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 * This file is released under the GPLv2.
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
46 * To use this allocator, arch code should do the followings.
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
56 #include <linux/bitmap.h>
57 #include <linux/bootmem.h>
58 #include <linux/err.h>
59 #include <linux/list.h>
60 #include <linux/log2.h>
62 #include <linux/module.h>
63 #include <linux/mutex.h>
64 #include <linux/percpu.h>
65 #include <linux/pfn.h>
66 #include <linux/slab.h>
67 #include <linux/spinlock.h>
68 #include <linux/vmalloc.h>
69 #include <linux/workqueue.h>
70 #include <linux/kmemleak.h>
72 #include <asm/cacheflush.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
77 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79 #define PCPU_ATOMIC_MAP_MARGIN_LOW 32
80 #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
81 #define PCPU_EMPTY_POP_PAGES_LOW 2
82 #define PCPU_EMPTY_POP_PAGES_HIGH 4
85 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
86 #ifndef __addr_to_pcpu_ptr
87 #define __addr_to_pcpu_ptr(addr) \
88 (void __percpu *)((unsigned long)(addr) - \
89 (unsigned long)pcpu_base_addr + \
90 (unsigned long)__per_cpu_start)
92 #ifndef __pcpu_ptr_to_addr
93 #define __pcpu_ptr_to_addr(ptr) \
94 (void __force *)((unsigned long)(ptr) + \
95 (unsigned long)pcpu_base_addr - \
96 (unsigned long)__per_cpu_start)
98 #else /* CONFIG_SMP */
99 /* on UP, it's always identity mapped */
100 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
101 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
102 #endif /* CONFIG_SMP */
105 struct list_head list
; /* linked to pcpu_slot lists */
106 int free_size
; /* free bytes in the chunk */
107 int contig_hint
; /* max contiguous size hint */
108 void *base_addr
; /* base address of this chunk */
110 int map_used
; /* # of map entries used before the sentry */
111 int map_alloc
; /* # of map entries allocated */
112 int *map
; /* allocation map */
113 struct list_head map_extend_list
;/* on pcpu_map_extend_chunks */
115 void *data
; /* chunk data */
116 int first_free
; /* no free below this */
117 bool immutable
; /* no [de]population allowed */
118 int nr_populated
; /* # of populated pages */
119 unsigned long populated
[]; /* populated bitmap */
122 static int pcpu_unit_pages __read_mostly
;
123 static int pcpu_unit_size __read_mostly
;
124 static int pcpu_nr_units __read_mostly
;
125 static int pcpu_atom_size __read_mostly
;
126 static int pcpu_nr_slots __read_mostly
;
127 static size_t pcpu_chunk_struct_size __read_mostly
;
129 /* cpus with the lowest and highest unit addresses */
130 static unsigned int pcpu_low_unit_cpu __read_mostly
;
131 static unsigned int pcpu_high_unit_cpu __read_mostly
;
133 /* the address of the first chunk which starts with the kernel static area */
134 void *pcpu_base_addr __read_mostly
;
135 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
137 static const int *pcpu_unit_map __read_mostly
; /* cpu -> unit */
138 const unsigned long *pcpu_unit_offsets __read_mostly
; /* cpu -> unit offset */
140 /* group information, used for vm allocation */
141 static int pcpu_nr_groups __read_mostly
;
142 static const unsigned long *pcpu_group_offsets __read_mostly
;
143 static const size_t *pcpu_group_sizes __read_mostly
;
146 * The first chunk which always exists. Note that unlike other
147 * chunks, this one can be allocated and mapped in several different
148 * ways and thus often doesn't live in the vmalloc area.
150 static struct pcpu_chunk
*pcpu_first_chunk
;
153 * Optional reserved chunk. This chunk reserves part of the first
154 * chunk and serves it for reserved allocations. The amount of
155 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
156 * area doesn't exist, the following variables contain NULL and 0
159 static struct pcpu_chunk
*pcpu_reserved_chunk
;
160 static int pcpu_reserved_chunk_limit
;
162 static DEFINE_SPINLOCK(pcpu_lock
); /* all internal data structures */
163 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* chunk create/destroy, [de]pop, map ext */
165 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
167 /* chunks which need their map areas extended, protected by pcpu_lock */
168 static LIST_HEAD(pcpu_map_extend_chunks
);
171 * The number of empty populated pages, protected by pcpu_lock. The
172 * reserved chunk doesn't contribute to the count.
174 static int pcpu_nr_empty_pop_pages
;
177 * Balance work is used to populate or destroy chunks asynchronously. We
178 * try to keep the number of populated free pages between
179 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
182 static void pcpu_balance_workfn(struct work_struct
*work
);
183 static DECLARE_WORK(pcpu_balance_work
, pcpu_balance_workfn
);
184 static bool pcpu_async_enabled __read_mostly
;
185 static bool pcpu_atomic_alloc_failed
;
187 static void pcpu_schedule_balance_work(void)
189 if (pcpu_async_enabled
)
190 schedule_work(&pcpu_balance_work
);
193 static bool pcpu_addr_in_first_chunk(void *addr
)
195 void *first_start
= pcpu_first_chunk
->base_addr
;
197 return addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
;
200 static bool pcpu_addr_in_reserved_chunk(void *addr
)
202 void *first_start
= pcpu_first_chunk
->base_addr
;
204 return addr
>= first_start
&&
205 addr
< first_start
+ pcpu_reserved_chunk_limit
;
208 static int __pcpu_size_to_slot(int size
)
210 int highbit
= fls(size
); /* size is in bytes */
211 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
214 static int pcpu_size_to_slot(int size
)
216 if (size
== pcpu_unit_size
)
217 return pcpu_nr_slots
- 1;
218 return __pcpu_size_to_slot(size
);
221 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
223 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
226 return pcpu_size_to_slot(chunk
->free_size
);
229 /* set the pointer to a chunk in a page struct */
230 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
232 page
->index
= (unsigned long)pcpu
;
235 /* obtain pointer to a chunk from a page struct */
236 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
238 return (struct pcpu_chunk
*)page
->index
;
241 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
243 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
246 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
247 unsigned int cpu
, int page_idx
)
249 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
250 (page_idx
<< PAGE_SHIFT
);
253 static void __maybe_unused
pcpu_next_unpop(struct pcpu_chunk
*chunk
,
254 int *rs
, int *re
, int end
)
256 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
257 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
260 static void __maybe_unused
pcpu_next_pop(struct pcpu_chunk
*chunk
,
261 int *rs
, int *re
, int end
)
263 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
264 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
268 * (Un)populated page region iterators. Iterate over (un)populated
269 * page regions between @start and @end in @chunk. @rs and @re should
270 * be integer variables and will be set to start and end page index of
271 * the current region.
273 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
274 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
276 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
278 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
279 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
281 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
284 * pcpu_mem_zalloc - allocate memory
285 * @size: bytes to allocate
287 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
288 * kzalloc() is used; otherwise, vzalloc() is used. The returned
289 * memory is always zeroed.
292 * Does GFP_KERNEL allocation.
295 * Pointer to the allocated area on success, NULL on failure.
