1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* mm/percpu.c - percpu memory allocator
4	*
5	* Copyright (C) 2009 SUSE Linux Products GmbH
6	* Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7	*
8	* Copyright (C) 2017 Facebook Inc.
9	* Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
10	*
11	* The percpu allocator handles both static and dynamic areas. Percpu
12	* areas are allocated in chunks which are divided into units. There is
13	* a 1-to-1 mapping for units to possible cpus. These units are grouped
14	* based on NUMA properties of the machine.
15	*
16	* c0 c1 c2
17	* ------------------- ------------------- ------------
18	* | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19	* ------------------- ...... ------------------- .... ------------
20	*
21	* Allocation is done by offsets into a unit's address space. Ie., an
22	* area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23	* c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24	* and even sparse. Access is handled by configuring percpu base
25	* registers according to the cpu to unit mappings and offsetting the
26	* base address using pcpu_unit_size.
27	*
28	* There is special consideration for the first chunk which must handle
29	* the static percpu variables in the kernel image as allocation services
30	* are not online yet. In short, the first chunk is structured like so:
31	*
32	* <Static | [Reserved] | Dynamic>
33	*
34	* The static data is copied from the original section managed by the
35	* linker. The reserved section, if non-zero, primarily manages static
36	* percpu variables from kernel modules. Finally, the dynamic section
37	* takes care of normal allocations.
38	*
39	* The allocator organizes chunks into lists according to free size and
40	* memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41	* flag should be passed. All memcg-aware allocations are sharing one set
42	* of chunks and all unaccounted allocations and allocations performed
43	* by processes belonging to the root memory cgroup are using the second set.
44	*
45	* The allocator tries to allocate from the fullest chunk first. Each chunk
46	* is managed by a bitmap with metadata blocks. The allocation map is updated
47	* on every allocation and free to reflect the current state while the boundary
48	* map is only updated on allocation. Each metadata block contains
49	* information to help mitigate the need to iterate over large portions
50	* of the bitmap. The reverse mapping from page to chunk is stored in
51	* the page's index. Lastly, units are lazily backed and grow in unison.
52	*
53	* There is a unique conversion that goes on here between bytes and bits.
54	* Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55	* tracks the number of pages it is responsible for in nr_pages. Helper
56	* functions are used to convert from between the bytes, bits, and blocks.
57	* All hints are managed in bits unless explicitly stated.
58	*
59	* To use this allocator, arch code should do the following:
60	*
61	* - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62	* regular address to percpu pointer and back if they need to be
63	* different from the default
64	*
65	* - use pcpu_setup_first_chunk() during percpu area initialization to
66	* setup the first chunk containing the kernel static percpu area
67	*/
68
69	#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70
71	#include <linux/bitmap.h>
72	#include <linux/cpumask.h>
73	#include <linux/memblock.h>
74	#include <linux/err.h>
75	#include <linux/list.h>
76	#include <linux/log2.h>
77	#include <linux/mm.h>
78	#include <linux/module.h>
79	#include <linux/mutex.h>
80	#include <linux/percpu.h>
81	#include <linux/pfn.h>
82	#include <linux/slab.h>
83	#include <linux/spinlock.h>
84	#include <linux/vmalloc.h>
85	#include <linux/workqueue.h>
86	#include <linux/kmemleak.h>
87	#include <linux/sched.h>
88	#include <linux/sched/mm.h>
89	#include <linux/memcontrol.h>
90
91	#include <asm/cacheflush.h>
92	#include <asm/sections.h>
93	#include <asm/tlbflush.h>
94	#include <asm/io.h>
95
96	#define CREATE_TRACE_POINTS
97	#include <trace/events/percpu.h>
98
99	#include "percpu-internal.h"
100
101	/*
102	* The slots are sorted by the size of the biggest continuous free area.
103	* 1-31 bytes share the same slot.
104	*/
105	#define PCPU_SLOT_BASE_SHIFT 5
106	/* chunks in slots below this are subject to being sidelined on failed alloc */
107	#define PCPU_SLOT_FAIL_THRESHOLD 3
108
109	#define PCPU_EMPTY_POP_PAGES_LOW 2
110	#define PCPU_EMPTY_POP_PAGES_HIGH 4
111
112	#ifdef CONFIG_SMP
113	/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
114	#ifndef __addr_to_pcpu_ptr
115	#define __addr_to_pcpu_ptr(addr) \
116	(void __percpu *)((unsigned long)(addr) - \
117	(unsigned long)pcpu_base_addr + \
118	(unsigned long)__per_cpu_start)
119	#endif
120	#ifndef __pcpu_ptr_to_addr
121	#define __pcpu_ptr_to_addr(ptr) \
122	(void __force *)((unsigned long)(ptr) + \
123	(unsigned long)pcpu_base_addr - \
124	(unsigned long)__per_cpu_start)
125	#endif
126	#else /* CONFIG_SMP */
127	/* on UP, it's always identity mapped */
128	#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
129	#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
130	#endif /* CONFIG_SMP */
131
132	static int pcpu_unit_pages __ro_after_init;
133	static int pcpu_unit_size __ro_after_init;
134	static int pcpu_nr_units __ro_after_init;
135	static int pcpu_atom_size __ro_after_init;
136	int pcpu_nr_slots __ro_after_init;
137	static int pcpu_free_slot __ro_after_init;
138	int pcpu_sidelined_slot __ro_after_init;
139	int pcpu_to_depopulate_slot __ro_after_init;
140	static size_t pcpu_chunk_struct_size __ro_after_init;
141
142	/* cpus with the lowest and highest unit addresses */
143	static unsigned int pcpu_low_unit_cpu __ro_after_init;
144	static unsigned int pcpu_high_unit_cpu __ro_after_init;
145
146	/* the address of the first chunk which starts with the kernel static area */
147	void *pcpu_base_addr __ro_after_init;
148
149	static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
150	const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
151
152	/* group information, used for vm allocation */
153	static int pcpu_nr_groups __ro_after_init;
154	static const unsigned long *pcpu_group_offsets __ro_after_init;
155	static const size_t *pcpu_group_sizes __ro_after_init;
156
157	/*
158	* The first chunk which always exists. Note that unlike other
159	* chunks, this one can be allocated and mapped in several different
160	* ways and thus often doesn't live in the vmalloc area.
161	*/
162	struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
163
164	/*
165	* Optional reserved chunk. This chunk reserves part of the first
166	* chunk and serves it for reserved allocations. When the reserved
167	* region doesn't exist, the following variable is NULL.
168	*/
169	struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
170
171	DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
172	static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
173
174	struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
175
176	/*
177	* The number of empty populated pages, protected by pcpu_lock.
178	* The reserved chunk doesn't contribute to the count.
179	*/
180	int pcpu_nr_empty_pop_pages;
181
182	/*
183	* The number of populated pages in use by the allocator, protected by
184	* pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
185	* allocated/deallocated, it is allocated/deallocated in all units of a chunk
186	* and increments/decrements this count by 1).
187	*/
188	static unsigned long pcpu_nr_populated;
189
190	/*
191	* Balance work is used to populate or destroy chunks asynchronously. We
192	* try to keep the number of populated free pages between
193	* PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
194	* empty chunk.
195	*/
196	static void pcpu_balance_workfn(struct work_struct *work);
197	static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
198	static bool pcpu_async_enabled __read_mostly;
199	static bool pcpu_atomic_alloc_failed;
200
201	static void pcpu_schedule_balance_work(void)
202	{
203	if (pcpu_async_enabled)
204	schedule_work(work: &pcpu_balance_work);
205	}
206
207	/**
208	* pcpu_addr_in_chunk - check if the address is served from this chunk
209	* @chunk: chunk of interest
210	* @addr: percpu address
211	*
212	* RETURNS:
213	* True if the address is served from this chunk.
214	*/
215	static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
216	{
217	void *start_addr, *end_addr;
218
219	if (!chunk)
220	return false;
221
222	start_addr = chunk->base_addr + chunk->start_offset;
223	end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
224	chunk->end_offset;
225
226	return addr >= start_addr && addr < end_addr;
227	}
228
229	static int __pcpu_size_to_slot(int size)
230	{
231	int highbit = fls(x: size); /* size is in bytes */
232	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
233	}
234
235	static int pcpu_size_to_slot(int size)
236	{
237	if (size == pcpu_unit_size)
238	return pcpu_free_slot;
239	return __pcpu_size_to_slot(size);
240	}
241
242	static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
243	{
244	const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
245
246	if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
247	chunk_md->contig_hint == 0)
248	return 0;
249
250	return pcpu_size_to_slot(size: chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
251	}
252
253	/* set the pointer to a chunk in a page struct */
254	static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
255	{
256	page->index = (unsigned long)pcpu;
257	}
258
259	/* obtain pointer to a chunk from a page struct */
260	static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
261	{
262	return (struct pcpu_chunk *)page->index;
263	}
264
265	static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
266	{
267	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
268	}
269
270	static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
271	{
272	return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
273	}
274
275	static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
276	unsigned int cpu, int page_idx)
277	{
278	return (unsigned long)chunk->base_addr +
279	pcpu_unit_page_offset(cpu, page_idx);
280	}
281
282	/*
283	* The following are helper functions to help access bitmaps and convert
284	* between bitmap offsets to address offsets.
285	*/
286	static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
287	{
288	return chunk->alloc_map +
289	(index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
290	}
291
292	static unsigned long pcpu_off_to_block_index(int off)
293	{
294	return off / PCPU_BITMAP_BLOCK_BITS;
295	}
296
297	static unsigned long pcpu_off_to_block_off(int off)
298	{
299	return off & (PCPU_BITMAP_BLOCK_BITS - 1);
300	}
301
302	static unsigned long pcpu_block_off_to_off(int index, int off)
303	{
304	return index * PCPU_BITMAP_BLOCK_BITS + off;
305	}
306
307	/**
308	* pcpu_check_block_hint - check against the contig hint
309	* @block: block of interest
310	* @bits: size of allocation
311	* @align: alignment of area (max PAGE_SIZE)
312	*
313	* Check to see if the allocation can fit in the block's contig hint.
314	* Note, a chunk uses the same hints as a block so this can also check against
315	* the chunk's contig hint.
316	*/
317	static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
318	size_t align)
319	{
320	int bit_off = ALIGN(block->contig_hint_start, align) -
321	block->contig_hint_start;
322
323	return bit_off + bits <= block->contig_hint;
324	}
325
326	/*
327	* pcpu_next_hint - determine which hint to use
328	* @block: block of interest
329	* @alloc_bits: size of allocation
330	*
331	* This determines if we should scan based on the scan_hint or first_free.
332	* In general, we want to scan from first_free to fulfill allocations by
333	* first fit. However, if we know a scan_hint at position scan_hint_start
334	* cannot fulfill an allocation, we can begin scanning from there knowing
335	* the contig_hint will be our fallback.
336	*/
337	static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
338	{
339	/*
340	* The three conditions below determine if we can skip past the
341	* scan_hint. First, does the scan hint exist. Second, is the
342	* contig_hint after the scan_hint (possibly not true iff
343	* contig_hint == scan_hint). Third, is the allocation request
344	* larger than the scan_hint.
345	*/
346	if (block->scan_hint &&
347	block->contig_hint_start > block->scan_hint_start &&
348	alloc_bits > block->scan_hint)
349	return block->scan_hint_start + block->scan_hint;
350
351	return block->first_free;
352	}
353
354	/**
355	* pcpu_next_md_free_region - finds the next hint free area
356	* @chunk: chunk of interest
357	* @bit_off: chunk offset
358	* @bits: size of free area
359	*
360	* Helper function for pcpu_for_each_md_free_region. It checks
361	* block->contig_hint and performs aggregation across blocks to find the
362	* next hint. It modifies bit_off and bits in-place to be consumed in the
363	* loop.
364	*/
365	static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
366	int *bits)
367	{
368	int i = pcpu_off_to_block_index(off: *bit_off);
369	int block_off = pcpu_off_to_block_off(off: *bit_off);
370	struct pcpu_block_md *block;
371
372	*bits = 0;
373	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
374	block++, i++) {
375	/* handles contig area across blocks */
376	if (*bits) {
377	*bits += block->left_free;
378	if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
379	continue;
380	return;
381	}
382
383	/*
384	* This checks three things. First is there a contig_hint to
385	* check. Second, have we checked this hint before by
386	* comparing the block_off. Third, is this the same as the
387	* right contig hint. In the last case, it spills over into
388	* the next block and should be handled by the contig area
389	* across blocks code.
390	*/
391	*bits = block->contig_hint;
392	if (*bits && block->contig_hint_start >= block_off &&
393	*bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
394	*bit_off = pcpu_block_off_to_off(index: i,
395	off: block->contig_hint_start);
396	return;
397	}
398	/* reset to satisfy the second predicate above */
399	block_off = 0;
400
401	*bits = block->right_free;
402	*bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
403	}
404	}
405
406	/**
407	* pcpu_next_fit_region - finds fit areas for a given allocation request
408	* @chunk: chunk of interest
409	* @alloc_bits: size of allocation
410	* @align: alignment of area (max PAGE_SIZE)
411	* @bit_off: chunk offset
412	* @bits: size of free area
413	*
414	* Finds the next free region that is viable for use with a given size and
415	* alignment. This only returns if there is a valid area to be used for this
416	* allocation. block->first_free is returned if the allocation request fits
417	* within the block to see if the request can be fulfilled prior to the contig
418	* hint.
419	*/
420	static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
421	int align, int *bit_off, int *bits)
422	{
423	int i = pcpu_off_to_block_index(off: *bit_off);
424	int block_off = pcpu_off_to_block_off(off: *bit_off);
425	struct pcpu_block_md *block;
426
427	*bits = 0;
428	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
429	block++, i++) {
430	/* handles contig area across blocks */
431	if (*bits) {
432	*bits += block->left_free;
433	if (*bits >= alloc_bits)
434	return;
435	if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
436	continue;
437	}
438
439	/* check block->contig_hint */
440	*bits = ALIGN(block->contig_hint_start, align) -
441	block->contig_hint_start;
442	/*
443	* This uses the block offset to determine if this has been
444	* checked in the prior iteration.
445	*/
446	if (block->contig_hint &&
447	block->contig_hint_start >= block_off &&
448	block->contig_hint >= *bits + alloc_bits) {
449	int start = pcpu_next_hint(block, alloc_bits);
450
451	*bits += alloc_bits + block->contig_hint_start -
452	start;
453	*bit_off = pcpu_block_off_to_off(index: i, off: start);
454	return;
455	}
456	/* reset to satisfy the second predicate above */
457	block_off = 0;
458
459	*bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
460	align);
461	*bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
462	*bit_off = pcpu_block_off_to_off(index: i, off: *bit_off);
463	if (*bits >= alloc_bits)
464	return;
465	}
466
467	/* no valid offsets were found - fail condition */
468	*bit_off = pcpu_chunk_map_bits(chunk);
469	}
470
471	/*
472	* Metadata free area iterators. These perform aggregation of free areas
473	* based on the metadata blocks and return the offset @bit_off and size in
474	* bits of the free area @bits. pcpu_for_each_fit_region only returns when
475	* a fit is found for the allocation request.
