2 * mm/percpu.c - percpu memory allocator
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 * This file is released under the GPLv2.
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
46 * To use this allocator, arch code should do the followings.
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
56 #include <linux/bitmap.h>
57 #include <linux/bootmem.h>
58 #include <linux/err.h>
59 #include <linux/list.h>
60 #include <linux/log2.h>
62 #include <linux/module.h>
63 #include <linux/mutex.h>
64 #include <linux/percpu.h>
65 #include <linux/pfn.h>
66 #include <linux/slab.h>
67 #include <linux/spinlock.h>
68 #include <linux/vmalloc.h>
69 #include <linux/workqueue.h>
70 #include <linux/kmemleak.h>
72 #include <asm/cacheflush.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
77 #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78 #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
81 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82 #ifndef __addr_to_pcpu_ptr
83 #define __addr_to_pcpu_ptr(addr) \
84 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
88 #ifndef __pcpu_ptr_to_addr
89 #define __pcpu_ptr_to_addr(ptr) \
90 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
94 #else /* CONFIG_SMP */
95 /* on UP, it's always identity mapped */
96 #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
97 #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
98 #endif /* CONFIG_SMP */
101 struct list_head list; /* linked to pcpu_slot lists */
102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
104 void *base_addr; /* base address of this chunk */
105 int map_used; /* # of map entries used */
106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
108 void *data; /* chunk data */
109 bool immutable; /* no [de]population allowed */
110 unsigned long populated[]; /* populated bitmap */
113 static int pcpu_unit_pages __read_mostly;
114 static int pcpu_unit_size __read_mostly;
115 static int pcpu_nr_units __read_mostly;
116 static int pcpu_atom_size __read_mostly;
117 static int pcpu_nr_slots __read_mostly;
118 static size_t pcpu_chunk_struct_size __read_mostly;
120 /* cpus with the lowest and highest unit addresses */
121 static unsigned int pcpu_low_unit_cpu __read_mostly;
122 static unsigned int pcpu_high_unit_cpu __read_mostly;
124 /* the address of the first chunk which starts with the kernel static area */
125 void *pcpu_base_addr __read_mostly;
126 EXPORT_SYMBOL_GPL(pcpu_base_addr);
128 static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
129 const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
131 /* group information, used for vm allocation */
132 static int pcpu_nr_groups __read_mostly;
133 static const unsigned long *pcpu_group_offsets __read_mostly;
134 static const size_t *pcpu_group_sizes __read_mostly;
137 * The first chunk which always exists. Note that unlike other
138 * chunks, this one can be allocated and mapped in several different
139 * ways and thus often doesn't live in the vmalloc area.
141 static struct pcpu_chunk *pcpu_first_chunk;
144 * Optional reserved chunk. This chunk reserves part of the first
145 * chunk and serves it for reserved allocations. The amount of
146 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
147 * area doesn't exist, the following variables contain NULL and 0
150 static struct pcpu_chunk *pcpu_reserved_chunk;
151 static int pcpu_reserved_chunk_limit;
154 * Synchronization rules.
156 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
157 * protects allocation/reclaim paths, chunks, populated bitmap and
158 * vmalloc mapping. The latter is a spinlock and protects the index
159 * data structures - chunk slots, chunks and area maps in chunks.
161 * During allocation, pcpu_alloc_mutex is kept locked all the time and
162 * pcpu_lock is grabbed and released as necessary. All actual memory
163 * allocations are done using GFP_KERNEL with pcpu_lock released. In
164 * general, percpu memory can't be allocated with irq off but
165 * irqsave/restore are still used in alloc path so that it can be used
166 * from early init path - sched_init() specifically.
168 * Free path accesses and alters only the index data structures, so it
169 * can be safely called from atomic context. When memory needs to be
170 * returned to the system, free path schedules reclaim_work which
171 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
172 * reclaimed, release both locks and frees the chunks. Note that it's
173 * necessary to grab both locks to remove a chunk from circulation as
174 * allocation path might be referencing the chunk with only
175 * pcpu_alloc_mutex locked.
