xref: /openbmc/linux/mm/percpu.c (revision f7018c21)
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 #include <linux/bitmap.h>
57 #include <linux/bootmem.h>
58 #include <linux/err.h>
59 #include <linux/list.h>
60 #include <linux/log2.h>
61 #include <linux/mm.h>
62 #include <linux/module.h>
63 #include <linux/mutex.h>
64 #include <linux/percpu.h>
65 #include <linux/pfn.h>
66 #include <linux/slab.h>
67 #include <linux/spinlock.h>
68 #include <linux/vmalloc.h>
69 #include <linux/workqueue.h>
70 #include <linux/kmemleak.h>
71 
72 #include <asm/cacheflush.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/io.h>
76 
77 #define PCPU_SLOT_BASE_SHIFT		5	/* 1-31 shares the same slot */
78 #define PCPU_DFL_MAP_ALLOC		16	/* start a map with 16 ents */
79 
80 #ifdef CONFIG_SMP
81 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82 #ifndef __addr_to_pcpu_ptr
83 #define __addr_to_pcpu_ptr(addr)					\
84 	(void __percpu *)((unsigned long)(addr) -			\
85 			  (unsigned long)pcpu_base_addr	+		\
86 			  (unsigned long)__per_cpu_start)
87 #endif
88 #ifndef __pcpu_ptr_to_addr
89 #define __pcpu_ptr_to_addr(ptr)						\
90 	(void __force *)((unsigned long)(ptr) +				\
91 			 (unsigned long)pcpu_base_addr -		\
92 			 (unsigned long)__per_cpu_start)
93 #endif
94 #else	/* CONFIG_SMP */
95 /* on UP, it's always identity mapped */
96 #define __addr_to_pcpu_ptr(addr)	(void __percpu *)(addr)
97 #define __pcpu_ptr_to_addr(ptr)		(void __force *)(ptr)
98 #endif	/* CONFIG_SMP */
99 
100 struct pcpu_chunk {
101 	struct list_head	list;		/* linked to pcpu_slot lists */
102 	int			free_size;	/* free bytes in the chunk */
103 	int			contig_hint;	/* max contiguous size hint */
104 	void			*base_addr;	/* base address of this chunk */
105 	int			map_used;	/* # of map entries used before the sentry */
106 	int			map_alloc;	/* # of map entries allocated */
107 	int			*map;		/* allocation map */
108 	void			*data;		/* chunk data */
109 	int			first_free;	/* no free below this */
110 	bool			immutable;	/* no [de]population allowed */
111 	unsigned long		populated[];	/* populated bitmap */
112 };
113 
114 static int pcpu_unit_pages __read_mostly;
115 static int pcpu_unit_size __read_mostly;
116 static int pcpu_nr_units __read_mostly;
117 static int pcpu_atom_size __read_mostly;
118 static int pcpu_nr_slots __read_mostly;
119 static size_t pcpu_chunk_struct_size __read_mostly;
120 
121 /* cpus with the lowest and highest unit addresses */
122 static unsigned int pcpu_low_unit_cpu __read_mostly;
123 static unsigned int pcpu_high_unit_cpu __read_mostly;
124 
125 /* the address of the first chunk which starts with the kernel static area */
126 void *pcpu_base_addr __read_mostly;
127 EXPORT_SYMBOL_GPL(pcpu_base_addr);
128 
129 static const int *pcpu_unit_map __read_mostly;		/* cpu -> unit */
130 const unsigned long *pcpu_unit_offsets __read_mostly;	/* cpu -> unit offset */
131 
132 /* group information, used for vm allocation */
133 static int pcpu_nr_groups __read_mostly;
134 static const unsigned long *pcpu_group_offsets __read_mostly;
135 static const size_t *pcpu_group_sizes __read_mostly;
136 
137 /*
138  * The first chunk which always exists.  Note that unlike other
139  * chunks, this one can be allocated and mapped in several different
140  * ways and thus often doesn't live in the vmalloc area.
141  */
142 static struct pcpu_chunk *pcpu_first_chunk;
143 
144 /*
145  * Optional reserved chunk.  This chunk reserves part of the first
146  * chunk and serves it for reserved allocations.  The amount of
147  * reserved offset is in pcpu_reserved_chunk_limit.  When reserved
148  * area doesn't exist, the following variables contain NULL and 0
149  * respectively.
150  */
151 static struct pcpu_chunk *pcpu_reserved_chunk;
152 static int pcpu_reserved_chunk_limit;
153 
154 /*
155  * Synchronization rules.
156  *
157  * There are two locks - pcpu_alloc_mutex and pcpu_lock.  The former
158  * protects allocation/reclaim paths, chunks, populated bitmap and
159  * vmalloc mapping.  The latter is a spinlock and protects the index
160  * data structures - chunk slots, chunks and area maps in chunks.
161  *
162  * During allocation, pcpu_alloc_mutex is kept locked all the time and
163  * pcpu_lock is grabbed and released as necessary.  All actual memory
164  * allocations are done using GFP_KERNEL with pcpu_lock released.  In
165  * general, percpu memory can't be allocated with irq off but
166  * irqsave/restore are still used in alloc path so that it can be used
167  * from early init path - sched_init() specifically.
168  *
169  * Free path accesses and alters only the index data structures, so it
170  * can be safely called from atomic context.  When memory needs to be
171  * returned to the system, free path schedules reclaim_work which
172  * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
173  * reclaimed, release both locks and frees the chunks.  Note that it's
174  * necessary to grab both locks to remove a chunk from circulation as
175  * allocation path might be referencing the chunk with only
176  * pcpu_alloc_mutex locked.
177  */
178 static DEFINE_MUTEX(pcpu_alloc_mutex);	/* protects whole alloc and reclaim */
179 static DEFINE_SPINLOCK(pcpu_lock);	/* protects index data structures */
180 
181 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182 
183 /* reclaim work to release fully free chunks, scheduled from free path */
184 static void pcpu_reclaim(struct work_struct *work);
185 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186 
187 static bool pcpu_addr_in_first_chunk(void *addr)
188 {
189 	void *first_start = pcpu_first_chunk->base_addr;
190 
191 	return addr >= first_start && addr < first_start + pcpu_unit_size;
192 }
193 
194 static bool pcpu_addr_in_reserved_chunk(void *addr)
195 {
196 	void *first_start = pcpu_first_chunk->base_addr;
197 
198 	return addr >= first_start &&
199 		addr < first_start + pcpu_reserved_chunk_limit;
200 }
201 
202 static int __pcpu_size_to_slot(int size)
203 {
204 	int highbit = fls(size);	/* size is in bytes */
205 	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
206 }
207 
208 static int pcpu_size_to_slot(int size)
209 {
210 	if (size == pcpu_unit_size)
211 		return pcpu_nr_slots - 1;
212 	return __pcpu_size_to_slot(size);
213 }
214 
215 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216 {
217 	if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
218 		return 0;
219 
220 	return pcpu_size_to_slot(chunk->free_size);
221 }
222 
223 /* set the pointer to a chunk in a page struct */
224 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225 {
226 	page->index = (unsigned long)pcpu;
227 }
228 
229 /* obtain pointer to a chunk from a page struct */
230 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231 {
232 	return (struct pcpu_chunk *)page->index;
233 }
234 
235 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
236 {
237 	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
238 }
239 
240 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
241 				     unsigned int cpu, int page_idx)
242 {
243 	return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
244 		(page_idx << PAGE_SHIFT);
245 }
246 
247 static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
248 					   int *rs, int *re, int end)
249 {
250 	*rs = find_next_zero_bit(chunk->populated, end, *rs);
251 	*re = find_next_bit(chunk->populated, end, *rs + 1);
252 }
253 
254 static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
255 					 int *rs, int *re, int end)
256 {
257 	*rs = find_next_bit(chunk->populated, end, *rs);
258 	*re = find_next_zero_bit(chunk->populated, end, *rs + 1);
259 }
260 
261 /*
262  * (Un)populated page region iterators.  Iterate over (un)populated
263  * page regions between @start and @end in @chunk.  @rs and @re should
264  * be integer variables and will be set to start and end page index of
265  * the current region.
266  */
267 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end)		    \
268 	for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 	     (rs) < (re);						    \
270 	     (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271 
272 #define pcpu_for_each_pop_region(chunk, rs, re, start, end)		    \
273 	for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end));   \
274 	     (rs) < (re);						    \
275 	     (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276 
277 /**
278  * pcpu_mem_zalloc - allocate memory
279  * @size: bytes to allocate
280  *
281  * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
282  * kzalloc() is used; otherwise, vzalloc() is used.  The returned
283  * memory is always zeroed.
284  *
285  * CONTEXT:
286  * Does GFP_KERNEL allocation.
287  *
288  * RETURNS:
289  * Pointer to the allocated area on success, NULL on failure.
290  */
291 static void *pcpu_mem_zalloc(size_t size)
292 {
293 	if (WARN_ON_ONCE(!slab_is_available()))
294 		return NULL;
295 
296 	if (size <= PAGE_SIZE)
297 		return kzalloc(size, GFP_KERNEL);
298 	else
299 		return vzalloc(size);
300 }
301 
302 /**
303  * pcpu_mem_free - free memory
304  * @ptr: memory to free
305  * @size: size of the area
306  *
307  * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
308  */
309 static void pcpu_mem_free(void *ptr, size_t size)
310 {
311 	if (size <= PAGE_SIZE)
312 		kfree(ptr);
313 	else
314 		vfree(ptr);
315 }
316 
317 /**
318  * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
319  * @chunk: chunk of interest
320  * @oslot: the previous slot it was on
321  *
322  * This function is called after an allocation or free changed @chunk.
323  * New slot according to the changed state is determined and @chunk is
324  * moved to the slot.  Note that the reserved chunk is never put on
325  * chunk slots.
326  *
327  * CONTEXT:
328  * pcpu_lock.
329  */
330 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331 {
332 	int nslot = pcpu_chunk_slot(chunk);
333 
334 	if (chunk != pcpu_reserved_chunk && oslot != nslot) {
335 		if (oslot < nslot)
336 			list_move(&chunk->list, &pcpu_slot[nslot]);
337 		else
338 			list_move_tail(&chunk->list, &pcpu_slot[nslot]);
339 	}
340 }
341 
342 /**
343  * pcpu_need_to_extend - determine whether chunk area map needs to be extended
344  * @chunk: chunk of interest
345  *
346  * Determine whether area map of @chunk needs to be extended to
347  * accommodate a new allocation.
348  *
349  * CONTEXT:
350  * pcpu_lock.
351  *
352  * RETURNS:
353  * New target map allocation length if extension is necessary, 0
354  * otherwise.
355  */
356 static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
357 {
358 	int new_alloc;
359 
360 	if (chunk->map_alloc >= chunk->map_used + 3)
361 		return 0;
362 
363 	new_alloc = PCPU_DFL_MAP_ALLOC;
364 	while (new_alloc < chunk->map_used + 3)
365 		new_alloc *= 2;
366 
367 	return new_alloc;
368 }
369 
370 /**
371  * pcpu_extend_area_map - extend area map of a chunk
372  * @chunk: chunk of interest
373  * @new_alloc: new target allocation length of the area map
374  *
375  * Extend area map of @chunk to have @new_alloc entries.
376  *
377  * CONTEXT:
378  * Does GFP_KERNEL allocation.  Grabs and releases pcpu_lock.
379  *
380  * RETURNS:
381  * 0 on success, -errno on failure.
382  */
383 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384 {
385 	int *old = NULL, *new = NULL;
386 	size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
387 	unsigned long flags;
388 
389 	new = pcpu_mem_zalloc(new_size);
390 	if (!new)
391 		return -ENOMEM;
392 
393 	/* acquire pcpu_lock and switch to new area map */
394 	spin_lock_irqsave(&pcpu_lock, flags);
395 
396 	if (new_alloc <= chunk->map_alloc)
397 		goto out_unlock;
398 
399 	old_size = chunk->map_alloc * sizeof(chunk->map[0]);
400 	old = chunk->map;
401 
402 	memcpy(new, old, old_size);
403 
404 	chunk->map_alloc = new_alloc;
405 	chunk->map = new;
406 	new = NULL;
407 
408 out_unlock:
409 	spin_unlock_irqrestore(&pcpu_lock, flags);
410 
411 	/*
412 	 * pcpu_mem_free() might end up calling vfree() which uses
413 	 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 	 */
415 	pcpu_mem_free(old, old_size);
416 	pcpu_mem_free(new, new_size);
417 
418 	return 0;
419 }
420 
421 /**
422  * pcpu_alloc_area - allocate area from a pcpu_chunk
423  * @chunk: chunk of interest
424  * @size: wanted size in bytes
425  * @align: wanted align
426  *
427  * Try to allocate @size bytes area aligned at @align from @chunk.
428  * Note that this function only allocates the offset.  It doesn't
429  * populate or map the area.
430  *
431  * @chunk->map must have at least two free slots.
432  *
433  * CONTEXT:
434  * pcpu_lock.
435  *
436  * RETURNS:
437  * Allocated offset in @chunk on success, -1 if no matching area is
438  * found.
439  */
440 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
441 {
442 	int oslot = pcpu_chunk_slot(chunk);
443 	int max_contig = 0;
444 	int i, off;
445 	bool seen_free = false;
446 	int *p;
447 
448 	for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
449 		int head, tail;
450 		int this_size;
451 
452 		off = *p;
453 		if (off & 1)
454 			continue;
455 
456 		/* extra for alignment requirement */
457 		head = ALIGN(off, align) - off;
458 
459 		this_size = (p[1] & ~1) - off;
460 		if (this_size < head + size) {
461 			if (!seen_free) {
462 				chunk->first_free = i;
463 				seen_free = true;
464 			}
465 			max_contig = max(this_size, max_contig);
466 			continue;
467 		}
468 
469 		/*
470 		 * If head is small or the previous block is free,
471 		 * merge'em.  Note that 'small' is defined as smaller
472 		 * than sizeof(int), which is very small but isn't too
473 		 * uncommon for percpu allocations.
474 		 */
475 		if (head && (head < sizeof(int) || !(p[-1] & 1))) {
476 			*p = off += head;
477 			if (p[-1] & 1)
478 				chunk->free_size -= head;
479 			else
480 				max_contig = max(*p - p[-1], max_contig);
481 			this_size -= head;
482 			head = 0;
483 		}
484 
485 		/* if tail is small, just keep it around */
486 		tail = this_size - head - size;
487 		if (tail < sizeof(int)) {
488 			tail = 0;
489 			size = this_size - head;
490 		}
491 
492 		/* split if warranted */
493 		if (head || tail) {
494 			int nr_extra = !!head + !!tail;
495 
496 			/* insert new subblocks */
497 			memmove(p + nr_extra + 1, p + 1,
498 				sizeof(chunk->map[0]) * (chunk->map_used - i));
499 			chunk->map_used += nr_extra;
500 
501 			if (head) {
502 				if (!seen_free) {
503 					chunk->first_free = i;
504 					seen_free = true;
505 				}
506 				*++p = off += head;
507 				++i;
508 				max_contig = max(head, max_contig);
509 			}
510 			if (tail) {
511 				p[1] = off + size;
512 				max_contig = max(tail, max_contig);
513 			}
514 		}
515 
516 		if (!seen_free)
517 			chunk->first_free = i + 1;
518 
519 		/* update hint and mark allocated */
520 		if (i + 1 == chunk->map_used)
521 			chunk->contig_hint = max_contig; /* fully scanned */
522 		else
523 			chunk->contig_hint = max(chunk->contig_hint,
524 						 max_contig);
525 
526 		chunk->free_size -= size;
527 		*p |= 1;
528 
529 		pcpu_chunk_relocate(chunk, oslot);
530 		return off;
531 	}
532 
533 	chunk->contig_hint = max_contig;	/* fully scanned */
534 	pcpu_chunk_relocate(chunk, oslot);
535 
536 	/* tell the upper layer that this chunk has no matching area */
537 	return -1;
538 }
539 
540 /**
541  * pcpu_free_area - free area to a pcpu_chunk
542  * @chunk: chunk of interest
543  * @freeme: offset of area to free
544  *
545  * Free area starting from @freeme to @chunk.  Note that this function
546  * only modifies the allocation map.  It doesn't depopulate or unmap
547  * the area.
548  *
549  * CONTEXT:
550  * pcpu_lock.
551  */
552 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
553 {
554 	int oslot = pcpu_chunk_slot(chunk);
555 	int off = 0;
556 	unsigned i, j;
557 	int to_free = 0;
558 	int *p;
559 
560 	freeme |= 1;	/* we are searching for <given offset, in use> pair */
561 
562 	i = 0;
563 	j = chunk->map_used;
564 	while (i != j) {
565 		unsigned k = (i + j) / 2;
566 		off = chunk->map[k];
567 		if (off < freeme)
568 			i = k + 1;
569 		else if (off > freeme)
570 			j = k;
571 		else
572 			i = j = k;
573 	}
574 	BUG_ON(off != freeme);
575 
576 	if (i < chunk->first_free)
577 		chunk->first_free = i;
578 
579 	p = chunk->map + i;
580 	*p = off &= ~1;
581 	chunk->free_size += (p[1] & ~1) - off;
582 
583 	/* merge with next? */
584 	if (!(p[1] & 1))
585 		to_free++;
586 	/* merge with previous? */
587 	if (i > 0 && !(p[-1] & 1)) {
588 		to_free++;
589 		i--;
590 		p--;
591 	}
592 	if (to_free) {
593 		chunk->map_used -= to_free;
594 		memmove(p + 1, p + 1 + to_free,
595 			(chunk->map_used - i) * sizeof(chunk->map[0]));
596 	}
597 
598 	chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
599 	pcpu_chunk_relocate(chunk, oslot);
600 }
601 
602 static struct pcpu_chunk *pcpu_alloc_chunk(void)
603 {
604 	struct pcpu_chunk *chunk;
605 
606 	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
607 	if (!chunk)
608 		return NULL;
609 
610 	chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
611 						sizeof(chunk->map[0]));
612 	if (!chunk->map) {
613 		kfree(chunk);
614 		return NULL;
615 	}
616 
617 	chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618 	chunk->map[0] = 0;
619 	chunk->map[1] = pcpu_unit_size | 1;
620 	chunk->map_used = 1;
621 
622 	INIT_LIST_HEAD(&chunk->list);
623 	chunk->free_size = pcpu_unit_size;
624 	chunk->contig_hint = pcpu_unit_size;
625 
626 	return chunk;
627 }
628 
629 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
630 {
631 	if (!chunk)
632 		return;
633 	pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
634 	pcpu_mem_free(chunk, pcpu_chunk_struct_size);
635 }
636 
637 /*
638  * Chunk management implementation.
639  *
640  * To allow different implementations, chunk alloc/free and
641  * [de]population are implemented in a separate file which is pulled
642  * into this file and compiled together.  The following functions
643  * should be implemented.
644  *
645  * pcpu_populate_chunk		- populate the specified range of a chunk
646  * pcpu_depopulate_chunk	- depopulate the specified range of a chunk
647  * pcpu_create_chunk		- create a new chunk
648  * pcpu_destroy_chunk		- destroy a chunk, always preceded by full depop
649  * pcpu_addr_to_page		- translate address to physical address
650  * pcpu_verify_alloc_info	- check alloc_info is acceptable during init
651  */
652 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654 static struct pcpu_chunk *pcpu_create_chunk(void);
655 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656 static struct page *pcpu_addr_to_page(void *addr);
657 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
658 
659 #ifdef CONFIG_NEED_PER_CPU_KM
660 #include "percpu-km.c"
661 #else
662 #include "percpu-vm.c"
663 #endif
664 
665 /**
666  * pcpu_chunk_addr_search - determine chunk containing specified address
667  * @addr: address for which the chunk needs to be determined.
668  *
669  * RETURNS:
670  * The address of the found chunk.
671  */
672 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
673 {
674 	/* is it in the first chunk? */
675 	if (pcpu_addr_in_first_chunk(addr)) {
676 		/* is it in the reserved area? */
677 		if (pcpu_addr_in_reserved_chunk(addr))
678 			return pcpu_reserved_chunk;
679 		return pcpu_first_chunk;
680 	}
681 
682 	/*
683 	 * The address is relative to unit0 which might be unused and
684 	 * thus unmapped.  Offset the address to the unit space of the
685 	 * current processor before looking it up in the vmalloc
686 	 * space.  Note that any possible cpu id can be used here, so
687 	 * there's no need to worry about preemption or cpu hotplug.
688 	 */
689 	addr += pcpu_unit_offsets[raw_smp_processor_id()];
690 	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
691 }
692 
693 /**
694  * pcpu_alloc - the percpu allocator
695  * @size: size of area to allocate in bytes
696  * @align: alignment of area (max PAGE_SIZE)
697  * @reserved: allocate from the reserved chunk if available
698  *
699  * Allocate percpu area of @size bytes aligned at @align.
700  *
701  * CONTEXT:
702  * Does GFP_KERNEL allocation.
703  *
704  * RETURNS:
705  * Percpu pointer to the allocated area on success, NULL on failure.
706  */
707 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
708 {
709 	static int warn_limit = 10;
710 	struct pcpu_chunk *chunk;
711 	const char *err;
712 	int slot, off, new_alloc;
713 	unsigned long flags;
714 	void __percpu *ptr;
715 
716 	/*
717 	 * We want the lowest bit of offset available for in-use/free
718 	 * indicator, so force >= 16bit alignment and make size even.
719 	 */
720 	if (unlikely(align < 2))
721 		align = 2;
722 
723 	if (unlikely(size & 1))
724 		size++;
725 
726 	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
727 		WARN(true, "illegal size (%zu) or align (%zu) for "
728 		     "percpu allocation\n", size, align);
729 		return NULL;
730 	}
731 
732 	mutex_lock(&pcpu_alloc_mutex);
733 	spin_lock_irqsave(&pcpu_lock, flags);
734 
735 	/* serve reserved allocations from the reserved chunk if available */
736 	if (reserved && pcpu_reserved_chunk) {
737 		chunk = pcpu_reserved_chunk;
738 
739 		if (size > chunk->contig_hint) {
740 			err = "alloc from reserved chunk failed";
741 			goto fail_unlock;
742 		}
743 
744 		while ((new_alloc = pcpu_need_to_extend(chunk))) {
745 			spin_unlock_irqrestore(&pcpu_lock, flags);
746 			if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
747 				err = "failed to extend area map of reserved chunk";
748 				goto fail_unlock_mutex;
749 			}
750 			spin_lock_irqsave(&pcpu_lock, flags);
751 		}
752 
753 		off = pcpu_alloc_area(chunk, size, align);
754 		if (off >= 0)
755 			goto area_found;
756 
757 		err = "alloc from reserved chunk failed";
758 		goto fail_unlock;
759 	}
760 
761 restart:
762 	/* search through normal chunks */
763 	for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
764 		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
765 			if (size > chunk->contig_hint)
766 				continue;
767 
768 			new_alloc = pcpu_need_to_extend(chunk);
769 			if (new_alloc) {
770 				spin_unlock_irqrestore(&pcpu_lock, flags);
771 				if (pcpu_extend_area_map(chunk,
772 							 new_alloc) < 0) {
773 					err = "failed to extend area map";
774 					goto fail_unlock_mutex;
775 				}
776 				spin_lock_irqsave(&pcpu_lock, flags);
777 				/*
778 				 * pcpu_lock has been dropped, need to
779 				 * restart cpu_slot list walking.
780 				 */
781 				goto restart;
782 			}
783 
784 			off = pcpu_alloc_area(chunk, size, align);
785 			if (off >= 0)
786 				goto area_found;
787 		}
788 	}
789 
790 	/* hmmm... no space left, create a new chunk */
791 	spin_unlock_irqrestore(&pcpu_lock, flags);
792 
793 	chunk = pcpu_create_chunk();
794 	if (!chunk) {
795 		err = "failed to allocate new chunk";
796 		goto fail_unlock_mutex;
797 	}
798 
799 	spin_lock_irqsave(&pcpu_lock, flags);
800 	pcpu_chunk_relocate(chunk, -1);
801 	goto restart;
802 
803 area_found:
804 	spin_unlock_irqrestore(&pcpu_lock, flags);
805 
806 	/* populate, map and clear the area */
807 	if (pcpu_populate_chunk(chunk, off, size)) {
808 		spin_lock_irqsave(&pcpu_lock, flags);
809 		pcpu_free_area(chunk, off);
810 		err = "failed to populate";
811 		goto fail_unlock;
812 	}
813 
814 	mutex_unlock(&pcpu_alloc_mutex);
815 
816 	/* return address relative to base address */
817 	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
818 	kmemleak_alloc_percpu(ptr, size);
819 	return ptr;
820 
821 fail_unlock:
822 	spin_unlock_irqrestore(&pcpu_lock, flags);
823 fail_unlock_mutex:
824 	mutex_unlock(&pcpu_alloc_mutex);
825 	if (warn_limit) {
826 		pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
827 			   "%s\n", size, align, err);
828 		dump_stack();
829 		if (!--warn_limit)
830 			pr_info("PERCPU: limit reached, disable warning\n");
831 	}
832 	return NULL;
833 }
834 
835 /**
836  * __alloc_percpu - allocate dynamic percpu area
837  * @size: size of area to allocate in bytes
838  * @align: alignment of area (max PAGE_SIZE)
839  *
840  * Allocate zero-filled percpu area of @size bytes aligned at @align.
841  * Might sleep.  Might trigger writeouts.
842  *
843  * CONTEXT:
844  * Does GFP_KERNEL allocation.
845  *
846  * RETURNS:
847  * Percpu pointer to the allocated area on success, NULL on failure.
848  */
849 void __percpu *__alloc_percpu(size_t size, size_t align)
850 {
851 	return pcpu_alloc(size, align, false);
852 }
853 EXPORT_SYMBOL_GPL(__alloc_percpu);
854 
855 /**
856  * __alloc_reserved_percpu - allocate reserved percpu area
857  * @size: size of area to allocate in bytes
858  * @align: alignment of area (max PAGE_SIZE)
859  *
860  * Allocate zero-filled percpu area of @size bytes aligned at @align
861  * from reserved percpu area if arch has set it up; otherwise,
862  * allocation is served from the same dynamic area.  Might sleep.
863  * Might trigger writeouts.
864  *
865  * CONTEXT:
866  * Does GFP_KERNEL allocation.
867  *
868  * RETURNS:
869  * Percpu pointer to the allocated area on success, NULL on failure.
870  */
871 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
872 {
873 	return pcpu_alloc(size, align, true);
874 }
875 
876 /**
877  * pcpu_reclaim - reclaim fully free chunks, workqueue function
878  * @work: unused
879  *
880  * Reclaim all fully free chunks except for the first one.
881  *
882  * CONTEXT:
883  * workqueue context.
884  */
885 static void pcpu_reclaim(struct work_struct *work)
886 {
887 	LIST_HEAD(todo);
888 	struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
889 	struct pcpu_chunk *chunk, *next;
890 
891 	mutex_lock(&pcpu_alloc_mutex);
892 	spin_lock_irq(&pcpu_lock);
893 
894 	list_for_each_entry_safe(chunk, next, head, list) {
895 		WARN_ON(chunk->immutable);
896 
897 		/* spare the first one */
898 		if (chunk == list_first_entry(head, struct pcpu_chunk, list))
899 			continue;
900 
901 		list_move(&chunk->list, &todo);
902 	}
903 
904 	spin_unlock_irq(&pcpu_lock);
905 
906 	list_for_each_entry_safe(chunk, next, &todo, list) {
907 		pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
908 		pcpu_destroy_chunk(chunk);
909 	}
910 
911 	mutex_unlock(&pcpu_alloc_mutex);
912 }
913 
914 /**
915  * free_percpu - free percpu area
916  * @ptr: pointer to area to free
917  *
918  * Free percpu area @ptr.
919  *
920  * CONTEXT:
921  * Can be called from atomic context.
922  */
923 void free_percpu(void __percpu *ptr)
924 {
925 	void *addr;
926 	struct pcpu_chunk *chunk;
927 	unsigned long flags;
928 	int off;
929 
930 	if (!ptr)
931 		return;
932 
933 	kmemleak_free_percpu(ptr);
934 
935 	addr = __pcpu_ptr_to_addr(ptr);
936 
937 	spin_lock_irqsave(&pcpu_lock, flags);
938 
939 	chunk = pcpu_chunk_addr_search(addr);
940 	off = addr - chunk->base_addr;
941 
942 	pcpu_free_area(chunk, off);
943 
944 	/* if there are more than one fully free chunks, wake up grim reaper */
945 	if (chunk->free_size == pcpu_unit_size) {
946 		struct pcpu_chunk *pos;
947 
948 		list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
949 			if (pos != chunk) {
950 				schedule_work(&pcpu_reclaim_work);
951 				break;
952 			}
953 	}
954 
955 	spin_unlock_irqrestore(&pcpu_lock, flags);
956 }
957 EXPORT_SYMBOL_GPL(free_percpu);
958 
959 /**
960  * is_kernel_percpu_address - test whether address is from static percpu area
961  * @addr: address to test
962  *
963  * Test whether @addr belongs to in-kernel static percpu area.  Module
964  * static percpu areas are not considered.  For those, use
965  * is_module_percpu_address().
966  *
967  * RETURNS:
968  * %true if @addr is from in-kernel static percpu area, %false otherwise.
969  */
970 bool is_kernel_percpu_address(unsigned long addr)
971 {
972 #ifdef CONFIG_SMP
973 	const size_t static_size = __per_cpu_end - __per_cpu_start;
974 	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
975 	unsigned int cpu;
976 
977 	for_each_possible_cpu(cpu) {
978 		void *start = per_cpu_ptr(base, cpu);
979 
980 		if ((void *)addr >= start && (void *)addr < start + static_size)
981 			return true;
982         }
983 #endif
984 	/* on UP, can't distinguish from other static vars, always false */
985 	return false;
986 }
987 
988 /**
989  * per_cpu_ptr_to_phys - convert translated percpu address to physical address
990  * @addr: the address to be converted to physical address
991  *
992  * Given @addr which is dereferenceable address obtained via one of
993  * percpu access macros, this function translates it into its physical
994  * address.  The caller is responsible for ensuring @addr stays valid
995  * until this function finishes.
996  *
997  * percpu allocator has special setup for the first chunk, which currently
998  * supports either embedding in linear address space or vmalloc mapping,
999  * and, from the second one, the backing allocator (currently either vm or
1000  * km) provides translation.
1001  *
1002  * The addr can be tranlated simply without checking if it falls into the
1003  * first chunk. But the current code reflects better how percpu allocator
1004  * actually works, and the verification can discover both bugs in percpu
1005  * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1006  * code.
1007  *
1008  * RETURNS:
1009  * The physical address for @addr.
1010  */
1011 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1012 {
1013 	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1014 	bool in_first_chunk = false;
1015 	unsigned long first_low, first_high;
1016 	unsigned int cpu;
1017 
1018 	/*
1019 	 * The following test on unit_low/high isn't strictly
1020 	 * necessary but will speed up lookups of addresses which
1021 	 * aren't in the first chunk.
1022 	 */
1023 	first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1024 	first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1025 				     pcpu_unit_pages);
1026 	if ((unsigned long)addr >= first_low &&
1027 	    (unsigned long)addr < first_high) {
1028 		for_each_possible_cpu(cpu) {
1029 			void *start = per_cpu_ptr(base, cpu);
1030 
1031 			if (addr >= start && addr < start + pcpu_unit_size) {
1032 				in_first_chunk = true;
1033 				break;
1034 			}
1035 		}
1036 	}
1037 
1038 	if (in_first_chunk) {
1039 		if (!is_vmalloc_addr(addr))
1040 			return __pa(addr);
1041 		else
1042 			return page_to_phys(vmalloc_to_page(addr)) +
1043 			       offset_in_page(addr);
1044 	} else
1045 		return page_to_phys(pcpu_addr_to_page(addr)) +
1046 		       offset_in_page(addr);
1047 }
1048 
1049 /**
1050  * pcpu_alloc_alloc_info - allocate percpu allocation info
1051  * @nr_groups: the number of groups
1052  * @nr_units: the number of units
1053  *
1054  * Allocate ai which is large enough for @nr_groups groups containing
1055  * @nr_units units.  The returned ai's groups[0].cpu_map points to the
1056  * cpu_map array which is long enough for @nr_units and filled with
1057  * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
1058  * pointer of other groups.
1059  *
1060  * RETURNS:
1061  * Pointer to the allocated pcpu_alloc_info on success, NULL on
1062  * failure.
1063  */
1064 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1065 						      int nr_units)
1066 {
1067 	struct pcpu_alloc_info *ai;
1068 	size_t base_size, ai_size;
1069 	void *ptr;
1070 	int unit;
1071 
1072 	base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1073 			  __alignof__(ai->groups[0].cpu_map[0]));
1074 	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1075 
1076 	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1077 	if (!ptr)
1078 		return NULL;
1079 	ai = ptr;
1080 	ptr += base_size;
1081 
1082 	ai->groups[0].cpu_map = ptr;
1083 
1084 	for (unit = 0; unit < nr_units; unit++)
1085 		ai->groups[0].cpu_map[unit] = NR_CPUS;
1086 
1087 	ai->nr_groups = nr_groups;
1088 	ai->__ai_size = PFN_ALIGN(ai_size);
1089 
1090 	return ai;
1091 }
1092 
1093 /**
1094  * pcpu_free_alloc_info - free percpu allocation info
1095  * @ai: pcpu_alloc_info to free
1096  *
1097  * Free @ai which was allocated by pcpu_alloc_alloc_info().
1098  */
1099 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1100 {
1101 	memblock_free_early(__pa(ai), ai->__ai_size);
1102 }
1103 
1104 /**
1105  * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1106  * @lvl: loglevel
1107  * @ai: allocation info to dump
1108  *
1109  * Print out information about @ai using loglevel @lvl.
1110  */
1111 static void pcpu_dump_alloc_info(const char *lvl,
1112 				 const struct pcpu_alloc_info *ai)
1113 {
1114 	int group_width = 1, cpu_width = 1, width;
1115 	char empty_str[] = "--------";
1116 	int alloc = 0, alloc_end = 0;
1117 	int group, v;
1118 	int upa, apl;	/* units per alloc, allocs per line */
1119 
1120 	v = ai->nr_groups;
1121 	while (v /= 10)
1122 		group_width++;
1123 
1124 	v = num_possible_cpus();
1125 	while (v /= 10)
1126 		cpu_width++;
1127 	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1128 
1129 	upa = ai->alloc_size / ai->unit_size;
1130 	width = upa * (cpu_width + 1) + group_width + 3;
1131 	apl = rounddown_pow_of_two(max(60 / width, 1));
1132 
1133 	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1134 	       lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1135 	       ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1136 
1137 	for (group = 0; group < ai->nr_groups; group++) {
1138 		const struct pcpu_group_info *gi = &ai->groups[group];
1139 		int unit = 0, unit_end = 0;
1140 
1141 		BUG_ON(gi->nr_units % upa);
1142 		for (alloc_end += gi->nr_units / upa;
1143 		     alloc < alloc_end; alloc++) {
1144 			if (!(alloc % apl)) {
1145 				printk(KERN_CONT "\n");
1146 				printk("%spcpu-alloc: ", lvl);
1147 			}
1148 			printk(KERN_CONT "[%0*d] ", group_width, group);
1149 
1150 			for (unit_end += upa; unit < unit_end; unit++)
1151 				if (gi->cpu_map[unit] != NR_CPUS)
1152 					printk(KERN_CONT "%0*d ", cpu_width,
1153 					       gi->cpu_map[unit]);
1154 				else
1155 					printk(KERN_CONT "%s ", empty_str);
1156 		}
1157 	}
1158 	printk(KERN_CONT "\n");
1159 }
1160 
1161 /**
1162  * pcpu_setup_first_chunk - initialize the first percpu chunk
1163  * @ai: pcpu_alloc_info describing how to percpu area is shaped
1164  * @base_addr: mapped address
1165  *
1166  * Initialize the first percpu chunk which contains the kernel static
1167  * perpcu area.  This function is to be called from arch percpu area
1168  * setup path.
1169  *
1170  * @ai contains all information necessary to initialize the first
1171  * chunk and prime the dynamic percpu allocator.
1172  *
1173  * @ai->static_size is the size of static percpu area.
1174  *
1175  * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1176  * reserve after the static area in the first chunk.  This reserves
1177  * the first chunk such that it's available only through reserved
1178  * percpu allocation.  This is primarily used to serve module percpu
1179  * static areas on architectures where the addressing model has
1180  * limited offset range for symbol relocations to guarantee module
1181  * percpu symbols fall inside the relocatable range.
1182  *
1183  * @ai->dyn_size determines the number of bytes available for dynamic
1184  * allocation in the first chunk.  The area between @ai->static_size +
1185  * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1186  *
1187  * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1188  * and equal to or larger than @ai->static_size + @ai->reserved_size +
1189  * @ai->dyn_size.
1190  *
1191  * @ai->atom_size is the allocation atom size and used as alignment
1192  * for vm areas.
1193  *
1194  * @ai->alloc_size is the allocation size and always multiple of
1195  * @ai->atom_size.  This is larger than @ai->atom_size if
1196  * @ai->unit_size is larger than @ai->atom_size.
1197  *
1198  * @ai->nr_groups and @ai->groups describe virtual memory layout of
1199  * percpu areas.  Units which should be colocated are put into the
1200  * same group.  Dynamic VM areas will be allocated according to these
1201  * groupings.  If @ai->nr_groups is zero, a single group containing
1202  * all units is assumed.
1203  *
1204  * The caller should have mapped the first chunk at @base_addr and
1205  * copied static data to each unit.
1206  *
1207  * If the first chunk ends up with both reserved and dynamic areas, it
1208  * is served by two chunks - one to serve the core static and reserved
1209  * areas and the other for the dynamic area.  They share the same vm
1210  * and page map but uses different area allocation map to stay away
1211  * from each other.  The latter chunk is circulated in the chunk slots
1212  * and available for dynamic allocation like any other chunks.
1213  *
1214  * RETURNS:
1215  * 0 on success, -errno on failure.
1216  */
1217 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1218 				  void *base_addr)
1219 {
1220 	static char cpus_buf[4096] __initdata;
1221 	static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1222 	static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1223 	size_t dyn_size = ai->dyn_size;
1224 	size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1225 	struct pcpu_chunk *schunk, *dchunk = NULL;
1226 	unsigned long *group_offsets;
1227 	size_t *group_sizes;
1228 	unsigned long *unit_off;
1229 	unsigned int cpu;
1230 	int *unit_map;
1231 	int group, unit, i;
1232 
1233 	cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1234 
1235 #define PCPU_SETUP_BUG_ON(cond)	do {					\
1236 	if (unlikely(cond)) {						\
1237 		pr_emerg("PERCPU: failed to initialize, %s", #cond);	\
1238 		pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf);	\
1239 		pcpu_dump_alloc_info(KERN_EMERG, ai);			\
1240 		BUG();							\
1241 	}								\
1242 } while (0)
1243 
1244 	/* sanity checks */
1245 	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1246 #ifdef CONFIG_SMP
1247 	PCPU_SETUP_BUG_ON(!ai->static_size);
1248 	PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
1249 #endif
1250 	PCPU_SETUP_BUG_ON(!base_addr);
1251 	PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
1252 	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1253 	PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1254 	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1255 	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1256 	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1257 
1258 	/* process group information and build config tables accordingly */
1259 	group_offsets = memblock_virt_alloc(ai->nr_groups *
1260 					     sizeof(group_offsets[0]), 0);
1261 	group_sizes = memblock_virt_alloc(ai->nr_groups *
1262 					   sizeof(group_sizes[0]), 0);
1263 	unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1264 	unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1265 
1266 	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1267 		unit_map[cpu] = UINT_MAX;
1268 
1269 	pcpu_low_unit_cpu = NR_CPUS;
1270 	pcpu_high_unit_cpu = NR_CPUS;
1271 
1272 	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1273 		const struct pcpu_group_info *gi = &ai->groups[group];
1274 
1275 		group_offsets[group] = gi->base_offset;
1276 		group_sizes[group] = gi->nr_units * ai->unit_size;
1277 
1278 		for (i = 0; i < gi->nr_units; i++) {
1279 			cpu = gi->cpu_map[i];
1280 			if (cpu == NR_CPUS)
1281 				continue;
1282 
1283 			PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1284 			PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1285 			PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1286 
1287 			unit_map[cpu] = unit + i;
1288 			unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1289 
1290 			/* determine low/high unit_cpu */
1291 			if (pcpu_low_unit_cpu == NR_CPUS ||
1292 			    unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1293 				pcpu_low_unit_cpu = cpu;
1294 			if (pcpu_high_unit_cpu == NR_CPUS ||
1295 			    unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1296 				pcpu_high_unit_cpu = cpu;
1297 		}
1298 	}
1299 	pcpu_nr_units = unit;
1300 
1301 	for_each_possible_cpu(cpu)
1302 		PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1303 
1304 	/* we're done parsing the input, undefine BUG macro and dump config */
1305 #undef PCPU_SETUP_BUG_ON
1306 	pcpu_dump_alloc_info(KERN_DEBUG, ai);
1307 
1308 	pcpu_nr_groups = ai->nr_groups;
1309 	pcpu_group_offsets = group_offsets;
1310 	pcpu_group_sizes = group_sizes;
1311 	pcpu_unit_map = unit_map;
1312 	pcpu_unit_offsets = unit_off;
1313 
1314 	/* determine basic parameters */
1315 	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1316 	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1317 	pcpu_atom_size = ai->atom_size;
1318 	pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1319 		BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1320 
1321 	/*
1322 	 * Allocate chunk slots.  The additional last slot is for
1323 	 * empty chunks.
1324 	 */
1325 	pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1326 	pcpu_slot = memblock_virt_alloc(
1327 			pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1328 	for (i = 0; i < pcpu_nr_slots; i++)
1329 		INIT_LIST_HEAD(&pcpu_slot[i]);
1330 
1331 	/*
1332 	 * Initialize static chunk.  If reserved_size is zero, the
1333 	 * static chunk covers static area + dynamic allocation area
1334 	 * in the first chunk.  If reserved_size is not zero, it
1335 	 * covers static area + reserved area (mostly used for module
1336 	 * static percpu allocation).
1337 	 */
1338 	schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1339 	INIT_LIST_HEAD(&schunk->list);
1340 	schunk->base_addr = base_addr;
1341 	schunk->map = smap;
1342 	schunk->map_alloc = ARRAY_SIZE(smap);
1343 	schunk->immutable = true;
1344 	bitmap_fill(schunk->populated, pcpu_unit_pages);
1345 
1346 	if (ai->reserved_size) {
1347 		schunk->free_size = ai->reserved_size;
1348 		pcpu_reserved_chunk = schunk;
1349 		pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1350 	} else {
1351 		schunk->free_size = dyn_size;
1352 		dyn_size = 0;			/* dynamic area covered */
1353 	}
1354 	schunk->contig_hint = schunk->free_size;
1355 
1356 	schunk->map[0] = 1;
1357 	schunk->map[1] = ai->static_size;
1358 	schunk->map_used = 1;
1359 	if (schunk->free_size)
1360 		schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1361 	else
1362 		schunk->map[1] |= 1;
1363 
1364 	/* init dynamic chunk if necessary */
1365 	if (dyn_size) {
1366 		dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1367 		INIT_LIST_HEAD(&dchunk->list);
1368 		dchunk->base_addr = base_addr;
1369 		dchunk->map = dmap;
1370 		dchunk->map_alloc = ARRAY_SIZE(dmap);
1371 		dchunk->immutable = true;
1372 		bitmap_fill(dchunk->populated, pcpu_unit_pages);
1373 
1374 		dchunk->contig_hint = dchunk->free_size = dyn_size;
1375 		dchunk->map[0] = 1;
1376 		dchunk->map[1] = pcpu_reserved_chunk_limit;
1377 		dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1378 		dchunk->map_used = 2;
1379 	}
1380 
1381 	/* link the first chunk in */
1382 	pcpu_first_chunk = dchunk ?: schunk;
1383 	pcpu_chunk_relocate(pcpu_first_chunk, -1);
1384 
1385 	/* we're done */
1386 	pcpu_base_addr = base_addr;
1387 	return 0;
1388 }
1389 
1390 #ifdef CONFIG_SMP
1391 
1392 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1393 	[PCPU_FC_AUTO]	= "auto",
1394 	[PCPU_FC_EMBED]	= "embed",
1395 	[PCPU_FC_PAGE]	= "page",
1396 };
1397 
1398 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1399 
1400 static int __init percpu_alloc_setup(char *str)
1401 {
1402 	if (!str)
1403 		return -EINVAL;
1404 
1405 	if (0)
1406 		/* nada */;
1407 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1408 	else if (!strcmp(str, "embed"))
1409 		pcpu_chosen_fc = PCPU_FC_EMBED;
1410 #endif
1411 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1412 	else if (!strcmp(str, "page"))
1413 		pcpu_chosen_fc = PCPU_FC_PAGE;
1414 #endif
1415 	else
1416 		pr_warning("PERCPU: unknown allocator %s specified\n", str);
1417 
1418 	return 0;
1419 }
1420 early_param("percpu_alloc", percpu_alloc_setup);
1421 
1422 /*
1423  * pcpu_embed_first_chunk() is used by the generic percpu setup.
1424  * Build it if needed by the arch config or the generic setup is going
1425  * to be used.
1426  */
1427 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1428 	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1429 #define BUILD_EMBED_FIRST_CHUNK
1430 #endif
1431 
1432 /* build pcpu_page_first_chunk() iff needed by the arch config */
1433 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1434 #define BUILD_PAGE_FIRST_CHUNK
1435 #endif
1436 
1437 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1438 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1439 /**
1440  * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1441  * @reserved_size: the size of reserved percpu area in bytes
1442  * @dyn_size: minimum free size for dynamic allocation in bytes
1443  * @atom_size: allocation atom size
1444  * @cpu_distance_fn: callback to determine distance between cpus, optional
1445  *
1446  * This function determines grouping of units, their mappings to cpus
1447  * and other parameters considering needed percpu size, allocation
1448  * atom size and distances between CPUs.
1449  *
1450  * Groups are always mutliples of atom size and CPUs which are of
1451  * LOCAL_DISTANCE both ways are grouped together and share space for
1452  * units in the same group.  The returned configuration is guaranteed
1453  * to have CPUs on different nodes on different groups and >=75% usage
1454  * of allocated virtual address space.
1455  *
1456  * RETURNS:
1457  * On success, pointer to the new allocation_info is returned.  On
1458  * failure, ERR_PTR value is returned.
1459  */
1460 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1461 				size_t reserved_size, size_t dyn_size,
1462 				size_t atom_size,
1463 				pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1464 {
1465 	static int group_map[NR_CPUS] __initdata;
1466 	static int group_cnt[NR_CPUS] __initdata;
1467 	const size_t static_size = __per_cpu_end - __per_cpu_start;
1468 	int nr_groups = 1, nr_units = 0;
1469 	size_t size_sum, min_unit_size, alloc_size;
1470 	int upa, max_upa, uninitialized_var(best_upa);	/* units_per_alloc */
1471 	int last_allocs, group, unit;
1472 	unsigned int cpu, tcpu;
1473 	struct pcpu_alloc_info *ai;
1474 	unsigned int *cpu_map;
1475 
1476 	/* this function may be called multiple times */
1477 	memset(group_map, 0, sizeof(group_map));
1478 	memset(group_cnt, 0, sizeof(group_cnt));
1479 
1480 	/* calculate size_sum and ensure dyn_size is enough for early alloc */
1481 	size_sum = PFN_ALIGN(static_size + reserved_size +
1482 			    max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1483 	dyn_size = size_sum - static_size - reserved_size;
1484 
1485 	/*
1486 	 * Determine min_unit_size, alloc_size and max_upa such that
1487 	 * alloc_size is multiple of atom_size and is the smallest
1488 	 * which can accommodate 4k aligned segments which are equal to
1489 	 * or larger than min_unit_size.
1490 	 */
1491 	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1492 
1493 	alloc_size = roundup(min_unit_size, atom_size);
1494 	upa = alloc_size / min_unit_size;
1495 	while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1496 		upa--;
1497 	max_upa = upa;
1498 
1499 	/* group cpus according to their proximity */
1500 	for_each_possible_cpu(cpu) {
1501 		group = 0;
1502 	next_group:
1503 		for_each_possible_cpu(tcpu) {
1504 			if (cpu == tcpu)
1505 				break;
1506 			if (group_map[tcpu] == group && cpu_distance_fn &&
1507 			    (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1508 			     cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1509 				group++;
1510 				nr_groups = max(nr_groups, group + 1);
1511 				goto next_group;
1512 			}
1513 		}
1514 		group_map[cpu] = group;
1515 		group_cnt[group]++;
1516 	}
1517 
1518 	/*
1519 	 * Expand unit size until address space usage goes over 75%
1520 	 * and then as much as possible without using more address
1521 	 * space.
1522 	 */
1523 	last_allocs = INT_MAX;
1524 	for (upa = max_upa; upa; upa--) {
1525 		int allocs = 0, wasted = 0;
1526 
1527 		if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1528 			continue;
1529 
1530 		for (group = 0; group < nr_groups; group++) {
1531 			int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1532 			allocs += this_allocs;
1533 			wasted += this_allocs * upa - group_cnt[group];
1534 		}
1535 
1536 		/*
1537 		 * Don't accept if wastage is over 1/3.  The
1538 		 * greater-than comparison ensures upa==1 always
1539 		 * passes the following check.
1540 		 */
1541 		if (wasted > num_possible_cpus() / 3)
1542 			continue;
1543 
1544 		/* and then don't consume more memory */
1545 		if (allocs > last_allocs)
1546 			break;
1547 		last_allocs = allocs;
1548 		best_upa = upa;
1549 	}
1550 	upa = best_upa;
1551 
1552 	/* allocate and fill alloc_info */
1553 	for (group = 0; group < nr_groups; group++)
1554 		nr_units += roundup(group_cnt[group], upa);
1555 
1556 	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1557 	if (!ai)
1558 		return ERR_PTR(-ENOMEM);
1559 	cpu_map = ai->groups[0].cpu_map;
1560 
1561 	for (group = 0; group < nr_groups; group++) {
1562 		ai->groups[group].cpu_map = cpu_map;
1563 		cpu_map += roundup(group_cnt[group], upa);
1564 	}
1565 
1566 	ai->static_size = static_size;
1567 	ai->reserved_size = reserved_size;
1568 	ai->dyn_size = dyn_size;
1569 	ai->unit_size = alloc_size / upa;
1570 	ai->atom_size = atom_size;
1571 	ai->alloc_size = alloc_size;
1572 
1573 	for (group = 0, unit = 0; group_cnt[group]; group++) {
1574 		struct pcpu_group_info *gi = &ai->groups[group];
1575 
1576 		/*
1577 		 * Initialize base_offset as if all groups are located
1578 		 * back-to-back.  The caller should update this to
1579 		 * reflect actual allocation.
1580 		 */
1581 		gi->base_offset = unit * ai->unit_size;
1582 
1583 		for_each_possible_cpu(cpu)
1584 			if (group_map[cpu] == group)
1585 				gi->cpu_map[gi->nr_units++] = cpu;
1586 		gi->nr_units = roundup(gi->nr_units, upa);
1587 		unit += gi->nr_units;
1588 	}
1589 	BUG_ON(unit != nr_units);
1590 
1591 	return ai;
1592 }
1593 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1594 
1595 #if defined(BUILD_EMBED_FIRST_CHUNK)
1596 /**
1597  * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1598  * @reserved_size: the size of reserved percpu area in bytes
1599  * @dyn_size: minimum free size for dynamic allocation in bytes
1600  * @atom_size: allocation atom size
1601  * @cpu_distance_fn: callback to determine distance between cpus, optional
1602  * @alloc_fn: function to allocate percpu page
1603  * @free_fn: function to free percpu page
1604  *
1605  * This is a helper to ease setting up embedded first percpu chunk and
1606  * can be called where pcpu_setup_first_chunk() is expected.
1607  *
1608  * If this function is used to setup the first chunk, it is allocated
1609  * by calling @alloc_fn and used as-is without being mapped into
1610  * vmalloc area.  Allocations are always whole multiples of @atom_size
1611  * aligned to @atom_size.
1612  *
1613  * This enables the first chunk to piggy back on the linear physical
1614  * mapping which often uses larger page size.  Please note that this
1615  * can result in very sparse cpu->unit mapping on NUMA machines thus
1616  * requiring large vmalloc address space.  Don't use this allocator if
1617  * vmalloc space is not orders of magnitude larger than distances
1618  * between node memory addresses (ie. 32bit NUMA machines).
1619  *
1620  * @dyn_size specifies the minimum dynamic area size.
1621  *
1622  * If the needed size is smaller than the minimum or specified unit
1623  * size, the leftover is returned using @free_fn.
1624  *
1625  * RETURNS:
1626  * 0 on success, -errno on failure.
1627  */
1628 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1629 				  size_t atom_size,
1630 				  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1631 				  pcpu_fc_alloc_fn_t alloc_fn,
1632 				  pcpu_fc_free_fn_t free_fn)
1633 {
1634 	void *base = (void *)ULONG_MAX;
1635 	void **areas = NULL;
1636 	struct pcpu_alloc_info *ai;
1637 	size_t size_sum, areas_size, max_distance;
1638 	int group, i, rc;
1639 
1640 	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1641 				   cpu_distance_fn);
1642 	if (IS_ERR(ai))
1643 		return PTR_ERR(ai);
1644 
1645 	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1646 	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1647 
1648 	areas = memblock_virt_alloc_nopanic(areas_size, 0);
1649 	if (!areas) {
1650 		rc = -ENOMEM;
1651 		goto out_free;
1652 	}
1653 
1654 	/* allocate, copy and determine base address */
1655 	for (group = 0; group < ai->nr_groups; group++) {
1656 		struct pcpu_group_info *gi = &ai->groups[group];
1657 		unsigned int cpu = NR_CPUS;
1658 		void *ptr;
1659 
1660 		for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1661 			cpu = gi->cpu_map[i];
1662 		BUG_ON(cpu == NR_CPUS);
1663 
1664 		/* allocate space for the whole group */
1665 		ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1666 		if (!ptr) {
1667 			rc = -ENOMEM;
1668 			goto out_free_areas;
1669 		}
1670 		/* kmemleak tracks the percpu allocations separately */
1671 		kmemleak_free(ptr);
1672 		areas[group] = ptr;
1673 
1674 		base = min(ptr, base);
1675 	}
1676 
1677 	/*
1678 	 * Copy data and free unused parts.  This should happen after all
1679 	 * allocations are complete; otherwise, we may end up with
1680 	 * overlapping groups.
1681 	 */
1682 	for (group = 0; group < ai->nr_groups; group++) {
1683 		struct pcpu_group_info *gi = &ai->groups[group];
1684 		void *ptr = areas[group];
1685 
1686 		for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1687 			if (gi->cpu_map[i] == NR_CPUS) {
1688 				/* unused unit, free whole */
1689 				free_fn(ptr, ai->unit_size);
1690 				continue;
1691 			}
1692 			/* copy and return the unused part */
1693 			memcpy(ptr, __per_cpu_load, ai->static_size);
1694 			free_fn(ptr + size_sum, ai->unit_size - size_sum);
1695 		}
1696 	}
1697 
1698 	/* base address is now known, determine group base offsets */
1699 	max_distance = 0;
1700 	for (group = 0; group < ai->nr_groups; group++) {
1701 		ai->groups[group].base_offset = areas[group] - base;
1702 		max_distance = max_t(size_t, max_distance,
1703 				     ai->groups[group].base_offset);
1704 	}
1705 	max_distance += ai->unit_size;
1706 
1707 	/* warn if maximum distance is further than 75% of vmalloc space */
1708 	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
1709 		pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1710 			   "space 0x%lx\n", max_distance,
1711 			   VMALLOC_TOTAL);
1712 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1713 		/* and fail if we have fallback */
1714 		rc = -EINVAL;
1715 		goto out_free;
1716 #endif
1717 	}
1718 
1719 	pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1720 		PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1721 		ai->dyn_size, ai->unit_size);
1722 
1723 	rc = pcpu_setup_first_chunk(ai, base);
1724 	goto out_free;
1725 
1726 out_free_areas:
1727 	for (group = 0; group < ai->nr_groups; group++)
1728 		if (areas[group])
1729 			free_fn(areas[group],
1730 				ai->groups[group].nr_units * ai->unit_size);
1731 out_free:
1732 	pcpu_free_alloc_info(ai);
1733 	if (areas)
1734 		memblock_free_early(__pa(areas), areas_size);
1735 	return rc;
1736 }
1737 #endif /* BUILD_EMBED_FIRST_CHUNK */
1738 
1739 #ifdef BUILD_PAGE_FIRST_CHUNK
1740 /**
1741  * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1742  * @reserved_size: the size of reserved percpu area in bytes
1743  * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1744  * @free_fn: function to free percpu page, always called with PAGE_SIZE
1745  * @populate_pte_fn: function to populate pte
1746  *
1747  * This is a helper to ease setting up page-remapped first percpu
1748  * chunk and can be called where pcpu_setup_first_chunk() is expected.
1749  *
1750  * This is the basic allocator.  Static percpu area is allocated
1751  * page-by-page into vmalloc area.
1752  *
1753  * RETURNS:
1754  * 0 on success, -errno on failure.
1755  */
1756 int __init pcpu_page_first_chunk(size_t reserved_size,
1757 				 pcpu_fc_alloc_fn_t alloc_fn,
1758 				 pcpu_fc_free_fn_t free_fn,
1759 				 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1760 {
1761 	static struct vm_struct vm;
1762 	struct pcpu_alloc_info *ai;
1763 	char psize_str[16];
1764 	int unit_pages;
1765 	size_t pages_size;
1766 	struct page **pages;
1767 	int unit, i, j, rc;
1768 
1769 	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1770 
1771 	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
1772 	if (IS_ERR(ai))
1773 		return PTR_ERR(ai);
1774 	BUG_ON(ai->nr_groups != 1);
1775 	BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1776 
1777 	unit_pages = ai->unit_size >> PAGE_SHIFT;
1778 
1779 	/* unaligned allocations can't be freed, round up to page size */
1780 	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1781 			       sizeof(pages[0]));
1782 	pages = memblock_virt_alloc(pages_size, 0);
1783 
1784 	/* allocate pages */
1785 	j = 0;
1786 	for (unit = 0; unit < num_possible_cpus(); unit++)
1787 		for (i = 0; i < unit_pages; i++) {
1788 			unsigned int cpu = ai->groups[0].cpu_map[unit];
1789 			void *ptr;
1790 
1791 			ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1792 			if (!ptr) {
1793 				pr_warning("PERCPU: failed to allocate %s page "
1794 					   "for cpu%u\n", psize_str, cpu);
1795 				goto enomem;
1796 			}
1797 			/* kmemleak tracks the percpu allocations separately */
1798 			kmemleak_free(ptr);
1799 			pages[j++] = virt_to_page(ptr);
1800 		}
1801 
1802 	/* allocate vm area, map the pages and copy static data */
1803 	vm.flags = VM_ALLOC;
1804 	vm.size = num_possible_cpus() * ai->unit_size;
1805 	vm_area_register_early(&vm, PAGE_SIZE);
1806 
1807 	for (unit = 0; unit < num_possible_cpus(); unit++) {
1808 		unsigned long unit_addr =
1809 			(unsigned long)vm.addr + unit * ai->unit_size;
1810 
1811 		for (i = 0; i < unit_pages; i++)
1812 			populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1813 
1814 		/* pte already populated, the following shouldn't fail */
1815 		rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1816 				      unit_pages);
1817 		if (rc < 0)
1818 			panic("failed to map percpu area, err=%d\n", rc);
1819 
1820 		/*
1821 		 * FIXME: Archs with virtual cache should flush local
1822 		 * cache for the linear mapping here - something
1823 		 * equivalent to flush_cache_vmap() on the local cpu.
1824 		 * flush_cache_vmap() can't be used as most supporting
1825 		 * data structures are not set up yet.
1826 		 */
1827 
1828 		/* copy static data */
1829 		memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1830 	}
1831 
1832 	/* we're ready, commit */
1833 	pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1834 		unit_pages, psize_str, vm.addr, ai->static_size,
1835 		ai->reserved_size, ai->dyn_size);
1836 
1837 	rc = pcpu_setup_first_chunk(ai, vm.addr);
1838 	goto out_free_ar;
1839 
1840 enomem:
1841 	while (--j >= 0)
1842 		free_fn(page_address(pages[j]), PAGE_SIZE);
1843 	rc = -ENOMEM;
1844 out_free_ar:
1845 	memblock_free_early(__pa(pages), pages_size);
1846 	pcpu_free_alloc_info(ai);
1847 	return rc;
1848 }
1849 #endif /* BUILD_PAGE_FIRST_CHUNK */
1850 
1851 #ifndef	CONFIG_HAVE_SETUP_PER_CPU_AREA
1852 /*
1853  * Generic SMP percpu area setup.
1854  *
1855  * The embedding helper is used because its behavior closely resembles
1856  * the original non-dynamic generic percpu area setup.  This is
1857  * important because many archs have addressing restrictions and might
1858  * fail if the percpu area is located far away from the previous
1859  * location.  As an added bonus, in non-NUMA cases, embedding is
1860  * generally a good idea TLB-wise because percpu area can piggy back
1861  * on the physical linear memory mapping which uses large page
1862  * mappings on applicable archs.
1863  */
1864 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1865 EXPORT_SYMBOL(__per_cpu_offset);
1866 
1867 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1868 				       size_t align)
1869 {
1870 	return  memblock_virt_alloc_from_nopanic(
1871 			size, align, __pa(MAX_DMA_ADDRESS));
1872 }
1873 
1874 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1875 {
1876 	memblock_free_early(__pa(ptr), size);
1877 }
1878 
1879 void __init setup_per_cpu_areas(void)
1880 {
1881 	unsigned long delta;
1882 	unsigned int cpu;
1883 	int rc;
1884 
1885 	/*
1886 	 * Always reserve area for module percpu variables.  That's
1887 	 * what the legacy allocator did.
1888 	 */
1889 	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1890 				    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1891 				    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1892 	if (rc < 0)
1893 		panic("Failed to initialize percpu areas.");
1894 
1895 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1896 	for_each_possible_cpu(cpu)
1897 		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1898 }
1899 #endif	/* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1900 
1901 #else	/* CONFIG_SMP */
1902 
1903 /*
1904  * UP percpu area setup.
1905  *
1906  * UP always uses km-based percpu allocator with identity mapping.
1907  * Static percpu variables are indistinguishable from the usual static
1908  * variables and don't require any special preparation.
1909  */
1910 void __init setup_per_cpu_areas(void)
1911 {
1912 	const size_t unit_size =
1913 		roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1914 					 PERCPU_DYNAMIC_RESERVE));
1915 	struct pcpu_alloc_info *ai;
1916 	void *fc;
1917 
1918 	ai = pcpu_alloc_alloc_info(1, 1);
1919 	fc = memblock_virt_alloc_from_nopanic(unit_size,
1920 					      PAGE_SIZE,
1921 					      __pa(MAX_DMA_ADDRESS));
1922 	if (!ai || !fc)
1923 		panic("Failed to allocate memory for percpu areas.");
1924 	/* kmemleak tracks the percpu allocations separately */
1925 	kmemleak_free(fc);
1926 
1927 	ai->dyn_size = unit_size;
1928 	ai->unit_size = unit_size;
1929 	ai->atom_size = unit_size;
1930 	ai->alloc_size = unit_size;
1931 	ai->groups[0].nr_units = 1;
1932 	ai->groups[0].cpu_map[0] = 0;
1933 
1934 	if (pcpu_setup_first_chunk(ai, fc) < 0)
1935 		panic("Failed to initialize percpu areas.");
1936 }
1937 
1938 #endif	/* CONFIG_SMP */
1939 
1940 /*
1941  * First and reserved chunks are initialized with temporary allocation
1942  * map in initdata so that they can be used before slab is online.
1943  * This function is called after slab is brought up and replaces those
1944  * with properly allocated maps.
1945  */
1946 void __init percpu_init_late(void)
1947 {
1948 	struct pcpu_chunk *target_chunks[] =
1949 		{ pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1950 	struct pcpu_chunk *chunk;
1951 	unsigned long flags;
1952 	int i;
1953 
1954 	for (i = 0; (chunk = target_chunks[i]); i++) {
1955 		int *map;
1956 		const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1957 
1958 		BUILD_BUG_ON(size > PAGE_SIZE);
1959 
1960 		map = pcpu_mem_zalloc(size);
1961 		BUG_ON(!map);
1962 
1963 		spin_lock_irqsave(&pcpu_lock, flags);
1964 		memcpy(map, chunk->map, size);
1965 		chunk->map = map;
1966 		spin_unlock_irqrestore(&pcpu_lock, flags);
1967 	}
1968 }
1969