Loading...
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
100struct 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
114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
116static int pcpu_nr_units __read_mostly;
117static int pcpu_atom_size __read_mostly;
118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
120
121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
124
125/* the address of the first chunk which starts with the kernel static area */
126void *pcpu_base_addr __read_mostly;
127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
131
132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static 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 */
142static 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 */
151static struct pcpu_chunk *pcpu_reserved_chunk;
152static 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 */
178static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
179static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
180
181static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182
183/* reclaim work to release fully free chunks, scheduled from free path */
184static void pcpu_reclaim(struct work_struct *work);
185static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186
187static 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
194static 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
202static 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
208static 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
215static 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 */
224static 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 */
230static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231{
232 return (struct pcpu_chunk *)page->index;
233}
234
235static 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
240static 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
247static 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
254static 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 */
291static 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 */
309static 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 */
330static 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 */
356static 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 */
383static 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
408out_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 */
440static 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 */
552static 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
602static 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 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
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
629static 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 */
652static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654static struct pcpu_chunk *pcpu_create_chunk(void);
655static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656static struct page *pcpu_addr_to_page(void *addr);
657static 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 */
672static 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 */
707static 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
761restart:
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
803area_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
821fail_unlock:
822 spin_unlock_irqrestore(&pcpu_lock, flags);
823fail_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 */
849void __percpu *__alloc_percpu(size_t size, size_t align)
850{
851 return pcpu_alloc(size, align, false);
852}
853EXPORT_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 */
871void __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 */
885static 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 */
923void 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}
957EXPORT_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 */
970bool 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 */
1011phys_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 */
1064struct 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 */
1099void __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 */
1111static 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 */
1217int __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
1392const 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
1398enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1399
1400static 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}
1420early_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 */
1460static 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 */
1628int __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
1726out_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);
1731out_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 */
1756int __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
1840enomem:
1841 while (--j >= 0)
1842 free_fn(page_address(pages[j]), PAGE_SIZE);
1843 rc = -ENOMEM;
1844out_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 */
1864unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1865EXPORT_SYMBOL(__per_cpu_offset);
1866
1867static 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
1874static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1875{
1876 memblock_free_early(__pa(ptr), size);
1877}
1878
1879void __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 */
1910void __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 */
1946void __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}
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
74#include <asm/cacheflush.h>
75#include <asm/sections.h>
76#include <asm/tlbflush.h>
77#include <asm/io.h>
78
79#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
80#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
81#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
82#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
83#define PCPU_EMPTY_POP_PAGES_LOW 2
84#define PCPU_EMPTY_POP_PAGES_HIGH 4
85
86#ifdef CONFIG_SMP
87/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
88#ifndef __addr_to_pcpu_ptr
89#define __addr_to_pcpu_ptr(addr) \
90 (void __percpu *)((unsigned long)(addr) - \
91 (unsigned long)pcpu_base_addr + \
92 (unsigned long)__per_cpu_start)
93#endif
94#ifndef __pcpu_ptr_to_addr
95#define __pcpu_ptr_to_addr(ptr) \
96 (void __force *)((unsigned long)(ptr) + \
97 (unsigned long)pcpu_base_addr - \
98 (unsigned long)__per_cpu_start)
99#endif
100#else /* CONFIG_SMP */
101/* on UP, it's always identity mapped */
102#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
103#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
104#endif /* CONFIG_SMP */
105
106struct pcpu_chunk {
107 struct list_head list; /* linked to pcpu_slot lists */
108 int free_size; /* free bytes in the chunk */
109 int contig_hint; /* max contiguous size hint */
110 void *base_addr; /* base address of this chunk */
111
112 int map_used; /* # of map entries used before the sentry */
113 int map_alloc; /* # of map entries allocated */
114 int *map; /* allocation map */
115 struct list_head map_extend_list;/* on pcpu_map_extend_chunks */
116
117 void *data; /* chunk data */
118 int first_free; /* no free below this */
119 bool immutable; /* no [de]population allowed */
120 int nr_populated; /* # of populated pages */
121 unsigned long populated[]; /* populated bitmap */
122};
123
124static int pcpu_unit_pages __read_mostly;
125static int pcpu_unit_size __read_mostly;
126static int pcpu_nr_units __read_mostly;
127static int pcpu_atom_size __read_mostly;
128static int pcpu_nr_slots __read_mostly;
129static size_t pcpu_chunk_struct_size __read_mostly;
130
131/* cpus with the lowest and highest unit addresses */
132static unsigned int pcpu_low_unit_cpu __read_mostly;
133static unsigned int pcpu_high_unit_cpu __read_mostly;
134
135/* the address of the first chunk which starts with the kernel static area */
136void *pcpu_base_addr __read_mostly;
137EXPORT_SYMBOL_GPL(pcpu_base_addr);
138
139static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
140const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
141
142/* group information, used for vm allocation */
143static int pcpu_nr_groups __read_mostly;
144static const unsigned long *pcpu_group_offsets __read_mostly;
145static const size_t *pcpu_group_sizes __read_mostly;
146
147/*
148 * The first chunk which always exists. Note that unlike other
149 * chunks, this one can be allocated and mapped in several different
150 * ways and thus often doesn't live in the vmalloc area.
