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/memblock.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>
83 #include <linux/sched.h>
85 #include <asm/cacheflush.h>
86 #include <asm/sections.h>
87 #include <asm/tlbflush.h>
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/percpu.h>
93 #include "percpu-internal.h"
95 /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
96 #define PCPU_SLOT_BASE_SHIFT 5
97 /* chunks in slots below this are subject to being sidelined on failed alloc */
98 #define PCPU_SLOT_FAIL_THRESHOLD 3
100 #define PCPU_EMPTY_POP_PAGES_LOW 2
101 #define PCPU_EMPTY_POP_PAGES_HIGH 4
104 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
105 #ifndef __addr_to_pcpu_ptr
106 #define __addr_to_pcpu_ptr(addr) \
107 (void __percpu *)((unsigned long)(addr) - \
108 (unsigned long)pcpu_base_addr + \
109 (unsigned long)__per_cpu_start)
111 #ifndef __pcpu_ptr_to_addr
112 #define __pcpu_ptr_to_addr(ptr) \
113 (void __force *)((unsigned long)(ptr) + \
114 (unsigned long)pcpu_base_addr - \
115 (unsigned long)__per_cpu_start)
117 #else /* CONFIG_SMP */
118 /* on UP, it's always identity mapped */
119 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
120 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
121 #endif /* CONFIG_SMP */
123 static int pcpu_unit_pages __ro_after_init
;
124 static int pcpu_unit_size __ro_after_init
;
125 static int pcpu_nr_units __ro_after_init
;
126 static int pcpu_atom_size __ro_after_init
;
127 int pcpu_nr_slots __ro_after_init
;
128 static size_t pcpu_chunk_struct_size __ro_after_init
;
130 /* cpus with the lowest and highest unit addresses */
131 static unsigned int pcpu_low_unit_cpu __ro_after_init
;
132 static unsigned int pcpu_high_unit_cpu __ro_after_init
;
134 /* the address of the first chunk which starts with the kernel static area */
135 void *pcpu_base_addr __ro_after_init
;
136 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
138 static const int *pcpu_unit_map __ro_after_init
; /* cpu -> unit */
139 const unsigned long *pcpu_unit_offsets __ro_after_init
; /* cpu -> unit offset */
141 /* group information, used for vm allocation */
142 static int pcpu_nr_groups __ro_after_init
;
143 static const unsigned long *pcpu_group_offsets __ro_after_init
;
144 static const size_t *pcpu_group_sizes __ro_after_init
;
147 * The first chunk which always exists. Note that unlike other
148 * chunks, this one can be allocated and mapped in several different
149 * ways and thus often doesn't live in the vmalloc area.
151 struct pcpu_chunk
*pcpu_first_chunk __ro_after_init
;
154 * Optional reserved chunk. This chunk reserves part of the first
155 * chunk and serves it for reserved allocations. When the reserved
156 * region doesn't exist, the following variable is NULL.
158 struct pcpu_chunk
*pcpu_reserved_chunk __ro_after_init
;
160 DEFINE_SPINLOCK(pcpu_lock
); /* all internal data structures */
161 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* chunk create/destroy, [de]pop, map ext */
163 struct list_head
*pcpu_slot __ro_after_init
; /* chunk list slots */
165 /* chunks which need their map areas extended, protected by pcpu_lock */
166 static LIST_HEAD(pcpu_map_extend_chunks
);
169 * The number of empty populated pages, protected by pcpu_lock. The
170 * reserved chunk doesn't contribute to the count.
172 int pcpu_nr_empty_pop_pages
;
175 * The number of populated pages in use by the allocator, protected by
176 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
177 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
178 * and increments/decrements this count by 1).
180 static unsigned long pcpu_nr_populated
;
183 * Balance work is used to populate or destroy chunks asynchronously. We
184 * try to keep the number of populated free pages between
185 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
188 static void pcpu_balance_workfn(struct work_struct
*work
);
189 static DECLARE_WORK(pcpu_balance_work
, pcpu_balance_workfn
);
190 static bool pcpu_async_enabled __read_mostly
;
191 static bool pcpu_atomic_alloc_failed
;
193 static void pcpu_schedule_balance_work(void)
195 if (pcpu_async_enabled
)
196 schedule_work(&pcpu_balance_work
);
200 * pcpu_addr_in_chunk - check if the address is served from this chunk
201 * @chunk: chunk of interest
202 * @addr: percpu address
205 * True if the address is served from this chunk.
207 static bool pcpu_addr_in_chunk(struct pcpu_chunk
*chunk
, void *addr
)
209 void *start_addr
, *end_addr
;
214 start_addr
= chunk
->base_addr
+ chunk
->start_offset
;
215 end_addr
= chunk
->base_addr
+ chunk
->nr_pages
* PAGE_SIZE
-
218 return addr
>= start_addr
&& addr
< end_addr
;
221 static int __pcpu_size_to_slot(int size
)
223 int highbit
= fls(size
); /* size is in bytes */
224 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
227 static int pcpu_size_to_slot(int size
)
229 if (size
== pcpu_unit_size
)
230 return pcpu_nr_slots
- 1;
231 return __pcpu_size_to_slot(size
);
234 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
236 const struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
238 if (chunk
->free_bytes
< PCPU_MIN_ALLOC_SIZE
||
239 chunk_md
->contig_hint
== 0)
242 return pcpu_size_to_slot(chunk_md
->contig_hint
* PCPU_MIN_ALLOC_SIZE
);
245 /* set the pointer to a chunk in a page struct */
246 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
248 page
->index
= (unsigned long)pcpu
;
251 /* obtain pointer to a chunk from a page struct */
252 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
254 return (struct pcpu_chunk
*)page
->index
;
257 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
259 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
262 static unsigned long pcpu_unit_page_offset(unsigned int cpu
, int page_idx
)
264 return pcpu_unit_offsets
[cpu
] + (page_idx
<< PAGE_SHIFT
);
267 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
268 unsigned int cpu
, int page_idx
)
270 return (unsigned long)chunk
->base_addr
+
271 pcpu_unit_page_offset(cpu
, page_idx
);
274 static void pcpu_next_unpop(unsigned long *bitmap
, int *rs
, int *re
, int end
)
276 *rs
= find_next_zero_bit(bitmap
, end
, *rs
);
277 *re
= find_next_bit(bitmap
, end
, *rs
+ 1);
280 static void pcpu_next_pop(unsigned long *bitmap
, int *rs
, int *re
, int end
)
282 *rs
= find_next_bit(bitmap
, end
, *rs
);
283 *re
= find_next_zero_bit(bitmap
, end
, *rs
+ 1);
287 * Bitmap region iterators. Iterates over the bitmap between
288 * [@start, @end) in @chunk. @rs and @re should be integer variables
289 * and will be set to start and end index of the current free region.
291 #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
292 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
294 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
296 #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
297 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
299 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
302 * The following are helper functions to help access bitmaps and convert
303 * between bitmap offsets to address offsets.
305 static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk
*chunk
, int index
)
307 return chunk
->alloc_map
+
308 (index
* PCPU_BITMAP_BLOCK_BITS
/ BITS_PER_LONG
);
311 static unsigned long pcpu_off_to_block_index(int off
)
313 return off
/ PCPU_BITMAP_BLOCK_BITS
;
316 static unsigned long pcpu_off_to_block_off(int off
)
318 return off
& (PCPU_BITMAP_BLOCK_BITS
- 1);
321 static unsigned long pcpu_block_off_to_off(int index
, int off
)
323 return index
* PCPU_BITMAP_BLOCK_BITS
+ off
;
327 * pcpu_next_hint - determine which hint to use
328 * @block: block of interest
329 * @alloc_bits: size of allocation
331 * This determines if we should scan based on the scan_hint or first_free.
332 * In general, we want to scan from first_free to fulfill allocations by
333 * first fit. However, if we know a scan_hint at position scan_hint_start
334 * cannot fulfill an allocation, we can begin scanning from there knowing
335 * the contig_hint will be our fallback.
337 static int pcpu_next_hint(struct pcpu_block_md
*block
, int alloc_bits
)
340 * The three conditions below determine if we can skip past the
341 * scan_hint. First, does the scan hint exist. Second, is the
342 * contig_hint after the scan_hint (possibly not true iff
343 * contig_hint == scan_hint). Third, is the allocation request
344 * larger than the scan_hint.
346 if (block
->scan_hint
&&
347 block
->contig_hint_start
> block
->scan_hint_start
&&
348 alloc_bits
> block
->scan_hint
)
349 return block
->scan_hint_start
+ block
->scan_hint
;
351 return block
->first_free
;
355 * pcpu_next_md_free_region - finds the next hint free area
356 * @chunk: chunk of interest
357 * @bit_off: chunk offset
358 * @bits: size of free area
360 * Helper function for pcpu_for_each_md_free_region. It checks
361 * block->contig_hint and performs aggregation across blocks to find the
362 * next hint. It modifies bit_off and bits in-place to be consumed in the
365 static void pcpu_next_md_free_region(struct pcpu_chunk
*chunk
, int *bit_off
,
368 int i
= pcpu_off_to_block_index(*bit_off
);
369 int block_off
= pcpu_off_to_block_off(*bit_off
);
370 struct pcpu_block_md
*block
;
373 for (block
= chunk
->md_blocks
+ i
; i
< pcpu_chunk_nr_blocks(chunk
);
375 /* handles contig area across blocks */
377 *bits
+= block
->left_free
;
378 if (block
->left_free
== PCPU_BITMAP_BLOCK_BITS
)
384 * This checks three things. First is there a contig_hint to
385 * check. Second, have we checked this hint before by
386 * comparing the block_off. Third, is this the same as the
387 * right contig hint. In the last case, it spills over into
388 * the next block and should be handled by the contig area
389 * across blocks code.
391 *bits
= block
->contig_hint
;
392 if (*bits
&& block
->contig_hint_start
>= block_off
&&
393 *bits
+ block
->contig_hint_start
< PCPU_BITMAP_BLOCK_BITS
) {
394 *bit_off
= pcpu_block_off_to_off(i
,
395 block
->contig_hint_start
);
398 /* reset to satisfy the second predicate above */
401 *bits
= block
->right_free
;
402 *bit_off
= (i
+ 1) * PCPU_BITMAP_BLOCK_BITS
- block
->right_free
;
407 * pcpu_next_fit_region - finds fit areas for a given allocation request
408 * @chunk: chunk of interest
409 * @alloc_bits: size of allocation
410 * @align: alignment of area (max PAGE_SIZE)
411 * @bit_off: chunk offset
412 * @bits: size of free area
414 * Finds the next free region that is viable for use with a given size and
415 * alignment. This only returns if there is a valid area to be used for this
416 * allocation. block->first_free is returned if the allocation request fits
417 * within the block to see if the request can be fulfilled prior to the contig
420 static void pcpu_next_fit_region(struct pcpu_chunk
*chunk
, int alloc_bits
,
421 int align
, int *bit_off
, int *bits
)
423 int i
= pcpu_off_to_block_index(*bit_off
);
424 int block_off
= pcpu_off_to_block_off(*bit_off
);
425 struct pcpu_block_md
*block
;
428 for (block
= chunk
->md_blocks
+ i
; i
< pcpu_chunk_nr_blocks(chunk
);
430 /* handles contig area across blocks */
432 *bits
+= block
->left_free
;
433 if (*bits
>= alloc_bits
)
435 if (block
->left_free
== PCPU_BITMAP_BLOCK_BITS
)
439 /* check block->contig_hint */
440 *bits
= ALIGN(block
->contig_hint_start
, align
) -
441 block
->contig_hint_start
;
443 * This uses the block offset to determine if this has been
444 * checked in the prior iteration.
446 if (block
->contig_hint
&&
447 block
->contig_hint_start
>= block_off
&&
448 block
->contig_hint
>= *bits
+ alloc_bits
) {
449 int start
= pcpu_next_hint(block
, alloc_bits
);
451 *bits
+= alloc_bits
+ block
->contig_hint_start
-
453 *bit_off
= pcpu_block_off_to_off(i
, start
);
456 /* reset to satisfy the second predicate above */
459 *bit_off
= ALIGN(PCPU_BITMAP_BLOCK_BITS
- block
->right_free
,
461 *bits
= PCPU_BITMAP_BLOCK_BITS
- *bit_off
;
462 *bit_off
= pcpu_block_off_to_off(i
, *bit_off
);
463 if (*bits
>= alloc_bits
)
467 /* no valid offsets were found - fail condition */
468 *bit_off
= pcpu_chunk_map_bits(chunk
);
472 * Metadata free area iterators. These perform aggregation of free areas
473 * based on the metadata blocks and return the offset @bit_off and size in
474 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
475 * a fit is found for the allocation request.
477 #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
478 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
479 (bit_off) < pcpu_chunk_map_bits((chunk)); \
480 (bit_off) += (bits) + 1, \
481 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
483 #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
484 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
486 (bit_off) < pcpu_chunk_map_bits((chunk)); \
487 (bit_off) += (bits), \
488 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
492 * pcpu_mem_zalloc - allocate memory
493 * @size: bytes to allocate
494 * @gfp: allocation flags
496 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
498 * This is to facilitate passing through whitelisted flags. The
499 * returned memory is always zeroed.
502 * Pointer to the allocated area on success, NULL on failure.
504 static void *pcpu_mem_zalloc(size_t size
, gfp_t gfp
)
506 if (WARN_ON_ONCE(!slab_is_available()))
509 if (size
<= PAGE_SIZE
)
510 return kzalloc(size
, gfp
);
512 return __vmalloc(size
, gfp
| __GFP_ZERO
, PAGE_KERNEL
);
516 * pcpu_mem_free - free memory
517 * @ptr: memory to free
519 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
521 static void pcpu_mem_free(void *ptr
)
526 static void __pcpu_chunk_move(struct pcpu_chunk
*chunk
, int slot
,
529 if (chunk
!= pcpu_reserved_chunk
) {
531 list_move(&chunk
->list
, &pcpu_slot
[slot
]);
533 list_move_tail(&chunk
->list
, &pcpu_slot
[slot
]);
537 static void pcpu_chunk_move(struct pcpu_chunk
*chunk
, int slot
)
539 __pcpu_chunk_move(chunk
, slot
, true);
543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
544 * @chunk: chunk of interest
545 * @oslot: the previous slot it was on
547 * This function is called after an allocation or free changed @chunk.
548 * New slot according to the changed state is determined and @chunk is
549 * moved to the slot. Note that the reserved chunk is never put on
555 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
557 int nslot
= pcpu_chunk_slot(chunk
);
560 __pcpu_chunk_move(chunk
, nslot
, oslot
< nslot
);
564 * pcpu_update_empty_pages - update empty page counters
565 * @chunk: chunk of interest
566 * @nr: nr of empty pages
568 * This is used to keep track of the empty pages now based on the premise
569 * a md_block covers a page. The hint update functions recognize if a block
570 * is made full or broken to calculate deltas for keeping track of free pages.
572 static inline void pcpu_update_empty_pages(struct pcpu_chunk
*chunk
, int nr
)
574 chunk
->nr_empty_pop_pages
+= nr
;
575 if (chunk
!= pcpu_reserved_chunk
)
576 pcpu_nr_empty_pop_pages
+= nr
;
580 * pcpu_region_overlap - determines if two regions overlap
581 * @a: start of first region, inclusive
582 * @b: end of first region, exclusive
583 * @x: start of second region, inclusive
584 * @y: end of second region, exclusive
586 * This is used to determine if the hint region [a, b) overlaps with the
587 * allocated region [x, y).
589 static inline bool pcpu_region_overlap(int a
, int b
, int x
, int y
)
591 return (a
< y
) && (x
< b
);
595 * pcpu_block_update - updates a block given a free area
596 * @block: block of interest
597 * @start: start offset in block
598 * @end: end offset in block
600 * Updates a block given a known free area. The region [start, end) is
601 * expected to be the entirety of the free area within a block. Chooses
602 * the best starting offset if the contig hints are equal.
604 static void pcpu_block_update(struct pcpu_block_md
*block
, int start
, int end
)
606 int contig
= end
- start
;
608 block
->first_free
= min(block
->first_free
, start
);
610 block
->left_free
= contig
;
612 if (end
== block
->nr_bits
)
613 block
->right_free
= contig
;
615 if (contig
> block
->contig_hint
) {
616 /* promote the old contig_hint to be the new scan_hint */
617 if (start
> block
->contig_hint_start
) {
618 if (block
->contig_hint
> block
->scan_hint
) {
619 block
->scan_hint_start
=
620 block
->contig_hint_start
;
621 block
->scan_hint
= block
->contig_hint
;
622 } else if (start
< block
->scan_hint_start
) {
624 * The old contig_hint == scan_hint. But, the
625 * new contig is larger so hold the invariant
626 * scan_hint_start < contig_hint_start.
628 block
->scan_hint
= 0;
631 block
->scan_hint
= 0;
633 block
->contig_hint_start
= start
;
634 block
->contig_hint
= contig
;
635 } else if (contig
== block
->contig_hint
) {
636 if (block
->contig_hint_start
&&
638 __ffs(start
) > __ffs(block
->contig_hint_start
))) {
639 /* start has a better alignment so use it */
640 block
->contig_hint_start
= start
;
641 if (start
< block
->scan_hint_start
&&
642 block
->contig_hint
> block
->scan_hint
)
643 block
->scan_hint
= 0;
644 } else if (start
> block
->scan_hint_start
||
645 block
->contig_hint
> block
->scan_hint
) {
647 * Knowing contig == contig_hint, update the scan_hint
648 * if it is farther than or larger than the current
651 block
->scan_hint_start
= start
;
652 block
->scan_hint
= contig
;
656 * The region is smaller than the contig_hint. So only update
657 * the scan_hint if it is larger than or equal and farther than
658 * the current scan_hint.
660 if ((start
< block
->contig_hint_start
&&
661 (contig
> block
->scan_hint
||
662 (contig
== block
->scan_hint
&&
663 start
> block
->scan_hint_start
)))) {
664 block
->scan_hint_start
= start
;
665 block
->scan_hint
= contig
;
671 * pcpu_block_update_scan - update a block given a free area from a scan
672 * @chunk: chunk of interest
673 * @bit_off: chunk offset
674 * @bits: size of free area
676 * Finding the final allocation spot first goes through pcpu_find_block_fit()
677 * to find a block that can hold the allocation and then pcpu_alloc_area()
678 * where a scan is used. When allocations require specific alignments,
679 * we can inadvertently create holes which will not be seen in the alloc
682 * This takes a given free area hole and updates a block as it may change the
683 * scan_hint. We need to scan backwards to ensure we don't miss free bits
686 static void pcpu_block_update_scan(struct pcpu_chunk
*chunk
, int bit_off
,
689 int s_off
= pcpu_off_to_block_off(bit_off
);
690 int e_off
= s_off
+ bits
;
692 struct pcpu_block_md
*block
;
694 if (e_off
> PCPU_BITMAP_BLOCK_BITS
)
697 s_index
= pcpu_off_to_block_index(bit_off
);
698 block
= chunk
->md_blocks
+ s_index
;
700 /* scan backwards in case of alignment skipping free bits */
701 l_bit
= find_last_bit(pcpu_index_alloc_map(chunk
, s_index
), s_off
);
702 s_off
= (s_off
== l_bit
) ? 0 : l_bit
+ 1;
704 pcpu_block_update(block
, s_off
, e_off
);
708 * pcpu_chunk_refresh_hint - updates metadata about a chunk
709 * @chunk: chunk of interest
710 * @full_scan: if we should scan from the beginning
712 * Iterates over the metadata blocks to find the largest contig area.
713 * A full scan can be avoided on the allocation path as this is triggered
714 * if we broke the contig_hint. In doing so, the scan_hint will be before
715 * the contig_hint or after if the scan_hint == contig_hint. This cannot
716 * be prevented on freeing as we want to find the largest area possibly
719 static void pcpu_chunk_refresh_hint(struct pcpu_chunk
*chunk
, bool full_scan
)
721 struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
724 /* promote scan_hint to contig_hint */
725 if (!full_scan
&& chunk_md
->scan_hint
) {
726 bit_off
= chunk_md
->scan_hint_start
+ chunk_md
->scan_hint
;
727 chunk_md
->contig_hint_start
= chunk_md
->scan_hint_start
;
728 chunk_md
->contig_hint
= chunk_md
->scan_hint
;
729 chunk_md
->scan_hint
= 0;
731 bit_off
= chunk_md
->first_free
;
732 chunk_md
->contig_hint
= 0;
736 pcpu_for_each_md_free_region(chunk
, bit_off
, bits
) {
737 pcpu_block_update(chunk_md
, bit_off
, bit_off
+ bits
);
742 * pcpu_block_refresh_hint
743 * @chunk: chunk of interest
744 * @index: index of the metadata block
746 * Scans over the block beginning at first_free and updates the block
747 * metadata accordingly.
749 static void pcpu_block_refresh_hint(struct pcpu_chunk
*chunk
, int index
)
751 struct pcpu_block_md
*block
= chunk
->md_blocks
+ index
;
752 unsigned long *alloc_map
= pcpu_index_alloc_map(chunk
, index
);
753 int rs
, re
, start
; /* region start, region end */
755 /* promote scan_hint to contig_hint */
756 if (block
->scan_hint
) {
757 start
= block
->scan_hint_start
+ block
->scan_hint
;
758 block
->contig_hint_start
= block
->scan_hint_start
;
759 block
->contig_hint
= block
->scan_hint
;
760 block
->scan_hint
= 0;
762 start
= block
->first_free
;
763 block
->contig_hint
= 0;
766 block
->right_free
= 0;
768 /* iterate over free areas and update the contig hints */
769 pcpu_for_each_unpop_region(alloc_map
, rs
, re
, start
,
770 PCPU_BITMAP_BLOCK_BITS
) {
771 pcpu_block_update(block
, rs
, re
);
776 * pcpu_block_update_hint_alloc - update hint on allocation path
777 * @chunk: chunk of interest
778 * @bit_off: chunk offset
779 * @bits: size of request
781 * Updates metadata for the allocation path. The metadata only has to be
782 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
783 * scans are required if the block's contig hint is broken.
785 static void pcpu_block_update_hint_alloc(struct pcpu_chunk
*chunk
, int bit_off
,
788 struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
789 int nr_empty_pages
= 0;
790 struct pcpu_block_md
*s_block
, *e_block
, *block
;
791 int s_index
, e_index
; /* block indexes of the freed allocation */
792 int s_off
, e_off
; /* block offsets of the freed allocation */
795 * Calculate per block offsets.
