usb: host: xhci-dbg: HCIVERSION should be a binary number
[GitHub/exynos8895/android_kernel_samsung_universal8895.git] / mm / page_alloc.c
1 /*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17 #include <linux/stddef.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
69 #include "internal.h"
70
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
78 #endif
79
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 /*
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
86 */
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
90 #endif
91
92 /*
93 * Array of node states.
94 */
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
98 #ifndef CONFIG_NUMA
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #endif
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
105 #endif
106 [N_CPU] = { { [0] = 1UL } },
107 #endif /* NUMA */
108 };
109 EXPORT_SYMBOL(node_states);
110
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
113
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
117 /*
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
122 */
123 unsigned long dirty_balance_reserve __read_mostly;
124
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127
128 /*
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
135 */
136 static inline int get_pcppage_migratetype(struct page *page)
137 {
138 return page->index;
139 }
140
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
142 {
143 page->index = migratetype;
144 }
145
146 #ifdef CONFIG_PM_SLEEP
147 /*
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
154 */
155
156 static gfp_t saved_gfp_mask;
157
158 void pm_restore_gfp_mask(void)
159 {
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
163 saved_gfp_mask = 0;
164 }
165 }
166
167 void pm_restrict_gfp_mask(void)
168 {
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
173 }
174
175 bool pm_suspended_storage(void)
176 {
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
178 return false;
179 return true;
180 }
181 #endif /* CONFIG_PM_SLEEP */
182
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
185 #endif
186
187 static void __free_pages_ok(struct page *page, unsigned int order);
188
189 /*
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
196 *
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
199 */
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
202 256,
203 #endif
204 #ifdef CONFIG_ZONE_DMA32
205 256,
206 #endif
207 #ifdef CONFIG_HIGHMEM
208 32,
209 #endif
210 32,
211 };
212
213 EXPORT_SYMBOL(totalram_pages);
214
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
217 "DMA",
218 #endif
219 #ifdef CONFIG_ZONE_DMA32
220 "DMA32",
221 #endif
222 "Normal",
223 #ifdef CONFIG_HIGHMEM
224 "HighMem",
225 #endif
226 "Movable",
227 #ifdef CONFIG_ZONE_DEVICE
228 "Device",
229 #endif
230 };
231
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
234 NULL,
235 free_compound_page,
236 #ifdef CONFIG_HUGETLB_PAGE
237 free_huge_page,
238 #endif
239 };
240
241 /*
242 * Try to keep at least this much lowmem free. Do not allow normal
243 * allocations below this point, only high priority ones. Automatically
244 * tuned according to the amount of memory in the system.
245 */
246 int min_free_kbytes = 1024;
247 int user_min_free_kbytes = -1;
248
249 /*
250 * Extra memory for the system to try freeing. Used to temporarily
251 * free memory, to make space for new workloads. Anyone can allocate
252 * down to the min watermarks controlled by min_free_kbytes above.
253 */
254 int extra_free_kbytes = 0;
255
256 static unsigned long __meminitdata nr_kernel_pages;
257 static unsigned long __meminitdata nr_all_pages;
258 static unsigned long __meminitdata dma_reserve;
259
260 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
261 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
262 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
263 static unsigned long __initdata required_kernelcore;
264 static unsigned long __initdata required_movablecore;
265 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
266
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 int movable_zone;
269 EXPORT_SYMBOL(movable_zone);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
271
272 #if MAX_NUMNODES > 1
273 int nr_node_ids __read_mostly = MAX_NUMNODES;
274 int nr_online_nodes __read_mostly = 1;
275 EXPORT_SYMBOL(nr_node_ids);
276 EXPORT_SYMBOL(nr_online_nodes);
277 #endif
278
279 int page_group_by_mobility_disabled __read_mostly;
280
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
282 static inline void reset_deferred_meminit(pg_data_t *pgdat)
283 {
284 pgdat->first_deferred_pfn = ULONG_MAX;
285 }
286
287 /* Returns true if the struct page for the pfn is uninitialised */
288 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
289 {
290 int nid = early_pfn_to_nid(pfn);
291
292 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
293 return true;
294
295 return false;
296 }
297
298 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
299 {
300 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
301 return true;
302
303 return false;
304 }
305
306 /*
307 * Returns false when the remaining initialisation should be deferred until
308 * later in the boot cycle when it can be parallelised.
309 */
310 static inline bool update_defer_init(pg_data_t *pgdat,
311 unsigned long pfn, unsigned long zone_end,
312 unsigned long *nr_initialised)
313 {
314 /* Always populate low zones for address-contrained allocations */
315 if (zone_end < pgdat_end_pfn(pgdat))
316 return true;
317
318 /* Initialise at least 2G of the highest zone */
319 (*nr_initialised)++;
320 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
321 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
322 pgdat->first_deferred_pfn = pfn;
323 return false;
324 }
325
326 return true;
327 }
328 #else
329 static inline void reset_deferred_meminit(pg_data_t *pgdat)
330 {
331 }
332
333 static inline bool early_page_uninitialised(unsigned long pfn)
334 {
335 return false;
336 }
337
338 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
339 {
340 return false;
341 }
342
343 static inline bool update_defer_init(pg_data_t *pgdat,
344 unsigned long pfn, unsigned long zone_end,
345 unsigned long *nr_initialised)
346 {
347 return true;
348 }
349 #endif
350
351
352 void set_pageblock_migratetype(struct page *page, int migratetype)
353 {
354 if (unlikely(page_group_by_mobility_disabled &&
355 migratetype < MIGRATE_PCPTYPES))
356 migratetype = MIGRATE_UNMOVABLE;
357
358 set_pageblock_flags_group(page, (unsigned long)migratetype,
359 PB_migrate, PB_migrate_end);
360 }
361
362 #ifdef CONFIG_DEBUG_VM
363 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
364 {
365 int ret = 0;
366 unsigned seq;
367 unsigned long pfn = page_to_pfn(page);
368 unsigned long sp, start_pfn;
369
370 do {
371 seq = zone_span_seqbegin(zone);
372 start_pfn = zone->zone_start_pfn;
373 sp = zone->spanned_pages;
374 if (!zone_spans_pfn(zone, pfn))
375 ret = 1;
376 } while (zone_span_seqretry(zone, seq));
377
378 if (ret)
379 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
380 pfn, zone_to_nid(zone), zone->name,
381 start_pfn, start_pfn + sp);
382
383 return ret;
384 }
385
386 static int page_is_consistent(struct zone *zone, struct page *page)
387 {
388 if (!pfn_valid_within(page_to_pfn(page)))
389 return 0;
390 if (zone != page_zone(page))
391 return 0;
392
393 return 1;
394 }
395 /*
396 * Temporary debugging check for pages not lying within a given zone.
397 */
398 static int bad_range(struct zone *zone, struct page *page)
399 {
400 if (page_outside_zone_boundaries(zone, page))
401 return 1;
402 if (!page_is_consistent(zone, page))
403 return 1;
404
405 return 0;
406 }
407 #else
408 static inline int bad_range(struct zone *zone, struct page *page)
409 {
410 return 0;
411 }
412 #endif
413
414 static void bad_page(struct page *page, const char *reason,
415 unsigned long bad_flags)
416 {
417 static unsigned long resume;
418 static unsigned long nr_shown;
419 static unsigned long nr_unshown;
420
421 /* Don't complain about poisoned pages */
422 if (PageHWPoison(page)) {
423 page_mapcount_reset(page); /* remove PageBuddy */
424 return;
425 }
426
427 /*
428 * Allow a burst of 60 reports, then keep quiet for that minute;
429 * or allow a steady drip of one report per second.
430 */
431 if (nr_shown == 60) {
432 if (time_before(jiffies, resume)) {
433 nr_unshown++;
434 goto out;
435 }
436 if (nr_unshown) {
437 printk(KERN_ALERT
438 "BUG: Bad page state: %lu messages suppressed\n",
439 nr_unshown);
440 nr_unshown = 0;
441 }
442 nr_shown = 0;
443 }
444 if (nr_shown++ == 0)
445 resume = jiffies + 60 * HZ;
446
447 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
448 current->comm, page_to_pfn(page));
449 dump_page_badflags(page, reason, bad_flags);
450
451 print_modules();
452 dump_stack();
453 out:
454 /* Leave bad fields for debug, except PageBuddy could make trouble */
455 page_mapcount_reset(page); /* remove PageBuddy */
456 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
457 }
458
459 /*
460 * Higher-order pages are called "compound pages". They are structured thusly:
461 *
462 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
463 *
464 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
465 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
466 *
467 * The first tail page's ->compound_dtor holds the offset in array of compound
468 * page destructors. See compound_page_dtors.
469 *
470 * The first tail page's ->compound_order holds the order of allocation.
471 * This usage means that zero-order pages may not be compound.
472 */
473
474 static void free_compound_page(struct page *page)
475 {
476 __free_pages_ok(page, compound_order(page));
477 }
478
479 void prep_compound_page(struct page *page, unsigned int order)
480 {
481 int i;
482 int nr_pages = 1 << order;
483
484 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
485 set_compound_order(page, order);
486 __SetPageHead(page);
487 for (i = 1; i < nr_pages; i++) {
488 struct page *p = page + i;
489 set_page_count(p, 0);
490 set_compound_head(p, page);
491 }
492 }
493
494 #ifdef CONFIG_DEBUG_PAGEALLOC
495 unsigned int _debug_guardpage_minorder;
496 bool _debug_pagealloc_enabled __read_mostly;
497 bool _debug_guardpage_enabled __read_mostly;
498
499 static int __init early_debug_pagealloc(char *buf)
500 {
501 if (!buf)
502 return -EINVAL;
503
504 if (strcmp(buf, "on") == 0)
505 _debug_pagealloc_enabled = true;
506
507 return 0;
508 }
509 early_param("debug_pagealloc", early_debug_pagealloc);
510
511 static bool need_debug_guardpage(void)
512 {
513 /* If we don't use debug_pagealloc, we don't need guard page */
514 if (!debug_pagealloc_enabled())
515 return false;
516
517 return true;
518 }
519
520 static void init_debug_guardpage(void)
521 {
522 if (!debug_pagealloc_enabled())
523 return;
524
525 _debug_guardpage_enabled = true;
526 }
527
528 struct page_ext_operations debug_guardpage_ops = {
529 .need = need_debug_guardpage,
530 .init = init_debug_guardpage,
531 };
532
533 static int __init debug_guardpage_minorder_setup(char *buf)
534 {
535 unsigned long res;
536
537 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
538 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
539 return 0;
540 }
541 _debug_guardpage_minorder = res;
542 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
543 return 0;
544 }
545 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
546
547 static inline void set_page_guard(struct zone *zone, struct page *page,
548 unsigned int order, int migratetype)
549 {
550 struct page_ext *page_ext;
551
552 if (!debug_guardpage_enabled())
553 return;
554
555 page_ext = lookup_page_ext(page);
556 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
557
558 INIT_LIST_HEAD(&page->lru);
559 set_page_private(page, order);
560 /* Guard pages are not available for any usage */
561 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
562 }
563
564 static inline void clear_page_guard(struct zone *zone, struct page *page,
565 unsigned int order, int migratetype)
566 {
567 struct page_ext *page_ext;
568
569 if (!debug_guardpage_enabled())
570 return;
571
572 page_ext = lookup_page_ext(page);
573 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
574
575 set_page_private(page, 0);
576 if (!is_migrate_isolate(migratetype))
577 __mod_zone_freepage_state(zone, (1 << order), migratetype);
578 }
579 #else
580 struct page_ext_operations debug_guardpage_ops = { NULL, };
581 static inline void set_page_guard(struct zone *zone, struct page *page,
582 unsigned int order, int migratetype) {}
583 static inline void clear_page_guard(struct zone *zone, struct page *page,
584 unsigned int order, int migratetype) {}
585 #endif
586
587 static inline void set_page_order(struct page *page, unsigned int order)
588 {
589 set_page_private(page, order);
590 __SetPageBuddy(page);
591 }
592
593 static inline void rmv_page_order(struct page *page)
594 {
595 __ClearPageBuddy(page);
596 set_page_private(page, 0);
597 }
598
599 /*
600 * This function checks whether a page is free && is the buddy
601 * we can do coalesce a page and its buddy if
602 * (a) the buddy is not in a hole &&
603 * (b) the buddy is in the buddy system &&
604 * (c) a page and its buddy have the same order &&
605 * (d) a page and its buddy are in the same zone.
606 *
607 * For recording whether a page is in the buddy system, we set ->_mapcount
608 * PAGE_BUDDY_MAPCOUNT_VALUE.
609 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
610 * serialized by zone->lock.
611 *
612 * For recording page's order, we use page_private(page).
613 */
614 static inline int page_is_buddy(struct page *page, struct page *buddy,
615 unsigned int order)
616 {
617 if (!pfn_valid_within(page_to_pfn(buddy)))
618 return 0;
619
620 if (page_is_guard(buddy) && page_order(buddy) == order) {
621 if (page_zone_id(page) != page_zone_id(buddy))
622 return 0;
623
624 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
625
626 return 1;
627 }
628
629 if (PageBuddy(buddy) && page_order(buddy) == order) {
630 /*
631 * zone check is done late to avoid uselessly
632 * calculating zone/node ids for pages that could
633 * never merge.
634 */
635 if (page_zone_id(page) != page_zone_id(buddy))
636 return 0;
637
638 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
639
640 return 1;
641 }
642 return 0;
643 }
644
645 /*
646 * Freeing function for a buddy system allocator.
647 *
648 * The concept of a buddy system is to maintain direct-mapped table
649 * (containing bit values) for memory blocks of various "orders".
650 * The bottom level table contains the map for the smallest allocatable
651 * units of memory (here, pages), and each level above it describes
652 * pairs of units from the levels below, hence, "buddies".
653 * At a high level, all that happens here is marking the table entry
654 * at the bottom level available, and propagating the changes upward
655 * as necessary, plus some accounting needed to play nicely with other
656 * parts of the VM system.
657 * At each level, we keep a list of pages, which are heads of continuous
658 * free pages of length of (1 << order) and marked with _mapcount
659 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
660 * field.
661 * So when we are allocating or freeing one, we can derive the state of the
662 * other. That is, if we allocate a small block, and both were
663 * free, the remainder of the region must be split into blocks.
664 * If a block is freed, and its buddy is also free, then this
665 * triggers coalescing into a block of larger size.
666 *
667 * -- nyc
668 */
669
670 static inline void __free_one_page(struct page *page,
671 unsigned long pfn,
672 struct zone *zone, unsigned int order,
673 int migratetype)
674 {
675 unsigned long page_idx;
676 unsigned long combined_idx;
677 unsigned long uninitialized_var(buddy_idx);
678 struct page *buddy;
679 unsigned int max_order;
680
681 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
682
683 VM_BUG_ON(!zone_is_initialized(zone));
684 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
685
686 VM_BUG_ON(migratetype == -1);
687 if (likely(!is_migrate_isolate(migratetype)))
688 __mod_zone_freepage_state(zone, 1 << order, migratetype);
689
690 page_idx = pfn & ((1 << MAX_ORDER) - 1);
691
692 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
693 VM_BUG_ON_PAGE(bad_range(zone, page), page);
694
695 continue_merging:
696 while (order < max_order - 1) {
697 buddy_idx = __find_buddy_index(page_idx, order);
698 buddy = page + (buddy_idx - page_idx);
699 if (!page_is_buddy(page, buddy, order))
700 goto done_merging;
701 /*
702 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
703 * merge with it and move up one order.
704 */
705 if (page_is_guard(buddy)) {
706 clear_page_guard(zone, buddy, order, migratetype);
707 } else {
708 list_del(&buddy->lru);
709 zone->free_area[order].nr_free--;
710 rmv_page_order(buddy);
711 }
712 combined_idx = buddy_idx & page_idx;
713 page = page + (combined_idx - page_idx);
714 page_idx = combined_idx;
715 order++;
716 }
717 if (max_order < MAX_ORDER) {
718 /* If we are here, it means order is >= pageblock_order.
719 * We want to prevent merge between freepages on isolate
720 * pageblock and normal pageblock. Without this, pageblock
721 * isolation could cause incorrect freepage or CMA accounting.
722 *
723 * We don't want to hit this code for the more frequent
724 * low-order merging.
725 */
726 if (unlikely(has_isolate_pageblock(zone))) {
727 int buddy_mt;
728
729 buddy_idx = __find_buddy_index(page_idx, order);
730 buddy = page + (buddy_idx - page_idx);
731 buddy_mt = get_pageblock_migratetype(buddy);
732
733 if (migratetype != buddy_mt
734 && (is_migrate_isolate(migratetype) ||
735 is_migrate_isolate(buddy_mt)))
736 goto done_merging;
737 }
738 max_order++;
739 goto continue_merging;
740 }
741
742 done_merging:
743 set_page_order(page, order);
744
745 /*
746 * If this is not the largest possible page, check if the buddy
747 * of the next-highest order is free. If it is, it's possible
748 * that pages are being freed that will coalesce soon. In case,
749 * that is happening, add the free page to the tail of the list
750 * so it's less likely to be used soon and more likely to be merged
751 * as a higher order page
752 */
753 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
754 struct page *higher_page, *higher_buddy;
755 combined_idx = buddy_idx & page_idx;
756 higher_page = page + (combined_idx - page_idx);
757 buddy_idx = __find_buddy_index(combined_idx, order + 1);
758 higher_buddy = higher_page + (buddy_idx - combined_idx);
759 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
760 list_add_tail(&page->lru,
761 &zone->free_area[order].free_list[migratetype]);
762 goto out;
763 }
764 }
765
766 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
767 out:
768 zone->free_area[order].nr_free++;
769 }
770
771 static inline int free_pages_check(struct page *page)
772 {
773 const char *bad_reason = NULL;
774 unsigned long bad_flags = 0;
775
776 if (unlikely(page_mapcount(page)))
777 bad_reason = "nonzero mapcount";
778 if (unlikely(page->mapping != NULL))
779 bad_reason = "non-NULL mapping";
780 if (unlikely(atomic_read(&page->_count) != 0))
781 bad_reason = "nonzero _count";
782 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
783 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
784 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
785 }
786 #ifdef CONFIG_MEMCG
787 if (unlikely(page->mem_cgroup))
788 bad_reason = "page still charged to cgroup";
789 #endif
790 if (unlikely(bad_reason)) {
791 bad_page(page, bad_reason, bad_flags);
792 return 1;
793 }
794 page_cpupid_reset_last(page);
795 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
796 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
797 return 0;
798 }
799
800 /*
801 * Frees a number of pages from the PCP lists
802 * Assumes all pages on list are in same zone, and of same order.
803 * count is the number of pages to free.
804 *
805 * If the zone was previously in an "all pages pinned" state then look to
806 * see if this freeing clears that state.
807 *
808 * And clear the zone's pages_scanned counter, to hold off the "all pages are
809 * pinned" detection logic.
810 */
811 static void free_pcppages_bulk(struct zone *zone, int count,
812 struct per_cpu_pages *pcp)
813 {
814 int migratetype = 0;
815 int batch_free = 0;
816 int to_free = count;
817 unsigned long nr_scanned;
818
819 spin_lock(&zone->lock);
820 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
821 if (nr_scanned)
822 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
823
824 while (to_free) {
825 struct page *page;
826 struct list_head *list;
827
828 /*
829 * Remove pages from lists in a round-robin fashion. A
830 * batch_free count is maintained that is incremented when an
831 * empty list is encountered. This is so more pages are freed
832 * off fuller lists instead of spinning excessively around empty
833 * lists
834 */
835 do {
836 batch_free++;
837 if (++migratetype == MIGRATE_PCPTYPES)
838 migratetype = 0;
839 list = &pcp->lists[migratetype];
840 } while (list_empty(list));
841
842 /* This is the only non-empty list. Free them all. */
843 if (batch_free == MIGRATE_PCPTYPES)
844 batch_free = to_free;
845
846 do {
847 int mt; /* migratetype of the to-be-freed page */
848
849 page = list_entry(list->prev, struct page, lru);
850 /* must delete as __free_one_page list manipulates */
851 list_del(&page->lru);
852
853 mt = get_pcppage_migratetype(page);
854 /* MIGRATE_ISOLATE page should not go to pcplists */
855 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
856 /* Pageblock could have been isolated meanwhile */
857 if (unlikely(has_isolate_pageblock(zone)))
858 mt = get_pageblock_migratetype(page);
859
860 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
861 trace_mm_page_pcpu_drain(page, 0, mt);
862 } while (--to_free && --batch_free && !list_empty(list));
863 }
864 spin_unlock(&zone->lock);
865 }
866
867 static void free_one_page(struct zone *zone,
868 struct page *page, unsigned long pfn,
869 unsigned int order,
870 int migratetype)
871 {
872 unsigned long nr_scanned;
873 spin_lock(&zone->lock);
874 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
875 if (nr_scanned)
876 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
877
878 if (unlikely(has_isolate_pageblock(zone) ||
879 is_migrate_isolate(migratetype))) {
880 migratetype = get_pfnblock_migratetype(page, pfn);
881 }
882 __free_one_page(page, pfn, zone, order, migratetype);
883 spin_unlock(&zone->lock);
884 }
885
886 static int free_tail_pages_check(struct page *head_page, struct page *page)
887 {
888 int ret = 1;
889
890 /*
891 * We rely page->lru.next never has bit 0 set, unless the page
892 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
893 */
894 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
895
896 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
897 ret = 0;
898 goto out;
899 }
900 if (unlikely(!PageTail(page))) {
901 bad_page(page, "PageTail not set", 0);
902 goto out;
903 }
904 if (unlikely(compound_head(page) != head_page)) {
905 bad_page(page, "compound_head not consistent", 0);
906 goto out;
907 }
908 ret = 0;
909 out:
910 clear_compound_head(page);
911 return ret;
912 }
913
914 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
915 unsigned long zone, int nid)
916 {
917 set_page_links(page, zone, nid, pfn);
918 init_page_count(page);
919 page_mapcount_reset(page);
920 page_cpupid_reset_last(page);
921
922 INIT_LIST_HEAD(&page->lru);
923 #ifdef WANT_PAGE_VIRTUAL
924 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
925 if (!is_highmem_idx(zone))
926 set_page_address(page, __va(pfn << PAGE_SHIFT));
927 #endif
928 }
929
930 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
931 int nid)
932 {
933 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
934 }
935
936 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
937 static void init_reserved_page(unsigned long pfn)
938 {
939 pg_data_t *pgdat;
940 int nid, zid;
941
942 if (!early_page_uninitialised(pfn))
943 return;
944
945 nid = early_pfn_to_nid(pfn);
946 pgdat = NODE_DATA(nid);
947
948 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
949 struct zone *zone = &pgdat->node_zones[zid];
950
951 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
952 break;
953 }
954 __init_single_pfn(pfn, zid, nid);
955 }
956 #else
957 static inline void init_reserved_page(unsigned long pfn)
958 {
959 }
960 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
961
962 /*
963 * Initialised pages do not have PageReserved set. This function is
964 * called for each range allocated by the bootmem allocator and
965 * marks the pages PageReserved. The remaining valid pages are later
966 * sent to the buddy page allocator.
