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 work_struct map_extend_work
;/* async ->map[] extension */
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 */
165 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
168 * The number of empty populated pages, protected by pcpu_lock. The
169 * reserved chunk doesn't contribute to the count.
171 static int pcpu_nr_empty_pop_pages
;
174 * Balance work is used to populate or destroy chunks asynchronously. We
175 * try to keep the number of populated free pages between
176 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
179 static void pcpu_balance_workfn(struct work_struct
*work
);
180 static DECLARE_WORK(pcpu_balance_work
, pcpu_balance_workfn
);
181 static bool pcpu_async_enabled __read_mostly
;
182 static bool pcpu_atomic_alloc_failed
;
184 static void pcpu_schedule_balance_work(void)
186 if (pcpu_async_enabled
)
187 schedule_work(&pcpu_balance_work
);
190 static bool pcpu_addr_in_first_chunk(void *addr
)
192 void *first_start
= pcpu_first_chunk
->base_addr
;
194 return addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
;
197 static bool pcpu_addr_in_reserved_chunk(void *addr
)
199 void *first_start
= pcpu_first_chunk
->base_addr
;
201 return addr
>= first_start
&&
202 addr
< first_start
+ pcpu_reserved_chunk_limit
;
205 static int __pcpu_size_to_slot(int size
)
207 int highbit
= fls(size
); /* size is in bytes */
208 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
211 static int pcpu_size_to_slot(int size
)
213 if (size
== pcpu_unit_size
)
214 return pcpu_nr_slots
- 1;
215 return __pcpu_size_to_slot(size
);
218 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
220 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
223 return pcpu_size_to_slot(chunk
->free_size
);
226 /* set the pointer to a chunk in a page struct */
227 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
229 page
->index
= (unsigned long)pcpu
;
232 /* obtain pointer to a chunk from a page struct */
233 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
235 return (struct pcpu_chunk
*)page
->index
;
238 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
240 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
243 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
244 unsigned int cpu
, int page_idx
)
246 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
247 (page_idx
<< PAGE_SHIFT
);
250 static void __maybe_unused
pcpu_next_unpop(struct pcpu_chunk
*chunk
,
251 int *rs
, int *re
, int end
)
253 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
254 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
257 static void __maybe_unused
pcpu_next_pop(struct pcpu_chunk
*chunk
,
258 int *rs
, int *re
, int end
)
260 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
261 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
265 * (Un)populated page region iterators. Iterate over (un)populated
266 * page regions between @start and @end in @chunk. @rs and @re should
267 * be integer variables and will be set to start and end page index of
268 * the current region.
270 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
271 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
273 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
275 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
276 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
278 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
281 * pcpu_mem_zalloc - allocate memory
282 * @size: bytes to allocate
284 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
285 * kzalloc() is used; otherwise, vzalloc() is used. The returned
286 * memory is always zeroed.
289 * Does GFP_KERNEL allocation.
292 * Pointer to the allocated area on success, NULL on failure.
294 static void *pcpu_mem_zalloc(size_t size
)
296 if (WARN_ON_ONCE(!slab_is_available()))
299 if (size
<= PAGE_SIZE
)
300 return kzalloc(size
, GFP_KERNEL
);
302 return vzalloc(size
);
306 * pcpu_mem_free - free memory
307 * @ptr: memory to free
309 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
311 static void pcpu_mem_free(void *ptr
)
317 * pcpu_count_occupied_pages - count the number of pages an area occupies
318 * @chunk: chunk of interest
319 * @i: index of the area in question
321 * Count the number of pages chunk's @i'th area occupies. When the area's
322 * start and/or end address isn't aligned to page boundary, the straddled
323 * page is included in the count iff the rest of the page is free.
325 static int pcpu_count_occupied_pages(struct pcpu_chunk
*chunk
, int i
)
327 int off
= chunk
->map
[i
] & ~1;
328 int end
= chunk
->map
[i
+ 1] & ~1;
330 if (!PAGE_ALIGNED(off
) && i
> 0) {
331 int prev
= chunk
->map
[i
- 1];
333 if (!(prev
& 1) && prev
<= round_down(off
, PAGE_SIZE
))
334 off
= round_down(off
, PAGE_SIZE
);
337 if (!PAGE_ALIGNED(end
) && i
+ 1 < chunk
->map_used
) {
338 int next
= chunk
->map
[i
+ 1];
339 int nend
= chunk
->map
[i
+ 2] & ~1;
341 if (!(next
& 1) && nend
>= round_up(end
, PAGE_SIZE
))
342 end
= round_up(end
, PAGE_SIZE
);
345 return max_t(int, PFN_DOWN(end
) - PFN_UP(off
), 0);
349 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
350 * @chunk: chunk of interest
351 * @oslot: the previous slot it was on
353 * This function is called after an allocation or free changed @chunk.
354 * New slot according to the changed state is determined and @chunk is
355 * moved to the slot. Note that the reserved chunk is never put on
361 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
363 int nslot
= pcpu_chunk_slot(chunk
);
365 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
367 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
369 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
374 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
375 * @chunk: chunk of interest
376 * @is_atomic: the allocation context
378 * Determine whether area map of @chunk needs to be extended. If
379 * @is_atomic, only the amount necessary for a new allocation is
380 * considered; however, async extension is scheduled if the left amount is
381 * low. If !@is_atomic, it aims for more empty space. Combined, this
382 * ensures that the map is likely to have enough available space to
383 * accomodate atomic allocations which can't extend maps directly.
389 * New target map allocation length if extension is necessary, 0
392 static int pcpu_need_to_extend(struct pcpu_chunk
*chunk
, bool is_atomic
)
394 int margin
, new_alloc
;
399 if (chunk
->map_alloc
<
400 chunk
->map_used
+ PCPU_ATOMIC_MAP_MARGIN_LOW
&&
402 schedule_work(&chunk
->map_extend_work
);
404 margin
= PCPU_ATOMIC_MAP_MARGIN_HIGH
;
407 if (chunk
->map_alloc
>= chunk
->map_used
+ margin
)
410 new_alloc
= PCPU_DFL_MAP_ALLOC
;
411 while (new_alloc
< chunk
->map_used
+ margin
)
418 * pcpu_extend_area_map - extend area map of a chunk
419 * @chunk: chunk of interest
420 * @new_alloc: new target allocation length of the area map
422 * Extend area map of @chunk to have @new_alloc entries.
425 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
428 * 0 on success, -errno on failure.
