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 following:
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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
58 #include <linux/bitmap.h>
59 #include <linux/bootmem.h>
60 #include <linux/err.h>
61 #include <linux/list.h>
62 #include <linux/log2.h>
64 #include <linux/module.h>
65 #include <linux/mutex.h>
66 #include <linux/percpu.h>
67 #include <linux/pfn.h>
68 #include <linux/slab.h>
69 #include <linux/spinlock.h>
70 #include <linux/vmalloc.h>
71 #include <linux/workqueue.h>
72 #include <linux/kmemleak.h>
74 #include <asm/cacheflush.h>
75 #include <asm/sections.h>
76 #include <asm/tlbflush.h>
79 #define CREATE_TRACE_POINTS
80 #include <trace/events/percpu.h>
82 #include "percpu-internal.h"
84 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
85 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
86 #define PCPU_ATOMIC_MAP_MARGIN_LOW 32
87 #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
88 #define PCPU_EMPTY_POP_PAGES_LOW 2
89 #define PCPU_EMPTY_POP_PAGES_HIGH 4
92 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
93 #ifndef __addr_to_pcpu_ptr
94 #define __addr_to_pcpu_ptr(addr) \
95 (void __percpu *)((unsigned long)(addr) - \
96 (unsigned long)pcpu_base_addr + \
97 (unsigned long)__per_cpu_start)
99 #ifndef __pcpu_ptr_to_addr
100 #define __pcpu_ptr_to_addr(ptr) \
101 (void __force *)((unsigned long)(ptr) + \
102 (unsigned long)pcpu_base_addr - \
103 (unsigned long)__per_cpu_start)
105 #else /* CONFIG_SMP */
106 /* on UP, it's always identity mapped */
107 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
108 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
109 #endif /* CONFIG_SMP */
111 static int pcpu_unit_pages __ro_after_init
;
112 static int pcpu_unit_size __ro_after_init
;
113 static int pcpu_nr_units __ro_after_init
;
114 static int pcpu_atom_size __ro_after_init
;
115 int pcpu_nr_slots __ro_after_init
;
116 static size_t pcpu_chunk_struct_size __ro_after_init
;
118 /* cpus with the lowest and highest unit addresses */
119 static unsigned int pcpu_low_unit_cpu __ro_after_init
;
120 static unsigned int pcpu_high_unit_cpu __ro_after_init
;
122 /* the address of the first chunk which starts with the kernel static area */
123 void *pcpu_base_addr __ro_after_init
;
124 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
126 static const int *pcpu_unit_map __ro_after_init
; /* cpu -> unit */
127 const unsigned long *pcpu_unit_offsets __ro_after_init
; /* cpu -> unit offset */
129 /* group information, used for vm allocation */
130 static int pcpu_nr_groups __ro_after_init
;
131 static const unsigned long *pcpu_group_offsets __ro_after_init
;
132 static const size_t *pcpu_group_sizes __ro_after_init
;
135 * The first chunk which always exists. Note that unlike other
136 * chunks, this one can be allocated and mapped in several different
137 * ways and thus often doesn't live in the vmalloc area.
139 struct pcpu_chunk
*pcpu_first_chunk __ro_after_init
;
142 * Optional reserved chunk. This chunk reserves part of the first
143 * chunk and serves it for reserved allocations. The amount of
144 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
145 * area doesn't exist, the following variables contain NULL and 0
148 struct pcpu_chunk
*pcpu_reserved_chunk __ro_after_init
;
149 static int pcpu_reserved_chunk_limit __ro_after_init
;
151 DEFINE_SPINLOCK(pcpu_lock
); /* all internal data structures */
152 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* chunk create/destroy, [de]pop, map ext */
154 struct list_head
*pcpu_slot __ro_after_init
; /* chunk list slots */
156 /* chunks which need their map areas extended, protected by pcpu_lock */
157 static LIST_HEAD(pcpu_map_extend_chunks
);
160 * The number of empty populated pages, protected by pcpu_lock. The
161 * reserved chunk doesn't contribute to the count.
163 static int pcpu_nr_empty_pop_pages
;
166 * Balance work is used to populate or destroy chunks asynchronously. We
167 * try to keep the number of populated free pages between
168 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
171 static void pcpu_balance_workfn(struct work_struct
*work
);
172 static DECLARE_WORK(pcpu_balance_work
, pcpu_balance_workfn
);
173 static bool pcpu_async_enabled __read_mostly
;
174 static bool pcpu_atomic_alloc_failed
;
176 static void pcpu_schedule_balance_work(void)
178 if (pcpu_async_enabled
)
179 schedule_work(&pcpu_balance_work
);
182 static bool pcpu_addr_in_first_chunk(void *addr
)
184 void *first_start
= pcpu_first_chunk
->base_addr
;
186 return addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
;
189 static bool pcpu_addr_in_reserved_chunk(void *addr
)
191 void *first_start
= pcpu_first_chunk
->base_addr
;
193 return addr
>= first_start
&&
194 addr
< first_start
+ pcpu_reserved_chunk_limit
;
197 static int __pcpu_size_to_slot(int size
)
199 int highbit
= fls(size
); /* size is in bytes */
200 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
203 static int pcpu_size_to_slot(int size
)
205 if (size
== pcpu_unit_size
)
206 return pcpu_nr_slots
- 1;
207 return __pcpu_size_to_slot(size
);
210 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
212 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
215 return pcpu_size_to_slot(chunk
->free_size
);
218 /* set the pointer to a chunk in a page struct */
219 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
221 page
->index
= (unsigned long)pcpu
;
224 /* obtain pointer to a chunk from a page struct */
225 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
227 return (struct pcpu_chunk
*)page
->index
;
230 static int __maybe_unused
pcpu_page_idx(unsigned int cpu
, int page_idx
)
232 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
235 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
236 unsigned int cpu
, int page_idx
)
238 return (unsigned long)chunk
->base_addr
+ pcpu_unit_offsets
[cpu
] +
239 (page_idx
<< PAGE_SHIFT
);
242 static void __maybe_unused
pcpu_next_unpop(struct pcpu_chunk
*chunk
,
243 int *rs
, int *re
, int end
)
245 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
246 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
249 static void __maybe_unused
pcpu_next_pop(struct pcpu_chunk
*chunk
,
250 int *rs
, int *re
, int end
)
252 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
253 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
257 * (Un)populated page region iterators. Iterate over (un)populated
258 * page regions between @start and @end in @chunk. @rs and @re should
259 * be integer variables and will be set to start and end page index of
260 * the current region.
262 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
263 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
265 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
267 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
270 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
273 * pcpu_mem_zalloc - allocate memory
274 * @size: bytes to allocate
276 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
277 * kzalloc() is used; otherwise, vzalloc() is used. The returned
278 * memory is always zeroed.
281 * Does GFP_KERNEL allocation.
284 * Pointer to the allocated area on success, NULL on failure.
286 static void *pcpu_mem_zalloc(size_t size
)
288 if (WARN_ON_ONCE(!slab_is_available()))
291 if (size
<= PAGE_SIZE
)
292 return kzalloc(size
, GFP_KERNEL
);
294 return vzalloc(size
);
298 * pcpu_mem_free - free memory
299 * @ptr: memory to free
301 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
303 static void pcpu_mem_free(void *ptr
)
309 * pcpu_count_occupied_pages - count the number of pages an area occupies
310 * @chunk: chunk of interest
311 * @i: index of the area in question
313 * Count the number of pages chunk's @i'th area occupies. When the area's
314 * start and/or end address isn't aligned to page boundary, the straddled
315 * page is included in the count iff the rest of the page is free.
