x86/speculation/mds: Fix documentation typo
[linux/fpc-iii.git] / mm / percpu.c
blob0c06e2f549a7bab4f2d4f4a9663d4e64d31a55b3
1 /*
2 * mm/percpu.c - percpu memory allocator
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 * Copyright (C) 2017 Facebook Inc.
8 * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com>
10 * This file is released under the GPLv2 license.
12 * The percpu allocator handles both static and dynamic areas. Percpu
13 * areas are allocated in chunks which are divided into units. There is
14 * a 1-to-1 mapping for units to possible cpus. These units are grouped
15 * based on NUMA properties of the machine.
17 * c0 c1 c2
18 * ------------------- ------------------- ------------
19 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
20 * ------------------- ...... ------------------- .... ------------
22 * Allocation is done by offsets into a unit's address space. Ie., an
23 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
24 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
25 * and even sparse. Access is handled by configuring percpu base
26 * registers according to the cpu to unit mappings and offsetting the
27 * base address using pcpu_unit_size.
29 * There is special consideration for the first chunk which must handle
30 * the static percpu variables in the kernel image as allocation services
31 * are not online yet. In short, the first chunk is structured like so:
33 * <Static | [Reserved] | Dynamic>
35 * The static data is copied from the original section managed by the
36 * linker. The reserved section, if non-zero, primarily manages static
37 * percpu variables from kernel modules. Finally, the dynamic section
38 * takes care of normal allocations.
40 * The allocator organizes chunks into lists according to free size and
41 * tries to allocate from the fullest chunk first. Each chunk is managed
42 * by a bitmap with metadata blocks. The allocation map is updated on
43 * every allocation and free to reflect the current state while the boundary
44 * map is only updated on allocation. Each metadata block contains
45 * information to help mitigate the need to iterate over large portions
46 * of the bitmap. The reverse mapping from page to chunk is stored in
47 * the page's index. Lastly, units are lazily backed and grow in unison.
49 * There is a unique conversion that goes on here between bytes and bits.
50 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
51 * tracks the number of pages it is responsible for in nr_pages. Helper
52 * functions are used to convert from between the bytes, bits, and blocks.
53 * All hints are managed in bits unless explicitly stated.
55 * To use this allocator, arch code should do the following:
57 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
58 * regular address to percpu pointer and back if they need to be
59 * different from the default
61 * - use pcpu_setup_first_chunk() during percpu area initialization to
62 * setup the first chunk containing the kernel static percpu area
65 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
67 #include <linux/bitmap.h>
68 #include <linux/bootmem.h>
69 #include <linux/err.h>
70 #include <linux/lcm.h>
71 #include <linux/list.h>
72 #include <linux/log2.h>
73 #include <linux/mm.h>
74 #include <linux/module.h>
75 #include <linux/mutex.h>
76 #include <linux/percpu.h>
77 #include <linux/pfn.h>
78 #include <linux/slab.h>
79 #include <linux/spinlock.h>
80 #include <linux/vmalloc.h>
81 #include <linux/workqueue.h>
82 #include <linux/kmemleak.h>
83 #include <linux/sched.h>
85 #include <asm/cacheflush.h>
86 #include <asm/sections.h>
87 #include <asm/tlbflush.h>
88 #include <asm/io.h>
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/percpu.h>
93 #include "percpu-internal.h"
95 /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
96 #define PCPU_SLOT_BASE_SHIFT 5
98 #define PCPU_EMPTY_POP_PAGES_LOW 2
99 #define PCPU_EMPTY_POP_PAGES_HIGH 4
101 #ifdef CONFIG_SMP
102 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
103 #ifndef __addr_to_pcpu_ptr
104 #define __addr_to_pcpu_ptr(addr) \
105 (void __percpu *)((unsigned long)(addr) - \
106 (unsigned long)pcpu_base_addr + \
107 (unsigned long)__per_cpu_start)
108 #endif
109 #ifndef __pcpu_ptr_to_addr
110 #define __pcpu_ptr_to_addr(ptr) \
111 (void __force *)((unsigned long)(ptr) + \
112 (unsigned long)pcpu_base_addr - \
113 (unsigned long)__per_cpu_start)
114 #endif
115 #else /* CONFIG_SMP */
116 /* on UP, it's always identity mapped */
117 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
118 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
119 #endif /* CONFIG_SMP */
121 static int pcpu_unit_pages __ro_after_init;
122 static int pcpu_unit_size __ro_after_init;
123 static int pcpu_nr_units __ro_after_init;
124 static int pcpu_atom_size __ro_after_init;
125 int pcpu_nr_slots __ro_after_init;
126 static size_t pcpu_chunk_struct_size __ro_after_init;
128 /* cpus with the lowest and highest unit addresses */
129 static unsigned int pcpu_low_unit_cpu __ro_after_init;
130 static unsigned int pcpu_high_unit_cpu __ro_after_init;
132 /* the address of the first chunk which starts with the kernel static area */
133 void *pcpu_base_addr __ro_after_init;
134 EXPORT_SYMBOL_GPL(pcpu_base_addr);
136 static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
137 const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
139 /* group information, used for vm allocation */
140 static int pcpu_nr_groups __ro_after_init;
141 static const unsigned long *pcpu_group_offsets __ro_after_init;
142 static const size_t *pcpu_group_sizes __ro_after_init;
145 * The first chunk which always exists. Note that unlike other
146 * chunks, this one can be allocated and mapped in several different
147 * ways and thus often doesn't live in the vmalloc area.
149 struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
152 * Optional reserved chunk. This chunk reserves part of the first
153 * chunk and serves it for reserved allocations. When the reserved
154 * region doesn't exist, the following variable is NULL.
156 struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
158 DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
159 static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
161 struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
163 /* chunks which need their map areas extended, protected by pcpu_lock */
164 static LIST_HEAD(pcpu_map_extend_chunks);
167 * The number of empty populated pages, protected by pcpu_lock. The
168 * reserved chunk doesn't contribute to the count.
170 int pcpu_nr_empty_pop_pages;
173 * Balance work is used to populate or destroy chunks asynchronously. We
174 * try to keep the number of populated free pages between
175 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
176 * empty chunk.
178 static void pcpu_balance_workfn(struct work_struct *work);
179 static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
180 static bool pcpu_async_enabled __read_mostly;
181 static bool pcpu_atomic_alloc_failed;
183 static void pcpu_schedule_balance_work(void)
185 if (pcpu_async_enabled)
186 schedule_work(&pcpu_balance_work);
190 * pcpu_addr_in_chunk - check if the address is served from this chunk
191 * @chunk: chunk of interest
192 * @addr: percpu address
194 * RETURNS:
195 * True if the address is served from this chunk.
197 static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
199 void *start_addr, *end_addr;
201 if (!chunk)
202 return false;
204 start_addr = chunk->base_addr + chunk->start_offset;
205 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
206 chunk->end_offset;
208 return addr >= start_addr && addr < end_addr;
211 static int __pcpu_size_to_slot(int size)
213 int highbit = fls(size); /* size is in bytes */
214 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
217 static int pcpu_size_to_slot(int size)
219 if (size == pcpu_unit_size)
220 return pcpu_nr_slots - 1;
221 return __pcpu_size_to_slot(size);
224 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
226 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0)
227 return 0;
229 return pcpu_size_to_slot(chunk->free_bytes);
232 /* set the pointer to a chunk in a page struct */
233 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
235 page->index = (unsigned long)pcpu;
238 /* obtain pointer to a chunk from a page struct */
239 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
241 return (struct pcpu_chunk *)page->index;
244 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
246 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
249 static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
251 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
254 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
255 unsigned int cpu, int page_idx)
257 return (unsigned long)chunk->base_addr +
258 pcpu_unit_page_offset(cpu, page_idx);
261 static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
263 *rs = find_next_zero_bit(bitmap, end, *rs);
264 *re = find_next_bit(bitmap, end, *rs + 1);
267 static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
269 *rs = find_next_bit(bitmap, end, *rs);
270 *re = find_next_zero_bit(bitmap, end, *rs + 1);
274 * Bitmap region iterators. Iterates over the bitmap between
275 * [@start, @end) in @chunk. @rs and @re should be integer variables
276 * and will be set to start and end index of the current free region.
278 #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
279 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
280 (rs) < (re); \
281 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
283 #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
284 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
285 (rs) < (re); \
286 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
289 * The following are helper functions to help access bitmaps and convert
290 * between bitmap offsets to address offsets.
292 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
294 return chunk->alloc_map +
295 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
298 static unsigned long pcpu_off_to_block_index(int off)
300 return off / PCPU_BITMAP_BLOCK_BITS;
303 static unsigned long pcpu_off_to_block_off(int off)
305 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
308 static unsigned long pcpu_block_off_to_off(int index, int off)
310 return index * PCPU_BITMAP_BLOCK_BITS + off;
314 * pcpu_next_md_free_region - finds the next hint free area
315 * @chunk: chunk of interest
316 * @bit_off: chunk offset
317 * @bits: size of free area
319 * Helper function for pcpu_for_each_md_free_region. It checks
320 * block->contig_hint and performs aggregation across blocks to find the
321 * next hint. It modifies bit_off and bits in-place to be consumed in the
322 * loop.
324 static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
325 int *bits)
327 int i = pcpu_off_to_block_index(*bit_off);
328 int block_off = pcpu_off_to_block_off(*bit_off);
329 struct pcpu_block_md *block;
331 *bits = 0;
332 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
333 block++, i++) {
334 /* handles contig area across blocks */
335 if (*bits) {
336 *bits += block->left_free;
337 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
338 continue;
339 return;
343 * This checks three things. First is there a contig_hint to
344 * check. Second, have we checked this hint before by
345 * comparing the block_off. Third, is this the same as the
346 * right contig hint. In the last case, it spills over into
347 * the next block and should be handled by the contig area
348 * across blocks code.