430 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
, int new_alloc
)
432 int *old
= NULL
, *new = NULL
;
433 size_t old_size
= 0, new_size
= new_alloc
* sizeof(new[0]);
436 new = pcpu_mem_zalloc(new_size
);
440 /* acquire pcpu_lock and switch to new area map */
441 spin_lock_irqsave(&pcpu_lock
, flags
);
443 if (new_alloc
<= chunk
->map_alloc
)
446 old_size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
449 memcpy(new, old
, old_size
);
451 chunk
->map_alloc
= new_alloc
;
456 spin_unlock_irqrestore(&pcpu_lock
, flags
);
459 * pcpu_mem_free() might end up calling vfree() which uses
460 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
468 static void pcpu_map_extend_workfn(struct work_struct
*work
)
470 struct pcpu_chunk
*chunk
= container_of(work
, struct pcpu_chunk
,
474 spin_lock_irq(&pcpu_lock
);
475 new_alloc
= pcpu_need_to_extend(chunk
, false);
476 spin_unlock_irq(&pcpu_lock
);
479 pcpu_extend_area_map(chunk
, new_alloc
);
483 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
484 * @chunk: chunk the candidate area belongs to
485 * @off: the offset to the start of the candidate area
486 * @this_size: the size of the candidate area
487 * @size: the size of the target allocation
488 * @align: the alignment of the target allocation
489 * @pop_only: only allocate from already populated region
491 * We're trying to allocate @size bytes aligned at @align. @chunk's area
492 * at @off sized @this_size is a candidate. This function determines
493 * whether the target allocation fits in the candidate area and returns the
494 * number of bytes to pad after @off. If the target area doesn't fit, -1
497 * If @pop_only is %true, this function only considers the already
498 * populated part of the candidate area.
500 static int pcpu_fit_in_area(struct pcpu_chunk
*chunk
, int off
, int this_size
,
501 int size
, int align
, bool pop_only
)
506 int head
= ALIGN(cand_off
, align
) - off
;
507 int page_start
, page_end
, rs
, re
;
509 if (this_size
< head
+ size
)
516 * If the first unpopulated page is beyond the end of the
517 * allocation, the whole allocation is populated;
518 * otherwise, retry from the end of the unpopulated area.
520 page_start
= PFN_DOWN(head
+ off
);
521 page_end
= PFN_UP(head
+ off
+ size
);
524 pcpu_next_unpop(chunk
, &rs
, &re
, PFN_UP(off
+ this_size
));
527 cand_off
= re
* PAGE_SIZE
;
532 * pcpu_alloc_area - allocate area from a pcpu_chunk
533 * @chunk: chunk of interest
534 * @size: wanted size in bytes
535 * @align: wanted align
536 * @pop_only: allocate only from the populated area
537 * @occ_pages_p: out param for the number of pages the area occupies
539 * Try to allocate @size bytes area aligned at @align from @chunk.
540 * Note that this function only allocates the offset. It doesn't
541 * populate or map the area.
543 * @chunk->map must have at least two free slots.
549 * Allocated offset in @chunk on success, -1 if no matching area is
552 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
,
553 bool pop_only
, int *occ_pages_p
)
555 int oslot
= pcpu_chunk_slot(chunk
);
558 bool seen_free
= false;
561 for (i
= chunk
->first_free
, p
= chunk
->map
+ i
; i
< chunk
->map_used
; i
++, p
++) {
569 this_size
= (p
[1] & ~1) - off
;
571 head
= pcpu_fit_in_area(chunk
, off
, this_size
, size
, align
,
575 chunk
->first_free
= i
;
578 max_contig
= max(this_size
, max_contig
);
583 * If head is small or the previous block is free,
584 * merge'em. Note that 'small' is defined as smaller
585 * than sizeof(int), which is very small but isn't too
586 * uncommon for percpu allocations.
588 if (head
&& (head
< sizeof(int) || !(p
[-1] & 1))) {
591 chunk
->free_size
-= head
;
593 max_contig
= max(*p
- p
[-1], max_contig
);
598 /* if tail is small, just keep it around */
599 tail
= this_size
- head
- size
;
600 if (tail
< sizeof(int)) {
602 size
= this_size
- head
;
605 /* split if warranted */
607 int nr_extra
= !!head
+ !!tail
;
609 /* insert new subblocks */
610 memmove(p
+ nr_extra
+ 1, p
+ 1,
611 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
612 chunk
->map_used
+= nr_extra
;
616 chunk
->first_free
= i
;
621 max_contig
= max(head
, max_contig
);
625 max_contig
= max(tail
, max_contig
);
630 chunk
->first_free
= i
+ 1;
632 /* update hint and mark allocated */
633 if (i
+ 1 == chunk
->map_used
)
634 chunk
->contig_hint
= max_contig
; /* fully scanned */
636 chunk
->contig_hint
= max(chunk
->contig_hint
,
639 chunk
->free_size
-= size
;
642 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
643 pcpu_chunk_relocate(chunk
, oslot
);
647 chunk
->contig_hint
= max_contig
; /* fully scanned */
648 pcpu_chunk_relocate(chunk
, oslot
);
650 /* tell the upper layer that this chunk has no matching area */
655 * pcpu_free_area - free area to a pcpu_chunk
656 * @chunk: chunk of interest
657 * @freeme: offset of area to free
658 * @occ_pages_p: out param for the number of pages the area occupies
660 * Free area starting from @freeme to @chunk. Note that this function
661 * only modifies the allocation map. It doesn't depopulate or unmap
667 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
,
670 int oslot
= pcpu_chunk_slot(chunk
);
676 freeme
|= 1; /* we are searching for <given offset, in use> pair */
681 unsigned k
= (i
+ j
) / 2;
685 else if (off
> freeme
)
690 BUG_ON(off
!= freeme
);
692 if (i
< chunk
->first_free
)
693 chunk
->first_free
= i
;
697 chunk
->free_size
+= (p
[1] & ~1) - off
;
699 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
701 /* merge with next? */
704 /* merge with previous? */
705 if (i
> 0 && !(p
[-1] & 1)) {
711 chunk
->map_used
-= to_free
;
712 memmove(p
+ 1, p
+ 1 + to_free
,
713 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
716 chunk
->contig_hint
= max(chunk
->map
[i
+ 1] - chunk
->map
[i
] - 1, chunk
->contig_hint
);
717 pcpu_chunk_relocate(chunk
, oslot
);
720 static struct pcpu_chunk
*pcpu_alloc_chunk(void)
722 struct pcpu_chunk
*chunk
;
724 chunk
= pcpu_mem_zalloc(pcpu_chunk_struct_size
);
728 chunk
->map
= pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC
*
729 sizeof(chunk
->map
[0]));
731 pcpu_mem_free(chunk
);
735 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
737 chunk
->map
[1] = pcpu_unit_size
| 1;
740 INIT_LIST_HEAD(&chunk
->list
);
741 INIT_WORK(&chunk
->map_extend_work
, pcpu_map_extend_workfn
);
742 chunk
->free_size
= pcpu_unit_size
;
743 chunk
->contig_hint
= pcpu_unit_size
;
748 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
752 pcpu_mem_free(chunk
->map
);
753 pcpu_mem_free(chunk
);
757 * pcpu_chunk_populated - post-population bookkeeping
758 * @chunk: pcpu_chunk which got populated
759 * @page_start: the start page
760 * @page_end: the end page
762 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
763 * the bookkeeping information accordingly. Must be called after each
764 * successful population.