297 static void *pcpu_mem_zalloc(size_t size
)
299 if (WARN_ON_ONCE(!slab_is_available()))
302 if (size
<= PAGE_SIZE
)
303 return kzalloc(size
, GFP_KERNEL
);
305 return vzalloc(size
);
309 * pcpu_mem_free - free memory
310 * @ptr: memory to free
311 * @size: size of the area
313 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
315 static void pcpu_mem_free(void *ptr
, size_t size
)
317 if (size
<= PAGE_SIZE
)
324 * pcpu_count_occupied_pages - count the number of pages an area occupies
325 * @chunk: chunk of interest
326 * @i: index of the area in question
328 * Count the number of pages chunk's @i'th area occupies. When the area's
329 * start and/or end address isn't aligned to page boundary, the straddled
330 * page is included in the count iff the rest of the page is free.
332 static int pcpu_count_occupied_pages(struct pcpu_chunk
*chunk
, int i
)
334 int off
= chunk
->map
[i
] & ~1;
335 int end
= chunk
->map
[i
+ 1] & ~1;
337 if (!PAGE_ALIGNED(off
) && i
> 0) {
338 int prev
= chunk
->map
[i
- 1];
340 if (!(prev
& 1) && prev
<= round_down(off
, PAGE_SIZE
))
341 off
= round_down(off
, PAGE_SIZE
);
344 if (!PAGE_ALIGNED(end
) && i
+ 1 < chunk
->map_used
) {
345 int next
= chunk
->map
[i
+ 1];
346 int nend
= chunk
->map
[i
+ 2] & ~1;
348 if (!(next
& 1) && nend
>= round_up(end
, PAGE_SIZE
))
349 end
= round_up(end
, PAGE_SIZE
);
352 return max_t(int, PFN_DOWN(end
) - PFN_UP(off
), 0);
356 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
357 * @chunk: chunk of interest
358 * @oslot: the previous slot it was on
360 * This function is called after an allocation or free changed @chunk.
361 * New slot according to the changed state is determined and @chunk is
362 * moved to the slot. Note that the reserved chunk is never put on
368 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
370 int nslot
= pcpu_chunk_slot(chunk
);
372 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
374 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
376 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
381 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
382 * @chunk: chunk of interest
383 * @is_atomic: the allocation context
385 * Determine whether area map of @chunk needs to be extended. If
386 * @is_atomic, only the amount necessary for a new allocation is
387 * considered; however, async extension is scheduled if the left amount is
388 * low. If !@is_atomic, it aims for more empty space. Combined, this
389 * ensures that the map is likely to have enough available space to
390 * accomodate atomic allocations which can't extend maps directly.
396 * New target map allocation length if extension is necessary, 0
399 static int pcpu_need_to_extend(struct pcpu_chunk
*chunk
, bool is_atomic
)
401 int margin
, new_alloc
;
403 lockdep_assert_held(&pcpu_lock
);
408 if (chunk
->map_alloc
<
409 chunk
->map_used
+ PCPU_ATOMIC_MAP_MARGIN_LOW
) {
410 if (list_empty(&chunk
->map_extend_list
)) {
411 list_add_tail(&chunk
->map_extend_list
,
412 &pcpu_map_extend_chunks
);
413 pcpu_schedule_balance_work();
417 margin
= PCPU_ATOMIC_MAP_MARGIN_HIGH
;
420 if (chunk
->map_alloc
>= chunk
->map_used
+ margin
)
423 new_alloc
= PCPU_DFL_MAP_ALLOC
;
424 while (new_alloc
< chunk
->map_used
+ margin
)
431 * pcpu_extend_area_map - extend area map of a chunk
432 * @chunk: chunk of interest
433 * @new_alloc: new target allocation length of the area map
435 * Extend area map of @chunk to have @new_alloc entries.
438 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
441 * 0 on success, -errno on failure.
443 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
, int new_alloc
)
445 int *old
= NULL
, *new = NULL
;
446 size_t old_size
= 0, new_size
= new_alloc
* sizeof(new[0]);
449 lockdep_assert_held(&pcpu_alloc_mutex
);
451 new = pcpu_mem_zalloc(new_size
);
455 /* acquire pcpu_lock and switch to new area map */
456 spin_lock_irqsave(&pcpu_lock
, flags
);
458 if (new_alloc
<= chunk
->map_alloc
)
461 old_size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
464 memcpy(new, old
, old_size
);
466 chunk
->map_alloc
= new_alloc
;
471 spin_unlock_irqrestore(&pcpu_lock
, flags
);
474 * pcpu_mem_free() might end up calling vfree() which uses
475 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
477 pcpu_mem_free(old
, old_size
);
478 pcpu_mem_free(new, new_size
);
484 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
485 * @chunk: chunk the candidate area belongs to
486 * @off: the offset to the start of the candidate area
487 * @this_size: the size of the candidate area
488 * @size: the size of the target allocation
489 * @align: the alignment of the target allocation
490 * @pop_only: only allocate from already populated region
492 * We're trying to allocate @size bytes aligned at @align. @chunk's area
493 * at @off sized @this_size is a candidate. This function determines
494 * whether the target allocation fits in the candidate area and returns the
495 * number of bytes to pad after @off. If the target area doesn't fit, -1
498 * If @pop_only is %true, this function only considers the already
499 * populated part of the candidate area.
501 static int pcpu_fit_in_area(struct pcpu_chunk
*chunk
, int off
, int this_size
,
502 int size
, int align
, bool pop_only
)
507 int head
= ALIGN(cand_off
, align
) - off
;
508 int page_start
, page_end
, rs
, re
;
510 if (this_size
< head
+ size
)
517 * If the first unpopulated page is beyond the end of the
518 * allocation, the whole allocation is populated;
519 * otherwise, retry from the end of the unpopulated area.
521 page_start
= PFN_DOWN(head
+ off
);
522 page_end
= PFN_UP(head
+ off
+ size
);
525 pcpu_next_unpop(chunk
, &rs
, &re
, PFN_UP(off
+ this_size
));
528 cand_off
= re
* PAGE_SIZE
;
533 * pcpu_alloc_area - allocate area from a pcpu_chunk
534 * @chunk: chunk of interest
535 * @size: wanted size in bytes
536 * @align: wanted align
537 * @pop_only: allocate only from the populated area
538 * @occ_pages_p: out param for the number of pages the area occupies
540 * Try to allocate @size bytes area aligned at @align from @chunk.
541 * Note that this function only allocates the offset. It doesn't
542 * populate or map the area.
544 * @chunk->map must have at least two free slots.
550 * Allocated offset in @chunk on success, -1 if no matching area is
553 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
,
554 bool pop_only
, int *occ_pages_p
)
556 int oslot
= pcpu_chunk_slot(chunk
);
559 bool seen_free
= false;
562 for (i
= chunk
->first_free
, p
= chunk
->map
+ i
; i
< chunk
->map_used
; i
++, p
++) {
570 this_size
= (p
[1] & ~1) - off
;
572 head
= pcpu_fit_in_area(chunk
, off
, this_size
, size
, align
,
576 chunk
->first_free
= i
;
579 max_contig
= max(this_size
, max_contig
);
584 * If head is small or the previous block is free,
585 * merge'em. Note that 'small' is defined as smaller
586 * than sizeof(int), which is very small but isn't too
587 * uncommon for percpu allocations.
