2 * linux/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 in vmalloc area. Each
11 * chunk is consisted of boot-time determined number of units and the
12 * first 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. ie. in
20 * ------------------- ------------------- ------------
21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
22 * ------------------- ...... ------------------- .... ------------
24 * Allocation is done in offset-size areas of single unit space. Ie,
25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
31 * There are usually many small percpu allocations many of them being
32 * as small as 4 bytes. The allocator organizes chunks into lists
33 * according to free size and tries to allocate from the fullest one.
34 * Each chunk keeps the maximum contiguous area size hint which is
35 * guaranteed to be eqaul to or larger than the maximum contiguous
36 * area in the chunk. This helps the allocator not to iterate the
37 * chunk maps unnecessarily.
39 * Allocation state in each chunk is kept using an array of integers
40 * on chunk->map. A positive value in the map represents a free
41 * region and negative allocated. Allocation inside a chunk is done
42 * by scanning this map sequentially and serving the first matching
43 * entry. This is mostly copied from the percpu_modalloc() allocator.
44 * Chunks can be determined from the address using the index field
45 * in the page struct. The index field contains a pointer to the chunk.
47 * To use this allocator, arch code should do the followings.
49 * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA
51 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
52 * regular address to percpu pointer and back if they need to be
53 * different from the default
55 * - use pcpu_setup_first_chunk() during percpu area initialization to
56 * setup the first chunk containing the kernel static percpu area
59 #include <linux/bitmap.h>
60 #include <linux/bootmem.h>
61 #include <linux/list.h>
63 #include <linux/module.h>
64 #include <linux/mutex.h>
65 #include <linux/percpu.h>
66 #include <linux/pfn.h>
67 #include <linux/slab.h>
68 #include <linux/spinlock.h>
69 #include <linux/vmalloc.h>
70 #include <linux/workqueue.h>
72 #include <asm/cacheflush.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
76 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
77 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
80 #ifndef __addr_to_pcpu_ptr
81 #define __addr_to_pcpu_ptr(addr) \
82 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
83 + (unsigned long)__per_cpu_start)
85 #ifndef __pcpu_ptr_to_addr
86 #define __pcpu_ptr_to_addr(ptr) \
87 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
88 - (unsigned long)__per_cpu_start)
92 struct list_head list
; /* linked to pcpu_slot lists */
93 int free_size
; /* free bytes in the chunk */
94 int contig_hint
; /* max contiguous size hint */
95 struct vm_struct
*vm
; /* mapped vmalloc region */
96 int map_used
; /* # of map entries used */
97 int map_alloc
; /* # of map entries allocated */
98 int *map
; /* allocation map */
99 bool immutable
; /* no [de]population allowed */
100 unsigned long populated
[]; /* populated bitmap */
103 static int pcpu_unit_pages __read_mostly
;
104 static int pcpu_unit_size __read_mostly
;
105 static int pcpu_nr_units __read_mostly
;
106 static int pcpu_chunk_size __read_mostly
;
107 static int pcpu_nr_slots __read_mostly
;
108 static size_t pcpu_chunk_struct_size __read_mostly
;
110 /* cpus with the lowest and highest unit numbers */
111 static unsigned int pcpu_first_unit_cpu __read_mostly
;
112 static unsigned int pcpu_last_unit_cpu __read_mostly
;
114 /* the address of the first chunk which starts with the kernel static area */
115 void *pcpu_base_addr __read_mostly
;
116 EXPORT_SYMBOL_GPL(pcpu_base_addr
);
118 /* cpu -> unit map */
119 const int *pcpu_unit_map __read_mostly
;
122 * The first chunk which always exists. Note that unlike other
123 * chunks, this one can be allocated and mapped in several different
124 * ways and thus often doesn't live in the vmalloc area.
126 static struct pcpu_chunk
*pcpu_first_chunk
;
129 * Optional reserved chunk. This chunk reserves part of the first
130 * chunk and serves it for reserved allocations. The amount of
131 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
132 * area doesn't exist, the following variables contain NULL and 0
135 static struct pcpu_chunk
*pcpu_reserved_chunk
;
136 static int pcpu_reserved_chunk_limit
;
139 * Synchronization rules.
141 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
142 * protects allocation/reclaim paths, chunks, populated bitmap and
143 * vmalloc mapping. The latter is a spinlock and protects the index
144 * data structures - chunk slots, chunks and area maps in chunks.
146 * During allocation, pcpu_alloc_mutex is kept locked all the time and
147 * pcpu_lock is grabbed and released as necessary. All actual memory
148 * allocations are done using GFP_KERNEL with pcpu_lock released.
150 * Free path accesses and alters only the index data structures, so it
151 * can be safely called from atomic context. When memory needs to be
152 * returned to the system, free path schedules reclaim_work which
153 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
154 * reclaimed, release both locks and frees the chunks. Note that it's
155 * necessary to grab both locks to remove a chunk from circulation as
156 * allocation path might be referencing the chunk with only
157 * pcpu_alloc_mutex locked.
159 static DEFINE_MUTEX(pcpu_alloc_mutex
); /* protects whole alloc and reclaim */
160 static DEFINE_SPINLOCK(pcpu_lock
); /* protects index data structures */
162 static struct list_head
*pcpu_slot __read_mostly
; /* chunk list slots */
164 /* reclaim work to release fully free chunks, scheduled from free path */
165 static void pcpu_reclaim(struct work_struct
*work
);
166 static DECLARE_WORK(pcpu_reclaim_work
, pcpu_reclaim
);
168 static int __pcpu_size_to_slot(int size
)
170 int highbit
= fls(size
); /* size is in bytes */
171 return max(highbit
- PCPU_SLOT_BASE_SHIFT
+ 2, 1);
174 static int pcpu_size_to_slot(int size
)
176 if (size
== pcpu_unit_size
)
177 return pcpu_nr_slots
- 1;
178 return __pcpu_size_to_slot(size
);
181 static int pcpu_chunk_slot(const struct pcpu_chunk
*chunk
)
183 if (chunk
->free_size
< sizeof(int) || chunk
->contig_hint
< sizeof(int))
186 return pcpu_size_to_slot(chunk
->free_size
);
189 static int pcpu_page_idx(unsigned int cpu
, int page_idx
)
191 return pcpu_unit_map
[cpu
] * pcpu_unit_pages
+ page_idx
;
194 static unsigned long pcpu_chunk_addr(struct pcpu_chunk
*chunk
,
195 unsigned int cpu
, int page_idx
)
197 return (unsigned long)chunk
->vm
->addr
+
198 (pcpu_page_idx(cpu
, page_idx
) << PAGE_SHIFT
);
201 static struct page
*pcpu_chunk_page(struct pcpu_chunk
*chunk
,
202 unsigned int cpu
, int page_idx
)
204 /* must not be used on pre-mapped chunk */
205 WARN_ON(chunk
->immutable
);
207 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk
, cpu
, page_idx
));
210 /* set the pointer to a chunk in a page struct */
211 static void pcpu_set_page_chunk(struct page
*page
, struct pcpu_chunk
*pcpu
)
213 page
->index
= (unsigned long)pcpu
;
216 /* obtain pointer to a chunk from a page struct */
217 static struct pcpu_chunk
*pcpu_get_page_chunk(struct page
*page
)
219 return (struct pcpu_chunk
*)page
->index
;
222 static void pcpu_next_unpop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
224 *rs
= find_next_zero_bit(chunk
->populated
, end
, *rs
);
225 *re
= find_next_bit(chunk
->populated
, end
, *rs
+ 1);
228 static void pcpu_next_pop(struct pcpu_chunk
*chunk
, int *rs
, int *re
, int end
)
230 *rs
= find_next_bit(chunk
->populated
, end
, *rs
);
231 *re
= find_next_zero_bit(chunk
->populated
, end
, *rs
+ 1);
235 * (Un)populated page region iterators. Iterate over (un)populated
236 * page regions betwen @start and @end in @chunk. @rs and @re should
237 * be integer variables and will be set to start and end page index of
238 * the current region.
