]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - mm/page_alloc.c
mm/page_alloc: fix counting of free pages after take off from buddy
[mirror_ubuntu-hirsute-kernel.git] / mm / page_alloc.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18 #include <linux/stddef.h>
19 #include <linux/mm.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.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/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75
76 #include <asm/sections.h>
77 #include <asm/tlbflush.h>
78 #include <asm/div64.h>
79 #include "internal.h"
80 #include "shuffle.h"
81 #include "page_reporting.h"
82
83 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
84 typedef int __bitwise fpi_t;
85
86 /* No special request */
87 #define FPI_NONE ((__force fpi_t)0)
88
89 /*
90 * Skip free page reporting notification for the (possibly merged) page.
91 * This does not hinder free page reporting from grabbing the page,
92 * reporting it and marking it "reported" - it only skips notifying
93 * the free page reporting infrastructure about a newly freed page. For
94 * example, used when temporarily pulling a page from a freelist and
95 * putting it back unmodified.
96 */
97 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
98
99 /*
100 * Place the (possibly merged) page to the tail of the freelist. Will ignore
101 * page shuffling (relevant code - e.g., memory onlining - is expected to
102 * shuffle the whole zone).
103 *
104 * Note: No code should rely on this flag for correctness - it's purely
105 * to allow for optimizations when handing back either fresh pages
106 * (memory onlining) or untouched pages (page isolation, free page
107 * reporting).
108 */
109 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
110
111 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
112 static DEFINE_MUTEX(pcp_batch_high_lock);
113 #define MIN_PERCPU_PAGELIST_FRACTION (8)
114
115 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
116 DEFINE_PER_CPU(int, numa_node);
117 EXPORT_PER_CPU_SYMBOL(numa_node);
118 #endif
119
120 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
121
122 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
123 /*
124 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
125 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
126 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
127 * defined in <linux/topology.h>.
128 */
129 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
130 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
131 #endif
132
133 /* work_structs for global per-cpu drains */
134 struct pcpu_drain {
135 struct zone *zone;
136 struct work_struct work;
137 };
138 static DEFINE_MUTEX(pcpu_drain_mutex);
139 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
140
141 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
142 volatile unsigned long latent_entropy __latent_entropy;
143 EXPORT_SYMBOL(latent_entropy);
144 #endif
145
146 /*
147 * Array of node states.
148 */
149 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
150 [N_POSSIBLE] = NODE_MASK_ALL,
151 [N_ONLINE] = { { [0] = 1UL } },
152 #ifndef CONFIG_NUMA
153 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
154 #ifdef CONFIG_HIGHMEM
155 [N_HIGH_MEMORY] = { { [0] = 1UL } },
156 #endif
157 [N_MEMORY] = { { [0] = 1UL } },
158 [N_CPU] = { { [0] = 1UL } },
159 #endif /* NUMA */
160 };
161 EXPORT_SYMBOL(node_states);
162
163 atomic_long_t _totalram_pages __read_mostly;
164 EXPORT_SYMBOL(_totalram_pages);
165 unsigned long totalreserve_pages __read_mostly;
166 unsigned long totalcma_pages __read_mostly;
167
168 int percpu_pagelist_fraction;
169 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
170 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
171 EXPORT_SYMBOL(init_on_alloc);
172
173 DEFINE_STATIC_KEY_FALSE(init_on_free);
174 EXPORT_SYMBOL(init_on_free);
175
176 static bool _init_on_alloc_enabled_early __read_mostly
177 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
178 static int __init early_init_on_alloc(char *buf)
179 {
180
181 return kstrtobool(buf, &_init_on_alloc_enabled_early);
182 }
183 early_param("init_on_alloc", early_init_on_alloc);
184
185 static bool _init_on_free_enabled_early __read_mostly
186 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
187 static int __init early_init_on_free(char *buf)
188 {
189 return kstrtobool(buf, &_init_on_free_enabled_early);
190 }
191 early_param("init_on_free", early_init_on_free);
192
193 /*
194 * A cached value of the page's pageblock's migratetype, used when the page is
195 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
196 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
197 * Also the migratetype set in the page does not necessarily match the pcplist
198 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
199 * other index - this ensures that it will be put on the correct CMA freelist.
200 */
201 static inline int get_pcppage_migratetype(struct page *page)
202 {
203 return page->index;
204 }
205
206 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
207 {
208 page->index = migratetype;
209 }
210
211 #ifdef CONFIG_PM_SLEEP
212 /*
213 * The following functions are used by the suspend/hibernate code to temporarily
214 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
215 * while devices are suspended. To avoid races with the suspend/hibernate code,
216 * they should always be called with system_transition_mutex held
217 * (gfp_allowed_mask also should only be modified with system_transition_mutex
218 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
219 * with that modification).
220 */
221
222 static gfp_t saved_gfp_mask;
223
224 void pm_restore_gfp_mask(void)
225 {
226 WARN_ON(!mutex_is_locked(&system_transition_mutex));
227 if (saved_gfp_mask) {
228 gfp_allowed_mask = saved_gfp_mask;
229 saved_gfp_mask = 0;
230 }
231 }
232
233 void pm_restrict_gfp_mask(void)
234 {
235 WARN_ON(!mutex_is_locked(&system_transition_mutex));
236 WARN_ON(saved_gfp_mask);
237 saved_gfp_mask = gfp_allowed_mask;
238 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
239 }
240
241 bool pm_suspended_storage(void)
242 {
243 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
244 return false;
245 return true;
246 }
247 #endif /* CONFIG_PM_SLEEP */
248
249 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
250 unsigned int pageblock_order __read_mostly;
251 #endif
252
253 static void __free_pages_ok(struct page *page, unsigned int order,
254 fpi_t fpi_flags);
255
256 /*
257 * results with 256, 32 in the lowmem_reserve sysctl:
258 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
259 * 1G machine -> (16M dma, 784M normal, 224M high)
260 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
261 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
262 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
263 *
264 * TBD: should special case ZONE_DMA32 machines here - in those we normally
265 * don't need any ZONE_NORMAL reservation
266 */
267 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
268 #ifdef CONFIG_ZONE_DMA
269 [ZONE_DMA] = 256,
270 #endif
271 #ifdef CONFIG_ZONE_DMA32
272 [ZONE_DMA32] = 256,
273 #endif
274 [ZONE_NORMAL] = 32,
275 #ifdef CONFIG_HIGHMEM
276 [ZONE_HIGHMEM] = 0,
277 #endif
278 [ZONE_MOVABLE] = 0,
279 };
280
281 static char * const zone_names[MAX_NR_ZONES] = {
282 #ifdef CONFIG_ZONE_DMA
283 "DMA",
284 #endif
285 #ifdef CONFIG_ZONE_DMA32
286 "DMA32",
287 #endif
288 "Normal",
289 #ifdef CONFIG_HIGHMEM
290 "HighMem",
291 #endif
292 "Movable",
293 #ifdef CONFIG_ZONE_DEVICE
294 "Device",
295 #endif
296 };
297
298 const char * const migratetype_names[MIGRATE_TYPES] = {
299 "Unmovable",
300 "Movable",
301 "Reclaimable",
302 "HighAtomic",
303 #ifdef CONFIG_CMA
304 "CMA",
305 #endif
306 #ifdef CONFIG_MEMORY_ISOLATION
307 "Isolate",
308 #endif
309 };
310
311 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
312 [NULL_COMPOUND_DTOR] = NULL,
313 [COMPOUND_PAGE_DTOR] = free_compound_page,
314 #ifdef CONFIG_HUGETLB_PAGE
315 [HUGETLB_PAGE_DTOR] = free_huge_page,
316 #endif
317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
318 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
319 #endif
320 };
321
322 int min_free_kbytes = 1024;
323 int user_min_free_kbytes = -1;
324 int watermark_boost_factor __read_mostly;
325 int watermark_scale_factor = 10;
326
327 static unsigned long nr_kernel_pages __initdata;
328 static unsigned long nr_all_pages __initdata;
329 static unsigned long dma_reserve __initdata;
330
331 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
332 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
333 static unsigned long required_kernelcore __initdata;
334 static unsigned long required_kernelcore_percent __initdata;
335 static unsigned long required_movablecore __initdata;
336 static unsigned long required_movablecore_percent __initdata;
337 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
338 static bool mirrored_kernelcore __meminitdata;
339
340 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
341 int movable_zone;
342 EXPORT_SYMBOL(movable_zone);
343
344 #if MAX_NUMNODES > 1
345 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
346 unsigned int nr_online_nodes __read_mostly = 1;
347 EXPORT_SYMBOL(nr_node_ids);
348 EXPORT_SYMBOL(nr_online_nodes);
349 #endif
350
351 int page_group_by_mobility_disabled __read_mostly;
352
353 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
354 /*
355 * During boot we initialize deferred pages on-demand, as needed, but once
356 * page_alloc_init_late() has finished, the deferred pages are all initialized,
357 * and we can permanently disable that path.
358 */
359 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
360
361 /*
362 * Calling kasan_free_pages() only after deferred memory initialization
363 * has completed. Poisoning pages during deferred memory init will greatly
364 * lengthen the process and cause problem in large memory systems as the
365 * deferred pages initialization is done with interrupt disabled.
366 *
367 * Assuming that there will be no reference to those newly initialized
368 * pages before they are ever allocated, this should have no effect on
369 * KASAN memory tracking as the poison will be properly inserted at page
370 * allocation time. The only corner case is when pages are allocated by
371 * on-demand allocation and then freed again before the deferred pages
372 * initialization is done, but this is not likely to happen.
373 */
374 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
375 {
376 if (!static_branch_unlikely(&deferred_pages))
377 kasan_free_pages(page, order);
378 }
379
380 /* Returns true if the struct page for the pfn is uninitialised */
381 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
382 {
383 int nid = early_pfn_to_nid(pfn);
384
385 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
386 return true;
387
388 return false;
389 }
390
391 /*
392 * Returns true when the remaining initialisation should be deferred until
393 * later in the boot cycle when it can be parallelised.
394 */
395 static bool __meminit
396 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
397 {
398 static unsigned long prev_end_pfn, nr_initialised;
399
400 /*
401 * prev_end_pfn static that contains the end of previous zone
402 * No need to protect because called very early in boot before smp_init.
403 */
404 if (prev_end_pfn != end_pfn) {
405 prev_end_pfn = end_pfn;
406 nr_initialised = 0;
407 }
408
409 /* Always populate low zones for address-constrained allocations */
410 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
411 return false;
412
413 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
414 return true;
415 /*
416 * We start only with one section of pages, more pages are added as
417 * needed until the rest of deferred pages are initialized.
418 */
419 nr_initialised++;
420 if ((nr_initialised > PAGES_PER_SECTION) &&
421 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
422 NODE_DATA(nid)->first_deferred_pfn = pfn;
423 return true;
424 }
425 return false;
426 }
427 #else
428 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
429
430 static inline bool early_page_uninitialised(unsigned long pfn)
431 {
432 return false;
433 }
434
435 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
436 {
437 return false;
438 }
439 #endif
440
441 /* Return a pointer to the bitmap storing bits affecting a block of pages */
442 static inline unsigned long *get_pageblock_bitmap(struct page *page,
443 unsigned long pfn)
444 {
445 #ifdef CONFIG_SPARSEMEM
446 return section_to_usemap(__pfn_to_section(pfn));
447 #else
448 return page_zone(page)->pageblock_flags;
449 #endif /* CONFIG_SPARSEMEM */
450 }
451
452 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
453 {
454 #ifdef CONFIG_SPARSEMEM
455 pfn &= (PAGES_PER_SECTION-1);
456 #else
457 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
458 #endif /* CONFIG_SPARSEMEM */
459 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
460 }
461
462 static __always_inline
463 unsigned long __get_pfnblock_flags_mask(struct page *page,
464 unsigned long pfn,
465 unsigned long mask)
466 {
467 unsigned long *bitmap;
468 unsigned long bitidx, word_bitidx;
469 unsigned long word;
470
471 bitmap = get_pageblock_bitmap(page, pfn);
472 bitidx = pfn_to_bitidx(page, pfn);
473 word_bitidx = bitidx / BITS_PER_LONG;
474 bitidx &= (BITS_PER_LONG-1);
475
476 word = bitmap[word_bitidx];
477 return (word >> bitidx) & mask;
478 }
479
480 /**
481 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
482 * @page: The page within the block of interest
483 * @pfn: The target page frame number
484 * @mask: mask of bits that the caller is interested in
485 *
486 * Return: pageblock_bits flags
487 */
488 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
489 unsigned long mask)
490 {
491 return __get_pfnblock_flags_mask(page, pfn, mask);
492 }
493
494 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
495 {
496 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
497 }
498
499 /**
500 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
501 * @page: The page within the block of interest
502 * @flags: The flags to set
503 * @pfn: The target page frame number
504 * @mask: mask of bits that the caller is interested in
505 */
506 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
507 unsigned long pfn,
508 unsigned long mask)
509 {
510 unsigned long *bitmap;
511 unsigned long bitidx, word_bitidx;
512 unsigned long old_word, word;
513
514 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
515 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
516
517 bitmap = get_pageblock_bitmap(page, pfn);
518 bitidx = pfn_to_bitidx(page, pfn);
519 word_bitidx = bitidx / BITS_PER_LONG;
520 bitidx &= (BITS_PER_LONG-1);
521
522 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
523
524 mask <<= bitidx;
525 flags <<= bitidx;
526
527 word = READ_ONCE(bitmap[word_bitidx]);
528 for (;;) {
529 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
530 if (word == old_word)
531 break;
532 word = old_word;
533 }
534 }
535
536 void set_pageblock_migratetype(struct page *page, int migratetype)
537 {
538 if (unlikely(page_group_by_mobility_disabled &&
539 migratetype < MIGRATE_PCPTYPES))
540 migratetype = MIGRATE_UNMOVABLE;
541
542 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
543 page_to_pfn(page), MIGRATETYPE_MASK);
544 }
545
546 #ifdef CONFIG_DEBUG_VM
547 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
548 {
549 int ret = 0;
550 unsigned seq;
551 unsigned long pfn = page_to_pfn(page);
552 unsigned long sp, start_pfn;
553
554 do {
555 seq = zone_span_seqbegin(zone);
556 start_pfn = zone->zone_start_pfn;
557 sp = zone->spanned_pages;
558 if (!zone_spans_pfn(zone, pfn))
559 ret = 1;
560 } while (zone_span_seqretry(zone, seq));
561
562 if (ret)
563 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
564 pfn, zone_to_nid(zone), zone->name,
565 start_pfn, start_pfn + sp);
566
567 return ret;
568 }
569
570 static int page_is_consistent(struct zone *zone, struct page *page)
571 {
572 if (!pfn_valid_within(page_to_pfn(page)))
573 return 0;
574 if (zone != page_zone(page))
575 return 0;
576
577 return 1;
578 }
579 /*
580 * Temporary debugging check for pages not lying within a given zone.
581 */
582 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
583 {
584 if (page_outside_zone_boundaries(zone, page))
585 return 1;
586 if (!page_is_consistent(zone, page))
587 return 1;
588
589 return 0;
590 }
591 #else
592 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
593 {
594 return 0;
595 }
596 #endif
597
598 static void bad_page(struct page *page, const char *reason)
599 {
600 static unsigned long resume;
601 static unsigned long nr_shown;
602 static unsigned long nr_unshown;
603
604 /*
605 * Allow a burst of 60 reports, then keep quiet for that minute;
606 * or allow a steady drip of one report per second.
607 */
608 if (nr_shown == 60) {
609 if (time_before(jiffies, resume)) {
610 nr_unshown++;
611 goto out;
612 }
613 if (nr_unshown) {
614 pr_alert(
615 "BUG: Bad page state: %lu messages suppressed\n",
616 nr_unshown);
617 nr_unshown = 0;
618 }
619 nr_shown = 0;
620 }
621 if (nr_shown++ == 0)
622 resume = jiffies + 60 * HZ;
623
624 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
625 current->comm, page_to_pfn(page));
626 __dump_page(page, reason);
627 dump_page_owner(page);
628
629 print_modules();
630 dump_stack();
631 out:
632 /* Leave bad fields for debug, except PageBuddy could make trouble */
633 page_mapcount_reset(page); /* remove PageBuddy */
634 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
635 }
636
637 /*
638 * Higher-order pages are called "compound pages". They are structured thusly:
639 *
640 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
641 *
642 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
643 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
644 *
645 * The first tail page's ->compound_dtor holds the offset in array of compound
646 * page destructors. See compound_page_dtors.
647 *
648 * The first tail page's ->compound_order holds the order of allocation.
649 * This usage means that zero-order pages may not be compound.
650 */
651
652 void free_compound_page(struct page *page)
653 {
654 mem_cgroup_uncharge(page);
655 __free_pages_ok(page, compound_order(page), FPI_NONE);
656 }
657
658 void prep_compound_page(struct page *page, unsigned int order)
659 {
660 int i;
661 int nr_pages = 1 << order;
662
663 __SetPageHead(page);
664 for (i = 1; i < nr_pages; i++) {
665 struct page *p = page + i;
666 set_page_count(p, 0);
667 p->mapping = TAIL_MAPPING;
668 set_compound_head(p, page);
669 }
670
671 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
672 set_compound_order(page, order);
673 atomic_set(compound_mapcount_ptr(page), -1);
674 if (hpage_pincount_available(page))
675 atomic_set(compound_pincount_ptr(page), 0);
676 }
677
678 #ifdef CONFIG_DEBUG_PAGEALLOC
679 unsigned int _debug_guardpage_minorder;
680
681 bool _debug_pagealloc_enabled_early __read_mostly
682 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
683 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
684 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
685 EXPORT_SYMBOL(_debug_pagealloc_enabled);
686
687 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
688
689 static int __init early_debug_pagealloc(char *buf)
690 {
691 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
692 }
693 early_param("debug_pagealloc", early_debug_pagealloc);
694
695 static int __init debug_guardpage_minorder_setup(char *buf)
696 {
697 unsigned long res;
698
699 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
700 pr_err("Bad debug_guardpage_minorder value\n");
701 return 0;
702 }
703 _debug_guardpage_minorder = res;
704 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
705 return 0;
706 }
707 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
708
709 static inline bool set_page_guard(struct zone *zone, struct page *page,
710 unsigned int order, int migratetype)
711 {
712 if (!debug_guardpage_enabled())
713 return false;
714
715 if (order >= debug_guardpage_minorder())
716 return false;
717
718 __SetPageGuard(page);
719 INIT_LIST_HEAD(&page->lru);
720 set_page_private(page, order);
721 /* Guard pages are not available for any usage */
722 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
723
724 return true;
725 }
726
727 static inline void clear_page_guard(struct zone *zone, struct page *page,
728 unsigned int order, int migratetype)
729 {
730 if (!debug_guardpage_enabled())
731 return;
732
733 __ClearPageGuard(page);
734
735 set_page_private(page, 0);
736 if (!is_migrate_isolate(migratetype))
737 __mod_zone_freepage_state(zone, (1 << order), migratetype);
738 }
739 #else
740 static inline bool set_page_guard(struct zone *zone, struct page *page,
741 unsigned int order, int migratetype) { return false; }
742 static inline void clear_page_guard(struct zone *zone, struct page *page,
743 unsigned int order, int migratetype) {}
744 #endif
745
746 /*
747 * Enable static keys related to various memory debugging and hardening options.
748 * Some override others, and depend on early params that are evaluated in the
749 * order of appearance. So we need to first gather the full picture of what was
750 * enabled, and then make decisions.
751 */
752 void init_mem_debugging_and_hardening(void)
753 {
754 bool page_poisoning_requested = false;
755
756 #ifdef CONFIG_PAGE_POISONING
757 /*
758 * Page poisoning is debug page alloc for some arches. If
759 * either of those options are enabled, enable poisoning.
760 */
761 if (page_poisoning_enabled() ||
762 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
763 debug_pagealloc_enabled())) {
764 static_branch_enable(&_page_poisoning_enabled);
765 page_poisoning_requested = true;
766 }
767 #endif
768
769 if (_init_on_alloc_enabled_early) {
770 if (page_poisoning_requested)
771 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
772 "will take precedence over init_on_alloc\n");
773 else
774 static_branch_enable(&init_on_alloc);
775 }
776 if (_init_on_free_enabled_early) {
777 if (page_poisoning_requested)
778 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
779 "will take precedence over init_on_free\n");
780 else
781 static_branch_enable(&init_on_free);
782 }
783
784 #ifdef CONFIG_DEBUG_PAGEALLOC
785 if (!debug_pagealloc_enabled())
786 return;
787
788 static_branch_enable(&_debug_pagealloc_enabled);
789
790 if (!debug_guardpage_minorder())
791 return;
792
793 static_branch_enable(&_debug_guardpage_enabled);
794 #endif
795 }
796
797 static inline void set_buddy_order(struct page *page, unsigned int order)
798 {
799 set_page_private(page, order);
800 __SetPageBuddy(page);
801 }
802
803 /*
804 * This function checks whether a page is free && is the buddy
805 * we can coalesce a page and its buddy if
806 * (a) the buddy is not in a hole (check before calling!) &&
807 * (b) the buddy is in the buddy system &&
808 * (c) a page and its buddy have the same order &&
809 * (d) a page and its buddy are in the same zone.
810 *
811 * For recording whether a page is in the buddy system, we set PageBuddy.
812 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
813 *
814 * For recording page's order, we use page_private(page).
815 */
816 static inline bool page_is_buddy(struct page *page, struct page *buddy,
817 unsigned int order)
818 {
819 if (!page_is_guard(buddy) && !PageBuddy(buddy))
820 return false;
821
822 if (buddy_order(buddy) != order)
823 return false;
824
825 /*
826 * zone check is done late to avoid uselessly calculating
827 * zone/node ids for pages that could never merge.
828 */
829 if (page_zone_id(page) != page_zone_id(buddy))
830 return false;
831
832 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
833
834 return true;
835 }
836
837 #ifdef CONFIG_COMPACTION
838 static inline struct capture_control *task_capc(struct zone *zone)
839 {
840 struct capture_control *capc = current->capture_control;
841
842 return unlikely(capc) &&
843 !(current->flags & PF_KTHREAD) &&
844 !capc->page &&
845 capc->cc->zone == zone ? capc : NULL;
846 }
847
848 static inline bool
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
851 {
852 if (!capc || order != capc->cc->order)
853 return false;
854
855 /* Do not accidentally pollute CMA or isolated regions*/
856 if (is_migrate_cma(migratetype) ||
857 is_migrate_isolate(migratetype))
858 return false;
859
860 /*
861 * Do not let lower order allocations polluate a movable pageblock.
862 * This might let an unmovable request use a reclaimable pageblock
863 * and vice-versa but no more than normal fallback logic which can
864 * have trouble finding a high-order free page.
865 */
866 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
867 return false;
868
869 capc->page = page;
870 return true;
871 }
872
873 #else
874 static inline struct capture_control *task_capc(struct zone *zone)
875 {
876 return NULL;
877 }
878
879 static inline bool
880 compaction_capture(struct capture_control *capc, struct page *page,
881 int order, int migratetype)
882 {
883 return false;
884 }
885 #endif /* CONFIG_COMPACTION */
886
887 /* Used for pages not on another list */
888 static inline void add_to_free_list(struct page *page, struct zone *zone,
889 unsigned int order, int migratetype)
890 {
891 struct free_area *area = &zone->free_area[order];
892
893 list_add(&page->lru, &area->free_list[migratetype]);
894 area->nr_free++;
895 }
896
897 /* Used for pages not on another list */
898 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
899 unsigned int order, int migratetype)
900 {
901 struct free_area *area = &zone->free_area[order];
902
903 list_add_tail(&page->lru, &area->free_list[migratetype]);
904 area->nr_free++;
905 }
906
907 /*
908 * Used for pages which are on another list. Move the pages to the tail
909 * of the list - so the moved pages won't immediately be considered for
910 * allocation again (e.g., optimization for memory onlining).
911 */
912 static inline void move_to_free_list(struct page *page, struct zone *zone,
913 unsigned int order, int migratetype)
914 {
915 struct free_area *area = &zone->free_area[order];
916
917 list_move_tail(&page->lru, &area->free_list[migratetype]);
918 }
919
920 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
921 unsigned int order)
922 {
923 /* clear reported state and update reported page count */
924 if (page_reported(page))
925 __ClearPageReported(page);
926
927 list_del(&page->lru);
928 __ClearPageBuddy(page);
929 set_page_private(page, 0);
930 zone->free_area[order].nr_free--;
931 }
932
933 /*
934 * If this is not the largest possible page, check if the buddy
935 * of the next-highest order is free. If it is, it's possible
936 * that pages are being freed that will coalesce soon. In case,
937 * that is happening, add the free page to the tail of the list
938 * so it's less likely to be used soon and more likely to be merged
939 * as a higher order page
940 */
941 static inline bool
942 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
943 struct page *page, unsigned int order)
944 {
945 struct page *higher_page, *higher_buddy;
946 unsigned long combined_pfn;
947
948 if (order >= MAX_ORDER - 2)
949 return false;
950
951 if (!pfn_valid_within(buddy_pfn))
952 return false;
953
954 combined_pfn = buddy_pfn & pfn;
955 higher_page = page + (combined_pfn - pfn);
956 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
957 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
958
959 return pfn_valid_within(buddy_pfn) &&
960 page_is_buddy(higher_page, higher_buddy, order + 1);
961 }
962
963 /*
964 * Freeing function for a buddy system allocator.
965 *
966 * The concept of a buddy system is to maintain direct-mapped table
967 * (containing bit values) for memory blocks of various "orders".
968 * The bottom level table contains the map for the smallest allocatable
969 * units of memory (here, pages), and each level above it describes
970 * pairs of units from the levels below, hence, "buddies".
971 * At a high level, all that happens here is marking the table entry
972 * at the bottom level available, and propagating the changes upward
973 * as necessary, plus some accounting needed to play nicely with other
974 * parts of the VM system.
975 * At each level, we keep a list of pages, which are heads of continuous
976 * free pages of length of (1 << order) and marked with PageBuddy.
977 * Page's order is recorded in page_private(page) field.
978 * So when we are allocating or freeing one, we can derive the state of the
979 * other. That is, if we allocate a small block, and both were
980 * free, the remainder of the region must be split into blocks.
981 * If a block is freed, and its buddy is also free, then this
982 * triggers coalescing into a block of larger size.
983 *
984 * -- nyc
985 */
986
987 static inline void __free_one_page(struct page *page,
988 unsigned long pfn,
989 struct zone *zone, unsigned int order,
990 int migratetype, fpi_t fpi_flags)
991 {
992 struct capture_control *capc = task_capc(zone);
993 unsigned long buddy_pfn;
994 unsigned long combined_pfn;
995 unsigned int max_order;
996 struct page *buddy;
997 bool to_tail;
998
999 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1000
1001 VM_BUG_ON(!zone_is_initialized(zone));
1002 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1003
1004 VM_BUG_ON(migratetype == -1);
1005 if (likely(!is_migrate_isolate(migratetype)))
1006 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1007
1008 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1009 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1010
1011 continue_merging:
1012 while (order < max_order) {
1013 if (compaction_capture(capc, page, order, migratetype)) {
1014 __mod_zone_freepage_state(zone, -(1 << order),
1015 migratetype);
1016 return;
1017 }
1018 buddy_pfn = __find_buddy_pfn(pfn, order);
1019 buddy = page + (buddy_pfn - pfn);
1020
1021 if (!pfn_valid_within(buddy_pfn))
1022 goto done_merging;
1023 if (!page_is_buddy(page, buddy, order))
1024 goto done_merging;
1025 /*
1026 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1027 * merge with it and move up one order.
1028 */
1029 if (page_is_guard(buddy))
1030 clear_page_guard(zone, buddy, order, migratetype);
1031 else
1032 del_page_from_free_list(buddy, zone, order);
1033 combined_pfn = buddy_pfn & pfn;
1034 page = page + (combined_pfn - pfn);
1035 pfn = combined_pfn;
1036 order++;
1037 }
1038 if (order < MAX_ORDER - 1) {
1039 /* If we are here, it means order is >= pageblock_order.
1040 * We want to prevent merge between freepages on isolate
1041 * pageblock and normal pageblock. Without this, pageblock
1042 * isolation could cause incorrect freepage or CMA accounting.
1043 *
1044 * We don't want to hit this code for the more frequent
1045 * low-order merging.
1046 */
1047 if (unlikely(has_isolate_pageblock(zone))) {
1048 int buddy_mt;
1049
1050 buddy_pfn = __find_buddy_pfn(pfn, order);
1051 buddy = page + (buddy_pfn - pfn);
1052 buddy_mt = get_pageblock_migratetype(buddy);
1053
1054 if (migratetype != buddy_mt
1055 && (is_migrate_isolate(migratetype) ||
1056 is_migrate_isolate(buddy_mt)))
1057 goto done_merging;
1058 }
1059 max_order = order + 1;
1060 goto continue_merging;
1061 }
1062
1063 done_merging:
1064 set_buddy_order(page, order);
1065
1066 if (fpi_flags & FPI_TO_TAIL)
1067 to_tail = true;
1068 else if (is_shuffle_order(order))
1069 to_tail = shuffle_pick_tail();
1070 else
1071 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1072
1073 if (to_tail)
1074 add_to_free_list_tail(page, zone, order, migratetype);
1075 else
1076 add_to_free_list(page, zone, order, migratetype);
1077
1078 /* Notify page reporting subsystem of freed page */
1079 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1080 page_reporting_notify_free(order);
1081 }
1082
1083 /*
1084 * A bad page could be due to a number of fields. Instead of multiple branches,
1085 * try and check multiple fields with one check. The caller must do a detailed
1086 * check if necessary.
1087 */
1088 static inline bool page_expected_state(struct page *page,
1089 unsigned long check_flags)
1090 {
1091 if (unlikely(atomic_read(&page->_mapcount) != -1))
1092 return false;
1093
1094 if (unlikely((unsigned long)page->mapping |
1095 page_ref_count(page) |
1096 #ifdef CONFIG_MEMCG
1097 (unsigned long)page_memcg(page) |
1098 #endif
1099 (page->flags & check_flags)))
1100 return false;
1101
1102 return true;
1103 }
1104
1105 static const char *page_bad_reason(struct page *page, unsigned long flags)
1106 {
1107 const char *bad_reason = NULL;
1108
1109 if (unlikely(atomic_read(&page->_mapcount) != -1))
1110 bad_reason = "nonzero mapcount";
1111 if (unlikely(page->mapping != NULL))
1112 bad_reason = "non-NULL mapping";
1113 if (unlikely(page_ref_count(page) != 0))
1114 bad_reason = "nonzero _refcount";
1115 if (unlikely(page->flags & flags)) {
1116 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1117 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1118 else
1119 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1120 }
1121 #ifdef CONFIG_MEMCG
1122 if (unlikely(page_memcg(page)))
1123 bad_reason = "page still charged to cgroup";
1124 #endif
1125 return bad_reason;
1126 }
1127
1128 static void check_free_page_bad(struct page *page)
1129 {
1130 bad_page(page,
1131 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1132 }
1133
1134 static inline int check_free_page(struct page *page)
1135 {
1136 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1137 return 0;
1138
1139 /* Something has gone sideways, find it */
1140 check_free_page_bad(page);
1141 return 1;
1142 }
1143
1144 static int free_tail_pages_check(struct page *head_page, struct page *page)
1145 {
1146 int ret = 1;
1147
1148 /*
1149 * We rely page->lru.next never has bit 0 set, unless the page
1150 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1151 */
1152 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1153
1154 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1155 ret = 0;
1156 goto out;
1157 }
1158 switch (page - head_page) {
1159 case 1:
1160 /* the first tail page: ->mapping may be compound_mapcount() */
1161 if (unlikely(compound_mapcount(page))) {
1162 bad_page(page, "nonzero compound_mapcount");
1163 goto out;
1164 }
1165 break;
1166 case 2:
1167 /*
1168 * the second tail page: ->mapping is
1169 * deferred_list.next -- ignore value.
1170 */
1171 break;
1172 default:
1173 if (page->mapping != TAIL_MAPPING) {
1174 bad_page(page, "corrupted mapping in tail page");
1175 goto out;
1176 }
1177 break;
1178 }
1179 if (unlikely(!PageTail(page))) {
1180 bad_page(page, "PageTail not set");
1181 goto out;
1182 }
1183 if (unlikely(compound_head(page) != head_page)) {
1184 bad_page(page, "compound_head not consistent");
1185 goto out;
1186 }
1187 ret = 0;
1188 out:
1189 page->mapping = NULL;
1190 clear_compound_head(page);
1191 return ret;
1192 }
1193
1194 static void kernel_init_free_pages(struct page *page, int numpages)
1195 {
1196 int i;
1197
1198 /* s390's use of memset() could override KASAN redzones. */
1199 kasan_disable_current();
1200 for (i = 0; i < numpages; i++) {
1201 u8 tag = page_kasan_tag(page + i);
1202 page_kasan_tag_reset(page + i);
1203 clear_highpage(page + i);
1204 page_kasan_tag_set(page + i, tag);
1205 }
1206 kasan_enable_current();
1207 }
1208
1209 static __always_inline bool free_pages_prepare(struct page *page,
1210 unsigned int order, bool check_free)
1211 {
1212 int bad = 0;
1213
1214 VM_BUG_ON_PAGE(PageTail(page), page);
1215
1216 trace_mm_page_free(page, order);
1217
1218 if (unlikely(PageHWPoison(page)) && !order) {
1219 /*
1220 * Do not let hwpoison pages hit pcplists/buddy
1221 * Untie memcg state and reset page's owner
1222 */
1223 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1224 __memcg_kmem_uncharge_page(page, order);
1225 reset_page_owner(page, order);
1226 return false;
1227 }
1228
1229 /*
1230 * Check tail pages before head page information is cleared to
1231 * avoid checking PageCompound for order-0 pages.