177 static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
178 static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
180 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182 /* reclaim work to release fully free chunks, scheduled from free path */
183 static void pcpu_reclaim(struct work_struct *work);
184 static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186 static bool pcpu_addr_in_first_chunk(void *addr)
188 void *first_start = pcpu_first_chunk->base_addr;
190 return addr >= first_start && addr < first_start + pcpu_unit_size;
193 static bool pcpu_addr_in_reserved_chunk(void *addr)
195 void *first_start = pcpu_first_chunk->base_addr;
197 return addr >= first_start &&
198 addr < first_start + pcpu_reserved_chunk_limit;
201 static int __pcpu_size_to_slot(int size)
203 int highbit = fls(size); /* size is in bytes */
204 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
207 static int pcpu_size_to_slot(int size)
209 if (size == pcpu_unit_size)
210 return pcpu_nr_slots - 1;
211 return __pcpu_size_to_slot(size);
214 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
219 return pcpu_size_to_slot(chunk->free_size);
222 /* set the pointer to a chunk in a page struct */
223 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225 page->index = (unsigned long)pcpu;
228 /* obtain pointer to a chunk from a page struct */
229 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231 return (struct pcpu_chunk *)page->index;
234 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
236 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
239 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
240 unsigned int cpu, int page_idx)
242 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
243 (page_idx << PAGE_SHIFT);
246 static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
247 int *rs, int *re, int end)
249 *rs = find_next_zero_bit(chunk->populated, end, *rs);
250 *re = find_next_bit(chunk->populated, end, *rs + 1);
253 static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
254 int *rs, int *re, int end)
256 *rs = find_next_bit(chunk->populated, end, *rs);
257 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
261 * (Un)populated page region iterators. Iterate over (un)populated
262 * page regions between @start and @end in @chunk. @rs and @re should
263 * be integer variables and will be set to start and end page index of
264 * the current region.
266 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
267 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271 #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
272 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
274 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
277 * pcpu_mem_zalloc - allocate memory
278 * @size: bytes to allocate
280 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
281 * kzalloc() is used; otherwise, vzalloc() is used. The returned
282 * memory is always zeroed.
285 * Does GFP_KERNEL allocation.
288 * Pointer to the allocated area on success, NULL on failure.
290 static void *pcpu_mem_zalloc(size_t size)
292 if (WARN_ON_ONCE(!slab_is_available()))
295 if (size <= PAGE_SIZE)
296 return kzalloc(size, GFP_KERNEL);
298 return vzalloc(size);
302 * pcpu_mem_free - free memory
303 * @ptr: memory to free
304 * @size: size of the area
306 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
308 static void pcpu_mem_free(void *ptr, size_t size)
310 if (size <= PAGE_SIZE)
317 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
318 * @chunk: chunk of interest
319 * @oslot: the previous slot it was on
321 * This function is called after an allocation or free changed @chunk.
322 * New slot according to the changed state is determined and @chunk is
323 * moved to the slot. Note that the reserved chunk is never put on
329 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331 int nslot = pcpu_chunk_slot(chunk);
333 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
335 list_move(&chunk->list, &pcpu_slot[nslot]);
337 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
342 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
343 * @chunk: chunk of interest
345 * Determine whether area map of @chunk needs to be extended to
346 * accommodate a new allocation.
352 * New target map allocation length if extension is necessary, 0
355 static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
359 if (chunk->map_alloc >= chunk->map_used + 2)
362 new_alloc = PCPU_DFL_MAP_ALLOC;
363 while (new_alloc < chunk->map_used + 2)
370 * pcpu_extend_area_map - extend area map of a chunk
371 * @chunk: chunk of interest
372 * @new_alloc: new target allocation length of the area map
374 * Extend area map of @chunk to have @new_alloc entries.
377 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
380 * 0 on success, -errno on failure.
382 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384 int *old = NULL, *new = NULL;
385 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
388 new = pcpu_mem_zalloc(new_size);
392 /* acquire pcpu_lock and switch to new area map */
393 spin_lock_irqsave(&pcpu_lock, flags);
395 if (new_alloc <= chunk->map_alloc)
398 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
401 memcpy(new, old, old_size);
403 chunk->map_alloc = new_alloc;
408 spin_unlock_irqrestore(&pcpu_lock, flags);
411 * pcpu_mem_free() might end up calling vfree() which uses
412 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 pcpu_mem_free(old, old_size);
415 pcpu_mem_free(new, new_size);
421 * pcpu_split_block - split a map block
422 * @chunk: chunk of interest
423 * @i: index of map block to split
424 * @head: head size in bytes (can be 0)
425 * @tail: tail size in bytes (can be 0)
427 * Split the @i'th map block into two or three blocks. If @head is
428 * non-zero, @head bytes block is inserted before block @i moving it
429 * to @i+1 and reducing its size by @head bytes.