240 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
241 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
243 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
245 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
246 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
248 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
251 * pcpu_mem_alloc - allocate memory
252 * @size: bytes to allocate
254 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
255 * kzalloc() is used; otherwise, vmalloc() is used. The returned
256 * memory is always zeroed.
259 * Does GFP_KERNEL allocation.
262 * Pointer to the allocated area on success, NULL on failure.
264 static void *pcpu_mem_alloc(size_t size
)
266 if (size
<= PAGE_SIZE
)
267 return kzalloc(size
, GFP_KERNEL
);
269 void *ptr
= vmalloc(size
);
271 memset(ptr
, 0, size
);
277 * pcpu_mem_free - free memory
278 * @ptr: memory to free
279 * @size: size of the area
281 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
283 static void pcpu_mem_free(void *ptr
, size_t size
)
285 if (size
<= PAGE_SIZE
)
292 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
293 * @chunk: chunk of interest
294 * @oslot: the previous slot it was on
296 * This function is called after an allocation or free changed @chunk.
297 * New slot according to the changed state is determined and @chunk is
298 * moved to the slot. Note that the reserved chunk is never put on
304 static void pcpu_chunk_relocate(struct pcpu_chunk
*chunk
, int oslot
)
306 int nslot
= pcpu_chunk_slot(chunk
);
308 if (chunk
!= pcpu_reserved_chunk
&& oslot
!= nslot
) {
310 list_move(&chunk
->list
, &pcpu_slot
[nslot
]);
312 list_move_tail(&chunk
->list
, &pcpu_slot
[nslot
]);
317 * pcpu_chunk_addr_search - determine chunk containing specified address
318 * @addr: address for which the chunk needs to be determined.
321 * The address of the found chunk.
323 static struct pcpu_chunk
*pcpu_chunk_addr_search(void *addr
)
325 void *first_start
= pcpu_first_chunk
->vm
->addr
;
327 /* is it in the first chunk? */
328 if (addr
>= first_start
&& addr
< first_start
+ pcpu_unit_size
) {
329 /* is it in the reserved area? */
330 if (addr
< first_start
+ pcpu_reserved_chunk_limit
)
331 return pcpu_reserved_chunk
;
332 return pcpu_first_chunk
;
336 * The address is relative to unit0 which might be unused and
337 * thus unmapped. Offset the address to the unit space of the
338 * current processor before looking it up in the vmalloc
339 * space. Note that any possible cpu id can be used here, so
340 * there's no need to worry about preemption or cpu hotplug.
342 addr
+= pcpu_unit_map
[smp_processor_id()] * pcpu_unit_size
;
343 return pcpu_get_page_chunk(vmalloc_to_page(addr
));
347 * pcpu_extend_area_map - extend area map for allocation
348 * @chunk: target chunk
350 * Extend area map of @chunk so that it can accomodate an allocation.
351 * A single allocation can split an area into three areas, so this
352 * function makes sure that @chunk->map has at least two extra slots.
355 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
356 * if area map is extended.
359 * 0 if noop, 1 if successfully extended, -errno on failure.
361 static int pcpu_extend_area_map(struct pcpu_chunk
*chunk
)
368 if (chunk
->map_alloc
>= chunk
->map_used
+ 2)
371 spin_unlock_irq(&pcpu_lock
);
373 new_alloc
= PCPU_DFL_MAP_ALLOC
;
374 while (new_alloc
< chunk
->map_used
+ 2)
377 new = pcpu_mem_alloc(new_alloc
* sizeof(new[0]));
379 spin_lock_irq(&pcpu_lock
);
384 * Acquire pcpu_lock and switch to new area map. Only free
385 * could have happened inbetween, so map_used couldn't have
388 spin_lock_irq(&pcpu_lock
);
389 BUG_ON(new_alloc
< chunk
->map_used
+ 2);
391 size
= chunk
->map_alloc
* sizeof(chunk
->map
[0]);
392 memcpy(new, chunk
->map
, size
);
395 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
396 * one of the first chunks and still using static map.
398 if (chunk
->map_alloc
>= PCPU_DFL_MAP_ALLOC
)
399 pcpu_mem_free(chunk
->map
, size
);
401 chunk
->map_alloc
= new_alloc
;
407 * pcpu_split_block - split a map block
408 * @chunk: chunk of interest
409 * @i: index of map block to split
410 * @head: head size in bytes (can be 0)
411 * @tail: tail size in bytes (can be 0)
413 * Split the @i'th map block into two or three blocks. If @head is
414 * non-zero, @head bytes block is inserted before block @i moving it
415 * to @i+1 and reducing its size by @head bytes.
417 * If @tail is non-zero, the target block, which can be @i or @i+1
418 * depending on @head, is reduced by @tail bytes and @tail byte block
419 * is inserted after the target block.
421 * @chunk->map must have enough free slots to accomodate the split.
426 static void pcpu_split_block(struct pcpu_chunk
*chunk
, int i
,
429 int nr_extra
= !!head
+ !!tail
;
431 BUG_ON(chunk
->map_alloc
< chunk
->map_used
+ nr_extra
);
433 /* insert new subblocks */
434 memmove(&chunk
->map
[i
+ nr_extra
], &chunk
->map
[i
],
435 sizeof(chunk
->map
[0]) * (chunk
->map_used
- i
));
436 chunk
->map_used
+= nr_extra
;
439 chunk
->map
[i
+ 1] = chunk
->map
[i
] - head
;
440 chunk
->map
[i
++] = head
;
443 chunk
->map
[i
++] -= tail
;
444 chunk
->map
[i
] = tail
;
449 * pcpu_alloc_area - allocate area from a pcpu_chunk
450 * @chunk: chunk of interest
451 * @size: wanted size in bytes
452 * @align: wanted align
454 * Try to allocate @size bytes area aligned at @align from @chunk.
455 * Note that this function only allocates the offset. It doesn't
456 * populate or map the area.
458 * @chunk->map must have at least two free slots.
464 * Allocated offset in @chunk on success, -1 if no matching area is
467 static int pcpu_alloc_area(struct pcpu_chunk
*chunk
, int size
, int align
)
469 int oslot
= pcpu_chunk_slot(chunk
);
473 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++])) {
474 bool is_last
= i
+ 1 == chunk
->map_used
;
477 /* extra for alignment requirement */
478 head
= ALIGN(off
, align
) - off
;
479 BUG_ON(i
== 0 && head
!= 0);
481 if (chunk
->map
[i
] < 0)
483 if (chunk
->map
[i
] < head
+ size
) {
484 max_contig
= max(chunk
->map
[i
], max_contig
);
489 * If head is small or the previous block is free,
490 * merge'em. Note that 'small' is defined as smaller
491 * than sizeof(int), which is very small but isn't too
492 * uncommon for percpu allocations.