766 static void pcpu_chunk_populated(struct pcpu_chunk
*chunk
,
767 int page_start
, int page_end
)
769 int nr
= page_end
- page_start
;
771 lockdep_assert_held(&pcpu_lock
);
773 bitmap_set(chunk
->populated
, page_start
, nr
);
774 chunk
->nr_populated
+= nr
;
775 pcpu_nr_empty_pop_pages
+= nr
;
779 * pcpu_chunk_depopulated - post-depopulation bookkeeping
780 * @chunk: pcpu_chunk which got depopulated
781 * @page_start: the start page
782 * @page_end: the end page
784 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
785 * Update the bookkeeping information accordingly. Must be called after
786 * each successful depopulation.
788 static void pcpu_chunk_depopulated(struct pcpu_chunk
*chunk
,
789 int page_start
, int page_end
)
791 int nr
= page_end
- page_start
;
793 lockdep_assert_held(&pcpu_lock
);
795 bitmap_clear(chunk
->populated
, page_start
, nr
);
796 chunk
->nr_populated
-= nr
;
797 pcpu_nr_empty_pop_pages
-= nr
;
801 * Chunk management implementation.
803 * To allow different implementations, chunk alloc/free and
804 * [de]population are implemented in a separate file which is pulled
805 * into this file and compiled together. The following functions
806 * should be implemented.
808 * pcpu_populate_chunk - populate the specified range of a chunk
809 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
810 * pcpu_create_chunk - create a new chunk
811 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
812 * pcpu_addr_to_page - translate address to physical address
813 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
815 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
816 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
817 static struct pcpu_chunk
*pcpu_create_chunk(void);
818 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
819 static struct page
*pcpu_addr_to_page(void *addr
);
820 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
822 #ifdef CONFIG_NEED_PER_CPU_KM
823 #include "percpu-km.c"
825 #include "percpu-vm.c"
829 * pcpu_chunk_addr_search - determine chunk containing specified address
830 * @addr: address for which the chunk needs to be determined.
833 * The address of the found chunk.
835 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
837 /* is it in the first chunk? */
838 if (pcpu_addr_in_first_chunk(addr
)) {
839 /* is it in the reserved area? */
840 if (pcpu_addr_in_reserved_chunk(addr
))
841 return pcpu_reserved_chunk
;
842 return pcpu_first_chunk
;
846 * The address is relative to unit0 which might be unused and
847 * thus unmapped. Offset the address to the unit space of the
848 * current processor before looking it up in the vmalloc
849 * space. Note that any possible cpu id can be used here, so
850 * there's no need to worry about preemption or cpu hotplug.
852 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
853 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
857 * pcpu_alloc - the percpu allocator
858 * @size: size of area to allocate in bytes
859 * @align: alignment of area (max PAGE_SIZE)
860 * @reserved: allocate from the reserved chunk if available
861 * @gfp: allocation flags
863 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
864 * contain %GFP_KERNEL, the allocation is atomic.
867 * Percpu pointer to the allocated area on success, NULL on failure.
869 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
,
872 static int warn_limit
= 10;
873 struct pcpu_chunk
*chunk
;
875 bool is_atomic
= (gfp
& GFP_KERNEL
) != GFP_KERNEL
;
877 int slot
, off
, new_alloc
, cpu
, ret
;
882 * We want the lowest bit of offset available for in-use/free
883 * indicator, so force >= 16bit alignment and make size even.
885 if (unlikely(align
< 2))
888 size
= ALIGN(size
, 2);
890 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
891 WARN(true, "illegal size (%zu) or align (%zu) for "
892 "percpu allocation\n", size
, align
);
896 spin_lock_irqsave(&pcpu_lock
, flags
);
898 /* serve reserved allocations from the reserved chunk if available */
899 if (reserved
&& pcpu_reserved_chunk
) {
900 chunk
= pcpu_reserved_chunk
;
902 if (size
> chunk
->contig_hint
) {
903 err
= "alloc from reserved chunk failed";
907 while ((new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
))) {
908 spin_unlock_irqrestore(&pcpu_lock
, flags
);
910 pcpu_extend_area_map(chunk
, new_alloc
) < 0) {
911 err
= "failed to extend area map of reserved chunk";
914 spin_lock_irqsave(&pcpu_lock
, flags
);
917 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
922 err
= "alloc from reserved chunk failed";
927 /* search through normal chunks */
928 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
929 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
930 if (size
> chunk
->contig_hint
)
933 new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
);
937 spin_unlock_irqrestore(&pcpu_lock
, flags
);
938 if (pcpu_extend_area_map(chunk
,
940 err
= "failed to extend area map";
943 spin_lock_irqsave(&pcpu_lock
, flags
);
945 * pcpu_lock has been dropped, need to
946 * restart cpu_slot list walking.
951 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
958 spin_unlock_irqrestore(&pcpu_lock
, flags
);
961 * No space left. Create a new chunk. We don't want multiple
962 * tasks to create chunks simultaneously. Serialize and create iff
963 * there's still no empty chunk after grabbing the mutex.
968 mutex_lock(&pcpu_alloc_mutex
);
970 if (list_empty(&pcpu_slot
[pcpu_nr_slots
- 1])) {
971 chunk
= pcpu_create_chunk();
973 mutex_unlock(&pcpu_alloc_mutex
);
974 err
= "failed to allocate new chunk";
978 spin_lock_irqsave(&pcpu_lock
, flags
);
979 pcpu_chunk_relocate(chunk
, -1);
981 spin_lock_irqsave(&pcpu_lock
, flags
);
984 mutex_unlock(&pcpu_alloc_mutex
);
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 mutex_lock(&pcpu_alloc_mutex
);
996 page_start
= PFN_DOWN(off
);
997 page_end
= PFN_UP(off
+ size
);
999 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
1000 WARN_ON(chunk
->immutable
);
1002 ret
= pcpu_populate_chunk(chunk
, rs
, re
);
1004 spin_lock_irqsave(&pcpu_lock
, flags
);
1006 mutex_unlock(&pcpu_alloc_mutex
);
1007 pcpu_free_area(chunk
, off
, &occ_pages
);
1008 err
= "failed to populate";
1011 pcpu_chunk_populated(chunk
, rs
, re
);
1012 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1015 mutex_unlock(&pcpu_alloc_mutex
);
1018 if (chunk
!= pcpu_reserved_chunk
)
1019 pcpu_nr_empty_pop_pages
-= occ_pages
;
1021 if (pcpu_nr_empty_pop_pages
< PCPU_EMPTY_POP_PAGES_LOW
)
1022 pcpu_schedule_balance_work();
1024 /* clear the areas and return address relative to base address */
1025 for_each_possible_cpu(cpu
)
1026 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1028 ptr
= __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1029 kmemleak_alloc_percpu(ptr
, size
, gfp
);
1033 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1035 if (!is_atomic
&& warn_limit
) {
1036 pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1037 size
, align
, is_atomic
, err
);
1040 pr_info("PERCPU: limit reached, disable warning\n");
1043 /* see the flag handling in pcpu_blance_workfn() */
1044 pcpu_atomic_alloc_failed
= true;
1045 pcpu_schedule_balance_work();
1051 * __alloc_percpu_gfp - allocate dynamic percpu area
1052 * @size: size of area to allocate in bytes
1053 * @align: alignment of area (max PAGE_SIZE)
1054 * @gfp: allocation flags
1056 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1057 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1058 * be called from any context but is a lot more likely to fail.