317 static int pcpu_count_occupied_pages(struct pcpu_chunk
*chunk
, int i
)
319 int off
= chunk
->map
[i
] & ~1;
320 int end
= chunk
->map
[i
+ 1] & ~1;
322 if (!PAGE_ALIGNED(off
) && i
> 0) {
323 int prev
= chunk
->map
[i
- 1];
325 if (!(prev
& 1) && prev
<= round_down(off
, PAGE_SIZE
))
326 off
= round_down(off
, PAGE_SIZE
);
329 if (!PAGE_ALIGNED(end
) && i
+ 1 < chunk
->map_used
) {
330 int next
= chunk
->map
[i
+ 1];
331 int nend
= chunk
->map
[i
+ 2] & ~1;
333 if (!(next
& 1) && nend
>= round_up(end
, PAGE_SIZE
))
334 end
= round_up(end
, PAGE_SIZE
);
337 return max_t(int, PFN_DOWN(end
) - PFN_UP(off
), 0);
341 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
342 * @chunk: chunk of interest
343 * @oslot: the previous slot it was on
345 * This function is called after an allocation or free changed @chunk.
346 * New slot according to the changed state is determined and @chunk is
347 * moved to the slot. Note that the reserved chunk is never put on
353 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
355 int nslot
= pcpu_chunk_slot(chunk
);
357 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
359 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
361 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
366 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
367 * @chunk: chunk of interest
368 * @is_atomic: the allocation context
370 * Determine whether area map of @chunk needs to be extended. If
371 * @is_atomic, only the amount necessary for a new allocation is
372 * considered; however, async extension is scheduled if the left amount is
373 * low. If !@is_atomic, it aims for more empty space. Combined, this
374 * ensures that the map is likely to have enough available space to
375 * accomodate atomic allocations which can't extend maps directly.
381 * New target map allocation length if extension is necessary, 0
384 static int pcpu_need_to_extend(struct pcpu_chunk
*chunk
, bool is_atomic
)
386 int margin
, new_alloc
;
388 lockdep_assert_held(&pcpu_lock
);
393 if (chunk
->map_alloc
<
394 chunk
->map_used
+ PCPU_ATOMIC_MAP_MARGIN_LOW
) {
395 if (list_empty(&chunk
->map_extend_list
)) {
396 list_add_tail(&chunk
->map_extend_list
,
397 &pcpu_map_extend_chunks
);
398 pcpu_schedule_balance_work();
402 margin
= PCPU_ATOMIC_MAP_MARGIN_HIGH
;
405 if (chunk
->map_alloc
>= chunk
->map_used
+ margin
)
408 new_alloc
= PCPU_DFL_MAP_ALLOC
;
409 while (new_alloc
< chunk
->map_used
+ margin
)
416 * pcpu_extend_area_map - extend area map of a chunk
417 * @chunk: chunk of interest
418 * @new_alloc: new target allocation length of the area map
420 * Extend area map of @chunk to have @new_alloc entries.
423 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
426 * 0 on success, -errno on failure.
428 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
, int new_alloc
)
430 int *old
= NULL
, *new = NULL
;
431 size_t old_size
= 0, new_size
= new_alloc
* sizeof(new[0]);
434 lockdep_assert_held(&pcpu_alloc_mutex
);
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.
469 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
470 * @chunk: chunk the candidate area belongs to
471 * @off: the offset to the start of the candidate area
472 * @this_size: the size of the candidate area
473 * @size: the size of the target allocation
474 * @align: the alignment of the target allocation
475 * @pop_only: only allocate from already populated region
477 * We're trying to allocate @size bytes aligned at @align. @chunk's area
478 * at @off sized @this_size is a candidate. This function determines
479 * whether the target allocation fits in the candidate area and returns the
480 * number of bytes to pad after @off. If the target area doesn't fit, -1
483 * If @pop_only is %true, this function only considers the already
484 * populated part of the candidate area.
486 static int pcpu_fit_in_area(struct pcpu_chunk
*chunk
, int off
, int this_size
,
487 int size
, int align
, bool pop_only
)
492 int head
= ALIGN(cand_off
, align
) - off
;
493 int page_start
, page_end
, rs
, re
;
495 if (this_size
< head
+ size
)
502 * If the first unpopulated page is beyond the end of the
503 * allocation, the whole allocation is populated;
504 * otherwise, retry from the end of the unpopulated area.
506 page_start
= PFN_DOWN(head
+ off
);
507 page_end
= PFN_UP(head
+ off
+ size
);
510 pcpu_next_unpop(chunk
, &rs
, &re
, PFN_UP(off
+ this_size
));
513 cand_off
= re
* PAGE_SIZE
;
518 * pcpu_alloc_area - allocate area from a pcpu_chunk
519 * @chunk: chunk of interest
520 * @size: wanted size in bytes
521 * @align: wanted align
522 * @pop_only: allocate only from the populated area
523 * @occ_pages_p: out param for the number of pages the area occupies
525 * Try to allocate @size bytes area aligned at @align from @chunk.
526 * Note that this function only allocates the offset. It doesn't
527 * populate or map the area.
529 * @chunk->map must have at least two free slots.
535 * Allocated offset in @chunk on success, -1 if no matching area is
538 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
,
539 bool pop_only
, int *occ_pages_p
)
541 int oslot
= pcpu_chunk_slot(chunk
);
544 bool seen_free
= false;
547 for (i
= chunk
->first_free
, p
= chunk
->map
+ i
; i
< chunk
->map_used
; i
++, p
++) {
555 this_size
= (p
[1] & ~1) - off
;
557 head
= pcpu_fit_in_area(chunk
, off
, this_size
, size
, align
,
561 chunk
->first_free
= i
;
564 max_contig
= max(this_size
, max_contig
);
569 * If head is small or the previous block is free,
570 * merge'em. Note that 'small' is defined as smaller
571 * than sizeof(int), which is very small but isn't too
572 * uncommon for percpu allocations.