494 if (head
&& (head
< sizeof(int) || chunk
->map
[i
- 1] > 0)) {
495 if (chunk
->map
[i
- 1] > 0)
496 chunk
->map
[i
- 1] += head
;
498 chunk
->map
[i
- 1] -= head
;
499 chunk
->free_size
-= head
;
501 chunk
->map
[i
] -= head
;
506 /* if tail is small, just keep it around */
507 tail
= chunk
->map
[i
] - head
- size
;
508 if (tail
< sizeof(int))
511 /* split if warranted */
513 pcpu_split_block(chunk
, i
, head
, tail
);
517 max_contig
= max(chunk
->map
[i
- 1], max_contig
);
520 max_contig
= max(chunk
->map
[i
+ 1], max_contig
);
523 /* update hint and mark allocated */
525 chunk
->contig_hint
= max_contig
; /* fully scanned */
527 chunk
->contig_hint
= max(chunk
->contig_hint
,
530 chunk
->free_size
-= chunk
->map
[i
];
531 chunk
->map
[i
] = -chunk
->map
[i
];
533 pcpu_chunk_relocate(chunk
, oslot
);
537 chunk
->contig_hint
= max_contig
; /* fully scanned */
538 pcpu_chunk_relocate(chunk
, oslot
);
540 /* tell the upper layer that this chunk has no matching area */
545 * pcpu_free_area - free area to a pcpu_chunk
546 * @chunk: chunk of interest
547 * @freeme: offset of area to free
549 * Free area starting from @freeme to @chunk. Note that this function
550 * only modifies the allocation map. It doesn't depopulate or unmap
556 static void pcpu_free_area(struct pcpu_chunk
*chunk
, int freeme
)
558 int oslot
= pcpu_chunk_slot(chunk
);
561 for (i
= 0, off
= 0; i
< chunk
->map_used
; off
+= abs(chunk
->map
[i
++]))
564 BUG_ON(off
!= freeme
);
565 BUG_ON(chunk
->map
[i
] > 0);
567 chunk
->map
[i
] = -chunk
->map
[i
];
568 chunk
->free_size
+= chunk
->map
[i
];
570 /* merge with previous? */
571 if (i
> 0 && chunk
->map
[i
- 1] >= 0) {
572 chunk
->map
[i
- 1] += chunk
->map
[i
];
574 memmove(&chunk
->map
[i
], &chunk
->map
[i
+ 1],
575 (chunk
->map_used
- i
) * sizeof(chunk
->map
[0]));
578 /* merge with next? */
579 if (i
+ 1 < chunk
->map_used
&& chunk
->map
[i
+ 1] >= 0) {
580 chunk
->map
[i
] += chunk
->map
[i
+ 1];
582 memmove(&chunk
->map
[i
+ 1], &chunk
->map
[i
+ 2],
583 (chunk
->map_used
- (i
+ 1)) * sizeof(chunk
->map
[0]));
586 chunk
->contig_hint
= max(chunk
->map
[i
], chunk
->contig_hint
);
587 pcpu_chunk_relocate(chunk
, oslot
);
591 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
592 * @chunk: chunk of interest
593 * @bitmapp: output parameter for bitmap
594 * @may_alloc: may allocate the array
596 * Returns pointer to array of pointers to struct page and bitmap,
597 * both of which can be indexed with pcpu_page_idx(). The returned
598 * array is cleared to zero and *@bitmapp is copied from
599 * @chunk->populated. Note that there is only one array and bitmap
600 * and access exclusion is the caller's responsibility.
603 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
604 * Otherwise, don't care.
607 * Pointer to temp pages array on success, NULL on failure.
609 static struct page
**pcpu_get_pages_and_bitmap(struct pcpu_chunk
*chunk
,
610 unsigned long **bitmapp
,
613 static struct page
**pages
;
614 static unsigned long *bitmap
;
615 size_t pages_size
= pcpu_nr_units
* pcpu_unit_pages
* sizeof(pages
[0]);
616 size_t bitmap_size
= BITS_TO_LONGS(pcpu_unit_pages
) *
617 sizeof(unsigned long);
619 if (!pages
|| !bitmap
) {
620 if (may_alloc
&& !pages
)
621 pages
= pcpu_mem_alloc(pages_size
);
622 if (may_alloc
&& !bitmap
)
623 bitmap
= pcpu_mem_alloc(bitmap_size
);
624 if (!pages
|| !bitmap
)
628 memset(pages
, 0, pages_size
);
629 bitmap_copy(bitmap
, chunk
->populated
, pcpu_unit_pages
);
636 * pcpu_free_pages - free pages which were allocated for @chunk
637 * @chunk: chunk pages were allocated for
638 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
639 * @populated: populated bitmap
640 * @page_start: page index of the first page to be freed
641 * @page_end: page index of the last page to be freed + 1
643 * Free pages [@page_start and @page_end) in @pages for all units.
644 * The pages were allocated for @chunk.
646 static void pcpu_free_pages(struct pcpu_chunk
*chunk
,
647 struct page
**pages
, unsigned long *populated
,
648 int page_start
, int page_end
)
653 for_each_possible_cpu(cpu
) {
654 for (i
= page_start
; i
< page_end
; i
++) {
655 struct page
*page
= pages
[pcpu_page_idx(cpu
, i
)];
664 * pcpu_alloc_pages - allocates pages for @chunk
665 * @chunk: target chunk
666 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
667 * @populated: populated bitmap
668 * @page_start: page index of the first page to be allocated
669 * @page_end: page index of the last page to be allocated + 1
671 * Allocate pages [@page_start,@page_end) into @pages for all units.
672 * The allocation is for @chunk. Percpu core doesn't care about the
673 * content of @pages and will pass it verbatim to pcpu_map_pages().
675 static int pcpu_alloc_pages(struct pcpu_chunk
*chunk
,
676 struct page
**pages
, unsigned long *populated
,
677 int page_start
, int page_end
)
679 const gfp_t gfp
= GFP_KERNEL
| __GFP_HIGHMEM
| __GFP_COLD
;
683 for_each_possible_cpu(cpu
) {
684 for (i
= page_start
; i
< page_end
; i
++) {
685 struct page
**pagep
= &pages
[pcpu_page_idx(cpu
, i
)];
687 *pagep
= alloc_pages_node(cpu_to_node(cpu
), gfp
, 0);
689 pcpu_free_pages(chunk
, pages
, populated
,
690 page_start
, page_end
);
699 * pcpu_pre_unmap_flush - flush cache prior to unmapping
700 * @chunk: chunk the regions to be flushed belongs to
701 * @page_start: page index of the first page to be flushed
702 * @page_end: page index of the last page to be flushed + 1
704 * Pages in [@page_start,@page_end) of @chunk are about to be
705 * unmapped. Flush cache. As each flushing trial can be very
706 * expensive, issue flush on the whole region at once rather than
707 * doing it for each cpu. This could be an overkill but is more
710 static void pcpu_pre_unmap_flush(struct pcpu_chunk
*chunk
,
711 int page_start
, int page_end
)
714 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
715 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
718 static void __pcpu_unmap_pages(unsigned long addr
, int nr_pages
)
720 unmap_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
);
724 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
725 * @chunk: chunk of interest
726 * @pages: pages array which can be used to pass information to free
727 * @populated: populated bitmap
728 * @page_start: page index of the first page to unmap
729 * @page_end: page index of the last page to unmap + 1
731 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
732 * Corresponding elements in @pages were cleared by the caller and can
733 * be used to carry information to pcpu_free_pages() which will be
734 * called after all unmaps are finished. The caller should call
735 * proper pre/post flush functions.