589 if (head
&& (head
< sizeof(int) || !(p
[-1] & 1))) {
592 chunk
->free_size
-= head
;
594 max_contig
= max(*p
- p
[-1], max_contig
);
599 /* if tail is small, just keep it around */
600 tail
= this_size
- head
- size
;
601 if (tail
< sizeof(int)) {
603 size
= this_size
- head
;
606 /* split if warranted */
608 int nr_extra
= !!head
+ !!tail
;
610 /* insert new subblocks */
611 memmove(p
+ nr_extra
+ 1, p
+ 1,
612 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
613 chunk
->map_used
+= nr_extra
;
617 chunk
->first_free
= i
;
622 max_contig
= max(head
, max_contig
);
626 max_contig
= max(tail
, max_contig
);
631 chunk
->first_free
= i
+ 1;
633 /* update hint and mark allocated */
634 if (i
+ 1 == chunk
->map_used
)
635 chunk
->contig_hint
= max_contig
; /* fully scanned */
637 chunk
->contig_hint
= max(chunk
->contig_hint
,
640 chunk
->free_size
-= size
;
643 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
644 pcpu_chunk_relocate(chunk
, oslot
);
648 chunk
->contig_hint
= max_contig
; /* fully scanned */
649 pcpu_chunk_relocate(chunk
, oslot
);
651 /* tell the upper layer that this chunk has no matching area */
656 * pcpu_free_area - free area to a pcpu_chunk
657 * @chunk: chunk of interest
658 * @freeme: offset of area to free
659 * @occ_pages_p: out param for the number of pages the area occupies
661 * Free area starting from @freeme to @chunk. Note that this function
662 * only modifies the allocation map. It doesn't depopulate or unmap
668 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
,
671 int oslot
= pcpu_chunk_slot(chunk
);
677 freeme
|= 1; /* we are searching for <given offset, in use> pair */
682 unsigned k
= (i
+ j
) / 2;
686 else if (off
> freeme
)
691 BUG_ON(off
!= freeme
);
693 if (i
< chunk
->first_free
)
694 chunk
->first_free
= i
;
698 chunk
->free_size
+= (p
[1] & ~1) - off
;
700 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
702 /* merge with next? */
705 /* merge with previous? */
706 if (i
> 0 && !(p
[-1] & 1)) {
712 chunk
->map_used
-= to_free
;
713 memmove(p
+ 1, p
+ 1 + to_free
,
714 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
717 chunk
->contig_hint
= max(chunk
->map
[i
+ 1] - chunk
->map
[i
] - 1, chunk
->contig_hint
);
718 pcpu_chunk_relocate(chunk
, oslot
);
721 static struct pcpu_chunk
*pcpu_alloc_chunk(void)
723 struct pcpu_chunk
*chunk
;
725 chunk
= pcpu_mem_zalloc(pcpu_chunk_struct_size
);
729 chunk
->map
= pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC
*
730 sizeof(chunk
->map
[0]));
732 pcpu_mem_free(chunk
, pcpu_chunk_struct_size
);
736 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
738 chunk
->map
[1] = pcpu_unit_size
| 1;
741 INIT_LIST_HEAD(&chunk
->list
);
742 INIT_LIST_HEAD(&chunk
->map_extend_list
);
743 chunk
->free_size
= pcpu_unit_size
;
744 chunk
->contig_hint
= pcpu_unit_size
;
749 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
753 pcpu_mem_free(chunk
->map
, chunk
->map_alloc
* sizeof(chunk
->map
[0]));
754 pcpu_mem_free(chunk
, pcpu_chunk_struct_size
);
758 * pcpu_chunk_populated - post-population bookkeeping
759 * @chunk: pcpu_chunk which got populated
760 * @page_start: the start page
761 * @page_end: the end page
763 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
764 * the bookkeeping information accordingly. Must be called after each
765 * successful population.
767 static void pcpu_chunk_populated(struct pcpu_chunk
*chunk
,
768 int page_start
, int page_end
)
770 int nr
= page_end
- page_start
;
772 lockdep_assert_held(&pcpu_lock
);
774 bitmap_set(chunk
->populated
, page_start
, nr
);
775 chunk
->nr_populated
+= nr
;
776 pcpu_nr_empty_pop_pages
+= nr
;
780 * pcpu_chunk_depopulated - post-depopulation bookkeeping
781 * @chunk: pcpu_chunk which got depopulated
782 * @page_start: the start page
783 * @page_end: the end page
785 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
786 * Update the bookkeeping information accordingly. Must be called after
787 * each successful depopulation.
789 static void pcpu_chunk_depopulated(struct pcpu_chunk
*chunk
,
790 int page_start
, int page_end
)
792 int nr
= page_end
- page_start
;
794 lockdep_assert_held(&pcpu_lock
);
796 bitmap_clear(chunk
->populated
, page_start
, nr
);
797 chunk
->nr_populated
-= nr
;
798 pcpu_nr_empty_pop_pages
-= nr
;
802 * Chunk management implementation.
804 * To allow different implementations, chunk alloc/free and
805 * [de]population are implemented in a separate file which is pulled
806 * into this file and compiled together. The following functions
807 * should be implemented.
809 * pcpu_populate_chunk - populate the specified range of a chunk
810 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
811 * pcpu_create_chunk - create a new chunk
812 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
813 * pcpu_addr_to_page - translate address to physical address
814 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
816 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
817 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
818 static struct pcpu_chunk
*pcpu_create_chunk(void);
819 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
820 static struct page
*pcpu_addr_to_page(void *addr
);
821 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
823 #ifdef CONFIG_NEED_PER_CPU_KM
824 #include "percpu-km.c"
826 #include "percpu-vm.c"
830 * pcpu_chunk_addr_search - determine chunk containing specified address
831 * @addr: address for which the chunk needs to be determined.
834 * The address of the found chunk.
836 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
838 /* is it in the first chunk? */
839 if (pcpu_addr_in_first_chunk(addr
)) {
840 /* is it in the reserved area? */
841 if (pcpu_addr_in_reserved_chunk(addr
))
842 return pcpu_reserved_chunk
;
843 return pcpu_first_chunk
;
847 * The address is relative to unit0 which might be unused and
848 * thus unmapped. Offset the address to the unit space of the
849 * current processor before looking it up in the vmalloc
850 * space. Note that any possible cpu id can be used here, so
851 * there's no need to worry about preemption or cpu hotplug.
853 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
854 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
858 * pcpu_alloc - the percpu allocator
859 * @size: size of area to allocate in bytes
860 * @align: alignment of area (max PAGE_SIZE)
861 * @reserved: allocate from the reserved chunk if available
862 * @gfp: allocation flags
864 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
865 * contain %GFP_KERNEL, the allocation is atomic.
868 * Percpu pointer to the allocated area on success, NULL on failure.
870 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
,
873 static int warn_limit
= 10;
874 struct pcpu_chunk
*chunk
;
876 bool is_atomic
= (gfp
& GFP_KERNEL
) != GFP_KERNEL
;
878 int slot
, off
, new_alloc
, cpu
, ret
;
883 * We want the lowest bit of offset available for in-use/free
884 * indicator, so force >= 16bit alignment and make size even.
886 if (unlikely(align
< 2))
889 size
= ALIGN(size
, 2);
891 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
892 WARN(true, "illegal size (%zu) or align (%zu) for "
893 "percpu allocation\n", size
, align
);
898 mutex_lock(&pcpu_alloc_mutex
);
900 spin_lock_irqsave(&pcpu_lock
, flags
);
902 /* serve reserved allocations from the reserved chunk if available */
903 if (reserved
&& pcpu_reserved_chunk
) {
904 chunk
= pcpu_reserved_chunk
;
906 if (size
> chunk
->contig_hint
) {
907 err
= "alloc from reserved chunk failed";
911 while ((new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
))) {
912 spin_unlock_irqrestore(&pcpu_lock
, flags
);
914 pcpu_extend_area_map(chunk
, new_alloc
) < 0) {
915 err
= "failed to extend area map of reserved chunk";
918 spin_lock_irqsave(&pcpu_lock
, flags
);
921 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
926 err
= "alloc from reserved chunk failed";
931 /* search through normal chunks */
932 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
933 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
934 if (size
> chunk
->contig_hint
)
937 new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
);
941 spin_unlock_irqrestore(&pcpu_lock
, flags
);
942 if (pcpu_extend_area_map(chunk
,
944 err
= "failed to extend area map";
947 spin_lock_irqsave(&pcpu_lock
, flags
);
949 * pcpu_lock has been dropped, need to
950 * restart cpu_slot list walking.