476	*/
477	#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
478	for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
479	(bit_off) < pcpu_chunk_map_bits((chunk)); \
480	(bit_off) += (bits) + 1, \
481	pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
482
483	#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
484	for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
485	&(bits)); \
486	(bit_off) < pcpu_chunk_map_bits((chunk)); \
487	(bit_off) += (bits), \
488	pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
489	&(bits)))
490
491	/**
492	* pcpu_mem_zalloc - allocate memory
493	* @size: bytes to allocate
494	* @gfp: allocation flags
495	*
496	* Allocate @size bytes. If @size is smaller than PAGE_SIZE,
497	* kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
498	* This is to facilitate passing through whitelisted flags. The
499	* returned memory is always zeroed.
500	*
501	* RETURNS:
502	* Pointer to the allocated area on success, NULL on failure.
503	*/
504	static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
505	{
506	if (WARN_ON_ONCE(!slab_is_available()))
507	return NULL;
508
509	if (size <= PAGE_SIZE)
510	return kzalloc(size, flags: gfp);
511	else
512	return __vmalloc(size, gfp_mask: gfp | __GFP_ZERO);
513	}
514
515	/**
516	* pcpu_mem_free - free memory
517	* @ptr: memory to free
518	*
519	* Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
520	*/
521	static void pcpu_mem_free(void *ptr)
522	{
523	kvfree(addr: ptr);
524	}
525
526	static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
527	bool move_front)
528	{
529	if (chunk != pcpu_reserved_chunk) {
530	if (move_front)
531	list_move(list: &chunk->list, head: &pcpu_chunk_lists[slot]);
532	else
533	list_move_tail(list: &chunk->list, head: &pcpu_chunk_lists[slot]);
534	}
535	}
536
537	static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
538	{
539	__pcpu_chunk_move(chunk, slot, move_front: true);
540	}
541
542	/**
543	* pcpu_chunk_relocate - put chunk in the appropriate chunk slot
544	* @chunk: chunk of interest
545	* @oslot: the previous slot it was on
546	*
547	* This function is called after an allocation or free changed @chunk.
548	* New slot according to the changed state is determined and @chunk is
549	* moved to the slot. Note that the reserved chunk is never put on
550	* chunk slots.
551	*
552	* CONTEXT:
553	* pcpu_lock.
554	*/
555	static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
556	{
557	int nslot = pcpu_chunk_slot(chunk);
558
559	/* leave isolated chunks in-place */
560	if (chunk->isolated)
561	return;
562
563	if (oslot != nslot)
564	__pcpu_chunk_move(chunk, slot: nslot, move_front: oslot < nslot);
565	}
566
567	static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
568	{
569	lockdep_assert_held(&pcpu_lock);
570
571	if (!chunk->isolated) {
572	chunk->isolated = true;
573	pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
574	}
575	list_move(list: &chunk->list, head: &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
576	}
577
578	static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
579	{
580	lockdep_assert_held(&pcpu_lock);
581
582	if (chunk->isolated) {
583	chunk->isolated = false;
584	pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
585	pcpu_chunk_relocate(chunk, oslot: -1);
586	}
587	}
588
589	/*
590	* pcpu_update_empty_pages - update empty page counters
591	* @chunk: chunk of interest
592	* @nr: nr of empty pages
593	*
594	* This is used to keep track of the empty pages now based on the premise
595	* a md_block covers a page. The hint update functions recognize if a block
596	* is made full or broken to calculate deltas for keeping track of free pages.
597	*/
598	static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
599	{
600	chunk->nr_empty_pop_pages += nr;
601	if (chunk != pcpu_reserved_chunk && !chunk->isolated)
602	pcpu_nr_empty_pop_pages += nr;
603	}
604
605	/*
606	* pcpu_region_overlap - determines if two regions overlap
607	* @a: start of first region, inclusive
608	* @b: end of first region, exclusive
609	* @x: start of second region, inclusive
610	* @y: end of second region, exclusive
611	*
612	* This is used to determine if the hint region [a, b) overlaps with the
613	* allocated region [x, y).
614	*/
615	static inline bool pcpu_region_overlap(int a, int b, int x, int y)
616	{
617	return (a < y) && (x < b);
618	}
619
620	/**
621	* pcpu_block_update - updates a block given a free area
622	* @block: block of interest
623	* @start: start offset in block
624	* @end: end offset in block
625	*
626	* Updates a block given a known free area. The region [start, end) is
627	* expected to be the entirety of the free area within a block. Chooses
628	* the best starting offset if the contig hints are equal.
629	*/
630	static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
631	{
632	int contig = end - start;
633
634	block->first_free = min(block->first_free, start);
635	if (start == 0)
636	block->left_free = contig;
637
638	if (end == block->nr_bits)
639	block->right_free = contig;
640
641	if (contig > block->contig_hint) {
642	/* promote the old contig_hint to be the new scan_hint */
643	if (start > block->contig_hint_start) {
644	if (block->contig_hint > block->scan_hint) {
645	block->scan_hint_start =
646	block->contig_hint_start;
647	block->scan_hint = block->contig_hint;
648	} else if (start < block->scan_hint_start) {
649	/*
650	* The old contig_hint == scan_hint. But, the
651	* new contig is larger so hold the invariant
652	* scan_hint_start < contig_hint_start.
653	*/
654	block->scan_hint = 0;
655	}
656	} else {
657	block->scan_hint = 0;
658	}
659	block->contig_hint_start = start;
660	block->contig_hint = contig;
661	} else if (contig == block->contig_hint) {
662	if (block->contig_hint_start &&
663	(!start ||
664	__ffs(start) > __ffs(block->contig_hint_start))) {
665	/* start has a better alignment so use it */
666	block->contig_hint_start = start;
667	if (start < block->scan_hint_start &&
668	block->contig_hint > block->scan_hint)
669	block->scan_hint = 0;
670	} else if (start > block->scan_hint_start ||
671	block->contig_hint > block->scan_hint) {
672	/*
673	* Knowing contig == contig_hint, update the scan_hint
674	* if it is farther than or larger than the current
675	* scan_hint.
676	*/
677	block->scan_hint_start = start;
678	block->scan_hint = contig;
679	}
680	} else {
681	/*
682	* The region is smaller than the contig_hint. So only update
683	* the scan_hint if it is larger than or equal and farther than
684	* the current scan_hint.
685	*/
686	if ((start < block->contig_hint_start &&
687	(contig > block->scan_hint ||
688	(contig == block->scan_hint &&
689	start > block->scan_hint_start)))) {
690	block->scan_hint_start = start;
691	block->scan_hint = contig;
692	}
693	}
694	}
695
696	/*
697	* pcpu_block_update_scan - update a block given a free area from a scan
698	* @chunk: chunk of interest
699	* @bit_off: chunk offset
700	* @bits: size of free area
701	*
702	* Finding the final allocation spot first goes through pcpu_find_block_fit()
703	* to find a block that can hold the allocation and then pcpu_alloc_area()
704	* where a scan is used. When allocations require specific alignments,
705	* we can inadvertently create holes which will not be seen in the alloc
706	* or free paths.
707	*
708	* This takes a given free area hole and updates a block as it may change the
709	* scan_hint. We need to scan backwards to ensure we don't miss free bits
710	* from alignment.
711	*/
712	static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
713	int bits)
714	{
715	int s_off = pcpu_off_to_block_off(off: bit_off);
716	int e_off = s_off + bits;
717	int s_index, l_bit;
718	struct pcpu_block_md *block;
719
720	if (e_off > PCPU_BITMAP_BLOCK_BITS)
721	return;
722
723	s_index = pcpu_off_to_block_index(off: bit_off);
724	block = chunk->md_blocks + s_index;
725
726	/* scan backwards in case of alignment skipping free bits */
727	l_bit = find_last_bit(addr: pcpu_index_alloc_map(chunk, index: s_index), size: s_off);
728	s_off = (s_off == l_bit) ? 0 : l_bit + 1;
729
730	pcpu_block_update(block, start: s_off, end: e_off);
731	}
732
733	/**
734	* pcpu_chunk_refresh_hint - updates metadata about a chunk
735	* @chunk: chunk of interest
736	* @full_scan: if we should scan from the beginning
737	*
738	* Iterates over the metadata blocks to find the largest contig area.
739	* A full scan can be avoided on the allocation path as this is triggered
740	* if we broke the contig_hint. In doing so, the scan_hint will be before
741	* the contig_hint or after if the scan_hint == contig_hint. This cannot
742	* be prevented on freeing as we want to find the largest area possibly
743	* spanning blocks.
744	*/
745	static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
746	{
747	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
748	int bit_off, bits;
749
750	/* promote scan_hint to contig_hint */
751	if (!full_scan && chunk_md->scan_hint) {
752	bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
753	chunk_md->contig_hint_start = chunk_md->scan_hint_start;
754	chunk_md->contig_hint = chunk_md->scan_hint;
755	chunk_md->scan_hint = 0;
756	} else {
757	bit_off = chunk_md->first_free;
758	chunk_md->contig_hint = 0;
759	}
760
761	bits = 0;
762	pcpu_for_each_md_free_region(chunk, bit_off, bits)
763	pcpu_block_update(block: chunk_md, start: bit_off, end: bit_off + bits);
764	}
765
766	/**
767	* pcpu_block_refresh_hint
768	* @chunk: chunk of interest
769	* @index: index of the metadata block
770	*
771	* Scans over the block beginning at first_free and updates the block
772	* metadata accordingly.
773	*/
774	static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
775	{
776	struct pcpu_block_md *block = chunk->md_blocks + index;
777	unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
778	unsigned int start, end; /* region start, region end */
779
780	/* promote scan_hint to contig_hint */
781	if (block->scan_hint) {
782	start = block->scan_hint_start + block->scan_hint;
783	block->contig_hint_start = block->scan_hint_start;
784	block->contig_hint = block->scan_hint;
785	block->scan_hint = 0;
786	} else {
787	start = block->first_free;
788	block->contig_hint = 0;
789	}
790
791	block->right_free = 0;
792
793	/* iterate over free areas and update the contig hints */
794	for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
795	pcpu_block_update(block, start, end);
796	}
797
798	/**
799	* pcpu_block_update_hint_alloc - update hint on allocation path
800	* @chunk: chunk of interest
801	* @bit_off: chunk offset
802	* @bits: size of request
803	*
804	* Updates metadata for the allocation path. The metadata only has to be
805	* refreshed by a full scan iff the chunk's contig hint is broken. Block level
806	* scans are required if the block's contig hint is broken.
807	*/
808	static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
809	int bits)
810	{
811	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
812	int nr_empty_pages = 0;
813	struct pcpu_block_md *s_block, *e_block, *block;
814	int s_index, e_index; /* block indexes of the freed allocation */
815	int s_off, e_off; /* block offsets of the freed allocation */
816
817	/*
818	* Calculate per block offsets.
819	* The calculation uses an inclusive range, but the resulting offsets
820	* are [start, end). e_index always points to the last block in the
821	* range.
822	*/
823	s_index = pcpu_off_to_block_index(off: bit_off);
824	e_index = pcpu_off_to_block_index(off: bit_off + bits - 1);
825	s_off = pcpu_off_to_block_off(off: bit_off);
826	e_off = pcpu_off_to_block_off(off: bit_off + bits - 1) + 1;
827
828	s_block = chunk->md_blocks + s_index;
829	e_block = chunk->md_blocks + e_index;
830
831	/*
832	* Update s_block.
833	*/
834	if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
835	nr_empty_pages++;
836
837	/*
838	* block->first_free must be updated if the allocation takes its place.
839	* If the allocation breaks the contig_hint, a scan is required to
840	* restore this hint.
841	*/
842	if (s_off == s_block->first_free)
843	s_block->first_free = find_next_zero_bit(
844	addr: pcpu_index_alloc_map(chunk, index: s_index),
845	PCPU_BITMAP_BLOCK_BITS,
846	offset: s_off + bits);
847
848	if (pcpu_region_overlap(a: s_block->scan_hint_start,
849	b: s_block->scan_hint_start + s_block->scan_hint,
850	x: s_off,
851	y: s_off + bits))
852	s_block->scan_hint = 0;
853
854	if (pcpu_region_overlap(a: s_block->contig_hint_start,
855	b: s_block->contig_hint_start +
856	s_block->contig_hint,
857	x: s_off,
858	y: s_off + bits)) {
859	/* block contig hint is broken - scan to fix it */
860	if (!s_off)
861	s_block->left_free = 0;
862	pcpu_block_refresh_hint(chunk, index: s_index);
863	} else {
864	/* update left and right contig manually */
865	s_block->left_free = min(s_block->left_free, s_off);
866	if (s_index == e_index)
867	s_block->right_free = min_t(int, s_block->right_free,
868	PCPU_BITMAP_BLOCK_BITS - e_off);
869	else
870	s_block->right_free = 0;
871	}
872
873	/*
874	* Update e_block.
875	*/
876	if (s_index != e_index) {
877	if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
878	nr_empty_pages++;
879
880	/*
881	* When the allocation is across blocks, the end is along
882	* the left part of the e_block.
883	*/
884	e_block->first_free = find_next_zero_bit(
885	addr: pcpu_index_alloc_map(chunk, index: e_index),
886	PCPU_BITMAP_BLOCK_BITS, offset: e_off);
887
888	if (e_off == PCPU_BITMAP_BLOCK_BITS) {
889	/* reset the block */
890	e_block++;
891	} else {
892	if (e_off > e_block->scan_hint_start)
893	e_block->scan_hint = 0;
894
895	e_block->left_free = 0;
896	if (e_off > e_block->contig_hint_start) {
897	/* contig hint is broken - scan to fix it */
898	pcpu_block_refresh_hint(chunk, index: e_index);
899	} else {
900	e_block->right_free =
901	min_t(int, e_block->right_free,
902	PCPU_BITMAP_BLOCK_BITS - e_off);
903	}
904	}
905
906	/* update in-between md_blocks */
907	nr_empty_pages += (e_index - s_index - 1);
908	for (block = s_block + 1; block < e_block; block++) {
909	block->scan_hint = 0;
910	block->contig_hint = 0;
911	block->left_free = 0;
912	block->right_free = 0;
913	}
914	}
915
916	/*
917	* If the allocation is not atomic, some blocks may not be
918	* populated with pages, while we account it here. The number
919	* of pages will be added back with pcpu_chunk_populated()
920	* when populating pages.