737 static void pcpu_unmap_pages(struct pcpu_chunk
*chunk
,
738 struct page
**pages
, unsigned long *populated
,
739 int page_start
, int page_end
)
744 for_each_possible_cpu(cpu
) {
745 for (i
= page_start
; i
< page_end
; i
++) {
748 page
= pcpu_chunk_page(chunk
, cpu
, i
);
750 pages
[pcpu_page_idx(cpu
, i
)] = page
;
752 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
753 page_end
- page_start
);
756 for (i
= page_start
; i
< page_end
; i
++)
757 __clear_bit(i
, populated
);
761 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
762 * @chunk: pcpu_chunk the regions to be flushed belong to
763 * @page_start: page index of the first page to be flushed
764 * @page_end: page index of the last page to be flushed + 1
766 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
767 * TLB for the regions. This can be skipped if the area is to be
768 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
770 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
771 * for the whole region.
773 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk
*chunk
,
774 int page_start
, int page_end
)
776 flush_tlb_kernel_range(
777 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
778 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
781 static int __pcpu_map_pages(unsigned long addr
, struct page
**pages
,
784 return map_kernel_range_noflush(addr
, nr_pages
<< PAGE_SHIFT
,
789 * pcpu_map_pages - map pages into a pcpu_chunk
790 * @chunk: chunk of interest
791 * @pages: pages array containing pages to be mapped
792 * @populated: populated bitmap
793 * @page_start: page index of the first page to map
794 * @page_end: page index of the last page to map + 1
796 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
797 * caller is responsible for calling pcpu_post_map_flush() after all
798 * mappings are complete.
800 * This function is responsible for setting corresponding bits in
801 * @chunk->populated bitmap and whatever is necessary for reverse
802 * lookup (addr -> chunk).
804 static int pcpu_map_pages(struct pcpu_chunk
*chunk
,
805 struct page
**pages
, unsigned long *populated
,
806 int page_start
, int page_end
)
808 unsigned int cpu
, tcpu
;
811 for_each_possible_cpu(cpu
) {
812 err
= __pcpu_map_pages(pcpu_chunk_addr(chunk
, cpu
, page_start
),
813 &pages
[pcpu_page_idx(cpu
, page_start
)],
814 page_end
- page_start
);
819 /* mapping successful, link chunk and mark populated */
820 for (i
= page_start
; i
< page_end
; i
++) {
821 for_each_possible_cpu(cpu
)
822 pcpu_set_page_chunk(pages
[pcpu_page_idx(cpu
, i
)],
824 __set_bit(i
, populated
);
830 for_each_possible_cpu(tcpu
) {
833 __pcpu_unmap_pages(pcpu_chunk_addr(chunk
, tcpu
, page_start
),
834 page_end
- page_start
);
840 * pcpu_post_map_flush - flush cache after mapping
841 * @chunk: pcpu_chunk the regions to be flushed belong to
842 * @page_start: page index of the first page to be flushed
843 * @page_end: page index of the last page to be flushed + 1
845 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
848 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
849 * for the whole region.
851 static void pcpu_post_map_flush(struct pcpu_chunk
*chunk
,
852 int page_start
, int page_end
)
855 pcpu_chunk_addr(chunk
, pcpu_first_unit_cpu
, page_start
),
856 pcpu_chunk_addr(chunk
, pcpu_last_unit_cpu
, page_end
));
860 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
861 * @chunk: chunk to depopulate
862 * @off: offset to the area to depopulate
863 * @size: size of the area to depopulate in bytes
864 * @flush: whether to flush cache and tlb or not
866 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
867 * from @chunk. If @flush is true, vcache is flushed before unmapping
873 static void pcpu_depopulate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
875 int page_start
= PFN_DOWN(off
);
876 int page_end
= PFN_UP(off
+ size
);
878 unsigned long *populated
;
881 /* quick path, check whether it's empty already */
882 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
883 if (rs
== page_start
&& re
== page_end
)
888 /* immutable chunks can't be depopulated */
889 WARN_ON(chunk
->immutable
);
892 * If control reaches here, there must have been at least one
893 * successful population attempt so the temp pages array must
896 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, false);
900 pcpu_pre_unmap_flush(chunk
, page_start
, page_end
);
902 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
903 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
905 /* no need to flush tlb, vmalloc will handle it lazily */
907 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
)
908 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
910 /* commit new bitmap */
911 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
915 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
916 * @chunk: chunk of interest
917 * @off: offset to the area to populate
918 * @size: size of the area to populate in bytes
920 * For each cpu, populate and map pages [@page_start,@page_end) into
921 * @chunk. The area is cleared on return.
924 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
926 static int pcpu_populate_chunk(struct pcpu_chunk
*chunk
, int off
, int size
)
928 int page_start
= PFN_DOWN(off
);
929 int page_end
= PFN_UP(off
+ size
);
930 int free_end
= page_start
, unmap_end
= page_start
;
932 unsigned long *populated
;
936 /* quick path, check whether all pages are already there */
937 pcpu_for_each_pop_region(chunk
, rs
, re
, page_start
, page_end
) {
938 if (rs
== page_start
&& re
== page_end
)
943 /* need to allocate and map pages, this chunk can't be immutable */
944 WARN_ON(chunk
->immutable
);
946 pages
= pcpu_get_pages_and_bitmap(chunk
, &populated
, true);
951 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
952 rc
= pcpu_alloc_pages(chunk
, pages
, populated
, rs
, re
);
958 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, page_end
) {
959 rc
= pcpu_map_pages(chunk
, pages
, populated
, rs
, re
);
964 pcpu_post_map_flush(chunk
, page_start
, page_end
);
966 /* commit new bitmap */
967 bitmap_copy(chunk
->populated
, populated
, pcpu_unit_pages
);
969 for_each_possible_cpu(cpu
)
970 memset((void *)pcpu_chunk_addr(chunk
, cpu
, 0) + off
, 0, size
);
974 pcpu_pre_unmap_flush(chunk
, page_start
, unmap_end
);
975 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, unmap_end
)
976 pcpu_unmap_pages(chunk
, pages
, populated
, rs
, re
);
977 pcpu_post_unmap_tlb_flush(chunk
, page_start
, unmap_end
);
979 pcpu_for_each_unpop_region(chunk
, rs
, re
, page_start
, free_end
)
980 pcpu_free_pages(chunk
, pages
, populated
, rs
, re
);
984 static void free_pcpu_chunk(struct pcpu_chunk
*chunk
)
989 free_vm_area(chunk
->vm
);
990 pcpu_mem_free(chunk
->map
, chunk
->map_alloc
* sizeof(chunk
->map
[0]));
994 static struct pcpu_chunk
*alloc_pcpu_chunk(void)
996 struct pcpu_chunk
*chunk
;
998 chunk
= kzalloc(pcpu_chunk_struct_size
, GFP_KERNEL
);
1002 chunk
->map
= pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC
* sizeof(chunk
->map
[0]));
1003 chunk
->map_alloc
= PCPU_DFL_MAP_ALLOC
;
1004 chunk
->map
[chunk
->map_used
++] = pcpu_unit_size
;
1006 chunk
->vm
= get_vm_area(pcpu_chunk_size
, GFP_KERNEL
);
1008 free_pcpu_chunk(chunk
);
1012 INIT_LIST_HEAD(&chunk
->list
);
1013 chunk
->free_size
= pcpu_unit_size
;
1014 chunk
->contig_hint
= pcpu_unit_size
;
1020 * pcpu_alloc - the percpu allocator
1021 * @size: size of area to allocate in bytes
1022 * @align: alignment of area (max PAGE_SIZE)
1023 * @reserved: allocate from the reserved chunk if available
1025 * Allocate percpu area of @size bytes aligned at @align.