955 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
962 spin_unlock_irqrestore(&pcpu_lock
, flags
);
965 * No space left. Create a new chunk. We don't want multiple
966 * tasks to create chunks simultaneously. Serialize and create iff
967 * there's still no empty chunk after grabbing the mutex.
972 if (list_empty(&pcpu_slot
[pcpu_nr_slots
- 1])) {
973 chunk
= pcpu_create_chunk();
975 err
= "failed to allocate new chunk";
979 spin_lock_irqsave(&pcpu_lock
, flags
);
980 pcpu_chunk_relocate(chunk
, -1);
982 spin_lock_irqsave(&pcpu_lock
, flags
);
988 spin_unlock_irqrestore(&pcpu_lock
, flags
);
990 /* populate if not all pages are already there */
992 int page_start
, page_end
, rs
, re
;
994 page_start
= PFN_DOWN(off
);
995 page_end
= PFN_UP(off
+ size
);
997 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
998 WARN_ON(chunk
->immutable
);
1000 ret
= pcpu_populate_chunk(chunk
, rs
, re
);
1002 spin_lock_irqsave(&pcpu_lock
, flags
);
1004 pcpu_free_area(chunk
, off
, &occ_pages
);
1005 err
= "failed to populate";
1008 pcpu_chunk_populated(chunk
, rs
, re
);
1009 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1012 mutex_unlock(&pcpu_alloc_mutex
);
1015 if (chunk
!= pcpu_reserved_chunk
)
1016 pcpu_nr_empty_pop_pages
-= occ_pages
;
1018 if (pcpu_nr_empty_pop_pages
< PCPU_EMPTY_POP_PAGES_LOW
)
1019 pcpu_schedule_balance_work();
1021 /* clear the areas and return address relative to base address */
1022 for_each_possible_cpu(cpu
)
1023 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1025 ptr
= __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1026 kmemleak_alloc_percpu(ptr
, size
, gfp
);
1030 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1032 if (!is_atomic
&& warn_limit
) {
1033 pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1034 size
, align
, is_atomic
, err
);
1037 pr_info("PERCPU: limit reached, disable warning\n");
1040 /* see the flag handling in pcpu_blance_workfn() */
1041 pcpu_atomic_alloc_failed
= true;
1042 pcpu_schedule_balance_work();
1044 mutex_unlock(&pcpu_alloc_mutex
);
1050 * __alloc_percpu_gfp - allocate dynamic percpu area
1051 * @size: size of area to allocate in bytes
1052 * @align: alignment of area (max PAGE_SIZE)
1053 * @gfp: allocation flags
1055 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1056 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1057 * be called from any context but is a lot more likely to fail.
1060 * Percpu pointer to the allocated area on success, NULL on failure.
1062 void __percpu
*__alloc_percpu_gfp(size_t size
, size_t align
, gfp_t gfp
)
1064 return pcpu_alloc(size
, align
, false, gfp
);
1066 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp
);
1069 * __alloc_percpu - allocate dynamic percpu area
1070 * @size: size of area to allocate in bytes
1071 * @align: alignment of area (max PAGE_SIZE)
1073 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1075 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
1077 return pcpu_alloc(size
, align
, false, GFP_KERNEL
);
1079 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1082 * __alloc_reserved_percpu - allocate reserved percpu area
1083 * @size: size of area to allocate in bytes
1084 * @align: alignment of area (max PAGE_SIZE)
1086 * Allocate zero-filled percpu area of @size bytes aligned at @align
1087 * from reserved percpu area if arch has set it up; otherwise,
1088 * allocation is served from the same dynamic area. Might sleep.
1089 * Might trigger writeouts.
1092 * Does GFP_KERNEL allocation.
1095 * Percpu pointer to the allocated area on success, NULL on failure.
1097 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
1099 return pcpu_alloc(size
, align
, true, GFP_KERNEL
);
1103 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1106 * Reclaim all fully free chunks except for the first one.
1108 static void pcpu_balance_workfn(struct work_struct
*work
)
1111 struct list_head
*free_head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1112 struct pcpu_chunk
*chunk
, *next
;
1113 int slot
, nr_to_pop
, ret
;
1116 * There's no reason to keep around multiple unused chunks and VM
1117 * areas can be scarce. Destroy all free chunks except for one.
1119 mutex_lock(&pcpu_alloc_mutex
);
1120 spin_lock_irq(&pcpu_lock
);
1122 list_for_each_entry_safe(chunk
, next
, free_head
, list
) {
1123 WARN_ON(chunk
->immutable
);
1125 /* spare the first one */
1126 if (chunk
== list_first_entry(free_head
, struct pcpu_chunk
, list
))
1129 list_del_init(&chunk
->map_extend_list
);
1130 list_move(&chunk
->list
, &to_free
);
1133 spin_unlock_irq(&pcpu_lock
);
1135 list_for_each_entry_safe(chunk
, next
, &to_free
, list
) {
1138 pcpu_for_each_pop_region(chunk
, rs
, re
, 0, pcpu_unit_pages
) {
1139 pcpu_depopulate_chunk(chunk
, rs
, re
);
1140 spin_lock_irq(&pcpu_lock
);
1141 pcpu_chunk_depopulated(chunk
, rs
, re
);
1142 spin_unlock_irq(&pcpu_lock
);
1144 pcpu_destroy_chunk(chunk
);
1147 /* service chunks which requested async area map extension */
1151 spin_lock_irq(&pcpu_lock
);
1153 chunk
= list_first_entry_or_null(&pcpu_map_extend_chunks
,
1154 struct pcpu_chunk
, map_extend_list
);
1156 list_del_init(&chunk
->map_extend_list
);
1157 new_alloc
= pcpu_need_to_extend(chunk
, false);
1160 spin_unlock_irq(&pcpu_lock
);
1163 pcpu_extend_area_map(chunk
, new_alloc
);
1167 * Ensure there are certain number of free populated pages for
1168 * atomic allocs. Fill up from the most packed so that atomic
1169 * allocs don't increase fragmentation. If atomic allocation
1170 * failed previously, always populate the maximum amount. This
1171 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1172 * failing indefinitely; however, large atomic allocs are not
1173 * something we support properly and can be highly unreliable and
1177 if (pcpu_atomic_alloc_failed
) {
1178 nr_to_pop
= PCPU_EMPTY_POP_PAGES_HIGH
;
1179 /* best effort anyway, don't worry about synchronization */
1180 pcpu_atomic_alloc_failed
= false;
1182 nr_to_pop
= clamp(PCPU_EMPTY_POP_PAGES_HIGH
-
1183 pcpu_nr_empty_pop_pages
,
1184 0, PCPU_EMPTY_POP_PAGES_HIGH
);
1187 for (slot
= pcpu_size_to_slot(PAGE_SIZE
); slot
< pcpu_nr_slots
; slot
++) {
1188 int nr_unpop
= 0, rs
, re
;
1193 spin_lock_irq(&pcpu_lock
);
1194 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1195 nr_unpop
= pcpu_unit_pages
- chunk
->nr_populated
;
1199 spin_unlock_irq(&pcpu_lock
);
1204 /* @chunk can't go away while pcpu_alloc_mutex is held */
1205 pcpu_for_each_unpop_region(chunk
, rs
, re
, 0, pcpu_unit_pages
) {
1206 int nr
= min(re
- rs
, nr_to_pop
);
1208 ret
= pcpu_populate_chunk(chunk
, rs
, rs
+ nr
);
1211 spin_lock_irq(&pcpu_lock
);
1212 pcpu_chunk_populated(chunk
, rs
, rs
+ nr
);
1213 spin_unlock_irq(&pcpu_lock
);
1224 /* ran out of chunks to populate, create a new one and retry */
1225 chunk
= pcpu_create_chunk();
1227 spin_lock_irq(&pcpu_lock
);
1228 pcpu_chunk_relocate(chunk
, -1);
1229 spin_unlock_irq(&pcpu_lock
);
1234 mutex_unlock(&pcpu_alloc_mutex
);
1238 * free_percpu - free percpu area
1239 * @ptr: pointer to area to free
1241 * Free percpu area @ptr.