967 */
968 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
969 {
970 unsigned long start_pfn = PFN_DOWN(start);
971 unsigned long end_pfn = PFN_UP(end);
972
973 for (; start_pfn < end_pfn; start_pfn++) {
974 if (pfn_valid(start_pfn)) {
975 struct page *page = pfn_to_page(start_pfn);
976
977 init_reserved_page(start_pfn);
978
979 /* Avoid false-positive PageTail() */
980 INIT_LIST_HEAD(&page->lru);
981
982 SetPageReserved(page);
983 }
984 }
985 }
986
987 static bool free_pages_prepare(struct page *page, unsigned int order)
988 {
989 bool compound = PageCompound(page);
990 int i, bad = 0;
991
992 VM_BUG_ON_PAGE(PageTail(page), page);
993 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
994
995 trace_mm_page_free(page, order);
996 kmemcheck_free_shadow(page, order);
997 kasan_free_pages(page, order);
998
999 if (PageAnon(page))
1000 page->mapping = NULL;
1001 bad += free_pages_check(page);
1002 for (i = 1; i < (1 << order); i++) {
1003 if (compound)
1004 bad += free_tail_pages_check(page, page + i);
1005 bad += free_pages_check(page + i);
1006 }
1007 if (bad)
1008 return false;
1009
1010 reset_page_owner(page, order);
1011
1012 if (!PageHighMem(page)) {
1013 debug_check_no_locks_freed(page_address(page),
1014 PAGE_SIZE << order);
1015 debug_check_no_obj_freed(page_address(page),
1016 PAGE_SIZE << order);
1017 }
1018 arch_free_page(page, order);
1019 kernel_map_pages(page, 1 << order, 0);
1020
1021 return true;
1022 }
1023
1024 static void __free_pages_ok(struct page *page, unsigned int order)
1025 {
1026 unsigned long flags;
1027 int migratetype;
1028 unsigned long pfn = page_to_pfn(page);
1029
1030 if (!free_pages_prepare(page, order))
1031 return;
1032
1033 migratetype = get_pfnblock_migratetype(page, pfn);
1034 local_irq_save(flags);
1035 __count_vm_events(PGFREE, 1 << order);
1036 free_one_page(page_zone(page), page, pfn, order, migratetype);
1037 local_irq_restore(flags);
1038 }
1039
1040 static void __init __free_pages_boot_core(struct page *page,
1041 unsigned long pfn, unsigned int order)
1042 {
1043 unsigned int nr_pages = 1 << order;
1044 struct page *p = page;
1045 unsigned int loop;
1046
1047 prefetchw(p);
1048 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1049 prefetchw(p + 1);
1050 __ClearPageReserved(p);
1051 set_page_count(p, 0);
1052 }
1053 __ClearPageReserved(p);
1054 set_page_count(p, 0);
1055
1056 page_zone(page)->managed_pages += nr_pages;
1057 set_page_refcounted(page);
1058 __free_pages(page, order);
1059 }
1060
1061 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1062 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1063
1064 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1065
1066 int __meminit early_pfn_to_nid(unsigned long pfn)
1067 {
1068 static DEFINE_SPINLOCK(early_pfn_lock);
1069 int nid;
1070
1071 spin_lock(&early_pfn_lock);
1072 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1073 if (nid < 0)
1074 nid = first_online_node;
1075 spin_unlock(&early_pfn_lock);
1076
1077 return nid;
1078 }
1079 #endif
1080
1081 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1082 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1083 struct mminit_pfnnid_cache *state)
1084 {
1085 int nid;
1086
1087 nid = __early_pfn_to_nid(pfn, state);
1088 if (nid >= 0 && nid != node)
1089 return false;
1090 return true;
1091 }
1092
1093 /* Only safe to use early in boot when initialisation is single-threaded */
1094 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1095 {
1096 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1097 }
1098
1099 #else
1100
1101 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1102 {
1103 return true;
1104 }
1105 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1106 struct mminit_pfnnid_cache *state)
1107 {
1108 return true;
1109 }
1110 #endif
1111
1112
1113 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1114 unsigned int order)
1115 {
1116 if (early_page_uninitialised(pfn))
1117 return;
1118 return __free_pages_boot_core(page, pfn, order);
1119 }
1120
1121 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1122 static void __init deferred_free_range(struct page *page,
1123 unsigned long pfn, int nr_pages)
1124 {
1125 int i;
1126
1127 if (!page)
1128 return;
1129
1130 /* Free a large naturally-aligned chunk if possible */
1131 if (nr_pages == MAX_ORDER_NR_PAGES &&
1132 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1133 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1134 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1135 return;
1136 }
1137
1138 for (i = 0; i < nr_pages; i++, page++, pfn++)
1139 __free_pages_boot_core(page, pfn, 0);
1140 }
1141
1142 /* Completion tracking for deferred_init_memmap() threads */
1143 static atomic_t pgdat_init_n_undone __initdata;
1144 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1145
1146 static inline void __init pgdat_init_report_one_done(void)
1147 {
1148 if (atomic_dec_and_test(&pgdat_init_n_undone))
1149 complete(&pgdat_init_all_done_comp);
1150 }
1151
1152 /* Initialise remaining memory on a node */
1153 static int __init deferred_init_memmap(void *data)
1154 {
1155 pg_data_t *pgdat = data;
1156 int nid = pgdat->node_id;
1157 struct mminit_pfnnid_cache nid_init_state = { };
1158 unsigned long start = jiffies;
1159 unsigned long nr_pages = 0;
1160 unsigned long walk_start, walk_end;
1161 int i, zid;
1162 struct zone *zone;
1163 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1164 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1165
1166 if (first_init_pfn == ULONG_MAX) {
1167 pgdat_init_report_one_done();
1168 return 0;
1169 }
1170
1171 /* Bind memory initialisation thread to a local node if possible */
1172 if (!cpumask_empty(cpumask))
1173 set_cpus_allowed_ptr(current, cpumask);
1174
1175 /* Sanity check boundaries */
1176 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1177 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1178 pgdat->first_deferred_pfn = ULONG_MAX;
1179
1180 /* Only the highest zone is deferred so find it */
1181 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1182 zone = pgdat->node_zones + zid;
1183 if (first_init_pfn < zone_end_pfn(zone))
1184 break;
1185 }
1186
1187 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1188 unsigned long pfn, end_pfn;
1189 struct page *page = NULL;
1190 struct page *free_base_page = NULL;
1191 unsigned long free_base_pfn = 0;
1192 int nr_to_free = 0;
1193
1194 end_pfn = min(walk_end, zone_end_pfn(zone));
1195 pfn = first_init_pfn;
1196 if (pfn < walk_start)
1197 pfn = walk_start;
1198 if (pfn < zone->zone_start_pfn)
1199 pfn = zone->zone_start_pfn;
1200
1201 for (; pfn < end_pfn; pfn++) {
1202 if (!pfn_valid_within(pfn))
1203 goto free_range;
1204
1205 /*
1206 * Ensure pfn_valid is checked every
1207 * MAX_ORDER_NR_PAGES for memory holes
1208 */
1209 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1210 if (!pfn_valid(pfn)) {
1211 page = NULL;
1212 goto free_range;
1213 }
1214 }
1215
1216 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1217 page = NULL;
1218 goto free_range;
1219 }
1220
1221 /* Minimise pfn page lookups and scheduler checks */
1222 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1223 page++;
1224 } else {
1225 nr_pages += nr_to_free;
1226 deferred_free_range(free_base_page,
1227 free_base_pfn, nr_to_free);
1228 free_base_page = NULL;
1229 free_base_pfn = nr_to_free = 0;
1230
1231 page = pfn_to_page(pfn);
1232 cond_resched();
1233 }
1234
1235 if (page->flags) {
1236 VM_BUG_ON(page_zone(page) != zone);
1237 goto free_range;
1238 }
1239
1240 __init_single_page(page, pfn, zid, nid);
1241 if (!free_base_page) {
1242 free_base_page = page;
1243 free_base_pfn = pfn;
1244 nr_to_free = 0;
1245 }
1246 nr_to_free++;
1247
1248 /* Where possible, batch up pages for a single free */
1249 continue;
1250 free_range:
1251 /* Free the current block of pages to allocator */
1252 nr_pages += nr_to_free;
1253 deferred_free_range(free_base_page, free_base_pfn,
1254 nr_to_free);
1255 free_base_page = NULL;
1256 free_base_pfn = nr_to_free = 0;
1257 }
1258
1259 first_init_pfn = max(end_pfn, first_init_pfn);
1260 }
1261
1262 /* Sanity check that the next zone really is unpopulated */
1263 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1264
1265 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1266 jiffies_to_msecs(jiffies - start));
1267
1268 pgdat_init_report_one_done();
1269 return 0;
1270 }
1271
1272 void __init page_alloc_init_late(void)
1273 {
1274 int nid;
1275
1276 /* There will be num_node_state(N_MEMORY) threads */
1277 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1278 for_each_node_state(nid, N_MEMORY) {
1279 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1280 }
1281
1282 /* Block until all are initialised */
1283 wait_for_completion(&pgdat_init_all_done_comp);
1284
1285 /* Reinit limits that are based on free pages after the kernel is up */
1286 files_maxfiles_init();
1287 }
1288 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1289
1290 #ifdef CONFIG_CMA
1291 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1292 void __init init_cma_reserved_pageblock(struct page *page)
1293 {
1294 unsigned i = pageblock_nr_pages;
1295 struct page *p = page;
1296
1297 do {
1298 __ClearPageReserved(p);
1299 set_page_count(p, 0);
1300 } while (++p, --i);
1301
1302 set_pageblock_migratetype(page, MIGRATE_CMA);
1303
1304 if (pageblock_order >= MAX_ORDER) {
1305 i = pageblock_nr_pages;
1306 p = page;
1307 do {
1308 set_page_refcounted(p);
1309 __free_pages(p, MAX_ORDER - 1);
1310 p += MAX_ORDER_NR_PAGES;
1311 } while (i -= MAX_ORDER_NR_PAGES);
1312 } else {
1313 set_page_refcounted(page);
1314 __free_pages(page, pageblock_order);
1315 }
1316
1317 adjust_managed_page_count(page, pageblock_nr_pages);
1318 }
1319 #endif
1320
1321 /*
1322 * The order of subdivision here is critical for the IO subsystem.
1323 * Please do not alter this order without good reasons and regression
1324 * testing. Specifically, as large blocks of memory are subdivided,
1325 * the order in which smaller blocks are delivered depends on the order
1326 * they're subdivided in this function. This is the primary factor
1327 * influencing the order in which pages are delivered to the IO
1328 * subsystem according to empirical testing, and this is also justified
1329 * by considering the behavior of a buddy system containing a single
1330 * large block of memory acted on by a series of small allocations.
1331 * This behavior is a critical factor in sglist merging's success.
1332 *
1333 * -- nyc
1334 */
1335 static inline void expand(struct zone *zone, struct page *page,
1336 int low, int high, struct free_area *area,
1337 int migratetype)
1338 {
1339 unsigned long size = 1 << high;
1340
1341 while (high > low) {
1342 area--;
1343 high--;
1344 size >>= 1;
1345 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1346
1347 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1348 debug_guardpage_enabled() &&
1349 high < debug_guardpage_minorder()) {
1350 /*
1351 * Mark as guard pages (or page), that will allow to
1352 * merge back to allocator when buddy will be freed.
1353 * Corresponding page table entries will not be touched,
1354 * pages will stay not present in virtual address space
1355 */
1356 set_page_guard(zone, &page[size], high, migratetype);
1357 continue;
1358 }
1359 list_add(&page[size].lru, &area->free_list[migratetype]);
1360 area->nr_free++;
1361 set_page_order(&page[size], high);
1362 }
1363 }
1364
1365 /*
1366 * This page is about to be returned from the page allocator
1367 */
1368 static inline int check_new_page(struct page *page)
1369 {
1370 const char *bad_reason = NULL;
1371 unsigned long bad_flags = 0;
1372
1373 if (unlikely(page_mapcount(page)))
1374 bad_reason = "nonzero mapcount";
1375 if (unlikely(page->mapping != NULL))
1376 bad_reason = "non-NULL mapping";
1377 if (unlikely(atomic_read(&page->_count) != 0))
1378 bad_reason = "nonzero _count";
1379 if (unlikely(page->flags & __PG_HWPOISON)) {
1380 bad_reason = "HWPoisoned (hardware-corrupted)";
1381 bad_flags = __PG_HWPOISON;
1382 }
1383 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1384 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1385 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1386 }
1387 #ifdef CONFIG_MEMCG
1388 if (unlikely(page->mem_cgroup))
1389 bad_reason = "page still charged to cgroup";
1390 #endif
1391 if (unlikely(bad_reason)) {
1392 bad_page(page, bad_reason, bad_flags);
1393 return 1;
1394 }
1395 return 0;
1396 }
1397
1398 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1399 int alloc_flags)
1400 {
1401 int i;
1402
1403 for (i = 0; i < (1 << order); i++) {
1404 struct page *p = page + i;
1405 if (unlikely(check_new_page(p)))
1406 return 1;
1407 }
1408
1409 set_page_private(page, 0);
1410 set_page_refcounted(page);
1411
1412 arch_alloc_page(page, order);
1413 kernel_map_pages(page, 1 << order, 1);
1414 kasan_alloc_pages(page, order);
1415
1416 if (gfp_flags & __GFP_ZERO)
1417 for (i = 0; i < (1 << order); i++)
1418 clear_highpage(page + i);
1419
1420 if (order && (gfp_flags & __GFP_COMP))
1421 prep_compound_page(page, order);
1422
1423 set_page_owner(page, order, gfp_flags);
1424
1425 /*
1426 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1427 * allocate the page. The expectation is that the caller is taking
1428 * steps that will free more memory. The caller should avoid the page
1429 * being used for !PFMEMALLOC purposes.
1430 */
1431 if (alloc_flags & ALLOC_NO_WATERMARKS)
1432 set_page_pfmemalloc(page);
1433 else
1434 clear_page_pfmemalloc(page);
1435
1436 return 0;
1437 }
1438
1439 /*
1440 * Go through the free lists for the given migratetype and remove
1441 * the smallest available page from the freelists
1442 */
1443 static inline
1444 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1445 int migratetype)
1446 {
1447 unsigned int current_order;
1448 struct free_area *area;
1449 struct page *page;
1450
1451 /* Find a page of the appropriate size in the preferred list */
1452 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1453 area = &(zone->free_area[current_order]);
1454 if (list_empty(&area->free_list[migratetype]))
1455 continue;
1456
1457 page = list_entry(area->free_list[migratetype].next,
1458 struct page, lru);
1459 list_del(&page->lru);
1460 rmv_page_order(page);
1461 area->nr_free--;
1462 expand(zone, page, order, current_order, area, migratetype);
1463 set_pcppage_migratetype(page, migratetype);
1464 return page;
1465 }
1466
1467 return NULL;
1468 }
1469
1470
1471 /*
1472 * This array describes the order lists are fallen back to when
1473 * the free lists for the desirable migrate type are depleted
1474 */
1475 static int fallbacks[MIGRATE_TYPES][4] = {
1476 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1477 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1478 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1479 #ifdef CONFIG_CMA
1480 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1481 #endif
1482 #ifdef CONFIG_MEMORY_ISOLATION
1483 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1484 #endif
1485 };
1486
1487 #ifdef CONFIG_CMA
1488 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1489 unsigned int order)
1490 {
1491 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1492 }
1493 #else
1494 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1495 unsigned int order) { return NULL; }
1496 #endif
1497
1498 /*
1499 * Move the free pages in a range to the free lists of the requested type.
1500 * Note that start_page and end_pages are not aligned on a pageblock
1501 * boundary. If alignment is required, use move_freepages_block()
1502 */
1503 int move_freepages(struct zone *zone,
1504 struct page *start_page, struct page *end_page,
1505 int migratetype)
1506 {
1507 struct page *page;
1508 unsigned int order;
1509 int pages_moved = 0;
1510
1511 #ifndef CONFIG_HOLES_IN_ZONE
1512 /*
1513 * page_zone is not safe to call in this context when
1514 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1515 * anyway as we check zone boundaries in move_freepages_block().
1516 * Remove at a later date when no bug reports exist related to
1517 * grouping pages by mobility
1518 */
1519 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1520 #endif
1521
1522 for (page = start_page; page <= end_page;) {
1523 /* Make sure we are not inadvertently changing nodes */
1524 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1525
1526 if (!pfn_valid_within(page_to_pfn(page))) {
1527 page++;
1528 continue;
1529 }
1530
1531 if (!PageBuddy(page)) {
1532 page++;
1533 continue;
1534 }
1535
1536 order = page_order(page);
1537 list_move(&page->lru,
1538 &zone->free_area[order].free_list[migratetype]);
1539 page += 1 << order;
1540 pages_moved += 1 << order;
1541 }
1542
1543 return pages_moved;
1544 }
1545
1546 int move_freepages_block(struct zone *zone, struct page *page,
1547 int migratetype)
1548 {
1549 unsigned long start_pfn, end_pfn;
1550 struct page *start_page, *end_page;
1551
1552 start_pfn = page_to_pfn(page);
1553 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1554 start_page = pfn_to_page(start_pfn);
1555 end_page = start_page + pageblock_nr_pages - 1;
1556 end_pfn = start_pfn + pageblock_nr_pages - 1;
1557
1558 /* Do not cross zone boundaries */
1559 if (!zone_spans_pfn(zone, start_pfn))
1560 start_page = page;
1561 if (!zone_spans_pfn(zone, end_pfn))
1562 return 0;
1563
1564 return move_freepages(zone, start_page, end_page, migratetype);
1565 }
1566
1567 static void change_pageblock_range(struct page *pageblock_page,
1568 int start_order, int migratetype)
1569 {
1570 int nr_pageblocks = 1 << (start_order - pageblock_order);
1571
1572 while (nr_pageblocks--) {
1573 set_pageblock_migratetype(pageblock_page, migratetype);
1574 pageblock_page += pageblock_nr_pages;
1575 }
1576 }
1577
1578 /*
1579 * When we are falling back to another migratetype during allocation, try to
1580 * steal extra free pages from the same pageblocks to satisfy further
1581 * allocations, instead of polluting multiple pageblocks.
1582 *
1583 * If we are stealing a relatively large buddy page, it is likely there will
1584 * be more free pages in the pageblock, so try to steal them all. For
1585 * reclaimable and unmovable allocations, we steal regardless of page size,
1586 * as fragmentation caused by those allocations polluting movable pageblocks
1587 * is worse than movable allocations stealing from unmovable and reclaimable
1588 * pageblocks.
1589 */
1590 static bool can_steal_fallback(unsigned int order, int start_mt)
1591 {
1592 /*
1593 * Leaving this order check is intended, although there is
1594 * relaxed order check in next check. The reason is that
1595 * we can actually steal whole pageblock if this condition met,
1596 * but, below check doesn't guarantee it and that is just heuristic
1597 * so could be changed anytime.
1598 */
1599 if (order >= pageblock_order)
1600 return true;
1601
1602 if (order >= pageblock_order / 2 ||
1603 start_mt == MIGRATE_RECLAIMABLE ||
1604 start_mt == MIGRATE_UNMOVABLE ||
1605 page_group_by_mobility_disabled)
1606 return true;
1607
1608 return false;
1609 }
1610
1611 /*
1612 * This function implements actual steal behaviour. If order is large enough,
1613 * we can steal whole pageblock. If not, we first move freepages in this
1614 * pageblock and check whether half of pages are moved or not. If half of
1615 * pages are moved, we can change migratetype of pageblock and permanently
1616 * use it's pages as requested migratetype in the future.
1617 */
1618 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1619 int start_type)
1620 {
1621 unsigned int current_order = page_order(page);
1622 int pages;
1623
1624 /* Take ownership for orders >= pageblock_order */
1625 if (current_order >= pageblock_order) {
1626 change_pageblock_range(page, current_order, start_type);
1627 return;
1628 }
1629
1630 pages = move_freepages_block(zone, page, start_type);
1631
1632 /* Claim the whole block if over half of it is free */
1633 if (pages >= (1 << (pageblock_order-1)) ||
1634 page_group_by_mobility_disabled)
1635 set_pageblock_migratetype(page, start_type);
1636 }
1637
1638 /*
1639 * Check whether there is a suitable fallback freepage with requested order.
1640 * If only_stealable is true, this function returns fallback_mt only if
1641 * we can steal other freepages all together. This would help to reduce
1642 * fragmentation due to mixed migratetype pages in one pageblock.
1643 */
1644 int find_suitable_fallback(struct free_area *area, unsigned int order,
1645 int migratetype, bool only_stealable, bool *can_steal)
1646 {
1647 int i;
1648 int fallback_mt;
1649
1650 if (area->nr_free == 0)
1651 return -1;
1652
1653 *can_steal = false;
1654 for (i = 0;; i++) {
1655 fallback_mt = fallbacks[migratetype][i];
1656 if (fallback_mt == MIGRATE_TYPES)
1657 break;
1658
1659 if (list_empty(&area->free_list[fallback_mt]))
1660 continue;
1661
1662 if (can_steal_fallback(order, migratetype))
1663 *can_steal = true;
1664
1665 if (!only_stealable)
1666 return fallback_mt;
1667
1668 if (*can_steal)
1669 return fallback_mt;
1670 }
1671
1672 return -1;
1673 }
1674
1675 /*
1676 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1677 * there are no empty page blocks that contain a page with a suitable order
1678 */
1679 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1680 unsigned int alloc_order)
1681 {
1682 int mt;
1683 unsigned long max_managed, flags;
1684
1685 /*
1686 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1687 * Check is race-prone but harmless.
1688 */
1689 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1690 if (zone->nr_reserved_highatomic >= max_managed)
1691 return;
1692
1693 spin_lock_irqsave(&zone->lock, flags);
1694
1695 /* Recheck the nr_reserved_highatomic limit under the lock */
1696 if (zone->nr_reserved_highatomic >= max_managed)
1697 goto out_unlock;
1698
1699 /* Yoink! */
1700 mt = get_pageblock_migratetype(page);
1701 if (mt != MIGRATE_HIGHATOMIC &&
1702 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1703 zone->nr_reserved_highatomic += pageblock_nr_pages;
1704 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1705 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1706 }
1707
1708 out_unlock:
1709 spin_unlock_irqrestore(&zone->lock, flags);
1710 }
1711
1712 /*
1713 * Used when an allocation is about to fail under memory pressure. This
1714 * potentially hurts the reliability of high-order allocations when under
1715 * intense memory pressure but failed atomic allocations should be easier
1716 * to recover from than an OOM.
1717 */
1718 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1719 {
1720 struct zonelist *zonelist = ac->zonelist;
1721 unsigned long flags;
1722 struct zoneref *z;
1723 struct zone *zone;
1724 struct page *page;
1725 int order;
1726
1727 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1728 ac->nodemask) {
1729 /* Preserve at least one pageblock */
1730 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1731 continue;
1732
1733 spin_lock_irqsave(&zone->lock, flags);
1734 for (order = 0; order < MAX_ORDER; order++) {
1735 struct free_area *area = &(zone->free_area[order]);
1736
1737 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1738 continue;
1739
1740 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1741 struct page, lru);
1742
1743 /*
1744 * It should never happen but changes to locking could
1745 * inadvertently allow a per-cpu drain to add pages
1746 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1747 * and watch for underflows.
1748 */
1749 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1750 zone->nr_reserved_highatomic);
1751
1752 /*
1753 * Convert to ac->migratetype and avoid the normal
1754 * pageblock stealing heuristics. Minimally, the caller
1755 * is doing the work and needs the pages. More
1756 * importantly, if the block was always converted to
1757 * MIGRATE_UNMOVABLE or another type then the number
1758 * of pageblocks that cannot be completely freed
1759 * may increase.
1760 */
1761 set_pageblock_migratetype(page, ac->migratetype);
1762 move_freepages_block(zone, page, ac->migratetype);
1763 spin_unlock_irqrestore(&zone->lock, flags);
1764 return;
1765 }
1766 spin_unlock_irqrestore(&zone->lock, flags);
1767 }
1768 }
1769
1770 /* Remove an element from the buddy allocator from the fallback list */
1771 static inline struct page *
1772 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1773 {
1774 struct free_area *area;
1775 unsigned int current_order;
1776 struct page *page;
1777 int fallback_mt;
1778 bool can_steal;
1779
1780 /* Find the largest possible block of pages in the other list */
1781 for (current_order = MAX_ORDER-1;
1782 current_order >= order && current_order <= MAX_ORDER-1;
1783 --current_order) {
1784 area = &(zone->free_area[current_order]);
1785 fallback_mt = find_suitable_fallback(area, current_order,
1786 start_migratetype, false, &can_steal);
1787 if (fallback_mt == -1)
1788 continue;
1789
1790 page = list_entry(area->free_list[fallback_mt].next,
1791 struct page, lru);
1792 if (can_steal)
1793 steal_suitable_fallback(zone, page, start_migratetype);
1794
1795 /* Remove the page from the freelists */
1796 area->nr_free--;
1797 list_del(&page->lru);
1798 rmv_page_order(page);
1799
1800 expand(zone, page, order, current_order, area,
1801 start_migratetype);
1802 /*
1803 * The pcppage_migratetype may differ from pageblock's
1804 * migratetype depending on the decisions in
1805 * find_suitable_fallback(). This is OK as long as it does not
1806 * differ for MIGRATE_CMA pageblocks. Those can be used as
1807 * fallback only via special __rmqueue_cma_fallback() function
1808 */
1809 set_pcppage_migratetype(page, start_migratetype);
1810
1811 trace_mm_page_alloc_extfrag(page, order, current_order,
1812 start_migratetype, fallback_mt);
1813
1814 return page;
1815 }
1816
1817 return NULL;
1818 }
1819
1820 /*
1821 * Do the hard work of removing an element from the buddy allocator.
1822 * Call me with the zone->lock already held.
1823 */
1824 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1825 int migratetype, gfp_t gfp_flags)
1826 {
1827 struct page *page;
1828
1829 page = __rmqueue_smallest(zone, order, migratetype);
1830 if (unlikely(!page)) {
1831 if (migratetype == MIGRATE_MOVABLE)
1832 page = __rmqueue_cma_fallback(zone, order);
1833
1834 if (!page)
1835 page = __rmqueue_fallback(zone, order, migratetype);
1836 }
1837
1838 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1839 return page;
1840 }
1841
1842 /*
1843 * Obtain a specified number of elements from the buddy allocator, all under
1844 * a single hold of the lock, for efficiency. Add them to the supplied list.
1845 * Returns the number of new pages which were placed at *list.
1846 */
1847 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1848 unsigned long count, struct list_head *list,
1849 int migratetype, bool cold)
1850 {
1851 int i;
1852
1853 spin_lock(&zone->lock);
1854 for (i = 0; i < count; ++i) {
1855 struct page *page = __rmqueue(zone, order, migratetype, 0);
1856 if (unlikely(page == NULL))
1857 break;
1858
1859 /*
1860 * Split buddy pages returned by expand() are received here
1861 * in physical page order. The page is added to the callers and
1862 * list and the list head then moves forward. From the callers
1863 * perspective, the linked list is ordered by page number in
1864 * some conditions. This is useful for IO devices that can
1865 * merge IO requests if the physical pages are ordered
1866 * properly.
1867 */
1868 if (likely(!cold))
1869 list_add(&page->lru, list);
1870 else
1871 list_add_tail(&page->lru, list);
1872 list = &page->lru;
1873 if (is_migrate_cma(get_pcppage_migratetype(page)))
1874 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1875 -(1 << order));
1876 }
1877 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1878 spin_unlock(&zone->lock);
1879 return i;
1880 }
1881
1882 #ifdef CONFIG_NUMA
1883 /*
1884 * Called from the vmstat counter updater to drain pagesets of this
1885 * currently executing processor on remote nodes after they have
1886 * expired.
1887 *
1888 * Note that this function must be called with the thread pinned to
1889 * a single processor.
1890 */
1891 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1892 {
1893 unsigned long flags;
1894 int to_drain, batch;
1895
1896 local_irq_save(flags);
1897 batch = READ_ONCE(pcp->batch);
1898 to_drain = min(pcp->count, batch);
1899 if (to_drain > 0) {
1900 free_pcppages_bulk(zone, to_drain, pcp);
1901 pcp->count -= to_drain;
1902 }
1903 local_irq_restore(flags);
1904 }
1905 #endif
1906
1907 /*
1908 * Drain pcplists of the indicated processor and zone.