1028 * Does GFP_KERNEL allocation.
1031 * Percpu pointer to the allocated area on success, NULL on failure.
1033 static void *pcpu_alloc(size_t size
, size_t align
, bool reserved
)
1035 struct pcpu_chunk
*chunk
;
1038 if (unlikely(!size
|| size
> PCPU_MIN_UNIT_SIZE
|| align
> PAGE_SIZE
)) {
1039 WARN(true, "illegal size (%zu) or align (%zu) for "
1040 "percpu allocation\n", size
, align
);
1044 mutex_lock(&pcpu_alloc_mutex
);
1045 spin_lock_irq(&pcpu_lock
);
1047 /* serve reserved allocations from the reserved chunk if available */
1048 if (reserved
&& pcpu_reserved_chunk
) {
1049 chunk
= pcpu_reserved_chunk
;
1050 if (size
> chunk
->contig_hint
||
1051 pcpu_extend_area_map(chunk
) < 0)
1053 off
= pcpu_alloc_area(chunk
, size
, align
);
1060 /* search through normal chunks */
1061 for (slot
= pcpu_size_to_slot(size
); slot
< pcpu_nr_slots
; slot
++) {
1062 list_for_each_entry(chunk
, &pcpu_slot
[slot
], list
) {
1063 if (size
> chunk
->contig_hint
)
1066 switch (pcpu_extend_area_map(chunk
)) {
1070 goto restart
; /* pcpu_lock dropped, restart */
1075 off
= pcpu_alloc_area(chunk
, size
, align
);
1081 /* hmmm... no space left, create a new chunk */
1082 spin_unlock_irq(&pcpu_lock
);
1084 chunk
= alloc_pcpu_chunk();
1086 goto fail_unlock_mutex
;
1088 spin_lock_irq(&pcpu_lock
);
1089 pcpu_chunk_relocate(chunk
, -1);
1093 spin_unlock_irq(&pcpu_lock
);
1095 /* populate, map and clear the area */
1096 if (pcpu_populate_chunk(chunk
, off
, size
)) {
1097 spin_lock_irq(&pcpu_lock
);
1098 pcpu_free_area(chunk
, off
);
1102 mutex_unlock(&pcpu_alloc_mutex
);
1104 /* return address relative to unit0 */
1105 return __addr_to_pcpu_ptr(chunk
->vm
->addr
+ off
);
1108 spin_unlock_irq(&pcpu_lock
);
1110 mutex_unlock(&pcpu_alloc_mutex
);
1115 * __alloc_percpu - allocate dynamic percpu area
1116 * @size: size of area to allocate in bytes
1117 * @align: alignment of area (max PAGE_SIZE)
1119 * Allocate percpu area of @size bytes aligned at @align. Might
1120 * sleep. Might trigger writeouts.
1123 * Does GFP_KERNEL allocation.
1126 * Percpu pointer to the allocated area on success, NULL on failure.
1128 void *__alloc_percpu(size_t size
, size_t align
)
1130 return pcpu_alloc(size
, align
, false);
1132 EXPORT_SYMBOL_GPL(__alloc_percpu
);
1135 * __alloc_reserved_percpu - allocate reserved percpu area
1136 * @size: size of area to allocate in bytes
1137 * @align: alignment of area (max PAGE_SIZE)
1139 * Allocate percpu area of @size bytes aligned at @align from reserved
1140 * percpu area if arch has set it up; otherwise, allocation is served
1141 * from the same dynamic area. Might sleep. Might trigger writeouts.
1144 * Does GFP_KERNEL allocation.
1147 * Percpu pointer to the allocated area on success, NULL on failure.
1149 void *__alloc_reserved_percpu(size_t size
, size_t align
)
1151 return pcpu_alloc(size
, align
, true);
1155 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1158 * Reclaim all fully free chunks except for the first one.
1161 * workqueue context.
1163 static void pcpu_reclaim(struct work_struct
*work
)
1166 struct list_head
*head
= &pcpu_slot
[pcpu_nr_slots
- 1];
1167 struct pcpu_chunk
*chunk
, *next
;
1169 mutex_lock(&pcpu_alloc_mutex
);
1170 spin_lock_irq(&pcpu_lock
);
1172 list_for_each_entry_safe(chunk
, next
, head
, list
) {
1173 WARN_ON(chunk
->immutable
);
1175 /* spare the first one */
1176 if (chunk
== list_first_entry(head
, struct pcpu_chunk
, list
))
1179 list_move(&chunk
->list
, &todo
);
1182 spin_unlock_irq(&pcpu_lock
);
1183 mutex_unlock(&pcpu_alloc_mutex
);
1185 list_for_each_entry_safe(chunk
, next
, &todo
, list
) {
1186 pcpu_depopulate_chunk(chunk
, 0, pcpu_unit_size
);
1187 free_pcpu_chunk(chunk
);
1192 * free_percpu - free percpu area
1193 * @ptr: pointer to area to free
1195 * Free percpu area @ptr.
1198 * Can be called from atomic context.
1200 void free_percpu(void *ptr
)
1202 void *addr
= __pcpu_ptr_to_addr(ptr
);
1203 struct pcpu_chunk
*chunk
;
1204 unsigned long flags
;
1210 spin_lock_irqsave(&pcpu_lock
, flags
);
1212 chunk
= pcpu_chunk_addr_search(addr
);
1213 off
= addr
- chunk
->vm
->addr
;
1215 pcpu_free_area(chunk
, off
);
1217 /* if there are more than one fully free chunks, wake up grim reaper */
1218 if (chunk
->free_size
== pcpu_unit_size
) {
1219 struct pcpu_chunk
*pos
;
1221 list_for_each_entry(pos
, &pcpu_slot
[pcpu_nr_slots
- 1], list
)
1223 schedule_work(&pcpu_reclaim_work
);
1228 spin_unlock_irqrestore(&pcpu_lock
, flags
);
1230 EXPORT_SYMBOL_GPL(free_percpu
);
1233 * pcpu_setup_first_chunk - initialize the first percpu chunk
1234 * @static_size: the size of static percpu area in bytes
1235 * @reserved_size: the size of reserved percpu area in bytes, 0 for none
1236 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1237 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE
1238 * @base_addr: mapped address
1239 * @unit_map: cpu -> unit map, NULL for sequential mapping
1241 * Initialize the first percpu chunk which contains the kernel static
1242 * perpcu area. This function is to be called from arch percpu area
1245 * @reserved_size, if non-zero, specifies the amount of bytes to
1246 * reserve after the static area in the first chunk. This reserves
1247 * the first chunk such that it's available only through reserved
1248 * percpu allocation. This is primarily used to serve module percpu
1249 * static areas on architectures where the addressing model has
1250 * limited offset range for symbol relocations to guarantee module
1251 * percpu symbols fall inside the relocatable range.
1253 * @dyn_size, if non-negative, determines the number of bytes
1254 * available for dynamic allocation in the first chunk. Specifying
1255 * non-negative value makes percpu leave alone the area beyond
1256 * @static_size + @reserved_size + @dyn_size.
1258 * @unit_size specifies unit size and must be aligned to PAGE_SIZE and
1259 * equal to or larger than @static_size + @reserved_size + if
1260 * non-negative, @dyn_size.
1262 * The caller should have mapped the first chunk at @base_addr and
1263 * copied static data to each unit.
1265 * If the first chunk ends up with both reserved and dynamic areas, it
1266 * is served by two chunks - one to serve the core static and reserved
1267 * areas and the other for the dynamic area. They share the same vm
1268 * and page map but uses different area allocation map to stay away
1269 * from each other. The latter chunk is circulated in the chunk slots
1270 * and available for dynamic allocation like any other chunks.
1273 * The determined pcpu_unit_size which can be used to initialize
1276 size_t __init
pcpu_setup_first_chunk(size_t static_size
, size_t reserved_size
,
1277 ssize_t dyn_size
, size_t unit_size
,
1278 void *base_addr
, const int *unit_map
)
1280 static struct vm_struct first_vm
;
1281 static int smap
[2], dmap
[2];
1282 size_t size_sum
= static_size
+ reserved_size
+
1283 (dyn_size
>= 0 ? dyn_size
: 0);
1284 struct pcpu_chunk
*schunk
, *dchunk
= NULL
;
1285 unsigned int cpu
, tcpu
;
1289 BUILD_BUG_ON(ARRAY_SIZE(smap
) >= PCPU_DFL_MAP_ALLOC
||
1290 ARRAY_SIZE(dmap
) >= PCPU_DFL_MAP_ALLOC
);
1291 BUG_ON(!static_size
);
1293 BUG_ON(unit_size
< size_sum
);
1294 BUG_ON(unit_size
& ~PAGE_MASK
);
1295 BUG_ON(unit_size
< PCPU_MIN_UNIT_SIZE
);
1297 /* determine number of units and verify and initialize pcpu_unit_map */
1299 int first_unit
= INT_MAX
, last_unit
= INT_MIN
;
1301 for_each_possible_cpu(cpu
) {
1302 int unit
= unit_map
[cpu
];
1305 for_each_possible_cpu(tcpu
) {
1308 /* the mapping should be one-to-one */
1309 BUG_ON(unit_map
[tcpu
] == unit
);
1312 if (unit
< first_unit
) {
1313 pcpu_first_unit_cpu
= cpu
;
1316 if (unit
> last_unit
) {
1317 pcpu_last_unit_cpu
= cpu
;
1321 pcpu_nr_units
= last_unit
+ 1;
1322 pcpu_unit_map
= unit_map
;
1326 /* #units == #cpus, identity mapped */
1327 identity_map
= alloc_bootmem(num_possible_cpus() *
1328 sizeof(identity_map
[0]));
1330 for_each_possible_cpu(cpu
)
1331 identity_map
[cpu
] = cpu
;
1333 pcpu_first_unit_cpu
= 0;
1334 pcpu_last_unit_cpu
= pcpu_nr_units
- 1;
1335 pcpu_nr_units
= num_possible_cpus();
1336 pcpu_unit_map
= identity_map
;
1339 /* determine basic parameters */
1340 pcpu_unit_pages
= unit_size
>> PAGE_SHIFT
;
1341 pcpu_unit_size
= pcpu_unit_pages
<< PAGE_SHIFT
;
1342 pcpu_chunk_size
= pcpu_nr_units
* pcpu_unit_size
;
1343 pcpu_chunk_struct_size
= sizeof(struct pcpu_chunk
) +
1344 BITS_TO_LONGS(pcpu_unit_pages
) * sizeof(unsigned long);
1347 dyn_size
= pcpu_unit_size
- static_size
- reserved_size
;
1349 first_vm
.flags
= VM_ALLOC
;
1350 first_vm
.size
= pcpu_chunk_size
;
1351 first_vm
.addr
= base_addr
;
1354 * Allocate chunk slots. The additional last slot is for
1357 pcpu_nr_slots
= __pcpu_size_to_slot(pcpu_unit_size
) + 2;
1358 pcpu_slot
= alloc_bootmem(pcpu_nr_slots
* sizeof(pcpu_slot
[0]));
1359 for (i
= 0; i
< pcpu_nr_slots
; i
++)
1360 INIT_LIST_HEAD(&pcpu_slot
[i
]);
1363 * Initialize static chunk. If reserved_size is zero, the
1364 * static chunk covers static area + dynamic allocation area
1365 * in the first chunk. If reserved_size is not zero, it
1366 * covers static area + reserved area (mostly used for module
1367 * static percpu allocation).
1369 schunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1370 INIT_LIST_HEAD(&schunk
->list
);
1371 schunk
->vm
= &first_vm
;
1373 schunk
->map_alloc
= ARRAY_SIZE(smap
);
1374 schunk
->immutable
= true;
1375 bitmap_fill(schunk
->populated
, pcpu_unit_pages
);
1377 if (reserved_size
) {
1378 schunk
->free_size
= reserved_size
;
1379 pcpu_reserved_chunk
= schunk
;
1380 pcpu_reserved_chunk_limit
= static_size
+ reserved_size
;
1382 schunk
->free_size
= dyn_size
;
1383 dyn_size
= 0; /* dynamic area covered */
1385 schunk
->contig_hint
= schunk
->free_size
;
1387 schunk
->map
[schunk
->map_used
++] = -static_size
;
1388 if (schunk
->free_size
)
1389 schunk
->map
[schunk
->map_used
++] = schunk
->free_size
;
1391 /* init dynamic chunk if necessary */
1393 dchunk
= alloc_bootmem(pcpu_chunk_struct_size
);
1394 INIT_LIST_HEAD(&dchunk
->list
);
1395 dchunk
->vm
= &first_vm
;
1397 dchunk
->map_alloc
= ARRAY_SIZE(dmap
);
1398 dchunk
->immutable
= true;
1399 bitmap_fill(dchunk
->populated
, pcpu_unit_pages
);
1401 dchunk
->contig_hint
= dchunk
->free_size
= dyn_size
;
1402 dchunk
->map
[dchunk
->map_used
++] = -pcpu_reserved_chunk_limit
;
1403 dchunk
->map
[dchunk
->map_used
++] = dchunk
->free_size
;
1406 /* link the first chunk in */
1407 pcpu_first_chunk
= dchunk
?: schunk
;
1408 pcpu_chunk_relocate(pcpu_first_chunk
, -1);
1411 pcpu_base_addr
= schunk
->vm
->addr
;
1412 return pcpu_unit_size
;
1415 static size_t pcpu_calc_fc_sizes(size_t static_size
, size_t reserved_size
,
1420 size_sum
= PFN_ALIGN(static_size
+ reserved_size
+
1421 (*dyn_sizep
>= 0 ? *dyn_sizep
: 0));
1422 if (*dyn_sizep
!= 0)
1423 *dyn_sizep
= size_sum
- static_size
- reserved_size
;
1429 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1430 * @static_size: the size of static percpu area in bytes
1431 * @reserved_size: the size of reserved percpu area in bytes
1432 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1434 * This is a helper to ease setting up embedded first percpu chunk and
1435 * can be called where pcpu_setup_first_chunk() is expected.
1437 * If this function is used to setup the first chunk, it is allocated
1438 * as a contiguous area using bootmem allocator and used as-is without
1439 * being mapped into vmalloc area. This enables the first chunk to
1440 * piggy back on the linear physical mapping which often uses larger
1443 * When @dyn_size is positive, dynamic area might be larger than
1444 * specified to fill page alignment. When @dyn_size is auto,
1445 * @dyn_size is just big enough to fill page alignment after static
1446 * and reserved areas.
1448 * If the needed size is smaller than the minimum or specified unit
1449 * size, the leftover is returned to the bootmem allocator.
1452 * The determined pcpu_unit_size which can be used to initialize
1453 * percpu access on success, -errno on failure.
1455 ssize_t __init
pcpu_embed_first_chunk(size_t static_size
, size_t reserved_size
,
1458 size_t size_sum
, unit_size
, chunk_size
;
1462 /* determine parameters and allocate */
1463 size_sum
= pcpu_calc_fc_sizes(static_size
, reserved_size
, &dyn_size
);
1465 unit_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1466 chunk_size
= unit_size
* num_possible_cpus();
1468 base
= __alloc_bootmem_nopanic(chunk_size
, PAGE_SIZE
,
1469 __pa(MAX_DMA_ADDRESS
));
1471 pr_warning("PERCPU: failed to allocate %zu bytes for "
1472 "embedding\n", chunk_size
);
1476 /* return the leftover and copy */
1477 for_each_possible_cpu(cpu
) {
1478 void *ptr
= base
+ cpu
* unit_size
;
1480 free_bootmem(__pa(ptr
+ size_sum
), unit_size
- size_sum
);
1481 memcpy(ptr
, __per_cpu_load
, static_size
);
1484 /* we're ready, commit */
1485 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
1486 size_sum
>> PAGE_SHIFT
, base
, static_size
);
1488 return pcpu_setup_first_chunk(static_size
, reserved_size
, dyn_size
,
1489 unit_size
, base
, NULL
);
1493 * pcpu_4k_first_chunk - map the first chunk using PAGE_SIZE pages
1494 * @static_size: the size of static percpu area in bytes
1495 * @reserved_size: the size of reserved percpu area in bytes
1496 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1497 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1498 * @populate_pte_fn: function to populate pte
1500 * This is a helper to ease setting up embedded first percpu chunk and
1501 * can be called where pcpu_setup_first_chunk() is expected.
1503 * This is the basic allocator. Static percpu area is allocated
1504 * page-by-page into vmalloc area.
1507 * The determined pcpu_unit_size which can be used to initialize
1508 * percpu access on success, -errno on failure.
1510 ssize_t __init
pcpu_4k_first_chunk(size_t static_size
, size_t reserved_size
,
1511 pcpu_fc_alloc_fn_t alloc_fn
,
1512 pcpu_fc_free_fn_t free_fn
,
1513 pcpu_fc_populate_pte_fn_t populate_pte_fn
)
1515 static struct vm_struct vm
;
1518 struct page
**pages
;
1523 unit_pages
= PFN_UP(max_t(size_t, static_size
+ reserved_size
,
1524 PCPU_MIN_UNIT_SIZE
));
1526 /* unaligned allocations can't be freed, round up to page size */
1527 pages_size
= PFN_ALIGN(unit_pages
* num_possible_cpus() *
1529 pages
= alloc_bootmem(pages_size
);
1531 /* allocate pages */
1533 for_each_possible_cpu(cpu
)
1534 for (i
= 0; i
< unit_pages
; i
++) {
1537 ptr
= alloc_fn(cpu
, PAGE_SIZE
);
1539 pr_warning("PERCPU: failed to allocate "
1540 "4k page for cpu%u\n", cpu
);
1543 pages
[j
++] = virt_to_page(ptr
);
1546 /* allocate vm area, map the pages and copy static data */
1547 vm
.flags
= VM_ALLOC
;
1548 vm
.size
= num_possible_cpus() * unit_pages
<< PAGE_SHIFT
;
1549 vm_area_register_early(&vm
, PAGE_SIZE
);
1551 for_each_possible_cpu(cpu
) {
1552 unsigned long unit_addr
= (unsigned long)vm
.addr
+
1553 (cpu
* unit_pages
<< PAGE_SHIFT
);
1555 for (i
= 0; i
< unit_pages
; i
++)
1556 populate_pte_fn(unit_addr
+ (i
<< PAGE_SHIFT
));
1558 /* pte already populated, the following shouldn't fail */
1559 ret
= __pcpu_map_pages(unit_addr
, &pages
[cpu
* unit_pages
],
1562 panic("failed to map percpu area, err=%zd\n", ret
);
1565 * FIXME: Archs with virtual cache should flush local
1566 * cache for the linear mapping here - something
1567 * equivalent to flush_cache_vmap() on the local cpu.
1568 * flush_cache_vmap() can't be used as most supporting
1569 * data structures are not set up yet.
1572 /* copy static data */
1573 memcpy((void *)unit_addr
, __per_cpu_load
, static_size
);
1576 /* we're ready, commit */
1577 pr_info("PERCPU: %d 4k pages per cpu, static data %zu bytes\n",
1578 unit_pages
, static_size
);
1580 ret
= pcpu_setup_first_chunk(static_size
, reserved_size
, -1,
1581 unit_pages
<< PAGE_SHIFT
, vm
.addr
, NULL
);
1586 free_fn(page_address(pages
[j
]), PAGE_SIZE
);
1589 free_bootmem(__pa(pages
), pages_size
);
1594 * Large page remapping first chunk setup helper
1596 #ifdef CONFIG_NEED_MULTIPLE_NODES
1602 static size_t pcpul_size
;
1603 static size_t pcpul_unit_size
;
1604 static struct pcpul_ent
*pcpul_map
;
1605 static struct vm_struct pcpul_vm
;
1608 * pcpu_lpage_first_chunk - remap the first percpu chunk using large page
1609 * @static_size: the size of static percpu area in bytes
1610 * @reserved_size: the size of reserved percpu area in bytes
1611 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1612 * @lpage_size: the size of a large page
1613 * @alloc_fn: function to allocate percpu lpage, always called with lpage_size
1614 * @free_fn: function to free percpu memory, @size <= lpage_size
1615 * @map_fn: function to map percpu lpage, always called with lpage_size
1617 * This allocator uses large page as unit. A large page is allocated
1618 * for each cpu and each is remapped into vmalloc area using large
1619 * page mapping. As large page can be quite large, only part of it is
1620 * used for the first chunk. Unused part is returned to the bootmem
1623 * So, the large pages are mapped twice - once to the physical mapping
1624 * and to the vmalloc area for the first percpu chunk. The double
1625 * mapping does add one more large TLB entry pressure but still is
1626 * much better than only using 4k mappings while still being NUMA
1630 * The determined pcpu_unit_size which can be used to initialize
1631 * percpu access on success, -errno on failure.
