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>
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 void *base_addr; /* base address of this chunk */
96 int map_used; /* # of map entries used */
97 int map_alloc; /* # of map entries allocated */
98 int *map; /* allocation map */
99 struct vm_struct **vms; /* mapped vmalloc regions */
100 bool immutable; /* no [de]population allowed */
101 unsigned long populated[]; /* populated bitmap */
104 static int pcpu_unit_pages __read_mostly;
105 static int pcpu_unit_size __read_mostly;
106 static int pcpu_nr_units __read_mostly;
107 static int pcpu_atom_size __read_mostly;
108 static int pcpu_nr_slots __read_mostly;
109 static size_t pcpu_chunk_struct_size __read_mostly;
111 /* cpus with the lowest and highest unit numbers */
112 static unsigned int pcpu_first_unit_cpu __read_mostly;
113 static unsigned int pcpu_last_unit_cpu __read_mostly;
115 /* the address of the first chunk which starts with the kernel static area */
116 void *pcpu_base_addr __read_mostly;
117 EXPORT_SYMBOL_GPL(pcpu_base_addr);
119 static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
120 const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
122 /* group information, used for vm allocation */
123 static int pcpu_nr_groups __read_mostly;
124 static const unsigned long *pcpu_group_offsets __read_mostly;
125 static const size_t *pcpu_group_sizes __read_mostly;
128 * The first chunk which always exists. Note that unlike other
129 * chunks, this one can be allocated and mapped in several different
130 * ways and thus often doesn't live in the vmalloc area.
132 static struct pcpu_chunk *pcpu_first_chunk;
135 * Optional reserved chunk. This chunk reserves part of the first
136 * chunk and serves it for reserved allocations. The amount of
137 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
138 * area doesn't exist, the following variables contain NULL and 0
141 static struct pcpu_chunk *pcpu_reserved_chunk;
142 static int pcpu_reserved_chunk_limit;
145 * Synchronization rules.
147 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
148 * protects allocation/reclaim paths, chunks, populated bitmap and
149 * vmalloc mapping. The latter is a spinlock and protects the index
150 * data structures - chunk slots, chunks and area maps in chunks.
152 * During allocation, pcpu_alloc_mutex is kept locked all the time and
153 * pcpu_lock is grabbed and released as necessary. All actual memory
154 * allocations are done using GFP_KERNEL with pcpu_lock released.
156 * Free path accesses and alters only the index data structures, so it
157 * can be safely called from atomic context. When memory needs to be
158 * returned to the system, free path schedules reclaim_work which
159 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
160 * reclaimed, release both locks and frees the chunks. Note that it's
161 * necessary to grab both locks to remove a chunk from circulation as
162 * allocation path might be referencing the chunk with only
163 * pcpu_alloc_mutex locked.
165 static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
166 static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
168 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
170 /* reclaim work to release fully free chunks, scheduled from free path */
171 static void pcpu_reclaim(struct work_struct *work);
172 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
174 static int __pcpu_size_to_slot(int size)
176 int highbit = fls(size); /* size is in bytes */
177 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
180 static int pcpu_size_to_slot(int size)
182 if (size == pcpu_unit_size)
183 return pcpu_nr_slots - 1;
184 return __pcpu_size_to_slot(size);
187 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
189 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
192 return pcpu_size_to_slot(chunk->free_size);
195 static int pcpu_page_idx(unsigned int cpu, int page_idx)
197 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
200 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
201 unsigned int cpu, int page_idx)
203 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
204 (page_idx << PAGE_SHIFT);
207 static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
208 unsigned int cpu, int page_idx)
210 /* must not be used on pre-mapped chunk */
211 WARN_ON(chunk->immutable);
213 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
216 /* set the pointer to a chunk in a page struct */
217 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
219 page->index = (unsigned long)pcpu;
222 /* obtain pointer to a chunk from a page struct */
223 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
225 return (struct pcpu_chunk *)page->index;
228 static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
230 *rs = find_next_zero_bit(chunk->populated, end, *rs);
231 *re = find_next_bit(chunk->populated, end, *rs + 1);
234 static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
236 *rs = find_next_bit(chunk->populated, end, *rs);
237 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
241 * (Un)populated page region iterators. Iterate over (un)populated
242 * page regions betwen @start and @end in @chunk. @rs and @re should
243 * be integer variables and will be set to start and end page index of
244 * the current region.
246 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
247 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
249 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
251 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
252 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
254 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
257 * pcpu_mem_alloc - allocate memory
258 * @size: bytes to allocate
260 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
261 * kzalloc() is used; otherwise, vmalloc() is used. The returned
262 * memory is always zeroed.
265 * Does GFP_KERNEL allocation.
268 * Pointer to the allocated area on success, NULL on failure.
270 static void *pcpu_mem_alloc(size_t size)
272 if (size <= PAGE_SIZE)
273 return kzalloc(size, GFP_KERNEL);
275 void *ptr = vmalloc(size);
277 memset(ptr, 0, size);
283 * pcpu_mem_free - free memory
284 * @ptr: memory to free
285 * @size: size of the area
287 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
289 static void pcpu_mem_free(void *ptr, size_t size)
291 if (size <= PAGE_SIZE)
298 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
299 * @chunk: chunk of interest
300 * @oslot: the previous slot it was on
302 * This function is called after an allocation or free changed @chunk.
303 * New slot according to the changed state is determined and @chunk is
304 * moved to the slot. Note that the reserved chunk is never put on
310 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
312 int nslot = pcpu_chunk_slot(chunk);
314 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
316 list_move(&chunk->list, &pcpu_slot[nslot]);
318 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
323 * pcpu_chunk_addr_search - determine chunk containing specified address
324 * @addr: address for which the chunk needs to be determined.
327 * The address of the found chunk.