1244 * Can be called from atomic context.
1246 void free_percpu(void __percpu
*ptr
)
1249 struct pcpu_chunk
*chunk
;
1250 unsigned long flags
;
1256 kmemleak_free_percpu(ptr
);
1258 addr
= __pcpu_ptr_to_addr(ptr
);
1260 spin_lock_irqsave(&pcpu_lock
, flags
);
1262 chunk
= pcpu_chunk_addr_search(addr
);
1263 off
= addr
- chunk
->base_addr
;
1265 pcpu_free_area(chunk
, off
, &occ_pages
);
1267 if (chunk
!= pcpu_reserved_chunk
)
1268 pcpu_nr_empty_pop_pages
+= occ_pages
;
1270 /* if there are more than one fully free chunks, wake up grim reaper */
1271 if (chunk
->free_size
== pcpu_unit_size
) {
1272 struct pcpu_chunk
*pos
;
1274 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1276 pcpu_schedule_balance_work();
1281 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1283 EXPORT_SYMBOL_GPL(free_percpu
);
1286 * is_kernel_percpu_address - test whether address is from static percpu area
1287 * @addr: address to test
1289 * Test whether @addr belongs to in-kernel static percpu area. Module
1290 * static percpu areas are not considered. For those, use
1291 * is_module_percpu_address().
1294 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1296 bool is_kernel_percpu_address(unsigned long addr
)
1299 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1300 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1303 for_each_possible_cpu(cpu
) {
1304 void *start
= per_cpu_ptr(base
, cpu
);
1306 if ((void *)addr
>= start
&& (void *)addr
< start
+ static_size
)
1310 /* on UP, can't distinguish from other static vars, always false */
1315 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1316 * @addr: the address to be converted to physical address
1318 * Given @addr which is dereferenceable address obtained via one of
1319 * percpu access macros, this function translates it into its physical
1320 * address. The caller is responsible for ensuring @addr stays valid
1321 * until this function finishes.
1323 * percpu allocator has special setup for the first chunk, which currently
1324 * supports either embedding in linear address space or vmalloc mapping,
1325 * and, from the second one, the backing allocator (currently either vm or
1326 * km) provides translation.
1328 * The addr can be translated simply without checking if it falls into the
1329 * first chunk. But the current code reflects better how percpu allocator
1330 * actually works, and the verification can discover both bugs in percpu
1331 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1335 * The physical address for @addr.
1337 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
1339 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1340 bool in_first_chunk
= false;
1341 unsigned long first_low
, first_high
;
1345 * The following test on unit_low/high isn't strictly
1346 * necessary but will speed up lookups of addresses which
1347 * aren't in the first chunk.
1349 first_low
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_low_unit_cpu
, 0);
1350 first_high
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_high_unit_cpu
,
1352 if ((unsigned long)addr
>= first_low
&&
1353 (unsigned long)addr
< first_high
) {
1354 for_each_possible_cpu(cpu
) {
1355 void *start
= per_cpu_ptr(base
, cpu
);
1357 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
1358 in_first_chunk
= true;
1364 if (in_first_chunk
) {
1365 if (!is_vmalloc_addr(addr
))
1368 return page_to_phys(vmalloc_to_page(addr
)) +
1369 offset_in_page(addr
);
1371 return page_to_phys(pcpu_addr_to_page(addr
)) +
1372 offset_in_page(addr
);
1376 * pcpu_alloc_alloc_info - allocate percpu allocation info
1377 * @nr_groups: the number of groups
1378 * @nr_units: the number of units
1380 * Allocate ai which is large enough for @nr_groups groups containing
1381 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1382 * cpu_map array which is long enough for @nr_units and filled with
1383 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1384 * pointer of other groups.
1387 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1390 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1393 struct pcpu_alloc_info
*ai
;
1394 size_t base_size
, ai_size
;
1398 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1399 __alignof__(ai
->groups
[0].cpu_map
[0]));
1400 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1402 ptr
= memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size
), 0);
1408 ai
->groups
[0].cpu_map
= ptr
;
1410 for (unit
= 0; unit
< nr_units
; unit
++)
1411 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1413 ai
->nr_groups
= nr_groups
;
1414 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1420 * pcpu_free_alloc_info - free percpu allocation info
1421 * @ai: pcpu_alloc_info to free
1423 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1425 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1427 memblock_free_early(__pa(ai
), ai
->__ai_size
);
1431 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1433 * @ai: allocation info to dump
1435 * Print out information about @ai using loglevel @lvl.
1437 static void pcpu_dump_alloc_info(const char *lvl
,
1438 const struct pcpu_alloc_info
*ai
)
1440 int group_width
= 1, cpu_width
= 1, width
;
1441 char empty_str
[] = "--------";
1442 int alloc
= 0, alloc_end
= 0;
1444 int upa
, apl
; /* units per alloc, allocs per line */
1450 v
= num_possible_cpus();
1453 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1455 upa
= ai
->alloc_size
/ ai
->unit_size
;
1456 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1457 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1459 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1460 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1461 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1463 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1464 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1465 int unit
= 0, unit_end
= 0;
1467 BUG_ON(gi
->nr_units
% upa
);
1468 for (alloc_end
+= gi
->nr_units
/ upa
;
1469 alloc
< alloc_end
; alloc
++) {
1470 if (!(alloc
% apl
)) {
1471 printk(KERN_CONT
"\n");
1472 printk("%spcpu-alloc: ", lvl
);
1474 printk(KERN_CONT
"[%0*d] ", group_width
, group
);
1476 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1477 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1478 printk(KERN_CONT
"%0*d ", cpu_width
,
1481 printk(KERN_CONT
"%s ", empty_str
);
1484 printk(KERN_CONT
"\n");
1488 * pcpu_setup_first_chunk - initialize the first percpu chunk
1489 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1490 * @base_addr: mapped address
1492 * Initialize the first percpu chunk which contains the kernel static
1493 * perpcu area. This function is to be called from arch percpu area
1496 * @ai contains all information necessary to initialize the first
1497 * chunk and prime the dynamic percpu allocator.
1499 * @ai->static_size is the size of static percpu area.