1061 * Percpu pointer to the allocated area on success, NULL on failure.
1063 void __percpu
*__alloc_percpu_gfp(size_t size
, size_t align
, gfp_t gfp
)
1065 return pcpu_alloc(size
, align
, false, gfp
);
1067 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp
);
1070 * __alloc_percpu - allocate dynamic percpu area
1071 * @size: size of area to allocate in bytes
1072 * @align: alignment of area (max PAGE_SIZE)
1074 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1076 void __percpu
*__alloc_percpu(size_t size
, size_t align
)
1078 return pcpu_alloc(size
, align
, false, GFP_KERNEL
);
1080 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1083 * __alloc_reserved_percpu - allocate reserved percpu area
1084 * @size: size of area to allocate in bytes
1085 * @align: alignment of area (max PAGE_SIZE)
1087 * Allocate zero-filled percpu area of @size bytes aligned at @align
1088 * from reserved percpu area if arch has set it up; otherwise,
1089 * allocation is served from the same dynamic area. Might sleep.
1090 * Might trigger writeouts.
1093 * Does GFP_KERNEL allocation.
1096 * Percpu pointer to the allocated area on success, NULL on failure.
1098 void __percpu
*__alloc_reserved_percpu(size_t size
, size_t align
)
1100 return pcpu_alloc(size
, align
, true, GFP_KERNEL
);
1104 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1107 * Reclaim all fully free chunks except for the first one.
1109 static void pcpu_balance_workfn(struct work_struct
*work
)
1112 struct list_head
*free_head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1113 struct pcpu_chunk
*chunk
, *next
;
1114 int slot
, nr_to_pop
, ret
;
1117 * There's no reason to keep around multiple unused chunks and VM
1118 * areas can be scarce. Destroy all free chunks except for one.
1120 mutex_lock(&pcpu_alloc_mutex
);
1121 spin_lock_irq(&pcpu_lock
);
1123 list_for_each_entry_safe(chunk
, next
, free_head
, list
) {
1124 WARN_ON(chunk
->immutable
);
1126 /* spare the first one */
1127 if (chunk
== list_first_entry(free_head
, struct pcpu_chunk
, 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
);
1148 * Ensure there are certain number of free populated pages for
1149 * atomic allocs. Fill up from the most packed so that atomic
1150 * allocs don't increase fragmentation. If atomic allocation
1151 * failed previously, always populate the maximum amount. This
1152 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1153 * failing indefinitely; however, large atomic allocs are not
1154 * something we support properly and can be highly unreliable and
1158 if (pcpu_atomic_alloc_failed
) {
1159 nr_to_pop
= PCPU_EMPTY_POP_PAGES_HIGH
;
1160 /* best effort anyway, don't worry about synchronization */
1161 pcpu_atomic_alloc_failed
= false;
1163 nr_to_pop
= clamp(PCPU_EMPTY_POP_PAGES_HIGH
-
1164 pcpu_nr_empty_pop_pages
,
1165 0, PCPU_EMPTY_POP_PAGES_HIGH
);
1168 for (slot
= pcpu_size_to_slot(PAGE_SIZE
); slot
< pcpu_nr_slots
; slot
++) {
1169 int nr_unpop
= 0, rs
, re
;
1174 spin_lock_irq(&pcpu_lock
);
1175 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1176 nr_unpop
= pcpu_unit_pages
- chunk
->nr_populated
;
1180 spin_unlock_irq(&pcpu_lock
);
1185 /* @chunk can't go away while pcpu_alloc_mutex is held */
1186 pcpu_for_each_unpop_region(chunk
, rs
, re
, 0, pcpu_unit_pages
) {
1187 int nr
= min(re
- rs
, nr_to_pop
);
1189 ret
= pcpu_populate_chunk(chunk
, rs
, rs
+ nr
);
1192 spin_lock_irq(&pcpu_lock
);
1193 pcpu_chunk_populated(chunk
, rs
, rs
+ nr
);
1194 spin_unlock_irq(&pcpu_lock
);
1205 /* ran out of chunks to populate, create a new one and retry */
1206 chunk
= pcpu_create_chunk();
1208 spin_lock_irq(&pcpu_lock
);
1209 pcpu_chunk_relocate(chunk
, -1);
1210 spin_unlock_irq(&pcpu_lock
);
1215 mutex_unlock(&pcpu_alloc_mutex
);
1219 * free_percpu - free percpu area
1220 * @ptr: pointer to area to free
1222 * Free percpu area @ptr.
1225 * Can be called from atomic context.
1227 void free_percpu(void __percpu
*ptr
)
1230 struct pcpu_chunk
*chunk
;
1231 unsigned long flags
;
1237 kmemleak_free_percpu(ptr
);
1239 addr
= __pcpu_ptr_to_addr(ptr
);
1241 spin_lock_irqsave(&pcpu_lock
, flags
);
1243 chunk
= pcpu_chunk_addr_search(addr
);
1244 off
= addr
- chunk
->base_addr
;
1246 pcpu_free_area(chunk
, off
, &occ_pages
);
1248 if (chunk
!= pcpu_reserved_chunk
)
1249 pcpu_nr_empty_pop_pages
+= occ_pages
;
1251 /* if there are more than one fully free chunks, wake up grim reaper */
1252 if (chunk
->free_size
== pcpu_unit_size
) {
1253 struct pcpu_chunk
*pos
;
1255 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1257 pcpu_schedule_balance_work();
1262 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1264 EXPORT_SYMBOL_GPL(free_percpu
);
1267 * is_kernel_percpu_address - test whether address is from static percpu area
1268 * @addr: address to test
1270 * Test whether @addr belongs to in-kernel static percpu area. Module
1271 * static percpu areas are not considered. For those, use
1272 * is_module_percpu_address().
1275 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1277 bool is_kernel_percpu_address(unsigned long addr
)
1280 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1281 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1284 for_each_possible_cpu(cpu
) {
1285 void *start
= per_cpu_ptr(base
, cpu
);
1287 if ((void *)addr
>= start
&& (void *)addr
< start
+ static_size
)
1291 /* on UP, can't distinguish from other static vars, always false */
1296 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1297 * @addr: the address to be converted to physical address
1299 * Given @addr which is dereferenceable address obtained via one of
1300 * percpu access macros, this function translates it into its physical
1301 * address. The caller is responsible for ensuring @addr stays valid
1302 * until this function finishes.
1304 * percpu allocator has special setup for the first chunk, which currently
1305 * supports either embedding in linear address space or vmalloc mapping,
1306 * and, from the second one, the backing allocator (currently either vm or
1307 * km) provides translation.
1309 * The addr can be translated simply without checking if it falls into the
1310 * first chunk. But the current code reflects better how percpu allocator
1311 * actually works, and the verification can discover both bugs in percpu
1312 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1316 * The physical address for @addr.
1318 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
1320 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1321 bool in_first_chunk
= false;
1322 unsigned long first_low
, first_high
;
1326 * The following test on unit_low/high isn't strictly
1327 * necessary but will speed up lookups of addresses which
1328 * aren't in the first chunk.
1330 first_low
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_low_unit_cpu
, 0);
1331 first_high
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_high_unit_cpu
,
1333 if ((unsigned long)addr
>= first_low
&&
1334 (unsigned long)addr
< first_high
) {
1335 for_each_possible_cpu(cpu
) {
1336 void *start
= per_cpu_ptr(base
, cpu
);
1338 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
1339 in_first_chunk
= true;
1345 if (in_first_chunk
) {
1346 if (!is_vmalloc_addr(addr
))
1349 return page_to_phys(vmalloc_to_page(addr
)) +
1350 offset_in_page(addr
);
1352 return page_to_phys(pcpu_addr_to_page(addr
)) +
1353 offset_in_page(addr
);
1357 * pcpu_alloc_alloc_info - allocate percpu allocation info
1358 * @nr_groups: the number of groups
1359 * @nr_units: the number of units
1361 * Allocate ai which is large enough for @nr_groups groups containing
1362 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1363 * cpu_map array which is long enough for @nr_units and filled with
1364 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1365 * pointer of other groups.
1368 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1371 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1374 struct pcpu_alloc_info
*ai
;
1375 size_t base_size
, ai_size
;
1379 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1380 __alignof__(ai
->groups
[0].cpu_map
[0]));
1381 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1383 ptr
= memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size
), 0);
1389 ai
->groups
[0].cpu_map
= ptr
;
1391 for (unit
= 0; unit
< nr_units
; unit
++)
1392 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1394 ai
->nr_groups
= nr_groups
;
1395 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1401 * pcpu_free_alloc_info - free percpu allocation info
1402 * @ai: pcpu_alloc_info to free
1404 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1406 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1408 memblock_free_early(__pa(ai
), ai
->__ai_size
);
1412 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1414 * @ai: allocation info to dump
1416 * Print out information about @ai using loglevel @lvl.
1418 static void pcpu_dump_alloc_info(const char *lvl
,
1419 const struct pcpu_alloc_info
*ai
)
1421 int group_width
= 1, cpu_width
= 1, width
;
1422 char empty_str
[] = "--------";
1423 int alloc
= 0, alloc_end
= 0;
1425 int upa
, apl
; /* units per alloc, allocs per line */
1431 v
= num_possible_cpus();
1434 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1436 upa
= ai
->alloc_size
/ ai
->unit_size
;
1437 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1438 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1440 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1441 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1442 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1444 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1445 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1446 int unit
= 0, unit_end
= 0;
1448 BUG_ON(gi
->nr_units
% upa
);
1449 for (alloc_end
+= gi
->nr_units
/ upa
;
1450 alloc
< alloc_end
; alloc
++) {
1451 if (!(alloc
% apl
)) {
1452 printk(KERN_CONT
"\n");
1453 printk("%spcpu-alloc: ", lvl
);
1455 printk(KERN_CONT
"[%0*d] ", group_width
, group
);
1457 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1458 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1459 printk(KERN_CONT
"%0*d ", cpu_width
,
1462 printk(KERN_CONT
"%s ", empty_str
);
1465 printk(KERN_CONT
"\n");
1469 * pcpu_setup_first_chunk - initialize the first percpu chunk
1470 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1471 * @base_addr: mapped address
1473 * Initialize the first percpu chunk which contains the kernel static
1474 * perpcu area. This function is to be called from arch percpu area
1477 * @ai contains all information necessary to initialize the first
1478 * chunk and prime the dynamic percpu allocator.
1480 * @ai->static_size is the size of static percpu area.
1482 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1483 * reserve after the static area in the first chunk. This reserves
1484 * the first chunk such that it's available only through reserved
1485 * percpu allocation. This is primarily used to serve module percpu
1486 * static areas on architectures where the addressing model has
1487 * limited offset range for symbol relocations to guarantee module
1488 * percpu symbols fall inside the relocatable range.
1490 * @ai->dyn_size determines the number of bytes available for dynamic
1491 * allocation in the first chunk. The area between @ai->static_size +
1492 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1494 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1495 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1498 * @ai->atom_size is the allocation atom size and used as alignment
1501 * @ai->alloc_size is the allocation size and always multiple of
1502 * @ai->atom_size. This is larger than @ai->atom_size if
1503 * @ai->unit_size is larger than @ai->atom_size.
1505 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1506 * percpu areas. Units which should be colocated are put into the
1507 * same group. Dynamic VM areas will be allocated according to these
1508 * groupings. If @ai->nr_groups is zero, a single group containing
1509 * all units is assumed.
1511 * The caller should have mapped the first chunk at @base_addr and
1512 * copied static data to each unit.
1514 * If the first chunk ends up with both reserved and dynamic areas, it
1515 * is served by two chunks - one to serve the core static and reserved
1516 * areas and the other for the dynamic area. They share the same vm
1517 * and page map but uses different area allocation map to stay away
1518 * from each other. The latter chunk is circulated in the chunk slots
1519 * and available for dynamic allocation like any other chunks.
1522 * 0 on success, -errno on failure.
