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 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41 * flag should be passed. All memcg-aware allocations are sharing one set
42 * of chunks and all unaccounted allocations and allocations performed
43 * by processes belonging to the root memory cgroup are using the second set.
45 * The allocator tries to allocate from the fullest chunk first. Each chunk
46 * is managed by a bitmap with metadata blocks. The allocation map is updated
47 * on every allocation and free to reflect the current state while the boundary
48 * map is only updated on allocation. Each metadata block contains
49 * information to help mitigate the need to iterate over large portions
50 * of the bitmap. The reverse mapping from page to chunk is stored in
51 * the page's index. Lastly, units are lazily backed and grow in unison.
53 * There is a unique conversion that goes on here between bytes and bits.
54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55 * tracks the number of pages it is responsible for in nr_pages. Helper
56 * functions are used to convert from between the bytes, bits, and blocks.
57 * All hints are managed in bits unless explicitly stated.
59 * To use this allocator, arch code should do the following:
61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62 * regular address to percpu pointer and back if they need to be
63 * different from the default
65 * - use pcpu_setup_first_chunk() during percpu area initialization to
66 * setup the first chunk containing the kernel static percpu area
69 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
71 #include <linux/bitmap.h>
72 #include <linux/cpumask.h>
73 #include <linux/memblock.h>
74 #include <linux/err.h>
75 #include <linux/lcm.h>
76 #include <linux/list.h>
77 #include <linux/log2.h>
79 #include <linux/module.h>
80 #include <linux/mutex.h>
81 #include <linux/percpu.h>
82 #include <linux/pfn.h>
83 #include <linux/slab.h>
84 #include <linux/spinlock.h>
85 #include <linux/vmalloc.h>
86 #include <linux/workqueue.h>
87 #include <linux/kmemleak.h>
88 #include <linux/sched.h>
89 #include <linux/sched/mm.h>
90 #include <linux/memcontrol.h>
92 #include <asm/cacheflush.h>
93 #include <asm/sections.h>
94 #include <asm/tlbflush.h>
97 #define CREATE_TRACE_POINTS
98 #include <trace/events/percpu.h>
100 #include "percpu-internal.h"
103 * The slots are sorted by the size of the biggest continuous free area.
104 * 1-31 bytes share the same slot.
106 #define PCPU_SLOT_BASE_SHIFT 5
107 /* chunks in slots below this are subject to being sidelined on failed alloc */
108 #define PCPU_SLOT_FAIL_THRESHOLD 3
110 #define PCPU_EMPTY_POP_PAGES_LOW 2
111 #define PCPU_EMPTY_POP_PAGES_HIGH 4
114 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
115 #ifndef __addr_to_pcpu_ptr
116 #define __addr_to_pcpu_ptr(addr) \
117 (void __percpu *)((unsigned long)(addr) - \
118 (unsigned long)pcpu_base_addr + \
119 (unsigned long)__per_cpu_start)
121 #ifndef __pcpu_ptr_to_addr
122 #define __pcpu_ptr_to_addr(ptr) \
123 (void __force *)((unsigned long)(ptr) + \
124 (unsigned long)pcpu_base_addr - \
125 (unsigned long)__per_cpu_start)
127 #else /* CONFIG_SMP */
128 /* on UP, it's always identity mapped */
129 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
130 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
131 #endif /* CONFIG_SMP */
133 static int pcpu_unit_pages __ro_after_init;
134 static int pcpu_unit_size __ro_after_init;
135 static int pcpu_nr_units __ro_after_init;
136 static int pcpu_atom_size __ro_after_init;
137 int pcpu_nr_slots __ro_after_init;
138 static int pcpu_free_slot __ro_after_init;
139 int pcpu_sidelined_slot __ro_after_init;
140 int pcpu_to_depopulate_slot __ro_after_init;
141 static size_t pcpu_chunk_struct_size __ro_after_init;
143 /* cpus with the lowest and highest unit addresses */
144 static unsigned int pcpu_low_unit_cpu __ro_after_init;
145 static unsigned int pcpu_high_unit_cpu __ro_after_init;
147 /* the address of the first chunk which starts with the kernel static area */
148 void *pcpu_base_addr __ro_after_init;
150 static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
151 const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
153 /* group information, used for vm allocation */
154 static int pcpu_nr_groups __ro_after_init;
155 static const unsigned long *pcpu_group_offsets __ro_after_init;
156 static const size_t *pcpu_group_sizes __ro_after_init;
159 * The first chunk which always exists. Note that unlike other
160 * chunks, this one can be allocated and mapped in several different
161 * ways and thus often doesn't live in the vmalloc area.
163 struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
166 * Optional reserved chunk. This chunk reserves part of the first
167 * chunk and serves it for reserved allocations. When the reserved
168 * region doesn't exist, the following variable is NULL.
170 struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
172 DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
173 static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
175 struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
177 /* chunks which need their map areas extended, protected by pcpu_lock */
178 static LIST_HEAD(pcpu_map_extend_chunks);
181 * The number of empty populated pages, protected by pcpu_lock.
182 * The reserved chunk doesn't contribute to the count.
184 int pcpu_nr_empty_pop_pages;
187 * The number of populated pages in use by the allocator, protected by
188 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
189 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
190 * and increments/decrements this count by 1).
192 static unsigned long pcpu_nr_populated;
195 * Balance work is used to populate or destroy chunks asynchronously. We
196 * try to keep the number of populated free pages between
197 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
200 static void pcpu_balance_workfn(struct work_struct *work);
201 static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
202 static bool pcpu_async_enabled __read_mostly;
203 static bool pcpu_atomic_alloc_failed;
205 static void pcpu_schedule_balance_work(void)
207 if (pcpu_async_enabled)
208 schedule_work(&pcpu_balance_work);
212 * pcpu_addr_in_chunk - check if the address is served from this chunk
213 * @chunk: chunk of interest
214 * @addr: percpu address
217 * True if the address is served from this chunk.
219 static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
221 void *start_addr, *end_addr;
226 start_addr = chunk->base_addr + chunk->start_offset;
227 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
230 return addr >= start_addr && addr < end_addr;
233 static int __pcpu_size_to_slot(int size)
235 int highbit = fls(size); /* size is in bytes */
236 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
239 static int pcpu_size_to_slot(int size)
241 if (size == pcpu_unit_size)
242 return pcpu_free_slot;
243 return __pcpu_size_to_slot(size);
246 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
248 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
250 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
251 chunk_md->contig_hint == 0)
254 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
257 /* set the pointer to a chunk in a page struct */
258 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
260 page->index = (unsigned long)pcpu;
263 /* obtain pointer to a chunk from a page struct */
264 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
266 return (struct pcpu_chunk *)page->index;
269 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
271 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
274 static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
276 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
279 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
280 unsigned int cpu, int page_idx)
282 return (unsigned long)chunk->base_addr +
283 pcpu_unit_page_offset(cpu, page_idx);
287 * The following are helper functions to help access bitmaps and convert
288 * between bitmap offsets to address offsets.
290 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
292 return chunk->alloc_map +
293 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
296 static unsigned long pcpu_off_to_block_index(int off)
298 return off / PCPU_BITMAP_BLOCK_BITS;
301 static unsigned long pcpu_off_to_block_off(int off)
303 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
306 static unsigned long pcpu_block_off_to_off(int index, int off)
308 return index * PCPU_BITMAP_BLOCK_BITS + off;
312 * pcpu_check_block_hint - check against the contig hint
313 * @block: block of interest
314 * @bits: size of allocation
315 * @align: alignment of area (max PAGE_SIZE)
317 * Check to see if the allocation can fit in the block's contig hint.
318 * Note, a chunk uses the same hints as a block so this can also check against
319 * the chunk's contig hint.
321 static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
324 int bit_off = ALIGN(block->contig_hint_start, align) -
325 block->contig_hint_start;
327 return bit_off + bits <= block->contig_hint;
331 * pcpu_next_hint - determine which hint to use
332 * @block: block of interest
333 * @alloc_bits: size of allocation
335 * This determines if we should scan based on the scan_hint or first_free.
336 * In general, we want to scan from first_free to fulfill allocations by
337 * first fit. However, if we know a scan_hint at position scan_hint_start
338 * cannot fulfill an allocation, we can begin scanning from there knowing
339 * the contig_hint will be our fallback.
341 static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
344 * The three conditions below determine if we can skip past the
345 * scan_hint. First, does the scan hint exist. Second, is the
346 * contig_hint after the scan_hint (possibly not true iff
347 * contig_hint == scan_hint). Third, is the allocation request
348 * larger than the scan_hint.
350 if (block->scan_hint &&
351 block->contig_hint_start > block->scan_hint_start &&
352 alloc_bits > block->scan_hint)
353 return block->scan_hint_start + block->scan_hint;
355 return block->first_free;
359 * pcpu_next_md_free_region - finds the next hint free area
360 * @chunk: chunk of interest
361 * @bit_off: chunk offset
362 * @bits: size of free area
364 * Helper function for pcpu_for_each_md_free_region. It checks
365 * block->contig_hint and performs aggregation across blocks to find the
366 * next hint. It modifies bit_off and bits in-place to be consumed in the
369 static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
372 int i = pcpu_off_to_block_index(*bit_off);
373 int block_off = pcpu_off_to_block_off(*bit_off);
374 struct pcpu_block_md *block;
377 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
379 /* handles contig area across blocks */
381 *bits += block->left_free;
382 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
388 * This checks three things. First is there a contig_hint to
389 * check. Second, have we checked this hint before by
390 * comparing the block_off. Third, is this the same as the
391 * right contig hint. In the last case, it spills over into
392 * the next block and should be handled by the contig area
393 * across blocks code.
395 *bits = block->contig_hint;
396 if (*bits && block->contig_hint_start >= block_off &&
397 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
398 *bit_off = pcpu_block_off_to_off(i,
399 block->contig_hint_start);
402 /* reset to satisfy the second predicate above */
405 *bits = block->right_free;
406 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
411 * pcpu_next_fit_region - finds fit areas for a given allocation request
412 * @chunk: chunk of interest
413 * @alloc_bits: size of allocation
414 * @align: alignment of area (max PAGE_SIZE)
415 * @bit_off: chunk offset
416 * @bits: size of free area
418 * Finds the next free region that is viable for use with a given size and
419 * alignment. This only returns if there is a valid area to be used for this
420 * allocation. block->first_free is returned if the allocation request fits
421 * within the block to see if the request can be fulfilled prior to the contig
424 static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
425 int align, int *bit_off, int *bits)
427 int i = pcpu_off_to_block_index(*bit_off);
428 int block_off = pcpu_off_to_block_off(*bit_off);
429 struct pcpu_block_md *block;
432 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
434 /* handles contig area across blocks */
436 *bits += block->left_free;
437 if (*bits >= alloc_bits)
439 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
443 /* check block->contig_hint */
444 *bits = ALIGN(block->contig_hint_start, align) -
445 block->contig_hint_start;
447 * This uses the block offset to determine if this has been
448 * checked in the prior iteration.