921	*/
922	if (nr_empty_pages)
923	pcpu_update_empty_pages(chunk, nr: -nr_empty_pages);
924
925	if (pcpu_region_overlap(a: chunk_md->scan_hint_start,
926	b: chunk_md->scan_hint_start +
927	chunk_md->scan_hint,
928	x: bit_off,
929	y: bit_off + bits))
930	chunk_md->scan_hint = 0;
931
932	/*
933	* The only time a full chunk scan is required is if the chunk
934	* contig hint is broken. Otherwise, it means a smaller space
935	* was used and therefore the chunk contig hint is still correct.
936	*/
937	if (pcpu_region_overlap(a: chunk_md->contig_hint_start,
938	b: chunk_md->contig_hint_start +
939	chunk_md->contig_hint,
940	x: bit_off,
941	y: bit_off + bits))
942	pcpu_chunk_refresh_hint(chunk, full_scan: false);
943	}
944
945	/**
946	* pcpu_block_update_hint_free - updates the block hints on the free path
947	* @chunk: chunk of interest
948	* @bit_off: chunk offset
949	* @bits: size of request
950	*
951	* Updates metadata for the allocation path. This avoids a blind block
952	* refresh by making use of the block contig hints. If this fails, it scans
953	* forward and backward to determine the extent of the free area. This is
954	* capped at the boundary of blocks.
955	*
956	* A chunk update is triggered if a page becomes free, a block becomes free,
957	* or the free spans across blocks. This tradeoff is to minimize iterating
958	* over the block metadata to update chunk_md->contig_hint.
959	* chunk_md->contig_hint may be off by up to a page, but it will never be more
960	* than the available space. If the contig hint is contained in one block, it
961	* will be accurate.
962	*/
963	static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
964	int bits)
965	{
966	int nr_empty_pages = 0;
967	struct pcpu_block_md *s_block, *e_block, *block;
968	int s_index, e_index; /* block indexes of the freed allocation */
969	int s_off, e_off; /* block offsets of the freed allocation */
970	int start, end; /* start and end of the whole free area */
971
972	/*
973	* Calculate per block offsets.
974	* The calculation uses an inclusive range, but the resulting offsets
975	* are [start, end). e_index always points to the last block in the
976	* range.
977	*/
978	s_index = pcpu_off_to_block_index(off: bit_off);
979	e_index = pcpu_off_to_block_index(off: bit_off + bits - 1);
980	s_off = pcpu_off_to_block_off(off: bit_off);
981	e_off = pcpu_off_to_block_off(off: bit_off + bits - 1) + 1;
982
983	s_block = chunk->md_blocks + s_index;
984	e_block = chunk->md_blocks + e_index;
985
986	/*
987	* Check if the freed area aligns with the block->contig_hint.
988	* If it does, then the scan to find the beginning/end of the
989	* larger free area can be avoided.
990	*
991	* start and end refer to beginning and end of the free area
992	* within each their respective blocks. This is not necessarily
993	* the entire free area as it may span blocks past the beginning
994	* or end of the block.
995	*/
996	start = s_off;
997	if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
998	start = s_block->contig_hint_start;
999	} else {
1000	/*
1001	* Scan backwards to find the extent of the free area.
1002	* find_last_bit returns the starting bit, so if the start bit
1003	* is returned, that means there was no last bit and the
1004	* remainder of the chunk is free.
1005	*/
1006	int l_bit = find_last_bit(addr: pcpu_index_alloc_map(chunk, index: s_index),
1007	size: start);
1008	start = (start == l_bit) ? 0 : l_bit + 1;
1009	}
1010
1011	end = e_off;
1012	if (e_off == e_block->contig_hint_start)
1013	end = e_block->contig_hint_start + e_block->contig_hint;
1014	else
1015	end = find_next_bit(addr: pcpu_index_alloc_map(chunk, index: e_index),
1016	PCPU_BITMAP_BLOCK_BITS, offset: end);
1017
1018	/* update s_block */
1019	e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1020	if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1021	nr_empty_pages++;
1022	pcpu_block_update(block: s_block, start, end: e_off);
1023
1024	/* freeing in the same block */
1025	if (s_index != e_index) {
1026	/* update e_block */
1027	if (end == PCPU_BITMAP_BLOCK_BITS)
1028	nr_empty_pages++;
1029	pcpu_block_update(block: e_block, start: 0, end);
1030
1031	/* reset md_blocks in the middle */
1032	nr_empty_pages += (e_index - s_index - 1);
1033	for (block = s_block + 1; block < e_block; block++) {
1034	block->first_free = 0;
1035	block->scan_hint = 0;
1036	block->contig_hint_start = 0;
1037	block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1038	block->left_free = PCPU_BITMAP_BLOCK_BITS;
1039	block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040	}
1041	}
1042
1043	if (nr_empty_pages)
1044	pcpu_update_empty_pages(chunk, nr: nr_empty_pages);
1045
1046	/*
1047	* Refresh chunk metadata when the free makes a block free or spans
1048	* across blocks. The contig_hint may be off by up to a page, but if
1049	* the contig_hint is contained in a block, it will be accurate with
1050	* the else condition below.
1051	*/
1052	if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1053	pcpu_chunk_refresh_hint(chunk, full_scan: true);
1054	else
1055	pcpu_block_update(block: &chunk->chunk_md,
1056	start: pcpu_block_off_to_off(index: s_index, off: start),
1057	end);
1058	}
1059
1060	/**
1061	* pcpu_is_populated - determines if the region is populated
1062	* @chunk: chunk of interest
1063	* @bit_off: chunk offset
1064	* @bits: size of area
1065	* @next_off: return value for the next offset to start searching
1066	*
1067	* For atomic allocations, check if the backing pages are populated.
1068	*
1069	* RETURNS:
1070	* Bool if the backing pages are populated.
1071	* next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1072	*/
1073	static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1074	int *next_off)
1075	{
1076	unsigned int start, end;
1077
1078	start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1079	end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1080
1081	start = find_next_zero_bit(addr: chunk->populated, size: end, offset: start);
1082	if (start >= end)
1083	return true;
1084
1085	end = find_next_bit(addr: chunk->populated, size: end, offset: start + 1);
1086
1087	*next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088	return false;
1089	}
1090
1091	/**
1092	* pcpu_find_block_fit - finds the block index to start searching
1093	* @chunk: chunk of interest
1094	* @alloc_bits: size of request in allocation units
1095	* @align: alignment of area (max PAGE_SIZE bytes)
1096	* @pop_only: use populated regions only
1097	*
1098	* Given a chunk and an allocation spec, find the offset to begin searching
1099	* for a free region. This iterates over the bitmap metadata blocks to
1100	* find an offset that will be guaranteed to fit the requirements. It is
1101	* not quite first fit as if the allocation does not fit in the contig hint
1102	* of a block or chunk, it is skipped. This errs on the side of caution
1103	* to prevent excess iteration. Poor alignment can cause the allocator to
1104	* skip over blocks and chunks that have valid free areas.
1105	*
1106	* RETURNS:
1107	* The offset in the bitmap to begin searching.
1108	* -1 if no offset is found.
1109	*/
1110	static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1111	size_t align, bool pop_only)
1112	{
1113	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1114	int bit_off, bits, next_off;
1115
1116	/*
1117	* This is an optimization to prevent scanning by assuming if the
1118	* allocation cannot fit in the global hint, there is memory pressure
1119	* and creating a new chunk would happen soon.
1120	*/
1121	if (!pcpu_check_block_hint(block: chunk_md, bits: alloc_bits, align))
1122	return -1;
1123
1124	bit_off = pcpu_next_hint(block: chunk_md, alloc_bits);
1125	bits = 0;
1126	pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1127	if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1128	next_off: &next_off))
1129	break;
1130
1131	bit_off = next_off;
1132	bits = 0;
1133	}
1134
1135	if (bit_off == pcpu_chunk_map_bits(chunk))
1136	return -1;
1137
1138	return bit_off;
1139	}
1140
1141	/*
1142	* pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1143	* @map: the address to base the search on
1144	* @size: the bitmap size in bits
1145	* @start: the bitnumber to start searching at
1146	* @nr: the number of zeroed bits we're looking for
1147	* @align_mask: alignment mask for zero area
1148	* @largest_off: offset of the largest area skipped
1149	* @largest_bits: size of the largest area skipped
1150	*
1151	* The @align_mask should be one less than a power of 2.
1152	*
1153	* This is a modified version of bitmap_find_next_zero_area_off() to remember
1154	* the largest area that was skipped. This is imperfect, but in general is
1155	* good enough. The largest remembered region is the largest failed region
1156	* seen. This does not include anything we possibly skipped due to alignment.
1157	* pcpu_block_update_scan() does scan backwards to try and recover what was
1158	* lost to alignment. While this can cause scanning to miss earlier possible
1159	* free areas, smaller allocations will eventually fill those holes.
1160	*/
1161	static unsigned long pcpu_find_zero_area(unsigned long *map,
1162	unsigned long size,
1163	unsigned long start,
1164	unsigned long nr,
1165	unsigned long align_mask,
1166	unsigned long *largest_off,
1167	unsigned long *largest_bits)
1168	{
1169	unsigned long index, end, i, area_off, area_bits;
1170	again:
1171	index = find_next_zero_bit(addr: map, size, offset: start);
1172
1173	/* Align allocation */
1174	index = __ALIGN_MASK(index, align_mask);
1175	area_off = index;
1176
1177	end = index + nr;
1178	if (end > size)
1179	return end;
1180	i = find_next_bit(addr: map, size: end, offset: index);
1181	if (i < end) {
1182	area_bits = i - area_off;
1183	/* remember largest unused area with best alignment */
1184	if (area_bits > *largest_bits ||
1185	(area_bits == *largest_bits && *largest_off &&
1186	(!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1187	*largest_off = area_off;
1188	*largest_bits = area_bits;
1189	}
1190
1191	start = i + 1;
1192	goto again;
1193	}
1194	return index;
1195	}
1196
1197	/**
1198	* pcpu_alloc_area - allocates an area from a pcpu_chunk
1199	* @chunk: chunk of interest
1200	* @alloc_bits: size of request in allocation units
1201	* @align: alignment of area (max PAGE_SIZE)
1202	* @start: bit_off to start searching
1203	*
1204	* This function takes in a @start offset to begin searching to fit an
1205	* allocation of @alloc_bits with alignment @align. It needs to scan
1206	* the allocation map because if it fits within the block's contig hint,
1207	* @start will be block->first_free. This is an attempt to fill the
1208	* allocation prior to breaking the contig hint. The allocation and
1209	* boundary maps are updated accordingly if it confirms a valid
1210	* free area.
1211	*
1212	* RETURNS:
1213	* Allocated addr offset in @chunk on success.
1214	* -1 if no matching area is found.
1215	*/
1216	static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1217	size_t align, int start)
1218	{
1219	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1220	size_t align_mask = (align) ? (align - 1) : 0;
1221	unsigned long area_off = 0, area_bits = 0;
1222	int bit_off, end, oslot;
1223
1224	lockdep_assert_held(&pcpu_lock);
1225
1226	oslot = pcpu_chunk_slot(chunk);
1227
1228	/*
1229	* Search to find a fit.
1230	*/
1231	end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1232	pcpu_chunk_map_bits(chunk));
1233	bit_off = pcpu_find_zero_area(map: chunk->alloc_map, size: end, start, nr: alloc_bits,
1234	align_mask, largest_off: &area_off, largest_bits: &area_bits);
1235	if (bit_off >= end)
1236	return -1;
1237
1238	if (area_bits)
1239	pcpu_block_update_scan(chunk, bit_off: area_off, bits: area_bits);
1240
1241	/* update alloc map */
1242	bitmap_set(map: chunk->alloc_map, start: bit_off, nbits: alloc_bits);
1243
1244	/* update boundary map */
1245	set_bit(nr: bit_off, addr: chunk->bound_map);
1246	bitmap_clear(map: chunk->bound_map, start: bit_off + 1, nbits: alloc_bits - 1);
1247	set_bit(nr: bit_off + alloc_bits, addr: chunk->bound_map);
1248
1249	chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1250
1251	/* update first free bit */
1252	if (bit_off == chunk_md->first_free)
1253	chunk_md->first_free = find_next_zero_bit(
1254	addr: chunk->alloc_map,
1255	size: pcpu_chunk_map_bits(chunk),
1256	offset: bit_off + alloc_bits);
1257
1258	pcpu_block_update_hint_alloc(chunk, bit_off, bits: alloc_bits);
1259
1260	pcpu_chunk_relocate(chunk, oslot);
1261
1262	return bit_off * PCPU_MIN_ALLOC_SIZE;
1263	}
1264
1265	/**
1266	* pcpu_free_area - frees the corresponding offset
1267	* @chunk: chunk of interest
1268	* @off: addr offset into chunk
1269	*
1270	* This function determines the size of an allocation to free using
1271	* the boundary bitmap and clears the allocation map.
1272	*
1273	* RETURNS:
1274	* Number of freed bytes.
1275	*/
1276	static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1277	{
1278	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1279	int bit_off, bits, end, oslot, freed;
1280
1281	lockdep_assert_held(&pcpu_lock);
1282	pcpu_stats_area_dealloc(chunk);
1283
1284	oslot = pcpu_chunk_slot(chunk);
1285
1286	bit_off = off / PCPU_MIN_ALLOC_SIZE;
1287
1288	/* find end index */
1289	end = find_next_bit(addr: chunk->bound_map, size: pcpu_chunk_map_bits(chunk),
1290	offset: bit_off + 1);
1291	bits = end - bit_off;
1292	bitmap_clear(map: chunk->alloc_map, start: bit_off, nbits: bits);
1293
1294	freed = bits * PCPU_MIN_ALLOC_SIZE;
1295
1296	/* update metadata */
1297	chunk->free_bytes += freed;
1298
1299	/* update first free bit */
1300	chunk_md->first_free = min(chunk_md->first_free, bit_off);
1301
1302	pcpu_block_update_hint_free(chunk, bit_off, bits);
1303
1304	pcpu_chunk_relocate(chunk, oslot);
1305
1306	return freed;
1307	}
1308
1309	static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1310	{
1311	block->scan_hint = 0;
1312	block->contig_hint = nr_bits;
1313	block->left_free = nr_bits;
1314	block->right_free = nr_bits;
1315	block->first_free = 0;
1316	block->nr_bits = nr_bits;
1317	}
1318
1319	static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1320	{
1321	struct pcpu_block_md *md_block;
1322
1323	/* init the chunk's block */
1324	pcpu_init_md_block(block: &chunk->chunk_md, nr_bits: pcpu_chunk_map_bits(chunk));
1325
1326	for (md_block = chunk->md_blocks;
1327	md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1328	md_block++)
1329	pcpu_init_md_block(block: md_block, PCPU_BITMAP_BLOCK_BITS);
1330	}
1331
1332	/**
1333	* pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1334	* @tmp_addr: the start of the region served
1335	* @map_size: size of the region served
1336	*
1337	* This is responsible for creating the chunks that serve the first chunk. The
1338	* base_addr is page aligned down of @tmp_addr while the region end is page
1339	* aligned up. Offsets are kept track of to determine the region served. All
1340	* this is done to appease the bitmap allocator in avoiding partial blocks.