1524 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1527 static int smap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1528 static int dmap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1529 size_t dyn_size
= ai
->dyn_size
;
1530 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1531 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1532 unsigned long *group_offsets
;
1533 size_t *group_sizes
;
1534 unsigned long *unit_off
;
1539 #define PCPU_SETUP_BUG_ON(cond) do { \
1540 if (unlikely(cond)) { \
1541 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1542 pr_emerg("PERCPU: cpu_possible_mask=%*pb\n", \
1543 cpumask_pr_args(cpu_possible_mask)); \
1544 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1550 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1552 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1553 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start
));
1555 PCPU_SETUP_BUG_ON(!base_addr
);
1556 PCPU_SETUP_BUG_ON(offset_in_page(base_addr
));
1557 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1558 PCPU_SETUP_BUG_ON(offset_in_page(ai
->unit_size
));
1559 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1560 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
1561 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
1563 /* process group information and build config tables accordingly */
1564 group_offsets
= memblock_virt_alloc(ai
->nr_groups
*
1565 sizeof(group_offsets
[0]), 0);
1566 group_sizes
= memblock_virt_alloc(ai
->nr_groups
*
1567 sizeof(group_sizes
[0]), 0);
1568 unit_map
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_map
[0]), 0);
1569 unit_off
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_off
[0]), 0);
1571 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1572 unit_map
[cpu
] = UINT_MAX
;
1574 pcpu_low_unit_cpu
= NR_CPUS
;
1575 pcpu_high_unit_cpu
= NR_CPUS
;
1577 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1578 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1580 group_offsets
[group
] = gi
->base_offset
;
1581 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1583 for (i
= 0; i
< gi
->nr_units
; i
++) {
1584 cpu
= gi
->cpu_map
[i
];
1588 PCPU_SETUP_BUG_ON(cpu
>= nr_cpu_ids
);
1589 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1590 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1592 unit_map
[cpu
] = unit
+ i
;
1593 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1595 /* determine low/high unit_cpu */
1596 if (pcpu_low_unit_cpu
== NR_CPUS
||
1597 unit_off
[cpu
] < unit_off
[pcpu_low_unit_cpu
])
1598 pcpu_low_unit_cpu
= cpu
;
1599 if (pcpu_high_unit_cpu
== NR_CPUS
||
1600 unit_off
[cpu
] > unit_off
[pcpu_high_unit_cpu
])
1601 pcpu_high_unit_cpu
= cpu
;
1604 pcpu_nr_units
= unit
;
1606 for_each_possible_cpu(cpu
)
1607 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1609 /* we're done parsing the input, undefine BUG macro and dump config */
1610 #undef PCPU_SETUP_BUG_ON
1611 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
1613 pcpu_nr_groups
= ai
->nr_groups
;
1614 pcpu_group_offsets
= group_offsets
;
1615 pcpu_group_sizes
= group_sizes
;
1616 pcpu_unit_map
= unit_map
;
1617 pcpu_unit_offsets
= unit_off
;
1619 /* determine basic parameters */
1620 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1621 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1622 pcpu_atom_size
= ai
->atom_size
;
1623 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1624 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1627 * Allocate chunk slots. The additional last slot is for
1630 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1631 pcpu_slot
= memblock_virt_alloc(
1632 pcpu_nr_slots
* sizeof(pcpu_slot
[0]), 0);
1633 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1634 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1637 * Initialize static chunk. If reserved_size is zero, the
1638 * static chunk covers static area + dynamic allocation area
1639 * in the first chunk. If reserved_size is not zero, it
1640 * covers static area + reserved area (mostly used for module
1641 * static percpu allocation).
1643 schunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1644 INIT_LIST_HEAD(&schunk
->list
);
1645 INIT_WORK(&schunk
->map_extend_work
, pcpu_map_extend_workfn
);
1646 schunk
->base_addr
= base_addr
;
1648 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1649 schunk
->immutable
= true;
1650 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1651 schunk
->nr_populated
= pcpu_unit_pages
;
1653 if (ai
->reserved_size
) {
1654 schunk
->free_size
= ai
->reserved_size
;
1655 pcpu_reserved_chunk
= schunk
;
1656 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1658 schunk
->free_size
= dyn_size
;
1659 dyn_size
= 0; /* dynamic area covered */
1661 schunk
->contig_hint
= schunk
->free_size
;
1664 schunk
->map
[1] = ai
->static_size
;
1665 schunk
->map_used
= 1;
1666 if (schunk
->free_size
)
1667 schunk
->map
[++schunk
->map_used
] = ai
->static_size
+ schunk
->free_size
;
1668 schunk
->map
[schunk
->map_used
] |= 1;
1670 /* init dynamic chunk if necessary */
1672 dchunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1673 INIT_LIST_HEAD(&dchunk
->list
);
1674 INIT_WORK(&dchunk
->map_extend_work
, pcpu_map_extend_workfn
);
1675 dchunk
->base_addr
= base_addr
;
1677 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1678 dchunk
->immutable
= true;
1679 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1680 dchunk
->nr_populated
= pcpu_unit_pages
;
1682 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1684 dchunk
->map
[1] = pcpu_reserved_chunk_limit
;
1685 dchunk
->map
[2] = (pcpu_reserved_chunk_limit
+ dchunk
->free_size
) | 1;
1686 dchunk
->map_used
= 2;
1689 /* link the first chunk in */
1690 pcpu_first_chunk
= dchunk
?: schunk
;
1691 pcpu_nr_empty_pop_pages
+=
1692 pcpu_count_occupied_pages(pcpu_first_chunk
, 1);
1693 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1696 pcpu_base_addr
= base_addr
;
1702 const char * const pcpu_fc_names
[PCPU_FC_NR
] __initconst
= {
1703 [PCPU_FC_AUTO
] = "auto",
1704 [PCPU_FC_EMBED
] = "embed",
1705 [PCPU_FC_PAGE
] = "page",
1708 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1710 static int __init
percpu_alloc_setup(char *str
)
1717 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1718 else if (!strcmp(str
, "embed"))
1719 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1721 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1722 else if (!strcmp(str
, "page"))
1723 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1726 pr_warning("PERCPU: unknown allocator %s specified\n", str
);
1730 early_param("percpu_alloc", percpu_alloc_setup
);
1733 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1734 * Build it if needed by the arch config or the generic setup is going
1737 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1738 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1739 #define BUILD_EMBED_FIRST_CHUNK
1742 /* build pcpu_page_first_chunk() iff needed by the arch config */
1743 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1744 #define BUILD_PAGE_FIRST_CHUNK
1747 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1748 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1750 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1751 * @reserved_size: the size of reserved percpu area in bytes
1752 * @dyn_size: minimum free size for dynamic allocation in bytes
1753 * @atom_size: allocation atom size
1754 * @cpu_distance_fn: callback to determine distance between cpus, optional
1756 * This function determines grouping of units, their mappings to cpus
1757 * and other parameters considering needed percpu size, allocation
1758 * atom size and distances between CPUs.
1760 * Groups are always multiples of atom size and CPUs which are of
1761 * LOCAL_DISTANCE both ways are grouped together and share space for
1762 * units in the same group. The returned configuration is guaranteed
1763 * to have CPUs on different nodes on different groups and >=75% usage
1764 * of allocated virtual address space.
1767 * On success, pointer to the new allocation_info is returned. On
1768 * failure, ERR_PTR value is returned.