1232 */
1233 if (unlikely(order)) {
1234 bool compound = PageCompound(page);
1235 int i;
1236
1237 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1238
1239 if (compound)
1240 ClearPageDoubleMap(page);
1241 for (i = 1; i < (1 << order); i++) {
1242 if (compound)
1243 bad += free_tail_pages_check(page, page + i);
1244 if (unlikely(check_free_page(page + i))) {
1245 bad++;
1246 continue;
1247 }
1248 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1249 }
1250 }
1251 if (PageMappingFlags(page))
1252 page->mapping = NULL;
1253 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1254 __memcg_kmem_uncharge_page(page, order);
1255 if (check_free)
1256 bad += check_free_page(page);
1257 if (bad)
1258 return false;
1259
1260 page_cpupid_reset_last(page);
1261 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1262 reset_page_owner(page, order);
1263
1264 if (!PageHighMem(page)) {
1265 debug_check_no_locks_freed(page_address(page),
1266 PAGE_SIZE << order);
1267 debug_check_no_obj_freed(page_address(page),
1268 PAGE_SIZE << order);
1269 }
1270 if (want_init_on_free())
1271 kernel_init_free_pages(page, 1 << order);
1272
1273 kernel_poison_pages(page, 1 << order);
1274
1275 /*
1276 * With hardware tag-based KASAN, memory tags must be set before the
1277 * page becomes unavailable via debug_pagealloc or arch_free_page.
1278 */
1279 kasan_free_nondeferred_pages(page, order);
1280
1281 /*
1282 * arch_free_page() can make the page's contents inaccessible. s390
1283 * does this. So nothing which can access the page's contents should
1284 * happen after this.
1285 */
1286 arch_free_page(page, order);
1287
1288 debug_pagealloc_unmap_pages(page, 1 << order);
1289
1290 return true;
1291 }
1292
1293 #ifdef CONFIG_DEBUG_VM
1294 /*
1295 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1296 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1297 * moved from pcp lists to free lists.
1298 */
1299 static bool free_pcp_prepare(struct page *page)
1300 {
1301 return free_pages_prepare(page, 0, true);
1302 }
1303
1304 static bool bulkfree_pcp_prepare(struct page *page)
1305 {
1306 if (debug_pagealloc_enabled_static())
1307 return check_free_page(page);
1308 else
1309 return false;
1310 }
1311 #else
1312 /*
1313 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1314 * moving from pcp lists to free list in order to reduce overhead. With
1315 * debug_pagealloc enabled, they are checked also immediately when being freed
1316 * to the pcp lists.
1317 */
1318 static bool free_pcp_prepare(struct page *page)
1319 {
1320 if (debug_pagealloc_enabled_static())
1321 return free_pages_prepare(page, 0, true);
1322 else
1323 return free_pages_prepare(page, 0, false);
1324 }
1325
1326 static bool bulkfree_pcp_prepare(struct page *page)
1327 {
1328 return check_free_page(page);
1329 }
1330 #endif /* CONFIG_DEBUG_VM */
1331
1332 static inline void prefetch_buddy(struct page *page)
1333 {
1334 unsigned long pfn = page_to_pfn(page);
1335 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1336 struct page *buddy = page + (buddy_pfn - pfn);
1337
1338 prefetch(buddy);
1339 }
1340
1341 /*
1342 * Frees a number of pages from the PCP lists
1343 * Assumes all pages on list are in same zone, and of same order.
1344 * count is the number of pages to free.
1345 *
1346 * If the zone was previously in an "all pages pinned" state then look to
1347 * see if this freeing clears that state.
1348 *
1349 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1350 * pinned" detection logic.
1351 */
1352 static void free_pcppages_bulk(struct zone *zone, int count,
1353 struct per_cpu_pages *pcp)
1354 {
1355 int migratetype = 0;
1356 int batch_free = 0;
1357 int prefetch_nr = READ_ONCE(pcp->batch);
1358 bool isolated_pageblocks;
1359 struct page *page, *tmp;
1360 LIST_HEAD(head);
1361
1362 /*
1363 * Ensure proper count is passed which otherwise would stuck in the
1364 * below while (list_empty(list)) loop.
1365 */
1366 count = min(pcp->count, count);
1367 while (count) {
1368 struct list_head *list;
1369
1370 /*
1371 * Remove pages from lists in a round-robin fashion. A
1372 * batch_free count is maintained that is incremented when an
1373 * empty list is encountered. This is so more pages are freed
1374 * off fuller lists instead of spinning excessively around empty
1375 * lists
1376 */
1377 do {
1378 batch_free++;
1379 if (++migratetype == MIGRATE_PCPTYPES)
1380 migratetype = 0;
1381 list = &pcp->lists[migratetype];
1382 } while (list_empty(list));
1383
1384 /* This is the only non-empty list. Free them all. */
1385 if (batch_free == MIGRATE_PCPTYPES)
1386 batch_free = count;
1387
1388 do {
1389 page = list_last_entry(list, struct page, lru);
1390 /* must delete to avoid corrupting pcp list */
1391 list_del(&page->lru);
1392 pcp->count--;
1393
1394 if (bulkfree_pcp_prepare(page))
1395 continue;
1396
1397 list_add_tail(&page->lru, &head);
1398
1399 /*
1400 * We are going to put the page back to the global
1401 * pool, prefetch its buddy to speed up later access
1402 * under zone->lock. It is believed the overhead of
1403 * an additional test and calculating buddy_pfn here
1404 * can be offset by reduced memory latency later. To
1405 * avoid excessive prefetching due to large count, only
1406 * prefetch buddy for the first pcp->batch nr of pages.
1407 */
1408 if (prefetch_nr) {
1409 prefetch_buddy(page);
1410 prefetch_nr--;
1411 }
1412 } while (--count && --batch_free && !list_empty(list));
1413 }
1414
1415 spin_lock(&zone->lock);
1416 isolated_pageblocks = has_isolate_pageblock(zone);
1417
1418 /*
1419 * Use safe version since after __free_one_page(),
1420 * page->lru.next will not point to original list.
1421 */
1422 list_for_each_entry_safe(page, tmp, &head, lru) {
1423 int mt = get_pcppage_migratetype(page);
1424 /* MIGRATE_ISOLATE page should not go to pcplists */
1425 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1426 /* Pageblock could have been isolated meanwhile */
1427 if (unlikely(isolated_pageblocks))
1428 mt = get_pageblock_migratetype(page);
1429
1430 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1431 trace_mm_page_pcpu_drain(page, 0, mt);
1432 }
1433 spin_unlock(&zone->lock);
1434 }
1435
1436 static void free_one_page(struct zone *zone,
1437 struct page *page, unsigned long pfn,
1438 unsigned int order,
1439 int migratetype, fpi_t fpi_flags)
1440 {
1441 spin_lock(&zone->lock);
1442 if (unlikely(has_isolate_pageblock(zone) ||
1443 is_migrate_isolate(migratetype))) {
1444 migratetype = get_pfnblock_migratetype(page, pfn);
1445 }
1446 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1447 spin_unlock(&zone->lock);
1448 }
1449
1450 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1451 unsigned long zone, int nid)
1452 {
1453 mm_zero_struct_page(page);
1454 set_page_links(page, zone, nid, pfn);
1455 init_page_count(page);
1456 page_mapcount_reset(page);
1457 page_cpupid_reset_last(page);
1458 page_kasan_tag_reset(page);
1459
1460 INIT_LIST_HEAD(&page->lru);
1461 #ifdef WANT_PAGE_VIRTUAL
1462 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1463 if (!is_highmem_idx(zone))
1464 set_page_address(page, __va(pfn << PAGE_SHIFT));
1465 #endif
1466 }
1467
1468 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1469 static void __meminit init_reserved_page(unsigned long pfn)
1470 {
1471 pg_data_t *pgdat;
1472 int nid, zid;
1473
1474 if (!early_page_uninitialised(pfn))
1475 return;
1476
1477 nid = early_pfn_to_nid(pfn);
1478 pgdat = NODE_DATA(nid);
1479
1480 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1481 struct zone *zone = &pgdat->node_zones[zid];
1482
1483 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1484 break;
1485 }
1486 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1487 }
1488 #else
1489 static inline void init_reserved_page(unsigned long pfn)
1490 {
1491 }
1492 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1493
1494 /*
1495 * Initialised pages do not have PageReserved set. This function is
1496 * called for each range allocated by the bootmem allocator and
1497 * marks the pages PageReserved. The remaining valid pages are later
1498 * sent to the buddy page allocator.
1499 */
1500 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1501 {
1502 unsigned long start_pfn = PFN_DOWN(start);
1503 unsigned long end_pfn = PFN_UP(end);
1504
1505 for (; start_pfn < end_pfn; start_pfn++) {
1506 if (pfn_valid(start_pfn)) {
1507 struct page *page = pfn_to_page(start_pfn);
1508
1509 init_reserved_page(start_pfn);
1510
1511 /* Avoid false-positive PageTail() */
1512 INIT_LIST_HEAD(&page->lru);
1513
1514 /*
1515 * no need for atomic set_bit because the struct
1516 * page is not visible yet so nobody should
1517 * access it yet.
1518 */
1519 __SetPageReserved(page);
1520 }
1521 }
1522 }
1523
1524 static void __free_pages_ok(struct page *page, unsigned int order,
1525 fpi_t fpi_flags)
1526 {
1527 unsigned long flags;
1528 int migratetype;
1529 unsigned long pfn = page_to_pfn(page);
1530
1531 if (!free_pages_prepare(page, order, true))
1532 return;
1533
1534 migratetype = get_pfnblock_migratetype(page, pfn);
1535 local_irq_save(flags);
1536 __count_vm_events(PGFREE, 1 << order);
1537 free_one_page(page_zone(page), page, pfn, order, migratetype,
1538 fpi_flags);
1539 local_irq_restore(flags);
1540 }
1541
1542 void __free_pages_core(struct page *page, unsigned int order)
1543 {
1544 unsigned int nr_pages = 1 << order;
1545 struct page *p = page;
1546 unsigned int loop;
1547
1548 /*
1549 * When initializing the memmap, __init_single_page() sets the refcount
1550 * of all pages to 1 ("allocated"/"not free"). We have to set the
1551 * refcount of all involved pages to 0.
1552 */
1553 prefetchw(p);
1554 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1555 prefetchw(p + 1);
1556 __ClearPageReserved(p);
1557 set_page_count(p, 0);
1558 }
1559 __ClearPageReserved(p);
1560 set_page_count(p, 0);
1561
1562 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1563
1564 /*
1565 * Bypass PCP and place fresh pages right to the tail, primarily
1566 * relevant for memory onlining.
1567 */
1568 __free_pages_ok(page, order, FPI_TO_TAIL);
1569 }
1570
1571 #ifdef CONFIG_NEED_MULTIPLE_NODES
1572
1573 /*
1574 * During memory init memblocks map pfns to nids. The search is expensive and
1575 * this caches recent lookups. The implementation of __early_pfn_to_nid
1576 * treats start/end as pfns.
1577 */
1578 struct mminit_pfnnid_cache {
1579 unsigned long last_start;
1580 unsigned long last_end;
1581 int last_nid;
1582 };
1583
1584 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1585
1586 /*
1587 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1588 */
1589 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1590 struct mminit_pfnnid_cache *state)
1591 {
1592 unsigned long start_pfn, end_pfn;
1593 int nid;
1594
1595 if (state->last_start <= pfn && pfn < state->last_end)
1596 return state->last_nid;
1597
1598 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1599 if (nid != NUMA_NO_NODE) {
1600 state->last_start = start_pfn;
1601 state->last_end = end_pfn;
1602 state->last_nid = nid;
1603 }
1604
1605 return nid;
1606 }
1607
1608 int __meminit early_pfn_to_nid(unsigned long pfn)
1609 {
1610 static DEFINE_SPINLOCK(early_pfn_lock);
1611 int nid;
1612
1613 spin_lock(&early_pfn_lock);
1614 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1615 if (nid < 0)
1616 nid = first_online_node;
1617 spin_unlock(&early_pfn_lock);
1618
1619 return nid;
1620 }
1621 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1622
1623 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1624 unsigned int order)
1625 {
1626 if (early_page_uninitialised(pfn))
1627 return;
1628 __free_pages_core(page, order);
1629 }
1630
1631 /*
1632 * Check that the whole (or subset of) a pageblock given by the interval of
1633 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1634 * with the migration of free compaction scanner. The scanners then need to
1635 * use only pfn_valid_within() check for arches that allow holes within
1636 * pageblocks.
1637 *
1638 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1639 *
1640 * It's possible on some configurations to have a setup like node0 node1 node0
1641 * i.e. it's possible that all pages within a zones range of pages do not
1642 * belong to a single zone. We assume that a border between node0 and node1
1643 * can occur within a single pageblock, but not a node0 node1 node0
1644 * interleaving within a single pageblock. It is therefore sufficient to check
1645 * the first and last page of a pageblock and avoid checking each individual
1646 * page in a pageblock.
1647 */
1648 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1649 unsigned long end_pfn, struct zone *zone)
1650 {
1651 struct page *start_page;
1652 struct page *end_page;
1653
1654 /* end_pfn is one past the range we are checking */
1655 end_pfn--;
1656
1657 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1658 return NULL;
1659
1660 start_page = pfn_to_online_page(start_pfn);
1661 if (!start_page)
1662 return NULL;
1663
1664 if (page_zone(start_page) != zone)
1665 return NULL;
1666
1667 end_page = pfn_to_page(end_pfn);
1668
1669 /* This gives a shorter code than deriving page_zone(end_page) */
1670 if (page_zone_id(start_page) != page_zone_id(end_page))
1671 return NULL;
1672
1673 return start_page;
1674 }
1675
1676 void set_zone_contiguous(struct zone *zone)
1677 {
1678 unsigned long block_start_pfn = zone->zone_start_pfn;
1679 unsigned long block_end_pfn;
1680
1681 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1682 for (; block_start_pfn < zone_end_pfn(zone);
1683 block_start_pfn = block_end_pfn,
1684 block_end_pfn += pageblock_nr_pages) {
1685
1686 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1687
1688 if (!__pageblock_pfn_to_page(block_start_pfn,
1689 block_end_pfn, zone))
1690 return;
1691 cond_resched();
1692 }
1693
1694 /* We confirm that there is no hole */
1695 zone->contiguous = true;
1696 }
1697
1698 void clear_zone_contiguous(struct zone *zone)
1699 {
1700 zone->contiguous = false;
1701 }
1702
1703 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1704 static void __init deferred_free_range(unsigned long pfn,
1705 unsigned long nr_pages)
1706 {
1707 struct page *page;
1708 unsigned long i;
1709
1710 if (!nr_pages)
1711 return;
1712
1713 page = pfn_to_page(pfn);
1714
1715 /* Free a large naturally-aligned chunk if possible */
1716 if (nr_pages == pageblock_nr_pages &&
1717 (pfn & (pageblock_nr_pages - 1)) == 0) {
1718 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1719 __free_pages_core(page, pageblock_order);
1720 return;
1721 }
1722
1723 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1724 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1725 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1726 __free_pages_core(page, 0);
1727 }
1728 }
1729
1730 /* Completion tracking for deferred_init_memmap() threads */
1731 static atomic_t pgdat_init_n_undone __initdata;
1732 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1733
1734 static inline void __init pgdat_init_report_one_done(void)
1735 {
1736 if (atomic_dec_and_test(&pgdat_init_n_undone))
1737 complete(&pgdat_init_all_done_comp);
1738 }
1739
1740 /*
1741 * Returns true if page needs to be initialized or freed to buddy allocator.
1742 *
1743 * First we check if pfn is valid on architectures where it is possible to have
1744 * holes within pageblock_nr_pages. On systems where it is not possible, this
1745 * function is optimized out.
1746 *
1747 * Then, we check if a current large page is valid by only checking the validity
1748 * of the head pfn.
1749 */
1750 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1751 {
1752 if (!pfn_valid_within(pfn))
1753 return false;
1754 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1755 return false;
1756 return true;
1757 }
1758
1759 /*
1760 * Free pages to buddy allocator. Try to free aligned pages in
1761 * pageblock_nr_pages sizes.
1762 */
1763 static void __init deferred_free_pages(unsigned long pfn,
1764 unsigned long end_pfn)
1765 {
1766 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1767 unsigned long nr_free = 0;
1768
1769 for (; pfn < end_pfn; pfn++) {
1770 if (!deferred_pfn_valid(pfn)) {
1771 deferred_free_range(pfn - nr_free, nr_free);
1772 nr_free = 0;
1773 } else if (!(pfn & nr_pgmask)) {
1774 deferred_free_range(pfn - nr_free, nr_free);
1775 nr_free = 1;
1776 } else {
1777 nr_free++;
1778 }
1779 }
1780 /* Free the last block of pages to allocator */
1781 deferred_free_range(pfn - nr_free, nr_free);
1782 }
1783
1784 /*
1785 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1786 * by performing it only once every pageblock_nr_pages.
1787 * Return number of pages initialized.
1788 */
1789 static unsigned long __init deferred_init_pages(struct zone *zone,
1790 unsigned long pfn,
1791 unsigned long end_pfn)
1792 {
1793 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1794 int nid = zone_to_nid(zone);
1795 unsigned long nr_pages = 0;
1796 int zid = zone_idx(zone);
1797 struct page *page = NULL;
1798
1799 for (; pfn < end_pfn; pfn++) {
1800 if (!deferred_pfn_valid(pfn)) {
1801 page = NULL;
1802 continue;
1803 } else if (!page || !(pfn & nr_pgmask)) {
1804 page = pfn_to_page(pfn);
1805 } else {
1806 page++;
1807 }
1808 __init_single_page(page, pfn, zid, nid);
1809 nr_pages++;
1810 }
1811 return (nr_pages);
1812 }
1813
1814 /*
1815 * This function is meant to pre-load the iterator for the zone init.
1816 * Specifically it walks through the ranges until we are caught up to the
1817 * first_init_pfn value and exits there. If we never encounter the value we
1818 * return false indicating there are no valid ranges left.
1819 */
1820 static bool __init
1821 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1822 unsigned long *spfn, unsigned long *epfn,
1823 unsigned long first_init_pfn)
1824 {
1825 u64 j;
1826
1827 /*
1828 * Start out by walking through the ranges in this zone that have
1829 * already been initialized. We don't need to do anything with them
1830 * so we just need to flush them out of the system.
1831 */
1832 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1833 if (*epfn <= first_init_pfn)
1834 continue;
1835 if (*spfn < first_init_pfn)
1836 *spfn = first_init_pfn;
1837 *i = j;
1838 return true;
1839 }
1840
1841 return false;
1842 }
1843
1844 /*
1845 * Initialize and free pages. We do it in two loops: first we initialize
1846 * struct page, then free to buddy allocator, because while we are
1847 * freeing pages we can access pages that are ahead (computing buddy
1848 * page in __free_one_page()).
1849 *
1850 * In order to try and keep some memory in the cache we have the loop
1851 * broken along max page order boundaries. This way we will not cause
1852 * any issues with the buddy page computation.
1853 */
1854 static unsigned long __init
1855 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1856 unsigned long *end_pfn)
1857 {
1858 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1859 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1860 unsigned long nr_pages = 0;
1861 u64 j = *i;
1862
1863 /* First we loop through and initialize the page values */
1864 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1865 unsigned long t;
1866
1867 if (mo_pfn <= *start_pfn)
1868 break;
1869
1870 t = min(mo_pfn, *end_pfn);
1871 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1872
1873 if (mo_pfn < *end_pfn) {
1874 *start_pfn = mo_pfn;
1875 break;
1876 }
1877 }
1878
1879 /* Reset values and now loop through freeing pages as needed */
1880 swap(j, *i);
1881
1882 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1883 unsigned long t;
1884
1885 if (mo_pfn <= spfn)
1886 break;
1887
1888 t = min(mo_pfn, epfn);
1889 deferred_free_pages(spfn, t);
1890
1891 if (mo_pfn <= epfn)
1892 break;
1893 }
1894
1895 return nr_pages;
1896 }
1897
1898 static void __init
1899 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1900 void *arg)
1901 {
1902 unsigned long spfn, epfn;
1903 struct zone *zone = arg;
1904 u64 i;
1905
1906 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1907
1908 /*
1909 * Initialize and free pages in MAX_ORDER sized increments so that we
1910 * can avoid introducing any issues with the buddy allocator.
1911 */
1912 while (spfn < end_pfn) {
1913 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1914 cond_resched();
1915 }
1916 }
1917
1918 /* An arch may override for more concurrency. */
1919 __weak int __init
1920 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1921 {
1922 return 1;
1923 }
1924
1925 /* Initialise remaining memory on a node */
1926 static int __init deferred_init_memmap(void *data)
1927 {
1928 pg_data_t *pgdat = data;
1929 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1930 unsigned long spfn = 0, epfn = 0;
1931 unsigned long first_init_pfn, flags;
1932 unsigned long start = jiffies;
1933 struct zone *zone;
1934 int zid, max_threads;
1935 u64 i;
1936
1937 /* Bind memory initialisation thread to a local node if possible */
1938 if (!cpumask_empty(cpumask))
1939 set_cpus_allowed_ptr(current, cpumask);
1940
1941 pgdat_resize_lock(pgdat, &flags);
1942 first_init_pfn = pgdat->first_deferred_pfn;
1943 if (first_init_pfn == ULONG_MAX) {
1944 pgdat_resize_unlock(pgdat, &flags);
1945 pgdat_init_report_one_done();
1946 return 0;
1947 }
1948
1949 /* Sanity check boundaries */
1950 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1951 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1952 pgdat->first_deferred_pfn = ULONG_MAX;
1953
1954 /*
1955 * Once we unlock here, the zone cannot be grown anymore, thus if an
1956 * interrupt thread must allocate this early in boot, zone must be
1957 * pre-grown prior to start of deferred page initialization.
1958 */
1959 pgdat_resize_unlock(pgdat, &flags);
1960
1961 /* Only the highest zone is deferred so find it */
1962 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1963 zone = pgdat->node_zones + zid;
1964 if (first_init_pfn < zone_end_pfn(zone))
1965 break;
1966 }
1967
1968 /* If the zone is empty somebody else may have cleared out the zone */
1969 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1970 first_init_pfn))
1971 goto zone_empty;
1972
1973 max_threads = deferred_page_init_max_threads(cpumask);
1974
1975 while (spfn < epfn) {
1976 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1977 struct padata_mt_job job = {
1978 .thread_fn = deferred_init_memmap_chunk,
1979 .fn_arg = zone,
1980 .start = spfn,
1981 .size = epfn_align - spfn,
1982 .align = PAGES_PER_SECTION,
1983 .min_chunk = PAGES_PER_SECTION,
1984 .max_threads = max_threads,
1985 };
1986
1987 padata_do_multithreaded(&job);
1988 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1989 epfn_align);
1990 }
1991 zone_empty:
1992 /* Sanity check that the next zone really is unpopulated */
1993 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1994
1995 pr_info("node %d deferred pages initialised in %ums\n",
1996 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1997
1998 pgdat_init_report_one_done();
1999 return 0;
2000 }
2001
2002 /*
2003 * If this zone has deferred pages, try to grow it by initializing enough
2004 * deferred pages to satisfy the allocation specified by order, rounded up to
2005 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2006 * of SECTION_SIZE bytes by initializing struct pages in increments of
2007 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2008 *
2009 * Return true when zone was grown, otherwise return false. We return true even
2010 * when we grow less than requested, to let the caller decide if there are
2011 * enough pages to satisfy the allocation.
2012 *
2013 * Note: We use noinline because this function is needed only during boot, and
2014 * it is called from a __ref function _deferred_grow_zone. This way we are
2015 * making sure that it is not inlined into permanent text section.
2016 */
2017 static noinline bool __init
2018 deferred_grow_zone(struct zone *zone, unsigned int order)
2019 {
2020 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2021 pg_data_t *pgdat = zone->zone_pgdat;
2022 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2023 unsigned long spfn, epfn, flags;
2024 unsigned long nr_pages = 0;
2025 u64 i;
2026
2027 /* Only the last zone may have deferred pages */
2028 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2029 return false;
2030
2031 pgdat_resize_lock(pgdat, &flags);
2032
2033 /*
2034 * If someone grew this zone while we were waiting for spinlock, return
2035 * true, as there might be enough pages already.
2036 */
2037 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2038 pgdat_resize_unlock(pgdat, &flags);
2039 return true;
2040 }
2041
2042 /* If the zone is empty somebody else may have cleared out the zone */
2043 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2044 first_deferred_pfn)) {
2045 pgdat->first_deferred_pfn = ULONG_MAX;
2046 pgdat_resize_unlock(pgdat, &flags);
2047 /* Retry only once. */
2048 return first_deferred_pfn != ULONG_MAX;
2049 }
2050
2051 /*
2052 * Initialize and free pages in MAX_ORDER sized increments so
2053 * that we can avoid introducing any issues with the buddy
2054 * allocator.
2055 */
2056 while (spfn < epfn) {
2057 /* update our first deferred PFN for this section */
2058 first_deferred_pfn = spfn;
2059
2060 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2061 touch_nmi_watchdog();
2062
2063 /* We should only stop along section boundaries */
2064 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2065 continue;
2066
2067 /* If our quota has been met we can stop here */
2068 if (nr_pages >= nr_pages_needed)
2069 break;
2070 }
2071
2072 pgdat->first_deferred_pfn = spfn;
2073 pgdat_resize_unlock(pgdat, &flags);
2074
2075 return nr_pages > 0;
2076 }
2077
2078 /*
2079 * deferred_grow_zone() is __init, but it is called from
2080 * get_page_from_freelist() during early boot until deferred_pages permanently
2081 * disables this call. This is why we have refdata wrapper to avoid warning,
2082 * and to ensure that the function body gets unloaded.
2083 */
2084 static bool __ref
2085 _deferred_grow_zone(struct zone *zone, unsigned int order)
2086 {
2087 return deferred_grow_zone(zone, order);
2088 }
2089
2090 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2091
2092 void __init page_alloc_init_late(void)
2093 {
2094 struct zone *zone;
2095 int nid;
2096
2097 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2098
2099 /* There will be num_node_state(N_MEMORY) threads */
2100 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2101 for_each_node_state(nid, N_MEMORY) {
2102 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2103 }
2104
2105 /* Block until all are initialised */
2106 wait_for_completion(&pgdat_init_all_done_comp);
2107
2108 /*
2109 * The number of managed pages has changed due to the initialisation
2110 * so the pcpu batch and high limits needs to be updated or the limits
2111 * will be artificially small.
2112 */
2113 for_each_populated_zone(zone)
2114 zone_pcp_update(zone);
2115
2116 /*
2117 * We initialized the rest of the deferred pages. Permanently disable
2118 * on-demand struct page initialization.
2119 */
2120 static_branch_disable(&deferred_pages);
2121
2122 /* Reinit limits that are based on free pages after the kernel is up */
2123 files_maxfiles_init();
2124 #endif
2125
2126 buffer_init();
2127
2128 /* Discard memblock private memory */
2129 memblock_discard();
2130
2131 for_each_node_state(nid, N_MEMORY)
2132 shuffle_free_memory(NODE_DATA(nid));
2133
2134 for_each_populated_zone(zone)
2135 set_zone_contiguous(zone);
2136 }
2137
2138 #ifdef CONFIG_CMA
2139 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2140 void __init init_cma_reserved_pageblock(struct page *page)
2141 {
2142 unsigned i = pageblock_nr_pages;
2143 struct page *p = page;
2144
2145 do {
2146 __ClearPageReserved(p);
2147 set_page_count(p, 0);
2148 } while (++p, --i);
2149
2150 set_pageblock_migratetype(page, MIGRATE_CMA);
2151
2152 if (pageblock_order >= MAX_ORDER) {
2153 i = pageblock_nr_pages;
2154 p = page;
2155 do {
2156 set_page_refcounted(p);
2157 __free_pages(p, MAX_ORDER - 1);
2158 p += MAX_ORDER_NR_PAGES;
2159 } while (i -= MAX_ORDER_NR_PAGES);
2160 } else {
2161 set_page_refcounted(page);
2162 __free_pages(page, pageblock_order);
2163 }
2164
2165 adjust_managed_page_count(page, pageblock_nr_pages);
2166 }
2167 #endif
2168
2169 /*
2170 * The order of subdivision here is critical for the IO subsystem.
2171 * Please do not alter this order without good reasons and regression
2172 * testing. Specifically, as large blocks of memory are subdivided,
2173 * the order in which smaller blocks are delivered depends on the order
2174 * they're subdivided in this function. This is the primary factor
2175 * influencing the order in which pages are delivered to the IO
2176 * subsystem according to empirical testing, and this is also justified
2177 * by considering the behavior of a buddy system containing a single
2178 * large block of memory acted on by a series of small allocations.
2179 * This behavior is a critical factor in sglist merging's success.
2180 *
2181 * -- nyc
2182 */
2183 static inline void expand(struct zone *zone, struct page *page,
2184 int low, int high, int migratetype)
2185 {
2186 unsigned long size = 1 << high;
2187
2188 while (high > low) {
2189 high--;
2190 size >>= 1;
2191 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2192
2193 /*
2194 * Mark as guard pages (or page), that will allow to
2195 * merge back to allocator when buddy will be freed.
2196 * Corresponding page table entries will not be touched,
2197 * pages will stay not present in virtual address space
2198 */
2199 if (set_page_guard(zone, &page[size], high, migratetype))
2200 continue;
2201
2202 add_to_free_list(&page[size], zone, high, migratetype);
2203 set_buddy_order(&page[size], high);
2204 }
2205 }
2206
2207 static void check_new_page_bad(struct page *page)
2208 {
2209 if (unlikely(page->flags & __PG_HWPOISON)) {
2210 /* Don't complain about hwpoisoned pages */
2211 page_mapcount_reset(page); /* remove PageBuddy */
2212 return;
2213 }
2214
2215 bad_page(page,
2216 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2217 }
2218
2219 /*
2220 * This page is about to be returned from the page allocator
2221 */
2222 static inline int check_new_page(struct page *page)
2223 {
2224 if (likely(page_expected_state(page,
2225 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2226 return 0;
2227
2228 check_new_page_bad(page);
2229 return 1;
2230 }
2231
2232 #ifdef CONFIG_DEBUG_VM
2233 /*
2234 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2235 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2236 * also checked when pcp lists are refilled from the free lists.
2237 */
2238 static inline bool check_pcp_refill(struct page *page)
2239 {
2240 if (debug_pagealloc_enabled_static())
2241 return check_new_page(page);
2242 else
2243 return false;
2244 }
2245
2246 static inline bool check_new_pcp(struct page *page)
2247 {
2248 return check_new_page(page);
2249 }
2250 #else
2251 /*
2252 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2253 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2254 * enabled, they are also checked when being allocated from the pcp lists.
2255 */
2256 static inline bool check_pcp_refill(struct page *page)
2257 {
2258 return check_new_page(page);
2259 }
2260 static inline bool check_new_pcp(struct page *page)
2261 {
2262 if (debug_pagealloc_enabled_static())
2263 return check_new_page(page);
2264 else
2265 return false;
2266 }
2267 #endif /* CONFIG_DEBUG_VM */
2268
2269 static bool check_new_pages(struct page *page, unsigned int order)
2270 {
2271 int i;
2272 for (i = 0; i < (1 << order); i++) {
2273 struct page *p = page + i;
2274
2275 if (unlikely(check_new_page(p)))
2276 return true;
2277 }
2278
2279 return false;
2280 }
2281
2282 inline void post_alloc_hook(struct page *page, unsigned int order,
2283 gfp_t gfp_flags)
2284 {
2285 set_page_private(page, 0);
2286 set_page_refcounted(page);
2287
2288 arch_alloc_page(page, order);
2289 debug_pagealloc_map_pages(page, 1 << order);
2290 kasan_alloc_pages(page, order);
2291 kernel_unpoison_pages(page, 1 << order);
2292 set_page_owner(page, order, gfp_flags);
2293
2294 if (!want_init_on_free() && want_init_on_alloc(gfp_flags))
2295 kernel_init_free_pages(page, 1 << order);
2296 }
2297
2298 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2299 unsigned int alloc_flags)
2300 {
2301 post_alloc_hook(page, order, gfp_flags);
2302
2303 if (order && (gfp_flags & __GFP_COMP))
2304 prep_compound_page(page, order);
2305
2306 /*
2307 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2308 * allocate the page. The expectation is that the caller is taking
2309 * steps that will free more memory. The caller should avoid the page
2310 * being used for !PFMEMALLOC purposes.
2311 */
2312 if (alloc_flags & ALLOC_NO_WATERMARKS)
2313 set_page_pfmemalloc(page);
2314 else
2315 clear_page_pfmemalloc(page);
2316 }
2317
2318 /*
2319 * Go through the free lists for the given migratetype and remove
2320 * the smallest available page from the freelists
2321 */
2322 static __always_inline
2323 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2324 int migratetype)
2325 {
2326 unsigned int current_order;
2327 struct free_area *area;
2328 struct page *page;
2329
2330 /* Find a page of the appropriate size in the preferred list */
2331 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2332 area = &(zone->free_area[current_order]);
2333 page = get_page_from_free_area(area, migratetype);
2334 if (!page)
2335 continue;
2336 del_page_from_free_list(page, zone, current_order);
2337 expand(zone, page, order, current_order, migratetype);
2338 set_pcppage_migratetype(page, migratetype);
2339 return page;
2340 }
2341
2342 return NULL;
2343 }
2344
2345
2346 /*
2347 * This array describes the order lists are fallen back to when
2348 * the free lists for the desirable migrate type are depleted
2349 */
2350 static int fallbacks[MIGRATE_TYPES][3] = {
2351 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2352 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2353 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2354 #ifdef CONFIG_CMA
2355 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2356 #endif
2357 #ifdef CONFIG_MEMORY_ISOLATION
2358 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2359 #endif
2360 };
2361
2362 #ifdef CONFIG_CMA
2363 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2364 unsigned int order)
2365 {
2366 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2367 }
2368 #else
2369 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2370 unsigned int order) { return NULL; }
2371 #endif
2372
2373 /*
2374 * Move the free pages in a range to the freelist tail of the requested type.
