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 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
50 * regular address to percpu pointer and back if they need to be
51 * different from the default
53 * - use pcpu_setup_first_chunk() during percpu area initialization to
54 * setup the first chunk containing the kernel static percpu area
57 #include <linux/bitmap.h>
58 #include <linux/bootmem.h>
59 #include <linux/err.h>
60 #include <linux/list.h>
61 #include <linux/log2.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>
77 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
80 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
81 #ifndef __addr_to_pcpu_ptr
82 #define __addr_to_pcpu_ptr(addr) \
83 (void __percpu *)((unsigned long)(addr) - \
84 (unsigned long)pcpu_base_addr + \
85 (unsigned long)__per_cpu_start)
87 #ifndef __pcpu_ptr_to_addr
88 #define __pcpu_ptr_to_addr(ptr) \
89 (void __force *)((unsigned long)(ptr) + \
90 (unsigned long)pcpu_base_addr - \
91 (unsigned long)__per_cpu_start)
95 struct list_head list; /* linked to pcpu_slot lists */
96 int free_size; /* free bytes in the chunk */
97 int contig_hint; /* max contiguous size hint */
98 void *base_addr; /* base address of this chunk */
99 int map_used; /* # of map entries used */
100 int map_alloc; /* # of map entries allocated */
101 int *map; /* allocation map */
102 struct vm_struct **vms; /* mapped vmalloc regions */
103 bool immutable; /* no [de]population allowed */
104 unsigned long populated[]; /* populated bitmap */
107 static int pcpu_unit_pages __read_mostly;
108 static int pcpu_unit_size __read_mostly;
109 static int pcpu_nr_units __read_mostly;
110 static int pcpu_atom_size __read_mostly;
111 static int pcpu_nr_slots __read_mostly;
112 static size_t pcpu_chunk_struct_size __read_mostly;
114 /* cpus with the lowest and highest unit numbers */
115 static unsigned int pcpu_first_unit_cpu __read_mostly;
116 static unsigned int pcpu_last_unit_cpu __read_mostly;
118 /* the address of the first chunk which starts with the kernel static area */
119 void *pcpu_base_addr __read_mostly;
120 EXPORT_SYMBOL_GPL(pcpu_base_addr);
122 static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
123 const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
125 /* group information, used for vm allocation */
126 static int pcpu_nr_groups __read_mostly;
127 static const unsigned long *pcpu_group_offsets __read_mostly;
128 static const size_t *pcpu_group_sizes __read_mostly;
131 * The first chunk which always exists. Note that unlike other
132 * chunks, this one can be allocated and mapped in several different
133 * ways and thus often doesn't live in the vmalloc area.
135 static struct pcpu_chunk *pcpu_first_chunk;
138 * Optional reserved chunk. This chunk reserves part of the first
139 * chunk and serves it for reserved allocations. The amount of
140 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
141 * area doesn't exist, the following variables contain NULL and 0
144 static struct pcpu_chunk *pcpu_reserved_chunk;
145 static int pcpu_reserved_chunk_limit;
148 * Synchronization rules.
150 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
151 * protects allocation/reclaim paths, chunks, populated bitmap and
152 * vmalloc mapping. The latter is a spinlock and protects the index
153 * data structures - chunk slots, chunks and area maps in chunks.
155 * During allocation, pcpu_alloc_mutex is kept locked all the time and
156 * pcpu_lock is grabbed and released as necessary. All actual memory
157 * allocations are done using GFP_KERNEL with pcpu_lock released. In
158 * general, percpu memory can't be allocated with irq off but
159 * irqsave/restore are still used in alloc path so that it can be used
160 * from early init path - sched_init() specifically.
162 * Free path accesses and alters only the index data structures, so it
163 * can be safely called from atomic context. When memory needs to be
164 * returned to the system, free path schedules reclaim_work which
165 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
166 * reclaimed, release both locks and frees the chunks. Note that it's
167 * necessary to grab both locks to remove a chunk from circulation as
168 * allocation path might be referencing the chunk with only
169 * pcpu_alloc_mutex locked.
171 static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
172 static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
174 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
176 /* reclaim work to release fully free chunks, scheduled from free path */
177 static void pcpu_reclaim(struct work_struct *work);
178 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
180 static bool pcpu_addr_in_first_chunk(void *addr)
182 void *first_start = pcpu_first_chunk->base_addr;
184 return addr >= first_start && addr < first_start + pcpu_unit_size;
187 static bool pcpu_addr_in_reserved_chunk(void *addr)
189 void *first_start = pcpu_first_chunk->base_addr;
191 return addr >= first_start &&
192 addr < first_start + pcpu_reserved_chunk_limit;
195 static int __pcpu_size_to_slot(int size)
197 int highbit = fls(size); /* size is in bytes */
198 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
201 static int pcpu_size_to_slot(int size)
203 if (size == pcpu_unit_size)
204 return pcpu_nr_slots - 1;
205 return __pcpu_size_to_slot(size);
208 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
210 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
213 return pcpu_size_to_slot(chunk->free_size);
216 static int pcpu_page_idx(unsigned int cpu, int page_idx)
218 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
221 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
222 unsigned int cpu, int page_idx)
224 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
225 (page_idx << PAGE_SHIFT);
228 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
229 unsigned int cpu, int page_idx)
231 /* must not be used on pre-mapped chunk */
232 WARN_ON(chunk->immutable);
234 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
237 /* set the pointer to a chunk in a page struct */
238 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
240 page->index = (unsigned long)pcpu;
243 /* obtain pointer to a chunk from a page struct */
244 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
246 return (struct pcpu_chunk *)page->index;
249 static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
251 *rs = find_next_zero_bit(chunk->populated, end, *rs);
252 *re = find_next_bit(chunk->populated, end, *rs + 1);
255 static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
257 *rs = find_next_bit(chunk->populated, end, *rs);
258 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
262 * (Un)populated page region iterators. Iterate over (un)populated
263 * page regions betwen @start and @end in @chunk. @rs and @re should
264 * be integer variables and will be set to start and end page index of
265 * the current region.
267 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
272 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
278 * pcpu_mem_alloc - allocate memory
279 * @size: bytes to allocate
281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
282 * kzalloc() is used; otherwise, vmalloc() is used. The returned
283 * memory is always zeroed.
286 * Does GFP_KERNEL allocation.
289 * Pointer to the allocated area on success, NULL on failure.
