |
6020 | else | |
6021 | numentries <<= (PAGE_SHIFT - scale); | |
6022 | ||
6023 | /* Make sure we've got at least a 0-order allocation.. */ | |
6024 | if (unlikely(flags & HASH_SMALL)) { | |
6025 | /* Makes no sense without HASH_EARLY */ | |
6026 | WARN_ON(!(flags & HASH_EARLY)); | |
6027 | if (!(numentries >> *_hash_shift)) { | |
6028 | numentries = 1UL << *_hash_shift; | |
6029 | BUG_ON(!numentries); | |
6030 | } | |
6031 | } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) | |
6032 | numentries = PAGE_SIZE / bucketsize; | |
6033 | } | |
6034 | numentries = roundup_pow_of_two(numentries); | |
6035 | ||
6036 | /* limit allocation size to 1/16 total memory by default */ | |
6037 | if (max == 0) { | |
6038 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; | |
6039 | do_div(max, bucketsize); | |
6040 | } | |
6041 | max = min(max, 0x80000000ULL); | |
6042 | ||
6043 | if (numentries < low_limit) | |
6044 | numentries = low_limit; | |
6045 | if (numentries > max) | |
6046 | numentries = max; | |
6047 | ||
6048 | log2qty = ilog2(numentries); | |
6049 | ||
6050 | do { | |
6051 | size = bucketsize << log2qty; | |
6052 | if (flags & HASH_EARLY) | |
6053 | table = memblock_virt_alloc_nopanic(size, 0); | |
6054 | else if (hashdist) | |
6055 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); | |
6056 | else { | |
6057 | /* | |
6058 | * If bucketsize is not a power-of-two, we may free | |
6059 | * some pages at the end of hash table which | |
6060 | * alloc_pages_exact() automatically does | |
6061 | */ | |
6062 | if (get_order(size) < MAX_ORDER) { | |
6063 | table = alloc_pages_exact(size, GFP_ATOMIC); | |
6064 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); | |
6065 | } | |
6066 | } | |
6067 | } while (!table && size > PAGE_SIZE && --log2qty); | |
6068 | ||
6069 | if (!table) | |
6070 | panic("Failed to allocate %s hash table\n", tablename); | |
6071 | ||
6072 | printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", | |
6073 | tablename, | |
6074 | (1UL << log2qty), | |
6075 | ilog2(size) - PAGE_SHIFT, | |
6076 | size); | |
6077 | ||
6078 | if (_hash_shift) | |
6079 | *_hash_shift = log2qty; | |
6080 | if (_hash_mask) | |
6081 | *_hash_mask = (1 << log2qty) - 1; | |
6082 | ||
6083 | return table; | |
6084 | } | |
6085 | ||
6086 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ | |
6087 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, | |
6088 | unsigned long pfn) | |
6089 | { | |
6090 | #ifdef CONFIG_SPARSEMEM | |
6091 | return __pfn_to_section(pfn)->pageblock_flags; | |
6092 | #else | |
6093 | return zone->pageblock_flags; | |
6094 | #endif /* CONFIG_SPARSEMEM */ | |
6095 | } | |
6096 | ||
6097 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) | |
6098 | { | |
6099 | #ifdef CONFIG_SPARSEMEM | |
6100 | pfn &= (PAGES_PER_SECTION-1); | |
6101 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
6102 | #else | |
6103 | pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); | |
6104 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
6105 | #endif /* CONFIG_SPARSEMEM */ | |
6106 | } | |
6107 | ||
6108 | /** | |
6109 | * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages | |
6110 | * @page: The page within the block of interest | |
6111 | * @pfn: The target page frame number | |
6112 | * @end_bitidx: The last bit of interest to retrieve | |
6113 | * @mask: mask of bits that the caller is interested in | |
6114 | * | |
6115 | * Return: pageblock_bits flags | |