1909 *
1910 * The processor must either be the current processor and the
1911 * thread pinned to the current processor or a processor that
1912 * is not online.
1913 */
1914 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1915 {
1916 unsigned long flags;
1917 struct per_cpu_pageset *pset;
1918 struct per_cpu_pages *pcp;
1919
1920 local_irq_save(flags);
1921 pset = per_cpu_ptr(zone->pageset, cpu);
1922
1923 pcp = &pset->pcp;
1924 if (pcp->count) {
1925 free_pcppages_bulk(zone, pcp->count, pcp);
1926 pcp->count = 0;
1927 }
1928 local_irq_restore(flags);
1929 }
1930
1931 /*
1932 * Drain pcplists of all zones on the indicated processor.
1933 *
1934 * The processor must either be the current processor and the
1935 * thread pinned to the current processor or a processor that
1936 * is not online.
1937 */
1938 static void drain_pages(unsigned int cpu)
1939 {
1940 struct zone *zone;
1941
1942 for_each_populated_zone(zone) {
1943 drain_pages_zone(cpu, zone);
1944 }
1945 }
1946
1947 /*
1948 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1949 *
1950 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1951 * the single zone's pages.
1952 */
1953 void drain_local_pages(struct zone *zone)
1954 {
1955 int cpu = smp_processor_id();
1956
1957 if (zone)
1958 drain_pages_zone(cpu, zone);
1959 else
1960 drain_pages(cpu);
1961 }
1962
1963 /*
1964 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1965 *
1966 * When zone parameter is non-NULL, spill just the single zone's pages.
1967 *
1968 * Note that this code is protected against sending an IPI to an offline
1969 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1970 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1971 * nothing keeps CPUs from showing up after we populated the cpumask and
1972 * before the call to on_each_cpu_mask().
1973 */
1974 void drain_all_pages(struct zone *zone)
1975 {
1976 int cpu;
1977
1978 /*
1979 * Allocate in the BSS so we wont require allocation in
1980 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1981 */
1982 static cpumask_t cpus_with_pcps;
1983
1984 /*
1985 * We don't care about racing with CPU hotplug event
1986 * as offline notification will cause the notified
1987 * cpu to drain that CPU pcps and on_each_cpu_mask
1988 * disables preemption as part of its processing
1989 */
1990 for_each_online_cpu(cpu) {
1991 struct per_cpu_pageset *pcp;
1992 struct zone *z;
1993 bool has_pcps = false;
1994
1995 if (zone) {
1996 pcp = per_cpu_ptr(zone->pageset, cpu);
1997 if (pcp->pcp.count)
1998 has_pcps = true;
1999 } else {
2000 for_each_populated_zone(z) {
2001 pcp = per_cpu_ptr(z->pageset, cpu);
2002 if (pcp->pcp.count) {
2003 has_pcps = true;
2004 break;
2005 }
2006 }
2007 }
2008
2009 if (has_pcps)
2010 cpumask_set_cpu(cpu, &cpus_with_pcps);
2011 else
2012 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2013 }
2014 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2015 zone, 1);
2016 }
2017
2018 #ifdef CONFIG_HIBERNATION
2019
2020 void mark_free_pages(struct zone *zone)
2021 {
2022 unsigned long pfn, max_zone_pfn;
2023 unsigned long flags;
2024 unsigned int order, t;
2025 struct list_head *curr;
2026
2027 if (zone_is_empty(zone))
2028 return;
2029
2030 spin_lock_irqsave(&zone->lock, flags);
2031
2032 max_zone_pfn = zone_end_pfn(zone);
2033 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2034 if (pfn_valid(pfn)) {
2035 struct page *page = pfn_to_page(pfn);
2036
2037 if (!swsusp_page_is_forbidden(page))
2038 swsusp_unset_page_free(page);
2039 }
2040
2041 for_each_migratetype_order(order, t) {
2042 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2043 unsigned long i;
2044
2045 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2046 for (i = 0; i < (1UL << order); i++)
2047 swsusp_set_page_free(pfn_to_page(pfn + i));
2048 }
2049 }
2050 spin_unlock_irqrestore(&zone->lock, flags);
2051 }
2052 #endif /* CONFIG_PM */
2053
2054 /*
2055 * Free a 0-order page
2056 * cold == true ? free a cold page : free a hot page
2057 */
2058 void free_hot_cold_page(struct page *page, bool cold)
2059 {
2060 struct zone *zone = page_zone(page);
2061 struct per_cpu_pages *pcp;
2062 unsigned long flags;
2063 unsigned long pfn = page_to_pfn(page);
2064 int migratetype;
2065
2066 if (!free_pages_prepare(page, 0))
2067 return;
2068
2069 migratetype = get_pfnblock_migratetype(page, pfn);
2070 set_pcppage_migratetype(page, migratetype);
2071 local_irq_save(flags);
2072 __count_vm_event(PGFREE);
2073
2074 /*
2075 * We only track unmovable, reclaimable and movable on pcp lists.
2076 * Free ISOLATE pages back to the allocator because they are being
2077 * offlined but treat RESERVE as movable pages so we can get those
2078 * areas back if necessary. Otherwise, we may have to free
2079 * excessively into the page allocator
2080 */
2081 if (migratetype >= MIGRATE_PCPTYPES) {
2082 if (unlikely(is_migrate_isolate(migratetype))) {
2083 free_one_page(zone, page, pfn, 0, migratetype);
2084 goto out;
2085 }
2086 migratetype = MIGRATE_MOVABLE;
2087 }
2088
2089 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2090 if (!cold)
2091 list_add(&page->lru, &pcp->lists[migratetype]);
2092 else
2093 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2094 pcp->count++;
2095 if (pcp->count >= pcp->high) {
2096 unsigned long batch = READ_ONCE(pcp->batch);
2097 free_pcppages_bulk(zone, batch, pcp);
2098 pcp->count -= batch;
2099 }
2100
2101 out:
2102 local_irq_restore(flags);
2103 }
2104
2105 /*
2106 * Free a list of 0-order pages
2107 */
2108 void free_hot_cold_page_list(struct list_head *list, bool cold)
2109 {
2110 struct page *page, *next;
2111
2112 list_for_each_entry_safe(page, next, list, lru) {
2113 trace_mm_page_free_batched(page, cold);
2114 free_hot_cold_page(page, cold);
2115 }
2116 }
2117
2118 /*
2119 * split_page takes a non-compound higher-order page, and splits it into
2120 * n (1<<order) sub-pages: page[0..n]
2121 * Each sub-page must be freed individually.
2122 *
2123 * Note: this is probably too low level an operation for use in drivers.
2124 * Please consult with lkml before using this in your driver.
2125 */
2126 void split_page(struct page *page, unsigned int order)
2127 {
2128 int i;
2129 gfp_t gfp_mask;
2130
2131 VM_BUG_ON_PAGE(PageCompound(page), page);
2132 VM_BUG_ON_PAGE(!page_count(page), page);
2133
2134 #ifdef CONFIG_KMEMCHECK
2135 /*
2136 * Split shadow pages too, because free(page[0]) would
2137 * otherwise free the whole shadow.
2138 */
2139 if (kmemcheck_page_is_tracked(page))
2140 split_page(virt_to_page(page[0].shadow), order);
2141 #endif
2142
2143 gfp_mask = get_page_owner_gfp(page);
2144 set_page_owner(page, 0, gfp_mask);
2145 for (i = 1; i < (1 << order); i++) {
2146 set_page_refcounted(page + i);
2147 set_page_owner(page + i, 0, gfp_mask);
2148 }
2149 }
2150 EXPORT_SYMBOL_GPL(split_page);
2151
2152 int __isolate_free_page(struct page *page, unsigned int order)
2153 {
2154 unsigned long watermark;
2155 struct zone *zone;
2156 int mt;
2157
2158 BUG_ON(!PageBuddy(page));
2159
2160 zone = page_zone(page);
2161 mt = get_pageblock_migratetype(page);
2162
2163 if (!is_migrate_isolate(mt)) {
2164 /* Obey watermarks as if the page was being allocated */
2165 watermark = low_wmark_pages(zone) + (1 << order);
2166 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2167 return 0;
2168
2169 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2170 }
2171
2172 /* Remove page from free list */
2173 list_del(&page->lru);
2174 zone->free_area[order].nr_free--;
2175 rmv_page_order(page);
2176
2177 set_page_owner(page, order, __GFP_MOVABLE);
2178
2179 /* Set the pageblock if the isolated page is at least a pageblock */
2180 if (order >= pageblock_order - 1) {
2181 struct page *endpage = page + (1 << order) - 1;
2182 for (; page < endpage; page += pageblock_nr_pages) {
2183 int mt = get_pageblock_migratetype(page);
2184 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2185 set_pageblock_migratetype(page,
2186 MIGRATE_MOVABLE);
2187 }
2188 }
2189
2190
2191 return 1UL << order;
2192 }
2193
2194 /*
2195 * Similar to split_page except the page is already free. As this is only
2196 * being used for migration, the migratetype of the block also changes.
2197 * As this is called with interrupts disabled, the caller is responsible
2198 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2199 * are enabled.
2200 *
2201 * Note: this is probably too low level an operation for use in drivers.
2202 * Please consult with lkml before using this in your driver.
2203 */
2204 int split_free_page(struct page *page)
2205 {
2206 unsigned int order;
2207 int nr_pages;
2208
2209 order = page_order(page);
2210
2211 nr_pages = __isolate_free_page(page, order);
2212 if (!nr_pages)
2213 return 0;
2214
2215 /* Split into individual pages */
2216 set_page_refcounted(page);
2217 split_page(page, order);
2218 return nr_pages;
2219 }
2220
2221 /*
2222 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2223 */
2224 static inline
2225 struct page *buffered_rmqueue(struct zone *preferred_zone,
2226 struct zone *zone, unsigned int order,
2227 gfp_t gfp_flags, int alloc_flags, int migratetype)
2228 {
2229 unsigned long flags;
2230 struct page *page;
2231 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2232
2233 if (likely(order == 0)) {
2234 struct per_cpu_pages *pcp;
2235 struct list_head *list;
2236
2237 local_irq_save(flags);
2238 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2239 list = &pcp->lists[migratetype];
2240 if (list_empty(list)) {
2241 pcp->count += rmqueue_bulk(zone, 0,
2242 pcp->batch, list,
2243 migratetype, cold);
2244 if (unlikely(list_empty(list)))
2245 goto failed;
2246 }
2247
2248 if (cold)
2249 page = list_entry(list->prev, struct page, lru);
2250 else
2251 page = list_entry(list->next, struct page, lru);
2252
2253 list_del(&page->lru);
2254 pcp->count--;
2255 } else {
2256 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2257 /*
2258 * __GFP_NOFAIL is not to be used in new code.
2259 *
2260 * All __GFP_NOFAIL callers should be fixed so that they
2261 * properly detect and handle allocation failures.
2262 *
2263 * We most definitely don't want callers attempting to
2264 * allocate greater than order-1 page units with
2265 * __GFP_NOFAIL.
2266 */
2267 WARN_ON_ONCE(order > 1);
2268 }
2269 spin_lock_irqsave(&zone->lock, flags);
2270
2271 page = NULL;
2272 if (alloc_flags & ALLOC_HARDER) {
2273 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2274 if (page)
2275 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2276 }
2277 if (!page)
2278 page = __rmqueue(zone, order, migratetype, gfp_flags);
2279 spin_unlock(&zone->lock);
2280 if (!page)
2281 goto failed;
2282 __mod_zone_freepage_state(zone, -(1 << order),
2283 get_pcppage_migratetype(page));
2284 }
2285
2286 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2287 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2288 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2289 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2290
2291 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2292 zone_statistics(preferred_zone, zone, gfp_flags);
2293 local_irq_restore(flags);
2294
2295 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2296 return page;
2297
2298 failed:
2299 local_irq_restore(flags);
2300 return NULL;
2301 }
2302
2303 #ifdef CONFIG_FAIL_PAGE_ALLOC
2304
2305 static struct {
2306 struct fault_attr attr;
2307
2308 bool ignore_gfp_highmem;
2309 bool ignore_gfp_reclaim;
2310 u32 min_order;
2311 } fail_page_alloc = {
2312 .attr = FAULT_ATTR_INITIALIZER,
2313 .ignore_gfp_reclaim = true,
2314 .ignore_gfp_highmem = true,
2315 .min_order = 1,
2316 };
2317
2318 static int __init setup_fail_page_alloc(char *str)
2319 {
2320 return setup_fault_attr(&fail_page_alloc.attr, str);
2321 }
2322 __setup("fail_page_alloc=", setup_fail_page_alloc);
2323
2324 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2325 {
2326 if (order < fail_page_alloc.min_order)
2327 return false;
2328 if (gfp_mask & __GFP_NOFAIL)
2329 return false;
2330 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2331 return false;
2332 if (fail_page_alloc.ignore_gfp_reclaim &&
2333 (gfp_mask & __GFP_DIRECT_RECLAIM))
2334 return false;
2335
2336 return should_fail(&fail_page_alloc.attr, 1 << order);
2337 }
2338
2339 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2340
2341 static int __init fail_page_alloc_debugfs(void)
2342 {
2343 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2344 struct dentry *dir;
2345
2346 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2347 &fail_page_alloc.attr);
2348 if (IS_ERR(dir))
2349 return PTR_ERR(dir);
2350
2351 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2352 &fail_page_alloc.ignore_gfp_reclaim))
2353 goto fail;
2354 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2355 &fail_page_alloc.ignore_gfp_highmem))
2356 goto fail;
2357 if (!debugfs_create_u32("min-order", mode, dir,
2358 &fail_page_alloc.min_order))
2359 goto fail;
2360
2361 return 0;
2362 fail:
2363 debugfs_remove_recursive(dir);
2364
2365 return -ENOMEM;
2366 }
2367
2368 late_initcall(fail_page_alloc_debugfs);
2369
2370 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2371
2372 #else /* CONFIG_FAIL_PAGE_ALLOC */
2373
2374 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2375 {
2376 return false;
2377 }
2378
2379 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2380
2381 /*
2382 * Return true if free base pages are above 'mark'. For high-order checks it
2383 * will return true of the order-0 watermark is reached and there is at least
2384 * one free page of a suitable size. Checking now avoids taking the zone lock
2385 * to check in the allocation paths if no pages are free.
2386 */
2387 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2388 unsigned long mark, int classzone_idx, int alloc_flags,
2389 long free_pages)
2390 {
2391 long min = mark;
2392 int o;
2393 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2394
2395 /* free_pages may go negative - that's OK */
2396 free_pages -= (1 << order) - 1;
2397
2398 if (alloc_flags & ALLOC_HIGH)
2399 min -= min / 2;
2400
2401 /*
2402 * If the caller does not have rights to ALLOC_HARDER then subtract
2403 * the high-atomic reserves. This will over-estimate the size of the
2404 * atomic reserve but it avoids a search.
2405 */
2406 if (likely(!alloc_harder))
2407 free_pages -= z->nr_reserved_highatomic;
2408 else
2409 min -= min / 4;
2410
2411 #ifdef CONFIG_CMA
2412 /* If allocation can't use CMA areas don't use free CMA pages */
2413 if (!(alloc_flags & ALLOC_CMA))
2414 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2415 #endif
2416
2417 /*
2418 * Check watermarks for an order-0 allocation request. If these
2419 * are not met, then a high-order request also cannot go ahead
2420 * even if a suitable page happened to be free.
2421 */
2422 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2423 return false;
2424
2425 /* If this is an order-0 request then the watermark is fine */
2426 if (!order)
2427 return true;
2428
2429 /* For a high-order request, check at least one suitable page is free */
2430 for (o = order; o < MAX_ORDER; o++) {
2431 struct free_area *area = &z->free_area[o];
2432 int mt;
2433
2434 if (!area->nr_free)
2435 continue;
2436
2437 if (alloc_harder)
2438 return true;
2439
2440 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2441 if (!list_empty(&area->free_list[mt]))
2442 return true;
2443 }
2444
2445 #ifdef CONFIG_CMA
2446 if ((alloc_flags & ALLOC_CMA) &&
2447 !list_empty(&area->free_list[MIGRATE_CMA])) {
2448 return true;
2449 }
2450 #endif
2451 }
2452 return false;
2453 }
2454
2455 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2456 int classzone_idx, int alloc_flags)
2457 {
2458 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2459 zone_page_state(z, NR_FREE_PAGES));
2460 }
2461
2462 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2463 unsigned long mark, int classzone_idx)
2464 {
2465 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2466
2467 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2468 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2469
2470 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2471 free_pages);
2472 }
2473
2474 #ifdef CONFIG_NUMA
2475 static bool zone_local(struct zone *local_zone, struct zone *zone)
2476 {
2477 return local_zone->node == zone->node;
2478 }
2479
2480 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2481 {
2482 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2483 RECLAIM_DISTANCE;
2484 }
2485 #else /* CONFIG_NUMA */
2486 static bool zone_local(struct zone *local_zone, struct zone *zone)
2487 {
2488 return true;
2489 }
2490
2491 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2492 {
2493 return true;
2494 }
2495 #endif /* CONFIG_NUMA */
2496
2497 static void reset_alloc_batches(struct zone *preferred_zone)
2498 {
2499 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2500
2501 do {
2502 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2503 high_wmark_pages(zone) - low_wmark_pages(zone) -
2504 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2505 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2506 } while (zone++ != preferred_zone);
2507 }
2508
2509 /*
2510 * get_page_from_freelist goes through the zonelist trying to allocate
2511 * a page.
2512 */
2513 static struct page *
2514 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2515 const struct alloc_context *ac)
2516 {
2517 struct zonelist *zonelist = ac->zonelist;
2518 struct zoneref *z;
2519 struct page *page = NULL;
2520 struct zone *zone;
2521 int nr_fair_skipped = 0;
2522 bool zonelist_rescan;
2523
2524 zonelist_scan:
2525 zonelist_rescan = false;
2526
2527 /*
2528 * Scan zonelist, looking for a zone with enough free.
2529 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2530 */
2531 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2532 ac->nodemask) {
2533 unsigned long mark;
2534
2535 if (cpusets_enabled() &&
2536 (alloc_flags & ALLOC_CPUSET) &&
2537 !cpuset_zone_allowed(zone, gfp_mask))
2538 continue;
2539 /*
2540 * Distribute pages in proportion to the individual
2541 * zone size to ensure fair page aging. The zone a
2542 * page was allocated in should have no effect on the
2543 * time the page has in memory before being reclaimed.
2544 */
2545 if (alloc_flags & ALLOC_FAIR) {
2546 if (!zone_local(ac->preferred_zone, zone))
2547 break;
2548 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2549 nr_fair_skipped++;
2550 continue;
2551 }
2552 }
2553 /*
2554 * When allocating a page cache page for writing, we
2555 * want to get it from a zone that is within its dirty
2556 * limit, such that no single zone holds more than its
2557 * proportional share of globally allowed dirty pages.
2558 * The dirty limits take into account the zone's
2559 * lowmem reserves and high watermark so that kswapd
2560 * should be able to balance it without having to
2561 * write pages from its LRU list.
2562 *
2563 * This may look like it could increase pressure on
2564 * lower zones by failing allocations in higher zones
2565 * before they are full. But the pages that do spill
2566 * over are limited as the lower zones are protected
2567 * by this very same mechanism. It should not become
2568 * a practical burden to them.
2569 *
2570 * XXX: For now, allow allocations to potentially
2571 * exceed the per-zone dirty limit in the slowpath
2572 * (spread_dirty_pages unset) before going into reclaim,
2573 * which is important when on a NUMA setup the allowed
2574 * zones are together not big enough to reach the
2575 * global limit. The proper fix for these situations
2576 * will require awareness of zones in the
2577 * dirty-throttling and the flusher threads.
2578 */
2579 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2580 continue;
2581
2582 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2583 if (!zone_watermark_ok(zone, order, mark,
2584 ac->classzone_idx, alloc_flags)) {
2585 int ret;
2586
2587 /* Checked here to keep the fast path fast */
2588 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2589 if (alloc_flags & ALLOC_NO_WATERMARKS)
2590 goto try_this_zone;
2591
2592 if (zone_reclaim_mode == 0 ||
2593 !zone_allows_reclaim(ac->preferred_zone, zone))
2594 continue;
2595
2596 ret = zone_reclaim(zone, gfp_mask, order);
2597 switch (ret) {
2598 case ZONE_RECLAIM_NOSCAN:
2599 /* did not scan */
2600 continue;
2601 case ZONE_RECLAIM_FULL:
2602 /* scanned but unreclaimable */
2603 continue;
2604 default:
2605 /* did we reclaim enough */
2606 if (zone_watermark_ok(zone, order, mark,
2607 ac->classzone_idx, alloc_flags))
2608 goto try_this_zone;
2609
2610 continue;
2611 }
2612 }
2613
2614 try_this_zone:
2615 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2616 gfp_mask, alloc_flags, ac->migratetype);
2617 if (page) {
2618 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2619 goto try_this_zone;
2620
2621 /*
2622 * If this is a high-order atomic allocation then check
2623 * if the pageblock should be reserved for the future
2624 */
2625 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2626 reserve_highatomic_pageblock(page, zone, order);
2627
2628 return page;
2629 }
2630 }
2631
2632 /*
2633 * The first pass makes sure allocations are spread fairly within the
2634 * local node. However, the local node might have free pages left
2635 * after the fairness batches are exhausted, and remote zones haven't
2636 * even been considered yet. Try once more without fairness, and
2637 * include remote zones now, before entering the slowpath and waking
2638 * kswapd: prefer spilling to a remote zone over swapping locally.
2639 */
2640 if (alloc_flags & ALLOC_FAIR) {
2641 alloc_flags &= ~ALLOC_FAIR;
2642 if (nr_fair_skipped) {
2643 zonelist_rescan = true;
2644 reset_alloc_batches(ac->preferred_zone);
2645 }
2646 if (nr_online_nodes > 1)
2647 zonelist_rescan = true;
2648 }
2649
2650 if (zonelist_rescan)
2651 goto zonelist_scan;
2652
2653 return NULL;
2654 }
2655
2656 /*
2657 * Large machines with many possible nodes should not always dump per-node
2658 * meminfo in irq context.
2659 */
2660 static inline bool should_suppress_show_mem(void)
2661 {
2662 bool ret = false;
2663
2664 #if NODES_SHIFT > 8
2665 ret = in_interrupt();
2666 #endif
2667 return ret;
2668 }
2669
2670 static DEFINE_RATELIMIT_STATE(nopage_rs,
2671 DEFAULT_RATELIMIT_INTERVAL,
2672 DEFAULT_RATELIMIT_BURST);
2673
2674 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2675 {
2676 unsigned int filter = SHOW_MEM_FILTER_NODES;
2677
2678 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2679 debug_guardpage_minorder() > 0)
2680 return;
2681
2682 /*
2683 * This documents exceptions given to allocations in certain
2684 * contexts that are allowed to allocate outside current's set
2685 * of allowed nodes.
2686 */
2687 if (!(gfp_mask & __GFP_NOMEMALLOC))
2688 if (test_thread_flag(TIF_MEMDIE) ||
2689 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2690 filter &= ~SHOW_MEM_FILTER_NODES;
2691 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2692 filter &= ~SHOW_MEM_FILTER_NODES;
2693
2694 if (fmt) {
2695 struct va_format vaf;
2696 va_list args;
2697
2698 va_start(args, fmt);
2699
2700 vaf.fmt = fmt;
2701 vaf.va = &args;
2702
2703 pr_warn("%pV", &vaf);
2704
2705 va_end(args);
2706 }
2707
2708 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2709 current->comm, order, gfp_mask);
2710
2711 dump_stack();
2712 if (!should_suppress_show_mem())
2713 show_mem(filter);
2714 }
2715
2716 static inline struct page *
2717 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2718 const struct alloc_context *ac, unsigned long *did_some_progress)
2719 {
2720 struct oom_control oc = {
2721 .zonelist = ac->zonelist,
2722 .nodemask = ac->nodemask,
2723 .gfp_mask = gfp_mask,
2724 .order = order,
2725 };
2726 struct page *page;
2727
2728 *did_some_progress = 0;
2729
2730 /*
2731 * Acquire the oom lock. If that fails, somebody else is
2732 * making progress for us.
2733 */
2734 if (!mutex_trylock(&oom_lock)) {
2735 *did_some_progress = 1;
2736 schedule_timeout_uninterruptible(1);
2737 return NULL;
2738 }
2739
2740 /*
2741 * Go through the zonelist yet one more time, keep very high watermark
2742 * here, this is only to catch a parallel oom killing, we must fail if
2743 * we're still under heavy pressure.
2744 */
2745 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2746 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2747 if (page)
2748 goto out;
2749
2750 if (!(gfp_mask & __GFP_NOFAIL)) {
2751 /* Coredumps can quickly deplete all memory reserves */
2752 if (current->flags & PF_DUMPCORE)
2753 goto out;
2754 /* The OOM killer will not help higher order allocs */
2755 if (order > PAGE_ALLOC_COSTLY_ORDER)
2756 goto out;
2757 /* The OOM killer does not needlessly kill tasks for lowmem */
2758 if (ac->high_zoneidx < ZONE_NORMAL)
2759 goto out;
2760 /* The OOM killer does not compensate for IO-less reclaim */
2761 if (!(gfp_mask & __GFP_FS)) {
2762 /*
2763 * XXX: Page reclaim didn't yield anything,
2764 * and the OOM killer can't be invoked, but
2765 * keep looping as per tradition.