291 static void *pcpu_mem_alloc(size_t size)
293 if (size <= PAGE_SIZE)
294 return kzalloc(size, GFP_KERNEL);
296 void *ptr = vmalloc(size);
298 memset(ptr, 0, size);
304 * pcpu_mem_free - free memory
305 * @ptr: memory to free
306 * @size: size of the area
308 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
310 static void pcpu_mem_free(void *ptr, size_t size)
312 if (size <= PAGE_SIZE)
319 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
320 * @chunk: chunk of interest
321 * @oslot: the previous slot it was on
323 * This function is called after an allocation or free changed @chunk.
324 * New slot according to the changed state is determined and @chunk is
325 * moved to the slot. Note that the reserved chunk is never put on
331 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
333 int nslot = pcpu_chunk_slot(chunk);
335 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
337 list_move(&chunk->list, &pcpu_slot[nslot]);
339 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
344 * pcpu_chunk_addr_search - determine chunk containing specified address
345 * @addr: address for which the chunk needs to be determined.
348 * The address of the found chunk.
350 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
352 /* is it in the first chunk? */
353 if (pcpu_addr_in_first_chunk(addr)) {
354 /* is it in the reserved area? */
355 if (pcpu_addr_in_reserved_chunk(addr))
356 return pcpu_reserved_chunk;
357 return pcpu_first_chunk;
361 * The address is relative to unit0 which might be unused and
362 * thus unmapped. Offset the address to the unit space of the
363 * current processor before looking it up in the vmalloc
364 * space. Note that any possible cpu id can be used here, so
365 * there's no need to worry about preemption or cpu hotplug.
367 addr += pcpu_unit_offsets[raw_smp_processor_id()];
368 return pcpu_get_page_chunk(vmalloc_to_page(addr));
372 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
373 * @chunk: chunk of interest
375 * Determine whether area map of @chunk needs to be extended to
376 * accomodate a new allocation.
382 * New target map allocation length if extension is necessary, 0
385 static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
389 if (chunk->map_alloc >= chunk->map_used + 2)
392 new_alloc = PCPU_DFL_MAP_ALLOC;
393 while (new_alloc < chunk->map_used + 2)
400 * pcpu_extend_area_map - extend area map of a chunk
401 * @chunk: chunk of interest
402 * @new_alloc: new target allocation length of the area map
404 * Extend area map of @chunk to have @new_alloc entries.
407 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
410 * 0 on success, -errno on failure.
412 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
414 int *old = NULL, *new = NULL;
415 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
418 new = pcpu_mem_alloc(new_size);
422 /* acquire pcpu_lock and switch to new area map */
423 spin_lock_irqsave(&pcpu_lock, flags);
425 if (new_alloc <= chunk->map_alloc)
428 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
429 memcpy(new, chunk->map, old_size);
432 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
433 * one of the first chunks and still using static map.
435 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
438 chunk->map_alloc = new_alloc;
443 spin_unlock_irqrestore(&pcpu_lock, flags);
446 * pcpu_mem_free() might end up calling vfree() which uses
447 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
449 pcpu_mem_free(old, old_size);
450 pcpu_mem_free(new, new_size);
456 * pcpu_split_block - split a map block
457 * @chunk: chunk of interest
458 * @i: index of map block to split
459 * @head: head size in bytes (can be 0)
460 * @tail: tail size in bytes (can be 0)
462 * Split the @i'th map block into two or three blocks. If @head is
463 * non-zero, @head bytes block is inserted before block @i moving it
464 * to @i+1 and reducing its size by @head bytes.
466 * If @tail is non-zero, the target block, which can be @i or @i+1
467 * depending on @head, is reduced by @tail bytes and @tail byte block
468 * is inserted after the target block.
470 * @chunk->map must have enough free slots to accomodate the split.
475 static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
478 int nr_extra = !!head + !!tail;
480 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
482 /* insert new subblocks */
483 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
484 sizeof(chunk->map[0]) * (chunk->map_used - i));
485 chunk->map_used += nr_extra;
488 chunk->map[i + 1] = chunk->map[i] - head;
489 chunk->map[i++] = head;
492 chunk->map[i++] -= tail;
493 chunk->map[i] = tail;
498 * pcpu_alloc_area - allocate area from a pcpu_chunk
499 * @chunk: chunk of interest
500 * @size: wanted size in bytes
501 * @align: wanted align
503 * Try to allocate @size bytes area aligned at @align from @chunk.
504 * Note that this function only allocates the offset. It doesn't
505 * populate or map the area.
507 * @chunk->map must have at least two free slots.
513 * Allocated offset in @chunk on success, -1 if no matching area is
516 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
518 int oslot = pcpu_chunk_slot(chunk);
522 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
523 bool is_last = i + 1 == chunk->map_used;
526 /* extra for alignment requirement */
527 head = ALIGN(off, align) - off;
528 BUG_ON(i == 0 && head != 0);
530 if (chunk->map[i] < 0)
532 if (chunk->map[i] < head + size) {
533 max_contig = max(chunk->map[i], max_contig);
538 * If head is small or the previous block is free,
539 * merge'em. Note that 'small' is defined as smaller
540 * than sizeof(int), which is very small but isn't too
541 * uncommon for percpu allocations.