431 * If @tail is non-zero, the target block, which can be @i or @i+1
432 * depending on @head, is reduced by @tail bytes and @tail byte block
433 * is inserted after the target block.
435 * @chunk->map must have enough free slots to accommodate the split.
440 static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
443 int nr_extra = !!head + !!tail;
445 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
447 /* insert new subblocks */
448 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
449 sizeof(chunk->map[0]) * (chunk->map_used - i));
450 chunk->map_used += nr_extra;
453 chunk->map[i + 1] = chunk->map[i] - head;
454 chunk->map[i++] = head;
457 chunk->map[i++] -= tail;
458 chunk->map[i] = tail;
463 * pcpu_alloc_area - allocate area from a pcpu_chunk
464 * @chunk: chunk of interest
465 * @size: wanted size in bytes
466 * @align: wanted align
468 * Try to allocate @size bytes area aligned at @align from @chunk.
469 * Note that this function only allocates the offset. It doesn't
470 * populate or map the area.
472 * @chunk->map must have at least two free slots.
478 * Allocated offset in @chunk on success, -1 if no matching area is
481 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
483 int oslot = pcpu_chunk_slot(chunk);
487 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
488 bool is_last = i + 1 == chunk->map_used;
491 /* extra for alignment requirement */
492 head = ALIGN(off, align) - off;
493 BUG_ON(i == 0 && head != 0);
495 if (chunk->map[i] < 0)
497 if (chunk->map[i] < head + size) {
498 max_contig = max(chunk->map[i], max_contig);
503 * If head is small or the previous block is free,
504 * merge'em. Note that 'small' is defined as smaller
505 * than sizeof(int), which is very small but isn't too
506 * uncommon for percpu allocations.
508 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
509 if (chunk->map[i - 1] > 0)
510 chunk->map[i - 1] += head;
512 chunk->map[i - 1] -= head;
513 chunk->free_size -= head;
515 chunk->map[i] -= head;
520 /* if tail is small, just keep it around */
521 tail = chunk->map[i] - head - size;
522 if (tail < sizeof(int))
525 /* split if warranted */
527 pcpu_split_block(chunk, i, head, tail);
531 max_contig = max(chunk->map[i - 1], max_contig);
534 max_contig = max(chunk->map[i + 1], max_contig);
537 /* update hint and mark allocated */
539 chunk->contig_hint = max_contig; /* fully scanned */
541 chunk->contig_hint = max(chunk->contig_hint,
544 chunk->free_size -= chunk->map[i];
545 chunk->map[i] = -chunk->map[i];
547 pcpu_chunk_relocate(chunk, oslot);
551 chunk->contig_hint = max_contig; /* fully scanned */
552 pcpu_chunk_relocate(chunk, oslot);
554 /* tell the upper layer that this chunk has no matching area */
559 * pcpu_free_area - free area to a pcpu_chunk
560 * @chunk: chunk of interest
561 * @freeme: offset of area to free
563 * Free area starting from @freeme to @chunk. Note that this function
564 * only modifies the allocation map. It doesn't depopulate or unmap
570 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
572 int oslot = pcpu_chunk_slot(chunk);
575 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
578 BUG_ON(off != freeme);
579 BUG_ON(chunk->map[i] > 0);
581 chunk->map[i] = -chunk->map[i];
582 chunk->free_size += chunk->map[i];
584 /* merge with previous? */
585 if (i > 0 && chunk->map[i - 1] >= 0) {
586 chunk->map[i - 1] += chunk->map[i];
588 memmove(&chunk->map[i], &chunk->map[i + 1],
589 (chunk->map_used - i) * sizeof(chunk->map[0]));
592 /* merge with next? */
593 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
594 chunk->map[i] += chunk->map[i + 1];
596 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
597 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
600 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
601 pcpu_chunk_relocate(chunk, oslot);
604 static struct pcpu_chunk *pcpu_alloc_chunk(void)
606 struct pcpu_chunk *chunk;
608 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
612 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
613 sizeof(chunk->map[0]));
619 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
620 chunk->map[chunk->map_used++] = pcpu_unit_size;
622 INIT_LIST_HEAD(&chunk->list);
623 chunk->free_size = pcpu_unit_size;
624 chunk->contig_hint = pcpu_unit_size;
629 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
633 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
638 * Chunk management implementation.