151 */
152static struct pcpu_chunk *pcpu_first_chunk;
153
154/*
155 * Optional reserved chunk. This chunk reserves part of the first
156 * chunk and serves it for reserved allocations. The amount of
157 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
158 * area doesn't exist, the following variables contain NULL and 0
159 * respectively.
160 */
161static struct pcpu_chunk *pcpu_reserved_chunk;
162static int pcpu_reserved_chunk_limit;
163
164static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
165static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
166
167static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
168
169/* chunks which need their map areas extended, protected by pcpu_lock */
170static LIST_HEAD(pcpu_map_extend_chunks);
171
172/*
173 * The number of empty populated pages, protected by pcpu_lock. The
174 * reserved chunk doesn't contribute to the count.
175 */
176static int pcpu_nr_empty_pop_pages;
177
178/*
179 * Balance work is used to populate or destroy chunks asynchronously. We
180 * try to keep the number of populated free pages between
181 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
182 * empty chunk.
183 */
184static void pcpu_balance_workfn(struct work_struct *work);
185static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
186static bool pcpu_async_enabled __read_mostly;
187static bool pcpu_atomic_alloc_failed;
188
189static void pcpu_schedule_balance_work(void)
190{
191 if (pcpu_async_enabled)
192 schedule_work(&pcpu_balance_work);
193}
194
195static bool pcpu_addr_in_first_chunk(void *addr)
196{
197 void *first_start = pcpu_first_chunk->base_addr;
198
199 return addr >= first_start && addr < first_start + pcpu_unit_size;
200}
201
202static bool pcpu_addr_in_reserved_chunk(void *addr)
203{
204 void *first_start = pcpu_first_chunk->base_addr;
205
206 return addr >= first_start &&
207 addr < first_start + pcpu_reserved_chunk_limit;
208}
209
210static int __pcpu_size_to_slot(int size)
211{
212 int highbit = fls(size); /* size is in bytes */
213 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
214}
215
216static int pcpu_size_to_slot(int size)
217{
218 if (size == pcpu_unit_size)
219 return pcpu_nr_slots - 1;
220 return __pcpu_size_to_slot(size);
221}
222
223static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
224{
225 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
226 return 0;
227
228 return pcpu_size_to_slot(chunk->free_size);
229}
230
231/* set the pointer to a chunk in a page struct */
232static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
233{
234 page->index = (unsigned long)pcpu;
235}
236
237/* obtain pointer to a chunk from a page struct */
238static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
239{
240 return (struct pcpu_chunk *)page->index;
241}
242
243static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
244{
245 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
246}
247
248static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
249 unsigned int cpu, int page_idx)
250{
251 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
252 (page_idx << PAGE_SHIFT);
253}
254
255static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
256 int *rs, int *re, int end)
257{
258 *rs = find_next_zero_bit(chunk->populated, end, *rs);
259 *re = find_next_bit(chunk->populated, end, *rs + 1);
260}
261
262static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
263 int *rs, int *re, int end)
264{
265 *rs = find_next_bit(chunk->populated, end, *rs);
266 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
267}
268
269/*
270 * (Un)populated page region iterators. Iterate over (un)populated
271 * page regions between @start and @end in @chunk. @rs and @re should
272 * be integer variables and will be set to start and end page index of
273 * the current region.
274 */
275#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
276 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
277 (rs) < (re); \
278 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
279
280#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
281 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
282 (rs) < (re); \
283 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
284
285/**
286 * pcpu_mem_zalloc - allocate memory
287 * @size: bytes to allocate
288 *
289 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
290 * kzalloc() is used; otherwise, vzalloc() is used. The returned
291 * memory is always zeroed.
292 *
293 * CONTEXT:
294 * Does GFP_KERNEL allocation.
295 *
296 * RETURNS:
297 * Pointer to the allocated area on success, NULL on failure.
298 */
299static void *pcpu_mem_zalloc(size_t size)
300{
301 if (WARN_ON_ONCE(!slab_is_available()))
302 return NULL;
303
304 if (size <= PAGE_SIZE)
305 return kzalloc(size, GFP_KERNEL);
306 else
307 return vzalloc(size);
308}
309
310/**
311 * pcpu_mem_free - free memory
312 * @ptr: memory to free
313 *
314 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
315 */
316static void pcpu_mem_free(void *ptr)
317{
318 kvfree(ptr);
319}
320
321/**
322 * pcpu_count_occupied_pages - count the number of pages an area occupies
323 * @chunk: chunk of interest
324 * @i: index of the area in question
325 *
326 * Count the number of pages chunk's @i'th area occupies. When the area's
327 * start and/or end address isn't aligned to page boundary, the straddled
328 * page is included in the count iff the rest of the page is free.
329 */
330static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
331{
332 int off = chunk->map[i] & ~1;
333 int end = chunk->map[i + 1] & ~1;
334
335 if (!PAGE_ALIGNED(off) && i > 0) {
336 int prev = chunk->map[i - 1];
337
338 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
339 off = round_down(off, PAGE_SIZE);
340 }
341
342 if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
343 int next = chunk->map[i + 1];
344 int nend = chunk->map[i + 2] & ~1;
345
346 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
347 end = round_up(end, PAGE_SIZE);
348 }
349
350 return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
351}
352
353/**
354 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
355 * @chunk: chunk of interest
356 * @oslot: the previous slot it was on
357 *
358 * This function is called after an allocation or free changed @chunk.
359 * New slot according to the changed state is determined and @chunk is
360 * moved to the slot. Note that the reserved chunk is never put on
361 * chunk slots.
362 *
363 * CONTEXT:
364 * pcpu_lock.
365 */
366static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
367{
368 int nslot = pcpu_chunk_slot(chunk);
369
370 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
371 if (oslot < nslot)
372 list_move(&chunk->list, &pcpu_slot[nslot]);
373 else
374 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
375 }
376}
377
378/**
379 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
380 * @chunk: chunk of interest
381 * @is_atomic: the allocation context
382 *
383 * Determine whether area map of @chunk needs to be extended. If
384 * @is_atomic, only the amount necessary for a new allocation is
385 * considered; however, async extension is scheduled if the left amount is
386 * low. If !@is_atomic, it aims for more empty space. Combined, this
387 * ensures that the map is likely to have enough available space to
388 * accomodate atomic allocations which can't extend maps directly.
389 *
390 * CONTEXT:
391 * pcpu_lock.
392 *
393 * RETURNS:
394 * New target map allocation length if extension is necessary, 0
395 * otherwise.
396 */
397static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
398{
399 int margin, new_alloc;
400
401 lockdep_assert_held(&pcpu_lock);
402
403 if (is_atomic) {
404 margin = 3;
405
406 if (chunk->map_alloc <
407 chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) {
408 if (list_empty(&chunk->map_extend_list)) {
409 list_add_tail(&chunk->map_extend_list,
410 &pcpu_map_extend_chunks);
411 pcpu_schedule_balance_work();
412 }
413 }
414 } else {
415 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
416 }
417
418 if (chunk->map_alloc >= chunk->map_used + margin)
419 return 0;
420
421 new_alloc = PCPU_DFL_MAP_ALLOC;
422 while (new_alloc < chunk->map_used + margin)
423 new_alloc *= 2;
424
425 return new_alloc;
426}
427
428/**
429 * pcpu_extend_area_map - extend area map of a chunk
430 * @chunk: chunk of interest
431 * @new_alloc: new target allocation length of the area map
432 *
433 * Extend area map of @chunk to have @new_alloc entries.
434 *
435 * CONTEXT:
436 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
437 *
438 * RETURNS:
439 * 0 on success, -errno on failure.
440 */
441static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
442{
443 int *old = NULL, *new = NULL;
444 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
445 unsigned long flags;
446
447 lockdep_assert_held(&pcpu_alloc_mutex);
448
449 new = pcpu_mem_zalloc(new_size);
450 if (!new)
451 return -ENOMEM;
452
453 /* acquire pcpu_lock and switch to new area map */
454 spin_lock_irqsave(&pcpu_lock, flags);
455
456 if (new_alloc <= chunk->map_alloc)
457 goto out_unlock;
458
459 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
460 old = chunk->map;
461
462 memcpy(new, old, old_size);
463
464 chunk->map_alloc = new_alloc;
465 chunk->map = new;
466 new = NULL;
467
468out_unlock:
469 spin_unlock_irqrestore(&pcpu_lock, flags);
470
471 /*
472 * pcpu_mem_free() might end up calling vfree() which uses
473 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
474 */
475 pcpu_mem_free(old);
476 pcpu_mem_free(new);
477
478 return 0;
479}
480
481/**
482 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
483 * @chunk: chunk the candidate area belongs to
484 * @off: the offset to the start of the candidate area
485 * @this_size: the size of the candidate area
486 * @size: the size of the target allocation
487 * @align: the alignment of the target allocation
488 * @pop_only: only allocate from already populated region
489 *
490 * We're trying to allocate @size bytes aligned at @align. @chunk's area
491 * at @off sized @this_size is a candidate. This function determines
492 * whether the target allocation fits in the candidate area and returns the
493 * number of bytes to pad after @off. If the target area doesn't fit, -1
494 * is returned.
495 *
496 * If @pop_only is %true, this function only considers the already
497 * populated part of the candidate area.
498 */
499static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
500 int size, int align, bool pop_only)
501{
502 int cand_off = off;
503
504 while (true) {
505 int head = ALIGN(cand_off, align) - off;
506 int page_start, page_end, rs, re;
507
508 if (this_size < head + size)
509 return -1;
510
511 if (!pop_only)
512 return head;
513
514 /*
515 * If the first unpopulated page is beyond the end of the
516 * allocation, the whole allocation is populated;
517 * otherwise, retry from the end of the unpopulated area.
518 */
519 page_start = PFN_DOWN(head + off);
520 page_end = PFN_UP(head + off + size);
521
522 rs = page_start;
523 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
524 if (rs >= page_end)
525 return head;
526 cand_off = re * PAGE_SIZE;
527 }
528}
529
530/**
531 * pcpu_alloc_area - allocate area from a pcpu_chunk
532 * @chunk: chunk of interest
533 * @size: wanted size in bytes
534 * @align: wanted align
535 * @pop_only: allocate only from the populated area
536 * @occ_pages_p: out param for the number of pages the area occupies
537 *
538 * Try to allocate @size bytes area aligned at @align from @chunk.
539 * Note that this function only allocates the offset. It doesn't
540 * populate or map the area.
541 *
542 * @chunk->map must have at least two free slots.
543 *
544 * CONTEXT:
545 * pcpu_lock.
546 *
547 * RETURNS:
548 * Allocated offset in @chunk on success, -1 if no matching area is
549 * found.
550 */
551static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
552 bool pop_only, int *occ_pages_p)
553{
554 int oslot = pcpu_chunk_slot(chunk);
555 int max_contig = 0;
556 int i, off;
557 bool seen_free = false;
558 int *p;
559
560 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
561 int head, tail;
562 int this_size;
563
564 off = *p;
565 if (off & 1)
566 continue;
567
568 this_size = (p[1] & ~1) - off;
569
570 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
571 pop_only);
572 if (head < 0) {
573 if (!seen_free) {
574 chunk->first_free = i;
575 seen_free = true;
576 }
577 max_contig = max(this_size, max_contig);
578 continue;
579 }
580
581 /*
582 * If head is small or the previous block is free,
583 * merge'em. Note that 'small' is defined as smaller
584 * than sizeof(int), which is very small but isn't too
585 * uncommon for percpu allocations.
586 */
587 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
588 *p = off += head;
589 if (p[-1] & 1)
590 chunk->free_size -= head;
591 else
592 max_contig = max(*p - p[-1], max_contig);
593 this_size -= head;
594 head = 0;
595 }
596
597 /* if tail is small, just keep it around */
598 tail = this_size - head - size;
599 if (tail < sizeof(int)) {
600 tail = 0;
601 size = this_size - head;
602 }
603
604 /* split if warranted */
605 if (head || tail) {
606 int nr_extra = !!head + !!tail;
607
608 /* insert new subblocks */
609 memmove(p + nr_extra + 1, p + 1,
610 sizeof(chunk->map[0]) * (chunk->map_used - i));
611 chunk->map_used += nr_extra;
612
613 if (head) {
614 if (!seen_free) {
615 chunk->first_free = i;
616 seen_free = true;
617 }
618 *++p = off += head;
619 ++i;
620 max_contig = max(head, max_contig);
621 }
622 if (tail) {
623 p[1] = off + size;
624 max_contig = max(tail, max_contig);
625 }
626 }
627
628 if (!seen_free)
629 chunk->first_free = i + 1;
630
631 /* update hint and mark allocated */
632 if (i + 1 == chunk->map_used)
633 chunk->contig_hint = max_contig; /* fully scanned */
634 else
635 chunk->contig_hint = max(chunk->contig_hint,
636 max_contig);
637
638 chunk->free_size -= size;
639 *p |= 1;
640
641 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
642 pcpu_chunk_relocate(chunk, oslot);
643 return off;
644 }
645
646 chunk->contig_hint = max_contig; /* fully scanned */
647 pcpu_chunk_relocate(chunk, oslot);
648
649 /* tell the upper layer that this chunk has no matching area */
650 return -1;
651}
652
653/**
654 * pcpu_free_area - free area to a pcpu_chunk
655 * @chunk: chunk of interest
656 * @freeme: offset of area to free
657 * @occ_pages_p: out param for the number of pages the area occupies
658 *
659 * Free area starting from @freeme to @chunk. Note that this function
660 * only modifies the allocation map. It doesn't depopulate or unmap
661 * the area.
662 *
663 * CONTEXT:
664 * pcpu_lock.
665 */
666static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
667 int *occ_pages_p)
668{
669 int oslot = pcpu_chunk_slot(chunk);
670 int off = 0;
671 unsigned i, j;
672 int to_free = 0;
673 int *p;
674
675 freeme |= 1; /* we are searching for <given offset, in use> pair */
676
677 i = 0;
678 j = chunk->map_used;
679 while (i != j) {
680 unsigned k = (i + j) / 2;
681 off = chunk->map[k];
682 if (off < freeme)
683 i = k + 1;
684 else if (off > freeme)
685 j = k;
686 else
687 i = j = k;
688 }
689 BUG_ON(off != freeme);
690
691 if (i < chunk->first_free)
692 chunk->first_free = i;
693
694 p = chunk->map + i;
695 *p = off &= ~1;
696 chunk->free_size += (p[1] & ~1) - off;
697
698 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
699
700 /* merge with next? */
701 if (!(p[1] & 1))
702 to_free++;
703 /* merge with previous? */
704 if (i > 0 && !(p[-1] & 1)) {
705 to_free++;
706 i--;
707 p--;
708 }
709 if (to_free) {
710 chunk->map_used -= to_free;
711 memmove(p + 1, p + 1 + to_free,
712 (chunk->map_used - i) * sizeof(chunk->map[0]));
713 }
714
715 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
716 pcpu_chunk_relocate(chunk, oslot);
717}
718
719static struct pcpu_chunk *pcpu_alloc_chunk(void)
720{
721 struct pcpu_chunk *chunk;
722
723 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
724 if (!chunk)
725 return NULL;
726
727 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
728 sizeof(chunk->map[0]));
729 if (!chunk->map) {
730 pcpu_mem_free(chunk);
731 return NULL;
732 }
733
734 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
735 chunk->map[0] = 0;
736 chunk->map[1] = pcpu_unit_size | 1;
737 chunk->map_used = 1;
738
739 INIT_LIST_HEAD(&chunk->list);
740 INIT_LIST_HEAD(&chunk->map_extend_list);
741 chunk->free_size = pcpu_unit_size;
742 chunk->contig_hint = pcpu_unit_size;
743
744 return chunk;
745}
746
747static void pcpu_free_chunk(struct pcpu_chunk *chunk)
748{
749 if (!chunk)
750 return;
751 pcpu_mem_free(chunk->map);
752 pcpu_mem_free(chunk);
753}
754
755/**
756 * pcpu_chunk_populated - post-population bookkeeping
757 * @chunk: pcpu_chunk which got populated
758 * @page_start: the start page
759 * @page_end: the end page
760 *
761 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
762 * the bookkeeping information accordingly. Must be called after each
763 * successful population.
764 */
765static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
766 int page_start, int page_end)
767{
768 int nr = page_end - page_start;
769
770 lockdep_assert_held(&pcpu_lock);
771
772 bitmap_set(chunk->populated, page_start, nr);
773 chunk->nr_populated += nr;
774 pcpu_nr_empty_pop_pages += nr;
775}
776
777/**
778 * pcpu_chunk_depopulated - post-depopulation bookkeeping
779 * @chunk: pcpu_chunk which got depopulated
780 * @page_start: the start page
781 * @page_end: the end page
782 *
783 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
784 * Update the bookkeeping information accordingly. Must be called after
785 * each successful depopulation.
786 */
787static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
788 int page_start, int page_end)
789{
790 int nr = page_end - page_start;
791
792 lockdep_assert_held(&pcpu_lock);
793
794 bitmap_clear(chunk->populated, page_start, nr);
795 chunk->nr_populated -= nr;
796 pcpu_nr_empty_pop_pages -= nr;
797}
798
799/*
800 * Chunk management implementation.
801 *
802 * To allow different implementations, chunk alloc/free and
803 * [de]population are implemented in a separate file which is pulled
804 * into this file and compiled together. The following functions
805 * should be implemented.
806 *
807 * pcpu_populate_chunk - populate the specified range of a chunk
808 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
809 * pcpu_create_chunk - create a new chunk
810 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
811 * pcpu_addr_to_page - translate address to physical address
812 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
813 */
814static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
815static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
816static struct pcpu_chunk *pcpu_create_chunk(void);
817static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
818static struct page *pcpu_addr_to_page(void *addr);
819static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
820
821#ifdef CONFIG_NEED_PER_CPU_KM
822#include "percpu-km.c"
823#else
824#include "percpu-vm.c"
825#endif
826
827/**
828 * pcpu_chunk_addr_search - determine chunk containing specified address
829 * @addr: address for which the chunk needs to be determined.
830 *
831 * RETURNS:
832 * The address of the found chunk.
833 */
834static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
835{
836 /* is it in the first chunk? */
837 if (pcpu_addr_in_first_chunk(addr)) {
838 /* is it in the reserved area? */
839 if (pcpu_addr_in_reserved_chunk(addr))
840 return pcpu_reserved_chunk;
841 return pcpu_first_chunk;
842 }
843
844 /*
845 * The address is relative to unit0 which might be unused and
846 * thus unmapped. Offset the address to the unit space of the
847 * current processor before looking it up in the vmalloc
848 * space. Note that any possible cpu id can be used here, so
849 * there's no need to worry about preemption or cpu hotplug.
850 */
851 addr += pcpu_unit_offsets[raw_smp_processor_id()];
852 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
853}
854
855/**
856 * pcpu_alloc - the percpu allocator
857 * @size: size of area to allocate in bytes
858 * @align: alignment of area (max PAGE_SIZE)
859 * @reserved: allocate from the reserved chunk if available
860 * @gfp: allocation flags
861 *
862 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
863 * contain %GFP_KERNEL, the allocation is atomic.
864 *
865 * RETURNS:
866 * Percpu pointer to the allocated area on success, NULL on failure.
867 */
868static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
869 gfp_t gfp)
870{
871 static int warn_limit = 10;
872 struct pcpu_chunk *chunk;
873 const char *err;
874 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
875 int occ_pages = 0;
876 int slot, off, new_alloc, cpu, ret;
877 unsigned long flags;
878 void __percpu *ptr;
879
880 /*
881 * We want the lowest bit of offset available for in-use/free
882 * indicator, so force >= 16bit alignment and make size even.
883 */
884 if (unlikely(align < 2))
885 align = 2;
886
887 size = ALIGN(size, 2);
888
889 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
890 !is_power_of_2(align))) {
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
929restart:
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
986area_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
1031fail_unlock:
1032 spin_unlock_irqrestore(&pcpu_lock, flags);
1033fail:
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 */
1064void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1065{
1066 return pcpu_alloc(size, align, false, gfp);
1067}
1068EXPORT_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 */
1077void __percpu *__alloc_percpu(size_t size, size_t align)
1078{
1079 return pcpu_alloc(size, align, false, GFP_KERNEL);
1080}
1081EXPORT_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 */
1099void __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 */
1110static 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 */
1178retry_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 */
1248void 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}
1285EXPORT_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 */
1298bool 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 */
1339phys_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 */
1392struct 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 */
1427void __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 */
1439static 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 */
1545int __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
1723const 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
1729enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1730
1731static 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}
1751early_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 */
1791static 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 */
1959int __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 @%p s%zu r%zu d%zu u%zu\n",
2052 PFN_DOWN(size_sum), base, 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
2058out_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);
2063out_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 */
2088int __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 int upa;
2101 int nr_g0_units;
2102
2103 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2104
2105 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2106 if (IS_ERR(ai))
2107 return PTR_ERR(ai);
2108 BUG_ON(ai->nr_groups != 1);
2109 upa = ai->alloc_size/ai->unit_size;
2110 nr_g0_units = roundup(num_possible_cpus(), upa);
2111 if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2112 pcpu_free_alloc_info(ai);
2113 return -EINVAL;
2114 }
2115
2116 unit_pages = ai->unit_size >> PAGE_SHIFT;
2117
2118 /* unaligned allocations can't be freed, round up to page size */
2119 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2120 sizeof(pages[0]));
2121 pages = memblock_virt_alloc(pages_size, 0);
2122
2123 /* allocate pages */
2124 j = 0;
2125 for (unit = 0; unit < num_possible_cpus(); unit++) {
2126 unsigned int cpu = ai->groups[0].cpu_map[unit];
2127 for (i = 0; i < unit_pages; i++) {
2128 void *ptr;
2129
2130 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2131 if (!ptr) {
2132 pr_warn("failed to allocate %s page for cpu%u\n",
2133 psize_str, cpu);
2134 goto enomem;
2135 }
2136 /* kmemleak tracks the percpu allocations separately */
2137 kmemleak_free(ptr);
2138 pages[j++] = virt_to_page(ptr);
2139 }
2140 }
2141
2142 /* allocate vm area, map the pages and copy static data */
2143 vm.flags = VM_ALLOC;
2144 vm.size = num_possible_cpus() * ai->unit_size;
2145 vm_area_register_early(&vm, PAGE_SIZE);
2146
2147 for (unit = 0; unit < num_possible_cpus(); unit++) {
2148 unsigned long unit_addr =
2149 (unsigned long)vm.addr + unit * ai->unit_size;
2150
2151 for (i = 0; i < unit_pages; i++)
2152 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2153
2154 /* pte already populated, the following shouldn't fail */
2155 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2156 unit_pages);
2157 if (rc < 0)
2158 panic("failed to map percpu area, err=%d\n", rc);
2159
2160 /*
2161 * FIXME: Archs with virtual cache should flush local
2162 * cache for the linear mapping here - something
2163 * equivalent to flush_cache_vmap() on the local cpu.
2164 * flush_cache_vmap() can't be used as most supporting
2165 * data structures are not set up yet.
2166 */
2167
2168 /* copy static data */
2169 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2170 }
2171
2172 /* we're ready, commit */
2173 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2174 unit_pages, psize_str, vm.addr, ai->static_size,
2175 ai->reserved_size, ai->dyn_size);
2176
2177 rc = pcpu_setup_first_chunk(ai, vm.addr);
2178 goto out_free_ar;
2179
2180enomem:
2181 while (--j >= 0)
2182 free_fn(page_address(pages[j]), PAGE_SIZE);
2183 rc = -ENOMEM;
2184out_free_ar:
2185 memblock_free_early(__pa(pages), pages_size);
2186 pcpu_free_alloc_info(ai);
2187 return rc;
2188}
2189#endif /* BUILD_PAGE_FIRST_CHUNK */
2190
2191#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2192/*
2193 * Generic SMP percpu area setup.
2194 *
2195 * The embedding helper is used because its behavior closely resembles
2196 * the original non-dynamic generic percpu area setup. This is
2197 * important because many archs have addressing restrictions and might
2198 * fail if the percpu area is located far away from the previous
2199 * location. As an added bonus, in non-NUMA cases, embedding is
2200 * generally a good idea TLB-wise because percpu area can piggy back
2201 * on the physical linear memory mapping which uses large page
2202 * mappings on applicable archs.
2203 */
2204unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2205EXPORT_SYMBOL(__per_cpu_offset);
2206
2207static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2208 size_t align)
2209{
2210 return memblock_virt_alloc_from_nopanic(
2211 size, align, __pa(MAX_DMA_ADDRESS));
2212}
2213
2214static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2215{
2216 memblock_free_early(__pa(ptr), size);
2217}
2218
2219void __init setup_per_cpu_areas(void)
2220{
2221 unsigned long delta;
2222 unsigned int cpu;
2223 int rc;
2224
2225 /*
2226 * Always reserve area for module percpu variables. That's
2227 * what the legacy allocator did.
2228 */
2229 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2230 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2231 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2232 if (rc < 0)
2233 panic("Failed to initialize percpu areas.");
2234
2235 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2236 for_each_possible_cpu(cpu)
2237 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2238}
2239#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2240
2241#else /* CONFIG_SMP */
2242
2243/*
2244 * UP percpu area setup.
2245 *
2246 * UP always uses km-based percpu allocator with identity mapping.
2247 * Static percpu variables are indistinguishable from the usual static
2248 * variables and don't require any special preparation.
2249 */
2250void __init setup_per_cpu_areas(void)
2251{
2252 const size_t unit_size =
2253 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2254 PERCPU_DYNAMIC_RESERVE));
2255 struct pcpu_alloc_info *ai;
2256 void *fc;
2257
2258 ai = pcpu_alloc_alloc_info(1, 1);
2259 fc = memblock_virt_alloc_from_nopanic(unit_size,
2260 PAGE_SIZE,
2261 __pa(MAX_DMA_ADDRESS));
2262 if (!ai || !fc)
2263 panic("Failed to allocate memory for percpu areas.");
2264 /* kmemleak tracks the percpu allocations separately */
2265 kmemleak_free(fc);
2266
2267 ai->dyn_size = unit_size;
2268 ai->unit_size = unit_size;
2269 ai->atom_size = unit_size;
2270 ai->alloc_size = unit_size;
2271 ai->groups[0].nr_units = 1;
2272 ai->groups[0].cpu_map[0] = 0;
2273
2274 if (pcpu_setup_first_chunk(ai, fc) < 0)
2275 panic("Failed to initialize percpu areas.");
2276}
2277
2278#endif /* CONFIG_SMP */
2279
2280/*
2281 * First and reserved chunks are initialized with temporary allocation
2282 * map in initdata so that they can be used before slab is online.
2283 * This function is called after slab is brought up and replaces those
2284 * with properly allocated maps.
2285 */
2286void __init percpu_init_late(void)
2287{
2288 struct pcpu_chunk *target_chunks[] =
2289 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2290 struct pcpu_chunk *chunk;
2291 unsigned long flags;
2292 int i;
2293
2294 for (i = 0; (chunk = target_chunks[i]); i++) {
2295 int *map;
2296 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2297
2298 BUILD_BUG_ON(size > PAGE_SIZE);
2299
2300 map = pcpu_mem_zalloc(size);
2301 BUG_ON(!map);
2302
2303 spin_lock_irqsave(&pcpu_lock, flags);
2304 memcpy(map, chunk->map, size);
2305 chunk->map = map;
2306 spin_unlock_irqrestore(&pcpu_lock, flags);
2307 }
2308}
2309
2310/*
2311 * Percpu allocator is initialized early during boot when neither slab or
2312 * workqueue is available. Plug async management until everything is up
2313 * and running.
2314 */
2315static int __init percpu_enable_async(void)
2316{
2317 pcpu_async_enabled = true;
2318 return 0;
2319}
2320subsys_initcall(percpu_enable_async);