543 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
544 if (chunk->map[i - 1] > 0)
545 chunk->map[i - 1] += head;
547 chunk->map[i - 1] -= head;
548 chunk->free_size -= head;
550 chunk->map[i] -= head;
555 /* if tail is small, just keep it around */
556 tail = chunk->map[i] - head - size;
557 if (tail < sizeof(int))
560 /* split if warranted */
562 pcpu_split_block(chunk, i, head, tail);
566 max_contig = max(chunk->map[i - 1], max_contig);
569 max_contig = max(chunk->map[i + 1], max_contig);
572 /* update hint and mark allocated */
574 chunk->contig_hint = max_contig; /* fully scanned */
576 chunk->contig_hint = max(chunk->contig_hint,
579 chunk->free_size -= chunk->map[i];
580 chunk->map[i] = -chunk->map[i];
582 pcpu_chunk_relocate(chunk, oslot);
586 chunk->contig_hint = max_contig; /* fully scanned */
587 pcpu_chunk_relocate(chunk, oslot);
589 /* tell the upper layer that this chunk has no matching area */
594 * pcpu_free_area - free area to a pcpu_chunk
595 * @chunk: chunk of interest
596 * @freeme: offset of area to free
598 * Free area starting from @freeme to @chunk. Note that this function
599 * only modifies the allocation map. It doesn't depopulate or unmap
605 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
607 int oslot = pcpu_chunk_slot(chunk);
610 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
613 BUG_ON(off != freeme);
614 BUG_ON(chunk->map[i] > 0);
616 chunk->map[i] = -chunk->map[i];
617 chunk->free_size += chunk->map[i];
619 /* merge with previous? */
620 if (i > 0 && chunk->map[i - 1] >= 0) {
621 chunk->map[i - 1] += chunk->map[i];
623 memmove(&chunk->map[i], &chunk->map[i + 1],
624 (chunk->map_used - i) * sizeof(chunk->map[0]));
627 /* merge with next? */
628 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
629 chunk->map[i] += chunk->map[i + 1];
631 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
632 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
635 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
636 pcpu_chunk_relocate(chunk, oslot);
640 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
641 * @chunk: chunk of interest
642 * @bitmapp: output parameter for bitmap
643 * @may_alloc: may allocate the array
645 * Returns pointer to array of pointers to struct page and bitmap,
646 * both of which can be indexed with pcpu_page_idx(). The returned
647 * array is cleared to zero and *@bitmapp is copied from
648 * @chunk->populated. Note that there is only one array and bitmap
649 * and access exclusion is the caller's responsibility.
652 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
653 * Otherwise, don't care.
656 * Pointer to temp pages array on success, NULL on failure.
658 static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
659 unsigned long **bitmapp,
662 static struct page **pages;
663 static unsigned long *bitmap;
664 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
665 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
666 sizeof(unsigned long);
668 if (!pages || !bitmap) {
669 if (may_alloc && !pages)
670 pages = pcpu_mem_alloc(pages_size);
671 if (may_alloc && !bitmap)
672 bitmap = pcpu_mem_alloc(bitmap_size);
673 if (!pages || !bitmap)
677 memset(pages, 0, pages_size);
678 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
685 * pcpu_free_pages - free pages which were allocated for @chunk
686 * @chunk: chunk pages were allocated for
687 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
688 * @populated: populated bitmap
689 * @page_start: page index of the first page to be freed
690 * @page_end: page index of the last page to be freed + 1
692 * Free pages [@page_start and @page_end) in @pages for all units.
693 * The pages were allocated for @chunk.
695 static void pcpu_free_pages(struct pcpu_chunk *chunk,
696 struct page **pages, unsigned long *populated,
697 int page_start, int page_end)
702 for_each_possible_cpu(cpu) {
703 for (i = page_start; i < page_end; i++) {
704 struct page *page = pages[pcpu_page_idx(cpu, i)];
713 * pcpu_alloc_pages - allocates pages for @chunk
714 * @chunk: target chunk
715 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
716 * @populated: populated bitmap
717 * @page_start: page index of the first page to be allocated
718 * @page_end: page index of the last page to be allocated + 1
720 * Allocate pages [@page_start,@page_end) into @pages for all units.
721 * The allocation is for @chunk. Percpu core doesn't care about the
722 * content of @pages and will pass it verbatim to pcpu_map_pages().
724 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
725 struct page **pages, unsigned long *populated,
726 int page_start, int page_end)
728 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
732 for_each_possible_cpu(cpu) {
733 for (i = page_start; i < page_end; i++) {
734 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
736 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
738 pcpu_free_pages(chunk, pages, populated,
739 page_start, page_end);
748 * pcpu_pre_unmap_flush - flush cache prior to unmapping
749 * @chunk: chunk the regions to be flushed belongs to
750 * @page_start: page index of the first page to be flushed
751 * @page_end: page index of the last page to be flushed + 1
753 * Pages in [@page_start,@page_end) of @chunk are about to be
754 * unmapped. Flush cache. As each flushing trial can be very
755 * expensive, issue flush on the whole region at once rather than
756 * doing it for each cpu. This could be an overkill but is more
759 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
760 int page_start, int page_end)
763 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
764 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
767 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
769 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
773 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
774 * @chunk: chunk of interest
775 * @pages: pages array which can be used to pass information to free
776 * @populated: populated bitmap
777 * @page_start: page index of the first page to unmap
778 * @page_end: page index of the last page to unmap + 1
780 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
781 * Corresponding elements in @pages were cleared by the caller and can
782 * be used to carry information to pcpu_free_pages() which will be
783 * called after all unmaps are finished. The caller should call
784 * proper pre/post flush functions.
786 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
787 struct page **pages, unsigned long *populated,
788 int page_start, int page_end)
793 for_each_possible_cpu(cpu) {
794 for (i = page_start; i < page_end; i++) {
797 page = pcpu_chunk_page(chunk, cpu, i);
799 pages[pcpu_page_idx(cpu, i)] = page;
801 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
802 page_end - page_start);
805 for (i = page_start; i < page_end; i++)
806 __clear_bit(i, populated);
810 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
811 * @chunk: pcpu_chunk the regions to be flushed belong to
812 * @page_start: page index of the first page to be flushed
813 * @page_end: page index of the last page to be flushed + 1
815 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
816 * TLB for the regions. This can be skipped if the area is to be
817 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
819 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
820 * for the whole region.
822 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
823 int page_start, int page_end)
825 flush_tlb_kernel_range(
826 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
827 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
830 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
833 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
838 * pcpu_map_pages - map pages into a pcpu_chunk
839 * @chunk: chunk of interest
840 * @pages: pages array containing pages to be mapped
841 * @populated: populated bitmap
842 * @page_start: page index of the first page to map
843 * @page_end: page index of the last page to map + 1
845 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
846 * caller is responsible for calling pcpu_post_map_flush() after all
847 * mappings are complete.
849 * This function is responsible for setting corresponding bits in
850 * @chunk->populated bitmap and whatever is necessary for reverse
851 * lookup (addr -> chunk).