1341	*
1342	* RETURNS:
1343	* Chunk serving the region at @tmp_addr of @map_size.
1344	*/
1345	static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1346	int map_size)
1347	{
1348	struct pcpu_chunk *chunk;
1349	unsigned long aligned_addr;
1350	int start_offset, offset_bits, region_size, region_bits;
1351	size_t alloc_size;
1352
1353	/* region calculations */
1354	aligned_addr = tmp_addr & PAGE_MASK;
1355
1356	start_offset = tmp_addr - aligned_addr;
1357	region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
1358
1359	/* allocate chunk */
1360	alloc_size = struct_size(chunk, populated,
1361	BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1362	chunk = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
1363	if (!chunk)
1364	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
1365	alloc_size);
1366
1367	INIT_LIST_HEAD(list: &chunk->list);
1368
1369	chunk->base_addr = (void *)aligned_addr;
1370	chunk->start_offset = start_offset;
1371	chunk->end_offset = region_size - chunk->start_offset - map_size;
1372
1373	chunk->nr_pages = region_size >> PAGE_SHIFT;
1374	region_bits = pcpu_chunk_map_bits(chunk);
1375
1376	alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1377	chunk->alloc_map = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
1378	if (!chunk->alloc_map)
1379	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
1380	alloc_size);
1381
1382	alloc_size =
1383	BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1384	chunk->bound_map = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
1385	if (!chunk->bound_map)
1386	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
1387	alloc_size);
1388
1389	alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1390	chunk->md_blocks = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
1391	if (!chunk->md_blocks)
1392	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
1393	alloc_size);
1394
1395	#ifdef CONFIG_MEMCG_KMEM
1396	/* first chunk is free to use */
1397	chunk->obj_cgroups = NULL;
1398	#endif
1399	pcpu_init_md_blocks(chunk);
1400
1401	/* manage populated page bitmap */
1402	chunk->immutable = true;
1403	bitmap_fill(dst: chunk->populated, nbits: chunk->nr_pages);
1404	chunk->nr_populated = chunk->nr_pages;
1405	chunk->nr_empty_pop_pages = chunk->nr_pages;
1406
1407	chunk->free_bytes = map_size;
1408
1409	if (chunk->start_offset) {
1410	/* hide the beginning of the bitmap */
1411	offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1412	bitmap_set(map: chunk->alloc_map, start: 0, nbits: offset_bits);
1413	set_bit(nr: 0, addr: chunk->bound_map);
1414	set_bit(nr: offset_bits, addr: chunk->bound_map);
1415
1416	chunk->chunk_md.first_free = offset_bits;
1417
1418	pcpu_block_update_hint_alloc(chunk, bit_off: 0, bits: offset_bits);
1419	}
1420
1421	if (chunk->end_offset) {
1422	/* hide the end of the bitmap */
1423	offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1424	bitmap_set(map: chunk->alloc_map,
1425	start: pcpu_chunk_map_bits(chunk) - offset_bits,
1426	nbits: offset_bits);
1427	set_bit(nr: (start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1428	addr: chunk->bound_map);
1429	set_bit(nr: region_bits, addr: chunk->bound_map);
1430
1431	pcpu_block_update_hint_alloc(chunk, bit_off: pcpu_chunk_map_bits(chunk)
1432	- offset_bits, bits: offset_bits);
1433	}
1434
1435	return chunk;
1436	}
1437
1438	static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1439	{
1440	struct pcpu_chunk *chunk;
1441	int region_bits;
1442
1443	chunk = pcpu_mem_zalloc(size: pcpu_chunk_struct_size, gfp);
1444	if (!chunk)
1445	return NULL;
1446
1447	INIT_LIST_HEAD(list: &chunk->list);
1448	chunk->nr_pages = pcpu_unit_pages;
1449	region_bits = pcpu_chunk_map_bits(chunk);
1450
1451	chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1452	sizeof(chunk->alloc_map[0]), gfp);
1453	if (!chunk->alloc_map)
1454	goto alloc_map_fail;
1455
1456	chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1457	sizeof(chunk->bound_map[0]), gfp);
1458	if (!chunk->bound_map)
1459	goto bound_map_fail;
1460
1461	chunk->md_blocks = pcpu_mem_zalloc(size: pcpu_chunk_nr_blocks(chunk) *
1462	sizeof(chunk->md_blocks[0]), gfp);
1463	if (!chunk->md_blocks)
1464	goto md_blocks_fail;
1465
1466	#ifdef CONFIG_MEMCG_KMEM
1467	if (!mem_cgroup_kmem_disabled()) {
1468	chunk->obj_cgroups =
1469	pcpu_mem_zalloc(size: pcpu_chunk_map_bits(chunk) *
1470	sizeof(struct obj_cgroup *), gfp);
1471	if (!chunk->obj_cgroups)
1472	goto objcg_fail;
1473	}
1474	#endif
1475
1476	pcpu_init_md_blocks(chunk);
1477
1478	/* init metadata */
1479	chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1480
1481	return chunk;
1482
1483	#ifdef CONFIG_MEMCG_KMEM
1484	objcg_fail:
1485	pcpu_mem_free(ptr: chunk->md_blocks);
1486	#endif
1487	md_blocks_fail:
1488	pcpu_mem_free(ptr: chunk->bound_map);
1489	bound_map_fail:
1490	pcpu_mem_free(ptr: chunk->alloc_map);
1491	alloc_map_fail:
1492	pcpu_mem_free(ptr: chunk);
1493
1494	return NULL;
1495	}
1496
1497	static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1498	{
1499	if (!chunk)
1500	return;
1501	#ifdef CONFIG_MEMCG_KMEM
1502	pcpu_mem_free(ptr: chunk->obj_cgroups);
1503	#endif
1504	pcpu_mem_free(ptr: chunk->md_blocks);
1505	pcpu_mem_free(ptr: chunk->bound_map);
1506	pcpu_mem_free(ptr: chunk->alloc_map);
1507	pcpu_mem_free(ptr: chunk);
1508	}
1509
1510	/**
1511	* pcpu_chunk_populated - post-population bookkeeping
1512	* @chunk: pcpu_chunk which got populated
1513	* @page_start: the start page
1514	* @page_end: the end page
1515	*
1516	* Pages in [@page_start,@page_end) have been populated to @chunk. Update
1517	* the bookkeeping information accordingly. Must be called after each
1518	* successful population.
1519	*/
1520	static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1521	int page_end)
1522	{
1523	int nr = page_end - page_start;
1524
1525	lockdep_assert_held(&pcpu_lock);
1526
1527	bitmap_set(map: chunk->populated, start: page_start, nbits: nr);
1528	chunk->nr_populated += nr;
1529	pcpu_nr_populated += nr;
1530
1531	pcpu_update_empty_pages(chunk, nr);
1532	}
1533
1534	/**
1535	* pcpu_chunk_depopulated - post-depopulation bookkeeping
1536	* @chunk: pcpu_chunk which got depopulated
1537	* @page_start: the start page
1538	* @page_end: the end page
1539	*
1540	* Pages in [@page_start,@page_end) have been depopulated from @chunk.
1541	* Update the bookkeeping information accordingly. Must be called after
1542	* each successful depopulation.
1543	*/
1544	static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1545	int page_start, int page_end)
1546	{
1547	int nr = page_end - page_start;
1548
1549	lockdep_assert_held(&pcpu_lock);
1550
1551	bitmap_clear(map: chunk->populated, start: page_start, nbits: nr);
1552	chunk->nr_populated -= nr;
1553	pcpu_nr_populated -= nr;
1554
1555	pcpu_update_empty_pages(chunk, nr: -nr);
1556	}
1557
1558	/*
1559	* Chunk management implementation.
1560	*
1561	* To allow different implementations, chunk alloc/free and
1562	* [de]population are implemented in a separate file which is pulled
1563	* into this file and compiled together. The following functions
1564	* should be implemented.
1565	*
1566	* pcpu_populate_chunk - populate the specified range of a chunk
1567	* pcpu_depopulate_chunk - depopulate the specified range of a chunk
1568	* pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk
1569	* pcpu_create_chunk - create a new chunk
1570	* pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1571	* pcpu_addr_to_page - translate address to physical address
1572	* pcpu_verify_alloc_info - check alloc_info is acceptable during init
1573	*/
1574	static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1575	int page_start, int page_end, gfp_t gfp);
1576	static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1577	int page_start, int page_end);
1578	static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1579	int page_start, int page_end);
1580	static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1581	static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1582	static struct page *pcpu_addr_to_page(void *addr);
1583	static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1584
1585	#ifdef CONFIG_NEED_PER_CPU_KM
1586	#include "percpu-km.c"
1587	#else
1588	#include "percpu-vm.c"
1589	#endif
1590
1591	/**
1592	* pcpu_chunk_addr_search - determine chunk containing specified address
1593	* @addr: address for which the chunk needs to be determined.
1594	*
1595	* This is an internal function that handles all but static allocations.
1596	* Static percpu address values should never be passed into the allocator.
1597	*
1598	* RETURNS:
1599	* The address of the found chunk.
1600	*/
1601	static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1602	{
1603	/* is it in the dynamic region (first chunk)? */
1604	if (pcpu_addr_in_chunk(chunk: pcpu_first_chunk, addr))
1605	return pcpu_first_chunk;
1606
1607	/* is it in the reserved region? */
1608	if (pcpu_addr_in_chunk(chunk: pcpu_reserved_chunk, addr))
1609	return pcpu_reserved_chunk;
1610
1611	/*
1612	* The address is relative to unit0 which might be unused and
1613	* thus unmapped. Offset the address to the unit space of the
1614	* current processor before looking it up in the vmalloc
1615	* space. Note that any possible cpu id can be used here, so
1616	* there's no need to worry about preemption or cpu hotplug.
1617	*/
1618	addr += pcpu_unit_offsets[raw_smp_processor_id()];
1619	return pcpu_get_page_chunk(page: pcpu_addr_to_page(addr));
1620	}
1621
1622	#ifdef CONFIG_MEMCG_KMEM
1623	static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1624	struct obj_cgroup **objcgp)
1625	{
1626	struct obj_cgroup *objcg;
1627
1628	if (!memcg_kmem_online() || !(gfp & __GFP_ACCOUNT))
1629	return true;
1630
1631	objcg = current_obj_cgroup();
1632	if (!objcg)
1633	return true;
1634
1635	if (obj_cgroup_charge(objcg, gfp, size: pcpu_obj_full_size(size)))
1636	return false;
1637
1638	*objcgp = objcg;
1639	return true;
1640	}
1641
1642	static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1643	struct pcpu_chunk *chunk, int off,
1644	size_t size)
1645	{
1646	if (!objcg)
1647	return;
1648
1649	if (likely(chunk && chunk->obj_cgroups)) {
1650	obj_cgroup_get(objcg);
1651	chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1652
1653	rcu_read_lock();
1654	mod_memcg_state(memcg: obj_cgroup_memcg(objcg), idx: MEMCG_PERCPU_B,
1655	val: pcpu_obj_full_size(size));
1656	rcu_read_unlock();
1657	} else {
1658	obj_cgroup_uncharge(objcg, size: pcpu_obj_full_size(size));
1659	}
1660	}
1661
1662	static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1663	{
1664	struct obj_cgroup *objcg;
1665
1666	if (unlikely(!chunk->obj_cgroups))
1667	return;
1668
1669	objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1670	if (!objcg)
1671	return;
1672	chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1673
1674	obj_cgroup_uncharge(objcg, size: pcpu_obj_full_size(size));
1675
1676	rcu_read_lock();
1677	mod_memcg_state(memcg: obj_cgroup_memcg(objcg), idx: MEMCG_PERCPU_B,
1678	val: -pcpu_obj_full_size(size));
1679	rcu_read_unlock();
1680
1681	obj_cgroup_put(objcg);
1682	}
1683
1684	#else /* CONFIG_MEMCG_KMEM */
1685	static bool
1686	pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1687	{
1688	return true;
1689	}
1690
1691	static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1692	struct pcpu_chunk *chunk, int off,
1693	size_t size)
1694	{
1695	}
1696
1697	static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1698	{
1699	}
1700	#endif /* CONFIG_MEMCG_KMEM */
1701
1702	/**
1703	* pcpu_alloc - the percpu allocator
1704	* @size: size of area to allocate in bytes
1705	* @align: alignment of area (max PAGE_SIZE)
1706	* @reserved: allocate from the reserved chunk if available
1707	* @gfp: allocation flags
1708	*
1709	* Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1710	* contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1711	* then no warning will be triggered on invalid or failed allocation
1712	* requests.
1713	*
1714	* RETURNS:
1715	* Percpu pointer to the allocated area on success, NULL on failure.
1716	*/
1717	static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1718	gfp_t gfp)
1719	{
1720	gfp_t pcpu_gfp;
1721	bool is_atomic;
1722	bool do_warn;
1723	struct obj_cgroup *objcg = NULL;
1724	static int warn_limit = 10;
1725	struct pcpu_chunk *chunk, *next;
1726	const char *err;
1727	int slot, off, cpu, ret;
1728	unsigned long flags;
1729	void __percpu *ptr;
1730	size_t bits, bit_align;
1731
1732	gfp = current_gfp_context(flags: gfp);
1733	/* whitelisted flags that can be passed to the backing allocators */
1734	pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1735	is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1736	do_warn = !(gfp & __GFP_NOWARN);
1737
1738	/*
1739	* There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1740	* therefore alignment must be a minimum of that many bytes.
1741	* An allocation may have internal fragmentation from rounding up
1742	* of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1743	*/
1744	if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1745	align = PCPU_MIN_ALLOC_SIZE;
1746
1747	size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1748	bits = size >> PCPU_MIN_ALLOC_SHIFT;
1749	bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1750
1751	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1752	!is_power_of_2(align))) {
1753	WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1754	size, align);
1755	return NULL;
1756	}
1757
1758	if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1759	return NULL;
1760
1761	if (!is_atomic) {
1762	/*
1763	* pcpu_balance_workfn() allocates memory under this mutex,
1764	* and it may wait for memory reclaim. Allow current task
1765	* to become OOM victim, in case of memory pressure.
1766	*/
1767	if (gfp & __GFP_NOFAIL) {
1768	mutex_lock(&pcpu_alloc_mutex);
1769	} else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1770	pcpu_memcg_post_alloc_hook(objcg, NULL, off: 0, size);
1771	return NULL;
1772	}
1773	}
1774
1775	spin_lock_irqsave(&pcpu_lock, flags);
1776
1777	/* serve reserved allocations from the reserved chunk if available */
1778	if (reserved && pcpu_reserved_chunk) {
1779	chunk = pcpu_reserved_chunk;
1780
1781	off = pcpu_find_block_fit(chunk, alloc_bits: bits, align: bit_align, pop_only: is_atomic);
1782	if (off < 0) {
1783	err = "alloc from reserved chunk failed";
1784	goto fail_unlock;
1785	}
1786
1787	off = pcpu_alloc_area(chunk, alloc_bits: bits, align: bit_align, start: off);
1788	if (off >= 0)
1789	goto area_found;
1790
1791	err = "alloc from reserved chunk failed";
1792	goto fail_unlock;
1793	}
1794
1795	restart:
1796	/* search through normal chunks */
1797	for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1798	list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1799	list) {
1800	off = pcpu_find_block_fit(chunk, alloc_bits: bits, align: bit_align,
1801	pop_only: is_atomic);
1802	if (off < 0) {
1803	if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1804	pcpu_chunk_move(chunk, slot: 0);
1805	continue;
1806	}
1807
1808	off = pcpu_alloc_area(chunk, alloc_bits: bits, align: bit_align, start: off);
1809	if (off >= 0) {
1810	pcpu_reintegrate_chunk(chunk);
1811	goto area_found;
1812	}
1813	}
1814	}
1815
1816	spin_unlock_irqrestore(lock: &pcpu_lock, flags);
1817
1818	if (is_atomic) {
1819	err = "atomic alloc failed, no space left";
1820	goto fail;
1821	}
1822
1823	/* No space left. Create a new chunk. */
1824	if (list_empty(head: &pcpu_chunk_lists[pcpu_free_slot])) {
1825	chunk = pcpu_create_chunk(gfp: pcpu_gfp);
1826	if (!chunk) {
1827	err = "failed to allocate new chunk";
1828	goto fail;
1829	}
1830
1831	spin_lock_irqsave(&pcpu_lock, flags);
1832	pcpu_chunk_relocate(chunk, oslot: -1);
1833	} else {
1834	spin_lock_irqsave(&pcpu_lock, flags);
1835	}
1836
1837	goto restart;
1838
1839	area_found:
1840	pcpu_stats_area_alloc(chunk, size);
1841	spin_unlock_irqrestore(lock: &pcpu_lock, flags);
1842
1843	/* populate if not all pages are already there */
1844	if (!is_atomic) {
1845	unsigned int page_end, rs, re;
1846
1847	rs = PFN_DOWN(off);
1848	page_end = PFN_UP(off + size);
1849
1850	for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
1851	WARN_ON(chunk->immutable);
1852
1853	ret = pcpu_populate_chunk(chunk, page_start: rs, page_end: re, gfp: pcpu_gfp);
1854
1855	spin_lock_irqsave(&pcpu_lock, flags);
1856	if (ret) {
1857	pcpu_free_area(chunk, off);
1858	err = "failed to populate";
1859	goto fail_unlock;
1860	}
1861	pcpu_chunk_populated(chunk, page_start: rs, page_end: re);
1862	spin_unlock_irqrestore(lock: &pcpu_lock, flags);
1863	}
1864
1865	mutex_unlock(lock: &pcpu_alloc_mutex);
1866	}
1867
1868	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1869	pcpu_schedule_balance_work();
1870
1871	/* clear the areas and return address relative to base address */
1872	for_each_possible_cpu(cpu)
1873	memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1874
1875	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1876	kmemleak_alloc_percpu(ptr, size, gfp);
1877
1878	trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1879	base_addr: chunk->base_addr, off, ptr,
1880	bytes_alloc: pcpu_obj_full_size(size), gfp_flags: gfp);
1881
1882	pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1883
1884	return ptr;
1885
1886	fail_unlock:
1887	spin_unlock_irqrestore(lock: &pcpu_lock, flags);
1888	fail:
1889	trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1890
1891	if (do_warn && warn_limit) {
1892	pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1893	size, align, is_atomic, err);
1894	if (!is_atomic)
1895	dump_stack();
1896	if (!--warn_limit)
1897	pr_info("limit reached, disable warning\n");
1898	}
1899
1900	if (is_atomic) {
1901	/* see the flag handling in pcpu_balance_workfn() */
1902	pcpu_atomic_alloc_failed = true;
1903	pcpu_schedule_balance_work();
1904	} else {
1905	mutex_unlock(lock: &pcpu_alloc_mutex);
1906	}
1907
1908	pcpu_memcg_post_alloc_hook(objcg, NULL, off: 0, size);
1909
1910	return NULL;
1911	}
1912
1913	/**
1914	* __alloc_percpu_gfp - allocate dynamic percpu area
1915	* @size: size of area to allocate in bytes
1916	* @align: alignment of area (max PAGE_SIZE)
1917	* @gfp: allocation flags
1918	*
1919	* Allocate zero-filled percpu area of @size bytes aligned at @align. If
1920	* @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1921	* be called from any context but is a lot more likely to fail. If @gfp
1922	* has __GFP_NOWARN then no warning will be triggered on invalid or failed
1923	* allocation requests.
1924	*
1925	* RETURNS:
1926	* Percpu pointer to the allocated area on success, NULL on failure.
1927	*/
1928	void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1929	{
1930	return pcpu_alloc(size, align, reserved: false, gfp);
1931	}
1932	EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1933
1934	/**
1935	* __alloc_percpu - allocate dynamic percpu area
1936	* @size: size of area to allocate in bytes
1937	* @align: alignment of area (max PAGE_SIZE)
1938	*
1939	* Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1940	*/
1941	void __percpu *__alloc_percpu(size_t size, size_t align)
1942	{
1943	return pcpu_alloc(size, align, reserved: false, GFP_KERNEL);
1944	}
1945	EXPORT_SYMBOL_GPL(__alloc_percpu);
1946
1947	/**
1948	* __alloc_reserved_percpu - allocate reserved percpu area
1949	* @size: size of area to allocate in bytes
1950	* @align: alignment of area (max PAGE_SIZE)
1951	*
1952	* Allocate zero-filled percpu area of @size bytes aligned at @align
1953	* from reserved percpu area if arch has set it up; otherwise,
1954	* allocation is served from the same dynamic area. Might sleep.
1955	* Might trigger writeouts.
1956	*
1957	* CONTEXT:
1958	* Does GFP_KERNEL allocation.
1959	*
1960	* RETURNS:
1961	* Percpu pointer to the allocated area on success, NULL on failure.
1962	*/
1963	void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1964	{
1965	return pcpu_alloc(size, align, reserved: true, GFP_KERNEL);
1966	}
1967
1968	/**
1969	* pcpu_balance_free - manage the amount of free chunks
1970	* @empty_only: free chunks only if there are no populated pages
1971	*
1972	* If empty_only is %false, reclaim all fully free chunks regardless of the
1973	* number of populated pages. Otherwise, only reclaim chunks that have no
1974	* populated pages.
1975	*
1976	* CONTEXT:
1977	* pcpu_lock (can be dropped temporarily)
1978	*/
1979	static void pcpu_balance_free(bool empty_only)
1980	{
1981	LIST_HEAD(to_free);
1982	struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1983	struct pcpu_chunk *chunk, *next;
1984
1985	lockdep_assert_held(&pcpu_lock);
1986
1987	/*
1988	* There's no reason to keep around multiple unused chunks and VM
1989	* areas can be scarce. Destroy all free chunks except for one.
1990	*/
1991	list_for_each_entry_safe(chunk, next, free_head, list) {
1992	WARN_ON(chunk->immutable);
1993
1994	/* spare the first one */
1995	if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1996	continue;
1997
1998	if (!empty_only || chunk->nr_empty_pop_pages == 0)
1999	list_move(list: &chunk->list, head: &to_free);
2000	}
2001
2002	if (list_empty(head: &to_free))
2003	return;
2004
2005	spin_unlock_irq(lock: &pcpu_lock);
2006	list_for_each_entry_safe(chunk, next, &to_free, list) {
2007	unsigned int rs, re;
2008
2009	for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2010	pcpu_depopulate_chunk(chunk, page_start: rs, page_end: re);
2011	spin_lock_irq(lock: &pcpu_lock);
2012	pcpu_chunk_depopulated(chunk, page_start: rs, page_end: re);
2013	spin_unlock_irq(lock: &pcpu_lock);
2014	}
2015	pcpu_destroy_chunk(chunk);
2016	cond_resched();
2017	}
2018	spin_lock_irq(lock: &pcpu_lock);
2019	}
2020
2021	/**
2022	* pcpu_balance_populated - manage the amount of populated pages
2023	*
2024	* Maintain a certain amount of populated pages to satisfy atomic allocations.
2025	* It is possible that this is called when physical memory is scarce causing
2026	* OOM killer to be triggered. We should avoid doing so until an actual
2027	* allocation causes the failure as it is possible that requests can be
2028	* serviced from already backed regions.
2029	*
2030	* CONTEXT:
2031	* pcpu_lock (can be dropped temporarily)
2032	*/
2033	static void pcpu_balance_populated(void)
2034	{
2035	/* gfp flags passed to underlying allocators */
2036	const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2037	struct pcpu_chunk *chunk;
2038	int slot, nr_to_pop, ret;
2039
2040	lockdep_assert_held(&pcpu_lock);
2041
2042	/*
2043	* Ensure there are certain number of free populated pages for
2044	* atomic allocs. Fill up from the most packed so that atomic
2045	* allocs don't increase fragmentation. If atomic allocation
2046	* failed previously, always populate the maximum amount. This
2047	* should prevent atomic allocs larger than PAGE_SIZE from keeping
2048	* failing indefinitely; however, large atomic allocs are not
2049	* something we support properly and can be highly unreliable and
2050	* inefficient.
2051	*/
2052	retry_pop:
2053	if (pcpu_atomic_alloc_failed) {
2054	nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2055	/* best effort anyway, don't worry about synchronization */
2056	pcpu_atomic_alloc_failed = false;
2057	} else {
2058	nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2059	pcpu_nr_empty_pop_pages,
2060	0, PCPU_EMPTY_POP_PAGES_HIGH);
2061	}
2062
2063	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2064	unsigned int nr_unpop = 0, rs, re;
2065
2066	if (!nr_to_pop)
2067	break;
2068
2069	list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2070	nr_unpop = chunk->nr_pages - chunk->nr_populated;
2071	if (nr_unpop)
2072	break;
2073	}
2074
2075	if (!nr_unpop)
2076	continue;
2077
2078	/* @chunk can't go away while pcpu_alloc_mutex is held */
2079	for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2080	int nr = min_t(int, re - rs, nr_to_pop);
2081
2082	spin_unlock_irq(lock: &pcpu_lock);
2083	ret = pcpu_populate_chunk(chunk, page_start: rs, page_end: rs + nr, gfp);
2084	cond_resched();
2085	spin_lock_irq(lock: &pcpu_lock);
2086	if (!ret) {
2087	nr_to_pop -= nr;
2088	pcpu_chunk_populated(chunk, page_start: rs, page_end: rs + nr);
2089	} else {
2090	nr_to_pop = 0;
2091	}
2092
2093	if (!nr_to_pop)
2094	break;
2095	}
2096	}
2097
2098	if (nr_to_pop) {
2099	/* ran out of chunks to populate, create a new one and retry */
2100	spin_unlock_irq(lock: &pcpu_lock);
2101	chunk = pcpu_create_chunk(gfp);
2102	cond_resched();
2103	spin_lock_irq(lock: &pcpu_lock);
2104	if (chunk) {
2105	pcpu_chunk_relocate(chunk, oslot: -1);
2106	goto retry_pop;
2107	}
2108	}
2109	}
2110
2111	/**
2112	* pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2113	*
2114	* Scan over chunks in the depopulate list and try to release unused populated
2115	* pages back to the system. Depopulated chunks are sidelined to prevent
2116	* repopulating these pages unless required. Fully free chunks are reintegrated
2117	* and freed accordingly (1 is kept around). If we drop below the empty
2118	* populated pages threshold, reintegrate the chunk if it has empty free pages.
2119	* Each chunk is scanned in the reverse order to keep populated pages close to
2120	* the beginning of the chunk.
2121	*
2122	* CONTEXT:
2123	* pcpu_lock (can be dropped temporarily)
2124	*
2125	*/
2126	static void pcpu_reclaim_populated(void)
2127	{
2128	struct pcpu_chunk *chunk;
2129	struct pcpu_block_md *block;
2130	int freed_page_start, freed_page_end;
2131	int i, end;
2132	bool reintegrate;
2133
2134	lockdep_assert_held(&pcpu_lock);
2135
2136	/*
2137	* Once a chunk is isolated to the to_depopulate list, the chunk is no
2138	* longer discoverable to allocations whom may populate pages. The only
2139	* other accessor is the free path which only returns area back to the
2140	* allocator not touching the populated bitmap.
2141	*/
2142	while ((chunk = list_first_entry_or_null(
2143	&pcpu_chunk_lists[pcpu_to_depopulate_slot],
2144	struct pcpu_chunk, list))) {
2145	WARN_ON(chunk->immutable);
2146
2147	/*
2148	* Scan chunk's pages in the reverse order to keep populated
2149	* pages close to the beginning of the chunk.
2150	*/
2151	freed_page_start = chunk->nr_pages;
2152	freed_page_end = 0;
2153	reintegrate = false;
2154	for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2155	/* no more work to do */
2156	if (chunk->nr_empty_pop_pages == 0)
2157	break;
2158
2159	/* reintegrate chunk to prevent atomic alloc failures */
2160	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2161	reintegrate = true;
2162	break;
2163	}
2164
2165	/*
2166	* If the page is empty and populated, start or
2167	* extend the (i, end) range. If i == 0, decrease
2168	* i and perform the depopulation to cover the last
2169	* (first) page in the chunk.
2170	*/
2171	block = chunk->md_blocks + i;
2172	if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2173	test_bit(i, chunk->populated)) {
2174	if (end == -1)
2175	end = i;
2176	if (i > 0)
2177	continue;
2178	i--;
2179	}
2180
2181	/* depopulate if there is an active range */
2182	if (end == -1)
2183	continue;
2184
2185	spin_unlock_irq(lock: &pcpu_lock);
2186	pcpu_depopulate_chunk(chunk, page_start: i + 1, page_end: end + 1);
2187	cond_resched();
2188	spin_lock_irq(lock: &pcpu_lock);
2189
2190	pcpu_chunk_depopulated(chunk, page_start: i + 1, page_end: end + 1);
2191	freed_page_start = min(freed_page_start, i + 1);
2192	freed_page_end = max(freed_page_end, end + 1);
2193
2194	/* reset the range and continue */
2195	end = -1;
2196	}
2197
2198	/* batch tlb flush per chunk to amortize cost */
2199	if (freed_page_start < freed_page_end) {
2200	spin_unlock_irq(lock: &pcpu_lock);
2201	pcpu_post_unmap_tlb_flush(chunk,
2202	page_start: freed_page_start,
2203	page_end: freed_page_end);
2204	cond_resched();
2205	spin_lock_irq(lock: &pcpu_lock);
2206	}
2207
2208	if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2209	pcpu_reintegrate_chunk(chunk);
2210	else
2211	list_move_tail(list: &chunk->list,
2212	head: &pcpu_chunk_lists[pcpu_sidelined_slot]);
2213	}
2214	}
2215
2216	/**
2217	* pcpu_balance_workfn - manage the amount of free chunks and populated pages
2218	* @work: unused
2219	*
2220	* For each chunk type, manage the number of fully free chunks and the number of
2221	* populated pages. An important thing to consider is when pages are freed and
2222	* how they contribute to the global counts.
