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.
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>
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>
84 #include <asm/cacheflush.h>
85 #include <asm/sections.h>
86 #include <asm/tlbflush.h>
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/percpu.h>
92 #include "percpu-internal.h"
94 /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
95 #define PCPU_SLOT_BASE_SHIFT 5
97 #define PCPU_EMPTY_POP_PAGES_LOW 2
98 #define PCPU_EMPTY_POP_PAGES_HIGH 4
101 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
102 #ifndef __addr_to_pcpu_ptr
103 #define __addr_to_pcpu_ptr(addr) \
104 (void __percpu *)((unsigned long)(addr) - \
105 (unsigned long)pcpu_base_addr + \
106 (unsigned long)__per_cpu_start)
108 #ifndef __pcpu_ptr_to_addr
109 #define __pcpu_ptr_to_addr(ptr) \
110 (void __force *)((unsigned long)(ptr) + \
111 (unsigned long)pcpu_base_addr - \
112 (unsigned long)__per_cpu_start)
114 #else /* CONFIG_SMP */
115 /* on UP, it's always identity mapped */
116 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
117 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
118 #endif /* CONFIG_SMP */
120 static int pcpu_unit_pages __ro_after_init
;
121 static int pcpu_unit_size __ro_after_init
;
122 static int pcpu_nr_units __ro_after_init
;
123 static int pcpu_atom_size __ro_after_init
;
124 int pcpu_nr_slots __ro_after_init
;
125 static size_t pcpu_chunk_struct_size __ro_after_init
;
127 /* cpus with the lowest and highest unit addresses */
128 static unsigned int pcpu_low_unit_cpu __ro_after_init
;
129 static unsigned int pcpu_high_unit_cpu __ro_after_init
;
131 /* the address of the first chunk which starts with the kernel static area */
132 void *pcpu_base_addr __ro_after_init
;
133 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
135 static const int *pcpu_unit_map __ro_after_init
; /* cpu -> unit */
136 const unsigned long *pcpu_unit_offsets __ro_after_init
; /* cpu -> unit offset */
138 /* group information, used for vm allocation */
139 static int pcpu_nr_groups __ro_after_init
;
140 static const unsigned long *pcpu_group_offsets __ro_after_init
;
141 static const size_t *pcpu_group_sizes __ro_after_init
;
144 * The first chunk which always exists. Note that unlike other
145 * chunks, this one can be allocated and mapped in several different
146 * ways and thus often doesn't live in the vmalloc area.
148 struct pcpu_chunk
*pcpu_first_chunk __ro_after_init
;
151 * Optional reserved chunk. This chunk reserves part of the first
152 * chunk and serves it for reserved allocations. When the reserved
153 * region doesn't exist, the following variable is NULL.
155 struct pcpu_chunk
*pcpu_reserved_chunk __ro_after_init
;
157 DEFINE_SPINLOCK(pcpu_lock
); /* all internal data structures */
158 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* chunk create/destroy, [de]pop, map ext */
160 struct list_head
*pcpu_slot __ro_after_init
; /* chunk list slots */
162 /* chunks which need their map areas extended, protected by pcpu_lock */
163 static LIST_HEAD(pcpu_map_extend_chunks
);
166 * The number of empty populated pages, protected by pcpu_lock. The
167 * reserved chunk doesn't contribute to the count.
169 int pcpu_nr_empty_pop_pages
;
172 * Balance work is used to populate or destroy chunks asynchronously. We
173 * try to keep the number of populated free pages between
174 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
177 static void pcpu_balance_workfn(struct work_struct
*work
);
178 static DECLARE_WORK(pcpu_balance_work
, pcpu_balance_workfn
);
179 static bool pcpu_async_enabled __read_mostly
;
180 static bool pcpu_atomic_alloc_failed
;
182 static void pcpu_schedule_balance_work(void)
184 if (pcpu_async_enabled
)
185 schedule_work(&pcpu_balance_work
);
189 * pcpu_addr_in_chunk - check if the address is served from this chunk
190 * @chunk: chunk of interest
191 * @addr: percpu address
194 * True if the address is served from this chunk.
196 static bool pcpu_addr_in_chunk(struct pcpu_chunk
*chunk
, void *addr
)
198 void *start_addr
, *end_addr
;
203 start_addr
= chunk
->base_addr
+ chunk
->start_offset
;
204 end_addr
= chunk
->base_addr
+ chunk
->nr_pages
* PAGE_SIZE
-
207 return addr
>= start_addr
&& addr
< end_addr
;
210 static int __pcpu_size_to_slot(int size
)
212 int highbit
= fls(size
); /* size is in bytes */
213 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
216 static int pcpu_size_to_slot(int size
)
218 if (size
== pcpu_unit_size
)
219 return pcpu_nr_slots
- 1;
220 return __pcpu_size_to_slot(size
);
223 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
225 if (chunk
->free_bytes
< PCPU_MIN_ALLOC_SIZE
|| chunk
->contig_bits
== 0)
228 return pcpu_size_to_slot(chunk
->free_bytes
);
231 /* set the pointer to a chunk in a page struct */
232 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
234 page
->index
= (unsigned long)pcpu
;
237 /* obtain pointer to a chunk from a page struct */
238 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
240 return (struct pcpu_chunk
*)page
->index
;
243 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
245 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
248 static unsigned long pcpu_unit_page_offset(unsigned int cpu
, int page_idx
)
250 return pcpu_unit_offsets
[cpu
] + (page_idx
<< PAGE_SHIFT
);
253 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
254 unsigned int cpu
, int page_idx
)
256 return (unsigned long)chunk
->base_addr
+
257 pcpu_unit_page_offset(cpu
, page_idx
);
260 static void pcpu_next_unpop(unsigned long *bitmap
, int *rs
, int *re
, int end
)
262 *rs
= find_next_zero_bit(bitmap
, end
, *rs
);
263 *re
= find_next_bit(bitmap
, end
, *rs
+ 1);
266 static void pcpu_next_pop(unsigned long *bitmap
, int *rs
, int *re
, int end
)
268 *rs
= find_next_bit(bitmap
, end
, *rs
);
269 *re
= find_next_zero_bit(bitmap
, end
, *rs
+ 1);
273 * Bitmap region iterators. Iterates over the bitmap between
274 * [@start, @end) in @chunk. @rs and @re should be integer variables
275 * and will be set to start and end index of the current free region.
277 #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
278 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
280 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
282 #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
283 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
285 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
288 * The following are helper functions to help access bitmaps and convert
289 * between bitmap offsets to address offsets.
291 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk
*chunk
, int index
)
293 return chunk
->alloc_map
+
294 (index
* PCPU_BITMAP_BLOCK_BITS
/ BITS_PER_LONG
);
297 static unsigned long pcpu_off_to_block_index(int off
)
299 return off
/ PCPU_BITMAP_BLOCK_BITS
;
302 static unsigned long pcpu_off_to_block_off(int off
)
304 return off
& (PCPU_BITMAP_BLOCK_BITS
- 1);
307 static unsigned long pcpu_block_off_to_off(int index
, int off
)
309 return index
* PCPU_BITMAP_BLOCK_BITS
+ off
;
313 * pcpu_next_md_free_region - finds the next hint free area
314 * @chunk: chunk of interest
315 * @bit_off: chunk offset
316 * @bits: size of free area
318 * Helper function for pcpu_for_each_md_free_region. It checks
319 * block->contig_hint and performs aggregation across blocks to find the
320 * next hint. It modifies bit_off and bits in-place to be consumed in the
323 static void pcpu_next_md_free_region(struct pcpu_chunk
*chunk
, int *bit_off
,
326 int i
= pcpu_off_to_block_index(*bit_off
);
327 int block_off
= pcpu_off_to_block_off(*bit_off
);
328 struct pcpu_block_md
*block
;
331 for (block
= chunk
->md_blocks
+ i
; i
< pcpu_chunk_nr_blocks(chunk
);
333 /* handles contig area across blocks */
335 *bits
+= block
->left_free
;
336 if (block
->left_free
== PCPU_BITMAP_BLOCK_BITS
)
342 * This checks three things. First is there a contig_hint to
343 * check. Second, have we checked this hint before by
344 * comparing the block_off. Third, is this the same as the
345 * right contig hint. In the last case, it spills over into
346 * the next block and should be handled by the contig area
347 * across blocks code.
349 *bits
= block
->contig_hint
;
350 if (*bits
&& block
->contig_hint_start
>= block_off
&&
351 *bits
+ block
->contig_hint_start
< PCPU_BITMAP_BLOCK_BITS
) {
352 *bit_off
= pcpu_block_off_to_off(i
,
353 block
->contig_hint_start
);
356 /* reset to satisfy the second predicate above */
359 *bits
= block
->right_free
;
360 *bit_off
= (i
+ 1) * PCPU_BITMAP_BLOCK_BITS
- block
->right_free
;
365 * pcpu_next_fit_region - finds fit areas for a given allocation request
366 * @chunk: chunk of interest
367 * @alloc_bits: size of allocation
368 * @align: alignment of area (max PAGE_SIZE)
369 * @bit_off: chunk offset
370 * @bits: size of free area
372 * Finds the next free region that is viable for use with a given size and
373 * alignment. This only returns if there is a valid area to be used for this
374 * allocation. block->first_free is returned if the allocation request fits
375 * within the block to see if the request can be fulfilled prior to the contig
378 static void pcpu_next_fit_region(struct pcpu_chunk
*chunk
, int alloc_bits
,
379 int align
, int *bit_off
, int *bits
)
381 int i
= pcpu_off_to_block_index(*bit_off
);
382 int block_off
= pcpu_off_to_block_off(*bit_off
);
383 struct pcpu_block_md
*block
;
386 for (block
= chunk
->md_blocks
+ i
; i
< pcpu_chunk_nr_blocks(chunk
);
388 /* handles contig area across blocks */
390 *bits
+= block
->left_free
;
391 if (*bits
>= alloc_bits
)
393 if (block
->left_free
== PCPU_BITMAP_BLOCK_BITS
)
397 /* check block->contig_hint */
398 *bits
= ALIGN(block
->contig_hint_start
, align
) -
399 block
->contig_hint_start
;
401 * This uses the block offset to determine if this has been
402 * checked in the prior iteration.