2375 * Note that start_page and end_pages are not aligned on a pageblock
2376 * boundary. If alignment is required, use move_freepages_block()
2377 */
2378 static int move_freepages(struct zone *zone,
2379 struct page *start_page, struct page *end_page,
2380 int migratetype, int *num_movable)
2381 {
2382 struct page *page;
2383 unsigned int order;
2384 int pages_moved = 0;
2385
2386 for (page = start_page; page <= end_page;) {
2387 if (!pfn_valid_within(page_to_pfn(page))) {
2388 page++;
2389 continue;
2390 }
2391
2392 if (!PageBuddy(page)) {
2393 /*
2394 * We assume that pages that could be isolated for
2395 * migration are movable. But we don't actually try
2396 * isolating, as that would be expensive.
2397 */
2398 if (num_movable &&
2399 (PageLRU(page) || __PageMovable(page)))
2400 (*num_movable)++;
2401
2402 page++;
2403 continue;
2404 }
2405
2406 /* Make sure we are not inadvertently changing nodes */
2407 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2408 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2409
2410 order = buddy_order(page);
2411 move_to_free_list(page, zone, order, migratetype);
2412 page += 1 << order;
2413 pages_moved += 1 << order;
2414 }
2415
2416 return pages_moved;
2417 }
2418
2419 int move_freepages_block(struct zone *zone, struct page *page,
2420 int migratetype, int *num_movable)
2421 {
2422 unsigned long start_pfn, end_pfn;
2423 struct page *start_page, *end_page;
2424
2425 if (num_movable)
2426 *num_movable = 0;
2427
2428 start_pfn = page_to_pfn(page);
2429 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2430 start_page = pfn_to_page(start_pfn);
2431 end_page = start_page + pageblock_nr_pages - 1;
2432 end_pfn = start_pfn + pageblock_nr_pages - 1;
2433
2434 /* Do not cross zone boundaries */
2435 if (!zone_spans_pfn(zone, start_pfn))
2436 start_page = page;
2437 if (!zone_spans_pfn(zone, end_pfn))
2438 return 0;
2439
2440 return move_freepages(zone, start_page, end_page, migratetype,
2441 num_movable);
2442 }
2443
2444 static void change_pageblock_range(struct page *pageblock_page,
2445 int start_order, int migratetype)
2446 {
2447 int nr_pageblocks = 1 << (start_order - pageblock_order);
2448
2449 while (nr_pageblocks--) {
2450 set_pageblock_migratetype(pageblock_page, migratetype);
2451 pageblock_page += pageblock_nr_pages;
2452 }
2453 }
2454
2455 /*
2456 * When we are falling back to another migratetype during allocation, try to
2457 * steal extra free pages from the same pageblocks to satisfy further
2458 * allocations, instead of polluting multiple pageblocks.
2459 *
2460 * If we are stealing a relatively large buddy page, it is likely there will
2461 * be more free pages in the pageblock, so try to steal them all. For
2462 * reclaimable and unmovable allocations, we steal regardless of page size,
2463 * as fragmentation caused by those allocations polluting movable pageblocks
2464 * is worse than movable allocations stealing from unmovable and reclaimable
2465 * pageblocks.
2466 */
2467 static bool can_steal_fallback(unsigned int order, int start_mt)
2468 {
2469 /*
2470 * Leaving this order check is intended, although there is
2471 * relaxed order check in next check. The reason is that
2472 * we can actually steal whole pageblock if this condition met,
2473 * but, below check doesn't guarantee it and that is just heuristic
2474 * so could be changed anytime.
2475 */
2476 if (order >= pageblock_order)
2477 return true;
2478
2479 if (order >= pageblock_order / 2 ||
2480 start_mt == MIGRATE_RECLAIMABLE ||
2481 start_mt == MIGRATE_UNMOVABLE ||
2482 page_group_by_mobility_disabled)
2483 return true;
2484
2485 return false;
2486 }
2487
2488 static inline bool boost_watermark(struct zone *zone)
2489 {
2490 unsigned long max_boost;
2491
2492 if (!watermark_boost_factor)
2493 return false;
2494 /*
2495 * Don't bother in zones that are unlikely to produce results.
2496 * On small machines, including kdump capture kernels running
2497 * in a small area, boosting the watermark can cause an out of
2498 * memory situation immediately.
2499 */
2500 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2501 return false;
2502
2503 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2504 watermark_boost_factor, 10000);
2505
2506 /*
2507 * high watermark may be uninitialised if fragmentation occurs
2508 * very early in boot so do not boost. We do not fall
2509 * through and boost by pageblock_nr_pages as failing
2510 * allocations that early means that reclaim is not going
2511 * to help and it may even be impossible to reclaim the
2512 * boosted watermark resulting in a hang.
2513 */
2514 if (!max_boost)
2515 return false;
2516
2517 max_boost = max(pageblock_nr_pages, max_boost);
2518
2519 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2520 max_boost);
2521
2522 return true;
2523 }
2524
2525 /*
2526 * This function implements actual steal behaviour. If order is large enough,
2527 * we can steal whole pageblock. If not, we first move freepages in this
2528 * pageblock to our migratetype and determine how many already-allocated pages
2529 * are there in the pageblock with a compatible migratetype. If at least half
2530 * of pages are free or compatible, we can change migratetype of the pageblock
2531 * itself, so pages freed in the future will be put on the correct free list.
2532 */
2533 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2534 unsigned int alloc_flags, int start_type, bool whole_block)
2535 {
2536 unsigned int current_order = buddy_order(page);
2537 int free_pages, movable_pages, alike_pages;
2538 int old_block_type;
2539
2540 old_block_type = get_pageblock_migratetype(page);
2541
2542 /*
2543 * This can happen due to races and we want to prevent broken
2544 * highatomic accounting.
2545 */
2546 if (is_migrate_highatomic(old_block_type))
2547 goto single_page;
2548
2549 /* Take ownership for orders >= pageblock_order */
2550 if (current_order >= pageblock_order) {
2551 change_pageblock_range(page, current_order, start_type);
2552 goto single_page;
2553 }
2554
2555 /*
2556 * Boost watermarks to increase reclaim pressure to reduce the
2557 * likelihood of future fallbacks. Wake kswapd now as the node
2558 * may be balanced overall and kswapd will not wake naturally.
2559 */
2560 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2561 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2562
2563 /* We are not allowed to try stealing from the whole block */
2564 if (!whole_block)
2565 goto single_page;
2566
2567 free_pages = move_freepages_block(zone, page, start_type,
2568 &movable_pages);
2569 /*
2570 * Determine how many pages are compatible with our allocation.
2571 * For movable allocation, it's the number of movable pages which
2572 * we just obtained. For other types it's a bit more tricky.
2573 */
2574 if (start_type == MIGRATE_MOVABLE) {
2575 alike_pages = movable_pages;
2576 } else {
2577 /*
2578 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2579 * to MOVABLE pageblock, consider all non-movable pages as
2580 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2581 * vice versa, be conservative since we can't distinguish the
2582 * exact migratetype of non-movable pages.
2583 */
2584 if (old_block_type == MIGRATE_MOVABLE)
2585 alike_pages = pageblock_nr_pages
2586 - (free_pages + movable_pages);
2587 else
2588 alike_pages = 0;
2589 }
2590
2591 /* moving whole block can fail due to zone boundary conditions */
2592 if (!free_pages)
2593 goto single_page;
2594
2595 /*
2596 * If a sufficient number of pages in the block are either free or of
2597 * comparable migratability as our allocation, claim the whole block.
2598 */
2599 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2600 page_group_by_mobility_disabled)
2601 set_pageblock_migratetype(page, start_type);
2602
2603 return;
2604
2605 single_page:
2606 move_to_free_list(page, zone, current_order, start_type);
2607 }
2608
2609 /*
2610 * Check whether there is a suitable fallback freepage with requested order.
2611 * If only_stealable is true, this function returns fallback_mt only if
2612 * we can steal other freepages all together. This would help to reduce
2613 * fragmentation due to mixed migratetype pages in one pageblock.
2614 */
2615 int find_suitable_fallback(struct free_area *area, unsigned int order,
2616 int migratetype, bool only_stealable, bool *can_steal)
2617 {
2618 int i;
2619 int fallback_mt;
2620
2621 if (area->nr_free == 0)
2622 return -1;
2623
2624 *can_steal = false;
2625 for (i = 0;; i++) {
2626 fallback_mt = fallbacks[migratetype][i];
2627 if (fallback_mt == MIGRATE_TYPES)
2628 break;
2629
2630 if (free_area_empty(area, fallback_mt))
2631 continue;
2632
2633 if (can_steal_fallback(order, migratetype))
2634 *can_steal = true;
2635
2636 if (!only_stealable)
2637 return fallback_mt;
2638
2639 if (*can_steal)
2640 return fallback_mt;
2641 }
2642
2643 return -1;
2644 }
2645
2646 /*
2647 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2648 * there are no empty page blocks that contain a page with a suitable order
2649 */
2650 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2651 unsigned int alloc_order)
2652 {
2653 int mt;
2654 unsigned long max_managed, flags;
2655
2656 /*
2657 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2658 * Check is race-prone but harmless.
2659 */
2660 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2661 if (zone->nr_reserved_highatomic >= max_managed)
2662 return;
2663
2664 spin_lock_irqsave(&zone->lock, flags);
2665
2666 /* Recheck the nr_reserved_highatomic limit under the lock */
2667 if (zone->nr_reserved_highatomic >= max_managed)
2668 goto out_unlock;
2669
2670 /* Yoink! */
2671 mt = get_pageblock_migratetype(page);
2672 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2673 && !is_migrate_cma(mt)) {
2674 zone->nr_reserved_highatomic += pageblock_nr_pages;
2675 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2676 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2677 }
2678
2679 out_unlock:
2680 spin_unlock_irqrestore(&zone->lock, flags);
2681 }
2682
2683 /*
2684 * Used when an allocation is about to fail under memory pressure. This
2685 * potentially hurts the reliability of high-order allocations when under
2686 * intense memory pressure but failed atomic allocations should be easier
2687 * to recover from than an OOM.
2688 *
2689 * If @force is true, try to unreserve a pageblock even though highatomic
2690 * pageblock is exhausted.
2691 */
2692 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2693 bool force)
2694 {
2695 struct zonelist *zonelist = ac->zonelist;
2696 unsigned long flags;
2697 struct zoneref *z;
2698 struct zone *zone;
2699 struct page *page;
2700 int order;
2701 bool ret;
2702
2703 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2704 ac->nodemask) {
2705 /*
2706 * Preserve at least one pageblock unless memory pressure
2707 * is really high.
2708 */
2709 if (!force && zone->nr_reserved_highatomic <=
2710 pageblock_nr_pages)
2711 continue;
2712
2713 spin_lock_irqsave(&zone->lock, flags);
2714 for (order = 0; order < MAX_ORDER; order++) {
2715 struct free_area *area = &(zone->free_area[order]);
2716
2717 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2718 if (!page)
2719 continue;
2720
2721 /*
2722 * In page freeing path, migratetype change is racy so
2723 * we can counter several free pages in a pageblock
2724 * in this loop althoug we changed the pageblock type
2725 * from highatomic to ac->migratetype. So we should
2726 * adjust the count once.
2727 */
2728 if (is_migrate_highatomic_page(page)) {
2729 /*
2730 * It should never happen but changes to
2731 * locking could inadvertently allow a per-cpu
2732 * drain to add pages to MIGRATE_HIGHATOMIC
2733 * while unreserving so be safe and watch for
2734 * underflows.
2735 */
2736 zone->nr_reserved_highatomic -= min(
2737 pageblock_nr_pages,
2738 zone->nr_reserved_highatomic);
2739 }
2740
2741 /*
2742 * Convert to ac->migratetype and avoid the normal
2743 * pageblock stealing heuristics. Minimally, the caller
2744 * is doing the work and needs the pages. More
2745 * importantly, if the block was always converted to
2746 * MIGRATE_UNMOVABLE or another type then the number
2747 * of pageblocks that cannot be completely freed
2748 * may increase.
2749 */
2750 set_pageblock_migratetype(page, ac->migratetype);
2751 ret = move_freepages_block(zone, page, ac->migratetype,
2752 NULL);
2753 if (ret) {
2754 spin_unlock_irqrestore(&zone->lock, flags);
2755 return ret;
2756 }
2757 }
2758 spin_unlock_irqrestore(&zone->lock, flags);
2759 }
2760
2761 return false;
2762 }
2763
2764 /*
2765 * Try finding a free buddy page on the fallback list and put it on the free
2766 * list of requested migratetype, possibly along with other pages from the same
2767 * block, depending on fragmentation avoidance heuristics. Returns true if
2768 * fallback was found so that __rmqueue_smallest() can grab it.
2769 *
2770 * The use of signed ints for order and current_order is a deliberate
2771 * deviation from the rest of this file, to make the for loop
2772 * condition simpler.
2773 */
2774 static __always_inline bool
2775 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2776 unsigned int alloc_flags)
2777 {
2778 struct free_area *area;
2779 int current_order;
2780 int min_order = order;
2781 struct page *page;
2782 int fallback_mt;
2783 bool can_steal;
2784
2785 /*
2786 * Do not steal pages from freelists belonging to other pageblocks
2787 * i.e. orders < pageblock_order. If there are no local zones free,
2788 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2789 */
2790 if (alloc_flags & ALLOC_NOFRAGMENT)
2791 min_order = pageblock_order;
2792
2793 /*
2794 * Find the largest available free page in the other list. This roughly
2795 * approximates finding the pageblock with the most free pages, which
2796 * would be too costly to do exactly.
2797 */
2798 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2799 --current_order) {
2800 area = &(zone->free_area[current_order]);
2801 fallback_mt = find_suitable_fallback(area, current_order,
2802 start_migratetype, false, &can_steal);
2803 if (fallback_mt == -1)
2804 continue;
2805
2806 /*
2807 * We cannot steal all free pages from the pageblock and the
2808 * requested migratetype is movable. In that case it's better to
2809 * steal and split the smallest available page instead of the
2810 * largest available page, because even if the next movable
2811 * allocation falls back into a different pageblock than this
2812 * one, it won't cause permanent fragmentation.
2813 */
2814 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2815 && current_order > order)
2816 goto find_smallest;
2817
2818 goto do_steal;
2819 }
2820
2821 return false;
2822
2823 find_smallest:
2824 for (current_order = order; current_order < MAX_ORDER;
2825 current_order++) {
2826 area = &(zone->free_area[current_order]);
2827 fallback_mt = find_suitable_fallback(area, current_order,
2828 start_migratetype, false, &can_steal);
2829 if (fallback_mt != -1)
2830 break;
2831 }
2832
2833 /*
2834 * This should not happen - we already found a suitable fallback
2835 * when looking for the largest page.
2836 */
2837 VM_BUG_ON(current_order == MAX_ORDER);
2838
2839 do_steal:
2840 page = get_page_from_free_area(area, fallback_mt);
2841
2842 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2843 can_steal);
2844
2845 trace_mm_page_alloc_extfrag(page, order, current_order,
2846 start_migratetype, fallback_mt);
2847
2848 return true;
2849
2850 }
2851
2852 /*
2853 * Do the hard work of removing an element from the buddy allocator.
2854 * Call me with the zone->lock already held.
2855 */
2856 static __always_inline struct page *
2857 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2858 unsigned int alloc_flags)
2859 {
2860 struct page *page;
2861
2862 if (IS_ENABLED(CONFIG_CMA)) {
2863 /*
2864 * Balance movable allocations between regular and CMA areas by
2865 * allocating from CMA when over half of the zone's free memory
2866 * is in the CMA area.
2867 */
2868 if (alloc_flags & ALLOC_CMA &&
2869 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2870 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2871 page = __rmqueue_cma_fallback(zone, order);
2872 if (page)
2873 goto out;
2874 }
2875 }
2876 retry:
2877 page = __rmqueue_smallest(zone, order, migratetype);
2878 if (unlikely(!page)) {
2879 if (alloc_flags & ALLOC_CMA)
2880 page = __rmqueue_cma_fallback(zone, order);
2881
2882 if (!page && __rmqueue_fallback(zone, order, migratetype,
2883 alloc_flags))
2884 goto retry;
2885 }
2886 out:
2887 if (page)
2888 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2889 return page;
2890 }
2891
2892 /*
2893 * Obtain a specified number of elements from the buddy allocator, all under
2894 * a single hold of the lock, for efficiency. Add them to the supplied list.
2895 * Returns the number of new pages which were placed at *list.
2896 */
2897 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2898 unsigned long count, struct list_head *list,
2899 int migratetype, unsigned int alloc_flags)
2900 {
2901 int i, alloced = 0;
2902
2903 spin_lock(&zone->lock);
2904 for (i = 0; i < count; ++i) {
2905 struct page *page = __rmqueue(zone, order, migratetype,
2906 alloc_flags);
2907 if (unlikely(page == NULL))
2908 break;
2909
2910 if (unlikely(check_pcp_refill(page)))
2911 continue;
2912
2913 /*
2914 * Split buddy pages returned by expand() are received here in
2915 * physical page order. The page is added to the tail of
2916 * caller's list. From the callers perspective, the linked list
2917 * is ordered by page number under some conditions. This is
2918 * useful for IO devices that can forward direction from the
2919 * head, thus also in the physical page order. This is useful
2920 * for IO devices that can merge IO requests if the physical
2921 * pages are ordered properly.
2922 */
2923 list_add_tail(&page->lru, list);
2924 alloced++;
2925 if (is_migrate_cma(get_pcppage_migratetype(page)))
2926 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2927 -(1 << order));
2928 }
2929
2930 /*
2931 * i pages were removed from the buddy list even if some leak due
2932 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2933 * on i. Do not confuse with 'alloced' which is the number of
2934 * pages added to the pcp list.
2935 */
2936 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2937 spin_unlock(&zone->lock);
2938 return alloced;
2939 }
2940
2941 #ifdef CONFIG_NUMA
2942 /*
2943 * Called from the vmstat counter updater to drain pagesets of this
2944 * currently executing processor on remote nodes after they have
2945 * expired.
2946 *
2947 * Note that this function must be called with the thread pinned to
2948 * a single processor.
2949 */
2950 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2951 {
2952 unsigned long flags;
2953 int to_drain, batch;
2954
2955 local_irq_save(flags);
2956 batch = READ_ONCE(pcp->batch);
2957 to_drain = min(pcp->count, batch);
2958 if (to_drain > 0)
2959 free_pcppages_bulk(zone, to_drain, pcp);
2960 local_irq_restore(flags);
2961 }
2962 #endif
2963
2964 /*
2965 * Drain pcplists of the indicated processor and zone.
2966 *
2967 * The processor must either be the current processor and the
2968 * thread pinned to the current processor or a processor that
2969 * is not online.
2970 */
2971 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2972 {
2973 unsigned long flags;
2974 struct per_cpu_pageset *pset;
2975 struct per_cpu_pages *pcp;
2976
2977 local_irq_save(flags);
2978 pset = per_cpu_ptr(zone->pageset, cpu);
2979
2980 pcp = &pset->pcp;
2981 if (pcp->count)
2982 free_pcppages_bulk(zone, pcp->count, pcp);
2983 local_irq_restore(flags);
2984 }
2985
2986 /*
2987 * Drain pcplists of all zones on the indicated processor.
2988 *
2989 * The processor must either be the current processor and the
2990 * thread pinned to the current processor or a processor that
2991 * is not online.
2992 */
2993 static void drain_pages(unsigned int cpu)
2994 {
2995 struct zone *zone;
2996
2997 for_each_populated_zone(zone) {
2998 drain_pages_zone(cpu, zone);
2999 }
3000 }
3001
3002 /*
3003 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3004 *
3005 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3006 * the single zone's pages.
3007 */
3008 void drain_local_pages(struct zone *zone)
3009 {
3010 int cpu = smp_processor_id();
3011
3012 if (zone)
3013 drain_pages_zone(cpu, zone);
3014 else
3015 drain_pages(cpu);
3016 }
3017
3018 static void drain_local_pages_wq(struct work_struct *work)
3019 {
3020 struct pcpu_drain *drain;
3021
3022 drain = container_of(work, struct pcpu_drain, work);
3023
3024 /*
3025 * drain_all_pages doesn't use proper cpu hotplug protection so
3026 * we can race with cpu offline when the WQ can move this from
3027 * a cpu pinned worker to an unbound one. We can operate on a different
3028 * cpu which is allright but we also have to make sure to not move to
3029 * a different one.
3030 */
3031 preempt_disable();
3032 drain_local_pages(drain->zone);
3033 preempt_enable();
3034 }
3035
3036 /*
3037 * The implementation of drain_all_pages(), exposing an extra parameter to
3038 * drain on all cpus.
3039 *
3040 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3041 * not empty. The check for non-emptiness can however race with a free to
3042 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3043 * that need the guarantee that every CPU has drained can disable the
3044 * optimizing racy check.
3045 */
3046 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3047 {
3048 int cpu;
3049
3050 /*
3051 * Allocate in the BSS so we wont require allocation in
3052 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3053 */
3054 static cpumask_t cpus_with_pcps;
3055
3056 /*
3057 * Make sure nobody triggers this path before mm_percpu_wq is fully
3058 * initialized.
3059 */
3060 if (WARN_ON_ONCE(!mm_percpu_wq))
3061 return;
3062
3063 /*
3064 * Do not drain if one is already in progress unless it's specific to
3065 * a zone. Such callers are primarily CMA and memory hotplug and need
3066 * the drain to be complete when the call returns.
3067 */
3068 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3069 if (!zone)
3070 return;
3071 mutex_lock(&pcpu_drain_mutex);
3072 }
3073
3074 /*
3075 * We don't care about racing with CPU hotplug event
3076 * as offline notification will cause the notified
3077 * cpu to drain that CPU pcps and on_each_cpu_mask
3078 * disables preemption as part of its processing
3079 */
3080 for_each_online_cpu(cpu) {
3081 struct per_cpu_pageset *pcp;
3082 struct zone *z;
3083 bool has_pcps = false;
3084
3085 if (force_all_cpus) {
3086 /*
3087 * The pcp.count check is racy, some callers need a
3088 * guarantee that no cpu is missed.
3089 */
3090 has_pcps = true;
3091 } else if (zone) {
3092 pcp = per_cpu_ptr(zone->pageset, cpu);
3093 if (pcp->pcp.count)
3094 has_pcps = true;
3095 } else {
3096 for_each_populated_zone(z) {
3097 pcp = per_cpu_ptr(z->pageset, cpu);
3098 if (pcp->pcp.count) {
3099 has_pcps = true;
3100 break;
3101 }
3102 }
3103 }
3104
3105 if (has_pcps)
3106 cpumask_set_cpu(cpu, &cpus_with_pcps);
3107 else
3108 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3109 }
3110
3111 for_each_cpu(cpu, &cpus_with_pcps) {
3112 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3113
3114 drain->zone = zone;
3115 INIT_WORK(&drain->work, drain_local_pages_wq);
3116 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3117 }
3118 for_each_cpu(cpu, &cpus_with_pcps)
3119 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3120
3121 mutex_unlock(&pcpu_drain_mutex);
3122 }
3123
3124 /*
3125 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3126 *
3127 * When zone parameter is non-NULL, spill just the single zone's pages.
3128 *
3129 * Note that this can be extremely slow as the draining happens in a workqueue.
3130 */
3131 void drain_all_pages(struct zone *zone)
3132 {
3133 __drain_all_pages(zone, false);
3134 }
3135
3136 #ifdef CONFIG_HIBERNATION
3137
3138 /*
3139 * Touch the watchdog for every WD_PAGE_COUNT pages.
3140 */
3141 #define WD_PAGE_COUNT (128*1024)
3142
3143 void mark_free_pages(struct zone *zone)
3144 {
3145 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3146 unsigned long flags;
3147 unsigned int order, t;
3148 struct page *page;
3149
3150 if (zone_is_empty(zone))
3151 return;
3152
3153 spin_lock_irqsave(&zone->lock, flags);
3154
3155 max_zone_pfn = zone_end_pfn(zone);
3156 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3157 if (pfn_valid(pfn)) {
3158 page = pfn_to_page(pfn);
3159
3160 if (!--page_count) {
3161 touch_nmi_watchdog();
3162 page_count = WD_PAGE_COUNT;
3163 }
3164
3165 if (page_zone(page) != zone)
3166 continue;
3167
3168 if (!swsusp_page_is_forbidden(page))
3169 swsusp_unset_page_free(page);
3170 }
3171
3172 for_each_migratetype_order(order, t) {
3173 list_for_each_entry(page,
3174 &zone->free_area[order].free_list[t], lru) {
3175 unsigned long i;
3176
3177 pfn = page_to_pfn(page);
3178 for (i = 0; i < (1UL << order); i++) {
3179 if (!--page_count) {
3180 touch_nmi_watchdog();
3181 page_count = WD_PAGE_COUNT;
3182 }
3183 swsusp_set_page_free(pfn_to_page(pfn + i));
3184 }
3185 }
3186 }
3187 spin_unlock_irqrestore(&zone->lock, flags);
3188 }
3189 #endif /* CONFIG_PM */
3190
3191 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3192 {
3193 int migratetype;
3194
3195 if (!free_pcp_prepare(page))
3196 return false;
3197
3198 migratetype = get_pfnblock_migratetype(page, pfn);
3199 set_pcppage_migratetype(page, migratetype);
3200 return true;
3201 }
3202
3203 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3204 {
3205 struct zone *zone = page_zone(page);
3206 struct per_cpu_pages *pcp;
3207 int migratetype;
3208
3209 migratetype = get_pcppage_migratetype(page);
3210 __count_vm_event(PGFREE);
3211
3212 /*
3213 * We only track unmovable, reclaimable and movable on pcp lists.
3214 * Free ISOLATE pages back to the allocator because they are being
3215 * offlined but treat HIGHATOMIC as movable pages so we can get those
3216 * areas back if necessary. Otherwise, we may have to free
3217 * excessively into the page allocator
3218 */
3219 if (migratetype >= MIGRATE_PCPTYPES) {
3220 if (unlikely(is_migrate_isolate(migratetype))) {
3221 free_one_page(zone, page, pfn, 0, migratetype,
3222 FPI_NONE);
3223 return;
3224 }
3225 migratetype = MIGRATE_MOVABLE;
3226 }
3227
3228 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3229 list_add(&page->lru, &pcp->lists[migratetype]);
3230 pcp->count++;
3231 if (pcp->count >= READ_ONCE(pcp->high))
3232 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3233 }
3234
3235 /*
3236 * Free a 0-order page
3237 */
3238 void free_unref_page(struct page *page)
3239 {
3240 unsigned long flags;
3241 unsigned long pfn = page_to_pfn(page);
3242
3243 if (!free_unref_page_prepare(page, pfn))
3244 return;
3245
3246 local_irq_save(flags);
3247 free_unref_page_commit(page, pfn);
3248 local_irq_restore(flags);
3249 }
3250
3251 /*
3252 * Free a list of 0-order pages
3253 */
3254 void free_unref_page_list(struct list_head *list)
3255 {
3256 struct page *page, *next;
3257 unsigned long flags, pfn;
3258 int batch_count = 0;
3259
3260 /* Prepare pages for freeing */
3261 list_for_each_entry_safe(page, next, list, lru) {
3262 pfn = page_to_pfn(page);
3263 if (!free_unref_page_prepare(page, pfn))
3264 list_del(&page->lru);
3265 set_page_private(page, pfn);
3266 }
3267
3268 local_irq_save(flags);
3269 list_for_each_entry_safe(page, next, list, lru) {
3270 unsigned long pfn = page_private(page);
3271
3272 set_page_private(page, 0);
3273 trace_mm_page_free_batched(page);
3274 free_unref_page_commit(page, pfn);
3275
3276 /*
3277 * Guard against excessive IRQ disabled times when we get
3278 * a large list of pages to free.
3279 */
3280 if (++batch_count == SWAP_CLUSTER_MAX) {
3281 local_irq_restore(flags);
3282 batch_count = 0;
3283 local_irq_save(flags);
3284 }
3285 }
3286 local_irq_restore(flags);
3287 }
3288
3289 /*
3290 * split_page takes a non-compound higher-order page, and splits it into
3291 * n (1<<order) sub-pages: page[0..n]
3292 * Each sub-page must be freed individually.
3293 *
3294 * Note: this is probably too low level an operation for use in drivers.
3295 * Please consult with lkml before using this in your driver.
3296 */
3297 void split_page(struct page *page, unsigned int order)
3298 {
3299 int i;
3300
3301 VM_BUG_ON_PAGE(PageCompound(page), page);
3302 VM_BUG_ON_PAGE(!page_count(page), page);
3303
3304 for (i = 1; i < (1 << order); i++)
3305 set_page_refcounted(page + i);
3306 split_page_owner(page, 1 << order);
3307 split_page_memcg(page, 1 << order);
3308 }
3309 EXPORT_SYMBOL_GPL(split_page);
3310
3311 int __isolate_free_page(struct page *page, unsigned int order)
3312 {
3313 unsigned long watermark;
3314 struct zone *zone;
3315 int mt;
3316
3317 BUG_ON(!PageBuddy(page));
3318
3319 zone = page_zone(page);
3320 mt = get_pageblock_migratetype(page);
3321
3322 if (!is_migrate_isolate(mt)) {
3323 /*
3324 * Obey watermarks as if the page was being allocated. We can
3325 * emulate a high-order watermark check with a raised order-0
3326 * watermark, because we already know our high-order page
3327 * exists.
3328 */
3329 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3330 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3331 return 0;
3332
3333 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3334 }
3335
3336 /* Remove page from free list */
3337
3338 del_page_from_free_list(page, zone, order);
3339
3340 /*
3341 * Set the pageblock if the isolated page is at least half of a
3342 * pageblock
3343 */
3344 if (order >= pageblock_order - 1) {
3345 struct page *endpage = page + (1 << order) - 1;
3346 for (; page < endpage; page += pageblock_nr_pages) {
3347 int mt = get_pageblock_migratetype(page);
3348 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3349 && !is_migrate_highatomic(mt))
3350 set_pageblock_migratetype(page,
3351 MIGRATE_MOVABLE);
3352 }
3353 }
3354
3355
3356 return 1UL << order;
3357 }
3358
3359 /**
3360 * __putback_isolated_page - Return a now-isolated page back where we got it
3361 * @page: Page that was isolated
3362 * @order: Order of the isolated page
3363 * @mt: The page's pageblock's migratetype
3364 *
3365 * This function is meant to return a page pulled from the free lists via
3366 * __isolate_free_page back to the free lists they were pulled from.
3367 */
3368 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3369 {
3370 struct zone *zone = page_zone(page);
3371
3372 /* zone lock should be held when this function is called */
3373 lockdep_assert_held(&zone->lock);
3374
3375 /* Return isolated page to tail of freelist. */
3376 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3377 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3378 }
3379
3380 /*
3381 * Update NUMA hit/miss statistics
3382 *
3383 * Must be called with interrupts disabled.
3384 */
3385 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3386 {
3387 #ifdef CONFIG_NUMA
3388 enum numa_stat_item local_stat = NUMA_LOCAL;
3389
3390 /* skip numa counters update if numa stats is disabled */
3391 if (!static_branch_likely(&vm_numa_stat_key))
3392 return;
3393
3394 if (zone_to_nid(z) != numa_node_id())
3395 local_stat = NUMA_OTHER;
3396
3397 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3398 __inc_numa_state(z, NUMA_HIT);
3399 else {
3400 __inc_numa_state(z, NUMA_MISS);
3401 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3402 }
3403 __inc_numa_state(z, local_stat);
3404 #endif
3405 }
3406
3407 /* Remove page from the per-cpu list, caller must protect the list */
3408 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3409 unsigned int alloc_flags,
3410 struct per_cpu_pages *pcp,
3411 struct list_head *list)
3412 {
3413 struct page *page;
3414
3415 do {
3416 if (list_empty(list)) {
3417 pcp->count += rmqueue_bulk(zone, 0,
3418 READ_ONCE(pcp->batch), list,
3419 migratetype, alloc_flags);
3420 if (unlikely(list_empty(list)))
3421 return NULL;
3422 }
3423
3424 page = list_first_entry(list, struct page, lru);
3425 list_del(&page->lru);
3426 pcp->count--;
3427 } while (check_new_pcp(page));
3428
3429 return page;
3430 }
3431
3432 /* Lock and remove page from the per-cpu list */
3433 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3434 struct zone *zone, gfp_t gfp_flags,
3435 int migratetype, unsigned int alloc_flags)
3436 {
3437 struct per_cpu_pages *pcp;
3438 struct list_head *list;
3439 struct page *page;
3440 unsigned long flags;
3441
3442 local_irq_save(flags);
3443 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3444 list = &pcp->lists[migratetype];
3445 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3446 if (page) {
3447 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3448 zone_statistics(preferred_zone, zone);
3449 }
3450 local_irq_restore(flags);
3451 return page;
3452 }
3453
3454 /*
3455 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3456 */
3457 static inline
3458 struct page *rmqueue(struct zone *preferred_zone,
3459 struct zone *zone, unsigned int order,
3460 gfp_t gfp_flags, unsigned int alloc_flags,
3461 int migratetype)
3462 {
3463 unsigned long flags;
3464 struct page *page;
3465
3466 if (likely(order == 0)) {
3467 /*
3468 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3469 * we need to skip it when CMA area isn't allowed.