1633 ssize_t __init
pcpu_lpage_first_chunk(size_t static_size
, size_t reserved_size
,
1634 ssize_t dyn_size
, size_t lpage_size
,
1635 pcpu_fc_alloc_fn_t alloc_fn
,
1636 pcpu_fc_free_fn_t free_fn
,
1637 pcpu_fc_map_fn_t map_fn
)
1646 * Currently supports only single page. Supporting multiple
1647 * pages won't be too difficult if it ever becomes necessary.
1649 size_sum
= pcpu_calc_fc_sizes(static_size
, reserved_size
, &dyn_size
);
1651 pcpul_unit_size
= lpage_size
;
1652 pcpul_size
= max_t(size_t, size_sum
, PCPU_MIN_UNIT_SIZE
);
1653 if (pcpul_size
> pcpul_unit_size
) {
1654 pr_warning("PERCPU: static data is larger than large page, "
1655 "can't use large page\n");
1659 /* allocate pointer array and alloc large pages */
1660 map_size
= PFN_ALIGN(num_possible_cpus() * sizeof(pcpul_map
[0]));
1661 pcpul_map
= alloc_bootmem(map_size
);
1663 for_each_possible_cpu(cpu
) {
1666 ptr
= alloc_fn(cpu
, lpage_size
);
1668 pr_warning("PERCPU: failed to allocate large page "
1669 "for cpu%u\n", cpu
);
1674 * Only use pcpul_size bytes and give back the rest.
1676 * Ingo: The lpage_size up-rounding bootmem is needed
1677 * to make sure the partial lpage is still fully RAM -
1678 * it's not well-specified to have a incompatible area
1679 * (unmapped RAM, device memory, etc.) in that hole.
1681 free_fn(ptr
+ pcpul_size
, lpage_size
- pcpul_size
);
1683 pcpul_map
[cpu
].cpu
= cpu
;
1684 pcpul_map
[cpu
].ptr
= ptr
;
1686 memcpy(ptr
, __per_cpu_load
, static_size
);
1689 /* allocate address and map */
1690 pcpul_vm
.flags
= VM_ALLOC
;
1691 pcpul_vm
.size
= num_possible_cpus() * pcpul_unit_size
;
1692 vm_area_register_early(&pcpul_vm
, pcpul_unit_size
);
1694 for_each_possible_cpu(cpu
)
1695 map_fn(pcpul_map
[cpu
].ptr
, pcpul_unit_size
,
1696 pcpul_vm
.addr
+ cpu
* pcpul_unit_size
);
1698 /* we're ready, commit */
1699 pr_info("PERCPU: Remapped at %p with large pages, static data "
1700 "%zu bytes\n", pcpul_vm
.addr
, static_size
);
1702 ret
= pcpu_setup_first_chunk(static_size
, reserved_size
, dyn_size
,
1703 pcpul_unit_size
, pcpul_vm
.addr
, NULL
);
1705 /* sort pcpul_map array for pcpu_lpage_remapped() */
1706 for (i
= 0; i
< num_possible_cpus() - 1; i
++)
1707 for (j
= i
+ 1; j
< num_possible_cpus(); j
++)
1708 if (pcpul_map
[i
].ptr
> pcpul_map
[j
].ptr
) {
1709 struct pcpul_ent tmp
= pcpul_map
[i
];
1710 pcpul_map
[i
] = pcpul_map
[j
];
1717 for_each_possible_cpu(cpu
)
1718 if (pcpul_map
[cpu
].ptr
)
1719 free_fn(pcpul_map
[cpu
].ptr
, pcpul_size
);
1720 free_bootmem(__pa(pcpul_map
), map_size
);
1725 * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area
1726 * @kaddr: the kernel address in question
1728 * Determine whether @kaddr falls in the pcpul recycled area. This is
1729 * used by pageattr to detect VM aliases and break up the pcpu large
1730 * page mapping such that the same physical page is not mapped under
1731 * different attributes.
1733 * The recycled area is always at the tail of a partially used large
1737 * Address of corresponding remapped pcpu address if match is found;
1740 void *pcpu_lpage_remapped(void *kaddr
)
1742 unsigned long unit_mask
= pcpul_unit_size
- 1;
1743 void *lpage_addr
= (void *)((unsigned long)kaddr
& ~unit_mask
);
1744 unsigned long offset
= (unsigned long)kaddr
& unit_mask
;
1745 int left
= 0, right
= num_possible_cpus() - 1;
1748 /* pcpul in use at all? */
1752 /* okay, perform binary search */
1753 while (left
<= right
) {
1754 pos
= (left
+ right
) / 2;
1756 if (pcpul_map
[pos
].ptr
< lpage_addr
)
1758 else if (pcpul_map
[pos
].ptr
> lpage_addr
)
1761 /* it shouldn't be in the area for the first chunk */
1762 WARN_ON(offset
< pcpul_size
);
1764 return pcpul_vm
.addr
+
1765 pcpul_map
[pos
].cpu
* pcpul_unit_size
+ offset
;
1774 * Generic percpu area setup.
1776 * The embedding helper is used because its behavior closely resembles
1777 * the original non-dynamic generic percpu area setup. This is
1778 * important because many archs have addressing restrictions and might
1779 * fail if the percpu area is located far away from the previous
1780 * location. As an added bonus, in non-NUMA cases, embedding is
1781 * generally a good idea TLB-wise because percpu area can piggy back
1782 * on the physical linear memory mapping which uses large page
1783 * mappings on applicable archs.
1785 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1786 unsigned long __per_cpu_offset
[NR_CPUS
] __read_mostly
;
1787 EXPORT_SYMBOL(__per_cpu_offset
);
1789 void __init
setup_per_cpu_areas(void)
1791 size_t static_size
= __per_cpu_end
- __per_cpu_start
;
1793 unsigned long delta
;
1797 * Always reserve area for module percpu variables. That's
1798 * what the legacy allocator did.
1800 unit_size
= pcpu_embed_first_chunk(static_size
, PERCPU_MODULE_RESERVE
,
1801 PERCPU_DYNAMIC_RESERVE
);
1803 panic("Failed to initialized percpu areas.");
1805 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
1806 for_each_possible_cpu(cpu
)
1807 __per_cpu_offset
[cpu
] = delta
+ cpu
* unit_size
;
1809 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */