path: root/mm/percpu.c (plain)
blob: 0462a2a00f0597e353cd91d268337e6607bafe24

1	/*
2	* mm/percpu.c - percpu memory allocator
3	*
4	* Copyright (C) 2009 SUSE Linux Products GmbH
5	* Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6	*
7	* This file is released under the GPLv2.
8	*
9	* This is percpu allocator which can handle both static and dynamic
10	* areas. Percpu areas are allocated in chunks. Each chunk is
11	* consisted of boot-time determined number of units and the first
12	* chunk is used for static percpu variables in the kernel image
13	* (special boot time alloc/init handling necessary as these areas
14	* need to be brought up before allocation services are running).
15	* Unit grows as necessary and all units grow or shrink in unison.
16	* When a chunk is filled up, another chunk is allocated.
17	*
18	* c0 c1 c2
19	* ------------------- ------------------- ------------
20	* | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21	* ------------------- ...... ------------------- .... ------------
22	*
23	* Allocation is done in offset-size areas of single unit space. Ie,
24	* an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25	* c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26	* cpus. On NUMA, the mapping can be non-linear and even sparse.
27	* Percpu access can be done by configuring percpu base registers
28	* according to cpu to unit mapping and pcpu_unit_size.
29	*
30	* There are usually many small percpu allocations many of them being
31	* as small as 4 bytes. The allocator organizes chunks into lists
32	* according to free size and tries to allocate from the fullest one.
33	* Each chunk keeps the maximum contiguous area size hint which is
34	* guaranteed to be equal to or larger than the maximum contiguous
35	* area in the chunk. This helps the allocator not to iterate the
36	* chunk maps unnecessarily.
37	*
38	* Allocation state in each chunk is kept using an array of integers
39	* on chunk->map. A positive value in the map represents a free
40	* region and negative allocated. Allocation inside a chunk is done
41	* by scanning this map sequentially and serving the first matching
42	* entry. This is mostly copied from the percpu_modalloc() allocator.
43	* Chunks can be determined from the address using the index field
44	* in the page struct. The index field contains a pointer to the chunk.
45	*
46	* To use this allocator, arch code should do the followings.
47	*
48	* - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49	* regular address to percpu pointer and back if they need to be
50	* different from the default
51	*
52	* - use pcpu_setup_first_chunk() during percpu area initialization to
53	* setup the first chunk containing the kernel static percpu area
54	*/
55
56	#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
57
58	#include <linux/bitmap.h>
59	#include <linux/bootmem.h>
60	#include <linux/err.h>
61	#include <linux/list.h>
62	#include <linux/log2.h>
63	#include <linux/mm.h>
64	#include <linux/module.h>
65	#include <linux/mutex.h>
66	#include <linux/percpu.h>
67	#include <linux/pfn.h>
68	#include <linux/slab.h>
69	#include <linux/spinlock.h>
70	#include <linux/vmalloc.h>
71	#include <linux/workqueue.h>
72	#include <linux/kmemleak.h>
73	#include <linux/sched.h>
74
75	#include <asm/cacheflush.h>
76	#include <asm/sections.h>
77	#include <asm/tlbflush.h>
78	#include <asm/io.h>
79
80	#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
81	#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
82	#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
83	#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
84	#define PCPU_EMPTY_POP_PAGES_LOW 2
85	#define PCPU_EMPTY_POP_PAGES_HIGH 4
86
87	#ifdef CONFIG_SMP
88	/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
89	#ifndef __addr_to_pcpu_ptr
90	#define __addr_to_pcpu_ptr(addr) \
91	(void __percpu *)((unsigned long)(addr) - \
92	(unsigned long)pcpu_base_addr + \
93	(unsigned long)__per_cpu_start)
94	#endif
95	#ifndef __pcpu_ptr_to_addr
96	#define __pcpu_ptr_to_addr(ptr) \
97	(void __force *)((unsigned long)(ptr) + \
98	(unsigned long)pcpu_base_addr - \
99	(unsigned long)__per_cpu_start)
100	#endif
101	#else /* CONFIG_SMP */
102	/* on UP, it's always identity mapped */
103	#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
104	#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
105	#endif /* CONFIG_SMP */
106
107	struct pcpu_chunk {
108	struct list_head list; /* linked to pcpu_slot lists */
109	int free_size; /* free bytes in the chunk */
110	int contig_hint; /* max contiguous size hint */
111	void *base_addr; /* base address of this chunk */
112
113	int map_used; /* # of map entries used before the sentry */
114	int map_alloc; /* # of map entries allocated */
115	int *map; /* allocation map */
116	struct list_head map_extend_list;/* on pcpu_map_extend_chunks */
117
118	void *data; /* chunk data */
119	int first_free; /* no free below this */
120	bool immutable; /* no [de]population allowed */
121	int nr_populated; /* # of populated pages */
122	unsigned long populated[]; /* populated bitmap */
123	};
124
125	static int pcpu_unit_pages __read_mostly;
126	static int pcpu_unit_size __read_mostly;
127	static int pcpu_nr_units __read_mostly;
128	static int pcpu_atom_size __read_mostly;
129	static int pcpu_nr_slots __read_mostly;
130	static size_t pcpu_chunk_struct_size __read_mostly;
131
132	/* cpus with the lowest and highest unit addresses */
133	static unsigned int pcpu_low_unit_cpu __read_mostly;
134	static unsigned int pcpu_high_unit_cpu __read_mostly;
135
136	/* the address of the first chunk which starts with the kernel static area */
137	void *pcpu_base_addr __read_mostly;
138	EXPORT_SYMBOL_GPL(pcpu_base_addr);
139
140	static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
141	const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
142
143	/* group information, used for vm allocation */
144	static int pcpu_nr_groups __read_mostly;
145	static const unsigned long *pcpu_group_offsets __read_mostly;
146	static const size_t *pcpu_group_sizes __read_mostly;
147
148	/*
149	* The first chunk which always exists. Note that unlike other
150	* chunks, this one can be allocated and mapped in several different
151	* ways and thus often doesn't live in the vmalloc area.
152	*/
153	static struct pcpu_chunk *pcpu_first_chunk;
154
155	/*
156	* Optional reserved chunk. This chunk reserves part of the first
157	* chunk and serves it for reserved allocations. The amount of
158	* reserved offset is in pcpu_reserved_chunk_limit. When reserved
159	* area doesn't exist, the following variables contain NULL and 0
160	* respectively.
161	*/
162	static struct pcpu_chunk *pcpu_reserved_chunk;
163	static int pcpu_reserved_chunk_limit;
164
165	static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
166	static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
167
168	static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
169
170	/* chunks which need their map areas extended, protected by pcpu_lock */
171	static LIST_HEAD(pcpu_map_extend_chunks);
172
173	/*
174	* The number of empty populated pages, protected by pcpu_lock. The
175	* reserved chunk doesn't contribute to the count.
176	*/
177	static int pcpu_nr_empty_pop_pages;
178
179	/*
180	* Balance work is used to populate or destroy chunks asynchronously. We
181	* try to keep the number of populated free pages between
182	* PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
183	* empty chunk.
184	*/
185	static void pcpu_balance_workfn(struct work_struct *work);
186	static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
187	static bool pcpu_async_enabled __read_mostly;
188	static bool pcpu_atomic_alloc_failed;
189
190	static void pcpu_schedule_balance_work(void)
191	{
192	if (pcpu_async_enabled)
193	schedule_work(&pcpu_balance_work);
194	}
195
196	static bool pcpu_addr_in_first_chunk(void *addr)
197	{
198	void *first_start = pcpu_first_chunk->base_addr;
199
200	return addr >= first_start && addr < first_start + pcpu_unit_size;
201	}
202
203	static bool pcpu_addr_in_reserved_chunk(void *addr)
204	{
205	void *first_start = pcpu_first_chunk->base_addr;
206
207	return addr >= first_start &&
208	addr < first_start + pcpu_reserved_chunk_limit;
209	}
210
211	static int __pcpu_size_to_slot(int size)
212	{
213	int highbit = fls(size); /* size is in bytes */
214	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
215	}
216
217	static int pcpu_size_to_slot(int size)
218	{
219	if (size == pcpu_unit_size)
220	return pcpu_nr_slots - 1;
221	return __pcpu_size_to_slot(size);
222	}
223
224	static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
225	{
226	if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
227	return 0;
228
229	return pcpu_size_to_slot(chunk->free_size);
230	}
231
232	/* set the pointer to a chunk in a page struct */
233	static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
234	{
235	page->index = (unsigned long)pcpu;
236	}
237
238	/* obtain pointer to a chunk from a page struct */
239	static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
240	{
241	return (struct pcpu_chunk *)page->index;
242	}
243
244	static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
245	{
246	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
247	}
248
249	static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
250	unsigned int cpu, int page_idx)
251	{
252	return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
253	(page_idx << PAGE_SHIFT);
254	}
255
256	static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
257	int *rs, int *re, int end)
258	{
259	*rs = find_next_zero_bit(chunk->populated, end, *rs);
260	*re = find_next_bit(chunk->populated, end, *rs + 1);
261	}
262
263	static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
264	int *rs, int *re, int end)
265	{
266	*rs = find_next_bit(chunk->populated, end, *rs);
267	*re = find_next_zero_bit(chunk->populated, end, *rs + 1);
268	}
269
270	/*
271	* (Un)populated page region iterators. Iterate over (un)populated
272	* page regions between @start and @end in @chunk. @rs and @re should
273	* be integer variables and will be set to start and end page index of
274	* the current region.
