1 // SPDX-License-Identifier: GPL-2.0-only
3 * mm/percpu.c - percpu memory allocator
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
32 * <Static | [Reserved] | Dynamic>
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
39 * The allocator organizes chunks into lists according to free size and
40 * tries to allocate from the fullest chunk first. Each chunk is managed
41 * by a bitmap with metadata blocks. The allocation map is updated on
42 * every allocation and free to reflect the current state while the boundary
43 * map is only updated on allocation. Each metadata block contains
44 * information to help mitigate the need to iterate over large portions
45 * of the bitmap. The reverse mapping from page to chunk is stored in
46 * the page's index. Lastly, units are lazily backed and grow in unison.
48 * There is a unique conversion that goes on here between bytes and bits.
49 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
50 * tracks the number of pages it is responsible for in nr_pages. Helper
51 * functions are used to convert from between the bytes, bits, and blocks.
52 * All hints are managed in bits unless explicitly stated.
54 * To use this allocator, arch code should do the following:
56 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
57 * regular address to percpu pointer and back if they need to be
58 * different from the default
60 * - use pcpu_setup_first_chunk() during percpu area initialization to
61 * setup the first chunk containing the kernel static percpu area
64 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
66 #include <linux/bitmap.h>
67 #include <linux/memblock.h>
68 #include <linux/err.h>
69 #include <linux/lcm.h>
70 #include <linux/list.h>
71 #include <linux/log2.h>
73 #include <linux/module.h>
74 #include <linux/mutex.h>
75 #include <linux/percpu.h>
76 #include <linux/pfn.h>
77 #include <linux/slab.h>
78 #include <linux/spinlock.h>
79 #include <linux/vmalloc.h>
80 #include <linux/workqueue.h>
81 #include <linux/kmemleak.h>
82 #include <linux/sched.h>
83 #include <linux/sched/mm.h>
85 #include <asm/cacheflush.h>
86 #include <asm/sections.h>
87 #include <asm/tlbflush.h>
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/percpu.h>
93 #include "percpu-internal.h"
95 /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
96 #define PCPU_SLOT_BASE_SHIFT 5
97 /* chunks in slots below this are subject to being sidelined on failed alloc */
98 #define PCPU_SLOT_FAIL_THRESHOLD 3
100 #define PCPU_EMPTY_POP_PAGES_LOW 2
101 #define PCPU_EMPTY_POP_PAGES_HIGH 4
104 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
105 #ifndef __addr_to_pcpu_ptr
106 #define __addr_to_pcpu_ptr(addr) \
107 (void __percpu *)((unsigned long)(addr) - \
108 (unsigned long)pcpu_base_addr + \
109 (unsigned long)__per_cpu_start)
111 #ifndef __pcpu_ptr_to_addr
112 #define __pcpu_ptr_to_addr(ptr) \
113 (void __force *)((unsigned long)(ptr) + \
114 (unsigned long)pcpu_base_addr - \
115 (unsigned long)__per_cpu_start)
117 #else /* CONFIG_SMP */
118 /* on UP, it's always identity mapped */
119 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
120 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
121 #endif /* CONFIG_SMP */
123 static int pcpu_unit_pages __ro_after_init;
124 static int pcpu_unit_size __ro_after_init;
125 static int pcpu_nr_units __ro_after_init;
126 static int pcpu_atom_size __ro_after_init;
127 int pcpu_nr_slots __ro_after_init;
128 static size_t pcpu_chunk_struct_size __ro_after_init;
130 /* cpus with the lowest and highest unit addresses */
131 static unsigned int pcpu_low_unit_cpu __ro_after_init;
132 static unsigned int pcpu_high_unit_cpu __ro_after_init;
134 /* the address of the first chunk which starts with the kernel static area */
135 void *pcpu_base_addr __ro_after_init;
136 EXPORT_SYMBOL_GPL(pcpu_base_addr);
138 static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
139 const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
141 /* group information, used for vm allocation */
142 static int pcpu_nr_groups __ro_after_init;
143 static const unsigned long *pcpu_group_offsets __ro_after_init;
144 static const size_t *pcpu_group_sizes __ro_after_init;
147 * The first chunk which always exists. Note that unlike other
148 * chunks, this one can be allocated and mapped in several different
149 * ways and thus often doesn't live in the vmalloc area.
151 struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
154 * Optional reserved chunk. This chunk reserves part of the first
155 * chunk and serves it for reserved allocations. When the reserved
156 * region doesn't exist, the following variable is NULL.
158 struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
160 DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
161 static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
163 struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
165 /* chunks which need their map areas extended, protected by pcpu_lock */
166 static LIST_HEAD(pcpu_map_extend_chunks);
169 * The number of empty populated pages, protected by pcpu_lock. The
170 * reserved chunk doesn't contribute to the count.
172 int pcpu_nr_empty_pop_pages;
175 * The number of populated pages in use by the allocator, protected by
176 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
177 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
178 * and increments/decrements this count by 1).
180 static unsigned long pcpu_nr_populated;
183 * Balance work is used to populate or destroy chunks asynchronously. We
184 * try to keep the number of populated free pages between
185 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
188 static void pcpu_balance_workfn(struct work_struct *work);
189 static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
190 static bool pcpu_async_enabled __read_mostly;
191 static bool pcpu_atomic_alloc_failed;
193 static void pcpu_schedule_balance_work(void)
195 if (pcpu_async_enabled)
196 schedule_work(&pcpu_balance_work);
200 * pcpu_addr_in_chunk - check if the address is served from this chunk
201 * @chunk: chunk of interest
202 * @addr: percpu address
205 * True if the address is served from this chunk.
207 static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
209 void *start_addr, *end_addr;
214 start_addr = chunk->base_addr + chunk->start_offset;
215 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
218 return addr >= start_addr && addr < end_addr;
221 static int __pcpu_size_to_slot(int size)
223 int highbit = fls(size); /* size is in bytes */
224 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
227 static int pcpu_size_to_slot(int size)
229 if (size == pcpu_unit_size)
230 return pcpu_nr_slots - 1;
231 return __pcpu_size_to_slot(size);
234 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
236 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
238 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
239 chunk_md->contig_hint == 0)
242 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
245 /* set the pointer to a chunk in a page struct */
246 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
248 page->index = (unsigned long)pcpu;
251 /* obtain pointer to a chunk from a page struct */
252 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
254 return (struct pcpu_chunk *)page->index;
257 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
259 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
262 static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
264 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
267 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
268 unsigned int cpu, int page_idx)
270 return (unsigned long)chunk->base_addr +
271 pcpu_unit_page_offset(cpu, page_idx);
275 * The following are helper functions to help access bitmaps and convert
276 * between bitmap offsets to address offsets.
278 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
280 return chunk->alloc_map +
281 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
284 static unsigned long pcpu_off_to_block_index(int off)
286 return off / PCPU_BITMAP_BLOCK_BITS;
289 static unsigned long pcpu_off_to_block_off(int off)
291 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
294 static unsigned long pcpu_block_off_to_off(int index, int off)
296 return index * PCPU_BITMAP_BLOCK_BITS + off;
300 * pcpu_next_hint - determine which hint to use
301 * @block: block of interest
302 * @alloc_bits: size of allocation
304 * This determines if we should scan based on the scan_hint or first_free.
305 * In general, we want to scan from first_free to fulfill allocations by
306 * first fit. However, if we know a scan_hint at position scan_hint_start
307 * cannot fulfill an allocation, we can begin scanning from there knowing
308 * the contig_hint will be our fallback.
310 static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
313 * The three conditions below determine if we can skip past the
314 * scan_hint. First, does the scan hint exist. Second, is the
315 * contig_hint after the scan_hint (possibly not true iff
316 * contig_hint == scan_hint). Third, is the allocation request
317 * larger than the scan_hint.
319 if (block->scan_hint &&
320 block->contig_hint_start > block->scan_hint_start &&
321 alloc_bits > block->scan_hint)
322 return block->scan_hint_start + block->scan_hint;
324 return block->first_free;
328 * pcpu_next_md_free_region - finds the next hint free area
329 * @chunk: chunk of interest
330 * @bit_off: chunk offset
331 * @bits: size of free area
333 * Helper function for pcpu_for_each_md_free_region. It checks
334 * block->contig_hint and performs aggregation across blocks to find the
335 * next hint. It modifies bit_off and bits in-place to be consumed in the
338 static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
341 int i = pcpu_off_to_block_index(*bit_off);
342 int block_off = pcpu_off_to_block_off(*bit_off);
343 struct pcpu_block_md *block;
346 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
348 /* handles contig area across blocks */
350 *bits += block->left_free;
351 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
357 * This checks three things. First is there a contig_hint to
358 * check. Second, have we checked this hint before by
359 * comparing the block_off. Third, is this the same as the
360 * right contig hint. In the last case, it spills over into
361 * the next block and should be handled by the contig area
362 * across blocks code.
364 *bits = block->contig_hint;
365 if (*bits && block->contig_hint_start >= block_off &&
366 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
367 *bit_off = pcpu_block_off_to_off(i,
368 block->contig_hint_start);
371 /* reset to satisfy the second predicate above */
374 *bits = block->right_free;
375 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
380 * pcpu_next_fit_region - finds fit areas for a given allocation request
381 * @chunk: chunk of interest
382 * @alloc_bits: size of allocation
383 * @align: alignment of area (max PAGE_SIZE)
384 * @bit_off: chunk offset
385 * @bits: size of free area
387 * Finds the next free region that is viable for use with a given size and
388 * alignment. This only returns if there is a valid area to be used for this
389 * allocation. block->first_free is returned if the allocation request fits
390 * within the block to see if the request can be fulfilled prior to the contig
393 static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
394 int align, int *bit_off, int *bits)
396 int i = pcpu_off_to_block_index(*bit_off);
397 int block_off = pcpu_off_to_block_off(*bit_off);
398 struct pcpu_block_md *block;
401 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
403 /* handles contig area across blocks */
405 *bits += block->left_free;
406 if (*bits >= alloc_bits)
408 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
412 /* check block->contig_hint */
413 *bits = ALIGN(block->contig_hint_start, align) -
414 block->contig_hint_start;
416 * This uses the block offset to determine if this has been
417 * checked in the prior iteration.