6116 | */ | |
6117 | unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, | |
6118 | unsigned long end_bitidx, | |
6119 | unsigned long mask) | |
6120 | { | |
6121 | struct zone *zone; | |
6122 | unsigned long *bitmap; | |
6123 | unsigned long bitidx, word_bitidx; | |
6124 | unsigned long word; | |
6125 | ||
6126 | zone = page_zone(page); | |
6127 | bitmap = get_pageblock_bitmap(zone, pfn); | |
6128 | bitidx = pfn_to_bitidx(zone, pfn); | |
6129 | word_bitidx = bitidx / BITS_PER_LONG; | |
6130 | bitidx &= (BITS_PER_LONG-1); | |
6131 | ||
6132 | word = bitmap[word_bitidx]; | |
6133 | bitidx += end_bitidx; | |
6134 | return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; | |
6135 | } | |
6136 | ||
6137 | /** | |
6138 | * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages | |
6139 | * @page: The page within the block of interest | |
6140 | * @flags: The flags to set | |
6141 | * @pfn: The target page frame number | |
6142 | * @end_bitidx: The last bit of interest | |
6143 | * @mask: mask of bits that the caller is interested in | |
6144 | */ | |
6145 | void set_pfnblock_flags_mask(struct page *page, unsigned long flags, | |
6146 | unsigned long pfn, | |
6147 | unsigned long end_bitidx, | |
6148 | unsigned long mask) | |
6149 | { | |
6150 | struct zone *zone; | |
6151 | unsigned long *bitmap; | |
6152 | unsigned long bitidx, word_bitidx; | |
6153 | unsigned long old_word, word; | |
6154 | ||
6155 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); | |
6156 | ||
6157 | zone = page_zone(page); | |
6158 | bitmap = get_pageblock_bitmap(zone, pfn); | |
6159 | bitidx = pfn_to_bitidx(zone, pfn); | |
6160 | word_bitidx = bitidx / BITS_PER_LONG; | |
6161 | bitidx &= (BITS_PER_LONG-1); | |
6162 | ||
6163 | VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page); | |
6164 | ||
6165 | bitidx += end_bitidx; | |
6166 | mask <<= (BITS_PER_LONG - bitidx - 1); | |
6167 | flags <<= (BITS_PER_LONG - bitidx - 1); | |
6168 | ||
6169 | word = ACCESS_ONCE(bitmap[word_bitidx]); | |
6170 | for (;;) { | |
6171 | old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); | |
6172 | if (word == old_word) | |
6173 | break; | |
6174 | word = old_word; | |
6175 | } | |
6176 | } | |
6177 | ||
6178 | /* | |
6179 | * This function checks whether pageblock includes unmovable pages or not. | |
6180 | * If @count is not zero, it is okay to include less @count unmovable pages | |
6181 | * | |
6182 | * PageLRU check without isolation or lru_lock could race so that | |
6183 | * MIGRATE_MOVABLE block might include unmovable pages. It means you can't | |
6184 | * expect this function should be exact. | |
6185 | */ | |
6186 | bool has_unmovable_pages(struct zone *zone, struct page *page, int count, | |
6187 | bool skip_hwpoisoned_pages) | |
6188 | { | |
6189 | unsigned long pfn, iter, found; | |
6190 | int mt; | |
6191 | ||
6192 | /* | |
6193 | * For avoiding noise data, lru_add_drain_all() should be called | |
6194 | * If ZONE_MOVABLE, the zone never contains unmovable pages | |
6195 | */ | |
6196 | if (zone_idx(zone) == ZONE_MOVABLE) | |
6197 | return false; | |
6198 | mt = get_pageblock_migratetype(page); | |
6199 | if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) | |
6200 | return false; | |
6201 | ||
6202 | pfn = page_to_pfn(page); | |
6203 | for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { | |
6204 | unsigned long check = pfn + iter; | |
6205 | ||
6206 | if (!pfn_valid_within(check)) | |
6207 | continue; | |
6208 | ||
6209 | page = pfn_to_page(check); | |
6210 | ||
6211 | /* | |
6212 | * Hugepages are not in LRU lists, but they're movable. | |
6213 | * We need not scan over tail pages bacause we don't | |
6214 | * handle each tail page individually in migration. | |
6215 | */ | |
6216 | if (PageHuge(page)) { | |
6217 | iter = round_up(iter + 1, 1<<compound_order(page)) - 1; | |
6218 | continue; | |
6219 | } | |
6220 | ||
6221 | /* | |
6222 | * We can't use page_count without pin a page | |
6223 | * because another CPU can free compound page. | |
6224 | * This check already skips compound tails of THP | |
6225 | * because their page->_count is zero at all time. | |
6226 | */ | |
6227 | if (!atomic_read(&page->_count)) { | |
6228 | if (PageBuddy(page)) | |
6229 | iter += (1 << page_order(page)) - 1; | |
6230 | continue; | |
6231 | } | |
6232 | ||
6233 | /* | |
6234 | * The HWPoisoned page may be not in buddy system, and | |
6235 | * page_count() is not 0. | |
6236 | */ | |
6237 | if (skip_hwpoisoned_pages && PageHWPoison(page)) | |
6238 | continue; | |
6239 | ||
6240 | if (!PageLRU(page)) | |
6241 | found++; | |
6242 | /* | |
6243 | * If there are RECLAIMABLE pages, we need to check | |
6244 | * it. But now, memory offline itself doesn't call | |
6245 | * shrink_node_slabs() and it still to be fixed. | |
6246 | */ | |
6247 | /* | |
6248 | * If the page is not RAM, page_count()should be 0. | |
6249 | * we don't need more check. This is an _used_ not-movable page. | |
6250 | * | |
6251 | * The problematic thing here is PG_reserved pages. PG_reserved | |
6252 | * is set to both of a memory hole page and a _used_ kernel | |
6253 | * page at boot. | |
6254 | */ | |
6255 | if (found > count) | |
6256 | return true; | |
6257 | } | |
6258 | return false; | |
6259 | } | |
6260 | ||
6261 | bool is_pageblock_removable_nolock(struct page *page) | |
6262 | { | |
6263 | struct zone *zone; | |
6264 | unsigned long pfn; | |
6265 | ||
6266 | /* | |
6267 | * We have to be careful here because we are iterating over memory | |
6268 | * sections which are not zone aware so we might end up outside of | |
6269 | * the zone but still within the section. | |
6270 | * We have to take care about the node as well. If the node is offline | |
6271 | * its NODE_DATA will be NULL - see page_zone. | |
6272 | */ | |
6273 | if (!node_online(page_to_nid(page))) | |
6274 | return false; | |
6275 | ||
6276 | zone = page_zone(page); | |
6277 | pfn = page_to_pfn(page); | |
6278 | if (!zone_spans_pfn(zone, pfn)) | |
6279 | return false; | |
6280 | ||
6281 | return !has_unmovable_pages(zone, page, 0, true); | |
6282 | } | |
6283 | ||
6284 | #ifdef CONFIG_CMA | |
6285 | ||
6286 | static unsigned long pfn_max_align_down(unsigned long pfn) | |
6287 | { | |
6288 | return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
6289 | pageblock_nr_pages) - 1); | |
6290 | } | |
6291 | ||
6292 | static unsigned long pfn_max_align_up(unsigned long pfn) | |
6293 | { | |
6294 | return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
6295 | pageblock_nr_pages)); | |
6296 | } | |
6297 | ||
6298 | /* [start, end) must belong to a single zone. */ | |
6299 | static int __alloc_contig_migrate_range(struct compact_control *cc, | |
6300 | unsigned long start, unsigned long end) | |
6301 | { | |
6302 | /* This function is based on compact_zone() from compaction.c. */ | |
6303 | unsigned long nr_reclaimed; | |
6304 | unsigned long pfn = start; | |