2223	*/
2224	static void pcpu_balance_workfn(struct work_struct *work)
2225	{
2226	/*
2227	* pcpu_balance_free() is called twice because the first time we may
2228	* trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2229	* to grow other chunks. This then gives pcpu_reclaim_populated() time
2230	* to move fully free chunks to the active list to be freed if
2231	* appropriate.
2232	*/
2233	mutex_lock(&pcpu_alloc_mutex);
2234	spin_lock_irq(lock: &pcpu_lock);
2235
2236	pcpu_balance_free(empty_only: false);
2237	pcpu_reclaim_populated();
2238	pcpu_balance_populated();
2239	pcpu_balance_free(empty_only: true);
2240
2241	spin_unlock_irq(lock: &pcpu_lock);
2242	mutex_unlock(lock: &pcpu_alloc_mutex);
2243	}
2244
2245	/**
2246	* pcpu_alloc_size - the size of the dynamic percpu area
2247	* @ptr: pointer to the dynamic percpu area
2248	*
2249	* Returns the size of the @ptr allocation. This is undefined for statically
2250	* defined percpu variables as there is no corresponding chunk->bound_map.
2251	*
2252	* RETURNS:
2253	* The size of the dynamic percpu area.
2254	*
2255	* CONTEXT:
2256	* Can be called from atomic context.
2257	*/
2258	size_t pcpu_alloc_size(void __percpu *ptr)
2259	{
2260	struct pcpu_chunk *chunk;
2261	unsigned long bit_off, end;
2262	void *addr;
2263
2264	if (!ptr)
2265	return 0;
2266
2267	addr = __pcpu_ptr_to_addr(ptr);
2268	/* No pcpu_lock here: ptr has not been freed, so chunk is still alive */
2269	chunk = pcpu_chunk_addr_search(addr);
2270	bit_off = (addr - chunk->base_addr) / PCPU_MIN_ALLOC_SIZE;
2271	end = find_next_bit(addr: chunk->bound_map, size: pcpu_chunk_map_bits(chunk),
2272	offset: bit_off + 1);
2273	return (end - bit_off) * PCPU_MIN_ALLOC_SIZE;
2274	}
2275
2276	/**
2277	* free_percpu - free percpu area
2278	* @ptr: pointer to area to free
2279	*
2280	* Free percpu area @ptr.
2281	*
2282	* CONTEXT:
2283	* Can be called from atomic context.
2284	*/
2285	void free_percpu(void __percpu *ptr)
2286	{
2287	void *addr;
2288	struct pcpu_chunk *chunk;
2289	unsigned long flags;
2290	int size, off;
2291	bool need_balance = false;
2292
2293	if (!ptr)
2294	return;
2295
2296	kmemleak_free_percpu(ptr);
2297
2298	addr = __pcpu_ptr_to_addr(ptr);
2299	chunk = pcpu_chunk_addr_search(addr);
2300	off = addr - chunk->base_addr;
2301
2302	spin_lock_irqsave(&pcpu_lock, flags);
2303	size = pcpu_free_area(chunk, off);
2304
2305	pcpu_memcg_free_hook(chunk, off, size);
2306
2307	/*
2308	* If there are more than one fully free chunks, wake up grim reaper.
2309	* If the chunk is isolated, it may be in the process of being
2310	* reclaimed. Let reclaim manage cleaning up of that chunk.
2311	*/
2312	if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2313	struct pcpu_chunk *pos;
2314
2315	list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2316	if (pos != chunk) {
2317	need_balance = true;
2318	break;
2319	}
2320	} else if (pcpu_should_reclaim_chunk(chunk)) {
2321	pcpu_isolate_chunk(chunk);
2322	need_balance = true;
2323	}
2324
2325	trace_percpu_free_percpu(base_addr: chunk->base_addr, off, ptr);
2326
2327	spin_unlock_irqrestore(lock: &pcpu_lock, flags);
2328
2329	if (need_balance)
2330	pcpu_schedule_balance_work();
2331	}
2332	EXPORT_SYMBOL_GPL(free_percpu);
2333
2334	bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2335	{
2336	#ifdef CONFIG_SMP
2337	const size_t static_size = __per_cpu_end - __per_cpu_start;
2338	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2339	unsigned int cpu;
2340
2341	for_each_possible_cpu(cpu) {
2342	void *start = per_cpu_ptr(base, cpu);
2343	void *va = (void *)addr;
2344
2345	if (va >= start && va < start + static_size) {
2346	if (can_addr) {
2347	*can_addr = (unsigned long) (va - start);
2348	*can_addr += (unsigned long)
2349	per_cpu_ptr(base, get_boot_cpu_id());
2350	}
2351	return true;
2352	}
2353	}
2354	#endif
2355	/* on UP, can't distinguish from other static vars, always false */
2356	return false;
2357	}
2358
2359	/**
2360	* is_kernel_percpu_address - test whether address is from static percpu area
2361	* @addr: address to test
2362	*
2363	* Test whether @addr belongs to in-kernel static percpu area. Module
2364	* static percpu areas are not considered. For those, use
2365	* is_module_percpu_address().
2366	*
2367	* RETURNS:
2368	* %true if @addr is from in-kernel static percpu area, %false otherwise.
2369	*/
2370	bool is_kernel_percpu_address(unsigned long addr)
2371	{
2372	return __is_kernel_percpu_address(addr, NULL);
2373	}
2374
2375	/**
2376	* per_cpu_ptr_to_phys - convert translated percpu address to physical address
2377	* @addr: the address to be converted to physical address
2378	*
2379	* Given @addr which is dereferenceable address obtained via one of
2380	* percpu access macros, this function translates it into its physical
2381	* address. The caller is responsible for ensuring @addr stays valid
2382	* until this function finishes.
2383	*
2384	* percpu allocator has special setup for the first chunk, which currently
2385	* supports either embedding in linear address space or vmalloc mapping,
2386	* and, from the second one, the backing allocator (currently either vm or
2387	* km) provides translation.
2388	*
2389	* The addr can be translated simply without checking if it falls into the
2390	* first chunk. But the current code reflects better how percpu allocator
2391	* actually works, and the verification can discover both bugs in percpu
2392	* allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2393	* code.
2394	*
2395	* RETURNS:
2396	* The physical address for @addr.
2397	*/
2398	phys_addr_t per_cpu_ptr_to_phys(void *addr)
2399	{
2400	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2401	bool in_first_chunk = false;
2402	unsigned long first_low, first_high;
2403	unsigned int cpu;
2404
2405	/*
2406	* The following test on unit_low/high isn't strictly
2407	* necessary but will speed up lookups of addresses which
2408	* aren't in the first chunk.
2409	*
2410	* The address check is against full chunk sizes. pcpu_base_addr
2411	* points to the beginning of the first chunk including the
2412	* static region. Assumes good intent as the first chunk may
2413	* not be full (ie. < pcpu_unit_pages in size).
2414	*/
2415	first_low = (unsigned long)pcpu_base_addr +
2416	pcpu_unit_page_offset(cpu: pcpu_low_unit_cpu, page_idx: 0);
2417	first_high = (unsigned long)pcpu_base_addr +
2418	pcpu_unit_page_offset(cpu: pcpu_high_unit_cpu, page_idx: pcpu_unit_pages);
2419	if ((unsigned long)addr >= first_low &&
2420	(unsigned long)addr < first_high) {
2421	for_each_possible_cpu(cpu) {
2422	void *start = per_cpu_ptr(base, cpu);
2423
2424	if (addr >= start && addr < start + pcpu_unit_size) {
2425	in_first_chunk = true;
2426	break;
2427	}
2428	}
2429	}
2430
2431	if (in_first_chunk) {
2432	if (!is_vmalloc_addr(x: addr))
2433	return __pa(addr);
2434	else
2435	return page_to_phys(vmalloc_to_page(addr)) +
2436	offset_in_page(addr);
2437	} else
2438	return page_to_phys(pcpu_addr_to_page(addr)) +
2439	offset_in_page(addr);
2440	}
2441
2442	/**
2443	* pcpu_alloc_alloc_info - allocate percpu allocation info
2444	* @nr_groups: the number of groups
2445	* @nr_units: the number of units
2446	*
2447	* Allocate ai which is large enough for @nr_groups groups containing
2448	* @nr_units units. The returned ai's groups[0].cpu_map points to the
2449	* cpu_map array which is long enough for @nr_units and filled with
2450	* NR_CPUS. It's the caller's responsibility to initialize cpu_map
2451	* pointer of other groups.
2452	*
2453	* RETURNS:
2454	* Pointer to the allocated pcpu_alloc_info on success, NULL on
2455	* failure.
2456	*/
2457	struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2458	int nr_units)
2459	{
2460	struct pcpu_alloc_info *ai;
2461	size_t base_size, ai_size;
2462	void *ptr;
2463	int unit;
2464
2465	base_size = ALIGN(struct_size(ai, groups, nr_groups),
2466	__alignof__(ai->groups[0].cpu_map[0]));
2467	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2468
2469	ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2470	if (!ptr)
2471	return NULL;
2472	ai = ptr;
2473	ptr += base_size;
2474
2475	ai->groups[0].cpu_map = ptr;
2476
2477	for (unit = 0; unit < nr_units; unit++)
2478	ai->groups[0].cpu_map[unit] = NR_CPUS;
2479
2480	ai->nr_groups = nr_groups;
2481	ai->__ai_size = PFN_ALIGN(ai_size);
2482
2483	return ai;
2484	}
2485
2486	/**
2487	* pcpu_free_alloc_info - free percpu allocation info
2488	* @ai: pcpu_alloc_info to free
2489	*
2490	* Free @ai which was allocated by pcpu_alloc_alloc_info().
2491	*/
2492	void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2493	{
2494	memblock_free(ptr: ai, size: ai->__ai_size);
2495	}
2496
2497	/**
2498	* pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2499	* @lvl: loglevel
2500	* @ai: allocation info to dump
2501	*
2502	* Print out information about @ai using loglevel @lvl.
2503	*/
2504	static void pcpu_dump_alloc_info(const char *lvl,
2505	const struct pcpu_alloc_info *ai)
2506	{
2507	int group_width = 1, cpu_width = 1, width;
2508	char empty_str[] = "--------";
2509	int alloc = 0, alloc_end = 0;
2510	int group, v;
2511	int upa, apl; /* units per alloc, allocs per line */
2512
2513	v = ai->nr_groups;
2514	while (v /= 10)
2515	group_width++;
2516
2517	v = num_possible_cpus();
2518	while (v /= 10)
2519	cpu_width++;
2520	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2521
2522	upa = ai->alloc_size / ai->unit_size;
2523	width = upa * (cpu_width + 1) + group_width + 3;
2524	apl = rounddown_pow_of_two(max(60 / width, 1));
2525
2526	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2527	lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2528	ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2529
2530	for (group = 0; group < ai->nr_groups; group++) {
2531	const struct pcpu_group_info *gi = &ai->groups[group];
2532	int unit = 0, unit_end = 0;
2533
2534	BUG_ON(gi->nr_units % upa);
2535	for (alloc_end += gi->nr_units / upa;
2536	alloc < alloc_end; alloc++) {
2537	if (!(alloc % apl)) {
2538	pr_cont("\n");
2539	printk("%spcpu-alloc: ", lvl);
2540	}
2541	pr_cont("[%0*d] ", group_width, group);
2542
2543	for (unit_end += upa; unit < unit_end; unit++)
2544	if (gi->cpu_map[unit] != NR_CPUS)
2545	pr_cont("%0*d ",
2546	cpu_width, gi->cpu_map[unit]);
2547	else
2548	pr_cont("%s ", empty_str);
2549	}
2550	}
2551	pr_cont("\n");
2552	}
2553
2554	/**
2555	* pcpu_setup_first_chunk - initialize the first percpu chunk
2556	* @ai: pcpu_alloc_info describing how to percpu area is shaped
2557	* @base_addr: mapped address
2558	*
2559	* Initialize the first percpu chunk which contains the kernel static
2560	* percpu area. This function is to be called from arch percpu area
2561	* setup path.
2562	*
2563	* @ai contains all information necessary to initialize the first
2564	* chunk and prime the dynamic percpu allocator.
2565	*
2566	* @ai->static_size is the size of static percpu area.
2567	*
2568	* @ai->reserved_size, if non-zero, specifies the amount of bytes to
2569	* reserve after the static area in the first chunk. This reserves
2570	* the first chunk such that it's available only through reserved
2571	* percpu allocation. This is primarily used to serve module percpu
2572	* static areas on architectures where the addressing model has
2573	* limited offset range for symbol relocations to guarantee module
2574	* percpu symbols fall inside the relocatable range.
2575	*
2576	* @ai->dyn_size determines the number of bytes available for dynamic
2577	* allocation in the first chunk. The area between @ai->static_size +
2578	* @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2579	*
2580	* @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2581	* and equal to or larger than @ai->static_size + @ai->reserved_size +
2582	* @ai->dyn_size.
2583	*
2584	* @ai->atom_size is the allocation atom size and used as alignment
2585	* for vm areas.
2586	*
2587	* @ai->alloc_size is the allocation size and always multiple of
2588	* @ai->atom_size. This is larger than @ai->atom_size if
2589	* @ai->unit_size is larger than @ai->atom_size.
2590	*
2591	* @ai->nr_groups and @ai->groups describe virtual memory layout of
2592	* percpu areas. Units which should be colocated are put into the
2593	* same group. Dynamic VM areas will be allocated according to these
2594	* groupings. If @ai->nr_groups is zero, a single group containing
2595	* all units is assumed.
2596	*
2597	* The caller should have mapped the first chunk at @base_addr and
2598	* copied static data to each unit.
2599	*
2600	* The first chunk will always contain a static and a dynamic region.
2601	* However, the static region is not managed by any chunk. If the first
2602	* chunk also contains a reserved region, it is served by two chunks -
2603	* one for the reserved region and one for the dynamic region. They
2604	* share the same vm, but use offset regions in the area allocation map.
2605	* The chunk serving the dynamic region is circulated in the chunk slots
2606	* and available for dynamic allocation like any other chunk.
2607	*/
2608	void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2609	void *base_addr)
2610	{
2611	size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2612	size_t static_size, dyn_size;
2613	unsigned long *group_offsets;
2614	size_t *group_sizes;
2615	unsigned long *unit_off;
2616	unsigned int cpu;
2617	int *unit_map;
2618	int group, unit, i;
2619	unsigned long tmp_addr;
2620	size_t alloc_size;
2621
2622	#define PCPU_SETUP_BUG_ON(cond) do { \
2623	if (unlikely(cond)) { \
2624	pr_emerg("failed to initialize, %s\n", #cond); \
2625	pr_emerg("cpu_possible_mask=%*pb\n", \
2626	cpumask_pr_args(cpu_possible_mask)); \
2627	pcpu_dump_alloc_info(KERN_EMERG, ai); \
2628	BUG(); \
2629	} \
2630	} while (0)
2631
2632	/* sanity checks */
2633	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2634	#ifdef CONFIG_SMP
2635	PCPU_SETUP_BUG_ON(!ai->static_size);
2636	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2637	#endif
2638	PCPU_SETUP_BUG_ON(!base_addr);
2639	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2640	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2641	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2642	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2643	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2644	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2645	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2646	PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2647	IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2648	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2649
2650	/* process group information and build config tables accordingly */
2651	alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2652	group_offsets = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
2653	if (!group_offsets)
2654	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
2655	alloc_size);
2656
2657	alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2658	group_sizes = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
2659	if (!group_sizes)
2660	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
2661	alloc_size);
2662
2663	alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2664	unit_map = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
2665	if (!unit_map)
2666	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
2667	alloc_size);
2668
2669	alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2670	unit_off = memblock_alloc(size: alloc_size, SMP_CACHE_BYTES);
2671	if (!unit_off)
2672	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
2673	alloc_size);
2674
2675	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2676	unit_map[cpu] = UINT_MAX;
2677
2678	pcpu_low_unit_cpu = NR_CPUS;
2679	pcpu_high_unit_cpu = NR_CPUS;
2680
2681	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2682	const struct pcpu_group_info *gi = &ai->groups[group];
2683
2684	group_offsets[group] = gi->base_offset;
2685	group_sizes[group] = gi->nr_units * ai->unit_size;
2686
2687	for (i = 0; i < gi->nr_units; i++) {
2688	cpu = gi->cpu_map[i];
2689	if (cpu == NR_CPUS)
2690	continue;
2691
2692	PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2693	PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2694	PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2695
2696	unit_map[cpu] = unit + i;
2697	unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2698
2699	/* determine low/high unit_cpu */
2700	if (pcpu_low_unit_cpu == NR_CPUS ||
2701	unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2702	pcpu_low_unit_cpu = cpu;
2703	if (pcpu_high_unit_cpu == NR_CPUS ||
2704	unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2705	pcpu_high_unit_cpu = cpu;
2706	}
2707	}
2708	pcpu_nr_units = unit;
2709
2710	for_each_possible_cpu(cpu)
2711	PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2712
2713	/* we're done parsing the input, undefine BUG macro and dump config */
2714	#undef PCPU_SETUP_BUG_ON
2715	pcpu_dump_alloc_info(KERN_DEBUG, ai);
2716
2717	pcpu_nr_groups = ai->nr_groups;
2718	pcpu_group_offsets = group_offsets;
2719	pcpu_group_sizes = group_sizes;
2720	pcpu_unit_map = unit_map;
2721	pcpu_unit_offsets = unit_off;
2722
2723	/* determine basic parameters */
2724	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2725	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2726	pcpu_atom_size = ai->atom_size;
2727	pcpu_chunk_struct_size = struct_size((struct pcpu_chunk *)0, populated,
2728	BITS_TO_LONGS(pcpu_unit_pages));
2729
2730	pcpu_stats_save_ai(ai);
2731
2732	/*
2733	* Allocate chunk slots. The slots after the active slots are:
2734	* sidelined_slot - isolated, depopulated chunks
2735	* free_slot - fully free chunks
2736	* to_depopulate_slot - isolated, chunks to depopulate
2737	*/
2738	pcpu_sidelined_slot = __pcpu_size_to_slot(size: pcpu_unit_size) + 1;
2739	pcpu_free_slot = pcpu_sidelined_slot + 1;
2740	pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2741	pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2742	pcpu_chunk_lists = memblock_alloc(size: pcpu_nr_slots *
2743	sizeof(pcpu_chunk_lists[0]),
2744	SMP_CACHE_BYTES);
2745	if (!pcpu_chunk_lists)
2746	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
2747	pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2748
2749	for (i = 0; i < pcpu_nr_slots; i++)
2750	INIT_LIST_HEAD(list: &pcpu_chunk_lists[i]);
2751
2752	/*
2753	* The end of the static region needs to be aligned with the
2754	* minimum allocation size as this offsets the reserved and
2755	* dynamic region. The first chunk ends page aligned by
2756	* expanding the dynamic region, therefore the dynamic region
2757	* can be shrunk to compensate while still staying above the
2758	* configured sizes.
2759	*/
2760	static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2761	dyn_size = ai->dyn_size - (static_size - ai->static_size);
2762
2763	/*
2764	* Initialize first chunk:
2765	* This chunk is broken up into 3 parts:
2766	* < static | [reserved] | dynamic >
2767	* - static - there is no backing chunk because these allocations can
2768	* never be freed.
2769	* - reserved (pcpu_reserved_chunk) - exists primarily to serve
2770	* allocations from module load.
2771	* - dynamic (pcpu_first_chunk) - serves the dynamic part of the first
2772	* chunk.
2773	*/
2774	tmp_addr = (unsigned long)base_addr + static_size;
2775	if (ai->reserved_size)
2776	pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr,
2777	map_size: ai->reserved_size);
2778	tmp_addr = (unsigned long)base_addr + static_size + ai->reserved_size;
2779	pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, map_size: dyn_size);
2780
2781	pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2782	pcpu_chunk_relocate(chunk: pcpu_first_chunk, oslot: -1);
2783
2784	/* include all regions of the first chunk */
2785	pcpu_nr_populated += PFN_DOWN(size_sum);
2786
2787	pcpu_stats_chunk_alloc();
2788	trace_percpu_create_chunk(base_addr);
2789
2790	/* we're done */
2791	pcpu_base_addr = base_addr;
2792	}
2793
2794	#ifdef CONFIG_SMP
2795
2796	const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2797	[PCPU_FC_AUTO] = "auto",
2798	[PCPU_FC_EMBED] = "embed",
2799	[PCPU_FC_PAGE] = "page",
2800	};
2801
2802	enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2803
2804	static int __init percpu_alloc_setup(char *str)
2805	{
2806	if (!str)
2807	return -EINVAL;
2808
2809	if (0)
2810	/* nada */;
2811	#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2812	else if (!strcmp(str, "embed"))
2813	pcpu_chosen_fc = PCPU_FC_EMBED;
2814	#endif
2815	#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2816	else if (!strcmp(str, "page"))
2817	pcpu_chosen_fc = PCPU_FC_PAGE;
2818	#endif
2819	else
2820	pr_warn("unknown allocator %s specified\n", str);
2821
2822	return 0;
2823	}
2824	early_param("percpu_alloc", percpu_alloc_setup);
2825
2826	/*
2827	* pcpu_embed_first_chunk() is used by the generic percpu setup.
2828	* Build it if needed by the arch config or the generic setup is going
2829	* to be used.
2830	*/
2831	#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2832	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2833	#define BUILD_EMBED_FIRST_CHUNK
2834	#endif
2835
2836	/* build pcpu_page_first_chunk() iff needed by the arch config */
2837	#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2838	#define BUILD_PAGE_FIRST_CHUNK
2839	#endif
2840
2841	/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2842	#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2843	/**
2844	* pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2845	* @reserved_size: the size of reserved percpu area in bytes
2846	* @dyn_size: minimum free size for dynamic allocation in bytes
2847	* @atom_size: allocation atom size
2848	* @cpu_distance_fn: callback to determine distance between cpus, optional
2849	*
2850	* This function determines grouping of units, their mappings to cpus
2851	* and other parameters considering needed percpu size, allocation
2852	* atom size and distances between CPUs.
2853	*
2854	* Groups are always multiples of atom size and CPUs which are of
2855	* LOCAL_DISTANCE both ways are grouped together and share space for
2856	* units in the same group. The returned configuration is guaranteed
2857	* to have CPUs on different nodes on different groups and >=75% usage
2858	* of allocated virtual address space.
2859	*
2860	* RETURNS:
2861	* On success, pointer to the new allocation_info is returned. On
2862	* failure, ERR_PTR value is returned.
2863	*/
2864	static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2865	size_t reserved_size, size_t dyn_size,
2866	size_t atom_size,
2867	pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2868	{
2869	static int group_map[NR_CPUS] __initdata;
2870	static int group_cnt[NR_CPUS] __initdata;
2871	static struct cpumask mask __initdata;
2872	const size_t static_size = __per_cpu_end - __per_cpu_start;
2873	int nr_groups = 1, nr_units = 0;
2874	size_t size_sum, min_unit_size, alloc_size;
2875	int upa, max_upa, best_upa; /* units_per_alloc */
2876	int last_allocs, group, unit;
2877	unsigned int cpu, tcpu;
2878	struct pcpu_alloc_info *ai;
2879	unsigned int *cpu_map;
2880
2881	/* this function may be called multiple times */
2882	memset(group_map, 0, sizeof(group_map));
2883	memset(group_cnt, 0, sizeof(group_cnt));
2884	cpumask_clear(dstp: &mask);
2885
2886	/* calculate size_sum and ensure dyn_size is enough for early alloc */
2887	size_sum = PFN_ALIGN(static_size + reserved_size +
2888	max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2889	dyn_size = size_sum - static_size - reserved_size;
2890
2891	/*
2892	* Determine min_unit_size, alloc_size and max_upa such that
2893	* alloc_size is multiple of atom_size and is the smallest
2894	* which can accommodate 4k aligned segments which are equal to
2895	* or larger than min_unit_size.
2896	*/
2897	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2898
2899	/* determine the maximum # of units that can fit in an allocation */
2900	alloc_size = roundup(min_unit_size, atom_size);
2901	upa = alloc_size / min_unit_size;
2902	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2903	upa--;
2904	max_upa = upa;
2905
2906	cpumask_copy(dstp: &mask, cpu_possible_mask);
2907
2908	/* group cpus according to their proximity */
2909	for (group = 0; !cpumask_empty(srcp: &mask); group++) {
2910	/* pop the group's first cpu */
2911	cpu = cpumask_first(srcp: &mask);
2912	group_map[cpu] = group;
2913	group_cnt[group]++;
2914	cpumask_clear_cpu(cpu, dstp: &mask);
2915
2916	for_each_cpu(tcpu, &mask) {
2917	if (!cpu_distance_fn ||
2918	(cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2919	cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2920	group_map[tcpu] = group;
2921	group_cnt[group]++;
2922	cpumask_clear_cpu(cpu: tcpu, dstp: &mask);
2923	}
2924	}
2925	}
2926	nr_groups = group;
2927
2928	/*
2929	* Wasted space is caused by a ratio imbalance of upa to group_cnt.
2930	* Expand the unit_size until we use >= 75% of the units allocated.
2931	* Related to atom_size, which could be much larger than the unit_size.
2932	*/
2933	last_allocs = INT_MAX;
2934	best_upa = 0;
2935	for (upa = max_upa; upa; upa--) {
2936	int allocs = 0, wasted = 0;
2937
2938	if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2939	continue;
2940
2941	for (group = 0; group < nr_groups; group++) {
2942	int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2943	allocs += this_allocs;
2944	wasted += this_allocs * upa - group_cnt[group];
2945	}
2946
2947	/*
2948	* Don't accept if wastage is over 1/3. The
2949	* greater-than comparison ensures upa==1 always
2950	* passes the following check.
2951	*/
2952	if (wasted > num_possible_cpus() / 3)
2953	continue;
2954
2955	/* and then don't consume more memory */
2956	if (allocs > last_allocs)
2957	break;
2958	last_allocs = allocs;
2959	best_upa = upa;
2960	}
2961	BUG_ON(!best_upa);
2962	upa = best_upa;
2963
2964	/* allocate and fill alloc_info */
2965	for (group = 0; group < nr_groups; group++)
2966	nr_units += roundup(group_cnt[group], upa);
2967
2968	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2969	if (!ai)
2970	return ERR_PTR(error: -ENOMEM);
2971	cpu_map = ai->groups[0].cpu_map;
2972
2973	for (group = 0; group < nr_groups; group++) {
2974	ai->groups[group].cpu_map = cpu_map;
2975	cpu_map += roundup(group_cnt[group], upa);
2976	}
2977
2978	ai->static_size = static_size;
2979	ai->reserved_size = reserved_size;
2980	ai->dyn_size = dyn_size;
2981	ai->unit_size = alloc_size / upa;
2982	ai->atom_size = atom_size;
2983	ai->alloc_size = alloc_size;
2984
2985	for (group = 0, unit = 0; group < nr_groups; group++) {
2986	struct pcpu_group_info *gi = &ai->groups[group];
2987
2988	/*
2989	* Initialize base_offset as if all groups are located
2990	* back-to-back. The caller should update this to
2991	* reflect actual allocation.
2992	*/
2993	gi->base_offset = unit * ai->unit_size;
2994
2995	for_each_possible_cpu(cpu)
2996	if (group_map[cpu] == group)
2997	gi->cpu_map[gi->nr_units++] = cpu;
2998	gi->nr_units = roundup(gi->nr_units, upa);
2999	unit += gi->nr_units;
3000	}
3001	BUG_ON(unit != nr_units);
3002
3003	return ai;
3004	}
3005
3006	static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
3007	pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3008	{
3009	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
3010	#ifdef CONFIG_NUMA
3011	int node = NUMA_NO_NODE;
3012	void *ptr;
3013
3014	if (cpu_to_nd_fn)
3015	node = cpu_to_nd_fn(cpu);
3016
3017	if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
3018	ptr = memblock_alloc_from(size, align, min_addr: goal);
3019	pr_info("cpu %d has no node %d or node-local memory\n",
3020	cpu, node);
3021	pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
3022	cpu, size, (u64)__pa(ptr));
3023	} else {
3024	ptr = memblock_alloc_try_nid(size, align, min_addr: goal,
3025	MEMBLOCK_ALLOC_ACCESSIBLE,
3026	nid: node);
3027
3028	pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
3029	cpu, size, node, (u64)__pa(ptr));
3030	}
3031	return ptr;
3032	#else
3033	return memblock_alloc_from(size, align, goal);
3034	#endif
3035	}
3036
3037	static void __init pcpu_fc_free(void *ptr, size_t size)
3038	{
3039	memblock_free(ptr, size);
3040	}
3041	#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3042
3043	#if defined(BUILD_EMBED_FIRST_CHUNK)
3044	/**
3045	* pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3046	* @reserved_size: the size of reserved percpu area in bytes
3047	* @dyn_size: minimum free size for dynamic allocation in bytes
3048	* @atom_size: allocation atom size
3049	* @cpu_distance_fn: callback to determine distance between cpus, optional
3050	* @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3051	*
3052	* This is a helper to ease setting up embedded first percpu chunk and
3053	* can be called where pcpu_setup_first_chunk() is expected.