640 * To allow different implementations, chunk alloc/free and
641 * [de]population are implemented in a separate file which is pulled
642 * into this file and compiled together. The following functions
643 * should be implemented.
645 * pcpu_populate_chunk - populate the specified range of a chunk
646 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
647 * pcpu_create_chunk - create a new chunk
648 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
649 * pcpu_addr_to_page - translate address to physical address
650 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
652 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654 static struct pcpu_chunk *pcpu_create_chunk(void);
655 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656 static struct page *pcpu_addr_to_page(void *addr);
657 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
659 #ifdef CONFIG_NEED_PER_CPU_KM
660 #include "percpu-km.c"
662 #include "percpu-vm.c"
666 * pcpu_chunk_addr_search - determine chunk containing specified address
667 * @addr: address for which the chunk needs to be determined.
670 * The address of the found chunk.
672 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
674 /* is it in the first chunk? */
675 if (pcpu_addr_in_first_chunk(addr)) {
676 /* is it in the reserved area? */
677 if (pcpu_addr_in_reserved_chunk(addr))
678 return pcpu_reserved_chunk;
679 return pcpu_first_chunk;
683 * The address is relative to unit0 which might be unused and
684 * thus unmapped. Offset the address to the unit space of the
685 * current processor before looking it up in the vmalloc
686 * space. Note that any possible cpu id can be used here, so
687 * there's no need to worry about preemption or cpu hotplug.
689 addr += pcpu_unit_offsets[raw_smp_processor_id()];
690 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
694 * pcpu_alloc - the percpu allocator
695 * @size: size of area to allocate in bytes
696 * @align: alignment of area (max PAGE_SIZE)
697 * @reserved: allocate from the reserved chunk if available
699 * Allocate percpu area of @size bytes aligned at @align.
702 * Does GFP_KERNEL allocation.
705 * Percpu pointer to the allocated area on success, NULL on failure.
707 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
709 static int warn_limit = 10;
710 struct pcpu_chunk *chunk;
712 int slot, off, new_alloc;
716 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
717 WARN(true, "illegal size (%zu) or align (%zu) for "
718 "percpu allocation\n", size, align);
722 mutex_lock(&pcpu_alloc_mutex);
723 spin_lock_irqsave(&pcpu_lock, flags);
725 /* serve reserved allocations from the reserved chunk if available */
726 if (reserved && pcpu_reserved_chunk) {
727 chunk = pcpu_reserved_chunk;
729 if (size > chunk->contig_hint) {
730 err = "alloc from reserved chunk failed";
734 while ((new_alloc = pcpu_need_to_extend(chunk))) {
735 spin_unlock_irqrestore(&pcpu_lock, flags);
736 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
737 err = "failed to extend area map of reserved chunk";
738 goto fail_unlock_mutex;
740 spin_lock_irqsave(&pcpu_lock, flags);
743 off = pcpu_alloc_area(chunk, size, align);
747 err = "alloc from reserved chunk failed";
752 /* search through normal chunks */
753 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
754 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
755 if (size > chunk->contig_hint)
758 new_alloc = pcpu_need_to_extend(chunk);
760 spin_unlock_irqrestore(&pcpu_lock, flags);
761 if (pcpu_extend_area_map(chunk,
763 err = "failed to extend area map";
764 goto fail_unlock_mutex;
766 spin_lock_irqsave(&pcpu_lock, flags);
768 * pcpu_lock has been dropped, need to
769 * restart cpu_slot list walking.
774 off = pcpu_alloc_area(chunk, size, align);
780 /* hmmm... no space left, create a new chunk */
781 spin_unlock_irqrestore(&pcpu_lock, flags);
783 chunk = pcpu_create_chunk();
785 err = "failed to allocate new chunk";
786 goto fail_unlock_mutex;
789 spin_lock_irqsave(&pcpu_lock, flags);
790 pcpu_chunk_relocate(chunk, -1);
794 spin_unlock_irqrestore(&pcpu_lock, flags);
796 /* populate, map and clear the area */
797 if (pcpu_populate_chunk(chunk, off, size)) {
798 spin_lock_irqsave(&pcpu_lock, flags);
799 pcpu_free_area(chunk, off);
800 err = "failed to populate";
804 mutex_unlock(&pcpu_alloc_mutex);
806 /* return address relative to base address */
807 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
808 kmemleak_alloc_percpu(ptr, size);
812 spin_unlock_irqrestore(&pcpu_lock, flags);
814 mutex_unlock(&pcpu_alloc_mutex);
816 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
817 "%s\n", size, align, err);
820 pr_info("PERCPU: limit reached, disable warning\n");
826 * __alloc_percpu - allocate dynamic percpu area
827 * @size: size of area to allocate in bytes
828 * @align: alignment of area (max PAGE_SIZE)
830 * Allocate zero-filled percpu area of @size bytes aligned at @align.
831 * Might sleep. Might trigger writeouts.
834 * Does GFP_KERNEL allocation.
837 * Percpu pointer to the allocated area on success, NULL on failure.
839 void __percpu *__alloc_percpu(size_t size, size_t align)
841 return pcpu_alloc(size, align, false);
843 EXPORT_SYMBOL_GPL(__alloc_percpu);
846 * __alloc_reserved_percpu - allocate reserved percpu area
847 * @size: size of area to allocate in bytes
848 * @align: alignment of area (max PAGE_SIZE)
850 * Allocate zero-filled percpu area of @size bytes aligned at @align
851 * from reserved percpu area if arch has set it up; otherwise,
852 * allocation is served from the same dynamic area. Might sleep.
853 * Might trigger writeouts.
856 * Does GFP_KERNEL allocation.
859 * Percpu pointer to the allocated area on success, NULL on failure.
861 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
863 return pcpu_alloc(size, align, true);
867 * pcpu_reclaim - reclaim fully free chunks, workqueue function
870 * Reclaim all fully free chunks except for the first one.
875 static void pcpu_reclaim(struct work_struct *work)
878 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
879 struct pcpu_chunk *chunk, *next;
881 mutex_lock(&pcpu_alloc_mutex);
882 spin_lock_irq(&pcpu_lock);
884 list_for_each_entry_safe(chunk, next, head, list) {
885 WARN_ON(chunk->immutable);
887 /* spare the first one */
888 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
891 list_move(&chunk->list, &todo);
894 spin_unlock_irq(&pcpu_lock);
896 list_for_each_entry_safe(chunk, next, &todo, list) {
897 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
898 pcpu_destroy_chunk(chunk);
901 mutex_unlock(&pcpu_alloc_mutex);
905 * free_percpu - free percpu area
906 * @ptr: pointer to area to free
908 * Free percpu area @ptr.
911 * Can be called from atomic context.
913 void free_percpu(void __percpu *ptr)
916 struct pcpu_chunk *chunk;
923 kmemleak_free_percpu(ptr);
925 addr = __pcpu_ptr_to_addr(ptr);
927 spin_lock_irqsave(&pcpu_lock, flags);
929 chunk = pcpu_chunk_addr_search(addr);
930 off = addr - chunk->base_addr;
932 pcpu_free_area(chunk, off);
934 /* if there are more than one fully free chunks, wake up grim reaper */
935 if (chunk->free_size == pcpu_unit_size) {
936 struct pcpu_chunk *pos;
938 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
940 schedule_work(&pcpu_reclaim_work);
945 spin_unlock_irqrestore(&pcpu_lock, flags);
947 EXPORT_SYMBOL_GPL(free_percpu);
950 * is_kernel_percpu_address - test whether address is from static percpu area
951 * @addr: address to test
953 * Test whether @addr belongs to in-kernel static percpu area. Module
954 * static percpu areas are not considered. For those, use
955 * is_module_percpu_address().
958 * %true if @addr is from in-kernel static percpu area, %false otherwise.
960 bool is_kernel_percpu_address(unsigned long addr)
963 const size_t static_size = __per_cpu_end - __per_cpu_start;
964 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
967 for_each_possible_cpu(cpu) {
968 void *start = per_cpu_ptr(base, cpu);
970 if ((void *)addr >= start && (void *)addr < start + static_size)
974 /* on UP, can't distinguish from other static vars, always false */
979 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
980 * @addr: the address to be converted to physical address
982 * Given @addr which is dereferenceable address obtained via one of
983 * percpu access macros, this function translates it into its physical
984 * address. The caller is responsible for ensuring @addr stays valid
985 * until this function finishes.