350 *bits = block->contig_hint;
351 if (*bits && block->contig_hint_start >= block_off &&
352 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
353 *bit_off = pcpu_block_off_to_off(i,
354 block->contig_hint_start);
355 return;
357 /* reset to satisfy the second predicate above */
358 block_off = 0;
360 *bits = block->right_free;
361 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
366 * pcpu_next_fit_region - finds fit areas for a given allocation request
367 * @chunk: chunk of interest
368 * @alloc_bits: size of allocation
369 * @align: alignment of area (max PAGE_SIZE)
370 * @bit_off: chunk offset
371 * @bits: size of free area
373 * Finds the next free region that is viable for use with a given size and
374 * alignment. This only returns if there is a valid area to be used for this
375 * allocation. block->first_free is returned if the allocation request fits
376 * within the block to see if the request can be fulfilled prior to the contig
377 * hint.
379 static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
380 int align, int *bit_off, int *bits)
382 int i = pcpu_off_to_block_index(*bit_off);
383 int block_off = pcpu_off_to_block_off(*bit_off);
384 struct pcpu_block_md *block;
386 *bits = 0;
387 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
388 block++, i++) {
389 /* handles contig area across blocks */
390 if (*bits) {
391 *bits += block->left_free;
392 if (*bits >= alloc_bits)
393 return;
394 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
395 continue;
398 /* check block->contig_hint */
399 *bits = ALIGN(block->contig_hint_start, align) -
400 block->contig_hint_start;
402 * This uses the block offset to determine if this has been
403 * checked in the prior iteration.
405 if (block->contig_hint &&
406 block->contig_hint_start >= block_off &&
407 block->contig_hint >= *bits + alloc_bits) {
408 *bits += alloc_bits + block->contig_hint_start -
409 block->first_free;
410 *bit_off = pcpu_block_off_to_off(i, block->first_free);
411 return;
413 /* reset to satisfy the second predicate above */
414 block_off = 0;
416 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
417 align);
418 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
419 *bit_off = pcpu_block_off_to_off(i, *bit_off);
420 if (*bits >= alloc_bits)
421 return;
424 /* no valid offsets were found - fail condition */
425 *bit_off = pcpu_chunk_map_bits(chunk);
429 * Metadata free area iterators. These perform aggregation of free areas
430 * based on the metadata blocks and return the offset @bit_off and size in
431 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
432 * a fit is found for the allocation request.
434 #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
435 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
436 (bit_off) < pcpu_chunk_map_bits((chunk)); \
437 (bit_off) += (bits) + 1, \
438 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
440 #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
441 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
442 &(bits)); \
443 (bit_off) < pcpu_chunk_map_bits((chunk)); \
444 (bit_off) += (bits), \
445 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
446 &(bits)))
449 * pcpu_mem_zalloc - allocate memory
450 * @size: bytes to allocate
451 * @gfp: allocation flags
453 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
454 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
455 * This is to facilitate passing through whitelisted flags. The
456 * returned memory is always zeroed.
458 * CONTEXT:
459 * Does GFP_KERNEL allocation.
461 * RETURNS:
462 * Pointer to the allocated area on success, NULL on failure.
464 static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
466 if (WARN_ON_ONCE(!slab_is_available()))
467 return NULL;
469 if (size <= PAGE_SIZE)
470 return kzalloc(size, gfp | GFP_KERNEL);
471 else
472 return __vmalloc(size, gfp | GFP_KERNEL | __GFP_ZERO,
473 PAGE_KERNEL);
477 * pcpu_mem_free - free memory
478 * @ptr: memory to free
480 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
482 static void pcpu_mem_free(void *ptr)
484 kvfree(ptr);
488 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
489 * @chunk: chunk of interest
490 * @oslot: the previous slot it was on
492 * This function is called after an allocation or free changed @chunk.
493 * New slot according to the changed state is determined and @chunk is
494 * moved to the slot. Note that the reserved chunk is never put on
495 * chunk slots.
497 * CONTEXT:
498 * pcpu_lock.
500 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
502 int nslot = pcpu_chunk_slot(chunk);
504 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
505 if (oslot < nslot)
506 list_move(&chunk->list, &pcpu_slot[nslot]);
507 else
508 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
513 * pcpu_cnt_pop_pages- counts populated backing pages in range
514 * @chunk: chunk of interest
515 * @bit_off: start offset
516 * @bits: size of area to check
518 * Calculates the number of populated pages in the region
519 * [page_start, page_end). This keeps track of how many empty populated
520 * pages are available and decide if async work should be scheduled.
522 * RETURNS:
523 * The nr of populated pages.
525 static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off,
526 int bits)
528 int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE);
529 int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
531 if (page_start >= page_end)
532 return 0;
535 * bitmap_weight counts the number of bits set in a bitmap up to
536 * the specified number of bits. This is counting the populated
537 * pages up to page_end and then subtracting the populated pages
538 * up to page_start to count the populated pages in
539 * [page_start, page_end).
541 return bitmap_weight(chunk->populated, page_end) -
542 bitmap_weight(chunk->populated, page_start);
546 * pcpu_chunk_update - updates the chunk metadata given a free area
547 * @chunk: chunk of interest
548 * @bit_off: chunk offset
549 * @bits: size of free area
551 * This updates the chunk's contig hint and starting offset given a free area.
552 * Choose the best starting offset if the contig hint is equal.
554 static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits)
556 if (bits > chunk->contig_bits) {
557 chunk->contig_bits_start = bit_off;
558 chunk->contig_bits = bits;
559 } else if (bits == chunk->contig_bits && chunk->contig_bits_start &&
560 (!bit_off ||
561 __ffs(bit_off) > __ffs(chunk->contig_bits_start))) {
562 /* use the start with the best alignment */
563 chunk->contig_bits_start = bit_off;
568 * pcpu_chunk_refresh_hint - updates metadata about a chunk
569 * @chunk: chunk of interest
571 * Iterates over the metadata blocks to find the largest contig area.
572 * It also counts the populated pages and uses the delta to update the
573 * global count.
575 * Updates:
576 * chunk->contig_bits
577 * chunk->contig_bits_start
578 * nr_empty_pop_pages (chunk and global)
580 static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk)
582 int bit_off, bits, nr_empty_pop_pages;
584 /* clear metadata */
585 chunk->contig_bits = 0;
587 bit_off = chunk->first_bit;
588 bits = nr_empty_pop_pages = 0;
589 pcpu_for_each_md_free_region(chunk, bit_off, bits) {
590 pcpu_chunk_update(chunk, bit_off, bits);
592 nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits);
596 * Keep track of nr_empty_pop_pages.
598 * The chunk maintains the previous number of free pages it held,
599 * so the delta is used to update the global counter. The reserved
600 * chunk is not part of the free page count as they are populated
601 * at init and are special to serving reserved allocations.
603 if (chunk != pcpu_reserved_chunk)
604 pcpu_nr_empty_pop_pages +=
605 (nr_empty_pop_pages - chunk->nr_empty_pop_pages);
607 chunk->nr_empty_pop_pages = nr_empty_pop_pages;
611 * pcpu_block_update - updates a block given a free area
612 * @block: block of interest
613 * @start: start offset in block
614 * @end: end offset in block
616 * Updates a block given a known free area. The region [start, end) is
617 * expected to be the entirety of the free area within a block. Chooses
618 * the best starting offset if the contig hints are equal.
620 static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
622 int contig = end - start;
624 block->first_free = min(block->first_free, start);
625 if (start == 0)
626 block->left_free = contig;
628 if (end == PCPU_BITMAP_BLOCK_BITS)
629 block->right_free = contig;
631 if (contig > block->contig_hint) {
632 block->contig_hint_start = start;
633 block->contig_hint = contig;
634 } else if (block->contig_hint_start && contig == block->contig_hint &&
635 (!start || __ffs(start) > __ffs(block->contig_hint_start))) {
636 /* use the start with the best alignment */
637 block->contig_hint_start = start;
642 * pcpu_block_refresh_hint
643 * @chunk: chunk of interest
644 * @index: index of the metadata block
646 * Scans over the block beginning at first_free and updates the block
647 * metadata accordingly.
649 static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
651 struct pcpu_block_md *block = chunk->md_blocks + index;
652 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
653 int rs, re; /* region start, region end */
655 /* clear hints */
656 block->contig_hint = 0;
657 block->left_free = block->right_free = 0;
659 /* iterate over free areas and update the contig hints */
660 pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free,
661 PCPU_BITMAP_BLOCK_BITS) {
662 pcpu_block_update(block, rs, re);
667 * pcpu_block_update_hint_alloc - update hint on allocation path
668 * @chunk: chunk of interest
669 * @bit_off: chunk offset
670 * @bits: size of request
672 * Updates metadata for the allocation path. The metadata only has to be
673 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
674 * scans are required if the block's contig hint is broken.
676 static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
677 int bits)
679 struct pcpu_block_md *s_block, *e_block, *block;
680 int s_index, e_index; /* block indexes of the freed allocation */
681 int s_off, e_off; /* block offsets of the freed allocation */
684 * Calculate per block offsets.
685 * The calculation uses an inclusive range, but the resulting offsets
686 * are [start, end). e_index always points to the last block in the
687 * range.
689 s_index = pcpu_off_to_block_index(bit_off);
690 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
691 s_off = pcpu_off_to_block_off(bit_off);
692 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
694 s_block = chunk->md_blocks + s_index;
695 e_block = chunk->md_blocks + e_index;
698 * Update s_block.
699 * block->first_free must be updated if the allocation takes its place.
700 * If the allocation breaks the contig_hint, a scan is required to
701 * restore this hint.