419 if (block->contig_hint &&
420 block->contig_hint_start >= block_off &&
421 block->contig_hint >= *bits + alloc_bits) {
422 int start = pcpu_next_hint(block, alloc_bits);
424 *bits += alloc_bits + block->contig_hint_start -
426 *bit_off = pcpu_block_off_to_off(i, start);
429 /* reset to satisfy the second predicate above */
432 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
434 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
435 *bit_off = pcpu_block_off_to_off(i, *bit_off);
436 if (*bits >= alloc_bits)
440 /* no valid offsets were found - fail condition */
441 *bit_off = pcpu_chunk_map_bits(chunk);
445 * Metadata free area iterators. These perform aggregation of free areas
446 * based on the metadata blocks and return the offset @bit_off and size in
447 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
448 * a fit is found for the allocation request.
450 #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
451 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
452 (bit_off) < pcpu_chunk_map_bits((chunk)); \
453 (bit_off) += (bits) + 1, \
454 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
456 #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
457 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
459 (bit_off) < pcpu_chunk_map_bits((chunk)); \
460 (bit_off) += (bits), \
461 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
465 * pcpu_mem_zalloc - allocate memory
466 * @size: bytes to allocate
467 * @gfp: allocation flags
469 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
470 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
471 * This is to facilitate passing through whitelisted flags. The
472 * returned memory is always zeroed.
475 * Pointer to the allocated area on success, NULL on failure.
477 static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
479 if (WARN_ON_ONCE(!slab_is_available()))
482 if (size <= PAGE_SIZE)
483 return kzalloc(size, gfp);
485 return __vmalloc(size, gfp | __GFP_ZERO);
489 * pcpu_mem_free - free memory
490 * @ptr: memory to free
492 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
494 static void pcpu_mem_free(void *ptr)
499 static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
502 if (chunk != pcpu_reserved_chunk) {
504 list_move(&chunk->list, &pcpu_slot[slot]);
506 list_move_tail(&chunk->list, &pcpu_slot[slot]);
510 static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
512 __pcpu_chunk_move(chunk, slot, true);
516 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
517 * @chunk: chunk of interest
518 * @oslot: the previous slot it was on
520 * This function is called after an allocation or free changed @chunk.
521 * New slot according to the changed state is determined and @chunk is
522 * moved to the slot. Note that the reserved chunk is never put on
528 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
530 int nslot = pcpu_chunk_slot(chunk);
533 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
537 * pcpu_update_empty_pages - update empty page counters
538 * @chunk: chunk of interest
539 * @nr: nr of empty pages
541 * This is used to keep track of the empty pages now based on the premise
542 * a md_block covers a page. The hint update functions recognize if a block
543 * is made full or broken to calculate deltas for keeping track of free pages.
545 static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
547 chunk->nr_empty_pop_pages += nr;
548 if (chunk != pcpu_reserved_chunk)
549 pcpu_nr_empty_pop_pages += nr;
553 * pcpu_region_overlap - determines if two regions overlap
554 * @a: start of first region, inclusive
555 * @b: end of first region, exclusive
556 * @x: start of second region, inclusive
557 * @y: end of second region, exclusive
559 * This is used to determine if the hint region [a, b) overlaps with the
560 * allocated region [x, y).
562 static inline bool pcpu_region_overlap(int a, int b, int x, int y)
564 return (a < y) && (x < b);
568 * pcpu_block_update - updates a block given a free area
569 * @block: block of interest
570 * @start: start offset in block
571 * @end: end offset in block
573 * Updates a block given a known free area. The region [start, end) is
574 * expected to be the entirety of the free area within a block. Chooses
575 * the best starting offset if the contig hints are equal.
577 static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
579 int contig = end - start;
581 block->first_free = min(block->first_free, start);
583 block->left_free = contig;
585 if (end == block->nr_bits)
586 block->right_free = contig;
588 if (contig > block->contig_hint) {
589 /* promote the old contig_hint to be the new scan_hint */
590 if (start > block->contig_hint_start) {
591 if (block->contig_hint > block->scan_hint) {
592 block->scan_hint_start =
593 block->contig_hint_start;
594 block->scan_hint = block->contig_hint;
595 } else if (start < block->scan_hint_start) {
597 * The old contig_hint == scan_hint. But, the
598 * new contig is larger so hold the invariant
599 * scan_hint_start < contig_hint_start.
601 block->scan_hint = 0;
604 block->scan_hint = 0;
606 block->contig_hint_start = start;
607 block->contig_hint = contig;
608 } else if (contig == block->contig_hint) {
609 if (block->contig_hint_start &&
611 __ffs(start) > __ffs(block->contig_hint_start))) {
612 /* start has a better alignment so use it */
613 block->contig_hint_start = start;
614 if (start < block->scan_hint_start &&
615 block->contig_hint > block->scan_hint)
616 block->scan_hint = 0;
617 } else if (start > block->scan_hint_start ||
618 block->contig_hint > block->scan_hint) {
620 * Knowing contig == contig_hint, update the scan_hint
621 * if it is farther than or larger than the current
624 block->scan_hint_start = start;
625 block->scan_hint = contig;
629 * The region is smaller than the contig_hint. So only update
630 * the scan_hint if it is larger than or equal and farther than
631 * the current scan_hint.
633 if ((start < block->contig_hint_start &&
634 (contig > block->scan_hint ||
635 (contig == block->scan_hint &&
636 start > block->scan_hint_start)))) {
637 block->scan_hint_start = start;
638 block->scan_hint = contig;
644 * pcpu_block_update_scan - update a block given a free area from a scan
645 * @chunk: chunk of interest
646 * @bit_off: chunk offset
647 * @bits: size of free area
649 * Finding the final allocation spot first goes through pcpu_find_block_fit()
650 * to find a block that can hold the allocation and then pcpu_alloc_area()
651 * where a scan is used. When allocations require specific alignments,
652 * we can inadvertently create holes which will not be seen in the alloc
655 * This takes a given free area hole and updates a block as it may change the
656 * scan_hint. We need to scan backwards to ensure we don't miss free bits
659 static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
662 int s_off = pcpu_off_to_block_off(bit_off);
663 int e_off = s_off + bits;
665 struct pcpu_block_md *block;
667 if (e_off > PCPU_BITMAP_BLOCK_BITS)
670 s_index = pcpu_off_to_block_index(bit_off);
671 block = chunk->md_blocks + s_index;
673 /* scan backwards in case of alignment skipping free bits */
674 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
675 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
677 pcpu_block_update(block, s_off, e_off);
681 * pcpu_chunk_refresh_hint - updates metadata about a chunk
682 * @chunk: chunk of interest
683 * @full_scan: if we should scan from the beginning
685 * Iterates over the metadata blocks to find the largest contig area.
686 * A full scan can be avoided on the allocation path as this is triggered
687 * if we broke the contig_hint. In doing so, the scan_hint will be before
688 * the contig_hint or after if the scan_hint == contig_hint. This cannot
689 * be prevented on freeing as we want to find the largest area possibly
692 static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
694 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
697 /* promote scan_hint to contig_hint */
698 if (!full_scan && chunk_md->scan_hint) {
699 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
700 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
701 chunk_md->contig_hint = chunk_md->scan_hint;
702 chunk_md->scan_hint = 0;
704 bit_off = chunk_md->first_free;
705 chunk_md->contig_hint = 0;
709 pcpu_for_each_md_free_region(chunk, bit_off, bits)
710 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
714 * pcpu_block_refresh_hint
715 * @chunk: chunk of interest
716 * @index: index of the metadata block
718 * Scans over the block beginning at first_free and updates the block
719 * metadata accordingly.
721 static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
723 struct pcpu_block_md *block = chunk->md_blocks + index;
724 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
725 unsigned int rs, re, start; /* region start, region end */
727 /* promote scan_hint to contig_hint */
728 if (block->scan_hint) {
729 start = block->scan_hint_start + block->scan_hint;
730 block->contig_hint_start = block->scan_hint_start;
731 block->contig_hint = block->scan_hint;
732 block->scan_hint = 0;
734 start = block->first_free;
735 block->contig_hint = 0;
738 block->right_free = 0;
740 /* iterate over free areas and update the contig hints */
741 bitmap_for_each_clear_region(alloc_map, rs, re, start,
742 PCPU_BITMAP_BLOCK_BITS)
743 pcpu_block_update(block, rs, re);
747 * pcpu_block_update_hint_alloc - update hint on allocation path
748 * @chunk: chunk of interest
749 * @bit_off: chunk offset
750 * @bits: size of request
752 * Updates metadata for the allocation path. The metadata only has to be
753 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
754 * scans are required if the block's contig hint is broken.
756 static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
759 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
760 int nr_empty_pages = 0;
761 struct pcpu_block_md *s_block, *e_block, *block;
762 int s_index, e_index; /* block indexes of the freed allocation */
763 int s_off, e_off; /* block offsets of the freed allocation */
766 * Calculate per block offsets.