1501 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1502 * reserve after the static area in the first chunk. This reserves
1503 * the first chunk such that it's available only through reserved
1504 * percpu allocation. This is primarily used to serve module percpu
1505 * static areas on architectures where the addressing model has
1506 * limited offset range for symbol relocations to guarantee module
1507 * percpu symbols fall inside the relocatable range.
1509 * @ai->dyn_size determines the number of bytes available for dynamic
1510 * allocation in the first chunk. The area between @ai->static_size +
1511 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1513 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1514 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1517 * @ai->atom_size is the allocation atom size and used as alignment
1520 * @ai->alloc_size is the allocation size and always multiple of
1521 * @ai->atom_size. This is larger than @ai->atom_size if
1522 * @ai->unit_size is larger than @ai->atom_size.
1524 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1525 * percpu areas. Units which should be colocated are put into the
1526 * same group. Dynamic VM areas will be allocated according to these
1527 * groupings. If @ai->nr_groups is zero, a single group containing
1528 * all units is assumed.
1530 * The caller should have mapped the first chunk at @base_addr and
1531 * copied static data to each unit.
1533 * If the first chunk ends up with both reserved and dynamic areas, it
1534 * is served by two chunks - one to serve the core static and reserved
1535 * areas and the other for the dynamic area. They share the same vm
1536 * and page map but uses different area allocation map to stay away
1537 * from each other. The latter chunk is circulated in the chunk slots
1538 * and available for dynamic allocation like any other chunks.
1541 * 0 on success, -errno on failure.
1543 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1546 static int smap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1547 static int dmap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1548 size_t dyn_size
= ai
->dyn_size
;
1549 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1550 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1551 unsigned long *group_offsets
;
1552 size_t *group_sizes
;
1553 unsigned long *unit_off
;
1558 #define PCPU_SETUP_BUG_ON(cond) do { \
1559 if (unlikely(cond)) { \
1560 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1561 pr_emerg("PERCPU: cpu_possible_mask=%*pb\n", \
1562 cpumask_pr_args(cpu_possible_mask)); \
1563 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1569 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1571 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1572 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start
));
1574 PCPU_SETUP_BUG_ON(!base_addr
);
1575 PCPU_SETUP_BUG_ON(offset_in_page(base_addr
));
1576 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1577 PCPU_SETUP_BUG_ON(offset_in_page(ai
->unit_size
));
1578 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1579 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
1580 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
1582 /* process group information and build config tables accordingly */
1583 group_offsets
= memblock_virt_alloc(ai
->nr_groups
*
1584 sizeof(group_offsets
[0]), 0);
1585 group_sizes
= memblock_virt_alloc(ai
->nr_groups
*
1586 sizeof(group_sizes
[0]), 0);
1587 unit_map
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_map
[0]), 0);
1588 unit_off
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_off
[0]), 0);
1590 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1591 unit_map
[cpu
] = UINT_MAX
;
1593 pcpu_low_unit_cpu
= NR_CPUS
;
1594 pcpu_high_unit_cpu
= NR_CPUS
;
1596 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1597 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1599 group_offsets
[group
] = gi
->base_offset
;
1600 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1602 for (i
= 0; i
< gi
->nr_units
; i
++) {
1603 cpu
= gi
->cpu_map
[i
];
1607 PCPU_SETUP_BUG_ON(cpu
>= nr_cpu_ids
);
1608 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1609 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1611 unit_map
[cpu
] = unit
+ i
;
1612 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1614 /* determine low/high unit_cpu */
1615 if (pcpu_low_unit_cpu
== NR_CPUS
||
1616 unit_off
[cpu
] < unit_off
[pcpu_low_unit_cpu
])
1617 pcpu_low_unit_cpu
= cpu
;
1618 if (pcpu_high_unit_cpu
== NR_CPUS
||
1619 unit_off
[cpu
] > unit_off
[pcpu_high_unit_cpu
])
1620 pcpu_high_unit_cpu
= cpu
;
1623 pcpu_nr_units
= unit
;
1625 for_each_possible_cpu(cpu
)
1626 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1628 /* we're done parsing the input, undefine BUG macro and dump config */
1629 #undef PCPU_SETUP_BUG_ON
1630 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
1632 pcpu_nr_groups
= ai
->nr_groups
;
1633 pcpu_group_offsets
= group_offsets
;
1634 pcpu_group_sizes
= group_sizes
;
1635 pcpu_unit_map
= unit_map
;
1636 pcpu_unit_offsets
= unit_off
;
1638 /* determine basic parameters */
1639 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1640 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1641 pcpu_atom_size
= ai
->atom_size
;
1642 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1643 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1646 * Allocate chunk slots. The additional last slot is for
1649 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1650 pcpu_slot
= memblock_virt_alloc(
1651 pcpu_nr_slots
* sizeof(pcpu_slot
[0]), 0);
1652 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1653 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1656 * Initialize static chunk. If reserved_size is zero, the
1657 * static chunk covers static area + dynamic allocation area
1658 * in the first chunk. If reserved_size is not zero, it
1659 * covers static area + reserved area (mostly used for module
1660 * static percpu allocation).
1662 schunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1663 INIT_LIST_HEAD(&schunk
->list
);
1664 INIT_LIST_HEAD(&schunk
->map_extend_list
);
1665 schunk
->base_addr
= base_addr
;
1667 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1668 schunk
->immutable
= true;
1669 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1670 schunk
->nr_populated
= pcpu_unit_pages
;
1672 if (ai
->reserved_size
) {
1673 schunk
->free_size
= ai
->reserved_size
;
1674 pcpu_reserved_chunk
= schunk
;
1675 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1677 schunk
->free_size
= dyn_size
;
1678 dyn_size
= 0; /* dynamic area covered */
1680 schunk
->contig_hint
= schunk
->free_size
;
1683 schunk
->map
[1] = ai
->static_size
;
1684 schunk
->map_used
= 1;
1685 if (schunk
->free_size
)
1686 schunk
->map
[++schunk
->map_used
] = ai
->static_size
+ schunk
->free_size
;
1687 schunk
->map
[schunk
->map_used
] |= 1;
1689 /* init dynamic chunk if necessary */
1691 dchunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1692 INIT_LIST_HEAD(&dchunk
->list
);
1693 INIT_LIST_HEAD(&dchunk
->map_extend_list
);
1694 dchunk
->base_addr
= base_addr
;
1696 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1697 dchunk
->immutable
= true;
1698 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1699 dchunk
->nr_populated
= pcpu_unit_pages
;
1701 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1703 dchunk
->map
[1] = pcpu_reserved_chunk_limit
;
1704 dchunk
->map
[2] = (pcpu_reserved_chunk_limit
+ dchunk
->free_size
) | 1;
1705 dchunk
->map_used
= 2;
1708 /* link the first chunk in */
1709 pcpu_first_chunk
= dchunk
?: schunk
;
1710 pcpu_nr_empty_pop_pages
+=
1711 pcpu_count_occupied_pages(pcpu_first_chunk
, 1);
1712 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1715 pcpu_base_addr
= base_addr
;
1721 const char * const pcpu_fc_names
[PCPU_FC_NR
] __initconst
= {
1722 [PCPU_FC_AUTO
] = "auto",
1723 [PCPU_FC_EMBED
] = "embed",
1724 [PCPU_FC_PAGE
] = "page",
1727 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1729 static int __init
percpu_alloc_setup(char *str
)
1736 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1737 else if (!strcmp(str
, "embed"))
1738 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1740 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1741 else if (!strcmp(str
, "page"))
1742 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1745 pr_warning("PERCPU: unknown allocator %s specified\n", str
);
1749 early_param("percpu_alloc", percpu_alloc_setup
);
1752 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1753 * Build it if needed by the arch config or the generic setup is going
1756 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1757 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1758 #define BUILD_EMBED_FIRST_CHUNK
1761 /* build pcpu_page_first_chunk() iff needed by the arch config */
1762 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1763 #define BUILD_PAGE_FIRST_CHUNK
1766 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1767 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1769 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1770 * @reserved_size: the size of reserved percpu area in bytes
1771 * @dyn_size: minimum free size for dynamic allocation in bytes
1772 * @atom_size: allocation atom size
1773 * @cpu_distance_fn: callback to determine distance between cpus, optional
1775 * This function determines grouping of units, their mappings to cpus
1776 * and other parameters considering needed percpu size, allocation
1777 * atom size and distances between CPUs.