574 if (head
&& (head
< sizeof(int) || !(p
[-1] & 1))) {
577 chunk
->free_size
-= head
;
579 max_contig
= max(*p
- p
[-1], max_contig
);
584 /* if tail is small, just keep it around */
585 tail
= this_size
- head
- size
;
586 if (tail
< sizeof(int)) {
588 size
= this_size
- head
;
591 /* split if warranted */
593 int nr_extra
= !!head
+ !!tail
;
595 /* insert new subblocks */
596 memmove(p
+ nr_extra
+ 1, p
+ 1,
597 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
598 chunk
->map_used
+= nr_extra
;
602 chunk
->first_free
= i
;
607 max_contig
= max(head
, max_contig
);
611 max_contig
= max(tail
, max_contig
);
616 chunk
->first_free
= i
+ 1;
618 /* update hint and mark allocated */
619 if (i
+ 1 == chunk
->map_used
)
620 chunk
->contig_hint
= max_contig
; /* fully scanned */
622 chunk
->contig_hint
= max(chunk
->contig_hint
,
625 chunk
->free_size
-= size
;
628 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
629 pcpu_chunk_relocate(chunk
, oslot
);
633 chunk
->contig_hint
= max_contig
; /* fully scanned */
634 pcpu_chunk_relocate(chunk
, oslot
);
636 /* tell the upper layer that this chunk has no matching area */
641 * pcpu_free_area - free area to a pcpu_chunk
642 * @chunk: chunk of interest
643 * @freeme: offset of area to free
644 * @occ_pages_p: out param for the number of pages the area occupies
646 * Free area starting from @freeme to @chunk. Note that this function
647 * only modifies the allocation map. It doesn't depopulate or unmap
653 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
,
656 int oslot
= pcpu_chunk_slot(chunk
);
662 lockdep_assert_held(&pcpu_lock
);
663 pcpu_stats_area_dealloc(chunk
);
665 freeme
|= 1; /* we are searching for <given offset, in use> pair */
670 unsigned k
= (i
+ j
) / 2;
674 else if (off
> freeme
)
679 BUG_ON(off
!= freeme
);
681 if (i
< chunk
->first_free
)
682 chunk
->first_free
= i
;
686 chunk
->free_size
+= (p
[1] & ~1) - off
;
688 *occ_pages_p
= pcpu_count_occupied_pages(chunk
, i
);
690 /* merge with next? */
693 /* merge with previous? */
694 if (i
> 0 && !(p
[-1] & 1)) {
700 chunk
->map_used
-= to_free
;
701 memmove(p
+ 1, p
+ 1 + to_free
,
702 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
705 chunk
->contig_hint
= max(chunk
->map
[i
+ 1] - chunk
->map
[i
] - 1, chunk
->contig_hint
);
706 pcpu_chunk_relocate(chunk
, oslot
);
709 static struct pcpu_chunk
*pcpu_alloc_chunk(void)
711 struct pcpu_chunk
*chunk
;
713 chunk
= pcpu_mem_zalloc(pcpu_chunk_struct_size
);
717 chunk
->map
= pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC
*
718 sizeof(chunk
->map
[0]));
720 pcpu_mem_free(chunk
);
724 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
726 chunk
->map
[1] = pcpu_unit_size
| 1;
728 chunk
->has_reserved
= false;
730 INIT_LIST_HEAD(&chunk
->list
);
731 INIT_LIST_HEAD(&chunk
->map_extend_list
);
732 chunk
->free_size
= pcpu_unit_size
;
733 chunk
->contig_hint
= pcpu_unit_size
;
738 static void pcpu_free_chunk(struct pcpu_chunk
*chunk
)
742 pcpu_mem_free(chunk
->map
);
743 pcpu_mem_free(chunk
);
747 * pcpu_chunk_populated - post-population bookkeeping
748 * @chunk: pcpu_chunk which got populated
749 * @page_start: the start page
750 * @page_end: the end page
752 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
753 * the bookkeeping information accordingly. Must be called after each
754 * successful population.
756 static void pcpu_chunk_populated(struct pcpu_chunk
*chunk
,
757 int page_start
, int page_end
)
759 int nr
= page_end
- page_start
;
761 lockdep_assert_held(&pcpu_lock
);
763 bitmap_set(chunk
->populated
, page_start
, nr
);
764 chunk
->nr_populated
+= nr
;
765 pcpu_nr_empty_pop_pages
+= nr
;
769 * pcpu_chunk_depopulated - post-depopulation bookkeeping
770 * @chunk: pcpu_chunk which got depopulated
771 * @page_start: the start page
772 * @page_end: the end page
774 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
775 * Update the bookkeeping information accordingly. Must be called after
776 * each successful depopulation.
778 static void pcpu_chunk_depopulated(struct pcpu_chunk
*chunk
,
779 int page_start
, int page_end
)
781 int nr
= page_end
- page_start
;
783 lockdep_assert_held(&pcpu_lock
);
785 bitmap_clear(chunk
->populated
, page_start
, nr
);
786 chunk
->nr_populated
-= nr
;
787 pcpu_nr_empty_pop_pages
-= nr
;
791 * Chunk management implementation.
793 * To allow different implementations, chunk alloc/free and
794 * [de]population are implemented in a separate file which is pulled
795 * into this file and compiled together. The following functions
796 * should be implemented.
798 * pcpu_populate_chunk - populate the specified range of a chunk
799 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
800 * pcpu_create_chunk - create a new chunk
801 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
802 * pcpu_addr_to_page - translate address to physical address
803 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
805 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
806 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
);
807 static struct pcpu_chunk
*pcpu_create_chunk(void);
808 static void pcpu_destroy_chunk(struct pcpu_chunk
*chunk
);
809 static struct page
*pcpu_addr_to_page(void *addr
);
810 static int __init
pcpu_verify_alloc_info(const struct pcpu_alloc_info
*ai
);
812 #ifdef CONFIG_NEED_PER_CPU_KM
813 #include "percpu-km.c"
815 #include "percpu-vm.c"
819 * pcpu_chunk_addr_search - determine chunk containing specified address
820 * @addr: address for which the chunk needs to be determined.
823 * The address of the found chunk.
825 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
827 /* is it in the first chunk? */
828 if (pcpu_addr_in_first_chunk(addr
)) {
829 /* is it in the reserved area? */
830 if (pcpu_addr_in_reserved_chunk(addr
))
831 return pcpu_reserved_chunk
;
832 return pcpu_first_chunk
;
836 * The address is relative to unit0 which might be unused and
837 * thus unmapped. Offset the address to the unit space of the
838 * current processor before looking it up in the vmalloc
839 * space. Note that any possible cpu id can be used here, so
840 * there's no need to worry about preemption or cpu hotplug.
842 addr
+= pcpu_unit_offsets
[raw_smp_processor_id()];
843 return pcpu_get_page_chunk(pcpu_addr_to_page(addr
));
847 * pcpu_alloc - the percpu allocator
848 * @size: size of area to allocate in bytes
849 * @align: alignment of area (max PAGE_SIZE)
850 * @reserved: allocate from the reserved chunk if available
851 * @gfp: allocation flags
853 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
854 * contain %GFP_KERNEL, the allocation is atomic.
857 * Percpu pointer to the allocated area on success, NULL on failure.
859 static void __percpu
*pcpu_alloc(size_t size
, size_t align
, bool reserved
,
862 static int warn_limit
= 10;
863 struct pcpu_chunk
*chunk
;
865 bool is_atomic
= (gfp
& GFP_KERNEL
) != GFP_KERNEL
;
867 int slot
, off
, new_alloc
, cpu
, ret
;
872 * We want the lowest bit of offset available for in-use/free
873 * indicator, so force >= 16bit alignment and make size even.
875 if (unlikely(align
< 2))
878 size
= ALIGN(size
, 2);
880 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
||
881 !is_power_of_2(align
))) {
882 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
888 mutex_lock(&pcpu_alloc_mutex
);
890 spin_lock_irqsave(&pcpu_lock
, flags
);
892 /* serve reserved allocations from the reserved chunk if available */
893 if (reserved
&& pcpu_reserved_chunk
) {
894 chunk
= pcpu_reserved_chunk
;
896 if (size
> chunk
->contig_hint
) {
897 err
= "alloc from reserved chunk failed";
901 while ((new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
))) {
902 spin_unlock_irqrestore(&pcpu_lock
, flags
);
904 pcpu_extend_area_map(chunk
, new_alloc
) < 0) {
905 err
= "failed to extend area map of reserved chunk";
908 spin_lock_irqsave(&pcpu_lock
, flags
);
911 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
916 err
= "alloc from reserved chunk failed";
921 /* search through normal chunks */
922 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
923 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
924 if (size
> chunk
->contig_hint
)
927 new_alloc
= pcpu_need_to_extend(chunk
, is_atomic
);
931 spin_unlock_irqrestore(&pcpu_lock
, flags
);
932 if (pcpu_extend_area_map(chunk
,
934 err
= "failed to extend area map";
937 spin_lock_irqsave(&pcpu_lock
, flags
);
939 * pcpu_lock has been dropped, need to
940 * restart cpu_slot list walking.
945 off
= pcpu_alloc_area(chunk
, size
, align
, is_atomic
,
952 spin_unlock_irqrestore(&pcpu_lock
, flags
);
955 * No space left. Create a new chunk. We don't want multiple
956 * tasks to create chunks simultaneously. Serialize and create iff
957 * there's still no empty chunk after grabbing the mutex.
962 if (list_empty(&pcpu_slot
[pcpu_nr_slots
- 1])) {
963 chunk
= pcpu_create_chunk();
965 err
= "failed to allocate new chunk";
969 spin_lock_irqsave(&pcpu_lock
, flags
);
970 pcpu_chunk_relocate(chunk
, -1);
972 spin_lock_irqsave(&pcpu_lock
, flags
);
978 pcpu_stats_area_alloc(chunk
, size
);
979 spin_unlock_irqrestore(&pcpu_lock
, flags
);
981 /* populate if not all pages are already there */
983 int page_start
, page_end
, rs
, re
;
985 page_start
= PFN_DOWN(off
);
986 page_end
= PFN_UP(off
+ size
);
988 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
989 WARN_ON(chunk
->immutable
);
991 ret
= pcpu_populate_chunk(chunk
, rs
, re
);
993 spin_lock_irqsave(&pcpu_lock
, flags
);
995 pcpu_free_area(chunk
, off
, &occ_pages
);
996 err
= "failed to populate";
999 pcpu_chunk_populated(chunk
, rs
, re
);
1000 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1003 mutex_unlock(&pcpu_alloc_mutex
);
1006 if (chunk
!= pcpu_reserved_chunk
) {
1007 spin_lock_irqsave(&pcpu_lock
, flags
);
1008 pcpu_nr_empty_pop_pages
-= occ_pages
;
1009 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1012 if (pcpu_nr_empty_pop_pages
< PCPU_EMPTY_POP_PAGES_LOW
)
1013 pcpu_schedule_balance_work();
1015 /* clear the areas and return address relative to base address */
1016 for_each_possible_cpu(cpu
)
1017 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
1019 ptr
= __addr_to_pcpu_ptr(chunk
->base_addr
+ off
);
1020 kmemleak_alloc_percpu(ptr
, size
, gfp
);
1022 trace_percpu_alloc_percpu(reserved
, is_atomic
, size
, align
,
1023 chunk
->base_addr
, off
, ptr
);
1028 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1030 trace_percpu_alloc_percpu_fail(reserved
, is_atomic
, size
, align
);
1032 if (!is_atomic
&& warn_limit
) {
1033 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1034 size
, align
, is_atomic
, err
);
1037 pr_info("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 trace_percpu_free_percpu(chunk
->base_addr
, off
, ptr
);
1283 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1285 EXPORT_SYMBOL_GPL(free_percpu
);
1287 bool __is_kernel_percpu_address(unsigned long addr
, unsigned long *can_addr
)
1290 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1291 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1294 for_each_possible_cpu(cpu
) {
1295 void *start
= per_cpu_ptr(base
, cpu
);
1296 void *va
= (void *)addr
;
1298 if (va
>= start
&& va
< start
+ static_size
) {
1300 *can_addr
= (unsigned long) (va
- start
);
1301 *can_addr
+= (unsigned long)
1302 per_cpu_ptr(base
, get_boot_cpu_id());
1308 /* on UP, can't distinguish from other static vars, always false */
1313 * is_kernel_percpu_address - test whether address is from static percpu area
1314 * @addr: address to test
1316 * Test whether @addr belongs to in-kernel static percpu area. Module
1317 * static percpu areas are not considered. For those, use
1318 * is_module_percpu_address().
1321 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1323 bool is_kernel_percpu_address(unsigned long addr
)
1325 return __is_kernel_percpu_address(addr
, NULL
);
1329 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1330 * @addr: the address to be converted to physical address
1332 * Given @addr which is dereferenceable address obtained via one of
1333 * percpu access macros, this function translates it into its physical
1334 * address. The caller is responsible for ensuring @addr stays valid
1335 * until this function finishes.
1337 * percpu allocator has special setup for the first chunk, which currently
1338 * supports either embedding in linear address space or vmalloc mapping,
1339 * and, from the second one, the backing allocator (currently either vm or
1340 * km) provides translation.
1342 * The addr can be translated simply without checking if it falls into the
1343 * first chunk. But the current code reflects better how percpu allocator
1344 * actually works, and the verification can discover both bugs in percpu
1345 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1349 * The physical address for @addr.
1351 phys_addr_t
per_cpu_ptr_to_phys(void *addr
)
1353 void __percpu
*base
= __addr_to_pcpu_ptr(pcpu_base_addr
);
1354 bool in_first_chunk
= false;
1355 unsigned long first_low
, first_high
;
1359 * The following test on unit_low/high isn't strictly
1360 * necessary but will speed up lookups of addresses which
1361 * aren't in the first chunk.
1363 first_low
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_low_unit_cpu
, 0);
1364 first_high
= pcpu_chunk_addr(pcpu_first_chunk
, pcpu_high_unit_cpu
,
1366 if ((unsigned long)addr
>= first_low
&&
1367 (unsigned long)addr
< first_high
) {
1368 for_each_possible_cpu(cpu
) {
1369 void *start
= per_cpu_ptr(base
, cpu
);
1371 if (addr
>= start
&& addr
< start
+ pcpu_unit_size
) {
1372 in_first_chunk
= true;
1378 if (in_first_chunk
) {
1379 if (!is_vmalloc_addr(addr
))
1382 return page_to_phys(vmalloc_to_page(addr
)) +
1383 offset_in_page(addr
);
1385 return page_to_phys(pcpu_addr_to_page(addr
)) +
1386 offset_in_page(addr
);
1390 * pcpu_alloc_alloc_info - allocate percpu allocation info
1391 * @nr_groups: the number of groups
1392 * @nr_units: the number of units
1394 * Allocate ai which is large enough for @nr_groups groups containing
1395 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1396 * cpu_map array which is long enough for @nr_units and filled with
1397 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1398 * pointer of other groups.
1401 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1404 struct pcpu_alloc_info
* __init
pcpu_alloc_alloc_info(int nr_groups
,
1407 struct pcpu_alloc_info
*ai
;
1408 size_t base_size
, ai_size
;
1412 base_size
= ALIGN(sizeof(*ai
) + nr_groups
* sizeof(ai
->groups
[0]),
1413 __alignof__(ai
->groups
[0].cpu_map
[0]));
1414 ai_size
= base_size
+ nr_units
* sizeof(ai
->groups
[0].cpu_map
[0]);
1416 ptr
= memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size
), 0);
1422 ai
->groups
[0].cpu_map
= ptr
;
1424 for (unit
= 0; unit
< nr_units
; unit
++)
1425 ai
->groups
[0].cpu_map
[unit
] = NR_CPUS
;
1427 ai
->nr_groups
= nr_groups
;
1428 ai
->__ai_size
= PFN_ALIGN(ai_size
);
1434 * pcpu_free_alloc_info - free percpu allocation info
1435 * @ai: pcpu_alloc_info to free
1437 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1439 void __init
pcpu_free_alloc_info(struct pcpu_alloc_info
*ai
)
1441 memblock_free_early(__pa(ai
), ai
->__ai_size
);
1445 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1447 * @ai: allocation info to dump
1449 * Print out information about @ai using loglevel @lvl.