275	*/
276	#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
277	for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
278	(rs) < (re); \
279	(rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
280
281	#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
282	for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
283	(rs) < (re); \
284	(rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
285
286	/**
287	* pcpu_mem_zalloc - allocate memory
288	* @size: bytes to allocate
289	*
290	* Allocate @size bytes. If @size is smaller than PAGE_SIZE,
291	* kzalloc() is used; otherwise, vzalloc() is used. The returned
292	* memory is always zeroed.
293	*
294	* CONTEXT:
295	* Does GFP_KERNEL allocation.
296	*
297	* RETURNS:
298	* Pointer to the allocated area on success, NULL on failure.
299	*/
300	static void *pcpu_mem_zalloc(size_t size)
301	{
302	if (WARN_ON_ONCE(!slab_is_available()))
303	return NULL;
304
305	if (size <= PAGE_SIZE)
306	return kzalloc(size, GFP_KERNEL);
307	else
308	return vzalloc(size);
309	}
310
311	/**
312	* pcpu_mem_free - free memory
313	* @ptr: memory to free
314	*
315	* Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
316	*/
317	static void pcpu_mem_free(void *ptr)
318	{
319	kvfree(ptr);
320	}
321
322	/**
323	* pcpu_count_occupied_pages - count the number of pages an area occupies
324	* @chunk: chunk of interest
325	* @i: index of the area in question
326	*
327	* Count the number of pages chunk's @i'th area occupies. When the area's
328	* start and/or end address isn't aligned to page boundary, the straddled
329	* page is included in the count iff the rest of the page is free.
330	*/
331	static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
332	{
333	int off = chunk->map[i] & ~1;
334	int end = chunk->map[i + 1] & ~1;
335
336	if (!PAGE_ALIGNED(off) && i > 0) {
337	int prev = chunk->map[i - 1];
338
339	if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
340	off = round_down(off, PAGE_SIZE);
341	}
342
343	if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
344	int next = chunk->map[i + 1];
345	int nend = chunk->map[i + 2] & ~1;
346
347	if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
348	end = round_up(end, PAGE_SIZE);
349	}
350
351	return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
352	}
353
354	/**
355	* pcpu_chunk_relocate - put chunk in the appropriate chunk slot
356	* @chunk: chunk of interest
357	* @oslot: the previous slot it was on
358	*
359	* This function is called after an allocation or free changed @chunk.
360	* New slot according to the changed state is determined and @chunk is
361	* moved to the slot. Note that the reserved chunk is never put on
362	* chunk slots.
363	*
364	* CONTEXT:
365	* pcpu_lock.
366	*/
367	static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
368	{
369	int nslot = pcpu_chunk_slot(chunk);
370
371	if (chunk != pcpu_reserved_chunk && oslot != nslot) {
372	if (oslot < nslot)
373	list_move(&chunk->list, &pcpu_slot[nslot]);
374	else
375	list_move_tail(&chunk->list, &pcpu_slot[nslot]);
376	}
377	}
378
379	/**
380	* pcpu_need_to_extend - determine whether chunk area map needs to be extended
381	* @chunk: chunk of interest
382	* @is_atomic: the allocation context
383	*
384	* Determine whether area map of @chunk needs to be extended. If
385	* @is_atomic, only the amount necessary for a new allocation is
386	* considered; however, async extension is scheduled if the left amount is
387	* low. If !@is_atomic, it aims for more empty space. Combined, this
388	* ensures that the map is likely to have enough available space to
389	* accomodate atomic allocations which can't extend maps directly.
390	*
391	* CONTEXT:
392	* pcpu_lock.
393	*
394	* RETURNS:
395	* New target map allocation length if extension is necessary, 0
396	* otherwise.
397	*/
398	static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
399	{
400	int margin, new_alloc;
401
402	lockdep_assert_held(&pcpu_lock);
403
404	if (is_atomic) {
405	margin = 3;
406
407	if (chunk->map_alloc <
408	chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) {
409	if (list_empty(&chunk->map_extend_list)) {
410	list_add_tail(&chunk->map_extend_list,
411	&pcpu_map_extend_chunks);
412	pcpu_schedule_balance_work();
413	}
414	}
415	} else {
416	margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
417	}
418
419	if (chunk->map_alloc >= chunk->map_used + margin)
420	return 0;
421
422	new_alloc = PCPU_DFL_MAP_ALLOC;
423	while (new_alloc < chunk->map_used + margin)
424	new_alloc *= 2;
425
426	return new_alloc;
427	}
428
429	/**
430	* pcpu_extend_area_map - extend area map of a chunk
431	* @chunk: chunk of interest
432	* @new_alloc: new target allocation length of the area map
433	*
434	* Extend area map of @chunk to have @new_alloc entries.
435	*
436	* CONTEXT:
437	* Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
438	*
439	* RETURNS:
440	* 0 on success, -errno on failure.
441	*/
442	static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
443	{
444	int *old = NULL, *new = NULL;
445	size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
446	unsigned long flags;
447
448	lockdep_assert_held(&pcpu_alloc_mutex);
449
450	new = pcpu_mem_zalloc(new_size);
451	if (!new)
452	return -ENOMEM;
453
454	/* acquire pcpu_lock and switch to new area map */
455	spin_lock_irqsave(&pcpu_lock, flags);
456
457	if (new_alloc <= chunk->map_alloc)
458	goto out_unlock;
459
460	old_size = chunk->map_alloc * sizeof(chunk->map[0]);
461	old = chunk->map;
462
463	memcpy(new, old, old_size);
464
465	chunk->map_alloc = new_alloc;
466	chunk->map = new;
467	new = NULL;
468
469	out_unlock:
470	spin_unlock_irqrestore(&pcpu_lock, flags);
471
472	/*
473	* pcpu_mem_free() might end up calling vfree() which uses
474	* IRQ-unsafe lock and thus can't be called under pcpu_lock.
475	*/
476	pcpu_mem_free(old);
477	pcpu_mem_free(new);
478
479	return 0;
480	}
481
482	/**
483	* pcpu_fit_in_area - try to fit the requested allocation in a candidate area
484	* @chunk: chunk the candidate area belongs to
485	* @off: the offset to the start of the candidate area
486	* @this_size: the size of the candidate area
487	* @size: the size of the target allocation
488	* @align: the alignment of the target allocation
489	* @pop_only: only allocate from already populated region
490	*
491	* We're trying to allocate @size bytes aligned at @align. @chunk's area
492	* at @off sized @this_size is a candidate. This function determines
493	* whether the target allocation fits in the candidate area and returns the
494	* number of bytes to pad after @off. If the target area doesn't fit, -1
495	* is returned.
496	*
497	* If @pop_only is %true, this function only considers the already
498	* populated part of the candidate area.
499	*/
500	static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
501	int size, int align, bool pop_only)
502	{
503	int cand_off = off;
504
505	while (true) {
506	int head = ALIGN(cand_off, align) - off;
507	int page_start, page_end, rs, re;
508
509	if (this_size < head + size)
510	return -1;
511
512	if (!pop_only)
513	return head;
514
515	/*
516	* If the first unpopulated page is beyond the end of the
517	* allocation, the whole allocation is populated;
518	* otherwise, retry from the end of the unpopulated area.
519	*/
520	page_start = PFN_DOWN(head + off);
521	page_end = PFN_UP(head + off + size);
522
523	rs = page_start;
524	pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
525	if (rs >= page_end)
526	return head;
527	cand_off = re * PAGE_SIZE;
528	}
529	}
530
531	/**
532	* pcpu_alloc_area - allocate area from a pcpu_chunk
533	* @chunk: chunk of interest
534	* @size: wanted size in bytes
535	* @align: wanted align
536	* @pop_only: allocate only from the populated area
537	* @occ_pages_p: out param for the number of pages the area occupies
538	*
539	* Try to allocate @size bytes area aligned at @align from @chunk.
540	* Note that this function only allocates the offset. It doesn't
541	* populate or map the area.
542	*
543	* @chunk->map must have at least two free slots.
544	*
545	* CONTEXT:
546	* pcpu_lock.
547	*
548	* RETURNS:
549	* Allocated offset in @chunk on success, -1 if no matching area is
550	* found.
551	*/
552	static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
553	bool pop_only, int *occ_pages_p)
554	{
555	int oslot = pcpu_chunk_slot(chunk);
556	int max_contig = 0;
557	int i, off;
558	bool seen_free = false;
559	int *p;
560
561	for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
562	int head, tail;
563	int this_size;
564
565	off = *p;
566	if (off & 1)
567	continue;
568
569	this_size = (p[1] & ~1) - off;
570
571	head = pcpu_fit_in_area(chunk, off, this_size, size, align,
572	pop_only);
573	if (head < 0) {
574	if (!seen_free) {
575	chunk->first_free = i;
576	seen_free = true;
577	}
578	max_contig = max(this_size, max_contig);
579	continue;
580	}
581
582	/*
583	* If head is small or the previous block is free,
584	* merge'em. Note that 'small' is defined as smaller
585	* than sizeof(int), which is very small but isn't too
586	* uncommon for percpu allocations.