703 if (s_off == s_block->first_free)
704 s_block->first_free = find_next_zero_bit(
705 pcpu_index_alloc_map(chunk, s_index),
706 PCPU_BITMAP_BLOCK_BITS,
707 s_off + bits);
709 if (s_off >= s_block->contig_hint_start &&
710 s_off < s_block->contig_hint_start + s_block->contig_hint) {
711 /* block contig hint is broken - scan to fix it */
712 pcpu_block_refresh_hint(chunk, s_index);
713 } else {
714 /* update left and right contig manually */
715 s_block->left_free = min(s_block->left_free, s_off);
716 if (s_index == e_index)
717 s_block->right_free = min_t(int, s_block->right_free,
718 PCPU_BITMAP_BLOCK_BITS - e_off);
719 else
720 s_block->right_free = 0;
724 * Update e_block.
726 if (s_index != e_index) {
728 * When the allocation is across blocks, the end is along
729 * the left part of the e_block.
731 e_block->first_free = find_next_zero_bit(
732 pcpu_index_alloc_map(chunk, e_index),
733 PCPU_BITMAP_BLOCK_BITS, e_off);
735 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
736 /* reset the block */
737 e_block++;
738 } else {
739 if (e_off > e_block->contig_hint_start) {
740 /* contig hint is broken - scan to fix it */
741 pcpu_block_refresh_hint(chunk, e_index);
742 } else {
743 e_block->left_free = 0;
744 e_block->right_free =
745 min_t(int, e_block->right_free,
746 PCPU_BITMAP_BLOCK_BITS - e_off);
750 /* update in-between md_blocks */
751 for (block = s_block + 1; block < e_block; block++) {
752 block->contig_hint = 0;
753 block->left_free = 0;
754 block->right_free = 0;
759 * The only time a full chunk scan is required is if the chunk
760 * contig hint is broken. Otherwise, it means a smaller space
761 * was used and therefore the chunk contig hint is still correct.
763 if (bit_off >= chunk->contig_bits_start &&
764 bit_off < chunk->contig_bits_start + chunk->contig_bits)
765 pcpu_chunk_refresh_hint(chunk);
769 * pcpu_block_update_hint_free - updates the block hints on the free path
770 * @chunk: chunk of interest
771 * @bit_off: chunk offset
772 * @bits: size of request
774 * Updates metadata for the allocation path. This avoids a blind block
775 * refresh by making use of the block contig hints. If this fails, it scans
776 * forward and backward to determine the extent of the free area. This is
777 * capped at the boundary of blocks.
779 * A chunk update is triggered if a page becomes free, a block becomes free,
780 * or the free spans across blocks. This tradeoff is to minimize iterating
781 * over the block metadata to update chunk->contig_bits. chunk->contig_bits
782 * may be off by up to a page, but it will never be more than the available
783 * space. If the contig hint is contained in one block, it will be accurate.
785 static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
786 int bits)
788 struct pcpu_block_md *s_block, *e_block, *block;
789 int s_index, e_index; /* block indexes of the freed allocation */
790 int s_off, e_off; /* block offsets of the freed allocation */
791 int start, end; /* start and end of the whole free area */
794 * Calculate per block offsets.
795 * The calculation uses an inclusive range, but the resulting offsets
796 * are [start, end). e_index always points to the last block in the
797 * range.
799 s_index = pcpu_off_to_block_index(bit_off);
800 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
801 s_off = pcpu_off_to_block_off(bit_off);
802 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
804 s_block = chunk->md_blocks + s_index;
805 e_block = chunk->md_blocks + e_index;
808 * Check if the freed area aligns with the block->contig_hint.
809 * If it does, then the scan to find the beginning/end of the
810 * larger free area can be avoided.
812 * start and end refer to beginning and end of the free area
813 * within each their respective blocks. This is not necessarily
814 * the entire free area as it may span blocks past the beginning
815 * or end of the block.
817 start = s_off;
818 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
819 start = s_block->contig_hint_start;
820 } else {
822 * Scan backwards to find the extent of the free area.
823 * find_last_bit returns the starting bit, so if the start bit
824 * is returned, that means there was no last bit and the
825 * remainder of the chunk is free.
827 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
828 start);
829 start = (start == l_bit) ? 0 : l_bit + 1;
832 end = e_off;
833 if (e_off == e_block->contig_hint_start)
834 end = e_block->contig_hint_start + e_block->contig_hint;
835 else
836 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
837 PCPU_BITMAP_BLOCK_BITS, end);
839 /* update s_block */
840 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
841 pcpu_block_update(s_block, start, e_off);
843 /* freeing in the same block */
844 if (s_index != e_index) {
845 /* update e_block */
846 pcpu_block_update(e_block, 0, end);
848 /* reset md_blocks in the middle */
849 for (block = s_block + 1; block < e_block; block++) {
850 block->first_free = 0;
851 block->contig_hint_start = 0;
852 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
853 block->left_free = PCPU_BITMAP_BLOCK_BITS;
854 block->right_free = PCPU_BITMAP_BLOCK_BITS;
859 * Refresh chunk metadata when the free makes a page free, a block
860 * free, or spans across blocks. The contig hint may be off by up to
861 * a page, but if the hint is contained in a block, it will be accurate
862 * with the else condition below.
864 if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) >
865 ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) ||
866 s_index != e_index)
867 pcpu_chunk_refresh_hint(chunk);
868 else
869 pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start),
870 s_block->contig_hint);
874 * pcpu_is_populated - determines if the region is populated
875 * @chunk: chunk of interest
876 * @bit_off: chunk offset
877 * @bits: size of area
878 * @next_off: return value for the next offset to start searching
880 * For atomic allocations, check if the backing pages are populated.
882 * RETURNS:
883 * Bool if the backing pages are populated.
884 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
886 static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
887 int *next_off)
889 int page_start, page_end, rs, re;
891 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
892 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
894 rs = page_start;
895 pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
896 if (rs >= page_end)
897 return true;
899 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
900 return false;
904 * pcpu_find_block_fit - finds the block index to start searching
905 * @chunk: chunk of interest
906 * @alloc_bits: size of request in allocation units
907 * @align: alignment of area (max PAGE_SIZE bytes)
908 * @pop_only: use populated regions only
910 * Given a chunk and an allocation spec, find the offset to begin searching
911 * for a free region. This iterates over the bitmap metadata blocks to
912 * find an offset that will be guaranteed to fit the requirements. It is
913 * not quite first fit as if the allocation does not fit in the contig hint
914 * of a block or chunk, it is skipped. This errs on the side of caution
915 * to prevent excess iteration. Poor alignment can cause the allocator to
916 * skip over blocks and chunks that have valid free areas.
918 * RETURNS:
919 * The offset in the bitmap to begin searching.
920 * -1 if no offset is found.
922 static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
923 size_t align, bool pop_only)
925 int bit_off, bits, next_off;
928 * Check to see if the allocation can fit in the chunk's contig hint.
929 * This is an optimization to prevent scanning by assuming if it
930 * cannot fit in the global hint, there is memory pressure and creating
931 * a new chunk would happen soon.
933 bit_off = ALIGN(chunk->contig_bits_start, align) -
934 chunk->contig_bits_start;
935 if (bit_off + alloc_bits > chunk->contig_bits)
936 return -1;
938 bit_off = chunk->first_bit;
939 bits = 0;
940 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
941 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
942 &next_off))
943 break;
945 bit_off = next_off;
946 bits = 0;
949 if (bit_off == pcpu_chunk_map_bits(chunk))
950 return -1;
952 return bit_off;
956 * pcpu_alloc_area - allocates an area from a pcpu_chunk
957 * @chunk: chunk of interest
958 * @alloc_bits: size of request in allocation units
959 * @align: alignment of area (max PAGE_SIZE)
960 * @start: bit_off to start searching
962 * This function takes in a @start offset to begin searching to fit an
963 * allocation of @alloc_bits with alignment @align. It needs to scan
964 * the allocation map because if it fits within the block's contig hint,
965 * @start will be block->first_free. This is an attempt to fill the
966 * allocation prior to breaking the contig hint. The allocation and
967 * boundary maps are updated accordingly if it confirms a valid
968 * free area.
970 * RETURNS:
971 * Allocated addr offset in @chunk on success.
972 * -1 if no matching area is found.
974 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
975 size_t align, int start)
977 size_t align_mask = (align) ? (align - 1) : 0;
978 int bit_off, end, oslot;
980 lockdep_assert_held(&pcpu_lock);
982 oslot = pcpu_chunk_slot(chunk);
985 * Search to find a fit.
987 end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS;
988 bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start,
989 alloc_bits, align_mask);
990 if (bit_off >= end)
991 return -1;
993 /* update alloc map */
994 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
996 /* update boundary map */
997 set_bit(bit_off, chunk->bound_map);
998 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
999 set_bit(bit_off + alloc_bits, chunk->bound_map);
1001 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1003 /* update first free bit */
1004 if (bit_off == chunk->first_bit)
1005 chunk->first_bit = find_next_zero_bit(
1006 chunk->alloc_map,
1007 pcpu_chunk_map_bits(chunk),
1008 bit_off + alloc_bits);
1010 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1012 pcpu_chunk_relocate(chunk, oslot);
1014 return bit_off * PCPU_MIN_ALLOC_SIZE;
1018 * pcpu_free_area - frees the corresponding offset
1019 * @chunk: chunk of interest
1020 * @off: addr offset into chunk
1022 * This function determines the size of an allocation to free using
1023 * the boundary bitmap and clears the allocation map.
1025 static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
1027 int bit_off, bits, end, oslot;
1029 lockdep_assert_held(&pcpu_lock);
1030 pcpu_stats_area_dealloc(chunk);
1032 oslot = pcpu_chunk_slot(chunk);
1034 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1036 /* find end index */
1037 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1038 bit_off + 1);
1039 bits = end - bit_off;
1040 bitmap_clear(chunk->alloc_map, bit_off, bits);
1042 /* update metadata */
1043 chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
1045 /* update first free bit */
1046 chunk->first_bit = min(chunk->first_bit, bit_off);
1048 pcpu_block_update_hint_free(chunk, bit_off, bits);
1050 pcpu_chunk_relocate(chunk, oslot);
1053 static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1055 struct pcpu_block_md *md_block;
1057 for (md_block = chunk->md_blocks;
1058 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1059 md_block++) {
1060 md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1061 md_block->left_free = PCPU_BITMAP_BLOCK_BITS;
1062 md_block->right_free = PCPU_BITMAP_BLOCK_BITS;
1067 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1068 * @tmp_addr: the start of the region served
1069 * @map_size: size of the region served
1071 * This is responsible for creating the chunks that serve the first chunk. The
1072 * base_addr is page aligned down of @tmp_addr while the region end is page
1073 * aligned up. Offsets are kept track of to determine the region served. All
1074 * this is done to appease the bitmap allocator in avoiding partial blocks.
1076 * RETURNS:
1077 * Chunk serving the region at @tmp_addr of @map_size.
1079 static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1080 int map_size)
1082 struct pcpu_chunk *chunk;
1083 unsigned long aligned_addr, lcm_align;
1084 int start_offset, offset_bits, region_size, region_bits;
1086 /* region calculations */
1087 aligned_addr = tmp_addr & PAGE_MASK;
1089 start_offset = tmp_addr - aligned_addr;
1092 * Align the end of the region with the LCM of PAGE_SIZE and
1093 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1094 * the other.
1096 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1097 region_size = ALIGN(start_offset + map_size, lcm_align);
1099 /* allocate chunk */
1100 chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) +
1101 BITS_TO_LONGS(region_size >> PAGE_SHIFT),
1104 INIT_LIST_HEAD(&chunk->list);
1106 chunk->base_addr = (void *)aligned_addr;
1107 chunk->start_offset = start_offset;
1108 chunk->end_offset = region_size - chunk->start_offset - map_size;
1110 chunk->nr_pages = region_size >> PAGE_SHIFT;
1111 region_bits = pcpu_chunk_map_bits(chunk);
1113 chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) *
1114 sizeof(chunk->alloc_map[0]), 0);
1115 chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) *
1116 sizeof(chunk->bound_map[0]), 0);
1117 chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) *
1118 sizeof(chunk->md_blocks[0]), 0);
1119 pcpu_init_md_blocks(chunk);
1121 /* manage populated page bitmap */
1122 chunk->immutable = true;
1123 bitmap_fill(chunk->populated, chunk->nr_pages);
1124 chunk->nr_populated = chunk->nr_pages;
1125 chunk->nr_empty_pop_pages =
1126 pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE,
1127 map_size / PCPU_MIN_ALLOC_SIZE);
1129 chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE;
1130 chunk->free_bytes = map_size;
1132 if (chunk->start_offset) {
1133 /* hide the beginning of the bitmap */
1134 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1135 bitmap_set(chunk->alloc_map, 0, offset_bits);
1136 set_bit(0, chunk->bound_map);
1137 set_bit(offset_bits, chunk->bound_map);
1139 chunk->first_bit = offset_bits;
1141 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1144 if (chunk->end_offset) {
1145 /* hide the end of the bitmap */
1146 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1147 bitmap_set(chunk->alloc_map,
1148 pcpu_chunk_map_bits(chunk) - offset_bits,
1149 offset_bits);
1150 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1151 chunk->bound_map);
1152 set_bit(region_bits, chunk->bound_map);
1154 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1155 - offset_bits, offset_bits);
1158 return chunk;
1161 static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1163 struct pcpu_chunk *chunk;
1164 int region_bits;
1166 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1167 if (!chunk)
1168 return NULL;
1170 INIT_LIST_HEAD(&chunk->list);
1171 chunk->nr_pages = pcpu_unit_pages;
1172 region_bits = pcpu_chunk_map_bits(chunk);
1174 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1175 sizeof(chunk->alloc_map[0]), gfp);
1176 if (!chunk->alloc_map)
1177 goto alloc_map_fail;
1179 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1180 sizeof(chunk->bound_map[0]), gfp);
1181 if (!chunk->bound_map)
1182 goto bound_map_fail;
1184 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1185 sizeof(chunk->md_blocks[0]), gfp);
1186 if (!chunk->md_blocks)
1187 goto md_blocks_fail;
1189 pcpu_init_md_blocks(chunk);
1191 /* init metadata */
1192 chunk->contig_bits = region_bits;
1193 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1195 return chunk;
1197 md_blocks_fail:
1198 pcpu_mem_free(chunk->bound_map);
1199 bound_map_fail:
1200 pcpu_mem_free(chunk->alloc_map);
1201 alloc_map_fail:
1202 pcpu_mem_free(chunk);
1204 return NULL;
1207 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1209 if (!chunk)
1210 return;
1211 pcpu_mem_free(chunk->md_blocks);
1212 pcpu_mem_free(chunk->bound_map);
1213 pcpu_mem_free(chunk->alloc_map);
1214 pcpu_mem_free(chunk);
1218 * pcpu_chunk_populated - post-population bookkeeping
1219 * @chunk: pcpu_chunk which got populated
1220 * @page_start: the start page
1221 * @page_end: the end page
1222 * @for_alloc: if this is to populate for allocation
1224 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1225 * the bookkeeping information accordingly. Must be called after each
1226 * successful population.
1228 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1229 * is to serve an allocation in that area.
1231 static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1232 int page_end, bool for_alloc)
1234 int nr = page_end - page_start;
1236 lockdep_assert_held(&pcpu_lock);
1238 bitmap_set(chunk->populated, page_start, nr);
1239 chunk->nr_populated += nr;
1241 if (!for_alloc) {
1242 chunk->nr_empty_pop_pages += nr;
1243 pcpu_nr_empty_pop_pages += nr;
1248 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1249 * @chunk: pcpu_chunk which got depopulated
1250 * @page_start: the start page
1251 * @page_end: the end page
1253 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1254 * Update the bookkeeping information accordingly. Must be called after
1255 * each successful depopulation.
1257 static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1258 int page_start, int page_end)
1260 int nr = page_end - page_start;
1262 lockdep_assert_held(&pcpu_lock);
1264 bitmap_clear(chunk->populated, page_start, nr);
1265 chunk->nr_populated -= nr;
1266 chunk->nr_empty_pop_pages -= nr;
1267 pcpu_nr_empty_pop_pages -= nr;
1271 * Chunk management implementation.
1273 * To allow different implementations, chunk alloc/free and
1274 * [de]population are implemented in a separate file which is pulled
1275 * into this file and compiled together. The following functions
1276 * should be implemented.
1278 * pcpu_populate_chunk - populate the specified range of a chunk
1279 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1280 * pcpu_create_chunk - create a new chunk
1281 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1282 * pcpu_addr_to_page - translate address to physical address
1283 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1285 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size,
1286 gfp_t gfp);
1287 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
1288 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1289 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1290 static struct page *pcpu_addr_to_page(void *addr);
1291 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1293 #ifdef CONFIG_NEED_PER_CPU_KM
1294 #include "percpu-km.c"
1295 #else
1296 #include "percpu-vm.c"
1297 #endif
1300 * pcpu_chunk_addr_search - determine chunk containing specified address
1301 * @addr: address for which the chunk needs to be determined.
1303 * This is an internal function that handles all but static allocations.
1304 * Static percpu address values should never be passed into the allocator.
1306 * RETURNS:
1307 * The address of the found chunk.
1309 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1311 /* is it in the dynamic region (first chunk)? */
1312 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1313 return pcpu_first_chunk;
1315 /* is it in the reserved region? */
1316 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1317 return pcpu_reserved_chunk;
1320 * The address is relative to unit0 which might be unused and
1321 * thus unmapped. Offset the address to the unit space of the
1322 * current processor before looking it up in the vmalloc
1323 * space. Note that any possible cpu id can be used here, so
1324 * there's no need to worry about preemption or cpu hotplug.
1326 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1327 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1331 * pcpu_alloc - the percpu allocator
1332 * @size: size of area to allocate in bytes
1333 * @align: alignment of area (max PAGE_SIZE)
1334 * @reserved: allocate from the reserved chunk if available
1335 * @gfp: allocation flags
1337 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1338 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1339 * then no warning will be triggered on invalid or failed allocation
1340 * requests.
1342 * RETURNS:
1343 * Percpu pointer to the allocated area on success, NULL on failure.
1345 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1346 gfp_t gfp)
1348 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1349 bool do_warn = !(gfp & __GFP_NOWARN);
1350 static int warn_limit = 10;
1351 struct pcpu_chunk *chunk;
1352 const char *err;
1353 int slot, off, cpu, ret;
1354 unsigned long flags;
1355 void __percpu *ptr;
1356 size_t bits, bit_align;
1359 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1360 * therefore alignment must be a minimum of that many bytes.
1361 * An allocation may have internal fragmentation from rounding up
1362 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1364 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1365 align = PCPU_MIN_ALLOC_SIZE;
1367 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1368 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1369 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1371 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1372 !is_power_of_2(align))) {
1373 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1374 size, align);
1375 return NULL;
1378 if (!is_atomic)
1379 mutex_lock(&pcpu_alloc_mutex);
1381 spin_lock_irqsave(&pcpu_lock, flags);
1383 /* serve reserved allocations from the reserved chunk if available */
1384 if (reserved && pcpu_reserved_chunk) {
1385 chunk = pcpu_reserved_chunk;
1387 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1388 if (off < 0) {
1389 err = "alloc from reserved chunk failed";
1390 goto fail_unlock;
1393 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1394 if (off >= 0)
1395 goto area_found;
1397 err = "alloc from reserved chunk failed";
1398 goto fail_unlock;
1401 restart:
1402 /* search through normal chunks */
1403 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1404 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1405 off = pcpu_find_block_fit(chunk, bits, bit_align,
1406 is_atomic);
1407 if (off < 0)
1408 continue;
1410 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1411 if (off >= 0)
1412 goto area_found;
1417 spin_unlock_irqrestore(&pcpu_lock, flags);
1420 * No space left. Create a new chunk. We don't want multiple
1421 * tasks to create chunks simultaneously. Serialize and create iff
1422 * there's still no empty chunk after grabbing the mutex.