1770 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1771 size_t reserved_size
, size_t dyn_size
,
1773 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1775 static int group_map
[NR_CPUS
] __initdata
;
1776 static int group_cnt
[NR_CPUS
] __initdata
;
1777 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1778 int nr_groups
= 1, nr_units
= 0;
1779 size_t size_sum
, min_unit_size
, alloc_size
;
1780 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1781 int last_allocs
, group
, unit
;
1782 unsigned int cpu
, tcpu
;
1783 struct pcpu_alloc_info
*ai
;
1784 unsigned int *cpu_map
;
1786 /* this function may be called multiple times */
1787 memset(group_map
, 0, sizeof(group_map
));
1788 memset(group_cnt
, 0, sizeof(group_cnt
));
1790 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1791 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1792 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
1793 dyn_size
= size_sum
- static_size
- reserved_size
;
1796 * Determine min_unit_size, alloc_size and max_upa such that
1797 * alloc_size is multiple of atom_size and is the smallest
1798 * which can accommodate 4k aligned segments which are equal to
1799 * or larger than min_unit_size.
1801 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1803 alloc_size
= roundup(min_unit_size
, atom_size
);
1804 upa
= alloc_size
/ min_unit_size
;
1805 while (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1809 /* group cpus according to their proximity */
1810 for_each_possible_cpu(cpu
) {
1813 for_each_possible_cpu(tcpu
) {
1816 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1817 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1818 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1820 nr_groups
= max(nr_groups
, group
+ 1);
1824 group_map
[cpu
] = group
;
1829 * Expand unit size until address space usage goes over 75%
1830 * and then as much as possible without using more address
1833 last_allocs
= INT_MAX
;
1834 for (upa
= max_upa
; upa
; upa
--) {
1835 int allocs
= 0, wasted
= 0;
1837 if (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1840 for (group
= 0; group
< nr_groups
; group
++) {
1841 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1842 allocs
+= this_allocs
;
1843 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1847 * Don't accept if wastage is over 1/3. The
1848 * greater-than comparison ensures upa==1 always
1849 * passes the following check.
1851 if (wasted
> num_possible_cpus() / 3)
1854 /* and then don't consume more memory */
1855 if (allocs
> last_allocs
)
1857 last_allocs
= allocs
;
1862 /* allocate and fill alloc_info */
1863 for (group
= 0; group
< nr_groups
; group
++)
1864 nr_units
+= roundup(group_cnt
[group
], upa
);
1866 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1868 return ERR_PTR(-ENOMEM
);
1869 cpu_map
= ai
->groups
[0].cpu_map
;
1871 for (group
= 0; group
< nr_groups
; group
++) {
1872 ai
->groups
[group
].cpu_map
= cpu_map
;
1873 cpu_map
+= roundup(group_cnt
[group
], upa
);
1876 ai
->static_size
= static_size
;
1877 ai
->reserved_size
= reserved_size
;
1878 ai
->dyn_size
= dyn_size
;
1879 ai
->unit_size
= alloc_size
/ upa
;
1880 ai
->atom_size
= atom_size
;
1881 ai
->alloc_size
= alloc_size
;
1883 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1884 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1887 * Initialize base_offset as if all groups are located
1888 * back-to-back. The caller should update this to
1889 * reflect actual allocation.
1891 gi
->base_offset
= unit
* ai
->unit_size
;
1893 for_each_possible_cpu(cpu
)
1894 if (group_map
[cpu
] == group
)
1895 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1896 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1897 unit
+= gi
->nr_units
;
1899 BUG_ON(unit
!= nr_units
);
1903 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1905 #if defined(BUILD_EMBED_FIRST_CHUNK)
1907 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1908 * @reserved_size: the size of reserved percpu area in bytes
1909 * @dyn_size: minimum free size for dynamic allocation in bytes
1910 * @atom_size: allocation atom size
1911 * @cpu_distance_fn: callback to determine distance between cpus, optional
1912 * @alloc_fn: function to allocate percpu page
1913 * @free_fn: function to free percpu page
1915 * This is a helper to ease setting up embedded first percpu chunk and
1916 * can be called where pcpu_setup_first_chunk() is expected.
1918 * If this function is used to setup the first chunk, it is allocated
1919 * by calling @alloc_fn and used as-is without being mapped into
1920 * vmalloc area. Allocations are always whole multiples of @atom_size
1921 * aligned to @atom_size.
1923 * This enables the first chunk to piggy back on the linear physical
1924 * mapping which often uses larger page size. Please note that this
1925 * can result in very sparse cpu->unit mapping on NUMA machines thus
1926 * requiring large vmalloc address space. Don't use this allocator if
1927 * vmalloc space is not orders of magnitude larger than distances
1928 * between node memory addresses (ie. 32bit NUMA machines).
1930 * @dyn_size specifies the minimum dynamic area size.
1932 * If the needed size is smaller than the minimum or specified unit
1933 * size, the leftover is returned using @free_fn.
1936 * 0 on success, -errno on failure.
1938 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
1940 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1941 pcpu_fc_alloc_fn_t alloc_fn
,
1942 pcpu_fc_free_fn_t free_fn
)
1944 void *base
= (void *)ULONG_MAX
;
1945 void **areas
= NULL
;
1946 struct pcpu_alloc_info
*ai
;
1947 size_t size_sum
, areas_size
, max_distance
;
1950 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1955 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1956 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1958 areas
= memblock_virt_alloc_nopanic(areas_size
, 0);
1964 /* allocate, copy and determine base address */
1965 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1966 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1967 unsigned int cpu
= NR_CPUS
;
1970 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
1971 cpu
= gi
->cpu_map
[i
];
1972 BUG_ON(cpu
== NR_CPUS
);
1974 /* allocate space for the whole group */
1975 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
1978 goto out_free_areas
;
1980 /* kmemleak tracks the percpu allocations separately */
1984 base
= min(ptr
, base
);
1988 * Copy data and free unused parts. This should happen after all
1989 * allocations are complete; otherwise, we may end up with
1990 * overlapping groups.