587	*/
588	if (head && (head < sizeof(int) || !(p[-1] & 1))) {
589	*p = off += head;
590	if (p[-1] & 1)
591	chunk->free_size -= head;
592	else
593	max_contig = max(*p - p[-1], max_contig);
594	this_size -= head;
595	head = 0;
596	}
597
598	/* if tail is small, just keep it around */
599	tail = this_size - head - size;
600	if (tail < sizeof(int)) {
601	tail = 0;
602	size = this_size - head;
603	}
604
605	/* split if warranted */
606	if (head || tail) {
607	int nr_extra = !!head + !!tail;
608
609	/* insert new subblocks */
610	memmove(p + nr_extra + 1, p + 1,
611	sizeof(chunk->map[0]) * (chunk->map_used - i));
612	chunk->map_used += nr_extra;
613
614	if (head) {
615	if (!seen_free) {
616	chunk->first_free = i;
617	seen_free = true;
618	}
619	*++p = off += head;
620	++i;
621	max_contig = max(head, max_contig);
622	}
623	if (tail) {
624	p[1] = off + size;
625	max_contig = max(tail, max_contig);
626	}
627	}
628
629	if (!seen_free)
630	chunk->first_free = i + 1;
631
632	/* update hint and mark allocated */
633	if (i + 1 == chunk->map_used)
634	chunk->contig_hint = max_contig; /* fully scanned */
635	else
636	chunk->contig_hint = max(chunk->contig_hint,
637	max_contig);
638
639	chunk->free_size -= size;
640	*p |= 1;
641
642	*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
643	pcpu_chunk_relocate(chunk, oslot);
644	return off;
645	}
646
647	chunk->contig_hint = max_contig; /* fully scanned */
648	pcpu_chunk_relocate(chunk, oslot);
649
650	/* tell the upper layer that this chunk has no matching area */
651	return -1;
652	}
653
654	/**
655	* pcpu_free_area - free area to a pcpu_chunk
656	* @chunk: chunk of interest
657	* @freeme: offset of area to free
658	* @occ_pages_p: out param for the number of pages the area occupies
659	*
660	* Free area starting from @freeme to @chunk. Note that this function
661	* only modifies the allocation map. It doesn't depopulate or unmap
662	* the area.
663	*
664	* CONTEXT:
665	* pcpu_lock.
666	*/
667	static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
668	int *occ_pages_p)
669	{
670	int oslot = pcpu_chunk_slot(chunk);
671	int off = 0;
672	unsigned i, j;
673	int to_free = 0;
674	int *p;
675
676	freeme |= 1; /* we are searching for <given offset, in use> pair */
677
678	i = 0;
679	j = chunk->map_used;
680	while (i != j) {
681	unsigned k = (i + j) / 2;
682	off = chunk->map[k];
683	if (off < freeme)
684	i = k + 1;
685	else if (off > freeme)
686	j = k;
687	else
688	i = j = k;
689	}
690	BUG_ON(off != freeme);
691
692	if (i < chunk->first_free)
693	chunk->first_free = i;
694
695	p = chunk->map + i;
696	*p = off &= ~1;
697	chunk->free_size += (p[1] & ~1) - off;
698
699	*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
700
701	/* merge with next? */
702	if (!(p[1] & 1))
703	to_free++;
704	/* merge with previous? */
705	if (i > 0 && !(p[-1] & 1)) {
706	to_free++;
707	i--;
708	p--;
709	}
710	if (to_free) {
711	chunk->map_used -= to_free;
712	memmove(p + 1, p + 1 + to_free,
713	(chunk->map_used - i) * sizeof(chunk->map[0]));
714	}
715
716	chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
717	pcpu_chunk_relocate(chunk, oslot);
718	}
719
720	static struct pcpu_chunk *pcpu_alloc_chunk(void)
721	{
722	struct pcpu_chunk *chunk;
723
724	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
725	if (!chunk)
726	return NULL;
727
728	chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
729	sizeof(chunk->map[0]));
730	if (!chunk->map) {
731	pcpu_mem_free(chunk);
732	return NULL;
733	}
734
735	chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
736	chunk->map[0] = 0;
737	chunk->map[1] = pcpu_unit_size | 1;
738	chunk->map_used = 1;
739
740	INIT_LIST_HEAD(&chunk->list);
741	INIT_LIST_HEAD(&chunk->map_extend_list);
742	chunk->free_size = pcpu_unit_size;
743	chunk->contig_hint = pcpu_unit_size;
744
745	return chunk;
746	}
747
748	static void pcpu_free_chunk(struct pcpu_chunk *chunk)
749	{
750	if (!chunk)
751	return;
752	pcpu_mem_free(chunk->map);
753	pcpu_mem_free(chunk);
754	}
755
756	/**
757	* pcpu_chunk_populated - post-population bookkeeping
758	* @chunk: pcpu_chunk which got populated
759	* @page_start: the start page
760	* @page_end: the end page
761	*
762	* Pages in [@page_start,@page_end) have been populated to @chunk. Update
763	* the bookkeeping information accordingly. Must be called after each
764	* successful population.
765	*/
766	static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
767	int page_start, int page_end)
768	{
769	int nr = page_end - page_start;
770
771	lockdep_assert_held(&pcpu_lock);
772
773	bitmap_set(chunk->populated, page_start, nr);
774	chunk->nr_populated += nr;
775	pcpu_nr_empty_pop_pages += nr;
776	}
777
778	/**
779	* pcpu_chunk_depopulated - post-depopulation bookkeeping
780	* @chunk: pcpu_chunk which got depopulated
781	* @page_start: the start page
782	* @page_end: the end page
783	*
784	* Pages in [@page_start,@page_end) have been depopulated from @chunk.
785	* Update the bookkeeping information accordingly. Must be called after
786	* each successful depopulation.
787	*/
788	static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
789	int page_start, int page_end)
790	{
791	int nr = page_end - page_start;
792
793	lockdep_assert_held(&pcpu_lock);
794
795	bitmap_clear(chunk->populated, page_start, nr);
796	chunk->nr_populated -= nr;
797	pcpu_nr_empty_pop_pages -= nr;
798	}
799
800	/*
801	* Chunk management implementation.
802	*
803	* To allow different implementations, chunk alloc/free and
804	* [de]population are implemented in a separate file which is pulled
805	* into this file and compiled together. The following functions
806	* should be implemented.
807	*
808	* pcpu_populate_chunk - populate the specified range of a chunk
809	* pcpu_depopulate_chunk - depopulate the specified range of a chunk
810	* pcpu_create_chunk - create a new chunk
811	* pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
812	* pcpu_addr_to_page - translate address to physical address
813	* pcpu_verify_alloc_info - check alloc_info is acceptable during init
814	*/
815	static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
816	static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
817	static struct pcpu_chunk *pcpu_create_chunk(void);
818	static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
819	static struct page *pcpu_addr_to_page(void *addr);
820	static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
821
822	#ifdef CONFIG_NEED_PER_CPU_KM
823	#include "percpu-km.c"
824	#else
825	#include "percpu-vm.c"
826	#endif
827
828	/**
829	* pcpu_chunk_addr_search - determine chunk containing specified address
830	* @addr: address for which the chunk needs to be determined.
831	*
832	* RETURNS:
833	* The address of the found chunk.
834	*/
835	static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
836	{
837	/* is it in the first chunk? */
838	if (pcpu_addr_in_first_chunk(addr)) {
839	/* is it in the reserved area? */
840	if (pcpu_addr_in_reserved_chunk(addr))
841	return pcpu_reserved_chunk;
842	return pcpu_first_chunk;
843	}
844
845	/*
846	* The address is relative to unit0 which might be unused and
847	* thus unmapped. Offset the address to the unit space of the
848	* current processor before looking it up in the vmalloc
849	* space. Note that any possible cpu id can be used here, so
850	* there's no need to worry about preemption or cpu hotplug.
851	*/
852	addr += pcpu_unit_offsets[raw_smp_processor_id()];
853	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
854	}
855
856	/**
857	* pcpu_alloc - the percpu allocator
858	* @size: size of area to allocate in bytes
859	* @align: alignment of area (max PAGE_SIZE)
860	* @reserved: allocate from the reserved chunk if available
861	* @gfp: allocation flags
862	*
863	* Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
864	* contain %GFP_KERNEL, the allocation is atomic.
865	*
866	* RETURNS:
867	* Percpu pointer to the allocated area on success, NULL on failure.
868	*/
869	static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
870	gfp_t gfp)
871	{
872	static int warn_limit = 10;
873	struct pcpu_chunk *chunk;
874	const char *err;
875	bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
876	int occ_pages = 0;
877	int slot, off, new_alloc, cpu, ret;
878	unsigned long flags;
879	void __percpu *ptr;
880
881	/*
882	* We want the lowest bit of offset available for in-use/free
883	* indicator, so force >= 16bit alignment and make size even.
884	*/
885	if (unlikely(align < 2))
886	align = 2;
887
888	size = ALIGN(size, 2);
889
890	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
891	WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
892	size, align);
893	return NULL;
894	}
895
896	if (!is_atomic)
897	mutex_lock(&pcpu_alloc_mutex);
898
899	spin_lock_irqsave(&pcpu_lock, flags);
900
901	/* serve reserved allocations from the reserved chunk if available */
902	if (reserved && pcpu_reserved_chunk) {
903	chunk = pcpu_reserved_chunk;
904
905	if (size > chunk->contig_hint) {
906	err = "alloc from reserved chunk failed";
907	goto fail_unlock;
908	}
909
910	while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
911	spin_unlock_irqrestore(&pcpu_lock, flags);
912	if (is_atomic ||
913	pcpu_extend_area_map(chunk, new_alloc) < 0) {
914	err = "failed to extend area map of reserved chunk";
915	goto fail;
916	}
917	spin_lock_irqsave(&pcpu_lock, flags);
918	}
919
920	off = pcpu_alloc_area(chunk, size, align, is_atomic,
921	&occ_pages);
922	if (off >= 0)
923	goto area_found;
924
925	err = "alloc from reserved chunk failed";
926	goto fail_unlock;
927	}
928
929	restart:
930	/* search through normal chunks */
931	for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
932	list_for_each_entry(chunk, &pcpu_slot[slot], list) {
933	if (size > chunk->contig_hint)
934	continue;
935
936	new_alloc = pcpu_need_to_extend(chunk, is_atomic);
937	if (new_alloc) {
938	if (is_atomic)
939	continue;
940	spin_unlock_irqrestore(&pcpu_lock, flags);
941	if (pcpu_extend_area_map(chunk,
942	new_alloc) < 0) {
943	err = "failed to extend area map";
944	goto fail;
945	}
946	spin_lock_irqsave(&pcpu_lock, flags);
947	/*
948	* pcpu_lock has been dropped, need to
949	* restart cpu_slot list walking.