2766 */
2767 *did_some_progress = 1;
2768 goto out;
2769 }
2770 if (pm_suspended_storage())
2771 goto out;
2772 /* The OOM killer may not free memory on a specific node */
2773 if (gfp_mask & __GFP_THISNODE)
2774 goto out;
2775 }
2776 /* Exhausted what can be done so it's blamo time */
2777 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2778 *did_some_progress = 1;
2779 out:
2780 mutex_unlock(&oom_lock);
2781 return page;
2782 }
2783
2784 #ifdef CONFIG_COMPACTION
2785 /* Try memory compaction for high-order allocations before reclaim */
2786 static struct page *
2787 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2788 int alloc_flags, const struct alloc_context *ac,
2789 enum migrate_mode mode, int *contended_compaction,
2790 bool *deferred_compaction)
2791 {
2792 unsigned long compact_result;
2793 struct page *page;
2794
2795 if (!order)
2796 return NULL;
2797
2798 current->flags |= PF_MEMALLOC;
2799 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2800 mode, contended_compaction);
2801 current->flags &= ~PF_MEMALLOC;
2802
2803 switch (compact_result) {
2804 case COMPACT_DEFERRED:
2805 *deferred_compaction = true;
2806 /* fall-through */
2807 case COMPACT_SKIPPED:
2808 return NULL;
2809 default:
2810 break;
2811 }
2812
2813 /*
2814 * At least in one zone compaction wasn't deferred or skipped, so let's
2815 * count a compaction stall
2816 */
2817 count_vm_event(COMPACTSTALL);
2818
2819 page = get_page_from_freelist(gfp_mask, order,
2820 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2821
2822 if (page) {
2823 struct zone *zone = page_zone(page);
2824
2825 zone->compact_blockskip_flush = false;
2826 compaction_defer_reset(zone, order, true);
2827 count_vm_event(COMPACTSUCCESS);
2828 return page;
2829 }
2830
2831 /*
2832 * It's bad if compaction run occurs and fails. The most likely reason
2833 * is that pages exist, but not enough to satisfy watermarks.
2834 */
2835 count_vm_event(COMPACTFAIL);
2836
2837 cond_resched();
2838
2839 return NULL;
2840 }
2841 #else
2842 static inline struct page *
2843 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2844 int alloc_flags, const struct alloc_context *ac,
2845 enum migrate_mode mode, int *contended_compaction,
2846 bool *deferred_compaction)
2847 {
2848 return NULL;
2849 }
2850 #endif /* CONFIG_COMPACTION */
2851
2852 /* Perform direct synchronous page reclaim */
2853 static int
2854 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2855 const struct alloc_context *ac)
2856 {
2857 struct reclaim_state reclaim_state;
2858 int progress;
2859
2860 cond_resched();
2861
2862 /* We now go into synchronous reclaim */
2863 cpuset_memory_pressure_bump();
2864 current->flags |= PF_MEMALLOC;
2865 lockdep_set_current_reclaim_state(gfp_mask);
2866 reclaim_state.reclaimed_slab = 0;
2867 current->reclaim_state = &reclaim_state;
2868
2869 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2870 ac->nodemask);
2871
2872 current->reclaim_state = NULL;
2873 lockdep_clear_current_reclaim_state();
2874 current->flags &= ~PF_MEMALLOC;
2875
2876 cond_resched();
2877
2878 return progress;
2879 }
2880
2881 /* The really slow allocator path where we enter direct reclaim */
2882 static inline struct page *
2883 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2884 int alloc_flags, const struct alloc_context *ac,
2885 unsigned long *did_some_progress)
2886 {
2887 struct page *page = NULL;
2888 bool drained = false;
2889
2890 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2891 if (unlikely(!(*did_some_progress)))
2892 return NULL;
2893
2894 retry:
2895 page = get_page_from_freelist(gfp_mask, order,
2896 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2897
2898 /*
2899 * If an allocation failed after direct reclaim, it could be because
2900 * pages are pinned on the per-cpu lists or in high alloc reserves.
2901 * Shrink them them and try again
2902 */
2903 if (!page && !drained) {
2904 unreserve_highatomic_pageblock(ac);
2905 drain_all_pages(NULL);
2906 drained = true;
2907 goto retry;
2908 }
2909
2910 return page;
2911 }
2912
2913 /*
2914 * This is called in the allocator slow-path if the allocation request is of
2915 * sufficient urgency to ignore watermarks and take other desperate measures
2916 */
2917 static inline struct page *
2918 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2919 const struct alloc_context *ac)
2920 {
2921 struct page *page;
2922
2923 do {
2924 page = get_page_from_freelist(gfp_mask, order,
2925 ALLOC_NO_WATERMARKS, ac);
2926
2927 if (!page && gfp_mask & __GFP_NOFAIL)
2928 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2929 HZ/50);
2930 } while (!page && (gfp_mask & __GFP_NOFAIL));
2931
2932 return page;
2933 }
2934
2935 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2936 {
2937 struct zoneref *z;
2938 struct zone *zone;
2939
2940 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2941 ac->high_zoneidx, ac->nodemask)
2942 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2943 }
2944
2945 static inline int
2946 gfp_to_alloc_flags(gfp_t gfp_mask)
2947 {
2948 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2949
2950 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2951 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2952
2953 /*
2954 * The caller may dip into page reserves a bit more if the caller
2955 * cannot run direct reclaim, or if the caller has realtime scheduling
2956 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2957 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2958 */
2959 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2960
2961 if (gfp_mask & __GFP_ATOMIC) {
2962 /*
2963 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2964 * if it can't schedule.
2965 */
2966 if (!(gfp_mask & __GFP_NOMEMALLOC))
2967 alloc_flags |= ALLOC_HARDER;
2968 /*
2969 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2970 * comment for __cpuset_node_allowed().
2971 */
2972 alloc_flags &= ~ALLOC_CPUSET;
2973 } else if (unlikely(rt_task(current)) && !in_interrupt())
2974 alloc_flags |= ALLOC_HARDER;
2975
2976 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2977 if (gfp_mask & __GFP_MEMALLOC)
2978 alloc_flags |= ALLOC_NO_WATERMARKS;
2979 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2980 alloc_flags |= ALLOC_NO_WATERMARKS;
2981 else if (!in_interrupt() &&
2982 ((current->flags & PF_MEMALLOC) ||
2983 unlikely(test_thread_flag(TIF_MEMDIE))))
2984 alloc_flags |= ALLOC_NO_WATERMARKS;
2985 }
2986 #ifdef CONFIG_CMA
2987 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2988 alloc_flags |= ALLOC_CMA;
2989 #endif
2990 return alloc_flags;
2991 }
2992
2993 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2994 {
2995 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2996 }
2997
2998 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2999 {
3000 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3001 }
3002
3003 static inline struct page *
3004 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3005 struct alloc_context *ac)
3006 {
3007 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3008 struct page *page = NULL;
3009 int alloc_flags;
3010 unsigned long pages_reclaimed = 0;
3011 unsigned long did_some_progress;
3012 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3013 bool deferred_compaction = false;
3014 int contended_compaction = COMPACT_CONTENDED_NONE;
3015
3016 /*
3017 * In the slowpath, we sanity check order to avoid ever trying to
3018 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3019 * be using allocators in order of preference for an area that is
3020 * too large.
3021 */
3022 if (order >= MAX_ORDER) {
3023 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3024 return NULL;
3025 }
3026
3027 /*
3028 * We also sanity check to catch abuse of atomic reserves being used by
3029 * callers that are not in atomic context.
3030 */
3031 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3032 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3033 gfp_mask &= ~__GFP_ATOMIC;
3034
3035 /*
3036 * If this allocation cannot block and it is for a specific node, then
3037 * fail early. There's no need to wakeup kswapd or retry for a
3038 * speculative node-specific allocation.
3039 */
3040 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3041 goto nopage;
3042
3043 retry:
3044 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3045 wake_all_kswapds(order, ac);
3046
3047 /*
3048 * OK, we're below the kswapd watermark and have kicked background
3049 * reclaim. Now things get more complex, so set up alloc_flags according
3050 * to how we want to proceed.
3051 */
3052 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3053
3054 /*
3055 * Find the true preferred zone if the allocation is unconstrained by
3056 * cpusets.
3057 */
3058 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3059 struct zoneref *preferred_zoneref;
3060 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3061 ac->high_zoneidx, NULL, &ac->preferred_zone);
3062 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3063 }
3064
3065 /* This is the last chance, in general, before the goto nopage. */
3066 page = get_page_from_freelist(gfp_mask, order,
3067 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3068 if (page)
3069 goto got_pg;
3070
3071 /* Allocate without watermarks if the context allows */
3072 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3073 /*
3074 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3075 * the allocation is high priority and these type of
3076 * allocations are system rather than user orientated
3077 */
3078 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3079
3080 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3081
3082 if (page) {
3083 goto got_pg;
3084 }
3085 }
3086
3087 /* Caller is not willing to reclaim, we can't balance anything */
3088 if (!can_direct_reclaim) {
3089 /*
3090 * All existing users of the deprecated __GFP_NOFAIL are
3091 * blockable, so warn of any new users that actually allow this
3092 * type of allocation to fail.
3093 */
3094 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3095 goto nopage;
3096 }
3097
3098 /* Avoid recursion of direct reclaim */
3099 if (current->flags & PF_MEMALLOC)
3100 goto nopage;
3101
3102 /* Avoid allocations with no watermarks from looping endlessly */
3103 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3104 goto nopage;
3105
3106 /*
3107 * Try direct compaction. The first pass is asynchronous. Subsequent
3108 * attempts after direct reclaim are synchronous
3109 */
3110 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3111 migration_mode,
3112 &contended_compaction,
3113 &deferred_compaction);
3114 if (page)
3115 goto got_pg;
3116
3117 /* Checks for THP-specific high-order allocations */
3118 if (is_thp_gfp_mask(gfp_mask)) {
3119 /*
3120 * If compaction is deferred for high-order allocations, it is
3121 * because sync compaction recently failed. If this is the case
3122 * and the caller requested a THP allocation, we do not want
3123 * to heavily disrupt the system, so we fail the allocation
3124 * instead of entering direct reclaim.
3125 */
3126 if (deferred_compaction)
3127 goto nopage;
3128
3129 /*
3130 * In all zones where compaction was attempted (and not
3131 * deferred or skipped), lock contention has been detected.
3132 * For THP allocation we do not want to disrupt the others
3133 * so we fallback to base pages instead.
3134 */
3135 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3136 goto nopage;
3137
3138 /*
3139 * If compaction was aborted due to need_resched(), we do not
3140 * want to further increase allocation latency, unless it is
3141 * khugepaged trying to collapse.
3142 */
3143 if (contended_compaction == COMPACT_CONTENDED_SCHED
3144 && !(current->flags & PF_KTHREAD))
3145 goto nopage;
3146 }
3147
3148 /*
3149 * It can become very expensive to allocate transparent hugepages at
3150 * fault, so use asynchronous memory compaction for THP unless it is
3151 * khugepaged trying to collapse.
3152 */
3153 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3154 migration_mode = MIGRATE_SYNC_LIGHT;
3155
3156 /* Try direct reclaim and then allocating */
3157 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3158 &did_some_progress);
3159 if (page)
3160 goto got_pg;
3161
3162 /* Do not loop if specifically requested */
3163 if (gfp_mask & __GFP_NORETRY)
3164 goto noretry;
3165
3166 /* Keep reclaiming pages as long as there is reasonable progress */
3167 pages_reclaimed += did_some_progress;
3168 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3169 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3170 /* Wait for some write requests to complete then retry */
3171 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3172 goto retry;
3173 }
3174
3175 /* Reclaim has failed us, start killing things */
3176 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3177 if (page)
3178 goto got_pg;
3179
3180 /* Retry as long as the OOM killer is making progress */
3181 if (did_some_progress)
3182 goto retry;
3183
3184 noretry:
3185 /*
3186 * High-order allocations do not necessarily loop after
3187 * direct reclaim and reclaim/compaction depends on compaction
3188 * being called after reclaim so call directly if necessary
3189 */
3190 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3191 ac, migration_mode,
3192 &contended_compaction,
3193 &deferred_compaction);
3194 if (page)
3195 goto got_pg;
3196 nopage:
3197 warn_alloc_failed(gfp_mask, order, NULL);
3198 got_pg:
3199 return page;
3200 }
3201
3202 /*
3203 * This is the 'heart' of the zoned buddy allocator.
3204 */
3205 struct page *
3206 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3207 struct zonelist *zonelist, nodemask_t *nodemask)
3208 {
3209 struct zoneref *preferred_zoneref;
3210 struct page *page = NULL;
3211 unsigned int cpuset_mems_cookie;
3212 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3213 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3214 struct alloc_context ac = {
3215 .high_zoneidx = gfp_zone(gfp_mask),
3216 .nodemask = nodemask,
3217 .migratetype = gfpflags_to_migratetype(gfp_mask),
3218 };
3219
3220 gfp_mask &= gfp_allowed_mask;
3221
3222 lockdep_trace_alloc(gfp_mask);
3223
3224 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3225
3226 if (should_fail_alloc_page(gfp_mask, order))
3227 return NULL;
3228
3229 /*
3230 * Check the zones suitable for the gfp_mask contain at least one
3231 * valid zone. It's possible to have an empty zonelist as a result
3232 * of __GFP_THISNODE and a memoryless node
3233 */
3234 if (unlikely(!zonelist->_zonerefs->zone))
3235 return NULL;
3236
3237 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3238 alloc_flags |= ALLOC_CMA;
3239
3240 retry_cpuset:
3241 cpuset_mems_cookie = read_mems_allowed_begin();
3242
3243 /* We set it here, as __alloc_pages_slowpath might have changed it */
3244 ac.zonelist = zonelist;
3245
3246 /* Dirty zone balancing only done in the fast path */
3247 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3248
3249 /* The preferred zone is used for statistics later */
3250 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3251 ac.nodemask ? : &cpuset_current_mems_allowed,
3252 &ac.preferred_zone);
3253 if (!ac.preferred_zone)
3254 goto out;
3255 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3256
3257 /* First allocation attempt */
3258 alloc_mask = gfp_mask|__GFP_HARDWALL;
3259 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3260 if (unlikely(!page)) {
3261 /*
3262 * Runtime PM, block IO and its error handling path
3263 * can deadlock because I/O on the device might not
3264 * complete.
3265 */
3266 alloc_mask = memalloc_noio_flags(gfp_mask);
3267 ac.spread_dirty_pages = false;
3268
3269 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3270 }
3271
3272 if (kmemcheck_enabled && page)
3273 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3274
3275 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3276
3277 out:
3278 /*
3279 * When updating a task's mems_allowed, it is possible to race with
3280 * parallel threads in such a way that an allocation can fail while
3281 * the mask is being updated. If a page allocation is about to fail,
3282 * check if the cpuset changed during allocation and if so, retry.
3283 */
3284 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3285 goto retry_cpuset;
3286
3287 return page;
3288 }
3289 EXPORT_SYMBOL(__alloc_pages_nodemask);
3290
3291 /*
3292 * Common helper functions.
3293 */
3294 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3295 {
3296 struct page *page;
3297
3298 /*
3299 * __get_free_pages() returns a 32-bit address, which cannot represent
3300 * a highmem page
3301 */
3302 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3303
3304 page = alloc_pages(gfp_mask, order);
3305 if (!page)
3306 return 0;
3307 return (unsigned long) page_address(page);
3308 }
3309 EXPORT_SYMBOL(__get_free_pages);
3310
3311 unsigned long get_zeroed_page(gfp_t gfp_mask)
3312 {
3313 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3314 }
3315 EXPORT_SYMBOL(get_zeroed_page);
3316
3317 void __free_pages(struct page *page, unsigned int order)
3318 {
3319 if (put_page_testzero(page)) {
3320 if (order == 0)
3321 free_hot_cold_page(page, false);
3322 else
3323 __free_pages_ok(page, order);
3324 }
3325 }
3326
3327 EXPORT_SYMBOL(__free_pages);
3328
3329 void free_pages(unsigned long addr, unsigned int order)
3330 {
3331 if (addr != 0) {
3332 VM_BUG_ON(!virt_addr_valid((void *)addr));
3333 __free_pages(virt_to_page((void *)addr), order);
3334 }
3335 }
3336
3337 EXPORT_SYMBOL(free_pages);
3338
3339 /*
3340 * Page Fragment:
3341 * An arbitrary-length arbitrary-offset area of memory which resides
3342 * within a 0 or higher order page. Multiple fragments within that page
3343 * are individually refcounted, in the page's reference counter.
3344 *
3345 * The page_frag functions below provide a simple allocation framework for
3346 * page fragments. This is used by the network stack and network device
3347 * drivers to provide a backing region of memory for use as either an
3348 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3349 */
3350 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3351 gfp_t gfp_mask)
3352 {
3353 struct page *page = NULL;
3354 gfp_t gfp = gfp_mask;
3355
3356 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3357 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3358 __GFP_NOMEMALLOC;
3359 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3360 PAGE_FRAG_CACHE_MAX_ORDER);
3361 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3362 #endif
3363 if (unlikely(!page))
3364 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3365
3366 nc->va = page ? page_address(page) : NULL;
3367
3368 return page;
3369 }
3370
3371 void *__alloc_page_frag(struct page_frag_cache *nc,
3372 unsigned int fragsz, gfp_t gfp_mask)
3373 {
3374 unsigned int size = PAGE_SIZE;
3375 struct page *page;
3376 int offset;
3377
3378 if (unlikely(!nc->va)) {
3379 refill:
3380 page = __page_frag_refill(nc, gfp_mask);
3381 if (!page)
3382 return NULL;
3383
3384 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3385 /* if size can vary use size else just use PAGE_SIZE */
3386 size = nc->size;
3387 #endif
3388 /* Even if we own the page, we do not use atomic_set().
3389 * This would break get_page_unless_zero() users.
3390 */
3391 atomic_add(size - 1, &page->_count);
3392
3393 /* reset page count bias and offset to start of new frag */
3394 nc->pfmemalloc = page_is_pfmemalloc(page);
3395 nc->pagecnt_bias = size;
3396 nc->offset = size;
3397 }
3398
3399 offset = nc->offset - fragsz;
3400 if (unlikely(offset < 0)) {
3401 page = virt_to_page(nc->va);
3402
3403 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3404 goto refill;
3405
3406 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3407 /* if size can vary use size else just use PAGE_SIZE */
3408 size = nc->size;
3409 #endif
3410 /* OK, page count is 0, we can safely set it */
3411 atomic_set(&page->_count, size);
3412
3413 /* reset page count bias and offset to start of new frag */
3414 nc->pagecnt_bias = size;
3415 offset = size - fragsz;
3416 }
3417
3418 nc->pagecnt_bias--;
3419 nc->offset = offset;
3420
3421 return nc->va + offset;
3422 }
3423 EXPORT_SYMBOL(__alloc_page_frag);
3424
3425 /*
3426 * Frees a page fragment allocated out of either a compound or order 0 page.
3427 */
3428 void __free_page_frag(void *addr)
3429 {
3430 struct page *page = virt_to_head_page(addr);
3431
3432 if (unlikely(put_page_testzero(page)))
3433 __free_pages_ok(page, compound_order(page));
3434 }
3435 EXPORT_SYMBOL(__free_page_frag);
3436
3437 /*
3438 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3439 * of the current memory cgroup.
3440 *
3441 * It should be used when the caller would like to use kmalloc, but since the
3442 * allocation is large, it has to fall back to the page allocator.
3443 */
3444 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3445 {
3446 struct page *page;
3447
3448 page = alloc_pages(gfp_mask, order);
3449 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3450 __free_pages(page, order);
3451 page = NULL;
3452 }
3453 return page;
3454 }
3455
3456 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3457 {
3458 struct page *page;
3459
3460 page = alloc_pages_node(nid, gfp_mask, order);
3461 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3462 __free_pages(page, order);
3463 page = NULL;
3464 }
3465 return page;
3466 }
3467
3468 /*
3469 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3470 * alloc_kmem_pages.
3471 */
3472 void __free_kmem_pages(struct page *page, unsigned int order)
3473 {
3474 memcg_kmem_uncharge(page, order);
3475 __free_pages(page, order);
3476 }
3477
3478 void free_kmem_pages(unsigned long addr, unsigned int order)
3479 {
3480 if (addr != 0) {
3481 VM_BUG_ON(!virt_addr_valid((void *)addr));
3482 __free_kmem_pages(virt_to_page((void *)addr), order);
3483 }
3484 }
3485
3486 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3487 size_t size)
3488 {
3489 if (addr) {
3490 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3491 unsigned long used = addr + PAGE_ALIGN(size);
3492
3493 split_page(virt_to_page((void *)addr), order);
3494 while (used < alloc_end) {
3495 free_page(used);
3496 used += PAGE_SIZE;
3497 }
3498 }
3499 return (void *)addr;
3500 }
3501
3502 /**
3503 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3504 * @size: the number of bytes to allocate
3505 * @gfp_mask: GFP flags for the allocation
3506 *
3507 * This function is similar to alloc_pages(), except that it allocates the
3508 * minimum number of pages to satisfy the request. alloc_pages() can only
3509 * allocate memory in power-of-two pages.
3510 *
3511 * This function is also limited by MAX_ORDER.
3512 *
3513 * Memory allocated by this function must be released by free_pages_exact().
3514 */
3515 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3516 {
3517 unsigned int order = get_order(size);
3518 unsigned long addr;
3519
3520 addr = __get_free_pages(gfp_mask, order);
3521 return make_alloc_exact(addr, order, size);
3522 }
3523 EXPORT_SYMBOL(alloc_pages_exact);
3524
3525 /**
3526 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3527 * pages on a node.
3528 * @nid: the preferred node ID where memory should be allocated
3529 * @size: the number of bytes to allocate
3530 * @gfp_mask: GFP flags for the allocation
3531 *
3532 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3533 * back.
3534 */
3535 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3536 {
3537 unsigned int order = get_order(size);
3538 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3539 if (!p)
3540 return NULL;
3541 return make_alloc_exact((unsigned long)page_address(p), order, size);
3542 }
3543
3544 /**
3545 * free_pages_exact - release memory allocated via alloc_pages_exact()
3546 * @virt: the value returned by alloc_pages_exact.
3547 * @size: size of allocation, same value as passed to alloc_pages_exact().
3548 *
3549 * Release the memory allocated by a previous call to alloc_pages_exact.
3550 */
3551 void free_pages_exact(void *virt, size_t size)
3552 {
3553 unsigned long addr = (unsigned long)virt;
3554 unsigned long end = addr + PAGE_ALIGN(size);
3555
3556 while (addr < end) {
3557 free_page(addr);
3558 addr += PAGE_SIZE;
3559 }
3560 }
3561 EXPORT_SYMBOL(free_pages_exact);
3562
3563 /**
3564 * nr_free_zone_pages - count number of pages beyond high watermark
3565 * @offset: The zone index of the highest zone
3566 *
3567 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3568 * high watermark within all zones at or below a given zone index. For each
3569 * zone, the number of pages is calculated as:
3570 * managed_pages - high_pages
3571 */
3572 static unsigned long nr_free_zone_pages(int offset)
3573 {
3574 struct zoneref *z;
3575 struct zone *zone;
3576
3577 /* Just pick one node, since fallback list is circular */
3578 unsigned long sum = 0;
3579
3580 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3581
3582 for_each_zone_zonelist(zone, z, zonelist, offset) {
3583 unsigned long size = zone->managed_pages;
3584 unsigned long high = high_wmark_pages(zone);
3585 if (size > high)
3586 sum += size - high;
3587 }
3588
3589 return sum;
3590 }
3591
3592 /**
3593 * nr_free_buffer_pages - count number of pages beyond high watermark
3594 *
3595 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3596 * watermark within ZONE_DMA and ZONE_NORMAL.
3597 */
3598 unsigned long nr_free_buffer_pages(void)
3599 {
3600 return nr_free_zone_pages(gfp_zone(GFP_USER));
3601 }
3602 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3603
3604 /**
3605 * nr_free_pagecache_pages - count number of pages beyond high watermark
3606 *
3607 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3608 * high watermark within all zones.
3609 */
3610 unsigned long nr_free_pagecache_pages(void)
3611 {
3612 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3613 }
3614
3615 static inline void show_node(struct zone *zone)
3616 {
3617 if (IS_ENABLED(CONFIG_NUMA))
3618 printk("Node %d ", zone_to_nid(zone));
3619 }
3620
3621 void si_meminfo(struct sysinfo *val)
3622 {
3623 val->totalram = totalram_pages;
3624 val->sharedram = global_page_state(NR_SHMEM);
3625 val->freeram = global_page_state(NR_FREE_PAGES);
3626 val->bufferram = nr_blockdev_pages();
3627 val->totalhigh = totalhigh_pages;
3628 val->freehigh = nr_free_highpages();
3629 val->mem_unit = PAGE_SIZE;
3630 }
3631
3632 EXPORT_SYMBOL(si_meminfo);
3633
3634 #ifdef CONFIG_NUMA
3635 void si_meminfo_node(struct sysinfo *val, int nid)
3636 {
3637 int zone_type; /* needs to be signed */
3638 unsigned long managed_pages = 0;
3639 pg_data_t *pgdat = NODE_DATA(nid);
3640
3641 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3642 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3643 val->totalram = managed_pages;
3644 val->sharedram = node_page_state(nid, NR_SHMEM);
3645 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3646 #ifdef CONFIG_HIGHMEM
3647 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3648 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3649 NR_FREE_PAGES);
3650 #else
3651 val->totalhigh = 0;
3652 val->freehigh = 0;
3653 #endif
3654 val->mem_unit = PAGE_SIZE;
3655 }
3656 #endif
3657
3658 /*
3659 * Determine whether the node should be displayed or not, depending on whether
3660 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3661 */
3662 bool skip_free_areas_node(unsigned int flags, int nid)
3663 {
3664 bool ret = false;
3665 unsigned int cpuset_mems_cookie;
3666
3667 if (!(flags & SHOW_MEM_FILTER_NODES))
3668 goto out;
3669
3670 do {
3671 cpuset_mems_cookie = read_mems_allowed_begin();
3672 ret = !node_isset(nid, cpuset_current_mems_allowed);
3673 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3674 out:
3675 return ret;
3676 }
3677
3678 #define K(x) ((x) << (PAGE_SHIFT-10))
3679
3680 static void show_migration_types(unsigned char type)
3681 {
3682 static const char types[MIGRATE_TYPES] = {
3683 [MIGRATE_UNMOVABLE] = 'U',
3684 [MIGRATE_MOVABLE] = 'M',
3685 [MIGRATE_RECLAIMABLE] = 'E',
3686 [MIGRATE_HIGHATOMIC] = 'H',
3687 #ifdef CONFIG_CMA
3688 [MIGRATE_CMA] = 'C',
3689 #endif
3690 #ifdef CONFIG_MEMORY_ISOLATION
3691 [MIGRATE_ISOLATE] = 'I',
3692 #endif
3693 };
3694 char tmp[MIGRATE_TYPES + 1];
3695 char *p = tmp;
3696 int i;
3697
3698 for (i = 0; i < MIGRATE_TYPES; i++) {
3699 if (type & (1 << i))
3700 *p++ = types[i];
3701 }
3702
3703 *p = '\0';
3704 printk("(%s) ", tmp);
3705 }
3706
3707 /*
3708 * Show free area list (used inside shift_scroll-lock stuff)
3709 * We also calculate the percentage fragmentation. We do this by counting the
3710 * memory on each free list with the exception of the first item on the list.