767 * The calculation uses an inclusive range, but the resulting offsets
768 * are [start, end). e_index always points to the last block in the
771 s_index = pcpu_off_to_block_index(bit_off);
772 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
773 s_off = pcpu_off_to_block_off(bit_off);
774 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
776 s_block = chunk->md_blocks + s_index;
777 e_block = chunk->md_blocks + e_index;
781 * block->first_free must be updated if the allocation takes its place.
782 * If the allocation breaks the contig_hint, a scan is required to
785 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
788 if (s_off == s_block->first_free)
789 s_block->first_free = find_next_zero_bit(
790 pcpu_index_alloc_map(chunk, s_index),
791 PCPU_BITMAP_BLOCK_BITS,
794 if (pcpu_region_overlap(s_block->scan_hint_start,
795 s_block->scan_hint_start + s_block->scan_hint,
798 s_block->scan_hint = 0;
800 if (pcpu_region_overlap(s_block->contig_hint_start,
801 s_block->contig_hint_start +
802 s_block->contig_hint,
805 /* block contig hint is broken - scan to fix it */
807 s_block->left_free = 0;
808 pcpu_block_refresh_hint(chunk, s_index);
810 /* update left and right contig manually */
811 s_block->left_free = min(s_block->left_free, s_off);
812 if (s_index == e_index)
813 s_block->right_free = min_t(int, s_block->right_free,
814 PCPU_BITMAP_BLOCK_BITS - e_off);
816 s_block->right_free = 0;
822 if (s_index != e_index) {
823 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
827 * When the allocation is across blocks, the end is along
828 * the left part of the e_block.
830 e_block->first_free = find_next_zero_bit(
831 pcpu_index_alloc_map(chunk, e_index),
832 PCPU_BITMAP_BLOCK_BITS, e_off);
834 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
835 /* reset the block */
838 if (e_off > e_block->scan_hint_start)
839 e_block->scan_hint = 0;
841 e_block->left_free = 0;
842 if (e_off > e_block->contig_hint_start) {
843 /* contig hint is broken - scan to fix it */
844 pcpu_block_refresh_hint(chunk, e_index);
846 e_block->right_free =
847 min_t(int, e_block->right_free,
848 PCPU_BITMAP_BLOCK_BITS - e_off);
852 /* update in-between md_blocks */
853 nr_empty_pages += (e_index - s_index - 1);
854 for (block = s_block + 1; block < e_block; block++) {
855 block->scan_hint = 0;
856 block->contig_hint = 0;
857 block->left_free = 0;
858 block->right_free = 0;
863 pcpu_update_empty_pages(chunk, -nr_empty_pages);
865 if (pcpu_region_overlap(chunk_md->scan_hint_start,
866 chunk_md->scan_hint_start +
870 chunk_md->scan_hint = 0;
873 * The only time a full chunk scan is required is if the chunk
874 * contig hint is broken. Otherwise, it means a smaller space
875 * was used and therefore the chunk contig hint is still correct.
877 if (pcpu_region_overlap(chunk_md->contig_hint_start,
878 chunk_md->contig_hint_start +
879 chunk_md->contig_hint,
882 pcpu_chunk_refresh_hint(chunk, false);
886 * pcpu_block_update_hint_free - updates the block hints on the free path
887 * @chunk: chunk of interest
888 * @bit_off: chunk offset
889 * @bits: size of request
891 * Updates metadata for the allocation path. This avoids a blind block
892 * refresh by making use of the block contig hints. If this fails, it scans
893 * forward and backward to determine the extent of the free area. This is
894 * capped at the boundary of blocks.
896 * A chunk update is triggered if a page becomes free, a block becomes free,
897 * or the free spans across blocks. This tradeoff is to minimize iterating
898 * over the block metadata to update chunk_md->contig_hint.
899 * chunk_md->contig_hint may be off by up to a page, but it will never be more
900 * than the available space. If the contig hint is contained in one block, it
903 static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
906 int nr_empty_pages = 0;
907 struct pcpu_block_md *s_block, *e_block, *block;
908 int s_index, e_index; /* block indexes of the freed allocation */
909 int s_off, e_off; /* block offsets of the freed allocation */
910 int start, end; /* start and end of the whole free area */
913 * Calculate per block offsets.
914 * The calculation uses an inclusive range, but the resulting offsets
915 * are [start, end). e_index always points to the last block in the
918 s_index = pcpu_off_to_block_index(bit_off);
919 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
920 s_off = pcpu_off_to_block_off(bit_off);
921 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
923 s_block = chunk->md_blocks + s_index;
924 e_block = chunk->md_blocks + e_index;
927 * Check if the freed area aligns with the block->contig_hint.
928 * If it does, then the scan to find the beginning/end of the
929 * larger free area can be avoided.
931 * start and end refer to beginning and end of the free area
932 * within each their respective blocks. This is not necessarily
933 * the entire free area as it may span blocks past the beginning
934 * or end of the block.
937 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
938 start = s_block->contig_hint_start;
941 * Scan backwards to find the extent of the free area.
942 * find_last_bit returns the starting bit, so if the start bit
943 * is returned, that means there was no last bit and the
944 * remainder of the chunk is free.
946 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
948 start = (start == l_bit) ? 0 : l_bit + 1;
952 if (e_off == e_block->contig_hint_start)
953 end = e_block->contig_hint_start + e_block->contig_hint;
955 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
956 PCPU_BITMAP_BLOCK_BITS, end);
959 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
960 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
962 pcpu_block_update(s_block, start, e_off);
964 /* freeing in the same block */
965 if (s_index != e_index) {
967 if (end == PCPU_BITMAP_BLOCK_BITS)
969 pcpu_block_update(e_block, 0, end);
971 /* reset md_blocks in the middle */
972 nr_empty_pages += (e_index - s_index - 1);
973 for (block = s_block + 1; block < e_block; block++) {
974 block->first_free = 0;
975 block->scan_hint = 0;
976 block->contig_hint_start = 0;
977 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
978 block->left_free = PCPU_BITMAP_BLOCK_BITS;
979 block->right_free = PCPU_BITMAP_BLOCK_BITS;
984 pcpu_update_empty_pages(chunk, nr_empty_pages);
987 * Refresh chunk metadata when the free makes a block free or spans
988 * across blocks. The contig_hint may be off by up to a page, but if
989 * the contig_hint is contained in a block, it will be accurate with
990 * the else condition below.
992 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
993 pcpu_chunk_refresh_hint(chunk, true);
995 pcpu_block_update(&chunk->chunk_md,
996 pcpu_block_off_to_off(s_index, start),
1001 * pcpu_is_populated - determines if the region is populated
1002 * @chunk: chunk of interest
1003 * @bit_off: chunk offset
1004 * @bits: size of area
1005 * @next_off: return value for the next offset to start searching
1007 * For atomic allocations, check if the backing pages are populated.
1010 * Bool if the backing pages are populated.
1011 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1013 static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1016 unsigned int page_start, page_end, rs, re;
1018 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1019 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1022 bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1026 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1031 * pcpu_find_block_fit - finds the block index to start searching
1032 * @chunk: chunk of interest
1033 * @alloc_bits: size of request in allocation units
1034 * @align: alignment of area (max PAGE_SIZE bytes)
1035 * @pop_only: use populated regions only
1037 * Given a chunk and an allocation spec, find the offset to begin searching
1038 * for a free region. This iterates over the bitmap metadata blocks to
1039 * find an offset that will be guaranteed to fit the requirements. It is
1040 * not quite first fit as if the allocation does not fit in the contig hint
1041 * of a block or chunk, it is skipped. This errs on the side of caution
1042 * to prevent excess iteration. Poor alignment can cause the allocator to
1043 * skip over blocks and chunks that have valid free areas.
1046 * The offset in the bitmap to begin searching.
1047 * -1 if no offset is found.
1049 static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1050 size_t align, bool pop_only)
1052 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1053 int bit_off, bits, next_off;
1056 * Check to see if the allocation can fit in the chunk's contig hint.
1057 * This is an optimization to prevent scanning by assuming if it
1058 * cannot fit in the global hint, there is memory pressure and creating
1059 * a new chunk would happen soon.
1061 bit_off = ALIGN(chunk_md->contig_hint_start, align) -
1062 chunk_md->contig_hint_start;
1063 if (bit_off + alloc_bits > chunk_md->contig_hint)
1066 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1068 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1069 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1077 if (bit_off == pcpu_chunk_map_bits(chunk))
1084 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1085 * @map: the address to base the search on
1086 * @size: the bitmap size in bits
1087 * @start: the bitnumber to start searching at
1088 * @nr: the number of zeroed bits we're looking for
1089 * @align_mask: alignment mask for zero area
1090 * @largest_off: offset of the largest area skipped
1091 * @largest_bits: size of the largest area skipped
1093 * The @align_mask should be one less than a power of 2.
1095 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1096 * the largest area that was skipped. This is imperfect, but in general is
1097 * good enough. The largest remembered region is the largest failed region
1098 * seen. This does not include anything we possibly skipped due to alignment.
1099 * pcpu_block_update_scan() does scan backwards to try and recover what was
1100 * lost to alignment. While this can cause scanning to miss earlier possible
1101 * free areas, smaller allocations will eventually fill those holes.