1451 static void pcpu_dump_alloc_info(const char *lvl
,
1452 const struct pcpu_alloc_info
*ai
)
1454 int group_width
= 1, cpu_width
= 1, width
;
1455 char empty_str
[] = "--------";
1456 int alloc
= 0, alloc_end
= 0;
1458 int upa
, apl
; /* units per alloc, allocs per line */
1464 v
= num_possible_cpus();
1467 empty_str
[min_t(int, cpu_width
, sizeof(empty_str
) - 1)] = '\0';
1469 upa
= ai
->alloc_size
/ ai
->unit_size
;
1470 width
= upa
* (cpu_width
+ 1) + group_width
+ 3;
1471 apl
= rounddown_pow_of_two(max(60 / width
, 1));
1473 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1474 lvl
, ai
->static_size
, ai
->reserved_size
, ai
->dyn_size
,
1475 ai
->unit_size
, ai
->alloc_size
/ ai
->atom_size
, ai
->atom_size
);
1477 for (group
= 0; group
< ai
->nr_groups
; group
++) {
1478 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1479 int unit
= 0, unit_end
= 0;
1481 BUG_ON(gi
->nr_units
% upa
);
1482 for (alloc_end
+= gi
->nr_units
/ upa
;
1483 alloc
< alloc_end
; alloc
++) {
1484 if (!(alloc
% apl
)) {
1486 printk("%spcpu-alloc: ", lvl
);
1488 pr_cont("[%0*d] ", group_width
, group
);
1490 for (unit_end
+= upa
; unit
< unit_end
; unit
++)
1491 if (gi
->cpu_map
[unit
] != NR_CPUS
)
1493 cpu_width
, gi
->cpu_map
[unit
]);
1495 pr_cont("%s ", empty_str
);
1502 * pcpu_setup_first_chunk - initialize the first percpu chunk
1503 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1504 * @base_addr: mapped address
1506 * Initialize the first percpu chunk which contains the kernel static
1507 * perpcu area. This function is to be called from arch percpu area
1510 * @ai contains all information necessary to initialize the first
1511 * chunk and prime the dynamic percpu allocator.
1513 * @ai->static_size is the size of static percpu area.
1515 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1516 * reserve after the static area in the first chunk. This reserves
1517 * the first chunk such that it's available only through reserved
1518 * percpu allocation. This is primarily used to serve module percpu
1519 * static areas on architectures where the addressing model has
1520 * limited offset range for symbol relocations to guarantee module
1521 * percpu symbols fall inside the relocatable range.
1523 * @ai->dyn_size determines the number of bytes available for dynamic
1524 * allocation in the first chunk. The area between @ai->static_size +
1525 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1527 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1528 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1531 * @ai->atom_size is the allocation atom size and used as alignment
1534 * @ai->alloc_size is the allocation size and always multiple of
1535 * @ai->atom_size. This is larger than @ai->atom_size if
1536 * @ai->unit_size is larger than @ai->atom_size.
1538 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1539 * percpu areas. Units which should be colocated are put into the
1540 * same group. Dynamic VM areas will be allocated according to these
1541 * groupings. If @ai->nr_groups is zero, a single group containing
1542 * all units is assumed.
1544 * The caller should have mapped the first chunk at @base_addr and
1545 * copied static data to each unit.
1547 * If the first chunk ends up with both reserved and dynamic areas, it
1548 * is served by two chunks - one to serve the core static and reserved
1549 * areas and the other for the dynamic area. They share the same vm
1550 * and page map but uses different area allocation map to stay away
1551 * from each other. The latter chunk is circulated in the chunk slots
1552 * and available for dynamic allocation like any other chunks.
1555 * 0 on success, -errno on failure.
1557 int __init
pcpu_setup_first_chunk(const struct pcpu_alloc_info
*ai
,
1560 static int smap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1561 static int dmap
[PERCPU_DYNAMIC_EARLY_SLOTS
] __initdata
;
1562 size_t dyn_size
= ai
->dyn_size
;
1563 size_t size_sum
= ai
->static_size
+ ai
->reserved_size
+ dyn_size
;
1564 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1565 unsigned long *group_offsets
;
1566 size_t *group_sizes
;
1567 unsigned long *unit_off
;
1572 #define PCPU_SETUP_BUG_ON(cond) do { \
1573 if (unlikely(cond)) { \
1574 pr_emerg("failed to initialize, %s\n", #cond); \
1575 pr_emerg("cpu_possible_mask=%*pb\n", \
1576 cpumask_pr_args(cpu_possible_mask)); \
1577 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1583 PCPU_SETUP_BUG_ON(ai
->nr_groups
<= 0);
1585 PCPU_SETUP_BUG_ON(!ai
->static_size
);
1586 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start
));
1588 PCPU_SETUP_BUG_ON(!base_addr
);
1589 PCPU_SETUP_BUG_ON(offset_in_page(base_addr
));
1590 PCPU_SETUP_BUG_ON(ai
->unit_size
< size_sum
);
1591 PCPU_SETUP_BUG_ON(offset_in_page(ai
->unit_size
));
1592 PCPU_SETUP_BUG_ON(ai
->unit_size
< PCPU_MIN_UNIT_SIZE
);
1593 PCPU_SETUP_BUG_ON(ai
->dyn_size
< PERCPU_DYNAMIC_EARLY_SIZE
);
1594 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai
) < 0);
1596 /* process group information and build config tables accordingly */
1597 group_offsets
= memblock_virt_alloc(ai
->nr_groups
*
1598 sizeof(group_offsets
[0]), 0);
1599 group_sizes
= memblock_virt_alloc(ai
->nr_groups
*
1600 sizeof(group_sizes
[0]), 0);
1601 unit_map
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_map
[0]), 0);
1602 unit_off
= memblock_virt_alloc(nr_cpu_ids
* sizeof(unit_off
[0]), 0);
1604 for (cpu
= 0; cpu
< nr_cpu_ids
; cpu
++)
1605 unit_map
[cpu
] = UINT_MAX
;
1607 pcpu_low_unit_cpu
= NR_CPUS
;
1608 pcpu_high_unit_cpu
= NR_CPUS
;
1610 for (group
= 0, unit
= 0; group
< ai
->nr_groups
; group
++, unit
+= i
) {
1611 const struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1613 group_offsets
[group
] = gi
->base_offset
;
1614 group_sizes
[group
] = gi
->nr_units
* ai
->unit_size
;
1616 for (i
= 0; i
< gi
->nr_units
; i
++) {
1617 cpu
= gi
->cpu_map
[i
];
1621 PCPU_SETUP_BUG_ON(cpu
>= nr_cpu_ids
);
1622 PCPU_SETUP_BUG_ON(!cpu_possible(cpu
));
1623 PCPU_SETUP_BUG_ON(unit_map
[cpu
] != UINT_MAX
);
1625 unit_map
[cpu
] = unit
+ i
;
1626 unit_off
[cpu
] = gi
->base_offset
+ i
* ai
->unit_size
;
1628 /* determine low/high unit_cpu */
1629 if (pcpu_low_unit_cpu
== NR_CPUS
||
1630 unit_off
[cpu
] < unit_off
[pcpu_low_unit_cpu
])
1631 pcpu_low_unit_cpu
= cpu
;
1632 if (pcpu_high_unit_cpu
== NR_CPUS
||
1633 unit_off
[cpu
] > unit_off
[pcpu_high_unit_cpu
])
1634 pcpu_high_unit_cpu
= cpu
;
1637 pcpu_nr_units
= unit
;
1639 for_each_possible_cpu(cpu
)
1640 PCPU_SETUP_BUG_ON(unit_map
[cpu
] == UINT_MAX
);
1642 /* we're done parsing the input, undefine BUG macro and dump config */
1643 #undef PCPU_SETUP_BUG_ON
1644 pcpu_dump_alloc_info(KERN_DEBUG
, ai
);
1646 pcpu_nr_groups
= ai
->nr_groups
;
1647 pcpu_group_offsets
= group_offsets
;
1648 pcpu_group_sizes
= group_sizes
;
1649 pcpu_unit_map
= unit_map
;
1650 pcpu_unit_offsets
= unit_off
;
1652 /* determine basic parameters */
1653 pcpu_unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
1654 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1655 pcpu_atom_size
= ai
->atom_size
;
1656 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1657 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1659 pcpu_stats_save_ai(ai
);
1662 * Allocate chunk slots. The additional last slot is for
1665 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1666 pcpu_slot
= memblock_virt_alloc(
1667 pcpu_nr_slots
* sizeof(pcpu_slot
[0]), 0);
1668 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1669 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1672 * Initialize static chunk. If reserved_size is zero, the
1673 * static chunk covers static area + dynamic allocation area
1674 * in the first chunk. If reserved_size is not zero, it
1675 * covers static area + reserved area (mostly used for module
1676 * static percpu allocation).
1678 schunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1679 INIT_LIST_HEAD(&schunk
->list
);
1680 INIT_LIST_HEAD(&schunk
->map_extend_list
);
1681 schunk
->base_addr
= base_addr
;
1683 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1684 schunk
->immutable
= true;
1685 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1686 schunk
->nr_populated
= pcpu_unit_pages
;
1688 if (ai
->reserved_size
) {
1689 schunk
->free_size
= ai
->reserved_size
;
1690 pcpu_reserved_chunk
= schunk
;
1691 pcpu_reserved_chunk_limit
= ai
->static_size
+ ai
->reserved_size
;
1693 schunk
->free_size
= dyn_size
;
1694 dyn_size
= 0; /* dynamic area covered */
1696 schunk
->contig_hint
= schunk
->free_size
;
1699 schunk
->map
[1] = ai
->static_size
;
1700 schunk
->map_used
= 1;
1701 if (schunk
->free_size
)
1702 schunk
->map
[++schunk
->map_used
] = ai
->static_size
+ schunk
->free_size
;
1703 schunk
->map
[schunk
->map_used
] |= 1;
1704 schunk
->has_reserved
= true;
1706 /* init dynamic chunk if necessary */
1708 dchunk
= memblock_virt_alloc(pcpu_chunk_struct_size
, 0);
1709 INIT_LIST_HEAD(&dchunk
->list
);
1710 INIT_LIST_HEAD(&dchunk
->map_extend_list
);
1711 dchunk
->base_addr
= base_addr
;
1713 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1714 dchunk
->immutable
= true;
1715 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1716 dchunk
->nr_populated
= pcpu_unit_pages
;
1718 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1720 dchunk
->map
[1] = pcpu_reserved_chunk_limit
;
1721 dchunk
->map
[2] = (pcpu_reserved_chunk_limit
+ dchunk
->free_size
) | 1;
1722 dchunk
->map_used
= 2;
1723 dchunk
->has_reserved
= true;
1726 /* link the first chunk in */
1727 pcpu_first_chunk
= dchunk
?: schunk
;
1728 pcpu_nr_empty_pop_pages
+=
1729 pcpu_count_occupied_pages(pcpu_first_chunk
, 1);
1730 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1732 pcpu_stats_chunk_alloc();
1733 trace_percpu_create_chunk(base_addr
);
1736 pcpu_base_addr
= base_addr
;
1742 const char * const pcpu_fc_names
[PCPU_FC_NR
] __initconst
= {
1743 [PCPU_FC_AUTO
] = "auto",
1744 [PCPU_FC_EMBED
] = "embed",
1745 [PCPU_FC_PAGE
] = "page",
1748 enum pcpu_fc pcpu_chosen_fc __initdata
= PCPU_FC_AUTO
;
1750 static int __init
percpu_alloc_setup(char *str
)
1757 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1758 else if (!strcmp(str
, "embed"))
1759 pcpu_chosen_fc
= PCPU_FC_EMBED
;
1761 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1762 else if (!strcmp(str
, "page"))
1763 pcpu_chosen_fc
= PCPU_FC_PAGE
;
1766 pr_warn("unknown allocator %s specified\n", str
);
1770 early_param("percpu_alloc", percpu_alloc_setup
);
1773 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1774 * Build it if needed by the arch config or the generic setup is going
1777 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1778 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1779 #define BUILD_EMBED_FIRST_CHUNK
1782 /* build pcpu_page_first_chunk() iff needed by the arch config */
1783 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1784 #define BUILD_PAGE_FIRST_CHUNK
1787 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1788 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1790 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1791 * @reserved_size: the size of reserved percpu area in bytes
1792 * @dyn_size: minimum free size for dynamic allocation in bytes
1793 * @atom_size: allocation atom size
1794 * @cpu_distance_fn: callback to determine distance between cpus, optional
1796 * This function determines grouping of units, their mappings to cpus
1797 * and other parameters considering needed percpu size, allocation
1798 * atom size and distances between CPUs.
1800 * Groups are always multiples of atom size and CPUs which are of
1801 * LOCAL_DISTANCE both ways are grouped together and share space for
1802 * units in the same group. The returned configuration is guaranteed
1803 * to have CPUs on different nodes on different groups and >=75% usage
1804 * of allocated virtual address space.
1807 * On success, pointer to the new allocation_info is returned. On
1808 * failure, ERR_PTR value is returned.
1810 static struct pcpu_alloc_info
* __init
pcpu_build_alloc_info(
1811 size_t reserved_size
, size_t dyn_size
,
1813 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
)
1815 static int group_map
[NR_CPUS
] __initdata
;
1816 static int group_cnt
[NR_CPUS
] __initdata
;
1817 const size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1818 int nr_groups
= 1, nr_units
= 0;
1819 size_t size_sum
, min_unit_size
, alloc_size
;
1820 int upa
, max_upa
, uninitialized_var(best_upa
); /* units_per_alloc */
1821 int last_allocs
, group
, unit
;
1822 unsigned int cpu
, tcpu
;
1823 struct pcpu_alloc_info
*ai
;
1824 unsigned int *cpu_map
;
1826 /* this function may be called multiple times */
1827 memset(group_map
, 0, sizeof(group_map
));
1828 memset(group_cnt
, 0, sizeof(group_cnt
));
1830 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1831 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1832 max_t(size_t, dyn_size
, PERCPU_DYNAMIC_EARLY_SIZE
));
1833 dyn_size
= size_sum
- static_size
- reserved_size
;
1836 * Determine min_unit_size, alloc_size and max_upa such that
1837 * alloc_size is multiple of atom_size and is the smallest
1838 * which can accommodate 4k aligned segments which are equal to
1839 * or larger than min_unit_size.
1841 min_unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1843 alloc_size
= roundup(min_unit_size
, atom_size
);
1844 upa
= alloc_size
/ min_unit_size
;
1845 while (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1849 /* group cpus according to their proximity */
1850 for_each_possible_cpu(cpu
) {
1853 for_each_possible_cpu(tcpu
) {
1856 if (group_map
[tcpu
] == group
&& cpu_distance_fn
&&
1857 (cpu_distance_fn(cpu
, tcpu
) > LOCAL_DISTANCE
||
1858 cpu_distance_fn(tcpu
, cpu
) > LOCAL_DISTANCE
)) {
1860 nr_groups
= max(nr_groups
, group
+ 1);
1864 group_map
[cpu
] = group
;
1869 * Expand unit size until address space usage goes over 75%
1870 * and then as much as possible without using more address
1873 last_allocs
= INT_MAX
;
1874 for (upa
= max_upa
; upa
; upa
--) {
1875 int allocs
= 0, wasted
= 0;
1877 if (alloc_size
% upa
|| (offset_in_page(alloc_size
/ upa
)))
1880 for (group
= 0; group
< nr_groups
; group
++) {
1881 int this_allocs
= DIV_ROUND_UP(group_cnt
[group
], upa
);
1882 allocs
+= this_allocs
;
1883 wasted
+= this_allocs
* upa
- group_cnt
[group
];
1887 * Don't accept if wastage is over 1/3. The
1888 * greater-than comparison ensures upa==1 always
1889 * passes the following check.
1891 if (wasted
> num_possible_cpus() / 3)
1894 /* and then don't consume more memory */
1895 if (allocs
> last_allocs
)
1897 last_allocs
= allocs
;
1902 /* allocate and fill alloc_info */
1903 for (group
= 0; group
< nr_groups
; group
++)
1904 nr_units
+= roundup(group_cnt
[group
], upa
);
1906 ai
= pcpu_alloc_alloc_info(nr_groups
, nr_units
);
1908 return ERR_PTR(-ENOMEM
);
1909 cpu_map
= ai
->groups
[0].cpu_map
;
1911 for (group
= 0; group
< nr_groups
; group
++) {
1912 ai
->groups
[group
].cpu_map
= cpu_map
;
1913 cpu_map
+= roundup(group_cnt
[group
], upa
);
1916 ai
->static_size
= static_size
;
1917 ai
->reserved_size
= reserved_size
;
1918 ai
->dyn_size
= dyn_size
;
1919 ai
->unit_size
= alloc_size
/ upa
;
1920 ai
->atom_size
= atom_size
;
1921 ai
->alloc_size
= alloc_size
;
1923 for (group
= 0, unit
= 0; group_cnt
[group
]; group
++) {
1924 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
1927 * Initialize base_offset as if all groups are located
1928 * back-to-back. The caller should update this to
1929 * reflect actual allocation.
1931 gi
->base_offset
= unit
* ai
->unit_size
;
1933 for_each_possible_cpu(cpu
)
1934 if (group_map
[cpu
] == group
)
1935 gi
->cpu_map
[gi
->nr_units
++] = cpu
;
1936 gi
->nr_units
= roundup(gi
->nr_units
, upa
);
1937 unit
+= gi
->nr_units
;
1939 BUG_ON(unit
!= nr_units
);
1943 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1945 #if defined(BUILD_EMBED_FIRST_CHUNK)
1947 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1948 * @reserved_size: the size of reserved percpu area in bytes
1949 * @dyn_size: minimum free size for dynamic allocation in bytes
1950 * @atom_size: allocation atom size
1951 * @cpu_distance_fn: callback to determine distance between cpus, optional
1952 * @alloc_fn: function to allocate percpu page
1953 * @free_fn: function to free percpu page
1955 * This is a helper to ease setting up embedded first percpu chunk and
1956 * can be called where pcpu_setup_first_chunk() is expected.
1958 * If this function is used to setup the first chunk, it is allocated
1959 * by calling @alloc_fn and used as-is without being mapped into
1960 * vmalloc area. Allocations are always whole multiples of @atom_size
1961 * aligned to @atom_size.
1963 * This enables the first chunk to piggy back on the linear physical
1964 * mapping which often uses larger page size. Please note that this
1965 * can result in very sparse cpu->unit mapping on NUMA machines thus
1966 * requiring large vmalloc address space. Don't use this allocator if
1967 * vmalloc space is not orders of magnitude larger than distances
1968 * between node memory addresses (ie. 32bit NUMA machines).
1970 * @dyn_size specifies the minimum dynamic area size.
1972 * If the needed size is smaller than the minimum or specified unit
1973 * size, the leftover is returned using @free_fn.
1976 * 0 on success, -errno on failure.
1978 int __init
pcpu_embed_first_chunk(size_t reserved_size
, size_t dyn_size
,
1980 pcpu_fc_cpu_distance_fn_t cpu_distance_fn
,
1981 pcpu_fc_alloc_fn_t alloc_fn
,
1982 pcpu_fc_free_fn_t free_fn
)
1984 void *base
= (void *)ULONG_MAX
;
1985 void **areas
= NULL
;
1986 struct pcpu_alloc_info
*ai
;
1987 size_t size_sum
, areas_size
;
1988 unsigned long max_distance
;
1989 int group
, i
, highest_group
, rc
;
1991 ai
= pcpu_build_alloc_info(reserved_size
, dyn_size
, atom_size
,
1996 size_sum
= ai
->static_size
+ ai
->reserved_size
+ ai
->dyn_size
;
1997 areas_size
= PFN_ALIGN(ai
->nr_groups
* sizeof(void *));
1999 areas
= memblock_virt_alloc_nopanic(areas_size
, 0);
2005 /* allocate, copy and determine base address & max_distance */
2007 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2008 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2009 unsigned int cpu
= NR_CPUS
;
2012 for (i
= 0; i
< gi
->nr_units
&& cpu
== NR_CPUS
; i
++)
2013 cpu
= gi
->cpu_map
[i
];
2014 BUG_ON(cpu
== NR_CPUS
);
2016 /* allocate space for the whole group */
2017 ptr
= alloc_fn(cpu
, gi
->nr_units
* ai
->unit_size
, atom_size
);
2020 goto out_free_areas
;
2022 /* kmemleak tracks the percpu allocations separately */
2026 base
= min(ptr
, base
);
2027 if (ptr
> areas
[highest_group
])
2028 highest_group
= group
;
2030 max_distance
= areas
[highest_group
] - base
;
2031 max_distance
+= ai
->unit_size
* ai
->groups
[highest_group
].nr_units
;
2033 /* warn if maximum distance is further than 75% of vmalloc space */
2034 if (max_distance
> VMALLOC_TOTAL
* 3 / 4) {
2035 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2036 max_distance
, VMALLOC_TOTAL
);
2037 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2038 /* and fail if we have fallback */
2040 goto out_free_areas
;
2045 * Copy data and free unused parts. This should happen after all
2046 * allocations are complete; otherwise, we may end up with
2047 * overlapping groups.
2049 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2050 struct pcpu_group_info
*gi
= &ai
->groups
[group
];
2051 void *ptr
= areas
[group
];
2053 for (i
= 0; i
< gi
->nr_units
; i
++, ptr
+= ai
->unit_size
) {
2054 if (gi
->cpu_map
[i
] == NR_CPUS
) {
2055 /* unused unit, free whole */
2056 free_fn(ptr
, ai
->unit_size
);
2059 /* copy and return the unused part */
2060 memcpy(ptr
, __per_cpu_load
, ai
->static_size
);
2061 free_fn(ptr
+ size_sum
, ai
->unit_size
- size_sum
);
2065 /* base address is now known, determine group base offsets */
2066 for (group
= 0; group
< ai
->nr_groups
; group
++) {
2067 ai
->groups
[group
].base_offset
= areas
[group
] - base
;
2070 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2071 PFN_DOWN(size_sum
), base
, ai
->static_size
, ai
->reserved_size
,
2072 ai
->dyn_size
, ai
->unit_size
);
2074 rc
= pcpu_setup_first_chunk(ai
, base
);
2078 for (group
= 0; group
< ai
->nr_groups
; group
++)
2080 free_fn(areas
[group
],
2081 ai
->groups
[group
].nr_units
* ai
->unit_size
);
2083 pcpu_free_alloc_info(ai
);
2085 memblock_free_early(__pa(areas
), areas_size
);
2088 #endif /* BUILD_EMBED_FIRST_CHUNK */
2090 #ifdef BUILD_PAGE_FIRST_CHUNK
2092 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2093 * @reserved_size: the size of reserved percpu area in bytes
2094 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2095 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2096 * @populate_pte_fn: function to populate pte
2098 * This is a helper to ease setting up page-remapped first percpu
2099 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2101 * This is the basic allocator. Static percpu area is allocated
2102 * page-by-page into vmalloc area.
2105 * 0 on success, -errno on failure.
2107 int __init
pcpu_page_first_chunk(size_t reserved_size
,
2108 pcpu_fc_alloc_fn_t alloc_fn
,
2109 pcpu_fc_free_fn_t free_fn
,
2110 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
2112 static struct vm_struct vm
;
2113 struct pcpu_alloc_info
*ai
;
2117 struct page
**pages
;
2122 snprintf(psize_str
, sizeof(psize_str
), "%luK", PAGE_SIZE
>> 10);
2124 ai
= pcpu_build_alloc_info(reserved_size
, 0, PAGE_SIZE
, NULL
);
2127 BUG_ON(ai
->nr_groups
!= 1);
2128 upa
= ai
->alloc_size
/ai
->unit_size
;
2129 nr_g0_units
= roundup(num_possible_cpus(), upa
);
2130 if (unlikely(WARN_ON(ai
->groups
[0].nr_units
!= nr_g0_units
))) {
2131 pcpu_free_alloc_info(ai
);
2135 unit_pages
= ai
->unit_size
>> PAGE_SHIFT
;
2137 /* unaligned allocations can't be freed, round up to page size */
2138 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
2140 pages
= memblock_virt_alloc(pages_size
, 0);
2142 /* allocate pages */
2144 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2145 unsigned int cpu
= ai
->groups
[0].cpu_map
[unit
];
2146 for (i
= 0; i
< unit_pages
; i
++) {
2149 ptr
= alloc_fn(cpu
, PAGE_SIZE
, PAGE_SIZE
);
2151 pr_warn("failed to allocate %s page for cpu%u\n",
2155 /* kmemleak tracks the percpu allocations separately */
2157 pages
[j
++] = virt_to_page(ptr
);
2161 /* allocate vm area, map the pages and copy static data */
2162 vm
.flags
= VM_ALLOC
;
2163 vm
.size
= num_possible_cpus() * ai
->unit_size
;
2164 vm_area_register_early(&vm
, PAGE_SIZE
);
2166 for (unit
= 0; unit
< num_possible_cpus(); unit
++) {
2167 unsigned long unit_addr
=
2168 (unsigned long)vm
.addr
+ unit
* ai
->unit_size
;
2170 for (i
= 0; i
< unit_pages
; i
++)
2171 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
2173 /* pte already populated, the following shouldn't fail */
2174 rc
= __pcpu_map_pages(unit_addr
, &pages
[unit
* unit_pages
],
2177 panic("failed to map percpu area, err=%d\n", rc
);
2180 * FIXME: Archs with virtual cache should flush local
2181 * cache for the linear mapping here - something
2182 * equivalent to flush_cache_vmap() on the local cpu.
2183 * flush_cache_vmap() can't be used as most supporting
2184 * data structures are not set up yet.
2187 /* copy static data */
2188 memcpy((void *)unit_addr
, __per_cpu_load
, ai
->static_size
);
2191 /* we're ready, commit */
2192 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2193 unit_pages
, psize_str
, vm
.addr
, ai
->static_size
,
2194 ai
->reserved_size
, ai
->dyn_size
);
2196 rc
= pcpu_setup_first_chunk(ai
, vm
.addr
);
2201 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
2204 memblock_free_early(__pa(pages
), pages_size
);
2205 pcpu_free_alloc_info(ai
);
2208 #endif /* BUILD_PAGE_FIRST_CHUNK */
2210 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2212 * Generic SMP percpu area setup.
2214 * The embedding helper is used because its behavior closely resembles
2215 * the original non-dynamic generic percpu area setup. This is
2216 * important because many archs have addressing restrictions and might
2217 * fail if the percpu area is located far away from the previous
2218 * location. As an added bonus, in non-NUMA cases, embedding is
2219 * generally a good idea TLB-wise because percpu area can piggy back
2220 * on the physical linear memory mapping which uses large page
2221 * mappings on applicable archs.
2223 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
2224 EXPORT_SYMBOL(__per_cpu_offset
);
2226 static void * __init
pcpu_dfl_fc_alloc(unsigned int cpu
, size_t size
,
2229 return memblock_virt_alloc_from_nopanic(
2230 size
, align
, __pa(MAX_DMA_ADDRESS
));
2233 static void __init
pcpu_dfl_fc_free(void *ptr
, size_t size
)
2235 memblock_free_early(__pa(ptr
), size
);
2238 void __init
setup_per_cpu_areas(void)
2240 unsigned long delta
;
2245 * Always reserve area for module percpu variables. That's
2246 * what the legacy allocator did.
2248 rc
= pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE
,
2249 PERCPU_DYNAMIC_RESERVE
, PAGE_SIZE
, NULL
,
2250 pcpu_dfl_fc_alloc
, pcpu_dfl_fc_free
);
2252 panic("Failed to initialize percpu areas.");
2254 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
2255 for_each_possible_cpu(cpu
)
2256 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
2258 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2260 #else /* CONFIG_SMP */
2263 * UP percpu area setup.
2265 * UP always uses km-based percpu allocator with identity mapping.
2266 * Static percpu variables are indistinguishable from the usual static
2267 * variables and don't require any special preparation.
2269 void __init
setup_per_cpu_areas(void)
2271 const size_t unit_size
=
2272 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE
,
2273 PERCPU_DYNAMIC_RESERVE
));
2274 struct pcpu_alloc_info
*ai
;
2277 ai
= pcpu_alloc_alloc_info(1, 1);
2278 fc
= memblock_virt_alloc_from_nopanic(unit_size
,
2280 __pa(MAX_DMA_ADDRESS
));
2282 panic("Failed to allocate memory for percpu areas.");
2283 /* kmemleak tracks the percpu allocations separately */
2286 ai
->dyn_size
= unit_size
;
2287 ai
->unit_size
= unit_size
;
2288 ai
->atom_size
= unit_size
;
2289 ai
->alloc_size
= unit_size
;
2290 ai
->groups
[0].nr_units
= 1;
2291 ai
->groups
[0].cpu_map
[0] = 0;
2293 if (pcpu_setup_first_chunk(ai
, fc
) < 0)
2294 panic("Failed to initialize percpu areas.");
2297 #endif /* CONFIG_SMP */
2300 * First and reserved chunks are initialized with temporary allocation
2301 * map in initdata so that they can be used before slab is online.
2302 * This function is called after slab is brought up and replaces those
2303 * with properly allocated maps.
2305 void __init
percpu_init_late(void)
2307 struct pcpu_chunk
*target_chunks
[] =
2308 { pcpu_first_chunk
, pcpu_reserved_chunk
, NULL
};
2309 struct pcpu_chunk
*chunk
;
2310 unsigned long flags
;
2313 for (i
= 0; (chunk
= target_chunks
[i
]); i
++) {
2315 const size_t size
= PERCPU_DYNAMIC_EARLY_SLOTS
* sizeof(map
[0]);
2317 BUILD_BUG_ON(size
> PAGE_SIZE
);
2319 map
= pcpu_mem_zalloc(size
);
2322 spin_lock_irqsave(&pcpu_lock
, flags
);
2323 memcpy(map
, chunk
->map
, size
);
2325 spin_unlock_irqrestore(&pcpu_lock
, flags
);
2330 * Percpu allocator is initialized early during boot when neither slab or
2331 * workqueue is available. Plug async management until everything is up
2334 static int __init
percpu_enable_async(void)
2336 pcpu_async_enabled
= true;
2339 subsys_initcall(percpu_enable_async
);