1424 if (is_atomic) {
1425 err = "atomic alloc failed, no space left";
1426 goto fail;
1429 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1430 chunk = pcpu_create_chunk(0);
1431 if (!chunk) {
1432 err = "failed to allocate new chunk";
1433 goto fail;
1436 spin_lock_irqsave(&pcpu_lock, flags);
1437 pcpu_chunk_relocate(chunk, -1);
1438 } else {
1439 spin_lock_irqsave(&pcpu_lock, flags);
1442 goto restart;
1444 area_found:
1445 pcpu_stats_area_alloc(chunk, size);
1446 spin_unlock_irqrestore(&pcpu_lock, flags);
1448 /* populate if not all pages are already there */
1449 if (!is_atomic) {
1450 int page_start, page_end, rs, re;
1452 page_start = PFN_DOWN(off);
1453 page_end = PFN_UP(off + size);
1455 pcpu_for_each_unpop_region(chunk->populated, rs, re,
1456 page_start, page_end) {
1457 WARN_ON(chunk->immutable);
1459 ret = pcpu_populate_chunk(chunk, rs, re, 0);
1461 spin_lock_irqsave(&pcpu_lock, flags);
1462 if (ret) {
1463 pcpu_free_area(chunk, off);
1464 err = "failed to populate";
1465 goto fail_unlock;
1467 pcpu_chunk_populated(chunk, rs, re, true);
1468 spin_unlock_irqrestore(&pcpu_lock, flags);
1471 mutex_unlock(&pcpu_alloc_mutex);
1474 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1475 pcpu_schedule_balance_work();
1477 /* clear the areas and return address relative to base address */
1478 for_each_possible_cpu(cpu)
1479 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1481 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1482 kmemleak_alloc_percpu(ptr, size, gfp);
1484 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1485 chunk->base_addr, off, ptr);
1487 return ptr;
1489 fail_unlock:
1490 spin_unlock_irqrestore(&pcpu_lock, flags);
1491 fail:
1492 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1494 if (!is_atomic && do_warn && warn_limit) {
1495 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1496 size, align, is_atomic, err);
1497 dump_stack();
1498 if (!--warn_limit)
1499 pr_info("limit reached, disable warning\n");
1501 if (is_atomic) {
1502 /* see the flag handling in pcpu_blance_workfn() */
1503 pcpu_atomic_alloc_failed = true;
1504 pcpu_schedule_balance_work();
1505 } else {
1506 mutex_unlock(&pcpu_alloc_mutex);
1508 return NULL;
1512 * __alloc_percpu_gfp - allocate dynamic percpu area
1513 * @size: size of area to allocate in bytes
1514 * @align: alignment of area (max PAGE_SIZE)
1515 * @gfp: allocation flags
1517 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1518 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1519 * be called from any context but is a lot more likely to fail. If @gfp
1520 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1521 * allocation requests.
1523 * RETURNS:
1524 * Percpu pointer to the allocated area on success, NULL on failure.
1526 void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1528 return pcpu_alloc(size, align, false, gfp);
1530 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1533 * __alloc_percpu - allocate dynamic percpu area
1534 * @size: size of area to allocate in bytes
1535 * @align: alignment of area (max PAGE_SIZE)
1537 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1539 void __percpu *__alloc_percpu(size_t size, size_t align)
1541 return pcpu_alloc(size, align, false, GFP_KERNEL);
1543 EXPORT_SYMBOL_GPL(__alloc_percpu);
1546 * __alloc_reserved_percpu - allocate reserved percpu area
1547 * @size: size of area to allocate in bytes
1548 * @align: alignment of area (max PAGE_SIZE)
1550 * Allocate zero-filled percpu area of @size bytes aligned at @align
1551 * from reserved percpu area if arch has set it up; otherwise,
1552 * allocation is served from the same dynamic area. Might sleep.
1553 * Might trigger writeouts.
1555 * CONTEXT:
1556 * Does GFP_KERNEL allocation.
1558 * RETURNS:
1559 * Percpu pointer to the allocated area on success, NULL on failure.
1561 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1563 return pcpu_alloc(size, align, true, GFP_KERNEL);
1567 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1568 * @work: unused
1570 * Reclaim all fully free chunks except for the first one. This is also
1571 * responsible for maintaining the pool of empty populated pages. However,
1572 * it is possible that this is called when physical memory is scarce causing
1573 * OOM killer to be triggered. We should avoid doing so until an actual
1574 * allocation causes the failure as it is possible that requests can be
1575 * serviced from already backed regions.
1577 static void pcpu_balance_workfn(struct work_struct *work)
1579 /* gfp flags passed to underlying allocators */
1580 const gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN;
1581 LIST_HEAD(to_free);
1582 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1583 struct pcpu_chunk *chunk, *next;
1584 int slot, nr_to_pop, ret;
1587 * There's no reason to keep around multiple unused chunks and VM
1588 * areas can be scarce. Destroy all free chunks except for one.
1590 mutex_lock(&pcpu_alloc_mutex);
1591 spin_lock_irq(&pcpu_lock);
1593 list_for_each_entry_safe(chunk, next, free_head, list) {
1594 WARN_ON(chunk->immutable);
1596 /* spare the first one */
1597 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1598 continue;
1600 list_move(&chunk->list, &to_free);
1603 spin_unlock_irq(&pcpu_lock);
1605 list_for_each_entry_safe(chunk, next, &to_free, list) {
1606 int rs, re;
1608 pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1609 chunk->nr_pages) {
1610 pcpu_depopulate_chunk(chunk, rs, re);
1611 spin_lock_irq(&pcpu_lock);
1612 pcpu_chunk_depopulated(chunk, rs, re);
1613 spin_unlock_irq(&pcpu_lock);
1615 pcpu_destroy_chunk(chunk);
1619 * Ensure there are certain number of free populated pages for
1620 * atomic allocs. Fill up from the most packed so that atomic
1621 * allocs don't increase fragmentation. If atomic allocation
1622 * failed previously, always populate the maximum amount. This
1623 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1624 * failing indefinitely; however, large atomic allocs are not
1625 * something we support properly and can be highly unreliable and
1626 * inefficient.
1628 retry_pop:
1629 if (pcpu_atomic_alloc_failed) {
1630 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1631 /* best effort anyway, don't worry about synchronization */
1632 pcpu_atomic_alloc_failed = false;
1633 } else {
1634 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1635 pcpu_nr_empty_pop_pages,
1636 0, PCPU_EMPTY_POP_PAGES_HIGH);
1639 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1640 int nr_unpop = 0, rs, re;
1642 if (!nr_to_pop)
1643 break;
1645 spin_lock_irq(&pcpu_lock);
1646 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1647 nr_unpop = chunk->nr_pages - chunk->nr_populated;
1648 if (nr_unpop)
1649 break;
1651 spin_unlock_irq(&pcpu_lock);
1653 if (!nr_unpop)
1654 continue;
1656 /* @chunk can't go away while pcpu_alloc_mutex is held */
1657 pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1658 chunk->nr_pages) {
1659 int nr = min(re - rs, nr_to_pop);
1661 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
1662 if (!ret) {
1663 nr_to_pop -= nr;
1664 spin_lock_irq(&pcpu_lock);
1665 pcpu_chunk_populated(chunk, rs, rs + nr, false);
1666 spin_unlock_irq(&pcpu_lock);
1667 } else {
1668 nr_to_pop = 0;
1671 if (!nr_to_pop)
1672 break;
1676 if (nr_to_pop) {
1677 /* ran out of chunks to populate, create a new one and retry */
1678 chunk = pcpu_create_chunk(gfp);
1679 if (chunk) {
1680 spin_lock_irq(&pcpu_lock);
1681 pcpu_chunk_relocate(chunk, -1);
1682 spin_unlock_irq(&pcpu_lock);
1683 goto retry_pop;
1687 mutex_unlock(&pcpu_alloc_mutex);
1691 * free_percpu - free percpu area
1692 * @ptr: pointer to area to free
1694 * Free percpu area @ptr.
1696 * CONTEXT:
1697 * Can be called from atomic context.
1699 void free_percpu(void __percpu *ptr)
1701 void *addr;
1702 struct pcpu_chunk *chunk;
1703 unsigned long flags;
1704 int off;
1706 if (!ptr)
1707 return;
1709 kmemleak_free_percpu(ptr);
1711 addr = __pcpu_ptr_to_addr(ptr);
1713 spin_lock_irqsave(&pcpu_lock, flags);
1715 chunk = pcpu_chunk_addr_search(addr);
1716 off = addr - chunk->base_addr;
1718 pcpu_free_area(chunk, off);
1720 /* if there are more than one fully free chunks, wake up grim reaper */
1721 if (chunk->free_bytes == pcpu_unit_size) {
1722 struct pcpu_chunk *pos;
1724 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1725 if (pos != chunk) {
1726 pcpu_schedule_balance_work();
1727 break;
1731 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1733 spin_unlock_irqrestore(&pcpu_lock, flags);
1735 EXPORT_SYMBOL_GPL(free_percpu);
1737 bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1739 #ifdef CONFIG_SMP
1740 const size_t static_size = __per_cpu_end - __per_cpu_start;
1741 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1742 unsigned int cpu;
1744 for_each_possible_cpu(cpu) {
1745 void *start = per_cpu_ptr(base, cpu);
1746 void *va = (void *)addr;
1748 if (va >= start && va < start + static_size) {
1749 if (can_addr) {
1750 *can_addr = (unsigned long) (va - start);
1751 *can_addr += (unsigned long)
1752 per_cpu_ptr(base, get_boot_cpu_id());
1754 return true;
1757 #endif
1758 /* on UP, can't distinguish from other static vars, always false */
1759 return false;
1763 * is_kernel_percpu_address - test whether address is from static percpu area
1764 * @addr: address to test
1766 * Test whether @addr belongs to in-kernel static percpu area. Module
1767 * static percpu areas are not considered. For those, use
1768 * is_module_percpu_address().
1770 * RETURNS:
1771 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1773 bool is_kernel_percpu_address(unsigned long addr)
1775 return __is_kernel_percpu_address(addr, NULL);
1779 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1780 * @addr: the address to be converted to physical address
1782 * Given @addr which is dereferenceable address obtained via one of
1783 * percpu access macros, this function translates it into its physical
1784 * address. The caller is responsible for ensuring @addr stays valid
1785 * until this function finishes.
1787 * percpu allocator has special setup for the first chunk, which currently
1788 * supports either embedding in linear address space or vmalloc mapping,
1789 * and, from the second one, the backing allocator (currently either vm or
1790 * km) provides translation.
1792 * The addr can be translated simply without checking if it falls into the
1793 * first chunk. But the current code reflects better how percpu allocator
1794 * actually works, and the verification can discover both bugs in percpu
1795 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1796 * code.
1798 * RETURNS:
1799 * The physical address for @addr.
1801 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1803 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1804 bool in_first_chunk = false;
1805 unsigned long first_low, first_high;
1806 unsigned int cpu;
1809 * The following test on unit_low/high isn't strictly
1810 * necessary but will speed up lookups of addresses which
1811 * aren't in the first chunk.
1813 * The address check is against full chunk sizes. pcpu_base_addr
1814 * points to the beginning of the first chunk including the
1815 * static region. Assumes good intent as the first chunk may
1816 * not be full (ie. < pcpu_unit_pages in size).
1818 first_low = (unsigned long)pcpu_base_addr +
1819 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
1820 first_high = (unsigned long)pcpu_base_addr +
1821 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
1822 if ((unsigned long)addr >= first_low &&
1823 (unsigned long)addr < first_high) {
1824 for_each_possible_cpu(cpu) {
1825 void *start = per_cpu_ptr(base, cpu);
1827 if (addr >= start && addr < start + pcpu_unit_size) {
1828 in_first_chunk = true;
1829 break;
1834 if (in_first_chunk) {
1835 if (!is_vmalloc_addr(addr))
1836 return __pa(addr);
1837 else
1838 return page_to_phys(vmalloc_to_page(addr)) +
1839 offset_in_page(addr);
1840 } else
1841 return page_to_phys(pcpu_addr_to_page(addr)) +
1842 offset_in_page(addr);
1846 * pcpu_alloc_alloc_info - allocate percpu allocation info
1847 * @nr_groups: the number of groups
1848 * @nr_units: the number of units
1850 * Allocate ai which is large enough for @nr_groups groups containing
1851 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1852 * cpu_map array which is long enough for @nr_units and filled with
1853 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1854 * pointer of other groups.
1856 * RETURNS:
1857 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1858 * failure.
1860 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1861 int nr_units)
1863 struct pcpu_alloc_info *ai;
1864 size_t base_size, ai_size;
1865 void *ptr;
1866 int unit;
1868 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1869 __alignof__(ai->groups[0].cpu_map[0]));
1870 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1872 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1873 if (!ptr)
1874 return NULL;
1875 ai = ptr;
1876 ptr += base_size;
1878 ai->groups[0].cpu_map = ptr;
1880 for (unit = 0; unit < nr_units; unit++)
1881 ai->groups[0].cpu_map[unit] = NR_CPUS;
1883 ai->nr_groups = nr_groups;
1884 ai->__ai_size = PFN_ALIGN(ai_size);
1886 return ai;
1890 * pcpu_free_alloc_info - free percpu allocation info
1891 * @ai: pcpu_alloc_info to free
1893 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1895 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1897 memblock_free_early(__pa(ai), ai->__ai_size);
1901 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1902 * @lvl: loglevel
1903 * @ai: allocation info to dump
1905 * Print out information about @ai using loglevel @lvl.
1907 static void pcpu_dump_alloc_info(const char *lvl,
1908 const struct pcpu_alloc_info *ai)
1910 int group_width = 1, cpu_width = 1, width;
1911 char empty_str[] = "--------";
1912 int alloc = 0, alloc_end = 0;
1913 int group, v;
1914 int upa, apl; /* units per alloc, allocs per line */
1916 v = ai->nr_groups;
1917 while (v /= 10)
1918 group_width++;
1920 v = num_possible_cpus();
1921 while (v /= 10)
1922 cpu_width++;
1923 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1925 upa = ai->alloc_size / ai->unit_size;
1926 width = upa * (cpu_width + 1) + group_width + 3;
1927 apl = rounddown_pow_of_two(max(60 / width, 1));
1929 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1930 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1931 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1933 for (group = 0; group < ai->nr_groups; group++) {
1934 const struct pcpu_group_info *gi = &ai->groups[group];
1935 int unit = 0, unit_end = 0;
1937 BUG_ON(gi->nr_units % upa);
1938 for (alloc_end += gi->nr_units / upa;
1939 alloc < alloc_end; alloc++) {
1940 if (!(alloc % apl)) {
1941 pr_cont("\n");
1942 printk("%spcpu-alloc: ", lvl);
1944 pr_cont("[%0*d] ", group_width, group);
1946 for (unit_end += upa; unit < unit_end; unit++)
1947 if (gi->cpu_map[unit] != NR_CPUS)
1948 pr_cont("%0*d ",
1949 cpu_width, gi->cpu_map[unit]);
1950 else
1951 pr_cont("%s ", empty_str);
1954 pr_cont("\n");
1958 * pcpu_setup_first_chunk - initialize the first percpu chunk
1959 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1960 * @base_addr: mapped address
1962 * Initialize the first percpu chunk which contains the kernel static
1963 * perpcu area. This function is to be called from arch percpu area
1964 * setup path.
1966 * @ai contains all information necessary to initialize the first
1967 * chunk and prime the dynamic percpu allocator.
1969 * @ai->static_size is the size of static percpu area.
1971 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1972 * reserve after the static area in the first chunk. This reserves
1973 * the first chunk such that it's available only through reserved
1974 * percpu allocation. This is primarily used to serve module percpu
1975 * static areas on architectures where the addressing model has
1976 * limited offset range for symbol relocations to guarantee module
1977 * percpu symbols fall inside the relocatable range.
1979 * @ai->dyn_size determines the number of bytes available for dynamic
1980 * allocation in the first chunk. The area between @ai->static_size +
1981 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1983 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1984 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1985 * @ai->dyn_size.
1987 * @ai->atom_size is the allocation atom size and used as alignment
1988 * for vm areas.
1990 * @ai->alloc_size is the allocation size and always multiple of
1991 * @ai->atom_size. This is larger than @ai->atom_size if
1992 * @ai->unit_size is larger than @ai->atom_size.
1994 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1995 * percpu areas. Units which should be colocated are put into the
1996 * same group. Dynamic VM areas will be allocated according to these
1997 * groupings. If @ai->nr_groups is zero, a single group containing
1998 * all units is assumed.
2000 * The caller should have mapped the first chunk at @base_addr and
2001 * copied static data to each unit.
2003 * The first chunk will always contain a static and a dynamic region.
2004 * However, the static region is not managed by any chunk. If the first
2005 * chunk also contains a reserved region, it is served by two chunks -
2006 * one for the reserved region and one for the dynamic region. They
2007 * share the same vm, but use offset regions in the area allocation map.
2008 * The chunk serving the dynamic region is circulated in the chunk slots
2009 * and available for dynamic allocation like any other chunk.
2011 * RETURNS:
2012 * 0 on success, -errno on failure.
2014 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2015 void *base_addr)
2017 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2018 size_t static_size, dyn_size;
2019 struct pcpu_chunk *chunk;
2020 unsigned long *group_offsets;
2021 size_t *group_sizes;
2022 unsigned long *unit_off;
2023 unsigned int cpu;
2024 int *unit_map;
2025 int group, unit, i;
2026 int map_size;
2027 unsigned long tmp_addr;
2029 #define PCPU_SETUP_BUG_ON(cond) do { \
2030 if (unlikely(cond)) { \
2031 pr_emerg("failed to initialize, %s\n", #cond); \
2032 pr_emerg("cpu_possible_mask=%*pb\n", \
2033 cpumask_pr_args(cpu_possible_mask)); \
2034 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2035 BUG(); \
2037 } while (0)
2039 /* sanity checks */
2040 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2041 #ifdef CONFIG_SMP
2042 PCPU_SETUP_BUG_ON(!ai->static_size);
2043 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2044 #endif
2045 PCPU_SETUP_BUG_ON(!base_addr);
2046 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2047 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2048 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2049 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2050 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2051 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2052 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2053 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2054 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2055 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2056 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2058 /* process group information and build config tables accordingly */
2059 group_offsets = memblock_virt_alloc(ai->nr_groups *
2060 sizeof(group_offsets[0]), 0);
2061 group_sizes = memblock_virt_alloc(ai->nr_groups *
2062 sizeof(group_sizes[0]), 0);
2063 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
2064 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
2066 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2067 unit_map[cpu] = UINT_MAX;
2069 pcpu_low_unit_cpu = NR_CPUS;
2070 pcpu_high_unit_cpu = NR_CPUS;
2072 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2073 const struct pcpu_group_info *gi = &ai->groups[group];
2075 group_offsets[group] = gi->base_offset;
2076 group_sizes[group] = gi->nr_units * ai->unit_size;
2078 for (i = 0; i < gi->nr_units; i++) {
2079 cpu = gi->cpu_map[i];
2080 if (cpu == NR_CPUS)
2081 continue;
2083 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2084 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2085 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2087 unit_map[cpu] = unit + i;
2088 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2090 /* determine low/high unit_cpu */
2091 if (pcpu_low_unit_cpu == NR_CPUS ||
2092 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2093 pcpu_low_unit_cpu = cpu;
2094 if (pcpu_high_unit_cpu == NR_CPUS ||
2095 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2096 pcpu_high_unit_cpu = cpu;
2099 pcpu_nr_units = unit;
2101 for_each_possible_cpu(cpu)
2102 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2104 /* we're done parsing the input, undefine BUG macro and dump config */
2105 #undef PCPU_SETUP_BUG_ON
2106 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2108 pcpu_nr_groups = ai->nr_groups;
2109 pcpu_group_offsets = group_offsets;
2110 pcpu_group_sizes = group_sizes;
2111 pcpu_unit_map = unit_map;
2112 pcpu_unit_offsets = unit_off;
2114 /* determine basic parameters */
2115 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2116 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2117 pcpu_atom_size = ai->atom_size;
2118 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2119 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2121 pcpu_stats_save_ai(ai);
2124 * Allocate chunk slots. The additional last slot is for
2125 * empty chunks.
2127 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2128 pcpu_slot = memblock_virt_alloc(
2129 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
2130 for (i = 0; i < pcpu_nr_slots; i++)
2131 INIT_LIST_HEAD(&pcpu_slot[i]);
2134 * The end of the static region needs to be aligned with the
2135 * minimum allocation size as this offsets the reserved and
2136 * dynamic region. The first chunk ends page aligned by
2137 * expanding the dynamic region, therefore the dynamic region
2138 * can be shrunk to compensate while still staying above the
2139 * configured sizes.
2141 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2142 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2145 * Initialize first chunk.
2146 * If the reserved_size is non-zero, this initializes the reserved
2147 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2148 * and the dynamic region is initialized here. The first chunk,
2149 * pcpu_first_chunk, will always point to the chunk that serves
2150 * the dynamic region.
2152 tmp_addr = (unsigned long)base_addr + static_size;
2153 map_size = ai->reserved_size ?: dyn_size;
2154 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2156 /* init dynamic chunk if necessary */
2157 if (ai->reserved_size) {
2158 pcpu_reserved_chunk = chunk;
2160 tmp_addr = (unsigned long)base_addr + static_size +
2161 ai->reserved_size;
2162 map_size = dyn_size;
2163 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2166 /* link the first chunk in */
2167 pcpu_first_chunk = chunk;
2168 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2169 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2171 pcpu_stats_chunk_alloc();
2172 trace_percpu_create_chunk(base_addr);
2174 /* we're done */
2175 pcpu_base_addr = base_addr;
2176 return 0;
2179 #ifdef CONFIG_SMP
2181 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2182 [PCPU_FC_AUTO] = "auto",
2183 [PCPU_FC_EMBED] = "embed",
2184 [PCPU_FC_PAGE] = "page",
2187 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2189 static int __init percpu_alloc_setup(char *str)
2191 if (!str)
2192 return -EINVAL;
2194 if (0)
2195 /* nada */;
2196 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2197 else if (!strcmp(str, "embed"))
2198 pcpu_chosen_fc = PCPU_FC_EMBED;
2199 #endif
2200 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2201 else if (!strcmp(str, "page"))
2202 pcpu_chosen_fc = PCPU_FC_PAGE;
2203 #endif
2204 else
2205 pr_warn("unknown allocator %s specified\n", str);
2207 return 0;
2209 early_param("percpu_alloc", percpu_alloc_setup);
2212 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2213 * Build it if needed by the arch config or the generic setup is going
2214 * to be used.
2216 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2217 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2218 #define BUILD_EMBED_FIRST_CHUNK
2219 #endif
2221 /* build pcpu_page_first_chunk() iff needed by the arch config */
2222 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2223 #define BUILD_PAGE_FIRST_CHUNK
2224 #endif
2226 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2227 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2229 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2230 * @reserved_size: the size of reserved percpu area in bytes
2231 * @dyn_size: minimum free size for dynamic allocation in bytes
2232 * @atom_size: allocation atom size
2233 * @cpu_distance_fn: callback to determine distance between cpus, optional
2235 * This function determines grouping of units, their mappings to cpus
2236 * and other parameters considering needed percpu size, allocation
2237 * atom size and distances between CPUs.
2239 * Groups are always multiples of atom size and CPUs which are of
2240 * LOCAL_DISTANCE both ways are grouped together and share space for
2241 * units in the same group. The returned configuration is guaranteed
2242 * to have CPUs on different nodes on different groups and >=75% usage
2243 * of allocated virtual address space.
2245 * RETURNS:
2246 * On success, pointer to the new allocation_info is returned. On
2247 * failure, ERR_PTR value is returned.
2249 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2250 size_t reserved_size, size_t dyn_size,
2251 size_t atom_size,
2252 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2254 static int group_map[NR_CPUS] __initdata;
2255 static int group_cnt[NR_CPUS] __initdata;
2256 const size_t static_size = __per_cpu_end - __per_cpu_start;
2257 int nr_groups = 1, nr_units = 0;
2258 size_t size_sum, min_unit_size, alloc_size;
2259 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
2260 int last_allocs, group, unit;
2261 unsigned int cpu, tcpu;
2262 struct pcpu_alloc_info *ai;
2263 unsigned int *cpu_map;
2265 /* this function may be called multiple times */
2266 memset(group_map, 0, sizeof(group_map));
2267 memset(group_cnt, 0, sizeof(group_cnt));
2269 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2270 size_sum = PFN_ALIGN(static_size + reserved_size +
2271 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2272 dyn_size = size_sum - static_size - reserved_size;
2275 * Determine min_unit_size, alloc_size and max_upa such that
2276 * alloc_size is multiple of atom_size and is the smallest
2277 * which can accommodate 4k aligned segments which are equal to
2278 * or larger than min_unit_size.
2280 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2282 /* determine the maximum # of units that can fit in an allocation */
2283 alloc_size = roundup(min_unit_size, atom_size);
2284 upa = alloc_size / min_unit_size;
2285 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2286 upa--;
2287 max_upa = upa;
2289 /* group cpus according to their proximity */
2290 for_each_possible_cpu(cpu) {
2291 group = 0;
2292 next_group:
2293 for_each_possible_cpu(tcpu) {
2294 if (cpu == tcpu)
2295 break;
2296 if (group_map[tcpu] == group && cpu_distance_fn &&
2297 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2298 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2299 group++;
2300 nr_groups = max(nr_groups, group + 1);
2301 goto next_group;
2304 group_map[cpu] = group;
2305 group_cnt[group]++;
2309 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2310 * Expand the unit_size until we use >= 75% of the units allocated.
2311 * Related to atom_size, which could be much larger than the unit_size.
2313 last_allocs = INT_MAX;
2314 for (upa = max_upa; upa; upa--) {
2315 int allocs = 0, wasted = 0;
2317 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2318 continue;
2320 for (group = 0; group < nr_groups; group++) {
2321 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2322 allocs += this_allocs;
2323 wasted += this_allocs * upa - group_cnt[group];
2327 * Don't accept if wastage is over 1/3. The
2328 * greater-than comparison ensures upa==1 always
2329 * passes the following check.
2331 if (wasted > num_possible_cpus() / 3)
2332 continue;
2334 /* and then don't consume more memory */
2335 if (allocs > last_allocs)
2336 break;
2337 last_allocs = allocs;
2338 best_upa = upa;
2340 upa = best_upa;
2342 /* allocate and fill alloc_info */
2343 for (group = 0; group < nr_groups; group++)
2344 nr_units += roundup(group_cnt[group], upa);
2346 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2347 if (!ai)
2348 return ERR_PTR(-ENOMEM);
2349 cpu_map = ai->groups[0].cpu_map;
2351 for (group = 0; group < nr_groups; group++) {
2352 ai->groups[group].cpu_map = cpu_map;
2353 cpu_map += roundup(group_cnt[group], upa);
2356 ai->static_size = static_size;
2357 ai->reserved_size = reserved_size;
2358 ai->dyn_size = dyn_size;
2359 ai->unit_size = alloc_size / upa;
2360 ai->atom_size = atom_size;
2361 ai->alloc_size = alloc_size;
2363 for (group = 0, unit = 0; group_cnt[group]; group++) {
2364 struct pcpu_group_info *gi = &ai->groups[group];
2367 * Initialize base_offset as if all groups are located
2368 * back-to-back. The caller should update this to
2369 * reflect actual allocation.
2371 gi->base_offset = unit * ai->unit_size;
2373 for_each_possible_cpu(cpu)
2374 if (group_map[cpu] == group)
2375 gi->cpu_map[gi->nr_units++] = cpu;
2376 gi->nr_units = roundup(gi->nr_units, upa);
2377 unit += gi->nr_units;
2379 BUG_ON(unit != nr_units);
2381 return ai;
2383 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2385 #if defined(BUILD_EMBED_FIRST_CHUNK)
2387 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2388 * @reserved_size: the size of reserved percpu area in bytes
2389 * @dyn_size: minimum free size for dynamic allocation in bytes
2390 * @atom_size: allocation atom size
2391 * @cpu_distance_fn: callback to determine distance between cpus, optional
2392 * @alloc_fn: function to allocate percpu page
2393 * @free_fn: function to free percpu page
2395 * This is a helper to ease setting up embedded first percpu chunk and
2396 * can be called where pcpu_setup_first_chunk() is expected.
2398 * If this function is used to setup the first chunk, it is allocated
2399 * by calling @alloc_fn and used as-is without being mapped into
2400 * vmalloc area. Allocations are always whole multiples of @atom_size
2401 * aligned to @atom_size.
2403 * This enables the first chunk to piggy back on the linear physical
2404 * mapping which often uses larger page size. Please note that this
2405 * can result in very sparse cpu->unit mapping on NUMA machines thus
2406 * requiring large vmalloc address space. Don't use this allocator if
2407 * vmalloc space is not orders of magnitude larger than distances
2408 * between node memory addresses (ie. 32bit NUMA machines).
2410 * @dyn_size specifies the minimum dynamic area size.
2412 * If the needed size is smaller than the minimum or specified unit
2413 * size, the leftover is returned using @free_fn.
2415 * RETURNS:
2416 * 0 on success, -errno on failure.
2418 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2419 size_t atom_size,
2420 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2421 pcpu_fc_alloc_fn_t alloc_fn,
2422 pcpu_fc_free_fn_t free_fn)
2424 void *base = (void *)ULONG_MAX;
2425 void **areas = NULL;
2426 struct pcpu_alloc_info *ai;
2427 size_t size_sum, areas_size;
2428 unsigned long max_distance;
2429 int group, i, highest_group, rc;
2431 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2432 cpu_distance_fn);
2433 if (IS_ERR(ai))
2434 return PTR_ERR(ai);
2436 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2437 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2439 areas = memblock_virt_alloc_nopanic(areas_size, 0);
2440 if (!areas) {
2441 rc = -ENOMEM;
2442 goto out_free;
2445 /* allocate, copy and determine base address & max_distance */
2446 highest_group = 0;
2447 for (group = 0; group < ai->nr_groups; group++) {
2448 struct pcpu_group_info *gi = &ai->groups[group];
2449 unsigned int cpu = NR_CPUS;
2450 void *ptr;
2452 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2453 cpu = gi->cpu_map[i];
2454 BUG_ON(cpu == NR_CPUS);
2456 /* allocate space for the whole group */
2457 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2458 if (!ptr) {
2459 rc = -ENOMEM;
2460 goto out_free_areas;
2462 /* kmemleak tracks the percpu allocations separately */
2463 kmemleak_free(ptr);
2464 areas[group] = ptr;
2466 base = min(ptr, base);
2467 if (ptr > areas[highest_group])
2468 highest_group = group;
2470 max_distance = areas[highest_group] - base;
2471 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2473 /* warn if maximum distance is further than 75% of vmalloc space */
2474 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2475 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2476 max_distance, VMALLOC_TOTAL);
2477 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2478 /* and fail if we have fallback */
2479 rc = -EINVAL;
2480 goto out_free_areas;
2481 #endif
2485 * Copy data and free unused parts. This should happen after all
2486 * allocations are complete; otherwise, we may end up with
2487 * overlapping groups.
2489 for (group = 0; group < ai->nr_groups; group++) {
2490 struct pcpu_group_info *gi = &ai->groups[group];
2491 void *ptr = areas[group];
2493 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2494 if (gi->cpu_map[i] == NR_CPUS) {
2495 /* unused unit, free whole */
2496 free_fn(ptr, ai->unit_size);
2497 continue;
2499 /* copy and return the unused part */
2500 memcpy(ptr, __per_cpu_load, ai->static_size);
2501 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2505 /* base address is now known, determine group base offsets */
2506 for (group = 0; group < ai->nr_groups; group++) {
2507 ai->groups[group].base_offset = areas[group] - base;
2510 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2511 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2512 ai->dyn_size, ai->unit_size);
2514 rc = pcpu_setup_first_chunk(ai, base);
2515 goto out_free;
2517 out_free_areas:
2518 for (group = 0; group < ai->nr_groups; group++)
2519 if (areas[group])
2520 free_fn(areas[group],
2521 ai->groups[group].nr_units * ai->unit_size);
2522 out_free:
2523 pcpu_free_alloc_info(ai);
2524 if (areas)
2525 memblock_free_early(__pa(areas), areas_size);
2526 return rc;
2528 #endif /* BUILD_EMBED_FIRST_CHUNK */
2530 #ifdef BUILD_PAGE_FIRST_CHUNK
2532 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2533 * @reserved_size: the size of reserved percpu area in bytes
2534 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2535 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2536 * @populate_pte_fn: function to populate pte
2538 * This is a helper to ease setting up page-remapped first percpu
2539 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2541 * This is the basic allocator. Static percpu area is allocated
2542 * page-by-page into vmalloc area.
2544 * RETURNS:
2545 * 0 on success, -errno on failure.
2547 int __init pcpu_page_first_chunk(size_t reserved_size,
2548 pcpu_fc_alloc_fn_t alloc_fn,
2549 pcpu_fc_free_fn_t free_fn,
2550 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2552 static struct vm_struct vm;
2553 struct pcpu_alloc_info *ai;
2554 char psize_str[16];
2555 int unit_pages;
2556 size_t pages_size;
2557 struct page **pages;
2558 int unit, i, j, rc;
2559 int upa;
2560 int nr_g0_units;
2562 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2564 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2565 if (IS_ERR(ai))
2566 return PTR_ERR(ai);
2567 BUG_ON(ai->nr_groups != 1);
2568 upa = ai->alloc_size/ai->unit_size;
2569 nr_g0_units = roundup(num_possible_cpus(), upa);
2570 if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2571 pcpu_free_alloc_info(ai);
2572 return -EINVAL;
2575 unit_pages = ai->unit_size >> PAGE_SHIFT;
2577 /* unaligned allocations can't be freed, round up to page size */
2578 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2579 sizeof(pages[0]));
2580 pages = memblock_virt_alloc(pages_size, 0);
2582 /* allocate pages */
2583 j = 0;
2584 for (unit = 0; unit < num_possible_cpus(); unit++) {
2585 unsigned int cpu = ai->groups[0].cpu_map[unit];
2586 for (i = 0; i < unit_pages; i++) {
2587 void *ptr;
2589 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2590 if (!ptr) {
2591 pr_warn("failed to allocate %s page for cpu%u\n",
2592 psize_str, cpu);
2593 goto enomem;
2595 /* kmemleak tracks the percpu allocations separately */
2596 kmemleak_free(ptr);
2597 pages[j++] = virt_to_page(ptr);
2601 /* allocate vm area, map the pages and copy static data */
2602 vm.flags = VM_ALLOC;
2603 vm.size = num_possible_cpus() * ai->unit_size;
2604 vm_area_register_early(&vm, PAGE_SIZE);
2606 for (unit = 0; unit < num_possible_cpus(); unit++) {
2607 unsigned long unit_addr =
2608 (unsigned long)vm.addr + unit * ai->unit_size;
2610 for (i = 0; i < unit_pages; i++)
2611 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2613 /* pte already populated, the following shouldn't fail */
2614 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2615 unit_pages);
2616 if (rc < 0)
2617 panic("failed to map percpu area, err=%d\n", rc);
2620 * FIXME: Archs with virtual cache should flush local
2621 * cache for the linear mapping here - something
2622 * equivalent to flush_cache_vmap() on the local cpu.
2623 * flush_cache_vmap() can't be used as most supporting
2624 * data structures are not set up yet.
2627 /* copy static data */
2628 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2631 /* we're ready, commit */
2632 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2633 unit_pages, psize_str, ai->static_size,
2634 ai->reserved_size, ai->dyn_size);
2636 rc = pcpu_setup_first_chunk(ai, vm.addr);
2637 goto out_free_ar;
2639 enomem:
2640 while (--j >= 0)
2641 free_fn(page_address(pages[j]), PAGE_SIZE);
2642 rc = -ENOMEM;
2643 out_free_ar:
2644 memblock_free_early(__pa(pages), pages_size);
2645 pcpu_free_alloc_info(ai);
2646 return rc;
2648 #endif /* BUILD_PAGE_FIRST_CHUNK */
2650 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2652 * Generic SMP percpu area setup.
2654 * The embedding helper is used because its behavior closely resembles
2655 * the original non-dynamic generic percpu area setup. This is
2656 * important because many archs have addressing restrictions and might
2657 * fail if the percpu area is located far away from the previous
2658 * location. As an added bonus, in non-NUMA cases, embedding is
2659 * generally a good idea TLB-wise because percpu area can piggy back
2660 * on the physical linear memory mapping which uses large page
2661 * mappings on applicable archs.
2663 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2664 EXPORT_SYMBOL(__per_cpu_offset);
2666 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2667 size_t align)
2669 return memblock_virt_alloc_from_nopanic(
2670 size, align, __pa(MAX_DMA_ADDRESS));
2673 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2675 memblock_free_early(__pa(ptr), size);
2678 void __init setup_per_cpu_areas(void)
2680 unsigned long delta;
2681 unsigned int cpu;
2682 int rc;
2685 * Always reserve area for module percpu variables. That's
2686 * what the legacy allocator did.
2688 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2689 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2690 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2691 if (rc < 0)
2692 panic("Failed to initialize percpu areas.");
2694 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2695 for_each_possible_cpu(cpu)
2696 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2698 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2700 #else /* CONFIG_SMP */
2703 * UP percpu area setup.
2705 * UP always uses km-based percpu allocator with identity mapping.
2706 * Static percpu variables are indistinguishable from the usual static
2707 * variables and don't require any special preparation.
2709 void __init setup_per_cpu_areas(void)
2711 const size_t unit_size =
2712 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2713 PERCPU_DYNAMIC_RESERVE));
2714 struct pcpu_alloc_info *ai;
2715 void *fc;
2717 ai = pcpu_alloc_alloc_info(1, 1);
2718 fc = memblock_virt_alloc_from_nopanic(unit_size,
2719 PAGE_SIZE,
2720 __pa(MAX_DMA_ADDRESS));
2721 if (!ai || !fc)
2722 panic("Failed to allocate memory for percpu areas.");
2723 /* kmemleak tracks the percpu allocations separately */
2724 kmemleak_free(fc);
2726 ai->dyn_size = unit_size;
2727 ai->unit_size = unit_size;
2728 ai->atom_size = unit_size;
2729 ai->alloc_size = unit_size;
2730 ai->groups[0].nr_units = 1;
2731 ai->groups[0].cpu_map[0] = 0;
2733 if (pcpu_setup_first_chunk(ai, fc) < 0)
2734 panic("Failed to initialize percpu areas.");
2737 #endif /* CONFIG_SMP */
2740 * Percpu allocator is initialized early during boot when neither slab or
2741 * workqueue is available. Plug async management until everything is up
2742 * and running.
2744 static int __init percpu_enable_async(void)
2746 pcpu_async_enabled = true;
2747 return 0;
2749 subsys_initcall(percpu_enable_async);