3711 *
3712 * Bits in @filter:
3713 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3714 * cpuset.
3715 */
3716 void show_free_areas(unsigned int filter)
3717 {
3718 unsigned long free_pcp = 0;
3719 int cpu;
3720 struct zone *zone;
3721
3722 for_each_populated_zone(zone) {
3723 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3724 continue;
3725
3726 for_each_online_cpu(cpu)
3727 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3728 }
3729
3730 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3731 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3732 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3733 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3734 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3735 " free:%lu free_pcp:%lu free_cma:%lu\n",
3736 global_page_state(NR_ACTIVE_ANON),
3737 global_page_state(NR_INACTIVE_ANON),
3738 global_page_state(NR_ISOLATED_ANON),
3739 global_page_state(NR_ACTIVE_FILE),
3740 global_page_state(NR_INACTIVE_FILE),
3741 global_page_state(NR_ISOLATED_FILE),
3742 global_page_state(NR_UNEVICTABLE),
3743 global_page_state(NR_FILE_DIRTY),
3744 global_page_state(NR_WRITEBACK),
3745 global_page_state(NR_UNSTABLE_NFS),
3746 global_page_state(NR_SLAB_RECLAIMABLE),
3747 global_page_state(NR_SLAB_UNRECLAIMABLE),
3748 global_page_state(NR_FILE_MAPPED),
3749 global_page_state(NR_SHMEM),
3750 global_page_state(NR_PAGETABLE),
3751 global_page_state(NR_BOUNCE),
3752 global_page_state(NR_FREE_PAGES),
3753 free_pcp,
3754 global_page_state(NR_FREE_CMA_PAGES));
3755
3756 for_each_populated_zone(zone) {
3757 int i;
3758
3759 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3760 continue;
3761
3762 free_pcp = 0;
3763 for_each_online_cpu(cpu)
3764 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3765
3766 show_node(zone);
3767 printk("%s"
3768 " free:%lukB"
3769 " min:%lukB"
3770 " low:%lukB"
3771 " high:%lukB"
3772 " active_anon:%lukB"
3773 " inactive_anon:%lukB"
3774 " active_file:%lukB"
3775 " inactive_file:%lukB"
3776 " unevictable:%lukB"
3777 " isolated(anon):%lukB"
3778 " isolated(file):%lukB"
3779 " present:%lukB"
3780 " managed:%lukB"
3781 " mlocked:%lukB"
3782 " dirty:%lukB"
3783 " writeback:%lukB"
3784 " mapped:%lukB"
3785 " shmem:%lukB"
3786 " slab_reclaimable:%lukB"
3787 " slab_unreclaimable:%lukB"
3788 " kernel_stack:%lukB"
3789 " pagetables:%lukB"
3790 " unstable:%lukB"
3791 " bounce:%lukB"
3792 " free_pcp:%lukB"
3793 " local_pcp:%ukB"
3794 " free_cma:%lukB"
3795 " writeback_tmp:%lukB"
3796 " pages_scanned:%lu"
3797 " all_unreclaimable? %s"
3798 "\n",
3799 zone->name,
3800 K(zone_page_state(zone, NR_FREE_PAGES)),
3801 K(min_wmark_pages(zone)),
3802 K(low_wmark_pages(zone)),
3803 K(high_wmark_pages(zone)),
3804 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3805 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3806 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3807 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3808 K(zone_page_state(zone, NR_UNEVICTABLE)),
3809 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3810 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3811 K(zone->present_pages),
3812 K(zone->managed_pages),
3813 K(zone_page_state(zone, NR_MLOCK)),
3814 K(zone_page_state(zone, NR_FILE_DIRTY)),
3815 K(zone_page_state(zone, NR_WRITEBACK)),
3816 K(zone_page_state(zone, NR_FILE_MAPPED)),
3817 K(zone_page_state(zone, NR_SHMEM)),
3818 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3819 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3820 zone_page_state(zone, NR_KERNEL_STACK) *
3821 THREAD_SIZE / 1024,
3822 K(zone_page_state(zone, NR_PAGETABLE)),
3823 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3824 K(zone_page_state(zone, NR_BOUNCE)),
3825 K(free_pcp),
3826 K(this_cpu_read(zone->pageset->pcp.count)),
3827 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3828 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3829 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3830 (!zone_reclaimable(zone) ? "yes" : "no")
3831 );
3832 printk("lowmem_reserve[]:");
3833 for (i = 0; i < MAX_NR_ZONES; i++)
3834 printk(" %ld", zone->lowmem_reserve[i]);
3835 printk("\n");
3836 }
3837
3838 for_each_populated_zone(zone) {
3839 unsigned int order;
3840 unsigned long nr[MAX_ORDER], flags, total = 0;
3841 unsigned char types[MAX_ORDER];
3842
3843 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3844 continue;
3845 show_node(zone);
3846 printk("%s: ", zone->name);
3847
3848 spin_lock_irqsave(&zone->lock, flags);
3849 for (order = 0; order < MAX_ORDER; order++) {
3850 struct free_area *area = &zone->free_area[order];
3851 int type;
3852
3853 nr[order] = area->nr_free;
3854 total += nr[order] << order;
3855
3856 types[order] = 0;
3857 for (type = 0; type < MIGRATE_TYPES; type++) {
3858 if (!list_empty(&area->free_list[type]))
3859 types[order] |= 1 << type;
3860 }
3861 }
3862 spin_unlock_irqrestore(&zone->lock, flags);
3863 for (order = 0; order < MAX_ORDER; order++) {
3864 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3865 if (nr[order])
3866 show_migration_types(types[order]);
3867 }
3868 printk("= %lukB\n", K(total));
3869 }
3870
3871 hugetlb_show_meminfo();
3872
3873 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3874
3875 show_swap_cache_info();
3876 }
3877
3878 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3879 {
3880 zoneref->zone = zone;
3881 zoneref->zone_idx = zone_idx(zone);
3882 }
3883
3884 /*
3885 * Builds allocation fallback zone lists.
3886 *
3887 * Add all populated zones of a node to the zonelist.
3888 */
3889 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3890 int nr_zones)
3891 {
3892 struct zone *zone;
3893 enum zone_type zone_type = MAX_NR_ZONES;
3894
3895 do {
3896 zone_type--;
3897 zone = pgdat->node_zones + zone_type;
3898 if (populated_zone(zone)) {
3899 zoneref_set_zone(zone,
3900 &zonelist->_zonerefs[nr_zones++]);
3901 check_highest_zone(zone_type);
3902 }
3903 } while (zone_type);
3904
3905 return nr_zones;
3906 }
3907
3908
3909 /*
3910 * zonelist_order:
3911 * 0 = automatic detection of better ordering.
3912 * 1 = order by ([node] distance, -zonetype)
3913 * 2 = order by (-zonetype, [node] distance)
3914 *
3915 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3916 * the same zonelist. So only NUMA can configure this param.
3917 */
3918 #define ZONELIST_ORDER_DEFAULT 0
3919 #define ZONELIST_ORDER_NODE 1
3920 #define ZONELIST_ORDER_ZONE 2
3921
3922 /* zonelist order in the kernel.
3923 * set_zonelist_order() will set this to NODE or ZONE.
3924 */
3925 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3926 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3927
3928
3929 #ifdef CONFIG_NUMA
3930 /* The value user specified ....changed by config */
3931 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3932 /* string for sysctl */
3933 #define NUMA_ZONELIST_ORDER_LEN 16
3934 char numa_zonelist_order[16] = "default";
3935
3936 /*
3937 * interface for configure zonelist ordering.
3938 * command line option "numa_zonelist_order"
3939 * = "[dD]efault - default, automatic configuration.
3940 * = "[nN]ode - order by node locality, then by zone within node
3941 * = "[zZ]one - order by zone, then by locality within zone
3942 */
3943
3944 static int __parse_numa_zonelist_order(char *s)
3945 {
3946 if (*s == 'd' || *s == 'D') {
3947 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3948 } else if (*s == 'n' || *s == 'N') {
3949 user_zonelist_order = ZONELIST_ORDER_NODE;
3950 } else if (*s == 'z' || *s == 'Z') {
3951 user_zonelist_order = ZONELIST_ORDER_ZONE;
3952 } else {
3953 printk(KERN_WARNING
3954 "Ignoring invalid numa_zonelist_order value: "
3955 "%s\n", s);
3956 return -EINVAL;
3957 }
3958 return 0;
3959 }
3960
3961 static __init int setup_numa_zonelist_order(char *s)
3962 {
3963 int ret;
3964
3965 if (!s)
3966 return 0;
3967
3968 ret = __parse_numa_zonelist_order(s);
3969 if (ret == 0)
3970 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3971
3972 return ret;
3973 }
3974 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3975
3976 /*
3977 * sysctl handler for numa_zonelist_order
3978 */
3979 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3980 void __user *buffer, size_t *length,
3981 loff_t *ppos)
3982 {
3983 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3984 int ret;
3985 static DEFINE_MUTEX(zl_order_mutex);
3986
3987 mutex_lock(&zl_order_mutex);
3988 if (write) {
3989 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3990 ret = -EINVAL;
3991 goto out;
3992 }
3993 strcpy(saved_string, (char *)table->data);
3994 }
3995 ret = proc_dostring(table, write, buffer, length, ppos);
3996 if (ret)
3997 goto out;
3998 if (write) {
3999 int oldval = user_zonelist_order;
4000
4001 ret = __parse_numa_zonelist_order((char *)table->data);
4002 if (ret) {
4003 /*
4004 * bogus value. restore saved string
4005 */
4006 strncpy((char *)table->data, saved_string,
4007 NUMA_ZONELIST_ORDER_LEN);
4008 user_zonelist_order = oldval;
4009 } else if (oldval != user_zonelist_order) {
4010 mutex_lock(&zonelists_mutex);
4011 build_all_zonelists(NULL, NULL);
4012 mutex_unlock(&zonelists_mutex);
4013 }
4014 }
4015 out:
4016 mutex_unlock(&zl_order_mutex);
4017 return ret;
4018 }
4019
4020
4021 #define MAX_NODE_LOAD (nr_online_nodes)
4022 static int node_load[MAX_NUMNODES];
4023
4024 /**
4025 * find_next_best_node - find the next node that should appear in a given node's fallback list
4026 * @node: node whose fallback list we're appending
4027 * @used_node_mask: nodemask_t of already used nodes
4028 *
4029 * We use a number of factors to determine which is the next node that should
4030 * appear on a given node's fallback list. The node should not have appeared
4031 * already in @node's fallback list, and it should be the next closest node
4032 * according to the distance array (which contains arbitrary distance values
4033 * from each node to each node in the system), and should also prefer nodes
4034 * with no CPUs, since presumably they'll have very little allocation pressure
4035 * on them otherwise.
4036 * It returns -1 if no node is found.
4037 */
4038 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4039 {
4040 int n, val;
4041 int min_val = INT_MAX;
4042 int best_node = NUMA_NO_NODE;
4043 const struct cpumask *tmp = cpumask_of_node(0);
4044
4045 /* Use the local node if we haven't already */
4046 if (!node_isset(node, *used_node_mask)) {
4047 node_set(node, *used_node_mask);
4048 return node;
4049 }
4050
4051 for_each_node_state(n, N_MEMORY) {
4052
4053 /* Don't want a node to appear more than once */
4054 if (node_isset(n, *used_node_mask))
4055 continue;
4056
4057 /* Use the distance array to find the distance */
4058 val = node_distance(node, n);
4059
4060 /* Penalize nodes under us ("prefer the next node") */
4061 val += (n < node);
4062
4063 /* Give preference to headless and unused nodes */
4064 tmp = cpumask_of_node(n);
4065 if (!cpumask_empty(tmp))
4066 val += PENALTY_FOR_NODE_WITH_CPUS;
4067
4068 /* Slight preference for less loaded node */
4069 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4070 val += node_load[n];
4071
4072 if (val < min_val) {
4073 min_val = val;
4074 best_node = n;
4075 }
4076 }
4077
4078 if (best_node >= 0)
4079 node_set(best_node, *used_node_mask);
4080
4081 return best_node;
4082 }
4083
4084
4085 /*
4086 * Build zonelists ordered by node and zones within node.
4087 * This results in maximum locality--normal zone overflows into local
4088 * DMA zone, if any--but risks exhausting DMA zone.
4089 */
4090 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4091 {
4092 int j;
4093 struct zonelist *zonelist;
4094
4095 zonelist = &pgdat->node_zonelists[0];
4096 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4097 ;
4098 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4099 zonelist->_zonerefs[j].zone = NULL;
4100 zonelist->_zonerefs[j].zone_idx = 0;
4101 }
4102
4103 /*
4104 * Build gfp_thisnode zonelists
4105 */
4106 static void build_thisnode_zonelists(pg_data_t *pgdat)
4107 {
4108 int j;
4109 struct zonelist *zonelist;
4110
4111 zonelist = &pgdat->node_zonelists[1];
4112 j = build_zonelists_node(pgdat, zonelist, 0);
4113 zonelist->_zonerefs[j].zone = NULL;
4114 zonelist->_zonerefs[j].zone_idx = 0;
4115 }
4116
4117 /*
4118 * Build zonelists ordered by zone and nodes within zones.
4119 * This results in conserving DMA zone[s] until all Normal memory is
4120 * exhausted, but results in overflowing to remote node while memory
4121 * may still exist in local DMA zone.
4122 */
4123 static int node_order[MAX_NUMNODES];
4124
4125 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4126 {
4127 int pos, j, node;
4128 int zone_type; /* needs to be signed */
4129 struct zone *z;
4130 struct zonelist *zonelist;
4131
4132 zonelist = &pgdat->node_zonelists[0];
4133 pos = 0;
4134 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4135 for (j = 0; j < nr_nodes; j++) {
4136 node = node_order[j];
4137 z = &NODE_DATA(node)->node_zones[zone_type];
4138 if (populated_zone(z)) {
4139 zoneref_set_zone(z,
4140 &zonelist->_zonerefs[pos++]);
4141 check_highest_zone(zone_type);
4142 }
4143 }
4144 }
4145 zonelist->_zonerefs[pos].zone = NULL;
4146 zonelist->_zonerefs[pos].zone_idx = 0;
4147 }
4148
4149 #if defined(CONFIG_64BIT)
4150 /*
4151 * Devices that require DMA32/DMA are relatively rare and do not justify a
4152 * penalty to every machine in case the specialised case applies. Default
4153 * to Node-ordering on 64-bit NUMA machines
4154 */
4155 static int default_zonelist_order(void)
4156 {
4157 return ZONELIST_ORDER_NODE;
4158 }
4159 #else
4160 /*
4161 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4162 * by the kernel. If processes running on node 0 deplete the low memory zone
4163 * then reclaim will occur more frequency increasing stalls and potentially
4164 * be easier to OOM if a large percentage of the zone is under writeback or
4165 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4166 * Hence, default to zone ordering on 32-bit.
4167 */
4168 static int default_zonelist_order(void)
4169 {
4170 return ZONELIST_ORDER_ZONE;
4171 }
4172 #endif /* CONFIG_64BIT */
4173
4174 static void set_zonelist_order(void)
4175 {
4176 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4177 current_zonelist_order = default_zonelist_order();
4178 else
4179 current_zonelist_order = user_zonelist_order;
4180 }
4181
4182 static void build_zonelists(pg_data_t *pgdat)
4183 {
4184 int j, node, load;
4185 enum zone_type i;
4186 nodemask_t used_mask;
4187 int local_node, prev_node;
4188 struct zonelist *zonelist;
4189 unsigned int order = current_zonelist_order;
4190
4191 /* initialize zonelists */
4192 for (i = 0; i < MAX_ZONELISTS; i++) {
4193 zonelist = pgdat->node_zonelists + i;
4194 zonelist->_zonerefs[0].zone = NULL;
4195 zonelist->_zonerefs[0].zone_idx = 0;
4196 }
4197
4198 /* NUMA-aware ordering of nodes */
4199 local_node = pgdat->node_id;
4200 load = nr_online_nodes;
4201 prev_node = local_node;
4202 nodes_clear(used_mask);
4203
4204 memset(node_order, 0, sizeof(node_order));
4205 j = 0;
4206
4207 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4208 /*
4209 * We don't want to pressure a particular node.
4210 * So adding penalty to the first node in same
4211 * distance group to make it round-robin.
4212 */
4213 if (node_distance(local_node, node) !=
4214 node_distance(local_node, prev_node))
4215 node_load[node] = load;
4216
4217 prev_node = node;
4218 load--;
4219 if (order == ZONELIST_ORDER_NODE)
4220 build_zonelists_in_node_order(pgdat, node);
4221 else
4222 node_order[j++] = node; /* remember order */
4223 }
4224
4225 if (order == ZONELIST_ORDER_ZONE) {
4226 /* calculate node order -- i.e., DMA last! */
4227 build_zonelists_in_zone_order(pgdat, j);
4228 }
4229
4230 build_thisnode_zonelists(pgdat);
4231 }
4232
4233 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4234 /*
4235 * Return node id of node used for "local" allocations.
4236 * I.e., first node id of first zone in arg node's generic zonelist.
4237 * Used for initializing percpu 'numa_mem', which is used primarily
4238 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4239 */
4240 int local_memory_node(int node)
4241 {
4242 struct zone *zone;
4243
4244 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4245 gfp_zone(GFP_KERNEL),
4246 NULL,
4247 &zone);
4248 return zone->node;
4249 }
4250 #endif
4251
4252 #else /* CONFIG_NUMA */
4253
4254 static void set_zonelist_order(void)
4255 {
4256 current_zonelist_order = ZONELIST_ORDER_ZONE;
4257 }
4258
4259 static void build_zonelists(pg_data_t *pgdat)
4260 {
4261 int node, local_node;
4262 enum zone_type j;
4263 struct zonelist *zonelist;
4264
4265 local_node = pgdat->node_id;
4266
4267 zonelist = &pgdat->node_zonelists[0];
4268 j = build_zonelists_node(pgdat, zonelist, 0);
4269
4270 /*
4271 * Now we build the zonelist so that it contains the zones
4272 * of all the other nodes.
4273 * We don't want to pressure a particular node, so when
4274 * building the zones for node N, we make sure that the
4275 * zones coming right after the local ones are those from
4276 * node N+1 (modulo N)
4277 */
4278 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4279 if (!node_online(node))
4280 continue;
4281 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4282 }
4283 for (node = 0; node < local_node; node++) {
4284 if (!node_online(node))
4285 continue;
4286 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4287 }
4288
4289 zonelist->_zonerefs[j].zone = NULL;
4290 zonelist->_zonerefs[j].zone_idx = 0;
4291 }
4292
4293 #endif /* CONFIG_NUMA */
4294
4295 /*
4296 * Boot pageset table. One per cpu which is going to be used for all
4297 * zones and all nodes. The parameters will be set in such a way
4298 * that an item put on a list will immediately be handed over to
4299 * the buddy list. This is safe since pageset manipulation is done
4300 * with interrupts disabled.
4301 *
4302 * The boot_pagesets must be kept even after bootup is complete for
4303 * unused processors and/or zones. They do play a role for bootstrapping
4304 * hotplugged processors.
4305 *
4306 * zoneinfo_show() and maybe other functions do
4307 * not check if the processor is online before following the pageset pointer.
4308 * Other parts of the kernel may not check if the zone is available.
4309 */
4310 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4311 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4312 static void setup_zone_pageset(struct zone *zone);
4313
4314 /*
4315 * Global mutex to protect against size modification of zonelists
4316 * as well as to serialize pageset setup for the new populated zone.
4317 */
4318 DEFINE_MUTEX(zonelists_mutex);
4319
4320 /* return values int ....just for stop_machine() */
4321 static int __build_all_zonelists(void *data)
4322 {
4323 int nid;
4324 int cpu;
4325 pg_data_t *self = data;
4326
4327 #ifdef CONFIG_NUMA
4328 memset(node_load, 0, sizeof(node_load));
4329 #endif
4330
4331 if (self && !node_online(self->node_id)) {
4332 build_zonelists(self);
4333 }
4334
4335 for_each_online_node(nid) {
4336 pg_data_t *pgdat = NODE_DATA(nid);
4337
4338 build_zonelists(pgdat);
4339 }
4340
4341 /*
4342 * Initialize the boot_pagesets that are going to be used
4343 * for bootstrapping processors. The real pagesets for
4344 * each zone will be allocated later when the per cpu
4345 * allocator is available.
4346 *
4347 * boot_pagesets are used also for bootstrapping offline
4348 * cpus if the system is already booted because the pagesets
4349 * are needed to initialize allocators on a specific cpu too.
4350 * F.e. the percpu allocator needs the page allocator which
4351 * needs the percpu allocator in order to allocate its pagesets
4352 * (a chicken-egg dilemma).
4353 */
4354 for_each_possible_cpu(cpu) {
4355 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4356
4357 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4358 /*
4359 * We now know the "local memory node" for each node--
4360 * i.e., the node of the first zone in the generic zonelist.
4361 * Set up numa_mem percpu variable for on-line cpus. During
4362 * boot, only the boot cpu should be on-line; we'll init the
4363 * secondary cpus' numa_mem as they come on-line. During
4364 * node/memory hotplug, we'll fixup all on-line cpus.
4365 */
4366 if (cpu_online(cpu))
4367 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4368 #endif
4369 }
4370
4371 return 0;
4372 }
4373
4374 static noinline void __init
4375 build_all_zonelists_init(void)
4376 {
4377 __build_all_zonelists(NULL);
4378 mminit_verify_zonelist();
4379 cpuset_init_current_mems_allowed();
4380 }
4381
4382 /*
4383 * Called with zonelists_mutex held always
4384 * unless system_state == SYSTEM_BOOTING.
4385 *
4386 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4387 * [we're only called with non-NULL zone through __meminit paths] and
4388 * (2) call of __init annotated helper build_all_zonelists_init
4389 * [protected by SYSTEM_BOOTING].
4390 */
4391 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4392 {
4393 set_zonelist_order();
4394
4395 if (system_state == SYSTEM_BOOTING) {
4396 build_all_zonelists_init();
4397 } else {
4398 #ifdef CONFIG_MEMORY_HOTPLUG
4399 if (zone)
4400 setup_zone_pageset(zone);
4401 #endif
4402 /* we have to stop all cpus to guarantee there is no user
4403 of zonelist */
4404 stop_machine(__build_all_zonelists, pgdat, NULL);
4405 /* cpuset refresh routine should be here */
4406 }
4407 vm_total_pages = nr_free_pagecache_pages();
4408 /*
4409 * Disable grouping by mobility if the number of pages in the
4410 * system is too low to allow the mechanism to work. It would be
4411 * more accurate, but expensive to check per-zone. This check is
4412 * made on memory-hotadd so a system can start with mobility
4413 * disabled and enable it later
4414 */
4415 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4416 page_group_by_mobility_disabled = 1;
4417 else
4418 page_group_by_mobility_disabled = 0;
4419
4420 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4421 "Total pages: %ld\n",
4422 nr_online_nodes,
4423 zonelist_order_name[current_zonelist_order],
4424 page_group_by_mobility_disabled ? "off" : "on",
4425 vm_total_pages);
4426 #ifdef CONFIG_NUMA
4427 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4428 #endif
4429 }
4430
4431 /*
4432 * Helper functions to size the waitqueue hash table.
4433 * Essentially these want to choose hash table sizes sufficiently
4434 * large so that collisions trying to wait on pages are rare.
4435 * But in fact, the number of active page waitqueues on typical
4436 * systems is ridiculously low, less than 200. So this is even
4437 * conservative, even though it seems large.
4438 *
4439 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4440 * waitqueues, i.e. the size of the waitq table given the number of pages.
4441 */
4442 #define PAGES_PER_WAITQUEUE 256
4443
4444 #ifndef CONFIG_MEMORY_HOTPLUG
4445 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4446 {
4447 unsigned long size = 1;
4448
4449 pages /= PAGES_PER_WAITQUEUE;
4450
4451 while (size < pages)
4452 size <<= 1;
4453
4454 /*
4455 * Once we have dozens or even hundreds of threads sleeping
4456 * on IO we've got bigger problems than wait queue collision.
4457 * Limit the size of the wait table to a reasonable size.
4458 */
4459 size = min(size, 4096UL);
4460
4461 return max(size, 4UL);
4462 }
4463 #else
4464 /*
4465 * A zone's size might be changed by hot-add, so it is not possible to determine
4466 * a suitable size for its wait_table. So we use the maximum size now.