1103 static unsigned long pcpu_find_zero_area(unsigned long *map,
1105 unsigned long start,
1107 unsigned long align_mask,
1108 unsigned long *largest_off,
1109 unsigned long *largest_bits)
1111 unsigned long index, end, i, area_off, area_bits;
1113 index = find_next_zero_bit(map, size, start);
1115 /* Align allocation */
1116 index = __ALIGN_MASK(index, align_mask);
1122 i = find_next_bit(map, end, index);
1124 area_bits = i - area_off;
1125 /* remember largest unused area with best alignment */
1126 if (area_bits > *largest_bits ||
1127 (area_bits == *largest_bits && *largest_off &&
1128 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1129 *largest_off = area_off;
1130 *largest_bits = area_bits;
1140 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1141 * @chunk: chunk of interest
1142 * @alloc_bits: size of request in allocation units
1143 * @align: alignment of area (max PAGE_SIZE)
1144 * @start: bit_off to start searching
1146 * This function takes in a @start offset to begin searching to fit an
1147 * allocation of @alloc_bits with alignment @align. It needs to scan
1148 * the allocation map because if it fits within the block's contig hint,
1149 * @start will be block->first_free. This is an attempt to fill the
1150 * allocation prior to breaking the contig hint. The allocation and
1151 * boundary maps are updated accordingly if it confirms a valid
1155 * Allocated addr offset in @chunk on success.
1156 * -1 if no matching area is found.
1158 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1159 size_t align, int start)
1161 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1162 size_t align_mask = (align) ? (align - 1) : 0;
1163 unsigned long area_off = 0, area_bits = 0;
1164 int bit_off, end, oslot;
1166 lockdep_assert_held(&pcpu_lock);
1168 oslot = pcpu_chunk_slot(chunk);
1171 * Search to find a fit.
1173 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1174 pcpu_chunk_map_bits(chunk));
1175 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1176 align_mask, &area_off, &area_bits);
1181 pcpu_block_update_scan(chunk, area_off, area_bits);
1183 /* update alloc map */
1184 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1186 /* update boundary map */
1187 set_bit(bit_off, chunk->bound_map);
1188 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1189 set_bit(bit_off + alloc_bits, chunk->bound_map);
1191 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1193 /* update first free bit */
1194 if (bit_off == chunk_md->first_free)
1195 chunk_md->first_free = find_next_zero_bit(
1197 pcpu_chunk_map_bits(chunk),
1198 bit_off + alloc_bits);
1200 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1202 pcpu_chunk_relocate(chunk, oslot);
1204 return bit_off * PCPU_MIN_ALLOC_SIZE;
1208 * pcpu_free_area - frees the corresponding offset
1209 * @chunk: chunk of interest
1210 * @off: addr offset into chunk
1212 * This function determines the size of an allocation to free using
1213 * the boundary bitmap and clears the allocation map.
1216 * Number of freed bytes.
1218 static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1220 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1221 int bit_off, bits, end, oslot, freed;
1223 lockdep_assert_held(&pcpu_lock);
1224 pcpu_stats_area_dealloc(chunk);
1226 oslot = pcpu_chunk_slot(chunk);
1228 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1230 /* find end index */
1231 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1233 bits = end - bit_off;
1234 bitmap_clear(chunk->alloc_map, bit_off, bits);
1236 freed = bits * PCPU_MIN_ALLOC_SIZE;
1238 /* update metadata */
1239 chunk->free_bytes += freed;
1241 /* update first free bit */
1242 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1244 pcpu_block_update_hint_free(chunk, bit_off, bits);
1246 pcpu_chunk_relocate(chunk, oslot);
1251 static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1253 block->scan_hint = 0;
1254 block->contig_hint = nr_bits;
1255 block->left_free = nr_bits;
1256 block->right_free = nr_bits;
1257 block->first_free = 0;
1258 block->nr_bits = nr_bits;
1261 static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1263 struct pcpu_block_md *md_block;
1265 /* init the chunk's block */
1266 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1268 for (md_block = chunk->md_blocks;
1269 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1271 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1275 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1276 * @tmp_addr: the start of the region served
1277 * @map_size: size of the region served
1279 * This is responsible for creating the chunks that serve the first chunk. The
1280 * base_addr is page aligned down of @tmp_addr while the region end is page
1281 * aligned up. Offsets are kept track of to determine the region served. All
1282 * this is done to appease the bitmap allocator in avoiding partial blocks.
1285 * Chunk serving the region at @tmp_addr of @map_size.
1287 static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1290 struct pcpu_chunk *chunk;
1291 unsigned long aligned_addr, lcm_align;
1292 int start_offset, offset_bits, region_size, region_bits;
1295 /* region calculations */
1296 aligned_addr = tmp_addr & PAGE_MASK;
1298 start_offset = tmp_addr - aligned_addr;
1301 * Align the end of the region with the LCM of PAGE_SIZE and
1302 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1305 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1306 region_size = ALIGN(start_offset + map_size, lcm_align);
1308 /* allocate chunk */
1309 alloc_size = sizeof(struct pcpu_chunk) +
1310 BITS_TO_LONGS(region_size >> PAGE_SHIFT);
1311 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1313 panic("%s: Failed to allocate %zu bytes\n", __func__,
1316 INIT_LIST_HEAD(&chunk->list);
1318 chunk->base_addr = (void *)aligned_addr;
1319 chunk->start_offset = start_offset;
1320 chunk->end_offset = region_size - chunk->start_offset - map_size;
1322 chunk->nr_pages = region_size >> PAGE_SHIFT;
1323 region_bits = pcpu_chunk_map_bits(chunk);
1325 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1326 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1327 if (!chunk->alloc_map)
1328 panic("%s: Failed to allocate %zu bytes\n", __func__,
1332 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1333 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1334 if (!chunk->bound_map)
1335 panic("%s: Failed to allocate %zu bytes\n", __func__,
1338 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1339 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1340 if (!chunk->md_blocks)
1341 panic("%s: Failed to allocate %zu bytes\n", __func__,
1344 pcpu_init_md_blocks(chunk);
1346 /* manage populated page bitmap */
1347 chunk->immutable = true;
1348 bitmap_fill(chunk->populated, chunk->nr_pages);
1349 chunk->nr_populated = chunk->nr_pages;
1350 chunk->nr_empty_pop_pages = chunk->nr_pages;
1352 chunk->free_bytes = map_size;
1354 if (chunk->start_offset) {
1355 /* hide the beginning of the bitmap */
1356 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1357 bitmap_set(chunk->alloc_map, 0, offset_bits);
1358 set_bit(0, chunk->bound_map);
1359 set_bit(offset_bits, chunk->bound_map);
1361 chunk->chunk_md.first_free = offset_bits;
1363 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1366 if (chunk->end_offset) {
1367 /* hide the end of the bitmap */
1368 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1369 bitmap_set(chunk->alloc_map,
1370 pcpu_chunk_map_bits(chunk) - offset_bits,
1372 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1374 set_bit(region_bits, chunk->bound_map);
1376 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1377 - offset_bits, offset_bits);
1383 static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1385 struct pcpu_chunk *chunk;
1388 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1392 INIT_LIST_HEAD(&chunk->list);
1393 chunk->nr_pages = pcpu_unit_pages;
1394 region_bits = pcpu_chunk_map_bits(chunk);
1396 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1397 sizeof(chunk->alloc_map[0]), gfp);
1398 if (!chunk->alloc_map)
1399 goto alloc_map_fail;
1401 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1402 sizeof(chunk->bound_map[0]), gfp);
1403 if (!chunk->bound_map)
1404 goto bound_map_fail;
1406 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1407 sizeof(chunk->md_blocks[0]), gfp);
1408 if (!chunk->md_blocks)
1409 goto md_blocks_fail;
1411 pcpu_init_md_blocks(chunk);
1414 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1419 pcpu_mem_free(chunk->bound_map);
1421 pcpu_mem_free(chunk->alloc_map);
1423 pcpu_mem_free(chunk);
1428 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1432 pcpu_mem_free(chunk->md_blocks);
1433 pcpu_mem_free(chunk->bound_map);
1434 pcpu_mem_free(chunk->alloc_map);
1435 pcpu_mem_free(chunk);
1439 * pcpu_chunk_populated - post-population bookkeeping
1440 * @chunk: pcpu_chunk which got populated
1441 * @page_start: the start page
1442 * @page_end: the end page
1444 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1445 * the bookkeeping information accordingly. Must be called after each
1446 * successful population.
1448 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1449 * is to serve an allocation in that area.
1451 static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1454 int nr = page_end - page_start;
1456 lockdep_assert_held(&pcpu_lock);
1458 bitmap_set(chunk->populated, page_start, nr);
1459 chunk->nr_populated += nr;
1460 pcpu_nr_populated += nr;
1462 pcpu_update_empty_pages(chunk, nr);
1466 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1467 * @chunk: pcpu_chunk which got depopulated
1468 * @page_start: the start page
1469 * @page_end: the end page
1471 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1472 * Update the bookkeeping information accordingly. Must be called after
1473 * each successful depopulation.
1475 static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1476 int page_start, int page_end)
1478 int nr = page_end - page_start;
1480 lockdep_assert_held(&pcpu_lock);
1482 bitmap_clear(chunk->populated, page_start, nr);
1483 chunk->nr_populated -= nr;
1484 pcpu_nr_populated -= nr;
1486 pcpu_update_empty_pages(chunk, -nr);
1490 * Chunk management implementation.
1492 * To allow different implementations, chunk alloc/free and
1493 * [de]population are implemented in a separate file which is pulled
1494 * into this file and compiled together. The following functions
1495 * should be implemented.
1497 * pcpu_populate_chunk - populate the specified range of a chunk
1498 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1499 * pcpu_create_chunk - create a new chunk
1500 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1501 * pcpu_addr_to_page - translate address to physical address
1502 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1504 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1505 int page_start, int page_end, gfp_t gfp);
1506 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1507 int page_start, int page_end);
1508 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1509 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1510 static struct page *pcpu_addr_to_page(void *addr);
1511 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1513 #ifdef CONFIG_NEED_PER_CPU_KM
1514 #include "percpu-km.c"
1516 #include "percpu-vm.c"
1520 * pcpu_chunk_addr_search - determine chunk containing specified address
1521 * @addr: address for which the chunk needs to be determined.
1523 * This is an internal function that handles all but static allocations.
1524 * Static percpu address values should never be passed into the allocator.
1527 * The address of the found chunk.
1529 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1531 /* is it in the dynamic region (first chunk)? */
1532 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1533 return pcpu_first_chunk;
1535 /* is it in the reserved region? */
1536 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1537 return pcpu_reserved_chunk;
1540 * The address is relative to unit0 which might be unused and
1541 * thus unmapped. Offset the address to the unit space of the
1542 * current processor before looking it up in the vmalloc
1543 * space. Note that any possible cpu id can be used here, so
1544 * there's no need to worry about preemption or cpu hotplug.
1546 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1547 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1551 * pcpu_alloc - the percpu allocator
1552 * @size: size of area to allocate in bytes
1553 * @align: alignment of area (max PAGE_SIZE)
1554 * @reserved: allocate from the reserved chunk if available
1555 * @gfp: allocation flags
1557 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1558 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1559 * then no warning will be triggered on invalid or failed allocation
1563 * Percpu pointer to the allocated area on success, NULL on failure.
1565 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1571 static int warn_limit = 10;
1572 struct pcpu_chunk *chunk, *next;
1574 int slot, off, cpu, ret;
1575 unsigned long flags;
1577 size_t bits, bit_align;
1579 gfp = current_gfp_context(gfp);
1580 /* whitelisted flags that can be passed to the backing allocators */
1581 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1582 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1583 do_warn = !(gfp & __GFP_NOWARN);
1586 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1587 * therefore alignment must be a minimum of that many bytes.
1588 * An allocation may have internal fragmentation from rounding up
1589 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1591 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1592 align = PCPU_MIN_ALLOC_SIZE;
1594 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1595 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1596 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1598 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1599 !is_power_of_2(align))) {
1600 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1607 * pcpu_balance_workfn() allocates memory under this mutex,
1608 * and it may wait for memory reclaim. Allow current task
1609 * to become OOM victim, in case of memory pressure.
1611 if (gfp & __GFP_NOFAIL)
1612 mutex_lock(&pcpu_alloc_mutex);
1613 else if (mutex_lock_killable(&pcpu_alloc_mutex))
1617 spin_lock_irqsave(&pcpu_lock, flags);
1619 /* serve reserved allocations from the reserved chunk if available */
1620 if (reserved && pcpu_reserved_chunk) {
1621 chunk = pcpu_reserved_chunk;
1623 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1625 err = "alloc from reserved chunk failed";
1629 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1633 err = "alloc from reserved chunk failed";
1638 /* search through normal chunks */
1639 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1640 list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) {
1641 off = pcpu_find_block_fit(chunk, bits, bit_align,
1644 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1645 pcpu_chunk_move(chunk, 0);
1649 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1656 spin_unlock_irqrestore(&pcpu_lock, flags);
1659 * No space left. Create a new chunk. We don't want multiple
1660 * tasks to create chunks simultaneously. Serialize and create iff
1661 * there's still no empty chunk after grabbing the mutex.
1664 err = "atomic alloc failed, no space left";
1668 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1669 chunk = pcpu_create_chunk(pcpu_gfp);
1671 err = "failed to allocate new chunk";
1675 spin_lock_irqsave(&pcpu_lock, flags);
1676 pcpu_chunk_relocate(chunk, -1);
1678 spin_lock_irqsave(&pcpu_lock, flags);
1684 pcpu_stats_area_alloc(chunk, size);
1685 spin_unlock_irqrestore(&pcpu_lock, flags);
1687 /* populate if not all pages are already there */
1689 unsigned int page_start, page_end, rs, re;
1691 page_start = PFN_DOWN(off);
1692 page_end = PFN_UP(off + size);
1694 bitmap_for_each_clear_region(chunk->populated, rs, re,
1695 page_start, page_end) {
1696 WARN_ON(chunk->immutable);
1698 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1700 spin_lock_irqsave(&pcpu_lock, flags);
1702 pcpu_free_area(chunk, off);
1703 err = "failed to populate";
1706 pcpu_chunk_populated(chunk, rs, re);
1707 spin_unlock_irqrestore(&pcpu_lock, flags);
1710 mutex_unlock(&pcpu_alloc_mutex);
1713 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1714 pcpu_schedule_balance_work();
1716 /* clear the areas and return address relative to base address */
1717 for_each_possible_cpu(cpu)
1718 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1720 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1721 kmemleak_alloc_percpu(ptr, size, gfp);
1723 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1724 chunk->base_addr, off, ptr);
1729 spin_unlock_irqrestore(&pcpu_lock, flags);
1731 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1733 if (!is_atomic && do_warn && warn_limit) {
1734 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1735 size, align, is_atomic, err);
1738 pr_info("limit reached, disable warning\n");
1741 /* see the flag handling in pcpu_blance_workfn() */
1742 pcpu_atomic_alloc_failed = true;
1743 pcpu_schedule_balance_work();
1745 mutex_unlock(&pcpu_alloc_mutex);
1751 * __alloc_percpu_gfp - allocate dynamic percpu area
1752 * @size: size of area to allocate in bytes
1753 * @align: alignment of area (max PAGE_SIZE)
1754 * @gfp: allocation flags
1756 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1757 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1758 * be called from any context but is a lot more likely to fail. If @gfp
1759 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1760 * allocation requests.
1763 * Percpu pointer to the allocated area on success, NULL on failure.
1765 void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1767 return pcpu_alloc(size, align, false, gfp);
1769 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1772 * __alloc_percpu - allocate dynamic percpu area
1773 * @size: size of area to allocate in bytes
1774 * @align: alignment of area (max PAGE_SIZE)
1776 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1778 void __percpu *__alloc_percpu(size_t size, size_t align)
1780 return pcpu_alloc(size, align, false, GFP_KERNEL);
1782 EXPORT_SYMBOL_GPL(__alloc_percpu);
1785 * __alloc_reserved_percpu - allocate reserved percpu area
1786 * @size: size of area to allocate in bytes
1787 * @align: alignment of area (max PAGE_SIZE)
1789 * Allocate zero-filled percpu area of @size bytes aligned at @align
1790 * from reserved percpu area if arch has set it up; otherwise,
1791 * allocation is served from the same dynamic area. Might sleep.
1792 * Might trigger writeouts.
1795 * Does GFP_KERNEL allocation.
1798 * Percpu pointer to the allocated area on success, NULL on failure.
1800 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1802 return pcpu_alloc(size, align, true, GFP_KERNEL);
1806 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1809 * Reclaim all fully free chunks except for the first one. This is also
1810 * responsible for maintaining the pool of empty populated pages. However,
1811 * it is possible that this is called when physical memory is scarce causing
1812 * OOM killer to be triggered. We should avoid doing so until an actual
1813 * allocation causes the failure as it is possible that requests can be
1814 * serviced from already backed regions.
1816 static void pcpu_balance_workfn(struct work_struct *work)
1818 /* gfp flags passed to underlying allocators */
1819 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1821 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1822 struct pcpu_chunk *chunk, *next;
1823 int slot, nr_to_pop, ret;
1826 * There's no reason to keep around multiple unused chunks and VM
1827 * areas can be scarce. Destroy all free chunks except for one.
1829 mutex_lock(&pcpu_alloc_mutex);
1830 spin_lock_irq(&pcpu_lock);
1832 list_for_each_entry_safe(chunk, next, free_head, list) {
1833 WARN_ON(chunk->immutable);
1835 /* spare the first one */
1836 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1839 list_move(&chunk->list, &to_free);
1842 spin_unlock_irq(&pcpu_lock);
1844 list_for_each_entry_safe(chunk, next, &to_free, list) {
1845 unsigned int rs, re;
1847 bitmap_for_each_set_region(chunk->populated, rs, re, 0,
1849 pcpu_depopulate_chunk(chunk, rs, re);
1850 spin_lock_irq(&pcpu_lock);
1851 pcpu_chunk_depopulated(chunk, rs, re);
1852 spin_unlock_irq(&pcpu_lock);
1854 pcpu_destroy_chunk(chunk);
1859 * Ensure there are certain number of free populated pages for
1860 * atomic allocs. Fill up from the most packed so that atomic
1861 * allocs don't increase fragmentation. If atomic allocation
1862 * failed previously, always populate the maximum amount. This
1863 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1864 * failing indefinitely; however, large atomic allocs are not
1865 * something we support properly and can be highly unreliable and
1869 if (pcpu_atomic_alloc_failed) {
1870 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1871 /* best effort anyway, don't worry about synchronization */
1872 pcpu_atomic_alloc_failed = false;
1874 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1875 pcpu_nr_empty_pop_pages,
1876 0, PCPU_EMPTY_POP_PAGES_HIGH);
1879 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1880 unsigned int nr_unpop = 0, rs, re;
1885 spin_lock_irq(&pcpu_lock);
1886 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1887 nr_unpop = chunk->nr_pages - chunk->nr_populated;
1891 spin_unlock_irq(&pcpu_lock);
1896 /* @chunk can't go away while pcpu_alloc_mutex is held */
1897 bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
1899 int nr = min_t(int, re - rs, nr_to_pop);
1901 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
1904 spin_lock_irq(&pcpu_lock);
1905 pcpu_chunk_populated(chunk, rs, rs + nr);
1906 spin_unlock_irq(&pcpu_lock);
1917 /* ran out of chunks to populate, create a new one and retry */
1918 chunk = pcpu_create_chunk(gfp);
1920 spin_lock_irq(&pcpu_lock);
1921 pcpu_chunk_relocate(chunk, -1);
1922 spin_unlock_irq(&pcpu_lock);
1927 mutex_unlock(&pcpu_alloc_mutex);
1931 * free_percpu - free percpu area
1932 * @ptr: pointer to area to free
1934 * Free percpu area @ptr.
1937 * Can be called from atomic context.
1939 void free_percpu(void __percpu *ptr)
1942 struct pcpu_chunk *chunk;
1943 unsigned long flags;
1945 bool need_balance = false;
1950 kmemleak_free_percpu(ptr);
1952 addr = __pcpu_ptr_to_addr(ptr);
1954 spin_lock_irqsave(&pcpu_lock, flags);
1956 chunk = pcpu_chunk_addr_search(addr);
1957 off = addr - chunk->base_addr;
1959 pcpu_free_area(chunk, off);
1961 /* if there are more than one fully free chunks, wake up grim reaper */
1962 if (chunk->free_bytes == pcpu_unit_size) {
1963 struct pcpu_chunk *pos;
1965 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1967 need_balance = true;
1972 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1974 spin_unlock_irqrestore(&pcpu_lock, flags);
1977 pcpu_schedule_balance_work();
1979 EXPORT_SYMBOL_GPL(free_percpu);
1981 bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1984 const size_t static_size = __per_cpu_end - __per_cpu_start;
1985 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1988 for_each_possible_cpu(cpu) {
1989 void *start = per_cpu_ptr(base, cpu);
1990 void *va = (void *)addr;
1992 if (va >= start && va < start + static_size) {
1994 *can_addr = (unsigned long) (va - start);
1995 *can_addr += (unsigned long)
1996 per_cpu_ptr(base, get_boot_cpu_id());
2002 /* on UP, can't distinguish from other static vars, always false */
2007 * is_kernel_percpu_address - test whether address is from static percpu area
2008 * @addr: address to test
2010 * Test whether @addr belongs to in-kernel static percpu area. Module
2011 * static percpu areas are not considered. For those, use
2012 * is_module_percpu_address().
2015 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2017 bool is_kernel_percpu_address(unsigned long addr)
2019 return __is_kernel_percpu_address(addr, NULL);
2023 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2024 * @addr: the address to be converted to physical address
2026 * Given @addr which is dereferenceable address obtained via one of
2027 * percpu access macros, this function translates it into its physical
2028 * address. The caller is responsible for ensuring @addr stays valid
2029 * until this function finishes.
2031 * percpu allocator has special setup for the first chunk, which currently
2032 * supports either embedding in linear address space or vmalloc mapping,
2033 * and, from the second one, the backing allocator (currently either vm or
2034 * km) provides translation.
2036 * The addr can be translated simply without checking if it falls into the
2037 * first chunk. But the current code reflects better how percpu allocator
2038 * actually works, and the verification can discover both bugs in percpu
2039 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2043 * The physical address for @addr.
2045 phys_addr_t per_cpu_ptr_to_phys(void *addr)
2047 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2048 bool in_first_chunk = false;
2049 unsigned long first_low, first_high;
2053 * The following test on unit_low/high isn't strictly
2054 * necessary but will speed up lookups of addresses which
2055 * aren't in the first chunk.
2057 * The address check is against full chunk sizes. pcpu_base_addr
2058 * points to the beginning of the first chunk including the
2059 * static region. Assumes good intent as the first chunk may
2060 * not be full (ie. < pcpu_unit_pages in size).
2062 first_low = (unsigned long)pcpu_base_addr +
2063 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2064 first_high = (unsigned long)pcpu_base_addr +
2065 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2066 if ((unsigned long)addr >= first_low &&
2067 (unsigned long)addr < first_high) {
2068 for_each_possible_cpu(cpu) {
2069 void *start = per_cpu_ptr(base, cpu);
2071 if (addr >= start && addr < start + pcpu_unit_size) {
2072 in_first_chunk = true;
2078 if (in_first_chunk) {
2079 if (!is_vmalloc_addr(addr))
2082 return page_to_phys(vmalloc_to_page(addr)) +
2083 offset_in_page(addr);
2085 return page_to_phys(pcpu_addr_to_page(addr)) +
2086 offset_in_page(addr);
2090 * pcpu_alloc_alloc_info - allocate percpu allocation info
2091 * @nr_groups: the number of groups
2092 * @nr_units: the number of units
2094 * Allocate ai which is large enough for @nr_groups groups containing
2095 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2096 * cpu_map array which is long enough for @nr_units and filled with
2097 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2098 * pointer of other groups.
2101 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2104 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2107 struct pcpu_alloc_info *ai;
2108 size_t base_size, ai_size;
2112 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2113 __alignof__(ai->groups[0].cpu_map[0]));
2114 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2116 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2122 ai->groups[0].cpu_map = ptr;
2124 for (unit = 0; unit < nr_units; unit++)
2125 ai->groups[0].cpu_map[unit] = NR_CPUS;
2127 ai->nr_groups = nr_groups;
2128 ai->__ai_size = PFN_ALIGN(ai_size);
2134 * pcpu_free_alloc_info - free percpu allocation info
2135 * @ai: pcpu_alloc_info to free
2137 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2139 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2141 memblock_free_early(__pa(ai), ai->__ai_size);
2145 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2147 * @ai: allocation info to dump
2149 * Print out information about @ai using loglevel @lvl.
2151 static void pcpu_dump_alloc_info(const char *lvl,
2152 const struct pcpu_alloc_info *ai)
2154 int group_width = 1, cpu_width = 1, width;
2155 char empty_str[] = "--------";
2156 int alloc = 0, alloc_end = 0;
2158 int upa, apl; /* units per alloc, allocs per line */
2164 v = num_possible_cpus();
2167 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2169 upa = ai->alloc_size / ai->unit_size;
2170 width = upa * (cpu_width + 1) + group_width + 3;
2171 apl = rounddown_pow_of_two(max(60 / width, 1));
2173 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2174 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2175 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2177 for (group = 0; group < ai->nr_groups; group++) {
2178 const struct pcpu_group_info *gi = &ai->groups[group];
2179 int unit = 0, unit_end = 0;
2181 BUG_ON(gi->nr_units % upa);
2182 for (alloc_end += gi->nr_units / upa;
2183 alloc < alloc_end; alloc++) {
2184 if (!(alloc % apl)) {
2186 printk("%spcpu-alloc: ", lvl);
2188 pr_cont("[%0*d] ", group_width, group);
2190 for (unit_end += upa; unit < unit_end; unit++)
2191 if (gi->cpu_map[unit] != NR_CPUS)
2193 cpu_width, gi->cpu_map[unit]);
2195 pr_cont("%s ", empty_str);
2202 * pcpu_setup_first_chunk - initialize the first percpu chunk
2203 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2204 * @base_addr: mapped address
2206 * Initialize the first percpu chunk which contains the kernel static
2207 * percpu area. This function is to be called from arch percpu area
2210 * @ai contains all information necessary to initialize the first
2211 * chunk and prime the dynamic percpu allocator.
2213 * @ai->static_size is the size of static percpu area.
2215 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2216 * reserve after the static area in the first chunk. This reserves
2217 * the first chunk such that it's available only through reserved
2218 * percpu allocation. This is primarily used to serve module percpu
2219 * static areas on architectures where the addressing model has
2220 * limited offset range for symbol relocations to guarantee module
2221 * percpu symbols fall inside the relocatable range.
2223 * @ai->dyn_size determines the number of bytes available for dynamic
2224 * allocation in the first chunk. The area between @ai->static_size +
2225 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2227 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2228 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2231 * @ai->atom_size is the allocation atom size and used as alignment
2234 * @ai->alloc_size is the allocation size and always multiple of
2235 * @ai->atom_size. This is larger than @ai->atom_size if
2236 * @ai->unit_size is larger than @ai->atom_size.
2238 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2239 * percpu areas. Units which should be colocated are put into the
2240 * same group. Dynamic VM areas will be allocated according to these
2241 * groupings. If @ai->nr_groups is zero, a single group containing
2242 * all units is assumed.
2244 * The caller should have mapped the first chunk at @base_addr and
2245 * copied static data to each unit.
2247 * The first chunk will always contain a static and a dynamic region.
2248 * However, the static region is not managed by any chunk. If the first
2249 * chunk also contains a reserved region, it is served by two chunks -
2250 * one for the reserved region and one for the dynamic region. They
2251 * share the same vm, but use offset regions in the area allocation map.
2252 * The chunk serving the dynamic region is circulated in the chunk slots
2253 * and available for dynamic allocation like any other chunk.