3470 */
3471 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3472 migratetype != MIGRATE_MOVABLE) {
3473 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3474 migratetype, alloc_flags);
3475 goto out;
3476 }
3477 }
3478
3479 /*
3480 * We most definitely don't want callers attempting to
3481 * allocate greater than order-1 page units with __GFP_NOFAIL.
3482 */
3483 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3484 spin_lock_irqsave(&zone->lock, flags);
3485
3486 do {
3487 page = NULL;
3488 /*
3489 * order-0 request can reach here when the pcplist is skipped
3490 * due to non-CMA allocation context. HIGHATOMIC area is
3491 * reserved for high-order atomic allocation, so order-0
3492 * request should skip it.
3493 */
3494 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3495 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3496 if (page)
3497 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3498 }
3499 if (!page)
3500 page = __rmqueue(zone, order, migratetype, alloc_flags);
3501 } while (page && check_new_pages(page, order));
3502 spin_unlock(&zone->lock);
3503 if (!page)
3504 goto failed;
3505 __mod_zone_freepage_state(zone, -(1 << order),
3506 get_pcppage_migratetype(page));
3507
3508 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3509 zone_statistics(preferred_zone, zone);
3510 local_irq_restore(flags);
3511
3512 out:
3513 /* Separate test+clear to avoid unnecessary atomics */
3514 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3515 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3516 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3517 }
3518
3519 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3520 return page;
3521
3522 failed:
3523 local_irq_restore(flags);
3524 return NULL;
3525 }
3526
3527 #ifdef CONFIG_FAIL_PAGE_ALLOC
3528
3529 static struct {
3530 struct fault_attr attr;
3531
3532 bool ignore_gfp_highmem;
3533 bool ignore_gfp_reclaim;
3534 u32 min_order;
3535 } fail_page_alloc = {
3536 .attr = FAULT_ATTR_INITIALIZER,
3537 .ignore_gfp_reclaim = true,
3538 .ignore_gfp_highmem = true,
3539 .min_order = 1,
3540 };
3541
3542 static int __init setup_fail_page_alloc(char *str)
3543 {
3544 return setup_fault_attr(&fail_page_alloc.attr, str);
3545 }
3546 __setup("fail_page_alloc=", setup_fail_page_alloc);
3547
3548 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3549 {
3550 if (order < fail_page_alloc.min_order)
3551 return false;
3552 if (gfp_mask & __GFP_NOFAIL)
3553 return false;
3554 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3555 return false;
3556 if (fail_page_alloc.ignore_gfp_reclaim &&
3557 (gfp_mask & __GFP_DIRECT_RECLAIM))
3558 return false;
3559
3560 return should_fail(&fail_page_alloc.attr, 1 << order);
3561 }
3562
3563 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3564
3565 static int __init fail_page_alloc_debugfs(void)
3566 {
3567 umode_t mode = S_IFREG | 0600;
3568 struct dentry *dir;
3569
3570 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3571 &fail_page_alloc.attr);
3572
3573 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3574 &fail_page_alloc.ignore_gfp_reclaim);
3575 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3576 &fail_page_alloc.ignore_gfp_highmem);
3577 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3578
3579 return 0;
3580 }
3581
3582 late_initcall(fail_page_alloc_debugfs);
3583
3584 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3585
3586 #else /* CONFIG_FAIL_PAGE_ALLOC */
3587
3588 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3589 {
3590 return false;
3591 }
3592
3593 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3594
3595 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3596 {
3597 return __should_fail_alloc_page(gfp_mask, order);
3598 }
3599 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3600
3601 static inline long __zone_watermark_unusable_free(struct zone *z,
3602 unsigned int order, unsigned int alloc_flags)
3603 {
3604 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3605 long unusable_free = (1 << order) - 1;
3606
3607 /*
3608 * If the caller does not have rights to ALLOC_HARDER then subtract
3609 * the high-atomic reserves. This will over-estimate the size of the
3610 * atomic reserve but it avoids a search.
3611 */
3612 if (likely(!alloc_harder))
3613 unusable_free += z->nr_reserved_highatomic;
3614
3615 #ifdef CONFIG_CMA
3616 /* If allocation can't use CMA areas don't use free CMA pages */
3617 if (!(alloc_flags & ALLOC_CMA))
3618 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3619 #endif
3620
3621 return unusable_free;
3622 }
3623
3624 /*
3625 * Return true if free base pages are above 'mark'. For high-order checks it
3626 * will return true of the order-0 watermark is reached and there is at least
3627 * one free page of a suitable size. Checking now avoids taking the zone lock
3628 * to check in the allocation paths if no pages are free.
3629 */
3630 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3631 int highest_zoneidx, unsigned int alloc_flags,
3632 long free_pages)
3633 {
3634 long min = mark;
3635 int o;
3636 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3637
3638 /* free_pages may go negative - that's OK */
3639 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3640
3641 if (alloc_flags & ALLOC_HIGH)
3642 min -= min / 2;
3643
3644 if (unlikely(alloc_harder)) {
3645 /*
3646 * OOM victims can try even harder than normal ALLOC_HARDER
3647 * users on the grounds that it's definitely going to be in
3648 * the exit path shortly and free memory. Any allocation it
3649 * makes during the free path will be small and short-lived.
3650 */
3651 if (alloc_flags & ALLOC_OOM)
3652 min -= min / 2;
3653 else
3654 min -= min / 4;
3655 }
3656
3657 /*
3658 * Check watermarks for an order-0 allocation request. If these
3659 * are not met, then a high-order request also cannot go ahead
3660 * even if a suitable page happened to be free.
3661 */
3662 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3663 return false;
3664
3665 /* If this is an order-0 request then the watermark is fine */
3666 if (!order)
3667 return true;
3668
3669 /* For a high-order request, check at least one suitable page is free */
3670 for (o = order; o < MAX_ORDER; o++) {
3671 struct free_area *area = &z->free_area[o];
3672 int mt;
3673
3674 if (!area->nr_free)
3675 continue;
3676
3677 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3678 if (!free_area_empty(area, mt))
3679 return true;
3680 }
3681
3682 #ifdef CONFIG_CMA
3683 if ((alloc_flags & ALLOC_CMA) &&
3684 !free_area_empty(area, MIGRATE_CMA)) {
3685 return true;
3686 }
3687 #endif
3688 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3689 return true;
3690 }
3691 return false;
3692 }
3693
3694 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3695 int highest_zoneidx, unsigned int alloc_flags)
3696 {
3697 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3698 zone_page_state(z, NR_FREE_PAGES));
3699 }
3700
3701 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3702 unsigned long mark, int highest_zoneidx,
3703 unsigned int alloc_flags, gfp_t gfp_mask)
3704 {
3705 long free_pages;
3706
3707 free_pages = zone_page_state(z, NR_FREE_PAGES);
3708
3709 /*
3710 * Fast check for order-0 only. If this fails then the reserves
3711 * need to be calculated.
3712 */
3713 if (!order) {
3714 long fast_free;
3715
3716 fast_free = free_pages;
3717 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3718 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3719 return true;
3720 }
3721
3722 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3723 free_pages))
3724 return true;
3725 /*
3726 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3727 * when checking the min watermark. The min watermark is the
3728 * point where boosting is ignored so that kswapd is woken up
3729 * when below the low watermark.
3730 */
3731 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3732 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3733 mark = z->_watermark[WMARK_MIN];
3734 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3735 alloc_flags, free_pages);
3736 }
3737
3738 return false;
3739 }
3740
3741 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3742 unsigned long mark, int highest_zoneidx)
3743 {
3744 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3745
3746 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3747 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3748
3749 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3750 free_pages);
3751 }
3752
3753 #ifdef CONFIG_NUMA
3754 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3755 {
3756 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3757 node_reclaim_distance;
3758 }
3759 #else /* CONFIG_NUMA */
3760 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3761 {
3762 return true;
3763 }
3764 #endif /* CONFIG_NUMA */
3765
3766 /*
3767 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3768 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3769 * premature use of a lower zone may cause lowmem pressure problems that
3770 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3771 * probably too small. It only makes sense to spread allocations to avoid
3772 * fragmentation between the Normal and DMA32 zones.
3773 */
3774 static inline unsigned int
3775 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3776 {
3777 unsigned int alloc_flags;
3778
3779 /*
3780 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3781 * to save a branch.
3782 */
3783 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3784
3785 #ifdef CONFIG_ZONE_DMA32
3786 if (!zone)
3787 return alloc_flags;
3788
3789 if (zone_idx(zone) != ZONE_NORMAL)
3790 return alloc_flags;
3791
3792 /*
3793 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3794 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3795 * on UMA that if Normal is populated then so is DMA32.
3796 */
3797 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3798 if (nr_online_nodes > 1 && !populated_zone(--zone))
3799 return alloc_flags;
3800
3801 alloc_flags |= ALLOC_NOFRAGMENT;
3802 #endif /* CONFIG_ZONE_DMA32 */
3803 return alloc_flags;
3804 }
3805
3806 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3807 unsigned int alloc_flags)
3808 {
3809 #ifdef CONFIG_CMA
3810 unsigned int pflags = current->flags;
3811
3812 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3813 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3814 alloc_flags |= ALLOC_CMA;
3815
3816 #endif
3817 return alloc_flags;
3818 }
3819
3820 /*
3821 * get_page_from_freelist goes through the zonelist trying to allocate
3822 * a page.
3823 */
3824 static struct page *
3825 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3826 const struct alloc_context *ac)
3827 {
3828 struct zoneref *z;
3829 struct zone *zone;
3830 struct pglist_data *last_pgdat_dirty_limit = NULL;
3831 bool no_fallback;
3832
3833 retry:
3834 /*
3835 * Scan zonelist, looking for a zone with enough free.
3836 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3837 */
3838 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3839 z = ac->preferred_zoneref;
3840 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3841 ac->nodemask) {
3842 struct page *page;
3843 unsigned long mark;
3844
3845 if (cpusets_enabled() &&
3846 (alloc_flags & ALLOC_CPUSET) &&
3847 !__cpuset_zone_allowed(zone, gfp_mask))
3848 continue;
3849 /*
3850 * When allocating a page cache page for writing, we
3851 * want to get it from a node that is within its dirty
3852 * limit, such that no single node holds more than its
3853 * proportional share of globally allowed dirty pages.
3854 * The dirty limits take into account the node's
3855 * lowmem reserves and high watermark so that kswapd
3856 * should be able to balance it without having to
3857 * write pages from its LRU list.
3858 *
3859 * XXX: For now, allow allocations to potentially
3860 * exceed the per-node dirty limit in the slowpath
3861 * (spread_dirty_pages unset) before going into reclaim,
3862 * which is important when on a NUMA setup the allowed
3863 * nodes are together not big enough to reach the
3864 * global limit. The proper fix for these situations
3865 * will require awareness of nodes in the
3866 * dirty-throttling and the flusher threads.
3867 */
3868 if (ac->spread_dirty_pages) {
3869 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3870 continue;
3871
3872 if (!node_dirty_ok(zone->zone_pgdat)) {
3873 last_pgdat_dirty_limit = zone->zone_pgdat;
3874 continue;
3875 }
3876 }
3877
3878 if (no_fallback && nr_online_nodes > 1 &&
3879 zone != ac->preferred_zoneref->zone) {
3880 int local_nid;
3881
3882 /*
3883 * If moving to a remote node, retry but allow
3884 * fragmenting fallbacks. Locality is more important
3885 * than fragmentation avoidance.
3886 */
3887 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3888 if (zone_to_nid(zone) != local_nid) {
3889 alloc_flags &= ~ALLOC_NOFRAGMENT;
3890 goto retry;
3891 }
3892 }
3893
3894 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3895 if (!zone_watermark_fast(zone, order, mark,
3896 ac->highest_zoneidx, alloc_flags,
3897 gfp_mask)) {
3898 int ret;
3899
3900 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3901 /*
3902 * Watermark failed for this zone, but see if we can
3903 * grow this zone if it contains deferred pages.
3904 */
3905 if (static_branch_unlikely(&deferred_pages)) {
3906 if (_deferred_grow_zone(zone, order))
3907 goto try_this_zone;
3908 }
3909 #endif
3910 /* Checked here to keep the fast path fast */
3911 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3912 if (alloc_flags & ALLOC_NO_WATERMARKS)
3913 goto try_this_zone;
3914
3915 if (node_reclaim_mode == 0 ||
3916 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3917 continue;
3918
3919 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3920 switch (ret) {
3921 case NODE_RECLAIM_NOSCAN:
3922 /* did not scan */
3923 continue;
3924 case NODE_RECLAIM_FULL:
3925 /* scanned but unreclaimable */
3926 continue;
3927 default:
3928 /* did we reclaim enough */
3929 if (zone_watermark_ok(zone, order, mark,
3930 ac->highest_zoneidx, alloc_flags))
3931 goto try_this_zone;
3932
3933 continue;
3934 }
3935 }
3936
3937 try_this_zone:
3938 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3939 gfp_mask, alloc_flags, ac->migratetype);
3940 if (page) {
3941 prep_new_page(page, order, gfp_mask, alloc_flags);
3942
3943 /*
3944 * If this is a high-order atomic allocation then check
3945 * if the pageblock should be reserved for the future
3946 */
3947 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3948 reserve_highatomic_pageblock(page, zone, order);
3949
3950 return page;
3951 } else {
3952 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3953 /* Try again if zone has deferred pages */
3954 if (static_branch_unlikely(&deferred_pages)) {
3955 if (_deferred_grow_zone(zone, order))
3956 goto try_this_zone;
3957 }
3958 #endif
3959 }
3960 }
3961
3962 /*
3963 * It's possible on a UMA machine to get through all zones that are
3964 * fragmented. If avoiding fragmentation, reset and try again.
3965 */
3966 if (no_fallback) {
3967 alloc_flags &= ~ALLOC_NOFRAGMENT;
3968 goto retry;
3969 }
3970
3971 return NULL;
3972 }
3973
3974 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3975 {
3976 unsigned int filter = SHOW_MEM_FILTER_NODES;
3977
3978 /*
3979 * This documents exceptions given to allocations in certain
3980 * contexts that are allowed to allocate outside current's set
3981 * of allowed nodes.
3982 */
3983 if (!(gfp_mask & __GFP_NOMEMALLOC))
3984 if (tsk_is_oom_victim(current) ||
3985 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3986 filter &= ~SHOW_MEM_FILTER_NODES;
3987 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3988 filter &= ~SHOW_MEM_FILTER_NODES;
3989
3990 show_mem(filter, nodemask);
3991 }
3992
3993 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3994 {
3995 struct va_format vaf;
3996 va_list args;
3997 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3998
3999 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4000 return;
4001
4002 va_start(args, fmt);
4003 vaf.fmt = fmt;
4004 vaf.va = &args;
4005 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4006 current->comm, &vaf, gfp_mask, &gfp_mask,
4007 nodemask_pr_args(nodemask));
4008 va_end(args);
4009
4010 cpuset_print_current_mems_allowed();
4011 pr_cont("\n");
4012 dump_stack();
4013 warn_alloc_show_mem(gfp_mask, nodemask);
4014 }
4015
4016 static inline struct page *
4017 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4018 unsigned int alloc_flags,
4019 const struct alloc_context *ac)
4020 {
4021 struct page *page;
4022
4023 page = get_page_from_freelist(gfp_mask, order,
4024 alloc_flags|ALLOC_CPUSET, ac);
4025 /*
4026 * fallback to ignore cpuset restriction if our nodes
4027 * are depleted
4028 */
4029 if (!page)
4030 page = get_page_from_freelist(gfp_mask, order,
4031 alloc_flags, ac);
4032
4033 return page;
4034 }
4035
4036 static inline struct page *
4037 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4038 const struct alloc_context *ac, unsigned long *did_some_progress)
4039 {
4040 struct oom_control oc = {
4041 .zonelist = ac->zonelist,
4042 .nodemask = ac->nodemask,
4043 .memcg = NULL,
4044 .gfp_mask = gfp_mask,
4045 .order = order,
4046 };
4047 struct page *page;
4048
4049 *did_some_progress = 0;
4050
4051 /*
4052 * Acquire the oom lock. If that fails, somebody else is
4053 * making progress for us.
4054 */
4055 if (!mutex_trylock(&oom_lock)) {
4056 *did_some_progress = 1;
4057 schedule_timeout_uninterruptible(1);
4058 return NULL;
4059 }
4060
4061 /*
4062 * Go through the zonelist yet one more time, keep very high watermark
4063 * here, this is only to catch a parallel oom killing, we must fail if
4064 * we're still under heavy pressure. But make sure that this reclaim
4065 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4066 * allocation which will never fail due to oom_lock already held.
4067 */
4068 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4069 ~__GFP_DIRECT_RECLAIM, order,
4070 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4071 if (page)
4072 goto out;
4073
4074 /* Coredumps can quickly deplete all memory reserves */
4075 if (current->flags & PF_DUMPCORE)
4076 goto out;
4077 /* The OOM killer will not help higher order allocs */
4078 if (order > PAGE_ALLOC_COSTLY_ORDER)
4079 goto out;
4080 /*
4081 * We have already exhausted all our reclaim opportunities without any
4082 * success so it is time to admit defeat. We will skip the OOM killer
4083 * because it is very likely that the caller has a more reasonable
4084 * fallback than shooting a random task.
4085 *
4086 * The OOM killer may not free memory on a specific node.
4087 */
4088 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4089 goto out;
4090 /* The OOM killer does not needlessly kill tasks for lowmem */
4091 if (ac->highest_zoneidx < ZONE_NORMAL)
4092 goto out;
4093 if (pm_suspended_storage())
4094 goto out;
4095 /*
4096 * XXX: GFP_NOFS allocations should rather fail than rely on
4097 * other request to make a forward progress.
4098 * We are in an unfortunate situation where out_of_memory cannot
4099 * do much for this context but let's try it to at least get
4100 * access to memory reserved if the current task is killed (see
4101 * out_of_memory). Once filesystems are ready to handle allocation
4102 * failures more gracefully we should just bail out here.
4103 */
4104
4105 /* Exhausted what can be done so it's blame time */
4106 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4107 *did_some_progress = 1;
4108
4109 /*
4110 * Help non-failing allocations by giving them access to memory
4111 * reserves
4112 */
4113 if (gfp_mask & __GFP_NOFAIL)
4114 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4115 ALLOC_NO_WATERMARKS, ac);
4116 }
4117 out:
4118 mutex_unlock(&oom_lock);
4119 return page;
4120 }
4121
4122 /*
4123 * Maximum number of compaction retries wit a progress before OOM
4124 * killer is consider as the only way to move forward.
4125 */
4126 #define MAX_COMPACT_RETRIES 16
4127
4128 #ifdef CONFIG_COMPACTION
4129 /* Try memory compaction for high-order allocations before reclaim */
4130 static struct page *
4131 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4132 unsigned int alloc_flags, const struct alloc_context *ac,
4133 enum compact_priority prio, enum compact_result *compact_result)
4134 {
4135 struct page *page = NULL;
4136 unsigned long pflags;
4137 unsigned int noreclaim_flag;
4138
4139 if (!order)
4140 return NULL;
4141
4142 psi_memstall_enter(&pflags);
4143 noreclaim_flag = memalloc_noreclaim_save();
4144
4145 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4146 prio, &page);
4147
4148 memalloc_noreclaim_restore(noreclaim_flag);
4149 psi_memstall_leave(&pflags);
4150
4151 /*
4152 * At least in one zone compaction wasn't deferred or skipped, so let's
4153 * count a compaction stall
4154 */
4155 count_vm_event(COMPACTSTALL);
4156
4157 /* Prep a captured page if available */
4158 if (page)
4159 prep_new_page(page, order, gfp_mask, alloc_flags);
4160
4161 /* Try get a page from the freelist if available */
4162 if (!page)
4163 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4164
4165 if (page) {
4166 struct zone *zone = page_zone(page);
4167
4168 zone->compact_blockskip_flush = false;
4169 compaction_defer_reset(zone, order, true);
4170 count_vm_event(COMPACTSUCCESS);
4171 return page;
4172 }
4173
4174 /*
4175 * It's bad if compaction run occurs and fails. The most likely reason
4176 * is that pages exist, but not enough to satisfy watermarks.
4177 */
4178 count_vm_event(COMPACTFAIL);
4179
4180 cond_resched();
4181
4182 return NULL;
4183 }
4184
4185 static inline bool
4186 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4187 enum compact_result compact_result,
4188 enum compact_priority *compact_priority,
4189 int *compaction_retries)
4190 {
4191 int max_retries = MAX_COMPACT_RETRIES;
4192 int min_priority;
4193 bool ret = false;
4194 int retries = *compaction_retries;
4195 enum compact_priority priority = *compact_priority;
4196
4197 if (!order)
4198 return false;
4199
4200 if (compaction_made_progress(compact_result))
4201 (*compaction_retries)++;
4202
4203 /*
4204 * compaction considers all the zone as desperately out of memory
4205 * so it doesn't really make much sense to retry except when the
4206 * failure could be caused by insufficient priority
4207 */
4208 if (compaction_failed(compact_result))
4209 goto check_priority;
4210
4211 /*
4212 * compaction was skipped because there are not enough order-0 pages
4213 * to work with, so we retry only if it looks like reclaim can help.
4214 */
4215 if (compaction_needs_reclaim(compact_result)) {
4216 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4217 goto out;
4218 }
4219
4220 /*
4221 * make sure the compaction wasn't deferred or didn't bail out early
4222 * due to locks contention before we declare that we should give up.
4223 * But the next retry should use a higher priority if allowed, so
4224 * we don't just keep bailing out endlessly.
4225 */
4226 if (compaction_withdrawn(compact_result)) {
4227 goto check_priority;
4228 }
4229
4230 /*
4231 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4232 * costly ones because they are de facto nofail and invoke OOM
4233 * killer to move on while costly can fail and users are ready
4234 * to cope with that. 1/4 retries is rather arbitrary but we
4235 * would need much more detailed feedback from compaction to
4236 * make a better decision.
4237 */
4238 if (order > PAGE_ALLOC_COSTLY_ORDER)
4239 max_retries /= 4;
4240 if (*compaction_retries <= max_retries) {
4241 ret = true;
4242 goto out;
4243 }
4244
4245 /*
4246 * Make sure there are attempts at the highest priority if we exhausted
4247 * all retries or failed at the lower priorities.
4248 */
4249 check_priority:
4250 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4251 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4252
4253 if (*compact_priority > min_priority) {
4254 (*compact_priority)--;
4255 *compaction_retries = 0;
4256 ret = true;
4257 }
4258 out:
4259 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4260 return ret;
4261 }
4262 #else
4263 static inline struct page *
4264 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4265 unsigned int alloc_flags, const struct alloc_context *ac,
4266 enum compact_priority prio, enum compact_result *compact_result)
4267 {
4268 *compact_result = COMPACT_SKIPPED;
4269 return NULL;
4270 }
4271
4272 static inline bool
4273 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4274 enum compact_result compact_result,
4275 enum compact_priority *compact_priority,
4276 int *compaction_retries)
4277 {
4278 struct zone *zone;
4279 struct zoneref *z;
4280
4281 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4282 return false;
4283
4284 /*
4285 * There are setups with compaction disabled which would prefer to loop
4286 * inside the allocator rather than hit the oom killer prematurely.
4287 * Let's give them a good hope and keep retrying while the order-0
4288 * watermarks are OK.
4289 */
4290 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4291 ac->highest_zoneidx, ac->nodemask) {
4292 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4293 ac->highest_zoneidx, alloc_flags))
4294 return true;
4295 }
4296 return false;
4297 }
4298 #endif /* CONFIG_COMPACTION */
4299
4300 #ifdef CONFIG_LOCKDEP
4301 static struct lockdep_map __fs_reclaim_map =
4302 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4303
4304 static bool __need_reclaim(gfp_t gfp_mask)
4305 {
4306 /* no reclaim without waiting on it */
4307 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4308 return false;
4309
4310 /* this guy won't enter reclaim */
4311 if (current->flags & PF_MEMALLOC)
4312 return false;
4313
4314 if (gfp_mask & __GFP_NOLOCKDEP)
4315 return false;
4316
4317 return true;
4318 }
4319
4320 void __fs_reclaim_acquire(void)
4321 {
4322 lock_map_acquire(&__fs_reclaim_map);
4323 }
4324
4325 void __fs_reclaim_release(void)
4326 {
4327 lock_map_release(&__fs_reclaim_map);
4328 }
4329
4330 void fs_reclaim_acquire(gfp_t gfp_mask)
4331 {
4332 gfp_mask = current_gfp_context(gfp_mask);
4333
4334 if (__need_reclaim(gfp_mask)) {
4335 if (gfp_mask & __GFP_FS)
4336 __fs_reclaim_acquire();
4337
4338 #ifdef CONFIG_MMU_NOTIFIER
4339 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4340 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4341 #endif
4342
4343 }
4344 }
4345 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4346
4347 void fs_reclaim_release(gfp_t gfp_mask)
4348 {
4349 gfp_mask = current_gfp_context(gfp_mask);
4350
4351 if (__need_reclaim(gfp_mask)) {
4352 if (gfp_mask & __GFP_FS)
4353 __fs_reclaim_release();
4354 }
4355 }
4356 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4357 #endif
4358
4359 /* Perform direct synchronous page reclaim */
4360 static unsigned long
4361 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4362 const struct alloc_context *ac)
4363 {
4364 unsigned int noreclaim_flag;
4365 unsigned long pflags, progress;
4366
4367 cond_resched();
4368
4369 /* We now go into synchronous reclaim */
4370 cpuset_memory_pressure_bump();
4371 psi_memstall_enter(&pflags);
4372 fs_reclaim_acquire(gfp_mask);
4373 noreclaim_flag = memalloc_noreclaim_save();
4374
4375 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4376 ac->nodemask);
4377
4378 memalloc_noreclaim_restore(noreclaim_flag);
4379 fs_reclaim_release(gfp_mask);
4380 psi_memstall_leave(&pflags);
4381
4382 cond_resched();
4383
4384 return progress;
4385 }
4386
4387 /* The really slow allocator path where we enter direct reclaim */
4388 static inline struct page *
4389 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4390 unsigned int alloc_flags, const struct alloc_context *ac,
4391 unsigned long *did_some_progress)
4392 {
4393 struct page *page = NULL;
4394 bool drained = false;
4395
4396 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4397 if (unlikely(!(*did_some_progress)))
4398 return NULL;
4399
4400 retry:
4401 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4402
4403 /*
4404 * If an allocation failed after direct reclaim, it could be because
4405 * pages are pinned on the per-cpu lists or in high alloc reserves.
4406 * Shrink them and try again
4407 */
4408 if (!page && !drained) {
4409 unreserve_highatomic_pageblock(ac, false);
4410 drain_all_pages(NULL);
4411 drained = true;
4412 goto retry;
4413 }
4414
4415 return page;
4416 }
4417
4418 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4419 const struct alloc_context *ac)
4420 {
4421 struct zoneref *z;
4422 struct zone *zone;
4423 pg_data_t *last_pgdat = NULL;
4424 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4425
4426 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4427 ac->nodemask) {
4428 if (last_pgdat != zone->zone_pgdat)
4429 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4430 last_pgdat = zone->zone_pgdat;
4431 }
4432 }
4433
4434 static inline unsigned int
4435 gfp_to_alloc_flags(gfp_t gfp_mask)
4436 {
4437 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4438
4439 /*
4440 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4441 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4442 * to save two branches.
4443 */
4444 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4445 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4446
4447 /*
4448 * The caller may dip into page reserves a bit more if the caller
4449 * cannot run direct reclaim, or if the caller has realtime scheduling
4450 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4451 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4452 */
4453 alloc_flags |= (__force int)
4454 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4455
4456 if (gfp_mask & __GFP_ATOMIC) {
4457 /*
4458 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4459 * if it can't schedule.
4460 */
4461 if (!(gfp_mask & __GFP_NOMEMALLOC))
4462 alloc_flags |= ALLOC_HARDER;
4463 /*
4464 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4465 * comment for __cpuset_node_allowed().
4466 */
4467 alloc_flags &= ~ALLOC_CPUSET;
4468 } else if (unlikely(rt_task(current)) && !in_interrupt())
4469 alloc_flags |= ALLOC_HARDER;
4470
4471 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4472
4473 return alloc_flags;
4474 }
4475
4476 static bool oom_reserves_allowed(struct task_struct *tsk)
4477 {
4478 if (!tsk_is_oom_victim(tsk))
4479 return false;
4480
4481 /*
4482 * !MMU doesn't have oom reaper so give access to memory reserves
4483 * only to the thread with TIF_MEMDIE set
4484 */
4485 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4486 return false;
4487
4488 return true;
4489 }
4490
4491 /*
4492 * Distinguish requests which really need access to full memory
4493 * reserves from oom victims which can live with a portion of it
4494 */
4495 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4496 {
4497 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4498 return 0;
4499 if (gfp_mask & __GFP_MEMALLOC)
4500 return ALLOC_NO_WATERMARKS;
4501 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4502 return ALLOC_NO_WATERMARKS;
4503 if (!in_interrupt()) {
4504 if (current->flags & PF_MEMALLOC)
4505 return ALLOC_NO_WATERMARKS;
4506 else if (oom_reserves_allowed(current))
4507 return ALLOC_OOM;
4508 }
4509
4510 return 0;
4511 }
4512
4513 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4514 {
4515 return !!__gfp_pfmemalloc_flags(gfp_mask);
4516 }
4517
4518 /*
4519 * Checks whether it makes sense to retry the reclaim to make a forward progress
4520 * for the given allocation request.
4521 *
4522 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4523 * without success, or when we couldn't even meet the watermark if we
4524 * reclaimed all remaining pages on the LRU lists.
4525 *
4526 * Returns true if a retry is viable or false to enter the oom path.
4527 */
4528 static inline bool
4529 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4530 struct alloc_context *ac, int alloc_flags,
4531 bool did_some_progress, int *no_progress_loops)
4532 {
4533 struct zone *zone;
4534 struct zoneref *z;
4535 bool ret = false;
4536
4537 /*
4538 * Costly allocations might have made a progress but this doesn't mean
4539 * their order will become available due to high fragmentation so
4540 * always increment the no progress counter for them
4541 */
4542 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4543 *no_progress_loops = 0;
4544 else
4545 (*no_progress_loops)++;
4546
4547 /*
4548 * Make sure we converge to OOM if we cannot make any progress
4549 * several times in the row.
4550 */
4551 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4552 /* Before OOM, exhaust highatomic_reserve */
4553 return unreserve_highatomic_pageblock(ac, true);
4554 }
4555
4556 /*
4557 * Keep reclaiming pages while there is a chance this will lead
4558 * somewhere. If none of the target zones can satisfy our allocation
4559 * request even if all reclaimable pages are considered then we are
4560 * screwed and have to go OOM.
4561 */
4562 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4563 ac->highest_zoneidx, ac->nodemask) {
4564 unsigned long available;
4565 unsigned long reclaimable;
4566 unsigned long min_wmark = min_wmark_pages(zone);
4567 bool wmark;
4568
4569 available = reclaimable = zone_reclaimable_pages(zone);
4570 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4571
4572 /*
4573 * Would the allocation succeed if we reclaimed all
4574 * reclaimable pages?
4575 */
4576 wmark = __zone_watermark_ok(zone, order, min_wmark,
4577 ac->highest_zoneidx, alloc_flags, available);
4578 trace_reclaim_retry_zone(z, order, reclaimable,
4579 available, min_wmark, *no_progress_loops, wmark);
4580 if (wmark) {
4581 /*
4582 * If we didn't make any progress and have a lot of
4583 * dirty + writeback pages then we should wait for
4584 * an IO to complete to slow down the reclaim and
4585 * prevent from pre mature OOM
4586 */
4587 if (!did_some_progress) {
4588 unsigned long write_pending;
4589
4590 write_pending = zone_page_state_snapshot(zone,
4591 NR_ZONE_WRITE_PENDING);
4592
4593 if (2 * write_pending > reclaimable) {
4594 congestion_wait(BLK_RW_ASYNC, HZ/10);
4595 return true;
4596 }
4597 }
4598
4599 ret = true;
4600 goto out;
4601 }
4602 }
4603
4604 out:
4605 /*
4606 * Memory allocation/reclaim might be called from a WQ context and the
4607 * current implementation of the WQ concurrency control doesn't
4608 * recognize that a particular WQ is congested if the worker thread is
4609 * looping without ever sleeping. Therefore we have to do a short sleep
4610 * here rather than calling cond_resched().
4611 */
4612 if (current->flags & PF_WQ_WORKER)
4613 schedule_timeout_uninterruptible(1);
4614 else
4615 cond_resched();
4616 return ret;
4617 }
4618
4619 static inline bool
4620 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4621 {
4622 /*
4623 * It's possible that cpuset's mems_allowed and the nodemask from
4624 * mempolicy don't intersect. This should be normally dealt with by
4625 * policy_nodemask(), but it's possible to race with cpuset update in
4626 * such a way the check therein was true, and then it became false
4627 * before we got our cpuset_mems_cookie here.
4628 * This assumes that for all allocations, ac->nodemask can come only
4629 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4630 * when it does not intersect with the cpuset restrictions) or the
4631 * caller can deal with a violated nodemask.
4632 */
4633 if (cpusets_enabled() && ac->nodemask &&
4634 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4635 ac->nodemask = NULL;
4636 return true;
4637 }
4638
4639 /*
4640 * When updating a task's mems_allowed or mempolicy nodemask, it is
4641 * possible to race with parallel threads in such a way that our
4642 * allocation can fail while the mask is being updated. If we are about
4643 * to fail, check if the cpuset changed during allocation and if so,
4644 * retry.
4645 */
4646 if (read_mems_allowed_retry(cpuset_mems_cookie))
4647 return true;
4648
4649 return false;
4650 }
4651
4652 static inline struct page *
4653 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4654 struct alloc_context *ac)
4655 {
4656 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4657 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4658 struct page *page = NULL;
4659 unsigned int alloc_flags;
4660 unsigned long did_some_progress;
4661 enum compact_priority compact_priority;
4662 enum compact_result compact_result;
4663 int compaction_retries;
4664 int no_progress_loops;
4665 unsigned int cpuset_mems_cookie;
4666 int reserve_flags;
4667
4668 /*
4669 * We also sanity check to catch abuse of atomic reserves being used by
4670 * callers that are not in atomic context.
4671 */
4672 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4673 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4674 gfp_mask &= ~__GFP_ATOMIC;
4675
4676 retry_cpuset:
4677 compaction_retries = 0;
4678 no_progress_loops = 0;
4679 compact_priority = DEF_COMPACT_PRIORITY;
4680 cpuset_mems_cookie = read_mems_allowed_begin();
4681
4682 /*
4683 * The fast path uses conservative alloc_flags to succeed only until
4684 * kswapd needs to be woken up, and to avoid the cost of setting up
4685 * alloc_flags precisely. So we do that now.
4686 */
4687 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4688
4689 /*
4690 * We need to recalculate the starting point for the zonelist iterator
4691 * because we might have used different nodemask in the fast path, or
4692 * there was a cpuset modification and we are retrying - otherwise we
4693 * could end up iterating over non-eligible zones endlessly.
4694 */
4695 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4696 ac->highest_zoneidx, ac->nodemask);
4697 if (!ac->preferred_zoneref->zone)
4698 goto nopage;
4699
4700 if (alloc_flags & ALLOC_KSWAPD)
4701 wake_all_kswapds(order, gfp_mask, ac);
4702
4703 /*
4704 * The adjusted alloc_flags might result in immediate success, so try
4705 * that first
4706 */
4707 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4708 if (page)
4709 goto got_pg;
4710
4711 /*
4712 * For costly allocations, try direct compaction first, as it's likely
4713 * that we have enough base pages and don't need to reclaim. For non-
4714 * movable high-order allocations, do that as well, as compaction will
4715 * try prevent permanent fragmentation by migrating from blocks of the
4716 * same migratetype.
4717 * Don't try this for allocations that are allowed to ignore
4718 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4719 */
4720 if (can_direct_reclaim &&
4721 (costly_order ||
4722 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4723 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4724 page = __alloc_pages_direct_compact(gfp_mask, order,
4725 alloc_flags, ac,
4726 INIT_COMPACT_PRIORITY,
4727 &compact_result);
4728 if (page)
4729 goto got_pg;
4730
4731 /*
4732 * Checks for costly allocations with __GFP_NORETRY, which
4733 * includes some THP page fault allocations
4734 */
4735 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4736 /*
4737 * If allocating entire pageblock(s) and compaction
4738 * failed because all zones are below low watermarks
4739 * or is prohibited because it recently failed at this
4740 * order, fail immediately unless the allocator has
4741 * requested compaction and reclaim retry.
4742 *
4743 * Reclaim is
4744 * - potentially very expensive because zones are far
4745 * below their low watermarks or this is part of very
4746 * bursty high order allocations,
4747 * - not guaranteed to help because isolate_freepages()
4748 * may not iterate over freed pages as part of its
4749 * linear scan, and
4750 * - unlikely to make entire pageblocks free on its
4751 * own.
4752 */
4753 if (compact_result == COMPACT_SKIPPED ||
4754 compact_result == COMPACT_DEFERRED)
4755 goto nopage;
4756
4757 /*
4758 * Looks like reclaim/compaction is worth trying, but
4759 * sync compaction could be very expensive, so keep
4760 * using async compaction.
4761 */
4762 compact_priority = INIT_COMPACT_PRIORITY;
4763 }
4764 }
4765
4766 retry:
4767 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4768 if (alloc_flags & ALLOC_KSWAPD)
4769 wake_all_kswapds(order, gfp_mask, ac);
4770
4771 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4772 if (reserve_flags)
4773 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4774
4775 /*
4776 * Reset the nodemask and zonelist iterators if memory policies can be
4777 * ignored. These allocations are high priority and system rather than
4778 * user oriented.
4779 */
4780 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4781 ac->nodemask = NULL;
4782 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4783 ac->highest_zoneidx, ac->nodemask);
4784 }
4785
4786 /* Attempt with potentially adjusted zonelist and alloc_flags */
4787 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4788 if (page)
4789 goto got_pg;
4790
4791 /* Caller is not willing to reclaim, we can't balance anything */
4792 if (!can_direct_reclaim)
4793 goto nopage;
4794
4795 /* Avoid recursion of direct reclaim */
4796 if (current->flags & PF_MEMALLOC)
4797 goto nopage;
4798
4799 /* Try direct reclaim and then allocating */
4800 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4801 &did_some_progress);
4802 if (page)
4803 goto got_pg;
4804
4805 /* Try direct compaction and then allocating */
4806 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4807 compact_priority, &compact_result);
4808 if (page)
4809 goto got_pg;
4810
4811 /* Do not loop if specifically requested */
4812 if (gfp_mask & __GFP_NORETRY)
4813 goto nopage;
4814
4815 /*
4816 * Do not retry costly high order allocations unless they are
4817 * __GFP_RETRY_MAYFAIL
4818 */
4819 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4820 goto nopage;
4821
4822 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4823 did_some_progress > 0, &no_progress_loops))
4824 goto retry;
4825
4826 /*
4827 * It doesn't make any sense to retry for the compaction if the order-0
4828 * reclaim is not able to make any progress because the current
4829 * implementation of the compaction depends on the sufficient amount
4830 * of free memory (see __compaction_suitable)
4831 */
4832 if (did_some_progress > 0 &&
4833 should_compact_retry(ac, order, alloc_flags,
4834 compact_result, &compact_priority,
4835 &compaction_retries))
4836 goto retry;
4837
4838
4839 /* Deal with possible cpuset update races before we start OOM killing */
4840 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4841 goto retry_cpuset;
4842
4843 /* Reclaim has failed us, start killing things */
4844 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4845 if (page)
4846 goto got_pg;
4847
4848 /* Avoid allocations with no watermarks from looping endlessly */
4849 if (tsk_is_oom_victim(current) &&
4850 (alloc_flags & ALLOC_OOM ||
4851 (gfp_mask & __GFP_NOMEMALLOC)))
4852 goto nopage;
4853
4854 /* Retry as long as the OOM killer is making progress */
4855 if (did_some_progress) {
4856 no_progress_loops = 0;
4857 goto retry;
4858 }
4859
4860 nopage:
4861 /* Deal with possible cpuset update races before we fail */
4862 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4863 goto retry_cpuset;
4864
4865 /*
4866 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4867 * we always retry
4868 */
4869 if (gfp_mask & __GFP_NOFAIL) {
4870 /*
4871 * All existing users of the __GFP_NOFAIL are blockable, so warn
4872 * of any new users that actually require GFP_NOWAIT
4873 */
4874 if (WARN_ON_ONCE(!can_direct_reclaim))
4875 goto fail;
4876
4877 /*
4878 * PF_MEMALLOC request from this context is rather bizarre
4879 * because we cannot reclaim anything and only can loop waiting
4880 * for somebody to do a work for us
4881 */
4882 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4883
4884 /*
4885 * non failing costly orders are a hard requirement which we
4886 * are not prepared for much so let's warn about these users
4887 * so that we can identify them and convert them to something
4888 * else.
4889 */
4890 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4891
4892 /*
4893 * Help non-failing allocations by giving them access to memory
4894 * reserves but do not use ALLOC_NO_WATERMARKS because this
4895 * could deplete whole memory reserves which would just make
4896 * the situation worse
4897 */
4898 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4899 if (page)
4900 goto got_pg;
4901
4902 cond_resched();
4903 goto retry;
4904 }
4905 fail:
4906 warn_alloc(gfp_mask, ac->nodemask,
4907 "page allocation failure: order:%u", order);
4908 got_pg:
4909 return page;
4910 }
4911
4912 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4913 int preferred_nid, nodemask_t *nodemask,
4914 struct alloc_context *ac, gfp_t *alloc_mask,
4915 unsigned int *alloc_flags)
4916 {
4917 ac->highest_zoneidx = gfp_zone(gfp_mask);
4918 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4919 ac->nodemask = nodemask;
4920 ac->migratetype = gfp_migratetype(gfp_mask);
4921
4922 if (cpusets_enabled()) {
4923 *alloc_mask |= __GFP_HARDWALL;
4924 /*
4925 * When we are in the interrupt context, it is irrelevant
4926 * to the current task context. It means that any node ok.
4927 */
4928 if (!in_interrupt() && !ac->nodemask)
4929 ac->nodemask = &cpuset_current_mems_allowed;
4930 else
4931 *alloc_flags |= ALLOC_CPUSET;
4932 }
4933
4934 fs_reclaim_acquire(gfp_mask);
4935 fs_reclaim_release(gfp_mask);
4936
4937 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4938
4939 if (should_fail_alloc_page(gfp_mask, order))
4940 return false;
4941
4942 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4943
4944 /* Dirty zone balancing only done in the fast path */
4945 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4946
4947 /*
4948 * The preferred zone is used for statistics but crucially it is
4949 * also used as the starting point for the zonelist iterator. It
4950 * may get reset for allocations that ignore memory policies.
4951 */
4952 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4953 ac->highest_zoneidx, ac->nodemask);
4954
4955 return true;
4956 }
4957
4958 /*
4959 * This is the 'heart' of the zoned buddy allocator.
4960 */
4961 struct page *
4962 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4963 nodemask_t *nodemask)
4964 {
4965 struct page *page;
4966 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4967 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4968 struct alloc_context ac = { };
4969
4970 /*
4971 * There are several places where we assume that the order value is sane
4972 * so bail out early if the request is out of bound.
4973 */
4974 if (unlikely(order >= MAX_ORDER)) {
4975 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4976 return NULL;
4977 }
4978
4979 gfp_mask &= gfp_allowed_mask;
4980 alloc_mask = gfp_mask;
4981 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4982 return NULL;
4983
4984 /*
4985 * Forbid the first pass from falling back to types that fragment
4986 * memory until all local zones are considered.
4987 */
4988 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4989
4990 /* First allocation attempt */
4991 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4992 if (likely(page))
4993 goto out;
4994
4995 /*
4996 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4997 * resp. GFP_NOIO which has to be inherited for all allocation requests
4998 * from a particular context which has been marked by
4999 * memalloc_no{fs,io}_{save,restore}.
5000 */
5001 alloc_mask = current_gfp_context(gfp_mask);
5002 ac.spread_dirty_pages = false;
5003
5004 /*
5005 * Restore the original nodemask if it was potentially replaced with
5006 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5007 */
5008 ac.nodemask = nodemask;
5009
5010 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5011
5012 out:
5013 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5014 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5015 __free_pages(page, order);
5016 page = NULL;
5017 }
5018
5019 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5020
5021 return page;
5022 }
5023 EXPORT_SYMBOL(__alloc_pages_nodemask);
5024
5025 /*
5026 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5027 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5028 * you need to access high mem.
5029 */
5030 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5031 {
5032 struct page *page;
5033
5034 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5035 if (!page)
5036 return 0;
5037 return (unsigned long) page_address(page);
5038 }
5039 EXPORT_SYMBOL(__get_free_pages);
5040
5041 unsigned long get_zeroed_page(gfp_t gfp_mask)
5042 {
5043 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5044 }
5045 EXPORT_SYMBOL(get_zeroed_page);
5046
5047 static inline void free_the_page(struct page *page, unsigned int order)
5048 {
5049 if (order == 0) /* Via pcp? */
5050 free_unref_page(page);
5051 else
5052 __free_pages_ok(page, order, FPI_NONE);
5053 }
5054
5055 /**
5056 * __free_pages - Free pages allocated with alloc_pages().
5057 * @page: The page pointer returned from alloc_pages().
5058 * @order: The order of the allocation.
5059 *
5060 * This function can free multi-page allocations that are not compound
5061 * pages. It does not check that the @order passed in matches that of
5062 * the allocation, so it is easy to leak memory. Freeing more memory
5063 * than was allocated will probably emit a warning.
5064 *
5065 * If the last reference to this page is speculative, it will be released
5066 * by put_page() which only frees the first page of a non-compound
5067 * allocation. To prevent the remaining pages from being leaked, we free
5068 * the subsequent pages here. If you want to use the page's reference
5069 * count to decide when to free the allocation, you should allocate a
5070 * compound page, and use put_page() instead of __free_pages().
5071 *
5072 * Context: May be called in interrupt context or while holding a normal
5073 * spinlock, but not in NMI context or while holding a raw spinlock.
5074 */
5075 void __free_pages(struct page *page, unsigned int order)
5076 {
5077 if (put_page_testzero(page))
5078 free_the_page(page, order);
5079 else if (!PageHead(page))
5080 while (order-- > 0)
5081 free_the_page(page + (1 << order), order);
5082 }
5083 EXPORT_SYMBOL(__free_pages);
5084
5085 void free_pages(unsigned long addr, unsigned int order)
5086 {
5087 if (addr != 0) {
5088 VM_BUG_ON(!virt_addr_valid((void *)addr));
5089 __free_pages(virt_to_page((void *)addr), order);
5090 }
5091 }
5092
5093 EXPORT_SYMBOL(free_pages);
5094
5095 /*
5096 * Page Fragment:
5097 * An arbitrary-length arbitrary-offset area of memory which resides
5098 * within a 0 or higher order page. Multiple fragments within that page
5099 * are individually refcounted, in the page's reference counter.
5100 *
5101 * The page_frag functions below provide a simple allocation framework for
5102 * page fragments. This is used by the network stack and network device
5103 * drivers to provide a backing region of memory for use as either an
5104 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5105 */
5106 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5107 gfp_t gfp_mask)
5108 {
5109 struct page *page = NULL;
5110 gfp_t gfp = gfp_mask;
5111
5112 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5113 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5114 __GFP_NOMEMALLOC;
5115 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5116 PAGE_FRAG_CACHE_MAX_ORDER);
5117 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5118 #endif
5119 if (unlikely(!page))
5120 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5121
5122 nc->va = page ? page_address(page) : NULL;
5123
5124 return page;
5125 }
5126
5127 void __page_frag_cache_drain(struct page *page, unsigned int count)
5128 {
5129 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5130
5131 if (page_ref_sub_and_test(page, count))
5132 free_the_page(page, compound_order(page));
5133 }
5134 EXPORT_SYMBOL(__page_frag_cache_drain);
5135
5136 void *page_frag_alloc(struct page_frag_cache *nc,
5137 unsigned int fragsz, gfp_t gfp_mask)
5138 {
5139 unsigned int size = PAGE_SIZE;
5140 struct page *page;
5141 int offset;
5142
5143 if (unlikely(!nc->va)) {
5144 refill:
5145 page = __page_frag_cache_refill(nc, gfp_mask);
5146 if (!page)
5147 return NULL;
5148
5149 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5150 /* if size can vary use size else just use PAGE_SIZE */
5151 size = nc->size;
5152 #endif
5153 /* Even if we own the page, we do not use atomic_set().
5154 * This would break get_page_unless_zero() users.
5155 */
5156 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5157
5158 /* reset page count bias and offset to start of new frag */
5159 nc->pfmemalloc = page_is_pfmemalloc(page);
5160 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5161 nc->offset = size;
5162 }
5163
5164 offset = nc->offset - fragsz;
5165 if (unlikely(offset < 0)) {
5166 page = virt_to_page(nc->va);
5167
5168 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5169 goto refill;
5170
5171 if (unlikely(nc->pfmemalloc)) {
5172 free_the_page(page, compound_order(page));
5173 goto refill;
5174 }
5175
5176 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5177 /* if size can vary use size else just use PAGE_SIZE */
5178 size = nc->size;
5179 #endif
5180 /* OK, page count is 0, we can safely set it */
5181 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5182
5183 /* reset page count bias and offset to start of new frag */
5184 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5185 offset = size - fragsz;
5186 }
5187
5188 nc->pagecnt_bias--;
5189 nc->offset = offset;
5190
5191 return nc->va + offset;
5192 }
5193 EXPORT_SYMBOL(page_frag_alloc);
5194
5195 /*
5196 * Frees a page fragment allocated out of either a compound or order 0 page.
5197 */
5198 void page_frag_free(void *addr)
5199 {
5200 struct page *page = virt_to_head_page(addr);
5201
5202 if (unlikely(put_page_testzero(page)))
5203 free_the_page(page, compound_order(page));
5204 }
5205 EXPORT_SYMBOL(page_frag_free);
5206
5207 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5208 size_t size)
5209 {
5210 if (addr) {
5211 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5212 unsigned long used = addr + PAGE_ALIGN(size);
5213
5214 split_page(virt_to_page((void *)addr), order);
5215 while (used < alloc_end) {
5216 free_page(used);
5217 used += PAGE_SIZE;
5218 }
5219 }
5220 return (void *)addr;
5221 }
5222
5223 /**
5224 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5225 * @size: the number of bytes to allocate
5226 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5227 *
5228 * This function is similar to alloc_pages(), except that it allocates the
5229 * minimum number of pages to satisfy the request. alloc_pages() can only
5230 * allocate memory in power-of-two pages.
5231 *
5232 * This function is also limited by MAX_ORDER.
5233 *
5234 * Memory allocated by this function must be released by free_pages_exact().
5235 *
5236 * Return: pointer to the allocated area or %NULL in case of error.
5237 */
5238 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5239 {
5240 unsigned int order = get_order(size);
5241 unsigned long addr;
5242
5243 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5244 gfp_mask &= ~__GFP_COMP;
5245
5246 addr = __get_free_pages(gfp_mask, order);
5247 return make_alloc_exact(addr, order, size);
5248 }
5249 EXPORT_SYMBOL(alloc_pages_exact);
5250
5251 /**
5252 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5253 * pages on a node.
5254 * @nid: the preferred node ID where memory should be allocated
5255 * @size: the number of bytes to allocate
5256 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5257 *
5258 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5259 * back.
5260 *
5261 * Return: pointer to the allocated area or %NULL in case of error.
5262 */
5263 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5264 {
5265 unsigned int order = get_order(size);
5266 struct page *p;
5267
5268 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5269 gfp_mask &= ~__GFP_COMP;
5270
5271 p = alloc_pages_node(nid, gfp_mask, order);
5272 if (!p)
5273 return NULL;
5274 return make_alloc_exact((unsigned long)page_address(p), order, size);
5275 }
5276
5277 /**
5278 * free_pages_exact - release memory allocated via alloc_pages_exact()
5279 * @virt: the value returned by alloc_pages_exact.
5280 * @size: size of allocation, same value as passed to alloc_pages_exact().
5281 *
5282 * Release the memory allocated by a previous call to alloc_pages_exact.
5283 */
5284 void free_pages_exact(void *virt, size_t size)
5285 {
5286 unsigned long addr = (unsigned long)virt;
5287 unsigned long end = addr + PAGE_ALIGN(size);
5288
5289 while (addr < end) {
5290 free_page(addr);
5291 addr += PAGE_SIZE;
5292 }
5293 }
5294 EXPORT_SYMBOL(free_pages_exact);
5295
5296 /**
5297 * nr_free_zone_pages - count number of pages beyond high watermark
5298 * @offset: The zone index of the highest zone
5299 *
5300 * nr_free_zone_pages() counts the number of pages which are beyond the
5301 * high watermark within all zones at or below a given zone index. For each
5302 * zone, the number of pages is calculated as:
5303 *
5304 * nr_free_zone_pages = managed_pages - high_pages
5305 *
5306 * Return: number of pages beyond high watermark.
5307 */
5308 static unsigned long nr_free_zone_pages(int offset)
5309 {
5310 struct zoneref *z;
5311 struct zone *zone;
5312
5313 /* Just pick one node, since fallback list is circular */
5314 unsigned long sum = 0;
5315
5316 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5317
5318 for_each_zone_zonelist(zone, z, zonelist, offset) {
5319 unsigned long size = zone_managed_pages(zone);
5320 unsigned long high = high_wmark_pages(zone);
5321 if (size > high)
5322 sum += size - high;
5323 }
5324
5325 return sum;
5326 }
5327
5328 /**
5329 * nr_free_buffer_pages - count number of pages beyond high watermark
5330 *
5331 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5332 * watermark within ZONE_DMA and ZONE_NORMAL.
5333 *
5334 * Return: number of pages beyond high watermark within ZONE_DMA and
5335 * ZONE_NORMAL.
5336 */
5337 unsigned long nr_free_buffer_pages(void)
5338 {
5339 return nr_free_zone_pages(gfp_zone(GFP_USER));
5340 }
5341 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5342
5343 static inline void show_node(struct zone *zone)
5344 {
5345 if (IS_ENABLED(CONFIG_NUMA))
5346 printk("Node %d ", zone_to_nid(zone));
5347 }
5348
5349 long si_mem_available(void)
5350 {
5351 long available;
5352 unsigned long pagecache;
5353 unsigned long wmark_low = 0;
5354 unsigned long pages[NR_LRU_LISTS];
5355 unsigned long reclaimable;
5356 struct zone *zone;
5357 int lru;
5358
5359 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5360 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5361
5362 for_each_zone(zone)
5363 wmark_low += low_wmark_pages(zone);
5364
5365 /*
5366 * Estimate the amount of memory available for userspace allocations,
5367 * without causing swapping.
5368 */
5369 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5370
5371 /*
5372 * Not all the page cache can be freed, otherwise the system will
5373 * start swapping. Assume at least half of the page cache, or the
5374 * low watermark worth of cache, needs to stay.
5375 */
5376 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5377 pagecache -= min(pagecache / 2, wmark_low);
5378 available += pagecache;
5379
5380 /*
5381 * Part of the reclaimable slab and other kernel memory consists of
5382 * items that are in use, and cannot be freed. Cap this estimate at the
5383 * low watermark.
5384 */
5385 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5386 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5387 available += reclaimable - min(reclaimable / 2, wmark_low);
5388
5389 if (available < 0)
5390 available = 0;
5391 return available;
5392 }
5393 EXPORT_SYMBOL_GPL(si_mem_available);
5394
5395 void si_meminfo(struct sysinfo *val)
5396 {
5397 val->totalram = totalram_pages();
5398 val->sharedram = global_node_page_state(NR_SHMEM);
5399 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5400 val->bufferram = nr_blockdev_pages();
5401 val->totalhigh = totalhigh_pages();
5402 val->freehigh = nr_free_highpages();
5403 val->mem_unit = PAGE_SIZE;
5404 }
5405
5406 EXPORT_SYMBOL(si_meminfo);
5407
5408 #ifdef CONFIG_NUMA
5409 void si_meminfo_node(struct sysinfo *val, int nid)
5410 {
5411 int zone_type; /* needs to be signed */
5412 unsigned long managed_pages = 0;
5413 unsigned long managed_highpages = 0;
5414 unsigned long free_highpages = 0;
5415 pg_data_t *pgdat = NODE_DATA(nid);
5416
5417 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5418 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5419 val->totalram = managed_pages;
5420 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5421 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5422 #ifdef CONFIG_HIGHMEM
5423 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5424 struct zone *zone = &pgdat->node_zones[zone_type];
5425
5426 if (is_highmem(zone)) {
5427 managed_highpages += zone_managed_pages(zone);
5428 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5429 }
5430 }
5431 val->totalhigh = managed_highpages;
5432 val->freehigh = free_highpages;
5433 #else
5434 val->totalhigh = managed_highpages;
5435 val->freehigh = free_highpages;
5436 #endif
5437 val->mem_unit = PAGE_SIZE;
5438 }
5439 #endif
5440
5441 /*
5442 * Determine whether the node should be displayed or not, depending on whether
5443 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5444 */
5445 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5446 {
5447 if (!(flags & SHOW_MEM_FILTER_NODES))
5448 return false;
5449
5450 /*
5451 * no node mask - aka implicit memory numa policy. Do not bother with
5452 * the synchronization - read_mems_allowed_begin - because we do not
5453 * have to be precise here.
5454 */
5455 if (!nodemask)
5456 nodemask = &cpuset_current_mems_allowed;
5457
5458 return !node_isset(nid, *nodemask);
5459 }
5460
5461 #define K(x) ((x) << (PAGE_SHIFT-10))
5462
5463 static void show_migration_types(unsigned char type)
5464 {
5465 static const char types[MIGRATE_TYPES] = {
5466 [MIGRATE_UNMOVABLE] = 'U',
5467 [MIGRATE_MOVABLE] = 'M',
5468 [MIGRATE_RECLAIMABLE] = 'E',
5469 [MIGRATE_HIGHATOMIC] = 'H',
5470 #ifdef CONFIG_CMA
5471 [MIGRATE_CMA] = 'C',
5472 #endif
5473 #ifdef CONFIG_MEMORY_ISOLATION
5474 [MIGRATE_ISOLATE] = 'I',
5475 #endif
5476 };
5477 char tmp[MIGRATE_TYPES + 1];
5478 char *p = tmp;
5479 int i;
5480
5481 for (i = 0; i < MIGRATE_TYPES; i++) {
5482 if (type & (1 << i))
5483 *p++ = types[i];
5484 }
5485
5486 *p = '\0';
5487 printk(KERN_CONT "(%s) ", tmp);
5488 }
5489
5490 /*
5491 * Show free area list (used inside shift_scroll-lock stuff)
5492 * We also calculate the percentage fragmentation. We do this by counting the
5493 * memory on each free list with the exception of the first item on the list.
5494 *
5495 * Bits in @filter:
5496 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5497 * cpuset.
5498 */
5499 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5500 {
5501 unsigned long free_pcp = 0;
5502 int cpu;
5503 struct zone *zone;
5504 pg_data_t *pgdat;
5505
5506 for_each_populated_zone(zone) {
5507 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5508 continue;
5509
5510 for_each_online_cpu(cpu)
5511 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5512 }
5513
5514 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5515 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5516 " unevictable:%lu dirty:%lu writeback:%lu\n"
5517 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5518 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5519 " free:%lu free_pcp:%lu free_cma:%lu\n",
5520 global_node_page_state(NR_ACTIVE_ANON),
5521 global_node_page_state(NR_INACTIVE_ANON),
5522 global_node_page_state(NR_ISOLATED_ANON),
5523 global_node_page_state(NR_ACTIVE_FILE),
5524 global_node_page_state(NR_INACTIVE_FILE),
5525 global_node_page_state(NR_ISOLATED_FILE),
5526 global_node_page_state(NR_UNEVICTABLE),
5527 global_node_page_state(NR_FILE_DIRTY),
5528 global_node_page_state(NR_WRITEBACK),
5529 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5530 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5531 global_node_page_state(NR_FILE_MAPPED),
5532 global_node_page_state(NR_SHMEM),
5533 global_node_page_state(NR_PAGETABLE),
5534 global_zone_page_state(NR_BOUNCE),
5535 global_zone_page_state(NR_FREE_PAGES),
5536 free_pcp,
5537 global_zone_page_state(NR_FREE_CMA_PAGES));
5538
5539 for_each_online_pgdat(pgdat) {
5540 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5541 continue;
5542
5543 printk("Node %d"
5544 " active_anon:%lukB"
5545 " inactive_anon:%lukB"
5546 " active_file:%lukB"
5547 " inactive_file:%lukB"
5548 " unevictable:%lukB"
5549 " isolated(anon):%lukB"
5550 " isolated(file):%lukB"
5551 " mapped:%lukB"
5552 " dirty:%lukB"
5553 " writeback:%lukB"
5554 " shmem:%lukB"
5555 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5556 " shmem_thp: %lukB"
5557 " shmem_pmdmapped: %lukB"
5558 " anon_thp: %lukB"
5559 #endif
5560 " writeback_tmp:%lukB"
5561 " kernel_stack:%lukB"
5562 #ifdef CONFIG_SHADOW_CALL_STACK
5563 " shadow_call_stack:%lukB"
5564 #endif
5565 " pagetables:%lukB"
5566 " all_unreclaimable? %s"
5567 "\n",
5568 pgdat->node_id,
5569 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5570 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5571 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5572 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5573 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5574 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5575 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5576 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5577 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5578 K(node_page_state(pgdat, NR_WRITEBACK)),
5579 K(node_page_state(pgdat, NR_SHMEM)),
5580 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5581 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5582 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5583 * HPAGE_PMD_NR),
5584 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5585 #endif
5586 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5587 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5588 #ifdef CONFIG_SHADOW_CALL_STACK
5589 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5590 #endif
5591 K(node_page_state(pgdat, NR_PAGETABLE)),
5592 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5593 "yes" : "no");
5594 }
5595
5596 for_each_populated_zone(zone) {
5597 int i;
5598
5599 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5600 continue;
5601
5602 free_pcp = 0;
5603 for_each_online_cpu(cpu)
5604 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5605
5606 show_node(zone);
5607 printk(KERN_CONT
5608 "%s"
5609 " free:%lukB"
5610 " min:%lukB"
5611 " low:%lukB"
5612 " high:%lukB"
5613 " reserved_highatomic:%luKB"
5614 " active_anon:%lukB"
5615 " inactive_anon:%lukB"
5616 " active_file:%lukB"
5617 " inactive_file:%lukB"
5618 " unevictable:%lukB"
5619 " writepending:%lukB"
5620 " present:%lukB"
5621 " managed:%lukB"
5622 " mlocked:%lukB"
5623 " bounce:%lukB"
5624 " free_pcp:%lukB"
5625 " local_pcp:%ukB"
5626 " free_cma:%lukB"
5627 "\n",
5628 zone->name,
5629 K(zone_page_state(zone, NR_FREE_PAGES)),
5630 K(min_wmark_pages(zone)),
5631 K(low_wmark_pages(zone)),
5632 K(high_wmark_pages(zone)),
5633 K(zone->nr_reserved_highatomic),
5634 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5635 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5636 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5637 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5638 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5639 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5640 K(zone->present_pages),
5641 K(zone_managed_pages(zone)),
5642 K(zone_page_state(zone, NR_MLOCK)),
5643 K(zone_page_state(zone, NR_BOUNCE)),
5644 K(free_pcp),
5645 K(this_cpu_read(zone->pageset->pcp.count)),
5646 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5647 printk("lowmem_reserve[]:");
5648 for (i = 0; i < MAX_NR_ZONES; i++)
5649 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5650 printk(KERN_CONT "\n");
5651 }
5652
5653 for_each_populated_zone(zone) {
5654 unsigned int order;
5655 unsigned long nr[MAX_ORDER], flags, total = 0;
5656 unsigned char types[MAX_ORDER];
5657
5658 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5659 continue;
5660 show_node(zone);
5661 printk(KERN_CONT "%s: ", zone->name);
5662
5663 spin_lock_irqsave(&zone->lock, flags);
5664 for (order = 0; order < MAX_ORDER; order++) {
5665 struct free_area *area = &zone->free_area[order];
5666 int type;
5667
5668 nr[order] = area->nr_free;
5669 total += nr[order] << order;
5670
5671 types[order] = 0;
5672 for (type = 0; type < MIGRATE_TYPES; type++) {
5673 if (!free_area_empty(area, type))
5674 types[order] |= 1 << type;
5675 }
5676 }
5677 spin_unlock_irqrestore(&zone->lock, flags);
5678 for (order = 0; order < MAX_ORDER; order++) {
5679 printk(KERN_CONT "%lu*%lukB ",
5680 nr[order], K(1UL) << order);
5681 if (nr[order])
5682 show_migration_types(types[order]);
5683 }
5684 printk(KERN_CONT "= %lukB\n", K(total));
5685 }
5686
5687 hugetlb_show_meminfo();
5688
5689 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5690
5691 show_swap_cache_info();
5692 }
5693
5694 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5695 {
5696 zoneref->zone = zone;
5697 zoneref->zone_idx = zone_idx(zone);
5698 }
5699
5700 /*
5701 * Builds allocation fallback zone lists.
5702 *
5703 * Add all populated zones of a node to the zonelist.
5704 */
5705 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5706 {
5707 struct zone *zone;
5708 enum zone_type zone_type = MAX_NR_ZONES;
5709 int nr_zones = 0;
5710
5711 do {
5712 zone_type--;
5713 zone = pgdat->node_zones + zone_type;
5714 if (managed_zone(zone)) {
5715 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5716 check_highest_zone(zone_type);
5717 }
5718 } while (zone_type);
5719
5720 return nr_zones;
5721 }
5722
5723 #ifdef CONFIG_NUMA
5724
5725 static int __parse_numa_zonelist_order(char *s)
5726 {
5727 /*
5728 * We used to support different zonlists modes but they turned
5729 * out to be just not useful. Let's keep the warning in place
5730 * if somebody still use the cmd line parameter so that we do
5731 * not fail it silently
5732 */
5733 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5734 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5735 return -EINVAL;
5736 }
5737 return 0;
5738 }
5739
5740 char numa_zonelist_order[] = "Node";
5741
5742 /*
5743 * sysctl handler for numa_zonelist_order
5744 */
5745 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5746 void *buffer, size_t *length, loff_t *ppos)
5747 {
5748 if (write)
5749 return __parse_numa_zonelist_order(buffer);
5750 return proc_dostring(table, write, buffer, length, ppos);
5751 }
5752
5753
5754 #define MAX_NODE_LOAD (nr_online_nodes)
5755 static int node_load[MAX_NUMNODES];
5756
5757 /**
5758 * find_next_best_node - find the next node that should appear in a given node's fallback list
5759 * @node: node whose fallback list we're appending
5760 * @used_node_mask: nodemask_t of already used nodes
5761 *
5762 * We use a number of factors to determine which is the next node that should
5763 * appear on a given node's fallback list. The node should not have appeared
5764 * already in @node's fallback list, and it should be the next closest node
5765 * according to the distance array (which contains arbitrary distance values
5766 * from each node to each node in the system), and should also prefer nodes
5767 * with no CPUs, since presumably they'll have very little allocation pressure
5768 * on them otherwise.