329 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
331 void *first_start = pcpu_first_chunk->base_addr;
333 /* is it in the first chunk? */
334 if (addr >= first_start && addr < first_start + pcpu_unit_size) {
335 /* is it in the reserved area? */
336 if (addr < first_start + pcpu_reserved_chunk_limit)
337 return pcpu_reserved_chunk;
338 return pcpu_first_chunk;
342 * The address is relative to unit0 which might be unused and
343 * thus unmapped. Offset the address to the unit space of the
344 * current processor before looking it up in the vmalloc
345 * space. Note that any possible cpu id can be used here, so
346 * there's no need to worry about preemption or cpu hotplug.
348 addr += pcpu_unit_offsets[raw_smp_processor_id()];
349 return pcpu_get_page_chunk(vmalloc_to_page(addr));
353 * pcpu_extend_area_map - extend area map for allocation
354 * @chunk: target chunk
356 * Extend area map of @chunk so that it can accomodate an allocation.
357 * A single allocation can split an area into three areas, so this
358 * function makes sure that @chunk->map has at least two extra slots.
361 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
362 * if area map is extended.
365 * 0 if noop, 1 if successfully extended, -errno on failure.
367 static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
368 __releases(lock) __acquires(lock)
375 if (chunk->map_alloc >= chunk->map_used + 2)
378 spin_unlock_irq(&pcpu_lock);
380 new_alloc = PCPU_DFL_MAP_ALLOC;
381 while (new_alloc < chunk->map_used + 2)
384 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
386 spin_lock_irq(&pcpu_lock);
391 * Acquire pcpu_lock and switch to new area map. Only free
392 * could have happened inbetween, so map_used couldn't have
395 spin_lock_irq(&pcpu_lock);
396 BUG_ON(new_alloc < chunk->map_used + 2);
398 size = chunk->map_alloc * sizeof(chunk->map[0]);
399 memcpy(new, chunk->map, size);
402 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
403 * one of the first chunks and still using static map.
405 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
406 pcpu_mem_free(chunk->map, size);
408 chunk->map_alloc = new_alloc;
414 * pcpu_split_block - split a map block
415 * @chunk: chunk of interest
416 * @i: index of map block to split
417 * @head: head size in bytes (can be 0)
418 * @tail: tail size in bytes (can be 0)
420 * Split the @i'th map block into two or three blocks. If @head is
421 * non-zero, @head bytes block is inserted before block @i moving it
422 * to @i+1 and reducing its size by @head bytes.
424 * If @tail is non-zero, the target block, which can be @i or @i+1
425 * depending on @head, is reduced by @tail bytes and @tail byte block
426 * is inserted after the target block.
428 * @chunk->map must have enough free slots to accomodate the split.
433 static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
436 int nr_extra = !!head + !!tail;
438 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
440 /* insert new subblocks */
441 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
442 sizeof(chunk->map[0]) * (chunk->map_used - i));
443 chunk->map_used += nr_extra;
446 chunk->map[i + 1] = chunk->map[i] - head;
447 chunk->map[i++] = head;
450 chunk->map[i++] -= tail;
451 chunk->map[i] = tail;
456 * pcpu_alloc_area - allocate area from a pcpu_chunk
457 * @chunk: chunk of interest
458 * @size: wanted size in bytes
459 * @align: wanted align
461 * Try to allocate @size bytes area aligned at @align from @chunk.
462 * Note that this function only allocates the offset. It doesn't
463 * populate or map the area.
465 * @chunk->map must have at least two free slots.
471 * Allocated offset in @chunk on success, -1 if no matching area is
474 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
476 int oslot = pcpu_chunk_slot(chunk);
480 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
481 bool is_last = i + 1 == chunk->map_used;
484 /* extra for alignment requirement */
485 head = ALIGN(off, align) - off;
486 BUG_ON(i == 0 && head != 0);
488 if (chunk->map[i] < 0)
490 if (chunk->map[i] < head + size) {
491 max_contig = max(chunk->map[i], max_contig);
496 * If head is small or the previous block is free,
497 * merge'em. Note that 'small' is defined as smaller
498 * than sizeof(int), which is very small but isn't too
499 * uncommon for percpu allocations.
501 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
502 if (chunk->map[i - 1] > 0)
503 chunk->map[i - 1] += head;
505 chunk->map[i - 1] -= head;
506 chunk->free_size -= head;
508 chunk->map[i] -= head;
513 /* if tail is small, just keep it around */
514 tail = chunk->map[i] - head - size;
515 if (tail < sizeof(int))
518 /* split if warranted */
520 pcpu_split_block(chunk, i, head, tail);
524 max_contig = max(chunk->map[i - 1], max_contig);
527 max_contig = max(chunk->map[i + 1], max_contig);
530 /* update hint and mark allocated */
532 chunk->contig_hint = max_contig; /* fully scanned */
534 chunk->contig_hint = max(chunk->contig_hint,
537 chunk->free_size -= chunk->map[i];
538 chunk->map[i] = -chunk->map[i];
540 pcpu_chunk_relocate(chunk, oslot);
544 chunk->contig_hint = max_contig; /* fully scanned */
545 pcpu_chunk_relocate(chunk, oslot);
547 /* tell the upper layer that this chunk has no matching area */
552 * pcpu_free_area - free area to a pcpu_chunk
553 * @chunk: chunk of interest
554 * @freeme: offset of area to free
556 * Free area starting from @freeme to @chunk. Note that this function
557 * only modifies the allocation map. It doesn't depopulate or unmap
563 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
565 int oslot = pcpu_chunk_slot(chunk);
568 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
571 BUG_ON(off != freeme);
572 BUG_ON(chunk->map[i] > 0);
574 chunk->map[i] = -chunk->map[i];
575 chunk->free_size += chunk->map[i];
577 /* merge with previous? */
578 if (i > 0 && chunk->map[i - 1] >= 0) {
579 chunk->map[i - 1] += chunk->map[i];
581 memmove(&chunk->map[i], &chunk->map[i + 1],
582 (chunk->map_used - i) * sizeof(chunk->map[0]));
585 /* merge with next? */
586 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
587 chunk->map[i] += chunk->map[i + 1];
589 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
590 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
593 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
594 pcpu_chunk_relocate(chunk, oslot);
598 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
599 * @chunk: chunk of interest
600 * @bitmapp: output parameter for bitmap
601 * @may_alloc: may allocate the array
603 * Returns pointer to array of pointers to struct page and bitmap,
604 * both of which can be indexed with pcpu_page_idx(). The returned
605 * array is cleared to zero and *@bitmapp is copied from
606 * @chunk->populated. Note that there is only one array and bitmap
607 * and access exclusion is the caller's responsibility.