450 if (block->contig_hint &&
451 block->contig_hint_start >= block_off &&
452 block->contig_hint >= *bits + alloc_bits) {
453 int start = pcpu_next_hint(block, alloc_bits);
455 *bits += alloc_bits + block->contig_hint_start -
457 *bit_off = pcpu_block_off_to_off(i, start);
460 /* reset to satisfy the second predicate above */
463 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
465 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
466 *bit_off = pcpu_block_off_to_off(i, *bit_off);
467 if (*bits >= alloc_bits)
471 /* no valid offsets were found - fail condition */
472 *bit_off = pcpu_chunk_map_bits(chunk);
476 * Metadata free area iterators. These perform aggregation of free areas
477 * based on the metadata blocks and return the offset @bit_off and size in
478 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
479 * a fit is found for the allocation request.
481 #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
482 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
483 (bit_off) < pcpu_chunk_map_bits((chunk)); \
484 (bit_off) += (bits) + 1, \
485 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
487 #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
488 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
490 (bit_off) < pcpu_chunk_map_bits((chunk)); \
491 (bit_off) += (bits), \
492 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
496 * pcpu_mem_zalloc - allocate memory
497 * @size: bytes to allocate
498 * @gfp: allocation flags
500 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
501 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
502 * This is to facilitate passing through whitelisted flags. The
503 * returned memory is always zeroed.
506 * Pointer to the allocated area on success, NULL on failure.
508 static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
510 if (WARN_ON_ONCE(!slab_is_available()))
513 if (size <= PAGE_SIZE)
514 return kzalloc(size, gfp);
516 return __vmalloc(size, gfp | __GFP_ZERO);
520 * pcpu_mem_free - free memory
521 * @ptr: memory to free
523 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
525 static void pcpu_mem_free(void *ptr)
530 static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
533 if (chunk != pcpu_reserved_chunk) {
535 list_move(&chunk->list, &pcpu_chunk_lists[slot]);
537 list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
541 static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
543 __pcpu_chunk_move(chunk, slot, true);
547 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
548 * @chunk: chunk of interest
549 * @oslot: the previous slot it was on
551 * This function is called after an allocation or free changed @chunk.
552 * New slot according to the changed state is determined and @chunk is
553 * moved to the slot. Note that the reserved chunk is never put on
559 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
561 int nslot = pcpu_chunk_slot(chunk);
563 /* leave isolated chunks in-place */
568 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
571 static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
573 lockdep_assert_held(&pcpu_lock);
575 if (!chunk->isolated) {
576 chunk->isolated = true;
577 pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
579 list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
582 static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
584 lockdep_assert_held(&pcpu_lock);
586 if (chunk->isolated) {
587 chunk->isolated = false;
588 pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
589 pcpu_chunk_relocate(chunk, -1);
594 * pcpu_update_empty_pages - update empty page counters
595 * @chunk: chunk of interest
596 * @nr: nr of empty pages
598 * This is used to keep track of the empty pages now based on the premise
599 * a md_block covers a page. The hint update functions recognize if a block
600 * is made full or broken to calculate deltas for keeping track of free pages.
602 static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
604 chunk->nr_empty_pop_pages += nr;
605 if (chunk != pcpu_reserved_chunk && !chunk->isolated)
606 pcpu_nr_empty_pop_pages += nr;
610 * pcpu_region_overlap - determines if two regions overlap
611 * @a: start of first region, inclusive
612 * @b: end of first region, exclusive
613 * @x: start of second region, inclusive
614 * @y: end of second region, exclusive
616 * This is used to determine if the hint region [a, b) overlaps with the
617 * allocated region [x, y).
619 static inline bool pcpu_region_overlap(int a, int b, int x, int y)
621 return (a < y) && (x < b);
625 * pcpu_block_update - updates a block given a free area
626 * @block: block of interest
627 * @start: start offset in block
628 * @end: end offset in block
630 * Updates a block given a known free area. The region [start, end) is
631 * expected to be the entirety of the free area within a block. Chooses
632 * the best starting offset if the contig hints are equal.
634 static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
636 int contig = end - start;
638 block->first_free = min(block->first_free, start);
640 block->left_free = contig;
642 if (end == block->nr_bits)
643 block->right_free = contig;
645 if (contig > block->contig_hint) {
646 /* promote the old contig_hint to be the new scan_hint */
647 if (start > block->contig_hint_start) {
648 if (block->contig_hint > block->scan_hint) {
649 block->scan_hint_start =
650 block->contig_hint_start;
651 block->scan_hint = block->contig_hint;
652 } else if (start < block->scan_hint_start) {
654 * The old contig_hint == scan_hint. But, the
655 * new contig is larger so hold the invariant
656 * scan_hint_start < contig_hint_start.
658 block->scan_hint = 0;
661 block->scan_hint = 0;
663 block->contig_hint_start = start;
664 block->contig_hint = contig;
665 } else if (contig == block->contig_hint) {
666 if (block->contig_hint_start &&
668 __ffs(start) > __ffs(block->contig_hint_start))) {
669 /* start has a better alignment so use it */
670 block->contig_hint_start = start;
671 if (start < block->scan_hint_start &&
672 block->contig_hint > block->scan_hint)
673 block->scan_hint = 0;
674 } else if (start > block->scan_hint_start ||
675 block->contig_hint > block->scan_hint) {
677 * Knowing contig == contig_hint, update the scan_hint
678 * if it is farther than or larger than the current
681 block->scan_hint_start = start;
682 block->scan_hint = contig;
686 * The region is smaller than the contig_hint. So only update
687 * the scan_hint if it is larger than or equal and farther than
688 * the current scan_hint.
690 if ((start < block->contig_hint_start &&
691 (contig > block->scan_hint ||
692 (contig == block->scan_hint &&
693 start > block->scan_hint_start)))) {
694 block->scan_hint_start = start;
695 block->scan_hint = contig;
701 * pcpu_block_update_scan - update a block given a free area from a scan
702 * @chunk: chunk of interest
703 * @bit_off: chunk offset
704 * @bits: size of free area
706 * Finding the final allocation spot first goes through pcpu_find_block_fit()
707 * to find a block that can hold the allocation and then pcpu_alloc_area()
708 * where a scan is used. When allocations require specific alignments,
709 * we can inadvertently create holes which will not be seen in the alloc
712 * This takes a given free area hole and updates a block as it may change the
713 * scan_hint. We need to scan backwards to ensure we don't miss free bits
716 static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
719 int s_off = pcpu_off_to_block_off(bit_off);
720 int e_off = s_off + bits;
722 struct pcpu_block_md *block;
724 if (e_off > PCPU_BITMAP_BLOCK_BITS)
727 s_index = pcpu_off_to_block_index(bit_off);
728 block = chunk->md_blocks + s_index;
730 /* scan backwards in case of alignment skipping free bits */
731 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
732 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
734 pcpu_block_update(block, s_off, e_off);
738 * pcpu_chunk_refresh_hint - updates metadata about a chunk
739 * @chunk: chunk of interest
740 * @full_scan: if we should scan from the beginning
742 * Iterates over the metadata blocks to find the largest contig area.
743 * A full scan can be avoided on the allocation path as this is triggered
744 * if we broke the contig_hint. In doing so, the scan_hint will be before
745 * the contig_hint or after if the scan_hint == contig_hint. This cannot
746 * be prevented on freeing as we want to find the largest area possibly
749 static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
751 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
754 /* promote scan_hint to contig_hint */
755 if (!full_scan && chunk_md->scan_hint) {
756 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
757 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
758 chunk_md->contig_hint = chunk_md->scan_hint;
759 chunk_md->scan_hint = 0;
761 bit_off = chunk_md->first_free;
762 chunk_md->contig_hint = 0;
766 pcpu_for_each_md_free_region(chunk, bit_off, bits)
767 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
771 * pcpu_block_refresh_hint
772 * @chunk: chunk of interest
773 * @index: index of the metadata block
775 * Scans over the block beginning at first_free and updates the block
776 * metadata accordingly.
778 static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
780 struct pcpu_block_md *block = chunk->md_blocks + index;
781 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
782 unsigned int start, end; /* region start, region end */
784 /* promote scan_hint to contig_hint */
785 if (block->scan_hint) {
786 start = block->scan_hint_start + block->scan_hint;
787 block->contig_hint_start = block->scan_hint_start;
788 block->contig_hint = block->scan_hint;
789 block->scan_hint = 0;
791 start = block->first_free;
792 block->contig_hint = 0;
795 block->right_free = 0;
797 /* iterate over free areas and update the contig hints */
798 for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
799 pcpu_block_update(block, start, end);
803 * pcpu_block_update_hint_alloc - update hint on allocation path
804 * @chunk: chunk of interest
805 * @bit_off: chunk offset
806 * @bits: size of request
808 * Updates metadata for the allocation path. The metadata only has to be
809 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
810 * scans are required if the block's contig hint is broken.
812 static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
815 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
816 int nr_empty_pages = 0;
817 struct pcpu_block_md *s_block, *e_block, *block;
818 int s_index, e_index; /* block indexes of the freed allocation */
819 int s_off, e_off; /* block offsets of the freed allocation */
822 * Calculate per block offsets.
823 * The calculation uses an inclusive range, but the resulting offsets
824 * are [start, end). e_index always points to the last block in the
827 s_index = pcpu_off_to_block_index(bit_off);
828 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
829 s_off = pcpu_off_to_block_off(bit_off);
830 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
832 s_block = chunk->md_blocks + s_index;
833 e_block = chunk->md_blocks + e_index;
837 * block->first_free must be updated if the allocation takes its place.
838 * If the allocation breaks the contig_hint, a scan is required to
841 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
844 if (s_off == s_block->first_free)
845 s_block->first_free = find_next_zero_bit(
846 pcpu_index_alloc_map(chunk, s_index),
847 PCPU_BITMAP_BLOCK_BITS,
850 if (pcpu_region_overlap(s_block->scan_hint_start,
851 s_block->scan_hint_start + s_block->scan_hint,
854 s_block->scan_hint = 0;
856 if (pcpu_region_overlap(s_block->contig_hint_start,
857 s_block->contig_hint_start +
858 s_block->contig_hint,
861 /* block contig hint is broken - scan to fix it */
863 s_block->left_free = 0;
864 pcpu_block_refresh_hint(chunk, s_index);
866 /* update left and right contig manually */
867 s_block->left_free = min(s_block->left_free, s_off);
868 if (s_index == e_index)
869 s_block->right_free = min_t(int, s_block->right_free,
870 PCPU_BITMAP_BLOCK_BITS - e_off);
872 s_block->right_free = 0;
878 if (s_index != e_index) {
879 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
883 * When the allocation is across blocks, the end is along
884 * the left part of the e_block.