853 static int pcpu_map_pages(struct pcpu_chunk *chunk,
854 struct page **pages, unsigned long *populated,
855 int page_start, int page_end)
857 unsigned int cpu, tcpu;
860 for_each_possible_cpu(cpu) {
861 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
862 &pages[pcpu_page_idx(cpu, page_start)],
863 page_end - page_start);
868 /* mapping successful, link chunk and mark populated */
869 for (i = page_start; i < page_end; i++) {
870 for_each_possible_cpu(cpu)
871 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
873 __set_bit(i, populated);
879 for_each_possible_cpu(tcpu) {
882 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
883 page_end - page_start);
889 * pcpu_post_map_flush - flush cache after mapping
890 * @chunk: pcpu_chunk the regions to be flushed belong to
891 * @page_start: page index of the first page to be flushed
892 * @page_end: page index of the last page to be flushed + 1
894 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
897 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
898 * for the whole region.
900 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
901 int page_start, int page_end)
904 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
905 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
909 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
910 * @chunk: chunk to depopulate
911 * @off: offset to the area to depopulate
912 * @size: size of the area to depopulate in bytes
913 * @flush: whether to flush cache and tlb or not
915 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
916 * from @chunk. If @flush is true, vcache is flushed before unmapping
922 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
924 int page_start = PFN_DOWN(off);
925 int page_end = PFN_UP(off + size);
927 unsigned long *populated;
930 /* quick path, check whether it's empty already */
932 pcpu_next_unpop(chunk, &rs, &re, page_end);
933 if (rs == page_start && re == page_end)
936 /* immutable chunks can't be depopulated */
937 WARN_ON(chunk->immutable);
940 * If control reaches here, there must have been at least one
941 * successful population attempt so the temp pages array must
944 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
948 pcpu_pre_unmap_flush(chunk, page_start, page_end);
950 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
951 pcpu_unmap_pages(chunk, pages, populated, rs, re);
953 /* no need to flush tlb, vmalloc will handle it lazily */
955 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
956 pcpu_free_pages(chunk, pages, populated, rs, re);
958 /* commit new bitmap */
959 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
963 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
964 * @chunk: chunk of interest
965 * @off: offset to the area to populate
966 * @size: size of the area to populate in bytes
968 * For each cpu, populate and map pages [@page_start,@page_end) into
969 * @chunk. The area is cleared on return.
972 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
974 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
976 int page_start = PFN_DOWN(off);
977 int page_end = PFN_UP(off + size);
978 int free_end = page_start, unmap_end = page_start;
980 unsigned long *populated;
984 /* quick path, check whether all pages are already there */
986 pcpu_next_pop(chunk, &rs, &re, page_end);
987 if (rs == page_start && re == page_end)
990 /* need to allocate and map pages, this chunk can't be immutable */
991 WARN_ON(chunk->immutable);
993 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
998 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
999 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
1005 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
1006 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
1011 pcpu_post_map_flush(chunk, page_start, page_end);
1013 /* commit new bitmap */
1014 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
1016 for_each_possible_cpu(cpu)
1017 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1021 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
1022 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
1023 pcpu_unmap_pages(chunk, pages, populated, rs, re);
1024 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
1026 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
1027 pcpu_free_pages(chunk, pages, populated, rs, re);
1031 static void free_pcpu_chunk(struct pcpu_chunk *chunk)
1036 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
1037 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
1041 static struct pcpu_chunk *alloc_pcpu_chunk(void)
1043 struct pcpu_chunk *chunk;
1045 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1049 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
1050 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1051 chunk->map[chunk->map_used++] = pcpu_unit_size;
1053 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1054 pcpu_nr_groups, pcpu_atom_size,
1057 free_pcpu_chunk(chunk);
1061 INIT_LIST_HEAD(&chunk->list);
1062 chunk->free_size = pcpu_unit_size;
1063 chunk->contig_hint = pcpu_unit_size;
1064 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
1070 * pcpu_alloc - the percpu allocator
1071 * @size: size of area to allocate in bytes
1072 * @align: alignment of area (max PAGE_SIZE)
1073 * @reserved: allocate from the reserved chunk if available
1075 * Allocate percpu area of @size bytes aligned at @align.
1078 * Does GFP_KERNEL allocation.
1081 * Percpu pointer to the allocated area on success, NULL on failure.
1083 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
1085 static int warn_limit = 10;
1086 struct pcpu_chunk *chunk;
1088 int slot, off, new_alloc;
1089 unsigned long flags;
1091 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
1092 WARN(true, "illegal size (%zu) or align (%zu) for "
1093 "percpu allocation\n", size, align);
1097 mutex_lock(&pcpu_alloc_mutex);
1098 spin_lock_irqsave(&pcpu_lock, flags);
1100 /* serve reserved allocations from the reserved chunk if available */
1101 if (reserved && pcpu_reserved_chunk) {
1102 chunk = pcpu_reserved_chunk;
1104 if (size > chunk->contig_hint) {
1105 err = "alloc from reserved chunk failed";
1109 while ((new_alloc = pcpu_need_to_extend(chunk))) {
1110 spin_unlock_irqrestore(&pcpu_lock, flags);
1111 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
1112 err = "failed to extend area map of reserved chunk";
1113 goto fail_unlock_mutex;
1115 spin_lock_irqsave(&pcpu_lock, flags);
1118 off = pcpu_alloc_area(chunk, size, align);
1122 err = "alloc from reserved chunk failed";
1127 /* search through normal chunks */
1128 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1129 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1130 if (size > chunk->contig_hint)
1133 new_alloc = pcpu_need_to_extend(chunk);
1135 spin_unlock_irqrestore(&pcpu_lock, flags);
1136 if (pcpu_extend_area_map(chunk,
1138 err = "failed to extend area map";
1139 goto fail_unlock_mutex;
1141 spin_lock_irqsave(&pcpu_lock, flags);
1143 * pcpu_lock has been dropped, need to
1144 * restart cpu_slot list walking.
1149 off = pcpu_alloc_area(chunk, size, align);
1155 /* hmmm... no space left, create a new chunk */
1156 spin_unlock_irqrestore(&pcpu_lock, flags);
1158 chunk = alloc_pcpu_chunk();
1160 err = "failed to allocate new chunk";
1161 goto fail_unlock_mutex;
1164 spin_lock_irqsave(&pcpu_lock, flags);
1165 pcpu_chunk_relocate(chunk, -1);
1169 spin_unlock_irqrestore(&pcpu_lock, flags);
1171 /* populate, map and clear the area */
1172 if (pcpu_populate_chunk(chunk, off, size)) {
1173 spin_lock_irqsave(&pcpu_lock, flags);
1174 pcpu_free_area(chunk, off);
1175 err = "failed to populate";
1179 mutex_unlock(&pcpu_alloc_mutex);
1181 /* return address relative to base address */
1182 return __addr_to_pcpu_ptr(chunk->base_addr + off);
1185 spin_unlock_irqrestore(&pcpu_lock, flags);
1187 mutex_unlock(&pcpu_alloc_mutex);
1189 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1190 "%s\n", size, align, err);
1193 pr_info("PERCPU: limit reached, disable warning\n");
1199 * __alloc_percpu - allocate dynamic percpu area
1200 * @size: size of area to allocate in bytes
1201 * @align: alignment of area (max PAGE_SIZE)
1203 * Allocate percpu area of @size bytes aligned at @align. Might
1204 * sleep. Might trigger writeouts.
1207 * Does GFP_KERNEL allocation.
1210 * Percpu pointer to the allocated area on success, NULL on failure.
1212 void __percpu *__alloc_percpu(size_t size, size_t align)
1214 return pcpu_alloc(size, align, false);
1216 EXPORT_SYMBOL_GPL(__alloc_percpu);
1219 * __alloc_reserved_percpu - allocate reserved percpu area
1220 * @size: size of area to allocate in bytes
1221 * @align: alignment of area (max PAGE_SIZE)
1223 * Allocate percpu area of @size bytes aligned at @align from reserved
1224 * percpu area if arch has set it up; otherwise, allocation is served
1225 * from the same dynamic area. Might sleep. Might trigger writeouts.
1228 * Does GFP_KERNEL allocation.
1231 * Percpu pointer to the allocated area on success, NULL on failure.
1233 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1235 return pcpu_alloc(size, align, true);
1239 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1242 * Reclaim all fully free chunks except for the first one.
1245 * workqueue context.
1247 static void pcpu_reclaim(struct work_struct *work)
1250 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1251 struct pcpu_chunk *chunk, *next;
1253 mutex_lock(&pcpu_alloc_mutex);
1254 spin_lock_irq(&pcpu_lock);
1256 list_for_each_entry_safe(chunk, next, head, list) {
1257 WARN_ON(chunk->immutable);
1259 /* spare the first one */
1260 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1263 list_move(&chunk->list, &todo);
1266 spin_unlock_irq(&pcpu_lock);
1268 list_for_each_entry_safe(chunk, next, &todo, list) {
1269 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
1270 free_pcpu_chunk(chunk);
1273 mutex_unlock(&pcpu_alloc_mutex);
1277 * free_percpu - free percpu area
1278 * @ptr: pointer to area to free
1280 * Free percpu area @ptr.
1283 * Can be called from atomic context.
1285 void free_percpu(void __percpu *ptr)
1288 struct pcpu_chunk *chunk;
1289 unsigned long flags;
1295 addr = __pcpu_ptr_to_addr(ptr);
1297 spin_lock_irqsave(&pcpu_lock, flags);
1299 chunk = pcpu_chunk_addr_search(addr);
1300 off = addr - chunk->base_addr;
1302 pcpu_free_area(chunk, off);
1304 /* if there are more than one fully free chunks, wake up grim reaper */
1305 if (chunk->free_size == pcpu_unit_size) {
1306 struct pcpu_chunk *pos;
1308 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1310 schedule_work(&pcpu_reclaim_work);
1315 spin_unlock_irqrestore(&pcpu_lock, flags);
1317 EXPORT_SYMBOL_GPL(free_percpu);
1320 * is_kernel_percpu_address - test whether address is from static percpu area
1321 * @addr: address to test
1323 * Test whether @addr belongs to in-kernel static percpu area. Module
1324 * static percpu areas are not considered. For those, use
1325 * is_module_percpu_address().
1328 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1330 bool is_kernel_percpu_address(unsigned long addr)
1332 const size_t static_size = __per_cpu_end - __per_cpu_start;
1333 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1336 for_each_possible_cpu(cpu) {
1337 void *start = per_cpu_ptr(base, cpu);
1339 if ((void *)addr >= start && (void *)addr < start + static_size)
1346 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1347 * @addr: the address to be converted to physical address
1349 * Given @addr which is dereferenceable address obtained via one of
1350 * percpu access macros, this function translates it into its physical
1351 * address. The caller is responsible for ensuring @addr stays valid
1352 * until this function finishes.
1355 * The physical address for @addr.
1357 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1359 if (pcpu_addr_in_first_chunk(addr)) {
1360 if ((unsigned long)addr < VMALLOC_START ||
1361 (unsigned long)addr >= VMALLOC_END)
1364 return page_to_phys(vmalloc_to_page(addr));
1366 return page_to_phys(vmalloc_to_page(addr));
1369 static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1370 size_t reserved_size,
1375 size_sum = PFN_ALIGN(static_size + reserved_size +
1376 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1377 if (*dyn_sizep != 0)
1378 *dyn_sizep = size_sum - static_size - reserved_size;
1384 * pcpu_alloc_alloc_info - allocate percpu allocation info
1385 * @nr_groups: the number of groups
1386 * @nr_units: the number of units
1388 * Allocate ai which is large enough for @nr_groups groups containing
1389 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1390 * cpu_map array which is long enough for @nr_units and filled with
1391 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1392 * pointer of other groups.
1395 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1398 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1401 struct pcpu_alloc_info *ai;
1402 size_t base_size, ai_size;
1406 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1407 __alignof__(ai->groups[0].cpu_map[0]));
1408 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1410 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1416 ai->groups[0].cpu_map = ptr;
1418 for (unit = 0; unit < nr_units; unit++)
1419 ai->groups[0].cpu_map[unit] = NR_CPUS;
1421 ai->nr_groups = nr_groups;
1422 ai->__ai_size = PFN_ALIGN(ai_size);
1428 * pcpu_free_alloc_info - free percpu allocation info
1429 * @ai: pcpu_alloc_info to free
1431 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1433 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1435 free_bootmem(__pa(ai), ai->__ai_size);
1439 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1440 * @reserved_size: the size of reserved percpu area in bytes
1441 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1442 * @atom_size: allocation atom size
1443 * @cpu_distance_fn: callback to determine distance between cpus, optional
1445 * This function determines grouping of units, their mappings to cpus
1446 * and other parameters considering needed percpu size, allocation
1447 * atom size and distances between CPUs.
1449 * Groups are always mutliples of atom size and CPUs which are of
1450 * LOCAL_DISTANCE both ways are grouped together and share space for
1451 * units in the same group. The returned configuration is guaranteed
1452 * to have CPUs on different nodes on different groups and >=75% usage
1453 * of allocated virtual address space.
1456 * On success, pointer to the new allocation_info is returned. On
1457 * failure, ERR_PTR value is returned.
1459 struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1460 size_t reserved_size, ssize_t dyn_size,
1462 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1464 static int group_map[NR_CPUS] __initdata;
1465 static int group_cnt[NR_CPUS] __initdata;
1466 const size_t static_size = __per_cpu_end - __per_cpu_start;
1467 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
1468 size_t size_sum, min_unit_size, alloc_size;
1469 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1470 int last_allocs, group, unit;
1471 unsigned int cpu, tcpu;
1472 struct pcpu_alloc_info *ai;
1473 unsigned int *cpu_map;
1475 /* this function may be called multiple times */
1476 memset(group_map, 0, sizeof(group_map));
1477 memset(group_cnt, 0, sizeof(group_map));
1480 * Determine min_unit_size, alloc_size and max_upa such that
1481 * alloc_size is multiple of atom_size and is the smallest
1482 * which can accomodate 4k aligned segments which are equal to
1483 * or larger than min_unit_size.
1485 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
1486 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1488 alloc_size = roundup(min_unit_size, atom_size);
1489 upa = alloc_size / min_unit_size;
1490 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1494 /* group cpus according to their proximity */
1495 for_each_possible_cpu(cpu) {
1498 for_each_possible_cpu(tcpu) {
1501 if (group_map[tcpu] == group && cpu_distance_fn &&
1502 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1503 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1505 nr_groups = max(nr_groups, group + 1);
1509 group_map[cpu] = group;
1511 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1515 * Expand unit size until address space usage goes over 75%
1516 * and then as much as possible without using more address
1519 last_allocs = INT_MAX;
1520 for (upa = max_upa; upa; upa--) {
1521 int allocs = 0, wasted = 0;
1523 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1526 for (group = 0; group < nr_groups; group++) {
1527 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1528 allocs += this_allocs;
1529 wasted += this_allocs * upa - group_cnt[group];
1533 * Don't accept if wastage is over 25%. The
1534 * greater-than comparison ensures upa==1 always
1535 * passes the following check.
1537 if (wasted > num_possible_cpus() / 3)
1540 /* and then don't consume more memory */
1541 if (allocs > last_allocs)
1543 last_allocs = allocs;
1548 /* allocate and fill alloc_info */
1549 for (group = 0; group < nr_groups; group++)
1550 nr_units += roundup(group_cnt[group], upa);
1552 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1554 return ERR_PTR(-ENOMEM);
1555 cpu_map = ai->groups[0].cpu_map;
1557 for (group = 0; group < nr_groups; group++) {
1558 ai->groups[group].cpu_map = cpu_map;
1559 cpu_map += roundup(group_cnt[group], upa);
1562 ai->static_size = static_size;
1563 ai->reserved_size = reserved_size;
1564 ai->dyn_size = dyn_size;
1565 ai->unit_size = alloc_size / upa;
1566 ai->atom_size = atom_size;
1567 ai->alloc_size = alloc_size;
1569 for (group = 0, unit = 0; group_cnt[group]; group++) {
1570 struct pcpu_group_info *gi = &ai->groups[group];
1573 * Initialize base_offset as if all groups are located
1574 * back-to-back. The caller should update this to
1575 * reflect actual allocation.
1577 gi->base_offset = unit * ai->unit_size;
1579 for_each_possible_cpu(cpu)
1580 if (group_map[cpu] == group)
1581 gi->cpu_map[gi->nr_units++] = cpu;
1582 gi->nr_units = roundup(gi->nr_units, upa);
1583 unit += gi->nr_units;
1585 BUG_ON(unit != nr_units);
1591 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1593 * @ai: allocation info to dump
1595 * Print out information about @ai using loglevel @lvl.
1597 static void pcpu_dump_alloc_info(const char *lvl,
1598 const struct pcpu_alloc_info *ai)
1600 int group_width = 1, cpu_width = 1, width;
1601 char empty_str[] = "--------";
1602 int alloc = 0, alloc_end = 0;
1604 int upa, apl; /* units per alloc, allocs per line */
1610 v = num_possible_cpus();
1613 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1615 upa = ai->alloc_size / ai->unit_size;
1616 width = upa * (cpu_width + 1) + group_width + 3;
1617 apl = rounddown_pow_of_two(max(60 / width, 1));
1619 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1620 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1621 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1623 for (group = 0; group < ai->nr_groups; group++) {
1624 const struct pcpu_group_info *gi = &ai->groups[group];
1625 int unit = 0, unit_end = 0;
1627 BUG_ON(gi->nr_units % upa);
1628 for (alloc_end += gi->nr_units / upa;
1629 alloc < alloc_end; alloc++) {
1630 if (!(alloc % apl)) {
1632 printk("%spcpu-alloc: ", lvl);
1634 printk("[%0*d] ", group_width, group);
1636 for (unit_end += upa; unit < unit_end; unit++)
1637 if (gi->cpu_map[unit] != NR_CPUS)
1638 printk("%0*d ", cpu_width,
1641 printk("%s ", empty_str);
1648 * pcpu_setup_first_chunk - initialize the first percpu chunk
1649 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1650 * @base_addr: mapped address
1652 * Initialize the first percpu chunk which contains the kernel static
1653 * perpcu area. This function is to be called from arch percpu area
1656 * @ai contains all information necessary to initialize the first
1657 * chunk and prime the dynamic percpu allocator.
1659 * @ai->static_size is the size of static percpu area.
1661 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1662 * reserve after the static area in the first chunk. This reserves
1663 * the first chunk such that it's available only through reserved
1664 * percpu allocation. This is primarily used to serve module percpu
1665 * static areas on architectures where the addressing model has
1666 * limited offset range for symbol relocations to guarantee module
1667 * percpu symbols fall inside the relocatable range.
1669 * @ai->dyn_size determines the number of bytes available for dynamic
1670 * allocation in the first chunk. The area between @ai->static_size +
1671 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1673 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1674 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1677 * @ai->atom_size is the allocation atom size and used as alignment
1680 * @ai->alloc_size is the allocation size and always multiple of
1681 * @ai->atom_size. This is larger than @ai->atom_size if
1682 * @ai->unit_size is larger than @ai->atom_size.
1684 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1685 * percpu areas. Units which should be colocated are put into the
1686 * same group. Dynamic VM areas will be allocated according to these
1687 * groupings. If @ai->nr_groups is zero, a single group containing
1688 * all units is assumed.
1690 * The caller should have mapped the first chunk at @base_addr and
1691 * copied static data to each unit.
1693 * If the first chunk ends up with both reserved and dynamic areas, it
1694 * is served by two chunks - one to serve the core static and reserved
1695 * areas and the other for the dynamic area. They share the same vm
1696 * and page map but uses different area allocation map to stay away
1697 * from each other. The latter chunk is circulated in the chunk slots
1698 * and available for dynamic allocation like any other chunks.
1701 * 0 on success, -errno on failure.
1703 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1706 static char cpus_buf[4096] __initdata;
1707 static int smap[2], dmap[2];
1708 size_t dyn_size = ai->dyn_size;
1709 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1710 struct pcpu_chunk *schunk, *dchunk = NULL;
1711 unsigned long *group_offsets;
1712 size_t *group_sizes;
1713 unsigned long *unit_off;
1718 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1720 #define PCPU_SETUP_BUG_ON(cond) do { \
1721 if (unlikely(cond)) { \
1722 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1723 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1724 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1730 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1731 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1732 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1733 PCPU_SETUP_BUG_ON(!ai->static_size);
1734 PCPU_SETUP_BUG_ON(!base_addr);
1735 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1736 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1737 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1739 /* process group information and build config tables accordingly */
1740 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1741 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1742 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1743 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1745 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1746 unit_map[cpu] = UINT_MAX;
1747 pcpu_first_unit_cpu = NR_CPUS;
1749 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1750 const struct pcpu_group_info *gi = &ai->groups[group];
1752 group_offsets[group] = gi->base_offset;
1753 group_sizes[group] = gi->nr_units * ai->unit_size;
1755 for (i = 0; i < gi->nr_units; i++) {
1756 cpu = gi->cpu_map[i];
1760 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1761 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1762 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1764 unit_map[cpu] = unit + i;
1765 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1767 if (pcpu_first_unit_cpu == NR_CPUS)
1768 pcpu_first_unit_cpu = cpu;
1771 pcpu_last_unit_cpu = cpu;
1772 pcpu_nr_units = unit;
1774 for_each_possible_cpu(cpu)
1775 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1777 /* we're done parsing the input, undefine BUG macro and dump config */
1778 #undef PCPU_SETUP_BUG_ON
1779 pcpu_dump_alloc_info(KERN_INFO, ai);
1781 pcpu_nr_groups = ai->nr_groups;
1782 pcpu_group_offsets = group_offsets;
1783 pcpu_group_sizes = group_sizes;
1784 pcpu_unit_map = unit_map;
1785 pcpu_unit_offsets = unit_off;
1787 /* determine basic parameters */
1788 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1789 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1790 pcpu_atom_size = ai->atom_size;
1791 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1792 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1795 * Allocate chunk slots. The additional last slot is for
1798 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1799 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1800 for (i = 0; i < pcpu_nr_slots; i++)
1801 INIT_LIST_HEAD(&pcpu_slot[i]);
1804 * Initialize static chunk. If reserved_size is zero, the
1805 * static chunk covers static area + dynamic allocation area
1806 * in the first chunk. If reserved_size is not zero, it
1807 * covers static area + reserved area (mostly used for module
1808 * static percpu allocation).
1810 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1811 INIT_LIST_HEAD(&schunk->list);
1812 schunk->base_addr = base_addr;
1814 schunk->map_alloc = ARRAY_SIZE(smap);
1815 schunk->immutable = true;
1816 bitmap_fill(schunk->populated, pcpu_unit_pages);
1818 if (ai->reserved_size) {
1819 schunk->free_size = ai->reserved_size;
1820 pcpu_reserved_chunk = schunk;
1821 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1823 schunk->free_size = dyn_size;
1824 dyn_size = 0; /* dynamic area covered */
1826 schunk->contig_hint = schunk->free_size;
1828 schunk->map[schunk->map_used++] = -ai->static_size;
1829 if (schunk->free_size)
1830 schunk->map[schunk->map_used++] = schunk->free_size;
1832 /* init dynamic chunk if necessary */
1834 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1835 INIT_LIST_HEAD(&dchunk->list);
1836 dchunk->base_addr = base_addr;
1838 dchunk->map_alloc = ARRAY_SIZE(dmap);
1839 dchunk->immutable = true;
1840 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1842 dchunk->contig_hint = dchunk->free_size = dyn_size;
1843 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1844 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1847 /* link the first chunk in */
1848 pcpu_first_chunk = dchunk ?: schunk;
1849 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1852 pcpu_base_addr = base_addr;
1856 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1857 [PCPU_FC_AUTO] = "auto",
1858 [PCPU_FC_EMBED] = "embed",
1859 [PCPU_FC_PAGE] = "page",
1862 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1864 static int __init percpu_alloc_setup(char *str)
1868 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1869 else if (!strcmp(str, "embed"))
1870 pcpu_chosen_fc = PCPU_FC_EMBED;
1872 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1873 else if (!strcmp(str, "page"))
1874 pcpu_chosen_fc = PCPU_FC_PAGE;
1877 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1881 early_param("percpu_alloc", percpu_alloc_setup);
1883 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1884 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1886 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1887 * @reserved_size: the size of reserved percpu area in bytes
1888 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1889 * @atom_size: allocation atom size
1890 * @cpu_distance_fn: callback to determine distance between cpus, optional
1891 * @alloc_fn: function to allocate percpu page
1892 * @free_fn: funtion to free percpu page
1894 * This is a helper to ease setting up embedded first percpu chunk and
1895 * can be called where pcpu_setup_first_chunk() is expected.
1897 * If this function is used to setup the first chunk, it is allocated
1898 * by calling @alloc_fn and used as-is without being mapped into
1899 * vmalloc area. Allocations are always whole multiples of @atom_size
1900 * aligned to @atom_size.
1902 * This enables the first chunk to piggy back on the linear physical
1903 * mapping which often uses larger page size. Please note that this
1904 * can result in very sparse cpu->unit mapping on NUMA machines thus
1905 * requiring large vmalloc address space. Don't use this allocator if
1906 * vmalloc space is not orders of magnitude larger than distances
1907 * between node memory addresses (ie. 32bit NUMA machines).
1909 * When @dyn_size is positive, dynamic area might be larger than
1910 * specified to fill page alignment. When @dyn_size is auto,
1911 * @dyn_size is just big enough to fill page alignment after static
1912 * and reserved areas.
1914 * If the needed size is smaller than the minimum or specified unit
1915 * size, the leftover is returned using @free_fn.
1918 * 0 on success, -errno on failure.
1920 int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1922 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1923 pcpu_fc_alloc_fn_t alloc_fn,
1924 pcpu_fc_free_fn_t free_fn)
1926 void *base = (void *)ULONG_MAX;
1927 void **areas = NULL;
1928 struct pcpu_alloc_info *ai;
1929 size_t size_sum, areas_size, max_distance;
1932 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1937 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1938 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1940 areas = alloc_bootmem_nopanic(areas_size);
1946 /* allocate, copy and determine base address */
1947 for (group = 0; group < ai->nr_groups; group++) {
1948 struct pcpu_group_info *gi = &ai->groups[group];
1949 unsigned int cpu = NR_CPUS;
1952 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1953 cpu = gi->cpu_map[i];
1954 BUG_ON(cpu == NR_CPUS);
1956 /* allocate space for the whole group */
1957 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1960 goto out_free_areas;
1964 base = min(ptr, base);
1966 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1967 if (gi->cpu_map[i] == NR_CPUS) {
1968 /* unused unit, free whole */
1969 free_fn(ptr, ai->unit_size);
1972 /* copy and return the unused part */
1973 memcpy(ptr, __per_cpu_load, ai->static_size);
1974 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1978 /* base address is now known, determine group base offsets */
1980 for (group = 0; group < ai->nr_groups; group++) {
1981 ai->groups[group].base_offset = areas[group] - base;
1982 max_distance = max_t(size_t, max_distance,
1983 ai->groups[group].base_offset);
1985 max_distance += ai->unit_size;
1987 /* warn if maximum distance is further than 75% of vmalloc space */
1988 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1989 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1991 max_distance, VMALLOC_END - VMALLOC_START);
1992 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1993 /* and fail if we have fallback */
1999 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2000 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2001 ai->dyn_size, ai->unit_size);
2003 rc = pcpu_setup_first_chunk(ai, base);
2007 for (group = 0; group < ai->nr_groups; group++)
2008 free_fn(areas[group],
2009 ai->groups[group].nr_units * ai->unit_size);
2011 pcpu_free_alloc_info(ai);
2013 free_bootmem(__pa(areas), areas_size);
2016 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
2017 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
2019 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2021 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2022 * @reserved_size: the size of reserved percpu area in bytes
2023 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2024 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
2025 * @populate_pte_fn: function to populate pte
2027 * This is a helper to ease setting up page-remapped first percpu
2028 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2030 * This is the basic allocator. Static percpu area is allocated
2031 * page-by-page into vmalloc area.
2034 * 0 on success, -errno on failure.
2036 int __init pcpu_page_first_chunk(size_t reserved_size,
2037 pcpu_fc_alloc_fn_t alloc_fn,
2038 pcpu_fc_free_fn_t free_fn,
2039 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2041 static struct vm_struct vm;
2042 struct pcpu_alloc_info *ai;
2046 struct page **pages;
2049 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2051 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
2054 BUG_ON(ai->nr_groups != 1);
2055 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2057 unit_pages = ai->unit_size >> PAGE_SHIFT;
2059 /* unaligned allocations can't be freed, round up to page size */
2060 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2062 pages = alloc_bootmem(pages_size);
2064 /* allocate pages */
2066 for (unit = 0; unit < num_possible_cpus(); unit++)
2067 for (i = 0; i < unit_pages; i++) {
2068 unsigned int cpu = ai->groups[0].cpu_map[unit];
2071 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2073 pr_warning("PERCPU: failed to allocate %s page "
2074 "for cpu%u\n", psize_str, cpu);
2077 pages[j++] = virt_to_page(ptr);
2080 /* allocate vm area, map the pages and copy static data */
2081 vm.flags = VM_ALLOC;
2082 vm.size = num_possible_cpus() * ai->unit_size;
2083 vm_area_register_early(&vm, PAGE_SIZE);
2085 for (unit = 0; unit < num_possible_cpus(); unit++) {
2086 unsigned long unit_addr =
2087 (unsigned long)vm.addr + unit * ai->unit_size;
2089 for (i = 0; i < unit_pages; i++)
2090 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2092 /* pte already populated, the following shouldn't fail */
2093 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2096 panic("failed to map percpu area, err=%d\n", rc);
2099 * FIXME: Archs with virtual cache should flush local
2100 * cache for the linear mapping here - something
2101 * equivalent to flush_cache_vmap() on the local cpu.
2102 * flush_cache_vmap() can't be used as most supporting
2103 * data structures are not set up yet.
2106 /* copy static data */
2107 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2110 /* we're ready, commit */
2111 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2112 unit_pages, psize_str, vm.addr, ai->static_size,
2113 ai->reserved_size, ai->dyn_size);
2115 rc = pcpu_setup_first_chunk(ai, vm.addr);
2120 free_fn(page_address(pages[j]), PAGE_SIZE);
2123 free_bootmem(__pa(pages), pages_size);
2124 pcpu_free_alloc_info(ai);
2127 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
2130 * Generic percpu area setup.
2132 * The embedding helper is used because its behavior closely resembles
2133 * the original non-dynamic generic percpu area setup. This is
2134 * important because many archs have addressing restrictions and might
2135 * fail if the percpu area is located far away from the previous
2136 * location. As an added bonus, in non-NUMA cases, embedding is
2137 * generally a good idea TLB-wise because percpu area can piggy back
2138 * on the physical linear memory mapping which uses large page
2139 * mappings on applicable archs.
2141 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2142 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2143 EXPORT_SYMBOL(__per_cpu_offset);
2145 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2148 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2151 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2153 free_bootmem(__pa(ptr), size);
2156 void __init setup_per_cpu_areas(void)
2158 unsigned long delta;
2163 * Always reserve area for module percpu variables. That's
2164 * what the legacy allocator did.
2166 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2167 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2168 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2170 panic("Failed to initialized percpu areas.");
2172 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2173 for_each_possible_cpu(cpu)
2174 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2176 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */