2 * mm/percpu.c - percpu memory allocator
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 * This file is released under the GPLv2 license.
9 * The percpu allocator handles both static and dynamic areas. Percpu
10 * areas are allocated in chunks which are divided into units. There is
11 * a 1-to-1 mapping for units to possible cpus. These units are grouped
12 * based on NUMA properties of the machine.
15 * ------------------- ------------------- ------------
16 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
17 * ------------------- ...... ------------------- .... ------------
19 * Allocation is done by offsets into a unit's address space. Ie., an
20 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
21 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
22 * and even sparse. Access is handled by configuring percpu base
23 * registers according to the cpu to unit mappings and offsetting the
24 * base address using pcpu_unit_size.
26 * There is special consideration for the first chunk which must handle
27 * the static percpu variables in the kernel image as allocation services
28 * are not online yet. In short, the first chunk is structure like so:
30 * <Static | [Reserved] | Dynamic>
32 * The static data is copied from the original section managed by the
33 * linker. The reserved section, if non-zero, primarily manages static
34 * percpu variables from kernel modules. Finally, the dynamic section
35 * takes care of normal allocations.
37 * Allocation state in each chunk is kept using an array of integers
38 * on chunk->map. A positive value in the map represents a free
39 * region and negative allocated. Allocation inside a chunk is done
40 * by scanning this map sequentially and serving the first matching
41 * entry. This is mostly copied from the percpu_modalloc() allocator.
42 * Chunks can be determined from the address using the index field
43 * in the page struct. The index field contains a pointer to the chunk.
45 * These chunks are organized into lists according to free_size and
46 * tries to allocate from the fullest chunk first. Each chunk maintains
47 * a maximum contiguous area size hint which is guaranteed to be equal
48 * to or larger than the maximum contiguous area in the chunk. This
49 * helps prevent the allocator from iterating over chunks unnecessarily.
51 * To use this allocator, arch code should do the following:
53 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
54 * regular address to percpu pointer and back if they need to be
55 * different from the default
57 * - use pcpu_setup_first_chunk() during percpu area initialization to
58 * setup the first chunk containing the kernel static percpu area
61 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
63 #include <linux/bitmap.h>
64 #include <linux/bootmem.h>
65 #include <linux/err.h>
66 #include <linux/lcm.h>
67 #include <linux/list.h>
68 #include <linux/log2.h>
70 #include <linux/module.h>
71 #include <linux/mutex.h>
72 #include <linux/percpu.h>
73 #include <linux/pfn.h>
74 #include <linux/slab.h>
75 #include <linux/spinlock.h>
76 #include <linux/vmalloc.h>
77 #include <linux/workqueue.h>
78 #include <linux/kmemleak.h>
80 #include <asm/cacheflush.h>
81 #include <asm/sections.h>
82 #include <asm/tlbflush.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/percpu.h>
88 #include "percpu-internal.h"
90 /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
91 #define PCPU_SLOT_BASE_SHIFT 5
93 #define PCPU_EMPTY_POP_PAGES_LOW 2
94 #define PCPU_EMPTY_POP_PAGES_HIGH 4
97 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
98 #ifndef __addr_to_pcpu_ptr
99 #define __addr_to_pcpu_ptr(addr) \
100 (void __percpu *)((unsigned long)(addr) - \
101 (unsigned long)pcpu_base_addr + \
102 (unsigned long)__per_cpu_start)
104 #ifndef __pcpu_ptr_to_addr
105 #define __pcpu_ptr_to_addr(ptr) \
106 (void __force *)((unsigned long)(ptr) + \
107 (unsigned long)pcpu_base_addr - \
108 (unsigned long)__per_cpu_start)
110 #else /* CONFIG_SMP */
111 /* on UP, it's always identity mapped */
112 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
113 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
114 #endif /* CONFIG_SMP */
116 static int pcpu_unit_pages __ro_after_init;
117 static int pcpu_unit_size __ro_after_init;
118 static int pcpu_nr_units __ro_after_init;
119 static int pcpu_atom_size __ro_after_init;
120 int pcpu_nr_slots __ro_after_init;
121 static size_t pcpu_chunk_struct_size __ro_after_init;
123 /* cpus with the lowest and highest unit addresses */
124 static unsigned int pcpu_low_unit_cpu __ro_after_init;
125 static unsigned int pcpu_high_unit_cpu __ro_after_init;
127 /* the address of the first chunk which starts with the kernel static area */
128 void *pcpu_base_addr __ro_after_init;
129 EXPORT_SYMBOL_GPL(pcpu_base_addr);
131 static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
132 const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
134 /* group information, used for vm allocation */
135 static int pcpu_nr_groups __ro_after_init;
136 static const unsigned long *pcpu_group_offsets __ro_after_init;
137 static const size_t *pcpu_group_sizes __ro_after_init;
140 * The first chunk which always exists. Note that unlike other
141 * chunks, this one can be allocated and mapped in several different
142 * ways and thus often doesn't live in the vmalloc area.
144 struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
147 * Optional reserved chunk. This chunk reserves part of the first
148 * chunk and serves it for reserved allocations. When the reserved
149 * region doesn't exist, the following variable is NULL.
151 struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
153 DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
154 static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
156 struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
158 /* chunks which need their map areas extended, protected by pcpu_lock */
159 static LIST_HEAD(pcpu_map_extend_chunks);
162 * The number of empty populated pages, protected by pcpu_lock. The
163 * reserved chunk doesn't contribute to the count.
165 int pcpu_nr_empty_pop_pages;
168 * Balance work is used to populate or destroy chunks asynchronously. We
169 * try to keep the number of populated free pages between
170 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
173 static void pcpu_balance_workfn(struct work_struct *work);
174 static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
175 static bool pcpu_async_enabled __read_mostly;
176 static bool pcpu_atomic_alloc_failed;
178 static void pcpu_schedule_balance_work(void)
180 if (pcpu_async_enabled)
181 schedule_work(&pcpu_balance_work);
185 * pcpu_addr_in_chunk - check if the address is served from this chunk
186 * @chunk: chunk of interest
187 * @addr: percpu address
190 * True if the address is served from this chunk.
192 static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
194 void *start_addr, *end_addr;
199 start_addr = chunk->base_addr + chunk->start_offset;
200 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
203 return addr >= start_addr && addr < end_addr;
206 static int __pcpu_size_to_slot(int size)
208 int highbit = fls(size); /* size is in bytes */
209 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
212 static int pcpu_size_to_slot(int size)
214 if (size == pcpu_unit_size)
215 return pcpu_nr_slots - 1;
216 return __pcpu_size_to_slot(size);
219 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
221 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0)
224 return pcpu_size_to_slot(chunk->free_bytes);
227 /* set the pointer to a chunk in a page struct */
228 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
230 page->index = (unsigned long)pcpu;
233 /* obtain pointer to a chunk from a page struct */
234 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
236 return (struct pcpu_chunk *)page->index;
239 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
241 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
244 static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
246 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
249 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
250 unsigned int cpu, int page_idx)
252 return (unsigned long)chunk->base_addr +
253 pcpu_unit_page_offset(cpu, page_idx);
256 static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
258 *rs = find_next_zero_bit(bitmap, end, *rs);
259 *re = find_next_bit(bitmap, end, *rs + 1);
262 static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
264 *rs = find_next_bit(bitmap, end, *rs);
265 *re = find_next_zero_bit(bitmap, end, *rs + 1);
269 * Bitmap region iterators. Iterates over the bitmap between
270 * [@start, @end) in @chunk. @rs and @re should be integer variables
271 * and will be set to start and end index of the current free region.
273 #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
274 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
276 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
278 #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
279 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
281 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
284 * The following are helper functions to help access bitmaps and convert
285 * between bitmap offsets to address offsets.
287 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
289 return chunk->alloc_map +
290 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
293 static unsigned long pcpu_off_to_block_index(int off)
295 return off / PCPU_BITMAP_BLOCK_BITS;
298 static unsigned long pcpu_off_to_block_off(int off)
300 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
304 * pcpu_mem_zalloc - allocate memory
305 * @size: bytes to allocate
307 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
308 * kzalloc() is used; otherwise, vzalloc() is used. The returned
309 * memory is always zeroed.
312 * Does GFP_KERNEL allocation.
315 * Pointer to the allocated area on success, NULL on failure.
317 static void *pcpu_mem_zalloc(size_t size)
319 if (WARN_ON_ONCE(!slab_is_available()))
322 if (size <= PAGE_SIZE)
323 return kzalloc(size, GFP_KERNEL);
325 return vzalloc(size);
329 * pcpu_mem_free - free memory
330 * @ptr: memory to free
332 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
334 static void pcpu_mem_free(void *ptr)
340 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
341 * @chunk: chunk of interest
342 * @oslot: the previous slot it was on
344 * This function is called after an allocation or free changed @chunk.
345 * New slot according to the changed state is determined and @chunk is
346 * moved to the slot. Note that the reserved chunk is never put on
352 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
354 int nslot = pcpu_chunk_slot(chunk);
356 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
358 list_move(&chunk->list, &pcpu_slot[nslot]);
360 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
365 * pcpu_cnt_pop_pages- counts populated backing pages in range
366 * @chunk: chunk of interest
367 * @bit_off: start offset
368 * @bits: size of area to check
370 * Calculates the number of populated pages in the region
371 * [page_start, page_end). This keeps track of how many empty populated
372 * pages are available and decide if async work should be scheduled.
375 * The nr of populated pages.
377 static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off,
380 int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE);
381 int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
383 if (page_start >= page_end)
387 * bitmap_weight counts the number of bits set in a bitmap up to
388 * the specified number of bits. This is counting the populated
389 * pages up to page_end and then subtracting the populated pages
390 * up to page_start to count the populated pages in
391 * [page_start, page_end).
393 return bitmap_weight(chunk->populated, page_end) -
394 bitmap_weight(chunk->populated, page_start);
398 * pcpu_chunk_update - updates the chunk metadata given a free area
399 * @chunk: chunk of interest
400 * @bit_off: chunk offset
401 * @bits: size of free area
403 * This updates the chunk's contig hint and starting offset given a free area.
405 static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits)
407 if (bits > chunk->contig_bits) {
408 chunk->contig_bits_start = bit_off;
409 chunk->contig_bits = bits;
414 * pcpu_chunk_refresh_hint - updates metadata about a chunk
415 * @chunk: chunk of interest
417 * Iterates over the chunk to find the largest free area.
421 * chunk->contig_bits_start
424 static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk)
426 int bits, nr_empty_pop_pages;
427 int rs, re; /* region start, region end */
430 chunk->contig_bits = 0;
432 bits = nr_empty_pop_pages = 0;
433 pcpu_for_each_unpop_region(chunk->alloc_map, rs, re, chunk->first_bit,
434 pcpu_chunk_map_bits(chunk)) {
437 pcpu_chunk_update(chunk, rs, bits);
439 nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, rs, bits);
443 * Keep track of nr_empty_pop_pages.
445 * The chunk maintains the previous number of free pages it held,
446 * so the delta is used to update the global counter. The reserved
447 * chunk is not part of the free page count as they are populated
448 * at init and are special to serving reserved allocations.
450 if (chunk != pcpu_reserved_chunk)
451 pcpu_nr_empty_pop_pages +=
452 (nr_empty_pop_pages - chunk->nr_empty_pop_pages);
454 chunk->nr_empty_pop_pages = nr_empty_pop_pages;
458 * pcpu_block_update - updates a block given a free area
459 * @block: block of interest
460 * @start: start offset in block
461 * @end: end offset in block
463 * Updates a block given a known free area. The region [start, end) is
464 * expected to be the entirety of the free area within a block.
466 static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
468 int contig = end - start;
470 block->first_free = min(block->first_free, start);
472 block->left_free = contig;
474 if (end == PCPU_BITMAP_BLOCK_BITS)
475 block->right_free = contig;
477 if (contig > block->contig_hint) {
478 block->contig_hint_start = start;
479 block->contig_hint = contig;
484 * pcpu_block_refresh_hint
485 * @chunk: chunk of interest
486 * @index: index of the metadata block
488 * Scans over the block beginning at first_free and updates the block
489 * metadata accordingly.
491 static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
493 struct pcpu_block_md *block = chunk->md_blocks + index;
494 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
495 int rs, re; /* region start, region end */
498 block->contig_hint = 0;
499 block->left_free = block->right_free = 0;
501 /* iterate over free areas and update the contig hints */
502 pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free,
503 PCPU_BITMAP_BLOCK_BITS) {
504 pcpu_block_update(block, rs, re);
509 * pcpu_block_update_hint_alloc - update hint on allocation path
510 * @chunk: chunk of interest
511 * @bit_off: chunk offset
512 * @bits: size of request
514 static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
517 struct pcpu_block_md *s_block, *e_block, *block;
518 int s_index, e_index; /* block indexes of the freed allocation */
519 int s_off, e_off; /* block offsets of the freed allocation */
522 * Calculate per block offsets.
523 * The calculation uses an inclusive range, but the resulting offsets
524 * are [start, end). e_index always points to the last block in the
527 s_index = pcpu_off_to_block_index(bit_off);
528 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
529 s_off = pcpu_off_to_block_off(bit_off);
530 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
532 s_block = chunk->md_blocks + s_index;
533 e_block = chunk->md_blocks + e_index;
538 pcpu_block_refresh_hint(chunk, s_index);
543 if (s_index != e_index) {
544 pcpu_block_refresh_hint(chunk, e_index);
546 /* update in-between md_blocks */
547 for (block = s_block + 1; block < e_block; block++) {
548 block->contig_hint = 0;
549 block->left_free = 0;
550 block->right_free = 0;
554 pcpu_chunk_refresh_hint(chunk);
558 * pcpu_block_update_hint_free - updates the block hints on the free path
559 * @chunk: chunk of interest
560 * @bit_off: chunk offset
561 * @bits: size of request
563 static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
566 struct pcpu_block_md *s_block, *e_block, *block;
567 int s_index, e_index; /* block indexes of the freed allocation */
568 int s_off, e_off; /* block offsets of the freed allocation */
571 * Calculate per block offsets.
572 * The calculation uses an inclusive range, but the resulting offsets
573 * are [start, end). e_index always points to the last block in the
576 s_index = pcpu_off_to_block_index(bit_off);
577 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
578 s_off = pcpu_off_to_block_off(bit_off);
579 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
581 s_block = chunk->md_blocks + s_index;
582 e_block = chunk->md_blocks + e_index;
585 pcpu_block_refresh_hint(chunk, s_index);
587 /* freeing in the same block */
588 if (s_index != e_index) {
590 pcpu_block_refresh_hint(chunk, e_index);
592 /* reset md_blocks in the middle */
593 for (block = s_block + 1; block < e_block; block++) {
594 block->first_free = 0;
595 block->contig_hint_start = 0;
596 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
597 block->left_free = PCPU_BITMAP_BLOCK_BITS;
598 block->right_free = PCPU_BITMAP_BLOCK_BITS;
602 pcpu_chunk_refresh_hint(chunk);
606 * pcpu_is_populated - determines if the region is populated
607 * @chunk: chunk of interest
608 * @bit_off: chunk offset
609 * @bits: size of area
610 * @next_off: return value for the next offset to start searching
612 * For atomic allocations, check if the backing pages are populated.
615 * Bool if the backing pages are populated.
616 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
618 static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
621 int page_start, page_end, rs, re;
623 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
624 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
627 pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
631 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
636 * pcpu_find_block_fit - finds the block index to start searching
637 * @chunk: chunk of interest
638 * @alloc_bits: size of request in allocation units
639 * @align: alignment of area (max PAGE_SIZE bytes)
640 * @pop_only: use populated regions only
643 * The offset in the bitmap to begin searching.
644 * -1 if no offset is found.
646 static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
647 size_t align, bool pop_only)
650 int re; /* region end */
653 * Check to see if the allocation can fit in the chunk's contig hint.
654 * This is an optimization to prevent scanning by assuming if it
655 * cannot fit in the global hint, there is memory pressure and creating
656 * a new chunk would happen soon.
658 bit_off = ALIGN(chunk->contig_bits_start, align) -
659 chunk->contig_bits_start;
660 if (bit_off + alloc_bits > chunk->contig_bits)
663 pcpu_for_each_unpop_region(chunk->alloc_map, bit_off, re,
665 pcpu_chunk_map_bits(chunk)) {
668 /* check alignment */
669 bits -= ALIGN(bit_off, align) - bit_off;
670 bit_off = ALIGN(bit_off, align);
671 if (bits < alloc_bits)
675 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
682 if (bit_off == pcpu_chunk_map_bits(chunk))
689 * pcpu_alloc_area - allocates an area from a pcpu_chunk
690 * @chunk: chunk of interest
691 * @alloc_bits: size of request in allocation units
692 * @align: alignment of area (max PAGE_SIZE)
693 * @start: bit_off to start searching
695 * This function takes in a @start offset to begin searching to fit an
696 * allocation of @alloc_bits with alignment @align. If it confirms a
697 * valid free area, it then updates the allocation and boundary maps
701 * Allocated addr offset in @chunk on success.
702 * -1 if no matching area is found.
704 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
705 size_t align, int start)
707 size_t align_mask = (align) ? (align - 1) : 0;
708 int bit_off, end, oslot;
710 lockdep_assert_held(&pcpu_lock);
712 oslot = pcpu_chunk_slot(chunk);
715 * Search to find a fit.
717 end = start + alloc_bits;
718 bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start,
719 alloc_bits, align_mask);
723 /* update alloc map */
724 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
726 /* update boundary map */
727 set_bit(bit_off, chunk->bound_map);
728 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
729 set_bit(bit_off + alloc_bits, chunk->bound_map);
731 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
733 /* update first free bit */
734 if (bit_off == chunk->first_bit)
735 chunk->first_bit = find_next_zero_bit(
737 pcpu_chunk_map_bits(chunk),
738 bit_off + alloc_bits);
740 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
742 pcpu_chunk_relocate(chunk, oslot);
744 return bit_off * PCPU_MIN_ALLOC_SIZE;
748 * pcpu_free_area - frees the corresponding offset
749 * @chunk: chunk of interest
750 * @off: addr offset into chunk
752 * This function determines the size of an allocation to free using
753 * the boundary bitmap and clears the allocation map.
755 static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
757 int bit_off, bits, end, oslot;
759 lockdep_assert_held(&pcpu_lock);
760 pcpu_stats_area_dealloc(chunk);
762 oslot = pcpu_chunk_slot(chunk);
764 bit_off = off / PCPU_MIN_ALLOC_SIZE;
767 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
769 bits = end - bit_off;
770 bitmap_clear(chunk->alloc_map, bit_off, bits);
772 /* update metadata */
773 chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
775 /* update first free bit */
776 chunk->first_bit = min(chunk->first_bit, bit_off);
778 pcpu_block_update_hint_free(chunk, bit_off, bits);
780 pcpu_chunk_relocate(chunk, oslot);
783 static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
785 struct pcpu_block_md *md_block;
787 for (md_block = chunk->md_blocks;
788 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
790 md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
791 md_block->left_free = PCPU_BITMAP_BLOCK_BITS;
792 md_block->right_free = PCPU_BITMAP_BLOCK_BITS;
797 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
798 * @tmp_addr: the start of the region served
799 * @map_size: size of the region served
801 * This is responsible for creating the chunks that serve the first chunk. The
802 * base_addr is page aligned down of @tmp_addr while the region end is page
803 * aligned up. Offsets are kept track of to determine the region served. All
804 * this is done to appease the bitmap allocator in avoiding partial blocks.
807 * Chunk serving the region at @tmp_addr of @map_size.
809 static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
812 struct pcpu_chunk *chunk;
813 unsigned long aligned_addr, lcm_align;
814 int start_offset, offset_bits, region_size, region_bits;
816 /* region calculations */
817 aligned_addr = tmp_addr & PAGE_MASK;
819 start_offset = tmp_addr - aligned_addr;
822 * Align the end of the region with the LCM of PAGE_SIZE and
823 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
826 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
827 region_size = ALIGN(start_offset + map_size, lcm_align);
830 chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) +
831 BITS_TO_LONGS(region_size >> PAGE_SHIFT),
834 INIT_LIST_HEAD(&chunk->list);
836 chunk->base_addr = (void *)aligned_addr;
837 chunk->start_offset = start_offset;
838 chunk->end_offset = region_size - chunk->start_offset - map_size;
840 chunk->nr_pages = region_size >> PAGE_SHIFT;
841 region_bits = pcpu_chunk_map_bits(chunk);
843 chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) *
844 sizeof(chunk->alloc_map[0]), 0);
845 chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) *
846 sizeof(chunk->bound_map[0]), 0);
847 chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) *
848 sizeof(chunk->md_blocks[0]), 0);
849 pcpu_init_md_blocks(chunk);
851 /* manage populated page bitmap */
852 chunk->immutable = true;
853 bitmap_fill(chunk->populated, chunk->nr_pages);
854 chunk->nr_populated = chunk->nr_pages;
855 chunk->nr_empty_pop_pages =
856 pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE,
857 map_size / PCPU_MIN_ALLOC_SIZE);
859 chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE;
860 chunk->free_bytes = map_size;
862 if (chunk->start_offset) {
863 /* hide the beginning of the bitmap */
864 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
865 bitmap_set(chunk->alloc_map, 0, offset_bits);
866 set_bit(0, chunk->bound_map);
867 set_bit(offset_bits, chunk->bound_map);
869 chunk->first_bit = offset_bits;
871 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
874 if (chunk->end_offset) {
875 /* hide the end of the bitmap */
876 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
877 bitmap_set(chunk->alloc_map,
878 pcpu_chunk_map_bits(chunk) - offset_bits,
880 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
882 set_bit(region_bits, chunk->bound_map);
884 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
885 - offset_bits, offset_bits);
891 static struct pcpu_chunk *pcpu_alloc_chunk(void)
893 struct pcpu_chunk *chunk;
896 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
900 INIT_LIST_HEAD(&chunk->list);
901 chunk->nr_pages = pcpu_unit_pages;
902 region_bits = pcpu_chunk_map_bits(chunk);
904 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
905 sizeof(chunk->alloc_map[0]));
906 if (!chunk->alloc_map)
909 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
910 sizeof(chunk->bound_map[0]));
911 if (!chunk->bound_map)
914 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
915 sizeof(chunk->md_blocks[0]));
916 if (!chunk->md_blocks)
919 pcpu_init_md_blocks(chunk);
922 chunk->contig_bits = region_bits;
923 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
928 pcpu_mem_free(chunk->bound_map);
930 pcpu_mem_free(chunk->alloc_map);
932 pcpu_mem_free(chunk);
937 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
941 pcpu_mem_free(chunk->bound_map);
942 pcpu_mem_free(chunk->alloc_map);
943 pcpu_mem_free(chunk);
947 * pcpu_chunk_populated - post-population bookkeeping
948 * @chunk: pcpu_chunk which got populated
949 * @page_start: the start page
950 * @page_end: the end page
951 * @for_alloc: if this is to populate for allocation
953 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
954 * the bookkeeping information accordingly. Must be called after each
955 * successful population.
957 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
958 * is to serve an allocation in that area.
960 static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
961 int page_end, bool for_alloc)
963 int nr = page_end - page_start;
965 lockdep_assert_held(&pcpu_lock);
967 bitmap_set(chunk->populated, page_start, nr);
968 chunk->nr_populated += nr;
971 chunk->nr_empty_pop_pages += nr;
972 pcpu_nr_empty_pop_pages += nr;
977 * pcpu_chunk_depopulated - post-depopulation bookkeeping
978 * @chunk: pcpu_chunk which got depopulated
979 * @page_start: the start page
980 * @page_end: the end page
982 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
983 * Update the bookkeeping information accordingly. Must be called after
984 * each successful depopulation.
986 static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
987 int page_start, int page_end)
989 int nr = page_end - page_start;
991 lockdep_assert_held(&pcpu_lock);
993 bitmap_clear(chunk->populated, page_start, nr);
994 chunk->nr_populated -= nr;
995 chunk->nr_empty_pop_pages -= nr;
996 pcpu_nr_empty_pop_pages -= nr;
1000 * Chunk management implementation.
1002 * To allow different implementations, chunk alloc/free and
1003 * [de]population are implemented in a separate file which is pulled
1004 * into this file and compiled together. The following functions
1005 * should be implemented.
1007 * pcpu_populate_chunk - populate the specified range of a chunk
1008 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1009 * pcpu_create_chunk - create a new chunk
1010 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1011 * pcpu_addr_to_page - translate address to physical address
1012 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1014 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
1015 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
1016 static struct pcpu_chunk *pcpu_create_chunk(void);
1017 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1018 static struct page *pcpu_addr_to_page(void *addr);
1019 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1021 #ifdef CONFIG_NEED_PER_CPU_KM
1022 #include "percpu-km.c"
1024 #include "percpu-vm.c"
1028 * pcpu_chunk_addr_search - determine chunk containing specified address
1029 * @addr: address for which the chunk needs to be determined.
1031 * This is an internal function that handles all but static allocations.
1032 * Static percpu address values should never be passed into the allocator.
1035 * The address of the found chunk.
1037 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1039 /* is it in the dynamic region (first chunk)? */
1040 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1041 return pcpu_first_chunk;
1043 /* is it in the reserved region? */
1044 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1045 return pcpu_reserved_chunk;
1048 * The address is relative to unit0 which might be unused and
1049 * thus unmapped. Offset the address to the unit space of the
1050 * current processor before looking it up in the vmalloc
1051 * space. Note that any possible cpu id can be used here, so
1052 * there's no need to worry about preemption or cpu hotplug.
1054 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1055 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1059 * pcpu_alloc - the percpu allocator
1060 * @size: size of area to allocate in bytes
1061 * @align: alignment of area (max PAGE_SIZE)
1062 * @reserved: allocate from the reserved chunk if available
1063 * @gfp: allocation flags
1065 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1066 * contain %GFP_KERNEL, the allocation is atomic.
1069 * Percpu pointer to the allocated area on success, NULL on failure.
1071 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1074 static int warn_limit = 10;
1075 struct pcpu_chunk *chunk;
1077 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1078 int slot, off, cpu, ret;
1079 unsigned long flags;
1081 size_t bits, bit_align;
1084 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1085 * therefore alignment must be a minimum of that many bytes.
1086 * An allocation may have internal fragmentation from rounding up
1087 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1089 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1090 align = PCPU_MIN_ALLOC_SIZE;
1092 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1093 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1094 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1096 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1097 !is_power_of_2(align))) {
1098 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1104 mutex_lock(&pcpu_alloc_mutex);
1106 spin_lock_irqsave(&pcpu_lock, flags);
1108 /* serve reserved allocations from the reserved chunk if available */
1109 if (reserved && pcpu_reserved_chunk) {
1110 chunk = pcpu_reserved_chunk;
1112 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1114 err = "alloc from reserved chunk failed";
1118 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1122 err = "alloc from reserved chunk failed";
1127 /* search through normal chunks */
1128 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1129 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1130 off = pcpu_find_block_fit(chunk, bits, bit_align,
1135 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1142 spin_unlock_irqrestore(&pcpu_lock, flags);
1145 * No space left. Create a new chunk. We don't want multiple
1146 * tasks to create chunks simultaneously. Serialize and create iff
1147 * there's still no empty chunk after grabbing the mutex.
1150 err = "atomic alloc failed, no space left";
1154 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1155 chunk = pcpu_create_chunk();
1157 err = "failed to allocate new chunk";
1161 spin_lock_irqsave(&pcpu_lock, flags);
1162 pcpu_chunk_relocate(chunk, -1);
1164 spin_lock_irqsave(&pcpu_lock, flags);
1170 pcpu_stats_area_alloc(chunk, size);
1171 spin_unlock_irqrestore(&pcpu_lock, flags);
1173 /* populate if not all pages are already there */
1175 int page_start, page_end, rs, re;
1177 page_start = PFN_DOWN(off);
1178 page_end = PFN_UP(off + size);
1180 pcpu_for_each_unpop_region(chunk->populated, rs, re,
1181 page_start, page_end) {
1182 WARN_ON(chunk->immutable);
1184 ret = pcpu_populate_chunk(chunk, rs, re);
1186 spin_lock_irqsave(&pcpu_lock, flags);
1188 pcpu_free_area(chunk, off);
1189 err = "failed to populate";
1192 pcpu_chunk_populated(chunk, rs, re, true);
1193 spin_unlock_irqrestore(&pcpu_lock, flags);
1196 mutex_unlock(&pcpu_alloc_mutex);
1199 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1200 pcpu_schedule_balance_work();
1202 /* clear the areas and return address relative to base address */
1203 for_each_possible_cpu(cpu)
1204 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1206 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1207 kmemleak_alloc_percpu(ptr, size, gfp);
1209 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1210 chunk->base_addr, off, ptr);
1215 spin_unlock_irqrestore(&pcpu_lock, flags);
1217 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1219 if (!is_atomic && warn_limit) {
1220 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1221 size, align, is_atomic, err);
1224 pr_info("limit reached, disable warning\n");
1227 /* see the flag handling in pcpu_blance_workfn() */
1228 pcpu_atomic_alloc_failed = true;
1229 pcpu_schedule_balance_work();
1231 mutex_unlock(&pcpu_alloc_mutex);
1237 * __alloc_percpu_gfp - allocate dynamic percpu area
1238 * @size: size of area to allocate in bytes
1239 * @align: alignment of area (max PAGE_SIZE)
1240 * @gfp: allocation flags
1242 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1243 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1244 * be called from any context but is a lot more likely to fail.
1247 * Percpu pointer to the allocated area on success, NULL on failure.
1249 void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1251 return pcpu_alloc(size, align, false, gfp);
1253 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1256 * __alloc_percpu - allocate dynamic percpu area
1257 * @size: size of area to allocate in bytes
1258 * @align: alignment of area (max PAGE_SIZE)
1260 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1262 void __percpu *__alloc_percpu(size_t size, size_t align)
1264 return pcpu_alloc(size, align, false, GFP_KERNEL);
1266 EXPORT_SYMBOL_GPL(__alloc_percpu);
1269 * __alloc_reserved_percpu - allocate reserved percpu area
1270 * @size: size of area to allocate in bytes
1271 * @align: alignment of area (max PAGE_SIZE)
1273 * Allocate zero-filled percpu area of @size bytes aligned at @align
1274 * from reserved percpu area if arch has set it up; otherwise,
1275 * allocation is served from the same dynamic area. Might sleep.
1276 * Might trigger writeouts.
1279 * Does GFP_KERNEL allocation.
1282 * Percpu pointer to the allocated area on success, NULL on failure.
1284 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1286 return pcpu_alloc(size, align, true, GFP_KERNEL);
1290 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1293 * Reclaim all fully free chunks except for the first one.
1295 static void pcpu_balance_workfn(struct work_struct *work)
1298 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1299 struct pcpu_chunk *chunk, *next;
1300 int slot, nr_to_pop, ret;
1303 * There's no reason to keep around multiple unused chunks and VM
1304 * areas can be scarce. Destroy all free chunks except for one.
1306 mutex_lock(&pcpu_alloc_mutex);
1307 spin_lock_irq(&pcpu_lock);
1309 list_for_each_entry_safe(chunk, next, free_head, list) {
1310 WARN_ON(chunk->immutable);
1312 /* spare the first one */
1313 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1316 list_move(&chunk->list, &to_free);
1319 spin_unlock_irq(&pcpu_lock);
1321 list_for_each_entry_safe(chunk, next, &to_free, list) {
1324 pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1326 pcpu_depopulate_chunk(chunk, rs, re);
1327 spin_lock_irq(&pcpu_lock);
1328 pcpu_chunk_depopulated(chunk, rs, re);
1329 spin_unlock_irq(&pcpu_lock);
1331 pcpu_destroy_chunk(chunk);
1335 * Ensure there are certain number of free populated pages for
1336 * atomic allocs. Fill up from the most packed so that atomic
1337 * allocs don't increase fragmentation. If atomic allocation
1338 * failed previously, always populate the maximum amount. This
1339 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1340 * failing indefinitely; however, large atomic allocs are not
1341 * something we support properly and can be highly unreliable and
1345 if (pcpu_atomic_alloc_failed) {
1346 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1347 /* best effort anyway, don't worry about synchronization */
1348 pcpu_atomic_alloc_failed = false;
1350 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1351 pcpu_nr_empty_pop_pages,
1352 0, PCPU_EMPTY_POP_PAGES_HIGH);
1355 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1356 int nr_unpop = 0, rs, re;
1361 spin_lock_irq(&pcpu_lock);
1362 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1363 nr_unpop = chunk->nr_pages - chunk->nr_populated;
1367 spin_unlock_irq(&pcpu_lock);
1372 /* @chunk can't go away while pcpu_alloc_mutex is held */
1373 pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1375 int nr = min(re - rs, nr_to_pop);
1377 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1380 spin_lock_irq(&pcpu_lock);
1381 pcpu_chunk_populated(chunk, rs, rs + nr, false);
1382 spin_unlock_irq(&pcpu_lock);
1393 /* ran out of chunks to populate, create a new one and retry */
1394 chunk = pcpu_create_chunk();
1396 spin_lock_irq(&pcpu_lock);
1397 pcpu_chunk_relocate(chunk, -1);
1398 spin_unlock_irq(&pcpu_lock);
1403 mutex_unlock(&pcpu_alloc_mutex);
1407 * free_percpu - free percpu area
1408 * @ptr: pointer to area to free
1410 * Free percpu area @ptr.
1413 * Can be called from atomic context.
1415 void free_percpu(void __percpu *ptr)
1418 struct pcpu_chunk *chunk;
1419 unsigned long flags;
1425 kmemleak_free_percpu(ptr);
1427 addr = __pcpu_ptr_to_addr(ptr);
1429 spin_lock_irqsave(&pcpu_lock, flags);
1431 chunk = pcpu_chunk_addr_search(addr);
1432 off = addr - chunk->base_addr;
1434 pcpu_free_area(chunk, off);
1436 /* if there are more than one fully free chunks, wake up grim reaper */
1437 if (chunk->free_bytes == pcpu_unit_size) {
1438 struct pcpu_chunk *pos;
1440 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1442 pcpu_schedule_balance_work();
1447 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1449 spin_unlock_irqrestore(&pcpu_lock, flags);
1451 EXPORT_SYMBOL_GPL(free_percpu);
1453 bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1456 const size_t static_size = __per_cpu_end - __per_cpu_start;
1457 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1460 for_each_possible_cpu(cpu) {
1461 void *start = per_cpu_ptr(base, cpu);
1462 void *va = (void *)addr;
1464 if (va >= start && va < start + static_size) {
1466 *can_addr = (unsigned long) (va - start);
1467 *can_addr += (unsigned long)
1468 per_cpu_ptr(base, get_boot_cpu_id());
1474 /* on UP, can't distinguish from other static vars, always false */
1479 * is_kernel_percpu_address - test whether address is from static percpu area
1480 * @addr: address to test
1482 * Test whether @addr belongs to in-kernel static percpu area. Module
1483 * static percpu areas are not considered. For those, use
1484 * is_module_percpu_address().
1487 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1489 bool is_kernel_percpu_address(unsigned long addr)
1491 return __is_kernel_percpu_address(addr, NULL);
1495 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1496 * @addr: the address to be converted to physical address
1498 * Given @addr which is dereferenceable address obtained via one of
1499 * percpu access macros, this function translates it into its physical
1500 * address. The caller is responsible for ensuring @addr stays valid
1501 * until this function finishes.
1503 * percpu allocator has special setup for the first chunk, which currently
1504 * supports either embedding in linear address space or vmalloc mapping,
1505 * and, from the second one, the backing allocator (currently either vm or
1506 * km) provides translation.
1508 * The addr can be translated simply without checking if it falls into the
1509 * first chunk. But the current code reflects better how percpu allocator
1510 * actually works, and the verification can discover both bugs in percpu
1511 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1515 * The physical address for @addr.
1517 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1519 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1520 bool in_first_chunk = false;
1521 unsigned long first_low, first_high;
1525 * The following test on unit_low/high isn't strictly
1526 * necessary but will speed up lookups of addresses which
1527 * aren't in the first chunk.
1529 * The address check is against full chunk sizes. pcpu_base_addr
1530 * points to the beginning of the first chunk including the
1531 * static region. Assumes good intent as the first chunk may
1532 * not be full (ie. < pcpu_unit_pages in size).
1534 first_low = (unsigned long)pcpu_base_addr +
1535 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
1536 first_high = (unsigned long)pcpu_base_addr +
1537 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
1538 if ((unsigned long)addr >= first_low &&
1539 (unsigned long)addr < first_high) {
1540 for_each_possible_cpu(cpu) {
1541 void *start = per_cpu_ptr(base, cpu);
1543 if (addr >= start && addr < start + pcpu_unit_size) {
1544 in_first_chunk = true;
1550 if (in_first_chunk) {
1551 if (!is_vmalloc_addr(addr))
1554 return page_to_phys(vmalloc_to_page(addr)) +
1555 offset_in_page(addr);
1557 return page_to_phys(pcpu_addr_to_page(addr)) +
1558 offset_in_page(addr);
1562 * pcpu_alloc_alloc_info - allocate percpu allocation info
1563 * @nr_groups: the number of groups
1564 * @nr_units: the number of units
1566 * Allocate ai which is large enough for @nr_groups groups containing
1567 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1568 * cpu_map array which is long enough for @nr_units and filled with
1569 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1570 * pointer of other groups.
1573 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1576 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1579 struct pcpu_alloc_info *ai;
1580 size_t base_size, ai_size;
1584 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1585 __alignof__(ai->groups[0].cpu_map[0]));
1586 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1588 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1594 ai->groups[0].cpu_map = ptr;
1596 for (unit = 0; unit < nr_units; unit++)
1597 ai->groups[0].cpu_map[unit] = NR_CPUS;
1599 ai->nr_groups = nr_groups;
1600 ai->__ai_size = PFN_ALIGN(ai_size);
1606 * pcpu_free_alloc_info - free percpu allocation info
1607 * @ai: pcpu_alloc_info to free
1609 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1611 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1613 memblock_free_early(__pa(ai), ai->__ai_size);
1617 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1619 * @ai: allocation info to dump
1621 * Print out information about @ai using loglevel @lvl.
1623 static void pcpu_dump_alloc_info(const char *lvl,
1624 const struct pcpu_alloc_info *ai)
1626 int group_width = 1, cpu_width = 1, width;
1627 char empty_str[] = "--------";
1628 int alloc = 0, alloc_end = 0;
1630 int upa, apl; /* units per alloc, allocs per line */
1636 v = num_possible_cpus();
1639 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1641 upa = ai->alloc_size / ai->unit_size;
1642 width = upa * (cpu_width + 1) + group_width + 3;
1643 apl = rounddown_pow_of_two(max(60 / width, 1));
1645 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1646 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1647 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1649 for (group = 0; group < ai->nr_groups; group++) {
1650 const struct pcpu_group_info *gi = &ai->groups[group];
1651 int unit = 0, unit_end = 0;
1653 BUG_ON(gi->nr_units % upa);
1654 for (alloc_end += gi->nr_units / upa;
1655 alloc < alloc_end; alloc++) {
1656 if (!(alloc % apl)) {
1658 printk("%spcpu-alloc: ", lvl);
1660 pr_cont("[%0*d] ", group_width, group);
1662 for (unit_end += upa; unit < unit_end; unit++)
1663 if (gi->cpu_map[unit] != NR_CPUS)
1665 cpu_width, gi->cpu_map[unit]);
1667 pr_cont("%s ", empty_str);
1674 * pcpu_setup_first_chunk - initialize the first percpu chunk
1675 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1676 * @base_addr: mapped address
1678 * Initialize the first percpu chunk which contains the kernel static
1679 * perpcu area. This function is to be called from arch percpu area
1682 * @ai contains all information necessary to initialize the first
1683 * chunk and prime the dynamic percpu allocator.
1685 * @ai->static_size is the size of static percpu area.
1687 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1688 * reserve after the static area in the first chunk. This reserves
1689 * the first chunk such that it's available only through reserved
1690 * percpu allocation. This is primarily used to serve module percpu
1691 * static areas on architectures where the addressing model has
1692 * limited offset range for symbol relocations to guarantee module
1693 * percpu symbols fall inside the relocatable range.
1695 * @ai->dyn_size determines the number of bytes available for dynamic
1696 * allocation in the first chunk. The area between @ai->static_size +
1697 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1699 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1700 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1703 * @ai->atom_size is the allocation atom size and used as alignment
1706 * @ai->alloc_size is the allocation size and always multiple of
1707 * @ai->atom_size. This is larger than @ai->atom_size if
1708 * @ai->unit_size is larger than @ai->atom_size.
1710 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1711 * percpu areas. Units which should be colocated are put into the
1712 * same group. Dynamic VM areas will be allocated according to these
1713 * groupings. If @ai->nr_groups is zero, a single group containing
1714 * all units is assumed.
1716 * The caller should have mapped the first chunk at @base_addr and
1717 * copied static data to each unit.
1719 * The first chunk will always contain a static and a dynamic region.
1720 * However, the static region is not managed by any chunk. If the first
1721 * chunk also contains a reserved region, it is served by two chunks -
1722 * one for the reserved region and one for the dynamic region. They
1723 * share the same vm, but use offset regions in the area allocation map.
1724 * The chunk serving the dynamic region is circulated in the chunk slots
1725 * and available for dynamic allocation like any other chunk.
1728 * 0 on success, -errno on failure.
1730 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1733 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1734 size_t static_size, dyn_size;
1735 struct pcpu_chunk *chunk;
1736 unsigned long *group_offsets;
1737 size_t *group_sizes;
1738 unsigned long *unit_off;
1743 unsigned long tmp_addr;
1745 #define PCPU_SETUP_BUG_ON(cond) do { \
1746 if (unlikely(cond)) { \
1747 pr_emerg("failed to initialize, %s\n", #cond); \
1748 pr_emerg("cpu_possible_mask=%*pb\n", \
1749 cpumask_pr_args(cpu_possible_mask)); \
1750 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1756 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1758 PCPU_SETUP_BUG_ON(!ai->static_size);
1759 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1761 PCPU_SETUP_BUG_ON(!base_addr);
1762 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1763 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1764 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1765 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1766 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
1767 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1768 PCPU_SETUP_BUG_ON(!ai->dyn_size);
1769 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
1770 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
1771 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
1772 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1774 /* process group information and build config tables accordingly */
1775 group_offsets = memblock_virt_alloc(ai->nr_groups *
1776 sizeof(group_offsets[0]), 0);
1777 group_sizes = memblock_virt_alloc(ai->nr_groups *
1778 sizeof(group_sizes[0]), 0);
1779 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1780 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1782 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1783 unit_map[cpu] = UINT_MAX;
1785 pcpu_low_unit_cpu = NR_CPUS;
1786 pcpu_high_unit_cpu = NR_CPUS;
1788 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1789 const struct pcpu_group_info *gi = &ai->groups[group];
1791 group_offsets[group] = gi->base_offset;
1792 group_sizes[group] = gi->nr_units * ai->unit_size;
1794 for (i = 0; i < gi->nr_units; i++) {
1795 cpu = gi->cpu_map[i];
1799 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1800 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1801 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1803 unit_map[cpu] = unit + i;
1804 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1806 /* determine low/high unit_cpu */
1807 if (pcpu_low_unit_cpu == NR_CPUS ||
1808 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1809 pcpu_low_unit_cpu = cpu;
1810 if (pcpu_high_unit_cpu == NR_CPUS ||
1811 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1812 pcpu_high_unit_cpu = cpu;
1815 pcpu_nr_units = unit;
1817 for_each_possible_cpu(cpu)
1818 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1820 /* we're done parsing the input, undefine BUG macro and dump config */
1821 #undef PCPU_SETUP_BUG_ON
1822 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1824 pcpu_nr_groups = ai->nr_groups;
1825 pcpu_group_offsets = group_offsets;
1826 pcpu_group_sizes = group_sizes;
1827 pcpu_unit_map = unit_map;
1828 pcpu_unit_offsets = unit_off;
1830 /* determine basic parameters */
1831 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1832 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1833 pcpu_atom_size = ai->atom_size;
1834 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1835 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1837 pcpu_stats_save_ai(ai);
1840 * Allocate chunk slots. The additional last slot is for
1843 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1844 pcpu_slot = memblock_virt_alloc(
1845 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1846 for (i = 0; i < pcpu_nr_slots; i++)
1847 INIT_LIST_HEAD(&pcpu_slot[i]);
1850 * The end of the static region needs to be aligned with the
1851 * minimum allocation size as this offsets the reserved and
1852 * dynamic region. The first chunk ends page aligned by
1853 * expanding the dynamic region, therefore the dynamic region
1854 * can be shrunk to compensate while still staying above the
1857 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
1858 dyn_size = ai->dyn_size - (static_size - ai->static_size);
1861 * Initialize first chunk.
1862 * If the reserved_size is non-zero, this initializes the reserved
1863 * chunk. If the reserved_size is zero, the reserved chunk is NULL
1864 * and the dynamic region is initialized here. The first chunk,
1865 * pcpu_first_chunk, will always point to the chunk that serves
1866 * the dynamic region.
1868 tmp_addr = (unsigned long)base_addr + static_size;
1869 map_size = ai->reserved_size ?: dyn_size;
1870 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
1872 /* init dynamic chunk if necessary */
1873 if (ai->reserved_size) {
1874 pcpu_reserved_chunk = chunk;
1876 tmp_addr = (unsigned long)base_addr + static_size +
1878 map_size = dyn_size;
1879 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
1882 /* link the first chunk in */
1883 pcpu_first_chunk = chunk;
1884 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
1885 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1887 pcpu_stats_chunk_alloc();
1888 trace_percpu_create_chunk(base_addr);
1891 pcpu_base_addr = base_addr;
1897 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1898 [PCPU_FC_AUTO] = "auto",
1899 [PCPU_FC_EMBED] = "embed",
1900 [PCPU_FC_PAGE] = "page",
1903 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1905 static int __init percpu_alloc_setup(char *str)
1912 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1913 else if (!strcmp(str, "embed"))
1914 pcpu_chosen_fc = PCPU_FC_EMBED;
1916 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1917 else if (!strcmp(str, "page"))
1918 pcpu_chosen_fc = PCPU_FC_PAGE;
1921 pr_warn("unknown allocator %s specified\n", str);
1925 early_param("percpu_alloc", percpu_alloc_setup);
1928 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1929 * Build it if needed by the arch config or the generic setup is going
1932 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1933 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1934 #define BUILD_EMBED_FIRST_CHUNK
1937 /* build pcpu_page_first_chunk() iff needed by the arch config */
1938 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1939 #define BUILD_PAGE_FIRST_CHUNK
1942 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1943 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1945 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1946 * @reserved_size: the size of reserved percpu area in bytes
1947 * @dyn_size: minimum free size for dynamic allocation in bytes
1948 * @atom_size: allocation atom size
1949 * @cpu_distance_fn: callback to determine distance between cpus, optional
1951 * This function determines grouping of units, their mappings to cpus
1952 * and other parameters considering needed percpu size, allocation
1953 * atom size and distances between CPUs.
1955 * Groups are always multiples of atom size and CPUs which are of
1956 * LOCAL_DISTANCE both ways are grouped together and share space for
1957 * units in the same group. The returned configuration is guaranteed
1958 * to have CPUs on different nodes on different groups and >=75% usage
1959 * of allocated virtual address space.
1962 * On success, pointer to the new allocation_info is returned. On
1963 * failure, ERR_PTR value is returned.
1965 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1966 size_t reserved_size, size_t dyn_size,
1968 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1970 static int group_map[NR_CPUS] __initdata;
1971 static int group_cnt[NR_CPUS] __initdata;
1972 const size_t static_size = __per_cpu_end - __per_cpu_start;
1973 int nr_groups = 1, nr_units = 0;
1974 size_t size_sum, min_unit_size, alloc_size;
1975 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1976 int last_allocs, group, unit;
1977 unsigned int cpu, tcpu;
1978 struct pcpu_alloc_info *ai;
1979 unsigned int *cpu_map;
1981 /* this function may be called multiple times */
1982 memset(group_map, 0, sizeof(group_map));
1983 memset(group_cnt, 0, sizeof(group_cnt));
1985 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1986 size_sum = PFN_ALIGN(static_size + reserved_size +
1987 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1988 dyn_size = size_sum - static_size - reserved_size;
1991 * Determine min_unit_size, alloc_size and max_upa such that
1992 * alloc_size is multiple of atom_size and is the smallest
1993 * which can accommodate 4k aligned segments which are equal to
1994 * or larger than min_unit_size.
1996 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1998 /* determine the maximum # of units that can fit in an allocation */
1999 alloc_size = roundup(min_unit_size, atom_size);
2000 upa = alloc_size / min_unit_size;
2001 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2005 /* group cpus according to their proximity */
2006 for_each_possible_cpu(cpu) {
2009 for_each_possible_cpu(tcpu) {
2012 if (group_map[tcpu] == group && cpu_distance_fn &&
2013 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2014 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2016 nr_groups = max(nr_groups, group + 1);
2020 group_map[cpu] = group;
2025 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2026 * Expand the unit_size until we use >= 75% of the units allocated.
2027 * Related to atom_size, which could be much larger than the unit_size.
2029 last_allocs = INT_MAX;
2030 for (upa = max_upa; upa; upa--) {
2031 int allocs = 0, wasted = 0;
2033 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2036 for (group = 0; group < nr_groups; group++) {
2037 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2038 allocs += this_allocs;
2039 wasted += this_allocs * upa - group_cnt[group];
2043 * Don't accept if wastage is over 1/3. The
2044 * greater-than comparison ensures upa==1 always
2045 * passes the following check.
2047 if (wasted > num_possible_cpus() / 3)
2050 /* and then don't consume more memory */
2051 if (allocs > last_allocs)
2053 last_allocs = allocs;
2058 /* allocate and fill alloc_info */
2059 for (group = 0; group < nr_groups; group++)
2060 nr_units += roundup(group_cnt[group], upa);
2062 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2064 return ERR_PTR(-ENOMEM);
2065 cpu_map = ai->groups[0].cpu_map;
2067 for (group = 0; group < nr_groups; group++) {
2068 ai->groups[group].cpu_map = cpu_map;
2069 cpu_map += roundup(group_cnt[group], upa);
2072 ai->static_size = static_size;
2073 ai->reserved_size = reserved_size;
2074 ai->dyn_size = dyn_size;
2075 ai->unit_size = alloc_size / upa;
2076 ai->atom_size = atom_size;
2077 ai->alloc_size = alloc_size;
2079 for (group = 0, unit = 0; group_cnt[group]; group++) {
2080 struct pcpu_group_info *gi = &ai->groups[group];
2083 * Initialize base_offset as if all groups are located
2084 * back-to-back. The caller should update this to
2085 * reflect actual allocation.
2087 gi->base_offset = unit * ai->unit_size;
2089 for_each_possible_cpu(cpu)
2090 if (group_map[cpu] == group)
2091 gi->cpu_map[gi->nr_units++] = cpu;
2092 gi->nr_units = roundup(gi->nr_units, upa);
2093 unit += gi->nr_units;
2095 BUG_ON(unit != nr_units);
2099 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2101 #if defined(BUILD_EMBED_FIRST_CHUNK)
2103 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2104 * @reserved_size: the size of reserved percpu area in bytes
2105 * @dyn_size: minimum free size for dynamic allocation in bytes
2106 * @atom_size: allocation atom size
2107 * @cpu_distance_fn: callback to determine distance between cpus, optional
2108 * @alloc_fn: function to allocate percpu page
2109 * @free_fn: function to free percpu page
2111 * This is a helper to ease setting up embedded first percpu chunk and
2112 * can be called where pcpu_setup_first_chunk() is expected.
2114 * If this function is used to setup the first chunk, it is allocated
2115 * by calling @alloc_fn and used as-is without being mapped into
2116 * vmalloc area. Allocations are always whole multiples of @atom_size
2117 * aligned to @atom_size.
2119 * This enables the first chunk to piggy back on the linear physical
2120 * mapping which often uses larger page size. Please note that this
2121 * can result in very sparse cpu->unit mapping on NUMA machines thus
2122 * requiring large vmalloc address space. Don't use this allocator if
2123 * vmalloc space is not orders of magnitude larger than distances
2124 * between node memory addresses (ie. 32bit NUMA machines).
2126 * @dyn_size specifies the minimum dynamic area size.
2128 * If the needed size is smaller than the minimum or specified unit
2129 * size, the leftover is returned using @free_fn.
2132 * 0 on success, -errno on failure.
2134 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2136 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2137 pcpu_fc_alloc_fn_t alloc_fn,
2138 pcpu_fc_free_fn_t free_fn)
2140 void *base = (void *)ULONG_MAX;
2141 void **areas = NULL;
2142 struct pcpu_alloc_info *ai;
2143 size_t size_sum, areas_size;
2144 unsigned long max_distance;
2145 int group, i, highest_group, rc;
2147 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2152 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2153 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2155 areas = memblock_virt_alloc_nopanic(areas_size, 0);
2161 /* allocate, copy and determine base address & max_distance */
2163 for (group = 0; group < ai->nr_groups; group++) {
2164 struct pcpu_group_info *gi = &ai->groups[group];
2165 unsigned int cpu = NR_CPUS;
2168 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2169 cpu = gi->cpu_map[i];
2170 BUG_ON(cpu == NR_CPUS);
2172 /* allocate space for the whole group */
2173 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2176 goto out_free_areas;
2178 /* kmemleak tracks the percpu allocations separately */
2182 base = min(ptr, base);
2183 if (ptr > areas[highest_group])
2184 highest_group = group;
2186 max_distance = areas[highest_group] - base;
2187 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2189 /* warn if maximum distance is further than 75% of vmalloc space */
2190 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2191 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2192 max_distance, VMALLOC_TOTAL);
2193 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2194 /* and fail if we have fallback */
2196 goto out_free_areas;
2201 * Copy data and free unused parts. This should happen after all
2202 * allocations are complete; otherwise, we may end up with
2203 * overlapping groups.
2205 for (group = 0; group < ai->nr_groups; group++) {
2206 struct pcpu_group_info *gi = &ai->groups[group];
2207 void *ptr = areas[group];
2209 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2210 if (gi->cpu_map[i] == NR_CPUS) {
2211 /* unused unit, free whole */
2212 free_fn(ptr, ai->unit_size);
2215 /* copy and return the unused part */
2216 memcpy(ptr, __per_cpu_load, ai->static_size);
2217 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2221 /* base address is now known, determine group base offsets */
2222 for (group = 0; group < ai->nr_groups; group++) {
2223 ai->groups[group].base_offset = areas[group] - base;
2226 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2227 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2228 ai->dyn_size, ai->unit_size);
2230 rc = pcpu_setup_first_chunk(ai, base);
2234 for (group = 0; group < ai->nr_groups; group++)
2236 free_fn(areas[group],
2237 ai->groups[group].nr_units * ai->unit_size);
2239 pcpu_free_alloc_info(ai);
2241 memblock_free_early(__pa(areas), areas_size);
2244 #endif /* BUILD_EMBED_FIRST_CHUNK */
2246 #ifdef BUILD_PAGE_FIRST_CHUNK
2248 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2249 * @reserved_size: the size of reserved percpu area in bytes
2250 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2251 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2252 * @populate_pte_fn: function to populate pte
2254 * This is a helper to ease setting up page-remapped first percpu
2255 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2257 * This is the basic allocator. Static percpu area is allocated
2258 * page-by-page into vmalloc area.
2261 * 0 on success, -errno on failure.
2263 int __init pcpu_page_first_chunk(size_t reserved_size,
2264 pcpu_fc_alloc_fn_t alloc_fn,
2265 pcpu_fc_free_fn_t free_fn,
2266 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2268 static struct vm_struct vm;
2269 struct pcpu_alloc_info *ai;
2273 struct page **pages;
2278 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2280 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2283 BUG_ON(ai->nr_groups != 1);
2284 upa = ai->alloc_size/ai->unit_size;
2285 nr_g0_units = roundup(num_possible_cpus(), upa);
2286 if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2287 pcpu_free_alloc_info(ai);
2291 unit_pages = ai->unit_size >> PAGE_SHIFT;
2293 /* unaligned allocations can't be freed, round up to page size */
2294 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2296 pages = memblock_virt_alloc(pages_size, 0);
2298 /* allocate pages */
2300 for (unit = 0; unit < num_possible_cpus(); unit++) {
2301 unsigned int cpu = ai->groups[0].cpu_map[unit];
2302 for (i = 0; i < unit_pages; i++) {
2305 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2307 pr_warn("failed to allocate %s page for cpu%u\n",
2311 /* kmemleak tracks the percpu allocations separately */
2313 pages[j++] = virt_to_page(ptr);
2317 /* allocate vm area, map the pages and copy static data */
2318 vm.flags = VM_ALLOC;
2319 vm.size = num_possible_cpus() * ai->unit_size;
2320 vm_area_register_early(&vm, PAGE_SIZE);
2322 for (unit = 0; unit < num_possible_cpus(); unit++) {
2323 unsigned long unit_addr =
2324 (unsigned long)vm.addr + unit * ai->unit_size;
2326 for (i = 0; i < unit_pages; i++)
2327 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2329 /* pte already populated, the following shouldn't fail */
2330 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2333 panic("failed to map percpu area, err=%d\n", rc);
2336 * FIXME: Archs with virtual cache should flush local
2337 * cache for the linear mapping here - something
2338 * equivalent to flush_cache_vmap() on the local cpu.
2339 * flush_cache_vmap() can't be used as most supporting
2340 * data structures are not set up yet.
2343 /* copy static data */
2344 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2347 /* we're ready, commit */
2348 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2349 unit_pages, psize_str, vm.addr, ai->static_size,
2350 ai->reserved_size, ai->dyn_size);
2352 rc = pcpu_setup_first_chunk(ai, vm.addr);
2357 free_fn(page_address(pages[j]), PAGE_SIZE);
2360 memblock_free_early(__pa(pages), pages_size);
2361 pcpu_free_alloc_info(ai);
2364 #endif /* BUILD_PAGE_FIRST_CHUNK */
2366 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2368 * Generic SMP percpu area setup.
2370 * The embedding helper is used because its behavior closely resembles
2371 * the original non-dynamic generic percpu area setup. This is
2372 * important because many archs have addressing restrictions and might
2373 * fail if the percpu area is located far away from the previous
2374 * location. As an added bonus, in non-NUMA cases, embedding is
2375 * generally a good idea TLB-wise because percpu area can piggy back
2376 * on the physical linear memory mapping which uses large page
2377 * mappings on applicable archs.
2379 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2380 EXPORT_SYMBOL(__per_cpu_offset);
2382 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2385 return memblock_virt_alloc_from_nopanic(
2386 size, align, __pa(MAX_DMA_ADDRESS));
2389 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2391 memblock_free_early(__pa(ptr), size);
2394 void __init setup_per_cpu_areas(void)
2396 unsigned long delta;
2401 * Always reserve area for module percpu variables. That's
2402 * what the legacy allocator did.
2404 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2405 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2406 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2408 panic("Failed to initialize percpu areas.");
2410 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2411 for_each_possible_cpu(cpu)
2412 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2414 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2416 #else /* CONFIG_SMP */
2419 * UP percpu area setup.
2421 * UP always uses km-based percpu allocator with identity mapping.
2422 * Static percpu variables are indistinguishable from the usual static
2423 * variables and don't require any special preparation.
2425 void __init setup_per_cpu_areas(void)
2427 const size_t unit_size =
2428 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2429 PERCPU_DYNAMIC_RESERVE));
2430 struct pcpu_alloc_info *ai;
2433 ai = pcpu_alloc_alloc_info(1, 1);
2434 fc = memblock_virt_alloc_from_nopanic(unit_size,
2436 __pa(MAX_DMA_ADDRESS));
2438 panic("Failed to allocate memory for percpu areas.");
2439 /* kmemleak tracks the percpu allocations separately */
2442 ai->dyn_size = unit_size;
2443 ai->unit_size = unit_size;
2444 ai->atom_size = unit_size;
2445 ai->alloc_size = unit_size;
2446 ai->groups[0].nr_units = 1;
2447 ai->groups[0].cpu_map[0] = 0;
2449 if (pcpu_setup_first_chunk(ai, fc) < 0)
2450 panic("Failed to initialize percpu areas.");
2453 #endif /* CONFIG_SMP */
2456 * Percpu allocator is initialized early during boot when neither slab or
2457 * workqueue is available. Plug async management until everything is up
2460 static int __init percpu_enable_async(void)
2462 pcpu_async_enabled = true;
2465 subsys_initcall(percpu_enable_async);