4467 *
4468 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4469 *
4470 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4471 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4472 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4473 *
4474 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4475 * or more by the traditional way. (See above). It equals:
4476 *
4477 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4478 * ia64(16K page size) : = ( 8G + 4M)byte.
4479 * powerpc (64K page size) : = (32G +16M)byte.
4480 */
4481 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4482 {
4483 return 4096UL;
4484 }
4485 #endif
4486
4487 /*
4488 * This is an integer logarithm so that shifts can be used later
4489 * to extract the more random high bits from the multiplicative
4490 * hash function before the remainder is taken.
4491 */
4492 static inline unsigned long wait_table_bits(unsigned long size)
4493 {
4494 return ffz(~size);
4495 }
4496
4497 /*
4498 * Initially all pages are reserved - free ones are freed
4499 * up by free_all_bootmem() once the early boot process is
4500 * done. Non-atomic initialization, single-pass.
4501 */
4502 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4503 unsigned long start_pfn, enum memmap_context context)
4504 {
4505 pg_data_t *pgdat = NODE_DATA(nid);
4506 unsigned long end_pfn = start_pfn + size;
4507 unsigned long pfn;
4508 struct zone *z;
4509 unsigned long nr_initialised = 0;
4510
4511 if (highest_memmap_pfn < end_pfn - 1)
4512 highest_memmap_pfn = end_pfn - 1;
4513
4514 z = &pgdat->node_zones[zone];
4515 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4516 /*
4517 * There can be holes in boot-time mem_map[]s
4518 * handed to this function. They do not
4519 * exist on hotplugged memory.
4520 */
4521 if (context == MEMMAP_EARLY) {
4522 if (!early_pfn_valid(pfn))
4523 continue;
4524 if (!early_pfn_in_nid(pfn, nid))
4525 continue;
4526 if (!update_defer_init(pgdat, pfn, end_pfn,
4527 &nr_initialised))
4528 break;
4529 }
4530
4531 /*
4532 * Mark the block movable so that blocks are reserved for
4533 * movable at startup. This will force kernel allocations
4534 * to reserve their blocks rather than leaking throughout
4535 * the address space during boot when many long-lived
4536 * kernel allocations are made.
4537 *
4538 * bitmap is created for zone's valid pfn range. but memmap
4539 * can be created for invalid pages (for alignment)
4540 * check here not to call set_pageblock_migratetype() against
4541 * pfn out of zone.
4542 */
4543 if (!(pfn & (pageblock_nr_pages - 1))) {
4544 struct page *page = pfn_to_page(pfn);
4545
4546 __init_single_page(page, pfn, zone, nid);
4547 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4548 } else {
4549 __init_single_pfn(pfn, zone, nid);
4550 }
4551 }
4552 }
4553
4554 static void __meminit zone_init_free_lists(struct zone *zone)
4555 {
4556 unsigned int order, t;
4557 for_each_migratetype_order(order, t) {
4558 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4559 zone->free_area[order].nr_free = 0;
4560 }
4561 }
4562
4563 #ifndef __HAVE_ARCH_MEMMAP_INIT
4564 #define memmap_init(size, nid, zone, start_pfn) \
4565 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4566 #endif
4567
4568 static int zone_batchsize(struct zone *zone)
4569 {
4570 #ifdef CONFIG_MMU
4571 int batch;
4572
4573 /*
4574 * The per-cpu-pages pools are set to around 1000th of the
4575 * size of the zone. But no more than 1/2 of a meg.
4576 *
4577 * OK, so we don't know how big the cache is. So guess.
4578 */
4579 batch = zone->managed_pages / 1024;
4580 if (batch * PAGE_SIZE > 512 * 1024)
4581 batch = (512 * 1024) / PAGE_SIZE;
4582 batch /= 4; /* We effectively *= 4 below */
4583 if (batch < 1)
4584 batch = 1;
4585
4586 /*
4587 * Clamp the batch to a 2^n - 1 value. Having a power
4588 * of 2 value was found to be more likely to have
4589 * suboptimal cache aliasing properties in some cases.
4590 *
4591 * For example if 2 tasks are alternately allocating
4592 * batches of pages, one task can end up with a lot
4593 * of pages of one half of the possible page colors
4594 * and the other with pages of the other colors.
4595 */
4596 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4597
4598 return batch;
4599
4600 #else
4601 /* The deferral and batching of frees should be suppressed under NOMMU
4602 * conditions.
4603 *
4604 * The problem is that NOMMU needs to be able to allocate large chunks
4605 * of contiguous memory as there's no hardware page translation to
4606 * assemble apparent contiguous memory from discontiguous pages.
4607 *
4608 * Queueing large contiguous runs of pages for batching, however,
4609 * causes the pages to actually be freed in smaller chunks. As there
4610 * can be a significant delay between the individual batches being
4611 * recycled, this leads to the once large chunks of space being
4612 * fragmented and becoming unavailable for high-order allocations.
4613 */
4614 return 0;
4615 #endif
4616 }
4617
4618 /*
4619 * pcp->high and pcp->batch values are related and dependent on one another:
4620 * ->batch must never be higher then ->high.
4621 * The following function updates them in a safe manner without read side
4622 * locking.
4623 *
4624 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4625 * those fields changing asynchronously (acording the the above rule).
4626 *
4627 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4628 * outside of boot time (or some other assurance that no concurrent updaters
4629 * exist).
4630 */
4631 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4632 unsigned long batch)
4633 {
4634 /* start with a fail safe value for batch */
4635 pcp->batch = 1;
4636 smp_wmb();
4637
4638 /* Update high, then batch, in order */
4639 pcp->high = high;
4640 smp_wmb();
4641
4642 pcp->batch = batch;
4643 }
4644
4645 /* a companion to pageset_set_high() */
4646 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4647 {
4648 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4649 }
4650
4651 static void pageset_init(struct per_cpu_pageset *p)
4652 {
4653 struct per_cpu_pages *pcp;
4654 int migratetype;
4655
4656 memset(p, 0, sizeof(*p));
4657
4658 pcp = &p->pcp;
4659 pcp->count = 0;
4660 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4661 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4662 }
4663
4664 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4665 {
4666 pageset_init(p);
4667 pageset_set_batch(p, batch);
4668 }
4669
4670 /*
4671 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4672 * to the value high for the pageset p.
4673 */
4674 static void pageset_set_high(struct per_cpu_pageset *p,
4675 unsigned long high)
4676 {
4677 unsigned long batch = max(1UL, high / 4);
4678 if ((high / 4) > (PAGE_SHIFT * 8))
4679 batch = PAGE_SHIFT * 8;
4680
4681 pageset_update(&p->pcp, high, batch);
4682 }
4683
4684 static void pageset_set_high_and_batch(struct zone *zone,
4685 struct per_cpu_pageset *pcp)
4686 {
4687 if (percpu_pagelist_fraction)
4688 pageset_set_high(pcp,
4689 (zone->managed_pages /
4690 percpu_pagelist_fraction));
4691 else
4692 pageset_set_batch(pcp, zone_batchsize(zone));
4693 }
4694
4695 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4696 {
4697 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4698
4699 pageset_init(pcp);
4700 pageset_set_high_and_batch(zone, pcp);
4701 }
4702
4703 static void __meminit setup_zone_pageset(struct zone *zone)
4704 {
4705 int cpu;
4706 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4707 for_each_possible_cpu(cpu)
4708 zone_pageset_init(zone, cpu);
4709 }
4710
4711 /*
4712 * Allocate per cpu pagesets and initialize them.
4713 * Before this call only boot pagesets were available.
4714 */
4715 void __init setup_per_cpu_pageset(void)
4716 {
4717 struct zone *zone;
4718
4719 for_each_populated_zone(zone)
4720 setup_zone_pageset(zone);
4721 }
4722
4723 static noinline __init_refok
4724 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4725 {
4726 int i;
4727 size_t alloc_size;
4728
4729 /*
4730 * The per-page waitqueue mechanism uses hashed waitqueues
4731 * per zone.
4732 */
4733 zone->wait_table_hash_nr_entries =
4734 wait_table_hash_nr_entries(zone_size_pages);
4735 zone->wait_table_bits =
4736 wait_table_bits(zone->wait_table_hash_nr_entries);
4737 alloc_size = zone->wait_table_hash_nr_entries
4738 * sizeof(wait_queue_head_t);
4739
4740 if (!slab_is_available()) {
4741 zone->wait_table = (wait_queue_head_t *)
4742 memblock_virt_alloc_node_nopanic(
4743 alloc_size, zone->zone_pgdat->node_id);
4744 } else {
4745 /*
4746 * This case means that a zone whose size was 0 gets new memory
4747 * via memory hot-add.
4748 * But it may be the case that a new node was hot-added. In
4749 * this case vmalloc() will not be able to use this new node's
4750 * memory - this wait_table must be initialized to use this new
4751 * node itself as well.
4752 * To use this new node's memory, further consideration will be
4753 * necessary.
4754 */
4755 zone->wait_table = vmalloc(alloc_size);
4756 }
4757 if (!zone->wait_table)
4758 return -ENOMEM;
4759
4760 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4761 init_waitqueue_head(zone->wait_table + i);
4762
4763 return 0;
4764 }
4765
4766 static __meminit void zone_pcp_init(struct zone *zone)
4767 {
4768 /*
4769 * per cpu subsystem is not up at this point. The following code
4770 * relies on the ability of the linker to provide the
4771 * offset of a (static) per cpu variable into the per cpu area.
4772 */
4773 zone->pageset = &boot_pageset;
4774
4775 if (populated_zone(zone))
4776 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4777 zone->name, zone->present_pages,
4778 zone_batchsize(zone));
4779 }
4780
4781 int __meminit init_currently_empty_zone(struct zone *zone,
4782 unsigned long zone_start_pfn,
4783 unsigned long size)
4784 {
4785 struct pglist_data *pgdat = zone->zone_pgdat;
4786 int ret;
4787 ret = zone_wait_table_init(zone, size);
4788 if (ret)
4789 return ret;
4790 pgdat->nr_zones = zone_idx(zone) + 1;
4791
4792 zone->zone_start_pfn = zone_start_pfn;
4793
4794 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4795 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4796 pgdat->node_id,
4797 (unsigned long)zone_idx(zone),
4798 zone_start_pfn, (zone_start_pfn + size));
4799
4800 zone_init_free_lists(zone);
4801
4802 return 0;
4803 }
4804
4805 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4806 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4807
4808 /*
4809 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4810 */
4811 int __meminit __early_pfn_to_nid(unsigned long pfn,
4812 struct mminit_pfnnid_cache *state)
4813 {
4814 unsigned long start_pfn, end_pfn;
4815 int nid;
4816
4817 if (state->last_start <= pfn && pfn < state->last_end)
4818 return state->last_nid;
4819
4820 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4821 if (nid != -1) {
4822 state->last_start = start_pfn;
4823 state->last_end = end_pfn;
4824 state->last_nid = nid;
4825 }
4826
4827 return nid;
4828 }
4829 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4830
4831 /**
4832 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4833 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4834 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4835 *
4836 * If an architecture guarantees that all ranges registered contain no holes
4837 * and may be freed, this this function may be used instead of calling
4838 * memblock_free_early_nid() manually.
4839 */
4840 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4841 {
4842 unsigned long start_pfn, end_pfn;
4843 int i, this_nid;
4844
4845 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4846 start_pfn = min(start_pfn, max_low_pfn);
4847 end_pfn = min(end_pfn, max_low_pfn);
4848
4849 if (start_pfn < end_pfn)
4850 memblock_free_early_nid(PFN_PHYS(start_pfn),
4851 (end_pfn - start_pfn) << PAGE_SHIFT,
4852 this_nid);
4853 }
4854 }
4855
4856 /**
4857 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4858 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4859 *
4860 * If an architecture guarantees that all ranges registered contain no holes and may
4861 * be freed, this function may be used instead of calling memory_present() manually.
4862 */
4863 void __init sparse_memory_present_with_active_regions(int nid)
4864 {
4865 unsigned long start_pfn, end_pfn;
4866 int i, this_nid;
4867
4868 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4869 memory_present(this_nid, start_pfn, end_pfn);
4870 }
4871
4872 /**
4873 * get_pfn_range_for_nid - Return the start and end page frames for a node
4874 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4875 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4876 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4877 *
4878 * It returns the start and end page frame of a node based on information
4879 * provided by memblock_set_node(). If called for a node
4880 * with no available memory, a warning is printed and the start and end
4881 * PFNs will be 0.
4882 */
4883 void __meminit get_pfn_range_for_nid(unsigned int nid,
4884 unsigned long *start_pfn, unsigned long *end_pfn)
4885 {
4886 unsigned long this_start_pfn, this_end_pfn;
4887 int i;
4888
4889 *start_pfn = -1UL;
4890 *end_pfn = 0;
4891
4892 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4893 *start_pfn = min(*start_pfn, this_start_pfn);
4894 *end_pfn = max(*end_pfn, this_end_pfn);
4895 }
4896
4897 if (*start_pfn == -1UL)
4898 *start_pfn = 0;
4899 }
4900
4901 /*
4902 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4903 * assumption is made that zones within a node are ordered in monotonic
4904 * increasing memory addresses so that the "highest" populated zone is used
4905 */
4906 static void __init find_usable_zone_for_movable(void)
4907 {
4908 int zone_index;
4909 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4910 if (zone_index == ZONE_MOVABLE)
4911 continue;
4912
4913 if (arch_zone_highest_possible_pfn[zone_index] >
4914 arch_zone_lowest_possible_pfn[zone_index])
4915 break;
4916 }
4917
4918 VM_BUG_ON(zone_index == -1);
4919 movable_zone = zone_index;
4920 }
4921
4922 /*
4923 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4924 * because it is sized independent of architecture. Unlike the other zones,
4925 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4926 * in each node depending on the size of each node and how evenly kernelcore
4927 * is distributed. This helper function adjusts the zone ranges
4928 * provided by the architecture for a given node by using the end of the
4929 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4930 * zones within a node are in order of monotonic increases memory addresses
4931 */
4932 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4933 unsigned long zone_type,
4934 unsigned long node_start_pfn,
4935 unsigned long node_end_pfn,
4936 unsigned long *zone_start_pfn,
4937 unsigned long *zone_end_pfn)
4938 {
4939 /* Only adjust if ZONE_MOVABLE is on this node */
4940 if (zone_movable_pfn[nid]) {
4941 /* Size ZONE_MOVABLE */
4942 if (zone_type == ZONE_MOVABLE) {
4943 *zone_start_pfn = zone_movable_pfn[nid];
4944 *zone_end_pfn = min(node_end_pfn,
4945 arch_zone_highest_possible_pfn[movable_zone]);
4946
4947 /* Adjust for ZONE_MOVABLE starting within this range */
4948 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4949 *zone_end_pfn > zone_movable_pfn[nid]) {
4950 *zone_end_pfn = zone_movable_pfn[nid];
4951
4952 /* Check if this whole range is within ZONE_MOVABLE */
4953 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4954 *zone_start_pfn = *zone_end_pfn;
4955 }
4956 }
4957
4958 /*
4959 * Return the number of pages a zone spans in a node, including holes
4960 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4961 */
4962 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4963 unsigned long zone_type,
4964 unsigned long node_start_pfn,
4965 unsigned long node_end_pfn,
4966 unsigned long *ignored)
4967 {
4968 unsigned long zone_start_pfn, zone_end_pfn;
4969
4970 /* When hotadd a new node from cpu_up(), the node should be empty */
4971 if (!node_start_pfn && !node_end_pfn)
4972 return 0;
4973
4974 /* Get the start and end of the zone */
4975 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4976 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4977 adjust_zone_range_for_zone_movable(nid, zone_type,
4978 node_start_pfn, node_end_pfn,
4979 &zone_start_pfn, &zone_end_pfn);
4980
4981 /* Check that this node has pages within the zone's required range */
4982 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4983 return 0;
4984
4985 /* Move the zone boundaries inside the node if necessary */
4986 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4987 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4988
4989 /* Return the spanned pages */
4990 return zone_end_pfn - zone_start_pfn;
4991 }
4992
4993 /*
4994 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4995 * then all holes in the requested range will be accounted for.
4996 */
4997 unsigned long __meminit __absent_pages_in_range(int nid,
4998 unsigned long range_start_pfn,
4999 unsigned long range_end_pfn)
5000 {
5001 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5002 unsigned long start_pfn, end_pfn;
5003 int i;
5004
5005 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5006 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5007 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5008 nr_absent -= end_pfn - start_pfn;
5009 }
5010 return nr_absent;
5011 }
5012
5013 /**
5014 * absent_pages_in_range - Return number of page frames in holes within a range
5015 * @start_pfn: The start PFN to start searching for holes
5016 * @end_pfn: The end PFN to stop searching for holes
5017 *
5018 * It returns the number of pages frames in memory holes within a range.
5019 */
5020 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5021 unsigned long end_pfn)
5022 {
5023 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5024 }
5025
5026 /* Return the number of page frames in holes in a zone on a node */
5027 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5028 unsigned long zone_type,
5029 unsigned long node_start_pfn,
5030 unsigned long node_end_pfn,
5031 unsigned long *ignored)
5032 {
5033 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5034 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5035 unsigned long zone_start_pfn, zone_end_pfn;
5036
5037 /* When hotadd a new node from cpu_up(), the node should be empty */
5038 if (!node_start_pfn && !node_end_pfn)
5039 return 0;
5040
5041 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5042 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5043
5044 adjust_zone_range_for_zone_movable(nid, zone_type,
5045 node_start_pfn, node_end_pfn,
5046 &zone_start_pfn, &zone_end_pfn);
5047 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5048 }
5049
5050 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5051 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5052 unsigned long zone_type,
5053 unsigned long node_start_pfn,
5054 unsigned long node_end_pfn,
5055 unsigned long *zones_size)
5056 {
5057 return zones_size[zone_type];
5058 }
5059
5060 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5061 unsigned long zone_type,
5062 unsigned long node_start_pfn,
5063 unsigned long node_end_pfn,
5064 unsigned long *zholes_size)
5065 {
5066 if (!zholes_size)
5067 return 0;
5068
5069 return zholes_size[zone_type];
5070 }
5071
5072 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5073
5074 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5075 unsigned long node_start_pfn,
5076 unsigned long node_end_pfn,
5077 unsigned long *zones_size,
5078 unsigned long *zholes_size)
5079 {
5080 unsigned long realtotalpages = 0, totalpages = 0;
5081 enum zone_type i;
5082
5083 for (i = 0; i < MAX_NR_ZONES; i++) {
5084 struct zone *zone = pgdat->node_zones + i;
5085 unsigned long size, real_size;
5086
5087 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5088 node_start_pfn,
5089 node_end_pfn,
5090 zones_size);
5091 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5092 node_start_pfn, node_end_pfn,
5093 zholes_size);
5094 zone->spanned_pages = size;
5095 zone->present_pages = real_size;
5096
5097 totalpages += size;
5098 realtotalpages += real_size;
5099 }
5100
5101 pgdat->node_spanned_pages = totalpages;
5102 pgdat->node_present_pages = realtotalpages;
5103 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5104 realtotalpages);
5105 }
5106
5107 #ifndef CONFIG_SPARSEMEM
5108 /*
5109 * Calculate the size of the zone->blockflags rounded to an unsigned long
5110 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5111 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5112 * round what is now in bits to nearest long in bits, then return it in
5113 * bytes.
5114 */
5115 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5116 {
5117 unsigned long usemapsize;
5118
5119 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5120 usemapsize = roundup(zonesize, pageblock_nr_pages);
5121 usemapsize = usemapsize >> pageblock_order;
5122 usemapsize *= NR_PAGEBLOCK_BITS;
5123 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5124
5125 return usemapsize / 8;
5126 }
5127
5128 static void __init setup_usemap(struct pglist_data *pgdat,
5129 struct zone *zone,
5130 unsigned long zone_start_pfn,
5131 unsigned long zonesize)
5132 {
5133 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5134 zone->pageblock_flags = NULL;
5135 if (usemapsize)
5136 zone->pageblock_flags =
5137 memblock_virt_alloc_node_nopanic(usemapsize,
5138 pgdat->node_id);
5139 }
5140 #else
5141 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5142 unsigned long zone_start_pfn, unsigned long zonesize) {}
5143 #endif /* CONFIG_SPARSEMEM */
5144
5145 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5146
5147 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5148 void __paginginit set_pageblock_order(void)
5149 {
5150 unsigned int order;
5151
5152 /* Check that pageblock_nr_pages has not already been setup */
5153 if (pageblock_order)
5154 return;
5155
5156 if (HPAGE_SHIFT > PAGE_SHIFT)
5157 order = HUGETLB_PAGE_ORDER;
5158 else
5159 order = MAX_ORDER - 1;
5160
5161 /*
5162 * Assume the largest contiguous order of interest is a huge page.
5163 * This value may be variable depending on boot parameters on IA64 and
5164 * powerpc.
5165 */
5166 pageblock_order = order;
5167 }
5168 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5169
5170 /*
5171 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5172 * is unused as pageblock_order is set at compile-time. See
5173 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5174 * the kernel config
5175 */
5176 void __paginginit set_pageblock_order(void)
5177 {
5178 }
5179
5180 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5181
5182 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5183 unsigned long present_pages)
5184 {
5185 unsigned long pages = spanned_pages;
5186
5187 /*
5188 * Provide a more accurate estimation if there are holes within
5189 * the zone and SPARSEMEM is in use. If there are holes within the
5190 * zone, each populated memory region may cost us one or two extra
5191 * memmap pages due to alignment because memmap pages for each
5192 * populated regions may not naturally algined on page boundary.
5193 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5194 */
5195 if (spanned_pages > present_pages + (present_pages >> 4) &&
5196 IS_ENABLED(CONFIG_SPARSEMEM))
5197 pages = present_pages;
5198
5199 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5200 }
5201
5202 /*
5203 * Set up the zone data structures:
5204 * - mark all pages reserved
5205 * - mark all memory queues empty
5206 * - clear the memory bitmaps
5207 *
5208 * NOTE: pgdat should get zeroed by caller.
5209 */
5210 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5211 {
5212 enum zone_type j;
5213 int nid = pgdat->node_id;
5214 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5215 int ret;
5216
5217 pgdat_resize_init(pgdat);
5218 #ifdef CONFIG_NUMA_BALANCING
5219 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5220 pgdat->numabalancing_migrate_nr_pages = 0;
5221 pgdat->numabalancing_migrate_next_window = jiffies;
5222 #endif
5223 init_waitqueue_head(&pgdat->kswapd_wait);
5224 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5225 pgdat_page_ext_init(pgdat);
5226
5227 for (j = 0; j < MAX_NR_ZONES; j++) {
5228 struct zone *zone = pgdat->node_zones + j;
5229 unsigned long size, realsize, freesize, memmap_pages;
5230
5231 size = zone->spanned_pages;
5232 realsize = freesize = zone->present_pages;
5233
5234 /*
5235 * Adjust freesize so that it accounts for how much memory
5236 * is used by this zone for memmap. This affects the watermark
5237 * and per-cpu initialisations
5238 */
5239 memmap_pages = calc_memmap_size(size, realsize);
5240 if (!is_highmem_idx(j)) {
5241 if (freesize >= memmap_pages) {
5242 freesize -= memmap_pages;
5243 if (memmap_pages)
5244 printk(KERN_DEBUG
5245 " %s zone: %lu pages used for memmap\n",
5246 zone_names[j], memmap_pages);
5247 } else
5248 printk(KERN_WARNING
5249 " %s zone: %lu pages exceeds freesize %lu\n",
5250 zone_names[j], memmap_pages, freesize);
5251 }
5252
5253 /* Account for reserved pages */
5254 if (j == 0 && freesize > dma_reserve) {
5255 freesize -= dma_reserve;
5256 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5257 zone_names[0], dma_reserve);
5258 }
5259
5260 if (!is_highmem_idx(j))
5261 nr_kernel_pages += freesize;
5262 /* Charge for highmem memmap if there are enough kernel pages */
5263 else if (nr_kernel_pages > memmap_pages * 2)
5264 nr_kernel_pages -= memmap_pages;
5265 nr_all_pages += freesize;
5266
5267 /*
5268 * Set an approximate value for lowmem here, it will be adjusted
5269 * when the bootmem allocator frees pages into the buddy system.
5270 * And all highmem pages will be managed by the buddy system.
5271 */
5272 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5273 #ifdef CONFIG_NUMA
5274 zone->node = nid;
5275 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5276 / 100;
5277 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5278 #endif
5279 zone->name = zone_names[j];
5280 spin_lock_init(&zone->lock);
5281 spin_lock_init(&zone->lru_lock);
5282 zone_seqlock_init(zone);
5283 zone->zone_pgdat = pgdat;
5284 zone_pcp_init(zone);
5285
5286 /* For bootup, initialized properly in watermark setup */
5287 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5288
5289 lruvec_init(&zone->lruvec);
5290 if (!size)
5291 continue;
5292
5293 set_pageblock_order();
5294 setup_usemap(pgdat, zone, zone_start_pfn, size);
5295 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5296 BUG_ON(ret);
5297 memmap_init(size, nid, j, zone_start_pfn);
5298 zone_start_pfn += size;
5299 }
5300 }
5301
5302 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5303 {
5304 unsigned long __maybe_unused start = 0;
5305 unsigned long __maybe_unused offset = 0;
5306
5307 /* Skip empty nodes */
5308 if (!pgdat->node_spanned_pages)
5309 return;
5310
5311 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5312 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5313 offset = pgdat->node_start_pfn - start;
5314 /* ia64 gets its own node_mem_map, before this, without bootmem */
5315 if (!pgdat->node_mem_map) {
5316 unsigned long size, end;
5317 struct page *map;
5318
5319 /*
5320 * The zone's endpoints aren't required to be MAX_ORDER
5321 * aligned but the node_mem_map endpoints must be in order
5322 * for the buddy allocator to function correctly.
5323 */
5324 end = pgdat_end_pfn(pgdat);
5325 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5326 size = (end - start) * sizeof(struct page);
5327 map = alloc_remap(pgdat->node_id, size);
5328 if (!map)
5329 map = memblock_virt_alloc_node_nopanic(size,
5330 pgdat->node_id);
5331 pgdat->node_mem_map = map + offset;
5332 }
5333 #ifndef CONFIG_NEED_MULTIPLE_NODES
5334 /*
5335 * With no DISCONTIG, the global mem_map is just set as node 0's
5336 */
5337 if (pgdat == NODE_DATA(0)) {
5338 mem_map = NODE_DATA(0)->node_mem_map;
5339 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5340 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5341 mem_map -= offset;
5342 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5343 }
5344 #endif
5345 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5346 }
5347
5348 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5349 unsigned long node_start_pfn, unsigned long *zholes_size)
5350 {
5351 pg_data_t *pgdat = NODE_DATA(nid);
5352 unsigned long start_pfn = 0;
5353 unsigned long end_pfn = 0;
5354
5355 /* pg_data_t should be reset to zero when it's allocated */
5356 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5357
5358 reset_deferred_meminit(pgdat);
5359 pgdat->node_id = nid;
5360 pgdat->node_start_pfn = node_start_pfn;
5361 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5362 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5363 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5364 (u64)start_pfn << PAGE_SHIFT,
5365 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5366 #endif
5367 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5368 zones_size, zholes_size);
5369
5370 alloc_node_mem_map(pgdat);
5371 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5372 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5373 nid, (unsigned long)pgdat,
5374 (unsigned long)pgdat->node_mem_map);
5375 #endif
5376
5377 free_area_init_core(pgdat);
5378 }
5379
5380 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5381
5382 #if MAX_NUMNODES > 1
5383 /*
5384 * Figure out the number of possible node ids.
5385 */
5386 void __init setup_nr_node_ids(void)
5387 {
5388 unsigned int highest;
5389
5390 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5391 nr_node_ids = highest + 1;
5392 }
5393 #endif
5394
5395 /**
5396 * node_map_pfn_alignment - determine the maximum internode alignment
5397 *
5398 * This function should be called after node map is populated and sorted.
5399 * It calculates the maximum power of two alignment which can distinguish
5400 * all the nodes.
5401 *
5402 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5403 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5404 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5405 * shifted, 1GiB is enough and this function will indicate so.
5406 *
5407 * This is used to test whether pfn -> nid mapping of the chosen memory
5408 * model has fine enough granularity to avoid incorrect mapping for the
5409 * populated node map.
5410 *
5411 * Returns the determined alignment in pfn's. 0 if there is no alignment
5412 * requirement (single node).
5413 */
5414 unsigned long __init node_map_pfn_alignment(void)
5415 {
5416 unsigned long accl_mask = 0, last_end = 0;
5417 unsigned long start, end, mask;
5418 int last_nid = -1;
5419 int i, nid;
5420
5421 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5422 if (!start || last_nid < 0 || last_nid == nid) {
5423 last_nid = nid;
5424 last_end = end;
5425 continue;
5426 }
5427
5428 /*
5429 * Start with a mask granular enough to pin-point to the
5430 * start pfn and tick off bits one-by-one until it becomes
5431 * too coarse to separate the current node from the last.
5432 */
5433 mask = ~((1 << __ffs(start)) - 1);
5434 while (mask && last_end <= (start & (mask << 1)))
5435 mask <<= 1;
5436
5437 /* accumulate all internode masks */
5438 accl_mask |= mask;
5439 }
5440
5441 /* convert mask to number of pages */
5442 return ~accl_mask + 1;
5443 }
5444
5445 /* Find the lowest pfn for a node */
5446 static unsigned long __init find_min_pfn_for_node(int nid)
5447 {
5448 unsigned long min_pfn = ULONG_MAX;
5449 unsigned long start_pfn;
5450 int i;
5451
5452 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5453 min_pfn = min(min_pfn, start_pfn);
5454
5455 if (min_pfn == ULONG_MAX) {
5456 printk(KERN_WARNING
5457 "Could not find start_pfn for node %d\n", nid);
5458 return 0;
5459 }
5460
5461 return min_pfn;
5462 }
5463
5464 /**
5465 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5466 *
5467 * It returns the minimum PFN based on information provided via
5468 * memblock_set_node().
5469 */
5470 unsigned long __init find_min_pfn_with_active_regions(void)
5471 {
5472 return find_min_pfn_for_node(MAX_NUMNODES);
5473 }
5474
5475 /*
5476 * early_calculate_totalpages()
5477 * Sum pages in active regions for movable zone.
5478 * Populate N_MEMORY for calculating usable_nodes.
5479 */
5480 static unsigned long __init early_calculate_totalpages(void)
5481 {
5482 unsigned long totalpages = 0;
5483 unsigned long start_pfn, end_pfn;
5484 int i, nid;
5485
5486 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5487 unsigned long pages = end_pfn - start_pfn;
5488
5489 totalpages += pages;
5490 if (pages)
5491 node_set_state(nid, N_MEMORY);
5492 }
5493 return totalpages;
5494 }
5495
5496 /*
5497 * Find the PFN the Movable zone begins in each node. Kernel memory
5498 * is spread evenly between nodes as long as the nodes have enough
5499 * memory. When they don't, some nodes will have more kernelcore than
5500 * others
5501 */
5502 static void __init find_zone_movable_pfns_for_nodes(void)
5503 {
5504 int i, nid;
5505 unsigned long usable_startpfn;
5506 unsigned long kernelcore_node, kernelcore_remaining;
5507 /* save the state before borrow the nodemask */
5508 nodemask_t saved_node_state = node_states[N_MEMORY];
5509 unsigned long totalpages = early_calculate_totalpages();
5510 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5511 struct memblock_region *r;
5512
5513 /* Need to find movable_zone earlier when movable_node is specified. */
5514 find_usable_zone_for_movable();
5515
5516 /*
5517 * If movable_node is specified, ignore kernelcore and movablecore
5518 * options.
5519 */
5520 if (movable_node_is_enabled()) {
5521 for_each_memblock(memory, r) {
5522 if (!memblock_is_hotpluggable(r))
5523 continue;
5524
5525 nid = r->nid;
5526
5527 usable_startpfn = PFN_DOWN(r->base);
5528 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5529 min(usable_startpfn, zone_movable_pfn[nid]) :
5530 usable_startpfn;
5531 }
5532
5533 goto out2;
5534 }
5535
5536 /*
5537 * If movablecore=nn[KMG] was specified, calculate what size of
5538 * kernelcore that corresponds so that memory usable for
5539 * any allocation type is evenly spread. If both kernelcore
5540 * and movablecore are specified, then the value of kernelcore
5541 * will be used for required_kernelcore if it's greater than
5542 * what movablecore would have allowed.
5543 */
5544 if (required_movablecore) {
5545 unsigned long corepages;
5546
5547 /*
5548 * Round-up so that ZONE_MOVABLE is at least as large as what
5549 * was requested by the user
5550 */
5551 required_movablecore =
5552 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5553 required_movablecore = min(totalpages, required_movablecore);
5554 corepages = totalpages - required_movablecore;
5555
5556 required_kernelcore = max(required_kernelcore, corepages);
5557 }
5558
5559 /*
5560 * If kernelcore was not specified or kernelcore size is larger
5561 * than totalpages, there is no ZONE_MOVABLE.
5562 */
5563 if (!required_kernelcore || required_kernelcore >= totalpages)
5564 goto out;
5565
5566 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5567 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5568
5569 restart:
5570 /* Spread kernelcore memory as evenly as possible throughout nodes */
5571 kernelcore_node = required_kernelcore / usable_nodes;
5572 for_each_node_state(nid, N_MEMORY) {
5573 unsigned long start_pfn, end_pfn;
5574
5575 /*
5576 * Recalculate kernelcore_node if the division per node
5577 * now exceeds what is necessary to satisfy the requested
5578 * amount of memory for the kernel
5579 */
5580 if (required_kernelcore < kernelcore_node)
5581 kernelcore_node = required_kernelcore / usable_nodes;
5582
5583 /*
5584 * As the map is walked, we track how much memory is usable
5585 * by the kernel using kernelcore_remaining. When it is
5586 * 0, the rest of the node is usable by ZONE_MOVABLE
5587 */
5588 kernelcore_remaining = kernelcore_node;
5589
5590 /* Go through each range of PFNs within this node */
5591 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5592 unsigned long size_pages;
5593
5594 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5595 if (start_pfn >= end_pfn)
5596 continue;
5597
5598 /* Account for what is only usable for kernelcore */
5599 if (start_pfn < usable_startpfn) {
5600 unsigned long kernel_pages;
5601 kernel_pages = min(end_pfn, usable_startpfn)
5602 - start_pfn;
5603
5604 kernelcore_remaining -= min(kernel_pages,
5605 kernelcore_remaining);
5606 required_kernelcore -= min(kernel_pages,
5607 required_kernelcore);
5608
5609 /* Continue if range is now fully accounted */
5610 if (end_pfn <= usable_startpfn) {
5611
5612 /*
5613 * Push zone_movable_pfn to the end so
5614 * that if we have to rebalance
5615 * kernelcore across nodes, we will
5616 * not double account here
5617 */
5618 zone_movable_pfn[nid] = end_pfn;
5619 continue;
5620 }
5621 start_pfn = usable_startpfn;
5622 }
5623
5624 /*
5625 * The usable PFN range for ZONE_MOVABLE is from
5626 * start_pfn->end_pfn. Calculate size_pages as the
5627 * number of pages used as kernelcore
5628 */
5629 size_pages = end_pfn - start_pfn;
5630 if (size_pages > kernelcore_remaining)
5631 size_pages = kernelcore_remaining;
5632 zone_movable_pfn[nid] = start_pfn + size_pages;
5633
5634 /*
5635 * Some kernelcore has been met, update counts and
5636 * break if the kernelcore for this node has been
5637 * satisfied
5638 */
5639 required_kernelcore -= min(required_kernelcore,
5640 size_pages);
5641 kernelcore_remaining -= size_pages;
5642 if (!kernelcore_remaining)
5643 break;
5644 }
5645 }
5646
5647 /*
5648 * If there is still required_kernelcore, we do another pass with one
5649 * less node in the count. This will push zone_movable_pfn[nid] further
5650 * along on the nodes that still have memory until kernelcore is
5651 * satisfied
5652 */
5653 usable_nodes--;
5654 if (usable_nodes && required_kernelcore > usable_nodes)
5655 goto restart;
5656
5657 out2:
5658 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5659 for (nid = 0; nid < MAX_NUMNODES; nid++)
5660 zone_movable_pfn[nid] =
5661 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5662
5663 out:
5664 /* restore the node_state */
5665 node_states[N_MEMORY] = saved_node_state;
5666 }
5667
5668 /* Any regular or high memory on that node ? */
5669 static void check_for_memory(pg_data_t *pgdat, int nid)
5670 {
5671 enum zone_type zone_type;
5672
5673 if (N_MEMORY == N_NORMAL_MEMORY)
5674 return;
5675
5676 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5677 struct zone *zone = &pgdat->node_zones[zone_type];
5678 if (populated_zone(zone)) {
5679 node_set_state(nid, N_HIGH_MEMORY);
5680 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5681 zone_type <= ZONE_NORMAL)
5682 node_set_state(nid, N_NORMAL_MEMORY);
5683 break;
5684 }
5685 }
5686 }
5687
5688 /**
5689 * free_area_init_nodes - Initialise all pg_data_t and zone data
5690 * @max_zone_pfn: an array of max PFNs for each zone
5691 *
5692 * This will call free_area_init_node() for each active node in the system.
5693 * Using the page ranges provided by memblock_set_node(), the size of each
5694 * zone in each node and their holes is calculated. If the maximum PFN
5695 * between two adjacent zones match, it is assumed that the zone is empty.
5696 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5697 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5698 * starts where the previous one ended. For example, ZONE_DMA32 starts
5699 * at arch_max_dma_pfn.
5700 */
5701 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5702 {
5703 unsigned long start_pfn, end_pfn;
5704 int i, nid;
5705
5706 /* Record where the zone boundaries are */
5707 memset(arch_zone_lowest_possible_pfn, 0,
5708 sizeof(arch_zone_lowest_possible_pfn));
5709 memset(arch_zone_highest_possible_pfn, 0,
5710 sizeof(arch_zone_highest_possible_pfn));
5711
5712 start_pfn = find_min_pfn_with_active_regions();
5713
5714 for (i = 0; i < MAX_NR_ZONES; i++) {
5715 if (i == ZONE_MOVABLE)
5716 continue;
5717
5718 end_pfn = max(max_zone_pfn[i], start_pfn);
5719 arch_zone_lowest_possible_pfn[i] = start_pfn;
5720 arch_zone_highest_possible_pfn[i] = end_pfn;
5721
5722 start_pfn = end_pfn;
5723 }
5724 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5725 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5726
5727 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5728 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5729 find_zone_movable_pfns_for_nodes();
5730
5731 /* Print out the zone ranges */
5732 pr_info("Zone ranges:\n");
5733 for (i = 0; i < MAX_NR_ZONES; i++) {
5734 if (i == ZONE_MOVABLE)
5735 continue;
5736 pr_info(" %-8s ", zone_names[i]);
5737 if (arch_zone_lowest_possible_pfn[i] ==
5738 arch_zone_highest_possible_pfn[i])
5739 pr_cont("empty\n");
5740 else
5741 pr_cont("[mem %#018Lx-%#018Lx]\n",
5742 (u64)arch_zone_lowest_possible_pfn[i]
5743 << PAGE_SHIFT,
5744 ((u64)arch_zone_highest_possible_pfn[i]
5745 << PAGE_SHIFT) - 1);
5746 }
5747
5748 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5749 pr_info("Movable zone start for each node\n");
5750 for (i = 0; i < MAX_NUMNODES; i++) {
5751 if (zone_movable_pfn[i])
5752 pr_info(" Node %d: %#018Lx\n", i,
5753 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5754 }
5755
5756 /* Print out the early node map */
5757 pr_info("Early memory node ranges\n");
5758 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5759 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5760 (u64)start_pfn << PAGE_SHIFT,
5761 ((u64)end_pfn << PAGE_SHIFT) - 1);
5762
5763 /* Initialise every node */
5764 mminit_verify_pageflags_layout();
5765 setup_nr_node_ids();
5766 for_each_online_node(nid) {
5767 pg_data_t *pgdat = NODE_DATA(nid);
5768 free_area_init_node(nid, NULL,
5769 find_min_pfn_for_node(nid), NULL);
5770
5771 /* Any memory on that node */
5772 if (pgdat->node_present_pages)
5773 node_set_state(nid, N_MEMORY);
5774 check_for_memory(pgdat, nid);
5775 }
5776 }
5777
5778 static int __init cmdline_parse_core(char *p, unsigned long *core)
5779 {
5780 unsigned long long coremem;
5781 if (!p)
5782 return -EINVAL;
5783
5784 coremem = memparse(p, &p);
5785 *core = coremem >> PAGE_SHIFT;
5786
5787 /* Paranoid check that UL is enough for the coremem value */
5788 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5789
5790 return 0;
5791 }
5792
5793 /*
5794 * kernelcore=size sets the amount of memory for use for allocations that
5795 * cannot be reclaimed or migrated.
5796 */
5797 static int __init cmdline_parse_kernelcore(char *p)
5798 {
5799 return cmdline_parse_core(p, &required_kernelcore);
5800 }
5801
5802 /*
5803 * movablecore=size sets the amount of memory for use for allocations that
5804 * can be reclaimed or migrated.
5805 */
5806 static int __init cmdline_parse_movablecore(char *p)
5807 {
5808 return cmdline_parse_core(p, &required_movablecore);
5809 }
5810
5811 early_param("kernelcore", cmdline_parse_kernelcore);
5812 early_param("movablecore", cmdline_parse_movablecore);
5813
5814 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5815
5816 void adjust_managed_page_count(struct page *page, long count)
5817 {
5818 spin_lock(&managed_page_count_lock);
5819 page_zone(page)->managed_pages += count;
5820 totalram_pages += count;
5821 #ifdef CONFIG_HIGHMEM
5822 if (PageHighMem(page))
5823 totalhigh_pages += count;
5824 #endif
5825 spin_unlock(&managed_page_count_lock);
5826 }
5827 EXPORT_SYMBOL(adjust_managed_page_count);
5828
5829 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5830 {
5831 void *pos;
5832 unsigned long pages = 0;
5833
5834 start = (void *)PAGE_ALIGN((unsigned long)start);
5835 end = (void *)((unsigned long)end & PAGE_MASK);
5836 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5837 if ((unsigned int)poison <= 0xFF)
5838 memset(pos, poison, PAGE_SIZE);
5839 free_reserved_page(virt_to_page(pos));
5840 }
5841
5842 if (pages && s)
5843 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5844 s, pages << (PAGE_SHIFT - 10), start, end);
5845
5846 return pages;
5847 }
5848 EXPORT_SYMBOL(free_reserved_area);
5849
5850 #ifdef CONFIG_HIGHMEM
5851 void free_highmem_page(struct page *page)
5852 {
5853 __free_reserved_page(page);
5854 totalram_pages++;
5855 page_zone(page)->managed_pages++;
5856 totalhigh_pages++;
5857 }
5858 #endif
5859
5860
5861 void __init mem_init_print_info(const char *str)
5862 {
5863 unsigned long physpages, codesize, datasize, rosize, bss_size;
5864 unsigned long init_code_size, init_data_size;
5865
5866 physpages = get_num_physpages();
5867 codesize = _etext - _stext;
5868 datasize = _edata - _sdata;
5869 rosize = __end_rodata - __start_rodata;
5870 bss_size = __bss_stop - __bss_start;
5871 init_data_size = __init_end - __init_begin;
5872 init_code_size = _einittext - _sinittext;
5873
5874 /*
5875 * Detect special cases and adjust section sizes accordingly:
5876 * 1) .init.* may be embedded into .data sections
5877 * 2) .init.text.* may be out of [__init_begin, __init_end],
5878 * please refer to arch/tile/kernel/vmlinux.lds.S.
5879 * 3) .rodata.* may be embedded into .text or .data sections.
5880 */
5881 #define adj_init_size(start, end, size, pos, adj) \
5882 do { \
5883 if (start <= pos && pos < end && size > adj) \
5884 size -= adj; \
5885 } while (0)
5886
5887 adj_init_size(__init_begin, __init_end, init_data_size,
5888 _sinittext, init_code_size);
5889 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5890 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5891 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5892 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5893
5894 #undef adj_init_size
5895
5896 pr_info("Memory: %luK/%luK available "
5897 "(%luK kernel code, %luK rwdata, %luK rodata, "
5898 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5899 #ifdef CONFIG_HIGHMEM
5900 ", %luK highmem"
5901 #endif
5902 "%s%s)\n",
5903 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5904 codesize >> 10, datasize >> 10, rosize >> 10,
5905 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5906 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5907 totalcma_pages << (PAGE_SHIFT-10),
5908 #ifdef CONFIG_HIGHMEM
5909 totalhigh_pages << (PAGE_SHIFT-10),
5910 #endif
5911 str ? ", " : "", str ? str : "");
5912 }
5913
5914 /**
5915 * set_dma_reserve - set the specified number of pages reserved in the first zone
5916 * @new_dma_reserve: The number of pages to mark reserved
5917 *
5918 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5919 * In the DMA zone, a significant percentage may be consumed by kernel image
5920 * and other unfreeable allocations which can skew the watermarks badly. This
5921 * function may optionally be used to account for unfreeable pages in the
5922 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5923 * smaller per-cpu batchsize.
5924 */
5925 void __init set_dma_reserve(unsigned long new_dma_reserve)
5926 {
5927 dma_reserve = new_dma_reserve;
5928 }
5929
5930 void __init free_area_init(unsigned long *zones_size)
5931 {
5932 free_area_init_node(0, zones_size,
5933 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5934 }
5935
5936 static int page_alloc_cpu_notify(struct notifier_block *self,
5937 unsigned long action, void *hcpu)
5938 {
5939 int cpu = (unsigned long)hcpu;
5940
5941 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5942 lru_add_drain_cpu(cpu);
5943 drain_pages(cpu);
5944
5945 /*
5946 * Spill the event counters of the dead processor
5947 * into the current processors event counters.
5948 * This artificially elevates the count of the current
5949 * processor.
5950 */
5951 vm_events_fold_cpu(cpu);
5952
5953 /*
5954 * Zero the differential counters of the dead processor
5955 * so that the vm statistics are consistent.
5956 *
5957 * This is only okay since the processor is dead and cannot
5958 * race with what we are doing.
5959 */
5960 cpu_vm_stats_fold(cpu);
5961 }
5962 return NOTIFY_OK;
5963 }
5964
5965 void __init page_alloc_init(void)
5966 {
5967 hotcpu_notifier(page_alloc_cpu_notify, 0);
5968 }
5969
5970 /*
5971 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5972 * or min_free_kbytes changes.
5973 */
5974 static void calculate_totalreserve_pages(void)
5975 {
5976 struct pglist_data *pgdat;
5977 unsigned long reserve_pages = 0;
5978 enum zone_type i, j;
5979
5980 for_each_online_pgdat(pgdat) {
5981 for (i = 0; i < MAX_NR_ZONES; i++) {
5982 struct zone *zone = pgdat->node_zones + i;
5983 long max = 0;
5984
5985 /* Find valid and maximum lowmem_reserve in the zone */
5986 for (j = i; j < MAX_NR_ZONES; j++) {
5987 if (zone->lowmem_reserve[j] > max)
5988 max = zone->lowmem_reserve[j];
5989 }
5990
5991 /* we treat the high watermark as reserved pages. */
5992 max += high_wmark_pages(zone);
5993
5994 if (max > zone->managed_pages)
5995 max = zone->managed_pages;
5996 reserve_pages += max;
5997 /*
5998 * Lowmem reserves are not available to
5999 * GFP_HIGHUSER page cache allocations and
6000 * kswapd tries to balance zones to their high
6001 * watermark. As a result, neither should be
6002 * regarded as dirtyable memory, to prevent a
6003 * situation where reclaim has to clean pages
6004 * in order to balance the zones.
6005 */
6006 zone->dirty_balance_reserve = max;
6007 }
6008 }
6009 dirty_balance_reserve = reserve_pages;
6010 totalreserve_pages = reserve_pages;
6011 }
6012
6013 /*
6014 * setup_per_zone_lowmem_reserve - called whenever
6015 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6016 * has a correct pages reserved value, so an adequate number of
6017 * pages are left in the zone after a successful __alloc_pages().
6018 */
6019 static void setup_per_zone_lowmem_reserve(void)
6020 {
6021 struct pglist_data *pgdat;
6022 enum zone_type j, idx;
6023
6024 for_each_online_pgdat(pgdat) {
6025 for (j = 0; j < MAX_NR_ZONES; j++) {
6026 struct zone *zone = pgdat->node_zones + j;
6027 unsigned long managed_pages = zone->managed_pages;
6028
6029 zone->lowmem_reserve[j] = 0;
6030
6031 idx = j;
6032 while (idx) {
6033 struct zone *lower_zone;
6034
6035 idx--;
6036
6037 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6038 sysctl_lowmem_reserve_ratio[idx] = 1;
6039
6040 lower_zone = pgdat->node_zones + idx;
6041 lower_zone->lowmem_reserve[j] = managed_pages /
6042 sysctl_lowmem_reserve_ratio[idx];
6043 managed_pages += lower_zone->managed_pages;
6044 }
6045 }
6046 }
6047
6048 /* update totalreserve_pages */
6049 calculate_totalreserve_pages();
6050 }
6051
6052 static void __setup_per_zone_wmarks(void)
6053 {
6054 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6055 unsigned long pages_low = extra_free_kbytes >> (PAGE_SHIFT - 10);
6056 unsigned long lowmem_pages = 0;
6057 struct zone *zone;
6058 unsigned long flags;
6059
6060 /* Calculate total number of !ZONE_HIGHMEM pages */
6061 for_each_zone(zone) {
6062 if (!is_highmem(zone))
6063 lowmem_pages += zone->managed_pages;
6064 }
6065
6066 for_each_zone(zone) {
6067 u64 min, low;
6068
6069 spin_lock_irqsave(&zone->lock, flags);
6070 min = (u64)pages_min * zone->managed_pages;
6071 do_div(min, lowmem_pages);
6072 low = (u64)pages_low * zone->managed_pages;
6073 do_div(low, vm_total_pages);
6074
6075 if (is_highmem(zone)) {
6076 /*
6077 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6078 * need highmem pages, so cap pages_min to a small
6079 * value here.
6080 *
6081 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6082 * deltas control asynch page reclaim, and so should
6083 * not be capped for highmem.
6084 */
6085 unsigned long min_pages;
6086
6087 min_pages = zone->managed_pages / 1024;
6088 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6089 zone->watermark[WMARK_MIN] = min_pages;
6090 } else {
6091 /*
6092 * If it's a lowmem zone, reserve a number of pages
6093 * proportionate to the zone's size.
6094 */
6095 zone->watermark[WMARK_MIN] = min;
6096 }
6097
6098 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) +
6099 low + (min >> 2);
6100 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) +
6101 low + (min >> 1);
6102
6103 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6104 high_wmark_pages(zone) - low_wmark_pages(zone) -
6105 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6106
6107 spin_unlock_irqrestore(&zone->lock, flags);
6108 }
6109
6110 /* update totalreserve_pages */
6111 calculate_totalreserve_pages();
6112 }
6113
6114 /**
6115 * setup_per_zone_wmarks - called when min_free_kbytes changes
6116 * or when memory is hot-{added|removed}
6117 *
6118 * Ensures that the watermark[min,low,high] values for each zone are set
6119 * correctly with respect to min_free_kbytes.
6120 */
6121 void setup_per_zone_wmarks(void)
6122 {
6123 mutex_lock(&zonelists_mutex);
6124 __setup_per_zone_wmarks();
6125 mutex_unlock(&zonelists_mutex);
6126 }
6127
6128 /*
6129 * The inactive anon list should be small enough that the VM never has to
6130 * do too much work, but large enough that each inactive page has a chance
6131 * to be referenced again before it is swapped out.
6132 *
6133 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6134 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6135 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6136 * the anonymous pages are kept on the inactive list.
6137 *
6138 * total target max
6139 * memory ratio inactive anon
6140 * -------------------------------------
6141 * 10MB 1 5MB
6142 * 100MB 1 50MB
6143 * 1GB 3 250MB
6144 * 10GB 10 0.9GB
6145 * 100GB 31 3GB
6146 * 1TB 101 10GB
6147 * 10TB 320 32GB
6148 */
6149 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6150 {
6151 unsigned int gb, ratio;
6152
6153 /* Zone size in gigabytes */
6154 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6155 if (gb)
6156 ratio = int_sqrt(10 * gb);
6157 else
6158 ratio = 1;
6159
6160 zone->inactive_ratio = ratio;
6161 }
6162
6163 static void __meminit setup_per_zone_inactive_ratio(void)
6164 {
6165 struct zone *zone;
6166
6167 for_each_zone(zone)
6168 calculate_zone_inactive_ratio(zone);
6169 }
6170
6171 /*
6172 * Initialise min_free_kbytes.
6173 *
6174 * For small machines we want it small (128k min). For large machines
6175 * we want it large (64MB max). But it is not linear, because network
6176 * bandwidth does not increase linearly with machine size. We use
6177 *
6178 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6179 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6180 *
6181 * which yields
6182 *
6183 * 16MB: 512k
6184 * 32MB: 724k
6185 * 64MB: 1024k
6186 * 128MB: 1448k
6187 * 256MB: 2048k
6188 * 512MB: 2896k
6189 * 1024MB: 4096k
6190 * 2048MB: 5792k
6191 * 4096MB: 8192k
6192 * 8192MB: 11584k
6193 * 16384MB: 16384k
6194 */
6195 int __meminit init_per_zone_wmark_min(void)
6196 {
6197 unsigned long lowmem_kbytes;
6198 int new_min_free_kbytes;
6199
6200 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6201 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6202
6203 if (new_min_free_kbytes > user_min_free_kbytes) {
6204 min_free_kbytes = new_min_free_kbytes;
6205 if (min_free_kbytes < 128)
6206 min_free_kbytes = 128;
6207 if (min_free_kbytes > 65536)
6208 min_free_kbytes = 65536;
6209 } else {
6210 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6211 new_min_free_kbytes, user_min_free_kbytes);
6212 }
6213 setup_per_zone_wmarks();
6214 refresh_zone_stat_thresholds();
6215 setup_per_zone_lowmem_reserve();
6216 setup_per_zone_inactive_ratio();
6217 return 0;
6218 }
6219 core_initcall(init_per_zone_wmark_min)
6220
6221 /*
6222 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6223 * that we can call two helper functions whenever min_free_kbytes
6224 * or extra_free_kbytes changes.
6225 */
6226 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6227 void __user *buffer, size_t *length, loff_t *ppos)
6228 {
6229 int rc;
6230
6231 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6232 if (rc)
6233 return rc;
6234
6235 if (write) {
6236 user_min_free_kbytes = min_free_kbytes;
6237 setup_per_zone_wmarks();
6238 }
6239 return 0;
6240 }
6241
6242 #ifdef CONFIG_NUMA
6243 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6244 void __user *buffer, size_t *length, loff_t *ppos)
6245 {
6246 struct zone *zone;
6247 int rc;
6248
6249 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6250 if (rc)
6251 return rc;
6252
6253 for_each_zone(zone)
6254 zone->min_unmapped_pages = (zone->managed_pages *
6255 sysctl_min_unmapped_ratio) / 100;
6256 return 0;
6257 }
6258
6259 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6260 void __user *buffer, size_t *length, loff_t *ppos)
6261 {
6262 struct zone *zone;
6263 int rc;
6264
6265 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6266 if (rc)
6267 return rc;
6268
6269 for_each_zone(zone)
6270 zone->min_slab_pages = (zone->managed_pages *
6271 sysctl_min_slab_ratio) / 100;
6272 return 0;
6273 }
6274 #endif
6275
6276 /*
6277 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6278 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6279 * whenever sysctl_lowmem_reserve_ratio changes.
6280 *
6281 * The reserve ratio obviously has absolutely no relation with the
6282 * minimum watermarks. The lowmem reserve ratio can only make sense
6283 * if in function of the boot time zone sizes.
6284 */
6285 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6286 void __user *buffer, size_t *length, loff_t *ppos)
6287 {
6288 proc_dointvec_minmax(table, write, buffer, length, ppos);
6289 setup_per_zone_lowmem_reserve();
6290 return 0;
6291 }
6292
6293 /*
6294 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6295 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6296 * pagelist can have before it gets flushed back to buddy allocator.
6297 */
6298 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6299 void __user *buffer, size_t *length, loff_t *ppos)
6300 {
6301 struct zone *zone;
6302 int old_percpu_pagelist_fraction;
6303 int ret;
6304
6305 mutex_lock(&pcp_batch_high_lock);
6306 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6307
6308 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6309 if (!write || ret < 0)
6310 goto out;
6311
6312 /* Sanity checking to avoid pcp imbalance */
6313 if (percpu_pagelist_fraction &&
6314 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6315 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6316 ret = -EINVAL;
6317 goto out;
6318 }
6319
6320 /* No change? */
6321 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6322 goto out;
6323
6324 for_each_populated_zone(zone) {
6325 unsigned int cpu;
6326
6327 for_each_possible_cpu(cpu)
6328 pageset_set_high_and_batch(zone,
6329 per_cpu_ptr(zone->pageset, cpu));
6330 }
6331 out:
6332 mutex_unlock(&pcp_batch_high_lock);
6333 return ret;
6334 }
6335
6336 #ifdef CONFIG_NUMA
6337 int hashdist = HASHDIST_DEFAULT;
6338
6339 static int __init set_hashdist(char *str)
6340 {
6341 if (!str)
6342 return 0;
6343 hashdist = simple_strtoul(str, &str, 0);
6344 return 1;
6345 }
6346 __setup("hashdist=", set_hashdist);
6347 #endif
6348
6349 /*
6350 * allocate a large system hash table from bootmem
6351 * - it is assumed that the hash table must contain an exact power-of-2
6352 * quantity of entries
6353 * - limit is the number of hash buckets, not the total allocation size
6354 */
6355 void *__init alloc_large_system_hash(const char *tablename,
6356 unsigned long bucketsize,
6357 unsigned long numentries,
6358 int scale,
6359 int flags,
6360 unsigned int *_hash_shift,
6361 unsigned int *_hash_mask,
6362 unsigned long low_limit,
6363 unsigned long high_limit)
6364 {
6365 unsigned long long max = high_limit;
6366 unsigned long log2qty, size;
6367 void *table = NULL;
6368
6369 /* allow the kernel cmdline to have a say */
6370 if (!numentries) {
6371 /* round applicable memory size up to nearest megabyte */
6372 numentries = nr_kernel_pages;
6373
6374 /* It isn't necessary when PAGE_SIZE >= 1MB */
6375 if (PAGE_SHIFT < 20)
6376 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6377
6378 /* limit to 1 bucket per 2^scale bytes of low memory */
6379 if (scale > PAGE_SHIFT)
6380 numentries >>= (scale - PAGE_SHIFT);
6381 else
6382 numentries <<= (PAGE_SHIFT - scale);
6383
6384 /* Make sure we've got at least a 0-order allocation.. */
6385 if (unlikely(flags & HASH_SMALL)) {
6386 /* Makes no sense without HASH_EARLY */
6387 WARN_ON(!(flags & HASH_EARLY));
6388 if (!(numentries >> *_hash_shift)) {
6389 numentries = 1UL << *_hash_shift;
6390 BUG_ON(!numentries);
6391 }
6392 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6393 numentries = PAGE_SIZE / bucketsize;
6394 }
6395 numentries = roundup_pow_of_two(numentries);
6396
6397 /* limit allocation size to 1/16 total memory by default */
6398 if (max == 0) {
6399 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6400 do_div(max, bucketsize);
6401 }
6402 max = min(max, 0x80000000ULL);
6403
6404 if (numentries < low_limit)
6405 numentries = low_limit;
6406 if (numentries > max)
6407 numentries = max;
6408
6409 log2qty = ilog2(numentries);
6410
6411 do {
6412 size = bucketsize << log2qty;
6413 if (flags & HASH_EARLY)
6414 table = memblock_virt_alloc_nopanic(size, 0);
6415 else if (hashdist)
6416 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6417 else {
6418 /*
6419 * If bucketsize is not a power-of-two, we may free
6420 * some pages at the end of hash table which
6421 * alloc_pages_exact() automatically does
6422 */
6423 if (get_order(size) < MAX_ORDER) {
6424 table = alloc_pages_exact(size, GFP_ATOMIC);
6425 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6426 }
6427 }
6428 } while (!table && size > PAGE_SIZE && --log2qty);
6429
6430 if (!table)
6431 panic("Failed to allocate %s hash table\n", tablename);
6432
6433 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6434 tablename,
6435 (1UL << log2qty),
6436 ilog2(size) - PAGE_SHIFT,
6437 size);
6438
6439 if (_hash_shift)
6440 *_hash_shift = log2qty;
6441 if (_hash_mask)
6442 *_hash_mask = (1 << log2qty) - 1;
6443
6444 return table;
6445 }
6446
6447 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6448 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6449 unsigned long pfn)
6450 {
6451 #ifdef CONFIG_SPARSEMEM
6452 return __pfn_to_section(pfn)->pageblock_flags;
6453 #else
6454 return zone->pageblock_flags;
6455 #endif /* CONFIG_SPARSEMEM */
6456 }
6457
6458 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6459 {
6460 #ifdef CONFIG_SPARSEMEM
6461 pfn &= (PAGES_PER_SECTION-1);
6462 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6463 #else
6464 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6465 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6466 #endif /* CONFIG_SPARSEMEM */
6467 }
6468
6469 /**
6470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6471 * @page: The page within the block of interest
6472 * @pfn: The target page frame number
6473 * @end_bitidx: The last bit of interest to retrieve
6474 * @mask: mask of bits that the caller is interested in
6475 *
6476 * Return: pageblock_bits flags
6477 */
6478 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6479 unsigned long end_bitidx,
6480 unsigned long mask)
6481 {
6482 struct zone *zone;
6483 unsigned long *bitmap;
6484 unsigned long bitidx, word_bitidx;
6485 unsigned long word;
6486
6487 zone = page_zone(page);
6488 bitmap = get_pageblock_bitmap(zone, pfn);
6489 bitidx = pfn_to_bitidx(zone, pfn);
6490 word_bitidx = bitidx / BITS_PER_LONG;
6491 bitidx &= (BITS_PER_LONG-1);
6492
6493 word = bitmap[word_bitidx];
6494 bitidx += end_bitidx;
6495 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6496 }
6497
6498 /**
6499 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6500 * @page: The page within the block of interest
6501 * @flags: The flags to set
6502 * @pfn: The target page frame number
6503 * @end_bitidx: The last bit of interest
6504 * @mask: mask of bits that the caller is interested in
6505 */
6506 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6507 unsigned long pfn,
6508 unsigned long end_bitidx,
6509 unsigned long mask)
6510 {
6511 struct zone *zone;
6512 unsigned long *bitmap;
6513 unsigned long bitidx, word_bitidx;
6514 unsigned long old_word, word;
6515
6516 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6517
6518 zone = page_zone(page);
6519 bitmap = get_pageblock_bitmap(zone, pfn);
6520 bitidx = pfn_to_bitidx(zone, pfn);
6521 word_bitidx = bitidx / BITS_PER_LONG;
6522 bitidx &= (BITS_PER_LONG-1);
6523
6524 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6525
6526 bitidx += end_bitidx;
6527 mask <<= (BITS_PER_LONG - bitidx - 1);
6528 flags <<= (BITS_PER_LONG - bitidx - 1);
6529
6530 word = READ_ONCE(bitmap[word_bitidx]);
6531 for (;;) {
6532 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6533 if (word == old_word)
6534 break;
6535 word = old_word;
6536 }
6537 }
6538
6539 /*
6540 * This function checks whether pageblock includes unmovable pages or not.
6541 * If @count is not zero, it is okay to include less @count unmovable pages
6542 *
6543 * PageLRU check without isolation or lru_lock could race so that
6544 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6545 * expect this function should be exact.
6546 */
6547 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6548 bool skip_hwpoisoned_pages)
6549 {
6550 unsigned long pfn, iter, found;
6551 int mt;
6552
6553 /*
6554 * For avoiding noise data, lru_add_drain_all() should be called
6555 * If ZONE_MOVABLE, the zone never contains unmovable pages
6556 */
6557 if (zone_idx(zone) == ZONE_MOVABLE)
6558 return false;
6559 mt = get_pageblock_migratetype(page);
6560 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6561 return false;
6562
6563 pfn = page_to_pfn(page);
6564 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6565 unsigned long check = pfn + iter;
6566
6567 if (!pfn_valid_within(check))
6568 continue;
6569
6570 page = pfn_to_page(check);
6571
6572 /*
6573 * Hugepages are not in LRU lists, but they're movable.
6574 * We need not scan over tail pages bacause we don't
6575 * handle each tail page individually in migration.
6576 */
6577 if (PageHuge(page)) {
6578 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6579 continue;
6580 }
6581
6582 /*
6583 * We can't use page_count without pin a page
6584 * because another CPU can free compound page.
6585 * This check already skips compound tails of THP
6586 * because their page->_count is zero at all time.
6587 */
6588 if (!atomic_read(&page->_count)) {
6589 if (PageBuddy(page))
6590 iter += (1 << page_order(page)) - 1;
6591 continue;
6592 }
6593
6594 /*
6595 * The HWPoisoned page may be not in buddy system, and
6596 * page_count() is not 0.
6597 */
6598 if (skip_hwpoisoned_pages && PageHWPoison(page))
6599 continue;
6600
6601 if (!PageLRU(page))
6602 found++;
6603 /*
6604 * If there are RECLAIMABLE pages, we need to check
6605 * it. But now, memory offline itself doesn't call
6606 * shrink_node_slabs() and it still to be fixed.
6607 */
6608 /*
6609 * If the page is not RAM, page_count()should be 0.
6610 * we don't need more check. This is an _used_ not-movable page.
6611 *
6612 * The problematic thing here is PG_reserved pages. PG_reserved
6613 * is set to both of a memory hole page and a _used_ kernel
6614 * page at boot.
6615 */
6616 if (found > count)
6617 return true;
6618 }
6619 return false;
6620 }
6621
6622 bool is_pageblock_removable_nolock(struct page *page)
6623 {
6624 struct zone *zone;
6625 unsigned long pfn;
6626
6627 /*
6628 * We have to be careful here because we are iterating over memory
6629 * sections which are not zone aware so we might end up outside of
6630 * the zone but still within the section.
6631 * We have to take care about the node as well. If the node is offline
6632 * its NODE_DATA will be NULL - see page_zone.
6633 */
6634 if (!node_online(page_to_nid(page)))
6635 return false;
6636
6637 zone = page_zone(page);
6638 pfn = page_to_pfn(page);
6639 if (!zone_spans_pfn(zone, pfn))
6640 return false;
6641
6642 return !has_unmovable_pages(zone, page, 0, true);
6643 }
6644
6645 #ifdef CONFIG_CMA
6646
6647 static unsigned long pfn_max_align_down(unsigned long pfn)
6648 {
6649 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6650 pageblock_nr_pages) - 1);
6651 }
6652
6653 static unsigned long pfn_max_align_up(unsigned long pfn)
6654 {
6655 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6656 pageblock_nr_pages));
6657 }
6658
6659 /* [start, end) must belong to a single zone. */
6660 static int __alloc_contig_migrate_range(struct compact_control *cc,
6661 unsigned long start, unsigned long end)
6662 {
6663 /* This function is based on compact_zone() from compaction.c. */
6664 unsigned long nr_reclaimed;
6665 unsigned long pfn = start;
6666 unsigned int tries = 0;
6667 int ret = 0;
6668
6669 migrate_prep();
6670
6671 while (pfn < end || !list_empty(&cc->migratepages)) {
6672 if (fatal_signal_pending(current)) {
6673 ret = -EINTR;
6674 break;
6675 }
6676
6677 if (list_empty(&cc->migratepages)) {
6678 cc->nr_migratepages = 0;
6679 pfn = isolate_migratepages_range(cc, pfn, end);
6680 if (!pfn) {
6681 ret = -EINTR;
6682 break;
6683 }
6684 tries = 0;
6685 } else if (++tries == 5) {
6686 ret = ret < 0 ? ret : -EBUSY;
6687 break;
6688 }
6689
6690 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6691 &cc->migratepages);
6692 cc->nr_migratepages -= nr_reclaimed;
6693
6694 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6695 NULL, 0, cc->mode, MR_CMA);
6696 }
6697 if (ret < 0) {
6698 putback_movable_pages(&cc->migratepages);
6699 return ret;
6700 }
6701 return 0;
6702 }
6703
6704 /**
6705 * alloc_contig_range() -- tries to allocate given range of pages
6706 * @start: start PFN to allocate
6707 * @end: one-past-the-last PFN to allocate
6708 * @migratetype: migratetype of the underlaying pageblocks (either
6709 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6710 * in range must have the same migratetype and it must
6711 * be either of the two.
6712 *
6713 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6714 * aligned, however it's the caller's responsibility to guarantee that
6715 * we are the only thread that changes migrate type of pageblocks the
6716 * pages fall in.
6717 *
6718 * The PFN range must belong to a single zone.
6719 *
6720 * Returns zero on success or negative error code. On success all
6721 * pages which PFN is in [start, end) are allocated for the caller and
6722 * need to be freed with free_contig_range().
6723 */
6724 int alloc_contig_range(unsigned long start, unsigned long end,
6725 unsigned migratetype)
6726 {
6727 unsigned long outer_start, outer_end;
6728 unsigned int order;
6729 int ret = 0;
6730
6731 struct compact_control cc = {
6732 .nr_migratepages = 0,
6733 .order = -1,
6734 .zone = page_zone(pfn_to_page(start)),
6735 .mode = MIGRATE_SYNC,
6736 .ignore_skip_hint = true,
6737 };
6738 INIT_LIST_HEAD(&cc.migratepages);
6739
6740 /*
6741 * What we do here is we mark all pageblocks in range as
6742 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6743 * have different sizes, and due to the way page allocator
6744 * work, we align the range to biggest of the two pages so
6745 * that page allocator won't try to merge buddies from
6746 * different pageblocks and change MIGRATE_ISOLATE to some
6747 * other migration type.
6748 *
6749 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6750 * migrate the pages from an unaligned range (ie. pages that
6751 * we are interested in). This will put all the pages in
6752 * range back to page allocator as MIGRATE_ISOLATE.
6753 *
6754 * When this is done, we take the pages in range from page
6755 * allocator removing them from the buddy system. This way
6756 * page allocator will never consider using them.
6757 *
6758 * This lets us mark the pageblocks back as
6759 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6760 * aligned range but not in the unaligned, original range are
6761 * put back to page allocator so that buddy can use them.
6762 */
6763
6764 ret = start_isolate_page_range(pfn_max_align_down(start),
6765 pfn_max_align_up(end), migratetype,
6766 false);
6767 if (ret)
6768 return ret;
6769
6770 ret = __alloc_contig_migrate_range(&cc, start, end);
6771 if (ret)
6772 goto done;
6773
6774 /*
6775 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6776 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6777 * more, all pages in [start, end) are free in page allocator.
6778 * What we are going to do is to allocate all pages from
6779 * [start, end) (that is remove them from page allocator).
6780 *
6781 * The only problem is that pages at the beginning and at the
6782 * end of interesting range may be not aligned with pages that
6783 * page allocator holds, ie. they can be part of higher order
6784 * pages. Because of this, we reserve the bigger range and
6785 * once this is done free the pages we are not interested in.
6786 *
6787 * We don't have to hold zone->lock here because the pages are
6788 * isolated thus they won't get removed from buddy.
6789 */
6790
6791 lru_add_drain_all();
6792 drain_all_pages(cc.zone);
6793
6794 order = 0;
6795 outer_start = start;
6796 while (!PageBuddy(pfn_to_page(outer_start))) {
6797 if (++order >= MAX_ORDER) {
6798 ret = -EBUSY;
6799 goto done;
6800 }
6801 outer_start &= ~0UL << order;
6802 }
6803
6804 /* Make sure the range is really isolated. */
6805 if (test_pages_isolated(outer_start, end, false)) {
6806 pr_info("%s: [%lx, %lx) PFNs busy\n",
6807 __func__, outer_start, end);
6808 ret = -EBUSY;
6809 goto done;
6810 }
6811
6812 /* Grab isolated pages from freelists. */
6813 outer_end = isolate_freepages_range(&cc, outer_start, end);
6814 if (!outer_end) {
6815 ret = -EBUSY;
6816 goto done;
6817 }
6818
6819 /* Free head and tail (if any) */
6820 if (start != outer_start)
6821 free_contig_range(outer_start, start - outer_start);
6822 if (end != outer_end)
6823 free_contig_range(end, outer_end - end);
6824
6825 done:
6826 undo_isolate_page_range(pfn_max_align_down(start),
6827 pfn_max_align_up(end), migratetype);
6828 return ret;
6829 }
6830
6831 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6832 {
6833 unsigned int count = 0;
6834
6835 for (; nr_pages--; pfn++) {
6836 struct page *page = pfn_to_page(pfn);
6837
6838 count += page_count(page) != 1;
6839 __free_page(page);
6840 }
6841 WARN(count != 0, "%d pages are still in use!\n", count);
6842 }
6843 #endif
6844
6845 #ifdef CONFIG_MEMORY_HOTPLUG
6846 /*
6847 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6848 * page high values need to be recalulated.
6849 */
6850 void __meminit zone_pcp_update(struct zone *zone)
6851 {
6852 unsigned cpu;
6853 mutex_lock(&pcp_batch_high_lock);
6854 for_each_possible_cpu(cpu)
6855 pageset_set_high_and_batch(zone,
6856 per_cpu_ptr(zone->pageset, cpu));
6857 mutex_unlock(&pcp_batch_high_lock);
6858 }
6859 #endif
6860
6861 void zone_pcp_reset(struct zone *zone)
6862 {
6863 unsigned long flags;
6864 int cpu;
6865 struct per_cpu_pageset *pset;
6866
6867 /* avoid races with drain_pages() */
6868 local_irq_save(flags);
6869 if (zone->pageset != &boot_pageset) {
6870 for_each_online_cpu(cpu) {
6871 pset = per_cpu_ptr(zone->pageset, cpu);
6872 drain_zonestat(zone, pset);
6873 }
6874 free_percpu(zone->pageset);
6875 zone->pageset = &boot_pageset;
6876 }
6877 local_irq_restore(flags);
6878 }
6879
6880 #ifdef CONFIG_MEMORY_HOTREMOVE
6881 /*
6882 * All pages in the range must be isolated before calling this.
6883 */
6884 void
6885 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6886 {
6887 struct page *page;
6888 struct zone *zone;
6889 unsigned int order, i;
6890 unsigned long pfn;
6891 unsigned long flags;
6892 /* find the first valid pfn */
6893 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6894 if (pfn_valid(pfn))
6895 break;
6896 if (pfn == end_pfn)
6897 return;
6898 zone = page_zone(pfn_to_page(pfn));
6899 spin_lock_irqsave(&zone->lock, flags);
6900 pfn = start_pfn;
6901 while (pfn < end_pfn) {
6902 if (!pfn_valid(pfn)) {
6903 pfn++;
6904 continue;
6905 }
6906 page = pfn_to_page(pfn);
6907 /*
6908 * The HWPoisoned page may be not in buddy system, and
6909 * page_count() is not 0.
6910 */
6911 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6912 pfn++;
6913 SetPageReserved(page);
6914 continue;
6915 }
6916
6917 BUG_ON(page_count(page));
6918 BUG_ON(!PageBuddy(page));
6919 order = page_order(page);
6920 #ifdef CONFIG_DEBUG_VM
6921 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6922 pfn, 1 << order, end_pfn);
6923 #endif
6924 list_del(&page->lru);
6925 rmv_page_order(page);
6926 zone->free_area[order].nr_free--;
6927 for (i = 0; i < (1 << order); i++)
6928 SetPageReserved((page+i));
6929 pfn += (1 << order);
6930 }
6931 spin_unlock_irqrestore(&zone->lock, flags);
6932 }
6933 #endif
6934
6935 #ifdef CONFIG_MEMORY_FAILURE
6936 bool is_free_buddy_page(struct page *page)
6937 {
6938 struct zone *zone = page_zone(page);
6939 unsigned long pfn = page_to_pfn(page);
6940 unsigned long flags;
6941 unsigned int order;
6942
6943 spin_lock_irqsave(&zone->lock, flags);
6944 for (order = 0; order < MAX_ORDER; order++) {
6945 struct page *page_head = page - (pfn & ((1 << order) - 1));
6946
6947 if (PageBuddy(page_head) && page_order(page_head) >= order)
6948 break;
6949 }
6950 spin_unlock_irqrestore(&zone->lock, flags);
6951
6952 return order < MAX_ORDER;
6953 }
6954 #endif