796 * The calculation uses an inclusive range, but the resulting offsets
797 * are [start, end). e_index always points to the last block in the
800 s_index
= pcpu_off_to_block_index(bit_off
);
801 e_index
= pcpu_off_to_block_index(bit_off
+ bits
- 1);
802 s_off
= pcpu_off_to_block_off(bit_off
);
803 e_off
= pcpu_off_to_block_off(bit_off
+ bits
- 1) + 1;
805 s_block
= chunk
->md_blocks
+ s_index
;
806 e_block
= chunk
->md_blocks
+ e_index
;
810 * block->first_free must be updated if the allocation takes its place.
811 * If the allocation breaks the contig_hint, a scan is required to
814 if (s_block
->contig_hint
== PCPU_BITMAP_BLOCK_BITS
)
817 if (s_off
== s_block
->first_free
)
818 s_block
->first_free
= find_next_zero_bit(
819 pcpu_index_alloc_map(chunk
, s_index
),
820 PCPU_BITMAP_BLOCK_BITS
,
823 if (pcpu_region_overlap(s_block
->scan_hint_start
,
824 s_block
->scan_hint_start
+ s_block
->scan_hint
,
827 s_block
->scan_hint
= 0;
829 if (pcpu_region_overlap(s_block
->contig_hint_start
,
830 s_block
->contig_hint_start
+
831 s_block
->contig_hint
,
834 /* block contig hint is broken - scan to fix it */
836 s_block
->left_free
= 0;
837 pcpu_block_refresh_hint(chunk
, s_index
);
839 /* update left and right contig manually */
840 s_block
->left_free
= min(s_block
->left_free
, s_off
);
841 if (s_index
== e_index
)
842 s_block
->right_free
= min_t(int, s_block
->right_free
,
843 PCPU_BITMAP_BLOCK_BITS
- e_off
);
845 s_block
->right_free
= 0;
851 if (s_index
!= e_index
) {
852 if (e_block
->contig_hint
== PCPU_BITMAP_BLOCK_BITS
)
856 * When the allocation is across blocks, the end is along
857 * the left part of the e_block.
859 e_block
->first_free
= find_next_zero_bit(
860 pcpu_index_alloc_map(chunk
, e_index
),
861 PCPU_BITMAP_BLOCK_BITS
, e_off
);
863 if (e_off
== PCPU_BITMAP_BLOCK_BITS
) {
864 /* reset the block */
867 if (e_off
> e_block
->scan_hint_start
)
868 e_block
->scan_hint
= 0;
870 e_block
->left_free
= 0;
871 if (e_off
> e_block
->contig_hint_start
) {
872 /* contig hint is broken - scan to fix it */
873 pcpu_block_refresh_hint(chunk
, e_index
);
875 e_block
->right_free
=
876 min_t(int, e_block
->right_free
,
877 PCPU_BITMAP_BLOCK_BITS
- e_off
);
881 /* update in-between md_blocks */
882 nr_empty_pages
+= (e_index
- s_index
- 1);
883 for (block
= s_block
+ 1; block
< e_block
; block
++) {
884 block
->scan_hint
= 0;
885 block
->contig_hint
= 0;
886 block
->left_free
= 0;
887 block
->right_free
= 0;
892 pcpu_update_empty_pages(chunk
, -nr_empty_pages
);
894 if (pcpu_region_overlap(chunk_md
->scan_hint_start
,
895 chunk_md
->scan_hint_start
+
899 chunk_md
->scan_hint
= 0;
902 * The only time a full chunk scan is required is if the chunk
903 * contig hint is broken. Otherwise, it means a smaller space
904 * was used and therefore the chunk contig hint is still correct.
906 if (pcpu_region_overlap(chunk_md
->contig_hint_start
,
907 chunk_md
->contig_hint_start
+
908 chunk_md
->contig_hint
,
911 pcpu_chunk_refresh_hint(chunk
, false);
915 * pcpu_block_update_hint_free - updates the block hints on the free path
916 * @chunk: chunk of interest
917 * @bit_off: chunk offset
918 * @bits: size of request
920 * Updates metadata for the allocation path. This avoids a blind block
921 * refresh by making use of the block contig hints. If this fails, it scans
922 * forward and backward to determine the extent of the free area. This is
923 * capped at the boundary of blocks.
925 * A chunk update is triggered if a page becomes free, a block becomes free,
926 * or the free spans across blocks. This tradeoff is to minimize iterating
927 * over the block metadata to update chunk_md->contig_hint.
928 * chunk_md->contig_hint may be off by up to a page, but it will never be more
929 * than the available space. If the contig hint is contained in one block, it
932 static void pcpu_block_update_hint_free(struct pcpu_chunk
*chunk
, int bit_off
,
935 int nr_empty_pages
= 0;
936 struct pcpu_block_md
*s_block
, *e_block
, *block
;
937 int s_index
, e_index
; /* block indexes of the freed allocation */
938 int s_off
, e_off
; /* block offsets of the freed allocation */
939 int start
, end
; /* start and end of the whole free area */
942 * Calculate per block offsets.
943 * The calculation uses an inclusive range, but the resulting offsets
944 * are [start, end). e_index always points to the last block in the
947 s_index
= pcpu_off_to_block_index(bit_off
);
948 e_index
= pcpu_off_to_block_index(bit_off
+ bits
- 1);
949 s_off
= pcpu_off_to_block_off(bit_off
);
950 e_off
= pcpu_off_to_block_off(bit_off
+ bits
- 1) + 1;
952 s_block
= chunk
->md_blocks
+ s_index
;
953 e_block
= chunk
->md_blocks
+ e_index
;
956 * Check if the freed area aligns with the block->contig_hint.
957 * If it does, then the scan to find the beginning/end of the
958 * larger free area can be avoided.
960 * start and end refer to beginning and end of the free area
961 * within each their respective blocks. This is not necessarily
962 * the entire free area as it may span blocks past the beginning
963 * or end of the block.
966 if (s_off
== s_block
->contig_hint
+ s_block
->contig_hint_start
) {
967 start
= s_block
->contig_hint_start
;
970 * Scan backwards to find the extent of the free area.
971 * find_last_bit returns the starting bit, so if the start bit
972 * is returned, that means there was no last bit and the
973 * remainder of the chunk is free.
975 int l_bit
= find_last_bit(pcpu_index_alloc_map(chunk
, s_index
),
977 start
= (start
== l_bit
) ? 0 : l_bit
+ 1;
981 if (e_off
== e_block
->contig_hint_start
)
982 end
= e_block
->contig_hint_start
+ e_block
->contig_hint
;
984 end
= find_next_bit(pcpu_index_alloc_map(chunk
, e_index
),
985 PCPU_BITMAP_BLOCK_BITS
, end
);
988 e_off
= (s_index
== e_index
) ? end
: PCPU_BITMAP_BLOCK_BITS
;
989 if (!start
&& e_off
== PCPU_BITMAP_BLOCK_BITS
)
991 pcpu_block_update(s_block
, start
, e_off
);
993 /* freeing in the same block */
994 if (s_index
!= e_index
) {
996 if (end
== PCPU_BITMAP_BLOCK_BITS
)
998 pcpu_block_update(e_block
, 0, end
);
1000 /* reset md_blocks in the middle */
1001 nr_empty_pages
+= (e_index
- s_index
- 1);
1002 for (block
= s_block
+ 1; block
< e_block
; block
++) {
1003 block
->first_free
= 0;
1004 block
->scan_hint
= 0;
1005 block
->contig_hint_start
= 0;
1006 block
->contig_hint
= PCPU_BITMAP_BLOCK_BITS
;
1007 block
->left_free
= PCPU_BITMAP_BLOCK_BITS
;
1008 block
->right_free
= PCPU_BITMAP_BLOCK_BITS
;
1013 pcpu_update_empty_pages(chunk
, nr_empty_pages
);
1016 * Refresh chunk metadata when the free makes a block free or spans
1017 * across blocks. The contig_hint may be off by up to a page, but if
1018 * the contig_hint is contained in a block, it will be accurate with
1019 * the else condition below.
1021 if (((end
- start
) >= PCPU_BITMAP_BLOCK_BITS
) || s_index
!= e_index
)
1022 pcpu_chunk_refresh_hint(chunk
, true);
1024 pcpu_block_update(&chunk
->chunk_md
,
1025 pcpu_block_off_to_off(s_index
, start
),
1030 * pcpu_is_populated - determines if the region is populated
1031 * @chunk: chunk of interest
1032 * @bit_off: chunk offset
1033 * @bits: size of area
1034 * @next_off: return value for the next offset to start searching
1036 * For atomic allocations, check if the backing pages are populated.
1039 * Bool if the backing pages are populated.
1040 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1042 static bool pcpu_is_populated(struct pcpu_chunk
*chunk
, int bit_off
, int bits
,
1045 int page_start
, page_end
, rs
, re
;
1047 page_start
= PFN_DOWN(bit_off
* PCPU_MIN_ALLOC_SIZE
);
1048 page_end
= PFN_UP((bit_off
+ bits
) * PCPU_MIN_ALLOC_SIZE
);
1051 pcpu_next_unpop(chunk
->populated
, &rs
, &re
, page_end
);
1055 *next_off
= re
* PAGE_SIZE
/ PCPU_MIN_ALLOC_SIZE
;
1060 * pcpu_find_block_fit - finds the block index to start searching
1061 * @chunk: chunk of interest
1062 * @alloc_bits: size of request in allocation units
1063 * @align: alignment of area (max PAGE_SIZE bytes)
1064 * @pop_only: use populated regions only
1066 * Given a chunk and an allocation spec, find the offset to begin searching
1067 * for a free region. This iterates over the bitmap metadata blocks to
1068 * find an offset that will be guaranteed to fit the requirements. It is
1069 * not quite first fit as if the allocation does not fit in the contig hint
1070 * of a block or chunk, it is skipped. This errs on the side of caution
1071 * to prevent excess iteration. Poor alignment can cause the allocator to
1072 * skip over blocks and chunks that have valid free areas.
1075 * The offset in the bitmap to begin searching.
1076 * -1 if no offset is found.
1078 static int pcpu_find_block_fit(struct pcpu_chunk
*chunk
, int alloc_bits
,
1079 size_t align
, bool pop_only
)
1081 struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
1082 int bit_off
, bits
, next_off
;
1085 * Check to see if the allocation can fit in the chunk's contig hint.
1086 * This is an optimization to prevent scanning by assuming if it
1087 * cannot fit in the global hint, there is memory pressure and creating
1088 * a new chunk would happen soon.
1090 bit_off
= ALIGN(chunk_md
->contig_hint_start
, align
) -
1091 chunk_md
->contig_hint_start
;
1092 if (bit_off
+ alloc_bits
> chunk_md
->contig_hint
)
1095 bit_off
= pcpu_next_hint(chunk_md
, alloc_bits
);
1097 pcpu_for_each_fit_region(chunk
, alloc_bits
, align
, bit_off
, bits
) {
1098 if (!pop_only
|| pcpu_is_populated(chunk
, bit_off
, bits
,
1106 if (bit_off
== pcpu_chunk_map_bits(chunk
))
1113 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1114 * @map: the address to base the search on
1115 * @size: the bitmap size in bits
1116 * @start: the bitnumber to start searching at
1117 * @nr: the number of zeroed bits we're looking for
1118 * @align_mask: alignment mask for zero area
1119 * @largest_off: offset of the largest area skipped
1120 * @largest_bits: size of the largest area skipped
1122 * The @align_mask should be one less than a power of 2.
1124 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1125 * the largest area that was skipped. This is imperfect, but in general is
1126 * good enough. The largest remembered region is the largest failed region
1127 * seen. This does not include anything we possibly skipped due to alignment.
1128 * pcpu_block_update_scan() does scan backwards to try and recover what was
1129 * lost to alignment. While this can cause scanning to miss earlier possible
1130 * free areas, smaller allocations will eventually fill those holes.
1132 static unsigned long pcpu_find_zero_area(unsigned long *map
,
1134 unsigned long start
,
1136 unsigned long align_mask
,
1137 unsigned long *largest_off
,
1138 unsigned long *largest_bits
)
1140 unsigned long index
, end
, i
, area_off
, area_bits
;
1142 index
= find_next_zero_bit(map
, size
, start
);
1144 /* Align allocation */
1145 index
= __ALIGN_MASK(index
, align_mask
);
1151 i
= find_next_bit(map
, end
, index
);
1153 area_bits
= i
- area_off
;
1154 /* remember largest unused area with best alignment */
1155 if (area_bits
> *largest_bits
||
1156 (area_bits
== *largest_bits
&& *largest_off
&&
1157 (!area_off
|| __ffs(area_off
) > __ffs(*largest_off
)))) {
1158 *largest_off
= area_off
;
1159 *largest_bits
= area_bits
;
1169 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1170 * @chunk: chunk of interest
1171 * @alloc_bits: size of request in allocation units
1172 * @align: alignment of area (max PAGE_SIZE)
1173 * @start: bit_off to start searching
1175 * This function takes in a @start offset to begin searching to fit an
1176 * allocation of @alloc_bits with alignment @align. It needs to scan
1177 * the allocation map because if it fits within the block's contig hint,
1178 * @start will be block->first_free. This is an attempt to fill the
1179 * allocation prior to breaking the contig hint. The allocation and
1180 * boundary maps are updated accordingly if it confirms a valid
1184 * Allocated addr offset in @chunk on success.
1185 * -1 if no matching area is found.
1187 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int alloc_bits
,
1188 size_t align
, int start
)
1190 struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
1191 size_t align_mask
= (align
) ? (align
- 1) : 0;
1192 unsigned long area_off
= 0, area_bits
= 0;
1193 int bit_off
, end
, oslot
;
1195 lockdep_assert_held(&pcpu_lock
);
1197 oslot
= pcpu_chunk_slot(chunk
);
1200 * Search to find a fit.
1202 end
= min_t(int, start
+ alloc_bits
+ PCPU_BITMAP_BLOCK_BITS
,
1203 pcpu_chunk_map_bits(chunk
));
1204 bit_off
= pcpu_find_zero_area(chunk
->alloc_map
, end
, start
, alloc_bits
,
1205 align_mask
, &area_off
, &area_bits
);
1210 pcpu_block_update_scan(chunk
, area_off
, area_bits
);
1212 /* update alloc map */
1213 bitmap_set(chunk
->alloc_map
, bit_off
, alloc_bits
);
1215 /* update boundary map */
1216 set_bit(bit_off
, chunk
->bound_map
);
1217 bitmap_clear(chunk
->bound_map
, bit_off
+ 1, alloc_bits
- 1);
1218 set_bit(bit_off
+ alloc_bits
, chunk
->bound_map
);
1220 chunk
->free_bytes
-= alloc_bits
* PCPU_MIN_ALLOC_SIZE
;
1222 /* update first free bit */
1223 if (bit_off
== chunk_md
->first_free
)
1224 chunk_md
->first_free
= find_next_zero_bit(
1226 pcpu_chunk_map_bits(chunk
),
1227 bit_off
+ alloc_bits
);
1229 pcpu_block_update_hint_alloc(chunk
, bit_off
, alloc_bits
);
1231 pcpu_chunk_relocate(chunk
, oslot
);
1233 return bit_off
* PCPU_MIN_ALLOC_SIZE
;
1237 * pcpu_free_area - frees the corresponding offset
1238 * @chunk: chunk of interest
1239 * @off: addr offset into chunk
1241 * This function determines the size of an allocation to free using
1242 * the boundary bitmap and clears the allocation map.
1244 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int off
)
1246 struct pcpu_block_md
*chunk_md
= &chunk
->chunk_md
;
1247 int bit_off
, bits
, end
, oslot
;
1249 lockdep_assert_held(&pcpu_lock
);
1250 pcpu_stats_area_dealloc(chunk
);
1252 oslot
= pcpu_chunk_slot(chunk
);
1254 bit_off
= off
/ PCPU_MIN_ALLOC_SIZE
;
1256 /* find end index */
1257 end
= find_next_bit(chunk
->bound_map
, pcpu_chunk_map_bits(chunk
),
1259 bits
= end
- bit_off
;
1260 bitmap_clear(chunk
->alloc_map
, bit_off
, bits
);
1262 /* update metadata */
1263 chunk
->free_bytes
+= bits
* PCPU_MIN_ALLOC_SIZE
;
1265 /* update first free bit */
1266 chunk_md
->first_free
= min(chunk_md
->first_free
, bit_off
);
1268 pcpu_block_update_hint_free(chunk
, bit_off
, bits
);
1270 pcpu_chunk_relocate(chunk
, oslot
);
1273 static void pcpu_init_md_block(struct pcpu_block_md
*block
, int nr_bits
)
1275 block
->scan_hint
= 0;
1276 block
->contig_hint
= nr_bits
;
1277 block
->left_free
= nr_bits
;
1278 block
->right_free
= nr_bits
;
1279 block
->first_free
= 0;
1280 block
->nr_bits
= nr_bits
;
1283 static void pcpu_init_md_blocks(struct pcpu_chunk
*chunk
)
1285 struct pcpu_block_md
*md_block
;
1287 /* init the chunk's block */
1288 pcpu_init_md_block(&chunk
->chunk_md
, pcpu_chunk_map_bits(chunk
));
1290 for (md_block
= chunk
->md_blocks
;
1291 md_block
!= chunk
->md_blocks
+ pcpu_chunk_nr_blocks(chunk
);
1293 pcpu_init_md_block(md_block
, PCPU_BITMAP_BLOCK_BITS
);
1297 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1298 * @tmp_addr: the start of the region served
1299 * @map_size: size of the region served
1301 * This is responsible for creating the chunks that serve the first chunk. The
1302 * base_addr is page aligned down of @tmp_addr while the region end is page
1303 * aligned up. Offsets are kept track of to determine the region served. All
1304 * this is done to appease the bitmap allocator in avoiding partial blocks.
1307 * Chunk serving the region at @tmp_addr of @map_size.
1309 static struct pcpu_chunk
* __init
pcpu_alloc_first_chunk(unsigned long tmp_addr
,
1312 struct pcpu_chunk
*chunk
;
1313 unsigned long aligned_addr
, lcm_align
;
1314 int start_offset
, offset_bits
, region_size
, region_bits
;
1317 /* region calculations */
1318 aligned_addr
= tmp_addr
& PAGE_MASK
;
1320 start_offset
= tmp_addr
- aligned_addr
;
1323 * Align the end of the region with the LCM of PAGE_SIZE and
1324 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1327 lcm_align
= lcm(PAGE_SIZE
, PCPU_BITMAP_BLOCK_SIZE
);
1328 region_size
= ALIGN(start_offset
+ map_size
, lcm_align
);
1330 /* allocate chunk */
1331 alloc_size
= sizeof(struct pcpu_chunk
) +
1332 BITS_TO_LONGS(region_size
>> PAGE_SHIFT
);
1333 chunk
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
1335 panic("%s: Failed to allocate %zu bytes\n", __func__
,
1338 INIT_LIST_HEAD(&chunk
->list
);
1340 chunk
->base_addr
= (void *)aligned_addr
;
1341 chunk
->start_offset
= start_offset
;
1342 chunk
->end_offset
= region_size
- chunk
->start_offset
- map_size
;
1344 chunk
->nr_pages
= region_size
>> PAGE_SHIFT
;
1345 region_bits
= pcpu_chunk_map_bits(chunk
);
1347 alloc_size
= BITS_TO_LONGS(region_bits
) * sizeof(chunk
->alloc_map
[0]);
1348 chunk
->alloc_map
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
1349 if (!chunk
->alloc_map
)
1350 panic("%s: Failed to allocate %zu bytes\n", __func__
,
1354 BITS_TO_LONGS(region_bits
+ 1) * sizeof(chunk
->bound_map
[0]);
1355 chunk
->bound_map
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
1356 if (!chunk
->bound_map
)
1357 panic("%s: Failed to allocate %zu bytes\n", __func__
,
1360 alloc_size
= pcpu_chunk_nr_blocks(chunk
) * sizeof(chunk
->md_blocks
[0]);
1361 chunk
->md_blocks
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
1362 if (!chunk
->md_blocks
)
1363 panic("%s: Failed to allocate %zu bytes\n", __func__
,
1366 pcpu_init_md_blocks(chunk
);
1368 /* manage populated page bitmap */
1369 chunk
->immutable
= true;
1370 bitmap_fill(chunk
->populated
, chunk
->nr_pages
);
1371 chunk
->nr_populated
= chunk
->nr_pages
;
1372 chunk
->nr_empty_pop_pages
= chunk
->nr_pages
;
1374 chunk
->free_bytes
= map_size
;
1376 if (chunk
->start_offset
) {
1377 /* hide the beginning of the bitmap */
1378 offset_bits
= chunk
->start_offset
/ PCPU_MIN_ALLOC_SIZE
;
1379 bitmap_set(chunk
->alloc_map
, 0, offset_bits
);
1380 set_bit(0, chunk
->bound_map
);
1381 set_bit(offset_bits
, chunk
->bound_map
);
1383 chunk
->chunk_md
.first_free
= offset_bits
;
1385 pcpu_block_update_hint_alloc(chunk
, 0, offset_bits
);
1388 if (chunk
->end_offset
) {
1389 /* hide the end of the bitmap */
1390 offset_bits
= chunk
->end_offset
/ PCPU_MIN_ALLOC_SIZE
;
1391 bitmap_set(chunk
->alloc_map
,
1392 pcpu_chunk_map_bits(chunk
) - offset_bits
,
1394 set_bit((start_offset
+ map_size
) / PCPU_MIN_ALLOC_SIZE
,
1396 set_bit(region_bits
, chunk
->bound_map
);
1398 pcpu_block_update_hint_alloc(chunk
, pcpu_chunk_map_bits(chunk
)
1399 - offset_bits
, offset_bits
);
1405 static struct pcpu_chunk
*pcpu_alloc_chunk(gfp_t gfp
)
1407 struct pcpu_chunk
*chunk
;
1410 chunk
= pcpu_mem_zalloc(pcpu_chunk_struct_size
, gfp
);
1414 INIT_LIST_HEAD(&chunk
->list
);
1415 chunk
->nr_pages
= pcpu_unit_pages
;
1416 region_bits
= pcpu_chunk_map_bits(chunk
);
1418 chunk
->alloc_map
= pcpu_mem_zalloc(BITS_TO_LONGS(region_bits
) *
1419 sizeof(chunk
->alloc_map
[0]), gfp
);
1420 if (!chunk
->alloc_map
)
1421 goto alloc_map_fail
;
1423 chunk
->bound_map
= pcpu_mem_zalloc(BITS_TO_LONGS(region_bits
+ 1) *
1424 sizeof(chunk
->bound_map
[0]), gfp
);
1425 if (!chunk
->bound_map
)
1426 goto bound_map_fail
;
1428 chunk
->md_blocks
= pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk
) *
1429 sizeof(chunk
->md_blocks
[0]), gfp
);
1430 if (!chunk
->md_blocks
)
1431 goto md_blocks_fail
;
1433 pcpu_init_md_blocks(chunk
);
1436 chunk
->free_bytes
= chunk
->nr_pages
* PAGE_SIZE
;
1441 pcpu_mem_free(chunk
->bound_map
);
1443 pcpu_mem_free(chunk
->alloc_map
);
1445 pcpu_mem_free(chunk
);
1450 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
1454 pcpu_mem_free(chunk
->md_blocks
);
1455 pcpu_mem_free(chunk
->bound_map
);
1456 pcpu_mem_free(chunk
->alloc_map
);
1457 pcpu_mem_free(chunk
);
1461 * pcpu_chunk_populated - post-population bookkeeping
1462 * @chunk: pcpu_chunk which got populated
1463 * @page_start: the start page
1464 * @page_end: the end page
1466 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1467 * the bookkeeping information accordingly. Must be called after each
1468 * successful population.
1470 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1471 * is to serve an allocation in that area.
1473 static void pcpu_chunk_populated(struct pcpu_chunk
*chunk
, int page_start
,
1476 int nr
= page_end
- page_start
;
1478 lockdep_assert_held(&pcpu_lock
);
1480 bitmap_set(chunk
->populated
, page_start
, nr
);
1481 chunk
->nr_populated
+= nr
;
1482 pcpu_nr_populated
+= nr
;
1484 pcpu_update_empty_pages(chunk
, nr
);
1488 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1489 * @chunk: pcpu_chunk which got depopulated
1490 * @page_start: the start page
1491 * @page_end: the end page
1493 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1494 * Update the bookkeeping information accordingly. Must be called after
1495 * each successful depopulation.
1497 static void pcpu_chunk_depopulated(struct pcpu_chunk
*chunk
,
1498 int page_start
, int page_end
)
1500 int nr
= page_end
- page_start
;
1502 lockdep_assert_held(&pcpu_lock
);
1504 bitmap_clear(chunk
->populated
, page_start
, nr
);
1505 chunk
->nr_populated
-= nr
;
1506 pcpu_nr_populated
-= nr
;
1508 pcpu_update_empty_pages(chunk
, -nr
);
1512 * Chunk management implementation.
1514 * To allow different implementations, chunk alloc/free and
1515 * [de]population are implemented in a separate file which is pulled
1516 * into this file and compiled together. The following functions
1517 * should be implemented.
1519 * pcpu_populate_chunk - populate the specified range of a chunk
1520 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1521 * pcpu_create_chunk - create a new chunk
1522 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1523 * pcpu_addr_to_page - translate address to physical address
1524 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1526 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
,
1527 int page_start
, int page_end
, gfp_t gfp
);
1528 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
,
1529 int page_start
, int page_end
);
1530 static struct pcpu_chunk
*pcpu_create_chunk(gfp_t gfp
);
1531 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
1532 static struct page
*pcpu_addr_to_page(void *addr
);
1533 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
1535 #ifdef CONFIG_NEED_PER_CPU_KM
1536 #include "percpu-km.c"
1538 #include "percpu-vm.c"
1542 * pcpu_chunk_addr_search - determine chunk containing specified address
1543 * @addr: address for which the chunk needs to be determined.
1545 * This is an internal function that handles all but static allocations.
1546 * Static percpu address values should never be passed into the allocator.
1549 * The address of the found chunk.
1551 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
1553 /* is it in the dynamic region (first chunk)? */
1554 if (pcpu_addr_in_chunk(pcpu_first_chunk
, addr
))
1555 return pcpu_first_chunk
;
1557 /* is it in the reserved region? */
1558 if (pcpu_addr_in_chunk(pcpu_reserved_chunk
, addr
))
1559 return pcpu_reserved_chunk
;
1562 * The address is relative to unit0 which might be unused and
1563 * thus unmapped. Offset the address to the unit space of the
1564 * current processor before looking it up in the vmalloc
1565 * space. Note that any possible cpu id can be used here, so
1566 * there's no need to worry about preemption or cpu hotplug.
1568 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
1569 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
1573 * pcpu_alloc - the percpu allocator
1574 * @size: size of area to allocate in bytes
1575 * @align: alignment of area (max PAGE_SIZE)
1576 * @reserved: allocate from the reserved chunk if available
1577 * @gfp: allocation flags
1579 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1580 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1581 * then no warning will be triggered on invalid or failed allocation
1585 * Percpu pointer to the allocated area on success, NULL on failure.
1587 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
,
1590 /* whitelisted flags that can be passed to the backing allocators */
1591 gfp_t pcpu_gfp
= gfp
& (GFP_KERNEL
| __GFP_NORETRY
| __GFP_NOWARN
);
1592 bool is_atomic
= (gfp
& GFP_KERNEL
) != GFP_KERNEL
;
1593 bool do_warn
= !(gfp
& __GFP_NOWARN
);
1594 static int warn_limit
= 10;
1595 struct pcpu_chunk
*chunk
, *next
;
1597 int slot
, off
, cpu
, ret
;
1598 unsigned long flags
;
1600 size_t bits
, bit_align
;
1603 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1604 * therefore alignment must be a minimum of that many bytes.
1605 * An allocation may have internal fragmentation from rounding up
1606 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1608 if (unlikely(align
< PCPU_MIN_ALLOC_SIZE
))
1609 align
= PCPU_MIN_ALLOC_SIZE
;
1611 size
= ALIGN(size
, PCPU_MIN_ALLOC_SIZE
);
1612 bits
= size
>> PCPU_MIN_ALLOC_SHIFT
;
1613 bit_align
= align
>> PCPU_MIN_ALLOC_SHIFT
;
1615 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
||
1616 !is_power_of_2(align
))) {
1617 WARN(do_warn
, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1624 * pcpu_balance_workfn() allocates memory under this mutex,
1625 * and it may wait for memory reclaim. Allow current task
1626 * to become OOM victim, in case of memory pressure.
1628 if (gfp
& __GFP_NOFAIL
)
1629 mutex_lock(&pcpu_alloc_mutex
);
1630 else if (mutex_lock_killable(&pcpu_alloc_mutex
))
1634 spin_lock_irqsave(&pcpu_lock
, flags
);
1636 /* serve reserved allocations from the reserved chunk if available */
1637 if (reserved
&& pcpu_reserved_chunk
) {
1638 chunk
= pcpu_reserved_chunk
;
1640 off
= pcpu_find_block_fit(chunk
, bits
, bit_align
, is_atomic
);
1642 err
= "alloc from reserved chunk failed";
1646 off
= pcpu_alloc_area(chunk
, bits
, bit_align
, off
);
1650 err
= "alloc from reserved chunk failed";
1655 /* search through normal chunks */
1656 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
1657 list_for_each_entry_safe(chunk
, next
, &pcpu_slot
[slot
], list
) {
1658 off
= pcpu_find_block_fit(chunk
, bits
, bit_align
,
1661 if (slot
< PCPU_SLOT_FAIL_THRESHOLD
)
1662 pcpu_chunk_move(chunk
, 0);
1666 off
= pcpu_alloc_area(chunk
, bits
, bit_align
, off
);
1673 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1676 * No space left. Create a new chunk. We don't want multiple
1677 * tasks to create chunks simultaneously. Serialize and create iff
1678 * there's still no empty chunk after grabbing the mutex.
1681 err
= "atomic alloc failed, no space left";
1685 if (list_empty(&pcpu_slot
[pcpu_nr_slots
- 1])) {
1686 chunk
= pcpu_create_chunk(pcpu_gfp
);
1688 err
= "failed to allocate new chunk";
1692 spin_lock_irqsave(&pcpu_lock
, flags
);
1693 pcpu_chunk_relocate(chunk
, -1);
1695 spin_lock_irqsave(&pcpu_lock
, flags
);
1701 pcpu_stats_area_alloc(chunk
, size
);
1702 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1704 /* populate if not all pages are already there */
1706 int page_start
, page_end
, rs
, re
;
1708 page_start
= PFN_DOWN(off
);
1709 page_end
= PFN_UP(off
+ size
);
1711 pcpu_for_each_unpop_region(chunk
->populated
, rs
, re
,
1712 page_start
, page_end
) {
1713 WARN_ON(chunk
->immutable
);
1715 ret
= pcpu_populate_chunk(chunk
, rs
, re
, pcpu_gfp
);
1717 spin_lock_irqsave(&pcpu_lock
, flags
);
1719 pcpu_free_area(chunk
, off
);
1720 err
= "failed to populate";
1723 pcpu_chunk_populated(chunk
, rs
, re
);
1724 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1727 mutex_unlock(&pcpu_alloc_mutex
);
1730 if (pcpu_nr_empty_pop_pages
< PCPU_EMPTY_POP_PAGES_LOW
)
1731 pcpu_schedule_balance_work();
1733 /* clear the areas and return address relative to base address */
1734 for_each_possible_cpu(cpu
)
1735 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1737 ptr
= __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1738 kmemleak_alloc_percpu(ptr
, size
, gfp
);
1740 trace_percpu_alloc_percpu(reserved
, is_atomic
, size
, align
,
1741 chunk
->base_addr
, off
, ptr
);
1746 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1748 trace_percpu_alloc_percpu_fail(reserved
, is_atomic
, size
, align
);
1750 if (!is_atomic
&& do_warn
&& warn_limit
) {
1751 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1752 size
, align
, is_atomic
, err
);
1755 pr_info("limit reached, disable warning\n");
1758 /* see the flag handling in pcpu_blance_workfn() */
1759 pcpu_atomic_alloc_failed
= true;
1760 pcpu_schedule_balance_work();
1762 mutex_unlock(&pcpu_alloc_mutex
);
1768 * __alloc_percpu_gfp - allocate dynamic percpu area
1769 * @size: size of area to allocate in bytes
1770 * @align: alignment of area (max PAGE_SIZE)
1771 * @gfp: allocation flags
1773 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1774 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1775 * be called from any context but is a lot more likely to fail. If @gfp
1776 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1777 * allocation requests.
1780 * Percpu pointer to the allocated area on success, NULL on failure.
1782 void __percpu
*__alloc_percpu_gfp(size_t size
, size_t align
, gfp_t gfp
)
1784 return pcpu_alloc(size
, align
, false, gfp
);
1786 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp
);
1789 * __alloc_percpu - allocate dynamic percpu area
1790 * @size: size of area to allocate in bytes
1791 * @align: alignment of area (max PAGE_SIZE)
1793 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1795 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
1797 return pcpu_alloc(size
, align
, false, GFP_KERNEL
);
1799 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1802 * __alloc_reserved_percpu - allocate reserved percpu area
1803 * @size: size of area to allocate in bytes
1804 * @align: alignment of area (max PAGE_SIZE)
1806 * Allocate zero-filled percpu area of @size bytes aligned at @align
1807 * from reserved percpu area if arch has set it up; otherwise,
1808 * allocation is served from the same dynamic area. Might sleep.
1809 * Might trigger writeouts.
1812 * Does GFP_KERNEL allocation.
1815 * Percpu pointer to the allocated area on success, NULL on failure.
1817 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
1819 return pcpu_alloc(size
, align
, true, GFP_KERNEL
);
1823 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1826 * Reclaim all fully free chunks except for the first one. This is also
1827 * responsible for maintaining the pool of empty populated pages. However,
1828 * it is possible that this is called when physical memory is scarce causing
1829 * OOM killer to be triggered. We should avoid doing so until an actual
1830 * allocation causes the failure as it is possible that requests can be
1831 * serviced from already backed regions.
1833 static void pcpu_balance_workfn(struct work_struct
*work
)
1835 /* gfp flags passed to underlying allocators */
1836 const gfp_t gfp
= GFP_KERNEL
| __GFP_NORETRY
| __GFP_NOWARN
;
1838 struct list_head
*free_head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1839 struct pcpu_chunk
*chunk
, *next
;
1840 int slot
, nr_to_pop
, ret
;
1843 * There's no reason to keep around multiple unused chunks and VM
1844 * areas can be scarce. Destroy all free chunks except for one.
1846 mutex_lock(&pcpu_alloc_mutex
);
1847 spin_lock_irq(&pcpu_lock
);
1849 list_for_each_entry_safe(chunk
, next
, free_head
, list
) {
1850 WARN_ON(chunk
->immutable
);
1852 /* spare the first one */
1853 if (chunk
== list_first_entry(free_head
, struct pcpu_chunk
, list
))
1856 list_move(&chunk
->list
, &to_free
);
1859 spin_unlock_irq(&pcpu_lock
);
1861 list_for_each_entry_safe(chunk
, next
, &to_free
, list
) {
1864 pcpu_for_each_pop_region(chunk
->populated
, rs
, re
, 0,
1866 pcpu_depopulate_chunk(chunk
, rs
, re
);
1867 spin_lock_irq(&pcpu_lock
);
1868 pcpu_chunk_depopulated(chunk
, rs
, re
);
1869 spin_unlock_irq(&pcpu_lock
);
1871 pcpu_destroy_chunk(chunk
);
1876 * Ensure there are certain number of free populated pages for
1877 * atomic allocs. Fill up from the most packed so that atomic
1878 * allocs don't increase fragmentation. If atomic allocation
1879 * failed previously, always populate the maximum amount. This
1880 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1881 * failing indefinitely; however, large atomic allocs are not
1882 * something we support properly and can be highly unreliable and
1886 if (pcpu_atomic_alloc_failed
) {
1887 nr_to_pop
= PCPU_EMPTY_POP_PAGES_HIGH
;
1888 /* best effort anyway, don't worry about synchronization */
1889 pcpu_atomic_alloc_failed
= false;
1891 nr_to_pop
= clamp(PCPU_EMPTY_POP_PAGES_HIGH
-
1892 pcpu_nr_empty_pop_pages
,
1893 0, PCPU_EMPTY_POP_PAGES_HIGH
);
1896 for (slot
= pcpu_size_to_slot(PAGE_SIZE
); slot
< pcpu_nr_slots
; slot
++) {
1897 int nr_unpop
= 0, rs
, re
;
1902 spin_lock_irq(&pcpu_lock
);
1903 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1904 nr_unpop
= chunk
->nr_pages
- chunk
->nr_populated
;
1908 spin_unlock_irq(&pcpu_lock
);
1913 /* @chunk can't go away while pcpu_alloc_mutex is held */
1914 pcpu_for_each_unpop_region(chunk
->populated
, rs
, re
, 0,
1916 int nr
= min(re
- rs
, nr_to_pop
);
1918 ret
= pcpu_populate_chunk(chunk
, rs
, rs
+ nr
, gfp
);
1921 spin_lock_irq(&pcpu_lock
);
1922 pcpu_chunk_populated(chunk
, rs
, rs
+ nr
);
1923 spin_unlock_irq(&pcpu_lock
);
1934 /* ran out of chunks to populate, create a new one and retry */
1935 chunk
= pcpu_create_chunk(gfp
);
1937 spin_lock_irq(&pcpu_lock
);
1938 pcpu_chunk_relocate(chunk
, -1);
1939 spin_unlock_irq(&pcpu_lock
);
1944 mutex_unlock(&pcpu_alloc_mutex
);
1948 * free_percpu - free percpu area
1949 * @ptr: pointer to area to free
1951 * Free percpu area @ptr.
1954 * Can be called from atomic context.
1956 void free_percpu(void __percpu
*ptr
)
1959 struct pcpu_chunk
*chunk
;
1960 unsigned long flags
;
1962 bool need_balance
= false;
1967 kmemleak_free_percpu(ptr
);
1969 addr
= __pcpu_ptr_to_addr(ptr
);
1971 spin_lock_irqsave(&pcpu_lock
, flags
);
1973 chunk
= pcpu_chunk_addr_search(addr
);
1974 off
= addr
- chunk
->base_addr
;
1976 pcpu_free_area(chunk
, off
);
1978 /* if there are more than one fully free chunks, wake up grim reaper */
1979 if (chunk
->free_bytes
== pcpu_unit_size
) {
1980 struct pcpu_chunk
*pos
;
1982 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1984 need_balance
= true;
1989 trace_percpu_free_percpu(chunk
->base_addr
, off
, ptr
);
1991 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1994 pcpu_schedule_balance_work();
1996 EXPORT_SYMBOL_GPL(free_percpu
);
1998 bool __is_kernel_percpu_address(unsigned long addr
, unsigned long *can_addr
)
2001 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
2002 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
2005 for_each_possible_cpu(cpu
) {
2006 void *start
= per_cpu_ptr(base
, cpu
);
2007 void *va
= (void *)addr
;
2009 if (va
>= start
&& va
< start
+ static_size
) {
2011 *can_addr
= (unsigned long) (va
- start
);
2012 *can_addr
+= (unsigned long)
2013 per_cpu_ptr(base
, get_boot_cpu_id());
2019 /* on UP, can't distinguish from other static vars, always false */
2024 * is_kernel_percpu_address - test whether address is from static percpu area
2025 * @addr: address to test
2027 * Test whether @addr belongs to in-kernel static percpu area. Module
2028 * static percpu areas are not considered. For those, use
2029 * is_module_percpu_address().
2032 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2034 bool is_kernel_percpu_address(unsigned long addr
)
2036 return __is_kernel_percpu_address(addr
, NULL
);
2040 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2041 * @addr: the address to be converted to physical address
2043 * Given @addr which is dereferenceable address obtained via one of
2044 * percpu access macros, this function translates it into its physical
2045 * address. The caller is responsible for ensuring @addr stays valid
2046 * until this function finishes.
2048 * percpu allocator has special setup for the first chunk, which currently
2049 * supports either embedding in linear address space or vmalloc mapping,
2050 * and, from the second one, the backing allocator (currently either vm or
2051 * km) provides translation.
2053 * The addr can be translated simply without checking if it falls into the
2054 * first chunk. But the current code reflects better how percpu allocator
2055 * actually works, and the verification can discover both bugs in percpu
2056 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2060 * The physical address for @addr.
2062 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
2064 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
2065 bool in_first_chunk
= false;
2066 unsigned long first_low
, first_high
;
2070 * The following test on unit_low/high isn't strictly
2071 * necessary but will speed up lookups of addresses which
2072 * aren't in the first chunk.
2074 * The address check is against full chunk sizes. pcpu_base_addr
2075 * points to the beginning of the first chunk including the
2076 * static region. Assumes good intent as the first chunk may
2077 * not be full (ie. < pcpu_unit_pages in size).
2079 first_low
= (unsigned long)pcpu_base_addr
+
2080 pcpu_unit_page_offset(pcpu_low_unit_cpu
, 0);
2081 first_high
= (unsigned long)pcpu_base_addr
+
2082 pcpu_unit_page_offset(pcpu_high_unit_cpu
, pcpu_unit_pages
);
2083 if ((unsigned long)addr
>= first_low
&&
2084 (unsigned long)addr
< first_high
) {
2085 for_each_possible_cpu(cpu
) {
2086 void *start
= per_cpu_ptr(base
, cpu
);
2088 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
2089 in_first_chunk
= true;
2095 if (in_first_chunk
) {
2096 if (!is_vmalloc_addr(addr
))
2099 return page_to_phys(vmalloc_to_page(addr
)) +
2100 offset_in_page(addr
);
2102 return page_to_phys(pcpu_addr_to_page(addr
)) +
2103 offset_in_page(addr
);
2107 * pcpu_alloc_alloc_info - allocate percpu allocation info
2108 * @nr_groups: the number of groups
2109 * @nr_units: the number of units
2111 * Allocate ai which is large enough for @nr_groups groups containing
2112 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2113 * cpu_map array which is long enough for @nr_units and filled with
2114 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2115 * pointer of other groups.
2118 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2121 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
2124 struct pcpu_alloc_info
*ai
;
2125 size_t base_size
, ai_size
;
2129 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
2130 __alignof__(ai
->groups
[0].cpu_map
[0]));
2131 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
2133 ptr
= memblock_alloc(PFN_ALIGN(ai_size
), PAGE_SIZE
);
2139 ai
->groups
[0].cpu_map
= ptr
;
2141 for (unit
= 0; unit
< nr_units
; unit
++)
2142 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
2144 ai
->nr_groups
= nr_groups
;
2145 ai
->__ai_size
= PFN_ALIGN(ai_size
);
2151 * pcpu_free_alloc_info - free percpu allocation info
2152 * @ai: pcpu_alloc_info to free
2154 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2156 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
2158 memblock_free_early(__pa(ai
), ai
->__ai_size
);
2162 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2164 * @ai: allocation info to dump
2166 * Print out information about @ai using loglevel @lvl.
2168 static void pcpu_dump_alloc_info(const char *lvl
,
2169 const struct pcpu_alloc_info
*ai
)
2171 int group_width
= 1, cpu_width
= 1, width
;
2172 char empty_str
[] = "--------";
2173 int alloc
= 0, alloc_end
= 0;
2175 int upa
, apl
; /* units per alloc, allocs per line */
2181 v
= num_possible_cpus();
2184 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
2186 upa
= ai
->alloc_size
/ ai
->unit_size
;
2187 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
2188 apl
= rounddown_pow_of_two(max(60 / width
, 1));
2190 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2191 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
2192 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
2194 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2195 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2196 int unit
= 0, unit_end
= 0;
2198 BUG_ON(gi
->nr_units
% upa
);
2199 for (alloc_end
+= gi
->nr_units
/ upa
;
2200 alloc
< alloc_end
; alloc
++) {
2201 if (!(alloc
% apl
)) {
2203 printk("%spcpu-alloc: ", lvl
);
2205 pr_cont("[%0*d] ", group_width
, group
);
2207 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
2208 if (gi
->cpu_map
[unit
] != NR_CPUS
)
2210 cpu_width
, gi
->cpu_map
[unit
]);
2212 pr_cont("%s ", empty_str
);
2219 * pcpu_setup_first_chunk - initialize the first percpu chunk
2220 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2221 * @base_addr: mapped address
2223 * Initialize the first percpu chunk which contains the kernel static
2224 * perpcu area. This function is to be called from arch percpu area
2227 * @ai contains all information necessary to initialize the first
2228 * chunk and prime the dynamic percpu allocator.
2230 * @ai->static_size is the size of static percpu area.
2232 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2233 * reserve after the static area in the first chunk. This reserves
2234 * the first chunk such that it's available only through reserved
2235 * percpu allocation. This is primarily used to serve module percpu
2236 * static areas on architectures where the addressing model has
2237 * limited offset range for symbol relocations to guarantee module
2238 * percpu symbols fall inside the relocatable range.
2240 * @ai->dyn_size determines the number of bytes available for dynamic
2241 * allocation in the first chunk. The area between @ai->static_size +
2242 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2244 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2245 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2248 * @ai->atom_size is the allocation atom size and used as alignment
2251 * @ai->alloc_size is the allocation size and always multiple of
2252 * @ai->atom_size. This is larger than @ai->atom_size if
2253 * @ai->unit_size is larger than @ai->atom_size.
2255 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2256 * percpu areas. Units which should be colocated are put into the
2257 * same group. Dynamic VM areas will be allocated according to these
2258 * groupings. If @ai->nr_groups is zero, a single group containing
2259 * all units is assumed.
2261 * The caller should have mapped the first chunk at @base_addr and
2262 * copied static data to each unit.
2264 * The first chunk will always contain a static and a dynamic region.
2265 * However, the static region is not managed by any chunk. If the first
2266 * chunk also contains a reserved region, it is served by two chunks -
2267 * one for the reserved region and one for the dynamic region. They
2268 * share the same vm, but use offset regions in the area allocation map.
2269 * The chunk serving the dynamic region is circulated in the chunk slots
2270 * and available for dynamic allocation like any other chunk.
2273 * 0 on success, -errno on failure.
2275 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
2278 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
2279 size_t static_size
, dyn_size
;
2280 struct pcpu_chunk
*chunk
;
2281 unsigned long *group_offsets
;
2282 size_t *group_sizes
;
2283 unsigned long *unit_off
;
2288 unsigned long tmp_addr
;
2291 #define PCPU_SETUP_BUG_ON(cond) do { \
2292 if (unlikely(cond)) { \
2293 pr_emerg("failed to initialize, %s\n", #cond); \
2294 pr_emerg("cpu_possible_mask=%*pb\n", \
2295 cpumask_pr_args(cpu_possible_mask)); \
2296 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2302 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
2304 PCPU_SETUP_BUG_ON(!ai
->static_size
);
2305 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start
));
2307 PCPU_SETUP_BUG_ON(!base_addr
);
2308 PCPU_SETUP_BUG_ON(offset_in_page(base_addr
));
2309 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
2310 PCPU_SETUP_BUG_ON(offset_in_page(ai
->unit_size
));
2311 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
2312 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai
->unit_size
, PCPU_BITMAP_BLOCK_SIZE
));
2313 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
2314 PCPU_SETUP_BUG_ON(!ai
->dyn_size
);
2315 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai
->reserved_size
, PCPU_MIN_ALLOC_SIZE
));
2316 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE
, PAGE_SIZE
) ||
2317 IS_ALIGNED(PAGE_SIZE
, PCPU_BITMAP_BLOCK_SIZE
)));
2318 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
2320 /* process group information and build config tables accordingly */
2321 alloc_size
= ai
->nr_groups
* sizeof(group_offsets
[0]);
2322 group_offsets
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
2324 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2327 alloc_size
= ai
->nr_groups
* sizeof(group_sizes
[0]);
2328 group_sizes
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
2330 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2333 alloc_size
= nr_cpu_ids
* sizeof(unit_map
[0]);
2334 unit_map
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
2336 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2339 alloc_size
= nr_cpu_ids
* sizeof(unit_off
[0]);
2340 unit_off
= memblock_alloc(alloc_size
, SMP_CACHE_BYTES
);
2342 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2345 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
2346 unit_map
[cpu
] = UINT_MAX
;
2348 pcpu_low_unit_cpu
= NR_CPUS
;
2349 pcpu_high_unit_cpu
= NR_CPUS
;
2351 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
2352 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2354 group_offsets
[group
] = gi
->base_offset
;
2355 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
2357 for (i
= 0; i
< gi
->nr_units
; i
++) {
2358 cpu
= gi
->cpu_map
[i
];
2362 PCPU_SETUP_BUG_ON(cpu
>= nr_cpu_ids
);
2363 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
2364 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
2366 unit_map
[cpu
] = unit
+ i
;
2367 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
2369 /* determine low/high unit_cpu */
2370 if (pcpu_low_unit_cpu
== NR_CPUS
||
2371 unit_off
[cpu
] < unit_off
[pcpu_low_unit_cpu
])
2372 pcpu_low_unit_cpu
= cpu
;
2373 if (pcpu_high_unit_cpu
== NR_CPUS
||
2374 unit_off
[cpu
] > unit_off
[pcpu_high_unit_cpu
])
2375 pcpu_high_unit_cpu
= cpu
;
2378 pcpu_nr_units
= unit
;
2380 for_each_possible_cpu(cpu
)
2381 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
2383 /* we're done parsing the input, undefine BUG macro and dump config */
2384 #undef PCPU_SETUP_BUG_ON
2385 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
2387 pcpu_nr_groups
= ai
->nr_groups
;
2388 pcpu_group_offsets
= group_offsets
;
2389 pcpu_group_sizes
= group_sizes
;
2390 pcpu_unit_map
= unit_map
;
2391 pcpu_unit_offsets
= unit_off
;
2393 /* determine basic parameters */
2394 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2395 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
2396 pcpu_atom_size
= ai
->atom_size
;
2397 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
2398 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
2400 pcpu_stats_save_ai(ai
);
2403 * Allocate chunk slots. The additional last slot is for
2406 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
2407 pcpu_slot
= memblock_alloc(pcpu_nr_slots
* sizeof(pcpu_slot
[0]),
2410 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2411 pcpu_nr_slots
* sizeof(pcpu_slot
[0]));
2412 for (i
= 0; i
< pcpu_nr_slots
; i
++)
2413 INIT_LIST_HEAD(&pcpu_slot
[i
]);
2416 * The end of the static region needs to be aligned with the
2417 * minimum allocation size as this offsets the reserved and
2418 * dynamic region. The first chunk ends page aligned by
2419 * expanding the dynamic region, therefore the dynamic region
2420 * can be shrunk to compensate while still staying above the
2423 static_size
= ALIGN(ai
->static_size
, PCPU_MIN_ALLOC_SIZE
);
2424 dyn_size
= ai
->dyn_size
- (static_size
- ai
->static_size
);
2427 * Initialize first chunk.
2428 * If the reserved_size is non-zero, this initializes the reserved
2429 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2430 * and the dynamic region is initialized here. The first chunk,
2431 * pcpu_first_chunk, will always point to the chunk that serves
2432 * the dynamic region.
2434 tmp_addr
= (unsigned long)base_addr
+ static_size
;
2435 map_size
= ai
->reserved_size
?: dyn_size
;
2436 chunk
= pcpu_alloc_first_chunk(tmp_addr
, map_size
);
2438 /* init dynamic chunk if necessary */
2439 if (ai
->reserved_size
) {
2440 pcpu_reserved_chunk
= chunk
;
2442 tmp_addr
= (unsigned long)base_addr
+ static_size
+
2444 map_size
= dyn_size
;
2445 chunk
= pcpu_alloc_first_chunk(tmp_addr
, map_size
);
2448 /* link the first chunk in */
2449 pcpu_first_chunk
= chunk
;
2450 pcpu_nr_empty_pop_pages
= pcpu_first_chunk
->nr_empty_pop_pages
;
2451 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
2453 /* include all regions of the first chunk */
2454 pcpu_nr_populated
+= PFN_DOWN(size_sum
);
2456 pcpu_stats_chunk_alloc();
2457 trace_percpu_create_chunk(base_addr
);
2460 pcpu_base_addr
= base_addr
;
2466 const char * const pcpu_fc_names
[PCPU_FC_NR
] __initconst
= {
2467 [PCPU_FC_AUTO
] = "auto",
2468 [PCPU_FC_EMBED
] = "embed",
2469 [PCPU_FC_PAGE
] = "page",
2472 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
2474 static int __init
percpu_alloc_setup(char *str
)
2481 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2482 else if (!strcmp(str
, "embed"))
2483 pcpu_chosen_fc
= PCPU_FC_EMBED
;
2485 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2486 else if (!strcmp(str
, "page"))
2487 pcpu_chosen_fc
= PCPU_FC_PAGE
;
2490 pr_warn("unknown allocator %s specified\n", str
);
2494 early_param("percpu_alloc", percpu_alloc_setup
);
2497 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2498 * Build it if needed by the arch config or the generic setup is going
2501 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2502 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2503 #define BUILD_EMBED_FIRST_CHUNK
2506 /* build pcpu_page_first_chunk() iff needed by the arch config */
2507 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2508 #define BUILD_PAGE_FIRST_CHUNK
2511 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2512 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2514 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2515 * @reserved_size: the size of reserved percpu area in bytes
2516 * @dyn_size: minimum free size for dynamic allocation in bytes
2517 * @atom_size: allocation atom size
2518 * @cpu_distance_fn: callback to determine distance between cpus, optional
2520 * This function determines grouping of units, their mappings to cpus
2521 * and other parameters considering needed percpu size, allocation
2522 * atom size and distances between CPUs.
2524 * Groups are always multiples of atom size and CPUs which are of
2525 * LOCAL_DISTANCE both ways are grouped together and share space for
2526 * units in the same group. The returned configuration is guaranteed
2527 * to have CPUs on different nodes on different groups and >=75% usage
2528 * of allocated virtual address space.
2531 * On success, pointer to the new allocation_info is returned. On
2532 * failure, ERR_PTR value is returned.
2534 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
2535 size_t reserved_size
, size_t dyn_size
,
2537 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
2539 static int group_map
[NR_CPUS
] __initdata
;
2540 static int group_cnt
[NR_CPUS
] __initdata
;
2541 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
2542 int nr_groups
= 1, nr_units
= 0;
2543 size_t size_sum
, min_unit_size
, alloc_size
;
2544 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
2545 int last_allocs
, group
, unit
;
2546 unsigned int cpu
, tcpu
;
2547 struct pcpu_alloc_info
*ai
;
2548 unsigned int *cpu_map
;
2550 /* this function may be called multiple times */
2551 memset(group_map
, 0, sizeof(group_map
));
2552 memset(group_cnt
, 0, sizeof(group_cnt
));
2554 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2555 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
2556 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
2557 dyn_size
= size_sum
- static_size
- reserved_size
;
2560 * Determine min_unit_size, alloc_size and max_upa such that
2561 * alloc_size is multiple of atom_size and is the smallest
2562 * which can accommodate 4k aligned segments which are equal to
2563 * or larger than min_unit_size.
2565 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
2567 /* determine the maximum # of units that can fit in an allocation */
2568 alloc_size
= roundup(min_unit_size
, atom_size
);
2569 upa
= alloc_size
/ min_unit_size
;
2570 while (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
2574 /* group cpus according to their proximity */
2575 for_each_possible_cpu(cpu
) {
2578 for_each_possible_cpu(tcpu
) {
2581 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
2582 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
2583 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
2585 nr_groups
= max(nr_groups
, group
+ 1);
2589 group_map
[cpu
] = group
;
2594 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2595 * Expand the unit_size until we use >= 75% of the units allocated.
2596 * Related to atom_size, which could be much larger than the unit_size.
2598 last_allocs
= INT_MAX
;
2599 for (upa
= max_upa
; upa
; upa
--) {
2600 int allocs
= 0, wasted
= 0;
2602 if (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
2605 for (group
= 0; group
< nr_groups
; group
++) {
2606 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
2607 allocs
+= this_allocs
;
2608 wasted
+= this_allocs
* upa
- group_cnt
[group
];
2612 * Don't accept if wastage is over 1/3. The
2613 * greater-than comparison ensures upa==1 always
2614 * passes the following check.
2616 if (wasted
> num_possible_cpus() / 3)
2619 /* and then don't consume more memory */
2620 if (allocs
> last_allocs
)
2622 last_allocs
= allocs
;
2627 /* allocate and fill alloc_info */
2628 for (group
= 0; group
< nr_groups
; group
++)
2629 nr_units
+= roundup(group_cnt
[group
], upa
);
2631 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
2633 return ERR_PTR(-ENOMEM
);
2634 cpu_map
= ai
->groups
[0].cpu_map
;
2636 for (group
= 0; group
< nr_groups
; group
++) {
2637 ai
->groups
[group
].cpu_map
= cpu_map
;
2638 cpu_map
+= roundup(group_cnt
[group
], upa
);
2641 ai
->static_size
= static_size
;
2642 ai
->reserved_size
= reserved_size
;
2643 ai
->dyn_size
= dyn_size
;
2644 ai
->unit_size
= alloc_size
/ upa
;
2645 ai
->atom_size
= atom_size
;
2646 ai
->alloc_size
= alloc_size
;
2648 for (group
= 0, unit
= 0; group
< nr_groups
; group
++) {
2649 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2652 * Initialize base_offset as if all groups are located
2653 * back-to-back. The caller should update this to
2654 * reflect actual allocation.
2656 gi
->base_offset
= unit
* ai
->unit_size
;
2658 for_each_possible_cpu(cpu
)
2659 if (group_map
[cpu
] == group
)
2660 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
2661 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
2662 unit
+= gi
->nr_units
;
2664 BUG_ON(unit
!= nr_units
);
2668 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2670 #if defined(BUILD_EMBED_FIRST_CHUNK)
2672 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2673 * @reserved_size: the size of reserved percpu area in bytes
2674 * @dyn_size: minimum free size for dynamic allocation in bytes
2675 * @atom_size: allocation atom size
2676 * @cpu_distance_fn: callback to determine distance between cpus, optional
2677 * @alloc_fn: function to allocate percpu page
2678 * @free_fn: function to free percpu page
2680 * This is a helper to ease setting up embedded first percpu chunk and
2681 * can be called where pcpu_setup_first_chunk() is expected.
2683 * If this function is used to setup the first chunk, it is allocated
2684 * by calling @alloc_fn and used as-is without being mapped into
2685 * vmalloc area. Allocations are always whole multiples of @atom_size
2686 * aligned to @atom_size.
2688 * This enables the first chunk to piggy back on the linear physical
2689 * mapping which often uses larger page size. Please note that this
2690 * can result in very sparse cpu->unit mapping on NUMA machines thus
2691 * requiring large vmalloc address space. Don't use this allocator if
2692 * vmalloc space is not orders of magnitude larger than distances
2693 * between node memory addresses (ie. 32bit NUMA machines).
2695 * @dyn_size specifies the minimum dynamic area size.
2697 * If the needed size is smaller than the minimum or specified unit
2698 * size, the leftover is returned using @free_fn.
2701 * 0 on success, -errno on failure.
2703 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
2705 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
2706 pcpu_fc_alloc_fn_t alloc_fn
,
2707 pcpu_fc_free_fn_t free_fn
)
2709 void *base
= (void *)ULONG_MAX
;
2710 void **areas
= NULL
;
2711 struct pcpu_alloc_info
*ai
;
2712 size_t size_sum
, areas_size
;
2713 unsigned long max_distance
;
2714 int group
, i
, highest_group
, rc
;
2716 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
2721 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
2722 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
2724 areas
= memblock_alloc(areas_size
, SMP_CACHE_BYTES
);
2730 /* allocate, copy and determine base address & max_distance */
2732 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2733 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2734 unsigned int cpu
= NR_CPUS
;
2737 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
2738 cpu
= gi
->cpu_map
[i
];
2739 BUG_ON(cpu
== NR_CPUS
);
2741 /* allocate space for the whole group */
2742 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
2745 goto out_free_areas
;
2747 /* kmemleak tracks the percpu allocations separately */
2751 base
= min(ptr
, base
);
2752 if (ptr
> areas
[highest_group
])
2753 highest_group
= group
;
2755 max_distance
= areas
[highest_group
] - base
;
2756 max_distance
+= ai
->unit_size
* ai
->groups
[highest_group
].nr_units
;
2758 /* warn if maximum distance is further than 75% of vmalloc space */
2759 if (max_distance
> VMALLOC_TOTAL
* 3 / 4) {
2760 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2761 max_distance
, VMALLOC_TOTAL
);
2762 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2763 /* and fail if we have fallback */
2765 goto out_free_areas
;
2770 * Copy data and free unused parts. This should happen after all
2771 * allocations are complete; otherwise, we may end up with
2772 * overlapping groups.
2774 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2775 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2776 void *ptr
= areas
[group
];
2778 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
2779 if (gi
->cpu_map
[i
] == NR_CPUS
) {
2780 /* unused unit, free whole */
2781 free_fn(ptr
, ai
->unit_size
);
2784 /* copy and return the unused part */
2785 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
2786 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2790 /* base address is now known, determine group base offsets */
2791 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2792 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2795 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2796 PFN_DOWN(size_sum
), ai
->static_size
, ai
->reserved_size
,
2797 ai
->dyn_size
, ai
->unit_size
);
2799 rc
= pcpu_setup_first_chunk(ai
, base
);
2803 for (group
= 0; group
< ai
->nr_groups
; group
++)
2805 free_fn(areas
[group
],
2806 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2808 pcpu_free_alloc_info(ai
);
2810 memblock_free_early(__pa(areas
), areas_size
);
2813 #endif /* BUILD_EMBED_FIRST_CHUNK */
2815 #ifdef BUILD_PAGE_FIRST_CHUNK
2817 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2818 * @reserved_size: the size of reserved percpu area in bytes
2819 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2820 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2821 * @populate_pte_fn: function to populate pte
2823 * This is a helper to ease setting up page-remapped first percpu
2824 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2826 * This is the basic allocator. Static percpu area is allocated
2827 * page-by-page into vmalloc area.
2830 * 0 on success, -errno on failure.
2832 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2833 pcpu_fc_alloc_fn_t alloc_fn
,
2834 pcpu_fc_free_fn_t free_fn
,
2835 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2837 static struct vm_struct vm
;
2838 struct pcpu_alloc_info
*ai
;
2842 struct page
**pages
;
2847 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2849 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
2852 BUG_ON(ai
->nr_groups
!= 1);
2853 upa
= ai
->alloc_size
/ai
->unit_size
;
2854 nr_g0_units
= roundup(num_possible_cpus(), upa
);
2855 if (WARN_ON(ai
->groups
[0].nr_units
!= nr_g0_units
)) {
2856 pcpu_free_alloc_info(ai
);
2860 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2862 /* unaligned allocations can't be freed, round up to page size */
2863 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2865 pages
= memblock_alloc(pages_size
, SMP_CACHE_BYTES
);
2867 panic("%s: Failed to allocate %zu bytes\n", __func__
,
2870 /* allocate pages */
2872 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2873 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2874 for (i
= 0; i
< unit_pages
; i
++) {
2877 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2879 pr_warn("failed to allocate %s page for cpu%u\n",
2883 /* kmemleak tracks the percpu allocations separately */
2885 pages
[j
++] = virt_to_page(ptr
);
2889 /* allocate vm area, map the pages and copy static data */
2890 vm
.flags
= VM_ALLOC
;
2891 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2892 vm_area_register_early(&vm
, PAGE_SIZE
);
2894 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2895 unsigned long unit_addr
=
2896 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2898 for (i
= 0; i
< unit_pages
; i
++)
2899 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2901 /* pte already populated, the following shouldn't fail */
2902 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2905 panic("failed to map percpu area, err=%d\n", rc
);
2908 * FIXME: Archs with virtual cache should flush local
2909 * cache for the linear mapping here - something
2910 * equivalent to flush_cache_vmap() on the local cpu.
2911 * flush_cache_vmap() can't be used as most supporting
2912 * data structures are not set up yet.
2915 /* copy static data */
2916 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2919 /* we're ready, commit */
2920 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2921 unit_pages
, psize_str
, ai
->static_size
,
2922 ai
->reserved_size
, ai
->dyn_size
);
2924 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2929 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2932 memblock_free_early(__pa(pages
), pages_size
);
2933 pcpu_free_alloc_info(ai
);
2936 #endif /* BUILD_PAGE_FIRST_CHUNK */
2938 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2940 * Generic SMP percpu area setup.
2942 * The embedding helper is used because its behavior closely resembles
2943 * the original non-dynamic generic percpu area setup. This is
2944 * important because many archs have addressing restrictions and might
2945 * fail if the percpu area is located far away from the previous
2946 * location. As an added bonus, in non-NUMA cases, embedding is
2947 * generally a good idea TLB-wise because percpu area can piggy back
2948 * on the physical linear memory mapping which uses large page
2949 * mappings on applicable archs.
2951 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2952 EXPORT_SYMBOL(__per_cpu_offset
);
2954 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2957 return memblock_alloc_from(size
, align
, __pa(MAX_DMA_ADDRESS
));
2960 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2962 memblock_free_early(__pa(ptr
), size
);
2965 void __init
setup_per_cpu_areas(void)
2967 unsigned long delta
;
2972 * Always reserve area for module percpu variables. That's
2973 * what the legacy allocator did.
2975 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2976 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2977 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2979 panic("Failed to initialize percpu areas.");
2981 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2982 for_each_possible_cpu(cpu
)
2983 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2985 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2987 #else /* CONFIG_SMP */
2990 * UP percpu area setup.
2992 * UP always uses km-based percpu allocator with identity mapping.
2993 * Static percpu variables are indistinguishable from the usual static
2994 * variables and don't require any special preparation.
2996 void __init
setup_per_cpu_areas(void)
2998 const size_t unit_size
=
2999 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
3000 PERCPU_DYNAMIC_RESERVE
));
3001 struct pcpu_alloc_info
*ai
;
3004 ai
= pcpu_alloc_alloc_info(1, 1);
3005 fc
= memblock_alloc_from(unit_size
, PAGE_SIZE
, __pa(MAX_DMA_ADDRESS
));
3007 panic("Failed to allocate memory for percpu areas.");
3008 /* kmemleak tracks the percpu allocations separately */
3011 ai
->dyn_size
= unit_size
;
3012 ai
->unit_size
= unit_size
;
3013 ai
->atom_size
= unit_size
;
3014 ai
->alloc_size
= unit_size
;
3015 ai
->groups
[0].nr_units
= 1;
3016 ai
->groups
[0].cpu_map
[0] = 0;
3018 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
3019 panic("Failed to initialize percpu areas.");
3020 pcpu_free_alloc_info(ai
);
3023 #endif /* CONFIG_SMP */
3026 * pcpu_nr_pages - calculate total number of populated backing pages
3028 * This reflects the number of pages populated to back chunks. Metadata is
3029 * excluded in the number exposed in meminfo as the number of backing pages
3030 * scales with the number of cpus and can quickly outweigh the memory used for
3031 * metadata. It also keeps this calculation nice and simple.
3034 * Total number of populated backing pages in use by the allocator.
3036 unsigned long pcpu_nr_pages(void)
3038 return pcpu_nr_populated
* pcpu_nr_units
;
3042 * Percpu allocator is initialized early during boot when neither slab or
3043 * workqueue is available. Plug async management until everything is up
3046 static int __init
percpu_enable_async(void)
3048 pcpu_async_enabled
= true;
3051 subsys_initcall(percpu_enable_async
);