5769 *
5770 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5771 */
5772 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5773 {
5774 int n, val;
5775 int min_val = INT_MAX;
5776 int best_node = NUMA_NO_NODE;
5777
5778 /* Use the local node if we haven't already */
5779 if (!node_isset(node, *used_node_mask)) {
5780 node_set(node, *used_node_mask);
5781 return node;
5782 }
5783
5784 for_each_node_state(n, N_MEMORY) {
5785
5786 /* Don't want a node to appear more than once */
5787 if (node_isset(n, *used_node_mask))
5788 continue;
5789
5790 /* Use the distance array to find the distance */
5791 val = node_distance(node, n);
5792
5793 /* Penalize nodes under us ("prefer the next node") */
5794 val += (n < node);
5795
5796 /* Give preference to headless and unused nodes */
5797 if (!cpumask_empty(cpumask_of_node(n)))
5798 val += PENALTY_FOR_NODE_WITH_CPUS;
5799
5800 /* Slight preference for less loaded node */
5801 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5802 val += node_load[n];
5803
5804 if (val < min_val) {
5805 min_val = val;
5806 best_node = n;
5807 }
5808 }
5809
5810 if (best_node >= 0)
5811 node_set(best_node, *used_node_mask);
5812
5813 return best_node;
5814 }
5815
5816
5817 /*
5818 * Build zonelists ordered by node and zones within node.
5819 * This results in maximum locality--normal zone overflows into local
5820 * DMA zone, if any--but risks exhausting DMA zone.
5821 */
5822 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5823 unsigned nr_nodes)
5824 {
5825 struct zoneref *zonerefs;
5826 int i;
5827
5828 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5829
5830 for (i = 0; i < nr_nodes; i++) {
5831 int nr_zones;
5832
5833 pg_data_t *node = NODE_DATA(node_order[i]);
5834
5835 nr_zones = build_zonerefs_node(node, zonerefs);
5836 zonerefs += nr_zones;
5837 }
5838 zonerefs->zone = NULL;
5839 zonerefs->zone_idx = 0;
5840 }
5841
5842 /*
5843 * Build gfp_thisnode zonelists
5844 */
5845 static void build_thisnode_zonelists(pg_data_t *pgdat)
5846 {
5847 struct zoneref *zonerefs;
5848 int nr_zones;
5849
5850 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5851 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5852 zonerefs += nr_zones;
5853 zonerefs->zone = NULL;
5854 zonerefs->zone_idx = 0;
5855 }
5856
5857 /*
5858 * Build zonelists ordered by zone and nodes within zones.
5859 * This results in conserving DMA zone[s] until all Normal memory is
5860 * exhausted, but results in overflowing to remote node while memory
5861 * may still exist in local DMA zone.
5862 */
5863
5864 static void build_zonelists(pg_data_t *pgdat)
5865 {
5866 static int node_order[MAX_NUMNODES];
5867 int node, load, nr_nodes = 0;
5868 nodemask_t used_mask = NODE_MASK_NONE;
5869 int local_node, prev_node;
5870
5871 /* NUMA-aware ordering of nodes */
5872 local_node = pgdat->node_id;
5873 load = nr_online_nodes;
5874 prev_node = local_node;
5875
5876 memset(node_order, 0, sizeof(node_order));
5877 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5878 /*
5879 * We don't want to pressure a particular node.
5880 * So adding penalty to the first node in same
5881 * distance group to make it round-robin.
5882 */
5883 if (node_distance(local_node, node) !=
5884 node_distance(local_node, prev_node))
5885 node_load[node] = load;
5886
5887 node_order[nr_nodes++] = node;
5888 prev_node = node;
5889 load--;
5890 }
5891
5892 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5893 build_thisnode_zonelists(pgdat);
5894 }
5895
5896 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5897 /*
5898 * Return node id of node used for "local" allocations.
5899 * I.e., first node id of first zone in arg node's generic zonelist.
5900 * Used for initializing percpu 'numa_mem', which is used primarily
5901 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5902 */
5903 int local_memory_node(int node)
5904 {
5905 struct zoneref *z;
5906
5907 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5908 gfp_zone(GFP_KERNEL),
5909 NULL);
5910 return zone_to_nid(z->zone);
5911 }
5912 #endif
5913
5914 static void setup_min_unmapped_ratio(void);
5915 static void setup_min_slab_ratio(void);
5916 #else /* CONFIG_NUMA */
5917
5918 static void build_zonelists(pg_data_t *pgdat)
5919 {
5920 int node, local_node;
5921 struct zoneref *zonerefs;
5922 int nr_zones;
5923
5924 local_node = pgdat->node_id;
5925
5926 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5927 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5928 zonerefs += nr_zones;
5929
5930 /*
5931 * Now we build the zonelist so that it contains the zones
5932 * of all the other nodes.
5933 * We don't want to pressure a particular node, so when
5934 * building the zones for node N, we make sure that the
5935 * zones coming right after the local ones are those from
5936 * node N+1 (modulo N)
5937 */
5938 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5939 if (!node_online(node))
5940 continue;
5941 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5942 zonerefs += nr_zones;
5943 }
5944 for (node = 0; node < local_node; node++) {
5945 if (!node_online(node))
5946 continue;
5947 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5948 zonerefs += nr_zones;
5949 }
5950
5951 zonerefs->zone = NULL;
5952 zonerefs->zone_idx = 0;
5953 }
5954
5955 #endif /* CONFIG_NUMA */
5956
5957 /*
5958 * Boot pageset table. One per cpu which is going to be used for all
5959 * zones and all nodes. The parameters will be set in such a way
5960 * that an item put on a list will immediately be handed over to
5961 * the buddy list. This is safe since pageset manipulation is done
5962 * with interrupts disabled.
5963 *
5964 * The boot_pagesets must be kept even after bootup is complete for
5965 * unused processors and/or zones. They do play a role for bootstrapping
5966 * hotplugged processors.
5967 *
5968 * zoneinfo_show() and maybe other functions do
5969 * not check if the processor is online before following the pageset pointer.
5970 * Other parts of the kernel may not check if the zone is available.
5971 */
5972 static void pageset_init(struct per_cpu_pageset *p);
5973 /* These effectively disable the pcplists in the boot pageset completely */
5974 #define BOOT_PAGESET_HIGH 0
5975 #define BOOT_PAGESET_BATCH 1
5976 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5977 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5978
5979 static void __build_all_zonelists(void *data)
5980 {
5981 int nid;
5982 int __maybe_unused cpu;
5983 pg_data_t *self = data;
5984 static DEFINE_SPINLOCK(lock);
5985
5986 spin_lock(&lock);
5987
5988 #ifdef CONFIG_NUMA
5989 memset(node_load, 0, sizeof(node_load));
5990 #endif
5991
5992 /*
5993 * This node is hotadded and no memory is yet present. So just
5994 * building zonelists is fine - no need to touch other nodes.
5995 */
5996 if (self && !node_online(self->node_id)) {
5997 build_zonelists(self);
5998 } else {
5999 for_each_online_node(nid) {
6000 pg_data_t *pgdat = NODE_DATA(nid);
6001
6002 build_zonelists(pgdat);
6003 }
6004
6005 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6006 /*
6007 * We now know the "local memory node" for each node--
6008 * i.e., the node of the first zone in the generic zonelist.
6009 * Set up numa_mem percpu variable for on-line cpus. During
6010 * boot, only the boot cpu should be on-line; we'll init the
6011 * secondary cpus' numa_mem as they come on-line. During
6012 * node/memory hotplug, we'll fixup all on-line cpus.
6013 */
6014 for_each_online_cpu(cpu)
6015 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6016 #endif
6017 }
6018
6019 spin_unlock(&lock);
6020 }
6021
6022 static noinline void __init
6023 build_all_zonelists_init(void)
6024 {
6025 int cpu;
6026
6027 __build_all_zonelists(NULL);
6028
6029 /*
6030 * Initialize the boot_pagesets that are going to be used
6031 * for bootstrapping processors. The real pagesets for
6032 * each zone will be allocated later when the per cpu
6033 * allocator is available.
6034 *
6035 * boot_pagesets are used also for bootstrapping offline
6036 * cpus if the system is already booted because the pagesets
6037 * are needed to initialize allocators on a specific cpu too.
6038 * F.e. the percpu allocator needs the page allocator which
6039 * needs the percpu allocator in order to allocate its pagesets
6040 * (a chicken-egg dilemma).
6041 */
6042 for_each_possible_cpu(cpu)
6043 pageset_init(&per_cpu(boot_pageset, cpu));
6044
6045 mminit_verify_zonelist();
6046 cpuset_init_current_mems_allowed();
6047 }
6048
6049 /*
6050 * unless system_state == SYSTEM_BOOTING.
6051 *
6052 * __ref due to call of __init annotated helper build_all_zonelists_init
6053 * [protected by SYSTEM_BOOTING].
6054 */
6055 void __ref build_all_zonelists(pg_data_t *pgdat)
6056 {
6057 unsigned long vm_total_pages;
6058
6059 if (system_state == SYSTEM_BOOTING) {
6060 build_all_zonelists_init();
6061 } else {
6062 __build_all_zonelists(pgdat);
6063 /* cpuset refresh routine should be here */
6064 }
6065 /* Get the number of free pages beyond high watermark in all zones. */
6066 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6067 /*
6068 * Disable grouping by mobility if the number of pages in the
6069 * system is too low to allow the mechanism to work. It would be
6070 * more accurate, but expensive to check per-zone. This check is
6071 * made on memory-hotadd so a system can start with mobility
6072 * disabled and enable it later
6073 */
6074 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6075 page_group_by_mobility_disabled = 1;
6076 else
6077 page_group_by_mobility_disabled = 0;
6078
6079 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6080 nr_online_nodes,
6081 page_group_by_mobility_disabled ? "off" : "on",
6082 vm_total_pages);
6083 #ifdef CONFIG_NUMA
6084 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6085 #endif
6086 }
6087
6088 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6089 static bool __meminit
6090 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6091 {
6092 static struct memblock_region *r;
6093
6094 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6095 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6096 for_each_mem_region(r) {
6097 if (*pfn < memblock_region_memory_end_pfn(r))
6098 break;
6099 }
6100 }
6101 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6102 memblock_is_mirror(r)) {
6103 *pfn = memblock_region_memory_end_pfn(r);
6104 return true;
6105 }
6106 }
6107 return false;
6108 }
6109
6110 /*
6111 * Initially all pages are reserved - free ones are freed
6112 * up by memblock_free_all() once the early boot process is
6113 * done. Non-atomic initialization, single-pass.
6114 *
6115 * All aligned pageblocks are initialized to the specified migratetype
6116 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6117 * zone stats (e.g., nr_isolate_pageblock) are touched.
6118 */
6119 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6120 unsigned long start_pfn, unsigned long zone_end_pfn,
6121 enum meminit_context context,
6122 struct vmem_altmap *altmap, int migratetype)
6123 {
6124 unsigned long pfn, end_pfn = start_pfn + size;
6125 struct page *page;
6126
6127 if (highest_memmap_pfn < end_pfn - 1)
6128 highest_memmap_pfn = end_pfn - 1;
6129
6130 #ifdef CONFIG_ZONE_DEVICE
6131 /*
6132 * Honor reservation requested by the driver for this ZONE_DEVICE
6133 * memory. We limit the total number of pages to initialize to just
6134 * those that might contain the memory mapping. We will defer the
6135 * ZONE_DEVICE page initialization until after we have released
6136 * the hotplug lock.
6137 */
6138 if (zone == ZONE_DEVICE) {
6139 if (!altmap)
6140 return;
6141
6142 if (start_pfn == altmap->base_pfn)
6143 start_pfn += altmap->reserve;
6144 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6145 }
6146 #endif
6147
6148 for (pfn = start_pfn; pfn < end_pfn; ) {
6149 /*
6150 * There can be holes in boot-time mem_map[]s handed to this
6151 * function. They do not exist on hotplugged memory.
6152 */
6153 if (context == MEMINIT_EARLY) {
6154 if (overlap_memmap_init(zone, &pfn))
6155 continue;
6156 if (defer_init(nid, pfn, zone_end_pfn))
6157 break;
6158 }
6159
6160 page = pfn_to_page(pfn);
6161 __init_single_page(page, pfn, zone, nid);
6162 if (context == MEMINIT_HOTPLUG)
6163 __SetPageReserved(page);
6164
6165 /*
6166 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6167 * such that unmovable allocations won't be scattered all
6168 * over the place during system boot.
6169 */
6170 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6171 set_pageblock_migratetype(page, migratetype);
6172 cond_resched();
6173 }
6174 pfn++;
6175 }
6176 }
6177
6178 #ifdef CONFIG_ZONE_DEVICE
6179 void __ref memmap_init_zone_device(struct zone *zone,
6180 unsigned long start_pfn,
6181 unsigned long nr_pages,
6182 struct dev_pagemap *pgmap)
6183 {
6184 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6185 struct pglist_data *pgdat = zone->zone_pgdat;
6186 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6187 unsigned long zone_idx = zone_idx(zone);
6188 unsigned long start = jiffies;
6189 int nid = pgdat->node_id;
6190
6191 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6192 return;
6193
6194 /*
6195 * The call to memmap_init_zone should have already taken care
6196 * of the pages reserved for the memmap, so we can just jump to
6197 * the end of that region and start processing the device pages.
6198 */
6199 if (altmap) {
6200 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6201 nr_pages = end_pfn - start_pfn;
6202 }
6203
6204 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6205 struct page *page = pfn_to_page(pfn);
6206
6207 __init_single_page(page, pfn, zone_idx, nid);
6208
6209 /*
6210 * Mark page reserved as it will need to wait for onlining
6211 * phase for it to be fully associated with a zone.
6212 *
6213 * We can use the non-atomic __set_bit operation for setting
6214 * the flag as we are still initializing the pages.
6215 */
6216 __SetPageReserved(page);
6217
6218 /*
6219 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6220 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6221 * ever freed or placed on a driver-private list.
6222 */
6223 page->pgmap = pgmap;
6224 page->zone_device_data = NULL;
6225
6226 /*
6227 * Mark the block movable so that blocks are reserved for
6228 * movable at startup. This will force kernel allocations
6229 * to reserve their blocks rather than leaking throughout
6230 * the address space during boot when many long-lived
6231 * kernel allocations are made.
6232 *
6233 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6234 * because this is done early in section_activate()
6235 */
6236 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6237 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6238 cond_resched();
6239 }
6240 }
6241
6242 pr_info("%s initialised %lu pages in %ums\n", __func__,
6243 nr_pages, jiffies_to_msecs(jiffies - start));
6244 }
6245
6246 #endif
6247 static void __meminit zone_init_free_lists(struct zone *zone)
6248 {
6249 unsigned int order, t;
6250 for_each_migratetype_order(order, t) {
6251 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6252 zone->free_area[order].nr_free = 0;
6253 }
6254 }
6255
6256 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6257 /*
6258 * Only struct pages that correspond to ranges defined by memblock.memory
6259 * are zeroed and initialized by going through __init_single_page() during
6260 * memmap_init_zone().
6261 *
6262 * But, there could be struct pages that correspond to holes in
6263 * memblock.memory. This can happen because of the following reasons:
6264 * - physical memory bank size is not necessarily the exact multiple of the
6265 * arbitrary section size
6266 * - early reserved memory may not be listed in memblock.memory
6267 * - memory layouts defined with memmap= kernel parameter may not align
6268 * nicely with memmap sections
6269 *
6270 * Explicitly initialize those struct pages so that:
6271 * - PG_Reserved is set
6272 * - zone and node links point to zone and node that span the page if the
6273 * hole is in the middle of a zone
6274 * - zone and node links point to adjacent zone/node if the hole falls on
6275 * the zone boundary; the pages in such holes will be prepended to the
6276 * zone/node above the hole except for the trailing pages in the last
6277 * section that will be appended to the zone/node below.
6278 */
6279 static u64 __meminit init_unavailable_range(unsigned long spfn,
6280 unsigned long epfn,
6281 int zone, int node)
6282 {
6283 unsigned long pfn;
6284 u64 pgcnt = 0;
6285
6286 for (pfn = spfn; pfn < epfn; pfn++) {
6287 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6288 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6289 + pageblock_nr_pages - 1;
6290 continue;
6291 }
6292 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6293 __SetPageReserved(pfn_to_page(pfn));
6294 pgcnt++;
6295 }
6296
6297 return pgcnt;
6298 }
6299 #else
6300 static inline u64 init_unavailable_range(unsigned long spfn, unsigned long epfn,
6301 int zone, int node)
6302 {
6303 return 0;
6304 }
6305 #endif
6306
6307 void __meminit __weak memmap_init(unsigned long size, int nid,
6308 unsigned long zone,
6309 unsigned long range_start_pfn)
6310 {
6311 static unsigned long hole_pfn;
6312 unsigned long start_pfn, end_pfn;
6313 unsigned long range_end_pfn = range_start_pfn + size;
6314 int i;
6315 u64 pgcnt = 0;
6316
6317 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6318 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6319 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6320
6321 if (end_pfn > start_pfn) {
6322 size = end_pfn - start_pfn;
6323 memmap_init_zone(size, nid, zone, start_pfn, range_end_pfn,
6324 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6325 }
6326
6327 if (hole_pfn < start_pfn)
6328 pgcnt += init_unavailable_range(hole_pfn, start_pfn,
6329 zone, nid);
6330 hole_pfn = end_pfn;
6331 }
6332
6333 #ifdef CONFIG_SPARSEMEM
6334 /*
6335 * Initialize the hole in the range [zone_end_pfn, section_end].
6336 * If zone boundary falls in the middle of a section, this hole
6337 * will be re-initialized during the call to this function for the
6338 * higher zone.
6339 */
6340 end_pfn = round_up(range_end_pfn, PAGES_PER_SECTION);
6341 if (hole_pfn < end_pfn)
6342 pgcnt += init_unavailable_range(hole_pfn, end_pfn,
6343 zone, nid);
6344 #endif
6345
6346 if (pgcnt)
6347 pr_info(" %s zone: %llu pages in unavailable ranges\n",
6348 zone_names[zone], pgcnt);
6349 }
6350
6351 static int zone_batchsize(struct zone *zone)
6352 {
6353 #ifdef CONFIG_MMU
6354 int batch;
6355
6356 /*
6357 * The per-cpu-pages pools are set to around 1000th of the
6358 * size of the zone.
6359 */
6360 batch = zone_managed_pages(zone) / 1024;
6361 /* But no more than a meg. */
6362 if (batch * PAGE_SIZE > 1024 * 1024)
6363 batch = (1024 * 1024) / PAGE_SIZE;
6364 batch /= 4; /* We effectively *= 4 below */
6365 if (batch < 1)
6366 batch = 1;
6367
6368 /*
6369 * Clamp the batch to a 2^n - 1 value. Having a power
6370 * of 2 value was found to be more likely to have
6371 * suboptimal cache aliasing properties in some cases.
6372 *
6373 * For example if 2 tasks are alternately allocating
6374 * batches of pages, one task can end up with a lot
6375 * of pages of one half of the possible page colors
6376 * and the other with pages of the other colors.
6377 */
6378 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6379
6380 return batch;
6381
6382 #else
6383 /* The deferral and batching of frees should be suppressed under NOMMU
6384 * conditions.
6385 *
6386 * The problem is that NOMMU needs to be able to allocate large chunks
6387 * of contiguous memory as there's no hardware page translation to
6388 * assemble apparent contiguous memory from discontiguous pages.
6389 *
6390 * Queueing large contiguous runs of pages for batching, however,
6391 * causes the pages to actually be freed in smaller chunks. As there
6392 * can be a significant delay between the individual batches being
6393 * recycled, this leads to the once large chunks of space being
6394 * fragmented and becoming unavailable for high-order allocations.
6395 */
6396 return 0;
6397 #endif
6398 }
6399
6400 /*
6401 * pcp->high and pcp->batch values are related and generally batch is lower
6402 * than high. They are also related to pcp->count such that count is lower
6403 * than high, and as soon as it reaches high, the pcplist is flushed.
6404 *
6405 * However, guaranteeing these relations at all times would require e.g. write
6406 * barriers here but also careful usage of read barriers at the read side, and
6407 * thus be prone to error and bad for performance. Thus the update only prevents
6408 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6409 * can cope with those fields changing asynchronously, and fully trust only the
6410 * pcp->count field on the local CPU with interrupts disabled.
6411 *
6412 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6413 * outside of boot time (or some other assurance that no concurrent updaters
6414 * exist).
6415 */
6416 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6417 unsigned long batch)
6418 {
6419 WRITE_ONCE(pcp->batch, batch);
6420 WRITE_ONCE(pcp->high, high);
6421 }
6422
6423 static void pageset_init(struct per_cpu_pageset *p)
6424 {
6425 struct per_cpu_pages *pcp;
6426 int migratetype;
6427
6428 memset(p, 0, sizeof(*p));
6429
6430 pcp = &p->pcp;
6431 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6432 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6433
6434 /*
6435 * Set batch and high values safe for a boot pageset. A true percpu
6436 * pageset's initialization will update them subsequently. Here we don't
6437 * need to be as careful as pageset_update() as nobody can access the
6438 * pageset yet.
6439 */
6440 pcp->high = BOOT_PAGESET_HIGH;
6441 pcp->batch = BOOT_PAGESET_BATCH;
6442 }
6443
6444 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6445 unsigned long batch)
6446 {
6447 struct per_cpu_pageset *p;
6448 int cpu;
6449
6450 for_each_possible_cpu(cpu) {
6451 p = per_cpu_ptr(zone->pageset, cpu);
6452 pageset_update(&p->pcp, high, batch);
6453 }
6454 }
6455
6456 /*
6457 * Calculate and set new high and batch values for all per-cpu pagesets of a
6458 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6459 */
6460 static void zone_set_pageset_high_and_batch(struct zone *zone)
6461 {
6462 unsigned long new_high, new_batch;
6463
6464 if (percpu_pagelist_fraction) {
6465 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6466 new_batch = max(1UL, new_high / 4);
6467 if ((new_high / 4) > (PAGE_SHIFT * 8))
6468 new_batch = PAGE_SHIFT * 8;
6469 } else {
6470 new_batch = zone_batchsize(zone);
6471 new_high = 6 * new_batch;
6472 new_batch = max(1UL, 1 * new_batch);
6473 }
6474
6475 if (zone->pageset_high == new_high &&
6476 zone->pageset_batch == new_batch)
6477 return;
6478
6479 zone->pageset_high = new_high;
6480 zone->pageset_batch = new_batch;
6481
6482 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6483 }
6484
6485 void __meminit setup_zone_pageset(struct zone *zone)
6486 {
6487 struct per_cpu_pageset *p;
6488 int cpu;
6489
6490 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6491 for_each_possible_cpu(cpu) {
6492 p = per_cpu_ptr(zone->pageset, cpu);
6493 pageset_init(p);
6494 }
6495
6496 zone_set_pageset_high_and_batch(zone);
6497 }
6498
6499 /*
6500 * Allocate per cpu pagesets and initialize them.
6501 * Before this call only boot pagesets were available.
6502 */
6503 void __init setup_per_cpu_pageset(void)
6504 {
6505 struct pglist_data *pgdat;
6506 struct zone *zone;
6507 int __maybe_unused cpu;
6508
6509 for_each_populated_zone(zone)
6510 setup_zone_pageset(zone);
6511
6512 #ifdef CONFIG_NUMA
6513 /*
6514 * Unpopulated zones continue using the boot pagesets.
6515 * The numa stats for these pagesets need to be reset.
6516 * Otherwise, they will end up skewing the stats of
6517 * the nodes these zones are associated with.
6518 */
6519 for_each_possible_cpu(cpu) {
6520 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6521 memset(pcp->vm_numa_stat_diff, 0,
6522 sizeof(pcp->vm_numa_stat_diff));
6523 }
6524 #endif
6525
6526 for_each_online_pgdat(pgdat)
6527 pgdat->per_cpu_nodestats =
6528 alloc_percpu(struct per_cpu_nodestat);
6529 }
6530
6531 static __meminit void zone_pcp_init(struct zone *zone)
6532 {
6533 /*
6534 * per cpu subsystem is not up at this point. The following code
6535 * relies on the ability of the linker to provide the
6536 * offset of a (static) per cpu variable into the per cpu area.
6537 */
6538 zone->pageset = &boot_pageset;
6539 zone->pageset_high = BOOT_PAGESET_HIGH;
6540 zone->pageset_batch = BOOT_PAGESET_BATCH;
6541
6542 if (populated_zone(zone))
6543 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6544 zone->name, zone->present_pages,
6545 zone_batchsize(zone));
6546 }
6547
6548 void __meminit init_currently_empty_zone(struct zone *zone,
6549 unsigned long zone_start_pfn,
6550 unsigned long size)
6551 {
6552 struct pglist_data *pgdat = zone->zone_pgdat;
6553 int zone_idx = zone_idx(zone) + 1;
6554
6555 if (zone_idx > pgdat->nr_zones)
6556 pgdat->nr_zones = zone_idx;
6557
6558 zone->zone_start_pfn = zone_start_pfn;
6559
6560 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6561 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6562 pgdat->node_id,
6563 (unsigned long)zone_idx(zone),
6564 zone_start_pfn, (zone_start_pfn + size));
6565
6566 zone_init_free_lists(zone);
6567 zone->initialized = 1;
6568 }
6569
6570 /**
6571 * get_pfn_range_for_nid - Return the start and end page frames for a node
6572 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6573 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6574 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6575 *
6576 * It returns the start and end page frame of a node based on information
6577 * provided by memblock_set_node(). If called for a node
6578 * with no available memory, a warning is printed and the start and end
6579 * PFNs will be 0.
6580 */
6581 void __init get_pfn_range_for_nid(unsigned int nid,
6582 unsigned long *start_pfn, unsigned long *end_pfn)
6583 {
6584 unsigned long this_start_pfn, this_end_pfn;
6585 int i;
6586
6587 *start_pfn = -1UL;
6588 *end_pfn = 0;
6589
6590 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6591 *start_pfn = min(*start_pfn, this_start_pfn);
6592 *end_pfn = max(*end_pfn, this_end_pfn);
6593 }
6594
6595 if (*start_pfn == -1UL)
6596 *start_pfn = 0;
6597 }
6598
6599 /*
6600 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6601 * assumption is made that zones within a node are ordered in monotonic
6602 * increasing memory addresses so that the "highest" populated zone is used
6603 */
6604 static void __init find_usable_zone_for_movable(void)
6605 {
6606 int zone_index;
6607 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6608 if (zone_index == ZONE_MOVABLE)
6609 continue;
6610
6611 if (arch_zone_highest_possible_pfn[zone_index] >
6612 arch_zone_lowest_possible_pfn[zone_index])
6613 break;
6614 }
6615
6616 VM_BUG_ON(zone_index == -1);
6617 movable_zone = zone_index;
6618 }
6619
6620 /*
6621 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6622 * because it is sized independent of architecture. Unlike the other zones,
6623 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6624 * in each node depending on the size of each node and how evenly kernelcore
6625 * is distributed. This helper function adjusts the zone ranges
6626 * provided by the architecture for a given node by using the end of the
6627 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6628 * zones within a node are in order of monotonic increases memory addresses
6629 */
6630 static void __init adjust_zone_range_for_zone_movable(int nid,
6631 unsigned long zone_type,
6632 unsigned long node_start_pfn,
6633 unsigned long node_end_pfn,
6634 unsigned long *zone_start_pfn,
6635 unsigned long *zone_end_pfn)
6636 {
6637 /* Only adjust if ZONE_MOVABLE is on this node */
6638 if (zone_movable_pfn[nid]) {
6639 /* Size ZONE_MOVABLE */
6640 if (zone_type == ZONE_MOVABLE) {
6641 *zone_start_pfn = zone_movable_pfn[nid];
6642 *zone_end_pfn = min(node_end_pfn,
6643 arch_zone_highest_possible_pfn[movable_zone]);
6644
6645 /* Adjust for ZONE_MOVABLE starting within this range */
6646 } else if (!mirrored_kernelcore &&
6647 *zone_start_pfn < zone_movable_pfn[nid] &&
6648 *zone_end_pfn > zone_movable_pfn[nid]) {
6649 *zone_end_pfn = zone_movable_pfn[nid];
6650
6651 /* Check if this whole range is within ZONE_MOVABLE */
6652 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6653 *zone_start_pfn = *zone_end_pfn;
6654 }
6655 }
6656
6657 /*
6658 * Return the number of pages a zone spans in a node, including holes
6659 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6660 */
6661 static unsigned long __init zone_spanned_pages_in_node(int nid,
6662 unsigned long zone_type,
6663 unsigned long node_start_pfn,
6664 unsigned long node_end_pfn,
6665 unsigned long *zone_start_pfn,
6666 unsigned long *zone_end_pfn)
6667 {
6668 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6669 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6670 /* When hotadd a new node from cpu_up(), the node should be empty */
6671 if (!node_start_pfn && !node_end_pfn)
6672 return 0;
6673
6674 /* Get the start and end of the zone */
6675 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6676 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6677 adjust_zone_range_for_zone_movable(nid, zone_type,
6678 node_start_pfn, node_end_pfn,
6679 zone_start_pfn, zone_end_pfn);
6680
6681 /* Check that this node has pages within the zone's required range */
6682 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6683 return 0;
6684
6685 /* Move the zone boundaries inside the node if necessary */
6686 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6687 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6688
6689 /* Return the spanned pages */
6690 return *zone_end_pfn - *zone_start_pfn;
6691 }
6692
6693 /*
6694 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6695 * then all holes in the requested range will be accounted for.
6696 */
6697 unsigned long __init __absent_pages_in_range(int nid,
6698 unsigned long range_start_pfn,
6699 unsigned long range_end_pfn)
6700 {
6701 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6702 unsigned long start_pfn, end_pfn;
6703 int i;
6704
6705 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6706 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6707 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6708 nr_absent -= end_pfn - start_pfn;
6709 }
6710 return nr_absent;
6711 }
6712
6713 /**
6714 * absent_pages_in_range - Return number of page frames in holes within a range
6715 * @start_pfn: The start PFN to start searching for holes
6716 * @end_pfn: The end PFN to stop searching for holes
6717 *
6718 * Return: the number of pages frames in memory holes within a range.
6719 */
6720 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6721 unsigned long end_pfn)
6722 {
6723 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6724 }
6725
6726 /* Return the number of page frames in holes in a zone on a node */
6727 static unsigned long __init zone_absent_pages_in_node(int nid,
6728 unsigned long zone_type,
6729 unsigned long node_start_pfn,
6730 unsigned long node_end_pfn)
6731 {
6732 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6733 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6734 unsigned long zone_start_pfn, zone_end_pfn;
6735 unsigned long nr_absent;
6736
6737 /* When hotadd a new node from cpu_up(), the node should be empty */
6738 if (!node_start_pfn && !node_end_pfn)
6739 return 0;
6740
6741 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6742 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6743
6744 adjust_zone_range_for_zone_movable(nid, zone_type,
6745 node_start_pfn, node_end_pfn,
6746 &zone_start_pfn, &zone_end_pfn);
6747 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6748
6749 /*
6750 * ZONE_MOVABLE handling.
6751 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6752 * and vice versa.
6753 */
6754 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6755 unsigned long start_pfn, end_pfn;
6756 struct memblock_region *r;
6757
6758 for_each_mem_region(r) {
6759 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6760 zone_start_pfn, zone_end_pfn);
6761 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6762 zone_start_pfn, zone_end_pfn);
6763
6764 if (zone_type == ZONE_MOVABLE &&
6765 memblock_is_mirror(r))
6766 nr_absent += end_pfn - start_pfn;
6767
6768 if (zone_type == ZONE_NORMAL &&
6769 !memblock_is_mirror(r))
6770 nr_absent += end_pfn - start_pfn;
6771 }
6772 }
6773
6774 return nr_absent;
6775 }
6776
6777 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6778 unsigned long node_start_pfn,
6779 unsigned long node_end_pfn)
6780 {
6781 unsigned long realtotalpages = 0, totalpages = 0;
6782 enum zone_type i;
6783
6784 for (i = 0; i < MAX_NR_ZONES; i++) {
6785 struct zone *zone = pgdat->node_zones + i;
6786 unsigned long zone_start_pfn, zone_end_pfn;
6787 unsigned long spanned, absent;
6788 unsigned long size, real_size;
6789
6790 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6791 node_start_pfn,
6792 node_end_pfn,
6793 &zone_start_pfn,
6794 &zone_end_pfn);
6795 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6796 node_start_pfn,
6797 node_end_pfn);
6798
6799 size = spanned;
6800 real_size = size - absent;
6801
6802 if (size)
6803 zone->zone_start_pfn = zone_start_pfn;
6804 else
6805 zone->zone_start_pfn = 0;
6806 zone->spanned_pages = size;
6807 zone->present_pages = real_size;
6808
6809 totalpages += size;
6810 realtotalpages += real_size;
6811 }
6812
6813 pgdat->node_spanned_pages = totalpages;
6814 pgdat->node_present_pages = realtotalpages;
6815 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6816 realtotalpages);
6817 }
6818
6819 #ifndef CONFIG_SPARSEMEM
6820 /*
6821 * Calculate the size of the zone->blockflags rounded to an unsigned long
6822 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6823 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6824 * round what is now in bits to nearest long in bits, then return it in
6825 * bytes.