6305 | unsigned int tries = 0; | |
6306 | int ret = 0; | |
6307 | ||
6308 | migrate_prep(); | |
6309 | ||
6310 | while (pfn < end || !list_empty(&cc->migratepages)) { | |
6311 | if (fatal_signal_pending(current)) { | |
6312 | ret = -EINTR; | |
6313 | break; | |
6314 | } | |
6315 | ||
6316 | if (list_empty(&cc->migratepages)) { | |
6317 | cc->nr_migratepages = 0; | |
6318 | pfn = isolate_migratepages_range(cc, pfn, end); | |
6319 | if (!pfn) { | |
6320 | ret = -EINTR; | |
6321 | break; | |
6322 | } | |
6323 | tries = 0; | |
6324 | } else if (++tries == 5) { | |
6325 | ret = ret < 0 ? ret : -EBUSY; | |
6326 | break; | |
6327 | } | |
6328 | ||
6329 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, | |
6330 | &cc->migratepages); | |
6331 | cc->nr_migratepages -= nr_reclaimed; | |
6332 | ||
6333 | ret = migrate_pages(&cc->migratepages, alloc_migrate_target, | |
6334 | NULL, 0, cc->mode, MR_CMA); | |
6335 | } | |
6336 | if (ret < 0) { | |
6337 | putback_movable_pages(&cc->migratepages); | |
6338 | return ret; | |
6339 | } | |
6340 | return 0; | |
6341 | } | |
6342 | ||
6343 | /** | |
6344 | * alloc_contig_range() -- tries to allocate given range of pages | |
6345 | * @start: start PFN to allocate | |
6346 | * @end: one-past-the-last PFN to allocate | |
6347 | * @migratetype: migratetype of the underlaying pageblocks (either | |
6348 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks | |
6349 | * in range must have the same migratetype and it must | |
6350 | * be either of the two. | |
6351 | * | |
6352 | * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES | |
6353 | * aligned, however it's the caller's responsibility to guarantee that | |
6354 | * we are the only thread that changes migrate type of pageblocks the | |
6355 | * pages fall in. | |
6356 | * | |
6357 | * The PFN range must belong to a single zone. | |
6358 | * | |
6359 | * Returns zero on success or negative error code. On success all | |
6360 | * pages which PFN is in [start, end) are allocated for the caller and | |
6361 | * need to be freed with free_contig_range(). | |
6362 | */ | |
6363 | int alloc_contig_range(unsigned long start, unsigned long end, | |
6364 | unsigned migratetype) | |
6365 | { | |
6366 | unsigned long outer_start, outer_end; | |
6367 | int ret = 0, order; | |
6368 | ||
6369 | struct compact_control cc = { | |
6370 | .nr_migratepages = 0, | |
6371 | .order = -1, | |
6372 | .zone = page_zone(pfn_to_page(start)), | |
6373 | .mode = MIGRATE_SYNC, | |
6374 | .ignore_skip_hint = true, | |
6375 | }; | |
6376 | INIT_LIST_HEAD(&cc.migratepages); | |
6377 | ||
6378 | /* | |
6379 | * What we do here is we mark all pageblocks in range as | |
6380 | * MIGRATE_ISOLATE. Because pageblock and max order pages may | |
6381 | * have different sizes, and due to the way page allocator | |
6382 | * work, we align the range to biggest of the two pages so | |
6383 | * that page allocator won't try to merge buddies from | |
6384 | * different pageblocks and change MIGRATE_ISOLATE to some | |
6385 | * other migration type. | |
6386 | * | |
6387 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we | |
6388 | * migrate the pages from an unaligned range (ie. pages that | |
6389 | * we are interested in). This will put all the pages in | |
6390 | * range back to page allocator as MIGRATE_ISOLATE. | |
6391 | * | |
6392 | * When this is done, we take the pages in range from page | |
6393 | * allocator removing them from the buddy system. This way | |
6394 | * page allocator will never consider using them. | |
6395 | * | |
6396 | * This lets us mark the pageblocks back as | |
6397 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the | |
6398 | * aligned range but not in the unaligned, original range are | |
6399 | * put back to page allocator so that buddy can use them. | |
6400 | */ | |
6401 | ||
6402 | ret = start_isolate_page_range(pfn_max_align_down(start), | |
6403 | pfn_max_align_up(end), migratetype, | |
6404 | false); | |
6405 | if (ret) | |
6406 | return ret; | |
6407 | ||
6408 | ret = __alloc_contig_migrate_range(&cc, start, end); | |
6409 | if (ret) | |
6410 | goto done; | |
6411 | ||
6412 | /* | |
6413 | * Pages from [start, end) are within a MAX_ORDER_NR_PAGES | |
6414 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's | |
6415 | * more, all pages in [start, end) are free in page allocator. | |
6416 | * What we are going to do is to allocate all pages from | |
6417 | * [start, end) (that is remove them from page allocator). | |
6418 | * | |
6419 | * The only problem is that pages at the beginning and at the | |
6420 | * end of interesting range may be not aligned with pages that | |
6421 | * page allocator holds, ie. they can be part of higher order | |
6422 | * pages. Because of this, we reserve the bigger range and | |
6423 | * once this is done free the pages we are not interested in. | |
6424 | * | |
6425 | * We don't have to hold zone->lock here because the pages are | |
6426 | * isolated thus they won't get removed from buddy. | |
6427 | */ | |
6428 | ||
6429 | lru_add_drain_all(); | |
6430 | drain_all_pages(cc.zone); | |
6431 | ||
6432 | order = 0; | |
6433 | outer_start = start; | |
6434 | while (!PageBuddy(pfn_to_page(outer_start))) { | |
6435 | if (++order >= MAX_ORDER) { | |
6436 | ret = -EBUSY; | |
6437 | goto done; | |
6438 | } | |
6439 | outer_start &= ~0UL << order; | |
6440 | } | |
6441 | ||
6442 | /* Make sure the range is really isolated. */ | |
6443 | if (test_pages_isolated(outer_start, end, false)) { | |
6444 | pr_info("%s: [%lx, %lx) PFNs busy\n", | |
6445 | __func__, outer_start, end); | |
6446 | ret = -EBUSY; | |
6447 | goto done; | |
6448 | } | |
6449 | ||
6450 | /* Grab isolated pages from freelists. */ | |
6451 | outer_end = isolate_freepages_range(&cc, outer_start, end); | |
6452 | if (!outer_end) { | |
6453 | ret = -EBUSY; | |
6454 | goto done; | |
6455 | } | |
6456 | ||
6457 | /* Free head and tail (if any) */ | |
6458 | if (start != outer_start) | |
6459 | free_contig_range(outer_start, start - outer_start); | |
6460 | if (end != outer_end) | |
6461 | free_contig_range(end, outer_end - end); | |
6462 | ||
6463 | done: | |
6464 | undo_isolate_page_range(pfn_max_align_down(start), | |
6465 | pfn_max_align_up(end), migratetype); | |
6466 | return ret; | |
6467 | } | |
6468 | ||
6469 | void free_contig_range(unsigned long pfn, unsigned nr_pages) | |
6470 | { | |
6471 | unsigned int count = 0; | |
6472 | ||
6473 | for (; nr_pages--; pfn++) { | |
6474 | struct page *page = pfn_to_page(pfn); | |
6475 | ||
6476 | count += page_count(page) != 1; | |
6477 | __free_page(page); | |
6478 | } | |
6479 | WARN(count != 0, "%d pages are still in use!\n", count); | |
6480 | } | |
6481 | #endif | |
6482 | ||
6483 | #ifdef CONFIG_MEMORY_HOTPLUG | |
6484 | /* | |
6485 | * The zone indicated has a new number of managed_pages; batch sizes and percpu | |
6486 | * page high values need to be recalulated. | |
6487 | */ | |
6488 | void __meminit zone_pcp_update(struct zone *zone) | |
6489 | { | |
6490 | unsigned cpu; | |
6491 | mutex_lock(&pcp_batch_high_lock); | |
6492 | for_each_possible_cpu(cpu) | |
6493 | pageset_set_high_and_batch(zone, | |
6494 | per_cpu_ptr(zone->pageset, cpu)); | |
6495 | mutex_unlock(&pcp_batch_high_lock); | |
6496 | } | |
6497 | #endif | |
6498 | ||
6499 | void zone_pcp_reset(struct zone *zone) | |
6500 | { | |
6501 | unsigned long flags; | |
6502 | int cpu; | |
6503 | struct per_cpu_pageset *pset; | |
6504 | ||
6505 | /* avoid races with drain_pages() */ | |
6506 | local_irq_save(flags); | |
6507 | if (zone->pageset != &boot_pageset) { | |
6508 | for_each_online_cpu(cpu) { | |
6509 | pset = per_cpu_ptr(zone->pageset, cpu); | |
6510 | drain_zonestat(zone, pset); | |
6511 | } | |
6512 | free_percpu(zone->pageset); | |
6513 | zone->pageset = &boot_pageset; | |
6514 | } | |
6515 | local_irq_restore(flags); | |
6516 | } | |
6517 | ||
6518 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
6519 | /* | |
6520 | * All pages in the range must be isolated before calling this. | |
6521 | */ | |
6522 | void | |
6523 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) | |
6524 | { | |
6525 | struct page *page; | |
6526 | struct zone *zone; | |
6527 | unsigned int order, i; | |
6528 | unsigned long pfn; | |
6529 | unsigned long flags; | |
6530 | /* find the first valid pfn */ | |
6531 | for (pfn = start_pfn; pfn < end_pfn; pfn++) | |
6532 | if (pfn_valid(pfn)) | |
6533 | break; | |
6534 | if (pfn == end_pfn) | |
6535 | return; | |
6536 | zone = page_zone(pfn_to_page(pfn)); | |
6537 | spin_lock_irqsave(&zone->lock, flags); | |
6538 | pfn = start_pfn; | |
6539 | while (pfn < end_pfn) { | |
6540 | if (!pfn_valid(pfn)) { | |
6541 | pfn++; | |
6542 | continue; | |
6543 | } | |
6544 | page = pfn_to_page(pfn); | |
6545 | /* | |
6546 | * The HWPoisoned page may be not in buddy system, and | |
6547 | * page_count() is not 0. | |
6548 | */ | |
6549 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { | |
6550 | pfn++; | |
6551 | SetPageReserved(page); | |
6552 | continue; | |
6553 | } | |
6554 | ||
6555 | BUG_ON(page_count(page)); | |
6556 | BUG_ON(!PageBuddy(page)); | |
6557 | order = page_order(page); | |
6558 | #ifdef CONFIG_DEBUG_VM | |
6559 | printk(KERN_INFO "remove from free list %lx %d %lx\n", | |
6560 | pfn, 1 << order, end_pfn); | |
6561 | #endif | |
6562 | list_del(&page->lru); | |
6563 | rmv_page_order(page); | |
6564 | zone->free_area[order].nr_free--; | |
6565 | for (i = 0; i < (1 << order); i++) | |
6566 | SetPageReserved((page+i)); | |
6567 | pfn += (1 << order); | |
6568 | } | |
6569 | spin_unlock_irqrestore(&zone->lock, flags); | |
6570 | } | |
6571 | #endif | |
6572 | ||
6573 | #ifdef CONFIG_MEMORY_FAILURE | |
6574 | bool is_free_buddy_page(struct page *page) | |
6575 | { | |
6576 | struct zone *zone = page_zone(page); | |
6577 | unsigned long pfn = page_to_pfn(page); | |
6578 | unsigned long flags; | |
6579 | unsigned int order; | |
6580 | ||
6581 | spin_lock_irqsave(&zone->lock, flags); | |
6582 | for (order = 0; order < MAX_ORDER; order++) { | |
6583 | struct page *page_head = page - (pfn & ((1 << order) - 1)); | |
6584 | ||
6585 | if (PageBuddy(page_head) && page_order(page_head) >= order) | |
6586 | break; | |
6587 | } | |
6588 | spin_unlock_irqrestore(&zone->lock, flags); | |
6589 | ||
6590 | return order < MAX_ORDER; | |
6591 | } | |
6592 | #endif |