2255 void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2258 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2259 size_t static_size, dyn_size;
2260 struct pcpu_chunk *chunk;
2261 unsigned long *group_offsets;
2262 size_t *group_sizes;
2263 unsigned long *unit_off;
2268 unsigned long tmp_addr;
2271 #define PCPU_SETUP_BUG_ON(cond) do { \
2272 if (unlikely(cond)) { \
2273 pr_emerg("failed to initialize, %s\n", #cond); \
2274 pr_emerg("cpu_possible_mask=%*pb\n", \
2275 cpumask_pr_args(cpu_possible_mask)); \
2276 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2282 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2284 PCPU_SETUP_BUG_ON(!ai->static_size);
2285 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2287 PCPU_SETUP_BUG_ON(!base_addr);
2288 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2289 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2290 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2291 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2292 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2293 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2294 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2295 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2296 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2297 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2298 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2300 /* process group information and build config tables accordingly */
2301 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2302 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2304 panic("%s: Failed to allocate %zu bytes\n", __func__,
2307 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2308 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2310 panic("%s: Failed to allocate %zu bytes\n", __func__,
2313 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2314 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2316 panic("%s: Failed to allocate %zu bytes\n", __func__,
2319 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2320 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2322 panic("%s: Failed to allocate %zu bytes\n", __func__,
2325 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2326 unit_map[cpu] = UINT_MAX;
2328 pcpu_low_unit_cpu = NR_CPUS;
2329 pcpu_high_unit_cpu = NR_CPUS;
2331 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2332 const struct pcpu_group_info *gi = &ai->groups[group];
2334 group_offsets[group] = gi->base_offset;
2335 group_sizes[group] = gi->nr_units * ai->unit_size;
2337 for (i = 0; i < gi->nr_units; i++) {
2338 cpu = gi->cpu_map[i];
2342 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2343 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2344 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2346 unit_map[cpu] = unit + i;
2347 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2349 /* determine low/high unit_cpu */
2350 if (pcpu_low_unit_cpu == NR_CPUS ||
2351 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2352 pcpu_low_unit_cpu = cpu;
2353 if (pcpu_high_unit_cpu == NR_CPUS ||
2354 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2355 pcpu_high_unit_cpu = cpu;
2358 pcpu_nr_units = unit;
2360 for_each_possible_cpu(cpu)
2361 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2363 /* we're done parsing the input, undefine BUG macro and dump config */
2364 #undef PCPU_SETUP_BUG_ON
2365 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2367 pcpu_nr_groups = ai->nr_groups;
2368 pcpu_group_offsets = group_offsets;
2369 pcpu_group_sizes = group_sizes;
2370 pcpu_unit_map = unit_map;
2371 pcpu_unit_offsets = unit_off;
2373 /* determine basic parameters */
2374 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2375 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2376 pcpu_atom_size = ai->atom_size;
2377 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2378 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2380 pcpu_stats_save_ai(ai);
2383 * Allocate chunk slots. The additional last slot is for
2386 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2387 pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]),
2390 panic("%s: Failed to allocate %zu bytes\n", __func__,
2391 pcpu_nr_slots * sizeof(pcpu_slot[0]));
2392 for (i = 0; i < pcpu_nr_slots; i++)
2393 INIT_LIST_HEAD(&pcpu_slot[i]);
2396 * The end of the static region needs to be aligned with the
2397 * minimum allocation size as this offsets the reserved and
2398 * dynamic region. The first chunk ends page aligned by
2399 * expanding the dynamic region, therefore the dynamic region
2400 * can be shrunk to compensate while still staying above the
2403 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2404 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2407 * Initialize first chunk.
2408 * If the reserved_size is non-zero, this initializes the reserved
2409 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2410 * and the dynamic region is initialized here. The first chunk,
2411 * pcpu_first_chunk, will always point to the chunk that serves
2412 * the dynamic region.
2414 tmp_addr = (unsigned long)base_addr + static_size;
2415 map_size = ai->reserved_size ?: dyn_size;
2416 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2418 /* init dynamic chunk if necessary */
2419 if (ai->reserved_size) {
2420 pcpu_reserved_chunk = chunk;
2422 tmp_addr = (unsigned long)base_addr + static_size +
2424 map_size = dyn_size;
2425 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2428 /* link the first chunk in */
2429 pcpu_first_chunk = chunk;
2430 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2431 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2433 /* include all regions of the first chunk */
2434 pcpu_nr_populated += PFN_DOWN(size_sum);
2436 pcpu_stats_chunk_alloc();
2437 trace_percpu_create_chunk(base_addr);
2440 pcpu_base_addr = base_addr;
2445 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2446 [PCPU_FC_AUTO] = "auto",
2447 [PCPU_FC_EMBED] = "embed",
2448 [PCPU_FC_PAGE] = "page",
2451 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2453 static int __init percpu_alloc_setup(char *str)
2460 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2461 else if (!strcmp(str, "embed"))
2462 pcpu_chosen_fc = PCPU_FC_EMBED;
2464 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2465 else if (!strcmp(str, "page"))
2466 pcpu_chosen_fc = PCPU_FC_PAGE;
2469 pr_warn("unknown allocator %s specified\n", str);
2473 early_param("percpu_alloc", percpu_alloc_setup);
2476 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2477 * Build it if needed by the arch config or the generic setup is going
2480 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2481 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2482 #define BUILD_EMBED_FIRST_CHUNK
2485 /* build pcpu_page_first_chunk() iff needed by the arch config */
2486 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2487 #define BUILD_PAGE_FIRST_CHUNK
2490 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2491 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2493 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2494 * @reserved_size: the size of reserved percpu area in bytes
2495 * @dyn_size: minimum free size for dynamic allocation in bytes
2496 * @atom_size: allocation atom size
2497 * @cpu_distance_fn: callback to determine distance between cpus, optional
2499 * This function determines grouping of units, their mappings to cpus
2500 * and other parameters considering needed percpu size, allocation
2501 * atom size and distances between CPUs.
2503 * Groups are always multiples of atom size and CPUs which are of
2504 * LOCAL_DISTANCE both ways are grouped together and share space for
2505 * units in the same group. The returned configuration is guaranteed
2506 * to have CPUs on different nodes on different groups and >=75% usage
2507 * of allocated virtual address space.
2510 * On success, pointer to the new allocation_info is returned. On
2511 * failure, ERR_PTR value is returned.
2513 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2514 size_t reserved_size, size_t dyn_size,
2516 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2518 static int group_map[NR_CPUS] __initdata;
2519 static int group_cnt[NR_CPUS] __initdata;
2520 const size_t static_size = __per_cpu_end - __per_cpu_start;
2521 int nr_groups = 1, nr_units = 0;
2522 size_t size_sum, min_unit_size, alloc_size;
2523 int upa, max_upa, best_upa; /* units_per_alloc */
2524 int last_allocs, group, unit;
2525 unsigned int cpu, tcpu;
2526 struct pcpu_alloc_info *ai;
2527 unsigned int *cpu_map;
2529 /* this function may be called multiple times */
2530 memset(group_map, 0, sizeof(group_map));
2531 memset(group_cnt, 0, sizeof(group_cnt));
2533 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2534 size_sum = PFN_ALIGN(static_size + reserved_size +
2535 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2536 dyn_size = size_sum - static_size - reserved_size;
2539 * Determine min_unit_size, alloc_size and max_upa such that
2540 * alloc_size is multiple of atom_size and is the smallest
2541 * which can accommodate 4k aligned segments which are equal to
2542 * or larger than min_unit_size.
2544 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2546 /* determine the maximum # of units that can fit in an allocation */
2547 alloc_size = roundup(min_unit_size, atom_size);
2548 upa = alloc_size / min_unit_size;
2549 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2553 /* group cpus according to their proximity */
2554 for_each_possible_cpu(cpu) {
2557 for_each_possible_cpu(tcpu) {
2560 if (group_map[tcpu] == group && cpu_distance_fn &&
2561 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2562 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2564 nr_groups = max(nr_groups, group + 1);
2568 group_map[cpu] = group;
2573 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2574 * Expand the unit_size until we use >= 75% of the units allocated.
2575 * Related to atom_size, which could be much larger than the unit_size.
2577 last_allocs = INT_MAX;
2578 for (upa = max_upa; upa; upa--) {
2579 int allocs = 0, wasted = 0;
2581 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2584 for (group = 0; group < nr_groups; group++) {
2585 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2586 allocs += this_allocs;
2587 wasted += this_allocs * upa - group_cnt[group];
2591 * Don't accept if wastage is over 1/3. The
2592 * greater-than comparison ensures upa==1 always
2593 * passes the following check.
2595 if (wasted > num_possible_cpus() / 3)
2598 /* and then don't consume more memory */
2599 if (allocs > last_allocs)
2601 last_allocs = allocs;
2606 /* allocate and fill alloc_info */
2607 for (group = 0; group < nr_groups; group++)
2608 nr_units += roundup(group_cnt[group], upa);
2610 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2612 return ERR_PTR(-ENOMEM);
2613 cpu_map = ai->groups[0].cpu_map;
2615 for (group = 0; group < nr_groups; group++) {
2616 ai->groups[group].cpu_map = cpu_map;
2617 cpu_map += roundup(group_cnt[group], upa);
2620 ai->static_size = static_size;
2621 ai->reserved_size = reserved_size;
2622 ai->dyn_size = dyn_size;
2623 ai->unit_size = alloc_size / upa;
2624 ai->atom_size = atom_size;
2625 ai->alloc_size = alloc_size;
2627 for (group = 0, unit = 0; group < nr_groups; group++) {
2628 struct pcpu_group_info *gi = &ai->groups[group];
2631 * Initialize base_offset as if all groups are located
2632 * back-to-back. The caller should update this to
2633 * reflect actual allocation.
2635 gi->base_offset = unit * ai->unit_size;
2637 for_each_possible_cpu(cpu)
2638 if (group_map[cpu] == group)
2639 gi->cpu_map[gi->nr_units++] = cpu;
2640 gi->nr_units = roundup(gi->nr_units, upa);
2641 unit += gi->nr_units;
2643 BUG_ON(unit != nr_units);
2647 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2649 #if defined(BUILD_EMBED_FIRST_CHUNK)
2651 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2652 * @reserved_size: the size of reserved percpu area in bytes
2653 * @dyn_size: minimum free size for dynamic allocation in bytes
2654 * @atom_size: allocation atom size
2655 * @cpu_distance_fn: callback to determine distance between cpus, optional
2656 * @alloc_fn: function to allocate percpu page
2657 * @free_fn: function to free percpu page
2659 * This is a helper to ease setting up embedded first percpu chunk and
2660 * can be called where pcpu_setup_first_chunk() is expected.
2662 * If this function is used to setup the first chunk, it is allocated
2663 * by calling @alloc_fn and used as-is without being mapped into
2664 * vmalloc area. Allocations are always whole multiples of @atom_size
2665 * aligned to @atom_size.
2667 * This enables the first chunk to piggy back on the linear physical
2668 * mapping which often uses larger page size. Please note that this
2669 * can result in very sparse cpu->unit mapping on NUMA machines thus
2670 * requiring large vmalloc address space. Don't use this allocator if
2671 * vmalloc space is not orders of magnitude larger than distances
2672 * between node memory addresses (ie. 32bit NUMA machines).
2674 * @dyn_size specifies the minimum dynamic area size.
2676 * If the needed size is smaller than the minimum or specified unit
2677 * size, the leftover is returned using @free_fn.
2680 * 0 on success, -errno on failure.
2682 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2684 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2685 pcpu_fc_alloc_fn_t alloc_fn,
2686 pcpu_fc_free_fn_t free_fn)
2688 void *base = (void *)ULONG_MAX;
2689 void **areas = NULL;
2690 struct pcpu_alloc_info *ai;
2691 size_t size_sum, areas_size;
2692 unsigned long max_distance;
2693 int group, i, highest_group, rc = 0;
2695 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2700 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2701 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2703 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
2709 /* allocate, copy and determine base address & max_distance */
2711 for (group = 0; group < ai->nr_groups; group++) {
2712 struct pcpu_group_info *gi = &ai->groups[group];
2713 unsigned int cpu = NR_CPUS;
2716 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2717 cpu = gi->cpu_map[i];
2718 BUG_ON(cpu == NR_CPUS);
2720 /* allocate space for the whole group */
2721 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2724 goto out_free_areas;
2726 /* kmemleak tracks the percpu allocations separately */
2730 base = min(ptr, base);
2731 if (ptr > areas[highest_group])
2732 highest_group = group;
2734 max_distance = areas[highest_group] - base;
2735 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2737 /* warn if maximum distance is further than 75% of vmalloc space */
2738 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2739 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2740 max_distance, VMALLOC_TOTAL);
2741 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2742 /* and fail if we have fallback */
2744 goto out_free_areas;
2749 * Copy data and free unused parts. This should happen after all
2750 * allocations are complete; otherwise, we may end up with
2751 * overlapping groups.
2753 for (group = 0; group < ai->nr_groups; group++) {
2754 struct pcpu_group_info *gi = &ai->groups[group];
2755 void *ptr = areas[group];
2757 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2758 if (gi->cpu_map[i] == NR_CPUS) {
2759 /* unused unit, free whole */
2760 free_fn(ptr, ai->unit_size);
2763 /* copy and return the unused part */
2764 memcpy(ptr, __per_cpu_load, ai->static_size);
2765 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2769 /* base address is now known, determine group base offsets */
2770 for (group = 0; group < ai->nr_groups; group++) {
2771 ai->groups[group].base_offset = areas[group] - base;
2774 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2775 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2776 ai->dyn_size, ai->unit_size);
2778 pcpu_setup_first_chunk(ai, base);
2782 for (group = 0; group < ai->nr_groups; group++)
2784 free_fn(areas[group],
2785 ai->groups[group].nr_units * ai->unit_size);
2787 pcpu_free_alloc_info(ai);
2789 memblock_free_early(__pa(areas), areas_size);
2792 #endif /* BUILD_EMBED_FIRST_CHUNK */
2794 #ifdef BUILD_PAGE_FIRST_CHUNK
2796 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2797 * @reserved_size: the size of reserved percpu area in bytes
2798 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2799 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2800 * @populate_pte_fn: function to populate pte
2802 * This is a helper to ease setting up page-remapped first percpu
2803 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2805 * This is the basic allocator. Static percpu area is allocated
2806 * page-by-page into vmalloc area.
2809 * 0 on success, -errno on failure.
2811 int __init pcpu_page_first_chunk(size_t reserved_size,
2812 pcpu_fc_alloc_fn_t alloc_fn,
2813 pcpu_fc_free_fn_t free_fn,
2814 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2816 static struct vm_struct vm;
2817 struct pcpu_alloc_info *ai;
2821 struct page **pages;
2822 int unit, i, j, rc = 0;
2826 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2828 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2831 BUG_ON(ai->nr_groups != 1);
2832 upa = ai->alloc_size/ai->unit_size;
2833 nr_g0_units = roundup(num_possible_cpus(), upa);
2834 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
2835 pcpu_free_alloc_info(ai);
2839 unit_pages = ai->unit_size >> PAGE_SHIFT;
2841 /* unaligned allocations can't be freed, round up to page size */
2842 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2844 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
2846 panic("%s: Failed to allocate %zu bytes\n", __func__,
2849 /* allocate pages */
2851 for (unit = 0; unit < num_possible_cpus(); unit++) {
2852 unsigned int cpu = ai->groups[0].cpu_map[unit];
2853 for (i = 0; i < unit_pages; i++) {
2856 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2858 pr_warn("failed to allocate %s page for cpu%u\n",
2862 /* kmemleak tracks the percpu allocations separately */
2864 pages[j++] = virt_to_page(ptr);
2868 /* allocate vm area, map the pages and copy static data */
2869 vm.flags = VM_ALLOC;
2870 vm.size = num_possible_cpus() * ai->unit_size;
2871 vm_area_register_early(&vm, PAGE_SIZE);
2873 for (unit = 0; unit < num_possible_cpus(); unit++) {
2874 unsigned long unit_addr =
2875 (unsigned long)vm.addr + unit * ai->unit_size;
2877 for (i = 0; i < unit_pages; i++)
2878 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2880 /* pte already populated, the following shouldn't fail */
2881 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2884 panic("failed to map percpu area, err=%d\n", rc);
2887 * FIXME: Archs with virtual cache should flush local
2888 * cache for the linear mapping here - something
2889 * equivalent to flush_cache_vmap() on the local cpu.
2890 * flush_cache_vmap() can't be used as most supporting
2891 * data structures are not set up yet.
2894 /* copy static data */
2895 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2898 /* we're ready, commit */
2899 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2900 unit_pages, psize_str, ai->static_size,
2901 ai->reserved_size, ai->dyn_size);
2903 pcpu_setup_first_chunk(ai, vm.addr);
2908 free_fn(page_address(pages[j]), PAGE_SIZE);
2911 memblock_free_early(__pa(pages), pages_size);
2912 pcpu_free_alloc_info(ai);
2915 #endif /* BUILD_PAGE_FIRST_CHUNK */
2917 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2919 * Generic SMP percpu area setup.
2921 * The embedding helper is used because its behavior closely resembles
2922 * the original non-dynamic generic percpu area setup. This is
2923 * important because many archs have addressing restrictions and might
2924 * fail if the percpu area is located far away from the previous
2925 * location. As an added bonus, in non-NUMA cases, embedding is
2926 * generally a good idea TLB-wise because percpu area can piggy back
2927 * on the physical linear memory mapping which uses large page
2928 * mappings on applicable archs.
2930 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2931 EXPORT_SYMBOL(__per_cpu_offset);
2933 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2936 return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
2939 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2941 memblock_free_early(__pa(ptr), size);
2944 void __init setup_per_cpu_areas(void)
2946 unsigned long delta;
2951 * Always reserve area for module percpu variables. That's
2952 * what the legacy allocator did.
2954 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2955 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2956 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2958 panic("Failed to initialize percpu areas.");
2960 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2961 for_each_possible_cpu(cpu)
2962 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2964 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2966 #else /* CONFIG_SMP */
2969 * UP percpu area setup.
2971 * UP always uses km-based percpu allocator with identity mapping.
2972 * Static percpu variables are indistinguishable from the usual static
2973 * variables and don't require any special preparation.
2975 void __init setup_per_cpu_areas(void)
2977 const size_t unit_size =
2978 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2979 PERCPU_DYNAMIC_RESERVE));
2980 struct pcpu_alloc_info *ai;
2983 ai = pcpu_alloc_alloc_info(1, 1);
2984 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
2986 panic("Failed to allocate memory for percpu areas.");
2987 /* kmemleak tracks the percpu allocations separately */
2990 ai->dyn_size = unit_size;
2991 ai->unit_size = unit_size;
2992 ai->atom_size = unit_size;
2993 ai->alloc_size = unit_size;
2994 ai->groups[0].nr_units = 1;
2995 ai->groups[0].cpu_map[0] = 0;
2997 pcpu_setup_first_chunk(ai, fc);
2998 pcpu_free_alloc_info(ai);
3001 #endif /* CONFIG_SMP */
3004 * pcpu_nr_pages - calculate total number of populated backing pages
3006 * This reflects the number of pages populated to back chunks. Metadata is
3007 * excluded in the number exposed in meminfo as the number of backing pages
3008 * scales with the number of cpus and can quickly outweigh the memory used for
3009 * metadata. It also keeps this calculation nice and simple.
3012 * Total number of populated backing pages in use by the allocator.
3014 unsigned long pcpu_nr_pages(void)
3016 return pcpu_nr_populated * pcpu_nr_units;
3020 * Percpu allocator is initialized early during boot when neither slab or
3021 * workqueue is available. Plug async management until everything is up
3024 static int __init percpu_enable_async(void)
3026 pcpu_async_enabled = true;
3029 subsys_initcall(percpu_enable_async);