6826 */
6827 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6828 {
6829 unsigned long usemapsize;
6830
6831 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6832 usemapsize = roundup(zonesize, pageblock_nr_pages);
6833 usemapsize = usemapsize >> pageblock_order;
6834 usemapsize *= NR_PAGEBLOCK_BITS;
6835 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6836
6837 return usemapsize / 8;
6838 }
6839
6840 static void __ref setup_usemap(struct pglist_data *pgdat,
6841 struct zone *zone,
6842 unsigned long zone_start_pfn,
6843 unsigned long zonesize)
6844 {
6845 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6846 zone->pageblock_flags = NULL;
6847 if (usemapsize) {
6848 zone->pageblock_flags =
6849 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6850 pgdat->node_id);
6851 if (!zone->pageblock_flags)
6852 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6853 usemapsize, zone->name, pgdat->node_id);
6854 }
6855 }
6856 #else
6857 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6858 unsigned long zone_start_pfn, unsigned long zonesize) {}
6859 #endif /* CONFIG_SPARSEMEM */
6860
6861 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6862
6863 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6864 void __init set_pageblock_order(void)
6865 {
6866 unsigned int order;
6867
6868 /* Check that pageblock_nr_pages has not already been setup */
6869 if (pageblock_order)
6870 return;
6871
6872 if (HPAGE_SHIFT > PAGE_SHIFT)
6873 order = HUGETLB_PAGE_ORDER;
6874 else
6875 order = MAX_ORDER - 1;
6876
6877 /*
6878 * Assume the largest contiguous order of interest is a huge page.
6879 * This value may be variable depending on boot parameters on IA64 and
6880 * powerpc.
6881 */
6882 pageblock_order = order;
6883 }
6884 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6885
6886 /*
6887 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6888 * is unused as pageblock_order is set at compile-time. See
6889 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6890 * the kernel config
6891 */
6892 void __init set_pageblock_order(void)
6893 {
6894 }
6895
6896 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6897
6898 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6899 unsigned long present_pages)
6900 {
6901 unsigned long pages = spanned_pages;
6902
6903 /*
6904 * Provide a more accurate estimation if there are holes within
6905 * the zone and SPARSEMEM is in use. If there are holes within the
6906 * zone, each populated memory region may cost us one or two extra
6907 * memmap pages due to alignment because memmap pages for each
6908 * populated regions may not be naturally aligned on page boundary.
6909 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6910 */
6911 if (spanned_pages > present_pages + (present_pages >> 4) &&
6912 IS_ENABLED(CONFIG_SPARSEMEM))
6913 pages = present_pages;
6914
6915 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6916 }
6917
6918 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6919 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6920 {
6921 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6922
6923 spin_lock_init(&ds_queue->split_queue_lock);
6924 INIT_LIST_HEAD(&ds_queue->split_queue);
6925 ds_queue->split_queue_len = 0;
6926 }
6927 #else
6928 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6929 #endif
6930
6931 #ifdef CONFIG_COMPACTION
6932 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6933 {
6934 init_waitqueue_head(&pgdat->kcompactd_wait);
6935 }
6936 #else
6937 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6938 #endif
6939
6940 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6941 {
6942 pgdat_resize_init(pgdat);
6943
6944 pgdat_init_split_queue(pgdat);
6945 pgdat_init_kcompactd(pgdat);
6946
6947 init_waitqueue_head(&pgdat->kswapd_wait);
6948 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6949
6950 pgdat_page_ext_init(pgdat);
6951 lruvec_init(&pgdat->__lruvec);
6952 }
6953
6954 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6955 unsigned long remaining_pages)
6956 {
6957 atomic_long_set(&zone->managed_pages, remaining_pages);
6958 zone_set_nid(zone, nid);
6959 zone->name = zone_names[idx];
6960 zone->zone_pgdat = NODE_DATA(nid);
6961 spin_lock_init(&zone->lock);
6962 zone_seqlock_init(zone);
6963 zone_pcp_init(zone);
6964 }
6965
6966 /*
6967 * Set up the zone data structures
6968 * - init pgdat internals
6969 * - init all zones belonging to this node
6970 *
6971 * NOTE: this function is only called during memory hotplug
6972 */
6973 #ifdef CONFIG_MEMORY_HOTPLUG
6974 void __ref free_area_init_core_hotplug(int nid)
6975 {
6976 enum zone_type z;
6977 pg_data_t *pgdat = NODE_DATA(nid);
6978
6979 pgdat_init_internals(pgdat);
6980 for (z = 0; z < MAX_NR_ZONES; z++)
6981 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6982 }
6983 #endif
6984
6985 /*
6986 * Set up the zone data structures:
6987 * - mark all pages reserved
6988 * - mark all memory queues empty
6989 * - clear the memory bitmaps
6990 *
6991 * NOTE: pgdat should get zeroed by caller.
6992 * NOTE: this function is only called during early init.
6993 */
6994 static void __init free_area_init_core(struct pglist_data *pgdat)
6995 {
6996 enum zone_type j;
6997 int nid = pgdat->node_id;
6998
6999 pgdat_init_internals(pgdat);
7000 pgdat->per_cpu_nodestats = &boot_nodestats;
7001
7002 for (j = 0; j < MAX_NR_ZONES; j++) {
7003 struct zone *zone = pgdat->node_zones + j;
7004 unsigned long size, freesize, memmap_pages;
7005 unsigned long zone_start_pfn = zone->zone_start_pfn;
7006
7007 size = zone->spanned_pages;
7008 freesize = zone->present_pages;
7009
7010 /*
7011 * Adjust freesize so that it accounts for how much memory
7012 * is used by this zone for memmap. This affects the watermark
7013 * and per-cpu initialisations
7014 */
7015 memmap_pages = calc_memmap_size(size, freesize);
7016 if (!is_highmem_idx(j)) {
7017 if (freesize >= memmap_pages) {
7018 freesize -= memmap_pages;
7019 if (memmap_pages)
7020 printk(KERN_DEBUG
7021 " %s zone: %lu pages used for memmap\n",
7022 zone_names[j], memmap_pages);
7023 } else
7024 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7025 zone_names[j], memmap_pages, freesize);
7026 }
7027
7028 /* Account for reserved pages */
7029 if (j == 0 && freesize > dma_reserve) {
7030 freesize -= dma_reserve;
7031 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7032 zone_names[0], dma_reserve);
7033 }
7034
7035 if (!is_highmem_idx(j))
7036 nr_kernel_pages += freesize;
7037 /* Charge for highmem memmap if there are enough kernel pages */
7038 else if (nr_kernel_pages > memmap_pages * 2)
7039 nr_kernel_pages -= memmap_pages;
7040 nr_all_pages += freesize;
7041
7042 /*
7043 * Set an approximate value for lowmem here, it will be adjusted
7044 * when the bootmem allocator frees pages into the buddy system.
7045 * And all highmem pages will be managed by the buddy system.
7046 */
7047 zone_init_internals(zone, j, nid, freesize);
7048
7049 if (!size)
7050 continue;
7051
7052 set_pageblock_order();
7053 setup_usemap(pgdat, zone, zone_start_pfn, size);
7054 init_currently_empty_zone(zone, zone_start_pfn, size);
7055 memmap_init(size, nid, j, zone_start_pfn);
7056 }
7057 }
7058
7059 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7060 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7061 {
7062 unsigned long __maybe_unused start = 0;
7063 unsigned long __maybe_unused offset = 0;
7064
7065 /* Skip empty nodes */
7066 if (!pgdat->node_spanned_pages)
7067 return;
7068
7069 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7070 offset = pgdat->node_start_pfn - start;
7071 /* ia64 gets its own node_mem_map, before this, without bootmem */
7072 if (!pgdat->node_mem_map) {
7073 unsigned long size, end;
7074 struct page *map;
7075
7076 /*
7077 * The zone's endpoints aren't required to be MAX_ORDER
7078 * aligned but the node_mem_map endpoints must be in order
7079 * for the buddy allocator to function correctly.
7080 */
7081 end = pgdat_end_pfn(pgdat);
7082 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7083 size = (end - start) * sizeof(struct page);
7084 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7085 pgdat->node_id);
7086 if (!map)
7087 panic("Failed to allocate %ld bytes for node %d memory map\n",
7088 size, pgdat->node_id);
7089 pgdat->node_mem_map = map + offset;
7090 }
7091 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7092 __func__, pgdat->node_id, (unsigned long)pgdat,
7093 (unsigned long)pgdat->node_mem_map);
7094 #ifndef CONFIG_NEED_MULTIPLE_NODES
7095 /*
7096 * With no DISCONTIG, the global mem_map is just set as node 0's
7097 */
7098 if (pgdat == NODE_DATA(0)) {
7099 mem_map = NODE_DATA(0)->node_mem_map;
7100 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7101 mem_map -= offset;
7102 }
7103 #endif
7104 }
7105 #else
7106 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7107 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7108
7109 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7110 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7111 {
7112 pgdat->first_deferred_pfn = ULONG_MAX;
7113 }
7114 #else
7115 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7116 #endif
7117
7118 static void __init free_area_init_node(int nid)
7119 {
7120 pg_data_t *pgdat = NODE_DATA(nid);
7121 unsigned long start_pfn = 0;
7122 unsigned long end_pfn = 0;
7123
7124 /* pg_data_t should be reset to zero when it's allocated */
7125 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7126
7127 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7128
7129 pgdat->node_id = nid;
7130 pgdat->node_start_pfn = start_pfn;
7131 pgdat->per_cpu_nodestats = NULL;
7132
7133 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7134 (u64)start_pfn << PAGE_SHIFT,
7135 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7136 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7137
7138 alloc_node_mem_map(pgdat);
7139 pgdat_set_deferred_range(pgdat);
7140
7141 free_area_init_core(pgdat);
7142 }
7143
7144 void __init free_area_init_memoryless_node(int nid)
7145 {
7146 free_area_init_node(nid);
7147 }
7148
7149 #if MAX_NUMNODES > 1
7150 /*
7151 * Figure out the number of possible node ids.
7152 */
7153 void __init setup_nr_node_ids(void)
7154 {
7155 unsigned int highest;
7156
7157 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7158 nr_node_ids = highest + 1;
7159 }
7160 #endif
7161
7162 /**
7163 * node_map_pfn_alignment - determine the maximum internode alignment
7164 *
7165 * This function should be called after node map is populated and sorted.
7166 * It calculates the maximum power of two alignment which can distinguish
7167 * all the nodes.
7168 *
7169 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7170 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7171 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7172 * shifted, 1GiB is enough and this function will indicate so.
7173 *
7174 * This is used to test whether pfn -> nid mapping of the chosen memory
7175 * model has fine enough granularity to avoid incorrect mapping for the
7176 * populated node map.
7177 *
7178 * Return: the determined alignment in pfn's. 0 if there is no alignment
7179 * requirement (single node).
7180 */
7181 unsigned long __init node_map_pfn_alignment(void)
7182 {
7183 unsigned long accl_mask = 0, last_end = 0;
7184 unsigned long start, end, mask;
7185 int last_nid = NUMA_NO_NODE;
7186 int i, nid;
7187
7188 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7189 if (!start || last_nid < 0 || last_nid == nid) {
7190 last_nid = nid;
7191 last_end = end;
7192 continue;
7193 }
7194
7195 /*
7196 * Start with a mask granular enough to pin-point to the
7197 * start pfn and tick off bits one-by-one until it becomes
7198 * too coarse to separate the current node from the last.
7199 */
7200 mask = ~((1 << __ffs(start)) - 1);
7201 while (mask && last_end <= (start & (mask << 1)))
7202 mask <<= 1;
7203
7204 /* accumulate all internode masks */
7205 accl_mask |= mask;
7206 }
7207
7208 /* convert mask to number of pages */
7209 return ~accl_mask + 1;
7210 }
7211
7212 /**
7213 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7214 *
7215 * Return: the minimum PFN based on information provided via
7216 * memblock_set_node().
7217 */
7218 unsigned long __init find_min_pfn_with_active_regions(void)
7219 {
7220 return PHYS_PFN(memblock_start_of_DRAM());
7221 }
7222
7223 /*
7224 * early_calculate_totalpages()
7225 * Sum pages in active regions for movable zone.
7226 * Populate N_MEMORY for calculating usable_nodes.
7227 */
7228 static unsigned long __init early_calculate_totalpages(void)
7229 {
7230 unsigned long totalpages = 0;
7231 unsigned long start_pfn, end_pfn;
7232 int i, nid;
7233
7234 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7235 unsigned long pages = end_pfn - start_pfn;
7236
7237 totalpages += pages;
7238 if (pages)
7239 node_set_state(nid, N_MEMORY);
7240 }
7241 return totalpages;
7242 }
7243
7244 /*
7245 * Find the PFN the Movable zone begins in each node. Kernel memory
7246 * is spread evenly between nodes as long as the nodes have enough
7247 * memory. When they don't, some nodes will have more kernelcore than
7248 * others
7249 */
7250 static void __init find_zone_movable_pfns_for_nodes(void)
7251 {
7252 int i, nid;
7253 unsigned long usable_startpfn;
7254 unsigned long kernelcore_node, kernelcore_remaining;
7255 /* save the state before borrow the nodemask */
7256 nodemask_t saved_node_state = node_states[N_MEMORY];
7257 unsigned long totalpages = early_calculate_totalpages();
7258 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7259 struct memblock_region *r;
7260
7261 /* Need to find movable_zone earlier when movable_node is specified. */
7262 find_usable_zone_for_movable();
7263
7264 /*
7265 * If movable_node is specified, ignore kernelcore and movablecore
7266 * options.
7267 */
7268 if (movable_node_is_enabled()) {
7269 for_each_mem_region(r) {
7270 if (!memblock_is_hotpluggable(r))
7271 continue;
7272
7273 nid = memblock_get_region_node(r);
7274
7275 usable_startpfn = PFN_DOWN(r->base);
7276 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7277 min(usable_startpfn, zone_movable_pfn[nid]) :
7278 usable_startpfn;
7279 }
7280
7281 goto out2;
7282 }
7283
7284 /*
7285 * If kernelcore=mirror is specified, ignore movablecore option
7286 */
7287 if (mirrored_kernelcore) {
7288 bool mem_below_4gb_not_mirrored = false;
7289
7290 for_each_mem_region(r) {
7291 if (memblock_is_mirror(r))
7292 continue;
7293
7294 nid = memblock_get_region_node(r);
7295
7296 usable_startpfn = memblock_region_memory_base_pfn(r);
7297
7298 if (usable_startpfn < 0x100000) {
7299 mem_below_4gb_not_mirrored = true;
7300 continue;
7301 }
7302
7303 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7304 min(usable_startpfn, zone_movable_pfn[nid]) :
7305 usable_startpfn;
7306 }
7307
7308 if (mem_below_4gb_not_mirrored)
7309 pr_warn("This configuration results in unmirrored kernel memory.\n");
7310
7311 goto out2;
7312 }
7313
7314 /*
7315 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7316 * amount of necessary memory.
7317 */
7318 if (required_kernelcore_percent)
7319 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7320 10000UL;
7321 if (required_movablecore_percent)
7322 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7323 10000UL;
7324
7325 /*
7326 * If movablecore= was specified, calculate what size of
7327 * kernelcore that corresponds so that memory usable for
7328 * any allocation type is evenly spread. If both kernelcore
7329 * and movablecore are specified, then the value of kernelcore
7330 * will be used for required_kernelcore if it's greater than
7331 * what movablecore would have allowed.
7332 */
7333 if (required_movablecore) {
7334 unsigned long corepages;
7335
7336 /*
7337 * Round-up so that ZONE_MOVABLE is at least as large as what
7338 * was requested by the user
7339 */
7340 required_movablecore =
7341 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7342 required_movablecore = min(totalpages, required_movablecore);
7343 corepages = totalpages - required_movablecore;
7344
7345 required_kernelcore = max(required_kernelcore, corepages);
7346 }
7347
7348 /*
7349 * If kernelcore was not specified or kernelcore size is larger
7350 * than totalpages, there is no ZONE_MOVABLE.
7351 */
7352 if (!required_kernelcore || required_kernelcore >= totalpages)
7353 goto out;
7354
7355 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7356 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7357
7358 restart:
7359 /* Spread kernelcore memory as evenly as possible throughout nodes */
7360 kernelcore_node = required_kernelcore / usable_nodes;
7361 for_each_node_state(nid, N_MEMORY) {
7362 unsigned long start_pfn, end_pfn;
7363
7364 /*
7365 * Recalculate kernelcore_node if the division per node
7366 * now exceeds what is necessary to satisfy the requested
7367 * amount of memory for the kernel
7368 */
7369 if (required_kernelcore < kernelcore_node)
7370 kernelcore_node = required_kernelcore / usable_nodes;
7371
7372 /*
7373 * As the map is walked, we track how much memory is usable
7374 * by the kernel using kernelcore_remaining. When it is
7375 * 0, the rest of the node is usable by ZONE_MOVABLE
7376 */
7377 kernelcore_remaining = kernelcore_node;
7378
7379 /* Go through each range of PFNs within this node */
7380 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7381 unsigned long size_pages;
7382
7383 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7384 if (start_pfn >= end_pfn)
7385 continue;
7386
7387 /* Account for what is only usable for kernelcore */
7388 if (start_pfn < usable_startpfn) {
7389 unsigned long kernel_pages;
7390 kernel_pages = min(end_pfn, usable_startpfn)
7391 - start_pfn;
7392
7393 kernelcore_remaining -= min(kernel_pages,
7394 kernelcore_remaining);
7395 required_kernelcore -= min(kernel_pages,
7396 required_kernelcore);
7397
7398 /* Continue if range is now fully accounted */
7399 if (end_pfn <= usable_startpfn) {
7400
7401 /*
7402 * Push zone_movable_pfn to the end so
7403 * that if we have to rebalance
7404 * kernelcore across nodes, we will
7405 * not double account here
7406 */
7407 zone_movable_pfn[nid] = end_pfn;
7408 continue;
7409 }
7410 start_pfn = usable_startpfn;
7411 }
7412
7413 /*
7414 * The usable PFN range for ZONE_MOVABLE is from
7415 * start_pfn->end_pfn. Calculate size_pages as the
7416 * number of pages used as kernelcore
7417 */
7418 size_pages = end_pfn - start_pfn;
7419 if (size_pages > kernelcore_remaining)
7420 size_pages = kernelcore_remaining;
7421 zone_movable_pfn[nid] = start_pfn + size_pages;
7422
7423 /*
7424 * Some kernelcore has been met, update counts and
7425 * break if the kernelcore for this node has been
7426 * satisfied
7427 */
7428 required_kernelcore -= min(required_kernelcore,
7429 size_pages);
7430 kernelcore_remaining -= size_pages;
7431 if (!kernelcore_remaining)
7432 break;
7433 }
7434 }
7435
7436 /*
7437 * If there is still required_kernelcore, we do another pass with one
7438 * less node in the count. This will push zone_movable_pfn[nid] further
7439 * along on the nodes that still have memory until kernelcore is
7440 * satisfied
7441 */
7442 usable_nodes--;
7443 if (usable_nodes && required_kernelcore > usable_nodes)
7444 goto restart;
7445
7446 out2:
7447 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7448 for (nid = 0; nid < MAX_NUMNODES; nid++)
7449 zone_movable_pfn[nid] =
7450 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7451
7452 out:
7453 /* restore the node_state */
7454 node_states[N_MEMORY] = saved_node_state;
7455 }
7456
7457 /* Any regular or high memory on that node ? */
7458 static void check_for_memory(pg_data_t *pgdat, int nid)
7459 {
7460 enum zone_type zone_type;
7461
7462 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7463 struct zone *zone = &pgdat->node_zones[zone_type];
7464 if (populated_zone(zone)) {
7465 if (IS_ENABLED(CONFIG_HIGHMEM))
7466 node_set_state(nid, N_HIGH_MEMORY);
7467 if (zone_type <= ZONE_NORMAL)
7468 node_set_state(nid, N_NORMAL_MEMORY);
7469 break;
7470 }
7471 }
7472 }
7473
7474 /*
7475 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7476 * such cases we allow max_zone_pfn sorted in the descending order
7477 */
7478 bool __weak arch_has_descending_max_zone_pfns(void)
7479 {
7480 return false;
7481 }
7482
7483 /**
7484 * free_area_init - Initialise all pg_data_t and zone data
7485 * @max_zone_pfn: an array of max PFNs for each zone
7486 *
7487 * This will call free_area_init_node() for each active node in the system.
7488 * Using the page ranges provided by memblock_set_node(), the size of each
7489 * zone in each node and their holes is calculated. If the maximum PFN
7490 * between two adjacent zones match, it is assumed that the zone is empty.
7491 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7492 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7493 * starts where the previous one ended. For example, ZONE_DMA32 starts
7494 * at arch_max_dma_pfn.
7495 */
7496 void __init free_area_init(unsigned long *max_zone_pfn)
7497 {
7498 unsigned long start_pfn, end_pfn;
7499 int i, nid, zone;
7500 bool descending;
7501
7502 /* Record where the zone boundaries are */
7503 memset(arch_zone_lowest_possible_pfn, 0,
7504 sizeof(arch_zone_lowest_possible_pfn));
7505 memset(arch_zone_highest_possible_pfn, 0,
7506 sizeof(arch_zone_highest_possible_pfn));
7507
7508 start_pfn = find_min_pfn_with_active_regions();
7509 descending = arch_has_descending_max_zone_pfns();
7510
7511 for (i = 0; i < MAX_NR_ZONES; i++) {
7512 if (descending)
7513 zone = MAX_NR_ZONES - i - 1;
7514 else
7515 zone = i;
7516
7517 if (zone == ZONE_MOVABLE)
7518 continue;
7519
7520 end_pfn = max(max_zone_pfn[zone], start_pfn);
7521 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7522 arch_zone_highest_possible_pfn[zone] = end_pfn;
7523
7524 start_pfn = end_pfn;
7525 }
7526
7527 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7528 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7529 find_zone_movable_pfns_for_nodes();
7530
7531 /* Print out the zone ranges */
7532 pr_info("Zone ranges:\n");
7533 for (i = 0; i < MAX_NR_ZONES; i++) {
7534 if (i == ZONE_MOVABLE)
7535 continue;
7536 pr_info(" %-8s ", zone_names[i]);
7537 if (arch_zone_lowest_possible_pfn[i] ==
7538 arch_zone_highest_possible_pfn[i])
7539 pr_cont("empty\n");
7540 else
7541 pr_cont("[mem %#018Lx-%#018Lx]\n",
7542 (u64)arch_zone_lowest_possible_pfn[i]
7543 << PAGE_SHIFT,
7544 ((u64)arch_zone_highest_possible_pfn[i]
7545 << PAGE_SHIFT) - 1);
7546 }
7547
7548 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7549 pr_info("Movable zone start for each node\n");
7550 for (i = 0; i < MAX_NUMNODES; i++) {
7551 if (zone_movable_pfn[i])
7552 pr_info(" Node %d: %#018Lx\n", i,
7553 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7554 }
7555
7556 /*
7557 * Print out the early node map, and initialize the
7558 * subsection-map relative to active online memory ranges to
7559 * enable future "sub-section" extensions of the memory map.
7560 */
7561 pr_info("Early memory node ranges\n");
7562 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7563 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7564 (u64)start_pfn << PAGE_SHIFT,
7565 ((u64)end_pfn << PAGE_SHIFT) - 1);
7566 subsection_map_init(start_pfn, end_pfn - start_pfn);
7567 }
7568
7569 /* Initialise every node */
7570 mminit_verify_pageflags_layout();
7571 setup_nr_node_ids();
7572 for_each_online_node(nid) {
7573 pg_data_t *pgdat = NODE_DATA(nid);
7574 free_area_init_node(nid);
7575
7576 /* Any memory on that node */
7577 if (pgdat->node_present_pages)
7578 node_set_state(nid, N_MEMORY);
7579 check_for_memory(pgdat, nid);
7580 }
7581 }
7582
7583 static int __init cmdline_parse_core(char *p, unsigned long *core,
7584 unsigned long *percent)
7585 {
7586 unsigned long long coremem;
7587 char *endptr;
7588
7589 if (!p)
7590 return -EINVAL;
7591
7592 /* Value may be a percentage of total memory, otherwise bytes */
7593 coremem = simple_strtoull(p, &endptr, 0);
7594 if (*endptr == '%') {
7595 /* Paranoid check for percent values greater than 100 */
7596 WARN_ON(coremem > 100);
7597
7598 *percent = coremem;
7599 } else {
7600 coremem = memparse(p, &p);
7601 /* Paranoid check that UL is enough for the coremem value */
7602 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7603
7604 *core = coremem >> PAGE_SHIFT;
7605 *percent = 0UL;
7606 }
7607 return 0;
7608 }
7609
7610 /*
7611 * kernelcore=size sets the amount of memory for use for allocations that
7612 * cannot be reclaimed or migrated.
7613 */
7614 static int __init cmdline_parse_kernelcore(char *p)
7615 {
7616 /* parse kernelcore=mirror */
7617 if (parse_option_str(p, "mirror")) {
7618 mirrored_kernelcore = true;
7619 return 0;
7620 }
7621
7622 return cmdline_parse_core(p, &required_kernelcore,
7623 &required_kernelcore_percent);
7624 }
7625
7626 /*
7627 * movablecore=size sets the amount of memory for use for allocations that
7628 * can be reclaimed or migrated.
7629 */
7630 static int __init cmdline_parse_movablecore(char *p)
7631 {
7632 return cmdline_parse_core(p, &required_movablecore,
7633 &required_movablecore_percent);
7634 }
7635
7636 early_param("kernelcore", cmdline_parse_kernelcore);
7637 early_param("movablecore", cmdline_parse_movablecore);
7638
7639 void adjust_managed_page_count(struct page *page, long count)
7640 {
7641 atomic_long_add(count, &page_zone(page)->managed_pages);
7642 totalram_pages_add(count);
7643 #ifdef CONFIG_HIGHMEM
7644 if (PageHighMem(page))
7645 totalhigh_pages_add(count);
7646 #endif
7647 }
7648 EXPORT_SYMBOL(adjust_managed_page_count);
7649
7650 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7651 {
7652 void *pos;
7653 unsigned long pages = 0;
7654
7655 start = (void *)PAGE_ALIGN((unsigned long)start);
7656 end = (void *)((unsigned long)end & PAGE_MASK);
7657 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7658 struct page *page = virt_to_page(pos);
7659 void *direct_map_addr;
7660
7661 /*
7662 * 'direct_map_addr' might be different from 'pos'
7663 * because some architectures' virt_to_page()
7664 * work with aliases. Getting the direct map
7665 * address ensures that we get a _writeable_
7666 * alias for the memset().
7667 */
7668 direct_map_addr = page_address(page);
7669 /*
7670 * Perform a kasan-unchecked memset() since this memory
7671 * has not been initialized.
7672 */
7673 direct_map_addr = kasan_reset_tag(direct_map_addr);
7674 if ((unsigned int)poison <= 0xFF)
7675 memset(direct_map_addr, poison, PAGE_SIZE);
7676
7677 free_reserved_page(page);
7678 }
7679
7680 if (pages && s)
7681 pr_info("Freeing %s memory: %ldK\n",
7682 s, pages << (PAGE_SHIFT - 10));
7683
7684 return pages;
7685 }
7686
7687 #ifdef CONFIG_HIGHMEM
7688 void free_highmem_page(struct page *page)
7689 {
7690 __free_reserved_page(page);
7691 totalram_pages_inc();
7692 atomic_long_inc(&page_zone(page)->managed_pages);
7693 totalhigh_pages_inc();
7694 }
7695 #endif
7696
7697
7698 void __init mem_init_print_info(const char *str)
7699 {
7700 unsigned long physpages, codesize, datasize, rosize, bss_size;
7701 unsigned long init_code_size, init_data_size;
7702
7703 physpages = get_num_physpages();
7704 codesize = _etext - _stext;
7705 datasize = _edata - _sdata;
7706 rosize = __end_rodata - __start_rodata;
7707 bss_size = __bss_stop - __bss_start;
7708 init_data_size = __init_end - __init_begin;
7709 init_code_size = _einittext - _sinittext;
7710
7711 /*
7712 * Detect special cases and adjust section sizes accordingly:
7713 * 1) .init.* may be embedded into .data sections
7714 * 2) .init.text.* may be out of [__init_begin, __init_end],
7715 * please refer to arch/tile/kernel/vmlinux.lds.S.
7716 * 3) .rodata.* may be embedded into .text or .data sections.
7717 */
7718 #define adj_init_size(start, end, size, pos, adj) \
7719 do { \
7720 if (start <= pos && pos < end && size > adj) \
7721 size -= adj; \
7722 } while (0)
7723
7724 adj_init_size(__init_begin, __init_end, init_data_size,
7725 _sinittext, init_code_size);
7726 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7727 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7728 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7729 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7730
7731 #undef adj_init_size
7732
7733 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7734 #ifdef CONFIG_HIGHMEM
7735 ", %luK highmem"
7736 #endif
7737 "%s%s)\n",
7738 nr_free_pages() << (PAGE_SHIFT - 10),
7739 physpages << (PAGE_SHIFT - 10),
7740 codesize >> 10, datasize >> 10, rosize >> 10,
7741 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7742 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7743 totalcma_pages << (PAGE_SHIFT - 10),
7744 #ifdef CONFIG_HIGHMEM
7745 totalhigh_pages() << (PAGE_SHIFT - 10),
7746 #endif
7747 str ? ", " : "", str ? str : "");
7748 }
7749
7750 /**
7751 * set_dma_reserve - set the specified number of pages reserved in the first zone
7752 * @new_dma_reserve: The number of pages to mark reserved
7753 *
7754 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7755 * In the DMA zone, a significant percentage may be consumed by kernel image
7756 * and other unfreeable allocations which can skew the watermarks badly. This
7757 * function may optionally be used to account for unfreeable pages in the
7758 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7759 * smaller per-cpu batchsize.
7760 */
7761 void __init set_dma_reserve(unsigned long new_dma_reserve)
7762 {
7763 dma_reserve = new_dma_reserve;
7764 }
7765
7766 static int page_alloc_cpu_dead(unsigned int cpu)
7767 {
7768
7769 lru_add_drain_cpu(cpu);
7770 drain_pages(cpu);
7771
7772 /*
7773 * Spill the event counters of the dead processor
7774 * into the current processors event counters.
7775 * This artificially elevates the count of the current
7776 * processor.
7777 */
7778 vm_events_fold_cpu(cpu);
7779
7780 /*
7781 * Zero the differential counters of the dead processor
7782 * so that the vm statistics are consistent.
7783 *
7784 * This is only okay since the processor is dead and cannot
7785 * race with what we are doing.
7786 */
7787 cpu_vm_stats_fold(cpu);
7788 return 0;
7789 }
7790
7791 #ifdef CONFIG_NUMA
7792 int hashdist = HASHDIST_DEFAULT;
7793
7794 static int __init set_hashdist(char *str)
7795 {
7796 if (!str)
7797 return 0;
7798 hashdist = simple_strtoul(str, &str, 0);
7799 return 1;
7800 }
7801 __setup("hashdist=", set_hashdist);
7802 #endif
7803
7804 void __init page_alloc_init(void)
7805 {
7806 int ret;
7807
7808 #ifdef CONFIG_NUMA
7809 if (num_node_state(N_MEMORY) == 1)
7810 hashdist = 0;
7811 #endif
7812
7813 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7814 "mm/page_alloc:dead", NULL,
7815 page_alloc_cpu_dead);
7816 WARN_ON(ret < 0);
7817 }
7818
7819 /*
7820 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7821 * or min_free_kbytes changes.
7822 */
7823 static void calculate_totalreserve_pages(void)
7824 {
7825 struct pglist_data *pgdat;
7826 unsigned long reserve_pages = 0;
7827 enum zone_type i, j;
7828
7829 for_each_online_pgdat(pgdat) {
7830
7831 pgdat->totalreserve_pages = 0;
7832
7833 for (i = 0; i < MAX_NR_ZONES; i++) {
7834 struct zone *zone = pgdat->node_zones + i;
7835 long max = 0;
7836 unsigned long managed_pages = zone_managed_pages(zone);
7837
7838 /* Find valid and maximum lowmem_reserve in the zone */
7839 for (j = i; j < MAX_NR_ZONES; j++) {
7840 if (zone->lowmem_reserve[j] > max)
7841 max = zone->lowmem_reserve[j];
7842 }
7843
7844 /* we treat the high watermark as reserved pages. */
7845 max += high_wmark_pages(zone);
7846
7847 if (max > managed_pages)
7848 max = managed_pages;
7849
7850 pgdat->totalreserve_pages += max;
7851
7852 reserve_pages += max;
7853 }
7854 }
7855 totalreserve_pages = reserve_pages;
7856 }
7857
7858 /*
7859 * setup_per_zone_lowmem_reserve - called whenever
7860 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7861 * has a correct pages reserved value, so an adequate number of
7862 * pages are left in the zone after a successful __alloc_pages().