950	*/
951	goto restart;
952	}
953
954	off = pcpu_alloc_area(chunk, size, align, is_atomic,
955	&occ_pages);
956	if (off >= 0)
957	goto area_found;
958	}
959	}
960
961	spin_unlock_irqrestore(&pcpu_lock, flags);
962
963	/*
964	* No space left. Create a new chunk. We don't want multiple
965	* tasks to create chunks simultaneously. Serialize and create iff
966	* there's still no empty chunk after grabbing the mutex.
967	*/
968	if (is_atomic)
969	goto fail;
970
971	if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
972	chunk = pcpu_create_chunk();
973	if (!chunk) {
974	err = "failed to allocate new chunk";
975	goto fail;
976	}
977
978	spin_lock_irqsave(&pcpu_lock, flags);
979	pcpu_chunk_relocate(chunk, -1);
980	} else {
981	spin_lock_irqsave(&pcpu_lock, flags);
982	}
983
984	goto restart;
985
986	area_found:
987	spin_unlock_irqrestore(&pcpu_lock, flags);
988
989	/* populate if not all pages are already there */
990	if (!is_atomic) {
991	int page_start, page_end, rs, re;
992
993	page_start = PFN_DOWN(off);
994	page_end = PFN_UP(off + size);
995
996	pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
997	WARN_ON(chunk->immutable);
998
999	ret = pcpu_populate_chunk(chunk, rs, re);
1000
1001	spin_lock_irqsave(&pcpu_lock, flags);
1002	if (ret) {
1003	pcpu_free_area(chunk, off, &occ_pages);
1004	err = "failed to populate";
1005	goto fail_unlock;
1006	}
1007	pcpu_chunk_populated(chunk, rs, re);
1008	spin_unlock_irqrestore(&pcpu_lock, flags);
1009	}
1010
1011	mutex_unlock(&pcpu_alloc_mutex);
1012	}
1013
1014	if (chunk != pcpu_reserved_chunk) {
1015	spin_lock_irqsave(&pcpu_lock, flags);
1016	pcpu_nr_empty_pop_pages -= occ_pages;
1017	spin_unlock_irqrestore(&pcpu_lock, flags);
1018	}
1019
1020	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1021	pcpu_schedule_balance_work();
1022
1023	/* clear the areas and return address relative to base address */
1024	for_each_possible_cpu(cpu)
1025	memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1026
1027	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1028	kmemleak_alloc_percpu(ptr, size, gfp);
1029	return ptr;
1030
1031	fail_unlock:
1032	spin_unlock_irqrestore(&pcpu_lock, flags);
1033	fail:
1034	if (!is_atomic && warn_limit) {
1035	pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1036	size, align, is_atomic, err);
1037	dump_stack();
1038	if (!--warn_limit)
1039	pr_info("limit reached, disable warning\n");
1040	}
1041	if (is_atomic) {
1042	/* see the flag handling in pcpu_blance_workfn() */
1043	pcpu_atomic_alloc_failed = true;
1044	pcpu_schedule_balance_work();
1045	} else {
1046	mutex_unlock(&pcpu_alloc_mutex);
1047	}
1048	return NULL;
1049	}
1050
1051	/**
1052	* __alloc_percpu_gfp - allocate dynamic percpu area
1053	* @size: size of area to allocate in bytes
1054	* @align: alignment of area (max PAGE_SIZE)
1055	* @gfp: allocation flags
1056	*
1057	* Allocate zero-filled percpu area of @size bytes aligned at @align. If
1058	* @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1059	* be called from any context but is a lot more likely to fail.
1060	*
1061	* RETURNS:
1062	* Percpu pointer to the allocated area on success, NULL on failure.
1063	*/
1064	void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1065	{
1066	return pcpu_alloc(size, align, false, gfp);
1067	}
1068	EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1069
1070	/**
1071	* __alloc_percpu - allocate dynamic percpu area
1072	* @size: size of area to allocate in bytes
1073	* @align: alignment of area (max PAGE_SIZE)
1074	*
1075	* Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1076	*/
1077	void __percpu *__alloc_percpu(size_t size, size_t align)
1078	{
1079	return pcpu_alloc(size, align, false, GFP_KERNEL);
1080	}
1081	EXPORT_SYMBOL_GPL(__alloc_percpu);
1082
1083	/**
1084	* __alloc_reserved_percpu - allocate reserved percpu area
1085	* @size: size of area to allocate in bytes
1086	* @align: alignment of area (max PAGE_SIZE)
1087	*
1088	* Allocate zero-filled percpu area of @size bytes aligned at @align
1089	* from reserved percpu area if arch has set it up; otherwise,
1090	* allocation is served from the same dynamic area. Might sleep.
1091	* Might trigger writeouts.
1092	*
1093	* CONTEXT:
1094	* Does GFP_KERNEL allocation.
1095	*
1096	* RETURNS:
1097	* Percpu pointer to the allocated area on success, NULL on failure.
1098	*/
1099	void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1100	{
1101	return pcpu_alloc(size, align, true, GFP_KERNEL);
1102	}
1103
1104	/**
1105	* pcpu_balance_workfn - manage the amount of free chunks and populated pages
1106	* @work: unused
1107	*
1108	* Reclaim all fully free chunks except for the first one.
1109	*/
1110	static void pcpu_balance_workfn(struct work_struct *work)
1111	{
1112	LIST_HEAD(to_free);
1113	struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1114	struct pcpu_chunk *chunk, *next;
1115	int slot, nr_to_pop, ret;
1116
1117	/*
1118	* There's no reason to keep around multiple unused chunks and VM
1119	* areas can be scarce. Destroy all free chunks except for one.
1120	*/
1121	mutex_lock(&pcpu_alloc_mutex);
1122	spin_lock_irq(&pcpu_lock);
1123
1124	list_for_each_entry_safe(chunk, next, free_head, list) {
1125	WARN_ON(chunk->immutable);
1126
1127	/* spare the first one */
1128	if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1129	continue;
1130
1131	list_del_init(&chunk->map_extend_list);
1132	list_move(&chunk->list, &to_free);
1133	}
1134
1135	spin_unlock_irq(&pcpu_lock);
1136
1137	list_for_each_entry_safe(chunk, next, &to_free, list) {
1138	int rs, re;
1139
1140	pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1141	pcpu_depopulate_chunk(chunk, rs, re);
1142	spin_lock_irq(&pcpu_lock);
1143	pcpu_chunk_depopulated(chunk, rs, re);
1144	spin_unlock_irq(&pcpu_lock);
1145	}
1146	pcpu_destroy_chunk(chunk);
1147	}
1148
1149	/* service chunks which requested async area map extension */
1150	do {
1151	int new_alloc = 0;
1152
1153	spin_lock_irq(&pcpu_lock);
1154
1155	chunk = list_first_entry_or_null(&pcpu_map_extend_chunks,
1156	struct pcpu_chunk, map_extend_list);
1157	if (chunk) {
1158	list_del_init(&chunk->map_extend_list);
1159	new_alloc = pcpu_need_to_extend(chunk, false);
1160	}
1161
1162	spin_unlock_irq(&pcpu_lock);
1163
1164	if (new_alloc)
1165	pcpu_extend_area_map(chunk, new_alloc);
1166	} while (chunk);
1167
1168	/*
1169	* Ensure there are certain number of free populated pages for
1170	* atomic allocs. Fill up from the most packed so that atomic
1171	* allocs don't increase fragmentation. If atomic allocation
1172	* failed previously, always populate the maximum amount. This
1173	* should prevent atomic allocs larger than PAGE_SIZE from keeping
1174	* failing indefinitely; however, large atomic allocs are not
1175	* something we support properly and can be highly unreliable and
1176	* inefficient.
1177	*/
1178	retry_pop:
1179	if (pcpu_atomic_alloc_failed) {
1180	nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1181	/* best effort anyway, don't worry about synchronization */
1182	pcpu_atomic_alloc_failed = false;
1183	} else {
1184	nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1185	pcpu_nr_empty_pop_pages,
1186	0, PCPU_EMPTY_POP_PAGES_HIGH);
1187	}
1188
1189	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1190	int nr_unpop = 0, rs, re;
1191
1192	if (!nr_to_pop)
1193	break;
1194
1195	spin_lock_irq(&pcpu_lock);
1196	list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1197	nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1198	if (nr_unpop)
1199	break;
1200	}
1201	spin_unlock_irq(&pcpu_lock);
1202
1203	if (!nr_unpop)
1204	continue;
1205
1206	/* @chunk can't go away while pcpu_alloc_mutex is held */
1207	pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1208	int nr = min(re - rs, nr_to_pop);
1209
1210	ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1211	if (!ret) {
1212	nr_to_pop -= nr;
1213	spin_lock_irq(&pcpu_lock);
1214	pcpu_chunk_populated(chunk, rs, rs + nr);
1215	spin_unlock_irq(&pcpu_lock);
1216	} else {
1217	nr_to_pop = 0;
1218	}
1219
1220	if (!nr_to_pop)
1221	break;
1222	}
1223	}
1224
1225	if (nr_to_pop) {
1226	/* ran out of chunks to populate, create a new one and retry */
1227	chunk = pcpu_create_chunk();
1228	if (chunk) {
1229	spin_lock_irq(&pcpu_lock);
1230	pcpu_chunk_relocate(chunk, -1);
1231	spin_unlock_irq(&pcpu_lock);
1232	goto retry_pop;
1233	}
1234	}
1235
1236	mutex_unlock(&pcpu_alloc_mutex);
1237	}
1238
1239	/**
1240	* free_percpu - free percpu area
1241	* @ptr: pointer to area to free
1242	*
1243	* Free percpu area @ptr.
1244	*
1245	* CONTEXT:
1246	* Can be called from atomic context.
1247	*/
1248	void free_percpu(void __percpu *ptr)
1249	{
1250	void *addr;
1251	struct pcpu_chunk *chunk;
1252	unsigned long flags;
1253	int off, occ_pages;
1254
1255	if (!ptr)
1256	return;
1257
1258	kmemleak_free_percpu(ptr);
1259
1260	addr = __pcpu_ptr_to_addr(ptr);
1261
1262	spin_lock_irqsave(&pcpu_lock, flags);
1263
1264	chunk = pcpu_chunk_addr_search(addr);
1265	off = addr - chunk->base_addr;
1266
1267	pcpu_free_area(chunk, off, &occ_pages);
1268
1269	if (chunk != pcpu_reserved_chunk)
1270	pcpu_nr_empty_pop_pages += occ_pages;
1271
1272	/* if there are more than one fully free chunks, wake up grim reaper */
1273	if (chunk->free_size == pcpu_unit_size) {
1274	struct pcpu_chunk *pos;
1275
1276	list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1277	if (pos != chunk) {
1278	pcpu_schedule_balance_work();
1279	break;
1280	}
1281	}
1282
1283	spin_unlock_irqrestore(&pcpu_lock, flags);
1284	}
1285	EXPORT_SYMBOL_GPL(free_percpu);
1286
1287	/**
1288	* is_kernel_percpu_address - test whether address is from static percpu area
1289	* @addr: address to test
1290	*
1291	* Test whether @addr belongs to in-kernel static percpu area. Module
1292	* static percpu areas are not considered. For those, use
1293	* is_module_percpu_address().
1294	*
1295	* RETURNS:
1296	* %true if @addr is from in-kernel static percpu area, %false otherwise.
1297	*/
1298	bool is_kernel_percpu_address(unsigned long addr)
1299	{
1300	#ifdef CONFIG_SMP
1301	const size_t static_size = __per_cpu_end - __per_cpu_start;
1302	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1303	unsigned int cpu;
1304
1305	for_each_possible_cpu(cpu) {
1306	void *start = per_cpu_ptr(base, cpu);
1307
1308	if ((void *)addr >= start && (void *)addr < start + static_size)
1309	return true;
1310	}
1311	#endif
1312	/* on UP, can't distinguish from other static vars, always false */
1313	return false;
1314	}
1315
1316	/**
1317	* per_cpu_ptr_to_phys - convert translated percpu address to physical address
1318	* @addr: the address to be converted to physical address
1319	*
1320	* Given @addr which is dereferenceable address obtained via one of
1321	* percpu access macros, this function translates it into its physical
1322	* address. The caller is responsible for ensuring @addr stays valid
1323	* until this function finishes.
1324	*
1325	* percpu allocator has special setup for the first chunk, which currently
1326	* supports either embedding in linear address space or vmalloc mapping,
1327	* and, from the second one, the backing allocator (currently either vm or
1328	* km) provides translation.
1329	*
1330	* The addr can be translated simply without checking if it falls into the
1331	* first chunk. But the current code reflects better how percpu allocator
1332	* actually works, and the verification can discover both bugs in percpu
1333	* allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1334	* code.
1335	*
1336	* RETURNS:
1337	* The physical address for @addr.
1338	*/
1339	phys_addr_t per_cpu_ptr_to_phys(void *addr)
1340	{
1341	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1342	bool in_first_chunk = false;
1343	unsigned long first_low, first_high;
1344	unsigned int cpu;
1345
1346	/*
1347	* The following test on unit_low/high isn't strictly
1348	* necessary but will speed up lookups of addresses which
1349	* aren't in the first chunk.
1350	*/
1351	first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1352	first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1353	pcpu_unit_pages);
1354	if ((unsigned long)addr >= first_low &&
1355	(unsigned long)addr < first_high) {
1356	for_each_possible_cpu(cpu) {
1357	void *start = per_cpu_ptr(base, cpu);
1358
1359	if (addr >= start && addr < start + pcpu_unit_size) {
1360	in_first_chunk = true;
1361	break;
1362	}
1363	}
1364	}
1365
1366	if (in_first_chunk) {
1367	if (!is_vmalloc_addr(addr))
1368	return __pa(addr);
1369	else
1370	return page_to_phys(vmalloc_to_page(addr)) +
1371	offset_in_page(addr);
1372	} else
1373	return page_to_phys(pcpu_addr_to_page(addr)) +
1374	offset_in_page(addr);
1375	}
1376
1377	/**
1378	* pcpu_alloc_alloc_info - allocate percpu allocation info
1379	* @nr_groups: the number of groups
1380	* @nr_units: the number of units
1381	*
1382	* Allocate ai which is large enough for @nr_groups groups containing
1383	* @nr_units units. The returned ai's groups[0].cpu_map points to the
1384	* cpu_map array which is long enough for @nr_units and filled with
1385	* NR_CPUS. It's the caller's responsibility to initialize cpu_map
1386	* pointer of other groups.
1387	*
1388	* RETURNS:
1389	* Pointer to the allocated pcpu_alloc_info on success, NULL on
1390	* failure.
1391	*/
1392	struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1393	int nr_units)
1394	{
1395	struct pcpu_alloc_info *ai;
1396	size_t base_size, ai_size;
1397	void *ptr;
1398	int unit;
1399
1400	base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1401	__alignof__(ai->groups[0].cpu_map[0]));
1402	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1403
1404	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1405	if (!ptr)
1406	return NULL;
1407	ai = ptr;
1408	ptr += base_size;
1409
1410	ai->groups[0].cpu_map = ptr;
1411
1412	for (unit = 0; unit < nr_units; unit++)
1413	ai->groups[0].cpu_map[unit] = NR_CPUS;
1414
1415	ai->nr_groups = nr_groups;
1416	ai->__ai_size = PFN_ALIGN(ai_size);
1417
1418	return ai;
1419	}
1420
1421	/**
1422	* pcpu_free_alloc_info - free percpu allocation info
1423	* @ai: pcpu_alloc_info to free
1424	*
1425	* Free @ai which was allocated by pcpu_alloc_alloc_info().
1426	*/
1427	void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1428	{
1429	memblock_free_early(__pa(ai), ai->__ai_size);
1430	}
1431
1432	/**
1433	* pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1434	* @lvl: loglevel
1435	* @ai: allocation info to dump
1436	*
1437	* Print out information about @ai using loglevel @lvl.
1438	*/
1439	static void pcpu_dump_alloc_info(const char *lvl,
1440	const struct pcpu_alloc_info *ai)
1441	{
1442	int group_width = 1, cpu_width = 1, width;
1443	char empty_str[] = "--------";
1444	int alloc = 0, alloc_end = 0;
1445	int group, v;
1446	int upa, apl; /* units per alloc, allocs per line */
1447
1448	v = ai->nr_groups;
1449	while (v /= 10)
1450	group_width++;
1451
1452	v = num_possible_cpus();
1453	while (v /= 10)
1454	cpu_width++;
1455	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1456
1457	upa = ai->alloc_size / ai->unit_size;
1458	width = upa * (cpu_width + 1) + group_width + 3;
1459	apl = rounddown_pow_of_two(max(60 / width, 1));
1460
1461	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1462	lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1463	ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1464
1465	for (group = 0; group < ai->nr_groups; group++) {
1466	const struct pcpu_group_info *gi = &ai->groups[group];
1467	int unit = 0, unit_end = 0;
1468
1469	BUG_ON(gi->nr_units % upa);
1470	for (alloc_end += gi->nr_units / upa;
1471	alloc < alloc_end; alloc++) {
1472	if (!(alloc % apl)) {
1473	pr_cont("\n");
1474	printk("%spcpu-alloc: ", lvl);
1475	}
1476	pr_cont("[%0*d] ", group_width, group);
1477
1478	for (unit_end += upa; unit < unit_end; unit++)
1479	if (gi->cpu_map[unit] != NR_CPUS)
1480	pr_cont("%0*d ",
1481	cpu_width, gi->cpu_map[unit]);
1482	else
1483	pr_cont("%s ", empty_str);
1484	}
1485	}
1486	pr_cont("\n");
1487	}
1488
1489	/**
1490	* pcpu_setup_first_chunk - initialize the first percpu chunk
1491	* @ai: pcpu_alloc_info describing how to percpu area is shaped
1492	* @base_addr: mapped address
1493	*
1494	* Initialize the first percpu chunk which contains the kernel static
1495	* perpcu area. This function is to be called from arch percpu area
1496	* setup path.
1497	*
1498	* @ai contains all information necessary to initialize the first
1499	* chunk and prime the dynamic percpu allocator.
1500	*
1501	* @ai->static_size is the size of static percpu area.
1502	*
1503	* @ai->reserved_size, if non-zero, specifies the amount of bytes to
1504	* reserve after the static area in the first chunk. This reserves
1505	* the first chunk such that it's available only through reserved
1506	* percpu allocation. This is primarily used to serve module percpu
1507	* static areas on architectures where the addressing model has
1508	* limited offset range for symbol relocations to guarantee module
1509	* percpu symbols fall inside the relocatable range.
1510	*
1511	* @ai->dyn_size determines the number of bytes available for dynamic
1512	* allocation in the first chunk. The area between @ai->static_size +
1513	* @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1514	*
1515	* @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1516	* and equal to or larger than @ai->static_size + @ai->reserved_size +
1517	* @ai->dyn_size.
1518	*
1519	* @ai->atom_size is the allocation atom size and used as alignment
1520	* for vm areas.
1521	*
1522	* @ai->alloc_size is the allocation size and always multiple of
1523	* @ai->atom_size. This is larger than @ai->atom_size if
1524	* @ai->unit_size is larger than @ai->atom_size.
1525	*
1526	* @ai->nr_groups and @ai->groups describe virtual memory layout of
1527	* percpu areas. Units which should be colocated are put into the
1528	* same group. Dynamic VM areas will be allocated according to these
1529	* groupings. If @ai->nr_groups is zero, a single group containing
1530	* all units is assumed.
1531	*
1532	* The caller should have mapped the first chunk at @base_addr and
1533	* copied static data to each unit.
1534	*
1535	* If the first chunk ends up with both reserved and dynamic areas, it
1536	* is served by two chunks - one to serve the core static and reserved
1537	* areas and the other for the dynamic area. They share the same vm
1538	* and page map but uses different area allocation map to stay away
1539	* from each other. The latter chunk is circulated in the chunk slots
1540	* and available for dynamic allocation like any other chunks.
1541	*
1542	* RETURNS:
1543	* 0 on success, -errno on failure.
1544	*/
1545	int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1546	void *base_addr)
1547	{
1548	static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1549	static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1550	size_t dyn_size = ai->dyn_size;
1551	size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1552	struct pcpu_chunk *schunk, *dchunk = NULL;
1553	unsigned long *group_offsets;
1554	size_t *group_sizes;
1555	unsigned long *unit_off;
1556	unsigned int cpu;
1557	int *unit_map;
1558	int group, unit, i;
1559
1560	#define PCPU_SETUP_BUG_ON(cond) do { \
1561	if (unlikely(cond)) { \
1562	pr_emerg("failed to initialize, %s\n", #cond); \
1563	pr_emerg("cpu_possible_mask=%*pb\n", \
1564	cpumask_pr_args(cpu_possible_mask)); \
1565	pcpu_dump_alloc_info(KERN_EMERG, ai); \
1566	BUG(); \
1567	} \
1568	} while (0)
1569
1570	/* sanity checks */
1571	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1572	#ifdef CONFIG_SMP
1573	PCPU_SETUP_BUG_ON(!ai->static_size);
1574	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1575	#endif
1576	PCPU_SETUP_BUG_ON(!base_addr);
1577	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1578	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1579	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1580	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1581	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1582	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1583
1584	/* process group information and build config tables accordingly */
1585	group_offsets = memblock_virt_alloc(ai->nr_groups *
1586	sizeof(group_offsets[0]), 0);
1587	group_sizes = memblock_virt_alloc(ai->nr_groups *
1588	sizeof(group_sizes[0]), 0);
1589	unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1590	unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1591
1592	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1593	unit_map[cpu] = UINT_MAX;
1594
1595	pcpu_low_unit_cpu = NR_CPUS;
1596	pcpu_high_unit_cpu = NR_CPUS;
1597
1598	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1599	const struct pcpu_group_info *gi = &ai->groups[group];
1600
1601	group_offsets[group] = gi->base_offset;
1602	group_sizes[group] = gi->nr_units * ai->unit_size;
1603
1604	for (i = 0; i < gi->nr_units; i++) {
1605	cpu = gi->cpu_map[i];
1606	if (cpu == NR_CPUS)
1607	continue;
1608
1609	PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1610	PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1611	PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1612
1613	unit_map[cpu] = unit + i;
1614	unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1615
1616	/* determine low/high unit_cpu */
1617	if (pcpu_low_unit_cpu == NR_CPUS ||
1618	unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1619	pcpu_low_unit_cpu = cpu;
1620	if (pcpu_high_unit_cpu == NR_CPUS ||
1621	unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1622	pcpu_high_unit_cpu = cpu;
1623	}
1624	}
1625	pcpu_nr_units = unit;
1626
1627	for_each_possible_cpu(cpu)
1628	PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1629
1630	/* we're done parsing the input, undefine BUG macro and dump config */
1631	#undef PCPU_SETUP_BUG_ON
1632	pcpu_dump_alloc_info(KERN_DEBUG, ai);
1633
1634	pcpu_nr_groups = ai->nr_groups;
1635	pcpu_group_offsets = group_offsets;
1636	pcpu_group_sizes = group_sizes;
1637	pcpu_unit_map = unit_map;
1638	pcpu_unit_offsets = unit_off;
1639
1640	/* determine basic parameters */
1641	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1642	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1643	pcpu_atom_size = ai->atom_size;
1644	pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1645	BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1646
1647	/*
1648	* Allocate chunk slots. The additional last slot is for
1649	* empty chunks.
1650	*/
1651	pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1652	pcpu_slot = memblock_virt_alloc(
1653	pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1654	for (i = 0; i < pcpu_nr_slots; i++)
1655	INIT_LIST_HEAD(&pcpu_slot[i]);
1656
1657	/*
1658	* Initialize static chunk. If reserved_size is zero, the
1659	* static chunk covers static area + dynamic allocation area
1660	* in the first chunk. If reserved_size is not zero, it
1661	* covers static area + reserved area (mostly used for module
1662	* static percpu allocation).
1663	*/
1664	schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1665	INIT_LIST_HEAD(&schunk->list);
1666	INIT_LIST_HEAD(&schunk->map_extend_list);
1667	schunk->base_addr = base_addr;
1668	schunk->map = smap;
1669	schunk->map_alloc = ARRAY_SIZE(smap);
1670	schunk->immutable = true;
1671	bitmap_fill(schunk->populated, pcpu_unit_pages);
1672	schunk->nr_populated = pcpu_unit_pages;
1673
1674	if (ai->reserved_size) {
1675	schunk->free_size = ai->reserved_size;
1676	pcpu_reserved_chunk = schunk;
1677	pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1678	} else {
1679	schunk->free_size = dyn_size;
1680	dyn_size = 0; /* dynamic area covered */
1681	}
1682	schunk->contig_hint = schunk->free_size;
1683
1684	schunk->map[0] = 1;
1685	schunk->map[1] = ai->static_size;
1686	schunk->map_used = 1;
1687	if (schunk->free_size)
1688	schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1689	schunk->map[schunk->map_used] |= 1;
1690
1691	/* init dynamic chunk if necessary */
1692	if (dyn_size) {
1693	dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1694	INIT_LIST_HEAD(&dchunk->list);
1695	INIT_LIST_HEAD(&dchunk->map_extend_list);
1696	dchunk->base_addr = base_addr;
1697	dchunk->map = dmap;
1698	dchunk->map_alloc = ARRAY_SIZE(dmap);
1699	dchunk->immutable = true;
1700	bitmap_fill(dchunk->populated, pcpu_unit_pages);
1701	dchunk->nr_populated = pcpu_unit_pages;
1702
1703	dchunk->contig_hint = dchunk->free_size = dyn_size;
1704	dchunk->map[0] = 1;
1705	dchunk->map[1] = pcpu_reserved_chunk_limit;
1706	dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1707	dchunk->map_used = 2;
1708	}
1709
1710	/* link the first chunk in */
1711	pcpu_first_chunk = dchunk ?: schunk;
1712	pcpu_nr_empty_pop_pages +=
1713	pcpu_count_occupied_pages(pcpu_first_chunk, 1);
1714	pcpu_chunk_relocate(pcpu_first_chunk, -1);
1715
1716	/* we're done */
1717	pcpu_base_addr = base_addr;
1718	return 0;
1719	}
1720
1721	#ifdef CONFIG_SMP
1722
1723	const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1724	[PCPU_FC_AUTO] = "auto",
1725	[PCPU_FC_EMBED] = "embed",
1726	[PCPU_FC_PAGE] = "page",
1727	};
1728
1729	enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1730
1731	static int __init percpu_alloc_setup(char *str)
1732	{
1733	if (!str)
1734	return -EINVAL;
1735
1736	if (0)
1737	/* nada */;
1738	#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1739	else if (!strcmp(str, "embed"))
1740	pcpu_chosen_fc = PCPU_FC_EMBED;
1741	#endif
1742	#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1743	else if (!strcmp(str, "page"))
1744	pcpu_chosen_fc = PCPU_FC_PAGE;
1745	#endif
1746	else
1747	pr_warn("unknown allocator %s specified\n", str);
1748
1749	return 0;
1750	}
1751	early_param("percpu_alloc", percpu_alloc_setup);
1752
1753	/*
1754	* pcpu_embed_first_chunk() is used by the generic percpu setup.
1755	* Build it if needed by the arch config or the generic setup is going
1756	* to be used.
1757	*/
1758	#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1759	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1760	#define BUILD_EMBED_FIRST_CHUNK
1761	#endif
1762
1763	/* build pcpu_page_first_chunk() iff needed by the arch config */
1764	#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1765	#define BUILD_PAGE_FIRST_CHUNK
1766	#endif
1767
1768	/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1769	#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1770	/**
1771	* pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1772	* @reserved_size: the size of reserved percpu area in bytes
1773	* @dyn_size: minimum free size for dynamic allocation in bytes
1774	* @atom_size: allocation atom size
1775	* @cpu_distance_fn: callback to determine distance between cpus, optional
1776	*
1777	* This function determines grouping of units, their mappings to cpus
1778	* and other parameters considering needed percpu size, allocation
1779	* atom size and distances between CPUs.
1780	*
1781	* Groups are always multiples of atom size and CPUs which are of
1782	* LOCAL_DISTANCE both ways are grouped together and share space for
1783	* units in the same group. The returned configuration is guaranteed
1784	* to have CPUs on different nodes on different groups and >=75% usage
1785	* of allocated virtual address space.
1786	*
1787	* RETURNS:
1788	* On success, pointer to the new allocation_info is returned. On
1789	* failure, ERR_PTR value is returned.
1790	*/
1791	static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1792	size_t reserved_size, size_t dyn_size,
1793	size_t atom_size,
1794	pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1795	{
1796	static int group_map[NR_CPUS] __initdata;
1797	static int group_cnt[NR_CPUS] __initdata;
1798	const size_t static_size = __per_cpu_end - __per_cpu_start;
1799	int nr_groups = 1, nr_units = 0;
1800	size_t size_sum, min_unit_size, alloc_size;
1801	int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1802	int last_allocs, group, unit;
1803	unsigned int cpu, tcpu;
1804	struct pcpu_alloc_info *ai;
1805	unsigned int *cpu_map;
1806
1807	/* this function may be called multiple times */
1808	memset(group_map, 0, sizeof(group_map));
1809	memset(group_cnt, 0, sizeof(group_cnt));
1810
1811	/* calculate size_sum and ensure dyn_size is enough for early alloc */
1812	size_sum = PFN_ALIGN(static_size + reserved_size +
1813	max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1814	dyn_size = size_sum - static_size - reserved_size;
1815
1816	/*
1817	* Determine min_unit_size, alloc_size and max_upa such that
1818	* alloc_size is multiple of atom_size and is the smallest
1819	* which can accommodate 4k aligned segments which are equal to
1820	* or larger than min_unit_size.
1821	*/
1822	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1823
1824	alloc_size = roundup(min_unit_size, atom_size);
1825	upa = alloc_size / min_unit_size;
1826	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1827	upa--;
1828	max_upa = upa;
1829
1830	/* group cpus according to their proximity */
1831	for_each_possible_cpu(cpu) {
1832	group = 0;
1833	next_group:
1834	for_each_possible_cpu(tcpu) {
1835	if (cpu == tcpu)
1836	break;
1837	if (group_map[tcpu] == group && cpu_distance_fn &&
1838	(cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1839	cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1840	group++;
1841	nr_groups = max(nr_groups, group + 1);
1842	goto next_group;
1843	}
1844	}
1845	group_map[cpu] = group;
1846	group_cnt[group]++;
1847	}
1848
1849	/*
1850	* Expand unit size until address space usage goes over 75%
1851	* and then as much as possible without using more address
1852	* space.
1853	*/
1854	last_allocs = INT_MAX;
1855	for (upa = max_upa; upa; upa--) {
1856	int allocs = 0, wasted = 0;
1857
1858	if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1859	continue;
1860
1861	for (group = 0; group < nr_groups; group++) {
1862	int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1863	allocs += this_allocs;
1864	wasted += this_allocs * upa - group_cnt[group];
1865	}
1866
1867	/*
1868	* Don't accept if wastage is over 1/3. The
1869	* greater-than comparison ensures upa==1 always
1870	* passes the following check.
1871	*/
1872	if (wasted > num_possible_cpus() / 3)
1873	continue;
1874
1875	/* and then don't consume more memory */
1876	if (allocs > last_allocs)
1877	break;
1878	last_allocs = allocs;
1879	best_upa = upa;
1880	}
1881	upa = best_upa;
1882
1883	/* allocate and fill alloc_info */
1884	for (group = 0; group < nr_groups; group++)
1885	nr_units += roundup(group_cnt[group], upa);
1886
1887	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1888	if (!ai)
1889	return ERR_PTR(-ENOMEM);
1890	cpu_map = ai->groups[0].cpu_map;
1891
1892	for (group = 0; group < nr_groups; group++) {
1893	ai->groups[group].cpu_map = cpu_map;
1894	cpu_map += roundup(group_cnt[group], upa);
1895	}
1896
1897	ai->static_size = static_size;
1898	ai->reserved_size = reserved_size;
1899	ai->dyn_size = dyn_size;
1900	ai->unit_size = alloc_size / upa;
1901	ai->atom_size = atom_size;
1902	ai->alloc_size = alloc_size;
1903
1904	for (group = 0, unit = 0; group_cnt[group]; group++) {
1905	struct pcpu_group_info *gi = &ai->groups[group];
1906
1907	/*
1908	* Initialize base_offset as if all groups are located
1909	* back-to-back. The caller should update this to
1910	* reflect actual allocation.
1911	*/
1912	gi->base_offset = unit * ai->unit_size;
1913
1914	for_each_possible_cpu(cpu)
1915	if (group_map[cpu] == group)
1916	gi->cpu_map[gi->nr_units++] = cpu;
1917	gi->nr_units = roundup(gi->nr_units, upa);
1918	unit += gi->nr_units;
1919	}
1920	BUG_ON(unit != nr_units);
1921
1922	return ai;
1923	}
1924	#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1925
1926	#if defined(BUILD_EMBED_FIRST_CHUNK)
1927	/**
1928	* pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1929	* @reserved_size: the size of reserved percpu area in bytes
1930	* @dyn_size: minimum free size for dynamic allocation in bytes
1931	* @atom_size: allocation atom size
1932	* @cpu_distance_fn: callback to determine distance between cpus, optional
1933	* @alloc_fn: function to allocate percpu page
1934	* @free_fn: function to free percpu page
1935	*
1936	* This is a helper to ease setting up embedded first percpu chunk and
1937	* can be called where pcpu_setup_first_chunk() is expected.
1938	*
1939	* If this function is used to setup the first chunk, it is allocated
1940	* by calling @alloc_fn and used as-is without being mapped into
1941	* vmalloc area. Allocations are always whole multiples of @atom_size
1942	* aligned to @atom_size.
1943	*
1944	* This enables the first chunk to piggy back on the linear physical
1945	* mapping which often uses larger page size. Please note that this
1946	* can result in very sparse cpu->unit mapping on NUMA machines thus
1947	* requiring large vmalloc address space. Don't use this allocator if
1948	* vmalloc space is not orders of magnitude larger than distances
1949	* between node memory addresses (ie. 32bit NUMA machines).
1950	*
1951	* @dyn_size specifies the minimum dynamic area size.
1952	*
1953	* If the needed size is smaller than the minimum or specified unit
1954	* size, the leftover is returned using @free_fn.
1955	*
1956	* RETURNS:
1957	* 0 on success, -errno on failure.
1958	*/
1959	int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1960	size_t atom_size,
1961	pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1962	pcpu_fc_alloc_fn_t alloc_fn,
1963	pcpu_fc_free_fn_t free_fn)
1964	{
1965	void *base = (void *)ULONG_MAX;
1966	void **areas = NULL;
1967	struct pcpu_alloc_info *ai;
1968	size_t size_sum, areas_size;
1969	unsigned long max_distance;
1970	int group, i, highest_group, rc;
1971
1972	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1973	cpu_distance_fn);
1974	if (IS_ERR(ai))
1975	return PTR_ERR(ai);
1976
1977	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1978	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1979
1980	areas = memblock_virt_alloc_nopanic(areas_size, 0);
1981	if (!areas) {
1982	rc = -ENOMEM;
1983	goto out_free;
1984	}
1985
1986	/* allocate, copy and determine base address & max_distance */
1987	highest_group = 0;
1988	for (group = 0; group < ai->nr_groups; group++) {
1989	struct pcpu_group_info *gi = &ai->groups[group];
1990	unsigned int cpu = NR_CPUS;
1991	void *ptr;
1992
1993	for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1994	cpu = gi->cpu_map[i];
1995	BUG_ON(cpu == NR_CPUS);
1996
1997	/* allocate space for the whole group */
1998	ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1999	if (!ptr) {
2000	rc = -ENOMEM;
2001	goto out_free_areas;
2002	}
2003	/* kmemleak tracks the percpu allocations separately */
2004	kmemleak_free(ptr);
2005	areas[group] = ptr;
2006
2007	base = min(ptr, base);
2008	if (ptr > areas[highest_group])
2009	highest_group = group;
2010	}
2011	max_distance = areas[highest_group] - base;
2012	max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2013
2014	/* warn if maximum distance is further than 75% of vmalloc space */
2015	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2016	pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2017	max_distance, VMALLOC_TOTAL);
2018	#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2019	/* and fail if we have fallback */
2020	rc = -EINVAL;
2021	goto out_free_areas;
2022	#endif
2023	}
2024
2025	/*
2026	* Copy data and free unused parts. This should happen after all
2027	* allocations are complete; otherwise, we may end up with
2028	* overlapping groups.