610 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
611 * Otherwise, don't care.
614 * Pointer to temp pages array on success, NULL on failure.
616 static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
617 unsigned long **bitmapp,
620 static struct page **pages;
621 static unsigned long *bitmap;
622 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
623 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
624 sizeof(unsigned long);
626 if (!pages || !bitmap) {
627 if (may_alloc && !pages)
628 pages = pcpu_mem_alloc(pages_size);
629 if (may_alloc && !bitmap)
630 bitmap = pcpu_mem_alloc(bitmap_size);
631 if (!pages || !bitmap)
635 memset(pages, 0, pages_size);
636 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
643 * pcpu_free_pages - free pages which were allocated for @chunk
644 * @chunk: chunk pages were allocated for
645 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
646 * @populated: populated bitmap
647 * @page_start: page index of the first page to be freed
648 * @page_end: page index of the last page to be freed + 1
650 * Free pages [@page_start and @page_end) in @pages for all units.
651 * The pages were allocated for @chunk.
653 static void pcpu_free_pages(struct pcpu_chunk *chunk,
654 struct page **pages, unsigned long *populated,
655 int page_start, int page_end)
660 for_each_possible_cpu(cpu) {
661 for (i = page_start; i < page_end; i++) {
662 struct page *page = pages[pcpu_page_idx(cpu, i)];
671 * pcpu_alloc_pages - allocates pages for @chunk
672 * @chunk: target chunk
673 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
674 * @populated: populated bitmap
675 * @page_start: page index of the first page to be allocated
676 * @page_end: page index of the last page to be allocated + 1
678 * Allocate pages [@page_start,@page_end) into @pages for all units.
679 * The allocation is for @chunk. Percpu core doesn't care about the
680 * content of @pages and will pass it verbatim to pcpu_map_pages().
682 static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
683 struct page **pages, unsigned long *populated,
684 int page_start, int page_end)
686 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
690 for_each_possible_cpu(cpu) {
691 for (i = page_start; i < page_end; i++) {
692 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
694 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
696 pcpu_free_pages(chunk, pages, populated,
697 page_start, page_end);
706 * pcpu_pre_unmap_flush - flush cache prior to unmapping
707 * @chunk: chunk the regions to be flushed belongs to
708 * @page_start: page index of the first page to be flushed
709 * @page_end: page index of the last page to be flushed + 1
711 * Pages in [@page_start,@page_end) of @chunk are about to be
712 * unmapped. Flush cache. As each flushing trial can be very
713 * expensive, issue flush on the whole region at once rather than
714 * doing it for each cpu. This could be an overkill but is more
717 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
718 int page_start, int page_end)
721 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
722 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
725 static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
727 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
731 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
732 * @chunk: chunk of interest
733 * @pages: pages array which can be used to pass information to free
734 * @populated: populated bitmap
735 * @page_start: page index of the first page to unmap
736 * @page_end: page index of the last page to unmap + 1
738 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
739 * Corresponding elements in @pages were cleared by the caller and can
740 * be used to carry information to pcpu_free_pages() which will be
741 * called after all unmaps are finished. The caller should call
742 * proper pre/post flush functions.
744 static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
745 struct page **pages, unsigned long *populated,
746 int page_start, int page_end)
751 for_each_possible_cpu(cpu) {
752 for (i = page_start; i < page_end; i++) {
755 page = pcpu_chunk_page(chunk, cpu, i);
757 pages[pcpu_page_idx(cpu, i)] = page;
759 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
760 page_end - page_start);
763 for (i = page_start; i < page_end; i++)
764 __clear_bit(i, populated);
768 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
769 * @chunk: pcpu_chunk the regions to be flushed belong to
770 * @page_start: page index of the first page to be flushed
771 * @page_end: page index of the last page to be flushed + 1
773 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
774 * TLB for the regions. This can be skipped if the area is to be
775 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
777 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
778 * for the whole region.
780 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
781 int page_start, int page_end)
783 flush_tlb_kernel_range(
784 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
785 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
788 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
791 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
796 * pcpu_map_pages - map pages into a pcpu_chunk
797 * @chunk: chunk of interest
798 * @pages: pages array containing pages to be mapped
799 * @populated: populated bitmap
800 * @page_start: page index of the first page to map
801 * @page_end: page index of the last page to map + 1
803 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
804 * caller is responsible for calling pcpu_post_map_flush() after all
805 * mappings are complete.
807 * This function is responsible for setting corresponding bits in
808 * @chunk->populated bitmap and whatever is necessary for reverse
809 * lookup (addr -> chunk).
811 static int pcpu_map_pages(struct pcpu_chunk *chunk,
812 struct page **pages, unsigned long *populated,
813 int page_start, int page_end)
815 unsigned int cpu, tcpu;
818 for_each_possible_cpu(cpu) {
819 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
820 &pages[pcpu_page_idx(cpu, page_start)],
821 page_end - page_start);
826 /* mapping successful, link chunk and mark populated */
827 for (i = page_start; i < page_end; i++) {
828 for_each_possible_cpu(cpu)
829 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
831 __set_bit(i, populated);
837 for_each_possible_cpu(tcpu) {
840 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
841 page_end - page_start);
847 * pcpu_post_map_flush - flush cache after mapping
848 * @chunk: pcpu_chunk the regions to be flushed belong to
849 * @page_start: page index of the first page to be flushed
850 * @page_end: page index of the last page to be flushed + 1
852 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
855 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
856 * for the whole region.
858 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
859 int page_start, int page_end)
862 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
863 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
867 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
868 * @chunk: chunk to depopulate
869 * @off: offset to the area to depopulate
870 * @size: size of the area to depopulate in bytes
871 * @flush: whether to flush cache and tlb or not
873 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
874 * from @chunk. If @flush is true, vcache is flushed before unmapping
880 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
882 int page_start = PFN_DOWN(off);
883 int page_end = PFN_UP(off + size);
885 unsigned long *populated;
888 /* quick path, check whether it's empty already */
890 pcpu_next_unpop(chunk, &rs, &re, page_end);
891 if (rs == page_start && re == page_end)
894 /* immutable chunks can't be depopulated */
895 WARN_ON(chunk->immutable);
898 * If control reaches here, there must have been at least one
899 * successful population attempt so the temp pages array must
902 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
906 pcpu_pre_unmap_flush(chunk, page_start, page_end);
908 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
909 pcpu_unmap_pages(chunk, pages, populated, rs, re);
911 /* no need to flush tlb, vmalloc will handle it lazily */
913 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
914 pcpu_free_pages(chunk, pages, populated, rs, re);
916 /* commit new bitmap */
917 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
921 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
922 * @chunk: chunk of interest
923 * @off: offset to the area to populate
924 * @size: size of the area to populate in bytes
926 * For each cpu, populate and map pages [@page_start,@page_end) into
927 * @chunk. The area is cleared on return.
930 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
932 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
934 int page_start = PFN_DOWN(off);
935 int page_end = PFN_UP(off + size);
936 int free_end = page_start, unmap_end = page_start;
938 unsigned long *populated;
942 /* quick path, check whether all pages are already there */
944 pcpu_next_pop(chunk, &rs, &re, page_end);
945 if (rs == page_start && re == page_end)
948 /* need to allocate and map pages, this chunk can't be immutable */
949 WARN_ON(chunk->immutable);
951 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
956 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
957 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
963 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
964 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
969 pcpu_post_map_flush(chunk, page_start, page_end);
971 /* commit new bitmap */
972 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
974 for_each_possible_cpu(cpu)
975 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
979 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
980 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
981 pcpu_unmap_pages(chunk, pages, populated, rs, re);
982 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
984 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
985 pcpu_free_pages(chunk, pages, populated, rs, re);
989 static void free_pcpu_chunk(struct pcpu_chunk *chunk)
994 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
995 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
999 static struct pcpu_chunk *alloc_pcpu_chunk(void)
1001 struct pcpu_chunk *chunk;
1003 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1007 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
1008 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1009 chunk->map[chunk->map_used++] = pcpu_unit_size;
1011 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1012 pcpu_nr_groups, pcpu_atom_size,
1015 free_pcpu_chunk(chunk);
1019 INIT_LIST_HEAD(&chunk->list);
1020 chunk->free_size = pcpu_unit_size;
1021 chunk->contig_hint = pcpu_unit_size;
1022 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
1028 * pcpu_alloc - the percpu allocator
1029 * @size: size of area to allocate in bytes
1030 * @align: alignment of area (max PAGE_SIZE)
1031 * @reserved: allocate from the reserved chunk if available
1033 * Allocate percpu area of @size bytes aligned at @align.
1036 * Does GFP_KERNEL allocation.
1039 * Percpu pointer to the allocated area on success, NULL on failure.
1041 static void *pcpu_alloc(size_t size, size_t align, bool reserved)
1043 static int warn_limit = 10;
1044 struct pcpu_chunk *chunk;
1048 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
1049 WARN(true, "illegal size (%zu) or align (%zu) for "
1050 "percpu allocation\n", size, align);
1054 mutex_lock(&pcpu_alloc_mutex);
1055 spin_lock_irq(&pcpu_lock);
1057 /* serve reserved allocations from the reserved chunk if available */
1058 if (reserved && pcpu_reserved_chunk) {
1059 chunk = pcpu_reserved_chunk;
1060 if (size > chunk->contig_hint ||
1061 pcpu_extend_area_map(chunk) < 0) {
1062 err = "failed to extend area map of reserved chunk";
1065 off = pcpu_alloc_area(chunk, size, align);
1068 err = "alloc from reserved chunk failed";
1073 /* search through normal chunks */
1074 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1075 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1076 if (size > chunk->contig_hint)
1079 switch (pcpu_extend_area_map(chunk)) {
1083 goto restart; /* pcpu_lock dropped, restart */
1085 err = "failed to extend area map";
1089 off = pcpu_alloc_area(chunk, size, align);
1095 /* hmmm... no space left, create a new chunk */
1096 spin_unlock_irq(&pcpu_lock);
1098 chunk = alloc_pcpu_chunk();
1100 err = "failed to allocate new chunk";
1101 goto fail_unlock_mutex;
1104 spin_lock_irq(&pcpu_lock);
1105 pcpu_chunk_relocate(chunk, -1);
1109 spin_unlock_irq(&pcpu_lock);
1111 /* populate, map and clear the area */
1112 if (pcpu_populate_chunk(chunk, off, size)) {
1113 spin_lock_irq(&pcpu_lock);
1114 pcpu_free_area(chunk, off);
1115 err = "failed to populate";
1119 mutex_unlock(&pcpu_alloc_mutex);
1121 /* return address relative to base address */
1122 return __addr_to_pcpu_ptr(chunk->base_addr + off);
1125 spin_unlock_irq(&pcpu_lock);
1127 mutex_unlock(&pcpu_alloc_mutex);
1129 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1130 "%s\n", size, align, err);
1133 pr_info("PERCPU: limit reached, disable warning\n");
1139 * __alloc_percpu - allocate dynamic percpu area
1140 * @size: size of area to allocate in bytes
1141 * @align: alignment of area (max PAGE_SIZE)
1143 * Allocate percpu area of @size bytes aligned at @align. Might
1144 * sleep. Might trigger writeouts.
1147 * Does GFP_KERNEL allocation.
1150 * Percpu pointer to the allocated area on success, NULL on failure.
1152 void *__alloc_percpu(size_t size, size_t align)
1154 return pcpu_alloc(size, align, false);
1156 EXPORT_SYMBOL_GPL(__alloc_percpu);
1159 * __alloc_reserved_percpu - allocate reserved percpu area
1160 * @size: size of area to allocate in bytes
1161 * @align: alignment of area (max PAGE_SIZE)
1163 * Allocate percpu area of @size bytes aligned at @align from reserved
1164 * percpu area if arch has set it up; otherwise, allocation is served
1165 * from the same dynamic area. Might sleep. Might trigger writeouts.
1168 * Does GFP_KERNEL allocation.
1171 * Percpu pointer to the allocated area on success, NULL on failure.
1173 void *__alloc_reserved_percpu(size_t size, size_t align)
1175 return pcpu_alloc(size, align, true);
1179 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1182 * Reclaim all fully free chunks except for the first one.
1185 * workqueue context.
1187 static void pcpu_reclaim(struct work_struct *work)
1190 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1191 struct pcpu_chunk *chunk, *next;
1193 mutex_lock(&pcpu_alloc_mutex);
1194 spin_lock_irq(&pcpu_lock);
1196 list_for_each_entry_safe(chunk, next, head, list) {
1197 WARN_ON(chunk->immutable);
1199 /* spare the first one */
1200 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1203 list_move(&chunk->list, &todo);
1206 spin_unlock_irq(&pcpu_lock);
1208 list_for_each_entry_safe(chunk, next, &todo, list) {
1209 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
1210 free_pcpu_chunk(chunk);
1213 mutex_unlock(&pcpu_alloc_mutex);
1217 * free_percpu - free percpu area
1218 * @ptr: pointer to area to free
1220 * Free percpu area @ptr.
1223 * Can be called from atomic context.
1225 void free_percpu(void *ptr)
1227 void *addr = __pcpu_ptr_to_addr(ptr);
1228 struct pcpu_chunk *chunk;
1229 unsigned long flags;
1235 spin_lock_irqsave(&pcpu_lock, flags);
1237 chunk = pcpu_chunk_addr_search(addr);
1238 off = addr - chunk->base_addr;
1240 pcpu_free_area(chunk, off);
1242 /* if there are more than one fully free chunks, wake up grim reaper */
1243 if (chunk->free_size == pcpu_unit_size) {
1244 struct pcpu_chunk *pos;
1246 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1248 schedule_work(&pcpu_reclaim_work);
1253 spin_unlock_irqrestore(&pcpu_lock, flags);
1255 EXPORT_SYMBOL_GPL(free_percpu);
1257 static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1258 size_t reserved_size,
1263 size_sum = PFN_ALIGN(static_size + reserved_size +
1264 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1265 if (*dyn_sizep != 0)
1266 *dyn_sizep = size_sum - static_size - reserved_size;
1272 * pcpu_alloc_alloc_info - allocate percpu allocation info
1273 * @nr_groups: the number of groups
1274 * @nr_units: the number of units
1276 * Allocate ai which is large enough for @nr_groups groups containing
1277 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1278 * cpu_map array which is long enough for @nr_units and filled with
1279 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1280 * pointer of other groups.
1283 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1286 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1289 struct pcpu_alloc_info *ai;
1290 size_t base_size, ai_size;
1294 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1295 __alignof__(ai->groups[0].cpu_map[0]));
1296 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1298 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1304 ai->groups[0].cpu_map = ptr;
1306 for (unit = 0; unit < nr_units; unit++)
1307 ai->groups[0].cpu_map[unit] = NR_CPUS;
1309 ai->nr_groups = nr_groups;
1310 ai->__ai_size = PFN_ALIGN(ai_size);
1316 * pcpu_free_alloc_info - free percpu allocation info
1317 * @ai: pcpu_alloc_info to free
1319 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1321 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1323 free_bootmem(__pa(ai), ai->__ai_size);
1327 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1328 * @reserved_size: the size of reserved percpu area in bytes
1329 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1330 * @atom_size: allocation atom size
1331 * @cpu_distance_fn: callback to determine distance between cpus, optional
1333 * This function determines grouping of units, their mappings to cpus
1334 * and other parameters considering needed percpu size, allocation
1335 * atom size and distances between CPUs.
1337 * Groups are always mutliples of atom size and CPUs which are of
1338 * LOCAL_DISTANCE both ways are grouped together and share space for
1339 * units in the same group. The returned configuration is guaranteed
1340 * to have CPUs on different nodes on different groups and >=75% usage
1341 * of allocated virtual address space.
1344 * On success, pointer to the new allocation_info is returned. On
1345 * failure, ERR_PTR value is returned.
1347 struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1348 size_t reserved_size, ssize_t dyn_size,
1350 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1352 static int group_map[NR_CPUS] __initdata;
1353 static int group_cnt[NR_CPUS] __initdata;
1354 const size_t static_size = __per_cpu_end - __per_cpu_start;
1355 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
1356 size_t size_sum, min_unit_size, alloc_size;
1357 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1358 int last_allocs, group, unit;
1359 unsigned int cpu, tcpu;
1360 struct pcpu_alloc_info *ai;
1361 unsigned int *cpu_map;
1363 /* this function may be called multiple times */
1364 memset(group_map, 0, sizeof(group_map));
1365 memset(group_cnt, 0, sizeof(group_map));
1368 * Determine min_unit_size, alloc_size and max_upa such that
1369 * alloc_size is multiple of atom_size and is the smallest
1370 * which can accomodate 4k aligned segments which are equal to
1371 * or larger than min_unit_size.
1373 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
1374 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1376 alloc_size = roundup(min_unit_size, atom_size);
1377 upa = alloc_size / min_unit_size;
1378 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1382 /* group cpus according to their proximity */
1383 for_each_possible_cpu(cpu) {
1386 for_each_possible_cpu(tcpu) {
1389 if (group_map[tcpu] == group && cpu_distance_fn &&
1390 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1391 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1393 nr_groups = max(nr_groups, group + 1);
1397 group_map[cpu] = group;
1399 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1403 * Expand unit size until address space usage goes over 75%
1404 * and then as much as possible without using more address
1407 last_allocs = INT_MAX;
1408 for (upa = max_upa; upa; upa--) {
1409 int allocs = 0, wasted = 0;
1411 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1414 for (group = 0; group < nr_groups; group++) {
1415 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1416 allocs += this_allocs;
1417 wasted += this_allocs * upa - group_cnt[group];
1421 * Don't accept if wastage is over 25%. The
1422 * greater-than comparison ensures upa==1 always
1423 * passes the following check.
1425 if (wasted > num_possible_cpus() / 3)
1428 /* and then don't consume more memory */
1429 if (allocs > last_allocs)
1431 last_allocs = allocs;
1436 /* allocate and fill alloc_info */
1437 for (group = 0; group < nr_groups; group++)
1438 nr_units += roundup(group_cnt[group], upa);
1440 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1442 return ERR_PTR(-ENOMEM);
1443 cpu_map = ai->groups[0].cpu_map;
1445 for (group = 0; group < nr_groups; group++) {
1446 ai->groups[group].cpu_map = cpu_map;
1447 cpu_map += roundup(group_cnt[group], upa);
1450 ai->static_size = static_size;
1451 ai->reserved_size = reserved_size;
1452 ai->dyn_size = dyn_size;
1453 ai->unit_size = alloc_size / upa;
1454 ai->atom_size = atom_size;
1455 ai->alloc_size = alloc_size;
1457 for (group = 0, unit = 0; group_cnt[group]; group++) {
1458 struct pcpu_group_info *gi = &ai->groups[group];
1461 * Initialize base_offset as if all groups are located
1462 * back-to-back. The caller should update this to
1463 * reflect actual allocation.
1465 gi->base_offset = unit * ai->unit_size;
1467 for_each_possible_cpu(cpu)
1468 if (group_map[cpu] == group)
1469 gi->cpu_map[gi->nr_units++] = cpu;
1470 gi->nr_units = roundup(gi->nr_units, upa);
1471 unit += gi->nr_units;
1473 BUG_ON(unit != nr_units);
1479 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1481 * @ai: allocation info to dump
1483 * Print out information about @ai using loglevel @lvl.
1485 static void pcpu_dump_alloc_info(const char *lvl,
1486 const struct pcpu_alloc_info *ai)
1488 int group_width = 1, cpu_width = 1, width;
1489 char empty_str[] = "--------";
1490 int alloc = 0, alloc_end = 0;
1492 int upa, apl; /* units per alloc, allocs per line */
1498 v = num_possible_cpus();
1501 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1503 upa = ai->alloc_size / ai->unit_size;
1504 width = upa * (cpu_width + 1) + group_width + 3;
1505 apl = rounddown_pow_of_two(max(60 / width, 1));
1507 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1508 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1509 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1511 for (group = 0; group < ai->nr_groups; group++) {
1512 const struct pcpu_group_info *gi = &ai->groups[group];
1513 int unit = 0, unit_end = 0;
1515 BUG_ON(gi->nr_units % upa);
1516 for (alloc_end += gi->nr_units / upa;
1517 alloc < alloc_end; alloc++) {
1518 if (!(alloc % apl)) {
1520 printk("%spcpu-alloc: ", lvl);
1522 printk("[%0*d] ", group_width, group);
1524 for (unit_end += upa; unit < unit_end; unit++)
1525 if (gi->cpu_map[unit] != NR_CPUS)
1526 printk("%0*d ", cpu_width,
1529 printk("%s ", empty_str);
1536 * pcpu_setup_first_chunk - initialize the first percpu chunk
1537 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1538 * @base_addr: mapped address
1540 * Initialize the first percpu chunk which contains the kernel static
1541 * perpcu area. This function is to be called from arch percpu area
1544 * @ai contains all information necessary to initialize the first
1545 * chunk and prime the dynamic percpu allocator.
1547 * @ai->static_size is the size of static percpu area.
1549 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1550 * reserve after the static area in the first chunk. This reserves
1551 * the first chunk such that it's available only through reserved
1552 * percpu allocation. This is primarily used to serve module percpu
1553 * static areas on architectures where the addressing model has
1554 * limited offset range for symbol relocations to guarantee module
1555 * percpu symbols fall inside the relocatable range.
1557 * @ai->dyn_size determines the number of bytes available for dynamic
1558 * allocation in the first chunk. The area between @ai->static_size +
1559 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1561 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1562 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1565 * @ai->atom_size is the allocation atom size and used as alignment
1568 * @ai->alloc_size is the allocation size and always multiple of
1569 * @ai->atom_size. This is larger than @ai->atom_size if
1570 * @ai->unit_size is larger than @ai->atom_size.
1572 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1573 * percpu areas. Units which should be colocated are put into the
1574 * same group. Dynamic VM areas will be allocated according to these
1575 * groupings. If @ai->nr_groups is zero, a single group containing
1576 * all units is assumed.
1578 * The caller should have mapped the first chunk at @base_addr and
1579 * copied static data to each unit.
1581 * If the first chunk ends up with both reserved and dynamic areas, it
1582 * is served by two chunks - one to serve the core static and reserved
1583 * areas and the other for the dynamic area. They share the same vm
1584 * and page map but uses different area allocation map to stay away
1585 * from each other. The latter chunk is circulated in the chunk slots
1586 * and available for dynamic allocation like any other chunks.
1589 * 0 on success, -errno on failure.
1591 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1594 static char cpus_buf[4096] __initdata;
1595 static int smap[2], dmap[2];
1596 size_t dyn_size = ai->dyn_size;
1597 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1598 struct pcpu_chunk *schunk, *dchunk = NULL;
1599 unsigned long *group_offsets;
1600 size_t *group_sizes;
1601 unsigned long *unit_off;
1606 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1608 #define PCPU_SETUP_BUG_ON(cond) do { \
1609 if (unlikely(cond)) { \
1610 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1611 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1612 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1618 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1619 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1620 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1621 PCPU_SETUP_BUG_ON(!ai->static_size);
1622 PCPU_SETUP_BUG_ON(!base_addr);
1623 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1624 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1625 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1627 /* process group information and build config tables accordingly */
1628 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1629 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1630 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1631 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1633 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1634 unit_map[cpu] = UINT_MAX;
1635 pcpu_first_unit_cpu = NR_CPUS;
1637 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1638 const struct pcpu_group_info *gi = &ai->groups[group];
1640 group_offsets[group] = gi->base_offset;
1641 group_sizes[group] = gi->nr_units * ai->unit_size;
1643 for (i = 0; i < gi->nr_units; i++) {
1644 cpu = gi->cpu_map[i];
1648 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1649 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1650 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1652 unit_map[cpu] = unit + i;
1653 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1655 if (pcpu_first_unit_cpu == NR_CPUS)
1656 pcpu_first_unit_cpu = cpu;
1659 pcpu_last_unit_cpu = cpu;
1660 pcpu_nr_units = unit;
1662 for_each_possible_cpu(cpu)
1663 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1665 /* we're done parsing the input, undefine BUG macro and dump config */
1666 #undef PCPU_SETUP_BUG_ON
1667 pcpu_dump_alloc_info(KERN_INFO, ai);
1669 pcpu_nr_groups = ai->nr_groups;
1670 pcpu_group_offsets = group_offsets;
1671 pcpu_group_sizes = group_sizes;
1672 pcpu_unit_map = unit_map;
1673 pcpu_unit_offsets = unit_off;
1675 /* determine basic parameters */
1676 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1677 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1678 pcpu_atom_size = ai->atom_size;
1679 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1680 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1683 * Allocate chunk slots. The additional last slot is for
1686 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1687 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1688 for (i = 0; i < pcpu_nr_slots; i++)
1689 INIT_LIST_HEAD(&pcpu_slot[i]);
1692 * Initialize static chunk. If reserved_size is zero, the
1693 * static chunk covers static area + dynamic allocation area
1694 * in the first chunk. If reserved_size is not zero, it
1695 * covers static area + reserved area (mostly used for module
1696 * static percpu allocation).
1698 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1699 INIT_LIST_HEAD(&schunk->list);
1700 schunk->base_addr = base_addr;
1702 schunk->map_alloc = ARRAY_SIZE(smap);
1703 schunk->immutable = true;
1704 bitmap_fill(schunk->populated, pcpu_unit_pages);
1706 if (ai->reserved_size) {
1707 schunk->free_size = ai->reserved_size;
1708 pcpu_reserved_chunk = schunk;
1709 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1711 schunk->free_size = dyn_size;
1712 dyn_size = 0; /* dynamic area covered */
1714 schunk->contig_hint = schunk->free_size;
1716 schunk->map[schunk->map_used++] = -ai->static_size;
1717 if (schunk->free_size)
1718 schunk->map[schunk->map_used++] = schunk->free_size;
1720 /* init dynamic chunk if necessary */
1722 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1723 INIT_LIST_HEAD(&dchunk->list);
1724 dchunk->base_addr = base_addr;
1726 dchunk->map_alloc = ARRAY_SIZE(dmap);
1727 dchunk->immutable = true;
1728 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1730 dchunk->contig_hint = dchunk->free_size = dyn_size;
1731 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1732 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1735 /* link the first chunk in */
1736 pcpu_first_chunk = dchunk ?: schunk;
1737 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1740 pcpu_base_addr = base_addr;
1744 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1745 [PCPU_FC_AUTO] = "auto",
1746 [PCPU_FC_EMBED] = "embed",
1747 [PCPU_FC_PAGE] = "page",
1750 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1752 static int __init percpu_alloc_setup(char *str)
1756 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1757 else if (!strcmp(str, "embed"))
1758 pcpu_chosen_fc = PCPU_FC_EMBED;
1760 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1761 else if (!strcmp(str, "page"))
1762 pcpu_chosen_fc = PCPU_FC_PAGE;
1765 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1769 early_param("percpu_alloc", percpu_alloc_setup);
1771 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1772 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1774 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1775 * @reserved_size: the size of reserved percpu area in bytes
1776 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1777 * @atom_size: allocation atom size
1778 * @cpu_distance_fn: callback to determine distance between cpus, optional
1779 * @alloc_fn: function to allocate percpu page
1780 * @free_fn: funtion to free percpu page
1782 * This is a helper to ease setting up embedded first percpu chunk and
1783 * can be called where pcpu_setup_first_chunk() is expected.
1785 * If this function is used to setup the first chunk, it is allocated
1786 * by calling @alloc_fn and used as-is without being mapped into
1787 * vmalloc area. Allocations are always whole multiples of @atom_size
1788 * aligned to @atom_size.
1790 * This enables the first chunk to piggy back on the linear physical
1791 * mapping which often uses larger page size. Please note that this
1792 * can result in very sparse cpu->unit mapping on NUMA machines thus
1793 * requiring large vmalloc address space. Don't use this allocator if
1794 * vmalloc space is not orders of magnitude larger than distances
1795 * between node memory addresses (ie. 32bit NUMA machines).
1797 * When @dyn_size is positive, dynamic area might be larger than
1798 * specified to fill page alignment. When @dyn_size is auto,
1799 * @dyn_size is just big enough to fill page alignment after static
1800 * and reserved areas.
1802 * If the needed size is smaller than the minimum or specified unit
1803 * size, the leftover is returned using @free_fn.
1806 * 0 on success, -errno on failure.