1779 * Groups are always multiples of atom size and CPUs which are of
1780 * LOCAL_DISTANCE both ways are grouped together and share space for
1781 * units in the same group. The returned configuration is guaranteed
1782 * to have CPUs on different nodes on different groups and >=75% usage
1783 * of allocated virtual address space.
1786 * On success, pointer to the new allocation_info is returned. On
1787 * failure, ERR_PTR value is returned.
1789 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1790 size_t reserved_size
, size_t dyn_size
,
1792 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1794 static int group_map
[NR_CPUS
] __initdata
;
1795 static int group_cnt
[NR_CPUS
] __initdata
;
1796 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1797 int nr_groups
= 1, nr_units
= 0;
1798 size_t size_sum
, min_unit_size
, alloc_size
;
1799 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1800 int last_allocs
, group
, unit
;
1801 unsigned int cpu
, tcpu
;
1802 struct pcpu_alloc_info
*ai
;
1803 unsigned int *cpu_map
;
1805 /* this function may be called multiple times */
1806 memset(group_map
, 0, sizeof(group_map
));
1807 memset(group_cnt
, 0, sizeof(group_cnt
));
1809 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1810 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1811 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
1812 dyn_size
= size_sum
- static_size
- reserved_size
;
1815 * Determine min_unit_size, alloc_size and max_upa such that
1816 * alloc_size is multiple of atom_size and is the smallest
1817 * which can accommodate 4k aligned segments which are equal to
1818 * or larger than min_unit_size.
1820 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1822 alloc_size
= roundup(min_unit_size
, atom_size
);
1823 upa
= alloc_size
/ min_unit_size
;
1824 while (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1828 /* group cpus according to their proximity */
1829 for_each_possible_cpu(cpu
) {
1832 for_each_possible_cpu(tcpu
) {
1835 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1836 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1837 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1839 nr_groups
= max(nr_groups
, group
+ 1);
1843 group_map
[cpu
] = group
;
1848 * Expand unit size until address space usage goes over 75%
1849 * and then as much as possible without using more address
1852 last_allocs
= INT_MAX
;
1853 for (upa
= max_upa
; upa
; upa
--) {
1854 int allocs
= 0, wasted
= 0;
1856 if (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1859 for (group
= 0; group
< nr_groups
; group
++) {
1860 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1861 allocs
+= this_allocs
;
1862 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1866 * Don't accept if wastage is over 1/3. The
1867 * greater-than comparison ensures upa==1 always
1868 * passes the following check.
1870 if (wasted
> num_possible_cpus() / 3)
1873 /* and then don't consume more memory */
1874 if (allocs
> last_allocs
)
1876 last_allocs
= allocs
;
1881 /* allocate and fill alloc_info */
1882 for (group
= 0; group
< nr_groups
; group
++)
1883 nr_units
+= roundup(group_cnt
[group
], upa
);
1885 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1887 return ERR_PTR(-ENOMEM
);
1888 cpu_map
= ai
->groups
[0].cpu_map
;
1890 for (group
= 0; group
< nr_groups
; group
++) {
1891 ai
->groups
[group
].cpu_map
= cpu_map
;
1892 cpu_map
+= roundup(group_cnt
[group
], upa
);
1895 ai
->static_size
= static_size
;
1896 ai
->reserved_size
= reserved_size
;
1897 ai
->dyn_size
= dyn_size
;
1898 ai
->unit_size
= alloc_size
/ upa
;
1899 ai
->atom_size
= atom_size
;
1900 ai
->alloc_size
= alloc_size
;
1902 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1903 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1906 * Initialize base_offset as if all groups are located
1907 * back-to-back. The caller should update this to
1908 * reflect actual allocation.
1910 gi
->base_offset
= unit
* ai
->unit_size
;
1912 for_each_possible_cpu(cpu
)
1913 if (group_map
[cpu
] == group
)
1914 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1915 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1916 unit
+= gi
->nr_units
;
1918 BUG_ON(unit
!= nr_units
);
1922 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1924 #if defined(BUILD_EMBED_FIRST_CHUNK)
1926 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1927 * @reserved_size: the size of reserved percpu area in bytes
1928 * @dyn_size: minimum free size for dynamic allocation in bytes
1929 * @atom_size: allocation atom size
1930 * @cpu_distance_fn: callback to determine distance between cpus, optional
1931 * @alloc_fn: function to allocate percpu page
1932 * @free_fn: function to free percpu page
1934 * This is a helper to ease setting up embedded first percpu chunk and
1935 * can be called where pcpu_setup_first_chunk() is expected.
1937 * If this function is used to setup the first chunk, it is allocated
1938 * by calling @alloc_fn and used as-is without being mapped into
1939 * vmalloc area. Allocations are always whole multiples of @atom_size
1940 * aligned to @atom_size.
1942 * This enables the first chunk to piggy back on the linear physical
1943 * mapping which often uses larger page size. Please note that this
1944 * can result in very sparse cpu->unit mapping on NUMA machines thus
1945 * requiring large vmalloc address space. Don't use this allocator if
1946 * vmalloc space is not orders of magnitude larger than distances
1947 * between node memory addresses (ie. 32bit NUMA machines).
1949 * @dyn_size specifies the minimum dynamic area size.
1951 * If the needed size is smaller than the minimum or specified unit
1952 * size, the leftover is returned using @free_fn.
1955 * 0 on success, -errno on failure.
1957 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
1959 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1960 pcpu_fc_alloc_fn_t alloc_fn
,
1961 pcpu_fc_free_fn_t free_fn
)
1963 void *base
= (void *)ULONG_MAX
;
1964 void **areas
= NULL
;
1965 struct pcpu_alloc_info
*ai
;
1966 size_t size_sum
, areas_size
, max_distance
;
1969 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1974 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1975 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1977 areas
= memblock_virt_alloc_nopanic(areas_size
, 0);
1983 /* allocate, copy and determine base address */
1984 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1985 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1986 unsigned int cpu
= NR_CPUS
;
1989 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
1990 cpu
= gi
->cpu_map
[i
];
1991 BUG_ON(cpu
== NR_CPUS
);
1993 /* allocate space for the whole group */
1994 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
1997 goto out_free_areas
;
1999 /* kmemleak tracks the percpu allocations separately */
2003 base
= min(ptr
, base
);
2007 * Copy data and free unused parts. This should happen after all
2008 * allocations are complete; otherwise, we may end up with
2009 * overlapping groups.
2011 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2012 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2013 void *ptr
= areas
[group
];
2015 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
2016 if (gi
->cpu_map
[i
] == NR_CPUS
) {
2017 /* unused unit, free whole */
2018 free_fn(ptr
, ai
->unit_size
);
2021 /* copy and return the unused part */
2022 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
2023 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2027 /* base address is now known, determine group base offsets */
2029 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2030 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2031 max_distance
= max_t(size_t, max_distance
,
2032 ai
->groups
[group
].base_offset
);
2034 max_distance
+= ai
->unit_size
;
2036 /* warn if maximum distance is further than 75% of vmalloc space */
2037 if (max_distance
> VMALLOC_TOTAL
* 3 / 4) {
2038 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
2039 "space 0x%lx\n", max_distance
,
2041 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2042 /* and fail if we have fallback */
2048 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2049 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
2050 ai
->dyn_size
, ai
->unit_size
);
2052 rc
= pcpu_setup_first_chunk(ai
, base
);
2056 for (group
= 0; group
< ai
->nr_groups
; group
++)
2058 free_fn(areas
[group
],
2059 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2061 pcpu_free_alloc_info(ai
);
2063 memblock_free_early(__pa(areas
), areas_size
);
2066 #endif /* BUILD_EMBED_FIRST_CHUNK */
2068 #ifdef BUILD_PAGE_FIRST_CHUNK
2070 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2071 * @reserved_size: the size of reserved percpu area in bytes
2072 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2073 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2074 * @populate_pte_fn: function to populate pte
2076 * This is a helper to ease setting up page-remapped first percpu
2077 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2079 * This is the basic allocator. Static percpu area is allocated
2080 * page-by-page into vmalloc area.
2083 * 0 on success, -errno on failure.
2085 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2086 pcpu_fc_alloc_fn_t alloc_fn
,
2087 pcpu_fc_free_fn_t free_fn
,
2088 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2090 static struct vm_struct vm
;
2091 struct pcpu_alloc_info
*ai
;
2095 struct page
**pages
;
2098 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2100 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
2103 BUG_ON(ai
->nr_groups
!= 1);
2104 BUG_ON(ai
->groups
[0].nr_units
!= num_possible_cpus());
2106 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2108 /* unaligned allocations can't be freed, round up to page size */
2109 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2111 pages
= memblock_virt_alloc(pages_size
, 0);
2113 /* allocate pages */
2115 for (unit
= 0; unit
< num_possible_cpus(); unit
++)
2116 for (i
= 0; i
< unit_pages
; i
++) {
2117 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2120 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2122 pr_warning("PERCPU: failed to allocate %s page "
2123 "for cpu%u\n", psize_str
, cpu
);
2126 /* kmemleak tracks the percpu allocations separately */
2128 pages
[j
++] = virt_to_page(ptr
);
2131 /* allocate vm area, map the pages and copy static data */
2132 vm
.flags
= VM_ALLOC
;
2133 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2134 vm_area_register_early(&vm
, PAGE_SIZE
);
2136 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2137 unsigned long unit_addr
=
2138 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2140 for (i
= 0; i
< unit_pages
; i
++)
2141 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2143 /* pte already populated, the following shouldn't fail */
2144 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2147 panic("failed to map percpu area, err=%d\n", rc
);
2150 * FIXME: Archs with virtual cache should flush local
2151 * cache for the linear mapping here - something
2152 * equivalent to flush_cache_vmap() on the local cpu.
2153 * flush_cache_vmap() can't be used as most supporting
2154 * data structures are not set up yet.
2157 /* copy static data */
2158 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2161 /* we're ready, commit */
2162 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2163 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2164 ai
->reserved_size
, ai
->dyn_size
);
2166 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2171 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2174 memblock_free_early(__pa(pages
), pages_size
);
2175 pcpu_free_alloc_info(ai
);
2178 #endif /* BUILD_PAGE_FIRST_CHUNK */
2180 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2182 * Generic SMP percpu area setup.
2184 * The embedding helper is used because its behavior closely resembles
2185 * the original non-dynamic generic percpu area setup. This is
2186 * important because many archs have addressing restrictions and might
2187 * fail if the percpu area is located far away from the previous
2188 * location. As an added bonus, in non-NUMA cases, embedding is
2189 * generally a good idea TLB-wise because percpu area can piggy back
2190 * on the physical linear memory mapping which uses large page
2191 * mappings on applicable archs.
2193 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2194 EXPORT_SYMBOL(__per_cpu_offset
);
2196 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2199 return memblock_virt_alloc_from_nopanic(
2200 size
, align
, __pa(MAX_DMA_ADDRESS
));
2203 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2205 memblock_free_early(__pa(ptr
), size
);
2208 void __init
setup_per_cpu_areas(void)
2210 unsigned long delta
;
2215 * Always reserve area for module percpu variables. That's
2216 * what the legacy allocator did.
2218 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2219 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2220 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2222 panic("Failed to initialize percpu areas.");
2224 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2225 for_each_possible_cpu(cpu
)
2226 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2228 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2230 #else /* CONFIG_SMP */
2233 * UP percpu area setup.
2235 * UP always uses km-based percpu allocator with identity mapping.
2236 * Static percpu variables are indistinguishable from the usual static
2237 * variables and don't require any special preparation.
2239 void __init
setup_per_cpu_areas(void)
2241 const size_t unit_size
=
2242 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
2243 PERCPU_DYNAMIC_RESERVE
));
2244 struct pcpu_alloc_info
*ai
;
2247 ai
= pcpu_alloc_alloc_info(1, 1);
2248 fc
= memblock_virt_alloc_from_nopanic(unit_size
,
2250 __pa(MAX_DMA_ADDRESS
));
2252 panic("Failed to allocate memory for percpu areas.");
2253 /* kmemleak tracks the percpu allocations separately */
2256 ai
->dyn_size
= unit_size
;
2257 ai
->unit_size
= unit_size
;
2258 ai
->atom_size
= unit_size
;
2259 ai
->alloc_size
= unit_size
;
2260 ai
->groups
[0].nr_units
= 1;
2261 ai
->groups
[0].cpu_map
[0] = 0;
2263 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
2264 panic("Failed to initialize percpu areas.");
2267 #endif /* CONFIG_SMP */
2270 * First and reserved chunks are initialized with temporary allocation
2271 * map in initdata so that they can be used before slab is online.
2272 * This function is called after slab is brought up and replaces those
2273 * with properly allocated maps.
2275 void __init
percpu_init_late(void)
2277 struct pcpu_chunk
*target_chunks
[] =
2278 { pcpu_first_chunk
, pcpu_reserved_chunk
, NULL
};
2279 struct pcpu_chunk
*chunk
;
2280 unsigned long flags
;
2283 for (i
= 0; (chunk
= target_chunks
[i
]); i
++) {
2285 const size_t size
= PERCPU_DYNAMIC_EARLY_SLOTS
* sizeof(map
[0]);
2287 BUILD_BUG_ON(size
> PAGE_SIZE
);
2289 map
= pcpu_mem_zalloc(size
);
2292 spin_lock_irqsave(&pcpu_lock
, flags
);
2293 memcpy(map
, chunk
->map
, size
);
2295 spin_unlock_irqrestore(&pcpu_lock
, flags
);
2300 * Percpu allocator is initialized early during boot when neither slab or
2301 * workqueue is available. Plug async management until everything is up
2304 static int __init
percpu_enable_async(void)
2306 pcpu_async_enabled
= true;
2309 subsys_initcall(percpu_enable_async
);