1992 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1993 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1994 void *ptr
= areas
[group
];
1996 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
1997 if (gi
->cpu_map
[i
] == NR_CPUS
) {
1998 /* unused unit, free whole */
1999 free_fn(ptr
, ai
->unit_size
);
2002 /* copy and return the unused part */
2003 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
2004 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2008 /* base address is now known, determine group base offsets */
2010 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2011 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2012 max_distance
= max_t(size_t, max_distance
,
2013 ai
->groups
[group
].base_offset
);
2015 max_distance
+= ai
->unit_size
;
2017 /* warn if maximum distance is further than 75% of vmalloc space */
2018 if (max_distance
> VMALLOC_TOTAL
* 3 / 4) {
2019 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
2020 "space 0x%lx\n", max_distance
,
2022 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2023 /* and fail if we have fallback */
2029 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2030 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
2031 ai
->dyn_size
, ai
->unit_size
);
2033 rc
= pcpu_setup_first_chunk(ai
, base
);
2037 for (group
= 0; group
< ai
->nr_groups
; group
++)
2039 free_fn(areas
[group
],
2040 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2042 pcpu_free_alloc_info(ai
);
2044 memblock_free_early(__pa(areas
), areas_size
);
2047 #endif /* BUILD_EMBED_FIRST_CHUNK */
2049 #ifdef BUILD_PAGE_FIRST_CHUNK
2051 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2052 * @reserved_size: the size of reserved percpu area in bytes
2053 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2054 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2055 * @populate_pte_fn: function to populate pte
2057 * This is a helper to ease setting up page-remapped first percpu
2058 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2060 * This is the basic allocator. Static percpu area is allocated
2061 * page-by-page into vmalloc area.
2064 * 0 on success, -errno on failure.
2066 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2067 pcpu_fc_alloc_fn_t alloc_fn
,
2068 pcpu_fc_free_fn_t free_fn
,
2069 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2071 static struct vm_struct vm
;
2072 struct pcpu_alloc_info
*ai
;
2076 struct page
**pages
;
2079 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2081 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
2084 BUG_ON(ai
->nr_groups
!= 1);
2085 BUG_ON(ai
->groups
[0].nr_units
!= num_possible_cpus());
2087 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2089 /* unaligned allocations can't be freed, round up to page size */
2090 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2092 pages
= memblock_virt_alloc(pages_size
, 0);
2094 /* allocate pages */
2096 for (unit
= 0; unit
< num_possible_cpus(); unit
++)
2097 for (i
= 0; i
< unit_pages
; i
++) {
2098 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2101 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2103 pr_warning("PERCPU: failed to allocate %s page "
2104 "for cpu%u\n", psize_str
, cpu
);
2107 /* kmemleak tracks the percpu allocations separately */
2109 pages
[j
++] = virt_to_page(ptr
);
2112 /* allocate vm area, map the pages and copy static data */
2113 vm
.flags
= VM_ALLOC
;
2114 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2115 vm_area_register_early(&vm
, PAGE_SIZE
);
2117 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2118 unsigned long unit_addr
=
2119 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2121 for (i
= 0; i
< unit_pages
; i
++)
2122 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2124 /* pte already populated, the following shouldn't fail */
2125 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2128 panic("failed to map percpu area, err=%d\n", rc
);
2131 * FIXME: Archs with virtual cache should flush local
2132 * cache for the linear mapping here - something
2133 * equivalent to flush_cache_vmap() on the local cpu.
2134 * flush_cache_vmap() can't be used as most supporting
2135 * data structures are not set up yet.
2138 /* copy static data */
2139 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2142 /* we're ready, commit */
2143 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2144 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2145 ai
->reserved_size
, ai
->dyn_size
);
2147 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2152 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2155 memblock_free_early(__pa(pages
), pages_size
);
2156 pcpu_free_alloc_info(ai
);
2159 #endif /* BUILD_PAGE_FIRST_CHUNK */
2161 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2163 * Generic SMP percpu area setup.
2165 * The embedding helper is used because its behavior closely resembles
2166 * the original non-dynamic generic percpu area setup. This is
2167 * important because many archs have addressing restrictions and might
2168 * fail if the percpu area is located far away from the previous
2169 * location. As an added bonus, in non-NUMA cases, embedding is
2170 * generally a good idea TLB-wise because percpu area can piggy back
2171 * on the physical linear memory mapping which uses large page
2172 * mappings on applicable archs.
2174 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2175 EXPORT_SYMBOL(__per_cpu_offset
);
2177 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2180 return memblock_virt_alloc_from_nopanic(
2181 size
, align
, __pa(MAX_DMA_ADDRESS
));
2184 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2186 memblock_free_early(__pa(ptr
), size
);
2189 void __init
setup_per_cpu_areas(void)
2191 unsigned long delta
;
2196 * Always reserve area for module percpu variables. That's
2197 * what the legacy allocator did.
2199 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2200 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2201 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2203 panic("Failed to initialize percpu areas.");
2205 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2206 for_each_possible_cpu(cpu
)
2207 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2209 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2211 #else /* CONFIG_SMP */
2214 * UP percpu area setup.
2216 * UP always uses km-based percpu allocator with identity mapping.
2217 * Static percpu variables are indistinguishable from the usual static
2218 * variables and don't require any special preparation.
2220 void __init
setup_per_cpu_areas(void)
2222 const size_t unit_size
=
2223 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
2224 PERCPU_DYNAMIC_RESERVE
));
2225 struct pcpu_alloc_info
*ai
;
2228 ai
= pcpu_alloc_alloc_info(1, 1);
2229 fc
= memblock_virt_alloc_from_nopanic(unit_size
,
2231 __pa(MAX_DMA_ADDRESS
));
2233 panic("Failed to allocate memory for percpu areas.");
2234 /* kmemleak tracks the percpu allocations separately */
2237 ai
->dyn_size
= unit_size
;
2238 ai
->unit_size
= unit_size
;
2239 ai
->atom_size
= unit_size
;
2240 ai
->alloc_size
= unit_size
;
2241 ai
->groups
[0].nr_units
= 1;
2242 ai
->groups
[0].cpu_map
[0] = 0;
2244 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
2245 panic("Failed to initialize percpu areas.");
2248 #endif /* CONFIG_SMP */
2251 * First and reserved chunks are initialized with temporary allocation
2252 * map in initdata so that they can be used before slab is online.
2253 * This function is called after slab is brought up and replaces those
2254 * with properly allocated maps.
2256 void __init
percpu_init_late(void)
2258 struct pcpu_chunk
*target_chunks
[] =
2259 { pcpu_first_chunk
, pcpu_reserved_chunk
, NULL
};
2260 struct pcpu_chunk
*chunk
;
2261 unsigned long flags
;
2264 for (i
= 0; (chunk
= target_chunks
[i
]); i
++) {
2266 const size_t size
= PERCPU_DYNAMIC_EARLY_SLOTS
* sizeof(map
[0]);
2268 BUILD_BUG_ON(size
> PAGE_SIZE
);
2270 map
= pcpu_mem_zalloc(size
);
2273 spin_lock_irqsave(&pcpu_lock
, flags
);
2274 memcpy(map
, chunk
->map
, size
);
2276 spin_unlock_irqrestore(&pcpu_lock
, flags
);
2281 * Percpu allocator is initialized early during boot when neither slab or
2282 * workqueue is available. Plug async management until everything is up
2285 static int __init
percpu_enable_async(void)
2287 pcpu_async_enabled
= true;
2290 subsys_initcall(percpu_enable_async
);