1808 int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1810 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1811 pcpu_fc_alloc_fn_t alloc_fn,
1812 pcpu_fc_free_fn_t free_fn)
1814 void *base = (void *)ULONG_MAX;
1815 void **areas = NULL;
1816 struct pcpu_alloc_info *ai;
1817 size_t size_sum, areas_size, max_distance;
1820 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1825 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1826 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1828 areas = alloc_bootmem_nopanic(areas_size);
1834 /* allocate, copy and determine base address */
1835 for (group = 0; group < ai->nr_groups; group++) {
1836 struct pcpu_group_info *gi = &ai->groups[group];
1837 unsigned int cpu = NR_CPUS;
1840 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1841 cpu = gi->cpu_map[i];
1842 BUG_ON(cpu == NR_CPUS);
1844 /* allocate space for the whole group */
1845 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1848 goto out_free_areas;
1852 base = min(ptr, base);
1854 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1855 if (gi->cpu_map[i] == NR_CPUS) {
1856 /* unused unit, free whole */
1857 free_fn(ptr, ai->unit_size);
1860 /* copy and return the unused part */
1861 memcpy(ptr, __per_cpu_load, ai->static_size);
1862 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1866 /* base address is now known, determine group base offsets */
1868 for (group = 0; group < ai->nr_groups; group++) {
1869 ai->groups[group].base_offset = areas[group] - base;
1870 max_distance = max_t(size_t, max_distance,
1871 ai->groups[group].base_offset);
1873 max_distance += ai->unit_size;
1875 /* warn if maximum distance is further than 75% of vmalloc space */
1876 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1877 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1879 max_distance, VMALLOC_END - VMALLOC_START);
1880 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1881 /* and fail if we have fallback */
1887 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1888 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1889 ai->dyn_size, ai->unit_size);
1891 rc = pcpu_setup_first_chunk(ai, base);
1895 for (group = 0; group < ai->nr_groups; group++)
1896 free_fn(areas[group],
1897 ai->groups[group].nr_units * ai->unit_size);
1899 pcpu_free_alloc_info(ai);
1901 free_bootmem(__pa(areas), areas_size);
1904 #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1905 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
1907 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1909 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1910 * @reserved_size: the size of reserved percpu area in bytes
1911 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1912 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1913 * @populate_pte_fn: function to populate pte
1915 * This is a helper to ease setting up page-remapped first percpu
1916 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1918 * This is the basic allocator. Static percpu area is allocated
1919 * page-by-page into vmalloc area.
1922 * 0 on success, -errno on failure.
1924 int __init pcpu_page_first_chunk(size_t reserved_size,
1925 pcpu_fc_alloc_fn_t alloc_fn,
1926 pcpu_fc_free_fn_t free_fn,
1927 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1929 static struct vm_struct vm;
1930 struct pcpu_alloc_info *ai;
1934 struct page **pages;
1937 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1939 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1942 BUG_ON(ai->nr_groups != 1);
1943 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1945 unit_pages = ai->unit_size >> PAGE_SHIFT;
1947 /* unaligned allocations can't be freed, round up to page size */
1948 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1950 pages = alloc_bootmem(pages_size);
1952 /* allocate pages */
1954 for (unit = 0; unit < num_possible_cpus(); unit++)
1955 for (i = 0; i < unit_pages; i++) {
1956 unsigned int cpu = ai->groups[0].cpu_map[unit];
1959 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1961 pr_warning("PERCPU: failed to allocate %s page "
1962 "for cpu%u\n", psize_str, cpu);
1965 pages[j++] = virt_to_page(ptr);
1968 /* allocate vm area, map the pages and copy static data */
1969 vm.flags = VM_ALLOC;
1970 vm.size = num_possible_cpus() * ai->unit_size;
1971 vm_area_register_early(&vm, PAGE_SIZE);
1973 for (unit = 0; unit < num_possible_cpus(); unit++) {
1974 unsigned long unit_addr =
1975 (unsigned long)vm.addr + unit * ai->unit_size;
1977 for (i = 0; i < unit_pages; i++)
1978 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1980 /* pte already populated, the following shouldn't fail */
1981 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1984 panic("failed to map percpu area, err=%d\n", rc);
1987 * FIXME: Archs with virtual cache should flush local
1988 * cache for the linear mapping here - something
1989 * equivalent to flush_cache_vmap() on the local cpu.
1990 * flush_cache_vmap() can't be used as most supporting
1991 * data structures are not set up yet.
1994 /* copy static data */
1995 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1998 /* we're ready, commit */
1999 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
2000 unit_pages, psize_str, vm.addr, ai->static_size,
2001 ai->reserved_size, ai->dyn_size);
2003 rc = pcpu_setup_first_chunk(ai, vm.addr);
2008 free_fn(page_address(pages[j]), PAGE_SIZE);
2011 free_bootmem(__pa(pages), pages_size);
2012 pcpu_free_alloc_info(ai);
2015 #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
2018 * Generic percpu area setup.
2020 * The embedding helper is used because its behavior closely resembles
2021 * the original non-dynamic generic percpu area setup. This is
2022 * important because many archs have addressing restrictions and might
2023 * fail if the percpu area is located far away from the previous
2024 * location. As an added bonus, in non-NUMA cases, embedding is
2025 * generally a good idea TLB-wise because percpu area can piggy back
2026 * on the physical linear memory mapping which uses large page
2027 * mappings on applicable archs.
2029 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2030 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2031 EXPORT_SYMBOL(__per_cpu_offset);
2033 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2036 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2039 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2041 free_bootmem(__pa(ptr), size);
2044 void __init setup_per_cpu_areas(void)
2046 unsigned long delta;
2051 * Always reserve area for module percpu variables. That's
2052 * what the legacy allocator did.
2054 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2055 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2056 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2058 panic("Failed to initialized percpu areas.");
2060 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2061 for_each_possible_cpu(cpu)
2062 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2064 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */