]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/page_alloc.c | |
3 | * | |
4 | * Manages the free list, the system allocates free pages here. | |
5 | * Note that kmalloc() lives in slab.c | |
6 | * | |
7 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
8 | * Swap reorganised 29.12.95, Stephen Tweedie | |
9 | * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 | |
10 | * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 | |
11 | * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 | |
12 | * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 | |
13 | * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 | |
14 | * (lots of bits borrowed from Ingo Molnar & Andrew Morton) | |
15 | */ | |
16 | ||
17 | #include <linux/stddef.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/swap.h> | |
20 | #include <linux/interrupt.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/jiffies.h> | |
23 | #include <linux/bootmem.h> | |
24 | #include <linux/memblock.h> | |
25 | #include <linux/compiler.h> | |
26 | #include <linux/kernel.h> | |
27 | #include <linux/kmemcheck.h> | |
28 | #include <linux/module.h> | |
29 | #include <linux/suspend.h> | |
30 | #include <linux/pagevec.h> | |
31 | #include <linux/blkdev.h> | |
32 | #include <linux/slab.h> | |
33 | #include <linux/ratelimit.h> | |
34 | #include <linux/oom.h> | |
35 | #include <linux/notifier.h> | |
36 | #include <linux/topology.h> | |
37 | #include <linux/sysctl.h> | |
38 | #include <linux/cpu.h> | |
39 | #include <linux/cpuset.h> | |
40 | #include <linux/memory_hotplug.h> | |
41 | #include <linux/nodemask.h> | |
42 | #include <linux/vmalloc.h> | |
43 | #include <linux/vmstat.h> | |
44 | #include <linux/mempolicy.h> | |
45 | #include <linux/stop_machine.h> | |
46 | #include <linux/sort.h> | |
47 | #include <linux/pfn.h> | |
48 | #include <linux/backing-dev.h> | |
49 | #include <linux/fault-inject.h> | |
50 | #include <linux/page-isolation.h> | |
51 | #include <linux/page_cgroup.h> | |
52 | #include <linux/debugobjects.h> | |
53 | #include <linux/kmemleak.h> | |
54 | #include <linux/compaction.h> | |
55 | #include <trace/events/kmem.h> | |
56 | #include <linux/ftrace_event.h> | |
57 | #include <linux/memcontrol.h> | |
58 | #include <linux/prefetch.h> | |
59 | #include <linux/migrate.h> | |
60 | #include <linux/page-debug-flags.h> | |
61 | #include <linux/sched/rt.h> | |
62 | ||
63 | #include <asm/tlbflush.h> | |
64 | #include <asm/div64.h> | |
65 | #include "internal.h" | |
66 | ||
67 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID | |
68 | DEFINE_PER_CPU(int, numa_node); | |
69 | EXPORT_PER_CPU_SYMBOL(numa_node); | |
70 | #endif | |
71 | ||
72 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
73 | /* | |
74 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. | |
75 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. | |
76 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() | |
77 | * defined in <linux/topology.h>. | |
78 | */ | |
79 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ | |
80 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); | |
81 | #endif | |
82 | ||
83 | /* | |
84 | * Array of node states. | |
85 | */ | |
86 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { | |
87 | [N_POSSIBLE] = NODE_MASK_ALL, | |
88 | [N_ONLINE] = { { [0] = 1UL } }, | |
89 | #ifndef CONFIG_NUMA | |
90 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
91 | #ifdef CONFIG_HIGHMEM | |
92 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
93 | #endif | |
94 | #ifdef CONFIG_MOVABLE_NODE | |
95 | [N_MEMORY] = { { [0] = 1UL } }, | |
96 | #endif | |
97 | [N_CPU] = { { [0] = 1UL } }, | |
98 | #endif /* NUMA */ | |
99 | }; | |
100 | EXPORT_SYMBOL(node_states); | |
101 | ||
102 | unsigned long totalram_pages __read_mostly; | |
103 | unsigned long totalreserve_pages __read_mostly; | |
104 | /* | |
105 | * When calculating the number of globally allowed dirty pages, there | |
106 | * is a certain number of per-zone reserves that should not be | |
107 | * considered dirtyable memory. This is the sum of those reserves | |
108 | * over all existing zones that contribute dirtyable memory. | |
109 | */ | |
110 | unsigned long dirty_balance_reserve __read_mostly; | |
111 | ||
112 | int percpu_pagelist_fraction; | |
113 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; | |
114 | ||
115 | #ifdef CONFIG_PM_SLEEP | |
116 | /* | |
117 | * The following functions are used by the suspend/hibernate code to temporarily | |
118 | * change gfp_allowed_mask in order to avoid using I/O during memory allocations | |
119 | * while devices are suspended. To avoid races with the suspend/hibernate code, | |
120 | * they should always be called with pm_mutex held (gfp_allowed_mask also should | |
121 | * only be modified with pm_mutex held, unless the suspend/hibernate code is | |
122 | * guaranteed not to run in parallel with that modification). | |
123 | */ | |
124 | ||
125 | static gfp_t saved_gfp_mask; | |
126 | ||
127 | void pm_restore_gfp_mask(void) | |
128 | { | |
129 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
130 | if (saved_gfp_mask) { | |
131 | gfp_allowed_mask = saved_gfp_mask; | |
132 | saved_gfp_mask = 0; | |
133 | } | |
134 | } | |
135 | ||
136 | void pm_restrict_gfp_mask(void) | |
137 | { | |
138 | WARN_ON(!mutex_is_locked(&pm_mutex)); | |
139 | WARN_ON(saved_gfp_mask); | |
140 | saved_gfp_mask = gfp_allowed_mask; | |
141 | gfp_allowed_mask &= ~GFP_IOFS; | |
142 | } | |
143 | ||
144 | bool pm_suspended_storage(void) | |
145 | { | |
146 | if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) | |
147 | return false; | |
148 | return true; | |
149 | } | |
150 | #endif /* CONFIG_PM_SLEEP */ | |
151 | ||
152 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
153 | int pageblock_order __read_mostly; | |
154 | #endif | |
155 | ||
156 | static void __free_pages_ok(struct page *page, unsigned int order); | |
157 | ||
158 | /* | |
159 | * results with 256, 32 in the lowmem_reserve sysctl: | |
160 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
161 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
162 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
163 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
164 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA | |
165 | * | |
166 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
167 | * don't need any ZONE_NORMAL reservation | |
168 | */ | |
169 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { | |
170 | #ifdef CONFIG_ZONE_DMA | |
171 | 256, | |
172 | #endif | |
173 | #ifdef CONFIG_ZONE_DMA32 | |
174 | 256, | |
175 | #endif | |
176 | #ifdef CONFIG_HIGHMEM | |
177 | 32, | |
178 | #endif | |
179 | 32, | |
180 | }; | |
181 | ||
182 | EXPORT_SYMBOL(totalram_pages); | |
183 | ||
184 | static char * const zone_names[MAX_NR_ZONES] = { | |
185 | #ifdef CONFIG_ZONE_DMA | |
186 | "DMA", | |
187 | #endif | |
188 | #ifdef CONFIG_ZONE_DMA32 | |
189 | "DMA32", | |
190 | #endif | |
191 | "Normal", | |
192 | #ifdef CONFIG_HIGHMEM | |
193 | "HighMem", | |
194 | #endif | |
195 | "Movable", | |
196 | }; | |
197 | ||
198 | int min_free_kbytes = 1024; | |
199 | ||
200 | static unsigned long __meminitdata nr_kernel_pages; | |
201 | static unsigned long __meminitdata nr_all_pages; | |
202 | static unsigned long __meminitdata dma_reserve; | |
203 | ||
204 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
205 | /* Movable memory ranges, will also be used by memblock subsystem. */ | |
206 | struct movablemem_map movablemem_map = { | |
207 | .acpi = false, | |
208 | .nr_map = 0, | |
209 | }; | |
210 | ||
211 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; | |
212 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; | |
213 | static unsigned long __initdata required_kernelcore; | |
214 | static unsigned long __initdata required_movablecore; | |
215 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; | |
216 | static unsigned long __meminitdata zone_movable_limit[MAX_NUMNODES]; | |
217 | ||
218 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
219 | int movable_zone; | |
220 | EXPORT_SYMBOL(movable_zone); | |
221 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
222 | ||
223 | #if MAX_NUMNODES > 1 | |
224 | int nr_node_ids __read_mostly = MAX_NUMNODES; | |
225 | int nr_online_nodes __read_mostly = 1; | |
226 | EXPORT_SYMBOL(nr_node_ids); | |
227 | EXPORT_SYMBOL(nr_online_nodes); | |
228 | #endif | |
229 | ||
230 | int page_group_by_mobility_disabled __read_mostly; | |
231 | ||
232 | void set_pageblock_migratetype(struct page *page, int migratetype) | |
233 | { | |
234 | ||
235 | if (unlikely(page_group_by_mobility_disabled)) | |
236 | migratetype = MIGRATE_UNMOVABLE; | |
237 | ||
238 | set_pageblock_flags_group(page, (unsigned long)migratetype, | |
239 | PB_migrate, PB_migrate_end); | |
240 | } | |
241 | ||
242 | bool oom_killer_disabled __read_mostly; | |
243 | ||
244 | #ifdef CONFIG_DEBUG_VM | |
245 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) | |
246 | { | |
247 | int ret = 0; | |
248 | unsigned seq; | |
249 | unsigned long pfn = page_to_pfn(page); | |
250 | ||
251 | do { | |
252 | seq = zone_span_seqbegin(zone); | |
253 | if (!zone_spans_pfn(zone, pfn)) | |
254 | ret = 1; | |
255 | } while (zone_span_seqretry(zone, seq)); | |
256 | ||
257 | return ret; | |
258 | } | |
259 | ||
260 | static int page_is_consistent(struct zone *zone, struct page *page) | |
261 | { | |
262 | if (!pfn_valid_within(page_to_pfn(page))) | |
263 | return 0; | |
264 | if (zone != page_zone(page)) | |
265 | return 0; | |
266 | ||
267 | return 1; | |
268 | } | |
269 | /* | |
270 | * Temporary debugging check for pages not lying within a given zone. | |
271 | */ | |
272 | static int bad_range(struct zone *zone, struct page *page) | |
273 | { | |
274 | if (page_outside_zone_boundaries(zone, page)) | |
275 | return 1; | |
276 | if (!page_is_consistent(zone, page)) | |
277 | return 1; | |
278 | ||
279 | return 0; | |
280 | } | |
281 | #else | |
282 | static inline int bad_range(struct zone *zone, struct page *page) | |
283 | { | |
284 | return 0; | |
285 | } | |
286 | #endif | |
287 | ||
288 | static void bad_page(struct page *page) | |
289 | { | |
290 | static unsigned long resume; | |
291 | static unsigned long nr_shown; | |
292 | static unsigned long nr_unshown; | |
293 | ||
294 | /* Don't complain about poisoned pages */ | |
295 | if (PageHWPoison(page)) { | |
296 | page_mapcount_reset(page); /* remove PageBuddy */ | |
297 | return; | |
298 | } | |
299 | ||
300 | /* | |
301 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
302 | * or allow a steady drip of one report per second. | |
303 | */ | |
304 | if (nr_shown == 60) { | |
305 | if (time_before(jiffies, resume)) { | |
306 | nr_unshown++; | |
307 | goto out; | |
308 | } | |
309 | if (nr_unshown) { | |
310 | printk(KERN_ALERT | |
311 | "BUG: Bad page state: %lu messages suppressed\n", | |
312 | nr_unshown); | |
313 | nr_unshown = 0; | |
314 | } | |
315 | nr_shown = 0; | |
316 | } | |
317 | if (nr_shown++ == 0) | |
318 | resume = jiffies + 60 * HZ; | |
319 | ||
320 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", | |
321 | current->comm, page_to_pfn(page)); | |
322 | dump_page(page); | |
323 | ||
324 | print_modules(); | |
325 | dump_stack(); | |
326 | out: | |
327 | /* Leave bad fields for debug, except PageBuddy could make trouble */ | |
328 | page_mapcount_reset(page); /* remove PageBuddy */ | |
329 | add_taint(TAINT_BAD_PAGE); | |
330 | } | |
331 | ||
332 | /* | |
333 | * Higher-order pages are called "compound pages". They are structured thusly: | |
334 | * | |
335 | * The first PAGE_SIZE page is called the "head page". | |
336 | * | |
337 | * The remaining PAGE_SIZE pages are called "tail pages". | |
338 | * | |
339 | * All pages have PG_compound set. All tail pages have their ->first_page | |
340 | * pointing at the head page. | |
341 | * | |
342 | * The first tail page's ->lru.next holds the address of the compound page's | |
343 | * put_page() function. Its ->lru.prev holds the order of allocation. | |
344 | * This usage means that zero-order pages may not be compound. | |
345 | */ | |
346 | ||
347 | static void free_compound_page(struct page *page) | |
348 | { | |
349 | __free_pages_ok(page, compound_order(page)); | |
350 | } | |
351 | ||
352 | void prep_compound_page(struct page *page, unsigned long order) | |
353 | { | |
354 | int i; | |
355 | int nr_pages = 1 << order; | |
356 | ||
357 | set_compound_page_dtor(page, free_compound_page); | |
358 | set_compound_order(page, order); | |
359 | __SetPageHead(page); | |
360 | for (i = 1; i < nr_pages; i++) { | |
361 | struct page *p = page + i; | |
362 | __SetPageTail(p); | |
363 | set_page_count(p, 0); | |
364 | p->first_page = page; | |
365 | } | |
366 | } | |
367 | ||
368 | /* update __split_huge_page_refcount if you change this function */ | |
369 | static int destroy_compound_page(struct page *page, unsigned long order) | |
370 | { | |
371 | int i; | |
372 | int nr_pages = 1 << order; | |
373 | int bad = 0; | |
374 | ||
375 | if (unlikely(compound_order(page) != order)) { | |
376 | bad_page(page); | |
377 | bad++; | |
378 | } | |
379 | ||
380 | __ClearPageHead(page); | |
381 | ||
382 | for (i = 1; i < nr_pages; i++) { | |
383 | struct page *p = page + i; | |
384 | ||
385 | if (unlikely(!PageTail(p) || (p->first_page != page))) { | |
386 | bad_page(page); | |
387 | bad++; | |
388 | } | |
389 | __ClearPageTail(p); | |
390 | } | |
391 | ||
392 | return bad; | |
393 | } | |
394 | ||
395 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) | |
396 | { | |
397 | int i; | |
398 | ||
399 | /* | |
400 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO | |
401 | * and __GFP_HIGHMEM from hard or soft interrupt context. | |
402 | */ | |
403 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); | |
404 | for (i = 0; i < (1 << order); i++) | |
405 | clear_highpage(page + i); | |
406 | } | |
407 | ||
408 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
409 | unsigned int _debug_guardpage_minorder; | |
410 | ||
411 | static int __init debug_guardpage_minorder_setup(char *buf) | |
412 | { | |
413 | unsigned long res; | |
414 | ||
415 | if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { | |
416 | printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); | |
417 | return 0; | |
418 | } | |
419 | _debug_guardpage_minorder = res; | |
420 | printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); | |
421 | return 0; | |
422 | } | |
423 | __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); | |
424 | ||
425 | static inline void set_page_guard_flag(struct page *page) | |
426 | { | |
427 | __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); | |
428 | } | |
429 | ||
430 | static inline void clear_page_guard_flag(struct page *page) | |
431 | { | |
432 | __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); | |
433 | } | |
434 | #else | |
435 | static inline void set_page_guard_flag(struct page *page) { } | |
436 | static inline void clear_page_guard_flag(struct page *page) { } | |
437 | #endif | |
438 | ||
439 | static inline void set_page_order(struct page *page, int order) | |
440 | { | |
441 | set_page_private(page, order); | |
442 | __SetPageBuddy(page); | |
443 | } | |
444 | ||
445 | static inline void rmv_page_order(struct page *page) | |
446 | { | |
447 | __ClearPageBuddy(page); | |
448 | set_page_private(page, 0); | |
449 | } | |
450 | ||
451 | /* | |
452 | * Locate the struct page for both the matching buddy in our | |
453 | * pair (buddy1) and the combined O(n+1) page they form (page). | |
454 | * | |
455 | * 1) Any buddy B1 will have an order O twin B2 which satisfies | |
456 | * the following equation: | |
457 | * B2 = B1 ^ (1 << O) | |
458 | * For example, if the starting buddy (buddy2) is #8 its order | |
459 | * 1 buddy is #10: | |
460 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 | |
461 | * | |
462 | * 2) Any buddy B will have an order O+1 parent P which | |
463 | * satisfies the following equation: | |
464 | * P = B & ~(1 << O) | |
465 | * | |
466 | * Assumption: *_mem_map is contiguous at least up to MAX_ORDER | |
467 | */ | |
468 | static inline unsigned long | |
469 | __find_buddy_index(unsigned long page_idx, unsigned int order) | |
470 | { | |
471 | return page_idx ^ (1 << order); | |
472 | } | |
473 | ||
474 | /* | |
475 | * This function checks whether a page is free && is the buddy | |
476 | * we can do coalesce a page and its buddy if | |
477 | * (a) the buddy is not in a hole && | |
478 | * (b) the buddy is in the buddy system && | |
479 | * (c) a page and its buddy have the same order && | |
480 | * (d) a page and its buddy are in the same zone. | |
481 | * | |
482 | * For recording whether a page is in the buddy system, we set ->_mapcount -2. | |
483 | * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. | |
484 | * | |
485 | * For recording page's order, we use page_private(page). | |
486 | */ | |
487 | static inline int page_is_buddy(struct page *page, struct page *buddy, | |
488 | int order) | |
489 | { | |
490 | if (!pfn_valid_within(page_to_pfn(buddy))) | |
491 | return 0; | |
492 | ||
493 | if (page_zone_id(page) != page_zone_id(buddy)) | |
494 | return 0; | |
495 | ||
496 | if (page_is_guard(buddy) && page_order(buddy) == order) { | |
497 | VM_BUG_ON(page_count(buddy) != 0); | |
498 | return 1; | |
499 | } | |
500 | ||
501 | if (PageBuddy(buddy) && page_order(buddy) == order) { | |
502 | VM_BUG_ON(page_count(buddy) != 0); | |
503 | return 1; | |
504 | } | |
505 | return 0; | |
506 | } | |
507 | ||
508 | /* | |
509 | * Freeing function for a buddy system allocator. | |
510 | * | |
511 | * The concept of a buddy system is to maintain direct-mapped table | |
512 | * (containing bit values) for memory blocks of various "orders". | |
513 | * The bottom level table contains the map for the smallest allocatable | |
514 | * units of memory (here, pages), and each level above it describes | |
515 | * pairs of units from the levels below, hence, "buddies". | |
516 | * At a high level, all that happens here is marking the table entry | |
517 | * at the bottom level available, and propagating the changes upward | |
518 | * as necessary, plus some accounting needed to play nicely with other | |
519 | * parts of the VM system. | |
520 | * At each level, we keep a list of pages, which are heads of continuous | |
521 | * free pages of length of (1 << order) and marked with _mapcount -2. Page's | |
522 | * order is recorded in page_private(page) field. | |
523 | * So when we are allocating or freeing one, we can derive the state of the | |
524 | * other. That is, if we allocate a small block, and both were | |
525 | * free, the remainder of the region must be split into blocks. | |
526 | * If a block is freed, and its buddy is also free, then this | |
527 | * triggers coalescing into a block of larger size. | |
528 | * | |
529 | * -- nyc | |
530 | */ | |
531 | ||
532 | static inline void __free_one_page(struct page *page, | |
533 | struct zone *zone, unsigned int order, | |
534 | int migratetype) | |
535 | { | |
536 | unsigned long page_idx; | |
537 | unsigned long combined_idx; | |
538 | unsigned long uninitialized_var(buddy_idx); | |
539 | struct page *buddy; | |
540 | ||
541 | VM_BUG_ON(!zone_is_initialized(zone)); | |
542 | ||
543 | if (unlikely(PageCompound(page))) | |
544 | if (unlikely(destroy_compound_page(page, order))) | |
545 | return; | |
546 | ||
547 | VM_BUG_ON(migratetype == -1); | |
548 | ||
549 | page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); | |
550 | ||
551 | VM_BUG_ON(page_idx & ((1 << order) - 1)); | |
552 | VM_BUG_ON(bad_range(zone, page)); | |
553 | ||
554 | while (order < MAX_ORDER-1) { | |
555 | buddy_idx = __find_buddy_index(page_idx, order); | |
556 | buddy = page + (buddy_idx - page_idx); | |
557 | if (!page_is_buddy(page, buddy, order)) | |
558 | break; | |
559 | /* | |
560 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, | |
561 | * merge with it and move up one order. | |
562 | */ | |
563 | if (page_is_guard(buddy)) { | |
564 | clear_page_guard_flag(buddy); | |
565 | set_page_private(page, 0); | |
566 | __mod_zone_freepage_state(zone, 1 << order, | |
567 | migratetype); | |
568 | } else { | |
569 | list_del(&buddy->lru); | |
570 | zone->free_area[order].nr_free--; | |
571 | rmv_page_order(buddy); | |
572 | } | |
573 | combined_idx = buddy_idx & page_idx; | |
574 | page = page + (combined_idx - page_idx); | |
575 | page_idx = combined_idx; | |
576 | order++; | |
577 | } | |
578 | set_page_order(page, order); | |
579 | ||
580 | /* | |
581 | * If this is not the largest possible page, check if the buddy | |
582 | * of the next-highest order is free. If it is, it's possible | |
583 | * that pages are being freed that will coalesce soon. In case, | |
584 | * that is happening, add the free page to the tail of the list | |
585 | * so it's less likely to be used soon and more likely to be merged | |
586 | * as a higher order page | |
587 | */ | |
588 | if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { | |
589 | struct page *higher_page, *higher_buddy; | |
590 | combined_idx = buddy_idx & page_idx; | |
591 | higher_page = page + (combined_idx - page_idx); | |
592 | buddy_idx = __find_buddy_index(combined_idx, order + 1); | |
593 | higher_buddy = higher_page + (buddy_idx - combined_idx); | |
594 | if (page_is_buddy(higher_page, higher_buddy, order + 1)) { | |
595 | list_add_tail(&page->lru, | |
596 | &zone->free_area[order].free_list[migratetype]); | |
597 | goto out; | |
598 | } | |
599 | } | |
600 | ||
601 | list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); | |
602 | out: | |
603 | zone->free_area[order].nr_free++; | |
604 | } | |
605 | ||
606 | static inline int free_pages_check(struct page *page) | |
607 | { | |
608 | if (unlikely(page_mapcount(page) | | |
609 | (page->mapping != NULL) | | |
610 | (atomic_read(&page->_count) != 0) | | |
611 | (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | | |
612 | (mem_cgroup_bad_page_check(page)))) { | |
613 | bad_page(page); | |
614 | return 1; | |
615 | } | |
616 | page_nid_reset_last(page); | |
617 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | |
618 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
619 | return 0; | |
620 | } | |
621 | ||
622 | /* | |
623 | * Frees a number of pages from the PCP lists | |
624 | * Assumes all pages on list are in same zone, and of same order. | |
625 | * count is the number of pages to free. | |
626 | * | |
627 | * If the zone was previously in an "all pages pinned" state then look to | |
628 | * see if this freeing clears that state. | |
629 | * | |
630 | * And clear the zone's pages_scanned counter, to hold off the "all pages are | |
631 | * pinned" detection logic. | |
632 | */ | |
633 | static void free_pcppages_bulk(struct zone *zone, int count, | |
634 | struct per_cpu_pages *pcp) | |
635 | { | |
636 | int migratetype = 0; | |
637 | int batch_free = 0; | |
638 | int to_free = count; | |
639 | ||
640 | spin_lock(&zone->lock); | |
641 | zone->all_unreclaimable = 0; | |
642 | zone->pages_scanned = 0; | |
643 | ||
644 | while (to_free) { | |
645 | struct page *page; | |
646 | struct list_head *list; | |
647 | ||
648 | /* | |
649 | * Remove pages from lists in a round-robin fashion. A | |
650 | * batch_free count is maintained that is incremented when an | |
651 | * empty list is encountered. This is so more pages are freed | |
652 | * off fuller lists instead of spinning excessively around empty | |
653 | * lists | |
654 | */ | |
655 | do { | |
656 | batch_free++; | |
657 | if (++migratetype == MIGRATE_PCPTYPES) | |
658 | migratetype = 0; | |
659 | list = &pcp->lists[migratetype]; | |
660 | } while (list_empty(list)); | |
661 | ||
662 | /* This is the only non-empty list. Free them all. */ | |
663 | if (batch_free == MIGRATE_PCPTYPES) | |
664 | batch_free = to_free; | |
665 | ||
666 | do { | |
667 | int mt; /* migratetype of the to-be-freed page */ | |
668 | ||
669 | page = list_entry(list->prev, struct page, lru); | |
670 | /* must delete as __free_one_page list manipulates */ | |
671 | list_del(&page->lru); | |
672 | mt = get_freepage_migratetype(page); | |
673 | /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ | |
674 | __free_one_page(page, zone, 0, mt); | |
675 | trace_mm_page_pcpu_drain(page, 0, mt); | |
676 | if (likely(!is_migrate_isolate_page(page))) { | |
677 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1); | |
678 | if (is_migrate_cma(mt)) | |
679 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1); | |
680 | } | |
681 | } while (--to_free && --batch_free && !list_empty(list)); | |
682 | } | |
683 | spin_unlock(&zone->lock); | |
684 | } | |
685 | ||
686 | static void free_one_page(struct zone *zone, struct page *page, int order, | |
687 | int migratetype) | |
688 | { | |
689 | spin_lock(&zone->lock); | |
690 | zone->all_unreclaimable = 0; | |
691 | zone->pages_scanned = 0; | |
692 | ||
693 | __free_one_page(page, zone, order, migratetype); | |
694 | if (unlikely(!is_migrate_isolate(migratetype))) | |
695 | __mod_zone_freepage_state(zone, 1 << order, migratetype); | |
696 | spin_unlock(&zone->lock); | |
697 | } | |
698 | ||
699 | static bool free_pages_prepare(struct page *page, unsigned int order) | |
700 | { | |
701 | int i; | |
702 | int bad = 0; | |
703 | ||
704 | trace_mm_page_free(page, order); | |
705 | kmemcheck_free_shadow(page, order); | |
706 | ||
707 | if (PageAnon(page)) | |
708 | page->mapping = NULL; | |
709 | for (i = 0; i < (1 << order); i++) | |
710 | bad += free_pages_check(page + i); | |
711 | if (bad) | |
712 | return false; | |
713 | ||
714 | if (!PageHighMem(page)) { | |
715 | debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); | |
716 | debug_check_no_obj_freed(page_address(page), | |
717 | PAGE_SIZE << order); | |
718 | } | |
719 | arch_free_page(page, order); | |
720 | kernel_map_pages(page, 1 << order, 0); | |
721 | ||
722 | return true; | |
723 | } | |
724 | ||
725 | static void __free_pages_ok(struct page *page, unsigned int order) | |
726 | { | |
727 | unsigned long flags; | |
728 | int migratetype; | |
729 | ||
730 | if (!free_pages_prepare(page, order)) | |
731 | return; | |
732 | ||
733 | local_irq_save(flags); | |
734 | __count_vm_events(PGFREE, 1 << order); | |
735 | migratetype = get_pageblock_migratetype(page); | |
736 | set_freepage_migratetype(page, migratetype); | |
737 | free_one_page(page_zone(page), page, order, migratetype); | |
738 | local_irq_restore(flags); | |
739 | } | |
740 | ||
741 | /* | |
742 | * Read access to zone->managed_pages is safe because it's unsigned long, | |
743 | * but we still need to serialize writers. Currently all callers of | |
744 | * __free_pages_bootmem() except put_page_bootmem() should only be used | |
745 | * at boot time. So for shorter boot time, we shift the burden to | |
746 | * put_page_bootmem() to serialize writers. | |
747 | */ | |
748 | void __meminit __free_pages_bootmem(struct page *page, unsigned int order) | |
749 | { | |
750 | unsigned int nr_pages = 1 << order; | |
751 | unsigned int loop; | |
752 | ||
753 | prefetchw(page); | |
754 | for (loop = 0; loop < nr_pages; loop++) { | |
755 | struct page *p = &page[loop]; | |
756 | ||
757 | if (loop + 1 < nr_pages) | |
758 | prefetchw(p + 1); | |
759 | __ClearPageReserved(p); | |
760 | set_page_count(p, 0); | |
761 | } | |
762 | ||
763 | page_zone(page)->managed_pages += 1 << order; | |
764 | set_page_refcounted(page); | |
765 | __free_pages(page, order); | |
766 | } | |
767 | ||
768 | #ifdef CONFIG_CMA | |
769 | /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */ | |
770 | void __init init_cma_reserved_pageblock(struct page *page) | |
771 | { | |
772 | unsigned i = pageblock_nr_pages; | |
773 | struct page *p = page; | |
774 | ||
775 | do { | |
776 | __ClearPageReserved(p); | |
777 | set_page_count(p, 0); | |
778 | } while (++p, --i); | |
779 | ||
780 | set_page_refcounted(page); | |
781 | set_pageblock_migratetype(page, MIGRATE_CMA); | |
782 | __free_pages(page, pageblock_order); | |
783 | totalram_pages += pageblock_nr_pages; | |
784 | #ifdef CONFIG_HIGHMEM | |
785 | if (PageHighMem(page)) | |
786 | totalhigh_pages += pageblock_nr_pages; | |
787 | #endif | |
788 | } | |
789 | #endif | |
790 | ||
791 | /* | |
792 | * The order of subdivision here is critical for the IO subsystem. | |
793 | * Please do not alter this order without good reasons and regression | |
794 | * testing. Specifically, as large blocks of memory are subdivided, | |
795 | * the order in which smaller blocks are delivered depends on the order | |
796 | * they're subdivided in this function. This is the primary factor | |
797 | * influencing the order in which pages are delivered to the IO | |
798 | * subsystem according to empirical testing, and this is also justified | |
799 | * by considering the behavior of a buddy system containing a single | |
800 | * large block of memory acted on by a series of small allocations. | |
801 | * This behavior is a critical factor in sglist merging's success. | |
802 | * | |
803 | * -- nyc | |
804 | */ | |
805 | static inline void expand(struct zone *zone, struct page *page, | |
806 | int low, int high, struct free_area *area, | |
807 | int migratetype) | |
808 | { | |
809 | unsigned long size = 1 << high; | |
810 | ||
811 | while (high > low) { | |
812 | area--; | |
813 | high--; | |
814 | size >>= 1; | |
815 | VM_BUG_ON(bad_range(zone, &page[size])); | |
816 | ||
817 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
818 | if (high < debug_guardpage_minorder()) { | |
819 | /* | |
820 | * Mark as guard pages (or page), that will allow to | |
821 | * merge back to allocator when buddy will be freed. | |
822 | * Corresponding page table entries will not be touched, | |
823 | * pages will stay not present in virtual address space | |
824 | */ | |
825 | INIT_LIST_HEAD(&page[size].lru); | |
826 | set_page_guard_flag(&page[size]); | |
827 | set_page_private(&page[size], high); | |
828 | /* Guard pages are not available for any usage */ | |
829 | __mod_zone_freepage_state(zone, -(1 << high), | |
830 | migratetype); | |
831 | continue; | |
832 | } | |
833 | #endif | |
834 | list_add(&page[size].lru, &area->free_list[migratetype]); | |
835 | area->nr_free++; | |
836 | set_page_order(&page[size], high); | |
837 | } | |
838 | } | |
839 | ||
840 | /* | |
841 | * This page is about to be returned from the page allocator | |
842 | */ | |
843 | static inline int check_new_page(struct page *page) | |
844 | { | |
845 | if (unlikely(page_mapcount(page) | | |
846 | (page->mapping != NULL) | | |
847 | (atomic_read(&page->_count) != 0) | | |
848 | (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | | |
849 | (mem_cgroup_bad_page_check(page)))) { | |
850 | bad_page(page); | |
851 | return 1; | |
852 | } | |
853 | return 0; | |
854 | } | |
855 | ||
856 | static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) | |
857 | { | |
858 | int i; | |
859 | ||
860 | for (i = 0; i < (1 << order); i++) { | |
861 | struct page *p = page + i; | |
862 | if (unlikely(check_new_page(p))) | |
863 | return 1; | |
864 | } | |
865 | ||
866 | set_page_private(page, 0); | |
867 | set_page_refcounted(page); | |
868 | ||
869 | arch_alloc_page(page, order); | |
870 | kernel_map_pages(page, 1 << order, 1); | |
871 | ||
872 | if (gfp_flags & __GFP_ZERO) | |
873 | prep_zero_page(page, order, gfp_flags); | |
874 | ||
875 | if (order && (gfp_flags & __GFP_COMP)) | |
876 | prep_compound_page(page, order); | |
877 | ||
878 | return 0; | |
879 | } | |
880 | ||
881 | /* | |
882 | * Go through the free lists for the given migratetype and remove | |
883 | * the smallest available page from the freelists | |
884 | */ | |
885 | static inline | |
886 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, | |
887 | int migratetype) | |
888 | { | |
889 | unsigned int current_order; | |
890 | struct free_area * area; | |
891 | struct page *page; | |
892 | ||
893 | /* Find a page of the appropriate size in the preferred list */ | |
894 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { | |
895 | area = &(zone->free_area[current_order]); | |
896 | if (list_empty(&area->free_list[migratetype])) | |
897 | continue; | |
898 | ||
899 | page = list_entry(area->free_list[migratetype].next, | |
900 | struct page, lru); | |
901 | list_del(&page->lru); | |
902 | rmv_page_order(page); | |
903 | area->nr_free--; | |
904 | expand(zone, page, order, current_order, area, migratetype); | |
905 | return page; | |
906 | } | |
907 | ||
908 | return NULL; | |
909 | } | |
910 | ||
911 | ||
912 | /* | |
913 | * This array describes the order lists are fallen back to when | |
914 | * the free lists for the desirable migrate type are depleted | |
915 | */ | |
916 | static int fallbacks[MIGRATE_TYPES][4] = { | |
917 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
918 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, | |
919 | #ifdef CONFIG_CMA | |
920 | [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
921 | [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ | |
922 | #else | |
923 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, | |
924 | #endif | |
925 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ | |
926 | #ifdef CONFIG_MEMORY_ISOLATION | |
927 | [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ | |
928 | #endif | |
929 | }; | |
930 | ||
931 | /* | |
932 | * Move the free pages in a range to the free lists of the requested type. | |
933 | * Note that start_page and end_pages are not aligned on a pageblock | |
934 | * boundary. If alignment is required, use move_freepages_block() | |
935 | */ | |
936 | int move_freepages(struct zone *zone, | |
937 | struct page *start_page, struct page *end_page, | |
938 | int migratetype) | |
939 | { | |
940 | struct page *page; | |
941 | unsigned long order; | |
942 | int pages_moved = 0; | |
943 | ||
944 | #ifndef CONFIG_HOLES_IN_ZONE | |
945 | /* | |
946 | * page_zone is not safe to call in this context when | |
947 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant | |
948 | * anyway as we check zone boundaries in move_freepages_block(). | |
949 | * Remove at a later date when no bug reports exist related to | |
950 | * grouping pages by mobility | |
951 | */ | |
952 | BUG_ON(page_zone(start_page) != page_zone(end_page)); | |
953 | #endif | |
954 | ||
955 | for (page = start_page; page <= end_page;) { | |
956 | /* Make sure we are not inadvertently changing nodes */ | |
957 | VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); | |
958 | ||
959 | if (!pfn_valid_within(page_to_pfn(page))) { | |
960 | page++; | |
961 | continue; | |
962 | } | |
963 | ||
964 | if (!PageBuddy(page)) { | |
965 | page++; | |
966 | continue; | |
967 | } | |
968 | ||
969 | order = page_order(page); | |
970 | list_move(&page->lru, | |
971 | &zone->free_area[order].free_list[migratetype]); | |
972 | set_freepage_migratetype(page, migratetype); | |
973 | page += 1 << order; | |
974 | pages_moved += 1 << order; | |
975 | } | |
976 | ||
977 | return pages_moved; | |
978 | } | |
979 | ||
980 | int move_freepages_block(struct zone *zone, struct page *page, | |
981 | int migratetype) | |
982 | { | |
983 | unsigned long start_pfn, end_pfn; | |
984 | struct page *start_page, *end_page; | |
985 | ||
986 | start_pfn = page_to_pfn(page); | |
987 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); | |
988 | start_page = pfn_to_page(start_pfn); | |
989 | end_page = start_page + pageblock_nr_pages - 1; | |
990 | end_pfn = start_pfn + pageblock_nr_pages - 1; | |
991 | ||
992 | /* Do not cross zone boundaries */ | |
993 | if (!zone_spans_pfn(zone, start_pfn)) | |
994 | start_page = page; | |
995 | if (!zone_spans_pfn(zone, end_pfn)) | |
996 | return 0; | |
997 | ||
998 | return move_freepages(zone, start_page, end_page, migratetype); | |
999 | } | |
1000 | ||
1001 | static void change_pageblock_range(struct page *pageblock_page, | |
1002 | int start_order, int migratetype) | |
1003 | { | |
1004 | int nr_pageblocks = 1 << (start_order - pageblock_order); | |
1005 | ||
1006 | while (nr_pageblocks--) { | |
1007 | set_pageblock_migratetype(pageblock_page, migratetype); | |
1008 | pageblock_page += pageblock_nr_pages; | |
1009 | } | |
1010 | } | |
1011 | ||
1012 | /* Remove an element from the buddy allocator from the fallback list */ | |
1013 | static inline struct page * | |
1014 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) | |
1015 | { | |
1016 | struct free_area * area; | |
1017 | int current_order; | |
1018 | struct page *page; | |
1019 | int migratetype, i; | |
1020 | ||
1021 | /* Find the largest possible block of pages in the other list */ | |
1022 | for (current_order = MAX_ORDER-1; current_order >= order; | |
1023 | --current_order) { | |
1024 | for (i = 0;; i++) { | |
1025 | migratetype = fallbacks[start_migratetype][i]; | |
1026 | ||
1027 | /* MIGRATE_RESERVE handled later if necessary */ | |
1028 | if (migratetype == MIGRATE_RESERVE) | |
1029 | break; | |
1030 | ||
1031 | area = &(zone->free_area[current_order]); | |
1032 | if (list_empty(&area->free_list[migratetype])) | |
1033 | continue; | |
1034 | ||
1035 | page = list_entry(area->free_list[migratetype].next, | |
1036 | struct page, lru); | |
1037 | area->nr_free--; | |
1038 | ||
1039 | /* | |
1040 | * If breaking a large block of pages, move all free | |
1041 | * pages to the preferred allocation list. If falling | |
1042 | * back for a reclaimable kernel allocation, be more | |
1043 | * aggressive about taking ownership of free pages | |
1044 | * | |
1045 | * On the other hand, never change migration | |
1046 | * type of MIGRATE_CMA pageblocks nor move CMA | |
1047 | * pages on different free lists. We don't | |
1048 | * want unmovable pages to be allocated from | |
1049 | * MIGRATE_CMA areas. | |
1050 | */ | |
1051 | if (!is_migrate_cma(migratetype) && | |
1052 | (unlikely(current_order >= pageblock_order / 2) || | |
1053 | start_migratetype == MIGRATE_RECLAIMABLE || | |
1054 | page_group_by_mobility_disabled)) { | |
1055 | int pages; | |
1056 | pages = move_freepages_block(zone, page, | |
1057 | start_migratetype); | |
1058 | ||
1059 | /* Claim the whole block if over half of it is free */ | |
1060 | if (pages >= (1 << (pageblock_order-1)) || | |
1061 | page_group_by_mobility_disabled) | |
1062 | set_pageblock_migratetype(page, | |
1063 | start_migratetype); | |
1064 | ||
1065 | migratetype = start_migratetype; | |
1066 | } | |
1067 | ||
1068 | /* Remove the page from the freelists */ | |
1069 | list_del(&page->lru); | |
1070 | rmv_page_order(page); | |
1071 | ||
1072 | /* Take ownership for orders >= pageblock_order */ | |
1073 | if (current_order >= pageblock_order && | |
1074 | !is_migrate_cma(migratetype)) | |
1075 | change_pageblock_range(page, current_order, | |
1076 | start_migratetype); | |
1077 | ||
1078 | expand(zone, page, order, current_order, area, | |
1079 | is_migrate_cma(migratetype) | |
1080 | ? migratetype : start_migratetype); | |
1081 | ||
1082 | trace_mm_page_alloc_extfrag(page, order, current_order, | |
1083 | start_migratetype, migratetype); | |
1084 | ||
1085 | return page; | |
1086 | } | |
1087 | } | |
1088 | ||
1089 | return NULL; | |
1090 | } | |
1091 | ||
1092 | /* | |
1093 | * Do the hard work of removing an element from the buddy allocator. | |
1094 | * Call me with the zone->lock already held. | |
1095 | */ | |
1096 | static struct page *__rmqueue(struct zone *zone, unsigned int order, | |
1097 | int migratetype) | |
1098 | { | |
1099 | struct page *page; | |
1100 | ||
1101 | retry_reserve: | |
1102 | page = __rmqueue_smallest(zone, order, migratetype); | |
1103 | ||
1104 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { | |
1105 | page = __rmqueue_fallback(zone, order, migratetype); | |
1106 | ||
1107 | /* | |
1108 | * Use MIGRATE_RESERVE rather than fail an allocation. goto | |
1109 | * is used because __rmqueue_smallest is an inline function | |
1110 | * and we want just one call site | |
1111 | */ | |
1112 | if (!page) { | |
1113 | migratetype = MIGRATE_RESERVE; | |
1114 | goto retry_reserve; | |
1115 | } | |
1116 | } | |
1117 | ||
1118 | trace_mm_page_alloc_zone_locked(page, order, migratetype); | |
1119 | return page; | |
1120 | } | |
1121 | ||
1122 | /* | |
1123 | * Obtain a specified number of elements from the buddy allocator, all under | |
1124 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
1125 | * Returns the number of new pages which were placed at *list. | |
1126 | */ | |
1127 | static int rmqueue_bulk(struct zone *zone, unsigned int order, | |
1128 | unsigned long count, struct list_head *list, | |
1129 | int migratetype, int cold) | |
1130 | { | |
1131 | int mt = migratetype, i; | |
1132 | ||
1133 | spin_lock(&zone->lock); | |
1134 | for (i = 0; i < count; ++i) { | |
1135 | struct page *page = __rmqueue(zone, order, migratetype); | |
1136 | if (unlikely(page == NULL)) | |
1137 | break; | |
1138 | ||
1139 | /* | |
1140 | * Split buddy pages returned by expand() are received here | |
1141 | * in physical page order. The page is added to the callers and | |
1142 | * list and the list head then moves forward. From the callers | |
1143 | * perspective, the linked list is ordered by page number in | |
1144 | * some conditions. This is useful for IO devices that can | |
1145 | * merge IO requests if the physical pages are ordered | |
1146 | * properly. | |
1147 | */ | |
1148 | if (likely(cold == 0)) | |
1149 | list_add(&page->lru, list); | |
1150 | else | |
1151 | list_add_tail(&page->lru, list); | |
1152 | if (IS_ENABLED(CONFIG_CMA)) { | |
1153 | mt = get_pageblock_migratetype(page); | |
1154 | if (!is_migrate_cma(mt) && !is_migrate_isolate(mt)) | |
1155 | mt = migratetype; | |
1156 | } | |
1157 | set_freepage_migratetype(page, mt); | |
1158 | list = &page->lru; | |
1159 | if (is_migrate_cma(mt)) | |
1160 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, | |
1161 | -(1 << order)); | |
1162 | } | |
1163 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); | |
1164 | spin_unlock(&zone->lock); | |
1165 | return i; | |
1166 | } | |
1167 | ||
1168 | #ifdef CONFIG_NUMA | |
1169 | /* | |
1170 | * Called from the vmstat counter updater to drain pagesets of this | |
1171 | * currently executing processor on remote nodes after they have | |
1172 | * expired. | |
1173 | * | |
1174 | * Note that this function must be called with the thread pinned to | |
1175 | * a single processor. | |
1176 | */ | |
1177 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) | |
1178 | { | |
1179 | unsigned long flags; | |
1180 | int to_drain; | |
1181 | ||
1182 | local_irq_save(flags); | |
1183 | if (pcp->count >= pcp->batch) | |
1184 | to_drain = pcp->batch; | |
1185 | else | |
1186 | to_drain = pcp->count; | |
1187 | if (to_drain > 0) { | |
1188 | free_pcppages_bulk(zone, to_drain, pcp); | |
1189 | pcp->count -= to_drain; | |
1190 | } | |
1191 | local_irq_restore(flags); | |
1192 | } | |
1193 | #endif | |
1194 | ||
1195 | /* | |
1196 | * Drain pages of the indicated processor. | |
1197 | * | |
1198 | * The processor must either be the current processor and the | |
1199 | * thread pinned to the current processor or a processor that | |
1200 | * is not online. | |
1201 | */ | |
1202 | static void drain_pages(unsigned int cpu) | |
1203 | { | |
1204 | unsigned long flags; | |
1205 | struct zone *zone; | |
1206 | ||
1207 | for_each_populated_zone(zone) { | |
1208 | struct per_cpu_pageset *pset; | |
1209 | struct per_cpu_pages *pcp; | |
1210 | ||
1211 | local_irq_save(flags); | |
1212 | pset = per_cpu_ptr(zone->pageset, cpu); | |
1213 | ||
1214 | pcp = &pset->pcp; | |
1215 | if (pcp->count) { | |
1216 | free_pcppages_bulk(zone, pcp->count, pcp); | |
1217 | pcp->count = 0; | |
1218 | } | |
1219 | local_irq_restore(flags); | |
1220 | } | |
1221 | } | |
1222 | ||
1223 | /* | |
1224 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
1225 | */ | |
1226 | void drain_local_pages(void *arg) | |
1227 | { | |
1228 | drain_pages(smp_processor_id()); | |
1229 | } | |
1230 | ||
1231 | /* | |
1232 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. | |
1233 | * | |
1234 | * Note that this code is protected against sending an IPI to an offline | |
1235 | * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: | |
1236 | * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but | |
1237 | * nothing keeps CPUs from showing up after we populated the cpumask and | |
1238 | * before the call to on_each_cpu_mask(). | |
1239 | */ | |
1240 | void drain_all_pages(void) | |
1241 | { | |
1242 | int cpu; | |
1243 | struct per_cpu_pageset *pcp; | |
1244 | struct zone *zone; | |
1245 | ||
1246 | /* | |
1247 | * Allocate in the BSS so we wont require allocation in | |
1248 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y | |
1249 | */ | |
1250 | static cpumask_t cpus_with_pcps; | |
1251 | ||
1252 | /* | |
1253 | * We don't care about racing with CPU hotplug event | |
1254 | * as offline notification will cause the notified | |
1255 | * cpu to drain that CPU pcps and on_each_cpu_mask | |
1256 | * disables preemption as part of its processing | |
1257 | */ | |
1258 | for_each_online_cpu(cpu) { | |
1259 | bool has_pcps = false; | |
1260 | for_each_populated_zone(zone) { | |
1261 | pcp = per_cpu_ptr(zone->pageset, cpu); | |
1262 | if (pcp->pcp.count) { | |
1263 | has_pcps = true; | |
1264 | break; | |
1265 | } | |
1266 | } | |
1267 | if (has_pcps) | |
1268 | cpumask_set_cpu(cpu, &cpus_with_pcps); | |
1269 | else | |
1270 | cpumask_clear_cpu(cpu, &cpus_with_pcps); | |
1271 | } | |
1272 | on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1); | |
1273 | } | |
1274 | ||
1275 | #ifdef CONFIG_HIBERNATION | |
1276 | ||
1277 | void mark_free_pages(struct zone *zone) | |
1278 | { | |
1279 | unsigned long pfn, max_zone_pfn; | |
1280 | unsigned long flags; | |
1281 | int order, t; | |
1282 | struct list_head *curr; | |
1283 | ||
1284 | if (!zone->spanned_pages) | |
1285 | return; | |
1286 | ||
1287 | spin_lock_irqsave(&zone->lock, flags); | |
1288 | ||
1289 | max_zone_pfn = zone_end_pfn(zone); | |
1290 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) | |
1291 | if (pfn_valid(pfn)) { | |
1292 | struct page *page = pfn_to_page(pfn); | |
1293 | ||
1294 | if (!swsusp_page_is_forbidden(page)) | |
1295 | swsusp_unset_page_free(page); | |
1296 | } | |
1297 | ||
1298 | for_each_migratetype_order(order, t) { | |
1299 | list_for_each(curr, &zone->free_area[order].free_list[t]) { | |
1300 | unsigned long i; | |
1301 | ||
1302 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); | |
1303 | for (i = 0; i < (1UL << order); i++) | |
1304 | swsusp_set_page_free(pfn_to_page(pfn + i)); | |
1305 | } | |
1306 | } | |
1307 | spin_unlock_irqrestore(&zone->lock, flags); | |
1308 | } | |
1309 | #endif /* CONFIG_PM */ | |
1310 | ||
1311 | /* | |
1312 | * Free a 0-order page | |
1313 | * cold == 1 ? free a cold page : free a hot page | |
1314 | */ | |
1315 | void free_hot_cold_page(struct page *page, int cold) | |
1316 | { | |
1317 | struct zone *zone = page_zone(page); | |
1318 | struct per_cpu_pages *pcp; | |
1319 | unsigned long flags; | |
1320 | int migratetype; | |
1321 | ||
1322 | if (!free_pages_prepare(page, 0)) | |
1323 | return; | |
1324 | ||
1325 | migratetype = get_pageblock_migratetype(page); | |
1326 | set_freepage_migratetype(page, migratetype); | |
1327 | local_irq_save(flags); | |
1328 | __count_vm_event(PGFREE); | |
1329 | ||
1330 | /* | |
1331 | * We only track unmovable, reclaimable and movable on pcp lists. | |
1332 | * Free ISOLATE pages back to the allocator because they are being | |
1333 | * offlined but treat RESERVE as movable pages so we can get those | |
1334 | * areas back if necessary. Otherwise, we may have to free | |
1335 | * excessively into the page allocator | |
1336 | */ | |
1337 | if (migratetype >= MIGRATE_PCPTYPES) { | |
1338 | if (unlikely(is_migrate_isolate(migratetype))) { | |
1339 | free_one_page(zone, page, 0, migratetype); | |
1340 | goto out; | |
1341 | } | |
1342 | migratetype = MIGRATE_MOVABLE; | |
1343 | } | |
1344 | ||
1345 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1346 | if (cold) | |
1347 | list_add_tail(&page->lru, &pcp->lists[migratetype]); | |
1348 | else | |
1349 | list_add(&page->lru, &pcp->lists[migratetype]); | |
1350 | pcp->count++; | |
1351 | if (pcp->count >= pcp->high) { | |
1352 | free_pcppages_bulk(zone, pcp->batch, pcp); | |
1353 | pcp->count -= pcp->batch; | |
1354 | } | |
1355 | ||
1356 | out: | |
1357 | local_irq_restore(flags); | |
1358 | } | |
1359 | ||
1360 | /* | |
1361 | * Free a list of 0-order pages | |
1362 | */ | |
1363 | void free_hot_cold_page_list(struct list_head *list, int cold) | |
1364 | { | |
1365 | struct page *page, *next; | |
1366 | ||
1367 | list_for_each_entry_safe(page, next, list, lru) { | |
1368 | trace_mm_page_free_batched(page, cold); | |
1369 | free_hot_cold_page(page, cold); | |
1370 | } | |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * split_page takes a non-compound higher-order page, and splits it into | |
1375 | * n (1<<order) sub-pages: page[0..n] | |
1376 | * Each sub-page must be freed individually. | |
1377 | * | |
1378 | * Note: this is probably too low level an operation for use in drivers. | |
1379 | * Please consult with lkml before using this in your driver. | |
1380 | */ | |
1381 | void split_page(struct page *page, unsigned int order) | |
1382 | { | |
1383 | int i; | |
1384 | ||
1385 | VM_BUG_ON(PageCompound(page)); | |
1386 | VM_BUG_ON(!page_count(page)); | |
1387 | ||
1388 | #ifdef CONFIG_KMEMCHECK | |
1389 | /* | |
1390 | * Split shadow pages too, because free(page[0]) would | |
1391 | * otherwise free the whole shadow. | |
1392 | */ | |
1393 | if (kmemcheck_page_is_tracked(page)) | |
1394 | split_page(virt_to_page(page[0].shadow), order); | |
1395 | #endif | |
1396 | ||
1397 | for (i = 1; i < (1 << order); i++) | |
1398 | set_page_refcounted(page + i); | |
1399 | } | |
1400 | ||
1401 | static int __isolate_free_page(struct page *page, unsigned int order) | |
1402 | { | |
1403 | unsigned long watermark; | |
1404 | struct zone *zone; | |
1405 | int mt; | |
1406 | ||
1407 | BUG_ON(!PageBuddy(page)); | |
1408 | ||
1409 | zone = page_zone(page); | |
1410 | mt = get_pageblock_migratetype(page); | |
1411 | ||
1412 | if (!is_migrate_isolate(mt)) { | |
1413 | /* Obey watermarks as if the page was being allocated */ | |
1414 | watermark = low_wmark_pages(zone) + (1 << order); | |
1415 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) | |
1416 | return 0; | |
1417 | ||
1418 | __mod_zone_freepage_state(zone, -(1UL << order), mt); | |
1419 | } | |
1420 | ||
1421 | /* Remove page from free list */ | |
1422 | list_del(&page->lru); | |
1423 | zone->free_area[order].nr_free--; | |
1424 | rmv_page_order(page); | |
1425 | ||
1426 | /* Set the pageblock if the isolated page is at least a pageblock */ | |
1427 | if (order >= pageblock_order - 1) { | |
1428 | struct page *endpage = page + (1 << order) - 1; | |
1429 | for (; page < endpage; page += pageblock_nr_pages) { | |
1430 | int mt = get_pageblock_migratetype(page); | |
1431 | if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) | |
1432 | set_pageblock_migratetype(page, | |
1433 | MIGRATE_MOVABLE); | |
1434 | } | |
1435 | } | |
1436 | ||
1437 | return 1UL << order; | |
1438 | } | |
1439 | ||
1440 | /* | |
1441 | * Similar to split_page except the page is already free. As this is only | |
1442 | * being used for migration, the migratetype of the block also changes. | |
1443 | * As this is called with interrupts disabled, the caller is responsible | |
1444 | * for calling arch_alloc_page() and kernel_map_page() after interrupts | |
1445 | * are enabled. | |
1446 | * | |
1447 | * Note: this is probably too low level an operation for use in drivers. | |
1448 | * Please consult with lkml before using this in your driver. | |
1449 | */ | |
1450 | int split_free_page(struct page *page) | |
1451 | { | |
1452 | unsigned int order; | |
1453 | int nr_pages; | |
1454 | ||
1455 | order = page_order(page); | |
1456 | ||
1457 | nr_pages = __isolate_free_page(page, order); | |
1458 | if (!nr_pages) | |
1459 | return 0; | |
1460 | ||
1461 | /* Split into individual pages */ | |
1462 | set_page_refcounted(page); | |
1463 | split_page(page, order); | |
1464 | return nr_pages; | |
1465 | } | |
1466 | ||
1467 | /* | |
1468 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But | |
1469 | * we cheat by calling it from here, in the order > 0 path. Saves a branch | |
1470 | * or two. | |
1471 | */ | |
1472 | static inline | |
1473 | struct page *buffered_rmqueue(struct zone *preferred_zone, | |
1474 | struct zone *zone, int order, gfp_t gfp_flags, | |
1475 | int migratetype) | |
1476 | { | |
1477 | unsigned long flags; | |
1478 | struct page *page; | |
1479 | int cold = !!(gfp_flags & __GFP_COLD); | |
1480 | ||
1481 | again: | |
1482 | if (likely(order == 0)) { | |
1483 | struct per_cpu_pages *pcp; | |
1484 | struct list_head *list; | |
1485 | ||
1486 | local_irq_save(flags); | |
1487 | pcp = &this_cpu_ptr(zone->pageset)->pcp; | |
1488 | list = &pcp->lists[migratetype]; | |
1489 | if (list_empty(list)) { | |
1490 | pcp->count += rmqueue_bulk(zone, 0, | |
1491 | pcp->batch, list, | |
1492 | migratetype, cold); | |
1493 | if (unlikely(list_empty(list))) | |
1494 | goto failed; | |
1495 | } | |
1496 | ||
1497 | if (cold) | |
1498 | page = list_entry(list->prev, struct page, lru); | |
1499 | else | |
1500 | page = list_entry(list->next, struct page, lru); | |
1501 | ||
1502 | list_del(&page->lru); | |
1503 | pcp->count--; | |
1504 | } else { | |
1505 | if (unlikely(gfp_flags & __GFP_NOFAIL)) { | |
1506 | /* | |
1507 | * __GFP_NOFAIL is not to be used in new code. | |
1508 | * | |
1509 | * All __GFP_NOFAIL callers should be fixed so that they | |
1510 | * properly detect and handle allocation failures. | |
1511 | * | |
1512 | * We most definitely don't want callers attempting to | |
1513 | * allocate greater than order-1 page units with | |
1514 | * __GFP_NOFAIL. | |
1515 | */ | |
1516 | WARN_ON_ONCE(order > 1); | |
1517 | } | |
1518 | spin_lock_irqsave(&zone->lock, flags); | |
1519 | page = __rmqueue(zone, order, migratetype); | |
1520 | spin_unlock(&zone->lock); | |
1521 | if (!page) | |
1522 | goto failed; | |
1523 | __mod_zone_freepage_state(zone, -(1 << order), | |
1524 | get_pageblock_migratetype(page)); | |
1525 | } | |
1526 | ||
1527 | __count_zone_vm_events(PGALLOC, zone, 1 << order); | |
1528 | zone_statistics(preferred_zone, zone, gfp_flags); | |
1529 | local_irq_restore(flags); | |
1530 | ||
1531 | VM_BUG_ON(bad_range(zone, page)); | |
1532 | if (prep_new_page(page, order, gfp_flags)) | |
1533 | goto again; | |
1534 | return page; | |
1535 | ||
1536 | failed: | |
1537 | local_irq_restore(flags); | |
1538 | return NULL; | |
1539 | } | |
1540 | ||
1541 | #ifdef CONFIG_FAIL_PAGE_ALLOC | |
1542 | ||
1543 | static struct { | |
1544 | struct fault_attr attr; | |
1545 | ||
1546 | u32 ignore_gfp_highmem; | |
1547 | u32 ignore_gfp_wait; | |
1548 | u32 min_order; | |
1549 | } fail_page_alloc = { | |
1550 | .attr = FAULT_ATTR_INITIALIZER, | |
1551 | .ignore_gfp_wait = 1, | |
1552 | .ignore_gfp_highmem = 1, | |
1553 | .min_order = 1, | |
1554 | }; | |
1555 | ||
1556 | static int __init setup_fail_page_alloc(char *str) | |
1557 | { | |
1558 | return setup_fault_attr(&fail_page_alloc.attr, str); | |
1559 | } | |
1560 | __setup("fail_page_alloc=", setup_fail_page_alloc); | |
1561 | ||
1562 | static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1563 | { | |
1564 | if (order < fail_page_alloc.min_order) | |
1565 | return false; | |
1566 | if (gfp_mask & __GFP_NOFAIL) | |
1567 | return false; | |
1568 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) | |
1569 | return false; | |
1570 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) | |
1571 | return false; | |
1572 | ||
1573 | return should_fail(&fail_page_alloc.attr, 1 << order); | |
1574 | } | |
1575 | ||
1576 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
1577 | ||
1578 | static int __init fail_page_alloc_debugfs(void) | |
1579 | { | |
1580 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
1581 | struct dentry *dir; | |
1582 | ||
1583 | dir = fault_create_debugfs_attr("fail_page_alloc", NULL, | |
1584 | &fail_page_alloc.attr); | |
1585 | if (IS_ERR(dir)) | |
1586 | return PTR_ERR(dir); | |
1587 | ||
1588 | if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
1589 | &fail_page_alloc.ignore_gfp_wait)) | |
1590 | goto fail; | |
1591 | if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, | |
1592 | &fail_page_alloc.ignore_gfp_highmem)) | |
1593 | goto fail; | |
1594 | if (!debugfs_create_u32("min-order", mode, dir, | |
1595 | &fail_page_alloc.min_order)) | |
1596 | goto fail; | |
1597 | ||
1598 | return 0; | |
1599 | fail: | |
1600 | debugfs_remove_recursive(dir); | |
1601 | ||
1602 | return -ENOMEM; | |
1603 | } | |
1604 | ||
1605 | late_initcall(fail_page_alloc_debugfs); | |
1606 | ||
1607 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
1608 | ||
1609 | #else /* CONFIG_FAIL_PAGE_ALLOC */ | |
1610 | ||
1611 | static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) | |
1612 | { | |
1613 | return false; | |
1614 | } | |
1615 | ||
1616 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ | |
1617 | ||
1618 | /* | |
1619 | * Return true if free pages are above 'mark'. This takes into account the order | |
1620 | * of the allocation. | |
1621 | */ | |
1622 | static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, | |
1623 | int classzone_idx, int alloc_flags, long free_pages) | |
1624 | { | |
1625 | /* free_pages my go negative - that's OK */ | |
1626 | long min = mark; | |
1627 | long lowmem_reserve = z->lowmem_reserve[classzone_idx]; | |
1628 | int o; | |
1629 | ||
1630 | free_pages -= (1 << order) - 1; | |
1631 | if (alloc_flags & ALLOC_HIGH) | |
1632 | min -= min / 2; | |
1633 | if (alloc_flags & ALLOC_HARDER) | |
1634 | min -= min / 4; | |
1635 | #ifdef CONFIG_CMA | |
1636 | /* If allocation can't use CMA areas don't use free CMA pages */ | |
1637 | if (!(alloc_flags & ALLOC_CMA)) | |
1638 | free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES); | |
1639 | #endif | |
1640 | if (free_pages <= min + lowmem_reserve) | |
1641 | return false; | |
1642 | for (o = 0; o < order; o++) { | |
1643 | /* At the next order, this order's pages become unavailable */ | |
1644 | free_pages -= z->free_area[o].nr_free << o; | |
1645 | ||
1646 | /* Require fewer higher order pages to be free */ | |
1647 | min >>= 1; | |
1648 | ||
1649 | if (free_pages <= min) | |
1650 | return false; | |
1651 | } | |
1652 | return true; | |
1653 | } | |
1654 | ||
1655 | bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, | |
1656 | int classzone_idx, int alloc_flags) | |
1657 | { | |
1658 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1659 | zone_page_state(z, NR_FREE_PAGES)); | |
1660 | } | |
1661 | ||
1662 | bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, | |
1663 | int classzone_idx, int alloc_flags) | |
1664 | { | |
1665 | long free_pages = zone_page_state(z, NR_FREE_PAGES); | |
1666 | ||
1667 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) | |
1668 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); | |
1669 | ||
1670 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, | |
1671 | free_pages); | |
1672 | } | |
1673 | ||
1674 | #ifdef CONFIG_NUMA | |
1675 | /* | |
1676 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to | |
1677 | * skip over zones that are not allowed by the cpuset, or that have | |
1678 | * been recently (in last second) found to be nearly full. See further | |
1679 | * comments in mmzone.h. Reduces cache footprint of zonelist scans | |
1680 | * that have to skip over a lot of full or unallowed zones. | |
1681 | * | |
1682 | * If the zonelist cache is present in the passed in zonelist, then | |
1683 | * returns a pointer to the allowed node mask (either the current | |
1684 | * tasks mems_allowed, or node_states[N_MEMORY].) | |
1685 | * | |
1686 | * If the zonelist cache is not available for this zonelist, does | |
1687 | * nothing and returns NULL. | |
1688 | * | |
1689 | * If the fullzones BITMAP in the zonelist cache is stale (more than | |
1690 | * a second since last zap'd) then we zap it out (clear its bits.) | |
1691 | * | |
1692 | * We hold off even calling zlc_setup, until after we've checked the | |
1693 | * first zone in the zonelist, on the theory that most allocations will | |
1694 | * be satisfied from that first zone, so best to examine that zone as | |
1695 | * quickly as we can. | |
1696 | */ | |
1697 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1698 | { | |
1699 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1700 | nodemask_t *allowednodes; /* zonelist_cache approximation */ | |
1701 | ||
1702 | zlc = zonelist->zlcache_ptr; | |
1703 | if (!zlc) | |
1704 | return NULL; | |
1705 | ||
1706 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { | |
1707 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1708 | zlc->last_full_zap = jiffies; | |
1709 | } | |
1710 | ||
1711 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? | |
1712 | &cpuset_current_mems_allowed : | |
1713 | &node_states[N_MEMORY]; | |
1714 | return allowednodes; | |
1715 | } | |
1716 | ||
1717 | /* | |
1718 | * Given 'z' scanning a zonelist, run a couple of quick checks to see | |
1719 | * if it is worth looking at further for free memory: | |
1720 | * 1) Check that the zone isn't thought to be full (doesn't have its | |
1721 | * bit set in the zonelist_cache fullzones BITMAP). | |
1722 | * 2) Check that the zones node (obtained from the zonelist_cache | |
1723 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. | |
1724 | * Return true (non-zero) if zone is worth looking at further, or | |
1725 | * else return false (zero) if it is not. | |
1726 | * | |
1727 | * This check -ignores- the distinction between various watermarks, | |
1728 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is | |
1729 | * found to be full for any variation of these watermarks, it will | |
1730 | * be considered full for up to one second by all requests, unless | |
1731 | * we are so low on memory on all allowed nodes that we are forced | |
1732 | * into the second scan of the zonelist. | |
1733 | * | |
1734 | * In the second scan we ignore this zonelist cache and exactly | |
1735 | * apply the watermarks to all zones, even it is slower to do so. | |
1736 | * We are low on memory in the second scan, and should leave no stone | |
1737 | * unturned looking for a free page. | |
1738 | */ | |
1739 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1740 | nodemask_t *allowednodes) | |
1741 | { | |
1742 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1743 | int i; /* index of *z in zonelist zones */ | |
1744 | int n; /* node that zone *z is on */ | |
1745 | ||
1746 | zlc = zonelist->zlcache_ptr; | |
1747 | if (!zlc) | |
1748 | return 1; | |
1749 | ||
1750 | i = z - zonelist->_zonerefs; | |
1751 | n = zlc->z_to_n[i]; | |
1752 | ||
1753 | /* This zone is worth trying if it is allowed but not full */ | |
1754 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); | |
1755 | } | |
1756 | ||
1757 | /* | |
1758 | * Given 'z' scanning a zonelist, set the corresponding bit in | |
1759 | * zlc->fullzones, so that subsequent attempts to allocate a page | |
1760 | * from that zone don't waste time re-examining it. | |
1761 | */ | |
1762 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1763 | { | |
1764 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1765 | int i; /* index of *z in zonelist zones */ | |
1766 | ||
1767 | zlc = zonelist->zlcache_ptr; | |
1768 | if (!zlc) | |
1769 | return; | |
1770 | ||
1771 | i = z - zonelist->_zonerefs; | |
1772 | ||
1773 | set_bit(i, zlc->fullzones); | |
1774 | } | |
1775 | ||
1776 | /* | |
1777 | * clear all zones full, called after direct reclaim makes progress so that | |
1778 | * a zone that was recently full is not skipped over for up to a second | |
1779 | */ | |
1780 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
1781 | { | |
1782 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ | |
1783 | ||
1784 | zlc = zonelist->zlcache_ptr; | |
1785 | if (!zlc) | |
1786 | return; | |
1787 | ||
1788 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
1789 | } | |
1790 | ||
1791 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) | |
1792 | { | |
1793 | return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes); | |
1794 | } | |
1795 | ||
1796 | static void __paginginit init_zone_allows_reclaim(int nid) | |
1797 | { | |
1798 | int i; | |
1799 | ||
1800 | for_each_online_node(i) | |
1801 | if (node_distance(nid, i) <= RECLAIM_DISTANCE) | |
1802 | node_set(i, NODE_DATA(nid)->reclaim_nodes); | |
1803 | else | |
1804 | zone_reclaim_mode = 1; | |
1805 | } | |
1806 | ||
1807 | #else /* CONFIG_NUMA */ | |
1808 | ||
1809 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) | |
1810 | { | |
1811 | return NULL; | |
1812 | } | |
1813 | ||
1814 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, | |
1815 | nodemask_t *allowednodes) | |
1816 | { | |
1817 | return 1; | |
1818 | } | |
1819 | ||
1820 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) | |
1821 | { | |
1822 | } | |
1823 | ||
1824 | static void zlc_clear_zones_full(struct zonelist *zonelist) | |
1825 | { | |
1826 | } | |
1827 | ||
1828 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) | |
1829 | { | |
1830 | return true; | |
1831 | } | |
1832 | ||
1833 | static inline void init_zone_allows_reclaim(int nid) | |
1834 | { | |
1835 | } | |
1836 | #endif /* CONFIG_NUMA */ | |
1837 | ||
1838 | /* | |
1839 | * get_page_from_freelist goes through the zonelist trying to allocate | |
1840 | * a page. | |
1841 | */ | |
1842 | static struct page * | |
1843 | get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, | |
1844 | struct zonelist *zonelist, int high_zoneidx, int alloc_flags, | |
1845 | struct zone *preferred_zone, int migratetype) | |
1846 | { | |
1847 | struct zoneref *z; | |
1848 | struct page *page = NULL; | |
1849 | int classzone_idx; | |
1850 | struct zone *zone; | |
1851 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ | |
1852 | int zlc_active = 0; /* set if using zonelist_cache */ | |
1853 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ | |
1854 | ||
1855 | classzone_idx = zone_idx(preferred_zone); | |
1856 | zonelist_scan: | |
1857 | /* | |
1858 | * Scan zonelist, looking for a zone with enough free. | |
1859 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
1860 | */ | |
1861 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
1862 | high_zoneidx, nodemask) { | |
1863 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && | |
1864 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
1865 | continue; | |
1866 | if ((alloc_flags & ALLOC_CPUSET) && | |
1867 | !cpuset_zone_allowed_softwall(zone, gfp_mask)) | |
1868 | continue; | |
1869 | /* | |
1870 | * When allocating a page cache page for writing, we | |
1871 | * want to get it from a zone that is within its dirty | |
1872 | * limit, such that no single zone holds more than its | |
1873 | * proportional share of globally allowed dirty pages. | |
1874 | * The dirty limits take into account the zone's | |
1875 | * lowmem reserves and high watermark so that kswapd | |
1876 | * should be able to balance it without having to | |
1877 | * write pages from its LRU list. | |
1878 | * | |
1879 | * This may look like it could increase pressure on | |
1880 | * lower zones by failing allocations in higher zones | |
1881 | * before they are full. But the pages that do spill | |
1882 | * over are limited as the lower zones are protected | |
1883 | * by this very same mechanism. It should not become | |
1884 | * a practical burden to them. | |
1885 | * | |
1886 | * XXX: For now, allow allocations to potentially | |
1887 | * exceed the per-zone dirty limit in the slowpath | |
1888 | * (ALLOC_WMARK_LOW unset) before going into reclaim, | |
1889 | * which is important when on a NUMA setup the allowed | |
1890 | * zones are together not big enough to reach the | |
1891 | * global limit. The proper fix for these situations | |
1892 | * will require awareness of zones in the | |
1893 | * dirty-throttling and the flusher threads. | |
1894 | */ | |
1895 | if ((alloc_flags & ALLOC_WMARK_LOW) && | |
1896 | (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone)) | |
1897 | goto this_zone_full; | |
1898 | ||
1899 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
1900 | if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { | |
1901 | unsigned long mark; | |
1902 | int ret; | |
1903 | ||
1904 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; | |
1905 | if (zone_watermark_ok(zone, order, mark, | |
1906 | classzone_idx, alloc_flags)) | |
1907 | goto try_this_zone; | |
1908 | ||
1909 | if (IS_ENABLED(CONFIG_NUMA) && | |
1910 | !did_zlc_setup && nr_online_nodes > 1) { | |
1911 | /* | |
1912 | * we do zlc_setup if there are multiple nodes | |
1913 | * and before considering the first zone allowed | |
1914 | * by the cpuset. | |
1915 | */ | |
1916 | allowednodes = zlc_setup(zonelist, alloc_flags); | |
1917 | zlc_active = 1; | |
1918 | did_zlc_setup = 1; | |
1919 | } | |
1920 | ||
1921 | if (zone_reclaim_mode == 0 || | |
1922 | !zone_allows_reclaim(preferred_zone, zone)) | |
1923 | goto this_zone_full; | |
1924 | ||
1925 | /* | |
1926 | * As we may have just activated ZLC, check if the first | |
1927 | * eligible zone has failed zone_reclaim recently. | |
1928 | */ | |
1929 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && | |
1930 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) | |
1931 | continue; | |
1932 | ||
1933 | ret = zone_reclaim(zone, gfp_mask, order); | |
1934 | switch (ret) { | |
1935 | case ZONE_RECLAIM_NOSCAN: | |
1936 | /* did not scan */ | |
1937 | continue; | |
1938 | case ZONE_RECLAIM_FULL: | |
1939 | /* scanned but unreclaimable */ | |
1940 | continue; | |
1941 | default: | |
1942 | /* did we reclaim enough */ | |
1943 | if (!zone_watermark_ok(zone, order, mark, | |
1944 | classzone_idx, alloc_flags)) | |
1945 | goto this_zone_full; | |
1946 | } | |
1947 | } | |
1948 | ||
1949 | try_this_zone: | |
1950 | page = buffered_rmqueue(preferred_zone, zone, order, | |
1951 | gfp_mask, migratetype); | |
1952 | if (page) | |
1953 | break; | |
1954 | this_zone_full: | |
1955 | if (IS_ENABLED(CONFIG_NUMA)) | |
1956 | zlc_mark_zone_full(zonelist, z); | |
1957 | } | |
1958 | ||
1959 | if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) { | |
1960 | /* Disable zlc cache for second zonelist scan */ | |
1961 | zlc_active = 0; | |
1962 | goto zonelist_scan; | |
1963 | } | |
1964 | ||
1965 | if (page) | |
1966 | /* | |
1967 | * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was | |
1968 | * necessary to allocate the page. The expectation is | |
1969 | * that the caller is taking steps that will free more | |
1970 | * memory. The caller should avoid the page being used | |
1971 | * for !PFMEMALLOC purposes. | |
1972 | */ | |
1973 | page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); | |
1974 | ||
1975 | return page; | |
1976 | } | |
1977 | ||
1978 | /* | |
1979 | * Large machines with many possible nodes should not always dump per-node | |
1980 | * meminfo in irq context. | |
1981 | */ | |
1982 | static inline bool should_suppress_show_mem(void) | |
1983 | { | |
1984 | bool ret = false; | |
1985 | ||
1986 | #if NODES_SHIFT > 8 | |
1987 | ret = in_interrupt(); | |
1988 | #endif | |
1989 | return ret; | |
1990 | } | |
1991 | ||
1992 | static DEFINE_RATELIMIT_STATE(nopage_rs, | |
1993 | DEFAULT_RATELIMIT_INTERVAL, | |
1994 | DEFAULT_RATELIMIT_BURST); | |
1995 | ||
1996 | void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) | |
1997 | { | |
1998 | unsigned int filter = SHOW_MEM_FILTER_NODES; | |
1999 | ||
2000 | if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || | |
2001 | debug_guardpage_minorder() > 0) | |
2002 | return; | |
2003 | ||
2004 | /* | |
2005 | * This documents exceptions given to allocations in certain | |
2006 | * contexts that are allowed to allocate outside current's set | |
2007 | * of allowed nodes. | |
2008 | */ | |
2009 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
2010 | if (test_thread_flag(TIF_MEMDIE) || | |
2011 | (current->flags & (PF_MEMALLOC | PF_EXITING))) | |
2012 | filter &= ~SHOW_MEM_FILTER_NODES; | |
2013 | if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) | |
2014 | filter &= ~SHOW_MEM_FILTER_NODES; | |
2015 | ||
2016 | if (fmt) { | |
2017 | struct va_format vaf; | |
2018 | va_list args; | |
2019 | ||
2020 | va_start(args, fmt); | |
2021 | ||
2022 | vaf.fmt = fmt; | |
2023 | vaf.va = &args; | |
2024 | ||
2025 | pr_warn("%pV", &vaf); | |
2026 | ||
2027 | va_end(args); | |
2028 | } | |
2029 | ||
2030 | pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", | |
2031 | current->comm, order, gfp_mask); | |
2032 | ||
2033 | dump_stack(); | |
2034 | if (!should_suppress_show_mem()) | |
2035 | show_mem(filter); | |
2036 | } | |
2037 | ||
2038 | static inline int | |
2039 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, | |
2040 | unsigned long did_some_progress, | |
2041 | unsigned long pages_reclaimed) | |
2042 | { | |
2043 | /* Do not loop if specifically requested */ | |
2044 | if (gfp_mask & __GFP_NORETRY) | |
2045 | return 0; | |
2046 | ||
2047 | /* Always retry if specifically requested */ | |
2048 | if (gfp_mask & __GFP_NOFAIL) | |
2049 | return 1; | |
2050 | ||
2051 | /* | |
2052 | * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim | |
2053 | * making forward progress without invoking OOM. Suspend also disables | |
2054 | * storage devices so kswapd will not help. Bail if we are suspending. | |
2055 | */ | |
2056 | if (!did_some_progress && pm_suspended_storage()) | |
2057 | return 0; | |
2058 | ||
2059 | /* | |
2060 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER | |
2061 | * means __GFP_NOFAIL, but that may not be true in other | |
2062 | * implementations. | |
2063 | */ | |
2064 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
2065 | return 1; | |
2066 | ||
2067 | /* | |
2068 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is | |
2069 | * specified, then we retry until we no longer reclaim any pages | |
2070 | * (above), or we've reclaimed an order of pages at least as | |
2071 | * large as the allocation's order. In both cases, if the | |
2072 | * allocation still fails, we stop retrying. | |
2073 | */ | |
2074 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) | |
2075 | return 1; | |
2076 | ||
2077 | return 0; | |
2078 | } | |
2079 | ||
2080 | static inline struct page * | |
2081 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
2082 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2083 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2084 | int migratetype) | |
2085 | { | |
2086 | struct page *page; | |
2087 | ||
2088 | /* Acquire the OOM killer lock for the zones in zonelist */ | |
2089 | if (!try_set_zonelist_oom(zonelist, gfp_mask)) { | |
2090 | schedule_timeout_uninterruptible(1); | |
2091 | return NULL; | |
2092 | } | |
2093 | ||
2094 | /* | |
2095 | * Go through the zonelist yet one more time, keep very high watermark | |
2096 | * here, this is only to catch a parallel oom killing, we must fail if | |
2097 | * we're still under heavy pressure. | |
2098 | */ | |
2099 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, | |
2100 | order, zonelist, high_zoneidx, | |
2101 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, | |
2102 | preferred_zone, migratetype); | |
2103 | if (page) | |
2104 | goto out; | |
2105 | ||
2106 | if (!(gfp_mask & __GFP_NOFAIL)) { | |
2107 | /* The OOM killer will not help higher order allocs */ | |
2108 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
2109 | goto out; | |
2110 | /* The OOM killer does not needlessly kill tasks for lowmem */ | |
2111 | if (high_zoneidx < ZONE_NORMAL) | |
2112 | goto out; | |
2113 | /* | |
2114 | * GFP_THISNODE contains __GFP_NORETRY and we never hit this. | |
2115 | * Sanity check for bare calls of __GFP_THISNODE, not real OOM. | |
2116 | * The caller should handle page allocation failure by itself if | |
2117 | * it specifies __GFP_THISNODE. | |
2118 | * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. | |
2119 | */ | |
2120 | if (gfp_mask & __GFP_THISNODE) | |
2121 | goto out; | |
2122 | } | |
2123 | /* Exhausted what can be done so it's blamo time */ | |
2124 | out_of_memory(zonelist, gfp_mask, order, nodemask, false); | |
2125 | ||
2126 | out: | |
2127 | clear_zonelist_oom(zonelist, gfp_mask); | |
2128 | return page; | |
2129 | } | |
2130 | ||
2131 | #ifdef CONFIG_COMPACTION | |
2132 | /* Try memory compaction for high-order allocations before reclaim */ | |
2133 | static struct page * | |
2134 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2135 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2136 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2137 | int migratetype, bool sync_migration, | |
2138 | bool *contended_compaction, bool *deferred_compaction, | |
2139 | unsigned long *did_some_progress) | |
2140 | { | |
2141 | if (!order) | |
2142 | return NULL; | |
2143 | ||
2144 | if (compaction_deferred(preferred_zone, order)) { | |
2145 | *deferred_compaction = true; | |
2146 | return NULL; | |
2147 | } | |
2148 | ||
2149 | current->flags |= PF_MEMALLOC; | |
2150 | *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, | |
2151 | nodemask, sync_migration, | |
2152 | contended_compaction); | |
2153 | current->flags &= ~PF_MEMALLOC; | |
2154 | ||
2155 | if (*did_some_progress != COMPACT_SKIPPED) { | |
2156 | struct page *page; | |
2157 | ||
2158 | /* Page migration frees to the PCP lists but we want merging */ | |
2159 | drain_pages(get_cpu()); | |
2160 | put_cpu(); | |
2161 | ||
2162 | page = get_page_from_freelist(gfp_mask, nodemask, | |
2163 | order, zonelist, high_zoneidx, | |
2164 | alloc_flags & ~ALLOC_NO_WATERMARKS, | |
2165 | preferred_zone, migratetype); | |
2166 | if (page) { | |
2167 | preferred_zone->compact_blockskip_flush = false; | |
2168 | preferred_zone->compact_considered = 0; | |
2169 | preferred_zone->compact_defer_shift = 0; | |
2170 | if (order >= preferred_zone->compact_order_failed) | |
2171 | preferred_zone->compact_order_failed = order + 1; | |
2172 | count_vm_event(COMPACTSUCCESS); | |
2173 | return page; | |
2174 | } | |
2175 | ||
2176 | /* | |
2177 | * It's bad if compaction run occurs and fails. | |
2178 | * The most likely reason is that pages exist, | |
2179 | * but not enough to satisfy watermarks. | |
2180 | */ | |
2181 | count_vm_event(COMPACTFAIL); | |
2182 | ||
2183 | /* | |
2184 | * As async compaction considers a subset of pageblocks, only | |
2185 | * defer if the failure was a sync compaction failure. | |
2186 | */ | |
2187 | if (sync_migration) | |
2188 | defer_compaction(preferred_zone, order); | |
2189 | ||
2190 | cond_resched(); | |
2191 | } | |
2192 | ||
2193 | return NULL; | |
2194 | } | |
2195 | #else | |
2196 | static inline struct page * | |
2197 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
2198 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2199 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2200 | int migratetype, bool sync_migration, | |
2201 | bool *contended_compaction, bool *deferred_compaction, | |
2202 | unsigned long *did_some_progress) | |
2203 | { | |
2204 | return NULL; | |
2205 | } | |
2206 | #endif /* CONFIG_COMPACTION */ | |
2207 | ||
2208 | /* Perform direct synchronous page reclaim */ | |
2209 | static int | |
2210 | __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist, | |
2211 | nodemask_t *nodemask) | |
2212 | { | |
2213 | struct reclaim_state reclaim_state; | |
2214 | int progress; | |
2215 | ||
2216 | cond_resched(); | |
2217 | ||
2218 | /* We now go into synchronous reclaim */ | |
2219 | cpuset_memory_pressure_bump(); | |
2220 | current->flags |= PF_MEMALLOC; | |
2221 | lockdep_set_current_reclaim_state(gfp_mask); | |
2222 | reclaim_state.reclaimed_slab = 0; | |
2223 | current->reclaim_state = &reclaim_state; | |
2224 | ||
2225 | progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); | |
2226 | ||
2227 | current->reclaim_state = NULL; | |
2228 | lockdep_clear_current_reclaim_state(); | |
2229 | current->flags &= ~PF_MEMALLOC; | |
2230 | ||
2231 | cond_resched(); | |
2232 | ||
2233 | return progress; | |
2234 | } | |
2235 | ||
2236 | /* The really slow allocator path where we enter direct reclaim */ | |
2237 | static inline struct page * | |
2238 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
2239 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2240 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, | |
2241 | int migratetype, unsigned long *did_some_progress) | |
2242 | { | |
2243 | struct page *page = NULL; | |
2244 | bool drained = false; | |
2245 | ||
2246 | *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist, | |
2247 | nodemask); | |
2248 | if (unlikely(!(*did_some_progress))) | |
2249 | return NULL; | |
2250 | ||
2251 | /* After successful reclaim, reconsider all zones for allocation */ | |
2252 | if (IS_ENABLED(CONFIG_NUMA)) | |
2253 | zlc_clear_zones_full(zonelist); | |
2254 | ||
2255 | retry: | |
2256 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
2257 | zonelist, high_zoneidx, | |
2258 | alloc_flags & ~ALLOC_NO_WATERMARKS, | |
2259 | preferred_zone, migratetype); | |
2260 | ||
2261 | /* | |
2262 | * If an allocation failed after direct reclaim, it could be because | |
2263 | * pages are pinned on the per-cpu lists. Drain them and try again | |
2264 | */ | |
2265 | if (!page && !drained) { | |
2266 | drain_all_pages(); | |
2267 | drained = true; | |
2268 | goto retry; | |
2269 | } | |
2270 | ||
2271 | return page; | |
2272 | } | |
2273 | ||
2274 | /* | |
2275 | * This is called in the allocator slow-path if the allocation request is of | |
2276 | * sufficient urgency to ignore watermarks and take other desperate measures | |
2277 | */ | |
2278 | static inline struct page * | |
2279 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, | |
2280 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2281 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2282 | int migratetype) | |
2283 | { | |
2284 | struct page *page; | |
2285 | ||
2286 | do { | |
2287 | page = get_page_from_freelist(gfp_mask, nodemask, order, | |
2288 | zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, | |
2289 | preferred_zone, migratetype); | |
2290 | ||
2291 | if (!page && gfp_mask & __GFP_NOFAIL) | |
2292 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); | |
2293 | } while (!page && (gfp_mask & __GFP_NOFAIL)); | |
2294 | ||
2295 | return page; | |
2296 | } | |
2297 | ||
2298 | static inline | |
2299 | void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, | |
2300 | enum zone_type high_zoneidx, | |
2301 | enum zone_type classzone_idx) | |
2302 | { | |
2303 | struct zoneref *z; | |
2304 | struct zone *zone; | |
2305 | ||
2306 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) | |
2307 | wakeup_kswapd(zone, order, classzone_idx); | |
2308 | } | |
2309 | ||
2310 | static inline int | |
2311 | gfp_to_alloc_flags(gfp_t gfp_mask) | |
2312 | { | |
2313 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; | |
2314 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
2315 | ||
2316 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ | |
2317 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); | |
2318 | ||
2319 | /* | |
2320 | * The caller may dip into page reserves a bit more if the caller | |
2321 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
2322 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
2323 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). | |
2324 | */ | |
2325 | alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); | |
2326 | ||
2327 | if (!wait) { | |
2328 | /* | |
2329 | * Not worth trying to allocate harder for | |
2330 | * __GFP_NOMEMALLOC even if it can't schedule. | |
2331 | */ | |
2332 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
2333 | alloc_flags |= ALLOC_HARDER; | |
2334 | /* | |
2335 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. | |
2336 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. | |
2337 | */ | |
2338 | alloc_flags &= ~ALLOC_CPUSET; | |
2339 | } else if (unlikely(rt_task(current)) && !in_interrupt()) | |
2340 | alloc_flags |= ALLOC_HARDER; | |
2341 | ||
2342 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { | |
2343 | if (gfp_mask & __GFP_MEMALLOC) | |
2344 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2345 | else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) | |
2346 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2347 | else if (!in_interrupt() && | |
2348 | ((current->flags & PF_MEMALLOC) || | |
2349 | unlikely(test_thread_flag(TIF_MEMDIE)))) | |
2350 | alloc_flags |= ALLOC_NO_WATERMARKS; | |
2351 | } | |
2352 | #ifdef CONFIG_CMA | |
2353 | if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) | |
2354 | alloc_flags |= ALLOC_CMA; | |
2355 | #endif | |
2356 | return alloc_flags; | |
2357 | } | |
2358 | ||
2359 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) | |
2360 | { | |
2361 | return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); | |
2362 | } | |
2363 | ||
2364 | static inline struct page * | |
2365 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
2366 | struct zonelist *zonelist, enum zone_type high_zoneidx, | |
2367 | nodemask_t *nodemask, struct zone *preferred_zone, | |
2368 | int migratetype) | |
2369 | { | |
2370 | const gfp_t wait = gfp_mask & __GFP_WAIT; | |
2371 | struct page *page = NULL; | |
2372 | int alloc_flags; | |
2373 | unsigned long pages_reclaimed = 0; | |
2374 | unsigned long did_some_progress; | |
2375 | bool sync_migration = false; | |
2376 | bool deferred_compaction = false; | |
2377 | bool contended_compaction = false; | |
2378 | ||
2379 | /* | |
2380 | * In the slowpath, we sanity check order to avoid ever trying to | |
2381 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may | |
2382 | * be using allocators in order of preference for an area that is | |
2383 | * too large. | |
2384 | */ | |
2385 | if (order >= MAX_ORDER) { | |
2386 | WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); | |
2387 | return NULL; | |
2388 | } | |
2389 | ||
2390 | /* | |
2391 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and | |
2392 | * __GFP_NOWARN set) should not cause reclaim since the subsystem | |
2393 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim | |
2394 | * using a larger set of nodes after it has established that the | |
2395 | * allowed per node queues are empty and that nodes are | |
2396 | * over allocated. | |
2397 | */ | |
2398 | if (IS_ENABLED(CONFIG_NUMA) && | |
2399 | (gfp_mask & GFP_THISNODE) == GFP_THISNODE) | |
2400 | goto nopage; | |
2401 | ||
2402 | restart: | |
2403 | if (!(gfp_mask & __GFP_NO_KSWAPD)) | |
2404 | wake_all_kswapd(order, zonelist, high_zoneidx, | |
2405 | zone_idx(preferred_zone)); | |
2406 | ||
2407 | /* | |
2408 | * OK, we're below the kswapd watermark and have kicked background | |
2409 | * reclaim. Now things get more complex, so set up alloc_flags according | |
2410 | * to how we want to proceed. | |
2411 | */ | |
2412 | alloc_flags = gfp_to_alloc_flags(gfp_mask); | |
2413 | ||
2414 | /* | |
2415 | * Find the true preferred zone if the allocation is unconstrained by | |
2416 | * cpusets. | |
2417 | */ | |
2418 | if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) | |
2419 | first_zones_zonelist(zonelist, high_zoneidx, NULL, | |
2420 | &preferred_zone); | |
2421 | ||
2422 | rebalance: | |
2423 | /* This is the last chance, in general, before the goto nopage. */ | |
2424 | page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, | |
2425 | high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, | |
2426 | preferred_zone, migratetype); | |
2427 | if (page) | |
2428 | goto got_pg; | |
2429 | ||
2430 | /* Allocate without watermarks if the context allows */ | |
2431 | if (alloc_flags & ALLOC_NO_WATERMARKS) { | |
2432 | /* | |
2433 | * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds | |
2434 | * the allocation is high priority and these type of | |
2435 | * allocations are system rather than user orientated | |
2436 | */ | |
2437 | zonelist = node_zonelist(numa_node_id(), gfp_mask); | |
2438 | ||
2439 | page = __alloc_pages_high_priority(gfp_mask, order, | |
2440 | zonelist, high_zoneidx, nodemask, | |
2441 | preferred_zone, migratetype); | |
2442 | if (page) { | |
2443 | goto got_pg; | |
2444 | } | |
2445 | } | |
2446 | ||
2447 | /* Atomic allocations - we can't balance anything */ | |
2448 | if (!wait) | |
2449 | goto nopage; | |
2450 | ||
2451 | /* Avoid recursion of direct reclaim */ | |
2452 | if (current->flags & PF_MEMALLOC) | |
2453 | goto nopage; | |
2454 | ||
2455 | /* Avoid allocations with no watermarks from looping endlessly */ | |
2456 | if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) | |
2457 | goto nopage; | |
2458 | ||
2459 | /* | |
2460 | * Try direct compaction. The first pass is asynchronous. Subsequent | |
2461 | * attempts after direct reclaim are synchronous | |
2462 | */ | |
2463 | page = __alloc_pages_direct_compact(gfp_mask, order, | |
2464 | zonelist, high_zoneidx, | |
2465 | nodemask, | |
2466 | alloc_flags, preferred_zone, | |
2467 | migratetype, sync_migration, | |
2468 | &contended_compaction, | |
2469 | &deferred_compaction, | |
2470 | &did_some_progress); | |
2471 | if (page) | |
2472 | goto got_pg; | |
2473 | sync_migration = true; | |
2474 | ||
2475 | /* | |
2476 | * If compaction is deferred for high-order allocations, it is because | |
2477 | * sync compaction recently failed. In this is the case and the caller | |
2478 | * requested a movable allocation that does not heavily disrupt the | |
2479 | * system then fail the allocation instead of entering direct reclaim. | |
2480 | */ | |
2481 | if ((deferred_compaction || contended_compaction) && | |
2482 | (gfp_mask & __GFP_NO_KSWAPD)) | |
2483 | goto nopage; | |
2484 | ||
2485 | /* Try direct reclaim and then allocating */ | |
2486 | page = __alloc_pages_direct_reclaim(gfp_mask, order, | |
2487 | zonelist, high_zoneidx, | |
2488 | nodemask, | |
2489 | alloc_flags, preferred_zone, | |
2490 | migratetype, &did_some_progress); | |
2491 | if (page) | |
2492 | goto got_pg; | |
2493 | ||
2494 | /* | |
2495 | * If we failed to make any progress reclaiming, then we are | |
2496 | * running out of options and have to consider going OOM | |
2497 | */ | |
2498 | if (!did_some_progress) { | |
2499 | if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { | |
2500 | if (oom_killer_disabled) | |
2501 | goto nopage; | |
2502 | /* Coredumps can quickly deplete all memory reserves */ | |
2503 | if ((current->flags & PF_DUMPCORE) && | |
2504 | !(gfp_mask & __GFP_NOFAIL)) | |
2505 | goto nopage; | |
2506 | page = __alloc_pages_may_oom(gfp_mask, order, | |
2507 | zonelist, high_zoneidx, | |
2508 | nodemask, preferred_zone, | |
2509 | migratetype); | |
2510 | if (page) | |
2511 | goto got_pg; | |
2512 | ||
2513 | if (!(gfp_mask & __GFP_NOFAIL)) { | |
2514 | /* | |
2515 | * The oom killer is not called for high-order | |
2516 | * allocations that may fail, so if no progress | |
2517 | * is being made, there are no other options and | |
2518 | * retrying is unlikely to help. | |
2519 | */ | |
2520 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
2521 | goto nopage; | |
2522 | /* | |
2523 | * The oom killer is not called for lowmem | |
2524 | * allocations to prevent needlessly killing | |
2525 | * innocent tasks. | |
2526 | */ | |
2527 | if (high_zoneidx < ZONE_NORMAL) | |
2528 | goto nopage; | |
2529 | } | |
2530 | ||
2531 | goto restart; | |
2532 | } | |
2533 | } | |
2534 | ||
2535 | /* Check if we should retry the allocation */ | |
2536 | pages_reclaimed += did_some_progress; | |
2537 | if (should_alloc_retry(gfp_mask, order, did_some_progress, | |
2538 | pages_reclaimed)) { | |
2539 | /* Wait for some write requests to complete then retry */ | |
2540 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); | |
2541 | goto rebalance; | |
2542 | } else { | |
2543 | /* | |
2544 | * High-order allocations do not necessarily loop after | |
2545 | * direct reclaim and reclaim/compaction depends on compaction | |
2546 | * being called after reclaim so call directly if necessary | |
2547 | */ | |
2548 | page = __alloc_pages_direct_compact(gfp_mask, order, | |
2549 | zonelist, high_zoneidx, | |
2550 | nodemask, | |
2551 | alloc_flags, preferred_zone, | |
2552 | migratetype, sync_migration, | |
2553 | &contended_compaction, | |
2554 | &deferred_compaction, | |
2555 | &did_some_progress); | |
2556 | if (page) | |
2557 | goto got_pg; | |
2558 | } | |
2559 | ||
2560 | nopage: | |
2561 | warn_alloc_failed(gfp_mask, order, NULL); | |
2562 | return page; | |
2563 | got_pg: | |
2564 | if (kmemcheck_enabled) | |
2565 | kmemcheck_pagealloc_alloc(page, order, gfp_mask); | |
2566 | ||
2567 | return page; | |
2568 | } | |
2569 | ||
2570 | /* | |
2571 | * This is the 'heart' of the zoned buddy allocator. | |
2572 | */ | |
2573 | struct page * | |
2574 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, | |
2575 | struct zonelist *zonelist, nodemask_t *nodemask) | |
2576 | { | |
2577 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); | |
2578 | struct zone *preferred_zone; | |
2579 | struct page *page = NULL; | |
2580 | int migratetype = allocflags_to_migratetype(gfp_mask); | |
2581 | unsigned int cpuset_mems_cookie; | |
2582 | int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET; | |
2583 | struct mem_cgroup *memcg = NULL; | |
2584 | ||
2585 | gfp_mask &= gfp_allowed_mask; | |
2586 | ||
2587 | lockdep_trace_alloc(gfp_mask); | |
2588 | ||
2589 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2590 | ||
2591 | if (should_fail_alloc_page(gfp_mask, order)) | |
2592 | return NULL; | |
2593 | ||
2594 | /* | |
2595 | * Check the zones suitable for the gfp_mask contain at least one | |
2596 | * valid zone. It's possible to have an empty zonelist as a result | |
2597 | * of GFP_THISNODE and a memoryless node | |
2598 | */ | |
2599 | if (unlikely(!zonelist->_zonerefs->zone)) | |
2600 | return NULL; | |
2601 | ||
2602 | /* | |
2603 | * Will only have any effect when __GFP_KMEMCG is set. This is | |
2604 | * verified in the (always inline) callee | |
2605 | */ | |
2606 | if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) | |
2607 | return NULL; | |
2608 | ||
2609 | retry_cpuset: | |
2610 | cpuset_mems_cookie = get_mems_allowed(); | |
2611 | ||
2612 | /* The preferred zone is used for statistics later */ | |
2613 | first_zones_zonelist(zonelist, high_zoneidx, | |
2614 | nodemask ? : &cpuset_current_mems_allowed, | |
2615 | &preferred_zone); | |
2616 | if (!preferred_zone) | |
2617 | goto out; | |
2618 | ||
2619 | #ifdef CONFIG_CMA | |
2620 | if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) | |
2621 | alloc_flags |= ALLOC_CMA; | |
2622 | #endif | |
2623 | /* First allocation attempt */ | |
2624 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, | |
2625 | zonelist, high_zoneidx, alloc_flags, | |
2626 | preferred_zone, migratetype); | |
2627 | if (unlikely(!page)) { | |
2628 | /* | |
2629 | * Runtime PM, block IO and its error handling path | |
2630 | * can deadlock because I/O on the device might not | |
2631 | * complete. | |
2632 | */ | |
2633 | gfp_mask = memalloc_noio_flags(gfp_mask); | |
2634 | page = __alloc_pages_slowpath(gfp_mask, order, | |
2635 | zonelist, high_zoneidx, nodemask, | |
2636 | preferred_zone, migratetype); | |
2637 | } | |
2638 | ||
2639 | trace_mm_page_alloc(page, order, gfp_mask, migratetype); | |
2640 | ||
2641 | out: | |
2642 | /* | |
2643 | * When updating a task's mems_allowed, it is possible to race with | |
2644 | * parallel threads in such a way that an allocation can fail while | |
2645 | * the mask is being updated. If a page allocation is about to fail, | |
2646 | * check if the cpuset changed during allocation and if so, retry. | |
2647 | */ | |
2648 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page)) | |
2649 | goto retry_cpuset; | |
2650 | ||
2651 | memcg_kmem_commit_charge(page, memcg, order); | |
2652 | ||
2653 | return page; | |
2654 | } | |
2655 | EXPORT_SYMBOL(__alloc_pages_nodemask); | |
2656 | ||
2657 | /* | |
2658 | * Common helper functions. | |
2659 | */ | |
2660 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) | |
2661 | { | |
2662 | struct page *page; | |
2663 | ||
2664 | /* | |
2665 | * __get_free_pages() returns a 32-bit address, which cannot represent | |
2666 | * a highmem page | |
2667 | */ | |
2668 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); | |
2669 | ||
2670 | page = alloc_pages(gfp_mask, order); | |
2671 | if (!page) | |
2672 | return 0; | |
2673 | return (unsigned long) page_address(page); | |
2674 | } | |
2675 | EXPORT_SYMBOL(__get_free_pages); | |
2676 | ||
2677 | unsigned long get_zeroed_page(gfp_t gfp_mask) | |
2678 | { | |
2679 | return __get_free_pages(gfp_mask | __GFP_ZERO, 0); | |
2680 | } | |
2681 | EXPORT_SYMBOL(get_zeroed_page); | |
2682 | ||
2683 | void __free_pages(struct page *page, unsigned int order) | |
2684 | { | |
2685 | if (put_page_testzero(page)) { | |
2686 | if (order == 0) | |
2687 | free_hot_cold_page(page, 0); | |
2688 | else | |
2689 | __free_pages_ok(page, order); | |
2690 | } | |
2691 | } | |
2692 | ||
2693 | EXPORT_SYMBOL(__free_pages); | |
2694 | ||
2695 | void free_pages(unsigned long addr, unsigned int order) | |
2696 | { | |
2697 | if (addr != 0) { | |
2698 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
2699 | __free_pages(virt_to_page((void *)addr), order); | |
2700 | } | |
2701 | } | |
2702 | ||
2703 | EXPORT_SYMBOL(free_pages); | |
2704 | ||
2705 | /* | |
2706 | * __free_memcg_kmem_pages and free_memcg_kmem_pages will free | |
2707 | * pages allocated with __GFP_KMEMCG. | |
2708 | * | |
2709 | * Those pages are accounted to a particular memcg, embedded in the | |
2710 | * corresponding page_cgroup. To avoid adding a hit in the allocator to search | |
2711 | * for that information only to find out that it is NULL for users who have no | |
2712 | * interest in that whatsoever, we provide these functions. | |
2713 | * | |
2714 | * The caller knows better which flags it relies on. | |
2715 | */ | |
2716 | void __free_memcg_kmem_pages(struct page *page, unsigned int order) | |
2717 | { | |
2718 | memcg_kmem_uncharge_pages(page, order); | |
2719 | __free_pages(page, order); | |
2720 | } | |
2721 | ||
2722 | void free_memcg_kmem_pages(unsigned long addr, unsigned int order) | |
2723 | { | |
2724 | if (addr != 0) { | |
2725 | VM_BUG_ON(!virt_addr_valid((void *)addr)); | |
2726 | __free_memcg_kmem_pages(virt_to_page((void *)addr), order); | |
2727 | } | |
2728 | } | |
2729 | ||
2730 | static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) | |
2731 | { | |
2732 | if (addr) { | |
2733 | unsigned long alloc_end = addr + (PAGE_SIZE << order); | |
2734 | unsigned long used = addr + PAGE_ALIGN(size); | |
2735 | ||
2736 | split_page(virt_to_page((void *)addr), order); | |
2737 | while (used < alloc_end) { | |
2738 | free_page(used); | |
2739 | used += PAGE_SIZE; | |
2740 | } | |
2741 | } | |
2742 | return (void *)addr; | |
2743 | } | |
2744 | ||
2745 | /** | |
2746 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
2747 | * @size: the number of bytes to allocate | |
2748 | * @gfp_mask: GFP flags for the allocation | |
2749 | * | |
2750 | * This function is similar to alloc_pages(), except that it allocates the | |
2751 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
2752 | * allocate memory in power-of-two pages. | |
2753 | * | |
2754 | * This function is also limited by MAX_ORDER. | |
2755 | * | |
2756 | * Memory allocated by this function must be released by free_pages_exact(). | |
2757 | */ | |
2758 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
2759 | { | |
2760 | unsigned int order = get_order(size); | |
2761 | unsigned long addr; | |
2762 | ||
2763 | addr = __get_free_pages(gfp_mask, order); | |
2764 | return make_alloc_exact(addr, order, size); | |
2765 | } | |
2766 | EXPORT_SYMBOL(alloc_pages_exact); | |
2767 | ||
2768 | /** | |
2769 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous | |
2770 | * pages on a node. | |
2771 | * @nid: the preferred node ID where memory should be allocated | |
2772 | * @size: the number of bytes to allocate | |
2773 | * @gfp_mask: GFP flags for the allocation | |
2774 | * | |
2775 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling | |
2776 | * back. | |
2777 | * Note this is not alloc_pages_exact_node() which allocates on a specific node, | |
2778 | * but is not exact. | |
2779 | */ | |
2780 | void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) | |
2781 | { | |
2782 | unsigned order = get_order(size); | |
2783 | struct page *p = alloc_pages_node(nid, gfp_mask, order); | |
2784 | if (!p) | |
2785 | return NULL; | |
2786 | return make_alloc_exact((unsigned long)page_address(p), order, size); | |
2787 | } | |
2788 | EXPORT_SYMBOL(alloc_pages_exact_nid); | |
2789 | ||
2790 | /** | |
2791 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
2792 | * @virt: the value returned by alloc_pages_exact. | |
2793 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
2794 | * | |
2795 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
2796 | */ | |
2797 | void free_pages_exact(void *virt, size_t size) | |
2798 | { | |
2799 | unsigned long addr = (unsigned long)virt; | |
2800 | unsigned long end = addr + PAGE_ALIGN(size); | |
2801 | ||
2802 | while (addr < end) { | |
2803 | free_page(addr); | |
2804 | addr += PAGE_SIZE; | |
2805 | } | |
2806 | } | |
2807 | EXPORT_SYMBOL(free_pages_exact); | |
2808 | ||
2809 | static unsigned int nr_free_zone_pages(int offset) | |
2810 | { | |
2811 | struct zoneref *z; | |
2812 | struct zone *zone; | |
2813 | ||
2814 | /* Just pick one node, since fallback list is circular */ | |
2815 | unsigned int sum = 0; | |
2816 | ||
2817 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); | |
2818 | ||
2819 | for_each_zone_zonelist(zone, z, zonelist, offset) { | |
2820 | unsigned long size = zone->managed_pages; | |
2821 | unsigned long high = high_wmark_pages(zone); | |
2822 | if (size > high) | |
2823 | sum += size - high; | |
2824 | } | |
2825 | ||
2826 | return sum; | |
2827 | } | |
2828 | ||
2829 | /* | |
2830 | * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL | |
2831 | */ | |
2832 | unsigned int nr_free_buffer_pages(void) | |
2833 | { | |
2834 | return nr_free_zone_pages(gfp_zone(GFP_USER)); | |
2835 | } | |
2836 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); | |
2837 | ||
2838 | /* | |
2839 | * Amount of free RAM allocatable within all zones | |
2840 | */ | |
2841 | unsigned int nr_free_pagecache_pages(void) | |
2842 | { | |
2843 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
2844 | } | |
2845 | ||
2846 | static inline void show_node(struct zone *zone) | |
2847 | { | |
2848 | if (IS_ENABLED(CONFIG_NUMA)) | |
2849 | printk("Node %d ", zone_to_nid(zone)); | |
2850 | } | |
2851 | ||
2852 | void si_meminfo(struct sysinfo *val) | |
2853 | { | |
2854 | val->totalram = totalram_pages; | |
2855 | val->sharedram = 0; | |
2856 | val->freeram = global_page_state(NR_FREE_PAGES); | |
2857 | val->bufferram = nr_blockdev_pages(); | |
2858 | val->totalhigh = totalhigh_pages; | |
2859 | val->freehigh = nr_free_highpages(); | |
2860 | val->mem_unit = PAGE_SIZE; | |
2861 | } | |
2862 | ||
2863 | EXPORT_SYMBOL(si_meminfo); | |
2864 | ||
2865 | #ifdef CONFIG_NUMA | |
2866 | void si_meminfo_node(struct sysinfo *val, int nid) | |
2867 | { | |
2868 | pg_data_t *pgdat = NODE_DATA(nid); | |
2869 | ||
2870 | val->totalram = pgdat->node_present_pages; | |
2871 | val->freeram = node_page_state(nid, NR_FREE_PAGES); | |
2872 | #ifdef CONFIG_HIGHMEM | |
2873 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; | |
2874 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], | |
2875 | NR_FREE_PAGES); | |
2876 | #else | |
2877 | val->totalhigh = 0; | |
2878 | val->freehigh = 0; | |
2879 | #endif | |
2880 | val->mem_unit = PAGE_SIZE; | |
2881 | } | |
2882 | #endif | |
2883 | ||
2884 | /* | |
2885 | * Determine whether the node should be displayed or not, depending on whether | |
2886 | * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). | |
2887 | */ | |
2888 | bool skip_free_areas_node(unsigned int flags, int nid) | |
2889 | { | |
2890 | bool ret = false; | |
2891 | unsigned int cpuset_mems_cookie; | |
2892 | ||
2893 | if (!(flags & SHOW_MEM_FILTER_NODES)) | |
2894 | goto out; | |
2895 | ||
2896 | do { | |
2897 | cpuset_mems_cookie = get_mems_allowed(); | |
2898 | ret = !node_isset(nid, cpuset_current_mems_allowed); | |
2899 | } while (!put_mems_allowed(cpuset_mems_cookie)); | |
2900 | out: | |
2901 | return ret; | |
2902 | } | |
2903 | ||
2904 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
2905 | ||
2906 | static void show_migration_types(unsigned char type) | |
2907 | { | |
2908 | static const char types[MIGRATE_TYPES] = { | |
2909 | [MIGRATE_UNMOVABLE] = 'U', | |
2910 | [MIGRATE_RECLAIMABLE] = 'E', | |
2911 | [MIGRATE_MOVABLE] = 'M', | |
2912 | [MIGRATE_RESERVE] = 'R', | |
2913 | #ifdef CONFIG_CMA | |
2914 | [MIGRATE_CMA] = 'C', | |
2915 | #endif | |
2916 | #ifdef CONFIG_MEMORY_ISOLATION | |
2917 | [MIGRATE_ISOLATE] = 'I', | |
2918 | #endif | |
2919 | }; | |
2920 | char tmp[MIGRATE_TYPES + 1]; | |
2921 | char *p = tmp; | |
2922 | int i; | |
2923 | ||
2924 | for (i = 0; i < MIGRATE_TYPES; i++) { | |
2925 | if (type & (1 << i)) | |
2926 | *p++ = types[i]; | |
2927 | } | |
2928 | ||
2929 | *p = '\0'; | |
2930 | printk("(%s) ", tmp); | |
2931 | } | |
2932 | ||
2933 | /* | |
2934 | * Show free area list (used inside shift_scroll-lock stuff) | |
2935 | * We also calculate the percentage fragmentation. We do this by counting the | |
2936 | * memory on each free list with the exception of the first item on the list. | |
2937 | * Suppresses nodes that are not allowed by current's cpuset if | |
2938 | * SHOW_MEM_FILTER_NODES is passed. | |
2939 | */ | |
2940 | void show_free_areas(unsigned int filter) | |
2941 | { | |
2942 | int cpu; | |
2943 | struct zone *zone; | |
2944 | ||
2945 | for_each_populated_zone(zone) { | |
2946 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
2947 | continue; | |
2948 | show_node(zone); | |
2949 | printk("%s per-cpu:\n", zone->name); | |
2950 | ||
2951 | for_each_online_cpu(cpu) { | |
2952 | struct per_cpu_pageset *pageset; | |
2953 | ||
2954 | pageset = per_cpu_ptr(zone->pageset, cpu); | |
2955 | ||
2956 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", | |
2957 | cpu, pageset->pcp.high, | |
2958 | pageset->pcp.batch, pageset->pcp.count); | |
2959 | } | |
2960 | } | |
2961 | ||
2962 | printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" | |
2963 | " active_file:%lu inactive_file:%lu isolated_file:%lu\n" | |
2964 | " unevictable:%lu" | |
2965 | " dirty:%lu writeback:%lu unstable:%lu\n" | |
2966 | " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" | |
2967 | " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" | |
2968 | " free_cma:%lu\n", | |
2969 | global_page_state(NR_ACTIVE_ANON), | |
2970 | global_page_state(NR_INACTIVE_ANON), | |
2971 | global_page_state(NR_ISOLATED_ANON), | |
2972 | global_page_state(NR_ACTIVE_FILE), | |
2973 | global_page_state(NR_INACTIVE_FILE), | |
2974 | global_page_state(NR_ISOLATED_FILE), | |
2975 | global_page_state(NR_UNEVICTABLE), | |
2976 | global_page_state(NR_FILE_DIRTY), | |
2977 | global_page_state(NR_WRITEBACK), | |
2978 | global_page_state(NR_UNSTABLE_NFS), | |
2979 | global_page_state(NR_FREE_PAGES), | |
2980 | global_page_state(NR_SLAB_RECLAIMABLE), | |
2981 | global_page_state(NR_SLAB_UNRECLAIMABLE), | |
2982 | global_page_state(NR_FILE_MAPPED), | |
2983 | global_page_state(NR_SHMEM), | |
2984 | global_page_state(NR_PAGETABLE), | |
2985 | global_page_state(NR_BOUNCE), | |
2986 | global_page_state(NR_FREE_CMA_PAGES)); | |
2987 | ||
2988 | for_each_populated_zone(zone) { | |
2989 | int i; | |
2990 | ||
2991 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
2992 | continue; | |
2993 | show_node(zone); | |
2994 | printk("%s" | |
2995 | " free:%lukB" | |
2996 | " min:%lukB" | |
2997 | " low:%lukB" | |
2998 | " high:%lukB" | |
2999 | " active_anon:%lukB" | |
3000 | " inactive_anon:%lukB" | |
3001 | " active_file:%lukB" | |
3002 | " inactive_file:%lukB" | |
3003 | " unevictable:%lukB" | |
3004 | " isolated(anon):%lukB" | |
3005 | " isolated(file):%lukB" | |
3006 | " present:%lukB" | |
3007 | " managed:%lukB" | |
3008 | " mlocked:%lukB" | |
3009 | " dirty:%lukB" | |
3010 | " writeback:%lukB" | |
3011 | " mapped:%lukB" | |
3012 | " shmem:%lukB" | |
3013 | " slab_reclaimable:%lukB" | |
3014 | " slab_unreclaimable:%lukB" | |
3015 | " kernel_stack:%lukB" | |
3016 | " pagetables:%lukB" | |
3017 | " unstable:%lukB" | |
3018 | " bounce:%lukB" | |
3019 | " free_cma:%lukB" | |
3020 | " writeback_tmp:%lukB" | |
3021 | " pages_scanned:%lu" | |
3022 | " all_unreclaimable? %s" | |
3023 | "\n", | |
3024 | zone->name, | |
3025 | K(zone_page_state(zone, NR_FREE_PAGES)), | |
3026 | K(min_wmark_pages(zone)), | |
3027 | K(low_wmark_pages(zone)), | |
3028 | K(high_wmark_pages(zone)), | |
3029 | K(zone_page_state(zone, NR_ACTIVE_ANON)), | |
3030 | K(zone_page_state(zone, NR_INACTIVE_ANON)), | |
3031 | K(zone_page_state(zone, NR_ACTIVE_FILE)), | |
3032 | K(zone_page_state(zone, NR_INACTIVE_FILE)), | |
3033 | K(zone_page_state(zone, NR_UNEVICTABLE)), | |
3034 | K(zone_page_state(zone, NR_ISOLATED_ANON)), | |
3035 | K(zone_page_state(zone, NR_ISOLATED_FILE)), | |
3036 | K(zone->present_pages), | |
3037 | K(zone->managed_pages), | |
3038 | K(zone_page_state(zone, NR_MLOCK)), | |
3039 | K(zone_page_state(zone, NR_FILE_DIRTY)), | |
3040 | K(zone_page_state(zone, NR_WRITEBACK)), | |
3041 | K(zone_page_state(zone, NR_FILE_MAPPED)), | |
3042 | K(zone_page_state(zone, NR_SHMEM)), | |
3043 | K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), | |
3044 | K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), | |
3045 | zone_page_state(zone, NR_KERNEL_STACK) * | |
3046 | THREAD_SIZE / 1024, | |
3047 | K(zone_page_state(zone, NR_PAGETABLE)), | |
3048 | K(zone_page_state(zone, NR_UNSTABLE_NFS)), | |
3049 | K(zone_page_state(zone, NR_BOUNCE)), | |
3050 | K(zone_page_state(zone, NR_FREE_CMA_PAGES)), | |
3051 | K(zone_page_state(zone, NR_WRITEBACK_TEMP)), | |
3052 | zone->pages_scanned, | |
3053 | (zone->all_unreclaimable ? "yes" : "no") | |
3054 | ); | |
3055 | printk("lowmem_reserve[]:"); | |
3056 | for (i = 0; i < MAX_NR_ZONES; i++) | |
3057 | printk(" %lu", zone->lowmem_reserve[i]); | |
3058 | printk("\n"); | |
3059 | } | |
3060 | ||
3061 | for_each_populated_zone(zone) { | |
3062 | unsigned long nr[MAX_ORDER], flags, order, total = 0; | |
3063 | unsigned char types[MAX_ORDER]; | |
3064 | ||
3065 | if (skip_free_areas_node(filter, zone_to_nid(zone))) | |
3066 | continue; | |
3067 | show_node(zone); | |
3068 | printk("%s: ", zone->name); | |
3069 | ||
3070 | spin_lock_irqsave(&zone->lock, flags); | |
3071 | for (order = 0; order < MAX_ORDER; order++) { | |
3072 | struct free_area *area = &zone->free_area[order]; | |
3073 | int type; | |
3074 | ||
3075 | nr[order] = area->nr_free; | |
3076 | total += nr[order] << order; | |
3077 | ||
3078 | types[order] = 0; | |
3079 | for (type = 0; type < MIGRATE_TYPES; type++) { | |
3080 | if (!list_empty(&area->free_list[type])) | |
3081 | types[order] |= 1 << type; | |
3082 | } | |
3083 | } | |
3084 | spin_unlock_irqrestore(&zone->lock, flags); | |
3085 | for (order = 0; order < MAX_ORDER; order++) { | |
3086 | printk("%lu*%lukB ", nr[order], K(1UL) << order); | |
3087 | if (nr[order]) | |
3088 | show_migration_types(types[order]); | |
3089 | } | |
3090 | printk("= %lukB\n", K(total)); | |
3091 | } | |
3092 | ||
3093 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); | |
3094 | ||
3095 | show_swap_cache_info(); | |
3096 | } | |
3097 | ||
3098 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) | |
3099 | { | |
3100 | zoneref->zone = zone; | |
3101 | zoneref->zone_idx = zone_idx(zone); | |
3102 | } | |
3103 | ||
3104 | /* | |
3105 | * Builds allocation fallback zone lists. | |
3106 | * | |
3107 | * Add all populated zones of a node to the zonelist. | |
3108 | */ | |
3109 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, | |
3110 | int nr_zones, enum zone_type zone_type) | |
3111 | { | |
3112 | struct zone *zone; | |
3113 | ||
3114 | BUG_ON(zone_type >= MAX_NR_ZONES); | |
3115 | zone_type++; | |
3116 | ||
3117 | do { | |
3118 | zone_type--; | |
3119 | zone = pgdat->node_zones + zone_type; | |
3120 | if (populated_zone(zone)) { | |
3121 | zoneref_set_zone(zone, | |
3122 | &zonelist->_zonerefs[nr_zones++]); | |
3123 | check_highest_zone(zone_type); | |
3124 | } | |
3125 | ||
3126 | } while (zone_type); | |
3127 | return nr_zones; | |
3128 | } | |
3129 | ||
3130 | ||
3131 | /* | |
3132 | * zonelist_order: | |
3133 | * 0 = automatic detection of better ordering. | |
3134 | * 1 = order by ([node] distance, -zonetype) | |
3135 | * 2 = order by (-zonetype, [node] distance) | |
3136 | * | |
3137 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create | |
3138 | * the same zonelist. So only NUMA can configure this param. | |
3139 | */ | |
3140 | #define ZONELIST_ORDER_DEFAULT 0 | |
3141 | #define ZONELIST_ORDER_NODE 1 | |
3142 | #define ZONELIST_ORDER_ZONE 2 | |
3143 | ||
3144 | /* zonelist order in the kernel. | |
3145 | * set_zonelist_order() will set this to NODE or ZONE. | |
3146 | */ | |
3147 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3148 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; | |
3149 | ||
3150 | ||
3151 | #ifdef CONFIG_NUMA | |
3152 | /* The value user specified ....changed by config */ | |
3153 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3154 | /* string for sysctl */ | |
3155 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
3156 | char numa_zonelist_order[16] = "default"; | |
3157 | ||
3158 | /* | |
3159 | * interface for configure zonelist ordering. | |
3160 | * command line option "numa_zonelist_order" | |
3161 | * = "[dD]efault - default, automatic configuration. | |
3162 | * = "[nN]ode - order by node locality, then by zone within node | |
3163 | * = "[zZ]one - order by zone, then by locality within zone | |
3164 | */ | |
3165 | ||
3166 | static int __parse_numa_zonelist_order(char *s) | |
3167 | { | |
3168 | if (*s == 'd' || *s == 'D') { | |
3169 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; | |
3170 | } else if (*s == 'n' || *s == 'N') { | |
3171 | user_zonelist_order = ZONELIST_ORDER_NODE; | |
3172 | } else if (*s == 'z' || *s == 'Z') { | |
3173 | user_zonelist_order = ZONELIST_ORDER_ZONE; | |
3174 | } else { | |
3175 | printk(KERN_WARNING | |
3176 | "Ignoring invalid numa_zonelist_order value: " | |
3177 | "%s\n", s); | |
3178 | return -EINVAL; | |
3179 | } | |
3180 | return 0; | |
3181 | } | |
3182 | ||
3183 | static __init int setup_numa_zonelist_order(char *s) | |
3184 | { | |
3185 | int ret; | |
3186 | ||
3187 | if (!s) | |
3188 | return 0; | |
3189 | ||
3190 | ret = __parse_numa_zonelist_order(s); | |
3191 | if (ret == 0) | |
3192 | strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); | |
3193 | ||
3194 | return ret; | |
3195 | } | |
3196 | early_param("numa_zonelist_order", setup_numa_zonelist_order); | |
3197 | ||
3198 | /* | |
3199 | * sysctl handler for numa_zonelist_order | |
3200 | */ | |
3201 | int numa_zonelist_order_handler(ctl_table *table, int write, | |
3202 | void __user *buffer, size_t *length, | |
3203 | loff_t *ppos) | |
3204 | { | |
3205 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; | |
3206 | int ret; | |
3207 | static DEFINE_MUTEX(zl_order_mutex); | |
3208 | ||
3209 | mutex_lock(&zl_order_mutex); | |
3210 | if (write) | |
3211 | strcpy(saved_string, (char*)table->data); | |
3212 | ret = proc_dostring(table, write, buffer, length, ppos); | |
3213 | if (ret) | |
3214 | goto out; | |
3215 | if (write) { | |
3216 | int oldval = user_zonelist_order; | |
3217 | if (__parse_numa_zonelist_order((char*)table->data)) { | |
3218 | /* | |
3219 | * bogus value. restore saved string | |
3220 | */ | |
3221 | strncpy((char*)table->data, saved_string, | |
3222 | NUMA_ZONELIST_ORDER_LEN); | |
3223 | user_zonelist_order = oldval; | |
3224 | } else if (oldval != user_zonelist_order) { | |
3225 | mutex_lock(&zonelists_mutex); | |
3226 | build_all_zonelists(NULL, NULL); | |
3227 | mutex_unlock(&zonelists_mutex); | |
3228 | } | |
3229 | } | |
3230 | out: | |
3231 | mutex_unlock(&zl_order_mutex); | |
3232 | return ret; | |
3233 | } | |
3234 | ||
3235 | ||
3236 | #define MAX_NODE_LOAD (nr_online_nodes) | |
3237 | static int node_load[MAX_NUMNODES]; | |
3238 | ||
3239 | /** | |
3240 | * find_next_best_node - find the next node that should appear in a given node's fallback list | |
3241 | * @node: node whose fallback list we're appending | |
3242 | * @used_node_mask: nodemask_t of already used nodes | |
3243 | * | |
3244 | * We use a number of factors to determine which is the next node that should | |
3245 | * appear on a given node's fallback list. The node should not have appeared | |
3246 | * already in @node's fallback list, and it should be the next closest node | |
3247 | * according to the distance array (which contains arbitrary distance values | |
3248 | * from each node to each node in the system), and should also prefer nodes | |
3249 | * with no CPUs, since presumably they'll have very little allocation pressure | |
3250 | * on them otherwise. | |
3251 | * It returns -1 if no node is found. | |
3252 | */ | |
3253 | static int find_next_best_node(int node, nodemask_t *used_node_mask) | |
3254 | { | |
3255 | int n, val; | |
3256 | int min_val = INT_MAX; | |
3257 | int best_node = -1; | |
3258 | const struct cpumask *tmp = cpumask_of_node(0); | |
3259 | ||
3260 | /* Use the local node if we haven't already */ | |
3261 | if (!node_isset(node, *used_node_mask)) { | |
3262 | node_set(node, *used_node_mask); | |
3263 | return node; | |
3264 | } | |
3265 | ||
3266 | for_each_node_state(n, N_MEMORY) { | |
3267 | ||
3268 | /* Don't want a node to appear more than once */ | |
3269 | if (node_isset(n, *used_node_mask)) | |
3270 | continue; | |
3271 | ||
3272 | /* Use the distance array to find the distance */ | |
3273 | val = node_distance(node, n); | |
3274 | ||
3275 | /* Penalize nodes under us ("prefer the next node") */ | |
3276 | val += (n < node); | |
3277 | ||
3278 | /* Give preference to headless and unused nodes */ | |
3279 | tmp = cpumask_of_node(n); | |
3280 | if (!cpumask_empty(tmp)) | |
3281 | val += PENALTY_FOR_NODE_WITH_CPUS; | |
3282 | ||
3283 | /* Slight preference for less loaded node */ | |
3284 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); | |
3285 | val += node_load[n]; | |
3286 | ||
3287 | if (val < min_val) { | |
3288 | min_val = val; | |
3289 | best_node = n; | |
3290 | } | |
3291 | } | |
3292 | ||
3293 | if (best_node >= 0) | |
3294 | node_set(best_node, *used_node_mask); | |
3295 | ||
3296 | return best_node; | |
3297 | } | |
3298 | ||
3299 | ||
3300 | /* | |
3301 | * Build zonelists ordered by node and zones within node. | |
3302 | * This results in maximum locality--normal zone overflows into local | |
3303 | * DMA zone, if any--but risks exhausting DMA zone. | |
3304 | */ | |
3305 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) | |
3306 | { | |
3307 | int j; | |
3308 | struct zonelist *zonelist; | |
3309 | ||
3310 | zonelist = &pgdat->node_zonelists[0]; | |
3311 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) | |
3312 | ; | |
3313 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3314 | MAX_NR_ZONES - 1); | |
3315 | zonelist->_zonerefs[j].zone = NULL; | |
3316 | zonelist->_zonerefs[j].zone_idx = 0; | |
3317 | } | |
3318 | ||
3319 | /* | |
3320 | * Build gfp_thisnode zonelists | |
3321 | */ | |
3322 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
3323 | { | |
3324 | int j; | |
3325 | struct zonelist *zonelist; | |
3326 | ||
3327 | zonelist = &pgdat->node_zonelists[1]; | |
3328 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
3329 | zonelist->_zonerefs[j].zone = NULL; | |
3330 | zonelist->_zonerefs[j].zone_idx = 0; | |
3331 | } | |
3332 | ||
3333 | /* | |
3334 | * Build zonelists ordered by zone and nodes within zones. | |
3335 | * This results in conserving DMA zone[s] until all Normal memory is | |
3336 | * exhausted, but results in overflowing to remote node while memory | |
3337 | * may still exist in local DMA zone. | |
3338 | */ | |
3339 | static int node_order[MAX_NUMNODES]; | |
3340 | ||
3341 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) | |
3342 | { | |
3343 | int pos, j, node; | |
3344 | int zone_type; /* needs to be signed */ | |
3345 | struct zone *z; | |
3346 | struct zonelist *zonelist; | |
3347 | ||
3348 | zonelist = &pgdat->node_zonelists[0]; | |
3349 | pos = 0; | |
3350 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { | |
3351 | for (j = 0; j < nr_nodes; j++) { | |
3352 | node = node_order[j]; | |
3353 | z = &NODE_DATA(node)->node_zones[zone_type]; | |
3354 | if (populated_zone(z)) { | |
3355 | zoneref_set_zone(z, | |
3356 | &zonelist->_zonerefs[pos++]); | |
3357 | check_highest_zone(zone_type); | |
3358 | } | |
3359 | } | |
3360 | } | |
3361 | zonelist->_zonerefs[pos].zone = NULL; | |
3362 | zonelist->_zonerefs[pos].zone_idx = 0; | |
3363 | } | |
3364 | ||
3365 | static int default_zonelist_order(void) | |
3366 | { | |
3367 | int nid, zone_type; | |
3368 | unsigned long low_kmem_size,total_size; | |
3369 | struct zone *z; | |
3370 | int average_size; | |
3371 | /* | |
3372 | * ZONE_DMA and ZONE_DMA32 can be very small area in the system. | |
3373 | * If they are really small and used heavily, the system can fall | |
3374 | * into OOM very easily. | |
3375 | * This function detect ZONE_DMA/DMA32 size and configures zone order. | |
3376 | */ | |
3377 | /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ | |
3378 | low_kmem_size = 0; | |
3379 | total_size = 0; | |
3380 | for_each_online_node(nid) { | |
3381 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
3382 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
3383 | if (populated_zone(z)) { | |
3384 | if (zone_type < ZONE_NORMAL) | |
3385 | low_kmem_size += z->present_pages; | |
3386 | total_size += z->present_pages; | |
3387 | } else if (zone_type == ZONE_NORMAL) { | |
3388 | /* | |
3389 | * If any node has only lowmem, then node order | |
3390 | * is preferred to allow kernel allocations | |
3391 | * locally; otherwise, they can easily infringe | |
3392 | * on other nodes when there is an abundance of | |
3393 | * lowmem available to allocate from. | |
3394 | */ | |
3395 | return ZONELIST_ORDER_NODE; | |
3396 | } | |
3397 | } | |
3398 | } | |
3399 | if (!low_kmem_size || /* there are no DMA area. */ | |
3400 | low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ | |
3401 | return ZONELIST_ORDER_NODE; | |
3402 | /* | |
3403 | * look into each node's config. | |
3404 | * If there is a node whose DMA/DMA32 memory is very big area on | |
3405 | * local memory, NODE_ORDER may be suitable. | |
3406 | */ | |
3407 | average_size = total_size / | |
3408 | (nodes_weight(node_states[N_MEMORY]) + 1); | |
3409 | for_each_online_node(nid) { | |
3410 | low_kmem_size = 0; | |
3411 | total_size = 0; | |
3412 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { | |
3413 | z = &NODE_DATA(nid)->node_zones[zone_type]; | |
3414 | if (populated_zone(z)) { | |
3415 | if (zone_type < ZONE_NORMAL) | |
3416 | low_kmem_size += z->present_pages; | |
3417 | total_size += z->present_pages; | |
3418 | } | |
3419 | } | |
3420 | if (low_kmem_size && | |
3421 | total_size > average_size && /* ignore small node */ | |
3422 | low_kmem_size > total_size * 70/100) | |
3423 | return ZONELIST_ORDER_NODE; | |
3424 | } | |
3425 | return ZONELIST_ORDER_ZONE; | |
3426 | } | |
3427 | ||
3428 | static void set_zonelist_order(void) | |
3429 | { | |
3430 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) | |
3431 | current_zonelist_order = default_zonelist_order(); | |
3432 | else | |
3433 | current_zonelist_order = user_zonelist_order; | |
3434 | } | |
3435 | ||
3436 | static void build_zonelists(pg_data_t *pgdat) | |
3437 | { | |
3438 | int j, node, load; | |
3439 | enum zone_type i; | |
3440 | nodemask_t used_mask; | |
3441 | int local_node, prev_node; | |
3442 | struct zonelist *zonelist; | |
3443 | int order = current_zonelist_order; | |
3444 | ||
3445 | /* initialize zonelists */ | |
3446 | for (i = 0; i < MAX_ZONELISTS; i++) { | |
3447 | zonelist = pgdat->node_zonelists + i; | |
3448 | zonelist->_zonerefs[0].zone = NULL; | |
3449 | zonelist->_zonerefs[0].zone_idx = 0; | |
3450 | } | |
3451 | ||
3452 | /* NUMA-aware ordering of nodes */ | |
3453 | local_node = pgdat->node_id; | |
3454 | load = nr_online_nodes; | |
3455 | prev_node = local_node; | |
3456 | nodes_clear(used_mask); | |
3457 | ||
3458 | memset(node_order, 0, sizeof(node_order)); | |
3459 | j = 0; | |
3460 | ||
3461 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { | |
3462 | /* | |
3463 | * We don't want to pressure a particular node. | |
3464 | * So adding penalty to the first node in same | |
3465 | * distance group to make it round-robin. | |
3466 | */ | |
3467 | if (node_distance(local_node, node) != | |
3468 | node_distance(local_node, prev_node)) | |
3469 | node_load[node] = load; | |
3470 | ||
3471 | prev_node = node; | |
3472 | load--; | |
3473 | if (order == ZONELIST_ORDER_NODE) | |
3474 | build_zonelists_in_node_order(pgdat, node); | |
3475 | else | |
3476 | node_order[j++] = node; /* remember order */ | |
3477 | } | |
3478 | ||
3479 | if (order == ZONELIST_ORDER_ZONE) { | |
3480 | /* calculate node order -- i.e., DMA last! */ | |
3481 | build_zonelists_in_zone_order(pgdat, j); | |
3482 | } | |
3483 | ||
3484 | build_thisnode_zonelists(pgdat); | |
3485 | } | |
3486 | ||
3487 | /* Construct the zonelist performance cache - see further mmzone.h */ | |
3488 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3489 | { | |
3490 | struct zonelist *zonelist; | |
3491 | struct zonelist_cache *zlc; | |
3492 | struct zoneref *z; | |
3493 | ||
3494 | zonelist = &pgdat->node_zonelists[0]; | |
3495 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; | |
3496 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); | |
3497 | for (z = zonelist->_zonerefs; z->zone; z++) | |
3498 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); | |
3499 | } | |
3500 | ||
3501 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3502 | /* | |
3503 | * Return node id of node used for "local" allocations. | |
3504 | * I.e., first node id of first zone in arg node's generic zonelist. | |
3505 | * Used for initializing percpu 'numa_mem', which is used primarily | |
3506 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. | |
3507 | */ | |
3508 | int local_memory_node(int node) | |
3509 | { | |
3510 | struct zone *zone; | |
3511 | ||
3512 | (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), | |
3513 | gfp_zone(GFP_KERNEL), | |
3514 | NULL, | |
3515 | &zone); | |
3516 | return zone->node; | |
3517 | } | |
3518 | #endif | |
3519 | ||
3520 | #else /* CONFIG_NUMA */ | |
3521 | ||
3522 | static void set_zonelist_order(void) | |
3523 | { | |
3524 | current_zonelist_order = ZONELIST_ORDER_ZONE; | |
3525 | } | |
3526 | ||
3527 | static void build_zonelists(pg_data_t *pgdat) | |
3528 | { | |
3529 | int node, local_node; | |
3530 | enum zone_type j; | |
3531 | struct zonelist *zonelist; | |
3532 | ||
3533 | local_node = pgdat->node_id; | |
3534 | ||
3535 | zonelist = &pgdat->node_zonelists[0]; | |
3536 | j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); | |
3537 | ||
3538 | /* | |
3539 | * Now we build the zonelist so that it contains the zones | |
3540 | * of all the other nodes. | |
3541 | * We don't want to pressure a particular node, so when | |
3542 | * building the zones for node N, we make sure that the | |
3543 | * zones coming right after the local ones are those from | |
3544 | * node N+1 (modulo N) | |
3545 | */ | |
3546 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
3547 | if (!node_online(node)) | |
3548 | continue; | |
3549 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3550 | MAX_NR_ZONES - 1); | |
3551 | } | |
3552 | for (node = 0; node < local_node; node++) { | |
3553 | if (!node_online(node)) | |
3554 | continue; | |
3555 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, | |
3556 | MAX_NR_ZONES - 1); | |
3557 | } | |
3558 | ||
3559 | zonelist->_zonerefs[j].zone = NULL; | |
3560 | zonelist->_zonerefs[j].zone_idx = 0; | |
3561 | } | |
3562 | ||
3563 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ | |
3564 | static void build_zonelist_cache(pg_data_t *pgdat) | |
3565 | { | |
3566 | pgdat->node_zonelists[0].zlcache_ptr = NULL; | |
3567 | } | |
3568 | ||
3569 | #endif /* CONFIG_NUMA */ | |
3570 | ||
3571 | /* | |
3572 | * Boot pageset table. One per cpu which is going to be used for all | |
3573 | * zones and all nodes. The parameters will be set in such a way | |
3574 | * that an item put on a list will immediately be handed over to | |
3575 | * the buddy list. This is safe since pageset manipulation is done | |
3576 | * with interrupts disabled. | |
3577 | * | |
3578 | * The boot_pagesets must be kept even after bootup is complete for | |
3579 | * unused processors and/or zones. They do play a role for bootstrapping | |
3580 | * hotplugged processors. | |
3581 | * | |
3582 | * zoneinfo_show() and maybe other functions do | |
3583 | * not check if the processor is online before following the pageset pointer. | |
3584 | * Other parts of the kernel may not check if the zone is available. | |
3585 | */ | |
3586 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); | |
3587 | static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); | |
3588 | static void setup_zone_pageset(struct zone *zone); | |
3589 | ||
3590 | /* | |
3591 | * Global mutex to protect against size modification of zonelists | |
3592 | * as well as to serialize pageset setup for the new populated zone. | |
3593 | */ | |
3594 | DEFINE_MUTEX(zonelists_mutex); | |
3595 | ||
3596 | /* return values int ....just for stop_machine() */ | |
3597 | static int __build_all_zonelists(void *data) | |
3598 | { | |
3599 | int nid; | |
3600 | int cpu; | |
3601 | pg_data_t *self = data; | |
3602 | ||
3603 | #ifdef CONFIG_NUMA | |
3604 | memset(node_load, 0, sizeof(node_load)); | |
3605 | #endif | |
3606 | ||
3607 | if (self && !node_online(self->node_id)) { | |
3608 | build_zonelists(self); | |
3609 | build_zonelist_cache(self); | |
3610 | } | |
3611 | ||
3612 | for_each_online_node(nid) { | |
3613 | pg_data_t *pgdat = NODE_DATA(nid); | |
3614 | ||
3615 | build_zonelists(pgdat); | |
3616 | build_zonelist_cache(pgdat); | |
3617 | } | |
3618 | ||
3619 | /* | |
3620 | * Initialize the boot_pagesets that are going to be used | |
3621 | * for bootstrapping processors. The real pagesets for | |
3622 | * each zone will be allocated later when the per cpu | |
3623 | * allocator is available. | |
3624 | * | |
3625 | * boot_pagesets are used also for bootstrapping offline | |
3626 | * cpus if the system is already booted because the pagesets | |
3627 | * are needed to initialize allocators on a specific cpu too. | |
3628 | * F.e. the percpu allocator needs the page allocator which | |
3629 | * needs the percpu allocator in order to allocate its pagesets | |
3630 | * (a chicken-egg dilemma). | |
3631 | */ | |
3632 | for_each_possible_cpu(cpu) { | |
3633 | setup_pageset(&per_cpu(boot_pageset, cpu), 0); | |
3634 | ||
3635 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES | |
3636 | /* | |
3637 | * We now know the "local memory node" for each node-- | |
3638 | * i.e., the node of the first zone in the generic zonelist. | |
3639 | * Set up numa_mem percpu variable for on-line cpus. During | |
3640 | * boot, only the boot cpu should be on-line; we'll init the | |
3641 | * secondary cpus' numa_mem as they come on-line. During | |
3642 | * node/memory hotplug, we'll fixup all on-line cpus. | |
3643 | */ | |
3644 | if (cpu_online(cpu)) | |
3645 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); | |
3646 | #endif | |
3647 | } | |
3648 | ||
3649 | return 0; | |
3650 | } | |
3651 | ||
3652 | /* | |
3653 | * Called with zonelists_mutex held always | |
3654 | * unless system_state == SYSTEM_BOOTING. | |
3655 | */ | |
3656 | void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) | |
3657 | { | |
3658 | set_zonelist_order(); | |
3659 | ||
3660 | if (system_state == SYSTEM_BOOTING) { | |
3661 | __build_all_zonelists(NULL); | |
3662 | mminit_verify_zonelist(); | |
3663 | cpuset_init_current_mems_allowed(); | |
3664 | } else { | |
3665 | /* we have to stop all cpus to guarantee there is no user | |
3666 | of zonelist */ | |
3667 | #ifdef CONFIG_MEMORY_HOTPLUG | |
3668 | if (zone) | |
3669 | setup_zone_pageset(zone); | |
3670 | #endif | |
3671 | stop_machine(__build_all_zonelists, pgdat, NULL); | |
3672 | /* cpuset refresh routine should be here */ | |
3673 | } | |
3674 | vm_total_pages = nr_free_pagecache_pages(); | |
3675 | /* | |
3676 | * Disable grouping by mobility if the number of pages in the | |
3677 | * system is too low to allow the mechanism to work. It would be | |
3678 | * more accurate, but expensive to check per-zone. This check is | |
3679 | * made on memory-hotadd so a system can start with mobility | |
3680 | * disabled and enable it later | |
3681 | */ | |
3682 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) | |
3683 | page_group_by_mobility_disabled = 1; | |
3684 | else | |
3685 | page_group_by_mobility_disabled = 0; | |
3686 | ||
3687 | printk("Built %i zonelists in %s order, mobility grouping %s. " | |
3688 | "Total pages: %ld\n", | |
3689 | nr_online_nodes, | |
3690 | zonelist_order_name[current_zonelist_order], | |
3691 | page_group_by_mobility_disabled ? "off" : "on", | |
3692 | vm_total_pages); | |
3693 | #ifdef CONFIG_NUMA | |
3694 | printk("Policy zone: %s\n", zone_names[policy_zone]); | |
3695 | #endif | |
3696 | } | |
3697 | ||
3698 | /* | |
3699 | * Helper functions to size the waitqueue hash table. | |
3700 | * Essentially these want to choose hash table sizes sufficiently | |
3701 | * large so that collisions trying to wait on pages are rare. | |
3702 | * But in fact, the number of active page waitqueues on typical | |
3703 | * systems is ridiculously low, less than 200. So this is even | |
3704 | * conservative, even though it seems large. | |
3705 | * | |
3706 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to | |
3707 | * waitqueues, i.e. the size of the waitq table given the number of pages. | |
3708 | */ | |
3709 | #define PAGES_PER_WAITQUEUE 256 | |
3710 | ||
3711 | #ifndef CONFIG_MEMORY_HOTPLUG | |
3712 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3713 | { | |
3714 | unsigned long size = 1; | |
3715 | ||
3716 | pages /= PAGES_PER_WAITQUEUE; | |
3717 | ||
3718 | while (size < pages) | |
3719 | size <<= 1; | |
3720 | ||
3721 | /* | |
3722 | * Once we have dozens or even hundreds of threads sleeping | |
3723 | * on IO we've got bigger problems than wait queue collision. | |
3724 | * Limit the size of the wait table to a reasonable size. | |
3725 | */ | |
3726 | size = min(size, 4096UL); | |
3727 | ||
3728 | return max(size, 4UL); | |
3729 | } | |
3730 | #else | |
3731 | /* | |
3732 | * A zone's size might be changed by hot-add, so it is not possible to determine | |
3733 | * a suitable size for its wait_table. So we use the maximum size now. | |
3734 | * | |
3735 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: | |
3736 | * | |
3737 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. | |
3738 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. | |
3739 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. | |
3740 | * | |
3741 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages | |
3742 | * or more by the traditional way. (See above). It equals: | |
3743 | * | |
3744 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. | |
3745 | * ia64(16K page size) : = ( 8G + 4M)byte. | |
3746 | * powerpc (64K page size) : = (32G +16M)byte. | |
3747 | */ | |
3748 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) | |
3749 | { | |
3750 | return 4096UL; | |
3751 | } | |
3752 | #endif | |
3753 | ||
3754 | /* | |
3755 | * This is an integer logarithm so that shifts can be used later | |
3756 | * to extract the more random high bits from the multiplicative | |
3757 | * hash function before the remainder is taken. | |
3758 | */ | |
3759 | static inline unsigned long wait_table_bits(unsigned long size) | |
3760 | { | |
3761 | return ffz(~size); | |
3762 | } | |
3763 | ||
3764 | #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) | |
3765 | ||
3766 | /* | |
3767 | * Check if a pageblock contains reserved pages | |
3768 | */ | |
3769 | static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) | |
3770 | { | |
3771 | unsigned long pfn; | |
3772 | ||
3773 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
3774 | if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) | |
3775 | return 1; | |
3776 | } | |
3777 | return 0; | |
3778 | } | |
3779 | ||
3780 | /* | |
3781 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number | |
3782 | * of blocks reserved is based on min_wmark_pages(zone). The memory within | |
3783 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes | |
3784 | * higher will lead to a bigger reserve which will get freed as contiguous | |
3785 | * blocks as reclaim kicks in | |
3786 | */ | |
3787 | static void setup_zone_migrate_reserve(struct zone *zone) | |
3788 | { | |
3789 | unsigned long start_pfn, pfn, end_pfn, block_end_pfn; | |
3790 | struct page *page; | |
3791 | unsigned long block_migratetype; | |
3792 | int reserve; | |
3793 | ||
3794 | /* | |
3795 | * Get the start pfn, end pfn and the number of blocks to reserve | |
3796 | * We have to be careful to be aligned to pageblock_nr_pages to | |
3797 | * make sure that we always check pfn_valid for the first page in | |
3798 | * the block. | |
3799 | */ | |
3800 | start_pfn = zone->zone_start_pfn; | |
3801 | end_pfn = zone_end_pfn(zone); | |
3802 | start_pfn = roundup(start_pfn, pageblock_nr_pages); | |
3803 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> | |
3804 | pageblock_order; | |
3805 | ||
3806 | /* | |
3807 | * Reserve blocks are generally in place to help high-order atomic | |
3808 | * allocations that are short-lived. A min_free_kbytes value that | |
3809 | * would result in more than 2 reserve blocks for atomic allocations | |
3810 | * is assumed to be in place to help anti-fragmentation for the | |
3811 | * future allocation of hugepages at runtime. | |
3812 | */ | |
3813 | reserve = min(2, reserve); | |
3814 | ||
3815 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { | |
3816 | if (!pfn_valid(pfn)) | |
3817 | continue; | |
3818 | page = pfn_to_page(pfn); | |
3819 | ||
3820 | /* Watch out for overlapping nodes */ | |
3821 | if (page_to_nid(page) != zone_to_nid(zone)) | |
3822 | continue; | |
3823 | ||
3824 | block_migratetype = get_pageblock_migratetype(page); | |
3825 | ||
3826 | /* Only test what is necessary when the reserves are not met */ | |
3827 | if (reserve > 0) { | |
3828 | /* | |
3829 | * Blocks with reserved pages will never free, skip | |
3830 | * them. | |
3831 | */ | |
3832 | block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); | |
3833 | if (pageblock_is_reserved(pfn, block_end_pfn)) | |
3834 | continue; | |
3835 | ||
3836 | /* If this block is reserved, account for it */ | |
3837 | if (block_migratetype == MIGRATE_RESERVE) { | |
3838 | reserve--; | |
3839 | continue; | |
3840 | } | |
3841 | ||
3842 | /* Suitable for reserving if this block is movable */ | |
3843 | if (block_migratetype == MIGRATE_MOVABLE) { | |
3844 | set_pageblock_migratetype(page, | |
3845 | MIGRATE_RESERVE); | |
3846 | move_freepages_block(zone, page, | |
3847 | MIGRATE_RESERVE); | |
3848 | reserve--; | |
3849 | continue; | |
3850 | } | |
3851 | } | |
3852 | ||
3853 | /* | |
3854 | * If the reserve is met and this is a previous reserved block, | |
3855 | * take it back | |
3856 | */ | |
3857 | if (block_migratetype == MIGRATE_RESERVE) { | |
3858 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
3859 | move_freepages_block(zone, page, MIGRATE_MOVABLE); | |
3860 | } | |
3861 | } | |
3862 | } | |
3863 | ||
3864 | /* | |
3865 | * Initially all pages are reserved - free ones are freed | |
3866 | * up by free_all_bootmem() once the early boot process is | |
3867 | * done. Non-atomic initialization, single-pass. | |
3868 | */ | |
3869 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, | |
3870 | unsigned long start_pfn, enum memmap_context context) | |
3871 | { | |
3872 | struct page *page; | |
3873 | unsigned long end_pfn = start_pfn + size; | |
3874 | unsigned long pfn; | |
3875 | struct zone *z; | |
3876 | ||
3877 | if (highest_memmap_pfn < end_pfn - 1) | |
3878 | highest_memmap_pfn = end_pfn - 1; | |
3879 | ||
3880 | z = &NODE_DATA(nid)->node_zones[zone]; | |
3881 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { | |
3882 | /* | |
3883 | * There can be holes in boot-time mem_map[]s | |
3884 | * handed to this function. They do not | |
3885 | * exist on hotplugged memory. | |
3886 | */ | |
3887 | if (context == MEMMAP_EARLY) { | |
3888 | if (!early_pfn_valid(pfn)) | |
3889 | continue; | |
3890 | if (!early_pfn_in_nid(pfn, nid)) | |
3891 | continue; | |
3892 | } | |
3893 | page = pfn_to_page(pfn); | |
3894 | set_page_links(page, zone, nid, pfn); | |
3895 | mminit_verify_page_links(page, zone, nid, pfn); | |
3896 | init_page_count(page); | |
3897 | page_mapcount_reset(page); | |
3898 | page_nid_reset_last(page); | |
3899 | SetPageReserved(page); | |
3900 | /* | |
3901 | * Mark the block movable so that blocks are reserved for | |
3902 | * movable at startup. This will force kernel allocations | |
3903 | * to reserve their blocks rather than leaking throughout | |
3904 | * the address space during boot when many long-lived | |
3905 | * kernel allocations are made. Later some blocks near | |
3906 | * the start are marked MIGRATE_RESERVE by | |
3907 | * setup_zone_migrate_reserve() | |
3908 | * | |
3909 | * bitmap is created for zone's valid pfn range. but memmap | |
3910 | * can be created for invalid pages (for alignment) | |
3911 | * check here not to call set_pageblock_migratetype() against | |
3912 | * pfn out of zone. | |
3913 | */ | |
3914 | if ((z->zone_start_pfn <= pfn) | |
3915 | && (pfn < zone_end_pfn(z)) | |
3916 | && !(pfn & (pageblock_nr_pages - 1))) | |
3917 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); | |
3918 | ||
3919 | INIT_LIST_HEAD(&page->lru); | |
3920 | #ifdef WANT_PAGE_VIRTUAL | |
3921 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ | |
3922 | if (!is_highmem_idx(zone)) | |
3923 | set_page_address(page, __va(pfn << PAGE_SHIFT)); | |
3924 | #endif | |
3925 | } | |
3926 | } | |
3927 | ||
3928 | static void __meminit zone_init_free_lists(struct zone *zone) | |
3929 | { | |
3930 | int order, t; | |
3931 | for_each_migratetype_order(order, t) { | |
3932 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); | |
3933 | zone->free_area[order].nr_free = 0; | |
3934 | } | |
3935 | } | |
3936 | ||
3937 | #ifndef __HAVE_ARCH_MEMMAP_INIT | |
3938 | #define memmap_init(size, nid, zone, start_pfn) \ | |
3939 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) | |
3940 | #endif | |
3941 | ||
3942 | static int __meminit zone_batchsize(struct zone *zone) | |
3943 | { | |
3944 | #ifdef CONFIG_MMU | |
3945 | int batch; | |
3946 | ||
3947 | /* | |
3948 | * The per-cpu-pages pools are set to around 1000th of the | |
3949 | * size of the zone. But no more than 1/2 of a meg. | |
3950 | * | |
3951 | * OK, so we don't know how big the cache is. So guess. | |
3952 | */ | |
3953 | batch = zone->managed_pages / 1024; | |
3954 | if (batch * PAGE_SIZE > 512 * 1024) | |
3955 | batch = (512 * 1024) / PAGE_SIZE; | |
3956 | batch /= 4; /* We effectively *= 4 below */ | |
3957 | if (batch < 1) | |
3958 | batch = 1; | |
3959 | ||
3960 | /* | |
3961 | * Clamp the batch to a 2^n - 1 value. Having a power | |
3962 | * of 2 value was found to be more likely to have | |
3963 | * suboptimal cache aliasing properties in some cases. | |
3964 | * | |
3965 | * For example if 2 tasks are alternately allocating | |
3966 | * batches of pages, one task can end up with a lot | |
3967 | * of pages of one half of the possible page colors | |
3968 | * and the other with pages of the other colors. | |
3969 | */ | |
3970 | batch = rounddown_pow_of_two(batch + batch/2) - 1; | |
3971 | ||
3972 | return batch; | |
3973 | ||
3974 | #else | |
3975 | /* The deferral and batching of frees should be suppressed under NOMMU | |
3976 | * conditions. | |
3977 | * | |
3978 | * The problem is that NOMMU needs to be able to allocate large chunks | |
3979 | * of contiguous memory as there's no hardware page translation to | |
3980 | * assemble apparent contiguous memory from discontiguous pages. | |
3981 | * | |
3982 | * Queueing large contiguous runs of pages for batching, however, | |
3983 | * causes the pages to actually be freed in smaller chunks. As there | |
3984 | * can be a significant delay between the individual batches being | |
3985 | * recycled, this leads to the once large chunks of space being | |
3986 | * fragmented and becoming unavailable for high-order allocations. | |
3987 | */ | |
3988 | return 0; | |
3989 | #endif | |
3990 | } | |
3991 | ||
3992 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) | |
3993 | { | |
3994 | struct per_cpu_pages *pcp; | |
3995 | int migratetype; | |
3996 | ||
3997 | memset(p, 0, sizeof(*p)); | |
3998 | ||
3999 | pcp = &p->pcp; | |
4000 | pcp->count = 0; | |
4001 | pcp->high = 6 * batch; | |
4002 | pcp->batch = max(1UL, 1 * batch); | |
4003 | for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) | |
4004 | INIT_LIST_HEAD(&pcp->lists[migratetype]); | |
4005 | } | |
4006 | ||
4007 | /* | |
4008 | * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist | |
4009 | * to the value high for the pageset p. | |
4010 | */ | |
4011 | ||
4012 | static void setup_pagelist_highmark(struct per_cpu_pageset *p, | |
4013 | unsigned long high) | |
4014 | { | |
4015 | struct per_cpu_pages *pcp; | |
4016 | ||
4017 | pcp = &p->pcp; | |
4018 | pcp->high = high; | |
4019 | pcp->batch = max(1UL, high/4); | |
4020 | if ((high/4) > (PAGE_SHIFT * 8)) | |
4021 | pcp->batch = PAGE_SHIFT * 8; | |
4022 | } | |
4023 | ||
4024 | static void __meminit setup_zone_pageset(struct zone *zone) | |
4025 | { | |
4026 | int cpu; | |
4027 | ||
4028 | zone->pageset = alloc_percpu(struct per_cpu_pageset); | |
4029 | ||
4030 | for_each_possible_cpu(cpu) { | |
4031 | struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); | |
4032 | ||
4033 | setup_pageset(pcp, zone_batchsize(zone)); | |
4034 | ||
4035 | if (percpu_pagelist_fraction) | |
4036 | setup_pagelist_highmark(pcp, | |
4037 | (zone->managed_pages / | |
4038 | percpu_pagelist_fraction)); | |
4039 | } | |
4040 | } | |
4041 | ||
4042 | /* | |
4043 | * Allocate per cpu pagesets and initialize them. | |
4044 | * Before this call only boot pagesets were available. | |
4045 | */ | |
4046 | void __init setup_per_cpu_pageset(void) | |
4047 | { | |
4048 | struct zone *zone; | |
4049 | ||
4050 | for_each_populated_zone(zone) | |
4051 | setup_zone_pageset(zone); | |
4052 | } | |
4053 | ||
4054 | static noinline __init_refok | |
4055 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) | |
4056 | { | |
4057 | int i; | |
4058 | struct pglist_data *pgdat = zone->zone_pgdat; | |
4059 | size_t alloc_size; | |
4060 | ||
4061 | /* | |
4062 | * The per-page waitqueue mechanism uses hashed waitqueues | |
4063 | * per zone. | |
4064 | */ | |
4065 | zone->wait_table_hash_nr_entries = | |
4066 | wait_table_hash_nr_entries(zone_size_pages); | |
4067 | zone->wait_table_bits = | |
4068 | wait_table_bits(zone->wait_table_hash_nr_entries); | |
4069 | alloc_size = zone->wait_table_hash_nr_entries | |
4070 | * sizeof(wait_queue_head_t); | |
4071 | ||
4072 | if (!slab_is_available()) { | |
4073 | zone->wait_table = (wait_queue_head_t *) | |
4074 | alloc_bootmem_node_nopanic(pgdat, alloc_size); | |
4075 | } else { | |
4076 | /* | |
4077 | * This case means that a zone whose size was 0 gets new memory | |
4078 | * via memory hot-add. | |
4079 | * But it may be the case that a new node was hot-added. In | |
4080 | * this case vmalloc() will not be able to use this new node's | |
4081 | * memory - this wait_table must be initialized to use this new | |
4082 | * node itself as well. | |
4083 | * To use this new node's memory, further consideration will be | |
4084 | * necessary. | |
4085 | */ | |
4086 | zone->wait_table = vmalloc(alloc_size); | |
4087 | } | |
4088 | if (!zone->wait_table) | |
4089 | return -ENOMEM; | |
4090 | ||
4091 | for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) | |
4092 | init_waitqueue_head(zone->wait_table + i); | |
4093 | ||
4094 | return 0; | |
4095 | } | |
4096 | ||
4097 | static __meminit void zone_pcp_init(struct zone *zone) | |
4098 | { | |
4099 | /* | |
4100 | * per cpu subsystem is not up at this point. The following code | |
4101 | * relies on the ability of the linker to provide the | |
4102 | * offset of a (static) per cpu variable into the per cpu area. | |
4103 | */ | |
4104 | zone->pageset = &boot_pageset; | |
4105 | ||
4106 | if (zone->present_pages) | |
4107 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", | |
4108 | zone->name, zone->present_pages, | |
4109 | zone_batchsize(zone)); | |
4110 | } | |
4111 | ||
4112 | int __meminit init_currently_empty_zone(struct zone *zone, | |
4113 | unsigned long zone_start_pfn, | |
4114 | unsigned long size, | |
4115 | enum memmap_context context) | |
4116 | { | |
4117 | struct pglist_data *pgdat = zone->zone_pgdat; | |
4118 | int ret; | |
4119 | ret = zone_wait_table_init(zone, size); | |
4120 | if (ret) | |
4121 | return ret; | |
4122 | pgdat->nr_zones = zone_idx(zone) + 1; | |
4123 | ||
4124 | zone->zone_start_pfn = zone_start_pfn; | |
4125 | ||
4126 | mminit_dprintk(MMINIT_TRACE, "memmap_init", | |
4127 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", | |
4128 | pgdat->node_id, | |
4129 | (unsigned long)zone_idx(zone), | |
4130 | zone_start_pfn, (zone_start_pfn + size)); | |
4131 | ||
4132 | zone_init_free_lists(zone); | |
4133 | ||
4134 | return 0; | |
4135 | } | |
4136 | ||
4137 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4138 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID | |
4139 | /* | |
4140 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. | |
4141 | * Architectures may implement their own version but if add_active_range() | |
4142 | * was used and there are no special requirements, this is a convenient | |
4143 | * alternative | |
4144 | */ | |
4145 | int __meminit __early_pfn_to_nid(unsigned long pfn) | |
4146 | { | |
4147 | unsigned long start_pfn, end_pfn; | |
4148 | int i, nid; | |
4149 | ||
4150 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) | |
4151 | if (start_pfn <= pfn && pfn < end_pfn) | |
4152 | return nid; | |
4153 | /* This is a memory hole */ | |
4154 | return -1; | |
4155 | } | |
4156 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ | |
4157 | ||
4158 | int __meminit early_pfn_to_nid(unsigned long pfn) | |
4159 | { | |
4160 | int nid; | |
4161 | ||
4162 | nid = __early_pfn_to_nid(pfn); | |
4163 | if (nid >= 0) | |
4164 | return nid; | |
4165 | /* just returns 0 */ | |
4166 | return 0; | |
4167 | } | |
4168 | ||
4169 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES | |
4170 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) | |
4171 | { | |
4172 | int nid; | |
4173 | ||
4174 | nid = __early_pfn_to_nid(pfn); | |
4175 | if (nid >= 0 && nid != node) | |
4176 | return false; | |
4177 | return true; | |
4178 | } | |
4179 | #endif | |
4180 | ||
4181 | /** | |
4182 | * free_bootmem_with_active_regions - Call free_bootmem_node for each active range | |
4183 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. | |
4184 | * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node | |
4185 | * | |
4186 | * If an architecture guarantees that all ranges registered with | |
4187 | * add_active_ranges() contain no holes and may be freed, this | |
4188 | * this function may be used instead of calling free_bootmem() manually. | |
4189 | */ | |
4190 | void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) | |
4191 | { | |
4192 | unsigned long start_pfn, end_pfn; | |
4193 | int i, this_nid; | |
4194 | ||
4195 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { | |
4196 | start_pfn = min(start_pfn, max_low_pfn); | |
4197 | end_pfn = min(end_pfn, max_low_pfn); | |
4198 | ||
4199 | if (start_pfn < end_pfn) | |
4200 | free_bootmem_node(NODE_DATA(this_nid), | |
4201 | PFN_PHYS(start_pfn), | |
4202 | (end_pfn - start_pfn) << PAGE_SHIFT); | |
4203 | } | |
4204 | } | |
4205 | ||
4206 | /** | |
4207 | * sparse_memory_present_with_active_regions - Call memory_present for each active range | |
4208 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. | |
4209 | * | |
4210 | * If an architecture guarantees that all ranges registered with | |
4211 | * add_active_ranges() contain no holes and may be freed, this | |
4212 | * function may be used instead of calling memory_present() manually. | |
4213 | */ | |
4214 | void __init sparse_memory_present_with_active_regions(int nid) | |
4215 | { | |
4216 | unsigned long start_pfn, end_pfn; | |
4217 | int i, this_nid; | |
4218 | ||
4219 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) | |
4220 | memory_present(this_nid, start_pfn, end_pfn); | |
4221 | } | |
4222 | ||
4223 | /** | |
4224 | * get_pfn_range_for_nid - Return the start and end page frames for a node | |
4225 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. | |
4226 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. | |
4227 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. | |
4228 | * | |
4229 | * It returns the start and end page frame of a node based on information | |
4230 | * provided by an arch calling add_active_range(). If called for a node | |
4231 | * with no available memory, a warning is printed and the start and end | |
4232 | * PFNs will be 0. | |
4233 | */ | |
4234 | void __meminit get_pfn_range_for_nid(unsigned int nid, | |
4235 | unsigned long *start_pfn, unsigned long *end_pfn) | |
4236 | { | |
4237 | unsigned long this_start_pfn, this_end_pfn; | |
4238 | int i; | |
4239 | ||
4240 | *start_pfn = -1UL; | |
4241 | *end_pfn = 0; | |
4242 | ||
4243 | for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { | |
4244 | *start_pfn = min(*start_pfn, this_start_pfn); | |
4245 | *end_pfn = max(*end_pfn, this_end_pfn); | |
4246 | } | |
4247 | ||
4248 | if (*start_pfn == -1UL) | |
4249 | *start_pfn = 0; | |
4250 | } | |
4251 | ||
4252 | /* | |
4253 | * This finds a zone that can be used for ZONE_MOVABLE pages. The | |
4254 | * assumption is made that zones within a node are ordered in monotonic | |
4255 | * increasing memory addresses so that the "highest" populated zone is used | |
4256 | */ | |
4257 | static void __init find_usable_zone_for_movable(void) | |
4258 | { | |
4259 | int zone_index; | |
4260 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { | |
4261 | if (zone_index == ZONE_MOVABLE) | |
4262 | continue; | |
4263 | ||
4264 | if (arch_zone_highest_possible_pfn[zone_index] > | |
4265 | arch_zone_lowest_possible_pfn[zone_index]) | |
4266 | break; | |
4267 | } | |
4268 | ||
4269 | VM_BUG_ON(zone_index == -1); | |
4270 | movable_zone = zone_index; | |
4271 | } | |
4272 | ||
4273 | /* | |
4274 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE | |
4275 | * because it is sized independent of architecture. Unlike the other zones, | |
4276 | * the starting point for ZONE_MOVABLE is not fixed. It may be different | |
4277 | * in each node depending on the size of each node and how evenly kernelcore | |
4278 | * is distributed. This helper function adjusts the zone ranges | |
4279 | * provided by the architecture for a given node by using the end of the | |
4280 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that | |
4281 | * zones within a node are in order of monotonic increases memory addresses | |
4282 | */ | |
4283 | static void __meminit adjust_zone_range_for_zone_movable(int nid, | |
4284 | unsigned long zone_type, | |
4285 | unsigned long node_start_pfn, | |
4286 | unsigned long node_end_pfn, | |
4287 | unsigned long *zone_start_pfn, | |
4288 | unsigned long *zone_end_pfn) | |
4289 | { | |
4290 | /* Only adjust if ZONE_MOVABLE is on this node */ | |
4291 | if (zone_movable_pfn[nid]) { | |
4292 | /* Size ZONE_MOVABLE */ | |
4293 | if (zone_type == ZONE_MOVABLE) { | |
4294 | *zone_start_pfn = zone_movable_pfn[nid]; | |
4295 | *zone_end_pfn = min(node_end_pfn, | |
4296 | arch_zone_highest_possible_pfn[movable_zone]); | |
4297 | ||
4298 | /* Adjust for ZONE_MOVABLE starting within this range */ | |
4299 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && | |
4300 | *zone_end_pfn > zone_movable_pfn[nid]) { | |
4301 | *zone_end_pfn = zone_movable_pfn[nid]; | |
4302 | ||
4303 | /* Check if this whole range is within ZONE_MOVABLE */ | |
4304 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) | |
4305 | *zone_start_pfn = *zone_end_pfn; | |
4306 | } | |
4307 | } | |
4308 | ||
4309 | /* | |
4310 | * Return the number of pages a zone spans in a node, including holes | |
4311 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() | |
4312 | */ | |
4313 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4314 | unsigned long zone_type, | |
4315 | unsigned long *ignored) | |
4316 | { | |
4317 | unsigned long node_start_pfn, node_end_pfn; | |
4318 | unsigned long zone_start_pfn, zone_end_pfn; | |
4319 | ||
4320 | /* Get the start and end of the node and zone */ | |
4321 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
4322 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; | |
4323 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; | |
4324 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4325 | node_start_pfn, node_end_pfn, | |
4326 | &zone_start_pfn, &zone_end_pfn); | |
4327 | ||
4328 | /* Check that this node has pages within the zone's required range */ | |
4329 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) | |
4330 | return 0; | |
4331 | ||
4332 | /* Move the zone boundaries inside the node if necessary */ | |
4333 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); | |
4334 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); | |
4335 | ||
4336 | /* Return the spanned pages */ | |
4337 | return zone_end_pfn - zone_start_pfn; | |
4338 | } | |
4339 | ||
4340 | /* | |
4341 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, | |
4342 | * then all holes in the requested range will be accounted for. | |
4343 | */ | |
4344 | unsigned long __meminit __absent_pages_in_range(int nid, | |
4345 | unsigned long range_start_pfn, | |
4346 | unsigned long range_end_pfn) | |
4347 | { | |
4348 | unsigned long nr_absent = range_end_pfn - range_start_pfn; | |
4349 | unsigned long start_pfn, end_pfn; | |
4350 | int i; | |
4351 | ||
4352 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
4353 | start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); | |
4354 | end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); | |
4355 | nr_absent -= end_pfn - start_pfn; | |
4356 | } | |
4357 | return nr_absent; | |
4358 | } | |
4359 | ||
4360 | /** | |
4361 | * absent_pages_in_range - Return number of page frames in holes within a range | |
4362 | * @start_pfn: The start PFN to start searching for holes | |
4363 | * @end_pfn: The end PFN to stop searching for holes | |
4364 | * | |
4365 | * It returns the number of pages frames in memory holes within a range. | |
4366 | */ | |
4367 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, | |
4368 | unsigned long end_pfn) | |
4369 | { | |
4370 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); | |
4371 | } | |
4372 | ||
4373 | /* Return the number of page frames in holes in a zone on a node */ | |
4374 | static unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4375 | unsigned long zone_type, | |
4376 | unsigned long *ignored) | |
4377 | { | |
4378 | unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; | |
4379 | unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; | |
4380 | unsigned long node_start_pfn, node_end_pfn; | |
4381 | unsigned long zone_start_pfn, zone_end_pfn; | |
4382 | ||
4383 | get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); | |
4384 | zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); | |
4385 | zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); | |
4386 | ||
4387 | adjust_zone_range_for_zone_movable(nid, zone_type, | |
4388 | node_start_pfn, node_end_pfn, | |
4389 | &zone_start_pfn, &zone_end_pfn); | |
4390 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); | |
4391 | } | |
4392 | ||
4393 | /** | |
4394 | * sanitize_zone_movable_limit - Sanitize the zone_movable_limit array. | |
4395 | * | |
4396 | * zone_movable_limit is initialized as 0. This function will try to get | |
4397 | * the first ZONE_MOVABLE pfn of each node from movablemem_map, and | |
4398 | * assigne them to zone_movable_limit. | |
4399 | * zone_movable_limit[nid] == 0 means no limit for the node. | |
4400 | * | |
4401 | * Note: Each range is represented as [start_pfn, end_pfn) | |
4402 | */ | |
4403 | static void __meminit sanitize_zone_movable_limit(void) | |
4404 | { | |
4405 | int map_pos = 0, i, nid; | |
4406 | unsigned long start_pfn, end_pfn; | |
4407 | ||
4408 | if (!movablemem_map.nr_map) | |
4409 | return; | |
4410 | ||
4411 | /* Iterate all ranges from minimum to maximum */ | |
4412 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { | |
4413 | /* | |
4414 | * If we have found lowest pfn of ZONE_MOVABLE of the node | |
4415 | * specified by user, just go on to check next range. | |
4416 | */ | |
4417 | if (zone_movable_limit[nid]) | |
4418 | continue; | |
4419 | ||
4420 | #ifdef CONFIG_ZONE_DMA | |
4421 | /* Skip DMA memory. */ | |
4422 | if (start_pfn < arch_zone_highest_possible_pfn[ZONE_DMA]) | |
4423 | start_pfn = arch_zone_highest_possible_pfn[ZONE_DMA]; | |
4424 | #endif | |
4425 | ||
4426 | #ifdef CONFIG_ZONE_DMA32 | |
4427 | /* Skip DMA32 memory. */ | |
4428 | if (start_pfn < arch_zone_highest_possible_pfn[ZONE_DMA32]) | |
4429 | start_pfn = arch_zone_highest_possible_pfn[ZONE_DMA32]; | |
4430 | #endif | |
4431 | ||
4432 | #ifdef CONFIG_HIGHMEM | |
4433 | /* Skip lowmem if ZONE_MOVABLE is highmem. */ | |
4434 | if (zone_movable_is_highmem() && | |
4435 | start_pfn < arch_zone_lowest_possible_pfn[ZONE_HIGHMEM]) | |
4436 | start_pfn = arch_zone_lowest_possible_pfn[ZONE_HIGHMEM]; | |
4437 | #endif | |
4438 | ||
4439 | if (start_pfn >= end_pfn) | |
4440 | continue; | |
4441 | ||
4442 | while (map_pos < movablemem_map.nr_map) { | |
4443 | if (end_pfn <= movablemem_map.map[map_pos].start_pfn) | |
4444 | break; | |
4445 | ||
4446 | if (start_pfn >= movablemem_map.map[map_pos].end_pfn) { | |
4447 | map_pos++; | |
4448 | continue; | |
4449 | } | |
4450 | ||
4451 | /* | |
4452 | * The start_pfn of ZONE_MOVABLE is either the minimum | |
4453 | * pfn specified by movablemem_map, or 0, which means | |
4454 | * the node has no ZONE_MOVABLE. | |
4455 | */ | |
4456 | zone_movable_limit[nid] = max(start_pfn, | |
4457 | movablemem_map.map[map_pos].start_pfn); | |
4458 | ||
4459 | break; | |
4460 | } | |
4461 | } | |
4462 | } | |
4463 | ||
4464 | #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4465 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, | |
4466 | unsigned long zone_type, | |
4467 | unsigned long *zones_size) | |
4468 | { | |
4469 | return zones_size[zone_type]; | |
4470 | } | |
4471 | ||
4472 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, | |
4473 | unsigned long zone_type, | |
4474 | unsigned long *zholes_size) | |
4475 | { | |
4476 | if (!zholes_size) | |
4477 | return 0; | |
4478 | ||
4479 | return zholes_size[zone_type]; | |
4480 | } | |
4481 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4482 | ||
4483 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, | |
4484 | unsigned long *zones_size, unsigned long *zholes_size) | |
4485 | { | |
4486 | unsigned long realtotalpages, totalpages = 0; | |
4487 | enum zone_type i; | |
4488 | ||
4489 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4490 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, | |
4491 | zones_size); | |
4492 | pgdat->node_spanned_pages = totalpages; | |
4493 | ||
4494 | realtotalpages = totalpages; | |
4495 | for (i = 0; i < MAX_NR_ZONES; i++) | |
4496 | realtotalpages -= | |
4497 | zone_absent_pages_in_node(pgdat->node_id, i, | |
4498 | zholes_size); | |
4499 | pgdat->node_present_pages = realtotalpages; | |
4500 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, | |
4501 | realtotalpages); | |
4502 | } | |
4503 | ||
4504 | #ifndef CONFIG_SPARSEMEM | |
4505 | /* | |
4506 | * Calculate the size of the zone->blockflags rounded to an unsigned long | |
4507 | * Start by making sure zonesize is a multiple of pageblock_order by rounding | |
4508 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally | |
4509 | * round what is now in bits to nearest long in bits, then return it in | |
4510 | * bytes. | |
4511 | */ | |
4512 | static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) | |
4513 | { | |
4514 | unsigned long usemapsize; | |
4515 | ||
4516 | zonesize += zone_start_pfn & (pageblock_nr_pages-1); | |
4517 | usemapsize = roundup(zonesize, pageblock_nr_pages); | |
4518 | usemapsize = usemapsize >> pageblock_order; | |
4519 | usemapsize *= NR_PAGEBLOCK_BITS; | |
4520 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); | |
4521 | ||
4522 | return usemapsize / 8; | |
4523 | } | |
4524 | ||
4525 | static void __init setup_usemap(struct pglist_data *pgdat, | |
4526 | struct zone *zone, | |
4527 | unsigned long zone_start_pfn, | |
4528 | unsigned long zonesize) | |
4529 | { | |
4530 | unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); | |
4531 | zone->pageblock_flags = NULL; | |
4532 | if (usemapsize) | |
4533 | zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, | |
4534 | usemapsize); | |
4535 | } | |
4536 | #else | |
4537 | static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, | |
4538 | unsigned long zone_start_pfn, unsigned long zonesize) {} | |
4539 | #endif /* CONFIG_SPARSEMEM */ | |
4540 | ||
4541 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE | |
4542 | ||
4543 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ | |
4544 | void __init set_pageblock_order(void) | |
4545 | { | |
4546 | unsigned int order; | |
4547 | ||
4548 | /* Check that pageblock_nr_pages has not already been setup */ | |
4549 | if (pageblock_order) | |
4550 | return; | |
4551 | ||
4552 | if (HPAGE_SHIFT > PAGE_SHIFT) | |
4553 | order = HUGETLB_PAGE_ORDER; | |
4554 | else | |
4555 | order = MAX_ORDER - 1; | |
4556 | ||
4557 | /* | |
4558 | * Assume the largest contiguous order of interest is a huge page. | |
4559 | * This value may be variable depending on boot parameters on IA64 and | |
4560 | * powerpc. | |
4561 | */ | |
4562 | pageblock_order = order; | |
4563 | } | |
4564 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4565 | ||
4566 | /* | |
4567 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() | |
4568 | * is unused as pageblock_order is set at compile-time. See | |
4569 | * include/linux/pageblock-flags.h for the values of pageblock_order based on | |
4570 | * the kernel config | |
4571 | */ | |
4572 | void __init set_pageblock_order(void) | |
4573 | { | |
4574 | } | |
4575 | ||
4576 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ | |
4577 | ||
4578 | static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, | |
4579 | unsigned long present_pages) | |
4580 | { | |
4581 | unsigned long pages = spanned_pages; | |
4582 | ||
4583 | /* | |
4584 | * Provide a more accurate estimation if there are holes within | |
4585 | * the zone and SPARSEMEM is in use. If there are holes within the | |
4586 | * zone, each populated memory region may cost us one or two extra | |
4587 | * memmap pages due to alignment because memmap pages for each | |
4588 | * populated regions may not naturally algined on page boundary. | |
4589 | * So the (present_pages >> 4) heuristic is a tradeoff for that. | |
4590 | */ | |
4591 | if (spanned_pages > present_pages + (present_pages >> 4) && | |
4592 | IS_ENABLED(CONFIG_SPARSEMEM)) | |
4593 | pages = present_pages; | |
4594 | ||
4595 | return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; | |
4596 | } | |
4597 | ||
4598 | /* | |
4599 | * Set up the zone data structures: | |
4600 | * - mark all pages reserved | |
4601 | * - mark all memory queues empty | |
4602 | * - clear the memory bitmaps | |
4603 | * | |
4604 | * NOTE: pgdat should get zeroed by caller. | |
4605 | */ | |
4606 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, | |
4607 | unsigned long *zones_size, unsigned long *zholes_size) | |
4608 | { | |
4609 | enum zone_type j; | |
4610 | int nid = pgdat->node_id; | |
4611 | unsigned long zone_start_pfn = pgdat->node_start_pfn; | |
4612 | int ret; | |
4613 | ||
4614 | pgdat_resize_init(pgdat); | |
4615 | #ifdef CONFIG_NUMA_BALANCING | |
4616 | spin_lock_init(&pgdat->numabalancing_migrate_lock); | |
4617 | pgdat->numabalancing_migrate_nr_pages = 0; | |
4618 | pgdat->numabalancing_migrate_next_window = jiffies; | |
4619 | #endif | |
4620 | init_waitqueue_head(&pgdat->kswapd_wait); | |
4621 | init_waitqueue_head(&pgdat->pfmemalloc_wait); | |
4622 | pgdat_page_cgroup_init(pgdat); | |
4623 | ||
4624 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
4625 | struct zone *zone = pgdat->node_zones + j; | |
4626 | unsigned long size, realsize, freesize, memmap_pages; | |
4627 | ||
4628 | size = zone_spanned_pages_in_node(nid, j, zones_size); | |
4629 | realsize = freesize = size - zone_absent_pages_in_node(nid, j, | |
4630 | zholes_size); | |
4631 | ||
4632 | /* | |
4633 | * Adjust freesize so that it accounts for how much memory | |
4634 | * is used by this zone for memmap. This affects the watermark | |
4635 | * and per-cpu initialisations | |
4636 | */ | |
4637 | memmap_pages = calc_memmap_size(size, realsize); | |
4638 | if (freesize >= memmap_pages) { | |
4639 | freesize -= memmap_pages; | |
4640 | if (memmap_pages) | |
4641 | printk(KERN_DEBUG | |
4642 | " %s zone: %lu pages used for memmap\n", | |
4643 | zone_names[j], memmap_pages); | |
4644 | } else | |
4645 | printk(KERN_WARNING | |
4646 | " %s zone: %lu pages exceeds freesize %lu\n", | |
4647 | zone_names[j], memmap_pages, freesize); | |
4648 | ||
4649 | /* Account for reserved pages */ | |
4650 | if (j == 0 && freesize > dma_reserve) { | |
4651 | freesize -= dma_reserve; | |
4652 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", | |
4653 | zone_names[0], dma_reserve); | |
4654 | } | |
4655 | ||
4656 | if (!is_highmem_idx(j)) | |
4657 | nr_kernel_pages += freesize; | |
4658 | /* Charge for highmem memmap if there are enough kernel pages */ | |
4659 | else if (nr_kernel_pages > memmap_pages * 2) | |
4660 | nr_kernel_pages -= memmap_pages; | |
4661 | nr_all_pages += freesize; | |
4662 | ||
4663 | zone->spanned_pages = size; | |
4664 | zone->present_pages = realsize; | |
4665 | /* | |
4666 | * Set an approximate value for lowmem here, it will be adjusted | |
4667 | * when the bootmem allocator frees pages into the buddy system. | |
4668 | * And all highmem pages will be managed by the buddy system. | |
4669 | */ | |
4670 | zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; | |
4671 | #ifdef CONFIG_NUMA | |
4672 | zone->node = nid; | |
4673 | zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) | |
4674 | / 100; | |
4675 | zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; | |
4676 | #endif | |
4677 | zone->name = zone_names[j]; | |
4678 | spin_lock_init(&zone->lock); | |
4679 | spin_lock_init(&zone->lru_lock); | |
4680 | zone_seqlock_init(zone); | |
4681 | zone->zone_pgdat = pgdat; | |
4682 | ||
4683 | zone_pcp_init(zone); | |
4684 | lruvec_init(&zone->lruvec); | |
4685 | if (!size) | |
4686 | continue; | |
4687 | ||
4688 | set_pageblock_order(); | |
4689 | setup_usemap(pgdat, zone, zone_start_pfn, size); | |
4690 | ret = init_currently_empty_zone(zone, zone_start_pfn, | |
4691 | size, MEMMAP_EARLY); | |
4692 | BUG_ON(ret); | |
4693 | memmap_init(size, nid, j, zone_start_pfn); | |
4694 | zone_start_pfn += size; | |
4695 | } | |
4696 | } | |
4697 | ||
4698 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) | |
4699 | { | |
4700 | /* Skip empty nodes */ | |
4701 | if (!pgdat->node_spanned_pages) | |
4702 | return; | |
4703 | ||
4704 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
4705 | /* ia64 gets its own node_mem_map, before this, without bootmem */ | |
4706 | if (!pgdat->node_mem_map) { | |
4707 | unsigned long size, start, end; | |
4708 | struct page *map; | |
4709 | ||
4710 | /* | |
4711 | * The zone's endpoints aren't required to be MAX_ORDER | |
4712 | * aligned but the node_mem_map endpoints must be in order | |
4713 | * for the buddy allocator to function correctly. | |
4714 | */ | |
4715 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); | |
4716 | end = pgdat_end_pfn(pgdat); | |
4717 | end = ALIGN(end, MAX_ORDER_NR_PAGES); | |
4718 | size = (end - start) * sizeof(struct page); | |
4719 | map = alloc_remap(pgdat->node_id, size); | |
4720 | if (!map) | |
4721 | map = alloc_bootmem_node_nopanic(pgdat, size); | |
4722 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); | |
4723 | } | |
4724 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
4725 | /* | |
4726 | * With no DISCONTIG, the global mem_map is just set as node 0's | |
4727 | */ | |
4728 | if (pgdat == NODE_DATA(0)) { | |
4729 | mem_map = NODE_DATA(0)->node_mem_map; | |
4730 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4731 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) | |
4732 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); | |
4733 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
4734 | } | |
4735 | #endif | |
4736 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ | |
4737 | } | |
4738 | ||
4739 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, | |
4740 | unsigned long node_start_pfn, unsigned long *zholes_size) | |
4741 | { | |
4742 | pg_data_t *pgdat = NODE_DATA(nid); | |
4743 | ||
4744 | /* pg_data_t should be reset to zero when it's allocated */ | |
4745 | WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); | |
4746 | ||
4747 | pgdat->node_id = nid; | |
4748 | pgdat->node_start_pfn = node_start_pfn; | |
4749 | init_zone_allows_reclaim(nid); | |
4750 | calculate_node_totalpages(pgdat, zones_size, zholes_size); | |
4751 | ||
4752 | alloc_node_mem_map(pgdat); | |
4753 | #ifdef CONFIG_FLAT_NODE_MEM_MAP | |
4754 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", | |
4755 | nid, (unsigned long)pgdat, | |
4756 | (unsigned long)pgdat->node_mem_map); | |
4757 | #endif | |
4758 | ||
4759 | free_area_init_core(pgdat, zones_size, zholes_size); | |
4760 | } | |
4761 | ||
4762 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP | |
4763 | ||
4764 | #if MAX_NUMNODES > 1 | |
4765 | /* | |
4766 | * Figure out the number of possible node ids. | |
4767 | */ | |
4768 | static void __init setup_nr_node_ids(void) | |
4769 | { | |
4770 | unsigned int node; | |
4771 | unsigned int highest = 0; | |
4772 | ||
4773 | for_each_node_mask(node, node_possible_map) | |
4774 | highest = node; | |
4775 | nr_node_ids = highest + 1; | |
4776 | } | |
4777 | #else | |
4778 | static inline void setup_nr_node_ids(void) | |
4779 | { | |
4780 | } | |
4781 | #endif | |
4782 | ||
4783 | /** | |
4784 | * node_map_pfn_alignment - determine the maximum internode alignment | |
4785 | * | |
4786 | * This function should be called after node map is populated and sorted. | |
4787 | * It calculates the maximum power of two alignment which can distinguish | |
4788 | * all the nodes. | |
4789 | * | |
4790 | * For example, if all nodes are 1GiB and aligned to 1GiB, the return value | |
4791 | * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the | |
4792 | * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is | |
4793 | * shifted, 1GiB is enough and this function will indicate so. | |
4794 | * | |
4795 | * This is used to test whether pfn -> nid mapping of the chosen memory | |
4796 | * model has fine enough granularity to avoid incorrect mapping for the | |
4797 | * populated node map. | |
4798 | * | |
4799 | * Returns the determined alignment in pfn's. 0 if there is no alignment | |
4800 | * requirement (single node). | |
4801 | */ | |
4802 | unsigned long __init node_map_pfn_alignment(void) | |
4803 | { | |
4804 | unsigned long accl_mask = 0, last_end = 0; | |
4805 | unsigned long start, end, mask; | |
4806 | int last_nid = -1; | |
4807 | int i, nid; | |
4808 | ||
4809 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { | |
4810 | if (!start || last_nid < 0 || last_nid == nid) { | |
4811 | last_nid = nid; | |
4812 | last_end = end; | |
4813 | continue; | |
4814 | } | |
4815 | ||
4816 | /* | |
4817 | * Start with a mask granular enough to pin-point to the | |
4818 | * start pfn and tick off bits one-by-one until it becomes | |
4819 | * too coarse to separate the current node from the last. | |
4820 | */ | |
4821 | mask = ~((1 << __ffs(start)) - 1); | |
4822 | while (mask && last_end <= (start & (mask << 1))) | |
4823 | mask <<= 1; | |
4824 | ||
4825 | /* accumulate all internode masks */ | |
4826 | accl_mask |= mask; | |
4827 | } | |
4828 | ||
4829 | /* convert mask to number of pages */ | |
4830 | return ~accl_mask + 1; | |
4831 | } | |
4832 | ||
4833 | /* Find the lowest pfn for a node */ | |
4834 | static unsigned long __init find_min_pfn_for_node(int nid) | |
4835 | { | |
4836 | unsigned long min_pfn = ULONG_MAX; | |
4837 | unsigned long start_pfn; | |
4838 | int i; | |
4839 | ||
4840 | for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) | |
4841 | min_pfn = min(min_pfn, start_pfn); | |
4842 | ||
4843 | if (min_pfn == ULONG_MAX) { | |
4844 | printk(KERN_WARNING | |
4845 | "Could not find start_pfn for node %d\n", nid); | |
4846 | return 0; | |
4847 | } | |
4848 | ||
4849 | return min_pfn; | |
4850 | } | |
4851 | ||
4852 | /** | |
4853 | * find_min_pfn_with_active_regions - Find the minimum PFN registered | |
4854 | * | |
4855 | * It returns the minimum PFN based on information provided via | |
4856 | * add_active_range(). | |
4857 | */ | |
4858 | unsigned long __init find_min_pfn_with_active_regions(void) | |
4859 | { | |
4860 | return find_min_pfn_for_node(MAX_NUMNODES); | |
4861 | } | |
4862 | ||
4863 | /* | |
4864 | * early_calculate_totalpages() | |
4865 | * Sum pages in active regions for movable zone. | |
4866 | * Populate N_MEMORY for calculating usable_nodes. | |
4867 | */ | |
4868 | static unsigned long __init early_calculate_totalpages(void) | |
4869 | { | |
4870 | unsigned long totalpages = 0; | |
4871 | unsigned long start_pfn, end_pfn; | |
4872 | int i, nid; | |
4873 | ||
4874 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { | |
4875 | unsigned long pages = end_pfn - start_pfn; | |
4876 | ||
4877 | totalpages += pages; | |
4878 | if (pages) | |
4879 | node_set_state(nid, N_MEMORY); | |
4880 | } | |
4881 | return totalpages; | |
4882 | } | |
4883 | ||
4884 | /* | |
4885 | * Find the PFN the Movable zone begins in each node. Kernel memory | |
4886 | * is spread evenly between nodes as long as the nodes have enough | |
4887 | * memory. When they don't, some nodes will have more kernelcore than | |
4888 | * others | |
4889 | */ | |
4890 | static void __init find_zone_movable_pfns_for_nodes(void) | |
4891 | { | |
4892 | int i, nid; | |
4893 | unsigned long usable_startpfn; | |
4894 | unsigned long kernelcore_node, kernelcore_remaining; | |
4895 | /* save the state before borrow the nodemask */ | |
4896 | nodemask_t saved_node_state = node_states[N_MEMORY]; | |
4897 | unsigned long totalpages = early_calculate_totalpages(); | |
4898 | int usable_nodes = nodes_weight(node_states[N_MEMORY]); | |
4899 | ||
4900 | /* | |
4901 | * If movablecore was specified, calculate what size of | |
4902 | * kernelcore that corresponds so that memory usable for | |
4903 | * any allocation type is evenly spread. If both kernelcore | |
4904 | * and movablecore are specified, then the value of kernelcore | |
4905 | * will be used for required_kernelcore if it's greater than | |
4906 | * what movablecore would have allowed. | |
4907 | */ | |
4908 | if (required_movablecore) { | |
4909 | unsigned long corepages; | |
4910 | ||
4911 | /* | |
4912 | * Round-up so that ZONE_MOVABLE is at least as large as what | |
4913 | * was requested by the user | |
4914 | */ | |
4915 | required_movablecore = | |
4916 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); | |
4917 | corepages = totalpages - required_movablecore; | |
4918 | ||
4919 | required_kernelcore = max(required_kernelcore, corepages); | |
4920 | } | |
4921 | ||
4922 | /* | |
4923 | * If neither kernelcore/movablecore nor movablemem_map is specified, | |
4924 | * there is no ZONE_MOVABLE. But if movablemem_map is specified, the | |
4925 | * start pfn of ZONE_MOVABLE has been stored in zone_movable_limit[]. | |
4926 | */ | |
4927 | if (!required_kernelcore) { | |
4928 | if (movablemem_map.nr_map) | |
4929 | memcpy(zone_movable_pfn, zone_movable_limit, | |
4930 | sizeof(zone_movable_pfn)); | |
4931 | goto out; | |
4932 | } | |
4933 | ||
4934 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ | |
4935 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; | |
4936 | ||
4937 | restart: | |
4938 | /* Spread kernelcore memory as evenly as possible throughout nodes */ | |
4939 | kernelcore_node = required_kernelcore / usable_nodes; | |
4940 | for_each_node_state(nid, N_MEMORY) { | |
4941 | unsigned long start_pfn, end_pfn; | |
4942 | ||
4943 | /* | |
4944 | * Recalculate kernelcore_node if the division per node | |
4945 | * now exceeds what is necessary to satisfy the requested | |
4946 | * amount of memory for the kernel | |
4947 | */ | |
4948 | if (required_kernelcore < kernelcore_node) | |
4949 | kernelcore_node = required_kernelcore / usable_nodes; | |
4950 | ||
4951 | /* | |
4952 | * As the map is walked, we track how much memory is usable | |
4953 | * by the kernel using kernelcore_remaining. When it is | |
4954 | * 0, the rest of the node is usable by ZONE_MOVABLE | |
4955 | */ | |
4956 | kernelcore_remaining = kernelcore_node; | |
4957 | ||
4958 | /* Go through each range of PFNs within this node */ | |
4959 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { | |
4960 | unsigned long size_pages; | |
4961 | ||
4962 | /* | |
4963 | * Find more memory for kernelcore in | |
4964 | * [zone_movable_pfn[nid], zone_movable_limit[nid]). | |
4965 | */ | |
4966 | start_pfn = max(start_pfn, zone_movable_pfn[nid]); | |
4967 | if (start_pfn >= end_pfn) | |
4968 | continue; | |
4969 | ||
4970 | if (zone_movable_limit[nid]) { | |
4971 | end_pfn = min(end_pfn, zone_movable_limit[nid]); | |
4972 | /* No range left for kernelcore in this node */ | |
4973 | if (start_pfn >= end_pfn) { | |
4974 | zone_movable_pfn[nid] = | |
4975 | zone_movable_limit[nid]; | |
4976 | break; | |
4977 | } | |
4978 | } | |
4979 | ||
4980 | /* Account for what is only usable for kernelcore */ | |
4981 | if (start_pfn < usable_startpfn) { | |
4982 | unsigned long kernel_pages; | |
4983 | kernel_pages = min(end_pfn, usable_startpfn) | |
4984 | - start_pfn; | |
4985 | ||
4986 | kernelcore_remaining -= min(kernel_pages, | |
4987 | kernelcore_remaining); | |
4988 | required_kernelcore -= min(kernel_pages, | |
4989 | required_kernelcore); | |
4990 | ||
4991 | /* Continue if range is now fully accounted */ | |
4992 | if (end_pfn <= usable_startpfn) { | |
4993 | ||
4994 | /* | |
4995 | * Push zone_movable_pfn to the end so | |
4996 | * that if we have to rebalance | |
4997 | * kernelcore across nodes, we will | |
4998 | * not double account here | |
4999 | */ | |
5000 | zone_movable_pfn[nid] = end_pfn; | |
5001 | continue; | |
5002 | } | |
5003 | start_pfn = usable_startpfn; | |
5004 | } | |
5005 | ||
5006 | /* | |
5007 | * The usable PFN range for ZONE_MOVABLE is from | |
5008 | * start_pfn->end_pfn. Calculate size_pages as the | |
5009 | * number of pages used as kernelcore | |
5010 | */ | |
5011 | size_pages = end_pfn - start_pfn; | |
5012 | if (size_pages > kernelcore_remaining) | |
5013 | size_pages = kernelcore_remaining; | |
5014 | zone_movable_pfn[nid] = start_pfn + size_pages; | |
5015 | ||
5016 | /* | |
5017 | * Some kernelcore has been met, update counts and | |
5018 | * break if the kernelcore for this node has been | |
5019 | * satisified | |
5020 | */ | |
5021 | required_kernelcore -= min(required_kernelcore, | |
5022 | size_pages); | |
5023 | kernelcore_remaining -= size_pages; | |
5024 | if (!kernelcore_remaining) | |
5025 | break; | |
5026 | } | |
5027 | } | |
5028 | ||
5029 | /* | |
5030 | * If there is still required_kernelcore, we do another pass with one | |
5031 | * less node in the count. This will push zone_movable_pfn[nid] further | |
5032 | * along on the nodes that still have memory until kernelcore is | |
5033 | * satisified | |
5034 | */ | |
5035 | usable_nodes--; | |
5036 | if (usable_nodes && required_kernelcore > usable_nodes) | |
5037 | goto restart; | |
5038 | ||
5039 | out: | |
5040 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ | |
5041 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
5042 | zone_movable_pfn[nid] = | |
5043 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); | |
5044 | ||
5045 | /* restore the node_state */ | |
5046 | node_states[N_MEMORY] = saved_node_state; | |
5047 | } | |
5048 | ||
5049 | /* Any regular or high memory on that node ? */ | |
5050 | static void check_for_memory(pg_data_t *pgdat, int nid) | |
5051 | { | |
5052 | enum zone_type zone_type; | |
5053 | ||
5054 | if (N_MEMORY == N_NORMAL_MEMORY) | |
5055 | return; | |
5056 | ||
5057 | for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { | |
5058 | struct zone *zone = &pgdat->node_zones[zone_type]; | |
5059 | if (zone->present_pages) { | |
5060 | node_set_state(nid, N_HIGH_MEMORY); | |
5061 | if (N_NORMAL_MEMORY != N_HIGH_MEMORY && | |
5062 | zone_type <= ZONE_NORMAL) | |
5063 | node_set_state(nid, N_NORMAL_MEMORY); | |
5064 | break; | |
5065 | } | |
5066 | } | |
5067 | } | |
5068 | ||
5069 | /** | |
5070 | * free_area_init_nodes - Initialise all pg_data_t and zone data | |
5071 | * @max_zone_pfn: an array of max PFNs for each zone | |
5072 | * | |
5073 | * This will call free_area_init_node() for each active node in the system. | |
5074 | * Using the page ranges provided by add_active_range(), the size of each | |
5075 | * zone in each node and their holes is calculated. If the maximum PFN | |
5076 | * between two adjacent zones match, it is assumed that the zone is empty. | |
5077 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed | |
5078 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone | |
5079 | * starts where the previous one ended. For example, ZONE_DMA32 starts | |
5080 | * at arch_max_dma_pfn. | |
5081 | */ | |
5082 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) | |
5083 | { | |
5084 | unsigned long start_pfn, end_pfn; | |
5085 | int i, nid; | |
5086 | ||
5087 | /* Record where the zone boundaries are */ | |
5088 | memset(arch_zone_lowest_possible_pfn, 0, | |
5089 | sizeof(arch_zone_lowest_possible_pfn)); | |
5090 | memset(arch_zone_highest_possible_pfn, 0, | |
5091 | sizeof(arch_zone_highest_possible_pfn)); | |
5092 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); | |
5093 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; | |
5094 | for (i = 1; i < MAX_NR_ZONES; i++) { | |
5095 | if (i == ZONE_MOVABLE) | |
5096 | continue; | |
5097 | arch_zone_lowest_possible_pfn[i] = | |
5098 | arch_zone_highest_possible_pfn[i-1]; | |
5099 | arch_zone_highest_possible_pfn[i] = | |
5100 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); | |
5101 | } | |
5102 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; | |
5103 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; | |
5104 | ||
5105 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ | |
5106 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); | |
5107 | find_usable_zone_for_movable(); | |
5108 | sanitize_zone_movable_limit(); | |
5109 | find_zone_movable_pfns_for_nodes(); | |
5110 | ||
5111 | /* Print out the zone ranges */ | |
5112 | printk("Zone ranges:\n"); | |
5113 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
5114 | if (i == ZONE_MOVABLE) | |
5115 | continue; | |
5116 | printk(KERN_CONT " %-8s ", zone_names[i]); | |
5117 | if (arch_zone_lowest_possible_pfn[i] == | |
5118 | arch_zone_highest_possible_pfn[i]) | |
5119 | printk(KERN_CONT "empty\n"); | |
5120 | else | |
5121 | printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n", | |
5122 | arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT, | |
5123 | (arch_zone_highest_possible_pfn[i] | |
5124 | << PAGE_SHIFT) - 1); | |
5125 | } | |
5126 | ||
5127 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ | |
5128 | printk("Movable zone start for each node\n"); | |
5129 | for (i = 0; i < MAX_NUMNODES; i++) { | |
5130 | if (zone_movable_pfn[i]) | |
5131 | printk(" Node %d: %#010lx\n", i, | |
5132 | zone_movable_pfn[i] << PAGE_SHIFT); | |
5133 | } | |
5134 | ||
5135 | /* Print out the early node map */ | |
5136 | printk("Early memory node ranges\n"); | |
5137 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) | |
5138 | printk(" node %3d: [mem %#010lx-%#010lx]\n", nid, | |
5139 | start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1); | |
5140 | ||
5141 | /* Initialise every node */ | |
5142 | mminit_verify_pageflags_layout(); | |
5143 | setup_nr_node_ids(); | |
5144 | for_each_online_node(nid) { | |
5145 | pg_data_t *pgdat = NODE_DATA(nid); | |
5146 | free_area_init_node(nid, NULL, | |
5147 | find_min_pfn_for_node(nid), NULL); | |
5148 | ||
5149 | /* Any memory on that node */ | |
5150 | if (pgdat->node_present_pages) | |
5151 | node_set_state(nid, N_MEMORY); | |
5152 | check_for_memory(pgdat, nid); | |
5153 | } | |
5154 | } | |
5155 | ||
5156 | static int __init cmdline_parse_core(char *p, unsigned long *core) | |
5157 | { | |
5158 | unsigned long long coremem; | |
5159 | if (!p) | |
5160 | return -EINVAL; | |
5161 | ||
5162 | coremem = memparse(p, &p); | |
5163 | *core = coremem >> PAGE_SHIFT; | |
5164 | ||
5165 | /* Paranoid check that UL is enough for the coremem value */ | |
5166 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); | |
5167 | ||
5168 | return 0; | |
5169 | } | |
5170 | ||
5171 | /* | |
5172 | * kernelcore=size sets the amount of memory for use for allocations that | |
5173 | * cannot be reclaimed or migrated. | |
5174 | */ | |
5175 | static int __init cmdline_parse_kernelcore(char *p) | |
5176 | { | |
5177 | return cmdline_parse_core(p, &required_kernelcore); | |
5178 | } | |
5179 | ||
5180 | /* | |
5181 | * movablecore=size sets the amount of memory for use for allocations that | |
5182 | * can be reclaimed or migrated. | |
5183 | */ | |
5184 | static int __init cmdline_parse_movablecore(char *p) | |
5185 | { | |
5186 | return cmdline_parse_core(p, &required_movablecore); | |
5187 | } | |
5188 | ||
5189 | early_param("kernelcore", cmdline_parse_kernelcore); | |
5190 | early_param("movablecore", cmdline_parse_movablecore); | |
5191 | ||
5192 | /** | |
5193 | * movablemem_map_overlap() - Check if a range overlaps movablemem_map.map[]. | |
5194 | * @start_pfn: start pfn of the range to be checked | |
5195 | * @end_pfn: end pfn of the range to be checked (exclusive) | |
5196 | * | |
5197 | * This function checks if a given memory range [start_pfn, end_pfn) overlaps | |
5198 | * the movablemem_map.map[] array. | |
5199 | * | |
5200 | * Return: index of the first overlapped element in movablemem_map.map[] | |
5201 | * or -1 if they don't overlap each other. | |
5202 | */ | |
5203 | int __init movablemem_map_overlap(unsigned long start_pfn, | |
5204 | unsigned long end_pfn) | |
5205 | { | |
5206 | int overlap; | |
5207 | ||
5208 | if (!movablemem_map.nr_map) | |
5209 | return -1; | |
5210 | ||
5211 | for (overlap = 0; overlap < movablemem_map.nr_map; overlap++) | |
5212 | if (start_pfn < movablemem_map.map[overlap].end_pfn) | |
5213 | break; | |
5214 | ||
5215 | if (overlap == movablemem_map.nr_map || | |
5216 | end_pfn <= movablemem_map.map[overlap].start_pfn) | |
5217 | return -1; | |
5218 | ||
5219 | return overlap; | |
5220 | } | |
5221 | ||
5222 | /** | |
5223 | * insert_movablemem_map - Insert a memory range in to movablemem_map.map. | |
5224 | * @start_pfn: start pfn of the range | |
5225 | * @end_pfn: end pfn of the range | |
5226 | * | |
5227 | * This function will also merge the overlapped ranges, and sort the array | |
5228 | * by start_pfn in monotonic increasing order. | |
5229 | */ | |
5230 | void __init insert_movablemem_map(unsigned long start_pfn, | |
5231 | unsigned long end_pfn) | |
5232 | { | |
5233 | int pos, overlap; | |
5234 | ||
5235 | /* | |
5236 | * pos will be at the 1st overlapped range, or the position | |
5237 | * where the element should be inserted. | |
5238 | */ | |
5239 | for (pos = 0; pos < movablemem_map.nr_map; pos++) | |
5240 | if (start_pfn <= movablemem_map.map[pos].end_pfn) | |
5241 | break; | |
5242 | ||
5243 | /* If there is no overlapped range, just insert the element. */ | |
5244 | if (pos == movablemem_map.nr_map || | |
5245 | end_pfn < movablemem_map.map[pos].start_pfn) { | |
5246 | /* | |
5247 | * If pos is not the end of array, we need to move all | |
5248 | * the rest elements backward. | |
5249 | */ | |
5250 | if (pos < movablemem_map.nr_map) | |
5251 | memmove(&movablemem_map.map[pos+1], | |
5252 | &movablemem_map.map[pos], | |
5253 | sizeof(struct movablemem_entry) * | |
5254 | (movablemem_map.nr_map - pos)); | |
5255 | movablemem_map.map[pos].start_pfn = start_pfn; | |
5256 | movablemem_map.map[pos].end_pfn = end_pfn; | |
5257 | movablemem_map.nr_map++; | |
5258 | return; | |
5259 | } | |
5260 | ||
5261 | /* overlap will be at the last overlapped range */ | |
5262 | for (overlap = pos + 1; overlap < movablemem_map.nr_map; overlap++) | |
5263 | if (end_pfn < movablemem_map.map[overlap].start_pfn) | |
5264 | break; | |
5265 | ||
5266 | /* | |
5267 | * If there are more ranges overlapped, we need to merge them, | |
5268 | * and move the rest elements forward. | |
5269 | */ | |
5270 | overlap--; | |
5271 | movablemem_map.map[pos].start_pfn = min(start_pfn, | |
5272 | movablemem_map.map[pos].start_pfn); | |
5273 | movablemem_map.map[pos].end_pfn = max(end_pfn, | |
5274 | movablemem_map.map[overlap].end_pfn); | |
5275 | ||
5276 | if (pos != overlap && overlap + 1 != movablemem_map.nr_map) | |
5277 | memmove(&movablemem_map.map[pos+1], | |
5278 | &movablemem_map.map[overlap+1], | |
5279 | sizeof(struct movablemem_entry) * | |
5280 | (movablemem_map.nr_map - overlap - 1)); | |
5281 | ||
5282 | movablemem_map.nr_map -= overlap - pos; | |
5283 | } | |
5284 | ||
5285 | /** | |
5286 | * movablemem_map_add_region - Add a memory range into movablemem_map. | |
5287 | * @start: physical start address of range | |
5288 | * @end: physical end address of range | |
5289 | * | |
5290 | * This function transform the physical address into pfn, and then add the | |
5291 | * range into movablemem_map by calling insert_movablemem_map(). | |
5292 | */ | |
5293 | static void __init movablemem_map_add_region(u64 start, u64 size) | |
5294 | { | |
5295 | unsigned long start_pfn, end_pfn; | |
5296 | ||
5297 | /* In case size == 0 or start + size overflows */ | |
5298 | if (start + size <= start) | |
5299 | return; | |
5300 | ||
5301 | if (movablemem_map.nr_map >= ARRAY_SIZE(movablemem_map.map)) { | |
5302 | pr_err("movablemem_map: too many entries;" | |
5303 | " ignoring [mem %#010llx-%#010llx]\n", | |
5304 | (unsigned long long) start, | |
5305 | (unsigned long long) (start + size - 1)); | |
5306 | return; | |
5307 | } | |
5308 | ||
5309 | start_pfn = PFN_DOWN(start); | |
5310 | end_pfn = PFN_UP(start + size); | |
5311 | insert_movablemem_map(start_pfn, end_pfn); | |
5312 | } | |
5313 | ||
5314 | /* | |
5315 | * cmdline_parse_movablemem_map - Parse boot option movablemem_map. | |
5316 | * @p: The boot option of the following format: | |
5317 | * movablemem_map=nn[KMG]@ss[KMG] | |
5318 | * | |
5319 | * This option sets the memory range [ss, ss+nn) to be used as movable memory. | |
5320 | * | |
5321 | * Return: 0 on success or -EINVAL on failure. | |
5322 | */ | |
5323 | static int __init cmdline_parse_movablemem_map(char *p) | |
5324 | { | |
5325 | char *oldp; | |
5326 | u64 start_at, mem_size; | |
5327 | ||
5328 | if (!p) | |
5329 | goto err; | |
5330 | ||
5331 | if (!strcmp(p, "acpi")) | |
5332 | movablemem_map.acpi = true; | |
5333 | ||
5334 | /* | |
5335 | * If user decide to use info from BIOS, all the other user specified | |
5336 | * ranges will be ingored. | |
5337 | */ | |
5338 | if (movablemem_map.acpi) { | |
5339 | if (movablemem_map.nr_map) { | |
5340 | memset(movablemem_map.map, 0, | |
5341 | sizeof(struct movablemem_entry) | |
5342 | * movablemem_map.nr_map); | |
5343 | movablemem_map.nr_map = 0; | |
5344 | } | |
5345 | return 0; | |
5346 | } | |
5347 | ||
5348 | oldp = p; | |
5349 | mem_size = memparse(p, &p); | |
5350 | if (p == oldp) | |
5351 | goto err; | |
5352 | ||
5353 | if (*p == '@') { | |
5354 | oldp = ++p; | |
5355 | start_at = memparse(p, &p); | |
5356 | if (p == oldp || *p != '\0') | |
5357 | goto err; | |
5358 | ||
5359 | movablemem_map_add_region(start_at, mem_size); | |
5360 | return 0; | |
5361 | } | |
5362 | err: | |
5363 | return -EINVAL; | |
5364 | } | |
5365 | early_param("movablemem_map", cmdline_parse_movablemem_map); | |
5366 | ||
5367 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ | |
5368 | ||
5369 | /** | |
5370 | * set_dma_reserve - set the specified number of pages reserved in the first zone | |
5371 | * @new_dma_reserve: The number of pages to mark reserved | |
5372 | * | |
5373 | * The per-cpu batchsize and zone watermarks are determined by present_pages. | |
5374 | * In the DMA zone, a significant percentage may be consumed by kernel image | |
5375 | * and other unfreeable allocations which can skew the watermarks badly. This | |
5376 | * function may optionally be used to account for unfreeable pages in the | |
5377 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and | |
5378 | * smaller per-cpu batchsize. | |
5379 | */ | |
5380 | void __init set_dma_reserve(unsigned long new_dma_reserve) | |
5381 | { | |
5382 | dma_reserve = new_dma_reserve; | |
5383 | } | |
5384 | ||
5385 | void __init free_area_init(unsigned long *zones_size) | |
5386 | { | |
5387 | free_area_init_node(0, zones_size, | |
5388 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); | |
5389 | } | |
5390 | ||
5391 | static int page_alloc_cpu_notify(struct notifier_block *self, | |
5392 | unsigned long action, void *hcpu) | |
5393 | { | |
5394 | int cpu = (unsigned long)hcpu; | |
5395 | ||
5396 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { | |
5397 | lru_add_drain_cpu(cpu); | |
5398 | drain_pages(cpu); | |
5399 | ||
5400 | /* | |
5401 | * Spill the event counters of the dead processor | |
5402 | * into the current processors event counters. | |
5403 | * This artificially elevates the count of the current | |
5404 | * processor. | |
5405 | */ | |
5406 | vm_events_fold_cpu(cpu); | |
5407 | ||
5408 | /* | |
5409 | * Zero the differential counters of the dead processor | |
5410 | * so that the vm statistics are consistent. | |
5411 | * | |
5412 | * This is only okay since the processor is dead and cannot | |
5413 | * race with what we are doing. | |
5414 | */ | |
5415 | refresh_cpu_vm_stats(cpu); | |
5416 | } | |
5417 | return NOTIFY_OK; | |
5418 | } | |
5419 | ||
5420 | void __init page_alloc_init(void) | |
5421 | { | |
5422 | hotcpu_notifier(page_alloc_cpu_notify, 0); | |
5423 | } | |
5424 | ||
5425 | /* | |
5426 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio | |
5427 | * or min_free_kbytes changes. | |
5428 | */ | |
5429 | static void calculate_totalreserve_pages(void) | |
5430 | { | |
5431 | struct pglist_data *pgdat; | |
5432 | unsigned long reserve_pages = 0; | |
5433 | enum zone_type i, j; | |
5434 | ||
5435 | for_each_online_pgdat(pgdat) { | |
5436 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
5437 | struct zone *zone = pgdat->node_zones + i; | |
5438 | unsigned long max = 0; | |
5439 | ||
5440 | /* Find valid and maximum lowmem_reserve in the zone */ | |
5441 | for (j = i; j < MAX_NR_ZONES; j++) { | |
5442 | if (zone->lowmem_reserve[j] > max) | |
5443 | max = zone->lowmem_reserve[j]; | |
5444 | } | |
5445 | ||
5446 | /* we treat the high watermark as reserved pages. */ | |
5447 | max += high_wmark_pages(zone); | |
5448 | ||
5449 | if (max > zone->managed_pages) | |
5450 | max = zone->managed_pages; | |
5451 | reserve_pages += max; | |
5452 | /* | |
5453 | * Lowmem reserves are not available to | |
5454 | * GFP_HIGHUSER page cache allocations and | |
5455 | * kswapd tries to balance zones to their high | |
5456 | * watermark. As a result, neither should be | |
5457 | * regarded as dirtyable memory, to prevent a | |
5458 | * situation where reclaim has to clean pages | |
5459 | * in order to balance the zones. | |
5460 | */ | |
5461 | zone->dirty_balance_reserve = max; | |
5462 | } | |
5463 | } | |
5464 | dirty_balance_reserve = reserve_pages; | |
5465 | totalreserve_pages = reserve_pages; | |
5466 | } | |
5467 | ||
5468 | /* | |
5469 | * setup_per_zone_lowmem_reserve - called whenever | |
5470 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone | |
5471 | * has a correct pages reserved value, so an adequate number of | |
5472 | * pages are left in the zone after a successful __alloc_pages(). | |
5473 | */ | |
5474 | static void setup_per_zone_lowmem_reserve(void) | |
5475 | { | |
5476 | struct pglist_data *pgdat; | |
5477 | enum zone_type j, idx; | |
5478 | ||
5479 | for_each_online_pgdat(pgdat) { | |
5480 | for (j = 0; j < MAX_NR_ZONES; j++) { | |
5481 | struct zone *zone = pgdat->node_zones + j; | |
5482 | unsigned long managed_pages = zone->managed_pages; | |
5483 | ||
5484 | zone->lowmem_reserve[j] = 0; | |
5485 | ||
5486 | idx = j; | |
5487 | while (idx) { | |
5488 | struct zone *lower_zone; | |
5489 | ||
5490 | idx--; | |
5491 | ||
5492 | if (sysctl_lowmem_reserve_ratio[idx] < 1) | |
5493 | sysctl_lowmem_reserve_ratio[idx] = 1; | |
5494 | ||
5495 | lower_zone = pgdat->node_zones + idx; | |
5496 | lower_zone->lowmem_reserve[j] = managed_pages / | |
5497 | sysctl_lowmem_reserve_ratio[idx]; | |
5498 | managed_pages += lower_zone->managed_pages; | |
5499 | } | |
5500 | } | |
5501 | } | |
5502 | ||
5503 | /* update totalreserve_pages */ | |
5504 | calculate_totalreserve_pages(); | |
5505 | } | |
5506 | ||
5507 | static void __setup_per_zone_wmarks(void) | |
5508 | { | |
5509 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
5510 | unsigned long lowmem_pages = 0; | |
5511 | struct zone *zone; | |
5512 | unsigned long flags; | |
5513 | ||
5514 | /* Calculate total number of !ZONE_HIGHMEM pages */ | |
5515 | for_each_zone(zone) { | |
5516 | if (!is_highmem(zone)) | |
5517 | lowmem_pages += zone->managed_pages; | |
5518 | } | |
5519 | ||
5520 | for_each_zone(zone) { | |
5521 | u64 tmp; | |
5522 | ||
5523 | spin_lock_irqsave(&zone->lock, flags); | |
5524 | tmp = (u64)pages_min * zone->managed_pages; | |
5525 | do_div(tmp, lowmem_pages); | |
5526 | if (is_highmem(zone)) { | |
5527 | /* | |
5528 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't | |
5529 | * need highmem pages, so cap pages_min to a small | |
5530 | * value here. | |
5531 | * | |
5532 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) | |
5533 | * deltas controls asynch page reclaim, and so should | |
5534 | * not be capped for highmem. | |
5535 | */ | |
5536 | unsigned long min_pages; | |
5537 | ||
5538 | min_pages = zone->managed_pages / 1024; | |
5539 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); | |
5540 | zone->watermark[WMARK_MIN] = min_pages; | |
5541 | } else { | |
5542 | /* | |
5543 | * If it's a lowmem zone, reserve a number of pages | |
5544 | * proportionate to the zone's size. | |
5545 | */ | |
5546 | zone->watermark[WMARK_MIN] = tmp; | |
5547 | } | |
5548 | ||
5549 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); | |
5550 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); | |
5551 | ||
5552 | setup_zone_migrate_reserve(zone); | |
5553 | spin_unlock_irqrestore(&zone->lock, flags); | |
5554 | } | |
5555 | ||
5556 | /* update totalreserve_pages */ | |
5557 | calculate_totalreserve_pages(); | |
5558 | } | |
5559 | ||
5560 | /** | |
5561 | * setup_per_zone_wmarks - called when min_free_kbytes changes | |
5562 | * or when memory is hot-{added|removed} | |
5563 | * | |
5564 | * Ensures that the watermark[min,low,high] values for each zone are set | |
5565 | * correctly with respect to min_free_kbytes. | |
5566 | */ | |
5567 | void setup_per_zone_wmarks(void) | |
5568 | { | |
5569 | mutex_lock(&zonelists_mutex); | |
5570 | __setup_per_zone_wmarks(); | |
5571 | mutex_unlock(&zonelists_mutex); | |
5572 | } | |
5573 | ||
5574 | /* | |
5575 | * The inactive anon list should be small enough that the VM never has to | |
5576 | * do too much work, but large enough that each inactive page has a chance | |
5577 | * to be referenced again before it is swapped out. | |
5578 | * | |
5579 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to | |
5580 | * INACTIVE_ANON pages on this zone's LRU, maintained by the | |
5581 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of | |
5582 | * the anonymous pages are kept on the inactive list. | |
5583 | * | |
5584 | * total target max | |
5585 | * memory ratio inactive anon | |
5586 | * ------------------------------------- | |
5587 | * 10MB 1 5MB | |
5588 | * 100MB 1 50MB | |
5589 | * 1GB 3 250MB | |
5590 | * 10GB 10 0.9GB | |
5591 | * 100GB 31 3GB | |
5592 | * 1TB 101 10GB | |
5593 | * 10TB 320 32GB | |
5594 | */ | |
5595 | static void __meminit calculate_zone_inactive_ratio(struct zone *zone) | |
5596 | { | |
5597 | unsigned int gb, ratio; | |
5598 | ||
5599 | /* Zone size in gigabytes */ | |
5600 | gb = zone->managed_pages >> (30 - PAGE_SHIFT); | |
5601 | if (gb) | |
5602 | ratio = int_sqrt(10 * gb); | |
5603 | else | |
5604 | ratio = 1; | |
5605 | ||
5606 | zone->inactive_ratio = ratio; | |
5607 | } | |
5608 | ||
5609 | static void __meminit setup_per_zone_inactive_ratio(void) | |
5610 | { | |
5611 | struct zone *zone; | |
5612 | ||
5613 | for_each_zone(zone) | |
5614 | calculate_zone_inactive_ratio(zone); | |
5615 | } | |
5616 | ||
5617 | /* | |
5618 | * Initialise min_free_kbytes. | |
5619 | * | |
5620 | * For small machines we want it small (128k min). For large machines | |
5621 | * we want it large (64MB max). But it is not linear, because network | |
5622 | * bandwidth does not increase linearly with machine size. We use | |
5623 | * | |
5624 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: | |
5625 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) | |
5626 | * | |
5627 | * which yields | |
5628 | * | |
5629 | * 16MB: 512k | |
5630 | * 32MB: 724k | |
5631 | * 64MB: 1024k | |
5632 | * 128MB: 1448k | |
5633 | * 256MB: 2048k | |
5634 | * 512MB: 2896k | |
5635 | * 1024MB: 4096k | |
5636 | * 2048MB: 5792k | |
5637 | * 4096MB: 8192k | |
5638 | * 8192MB: 11584k | |
5639 | * 16384MB: 16384k | |
5640 | */ | |
5641 | int __meminit init_per_zone_wmark_min(void) | |
5642 | { | |
5643 | unsigned long lowmem_kbytes; | |
5644 | ||
5645 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
5646 | ||
5647 | min_free_kbytes = int_sqrt(lowmem_kbytes * 16); | |
5648 | if (min_free_kbytes < 128) | |
5649 | min_free_kbytes = 128; | |
5650 | if (min_free_kbytes > 65536) | |
5651 | min_free_kbytes = 65536; | |
5652 | setup_per_zone_wmarks(); | |
5653 | refresh_zone_stat_thresholds(); | |
5654 | setup_per_zone_lowmem_reserve(); | |
5655 | setup_per_zone_inactive_ratio(); | |
5656 | return 0; | |
5657 | } | |
5658 | module_init(init_per_zone_wmark_min) | |
5659 | ||
5660 | /* | |
5661 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so | |
5662 | * that we can call two helper functions whenever min_free_kbytes | |
5663 | * changes. | |
5664 | */ | |
5665 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, | |
5666 | void __user *buffer, size_t *length, loff_t *ppos) | |
5667 | { | |
5668 | proc_dointvec(table, write, buffer, length, ppos); | |
5669 | if (write) | |
5670 | setup_per_zone_wmarks(); | |
5671 | return 0; | |
5672 | } | |
5673 | ||
5674 | #ifdef CONFIG_NUMA | |
5675 | int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, | |
5676 | void __user *buffer, size_t *length, loff_t *ppos) | |
5677 | { | |
5678 | struct zone *zone; | |
5679 | int rc; | |
5680 | ||
5681 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5682 | if (rc) | |
5683 | return rc; | |
5684 | ||
5685 | for_each_zone(zone) | |
5686 | zone->min_unmapped_pages = (zone->managed_pages * | |
5687 | sysctl_min_unmapped_ratio) / 100; | |
5688 | return 0; | |
5689 | } | |
5690 | ||
5691 | int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, | |
5692 | void __user *buffer, size_t *length, loff_t *ppos) | |
5693 | { | |
5694 | struct zone *zone; | |
5695 | int rc; | |
5696 | ||
5697 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5698 | if (rc) | |
5699 | return rc; | |
5700 | ||
5701 | for_each_zone(zone) | |
5702 | zone->min_slab_pages = (zone->managed_pages * | |
5703 | sysctl_min_slab_ratio) / 100; | |
5704 | return 0; | |
5705 | } | |
5706 | #endif | |
5707 | ||
5708 | /* | |
5709 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
5710 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
5711 | * whenever sysctl_lowmem_reserve_ratio changes. | |
5712 | * | |
5713 | * The reserve ratio obviously has absolutely no relation with the | |
5714 | * minimum watermarks. The lowmem reserve ratio can only make sense | |
5715 | * if in function of the boot time zone sizes. | |
5716 | */ | |
5717 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, | |
5718 | void __user *buffer, size_t *length, loff_t *ppos) | |
5719 | { | |
5720 | proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5721 | setup_per_zone_lowmem_reserve(); | |
5722 | return 0; | |
5723 | } | |
5724 | ||
5725 | /* | |
5726 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each | |
5727 | * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist | |
5728 | * can have before it gets flushed back to buddy allocator. | |
5729 | */ | |
5730 | ||
5731 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, | |
5732 | void __user *buffer, size_t *length, loff_t *ppos) | |
5733 | { | |
5734 | struct zone *zone; | |
5735 | unsigned int cpu; | |
5736 | int ret; | |
5737 | ||
5738 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5739 | if (!write || (ret < 0)) | |
5740 | return ret; | |
5741 | for_each_populated_zone(zone) { | |
5742 | for_each_possible_cpu(cpu) { | |
5743 | unsigned long high; | |
5744 | high = zone->managed_pages / percpu_pagelist_fraction; | |
5745 | setup_pagelist_highmark( | |
5746 | per_cpu_ptr(zone->pageset, cpu), high); | |
5747 | } | |
5748 | } | |
5749 | return 0; | |
5750 | } | |
5751 | ||
5752 | int hashdist = HASHDIST_DEFAULT; | |
5753 | ||
5754 | #ifdef CONFIG_NUMA | |
5755 | static int __init set_hashdist(char *str) | |
5756 | { | |
5757 | if (!str) | |
5758 | return 0; | |
5759 | hashdist = simple_strtoul(str, &str, 0); | |
5760 | return 1; | |
5761 | } | |
5762 | __setup("hashdist=", set_hashdist); | |
5763 | #endif | |
5764 | ||
5765 | /* | |
5766 | * allocate a large system hash table from bootmem | |
5767 | * - it is assumed that the hash table must contain an exact power-of-2 | |
5768 | * quantity of entries | |
5769 | * - limit is the number of hash buckets, not the total allocation size | |
5770 | */ | |
5771 | void *__init alloc_large_system_hash(const char *tablename, | |
5772 | unsigned long bucketsize, | |
5773 | unsigned long numentries, | |
5774 | int scale, | |
5775 | int flags, | |
5776 | unsigned int *_hash_shift, | |
5777 | unsigned int *_hash_mask, | |
5778 | unsigned long low_limit, | |
5779 | unsigned long high_limit) | |
5780 | { | |
5781 | unsigned long long max = high_limit; | |
5782 | unsigned long log2qty, size; | |
5783 | void *table = NULL; | |
5784 | ||
5785 | /* allow the kernel cmdline to have a say */ | |
5786 | if (!numentries) { | |
5787 | /* round applicable memory size up to nearest megabyte */ | |
5788 | numentries = nr_kernel_pages; | |
5789 | numentries += (1UL << (20 - PAGE_SHIFT)) - 1; | |
5790 | numentries >>= 20 - PAGE_SHIFT; | |
5791 | numentries <<= 20 - PAGE_SHIFT; | |
5792 | ||
5793 | /* limit to 1 bucket per 2^scale bytes of low memory */ | |
5794 | if (scale > PAGE_SHIFT) | |
5795 | numentries >>= (scale - PAGE_SHIFT); | |
5796 | else | |
5797 | numentries <<= (PAGE_SHIFT - scale); | |
5798 | ||
5799 | /* Make sure we've got at least a 0-order allocation.. */ | |
5800 | if (unlikely(flags & HASH_SMALL)) { | |
5801 | /* Makes no sense without HASH_EARLY */ | |
5802 | WARN_ON(!(flags & HASH_EARLY)); | |
5803 | if (!(numentries >> *_hash_shift)) { | |
5804 | numentries = 1UL << *_hash_shift; | |
5805 | BUG_ON(!numentries); | |
5806 | } | |
5807 | } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) | |
5808 | numentries = PAGE_SIZE / bucketsize; | |
5809 | } | |
5810 | numentries = roundup_pow_of_two(numentries); | |
5811 | ||
5812 | /* limit allocation size to 1/16 total memory by default */ | |
5813 | if (max == 0) { | |
5814 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; | |
5815 | do_div(max, bucketsize); | |
5816 | } | |
5817 | max = min(max, 0x80000000ULL); | |
5818 | ||
5819 | if (numentries < low_limit) | |
5820 | numentries = low_limit; | |
5821 | if (numentries > max) | |
5822 | numentries = max; | |
5823 | ||
5824 | log2qty = ilog2(numentries); | |
5825 | ||
5826 | do { | |
5827 | size = bucketsize << log2qty; | |
5828 | if (flags & HASH_EARLY) | |
5829 | table = alloc_bootmem_nopanic(size); | |
5830 | else if (hashdist) | |
5831 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); | |
5832 | else { | |
5833 | /* | |
5834 | * If bucketsize is not a power-of-two, we may free | |
5835 | * some pages at the end of hash table which | |
5836 | * alloc_pages_exact() automatically does | |
5837 | */ | |
5838 | if (get_order(size) < MAX_ORDER) { | |
5839 | table = alloc_pages_exact(size, GFP_ATOMIC); | |
5840 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); | |
5841 | } | |
5842 | } | |
5843 | } while (!table && size > PAGE_SIZE && --log2qty); | |
5844 | ||
5845 | if (!table) | |
5846 | panic("Failed to allocate %s hash table\n", tablename); | |
5847 | ||
5848 | printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", | |
5849 | tablename, | |
5850 | (1UL << log2qty), | |
5851 | ilog2(size) - PAGE_SHIFT, | |
5852 | size); | |
5853 | ||
5854 | if (_hash_shift) | |
5855 | *_hash_shift = log2qty; | |
5856 | if (_hash_mask) | |
5857 | *_hash_mask = (1 << log2qty) - 1; | |
5858 | ||
5859 | return table; | |
5860 | } | |
5861 | ||
5862 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ | |
5863 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, | |
5864 | unsigned long pfn) | |
5865 | { | |
5866 | #ifdef CONFIG_SPARSEMEM | |
5867 | return __pfn_to_section(pfn)->pageblock_flags; | |
5868 | #else | |
5869 | return zone->pageblock_flags; | |
5870 | #endif /* CONFIG_SPARSEMEM */ | |
5871 | } | |
5872 | ||
5873 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) | |
5874 | { | |
5875 | #ifdef CONFIG_SPARSEMEM | |
5876 | pfn &= (PAGES_PER_SECTION-1); | |
5877 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
5878 | #else | |
5879 | pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); | |
5880 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; | |
5881 | #endif /* CONFIG_SPARSEMEM */ | |
5882 | } | |
5883 | ||
5884 | /** | |
5885 | * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages | |
5886 | * @page: The page within the block of interest | |
5887 | * @start_bitidx: The first bit of interest to retrieve | |
5888 | * @end_bitidx: The last bit of interest | |
5889 | * returns pageblock_bits flags | |
5890 | */ | |
5891 | unsigned long get_pageblock_flags_group(struct page *page, | |
5892 | int start_bitidx, int end_bitidx) | |
5893 | { | |
5894 | struct zone *zone; | |
5895 | unsigned long *bitmap; | |
5896 | unsigned long pfn, bitidx; | |
5897 | unsigned long flags = 0; | |
5898 | unsigned long value = 1; | |
5899 | ||
5900 | zone = page_zone(page); | |
5901 | pfn = page_to_pfn(page); | |
5902 | bitmap = get_pageblock_bitmap(zone, pfn); | |
5903 | bitidx = pfn_to_bitidx(zone, pfn); | |
5904 | ||
5905 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
5906 | if (test_bit(bitidx + start_bitidx, bitmap)) | |
5907 | flags |= value; | |
5908 | ||
5909 | return flags; | |
5910 | } | |
5911 | ||
5912 | /** | |
5913 | * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages | |
5914 | * @page: The page within the block of interest | |
5915 | * @start_bitidx: The first bit of interest | |
5916 | * @end_bitidx: The last bit of interest | |
5917 | * @flags: The flags to set | |
5918 | */ | |
5919 | void set_pageblock_flags_group(struct page *page, unsigned long flags, | |
5920 | int start_bitidx, int end_bitidx) | |
5921 | { | |
5922 | struct zone *zone; | |
5923 | unsigned long *bitmap; | |
5924 | unsigned long pfn, bitidx; | |
5925 | unsigned long value = 1; | |
5926 | ||
5927 | zone = page_zone(page); | |
5928 | pfn = page_to_pfn(page); | |
5929 | bitmap = get_pageblock_bitmap(zone, pfn); | |
5930 | bitidx = pfn_to_bitidx(zone, pfn); | |
5931 | VM_BUG_ON(!zone_spans_pfn(zone, pfn)); | |
5932 | ||
5933 | for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) | |
5934 | if (flags & value) | |
5935 | __set_bit(bitidx + start_bitidx, bitmap); | |
5936 | else | |
5937 | __clear_bit(bitidx + start_bitidx, bitmap); | |
5938 | } | |
5939 | ||
5940 | /* | |
5941 | * This function checks whether pageblock includes unmovable pages or not. | |
5942 | * If @count is not zero, it is okay to include less @count unmovable pages | |
5943 | * | |
5944 | * PageLRU check wihtout isolation or lru_lock could race so that | |
5945 | * MIGRATE_MOVABLE block might include unmovable pages. It means you can't | |
5946 | * expect this function should be exact. | |
5947 | */ | |
5948 | bool has_unmovable_pages(struct zone *zone, struct page *page, int count, | |
5949 | bool skip_hwpoisoned_pages) | |
5950 | { | |
5951 | unsigned long pfn, iter, found; | |
5952 | int mt; | |
5953 | ||
5954 | /* | |
5955 | * For avoiding noise data, lru_add_drain_all() should be called | |
5956 | * If ZONE_MOVABLE, the zone never contains unmovable pages | |
5957 | */ | |
5958 | if (zone_idx(zone) == ZONE_MOVABLE) | |
5959 | return false; | |
5960 | mt = get_pageblock_migratetype(page); | |
5961 | if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) | |
5962 | return false; | |
5963 | ||
5964 | pfn = page_to_pfn(page); | |
5965 | for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { | |
5966 | unsigned long check = pfn + iter; | |
5967 | ||
5968 | if (!pfn_valid_within(check)) | |
5969 | continue; | |
5970 | ||
5971 | page = pfn_to_page(check); | |
5972 | /* | |
5973 | * We can't use page_count without pin a page | |
5974 | * because another CPU can free compound page. | |
5975 | * This check already skips compound tails of THP | |
5976 | * because their page->_count is zero at all time. | |
5977 | */ | |
5978 | if (!atomic_read(&page->_count)) { | |
5979 | if (PageBuddy(page)) | |
5980 | iter += (1 << page_order(page)) - 1; | |
5981 | continue; | |
5982 | } | |
5983 | ||
5984 | /* | |
5985 | * The HWPoisoned page may be not in buddy system, and | |
5986 | * page_count() is not 0. | |
5987 | */ | |
5988 | if (skip_hwpoisoned_pages && PageHWPoison(page)) | |
5989 | continue; | |
5990 | ||
5991 | if (!PageLRU(page)) | |
5992 | found++; | |
5993 | /* | |
5994 | * If there are RECLAIMABLE pages, we need to check it. | |
5995 | * But now, memory offline itself doesn't call shrink_slab() | |
5996 | * and it still to be fixed. | |
5997 | */ | |
5998 | /* | |
5999 | * If the page is not RAM, page_count()should be 0. | |
6000 | * we don't need more check. This is an _used_ not-movable page. | |
6001 | * | |
6002 | * The problematic thing here is PG_reserved pages. PG_reserved | |
6003 | * is set to both of a memory hole page and a _used_ kernel | |
6004 | * page at boot. | |
6005 | */ | |
6006 | if (found > count) | |
6007 | return true; | |
6008 | } | |
6009 | return false; | |
6010 | } | |
6011 | ||
6012 | bool is_pageblock_removable_nolock(struct page *page) | |
6013 | { | |
6014 | struct zone *zone; | |
6015 | unsigned long pfn; | |
6016 | ||
6017 | /* | |
6018 | * We have to be careful here because we are iterating over memory | |
6019 | * sections which are not zone aware so we might end up outside of | |
6020 | * the zone but still within the section. | |
6021 | * We have to take care about the node as well. If the node is offline | |
6022 | * its NODE_DATA will be NULL - see page_zone. | |
6023 | */ | |
6024 | if (!node_online(page_to_nid(page))) | |
6025 | return false; | |
6026 | ||
6027 | zone = page_zone(page); | |
6028 | pfn = page_to_pfn(page); | |
6029 | if (!zone_spans_pfn(zone, pfn)) | |
6030 | return false; | |
6031 | ||
6032 | return !has_unmovable_pages(zone, page, 0, true); | |
6033 | } | |
6034 | ||
6035 | #ifdef CONFIG_CMA | |
6036 | ||
6037 | static unsigned long pfn_max_align_down(unsigned long pfn) | |
6038 | { | |
6039 | return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
6040 | pageblock_nr_pages) - 1); | |
6041 | } | |
6042 | ||
6043 | static unsigned long pfn_max_align_up(unsigned long pfn) | |
6044 | { | |
6045 | return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, | |
6046 | pageblock_nr_pages)); | |
6047 | } | |
6048 | ||
6049 | /* [start, end) must belong to a single zone. */ | |
6050 | static int __alloc_contig_migrate_range(struct compact_control *cc, | |
6051 | unsigned long start, unsigned long end) | |
6052 | { | |
6053 | /* This function is based on compact_zone() from compaction.c. */ | |
6054 | unsigned long nr_reclaimed; | |
6055 | unsigned long pfn = start; | |
6056 | unsigned int tries = 0; | |
6057 | int ret = 0; | |
6058 | ||
6059 | migrate_prep(); | |
6060 | ||
6061 | while (pfn < end || !list_empty(&cc->migratepages)) { | |
6062 | if (fatal_signal_pending(current)) { | |
6063 | ret = -EINTR; | |
6064 | break; | |
6065 | } | |
6066 | ||
6067 | if (list_empty(&cc->migratepages)) { | |
6068 | cc->nr_migratepages = 0; | |
6069 | pfn = isolate_migratepages_range(cc->zone, cc, | |
6070 | pfn, end, true); | |
6071 | if (!pfn) { | |
6072 | ret = -EINTR; | |
6073 | break; | |
6074 | } | |
6075 | tries = 0; | |
6076 | } else if (++tries == 5) { | |
6077 | ret = ret < 0 ? ret : -EBUSY; | |
6078 | break; | |
6079 | } | |
6080 | ||
6081 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, | |
6082 | &cc->migratepages); | |
6083 | cc->nr_migratepages -= nr_reclaimed; | |
6084 | ||
6085 | ret = migrate_pages(&cc->migratepages, alloc_migrate_target, | |
6086 | 0, MIGRATE_SYNC, MR_CMA); | |
6087 | } | |
6088 | if (ret < 0) { | |
6089 | putback_movable_pages(&cc->migratepages); | |
6090 | return ret; | |
6091 | } | |
6092 | return 0; | |
6093 | } | |
6094 | ||
6095 | /** | |
6096 | * alloc_contig_range() -- tries to allocate given range of pages | |
6097 | * @start: start PFN to allocate | |
6098 | * @end: one-past-the-last PFN to allocate | |
6099 | * @migratetype: migratetype of the underlaying pageblocks (either | |
6100 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks | |
6101 | * in range must have the same migratetype and it must | |
6102 | * be either of the two. | |
6103 | * | |
6104 | * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES | |
6105 | * aligned, however it's the caller's responsibility to guarantee that | |
6106 | * we are the only thread that changes migrate type of pageblocks the | |
6107 | * pages fall in. | |
6108 | * | |
6109 | * The PFN range must belong to a single zone. | |
6110 | * | |
6111 | * Returns zero on success or negative error code. On success all | |
6112 | * pages which PFN is in [start, end) are allocated for the caller and | |
6113 | * need to be freed with free_contig_range(). | |
6114 | */ | |
6115 | int alloc_contig_range(unsigned long start, unsigned long end, | |
6116 | unsigned migratetype) | |
6117 | { | |
6118 | unsigned long outer_start, outer_end; | |
6119 | int ret = 0, order; | |
6120 | ||
6121 | struct compact_control cc = { | |
6122 | .nr_migratepages = 0, | |
6123 | .order = -1, | |
6124 | .zone = page_zone(pfn_to_page(start)), | |
6125 | .sync = true, | |
6126 | .ignore_skip_hint = true, | |
6127 | }; | |
6128 | INIT_LIST_HEAD(&cc.migratepages); | |
6129 | ||
6130 | /* | |
6131 | * What we do here is we mark all pageblocks in range as | |
6132 | * MIGRATE_ISOLATE. Because pageblock and max order pages may | |
6133 | * have different sizes, and due to the way page allocator | |
6134 | * work, we align the range to biggest of the two pages so | |
6135 | * that page allocator won't try to merge buddies from | |
6136 | * different pageblocks and change MIGRATE_ISOLATE to some | |
6137 | * other migration type. | |
6138 | * | |
6139 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we | |
6140 | * migrate the pages from an unaligned range (ie. pages that | |
6141 | * we are interested in). This will put all the pages in | |
6142 | * range back to page allocator as MIGRATE_ISOLATE. | |
6143 | * | |
6144 | * When this is done, we take the pages in range from page | |
6145 | * allocator removing them from the buddy system. This way | |
6146 | * page allocator will never consider using them. | |
6147 | * | |
6148 | * This lets us mark the pageblocks back as | |
6149 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the | |
6150 | * aligned range but not in the unaligned, original range are | |
6151 | * put back to page allocator so that buddy can use them. | |
6152 | */ | |
6153 | ||
6154 | ret = start_isolate_page_range(pfn_max_align_down(start), | |
6155 | pfn_max_align_up(end), migratetype, | |
6156 | false); | |
6157 | if (ret) | |
6158 | return ret; | |
6159 | ||
6160 | ret = __alloc_contig_migrate_range(&cc, start, end); | |
6161 | if (ret) | |
6162 | goto done; | |
6163 | ||
6164 | /* | |
6165 | * Pages from [start, end) are within a MAX_ORDER_NR_PAGES | |
6166 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's | |
6167 | * more, all pages in [start, end) are free in page allocator. | |
6168 | * What we are going to do is to allocate all pages from | |
6169 | * [start, end) (that is remove them from page allocator). | |
6170 | * | |
6171 | * The only problem is that pages at the beginning and at the | |
6172 | * end of interesting range may be not aligned with pages that | |
6173 | * page allocator holds, ie. they can be part of higher order | |
6174 | * pages. Because of this, we reserve the bigger range and | |
6175 | * once this is done free the pages we are not interested in. | |
6176 | * | |
6177 | * We don't have to hold zone->lock here because the pages are | |
6178 | * isolated thus they won't get removed from buddy. | |
6179 | */ | |
6180 | ||
6181 | lru_add_drain_all(); | |
6182 | drain_all_pages(); | |
6183 | ||
6184 | order = 0; | |
6185 | outer_start = start; | |
6186 | while (!PageBuddy(pfn_to_page(outer_start))) { | |
6187 | if (++order >= MAX_ORDER) { | |
6188 | ret = -EBUSY; | |
6189 | goto done; | |
6190 | } | |
6191 | outer_start &= ~0UL << order; | |
6192 | } | |
6193 | ||
6194 | /* Make sure the range is really isolated. */ | |
6195 | if (test_pages_isolated(outer_start, end, false)) { | |
6196 | pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n", | |
6197 | outer_start, end); | |
6198 | ret = -EBUSY; | |
6199 | goto done; | |
6200 | } | |
6201 | ||
6202 | ||
6203 | /* Grab isolated pages from freelists. */ | |
6204 | outer_end = isolate_freepages_range(&cc, outer_start, end); | |
6205 | if (!outer_end) { | |
6206 | ret = -EBUSY; | |
6207 | goto done; | |
6208 | } | |
6209 | ||
6210 | /* Free head and tail (if any) */ | |
6211 | if (start != outer_start) | |
6212 | free_contig_range(outer_start, start - outer_start); | |
6213 | if (end != outer_end) | |
6214 | free_contig_range(end, outer_end - end); | |
6215 | ||
6216 | done: | |
6217 | undo_isolate_page_range(pfn_max_align_down(start), | |
6218 | pfn_max_align_up(end), migratetype); | |
6219 | return ret; | |
6220 | } | |
6221 | ||
6222 | void free_contig_range(unsigned long pfn, unsigned nr_pages) | |
6223 | { | |
6224 | unsigned int count = 0; | |
6225 | ||
6226 | for (; nr_pages--; pfn++) { | |
6227 | struct page *page = pfn_to_page(pfn); | |
6228 | ||
6229 | count += page_count(page) != 1; | |
6230 | __free_page(page); | |
6231 | } | |
6232 | WARN(count != 0, "%d pages are still in use!\n", count); | |
6233 | } | |
6234 | #endif | |
6235 | ||
6236 | #ifdef CONFIG_MEMORY_HOTPLUG | |
6237 | static int __meminit __zone_pcp_update(void *data) | |
6238 | { | |
6239 | struct zone *zone = data; | |
6240 | int cpu; | |
6241 | unsigned long batch = zone_batchsize(zone), flags; | |
6242 | ||
6243 | for_each_possible_cpu(cpu) { | |
6244 | struct per_cpu_pageset *pset; | |
6245 | struct per_cpu_pages *pcp; | |
6246 | ||
6247 | pset = per_cpu_ptr(zone->pageset, cpu); | |
6248 | pcp = &pset->pcp; | |
6249 | ||
6250 | local_irq_save(flags); | |
6251 | if (pcp->count > 0) | |
6252 | free_pcppages_bulk(zone, pcp->count, pcp); | |
6253 | drain_zonestat(zone, pset); | |
6254 | setup_pageset(pset, batch); | |
6255 | local_irq_restore(flags); | |
6256 | } | |
6257 | return 0; | |
6258 | } | |
6259 | ||
6260 | void __meminit zone_pcp_update(struct zone *zone) | |
6261 | { | |
6262 | stop_machine(__zone_pcp_update, zone, NULL); | |
6263 | } | |
6264 | #endif | |
6265 | ||
6266 | void zone_pcp_reset(struct zone *zone) | |
6267 | { | |
6268 | unsigned long flags; | |
6269 | int cpu; | |
6270 | struct per_cpu_pageset *pset; | |
6271 | ||
6272 | /* avoid races with drain_pages() */ | |
6273 | local_irq_save(flags); | |
6274 | if (zone->pageset != &boot_pageset) { | |
6275 | for_each_online_cpu(cpu) { | |
6276 | pset = per_cpu_ptr(zone->pageset, cpu); | |
6277 | drain_zonestat(zone, pset); | |
6278 | } | |
6279 | free_percpu(zone->pageset); | |
6280 | zone->pageset = &boot_pageset; | |
6281 | } | |
6282 | local_irq_restore(flags); | |
6283 | } | |
6284 | ||
6285 | #ifdef CONFIG_MEMORY_HOTREMOVE | |
6286 | /* | |
6287 | * All pages in the range must be isolated before calling this. | |
6288 | */ | |
6289 | void | |
6290 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) | |
6291 | { | |
6292 | struct page *page; | |
6293 | struct zone *zone; | |
6294 | int order, i; | |
6295 | unsigned long pfn; | |
6296 | unsigned long flags; | |
6297 | /* find the first valid pfn */ | |
6298 | for (pfn = start_pfn; pfn < end_pfn; pfn++) | |
6299 | if (pfn_valid(pfn)) | |
6300 | break; | |
6301 | if (pfn == end_pfn) | |
6302 | return; | |
6303 | zone = page_zone(pfn_to_page(pfn)); | |
6304 | spin_lock_irqsave(&zone->lock, flags); | |
6305 | pfn = start_pfn; | |
6306 | while (pfn < end_pfn) { | |
6307 | if (!pfn_valid(pfn)) { | |
6308 | pfn++; | |
6309 | continue; | |
6310 | } | |
6311 | page = pfn_to_page(pfn); | |
6312 | /* | |
6313 | * The HWPoisoned page may be not in buddy system, and | |
6314 | * page_count() is not 0. | |
6315 | */ | |
6316 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { | |
6317 | pfn++; | |
6318 | SetPageReserved(page); | |
6319 | continue; | |
6320 | } | |
6321 | ||
6322 | BUG_ON(page_count(page)); | |
6323 | BUG_ON(!PageBuddy(page)); | |
6324 | order = page_order(page); | |
6325 | #ifdef CONFIG_DEBUG_VM | |
6326 | printk(KERN_INFO "remove from free list %lx %d %lx\n", | |
6327 | pfn, 1 << order, end_pfn); | |
6328 | #endif | |
6329 | list_del(&page->lru); | |
6330 | rmv_page_order(page); | |
6331 | zone->free_area[order].nr_free--; | |
6332 | for (i = 0; i < (1 << order); i++) | |
6333 | SetPageReserved((page+i)); | |
6334 | pfn += (1 << order); | |
6335 | } | |
6336 | spin_unlock_irqrestore(&zone->lock, flags); | |
6337 | } | |
6338 | #endif | |
6339 | ||
6340 | #ifdef CONFIG_MEMORY_FAILURE | |
6341 | bool is_free_buddy_page(struct page *page) | |
6342 | { | |
6343 | struct zone *zone = page_zone(page); | |
6344 | unsigned long pfn = page_to_pfn(page); | |
6345 | unsigned long flags; | |
6346 | int order; | |
6347 | ||
6348 | spin_lock_irqsave(&zone->lock, flags); | |
6349 | for (order = 0; order < MAX_ORDER; order++) { | |
6350 | struct page *page_head = page - (pfn & ((1 << order) - 1)); | |
6351 | ||
6352 | if (PageBuddy(page_head) && page_order(page_head) >= order) | |
6353 | break; | |
6354 | } | |
6355 | spin_unlock_irqrestore(&zone->lock, flags); | |
6356 | ||
6357 | return order < MAX_ORDER; | |
6358 | } | |
6359 | #endif | |
6360 | ||
6361 | static const struct trace_print_flags pageflag_names[] = { | |
6362 | {1UL << PG_locked, "locked" }, | |
6363 | {1UL << PG_error, "error" }, | |
6364 | {1UL << PG_referenced, "referenced" }, | |
6365 | {1UL << PG_uptodate, "uptodate" }, | |
6366 | {1UL << PG_dirty, "dirty" }, | |
6367 | {1UL << PG_lru, "lru" }, | |
6368 | {1UL << PG_active, "active" }, | |
6369 | {1UL << PG_slab, "slab" }, | |
6370 | {1UL << PG_owner_priv_1, "owner_priv_1" }, | |
6371 | {1UL << PG_arch_1, "arch_1" }, | |
6372 | {1UL << PG_reserved, "reserved" }, | |
6373 | {1UL << PG_private, "private" }, | |
6374 | {1UL << PG_private_2, "private_2" }, | |
6375 | {1UL << PG_writeback, "writeback" }, | |
6376 | #ifdef CONFIG_PAGEFLAGS_EXTENDED | |
6377 | {1UL << PG_head, "head" }, | |
6378 | {1UL << PG_tail, "tail" }, | |
6379 | #else | |
6380 | {1UL << PG_compound, "compound" }, | |
6381 | #endif | |
6382 | {1UL << PG_swapcache, "swapcache" }, | |
6383 | {1UL << PG_mappedtodisk, "mappedtodisk" }, | |
6384 | {1UL << PG_reclaim, "reclaim" }, | |
6385 | {1UL << PG_swapbacked, "swapbacked" }, | |
6386 | {1UL << PG_unevictable, "unevictable" }, | |
6387 | #ifdef CONFIG_MMU | |
6388 | {1UL << PG_mlocked, "mlocked" }, | |
6389 | #endif | |
6390 | #ifdef CONFIG_ARCH_USES_PG_UNCACHED | |
6391 | {1UL << PG_uncached, "uncached" }, | |
6392 | #endif | |
6393 | #ifdef CONFIG_MEMORY_FAILURE | |
6394 | {1UL << PG_hwpoison, "hwpoison" }, | |
6395 | #endif | |
6396 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
6397 | {1UL << PG_compound_lock, "compound_lock" }, | |
6398 | #endif | |
6399 | }; | |
6400 | ||
6401 | static void dump_page_flags(unsigned long flags) | |
6402 | { | |
6403 | const char *delim = ""; | |
6404 | unsigned long mask; | |
6405 | int i; | |
6406 | ||
6407 | BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS); | |
6408 | ||
6409 | printk(KERN_ALERT "page flags: %#lx(", flags); | |
6410 | ||
6411 | /* remove zone id */ | |
6412 | flags &= (1UL << NR_PAGEFLAGS) - 1; | |
6413 | ||
6414 | for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) { | |
6415 | ||
6416 | mask = pageflag_names[i].mask; | |
6417 | if ((flags & mask) != mask) | |
6418 | continue; | |
6419 | ||
6420 | flags &= ~mask; | |
6421 | printk("%s%s", delim, pageflag_names[i].name); | |
6422 | delim = "|"; | |
6423 | } | |
6424 | ||
6425 | /* check for left over flags */ | |
6426 | if (flags) | |
6427 | printk("%s%#lx", delim, flags); | |
6428 | ||
6429 | printk(")\n"); | |
6430 | } | |
6431 | ||
6432 | void dump_page(struct page *page) | |
6433 | { | |
6434 | printk(KERN_ALERT | |
6435 | "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", | |
6436 | page, atomic_read(&page->_count), page_mapcount(page), | |
6437 | page->mapping, page->index); | |
6438 | dump_page_flags(page->flags); | |
6439 | mem_cgroup_print_bad_page(page); | |
6440 | } |