886 e_block->first_free = find_next_zero_bit(
887 pcpu_index_alloc_map(chunk, e_index),
888 PCPU_BITMAP_BLOCK_BITS, e_off);
890 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
891 /* reset the block */
894 if (e_off > e_block->scan_hint_start)
895 e_block->scan_hint = 0;
897 e_block->left_free = 0;
898 if (e_off > e_block->contig_hint_start) {
899 /* contig hint is broken - scan to fix it */
900 pcpu_block_refresh_hint(chunk, e_index);
902 e_block->right_free =
903 min_t(int, e_block->right_free,
904 PCPU_BITMAP_BLOCK_BITS - e_off);
908 /* update in-between md_blocks */
909 nr_empty_pages += (e_index - s_index - 1);
910 for (block = s_block + 1; block < e_block; block++) {
911 block->scan_hint = 0;
912 block->contig_hint = 0;
913 block->left_free = 0;
914 block->right_free = 0;
919 pcpu_update_empty_pages(chunk, -nr_empty_pages);
921 if (pcpu_region_overlap(chunk_md->scan_hint_start,
922 chunk_md->scan_hint_start +
926 chunk_md->scan_hint = 0;
929 * The only time a full chunk scan is required is if the chunk
930 * contig hint is broken. Otherwise, it means a smaller space
931 * was used and therefore the chunk contig hint is still correct.
933 if (pcpu_region_overlap(chunk_md->contig_hint_start,
934 chunk_md->contig_hint_start +
935 chunk_md->contig_hint,
938 pcpu_chunk_refresh_hint(chunk, false);
942 * pcpu_block_update_hint_free - updates the block hints on the free path
943 * @chunk: chunk of interest
944 * @bit_off: chunk offset
945 * @bits: size of request
947 * Updates metadata for the allocation path. This avoids a blind block
948 * refresh by making use of the block contig hints. If this fails, it scans
949 * forward and backward to determine the extent of the free area. This is
950 * capped at the boundary of blocks.
952 * A chunk update is triggered if a page becomes free, a block becomes free,
953 * or the free spans across blocks. This tradeoff is to minimize iterating
954 * over the block metadata to update chunk_md->contig_hint.
955 * chunk_md->contig_hint may be off by up to a page, but it will never be more
956 * than the available space. If the contig hint is contained in one block, it
959 static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
962 int nr_empty_pages = 0;
963 struct pcpu_block_md *s_block, *e_block, *block;
964 int s_index, e_index; /* block indexes of the freed allocation */
965 int s_off, e_off; /* block offsets of the freed allocation */
966 int start, end; /* start and end of the whole free area */
969 * Calculate per block offsets.
970 * The calculation uses an inclusive range, but the resulting offsets
971 * are [start, end). e_index always points to the last block in the
974 s_index = pcpu_off_to_block_index(bit_off);
975 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
976 s_off = pcpu_off_to_block_off(bit_off);
977 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
979 s_block = chunk->md_blocks + s_index;
980 e_block = chunk->md_blocks + e_index;
983 * Check if the freed area aligns with the block->contig_hint.
984 * If it does, then the scan to find the beginning/end of the
985 * larger free area can be avoided.
987 * start and end refer to beginning and end of the free area
988 * within each their respective blocks. This is not necessarily
989 * the entire free area as it may span blocks past the beginning
990 * or end of the block.
993 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
994 start = s_block->contig_hint_start;
997 * Scan backwards to find the extent of the free area.
998 * find_last_bit returns the starting bit, so if the start bit
999 * is returned, that means there was no last bit and the
1000 * remainder of the chunk is free.
1002 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1004 start = (start == l_bit) ? 0 : l_bit + 1;
1008 if (e_off == e_block->contig_hint_start)
1009 end = e_block->contig_hint_start + e_block->contig_hint;
1011 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1012 PCPU_BITMAP_BLOCK_BITS, end);
1014 /* update s_block */
1015 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1016 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1018 pcpu_block_update(s_block, start, e_off);
1020 /* freeing in the same block */
1021 if (s_index != e_index) {
1022 /* update e_block */
1023 if (end == PCPU_BITMAP_BLOCK_BITS)
1025 pcpu_block_update(e_block, 0, end);
1027 /* reset md_blocks in the middle */
1028 nr_empty_pages += (e_index - s_index - 1);
1029 for (block = s_block + 1; block < e_block; block++) {
1030 block->first_free = 0;
1031 block->scan_hint = 0;
1032 block->contig_hint_start = 0;
1033 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1034 block->left_free = PCPU_BITMAP_BLOCK_BITS;
1035 block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040 pcpu_update_empty_pages(chunk, nr_empty_pages);
1043 * Refresh chunk metadata when the free makes a block free or spans
1044 * across blocks. The contig_hint may be off by up to a page, but if
1045 * the contig_hint is contained in a block, it will be accurate with
1046 * the else condition below.
1048 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1049 pcpu_chunk_refresh_hint(chunk, true);
1051 pcpu_block_update(&chunk->chunk_md,
1052 pcpu_block_off_to_off(s_index, start),
1057 * pcpu_is_populated - determines if the region is populated
1058 * @chunk: chunk of interest
1059 * @bit_off: chunk offset
1060 * @bits: size of area
1061 * @next_off: return value for the next offset to start searching
1063 * For atomic allocations, check if the backing pages are populated.
1066 * Bool if the backing pages are populated.
1067 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1069 static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1072 unsigned int start, end;
1074 start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1075 end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1077 start = find_next_zero_bit(chunk->populated, end, start);
1081 end = find_next_bit(chunk->populated, end, start + 1);
1083 *next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088 * pcpu_find_block_fit - finds the block index to start searching
1089 * @chunk: chunk of interest
1090 * @alloc_bits: size of request in allocation units
1091 * @align: alignment of area (max PAGE_SIZE bytes)
1092 * @pop_only: use populated regions only
1094 * Given a chunk and an allocation spec, find the offset to begin searching
1095 * for a free region. This iterates over the bitmap metadata blocks to
1096 * find an offset that will be guaranteed to fit the requirements. It is
1097 * not quite first fit as if the allocation does not fit in the contig hint
1098 * of a block or chunk, it is skipped. This errs on the side of caution
1099 * to prevent excess iteration. Poor alignment can cause the allocator to
1100 * skip over blocks and chunks that have valid free areas.
1103 * The offset in the bitmap to begin searching.
1104 * -1 if no offset is found.
1106 static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1107 size_t align, bool pop_only)
1109 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1110 int bit_off, bits, next_off;
1113 * This is an optimization to prevent scanning by assuming if the
1114 * allocation cannot fit in the global hint, there is memory pressure
1115 * and creating a new chunk would happen soon.
1117 if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1120 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1122 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1123 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1131 if (bit_off == pcpu_chunk_map_bits(chunk))
1138 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1139 * @map: the address to base the search on
1140 * @size: the bitmap size in bits
1141 * @start: the bitnumber to start searching at
1142 * @nr: the number of zeroed bits we're looking for
1143 * @align_mask: alignment mask for zero area
1144 * @largest_off: offset of the largest area skipped
1145 * @largest_bits: size of the largest area skipped
1147 * The @align_mask should be one less than a power of 2.
1149 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1150 * the largest area that was skipped. This is imperfect, but in general is
1151 * good enough. The largest remembered region is the largest failed region
1152 * seen. This does not include anything we possibly skipped due to alignment.
1153 * pcpu_block_update_scan() does scan backwards to try and recover what was
1154 * lost to alignment. While this can cause scanning to miss earlier possible
1155 * free areas, smaller allocations will eventually fill those holes.
1157 static unsigned long pcpu_find_zero_area(unsigned long *map,
1159 unsigned long start,
1161 unsigned long align_mask,
1162 unsigned long *largest_off,
1163 unsigned long *largest_bits)
1165 unsigned long index, end, i, area_off, area_bits;
1167 index = find_next_zero_bit(map, size, start);
1169 /* Align allocation */
1170 index = __ALIGN_MASK(index, align_mask);
1176 i = find_next_bit(map, end, index);
1178 area_bits = i - area_off;
1179 /* remember largest unused area with best alignment */
1180 if (area_bits > *largest_bits ||
1181 (area_bits == *largest_bits && *largest_off &&
1182 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1183 *largest_off = area_off;
1184 *largest_bits = area_bits;
1194 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1195 * @chunk: chunk of interest
1196 * @alloc_bits: size of request in allocation units
1197 * @align: alignment of area (max PAGE_SIZE)
1198 * @start: bit_off to start searching
1200 * This function takes in a @start offset to begin searching to fit an
1201 * allocation of @alloc_bits with alignment @align. It needs to scan
1202 * the allocation map because if it fits within the block's contig hint,
1203 * @start will be block->first_free. This is an attempt to fill the
1204 * allocation prior to breaking the contig hint. The allocation and
1205 * boundary maps are updated accordingly if it confirms a valid
1209 * Allocated addr offset in @chunk on success.
1210 * -1 if no matching area is found.
1212 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1213 size_t align, int start)
1215 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1216 size_t align_mask = (align) ? (align - 1) : 0;
1217 unsigned long area_off = 0, area_bits = 0;
1218 int bit_off, end, oslot;
1220 lockdep_assert_held(&pcpu_lock);
1222 oslot = pcpu_chunk_slot(chunk);
1225 * Search to find a fit.
1227 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1228 pcpu_chunk_map_bits(chunk));
1229 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1230 align_mask, &area_off, &area_bits);
1235 pcpu_block_update_scan(chunk, area_off, area_bits);
1237 /* update alloc map */
1238 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1240 /* update boundary map */
1241 set_bit(bit_off, chunk->bound_map);
1242 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1243 set_bit(bit_off + alloc_bits, chunk->bound_map);
1245 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1247 /* update first free bit */
1248 if (bit_off == chunk_md->first_free)
1249 chunk_md->first_free = find_next_zero_bit(
1251 pcpu_chunk_map_bits(chunk),
1252 bit_off + alloc_bits);
1254 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1256 pcpu_chunk_relocate(chunk, oslot);
1258 return bit_off * PCPU_MIN_ALLOC_SIZE;
1262 * pcpu_free_area - frees the corresponding offset
1263 * @chunk: chunk of interest
1264 * @off: addr offset into chunk
1266 * This function determines the size of an allocation to free using
1267 * the boundary bitmap and clears the allocation map.
1270 * Number of freed bytes.
1272 static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1274 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1275 int bit_off, bits, end, oslot, freed;
1277 lockdep_assert_held(&pcpu_lock);
1278 pcpu_stats_area_dealloc(chunk);
1280 oslot = pcpu_chunk_slot(chunk);
1282 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1284 /* find end index */
1285 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1287 bits = end - bit_off;
1288 bitmap_clear(chunk->alloc_map, bit_off, bits);
1290 freed = bits * PCPU_MIN_ALLOC_SIZE;
1292 /* update metadata */
1293 chunk->free_bytes += freed;
1295 /* update first free bit */
1296 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1298 pcpu_block_update_hint_free(chunk, bit_off, bits);
1300 pcpu_chunk_relocate(chunk, oslot);
1305 static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1307 block->scan_hint = 0;
1308 block->contig_hint = nr_bits;
1309 block->left_free = nr_bits;
1310 block->right_free = nr_bits;
1311 block->first_free = 0;
1312 block->nr_bits = nr_bits;
1315 static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1317 struct pcpu_block_md *md_block;
1319 /* init the chunk's block */
1320 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1322 for (md_block = chunk->md_blocks;
1323 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1325 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1329 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1330 * @tmp_addr: the start of the region served
1331 * @map_size: size of the region served
1333 * This is responsible for creating the chunks that serve the first chunk. The
1334 * base_addr is page aligned down of @tmp_addr while the region end is page
1335 * aligned up. Offsets are kept track of to determine the region served. All
1336 * this is done to appease the bitmap allocator in avoiding partial blocks.
1339 * Chunk serving the region at @tmp_addr of @map_size.
1341 static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1344 struct pcpu_chunk *chunk;
1345 unsigned long aligned_addr, lcm_align;
1346 int start_offset, offset_bits, region_size, region_bits;
1349 /* region calculations */
1350 aligned_addr = tmp_addr & PAGE_MASK;
1352 start_offset = tmp_addr - aligned_addr;
1355 * Align the end of the region with the LCM of PAGE_SIZE and
1356 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1359 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1360 region_size = ALIGN(start_offset + map_size, lcm_align);
1362 /* allocate chunk */
1363 alloc_size = struct_size(chunk, populated,
1364 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1365 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1367 panic("%s: Failed to allocate %zu bytes\n", __func__,
1370 INIT_LIST_HEAD(&chunk->list);
1372 chunk->base_addr = (void *)aligned_addr;
1373 chunk->start_offset = start_offset;
1374 chunk->end_offset = region_size - chunk->start_offset - map_size;
1376 chunk->nr_pages = region_size >> PAGE_SHIFT;
1377 region_bits = pcpu_chunk_map_bits(chunk);
1379 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1380 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1381 if (!chunk->alloc_map)
1382 panic("%s: Failed to allocate %zu bytes\n", __func__,
1386 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1387 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1388 if (!chunk->bound_map)
1389 panic("%s: Failed to allocate %zu bytes\n", __func__,
1392 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1393 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1394 if (!chunk->md_blocks)
1395 panic("%s: Failed to allocate %zu bytes\n", __func__,
1398 #ifdef CONFIG_MEMCG_KMEM
1399 /* first chunk is free to use */
1400 chunk->obj_cgroups = NULL;
1402 pcpu_init_md_blocks(chunk);
1404 /* manage populated page bitmap */
1405 chunk->immutable = true;
1406 bitmap_fill(chunk->populated, chunk->nr_pages);
1407 chunk->nr_populated = chunk->nr_pages;
1408 chunk->nr_empty_pop_pages = chunk->nr_pages;
1410 chunk->free_bytes = map_size;
1412 if (chunk->start_offset) {
1413 /* hide the beginning of the bitmap */
1414 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1415 bitmap_set(chunk->alloc_map, 0, offset_bits);
1416 set_bit(0, chunk->bound_map);
1417 set_bit(offset_bits, chunk->bound_map);
1419 chunk->chunk_md.first_free = offset_bits;
1421 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1424 if (chunk->end_offset) {
1425 /* hide the end of the bitmap */
1426 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1427 bitmap_set(chunk->alloc_map,
1428 pcpu_chunk_map_bits(chunk) - offset_bits,
1430 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1432 set_bit(region_bits, chunk->bound_map);
1434 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1435 - offset_bits, offset_bits);
1441 static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1443 struct pcpu_chunk *chunk;
1446 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1450 INIT_LIST_HEAD(&chunk->list);
1451 chunk->nr_pages = pcpu_unit_pages;
1452 region_bits = pcpu_chunk_map_bits(chunk);
1454 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1455 sizeof(chunk->alloc_map[0]), gfp);
1456 if (!chunk->alloc_map)
1457 goto alloc_map_fail;
1459 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1460 sizeof(chunk->bound_map[0]), gfp);
1461 if (!chunk->bound_map)
1462 goto bound_map_fail;
1464 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1465 sizeof(chunk->md_blocks[0]), gfp);
1466 if (!chunk->md_blocks)
1467 goto md_blocks_fail;
1469 #ifdef CONFIG_MEMCG_KMEM
1470 if (!mem_cgroup_kmem_disabled()) {
1471 chunk->obj_cgroups =
1472 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1473 sizeof(struct obj_cgroup *), gfp);
1474 if (!chunk->obj_cgroups)
1479 pcpu_init_md_blocks(chunk);
1482 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1486 #ifdef CONFIG_MEMCG_KMEM
1488 pcpu_mem_free(chunk->md_blocks);
1491 pcpu_mem_free(chunk->bound_map);
1493 pcpu_mem_free(chunk->alloc_map);
1495 pcpu_mem_free(chunk);
1500 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1504 #ifdef CONFIG_MEMCG_KMEM
1505 pcpu_mem_free(chunk->obj_cgroups);
1507 pcpu_mem_free(chunk->md_blocks);
1508 pcpu_mem_free(chunk->bound_map);
1509 pcpu_mem_free(chunk->alloc_map);
1510 pcpu_mem_free(chunk);
1514 * pcpu_chunk_populated - post-population bookkeeping
1515 * @chunk: pcpu_chunk which got populated
1516 * @page_start: the start page
1517 * @page_end: the end page
1519 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1520 * the bookkeeping information accordingly. Must be called after each
1521 * successful population.
1523 static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1526 int nr = page_end - page_start;
1528 lockdep_assert_held(&pcpu_lock);
1530 bitmap_set(chunk->populated, page_start, nr);
1531 chunk->nr_populated += nr;
1532 pcpu_nr_populated += nr;
1534 pcpu_update_empty_pages(chunk, nr);
1538 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1539 * @chunk: pcpu_chunk which got depopulated
1540 * @page_start: the start page
1541 * @page_end: the end page
1543 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1544 * Update the bookkeeping information accordingly. Must be called after
1545 * each successful depopulation.
1547 static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1548 int page_start, int page_end)
1550 int nr = page_end - page_start;
1552 lockdep_assert_held(&pcpu_lock);
1554 bitmap_clear(chunk->populated, page_start, nr);
1555 chunk->nr_populated -= nr;
1556 pcpu_nr_populated -= nr;
1558 pcpu_update_empty_pages(chunk, -nr);
1562 * Chunk management implementation.
1564 * To allow different implementations, chunk alloc/free and
1565 * [de]population are implemented in a separate file which is pulled
1566 * into this file and compiled together. The following functions
1567 * should be implemented.
1569 * pcpu_populate_chunk - populate the specified range of a chunk
1570 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1571 * pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk
1572 * pcpu_create_chunk - create a new chunk
1573 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1574 * pcpu_addr_to_page - translate address to physical address
1575 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1577 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1578 int page_start, int page_end, gfp_t gfp);
1579 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1580 int page_start, int page_end);
1581 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1582 int page_start, int page_end);
1583 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1584 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1585 static struct page *pcpu_addr_to_page(void *addr);
1586 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1588 #ifdef CONFIG_NEED_PER_CPU_KM
1589 #include "percpu-km.c"
1591 #include "percpu-vm.c"
1595 * pcpu_chunk_addr_search - determine chunk containing specified address
1596 * @addr: address for which the chunk needs to be determined.
1598 * This is an internal function that handles all but static allocations.
1599 * Static percpu address values should never be passed into the allocator.
1602 * The address of the found chunk.
1604 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1606 /* is it in the dynamic region (first chunk)? */
1607 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1608 return pcpu_first_chunk;
1610 /* is it in the reserved region? */
1611 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1612 return pcpu_reserved_chunk;
1615 * The address is relative to unit0 which might be unused and
1616 * thus unmapped. Offset the address to the unit space of the
1617 * current processor before looking it up in the vmalloc
1618 * space. Note that any possible cpu id can be used here, so
1619 * there's no need to worry about preemption or cpu hotplug.
1621 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1622 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1625 #ifdef CONFIG_MEMCG_KMEM
1626 static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1627 struct obj_cgroup **objcgp)
1629 struct obj_cgroup *objcg;
1631 if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT))
1634 objcg = get_obj_cgroup_from_current();
1638 if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size))) {
1639 obj_cgroup_put(objcg);
1647 static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1648 struct pcpu_chunk *chunk, int off,
1654 if (likely(chunk && chunk->obj_cgroups)) {
1655 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1658 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1659 pcpu_obj_full_size(size));
1662 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1663 obj_cgroup_put(objcg);
1667 static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1669 struct obj_cgroup *objcg;
1671 if (unlikely(!chunk->obj_cgroups))
1674 objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1677 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1679 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1682 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1683 -pcpu_obj_full_size(size));
1686 obj_cgroup_put(objcg);
1689 #else /* CONFIG_MEMCG_KMEM */
1691 pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1696 static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1697 struct pcpu_chunk *chunk, int off,
1702 static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1705 #endif /* CONFIG_MEMCG_KMEM */
1708 * pcpu_alloc - the percpu allocator
1709 * @size: size of area to allocate in bytes
1710 * @align: alignment of area (max PAGE_SIZE)
1711 * @reserved: allocate from the reserved chunk if available
1712 * @gfp: allocation flags
1714 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1715 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1716 * then no warning will be triggered on invalid or failed allocation
1720 * Percpu pointer to the allocated area on success, NULL on failure.
1722 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1728 struct obj_cgroup *objcg = NULL;
1729 static int warn_limit = 10;
1730 struct pcpu_chunk *chunk, *next;
1732 int slot, off, cpu, ret;
1733 unsigned long flags;
1735 size_t bits, bit_align;
1737 gfp = current_gfp_context(gfp);
1738 /* whitelisted flags that can be passed to the backing allocators */
1739 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1740 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1741 do_warn = !(gfp & __GFP_NOWARN);
1744 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1745 * therefore alignment must be a minimum of that many bytes.
1746 * An allocation may have internal fragmentation from rounding up
1747 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1749 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1750 align = PCPU_MIN_ALLOC_SIZE;
1752 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1753 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1754 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1756 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1757 !is_power_of_2(align))) {
1758 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1763 if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1768 * pcpu_balance_workfn() allocates memory under this mutex,
1769 * and it may wait for memory reclaim. Allow current task
1770 * to become OOM victim, in case of memory pressure.
1772 if (gfp & __GFP_NOFAIL) {
1773 mutex_lock(&pcpu_alloc_mutex);
1774 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1775 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1780 spin_lock_irqsave(&pcpu_lock, flags);
1782 /* serve reserved allocations from the reserved chunk if available */
1783 if (reserved && pcpu_reserved_chunk) {
1784 chunk = pcpu_reserved_chunk;
1786 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1788 err = "alloc from reserved chunk failed";
1792 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1796 err = "alloc from reserved chunk failed";
1801 /* search through normal chunks */
1802 for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1803 list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1805 off = pcpu_find_block_fit(chunk, bits, bit_align,
1808 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1809 pcpu_chunk_move(chunk, 0);
1813 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1815 pcpu_reintegrate_chunk(chunk);
1821 spin_unlock_irqrestore(&pcpu_lock, flags);
1824 * No space left. Create a new chunk. We don't want multiple
1825 * tasks to create chunks simultaneously. Serialize and create iff
1826 * there's still no empty chunk after grabbing the mutex.
1829 err = "atomic alloc failed, no space left";
1833 if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1834 chunk = pcpu_create_chunk(pcpu_gfp);
1836 err = "failed to allocate new chunk";
1840 spin_lock_irqsave(&pcpu_lock, flags);
1841 pcpu_chunk_relocate(chunk, -1);
1843 spin_lock_irqsave(&pcpu_lock, flags);
1849 pcpu_stats_area_alloc(chunk, size);
1850 spin_unlock_irqrestore(&pcpu_lock, flags);
1852 /* populate if not all pages are already there */
1854 unsigned int page_end, rs, re;
1857 page_end = PFN_UP(off + size);
1859 for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
1860 WARN_ON(chunk->immutable);
1862 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1864 spin_lock_irqsave(&pcpu_lock, flags);
1866 pcpu_free_area(chunk, off);
1867 err = "failed to populate";
1870 pcpu_chunk_populated(chunk, rs, re);
1871 spin_unlock_irqrestore(&pcpu_lock, flags);
1874 mutex_unlock(&pcpu_alloc_mutex);
1877 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1878 pcpu_schedule_balance_work();
1880 /* clear the areas and return address relative to base address */
1881 for_each_possible_cpu(cpu)
1882 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1884 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1885 kmemleak_alloc_percpu(ptr, size, gfp);
1887 trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1888 chunk->base_addr, off, ptr,
1889 pcpu_obj_full_size(size), gfp);
1891 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1896 spin_unlock_irqrestore(&pcpu_lock, flags);
1898 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1900 if (!is_atomic && do_warn && warn_limit) {
1901 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1902 size, align, is_atomic, err);
1905 pr_info("limit reached, disable warning\n");
1908 /* see the flag handling in pcpu_balance_workfn() */
1909 pcpu_atomic_alloc_failed = true;
1910 pcpu_schedule_balance_work();
1912 mutex_unlock(&pcpu_alloc_mutex);
1915 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1921 * __alloc_percpu_gfp - allocate dynamic percpu area
1922 * @size: size of area to allocate in bytes
1923 * @align: alignment of area (max PAGE_SIZE)
1924 * @gfp: allocation flags
1926 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1927 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1928 * be called from any context but is a lot more likely to fail. If @gfp
1929 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1930 * allocation requests.
1933 * Percpu pointer to the allocated area on success, NULL on failure.
1935 void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1937 return pcpu_alloc(size, align, false, gfp);
1939 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1942 * __alloc_percpu - allocate dynamic percpu area
1943 * @size: size of area to allocate in bytes
1944 * @align: alignment of area (max PAGE_SIZE)
1946 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1948 void __percpu *__alloc_percpu(size_t size, size_t align)
1950 return pcpu_alloc(size, align, false, GFP_KERNEL);
1952 EXPORT_SYMBOL_GPL(__alloc_percpu);
1955 * __alloc_reserved_percpu - allocate reserved percpu area
1956 * @size: size of area to allocate in bytes
1957 * @align: alignment of area (max PAGE_SIZE)
1959 * Allocate zero-filled percpu area of @size bytes aligned at @align
1960 * from reserved percpu area if arch has set it up; otherwise,
1961 * allocation is served from the same dynamic area. Might sleep.
1962 * Might trigger writeouts.
1965 * Does GFP_KERNEL allocation.
1968 * Percpu pointer to the allocated area on success, NULL on failure.
1970 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1972 return pcpu_alloc(size, align, true, GFP_KERNEL);
1976 * pcpu_balance_free - manage the amount of free chunks
1977 * @empty_only: free chunks only if there are no populated pages
1979 * If empty_only is %false, reclaim all fully free chunks regardless of the
1980 * number of populated pages. Otherwise, only reclaim chunks that have no
1984 * pcpu_lock (can be dropped temporarily)
1986 static void pcpu_balance_free(bool empty_only)
1989 struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1990 struct pcpu_chunk *chunk, *next;
1992 lockdep_assert_held(&pcpu_lock);
1995 * There's no reason to keep around multiple unused chunks and VM
1996 * areas can be scarce. Destroy all free chunks except for one.
1998 list_for_each_entry_safe(chunk, next, free_head, list) {
1999 WARN_ON(chunk->immutable);
2001 /* spare the first one */
2002 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
2005 if (!empty_only || chunk->nr_empty_pop_pages == 0)
2006 list_move(&chunk->list, &to_free);
2009 if (list_empty(&to_free))
2012 spin_unlock_irq(&pcpu_lock);
2013 list_for_each_entry_safe(chunk, next, &to_free, list) {
2014 unsigned int rs, re;
2016 for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2017 pcpu_depopulate_chunk(chunk, rs, re);
2018 spin_lock_irq(&pcpu_lock);
2019 pcpu_chunk_depopulated(chunk, rs, re);
2020 spin_unlock_irq(&pcpu_lock);
2022 pcpu_destroy_chunk(chunk);
2025 spin_lock_irq(&pcpu_lock);
2029 * pcpu_balance_populated - manage the amount of populated pages
2031 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2032 * It is possible that this is called when physical memory is scarce causing
2033 * OOM killer to be triggered. We should avoid doing so until an actual
2034 * allocation causes the failure as it is possible that requests can be
2035 * serviced from already backed regions.
2038 * pcpu_lock (can be dropped temporarily)
2040 static void pcpu_balance_populated(void)
2042 /* gfp flags passed to underlying allocators */
2043 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2044 struct pcpu_chunk *chunk;
2045 int slot, nr_to_pop, ret;
2047 lockdep_assert_held(&pcpu_lock);
2050 * Ensure there are certain number of free populated pages for
2051 * atomic allocs. Fill up from the most packed so that atomic
2052 * allocs don't increase fragmentation. If atomic allocation
2053 * failed previously, always populate the maximum amount. This
2054 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2055 * failing indefinitely; however, large atomic allocs are not
2056 * something we support properly and can be highly unreliable and
2060 if (pcpu_atomic_alloc_failed) {
2061 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2062 /* best effort anyway, don't worry about synchronization */
2063 pcpu_atomic_alloc_failed = false;
2065 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2066 pcpu_nr_empty_pop_pages,
2067 0, PCPU_EMPTY_POP_PAGES_HIGH);
2070 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2071 unsigned int nr_unpop = 0, rs, re;
2076 list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2077 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2085 /* @chunk can't go away while pcpu_alloc_mutex is held */
2086 for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2087 int nr = min_t(int, re - rs, nr_to_pop);
2089 spin_unlock_irq(&pcpu_lock);
2090 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2092 spin_lock_irq(&pcpu_lock);
2095 pcpu_chunk_populated(chunk, rs, rs + nr);
2106 /* ran out of chunks to populate, create a new one and retry */
2107 spin_unlock_irq(&pcpu_lock);
2108 chunk = pcpu_create_chunk(gfp);
2110 spin_lock_irq(&pcpu_lock);
2112 pcpu_chunk_relocate(chunk, -1);
2119 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2121 * Scan over chunks in the depopulate list and try to release unused populated
2122 * pages back to the system. Depopulated chunks are sidelined to prevent
2123 * repopulating these pages unless required. Fully free chunks are reintegrated
2124 * and freed accordingly (1 is kept around). If we drop below the empty
2125 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2126 * Each chunk is scanned in the reverse order to keep populated pages close to
2127 * the beginning of the chunk.
2130 * pcpu_lock (can be dropped temporarily)
2133 static void pcpu_reclaim_populated(void)
2135 struct pcpu_chunk *chunk;
2136 struct pcpu_block_md *block;
2137 int freed_page_start, freed_page_end;
2141 lockdep_assert_held(&pcpu_lock);
2144 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2145 * longer discoverable to allocations whom may populate pages. The only
2146 * other accessor is the free path which only returns area back to the
2147 * allocator not touching the populated bitmap.
2149 while (!list_empty(&pcpu_chunk_lists[pcpu_to_depopulate_slot])) {
2150 chunk = list_first_entry(&pcpu_chunk_lists[pcpu_to_depopulate_slot],
2151 struct pcpu_chunk, list);
2152 WARN_ON(chunk->immutable);
2155 * Scan chunk's pages in the reverse order to keep populated
2156 * pages close to the beginning of the chunk.
2158 freed_page_start = chunk->nr_pages;
2160 reintegrate = false;
2161 for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2162 /* no more work to do */
2163 if (chunk->nr_empty_pop_pages == 0)
2166 /* reintegrate chunk to prevent atomic alloc failures */
2167 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2173 * If the page is empty and populated, start or
2174 * extend the (i, end) range. If i == 0, decrease
2175 * i and perform the depopulation to cover the last
2176 * (first) page in the chunk.
2178 block = chunk->md_blocks + i;
2179 if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2180 test_bit(i, chunk->populated)) {
2188 /* depopulate if there is an active range */
2192 spin_unlock_irq(&pcpu_lock);
2193 pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2195 spin_lock_irq(&pcpu_lock);
2197 pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2198 freed_page_start = min(freed_page_start, i + 1);
2199 freed_page_end = max(freed_page_end, end + 1);
2201 /* reset the range and continue */
2206 /* batch tlb flush per chunk to amortize cost */
2207 if (freed_page_start < freed_page_end) {
2208 spin_unlock_irq(&pcpu_lock);
2209 pcpu_post_unmap_tlb_flush(chunk,
2213 spin_lock_irq(&pcpu_lock);
2216 if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2217 pcpu_reintegrate_chunk(chunk);
2219 list_move_tail(&chunk->list,
2220 &pcpu_chunk_lists[pcpu_sidelined_slot]);
2225 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2228 * For each chunk type, manage the number of fully free chunks and the number of
2229 * populated pages. An important thing to consider is when pages are freed and
2230 * how they contribute to the global counts.
2232 static void pcpu_balance_workfn(struct work_struct *work)
2235 * pcpu_balance_free() is called twice because the first time we may
2236 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2237 * to grow other chunks. This then gives pcpu_reclaim_populated() time
2238 * to move fully free chunks to the active list to be freed if
2241 mutex_lock(&pcpu_alloc_mutex);
2242 spin_lock_irq(&pcpu_lock);
2244 pcpu_balance_free(false);
2245 pcpu_reclaim_populated();
2246 pcpu_balance_populated();
2247 pcpu_balance_free(true);
2249 spin_unlock_irq(&pcpu_lock);
2250 mutex_unlock(&pcpu_alloc_mutex);
2254 * free_percpu - free percpu area
2255 * @ptr: pointer to area to free
2257 * Free percpu area @ptr.
2260 * Can be called from atomic context.
2262 void free_percpu(void __percpu *ptr)
2265 struct pcpu_chunk *chunk;
2266 unsigned long flags;
2268 bool need_balance = false;
2273 kmemleak_free_percpu(ptr);
2275 addr = __pcpu_ptr_to_addr(ptr);
2277 spin_lock_irqsave(&pcpu_lock, flags);
2279 chunk = pcpu_chunk_addr_search(addr);
2280 off = addr - chunk->base_addr;
2282 size = pcpu_free_area(chunk, off);
2284 pcpu_memcg_free_hook(chunk, off, size);
2287 * If there are more than one fully free chunks, wake up grim reaper.
2288 * If the chunk is isolated, it may be in the process of being
2289 * reclaimed. Let reclaim manage cleaning up of that chunk.
2291 if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2292 struct pcpu_chunk *pos;
2294 list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2296 need_balance = true;
2299 } else if (pcpu_should_reclaim_chunk(chunk)) {
2300 pcpu_isolate_chunk(chunk);
2301 need_balance = true;
2304 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2306 spin_unlock_irqrestore(&pcpu_lock, flags);
2309 pcpu_schedule_balance_work();
2311 EXPORT_SYMBOL_GPL(free_percpu);
2313 bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2316 const size_t static_size = __per_cpu_end - __per_cpu_start;
2317 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2320 for_each_possible_cpu(cpu) {
2321 void *start = per_cpu_ptr(base, cpu);
2322 void *va = (void *)addr;
2324 if (va >= start && va < start + static_size) {
2326 *can_addr = (unsigned long) (va - start);
2327 *can_addr += (unsigned long)
2328 per_cpu_ptr(base, get_boot_cpu_id());
2334 /* on UP, can't distinguish from other static vars, always false */
2339 * is_kernel_percpu_address - test whether address is from static percpu area
2340 * @addr: address to test
2342 * Test whether @addr belongs to in-kernel static percpu area. Module
2343 * static percpu areas are not considered. For those, use
2344 * is_module_percpu_address().
2347 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2349 bool is_kernel_percpu_address(unsigned long addr)
2351 return __is_kernel_percpu_address(addr, NULL);
2355 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2356 * @addr: the address to be converted to physical address
2358 * Given @addr which is dereferenceable address obtained via one of
2359 * percpu access macros, this function translates it into its physical
2360 * address. The caller is responsible for ensuring @addr stays valid
2361 * until this function finishes.
2363 * percpu allocator has special setup for the first chunk, which currently
2364 * supports either embedding in linear address space or vmalloc mapping,
2365 * and, from the second one, the backing allocator (currently either vm or
2366 * km) provides translation.
2368 * The addr can be translated simply without checking if it falls into the
2369 * first chunk. But the current code reflects better how percpu allocator
2370 * actually works, and the verification can discover both bugs in percpu
2371 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2375 * The physical address for @addr.
2377 phys_addr_t per_cpu_ptr_to_phys(void *addr)
2379 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2380 bool in_first_chunk = false;
2381 unsigned long first_low, first_high;
2385 * The following test on unit_low/high isn't strictly
2386 * necessary but will speed up lookups of addresses which
2387 * aren't in the first chunk.
2389 * The address check is against full chunk sizes. pcpu_base_addr
2390 * points to the beginning of the first chunk including the
2391 * static region. Assumes good intent as the first chunk may
2392 * not be full (ie. < pcpu_unit_pages in size).
2394 first_low = (unsigned long)pcpu_base_addr +
2395 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2396 first_high = (unsigned long)pcpu_base_addr +
2397 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2398 if ((unsigned long)addr >= first_low &&
2399 (unsigned long)addr < first_high) {
2400 for_each_possible_cpu(cpu) {
2401 void *start = per_cpu_ptr(base, cpu);
2403 if (addr >= start && addr < start + pcpu_unit_size) {
2404 in_first_chunk = true;
2410 if (in_first_chunk) {
2411 if (!is_vmalloc_addr(addr))
2414 return page_to_phys(vmalloc_to_page(addr)) +
2415 offset_in_page(addr);
2417 return page_to_phys(pcpu_addr_to_page(addr)) +
2418 offset_in_page(addr);
2422 * pcpu_alloc_alloc_info - allocate percpu allocation info
2423 * @nr_groups: the number of groups
2424 * @nr_units: the number of units
2426 * Allocate ai which is large enough for @nr_groups groups containing
2427 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2428 * cpu_map array which is long enough for @nr_units and filled with
2429 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2430 * pointer of other groups.
2433 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2436 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2439 struct pcpu_alloc_info *ai;
2440 size_t base_size, ai_size;
2444 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2445 __alignof__(ai->groups[0].cpu_map[0]));
2446 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2448 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2454 ai->groups[0].cpu_map = ptr;
2456 for (unit = 0; unit < nr_units; unit++)
2457 ai->groups[0].cpu_map[unit] = NR_CPUS;
2459 ai->nr_groups = nr_groups;
2460 ai->__ai_size = PFN_ALIGN(ai_size);
2466 * pcpu_free_alloc_info - free percpu allocation info
2467 * @ai: pcpu_alloc_info to free
2469 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2471 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2473 memblock_free(ai, ai->__ai_size);
2477 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2479 * @ai: allocation info to dump
2481 * Print out information about @ai using loglevel @lvl.
2483 static void pcpu_dump_alloc_info(const char *lvl,
2484 const struct pcpu_alloc_info *ai)
2486 int group_width = 1, cpu_width = 1, width;
2487 char empty_str[] = "--------";
2488 int alloc = 0, alloc_end = 0;
2490 int upa, apl; /* units per alloc, allocs per line */
2496 v = num_possible_cpus();
2499 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2501 upa = ai->alloc_size / ai->unit_size;
2502 width = upa * (cpu_width + 1) + group_width + 3;
2503 apl = rounddown_pow_of_two(max(60 / width, 1));
2505 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2506 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2507 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2509 for (group = 0; group < ai->nr_groups; group++) {
2510 const struct pcpu_group_info *gi = &ai->groups[group];
2511 int unit = 0, unit_end = 0;
2513 BUG_ON(gi->nr_units % upa);
2514 for (alloc_end += gi->nr_units / upa;
2515 alloc < alloc_end; alloc++) {
2516 if (!(alloc % apl)) {
2518 printk("%spcpu-alloc: ", lvl);
2520 pr_cont("[%0*d] ", group_width, group);
2522 for (unit_end += upa; unit < unit_end; unit++)
2523 if (gi->cpu_map[unit] != NR_CPUS)
2525 cpu_width, gi->cpu_map[unit]);
2527 pr_cont("%s ", empty_str);
2534 * pcpu_setup_first_chunk - initialize the first percpu chunk
2535 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2536 * @base_addr: mapped address
2538 * Initialize the first percpu chunk which contains the kernel static
2539 * percpu area. This function is to be called from arch percpu area
2542 * @ai contains all information necessary to initialize the first
2543 * chunk and prime the dynamic percpu allocator.
2545 * @ai->static_size is the size of static percpu area.
2547 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2548 * reserve after the static area in the first chunk. This reserves
2549 * the first chunk such that it's available only through reserved
2550 * percpu allocation. This is primarily used to serve module percpu
2551 * static areas on architectures where the addressing model has
2552 * limited offset range for symbol relocations to guarantee module
2553 * percpu symbols fall inside the relocatable range.
2555 * @ai->dyn_size determines the number of bytes available for dynamic
2556 * allocation in the first chunk. The area between @ai->static_size +
2557 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2559 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2560 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2563 * @ai->atom_size is the allocation atom size and used as alignment
2566 * @ai->alloc_size is the allocation size and always multiple of
2567 * @ai->atom_size. This is larger than @ai->atom_size if
2568 * @ai->unit_size is larger than @ai->atom_size.
2570 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2571 * percpu areas. Units which should be colocated are put into the
2572 * same group. Dynamic VM areas will be allocated according to these
2573 * groupings. If @ai->nr_groups is zero, a single group containing
2574 * all units is assumed.
2576 * The caller should have mapped the first chunk at @base_addr and
2577 * copied static data to each unit.
2579 * The first chunk will always contain a static and a dynamic region.
2580 * However, the static region is not managed by any chunk. If the first
2581 * chunk also contains a reserved region, it is served by two chunks -
2582 * one for the reserved region and one for the dynamic region. They
2583 * share the same vm, but use offset regions in the area allocation map.
2584 * The chunk serving the dynamic region is circulated in the chunk slots
2585 * and available for dynamic allocation like any other chunk.
2587 void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2590 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2591 size_t static_size, dyn_size;
2592 struct pcpu_chunk *chunk;
2593 unsigned long *group_offsets;
2594 size_t *group_sizes;
2595 unsigned long *unit_off;
2600 unsigned long tmp_addr;
2603 #define PCPU_SETUP_BUG_ON(cond) do { \
2604 if (unlikely(cond)) { \
2605 pr_emerg("failed to initialize, %s\n", #cond); \
2606 pr_emerg("cpu_possible_mask=%*pb\n", \
2607 cpumask_pr_args(cpu_possible_mask)); \
2608 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2614 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2616 PCPU_SETUP_BUG_ON(!ai->static_size);
2617 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2619 PCPU_SETUP_BUG_ON(!base_addr);
2620 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2621 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2622 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2623 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2624 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2625 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2626 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2627 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2628 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2629 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2630 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2632 /* process group information and build config tables accordingly */
2633 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2634 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2636 panic("%s: Failed to allocate %zu bytes\n", __func__,
2639 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2640 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2642 panic("%s: Failed to allocate %zu bytes\n", __func__,
2645 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2646 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2648 panic("%s: Failed to allocate %zu bytes\n", __func__,
2651 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2652 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2654 panic("%s: Failed to allocate %zu bytes\n", __func__,
2657 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2658 unit_map[cpu] = UINT_MAX;
2660 pcpu_low_unit_cpu = NR_CPUS;
2661 pcpu_high_unit_cpu = NR_CPUS;
2663 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2664 const struct pcpu_group_info *gi = &ai->groups[group];
2666 group_offsets[group] = gi->base_offset;
2667 group_sizes[group] = gi->nr_units * ai->unit_size;
2669 for (i = 0; i < gi->nr_units; i++) {
2670 cpu = gi->cpu_map[i];
2674 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2675 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2676 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2678 unit_map[cpu] = unit + i;
2679 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2681 /* determine low/high unit_cpu */
2682 if (pcpu_low_unit_cpu == NR_CPUS ||
2683 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2684 pcpu_low_unit_cpu = cpu;
2685 if (pcpu_high_unit_cpu == NR_CPUS ||
2686 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2687 pcpu_high_unit_cpu = cpu;
2690 pcpu_nr_units = unit;
2692 for_each_possible_cpu(cpu)
2693 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2695 /* we're done parsing the input, undefine BUG macro and dump config */
2696 #undef PCPU_SETUP_BUG_ON
2697 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2699 pcpu_nr_groups = ai->nr_groups;
2700 pcpu_group_offsets = group_offsets;
2701 pcpu_group_sizes = group_sizes;
2702 pcpu_unit_map = unit_map;
2703 pcpu_unit_offsets = unit_off;
2705 /* determine basic parameters */
2706 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2707 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2708 pcpu_atom_size = ai->atom_size;
2709 pcpu_chunk_struct_size = struct_size(chunk, populated,
2710 BITS_TO_LONGS(pcpu_unit_pages));
2712 pcpu_stats_save_ai(ai);
2715 * Allocate chunk slots. The slots after the active slots are:
2716 * sidelined_slot - isolated, depopulated chunks
2717 * free_slot - fully free chunks
2718 * to_depopulate_slot - isolated, chunks to depopulate
2720 pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2721 pcpu_free_slot = pcpu_sidelined_slot + 1;
2722 pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2723 pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2724 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2725 sizeof(pcpu_chunk_lists[0]),
2727 if (!pcpu_chunk_lists)
2728 panic("%s: Failed to allocate %zu bytes\n", __func__,
2729 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2731 for (i = 0; i < pcpu_nr_slots; i++)
2732 INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2735 * The end of the static region needs to be aligned with the
2736 * minimum allocation size as this offsets the reserved and
2737 * dynamic region. The first chunk ends page aligned by
2738 * expanding the dynamic region, therefore the dynamic region
2739 * can be shrunk to compensate while still staying above the
2742 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2743 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2746 * Initialize first chunk.
2747 * If the reserved_size is non-zero, this initializes the reserved
2748 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2749 * and the dynamic region is initialized here. The first chunk,
2750 * pcpu_first_chunk, will always point to the chunk that serves
2751 * the dynamic region.
2753 tmp_addr = (unsigned long)base_addr + static_size;
2754 map_size = ai->reserved_size ?: dyn_size;
2755 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2757 /* init dynamic chunk if necessary */
2758 if (ai->reserved_size) {
2759 pcpu_reserved_chunk = chunk;
2761 tmp_addr = (unsigned long)base_addr + static_size +
2763 map_size = dyn_size;
2764 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2767 /* link the first chunk in */
2768 pcpu_first_chunk = chunk;
2769 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2770 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2772 /* include all regions of the first chunk */
2773 pcpu_nr_populated += PFN_DOWN(size_sum);
2775 pcpu_stats_chunk_alloc();
2776 trace_percpu_create_chunk(base_addr);
2779 pcpu_base_addr = base_addr;
2784 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2785 [PCPU_FC_AUTO] = "auto",
2786 [PCPU_FC_EMBED] = "embed",
2787 [PCPU_FC_PAGE] = "page",
2790 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2792 static int __init percpu_alloc_setup(char *str)
2799 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2800 else if (!strcmp(str, "embed"))
2801 pcpu_chosen_fc = PCPU_FC_EMBED;
2803 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2804 else if (!strcmp(str, "page"))
2805 pcpu_chosen_fc = PCPU_FC_PAGE;
2808 pr_warn("unknown allocator %s specified\n", str);
2812 early_param("percpu_alloc", percpu_alloc_setup);
2815 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2816 * Build it if needed by the arch config or the generic setup is going
2819 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2820 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2821 #define BUILD_EMBED_FIRST_CHUNK
2824 /* build pcpu_page_first_chunk() iff needed by the arch config */
2825 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2826 #define BUILD_PAGE_FIRST_CHUNK
2829 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2830 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2832 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2833 * @reserved_size: the size of reserved percpu area in bytes
2834 * @dyn_size: minimum free size for dynamic allocation in bytes
2835 * @atom_size: allocation atom size
2836 * @cpu_distance_fn: callback to determine distance between cpus, optional
2838 * This function determines grouping of units, their mappings to cpus
2839 * and other parameters considering needed percpu size, allocation
2840 * atom size and distances between CPUs.
2842 * Groups are always multiples of atom size and CPUs which are of
2843 * LOCAL_DISTANCE both ways are grouped together and share space for
2844 * units in the same group. The returned configuration is guaranteed
2845 * to have CPUs on different nodes on different groups and >=75% usage
2846 * of allocated virtual address space.
2849 * On success, pointer to the new allocation_info is returned. On
2850 * failure, ERR_PTR value is returned.
2852 static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2853 size_t reserved_size, size_t dyn_size,
2855 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2857 static int group_map[NR_CPUS] __initdata;
2858 static int group_cnt[NR_CPUS] __initdata;
2859 static struct cpumask mask __initdata;
2860 const size_t static_size = __per_cpu_end - __per_cpu_start;
2861 int nr_groups = 1, nr_units = 0;
2862 size_t size_sum, min_unit_size, alloc_size;
2863 int upa, max_upa, best_upa; /* units_per_alloc */
2864 int last_allocs, group, unit;
2865 unsigned int cpu, tcpu;
2866 struct pcpu_alloc_info *ai;
2867 unsigned int *cpu_map;
2869 /* this function may be called multiple times */
2870 memset(group_map, 0, sizeof(group_map));
2871 memset(group_cnt, 0, sizeof(group_cnt));
2872 cpumask_clear(&mask);
2874 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2875 size_sum = PFN_ALIGN(static_size + reserved_size +
2876 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2877 dyn_size = size_sum - static_size - reserved_size;
2880 * Determine min_unit_size, alloc_size and max_upa such that
2881 * alloc_size is multiple of atom_size and is the smallest
2882 * which can accommodate 4k aligned segments which are equal to
2883 * or larger than min_unit_size.
2885 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2887 /* determine the maximum # of units that can fit in an allocation */
2888 alloc_size = roundup(min_unit_size, atom_size);
2889 upa = alloc_size / min_unit_size;
2890 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2894 cpumask_copy(&mask, cpu_possible_mask);
2896 /* group cpus according to their proximity */
2897 for (group = 0; !cpumask_empty(&mask); group++) {
2898 /* pop the group's first cpu */
2899 cpu = cpumask_first(&mask);
2900 group_map[cpu] = group;
2902 cpumask_clear_cpu(cpu, &mask);
2904 for_each_cpu(tcpu, &mask) {
2905 if (!cpu_distance_fn ||
2906 (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2907 cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2908 group_map[tcpu] = group;
2910 cpumask_clear_cpu(tcpu, &mask);
2917 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2918 * Expand the unit_size until we use >= 75% of the units allocated.
2919 * Related to atom_size, which could be much larger than the unit_size.
2921 last_allocs = INT_MAX;
2923 for (upa = max_upa; upa; upa--) {
2924 int allocs = 0, wasted = 0;
2926 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2929 for (group = 0; group < nr_groups; group++) {
2930 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2931 allocs += this_allocs;
2932 wasted += this_allocs * upa - group_cnt[group];
2936 * Don't accept if wastage is over 1/3. The
2937 * greater-than comparison ensures upa==1 always
2938 * passes the following check.
2940 if (wasted > num_possible_cpus() / 3)
2943 /* and then don't consume more memory */
2944 if (allocs > last_allocs)
2946 last_allocs = allocs;
2952 /* allocate and fill alloc_info */
2953 for (group = 0; group < nr_groups; group++)
2954 nr_units += roundup(group_cnt[group], upa);
2956 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2958 return ERR_PTR(-ENOMEM);
2959 cpu_map = ai->groups[0].cpu_map;
2961 for (group = 0; group < nr_groups; group++) {
2962 ai->groups[group].cpu_map = cpu_map;
2963 cpu_map += roundup(group_cnt[group], upa);
2966 ai->static_size = static_size;
2967 ai->reserved_size = reserved_size;
2968 ai->dyn_size = dyn_size;
2969 ai->unit_size = alloc_size / upa;
2970 ai->atom_size = atom_size;
2971 ai->alloc_size = alloc_size;
2973 for (group = 0, unit = 0; group < nr_groups; group++) {
2974 struct pcpu_group_info *gi = &ai->groups[group];
2977 * Initialize base_offset as if all groups are located
2978 * back-to-back. The caller should update this to
2979 * reflect actual allocation.
2981 gi->base_offset = unit * ai->unit_size;
2983 for_each_possible_cpu(cpu)
2984 if (group_map[cpu] == group)
2985 gi->cpu_map[gi->nr_units++] = cpu;
2986 gi->nr_units = roundup(gi->nr_units, upa);
2987 unit += gi->nr_units;
2989 BUG_ON(unit != nr_units);
2994 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
2995 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
2997 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
2999 int node = NUMA_NO_NODE;
3003 node = cpu_to_nd_fn(cpu);
3005 if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
3006 ptr = memblock_alloc_from(size, align, goal);
3007 pr_info("cpu %d has no node %d or node-local memory\n",
3009 pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
3010 cpu, size, (u64)__pa(ptr));
3012 ptr = memblock_alloc_try_nid(size, align, goal,
3013 MEMBLOCK_ALLOC_ACCESSIBLE,
3016 pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
3017 cpu, size, node, (u64)__pa(ptr));
3021 return memblock_alloc_from(size, align, goal);
3025 static void __init pcpu_fc_free(void *ptr, size_t size)
3027 memblock_free(ptr, size);
3029 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3031 #if defined(BUILD_EMBED_FIRST_CHUNK)
3033 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3034 * @reserved_size: the size of reserved percpu area in bytes
3035 * @dyn_size: minimum free size for dynamic allocation in bytes
3036 * @atom_size: allocation atom size
3037 * @cpu_distance_fn: callback to determine distance between cpus, optional
3038 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3040 * This is a helper to ease setting up embedded first percpu chunk and
3041 * can be called where pcpu_setup_first_chunk() is expected.
3043 * If this function is used to setup the first chunk, it is allocated
3044 * by calling pcpu_fc_alloc and used as-is without being mapped into
3045 * vmalloc area. Allocations are always whole multiples of @atom_size
3046 * aligned to @atom_size.
3048 * This enables the first chunk to piggy back on the linear physical
3049 * mapping which often uses larger page size. Please note that this
3050 * can result in very sparse cpu->unit mapping on NUMA machines thus
3051 * requiring large vmalloc address space. Don't use this allocator if
3052 * vmalloc space is not orders of magnitude larger than distances
3053 * between node memory addresses (ie. 32bit NUMA machines).
3055 * @dyn_size specifies the minimum dynamic area size.
3057 * If the needed size is smaller than the minimum or specified unit
3058 * size, the leftover is returned using pcpu_fc_free.
3061 * 0 on success, -errno on failure.
3063 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3065 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3066 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3068 void *base = (void *)ULONG_MAX;
3069 void **areas = NULL;
3070 struct pcpu_alloc_info *ai;
3071 size_t size_sum, areas_size;
3072 unsigned long max_distance;
3073 int group, i, highest_group, rc = 0;
3075 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3080 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3081 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3083 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3089 /* allocate, copy and determine base address & max_distance */
3091 for (group = 0; group < ai->nr_groups; group++) {
3092 struct pcpu_group_info *gi = &ai->groups[group];
3093 unsigned int cpu = NR_CPUS;
3096 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3097 cpu = gi->cpu_map[i];
3098 BUG_ON(cpu == NR_CPUS);
3100 /* allocate space for the whole group */
3101 ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn);
3104 goto out_free_areas;
3106 /* kmemleak tracks the percpu allocations separately */
3110 base = min(ptr, base);
3111 if (ptr > areas[highest_group])
3112 highest_group = group;
3114 max_distance = areas[highest_group] - base;
3115 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3117 /* warn if maximum distance is further than 75% of vmalloc space */
3118 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3119 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3120 max_distance, VMALLOC_TOTAL);
3121 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3122 /* and fail if we have fallback */
3124 goto out_free_areas;
3129 * Copy data and free unused parts. This should happen after all
3130 * allocations are complete; otherwise, we may end up with
3131 * overlapping groups.
3133 for (group = 0; group < ai->nr_groups; group++) {
3134 struct pcpu_group_info *gi = &ai->groups[group];
3135 void *ptr = areas[group];
3137 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3138 if (gi->cpu_map[i] == NR_CPUS) {
3139 /* unused unit, free whole */
3140 pcpu_fc_free(ptr, ai->unit_size);
3143 /* copy and return the unused part */
3144 memcpy(ptr, __per_cpu_load, ai->static_size);
3145 pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum);
3149 /* base address is now known, determine group base offsets */
3150 for (group = 0; group < ai->nr_groups; group++) {
3151 ai->groups[group].base_offset = areas[group] - base;
3154 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3155 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3156 ai->dyn_size, ai->unit_size);
3158 pcpu_setup_first_chunk(ai, base);
3162 for (group = 0; group < ai->nr_groups; group++)
3164 pcpu_fc_free(areas[group],
3165 ai->groups[group].nr_units * ai->unit_size);
3167 pcpu_free_alloc_info(ai);
3169 memblock_free(areas, areas_size);
3172 #endif /* BUILD_EMBED_FIRST_CHUNK */
3174 #ifdef BUILD_PAGE_FIRST_CHUNK
3175 #include <asm/pgalloc.h>
3177 #ifndef P4D_TABLE_SIZE
3178 #define P4D_TABLE_SIZE PAGE_SIZE
3181 #ifndef PUD_TABLE_SIZE
3182 #define PUD_TABLE_SIZE PAGE_SIZE
3185 #ifndef PMD_TABLE_SIZE
3186 #define PMD_TABLE_SIZE PAGE_SIZE
3189 #ifndef PTE_TABLE_SIZE
3190 #define PTE_TABLE_SIZE PAGE_SIZE
3192 void __init __weak pcpu_populate_pte(unsigned long addr)
3194 pgd_t *pgd = pgd_offset_k(addr);
3199 if (pgd_none(*pgd)) {
3202 new = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3205 pgd_populate(&init_mm, pgd, new);
3208 p4d = p4d_offset(pgd, addr);
3209 if (p4d_none(*p4d)) {
3212 new = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3215 p4d_populate(&init_mm, p4d, new);
3218 pud = pud_offset(p4d, addr);
3219 if (pud_none(*pud)) {
3222 new = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3225 pud_populate(&init_mm, pud, new);
3228 pmd = pmd_offset(pud, addr);
3229 if (!pmd_present(*pmd)) {
3232 new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3235 pmd_populate_kernel(&init_mm, pmd, new);
3241 panic("%s: Failed to allocate memory\n", __func__);
3245 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3246 * @reserved_size: the size of reserved percpu area in bytes
3247 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3249 * This is a helper to ease setting up page-remapped first percpu
3250 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3252 * This is the basic allocator. Static percpu area is allocated
3253 * page-by-page into vmalloc area.
3256 * 0 on success, -errno on failure.
3258 int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3260 static struct vm_struct vm;
3261 struct pcpu_alloc_info *ai;
3265 struct page **pages;
3266 int unit, i, j, rc = 0;
3270 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3272 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3275 BUG_ON(ai->nr_groups != 1);
3276 upa = ai->alloc_size/ai->unit_size;
3277 nr_g0_units = roundup(num_possible_cpus(), upa);
3278 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3279 pcpu_free_alloc_info(ai);
3283 unit_pages = ai->unit_size >> PAGE_SHIFT;
3285 /* unaligned allocations can't be freed, round up to page size */
3286 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3288 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3290 panic("%s: Failed to allocate %zu bytes\n", __func__,
3293 /* allocate pages */
3295 for (unit = 0; unit < num_possible_cpus(); unit++) {
3296 unsigned int cpu = ai->groups[0].cpu_map[unit];
3297 for (i = 0; i < unit_pages; i++) {
3300 ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3302 pr_warn("failed to allocate %s page for cpu%u\n",
3306 /* kmemleak tracks the percpu allocations separately */
3308 pages[j++] = virt_to_page(ptr);
3312 /* allocate vm area, map the pages and copy static data */
3313 vm.flags = VM_ALLOC;
3314 vm.size = num_possible_cpus() * ai->unit_size;
3315 vm_area_register_early(&vm, PAGE_SIZE);
3317 for (unit = 0; unit < num_possible_cpus(); unit++) {
3318 unsigned long unit_addr =
3319 (unsigned long)vm.addr + unit * ai->unit_size;
3321 for (i = 0; i < unit_pages; i++)
3322 pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT));
3324 /* pte already populated, the following shouldn't fail */
3325 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3328 panic("failed to map percpu area, err=%d\n", rc);
3331 * FIXME: Archs with virtual cache should flush local
3332 * cache for the linear mapping here - something
3333 * equivalent to flush_cache_vmap() on the local cpu.
3334 * flush_cache_vmap() can't be used as most supporting
3335 * data structures are not set up yet.
3338 /* copy static data */
3339 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3342 /* we're ready, commit */
3343 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3344 unit_pages, psize_str, ai->static_size,
3345 ai->reserved_size, ai->dyn_size);
3347 pcpu_setup_first_chunk(ai, vm.addr);
3352 pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3355 memblock_free(pages, pages_size);
3356 pcpu_free_alloc_info(ai);
3359 #endif /* BUILD_PAGE_FIRST_CHUNK */
3361 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3363 * Generic SMP percpu area setup.
3365 * The embedding helper is used because its behavior closely resembles
3366 * the original non-dynamic generic percpu area setup. This is
3367 * important because many archs have addressing restrictions and might
3368 * fail if the percpu area is located far away from the previous
3369 * location. As an added bonus, in non-NUMA cases, embedding is
3370 * generally a good idea TLB-wise because percpu area can piggy back
3371 * on the physical linear memory mapping which uses large page
3372 * mappings on applicable archs.
3374 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3375 EXPORT_SYMBOL(__per_cpu_offset);
3377 void __init setup_per_cpu_areas(void)
3379 unsigned long delta;
3384 * Always reserve area for module percpu variables. That's
3385 * what the legacy allocator did.
3387 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3388 PAGE_SIZE, NULL, NULL);
3390 panic("Failed to initialize percpu areas.");
3392 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3393 for_each_possible_cpu(cpu)
3394 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3396 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3398 #else /* CONFIG_SMP */
3401 * UP percpu area setup.
3403 * UP always uses km-based percpu allocator with identity mapping.
3404 * Static percpu variables are indistinguishable from the usual static
3405 * variables and don't require any special preparation.
3407 void __init setup_per_cpu_areas(void)
3409 const size_t unit_size =
3410 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3411 PERCPU_DYNAMIC_RESERVE));
3412 struct pcpu_alloc_info *ai;
3415 ai = pcpu_alloc_alloc_info(1, 1);
3416 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3418 panic("Failed to allocate memory for percpu areas.");
3419 /* kmemleak tracks the percpu allocations separately */
3422 ai->dyn_size = unit_size;
3423 ai->unit_size = unit_size;
3424 ai->atom_size = unit_size;
3425 ai->alloc_size = unit_size;
3426 ai->groups[0].nr_units = 1;
3427 ai->groups[0].cpu_map[0] = 0;
3429 pcpu_setup_first_chunk(ai, fc);
3430 pcpu_free_alloc_info(ai);
3433 #endif /* CONFIG_SMP */
3436 * pcpu_nr_pages - calculate total number of populated backing pages
3438 * This reflects the number of pages populated to back chunks. Metadata is
3439 * excluded in the number exposed in meminfo as the number of backing pages
3440 * scales with the number of cpus and can quickly outweigh the memory used for
3441 * metadata. It also keeps this calculation nice and simple.
3444 * Total number of populated backing pages in use by the allocator.
3446 unsigned long pcpu_nr_pages(void)
3448 return pcpu_nr_populated * pcpu_nr_units;
3452 * Percpu allocator is initialized early during boot when neither slab or
3453 * workqueue is available. Plug async management until everything is up
3456 static int __init percpu_enable_async(void)
3458 pcpu_async_enabled = true;
3461 subsys_initcall(percpu_enable_async);