404 if (block
->contig_hint
&&
405 block
->contig_hint_start
>= block_off
&&
406 block
->contig_hint
>= *bits
+ alloc_bits
) {
407 *bits
+= alloc_bits
+ block
->contig_hint_start
-
409 *bit_off
= pcpu_block_off_to_off(i
, block
->first_free
);
412 /* reset to satisfy the second predicate above */
415 *bit_off
= ALIGN(PCPU_BITMAP_BLOCK_BITS
- block
->right_free
,
417 *bits
= PCPU_BITMAP_BLOCK_BITS
- *bit_off
;
418 *bit_off
= pcpu_block_off_to_off(i
, *bit_off
);
419 if (*bits
>= alloc_bits
)
423 /* no valid offsets were found - fail condition */
424 *bit_off
= pcpu_chunk_map_bits(chunk
);
428 * Metadata free area iterators. These perform aggregation of free areas
429 * based on the metadata blocks and return the offset @bit_off and size in
430 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
431 * a fit is found for the allocation request.
433 #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
434 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
435 (bit_off) < pcpu_chunk_map_bits((chunk)); \
436 (bit_off) += (bits) + 1, \
437 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
439 #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
440 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
442 (bit_off) < pcpu_chunk_map_bits((chunk)); \
443 (bit_off) += (bits), \
444 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
448 * pcpu_mem_zalloc - allocate memory
449 * @size: bytes to allocate
450 * @gfp: allocation flags
452 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
453 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
454 * This is to facilitate passing through whitelisted flags. The
455 * returned memory is always zeroed.
458 * Does GFP_KERNEL allocation.
461 * Pointer to the allocated area on success, NULL on failure.
463 static void *pcpu_mem_zalloc(size_t size
, gfp_t gfp
)
465 if (WARN_ON_ONCE(!slab_is_available()))
468 if (size
<= PAGE_SIZE
)
469 return kzalloc(size
, gfp
| GFP_KERNEL
);
471 return __vmalloc(size
, gfp
| GFP_KERNEL
| __GFP_ZERO
,
476 * pcpu_mem_free - free memory
477 * @ptr: memory to free
479 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
481 static void pcpu_mem_free(void *ptr
)
487 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
488 * @chunk: chunk of interest
489 * @oslot: the previous slot it was on
491 * This function is called after an allocation or free changed @chunk.
492 * New slot according to the changed state is determined and @chunk is
493 * moved to the slot. Note that the reserved chunk is never put on
499 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
501 int nslot
= pcpu_chunk_slot(chunk
);
503 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
505 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
507 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
512 * pcpu_cnt_pop_pages- counts populated backing pages in range
513 * @chunk: chunk of interest
514 * @bit_off: start offset
515 * @bits: size of area to check
517 * Calculates the number of populated pages in the region
518 * [page_start, page_end). This keeps track of how many empty populated
519 * pages are available and decide if async work should be scheduled.
522 * The nr of populated pages.
524 static inline int pcpu_cnt_pop_pages(struct pcpu_chunk
*chunk
, int bit_off
,
527 int page_start
= PFN_UP(bit_off
* PCPU_MIN_ALLOC_SIZE
);
528 int page_end
= PFN_DOWN((bit_off
+ bits
) * PCPU_MIN_ALLOC_SIZE
);
530 if (page_start
>= page_end
)
534 * bitmap_weight counts the number of bits set in a bitmap up to
535 * the specified number of bits. This is counting the populated
536 * pages up to page_end and then subtracting the populated pages
537 * up to page_start to count the populated pages in
538 * [page_start, page_end).
540 return bitmap_weight(chunk
->populated
, page_end
) -
541 bitmap_weight(chunk
->populated
, page_start
);
545 * pcpu_chunk_update - updates the chunk metadata given a free area
546 * @chunk: chunk of interest
547 * @bit_off: chunk offset
548 * @bits: size of free area
550 * This updates the chunk's contig hint and starting offset given a free area.
551 * Choose the best starting offset if the contig hint is equal.
553 static void pcpu_chunk_update(struct pcpu_chunk
*chunk
, int bit_off
, int bits
)
555 if (bits
> chunk
->contig_bits
) {
556 chunk
->contig_bits_start
= bit_off
;
557 chunk
->contig_bits
= bits
;
558 } else if (bits
== chunk
->contig_bits
&& chunk
->contig_bits_start
&&
560 __ffs(bit_off
) > __ffs(chunk
->contig_bits_start
))) {
561 /* use the start with the best alignment */
562 chunk
->contig_bits_start
= bit_off
;
567 * pcpu_chunk_refresh_hint - updates metadata about a chunk
568 * @chunk: chunk of interest
570 * Iterates over the metadata blocks to find the largest contig area.
571 * It also counts the populated pages and uses the delta to update the
576 * chunk->contig_bits_start
577 * nr_empty_pop_pages (chunk and global)
579 static void pcpu_chunk_refresh_hint(struct pcpu_chunk
*chunk
)
581 int bit_off
, bits
, nr_empty_pop_pages
;
584 chunk
->contig_bits
= 0;
586 bit_off
= chunk
->first_bit
;
587 bits
= nr_empty_pop_pages
= 0;
588 pcpu_for_each_md_free_region(chunk
, bit_off
, bits
) {
589 pcpu_chunk_update(chunk
, bit_off
, bits
);
591 nr_empty_pop_pages
+= pcpu_cnt_pop_pages(chunk
, bit_off
, bits
);
595 * Keep track of nr_empty_pop_pages.
597 * The chunk maintains the previous number of free pages it held,
598 * so the delta is used to update the global counter. The reserved
599 * chunk is not part of the free page count as they are populated
600 * at init and are special to serving reserved allocations.
602 if (chunk
!= pcpu_reserved_chunk
)
603 pcpu_nr_empty_pop_pages
+=
604 (nr_empty_pop_pages
- chunk
->nr_empty_pop_pages
);
606 chunk
->nr_empty_pop_pages
= nr_empty_pop_pages
;
610 * pcpu_block_update - updates a block given a free area
611 * @block: block of interest
612 * @start: start offset in block
613 * @end: end offset in block
615 * Updates a block given a known free area. The region [start, end) is
616 * expected to be the entirety of the free area within a block. Chooses
617 * the best starting offset if the contig hints are equal.
619 static void pcpu_block_update(struct pcpu_block_md
*block
, int start
, int end
)
621 int contig
= end
- start
;
623 block
->first_free
= min(block
->first_free
, start
);
625 block
->left_free
= contig
;
627 if (end
== PCPU_BITMAP_BLOCK_BITS
)
628 block
->right_free
= contig
;
630 if (contig
> block
->contig_hint
) {
631 block
->contig_hint_start
= start
;
632 block
->contig_hint
= contig
;
633 } else if (block
->contig_hint_start
&& contig
== block
->contig_hint
&&
634 (!start
|| __ffs(start
) > __ffs(block
->contig_hint_start
))) {
635 /* use the start with the best alignment */
636 block
->contig_hint_start
= start
;
641 * pcpu_block_refresh_hint
642 * @chunk: chunk of interest
643 * @index: index of the metadata block
645 * Scans over the block beginning at first_free and updates the block
646 * metadata accordingly.
648 static void pcpu_block_refresh_hint(struct pcpu_chunk
*chunk
, int index
)
650 struct pcpu_block_md
*block
= chunk
->md_blocks
+ index
;
651 unsigned long *alloc_map
= pcpu_index_alloc_map(chunk
, index
);
652 int rs
, re
; /* region start, region end */
655 block
->contig_hint
= 0;
656 block
->left_free
= block
->right_free
= 0;
658 /* iterate over free areas and update the contig hints */
659 pcpu_for_each_unpop_region(alloc_map
, rs
, re
, block
->first_free
,
660 PCPU_BITMAP_BLOCK_BITS
) {
661 pcpu_block_update(block
, rs
, re
);
666 * pcpu_block_update_hint_alloc - update hint on allocation path
667 * @chunk: chunk of interest
668 * @bit_off: chunk offset
669 * @bits: size of request
671 * Updates metadata for the allocation path. The metadata only has to be
672 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
673 * scans are required if the block's contig hint is broken.
675 static void pcpu_block_update_hint_alloc(struct pcpu_chunk
*chunk
, int bit_off
,
678 struct pcpu_block_md
*s_block
, *e_block
, *block
;
679 int s_index
, e_index
; /* block indexes of the freed allocation */
680 int s_off
, e_off
; /* block offsets of the freed allocation */
683 * Calculate per block offsets.
684 * The calculation uses an inclusive range, but the resulting offsets
685 * are [start, end). e_index always points to the last block in the
688 s_index
= pcpu_off_to_block_index(bit_off
);
689 e_index
= pcpu_off_to_block_index(bit_off
+ bits
- 1);
690 s_off
= pcpu_off_to_block_off(bit_off
);
691 e_off
= pcpu_off_to_block_off(bit_off
+ bits
- 1) + 1;
693 s_block
= chunk
->md_blocks
+ s_index
;
694 e_block
= chunk
->md_blocks
+ e_index
;
698 * block->first_free must be updated if the allocation takes its place.
699 * If the allocation breaks the contig_hint, a scan is required to
702 if (s_off
== s_block
->first_free
)
703 s_block
->first_free
= find_next_zero_bit(
704 pcpu_index_alloc_map(chunk
, s_index
),
705 PCPU_BITMAP_BLOCK_BITS
,
708 if (s_off
>= s_block
->contig_hint_start
&&
709 s_off
< s_block
->contig_hint_start
+ s_block
->contig_hint
) {
710 /* block contig hint is broken - scan to fix it */
711 pcpu_block_refresh_hint(chunk
, s_index
);
713 /* update left and right contig manually */
714 s_block
->left_free
= min(s_block
->left_free
, s_off
);
715 if (s_index
== e_index
)
716 s_block
->right_free
= min_t(int, s_block
->right_free
,
717 PCPU_BITMAP_BLOCK_BITS
- e_off
);
719 s_block
->right_free
= 0;
725 if (s_index
!= e_index
) {
727 * When the allocation is across blocks, the end is along
728 * the left part of the e_block.
730 e_block
->first_free
= find_next_zero_bit(
731 pcpu_index_alloc_map(chunk
, e_index
),
732 PCPU_BITMAP_BLOCK_BITS
, e_off
);
734 if (e_off
== PCPU_BITMAP_BLOCK_BITS
) {
735 /* reset the block */
738 if (e_off
> e_block
->contig_hint_start
) {
739 /* contig hint is broken - scan to fix it */
740 pcpu_block_refresh_hint(chunk
, e_index
);
742 e_block
->left_free
= 0;
743 e_block
->right_free
=
744 min_t(int, e_block
->right_free
,
745 PCPU_BITMAP_BLOCK_BITS
- e_off
);
749 /* update in-between md_blocks */
750 for (block
= s_block
+ 1; block
< e_block
; block
++) {
751 block
->contig_hint
= 0;
752 block
->left_free
= 0;
753 block
->right_free
= 0;
758 * The only time a full chunk scan is required is if the chunk
759 * contig hint is broken. Otherwise, it means a smaller space
760 * was used and therefore the chunk contig hint is still correct.
762 if (bit_off
>= chunk
->contig_bits_start
&&
763 bit_off
< chunk
->contig_bits_start
+ chunk
->contig_bits
)
764 pcpu_chunk_refresh_hint(chunk
);
768 * pcpu_block_update_hint_free - updates the block hints on the free path
769 * @chunk: chunk of interest
770 * @bit_off: chunk offset
771 * @bits: size of request
773 * Updates metadata for the allocation path. This avoids a blind block
774 * refresh by making use of the block contig hints. If this fails, it scans
775 * forward and backward to determine the extent of the free area. This is
776 * capped at the boundary of blocks.
778 * A chunk update is triggered if a page becomes free, a block becomes free,
779 * or the free spans across blocks. This tradeoff is to minimize iterating
780 * over the block metadata to update chunk->contig_bits. chunk->contig_bits
781 * may be off by up to a page, but it will never be more than the available
782 * space. If the contig hint is contained in one block, it will be accurate.
784 static void pcpu_block_update_hint_free(struct pcpu_chunk
*chunk
, int bit_off
,
787 struct pcpu_block_md
*s_block
, *e_block
, *block
;
788 int s_index
, e_index
; /* block indexes of the freed allocation */
789 int s_off
, e_off
; /* block offsets of the freed allocation */
790 int start
, end
; /* start and end of the whole free area */
793 * Calculate per block offsets.
794 * The calculation uses an inclusive range, but the resulting offsets
795 * are [start, end). e_index always points to the last block in the
798 s_index
= pcpu_off_to_block_index(bit_off
);
799 e_index
= pcpu_off_to_block_index(bit_off
+ bits
- 1);
800 s_off
= pcpu_off_to_block_off(bit_off
);
801 e_off
= pcpu_off_to_block_off(bit_off
+ bits
- 1) + 1;
803 s_block
= chunk
->md_blocks
+ s_index
;
804 e_block
= chunk
->md_blocks
+ e_index
;
807 * Check if the freed area aligns with the block->contig_hint.
808 * If it does, then the scan to find the beginning/end of the
809 * larger free area can be avoided.
811 * start and end refer to beginning and end of the free area
812 * within each their respective blocks. This is not necessarily
813 * the entire free area as it may span blocks past the beginning
814 * or end of the block.
817 if (s_off
== s_block
->contig_hint
+ s_block
->contig_hint_start
) {
818 start
= s_block
->contig_hint_start
;
821 * Scan backwards to find the extent of the free area.
822 * find_last_bit returns the starting bit, so if the start bit
823 * is returned, that means there was no last bit and the
824 * remainder of the chunk is free.
826 int l_bit
= find_last_bit(pcpu_index_alloc_map(chunk
, s_index
),
828 start
= (start
== l_bit
) ? 0 : l_bit
+ 1;
832 if (e_off
== e_block
->contig_hint_start
)
833 end
= e_block
->contig_hint_start
+ e_block
->contig_hint
;
835 end
= find_next_bit(pcpu_index_alloc_map(chunk
, e_index
),
836 PCPU_BITMAP_BLOCK_BITS
, end
);
839 e_off
= (s_index
== e_index
) ? end
: PCPU_BITMAP_BLOCK_BITS
;
840 pcpu_block_update(s_block
, start
, e_off
);
842 /* freeing in the same block */
843 if (s_index
!= e_index
) {
845 pcpu_block_update(e_block
, 0, end
);
847 /* reset md_blocks in the middle */
848 for (block
= s_block
+ 1; block
< e_block
; block
++) {
849 block
->first_free
= 0;
850 block
->contig_hint_start
= 0;
851 block
->contig_hint
= PCPU_BITMAP_BLOCK_BITS
;
852 block
->left_free
= PCPU_BITMAP_BLOCK_BITS
;
853 block
->right_free
= PCPU_BITMAP_BLOCK_BITS
;
858 * Refresh chunk metadata when the free makes a page free, a block
859 * free, or spans across blocks. The contig hint may be off by up to
860 * a page, but if the hint is contained in a block, it will be accurate
861 * with the else condition below.
863 if ((ALIGN_DOWN(end
, min(PCPU_BITS_PER_PAGE
, PCPU_BITMAP_BLOCK_BITS
)) >
864 ALIGN(start
, min(PCPU_BITS_PER_PAGE
, PCPU_BITMAP_BLOCK_BITS
))) ||
866 pcpu_chunk_refresh_hint(chunk
);
868 pcpu_chunk_update(chunk
, pcpu_block_off_to_off(s_index
, start
),
869 s_block
->contig_hint
);
873 * pcpu_is_populated - determines if the region is populated
874 * @chunk: chunk of interest
875 * @bit_off: chunk offset
876 * @bits: size of area
877 * @next_off: return value for the next offset to start searching
879 * For atomic allocations, check if the backing pages are populated.
882 * Bool if the backing pages are populated.
883 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
885 static bool pcpu_is_populated(struct pcpu_chunk
*chunk
, int bit_off
, int bits
,
888 int page_start
, page_end
, rs
, re
;
890 page_start
= PFN_DOWN(bit_off
* PCPU_MIN_ALLOC_SIZE
);
891 page_end
= PFN_UP((bit_off
+ bits
) * PCPU_MIN_ALLOC_SIZE
);
894 pcpu_next_unpop(chunk
->populated
, &rs
, &re
, page_end
);
898 *next_off
= re
* PAGE_SIZE
/ PCPU_MIN_ALLOC_SIZE
;
903 * pcpu_find_block_fit - finds the block index to start searching
904 * @chunk: chunk of interest
905 * @alloc_bits: size of request in allocation units
906 * @align: alignment of area (max PAGE_SIZE bytes)
907 * @pop_only: use populated regions only
909 * Given a chunk and an allocation spec, find the offset to begin searching
910 * for a free region. This iterates over the bitmap metadata blocks to
911 * find an offset that will be guaranteed to fit the requirements. It is
912 * not quite first fit as if the allocation does not fit in the contig hint
913 * of a block or chunk, it is skipped. This errs on the side of caution
914 * to prevent excess iteration. Poor alignment can cause the allocator to
915 * skip over blocks and chunks that have valid free areas.
918 * The offset in the bitmap to begin searching.
919 * -1 if no offset is found.
921 static int pcpu_find_block_fit(struct pcpu_chunk
*chunk
, int alloc_bits
,
922 size_t align
, bool pop_only
)
924 int bit_off
, bits
, next_off
;
927 * Check to see if the allocation can fit in the chunk's contig hint.
928 * This is an optimization to prevent scanning by assuming if it
929 * cannot fit in the global hint, there is memory pressure and creating
930 * a new chunk would happen soon.
932 bit_off
= ALIGN(chunk
->contig_bits_start
, align
) -
933 chunk
->contig_bits_start
;
934 if (bit_off
+ alloc_bits
> chunk
->contig_bits
)
937 bit_off
= chunk
->first_bit
;
939 pcpu_for_each_fit_region(chunk
, alloc_bits
, align
, bit_off
, bits
) {
940 if (!pop_only
|| pcpu_is_populated(chunk
, bit_off
, bits
,
948 if (bit_off
== pcpu_chunk_map_bits(chunk
))
955 * pcpu_alloc_area - allocates an area from a pcpu_chunk
956 * @chunk: chunk of interest
957 * @alloc_bits: size of request in allocation units
958 * @align: alignment of area (max PAGE_SIZE)
959 * @start: bit_off to start searching
961 * This function takes in a @start offset to begin searching to fit an
962 * allocation of @alloc_bits with alignment @align. It needs to scan
963 * the allocation map because if it fits within the block's contig hint,
964 * @start will be block->first_free. This is an attempt to fill the
965 * allocation prior to breaking the contig hint. The allocation and
966 * boundary maps are updated accordingly if it confirms a valid
970 * Allocated addr offset in @chunk on success.
971 * -1 if no matching area is found.
973 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int alloc_bits
,
974 size_t align
, int start
)
976 size_t align_mask
= (align
) ? (align
- 1) : 0;
977 int bit_off
, end
, oslot
;
979 lockdep_assert_held(&pcpu_lock
);
981 oslot
= pcpu_chunk_slot(chunk
);
984 * Search to find a fit.
986 end
= start
+ alloc_bits
+ PCPU_BITMAP_BLOCK_BITS
;
987 bit_off
= bitmap_find_next_zero_area(chunk
->alloc_map
, end
, start
,
988 alloc_bits
, align_mask
);
992 /* update alloc map */
993 bitmap_set(chunk
->alloc_map
, bit_off
, alloc_bits
);
995 /* update boundary map */
996 set_bit(bit_off
, chunk
->bound_map
);
997 bitmap_clear(chunk
->bound_map
, bit_off
+ 1, alloc_bits
- 1);
998 set_bit(bit_off
+ alloc_bits
, chunk
->bound_map
);
1000 chunk
->free_bytes
-= alloc_bits
* PCPU_MIN_ALLOC_SIZE
;
1002 /* update first free bit */
1003 if (bit_off
== chunk
->first_bit
)
1004 chunk
->first_bit
= find_next_zero_bit(
1006 pcpu_chunk_map_bits(chunk
),
1007 bit_off
+ alloc_bits
);
1009 pcpu_block_update_hint_alloc(chunk
, bit_off
, alloc_bits
);
1011 pcpu_chunk_relocate(chunk
, oslot
);
1013 return bit_off
* PCPU_MIN_ALLOC_SIZE
;
1017 * pcpu_free_area - frees the corresponding offset
1018 * @chunk: chunk of interest
1019 * @off: addr offset into chunk
1021 * This function determines the size of an allocation to free using
1022 * the boundary bitmap and clears the allocation map.
1024 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int off
)
1026 int bit_off
, bits
, end
, oslot
;
1028 lockdep_assert_held(&pcpu_lock
);
1029 pcpu_stats_area_dealloc(chunk
);
1031 oslot
= pcpu_chunk_slot(chunk
);
1033 bit_off
= off
/ PCPU_MIN_ALLOC_SIZE
;
1035 /* find end index */
1036 end
= find_next_bit(chunk
->bound_map
, pcpu_chunk_map_bits(chunk
),
1038 bits
= end
- bit_off
;
1039 bitmap_clear(chunk
->alloc_map
, bit_off
, bits
);
1041 /* update metadata */
1042 chunk
->free_bytes
+= bits
* PCPU_MIN_ALLOC_SIZE
;
1044 /* update first free bit */
1045 chunk
->first_bit
= min(chunk
->first_bit
, bit_off
);
1047 pcpu_block_update_hint_free(chunk
, bit_off
, bits
);
1049 pcpu_chunk_relocate(chunk
, oslot
);
1052 static void pcpu_init_md_blocks(struct pcpu_chunk
*chunk
)
1054 struct pcpu_block_md
*md_block
;
1056 for (md_block
= chunk
->md_blocks
;
1057 md_block
!= chunk
->md_blocks
+ pcpu_chunk_nr_blocks(chunk
);
1059 md_block
->contig_hint
= PCPU_BITMAP_BLOCK_BITS
;
1060 md_block
->left_free
= PCPU_BITMAP_BLOCK_BITS
;
1061 md_block
->right_free
= PCPU_BITMAP_BLOCK_BITS
;
1066 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1067 * @tmp_addr: the start of the region served
1068 * @map_size: size of the region served
1070 * This is responsible for creating the chunks that serve the first chunk. The
1071 * base_addr is page aligned down of @tmp_addr while the region end is page
1072 * aligned up. Offsets are kept track of to determine the region served. All
1073 * this is done to appease the bitmap allocator in avoiding partial blocks.
1076 * Chunk serving the region at @tmp_addr of @map_size.
1078 static struct pcpu_chunk
* __init
pcpu_alloc_first_chunk(unsigned long tmp_addr
,
1081 struct pcpu_chunk
*chunk
;
1082 unsigned long aligned_addr
, lcm_align
;
1083 int start_offset
, offset_bits
, region_size
, region_bits
;
1085 /* region calculations */
1086 aligned_addr
= tmp_addr
& PAGE_MASK
;
1088 start_offset
= tmp_addr
- aligned_addr
;
1091 * Align the end of the region with the LCM of PAGE_SIZE and
1092 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1095 lcm_align
= lcm(PAGE_SIZE
, PCPU_BITMAP_BLOCK_SIZE
);
1096 region_size
= ALIGN(start_offset
+ map_size
, lcm_align
);
1098 /* allocate chunk */
1099 chunk
= memblock_virt_alloc(sizeof(struct pcpu_chunk
) +
1100 BITS_TO_LONGS(region_size
>> PAGE_SHIFT
),
1103 INIT_LIST_HEAD(&chunk
->list
);
1105 chunk
->base_addr
= (void *)aligned_addr
;
1106 chunk
->start_offset
= start_offset
;
1107 chunk
->end_offset
= region_size
- chunk
->start_offset
- map_size
;
1109 chunk
->nr_pages
= region_size
>> PAGE_SHIFT
;
1110 region_bits
= pcpu_chunk_map_bits(chunk
);
1112 chunk
->alloc_map
= memblock_virt_alloc(BITS_TO_LONGS(region_bits
) *
1113 sizeof(chunk
->alloc_map
[0]), 0);
1114 chunk
->bound_map
= memblock_virt_alloc(BITS_TO_LONGS(region_bits
+ 1) *
1115 sizeof(chunk
->bound_map
[0]), 0);
1116 chunk
->md_blocks
= memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk
) *
1117 sizeof(chunk
->md_blocks
[0]), 0);
1118 pcpu_init_md_blocks(chunk
);
1120 /* manage populated page bitmap */
1121 chunk
->immutable
= true;
1122 bitmap_fill(chunk
->populated
, chunk
->nr_pages
);
1123 chunk
->nr_populated
= chunk
->nr_pages
;
1124 chunk
->nr_empty_pop_pages
=
1125 pcpu_cnt_pop_pages(chunk
, start_offset
/ PCPU_MIN_ALLOC_SIZE
,
1126 map_size
/ PCPU_MIN_ALLOC_SIZE
);
1128 chunk
->contig_bits
= map_size
/ PCPU_MIN_ALLOC_SIZE
;
1129 chunk
->free_bytes
= map_size
;
1131 if (chunk
->start_offset
) {
1132 /* hide the beginning of the bitmap */
1133 offset_bits
= chunk
->start_offset
/ PCPU_MIN_ALLOC_SIZE
;
1134 bitmap_set(chunk
->alloc_map
, 0, offset_bits
);
1135 set_bit(0, chunk
->bound_map
);
1136 set_bit(offset_bits
, chunk
->bound_map
);
1138 chunk
->first_bit
= offset_bits
;
1140 pcpu_block_update_hint_alloc(chunk
, 0, offset_bits
);
1143 if (chunk
->end_offset
) {
1144 /* hide the end of the bitmap */
1145 offset_bits
= chunk
->end_offset
/ PCPU_MIN_ALLOC_SIZE
;
1146 bitmap_set(chunk
->alloc_map
,
1147 pcpu_chunk_map_bits(chunk
) - offset_bits
,
1149 set_bit((start_offset
+ map_size
) / PCPU_MIN_ALLOC_SIZE
,
1151 set_bit(region_bits
, chunk
->bound_map
);
1153 pcpu_block_update_hint_alloc(chunk
, pcpu_chunk_map_bits(chunk
)
1154 - offset_bits
, offset_bits
);
1160 static struct pcpu_chunk
*pcpu_alloc_chunk(gfp_t gfp
)
1162 struct pcpu_chunk
*chunk
;
1165 chunk
= pcpu_mem_zalloc(pcpu_chunk_struct_size
, gfp
);
1169 INIT_LIST_HEAD(&chunk
->list
);
1170 chunk
->nr_pages
= pcpu_unit_pages
;
1171 region_bits
= pcpu_chunk_map_bits(chunk
);
1173 chunk
->alloc_map
= pcpu_mem_zalloc(BITS_TO_LONGS(region_bits
) *
1174 sizeof(chunk
->alloc_map
[0]), gfp
);
1175 if (!chunk
->alloc_map
)
1176 goto alloc_map_fail
;
1178 chunk
->bound_map
= pcpu_mem_zalloc(BITS_TO_LONGS(region_bits
+ 1) *
1179 sizeof(chunk
->bound_map
[0]), gfp
);
1180 if (!chunk
->bound_map
)
1181 goto bound_map_fail
;
1183 chunk
->md_blocks
= pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk
) *
1184 sizeof(chunk
->md_blocks
[0]), gfp
);
1185 if (!chunk
->md_blocks
)
1186 goto md_blocks_fail
;
1188 pcpu_init_md_blocks(chunk
);
1191 chunk
->contig_bits
= region_bits
;
1192 chunk
->free_bytes
= chunk
->nr_pages
* PAGE_SIZE
;
1197 pcpu_mem_free(chunk
->bound_map
);
1199 pcpu_mem_free(chunk
->alloc_map
);
1201 pcpu_mem_free(chunk
);
1206 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
1210 pcpu_mem_free(chunk
->bound_map
);
1211 pcpu_mem_free(chunk
->alloc_map
);
1212 pcpu_mem_free(chunk
);
1216 * pcpu_chunk_populated - post-population bookkeeping
1217 * @chunk: pcpu_chunk which got populated
1218 * @page_start: the start page
1219 * @page_end: the end page
1220 * @for_alloc: if this is to populate for allocation
1222 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1223 * the bookkeeping information accordingly. Must be called after each
1224 * successful population.
1226 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1227 * is to serve an allocation in that area.
1229 static void pcpu_chunk_populated(struct pcpu_chunk
*chunk
, int page_start
,
1230 int page_end
, bool for_alloc
)
1232 int nr
= page_end
- page_start
;
1234 lockdep_assert_held(&pcpu_lock
);
1236 bitmap_set(chunk
->populated
, page_start
, nr
);
1237 chunk
->nr_populated
+= nr
;
1240 chunk
->nr_empty_pop_pages
+= nr
;
1241 pcpu_nr_empty_pop_pages
+= nr
;
1246 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1247 * @chunk: pcpu_chunk which got depopulated
1248 * @page_start: the start page
1249 * @page_end: the end page
1251 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1252 * Update the bookkeeping information accordingly. Must be called after
1253 * each successful depopulation.
1255 static void pcpu_chunk_depopulated(struct pcpu_chunk
*chunk
,
1256 int page_start
, int page_end
)
1258 int nr
= page_end
- page_start
;
1260 lockdep_assert_held(&pcpu_lock
);
1262 bitmap_clear(chunk
->populated
, page_start
, nr
);
1263 chunk
->nr_populated
-= nr
;
1264 chunk
->nr_empty_pop_pages
-= nr
;
1265 pcpu_nr_empty_pop_pages
-= nr
;
1269 * Chunk management implementation.
1271 * To allow different implementations, chunk alloc/free and
1272 * [de]population are implemented in a separate file which is pulled
1273 * into this file and compiled together. The following functions
1274 * should be implemented.
1276 * pcpu_populate_chunk - populate the specified range of a chunk
1277 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1278 * pcpu_create_chunk - create a new chunk
1279 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1280 * pcpu_addr_to_page - translate address to physical address
1281 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1283 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
,
1285 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
1286 static struct pcpu_chunk
*pcpu_create_chunk(gfp_t gfp
);
1287 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
1288 static struct page
*pcpu_addr_to_page(void *addr
);
1289 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
1291 #ifdef CONFIG_NEED_PER_CPU_KM
1292 #include "percpu-km.c"
1294 #include "percpu-vm.c"
1298 * pcpu_chunk_addr_search - determine chunk containing specified address
1299 * @addr: address for which the chunk needs to be determined.
1301 * This is an internal function that handles all but static allocations.
1302 * Static percpu address values should never be passed into the allocator.
1305 * The address of the found chunk.
1307 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
1309 /* is it in the dynamic region (first chunk)? */
1310 if (pcpu_addr_in_chunk(pcpu_first_chunk
, addr
))
1311 return pcpu_first_chunk
;
1313 /* is it in the reserved region? */
1314 if (pcpu_addr_in_chunk(pcpu_reserved_chunk
, addr
))
1315 return pcpu_reserved_chunk
;
1318 * The address is relative to unit0 which might be unused and
1319 * thus unmapped. Offset the address to the unit space of the
1320 * current processor before looking it up in the vmalloc
1321 * space. Note that any possible cpu id can be used here, so
1322 * there's no need to worry about preemption or cpu hotplug.
1324 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
1325 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
1329 * pcpu_alloc - the percpu allocator
1330 * @size: size of area to allocate in bytes
1331 * @align: alignment of area (max PAGE_SIZE)
1332 * @reserved: allocate from the reserved chunk if available
1333 * @gfp: allocation flags
1335 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1336 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1337 * then no warning will be triggered on invalid or failed allocation
1341 * Percpu pointer to the allocated area on success, NULL on failure.
1343 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
,
1346 bool is_atomic
= (gfp
& GFP_KERNEL
) != GFP_KERNEL
;
1347 bool do_warn
= !(gfp
& __GFP_NOWARN
);
1348 static int warn_limit
= 10;
1349 struct pcpu_chunk
*chunk
;
1351 int slot
, off
, cpu
, ret
;
1352 unsigned long flags
;
1354 size_t bits
, bit_align
;
1357 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1358 * therefore alignment must be a minimum of that many bytes.
1359 * An allocation may have internal fragmentation from rounding up
1360 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1362 if (unlikely(align
< PCPU_MIN_ALLOC_SIZE
))
1363 align
= PCPU_MIN_ALLOC_SIZE
;
1365 size
= ALIGN(size
, PCPU_MIN_ALLOC_SIZE
);
1366 bits
= size
>> PCPU_MIN_ALLOC_SHIFT
;
1367 bit_align
= align
>> PCPU_MIN_ALLOC_SHIFT
;
1369 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
||
1370 !is_power_of_2(align
))) {
1371 WARN(do_warn
, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1377 mutex_lock(&pcpu_alloc_mutex
);
1379 spin_lock_irqsave(&pcpu_lock
, flags
);
1381 /* serve reserved allocations from the reserved chunk if available */
1382 if (reserved
&& pcpu_reserved_chunk
) {
1383 chunk
= pcpu_reserved_chunk
;
1385 off
= pcpu_find_block_fit(chunk
, bits
, bit_align
, is_atomic
);
1387 err
= "alloc from reserved chunk failed";
1391 off
= pcpu_alloc_area(chunk
, bits
, bit_align
, off
);
1395 err
= "alloc from reserved chunk failed";
1400 /* search through normal chunks */
1401 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
1402 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1403 off
= pcpu_find_block_fit(chunk
, bits
, bit_align
,
1408 off
= pcpu_alloc_area(chunk
, bits
, bit_align
, off
);
1415 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1418 * No space left. Create a new chunk. We don't want multiple
1419 * tasks to create chunks simultaneously. Serialize and create iff
1420 * there's still no empty chunk after grabbing the mutex.
1423 err
= "atomic alloc failed, no space left";
1427 if (list_empty(&pcpu_slot
[pcpu_nr_slots
- 1])) {
1428 chunk
= pcpu_create_chunk(0);
1430 err
= "failed to allocate new chunk";
1434 spin_lock_irqsave(&pcpu_lock
, flags
);
1435 pcpu_chunk_relocate(chunk
, -1);
1437 spin_lock_irqsave(&pcpu_lock
, flags
);
1443 pcpu_stats_area_alloc(chunk
, size
);
1444 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1446 /* populate if not all pages are already there */
1448 int page_start
, page_end
, rs
, re
;
1450 page_start
= PFN_DOWN(off
);
1451 page_end
= PFN_UP(off
+ size
);
1453 pcpu_for_each_unpop_region(chunk
->populated
, rs
, re
,
1454 page_start
, page_end
) {
1455 WARN_ON(chunk
->immutable
);
1457 ret
= pcpu_populate_chunk(chunk
, rs
, re
, 0);
1459 spin_lock_irqsave(&pcpu_lock
, flags
);
1461 pcpu_free_area(chunk
, off
);
1462 err
= "failed to populate";
1465 pcpu_chunk_populated(chunk
, rs
, re
, true);
1466 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1469 mutex_unlock(&pcpu_alloc_mutex
);
1472 if (pcpu_nr_empty_pop_pages
< PCPU_EMPTY_POP_PAGES_LOW
)
1473 pcpu_schedule_balance_work();
1475 /* clear the areas and return address relative to base address */
1476 for_each_possible_cpu(cpu
)
1477 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1479 ptr
= __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1480 kmemleak_alloc_percpu(ptr
, size
, gfp
);
1482 trace_percpu_alloc_percpu(reserved
, is_atomic
, size
, align
,
1483 chunk
->base_addr
, off
, ptr
);
1488 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1490 trace_percpu_alloc_percpu_fail(reserved
, is_atomic
, size
, align
);
1492 if (!is_atomic
&& do_warn
&& warn_limit
) {
1493 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1494 size
, align
, is_atomic
, err
);
1497 pr_info("limit reached, disable warning\n");
1500 /* see the flag handling in pcpu_blance_workfn() */
1501 pcpu_atomic_alloc_failed
= true;
1502 pcpu_schedule_balance_work();
1504 mutex_unlock(&pcpu_alloc_mutex
);
1510 * __alloc_percpu_gfp - allocate dynamic percpu area
1511 * @size: size of area to allocate in bytes
1512 * @align: alignment of area (max PAGE_SIZE)
1513 * @gfp: allocation flags
1515 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1516 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1517 * be called from any context but is a lot more likely to fail. If @gfp
1518 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1519 * allocation requests.
1522 * Percpu pointer to the allocated area on success, NULL on failure.
1524 void __percpu
*__alloc_percpu_gfp(size_t size
, size_t align
, gfp_t gfp
)
1526 return pcpu_alloc(size
, align
, false, gfp
);
1528 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp
);
1531 * __alloc_percpu - allocate dynamic percpu area
1532 * @size: size of area to allocate in bytes
1533 * @align: alignment of area (max PAGE_SIZE)
1535 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1537 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
1539 return pcpu_alloc(size
, align
, false, GFP_KERNEL
);
1541 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1544 * __alloc_reserved_percpu - allocate reserved percpu area
1545 * @size: size of area to allocate in bytes
1546 * @align: alignment of area (max PAGE_SIZE)
1548 * Allocate zero-filled percpu area of @size bytes aligned at @align
1549 * from reserved percpu area if arch has set it up; otherwise,
1550 * allocation is served from the same dynamic area. Might sleep.
1551 * Might trigger writeouts.
1554 * Does GFP_KERNEL allocation.
1557 * Percpu pointer to the allocated area on success, NULL on failure.
1559 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
1561 return pcpu_alloc(size
, align
, true, GFP_KERNEL
);
1565 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1568 * Reclaim all fully free chunks except for the first one. This is also
1569 * responsible for maintaining the pool of empty populated pages. However,
1570 * it is possible that this is called when physical memory is scarce causing
1571 * OOM killer to be triggered. We should avoid doing so until an actual
1572 * allocation causes the failure as it is possible that requests can be
1573 * serviced from already backed regions.
1575 static void pcpu_balance_workfn(struct work_struct
*work
)
1577 /* gfp flags passed to underlying allocators */
1578 const gfp_t gfp
= __GFP_NORETRY
| __GFP_NOWARN
;
1580 struct list_head
*free_head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1581 struct pcpu_chunk
*chunk
, *next
;
1582 int slot
, nr_to_pop
, ret
;
1585 * There's no reason to keep around multiple unused chunks and VM
1586 * areas can be scarce. Destroy all free chunks except for one.
1588 mutex_lock(&pcpu_alloc_mutex
);
1589 spin_lock_irq(&pcpu_lock
);
1591 list_for_each_entry_safe(chunk
, next
, free_head
, list
) {
1592 WARN_ON(chunk
->immutable
);
1594 /* spare the first one */
1595 if (chunk
== list_first_entry(free_head
, struct pcpu_chunk
, list
))
1598 list_move(&chunk
->list
, &to_free
);
1601 spin_unlock_irq(&pcpu_lock
);
1603 list_for_each_entry_safe(chunk
, next
, &to_free
, list
) {
1606 pcpu_for_each_pop_region(chunk
->populated
, rs
, re
, 0,
1608 pcpu_depopulate_chunk(chunk
, rs
, re
);
1609 spin_lock_irq(&pcpu_lock
);
1610 pcpu_chunk_depopulated(chunk
, rs
, re
);
1611 spin_unlock_irq(&pcpu_lock
);
1613 pcpu_destroy_chunk(chunk
);
1617 * Ensure there are certain number of free populated pages for
1618 * atomic allocs. Fill up from the most packed so that atomic
1619 * allocs don't increase fragmentation. If atomic allocation
1620 * failed previously, always populate the maximum amount. This
1621 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1622 * failing indefinitely; however, large atomic allocs are not
1623 * something we support properly and can be highly unreliable and
1627 if (pcpu_atomic_alloc_failed
) {
1628 nr_to_pop
= PCPU_EMPTY_POP_PAGES_HIGH
;
1629 /* best effort anyway, don't worry about synchronization */
1630 pcpu_atomic_alloc_failed
= false;
1632 nr_to_pop
= clamp(PCPU_EMPTY_POP_PAGES_HIGH
-
1633 pcpu_nr_empty_pop_pages
,
1634 0, PCPU_EMPTY_POP_PAGES_HIGH
);
1637 for (slot
= pcpu_size_to_slot(PAGE_SIZE
); slot
< pcpu_nr_slots
; slot
++) {
1638 int nr_unpop
= 0, rs
, re
;
1643 spin_lock_irq(&pcpu_lock
);
1644 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1645 nr_unpop
= chunk
->nr_pages
- chunk
->nr_populated
;
1649 spin_unlock_irq(&pcpu_lock
);
1654 /* @chunk can't go away while pcpu_alloc_mutex is held */
1655 pcpu_for_each_unpop_region(chunk
->populated
, rs
, re
, 0,
1657 int nr
= min(re
- rs
, nr_to_pop
);
1659 ret
= pcpu_populate_chunk(chunk
, rs
, rs
+ nr
, gfp
);
1662 spin_lock_irq(&pcpu_lock
);
1663 pcpu_chunk_populated(chunk
, rs
, rs
+ nr
, false);
1664 spin_unlock_irq(&pcpu_lock
);
1675 /* ran out of chunks to populate, create a new one and retry */
1676 chunk
= pcpu_create_chunk(gfp
);
1678 spin_lock_irq(&pcpu_lock
);
1679 pcpu_chunk_relocate(chunk
, -1);
1680 spin_unlock_irq(&pcpu_lock
);
1685 mutex_unlock(&pcpu_alloc_mutex
);
1689 * free_percpu - free percpu area
1690 * @ptr: pointer to area to free
1692 * Free percpu area @ptr.
1695 * Can be called from atomic context.
1697 void free_percpu(void __percpu
*ptr
)
1700 struct pcpu_chunk
*chunk
;
1701 unsigned long flags
;
1707 kmemleak_free_percpu(ptr
);
1709 addr
= __pcpu_ptr_to_addr(ptr
);
1711 spin_lock_irqsave(&pcpu_lock
, flags
);
1713 chunk
= pcpu_chunk_addr_search(addr
);
1714 off
= addr
- chunk
->base_addr
;
1716 pcpu_free_area(chunk
, off
);
1718 /* if there are more than one fully free chunks, wake up grim reaper */
1719 if (chunk
->free_bytes
== pcpu_unit_size
) {
1720 struct pcpu_chunk
*pos
;
1722 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1724 pcpu_schedule_balance_work();
1729 trace_percpu_free_percpu(chunk
->base_addr
, off
, ptr
);
1731 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1733 EXPORT_SYMBOL_GPL(free_percpu
);
1735 bool __is_kernel_percpu_address(unsigned long addr
, unsigned long *can_addr
)
1738 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1739 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1742 for_each_possible_cpu(cpu
) {
1743 void *start
= per_cpu_ptr(base
, cpu
);
1744 void *va
= (void *)addr
;
1746 if (va
>= start
&& va
< start
+ static_size
) {
1748 *can_addr
= (unsigned long) (va
- start
);
1749 *can_addr
+= (unsigned long)
1750 per_cpu_ptr(base
, get_boot_cpu_id());
1756 /* on UP, can't distinguish from other static vars, always false */
1761 * is_kernel_percpu_address - test whether address is from static percpu area
1762 * @addr: address to test
1764 * Test whether @addr belongs to in-kernel static percpu area. Module
1765 * static percpu areas are not considered. For those, use
1766 * is_module_percpu_address().
1769 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1771 bool is_kernel_percpu_address(unsigned long addr
)
1773 return __is_kernel_percpu_address(addr
, NULL
);
1777 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1778 * @addr: the address to be converted to physical address
1780 * Given @addr which is dereferenceable address obtained via one of
1781 * percpu access macros, this function translates it into its physical
1782 * address. The caller is responsible for ensuring @addr stays valid
1783 * until this function finishes.
1785 * percpu allocator has special setup for the first chunk, which currently
1786 * supports either embedding in linear address space or vmalloc mapping,
1787 * and, from the second one, the backing allocator (currently either vm or
1788 * km) provides translation.
1790 * The addr can be translated simply without checking if it falls into the
1791 * first chunk. But the current code reflects better how percpu allocator
1792 * actually works, and the verification can discover both bugs in percpu
1793 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1797 * The physical address for @addr.
1799 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
1801 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1802 bool in_first_chunk
= false;
1803 unsigned long first_low
, first_high
;
1807 * The following test on unit_low/high isn't strictly
1808 * necessary but will speed up lookups of addresses which
1809 * aren't in the first chunk.
1811 * The address check is against full chunk sizes. pcpu_base_addr
1812 * points to the beginning of the first chunk including the
1813 * static region. Assumes good intent as the first chunk may
1814 * not be full (ie. < pcpu_unit_pages in size).
1816 first_low
= (unsigned long)pcpu_base_addr
+
1817 pcpu_unit_page_offset(pcpu_low_unit_cpu
, 0);
1818 first_high
= (unsigned long)pcpu_base_addr
+
1819 pcpu_unit_page_offset(pcpu_high_unit_cpu
, pcpu_unit_pages
);
1820 if ((unsigned long)addr
>= first_low
&&
1821 (unsigned long)addr
< first_high
) {
1822 for_each_possible_cpu(cpu
) {
1823 void *start
= per_cpu_ptr(base
, cpu
);
1825 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
1826 in_first_chunk
= true;
1832 if (in_first_chunk
) {
1833 if (!is_vmalloc_addr(addr
))
1836 return page_to_phys(vmalloc_to_page(addr
)) +
1837 offset_in_page(addr
);
1839 return page_to_phys(pcpu_addr_to_page(addr
)) +
1840 offset_in_page(addr
);
1844 * pcpu_alloc_alloc_info - allocate percpu allocation info
1845 * @nr_groups: the number of groups
1846 * @nr_units: the number of units
1848 * Allocate ai which is large enough for @nr_groups groups containing
1849 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1850 * cpu_map array which is long enough for @nr_units and filled with
1851 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1852 * pointer of other groups.
1855 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1858 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1861 struct pcpu_alloc_info
*ai
;
1862 size_t base_size
, ai_size
;
1866 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1867 __alignof__(ai
->groups
[0].cpu_map
[0]));
1868 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1870 ptr
= memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size
), PAGE_SIZE
);
1876 ai
->groups
[0].cpu_map
= ptr
;
1878 for (unit
= 0; unit
< nr_units
; unit
++)
1879 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1881 ai
->nr_groups
= nr_groups
;
1882 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1888 * pcpu_free_alloc_info - free percpu allocation info
1889 * @ai: pcpu_alloc_info to free
1891 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1893 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1895 memblock_free_early(__pa(ai
), ai
->__ai_size
);
1899 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1901 * @ai: allocation info to dump
1903 * Print out information about @ai using loglevel @lvl.
1905 static void pcpu_dump_alloc_info(const char *lvl
,
1906 const struct pcpu_alloc_info
*ai
)
1908 int group_width
= 1, cpu_width
= 1, width
;
1909 char empty_str
[] = "--------";
1910 int alloc
= 0, alloc_end
= 0;
1912 int upa
, apl
; /* units per alloc, allocs per line */
1918 v
= num_possible_cpus();
1921 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1923 upa
= ai
->alloc_size
/ ai
->unit_size
;
1924 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1925 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1927 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1928 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1929 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1931 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1932 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1933 int unit
= 0, unit_end
= 0;
1935 BUG_ON(gi
->nr_units
% upa
);
1936 for (alloc_end
+= gi
->nr_units
/ upa
;
1937 alloc
< alloc_end
; alloc
++) {
1938 if (!(alloc
% apl
)) {
1940 printk("%spcpu-alloc: ", lvl
);
1942 pr_cont("[%0*d] ", group_width
, group
);
1944 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1945 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1947 cpu_width
, gi
->cpu_map
[unit
]);
1949 pr_cont("%s ", empty_str
);
1956 * pcpu_setup_first_chunk - initialize the first percpu chunk
1957 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1958 * @base_addr: mapped address
1960 * Initialize the first percpu chunk which contains the kernel static
1961 * perpcu area. This function is to be called from arch percpu area
1964 * @ai contains all information necessary to initialize the first
1965 * chunk and prime the dynamic percpu allocator.
1967 * @ai->static_size is the size of static percpu area.
1969 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1970 * reserve after the static area in the first chunk. This reserves
1971 * the first chunk such that it's available only through reserved
1972 * percpu allocation. This is primarily used to serve module percpu
1973 * static areas on architectures where the addressing model has
1974 * limited offset range for symbol relocations to guarantee module
1975 * percpu symbols fall inside the relocatable range.
1977 * @ai->dyn_size determines the number of bytes available for dynamic
1978 * allocation in the first chunk. The area between @ai->static_size +
1979 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1981 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1982 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1985 * @ai->atom_size is the allocation atom size and used as alignment
1988 * @ai->alloc_size is the allocation size and always multiple of
1989 * @ai->atom_size. This is larger than @ai->atom_size if
1990 * @ai->unit_size is larger than @ai->atom_size.
1992 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1993 * percpu areas. Units which should be colocated are put into the
1994 * same group. Dynamic VM areas will be allocated according to these
1995 * groupings. If @ai->nr_groups is zero, a single group containing
1996 * all units is assumed.
1998 * The caller should have mapped the first chunk at @base_addr and
1999 * copied static data to each unit.
2001 * The first chunk will always contain a static and a dynamic region.
2002 * However, the static region is not managed by any chunk. If the first
2003 * chunk also contains a reserved region, it is served by two chunks -
2004 * one for the reserved region and one for the dynamic region. They
2005 * share the same vm, but use offset regions in the area allocation map.
2006 * The chunk serving the dynamic region is circulated in the chunk slots
2007 * and available for dynamic allocation like any other chunk.
2010 * 0 on success, -errno on failure.
2012 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
2015 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
2016 size_t static_size
, dyn_size
;
2017 struct pcpu_chunk
*chunk
;
2018 unsigned long *group_offsets
;
2019 size_t *group_sizes
;
2020 unsigned long *unit_off
;
2025 unsigned long tmp_addr
;
2027 #define PCPU_SETUP_BUG_ON(cond) do { \
2028 if (unlikely(cond)) { \
2029 pr_emerg("failed to initialize, %s\n", #cond); \
2030 pr_emerg("cpu_possible_mask=%*pb\n", \
2031 cpumask_pr_args(cpu_possible_mask)); \
2032 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2038 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
2040 PCPU_SETUP_BUG_ON(!ai
->static_size
);
2041 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start
));
2043 PCPU_SETUP_BUG_ON(!base_addr
);
2044 PCPU_SETUP_BUG_ON(offset_in_page(base_addr
));
2045 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
2046 PCPU_SETUP_BUG_ON(offset_in_page(ai
->unit_size
));
2047 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
2048 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai
->unit_size
, PCPU_BITMAP_BLOCK_SIZE
));
2049 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
2050 PCPU_SETUP_BUG_ON(!ai
->dyn_size
);
2051 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai
->reserved_size
, PCPU_MIN_ALLOC_SIZE
));
2052 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE
, PAGE_SIZE
) ||
2053 IS_ALIGNED(PAGE_SIZE
, PCPU_BITMAP_BLOCK_SIZE
)));
2054 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
2056 /* process group information and build config tables accordingly */
2057 group_offsets
= memblock_virt_alloc(ai
->nr_groups
*
2058 sizeof(group_offsets
[0]), 0);
2059 group_sizes
= memblock_virt_alloc(ai
->nr_groups
*
2060 sizeof(group_sizes
[0]), 0);
2061 unit_map
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_map
[0]), 0);
2062 unit_off
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_off
[0]), 0);
2064 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
2065 unit_map
[cpu
] = UINT_MAX
;
2067 pcpu_low_unit_cpu
= NR_CPUS
;
2068 pcpu_high_unit_cpu
= NR_CPUS
;
2070 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
2071 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2073 group_offsets
[group
] = gi
->base_offset
;
2074 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
2076 for (i
= 0; i
< gi
->nr_units
; i
++) {
2077 cpu
= gi
->cpu_map
[i
];
2081 PCPU_SETUP_BUG_ON(cpu
>= nr_cpu_ids
);
2082 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
2083 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
2085 unit_map
[cpu
] = unit
+ i
;
2086 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
2088 /* determine low/high unit_cpu */
2089 if (pcpu_low_unit_cpu
== NR_CPUS
||
2090 unit_off
[cpu
] < unit_off
[pcpu_low_unit_cpu
])
2091 pcpu_low_unit_cpu
= cpu
;
2092 if (pcpu_high_unit_cpu
== NR_CPUS
||
2093 unit_off
[cpu
] > unit_off
[pcpu_high_unit_cpu
])
2094 pcpu_high_unit_cpu
= cpu
;
2097 pcpu_nr_units
= unit
;
2099 for_each_possible_cpu(cpu
)
2100 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
2102 /* we're done parsing the input, undefine BUG macro and dump config */
2103 #undef PCPU_SETUP_BUG_ON
2104 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
2106 pcpu_nr_groups
= ai
->nr_groups
;
2107 pcpu_group_offsets
= group_offsets
;
2108 pcpu_group_sizes
= group_sizes
;
2109 pcpu_unit_map
= unit_map
;
2110 pcpu_unit_offsets
= unit_off
;
2112 /* determine basic parameters */
2113 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2114 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
2115 pcpu_atom_size
= ai
->atom_size
;
2116 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
2117 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
2119 pcpu_stats_save_ai(ai
);
2122 * Allocate chunk slots. The additional last slot is for
2125 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
2126 pcpu_slot
= memblock_virt_alloc(
2127 pcpu_nr_slots
* sizeof(pcpu_slot
[0]), 0);
2128 for (i
= 0; i
< pcpu_nr_slots
; i
++)
2129 INIT_LIST_HEAD(&pcpu_slot
[i
]);
2132 * The end of the static region needs to be aligned with the
2133 * minimum allocation size as this offsets the reserved and
2134 * dynamic region. The first chunk ends page aligned by
2135 * expanding the dynamic region, therefore the dynamic region
2136 * can be shrunk to compensate while still staying above the
2139 static_size
= ALIGN(ai
->static_size
, PCPU_MIN_ALLOC_SIZE
);
2140 dyn_size
= ai
->dyn_size
- (static_size
- ai
->static_size
);
2143 * Initialize first chunk.
2144 * If the reserved_size is non-zero, this initializes the reserved
2145 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2146 * and the dynamic region is initialized here. The first chunk,
2147 * pcpu_first_chunk, will always point to the chunk that serves
2148 * the dynamic region.
2150 tmp_addr
= (unsigned long)base_addr
+ static_size
;
2151 map_size
= ai
->reserved_size
?: dyn_size
;
2152 chunk
= pcpu_alloc_first_chunk(tmp_addr
, map_size
);
2154 /* init dynamic chunk if necessary */
2155 if (ai
->reserved_size
) {
2156 pcpu_reserved_chunk
= chunk
;
2158 tmp_addr
= (unsigned long)base_addr
+ static_size
+
2160 map_size
= dyn_size
;
2161 chunk
= pcpu_alloc_first_chunk(tmp_addr
, map_size
);
2164 /* link the first chunk in */
2165 pcpu_first_chunk
= chunk
;
2166 pcpu_nr_empty_pop_pages
= pcpu_first_chunk
->nr_empty_pop_pages
;
2167 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
2169 pcpu_stats_chunk_alloc();
2170 trace_percpu_create_chunk(base_addr
);
2173 pcpu_base_addr
= base_addr
;
2179 const char * const pcpu_fc_names
[PCPU_FC_NR
] __initconst
= {
2180 [PCPU_FC_AUTO
] = "auto",
2181 [PCPU_FC_EMBED
] = "embed",
2182 [PCPU_FC_PAGE
] = "page",
2185 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
2187 static int __init
percpu_alloc_setup(char *str
)
2194 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2195 else if (!strcmp(str
, "embed"))
2196 pcpu_chosen_fc
= PCPU_FC_EMBED
;
2198 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2199 else if (!strcmp(str
, "page"))
2200 pcpu_chosen_fc
= PCPU_FC_PAGE
;
2203 pr_warn("unknown allocator %s specified\n", str
);
2207 early_param("percpu_alloc", percpu_alloc_setup
);
2210 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2211 * Build it if needed by the arch config or the generic setup is going
2214 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2215 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2216 #define BUILD_EMBED_FIRST_CHUNK
2219 /* build pcpu_page_first_chunk() iff needed by the arch config */
2220 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2221 #define BUILD_PAGE_FIRST_CHUNK
2224 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2225 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2227 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2228 * @reserved_size: the size of reserved percpu area in bytes
2229 * @dyn_size: minimum free size for dynamic allocation in bytes
2230 * @atom_size: allocation atom size
2231 * @cpu_distance_fn: callback to determine distance between cpus, optional
2233 * This function determines grouping of units, their mappings to cpus
2234 * and other parameters considering needed percpu size, allocation
2235 * atom size and distances between CPUs.
2237 * Groups are always multiples of atom size and CPUs which are of
2238 * LOCAL_DISTANCE both ways are grouped together and share space for
2239 * units in the same group. The returned configuration is guaranteed
2240 * to have CPUs on different nodes on different groups and >=75% usage
2241 * of allocated virtual address space.
2244 * On success, pointer to the new allocation_info is returned. On
2245 * failure, ERR_PTR value is returned.
2247 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
2248 size_t reserved_size
, size_t dyn_size
,
2250 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
2252 static int group_map
[NR_CPUS
] __initdata
;
2253 static int group_cnt
[NR_CPUS
] __initdata
;
2254 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
2255 int nr_groups
= 1, nr_units
= 0;
2256 size_t size_sum
, min_unit_size
, alloc_size
;
2257 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
2258 int last_allocs
, group
, unit
;
2259 unsigned int cpu
, tcpu
;
2260 struct pcpu_alloc_info
*ai
;
2261 unsigned int *cpu_map
;
2263 /* this function may be called multiple times */
2264 memset(group_map
, 0, sizeof(group_map
));
2265 memset(group_cnt
, 0, sizeof(group_cnt
));
2267 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2268 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
2269 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
2270 dyn_size
= size_sum
- static_size
- reserved_size
;
2273 * Determine min_unit_size, alloc_size and max_upa such that
2274 * alloc_size is multiple of atom_size and is the smallest
2275 * which can accommodate 4k aligned segments which are equal to
2276 * or larger than min_unit_size.
2278 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
2280 /* determine the maximum # of units that can fit in an allocation */
2281 alloc_size
= roundup(min_unit_size
, atom_size
);
2282 upa
= alloc_size
/ min_unit_size
;
2283 while (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
2287 /* group cpus according to their proximity */
2288 for_each_possible_cpu(cpu
) {
2291 for_each_possible_cpu(tcpu
) {
2294 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
2295 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
2296 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
2298 nr_groups
= max(nr_groups
, group
+ 1);
2302 group_map
[cpu
] = group
;
2307 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2308 * Expand the unit_size until we use >= 75% of the units allocated.
2309 * Related to atom_size, which could be much larger than the unit_size.
2311 last_allocs
= INT_MAX
;
2312 for (upa
= max_upa
; upa
; upa
--) {
2313 int allocs
= 0, wasted
= 0;
2315 if (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
2318 for (group
= 0; group
< nr_groups
; group
++) {
2319 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
2320 allocs
+= this_allocs
;
2321 wasted
+= this_allocs
* upa
- group_cnt
[group
];
2325 * Don't accept if wastage is over 1/3. The
2326 * greater-than comparison ensures upa==1 always
2327 * passes the following check.
2329 if (wasted
> num_possible_cpus() / 3)
2332 /* and then don't consume more memory */
2333 if (allocs
> last_allocs
)
2335 last_allocs
= allocs
;
2340 /* allocate and fill alloc_info */
2341 for (group
= 0; group
< nr_groups
; group
++)
2342 nr_units
+= roundup(group_cnt
[group
], upa
);
2344 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
2346 return ERR_PTR(-ENOMEM
);
2347 cpu_map
= ai
->groups
[0].cpu_map
;
2349 for (group
= 0; group
< nr_groups
; group
++) {
2350 ai
->groups
[group
].cpu_map
= cpu_map
;
2351 cpu_map
+= roundup(group_cnt
[group
], upa
);
2354 ai
->static_size
= static_size
;
2355 ai
->reserved_size
= reserved_size
;
2356 ai
->dyn_size
= dyn_size
;
2357 ai
->unit_size
= alloc_size
/ upa
;
2358 ai
->atom_size
= atom_size
;
2359 ai
->alloc_size
= alloc_size
;
2361 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
2362 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2365 * Initialize base_offset as if all groups are located
2366 * back-to-back. The caller should update this to
2367 * reflect actual allocation.
2369 gi
->base_offset
= unit
* ai
->unit_size
;
2371 for_each_possible_cpu(cpu
)
2372 if (group_map
[cpu
] == group
)
2373 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
2374 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
2375 unit
+= gi
->nr_units
;
2377 BUG_ON(unit
!= nr_units
);
2381 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2383 #if defined(BUILD_EMBED_FIRST_CHUNK)
2385 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2386 * @reserved_size: the size of reserved percpu area in bytes
2387 * @dyn_size: minimum free size for dynamic allocation in bytes
2388 * @atom_size: allocation atom size
2389 * @cpu_distance_fn: callback to determine distance between cpus, optional
2390 * @alloc_fn: function to allocate percpu page
2391 * @free_fn: function to free percpu page
2393 * This is a helper to ease setting up embedded first percpu chunk and
2394 * can be called where pcpu_setup_first_chunk() is expected.
2396 * If this function is used to setup the first chunk, it is allocated
2397 * by calling @alloc_fn and used as-is without being mapped into
2398 * vmalloc area. Allocations are always whole multiples of @atom_size
2399 * aligned to @atom_size.
2401 * This enables the first chunk to piggy back on the linear physical
2402 * mapping which often uses larger page size. Please note that this
2403 * can result in very sparse cpu->unit mapping on NUMA machines thus
2404 * requiring large vmalloc address space. Don't use this allocator if
2405 * vmalloc space is not orders of magnitude larger than distances
2406 * between node memory addresses (ie. 32bit NUMA machines).
2408 * @dyn_size specifies the minimum dynamic area size.
2410 * If the needed size is smaller than the minimum or specified unit
2411 * size, the leftover is returned using @free_fn.
2414 * 0 on success, -errno on failure.
2416 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
2418 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
2419 pcpu_fc_alloc_fn_t alloc_fn
,
2420 pcpu_fc_free_fn_t free_fn
)
2422 void *base
= (void *)ULONG_MAX
;
2423 void **areas
= NULL
;
2424 struct pcpu_alloc_info
*ai
;
2425 size_t size_sum
, areas_size
;
2426 unsigned long max_distance
;
2427 int group
, i
, highest_group
, rc
;
2429 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
2434 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
2435 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
2437 areas
= memblock_virt_alloc_nopanic(areas_size
, 0);
2443 /* allocate, copy and determine base address & max_distance */
2445 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2446 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2447 unsigned int cpu
= NR_CPUS
;
2450 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
2451 cpu
= gi
->cpu_map
[i
];
2452 BUG_ON(cpu
== NR_CPUS
);
2454 /* allocate space for the whole group */
2455 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
2458 goto out_free_areas
;
2460 /* kmemleak tracks the percpu allocations separately */
2464 base
= min(ptr
, base
);
2465 if (ptr
> areas
[highest_group
])
2466 highest_group
= group
;
2468 max_distance
= areas
[highest_group
] - base
;
2469 max_distance
+= ai
->unit_size
* ai
->groups
[highest_group
].nr_units
;
2471 /* warn if maximum distance is further than 75% of vmalloc space */
2472 if (max_distance
> VMALLOC_TOTAL
* 3 / 4) {
2473 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2474 max_distance
, VMALLOC_TOTAL
);
2475 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2476 /* and fail if we have fallback */
2478 goto out_free_areas
;
2483 * Copy data and free unused parts. This should happen after all
2484 * allocations are complete; otherwise, we may end up with
2485 * overlapping groups.
2487 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2488 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2489 void *ptr
= areas
[group
];
2491 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
2492 if (gi
->cpu_map
[i
] == NR_CPUS
) {
2493 /* unused unit, free whole */
2494 free_fn(ptr
, ai
->unit_size
);
2497 /* copy and return the unused part */
2498 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
2499 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2503 /* base address is now known, determine group base offsets */
2504 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2505 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2508 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2509 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
2510 ai
->dyn_size
, ai
->unit_size
);
2512 rc
= pcpu_setup_first_chunk(ai
, base
);
2516 for (group
= 0; group
< ai
->nr_groups
; group
++)
2518 free_fn(areas
[group
],
2519 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2521 pcpu_free_alloc_info(ai
);
2523 memblock_free_early(__pa(areas
), areas_size
);
2526 #endif /* BUILD_EMBED_FIRST_CHUNK */
2528 #ifdef BUILD_PAGE_FIRST_CHUNK
2530 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2531 * @reserved_size: the size of reserved percpu area in bytes
2532 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2533 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2534 * @populate_pte_fn: function to populate pte
2536 * This is a helper to ease setting up page-remapped first percpu
2537 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2539 * This is the basic allocator. Static percpu area is allocated
2540 * page-by-page into vmalloc area.
2543 * 0 on success, -errno on failure.
2545 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2546 pcpu_fc_alloc_fn_t alloc_fn
,
2547 pcpu_fc_free_fn_t free_fn
,
2548 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2550 static struct vm_struct vm
;
2551 struct pcpu_alloc_info
*ai
;
2555 struct page
**pages
;
2560 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2562 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
2565 BUG_ON(ai
->nr_groups
!= 1);
2566 upa
= ai
->alloc_size
/ai
->unit_size
;
2567 nr_g0_units
= roundup(num_possible_cpus(), upa
);
2568 if (unlikely(WARN_ON(ai
->groups
[0].nr_units
!= nr_g0_units
))) {
2569 pcpu_free_alloc_info(ai
);
2573 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2575 /* unaligned allocations can't be freed, round up to page size */
2576 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2578 pages
= memblock_virt_alloc(pages_size
, 0);
2580 /* allocate pages */
2582 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2583 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2584 for (i
= 0; i
< unit_pages
; i
++) {
2587 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2589 pr_warn("failed to allocate %s page for cpu%u\n",
2593 /* kmemleak tracks the percpu allocations separately */
2595 pages
[j
++] = virt_to_page(ptr
);
2599 /* allocate vm area, map the pages and copy static data */
2600 vm
.flags
= VM_ALLOC
;
2601 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2602 vm_area_register_early(&vm
, PAGE_SIZE
);
2604 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2605 unsigned long unit_addr
=
2606 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2608 for (i
= 0; i
< unit_pages
; i
++)
2609 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2611 /* pte already populated, the following shouldn't fail */
2612 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2615 panic("failed to map percpu area, err=%d\n", rc
);
2618 * FIXME: Archs with virtual cache should flush local
2619 * cache for the linear mapping here - something
2620 * equivalent to flush_cache_vmap() on the local cpu.
2621 * flush_cache_vmap() can't be used as most supporting
2622 * data structures are not set up yet.
2625 /* copy static data */
2626 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2629 /* we're ready, commit */
2630 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2631 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2632 ai
->reserved_size
, ai
->dyn_size
);
2634 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2639 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2642 memblock_free_early(__pa(pages
), pages_size
);
2643 pcpu_free_alloc_info(ai
);
2646 #endif /* BUILD_PAGE_FIRST_CHUNK */
2648 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2650 * Generic SMP percpu area setup.
2652 * The embedding helper is used because its behavior closely resembles
2653 * the original non-dynamic generic percpu area setup. This is
2654 * important because many archs have addressing restrictions and might
2655 * fail if the percpu area is located far away from the previous
2656 * location. As an added bonus, in non-NUMA cases, embedding is
2657 * generally a good idea TLB-wise because percpu area can piggy back
2658 * on the physical linear memory mapping which uses large page
2659 * mappings on applicable archs.
2661 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2662 EXPORT_SYMBOL(__per_cpu_offset
);
2664 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2667 return memblock_virt_alloc_from_nopanic(
2668 size
, align
, __pa(MAX_DMA_ADDRESS
));
2671 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2673 memblock_free_early(__pa(ptr
), size
);
2676 void __init
setup_per_cpu_areas(void)
2678 unsigned long delta
;
2683 * Always reserve area for module percpu variables. That's
2684 * what the legacy allocator did.
2686 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2687 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2688 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2690 panic("Failed to initialize percpu areas.");
2692 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2693 for_each_possible_cpu(cpu
)
2694 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2696 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2698 #else /* CONFIG_SMP */
2701 * UP percpu area setup.
2703 * UP always uses km-based percpu allocator with identity mapping.
2704 * Static percpu variables are indistinguishable from the usual static
2705 * variables and don't require any special preparation.
2707 void __init
setup_per_cpu_areas(void)
2709 const size_t unit_size
=
2710 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
2711 PERCPU_DYNAMIC_RESERVE
));
2712 struct pcpu_alloc_info
*ai
;
2715 ai
= pcpu_alloc_alloc_info(1, 1);
2716 fc
= memblock_virt_alloc_from_nopanic(unit_size
,
2718 __pa(MAX_DMA_ADDRESS
));
2720 panic("Failed to allocate memory for percpu areas.");
2721 /* kmemleak tracks the percpu allocations separately */
2724 ai
->dyn_size
= unit_size
;
2725 ai
->unit_size
= unit_size
;
2726 ai
->atom_size
= unit_size
;
2727 ai
->alloc_size
= unit_size
;
2728 ai
->groups
[0].nr_units
= 1;
2729 ai
->groups
[0].cpu_map
[0] = 0;
2731 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
2732 panic("Failed to initialize percpu areas.");
2734 #warning "the CRIS architecture has physical and virtual addresses confused"
2736 pcpu_free_alloc_info(ai
);
2740 #endif /* CONFIG_SMP */
2743 * Percpu allocator is initialized early during boot when neither slab or
2744 * workqueue is available. Plug async management until everything is up
2747 static int __init
percpu_enable_async(void)
2749 pcpu_async_enabled
= true;
2752 subsys_initcall(percpu_enable_async
);