7863 */
7864 static void setup_per_zone_lowmem_reserve(void)
7865 {
7866 struct pglist_data *pgdat;
7867 enum zone_type i, j;
7868
7869 for_each_online_pgdat(pgdat) {
7870 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7871 struct zone *zone = &pgdat->node_zones[i];
7872 int ratio = sysctl_lowmem_reserve_ratio[i];
7873 bool clear = !ratio || !zone_managed_pages(zone);
7874 unsigned long managed_pages = 0;
7875
7876 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7877 if (clear) {
7878 zone->lowmem_reserve[j] = 0;
7879 } else {
7880 struct zone *upper_zone = &pgdat->node_zones[j];
7881
7882 managed_pages += zone_managed_pages(upper_zone);
7883 zone->lowmem_reserve[j] = managed_pages / ratio;
7884 }
7885 }
7886 }
7887 }
7888
7889 /* update totalreserve_pages */
7890 calculate_totalreserve_pages();
7891 }
7892
7893 static void __setup_per_zone_wmarks(void)
7894 {
7895 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7896 unsigned long lowmem_pages = 0;
7897 struct zone *zone;
7898 unsigned long flags;
7899
7900 /* Calculate total number of !ZONE_HIGHMEM pages */
7901 for_each_zone(zone) {
7902 if (!is_highmem(zone))
7903 lowmem_pages += zone_managed_pages(zone);
7904 }
7905
7906 for_each_zone(zone) {
7907 u64 tmp;
7908
7909 spin_lock_irqsave(&zone->lock, flags);
7910 tmp = (u64)pages_min * zone_managed_pages(zone);
7911 do_div(tmp, lowmem_pages);
7912 if (is_highmem(zone)) {
7913 /*
7914 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7915 * need highmem pages, so cap pages_min to a small
7916 * value here.
7917 *
7918 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7919 * deltas control async page reclaim, and so should
7920 * not be capped for highmem.
7921 */
7922 unsigned long min_pages;
7923
7924 min_pages = zone_managed_pages(zone) / 1024;
7925 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7926 zone->_watermark[WMARK_MIN] = min_pages;
7927 } else {
7928 /*
7929 * If it's a lowmem zone, reserve a number of pages
7930 * proportionate to the zone's size.
7931 */
7932 zone->_watermark[WMARK_MIN] = tmp;
7933 }
7934
7935 /*
7936 * Set the kswapd watermarks distance according to the
7937 * scale factor in proportion to available memory, but
7938 * ensure a minimum size on small systems.
7939 */
7940 tmp = max_t(u64, tmp >> 2,
7941 mult_frac(zone_managed_pages(zone),
7942 watermark_scale_factor, 10000));
7943
7944 zone->watermark_boost = 0;
7945 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7946 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7947
7948 spin_unlock_irqrestore(&zone->lock, flags);
7949 }
7950
7951 /* update totalreserve_pages */
7952 calculate_totalreserve_pages();
7953 }
7954
7955 /**
7956 * setup_per_zone_wmarks - called when min_free_kbytes changes
7957 * or when memory is hot-{added|removed}
7958 *
7959 * Ensures that the watermark[min,low,high] values for each zone are set
7960 * correctly with respect to min_free_kbytes.
7961 */
7962 void setup_per_zone_wmarks(void)
7963 {
7964 static DEFINE_SPINLOCK(lock);
7965
7966 spin_lock(&lock);
7967 __setup_per_zone_wmarks();
7968 spin_unlock(&lock);
7969 }
7970
7971 /*
7972 * Initialise min_free_kbytes.
7973 *
7974 * For small machines we want it small (128k min). For large machines
7975 * we want it large (256MB max). But it is not linear, because network
7976 * bandwidth does not increase linearly with machine size. We use
7977 *
7978 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7979 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7980 *
7981 * which yields
7982 *
7983 * 16MB: 512k
7984 * 32MB: 724k
7985 * 64MB: 1024k
7986 * 128MB: 1448k
7987 * 256MB: 2048k
7988 * 512MB: 2896k
7989 * 1024MB: 4096k
7990 * 2048MB: 5792k
7991 * 4096MB: 8192k
7992 * 8192MB: 11584k
7993 * 16384MB: 16384k
7994 */
7995 int __meminit init_per_zone_wmark_min(void)
7996 {
7997 unsigned long lowmem_kbytes;
7998 int new_min_free_kbytes;
7999
8000 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8001 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8002
8003 if (new_min_free_kbytes > user_min_free_kbytes) {
8004 min_free_kbytes = new_min_free_kbytes;
8005 if (min_free_kbytes < 128)
8006 min_free_kbytes = 128;
8007 if (min_free_kbytes > 262144)
8008 min_free_kbytes = 262144;
8009 } else {
8010 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8011 new_min_free_kbytes, user_min_free_kbytes);
8012 }
8013 setup_per_zone_wmarks();
8014 refresh_zone_stat_thresholds();
8015 setup_per_zone_lowmem_reserve();
8016
8017 #ifdef CONFIG_NUMA
8018 setup_min_unmapped_ratio();
8019 setup_min_slab_ratio();
8020 #endif
8021
8022 khugepaged_min_free_kbytes_update();
8023
8024 return 0;
8025 }
8026 postcore_initcall(init_per_zone_wmark_min)
8027
8028 /*
8029 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8030 * that we can call two helper functions whenever min_free_kbytes
8031 * changes.
8032 */
8033 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8034 void *buffer, size_t *length, loff_t *ppos)
8035 {
8036 int rc;
8037
8038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8039 if (rc)
8040 return rc;
8041
8042 if (write) {
8043 user_min_free_kbytes = min_free_kbytes;
8044 setup_per_zone_wmarks();
8045 }
8046 return 0;
8047 }
8048
8049 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8050 void *buffer, size_t *length, loff_t *ppos)
8051 {
8052 int rc;
8053
8054 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8055 if (rc)
8056 return rc;
8057
8058 if (write)
8059 setup_per_zone_wmarks();
8060
8061 return 0;
8062 }
8063
8064 #ifdef CONFIG_NUMA
8065 static void setup_min_unmapped_ratio(void)
8066 {
8067 pg_data_t *pgdat;
8068 struct zone *zone;
8069
8070 for_each_online_pgdat(pgdat)
8071 pgdat->min_unmapped_pages = 0;
8072
8073 for_each_zone(zone)
8074 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8075 sysctl_min_unmapped_ratio) / 100;
8076 }
8077
8078
8079 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8080 void *buffer, size_t *length, loff_t *ppos)
8081 {
8082 int rc;
8083
8084 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8085 if (rc)
8086 return rc;
8087
8088 setup_min_unmapped_ratio();
8089
8090 return 0;
8091 }
8092
8093 static void setup_min_slab_ratio(void)
8094 {
8095 pg_data_t *pgdat;
8096 struct zone *zone;
8097
8098 for_each_online_pgdat(pgdat)
8099 pgdat->min_slab_pages = 0;
8100
8101 for_each_zone(zone)
8102 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8103 sysctl_min_slab_ratio) / 100;
8104 }
8105
8106 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8107 void *buffer, size_t *length, loff_t *ppos)
8108 {
8109 int rc;
8110
8111 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8112 if (rc)
8113 return rc;
8114
8115 setup_min_slab_ratio();
8116
8117 return 0;
8118 }
8119 #endif
8120
8121 /*
8122 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8123 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8124 * whenever sysctl_lowmem_reserve_ratio changes.
8125 *
8126 * The reserve ratio obviously has absolutely no relation with the
8127 * minimum watermarks. The lowmem reserve ratio can only make sense
8128 * if in function of the boot time zone sizes.
8129 */
8130 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8131 void *buffer, size_t *length, loff_t *ppos)
8132 {
8133 int i;
8134
8135 proc_dointvec_minmax(table, write, buffer, length, ppos);
8136
8137 for (i = 0; i < MAX_NR_ZONES; i++) {
8138 if (sysctl_lowmem_reserve_ratio[i] < 1)
8139 sysctl_lowmem_reserve_ratio[i] = 0;
8140 }
8141
8142 setup_per_zone_lowmem_reserve();
8143 return 0;
8144 }
8145
8146 /*
8147 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8148 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8149 * pagelist can have before it gets flushed back to buddy allocator.
8150 */
8151 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8152 void *buffer, size_t *length, loff_t *ppos)
8153 {
8154 struct zone *zone;
8155 int old_percpu_pagelist_fraction;
8156 int ret;
8157
8158 mutex_lock(&pcp_batch_high_lock);
8159 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8160
8161 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8162 if (!write || ret < 0)
8163 goto out;
8164
8165 /* Sanity checking to avoid pcp imbalance */
8166 if (percpu_pagelist_fraction &&
8167 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8168 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8169 ret = -EINVAL;
8170 goto out;
8171 }
8172
8173 /* No change? */
8174 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8175 goto out;
8176
8177 for_each_populated_zone(zone)
8178 zone_set_pageset_high_and_batch(zone);
8179 out:
8180 mutex_unlock(&pcp_batch_high_lock);
8181 return ret;
8182 }
8183
8184 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8185 /*
8186 * Returns the number of pages that arch has reserved but
8187 * is not known to alloc_large_system_hash().
8188 */
8189 static unsigned long __init arch_reserved_kernel_pages(void)
8190 {
8191 return 0;
8192 }
8193 #endif
8194
8195 /*
8196 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8197 * machines. As memory size is increased the scale is also increased but at
8198 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8199 * quadruples the scale is increased by one, which means the size of hash table
8200 * only doubles, instead of quadrupling as well.
8201 * Because 32-bit systems cannot have large physical memory, where this scaling
8202 * makes sense, it is disabled on such platforms.
8203 */
8204 #if __BITS_PER_LONG > 32
8205 #define ADAPT_SCALE_BASE (64ul << 30)
8206 #define ADAPT_SCALE_SHIFT 2
8207 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8208 #endif
8209
8210 /*
8211 * allocate a large system hash table from bootmem
8212 * - it is assumed that the hash table must contain an exact power-of-2
8213 * quantity of entries
8214 * - limit is the number of hash buckets, not the total allocation size
8215 */
8216 void *__init alloc_large_system_hash(const char *tablename,
8217 unsigned long bucketsize,
8218 unsigned long numentries,
8219 int scale,
8220 int flags,
8221 unsigned int *_hash_shift,
8222 unsigned int *_hash_mask,
8223 unsigned long low_limit,
8224 unsigned long high_limit)
8225 {
8226 unsigned long long max = high_limit;
8227 unsigned long log2qty, size;
8228 void *table = NULL;
8229 gfp_t gfp_flags;
8230 bool virt;
8231
8232 /* allow the kernel cmdline to have a say */
8233 if (!numentries) {
8234 /* round applicable memory size up to nearest megabyte */
8235 numentries = nr_kernel_pages;
8236 numentries -= arch_reserved_kernel_pages();
8237
8238 /* It isn't necessary when PAGE_SIZE >= 1MB */
8239 if (PAGE_SHIFT < 20)
8240 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8241
8242 #if __BITS_PER_LONG > 32
8243 if (!high_limit) {
8244 unsigned long adapt;
8245
8246 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8247 adapt <<= ADAPT_SCALE_SHIFT)
8248 scale++;
8249 }
8250 #endif
8251
8252 /* limit to 1 bucket per 2^scale bytes of low memory */
8253 if (scale > PAGE_SHIFT)
8254 numentries >>= (scale - PAGE_SHIFT);
8255 else
8256 numentries <<= (PAGE_SHIFT - scale);
8257
8258 /* Make sure we've got at least a 0-order allocation.. */
8259 if (unlikely(flags & HASH_SMALL)) {
8260 /* Makes no sense without HASH_EARLY */
8261 WARN_ON(!(flags & HASH_EARLY));
8262 if (!(numentries >> *_hash_shift)) {
8263 numentries = 1UL << *_hash_shift;
8264 BUG_ON(!numentries);
8265 }
8266 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8267 numentries = PAGE_SIZE / bucketsize;
8268 }
8269 numentries = roundup_pow_of_two(numentries);
8270
8271 /* limit allocation size to 1/16 total memory by default */
8272 if (max == 0) {
8273 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8274 do_div(max, bucketsize);
8275 }
8276 max = min(max, 0x80000000ULL);
8277
8278 if (numentries < low_limit)
8279 numentries = low_limit;
8280 if (numentries > max)
8281 numentries = max;
8282
8283 log2qty = ilog2(numentries);
8284
8285 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8286 do {
8287 virt = false;
8288 size = bucketsize << log2qty;
8289 if (flags & HASH_EARLY) {
8290 if (flags & HASH_ZERO)
8291 table = memblock_alloc(size, SMP_CACHE_BYTES);
8292 else
8293 table = memblock_alloc_raw(size,
8294 SMP_CACHE_BYTES);
8295 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8296 table = __vmalloc(size, gfp_flags);
8297 virt = true;
8298 } else {
8299 /*
8300 * If bucketsize is not a power-of-two, we may free
8301 * some pages at the end of hash table which
8302 * alloc_pages_exact() automatically does
8303 */
8304 table = alloc_pages_exact(size, gfp_flags);
8305 kmemleak_alloc(table, size, 1, gfp_flags);
8306 }
8307 } while (!table && size > PAGE_SIZE && --log2qty);
8308
8309 if (!table)
8310 panic("Failed to allocate %s hash table\n", tablename);
8311
8312 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8313 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8314 virt ? "vmalloc" : "linear");
8315
8316 if (_hash_shift)
8317 *_hash_shift = log2qty;
8318 if (_hash_mask)
8319 *_hash_mask = (1 << log2qty) - 1;
8320
8321 return table;
8322 }
8323
8324 /*
8325 * This function checks whether pageblock includes unmovable pages or not.
8326 *
8327 * PageLRU check without isolation or lru_lock could race so that
8328 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8329 * check without lock_page also may miss some movable non-lru pages at
8330 * race condition. So you can't expect this function should be exact.
8331 *
8332 * Returns a page without holding a reference. If the caller wants to
8333 * dereference that page (e.g., dumping), it has to make sure that it
8334 * cannot get removed (e.g., via memory unplug) concurrently.
8335 *
8336 */
8337 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8338 int migratetype, int flags)
8339 {
8340 unsigned long iter = 0;
8341 unsigned long pfn = page_to_pfn(page);
8342 unsigned long offset = pfn % pageblock_nr_pages;
8343
8344 if (is_migrate_cma_page(page)) {
8345 /*
8346 * CMA allocations (alloc_contig_range) really need to mark
8347 * isolate CMA pageblocks even when they are not movable in fact
8348 * so consider them movable here.
8349 */
8350 if (is_migrate_cma(migratetype))
8351 return NULL;
8352
8353 return page;
8354 }
8355
8356 for (; iter < pageblock_nr_pages - offset; iter++) {
8357 if (!pfn_valid_within(pfn + iter))
8358 continue;
8359
8360 page = pfn_to_page(pfn + iter);
8361
8362 /*
8363 * Both, bootmem allocations and memory holes are marked
8364 * PG_reserved and are unmovable. We can even have unmovable
8365 * allocations inside ZONE_MOVABLE, for example when
8366 * specifying "movablecore".
8367 */
8368 if (PageReserved(page))
8369 return page;
8370
8371 /*
8372 * If the zone is movable and we have ruled out all reserved
8373 * pages then it should be reasonably safe to assume the rest
8374 * is movable.
8375 */
8376 if (zone_idx(zone) == ZONE_MOVABLE)
8377 continue;
8378
8379 /*
8380 * Hugepages are not in LRU lists, but they're movable.
8381 * THPs are on the LRU, but need to be counted as #small pages.
8382 * We need not scan over tail pages because we don't
8383 * handle each tail page individually in migration.
8384 */
8385 if (PageHuge(page) || PageTransCompound(page)) {
8386 struct page *head = compound_head(page);
8387 unsigned int skip_pages;
8388
8389 if (PageHuge(page)) {
8390 if (!hugepage_migration_supported(page_hstate(head)))
8391 return page;
8392 } else if (!PageLRU(head) && !__PageMovable(head)) {
8393 return page;
8394 }
8395
8396 skip_pages = compound_nr(head) - (page - head);
8397 iter += skip_pages - 1;
8398 continue;
8399 }
8400
8401 /*
8402 * We can't use page_count without pin a page
8403 * because another CPU can free compound page.
8404 * This check already skips compound tails of THP
8405 * because their page->_refcount is zero at all time.
8406 */
8407 if (!page_ref_count(page)) {
8408 if (PageBuddy(page))
8409 iter += (1 << buddy_order(page)) - 1;
8410 continue;
8411 }
8412
8413 /*
8414 * The HWPoisoned page may be not in buddy system, and
8415 * page_count() is not 0.
8416 */
8417 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8418 continue;
8419
8420 /*
8421 * We treat all PageOffline() pages as movable when offlining
8422 * to give drivers a chance to decrement their reference count
8423 * in MEM_GOING_OFFLINE in order to indicate that these pages
8424 * can be offlined as there are no direct references anymore.
8425 * For actually unmovable PageOffline() where the driver does
8426 * not support this, we will fail later when trying to actually
8427 * move these pages that still have a reference count > 0.
8428 * (false negatives in this function only)
8429 */
8430 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8431 continue;
8432
8433 if (__PageMovable(page) || PageLRU(page))
8434 continue;
8435
8436 /*
8437 * If there are RECLAIMABLE pages, we need to check
8438 * it. But now, memory offline itself doesn't call
8439 * shrink_node_slabs() and it still to be fixed.
8440 */
8441 return page;
8442 }
8443 return NULL;
8444 }
8445
8446 #ifdef CONFIG_CONTIG_ALLOC
8447 static unsigned long pfn_max_align_down(unsigned long pfn)
8448 {
8449 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8450 pageblock_nr_pages) - 1);
8451 }
8452
8453 static unsigned long pfn_max_align_up(unsigned long pfn)
8454 {
8455 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8456 pageblock_nr_pages));
8457 }
8458
8459 /* [start, end) must belong to a single zone. */
8460 static int __alloc_contig_migrate_range(struct compact_control *cc,
8461 unsigned long start, unsigned long end)
8462 {
8463 /* This function is based on compact_zone() from compaction.c. */
8464 unsigned int nr_reclaimed;
8465 unsigned long pfn = start;
8466 unsigned int tries = 0;
8467 int ret = 0;
8468 struct migration_target_control mtc = {
8469 .nid = zone_to_nid(cc->zone),
8470 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8471 };
8472
8473 migrate_prep();
8474
8475 while (pfn < end || !list_empty(&cc->migratepages)) {
8476 if (fatal_signal_pending(current)) {
8477 ret = -EINTR;
8478 break;
8479 }
8480
8481 if (list_empty(&cc->migratepages)) {
8482 cc->nr_migratepages = 0;
8483 pfn = isolate_migratepages_range(cc, pfn, end);
8484 if (!pfn) {
8485 ret = -EINTR;
8486 break;
8487 }
8488 tries = 0;
8489 } else if (++tries == 5) {
8490 ret = ret < 0 ? ret : -EBUSY;
8491 break;
8492 }
8493
8494 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8495 &cc->migratepages);
8496 cc->nr_migratepages -= nr_reclaimed;
8497
8498 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8499 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8500 }
8501 if (ret < 0) {
8502 putback_movable_pages(&cc->migratepages);
8503 return ret;
8504 }
8505 return 0;
8506 }
8507
8508 /**
8509 * alloc_contig_range() -- tries to allocate given range of pages
8510 * @start: start PFN to allocate
8511 * @end: one-past-the-last PFN to allocate
8512 * @migratetype: migratetype of the underlaying pageblocks (either
8513 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8514 * in range must have the same migratetype and it must
8515 * be either of the two.
8516 * @gfp_mask: GFP mask to use during compaction
8517 *
8518 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8519 * aligned. The PFN range must belong to a single zone.
8520 *
8521 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8522 * pageblocks in the range. Once isolated, the pageblocks should not
8523 * be modified by others.
8524 *
8525 * Return: zero on success or negative error code. On success all
8526 * pages which PFN is in [start, end) are allocated for the caller and
8527 * need to be freed with free_contig_range().
8528 */
8529 int alloc_contig_range(unsigned long start, unsigned long end,
8530 unsigned migratetype, gfp_t gfp_mask)
8531 {
8532 unsigned long outer_start, outer_end;
8533 unsigned int order;
8534 int ret = 0;
8535
8536 struct compact_control cc = {
8537 .nr_migratepages = 0,
8538 .order = -1,
8539 .zone = page_zone(pfn_to_page(start)),
8540 .mode = MIGRATE_SYNC,
8541 .ignore_skip_hint = true,
8542 .no_set_skip_hint = true,
8543 .gfp_mask = current_gfp_context(gfp_mask),
8544 .alloc_contig = true,
8545 };
8546 INIT_LIST_HEAD(&cc.migratepages);
8547
8548 /*
8549 * What we do here is we mark all pageblocks in range as
8550 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8551 * have different sizes, and due to the way page allocator
8552 * work, we align the range to biggest of the two pages so
8553 * that page allocator won't try to merge buddies from
8554 * different pageblocks and change MIGRATE_ISOLATE to some
8555 * other migration type.
8556 *
8557 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8558 * migrate the pages from an unaligned range (ie. pages that
8559 * we are interested in). This will put all the pages in
8560 * range back to page allocator as MIGRATE_ISOLATE.
8561 *
8562 * When this is done, we take the pages in range from page
8563 * allocator removing them from the buddy system. This way
8564 * page allocator will never consider using them.
8565 *
8566 * This lets us mark the pageblocks back as
8567 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8568 * aligned range but not in the unaligned, original range are
8569 * put back to page allocator so that buddy can use them.
8570 */
8571
8572 ret = start_isolate_page_range(pfn_max_align_down(start),
8573 pfn_max_align_up(end), migratetype, 0);
8574 if (ret)
8575 return ret;
8576
8577 drain_all_pages(cc.zone);
8578
8579 /*
8580 * In case of -EBUSY, we'd like to know which page causes problem.
8581 * So, just fall through. test_pages_isolated() has a tracepoint
8582 * which will report the busy page.
8583 *
8584 * It is possible that busy pages could become available before
8585 * the call to test_pages_isolated, and the range will actually be
8586 * allocated. So, if we fall through be sure to clear ret so that
8587 * -EBUSY is not accidentally used or returned to caller.
8588 */
8589 ret = __alloc_contig_migrate_range(&cc, start, end);
8590 if (ret && ret != -EBUSY)
8591 goto done;
8592 ret =0;
8593
8594 /*
8595 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8596 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8597 * more, all pages in [start, end) are free in page allocator.
8598 * What we are going to do is to allocate all pages from
8599 * [start, end) (that is remove them from page allocator).
8600 *
8601 * The only problem is that pages at the beginning and at the
8602 * end of interesting range may be not aligned with pages that
8603 * page allocator holds, ie. they can be part of higher order
8604 * pages. Because of this, we reserve the bigger range and
8605 * once this is done free the pages we are not interested in.
8606 *
8607 * We don't have to hold zone->lock here because the pages are
8608 * isolated thus they won't get removed from buddy.
8609 */
8610
8611 lru_add_drain_all();
8612
8613 order = 0;
8614 outer_start = start;
8615 while (!PageBuddy(pfn_to_page(outer_start))) {
8616 if (++order >= MAX_ORDER) {
8617 outer_start = start;
8618 break;
8619 }
8620 outer_start &= ~0UL << order;
8621 }
8622
8623 if (outer_start != start) {
8624 order = buddy_order(pfn_to_page(outer_start));
8625
8626 /*
8627 * outer_start page could be small order buddy page and
8628 * it doesn't include start page. Adjust outer_start
8629 * in this case to report failed page properly
8630 * on tracepoint in test_pages_isolated()
8631 */
8632 if (outer_start + (1UL << order) <= start)
8633 outer_start = start;
8634 }
8635
8636 /* Make sure the range is really isolated. */
8637 if (test_pages_isolated(outer_start, end, 0)) {
8638 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8639 __func__, outer_start, end);
8640 ret = -EBUSY;
8641 goto done;
8642 }
8643
8644 /* Grab isolated pages from freelists. */
8645 outer_end = isolate_freepages_range(&cc, outer_start, end);
8646 if (!outer_end) {
8647 ret = -EBUSY;
8648 goto done;
8649 }
8650
8651 /* Free head and tail (if any) */
8652 if (start != outer_start)
8653 free_contig_range(outer_start, start - outer_start);
8654 if (end != outer_end)
8655 free_contig_range(end, outer_end - end);
8656
8657 done:
8658 undo_isolate_page_range(pfn_max_align_down(start),
8659 pfn_max_align_up(end), migratetype);
8660 return ret;
8661 }
8662 EXPORT_SYMBOL(alloc_contig_range);
8663
8664 static int __alloc_contig_pages(unsigned long start_pfn,
8665 unsigned long nr_pages, gfp_t gfp_mask)
8666 {
8667 unsigned long end_pfn = start_pfn + nr_pages;
8668
8669 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8670 gfp_mask);
8671 }
8672
8673 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8674 unsigned long nr_pages)
8675 {
8676 unsigned long i, end_pfn = start_pfn + nr_pages;
8677 struct page *page;
8678
8679 for (i = start_pfn; i < end_pfn; i++) {
8680 page = pfn_to_online_page(i);
8681 if (!page)
8682 return false;
8683
8684 if (page_zone(page) != z)
8685 return false;
8686
8687 if (PageReserved(page))
8688 return false;
8689
8690 if (page_count(page) > 0)
8691 return false;
8692
8693 if (PageHuge(page))
8694 return false;
8695 }
8696 return true;
8697 }
8698
8699 static bool zone_spans_last_pfn(const struct zone *zone,
8700 unsigned long start_pfn, unsigned long nr_pages)
8701 {
8702 unsigned long last_pfn = start_pfn + nr_pages - 1;
8703
8704 return zone_spans_pfn(zone, last_pfn);
8705 }
8706
8707 /**
8708 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8709 * @nr_pages: Number of contiguous pages to allocate
8710 * @gfp_mask: GFP mask to limit search and used during compaction
8711 * @nid: Target node
8712 * @nodemask: Mask for other possible nodes
8713 *
8714 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8715 * on an applicable zonelist to find a contiguous pfn range which can then be
8716 * tried for allocation with alloc_contig_range(). This routine is intended
8717 * for allocation requests which can not be fulfilled with the buddy allocator.
8718 *
8719 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8720 * power of two then the alignment is guaranteed to be to the given nr_pages
8721 * (e.g. 1GB request would be aligned to 1GB).
8722 *
8723 * Allocated pages can be freed with free_contig_range() or by manually calling
8724 * __free_page() on each allocated page.
8725 *
8726 * Return: pointer to contiguous pages on success, or NULL if not successful.
8727 */
8728 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8729 int nid, nodemask_t *nodemask)
8730 {
8731 unsigned long ret, pfn, flags;
8732 struct zonelist *zonelist;
8733 struct zone *zone;
8734 struct zoneref *z;
8735
8736 zonelist = node_zonelist(nid, gfp_mask);
8737 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8738 gfp_zone(gfp_mask), nodemask) {
8739 spin_lock_irqsave(&zone->lock, flags);
8740
8741 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8742 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8743 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8744 /*
8745 * We release the zone lock here because
8746 * alloc_contig_range() will also lock the zone
8747 * at some point. If there's an allocation
8748 * spinning on this lock, it may win the race
8749 * and cause alloc_contig_range() to fail...
8750 */
8751 spin_unlock_irqrestore(&zone->lock, flags);
8752 ret = __alloc_contig_pages(pfn, nr_pages,
8753 gfp_mask);
8754 if (!ret)
8755 return pfn_to_page(pfn);
8756 spin_lock_irqsave(&zone->lock, flags);
8757 }
8758 pfn += nr_pages;
8759 }
8760 spin_unlock_irqrestore(&zone->lock, flags);
8761 }
8762 return NULL;
8763 }
8764 #endif /* CONFIG_CONTIG_ALLOC */
8765
8766 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8767 {
8768 unsigned int count = 0;
8769
8770 for (; nr_pages--; pfn++) {
8771 struct page *page = pfn_to_page(pfn);
8772
8773 count += page_count(page) != 1;
8774 __free_page(page);
8775 }
8776 WARN(count != 0, "%d pages are still in use!\n", count);
8777 }
8778 EXPORT_SYMBOL(free_contig_range);
8779
8780 /*
8781 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8782 * page high values need to be recalulated.
8783 */
8784 void __meminit zone_pcp_update(struct zone *zone)
8785 {
8786 mutex_lock(&pcp_batch_high_lock);
8787 zone_set_pageset_high_and_batch(zone);
8788 mutex_unlock(&pcp_batch_high_lock);
8789 }
8790
8791 /*
8792 * Effectively disable pcplists for the zone by setting the high limit to 0
8793 * and draining all cpus. A concurrent page freeing on another CPU that's about
8794 * to put the page on pcplist will either finish before the drain and the page
8795 * will be drained, or observe the new high limit and skip the pcplist.
8796 *
8797 * Must be paired with a call to zone_pcp_enable().
8798 */
8799 void zone_pcp_disable(struct zone *zone)
8800 {
8801 mutex_lock(&pcp_batch_high_lock);
8802 __zone_set_pageset_high_and_batch(zone, 0, 1);
8803 __drain_all_pages(zone, true);
8804 }
8805
8806 void zone_pcp_enable(struct zone *zone)
8807 {
8808 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8809 mutex_unlock(&pcp_batch_high_lock);
8810 }
8811
8812 void zone_pcp_reset(struct zone *zone)
8813 {
8814 unsigned long flags;
8815 int cpu;
8816 struct per_cpu_pageset *pset;
8817
8818 /* avoid races with drain_pages() */
8819 local_irq_save(flags);
8820 if (zone->pageset != &boot_pageset) {
8821 for_each_online_cpu(cpu) {
8822 pset = per_cpu_ptr(zone->pageset, cpu);
8823 drain_zonestat(zone, pset);
8824 }
8825 free_percpu(zone->pageset);
8826 zone->pageset = &boot_pageset;
8827 }
8828 local_irq_restore(flags);
8829 }
8830
8831 #ifdef CONFIG_MEMORY_HOTREMOVE
8832 /*
8833 * All pages in the range must be in a single zone, must not contain holes,
8834 * must span full sections, and must be isolated before calling this function.
8835 */
8836 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8837 {
8838 unsigned long pfn = start_pfn;
8839 struct page *page;
8840 struct zone *zone;
8841 unsigned int order;
8842 unsigned long flags;
8843
8844 offline_mem_sections(pfn, end_pfn);
8845 zone = page_zone(pfn_to_page(pfn));
8846 spin_lock_irqsave(&zone->lock, flags);
8847 while (pfn < end_pfn) {
8848 page = pfn_to_page(pfn);
8849 /*
8850 * The HWPoisoned page may be not in buddy system, and
8851 * page_count() is not 0.
8852 */
8853 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8854 pfn++;
8855 continue;
8856 }
8857 /*
8858 * At this point all remaining PageOffline() pages have a
8859 * reference count of 0 and can simply be skipped.
8860 */
8861 if (PageOffline(page)) {
8862 BUG_ON(page_count(page));
8863 BUG_ON(PageBuddy(page));
8864 pfn++;
8865 continue;
8866 }
8867
8868 BUG_ON(page_count(page));
8869 BUG_ON(!PageBuddy(page));
8870 order = buddy_order(page);
8871 del_page_from_free_list(page, zone, order);
8872 pfn += (1 << order);
8873 }
8874 spin_unlock_irqrestore(&zone->lock, flags);
8875 }
8876 #endif
8877
8878 bool is_free_buddy_page(struct page *page)
8879 {
8880 struct zone *zone = page_zone(page);
8881 unsigned long pfn = page_to_pfn(page);
8882 unsigned long flags;
8883 unsigned int order;
8884
8885 spin_lock_irqsave(&zone->lock, flags);
8886 for (order = 0; order < MAX_ORDER; order++) {
8887 struct page *page_head = page - (pfn & ((1 << order) - 1));
8888
8889 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8890 break;
8891 }
8892 spin_unlock_irqrestore(&zone->lock, flags);
8893
8894 return order < MAX_ORDER;
8895 }
8896
8897 #ifdef CONFIG_MEMORY_FAILURE
8898 /*
8899 * Break down a higher-order page in sub-pages, and keep our target out of
8900 * buddy allocator.
8901 */
8902 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8903 struct page *target, int low, int high,
8904 int migratetype)
8905 {
8906 unsigned long size = 1 << high;
8907 struct page *current_buddy, *next_page;
8908
8909 while (high > low) {
8910 high--;
8911 size >>= 1;
8912
8913 if (target >= &page[size]) {
8914 next_page = page + size;
8915 current_buddy = page;
8916 } else {
8917 next_page = page;
8918 current_buddy = page + size;
8919 }
8920
8921 if (set_page_guard(zone, current_buddy, high, migratetype))
8922 continue;
8923
8924 if (current_buddy != target) {
8925 add_to_free_list(current_buddy, zone, high, migratetype);
8926 set_buddy_order(current_buddy, high);
8927 page = next_page;
8928 }
8929 }
8930 }
8931
8932 /*
8933 * Take a page that will be marked as poisoned off the buddy allocator.
8934 */
8935 bool take_page_off_buddy(struct page *page)
8936 {
8937 struct zone *zone = page_zone(page);
8938 unsigned long pfn = page_to_pfn(page);
8939 unsigned long flags;
8940 unsigned int order;
8941 bool ret = false;
8942
8943 spin_lock_irqsave(&zone->lock, flags);
8944 for (order = 0; order < MAX_ORDER; order++) {
8945 struct page *page_head = page - (pfn & ((1 << order) - 1));
8946 int page_order = buddy_order(page_head);
8947
8948 if (PageBuddy(page_head) && page_order >= order) {
8949 unsigned long pfn_head = page_to_pfn(page_head);
8950 int migratetype = get_pfnblock_migratetype(page_head,
8951 pfn_head);
8952
8953 del_page_from_free_list(page_head, zone, page_order);
8954 break_down_buddy_pages(zone, page_head, page, 0,
8955 page_order, migratetype);
8956 if (!is_migrate_isolate(migratetype))
8957 __mod_zone_freepage_state(zone, -1, migratetype);
8958 ret = true;
8959 break;
8960 }
8961 if (page_count(page_head) > 0)
8962 break;
8963 }
8964 spin_unlock_irqrestore(&zone->lock, flags);
8965 return ret;
8966 }
8967 #endif