2029	*/
2030	for (group = 0; group < ai->nr_groups; group++) {
2031	struct pcpu_group_info *gi = &ai->groups[group];
2032	void *ptr = areas[group];
2033
2034	for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2035	if (gi->cpu_map[i] == NR_CPUS) {
2036	/* unused unit, free whole */
2037	free_fn(ptr, ai->unit_size);
2038	continue;
2039	}
2040	/* copy and return the unused part */
2041	memcpy(ptr, __per_cpu_load, ai->static_size);
2042	free_fn(ptr + size_sum, ai->unit_size - size_sum);
2043	}
2044	}
2045
2046	/* base address is now known, determine group base offsets */
2047	for (group = 0; group < ai->nr_groups; group++) {
2048	ai->groups[group].base_offset = areas[group] - base;
2049	}
2050
2051	pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2052	PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2053	ai->dyn_size, ai->unit_size);
2054
2055	rc = pcpu_setup_first_chunk(ai, base);
2056	goto out_free;
2057
2058	out_free_areas:
2059	for (group = 0; group < ai->nr_groups; group++)
2060	if (areas[group])
2061	free_fn(areas[group],
2062	ai->groups[group].nr_units * ai->unit_size);
2063	out_free:
2064	pcpu_free_alloc_info(ai);
2065	if (areas)
2066	memblock_free_early(__pa(areas), areas_size);
2067	return rc;
2068	}
2069	#endif /* BUILD_EMBED_FIRST_CHUNK */
2070
2071	#ifdef BUILD_PAGE_FIRST_CHUNK
2072	/**
2073	* pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2074	* @reserved_size: the size of reserved percpu area in bytes
2075	* @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2076	* @free_fn: function to free percpu page, always called with PAGE_SIZE
2077	* @populate_pte_fn: function to populate pte
2078	*
2079	* This is a helper to ease setting up page-remapped first percpu
2080	* chunk and can be called where pcpu_setup_first_chunk() is expected.
2081	*
2082	* This is the basic allocator. Static percpu area is allocated
2083	* page-by-page into vmalloc area.
2084	*
2085	* RETURNS:
2086	* 0 on success, -errno on failure.
2087	*/
2088	int __init pcpu_page_first_chunk(size_t reserved_size,
2089	pcpu_fc_alloc_fn_t alloc_fn,
2090	pcpu_fc_free_fn_t free_fn,
2091	pcpu_fc_populate_pte_fn_t populate_pte_fn)
2092	{
2093	static struct vm_struct vm;
2094	struct pcpu_alloc_info *ai;
2095	char psize_str[16];
2096	int unit_pages;
2097	size_t pages_size;
2098	struct page **pages;
2099	int unit, i, j, rc;
2100
2101	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2102
2103	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2104	if (IS_ERR(ai))
2105	return PTR_ERR(ai);
2106	BUG_ON(ai->nr_groups != 1);
2107	BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2108
2109	unit_pages = ai->unit_size >> PAGE_SHIFT;
2110
2111	/* unaligned allocations can't be freed, round up to page size */
2112	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2113	sizeof(pages[0]));
2114	pages = memblock_virt_alloc(pages_size, 0);
2115
2116	/* allocate pages */
2117	j = 0;
2118	for (unit = 0; unit < num_possible_cpus(); unit++)
2119	for (i = 0; i < unit_pages; i++) {
2120	unsigned int cpu = ai->groups[0].cpu_map[unit];
2121	void *ptr;
2122
2123	ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2124	if (!ptr) {
2125	pr_warn("failed to allocate %s page for cpu%u\n",
2126	psize_str, cpu);
2127	goto enomem;
2128	}
2129	/* kmemleak tracks the percpu allocations separately */
2130	kmemleak_free(ptr);
2131	pages[j++] = virt_to_page(ptr);
2132	}
2133
2134	/* allocate vm area, map the pages and copy static data */
2135	vm.flags = VM_ALLOC;
2136	vm.size = num_possible_cpus() * ai->unit_size;
2137	vm_area_register_early(&vm, PAGE_SIZE);
2138
2139	for (unit = 0; unit < num_possible_cpus(); unit++) {
2140	unsigned long unit_addr =
2141	(unsigned long)vm.addr + unit * ai->unit_size;
2142
2143	for (i = 0; i < unit_pages; i++)
2144	populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2145
2146	/* pte already populated, the following shouldn't fail */
2147	rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2148	unit_pages);
2149	if (rc < 0)
2150	panic("failed to map percpu area, err=%d\n", rc);
2151
2152	/*
2153	* FIXME: Archs with virtual cache should flush local
2154	* cache for the linear mapping here - something
2155	* equivalent to flush_cache_vmap() on the local cpu.
2156	* flush_cache_vmap() can't be used as most supporting
2157	* data structures are not set up yet.
2158	*/
2159
2160	/* copy static data */
2161	memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2162	}
2163
2164	/* we're ready, commit */
2165	pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2166	unit_pages, psize_str, ai->static_size,
2167	ai->reserved_size, ai->dyn_size);
2168
2169	rc = pcpu_setup_first_chunk(ai, vm.addr);
2170	goto out_free_ar;
2171
2172	enomem:
2173	while (--j >= 0)
2174	free_fn(page_address(pages[j]), PAGE_SIZE);
2175	rc = -ENOMEM;
2176	out_free_ar:
2177	memblock_free_early(__pa(pages), pages_size);
2178	pcpu_free_alloc_info(ai);
2179	return rc;
2180	}
2181	#endif /* BUILD_PAGE_FIRST_CHUNK */
2182
2183	#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2184	/*
2185	* Generic SMP percpu area setup.
2186	*
2187	* The embedding helper is used because its behavior closely resembles
2188	* the original non-dynamic generic percpu area setup. This is
2189	* important because many archs have addressing restrictions and might
2190	* fail if the percpu area is located far away from the previous
2191	* location. As an added bonus, in non-NUMA cases, embedding is
2192	* generally a good idea TLB-wise because percpu area can piggy back
2193	* on the physical linear memory mapping which uses large page
2194	* mappings on applicable archs.
2195	*/
2196	unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2197	EXPORT_SYMBOL(__per_cpu_offset);
2198
2199	static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2200	size_t align)
2201	{
2202	return memblock_virt_alloc_from_nopanic(
2203	size, align, __pa(MAX_DMA_ADDRESS));
2204	}
2205
2206	static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2207	{
2208	memblock_free_early(__pa(ptr), size);
2209	}
2210
2211	void __init setup_per_cpu_areas(void)
2212	{
2213	unsigned long delta;
2214	unsigned int cpu;
2215	int rc;
2216
2217	/*
2218	* Always reserve area for module percpu variables. That's
2219	* what the legacy allocator did.
2220	*/
2221	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2222	PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2223	pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2224	if (rc < 0)
2225	panic("Failed to initialize percpu areas.");
2226
2227	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2228	for_each_possible_cpu(cpu)
2229	__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2230	}
2231	#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2232
2233	#else /* CONFIG_SMP */
2234
2235	/*
2236	* UP percpu area setup.
2237	*
2238	* UP always uses km-based percpu allocator with identity mapping.
2239	* Static percpu variables are indistinguishable from the usual static
2240	* variables and don't require any special preparation.
2241	*/
2242	void __init setup_per_cpu_areas(void)
2243	{
2244	const size_t unit_size =
2245	roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2246	PERCPU_DYNAMIC_RESERVE));
2247	struct pcpu_alloc_info *ai;
2248	void *fc;
2249
2250	ai = pcpu_alloc_alloc_info(1, 1);
2251	fc = memblock_virt_alloc_from_nopanic(unit_size,
2252	PAGE_SIZE,
2253	__pa(MAX_DMA_ADDRESS));
2254	if (!ai || !fc)
2255	panic("Failed to allocate memory for percpu areas.");
2256	/* kmemleak tracks the percpu allocations separately */
2257	kmemleak_free(fc);
2258
2259	ai->dyn_size = unit_size;
2260	ai->unit_size = unit_size;
2261	ai->atom_size = unit_size;
2262	ai->alloc_size = unit_size;
2263	ai->groups[0].nr_units = 1;
2264	ai->groups[0].cpu_map[0] = 0;
2265
2266	if (pcpu_setup_first_chunk(ai, fc) < 0)
2267	panic("Failed to initialize percpu areas.");
2268	}
2269
2270	#endif /* CONFIG_SMP */
2271
2272	/*
2273	* First and reserved chunks are initialized with temporary allocation
2274	* map in initdata so that they can be used before slab is online.
2275	* This function is called after slab is brought up and replaces those
2276	* with properly allocated maps.
2277	*/
2278	void __init percpu_init_late(void)
2279	{
2280	struct pcpu_chunk *target_chunks[] =
2281	{ pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2282	struct pcpu_chunk *chunk;
2283	unsigned long flags;
2284	int i;
2285
2286	for (i = 0; (chunk = target_chunks[i]); i++) {
2287	int *map;
2288	const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2289
2290	BUILD_BUG_ON(size > PAGE_SIZE);
2291
2292	map = pcpu_mem_zalloc(size);
2293	BUG_ON(!map);
2294
2295	spin_lock_irqsave(&pcpu_lock, flags);
2296	memcpy(map, chunk->map, size);
2297	chunk->map = map;
2298	spin_unlock_irqrestore(&pcpu_lock, flags);
2299	}
2300	}
2301
2302	/*
2303	* Percpu allocator is initialized early during boot when neither slab or
2304	* workqueue is available. Plug async management until everything is up
2305	* and running.
2306	*/
2307	static int __init percpu_enable_async(void)
2308	{
2309	pcpu_async_enabled = true;
2310	return 0;
2311	}
2312	subsys_initcall(percpu_enable_async);
2313