3054	*
3055	* If this function is used to setup the first chunk, it is allocated
3056	* by calling pcpu_fc_alloc and used as-is without being mapped into
3057	* vmalloc area. Allocations are always whole multiples of @atom_size
3058	* aligned to @atom_size.
3059	*
3060	* This enables the first chunk to piggy back on the linear physical
3061	* mapping which often uses larger page size. Please note that this
3062	* can result in very sparse cpu->unit mapping on NUMA machines thus
3063	* requiring large vmalloc address space. Don't use this allocator if
3064	* vmalloc space is not orders of magnitude larger than distances
3065	* between node memory addresses (ie. 32bit NUMA machines).
3066	*
3067	* @dyn_size specifies the minimum dynamic area size.
3068	*
3069	* If the needed size is smaller than the minimum or specified unit
3070	* size, the leftover is returned using pcpu_fc_free.
3071	*
3072	* RETURNS:
3073	* 0 on success, -errno on failure.
3074	*/
3075	int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3076	size_t atom_size,
3077	pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3078	pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3079	{
3080	void *base = (void *)ULONG_MAX;
3081	void **areas = NULL;
3082	struct pcpu_alloc_info *ai;
3083	size_t size_sum, areas_size;
3084	unsigned long max_distance;
3085	int group, i, highest_group, rc = 0;
3086
3087	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3088	cpu_distance_fn);
3089	if (IS_ERR(ptr: ai))
3090	return PTR_ERR(ptr: ai);
3091
3092	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3093	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3094
3095	areas = memblock_alloc(size: areas_size, SMP_CACHE_BYTES);
3096	if (!areas) {
3097	rc = -ENOMEM;
3098	goto out_free;
3099	}
3100
3101	/* allocate, copy and determine base address & max_distance */
3102	highest_group = 0;
3103	for (group = 0; group < ai->nr_groups; group++) {
3104	struct pcpu_group_info *gi = &ai->groups[group];
3105	unsigned int cpu = NR_CPUS;
3106	void *ptr;
3107
3108	for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3109	cpu = gi->cpu_map[i];
3110	BUG_ON(cpu == NR_CPUS);
3111
3112	/* allocate space for the whole group */
3113	ptr = pcpu_fc_alloc(cpu, size: gi->nr_units * ai->unit_size, align: atom_size, cpu_to_nd_fn);
3114	if (!ptr) {
3115	rc = -ENOMEM;
3116	goto out_free_areas;
3117	}
3118	/* kmemleak tracks the percpu allocations separately */
3119	kmemleak_ignore_phys(__pa(ptr));
3120	areas[group] = ptr;
3121
3122	base = min(ptr, base);
3123	if (ptr > areas[highest_group])
3124	highest_group = group;
3125	}
3126	max_distance = areas[highest_group] - base;
3127	max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3128
3129	/* warn if maximum distance is further than 75% of vmalloc space */
3130	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3131	pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3132	max_distance, VMALLOC_TOTAL);
3133	#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3134	/* and fail if we have fallback */
3135	rc = -EINVAL;
3136	goto out_free_areas;
3137	#endif
3138	}
3139
3140	/*
3141	* Copy data and free unused parts. This should happen after all
3142	* allocations are complete; otherwise, we may end up with
3143	* overlapping groups.
3144	*/
3145	for (group = 0; group < ai->nr_groups; group++) {
3146	struct pcpu_group_info *gi = &ai->groups[group];
3147	void *ptr = areas[group];
3148
3149	for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3150	if (gi->cpu_map[i] == NR_CPUS) {
3151	/* unused unit, free whole */
3152	pcpu_fc_free(ptr, size: ai->unit_size);
3153	continue;
3154	}
3155	/* copy and return the unused part */
3156	memcpy(ptr, __per_cpu_load, ai->static_size);
3157	pcpu_fc_free(ptr: ptr + size_sum, size: ai->unit_size - size_sum);
3158	}
3159	}
3160
3161	/* base address is now known, determine group base offsets */
3162	for (group = 0; group < ai->nr_groups; group++) {
3163	ai->groups[group].base_offset = areas[group] - base;
3164	}
3165
3166	pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3167	PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3168	ai->dyn_size, ai->unit_size);
3169
3170	pcpu_setup_first_chunk(ai, base_addr: base);
3171	goto out_free;
3172
3173	out_free_areas:
3174	for (group = 0; group < ai->nr_groups; group++)
3175	if (areas[group])
3176	pcpu_fc_free(ptr: areas[group],
3177	size: ai->groups[group].nr_units * ai->unit_size);
3178	out_free:
3179	pcpu_free_alloc_info(ai);
3180	if (areas)
3181	memblock_free(ptr: areas, size: areas_size);
3182	return rc;
3183	}
3184	#endif /* BUILD_EMBED_FIRST_CHUNK */
3185
3186	#ifdef BUILD_PAGE_FIRST_CHUNK
3187	#include <asm/pgalloc.h>
3188
3189	#ifndef P4D_TABLE_SIZE
3190	#define P4D_TABLE_SIZE PAGE_SIZE
3191	#endif
3192
3193	#ifndef PUD_TABLE_SIZE
3194	#define PUD_TABLE_SIZE PAGE_SIZE
3195	#endif
3196
3197	#ifndef PMD_TABLE_SIZE
3198	#define PMD_TABLE_SIZE PAGE_SIZE
3199	#endif
3200
3201	#ifndef PTE_TABLE_SIZE
3202	#define PTE_TABLE_SIZE PAGE_SIZE
3203	#endif
3204	void __init __weak pcpu_populate_pte(unsigned long addr)
3205	{
3206	pgd_t *pgd = pgd_offset_k(addr);
3207	p4d_t *p4d;
3208	pud_t *pud;
3209	pmd_t *pmd;
3210
3211	if (pgd_none(pgd: *pgd)) {
3212	p4d = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3213	if (!p4d)
3214	goto err_alloc;
3215	pgd_populate(mm: &init_mm, pgd, p4d);
3216	}
3217
3218	p4d = p4d_offset(pgd, address: addr);
3219	if (p4d_none(p4d: *p4d)) {
3220	pud = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3221	if (!pud)
3222	goto err_alloc;
3223	p4d_populate(mm: &init_mm, p4d, pud);
3224	}
3225
3226	pud = pud_offset(p4d, address: addr);
3227	if (pud_none(pud: *pud)) {
3228	pmd = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3229	if (!pmd)
3230	goto err_alloc;
3231	pud_populate(mm: &init_mm, pud, pmd);
3232	}
3233
3234	pmd = pmd_offset(pud, address: addr);
3235	if (!pmd_present(pmd: *pmd)) {
3236	pte_t *new;
3237
3238	new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3239	if (!new)
3240	goto err_alloc;
3241	pmd_populate_kernel(mm: &init_mm, pmd, pte: new);
3242	}
3243
3244	return;
3245
3246	err_alloc:
3247	panic(fmt: "%s: Failed to allocate memory\n", __func__);
3248	}
3249
3250	/**
3251	* pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3252	* @reserved_size: the size of reserved percpu area in bytes
3253	* @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3254	*
3255	* This is a helper to ease setting up page-remapped first percpu
3256	* chunk and can be called where pcpu_setup_first_chunk() is expected.
3257	*
3258	* This is the basic allocator. Static percpu area is allocated
3259	* page-by-page into vmalloc area.
3260	*
3261	* RETURNS:
3262	* 0 on success, -errno on failure.
3263	*/
3264	int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3265	{
3266	static struct vm_struct vm;
3267	struct pcpu_alloc_info *ai;
3268	char psize_str[16];
3269	int unit_pages;
3270	size_t pages_size;
3271	struct page **pages;
3272	int unit, i, j, rc = 0;
3273	int upa;
3274	int nr_g0_units;
3275
3276	snprintf(buf: psize_str, size: sizeof(psize_str), fmt: "%luK", PAGE_SIZE >> 10);
3277
3278	ai = pcpu_build_alloc_info(reserved_size, dyn_size: 0, PAGE_SIZE, NULL);
3279	if (IS_ERR(ptr: ai))
3280	return PTR_ERR(ptr: ai);
3281	BUG_ON(ai->nr_groups != 1);
3282	upa = ai->alloc_size/ai->unit_size;
3283	nr_g0_units = roundup(num_possible_cpus(), upa);
3284	if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3285	pcpu_free_alloc_info(ai);
3286	return -EINVAL;
3287	}
3288
3289	unit_pages = ai->unit_size >> PAGE_SHIFT;
3290
3291	/* unaligned allocations can't be freed, round up to page size */
3292	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3293	sizeof(pages[0]));
3294	pages = memblock_alloc(size: pages_size, SMP_CACHE_BYTES);
3295	if (!pages)
3296	panic(fmt: "%s: Failed to allocate %zu bytes\n", __func__,
3297	pages_size);
3298
3299	/* allocate pages */
3300	j = 0;
3301	for (unit = 0; unit < num_possible_cpus(); unit++) {
3302	unsigned int cpu = ai->groups[0].cpu_map[unit];
3303	for (i = 0; i < unit_pages; i++) {
3304	void *ptr;
3305
3306	ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3307	if (!ptr) {
3308	pr_warn("failed to allocate %s page for cpu%u\n",
3309	psize_str, cpu);
3310	goto enomem;
3311	}
3312	/* kmemleak tracks the percpu allocations separately */
3313	kmemleak_ignore_phys(__pa(ptr));
3314	pages[j++] = virt_to_page(ptr);
3315	}
3316	}
3317
3318	/* allocate vm area, map the pages and copy static data */
3319	vm.flags = VM_ALLOC;
3320	vm.size = num_possible_cpus() * ai->unit_size;
3321	vm_area_register_early(vm: &vm, PAGE_SIZE);
3322
3323	for (unit = 0; unit < num_possible_cpus(); unit++) {
3324	unsigned long unit_addr =
3325	(unsigned long)vm.addr + unit * ai->unit_size;
3326
3327	for (i = 0; i < unit_pages; i++)
3328	pcpu_populate_pte(addr: unit_addr + (i << PAGE_SHIFT));
3329
3330	/* pte already populated, the following shouldn't fail */
3331	rc = __pcpu_map_pages(addr: unit_addr, pages: &pages[unit * unit_pages],
3332	nr_pages: unit_pages);
3333	if (rc < 0)
3334	panic(fmt: "failed to map percpu area, err=%d\n", rc);
3335
3336	/*
3337	* FIXME: Archs with virtual cache should flush local
3338	* cache for the linear mapping here - something
3339	* equivalent to flush_cache_vmap() on the local cpu.
3340	* flush_cache_vmap() can't be used as most supporting
3341	* data structures are not set up yet.
3342	*/
3343
3344	/* copy static data */
3345	memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3346	}
3347
3348	/* we're ready, commit */
3349	pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3350	unit_pages, psize_str, ai->static_size,
3351	ai->reserved_size, ai->dyn_size);
3352
3353	pcpu_setup_first_chunk(ai, base_addr: vm.addr);
3354	goto out_free_ar;
3355
3356	enomem:
3357	while (--j >= 0)
3358	pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3359	rc = -ENOMEM;
3360	out_free_ar:
3361	memblock_free(ptr: pages, size: pages_size);
3362	pcpu_free_alloc_info(ai);
3363	return rc;
3364	}
3365	#endif /* BUILD_PAGE_FIRST_CHUNK */
3366
3367	#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3368	/*
3369	* Generic SMP percpu area setup.
3370	*
3371	* The embedding helper is used because its behavior closely resembles
3372	* the original non-dynamic generic percpu area setup. This is
3373	* important because many archs have addressing restrictions and might
3374	* fail if the percpu area is located far away from the previous
3375	* location. As an added bonus, in non-NUMA cases, embedding is
3376	* generally a good idea TLB-wise because percpu area can piggy back
3377	* on the physical linear memory mapping which uses large page
3378	* mappings on applicable archs.
3379	*/
3380	unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3381	EXPORT_SYMBOL(__per_cpu_offset);
3382
3383	void __init setup_per_cpu_areas(void)
3384	{
3385	unsigned long delta;
3386	unsigned int cpu;
3387	int rc;
3388
3389	/*
3390	* Always reserve area for module percpu variables. That's
3391	* what the legacy allocator did.
3392	*/
3393	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3394	PAGE_SIZE, NULL, NULL);
3395	if (rc < 0)
3396	panic("Failed to initialize percpu areas.");
3397
3398	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3399	for_each_possible_cpu(cpu)
3400	__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3401	}
3402	#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3403
3404	#else /* CONFIG_SMP */
3405
3406	/*
3407	* UP percpu area setup.
3408	*
3409	* UP always uses km-based percpu allocator with identity mapping.
3410	* Static percpu variables are indistinguishable from the usual static
3411	* variables and don't require any special preparation.
3412	*/
3413	void __init setup_per_cpu_areas(void)
3414	{
3415	const size_t unit_size =
3416	roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3417	PERCPU_DYNAMIC_RESERVE));
3418	struct pcpu_alloc_info *ai;
3419	void *fc;
3420
3421	ai = pcpu_alloc_alloc_info(1, 1);
3422	fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3423	if (!ai || !fc)
3424	panic("Failed to allocate memory for percpu areas.");
3425	/* kmemleak tracks the percpu allocations separately */
3426	kmemleak_ignore_phys(__pa(fc));
3427
3428	ai->dyn_size = unit_size;
3429	ai->unit_size = unit_size;
3430	ai->atom_size = unit_size;
3431	ai->alloc_size = unit_size;
3432	ai->groups[0].nr_units = 1;
3433	ai->groups[0].cpu_map[0] = 0;
3434
3435	pcpu_setup_first_chunk(ai, fc);
3436	pcpu_free_alloc_info(ai);
3437	}
3438
3439	#endif /* CONFIG_SMP */
3440
3441	/*
3442	* pcpu_nr_pages - calculate total number of populated backing pages
3443	*
3444	* This reflects the number of pages populated to back chunks. Metadata is
3445	* excluded in the number exposed in meminfo as the number of backing pages
3446	* scales with the number of cpus and can quickly outweigh the memory used for
3447	* metadata. It also keeps this calculation nice and simple.
3448	*
3449	* RETURNS:
3450	* Total number of populated backing pages in use by the allocator.
3451	*/
3452	unsigned long pcpu_nr_pages(void)
3453	{
3454	return pcpu_nr_populated * pcpu_nr_units;
3455	}
3456
3457	/*
3458	* Percpu allocator is initialized early during boot when neither slab or
3459	* workqueue is available. Plug async management until everything is up
3460	* and running.
3461	*/
3462	static int __init percpu_enable_async(void)
3463	{
3464	pcpu_async_enabled = true;
3465	return 0;
3466	}
3467	subsys_initcall(percpu_enable_async);
3468