987 * percpu allocator has special setup for the first chunk, which currently
988 * supports either embedding in linear address space or vmalloc mapping,
989 * and, from the second one, the backing allocator (currently either vm or
990 * km) provides translation.
992 * The addr can be tranlated simply without checking if it falls into the
993 * first chunk. But the current code reflects better how percpu allocator
994 * actually works, and the verification can discover both bugs in percpu
995 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
999 * The physical address for @addr.
1001 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1003 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1004 bool in_first_chunk = false;
1005 unsigned long first_low, first_high;
1009 * The following test on unit_low/high isn't strictly
1010 * necessary but will speed up lookups of addresses which
1011 * aren't in the first chunk.
1013 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1014 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1016 if ((unsigned long)addr >= first_low &&
1017 (unsigned long)addr < first_high) {
1018 for_each_possible_cpu(cpu) {
1019 void *start = per_cpu_ptr(base, cpu);
1021 if (addr >= start && addr < start + pcpu_unit_size) {
1022 in_first_chunk = true;
1028 if (in_first_chunk) {
1029 if (!is_vmalloc_addr(addr))
1032 return page_to_phys(vmalloc_to_page(addr)) +
1033 offset_in_page(addr);
1035 return page_to_phys(pcpu_addr_to_page(addr)) +
1036 offset_in_page(addr);
1040 * pcpu_alloc_alloc_info - allocate percpu allocation info
1041 * @nr_groups: the number of groups
1042 * @nr_units: the number of units
1044 * Allocate ai which is large enough for @nr_groups groups containing
1045 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1046 * cpu_map array which is long enough for @nr_units and filled with
1047 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1048 * pointer of other groups.
1051 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1054 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1057 struct pcpu_alloc_info *ai;
1058 size_t base_size, ai_size;
1062 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1063 __alignof__(ai->groups[0].cpu_map[0]));
1064 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1066 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1072 ai->groups[0].cpu_map = ptr;
1074 for (unit = 0; unit < nr_units; unit++)
1075 ai->groups[0].cpu_map[unit] = NR_CPUS;
1077 ai->nr_groups = nr_groups;
1078 ai->__ai_size = PFN_ALIGN(ai_size);
1084 * pcpu_free_alloc_info - free percpu allocation info
1085 * @ai: pcpu_alloc_info to free
1087 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1089 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1091 free_bootmem(__pa(ai), ai->__ai_size);
1095 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1097 * @ai: allocation info to dump
1099 * Print out information about @ai using loglevel @lvl.
1101 static void pcpu_dump_alloc_info(const char *lvl,
1102 const struct pcpu_alloc_info *ai)
1104 int group_width = 1, cpu_width = 1, width;
1105 char empty_str[] = "--------";
1106 int alloc = 0, alloc_end = 0;
1108 int upa, apl; /* units per alloc, allocs per line */
1114 v = num_possible_cpus();
1117 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1119 upa = ai->alloc_size / ai->unit_size;
1120 width = upa * (cpu_width + 1) + group_width + 3;
1121 apl = rounddown_pow_of_two(max(60 / width, 1));
1123 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1124 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1125 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1127 for (group = 0; group < ai->nr_groups; group++) {
1128 const struct pcpu_group_info *gi = &ai->groups[group];
1129 int unit = 0, unit_end = 0;
1131 BUG_ON(gi->nr_units % upa);
1132 for (alloc_end += gi->nr_units / upa;
1133 alloc < alloc_end; alloc++) {
1134 if (!(alloc % apl)) {
1135 printk(KERN_CONT "\n");
1136 printk("%spcpu-alloc: ", lvl);
1138 printk(KERN_CONT "[%0*d] ", group_width, group);
1140 for (unit_end += upa; unit < unit_end; unit++)
1141 if (gi->cpu_map[unit] != NR_CPUS)
1142 printk(KERN_CONT "%0*d ", cpu_width,
1145 printk(KERN_CONT "%s ", empty_str);
1148 printk(KERN_CONT "\n");
1152 * pcpu_setup_first_chunk - initialize the first percpu chunk
1153 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1154 * @base_addr: mapped address
1156 * Initialize the first percpu chunk which contains the kernel static
1157 * perpcu area. This function is to be called from arch percpu area
1160 * @ai contains all information necessary to initialize the first
1161 * chunk and prime the dynamic percpu allocator.
1163 * @ai->static_size is the size of static percpu area.
1165 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1166 * reserve after the static area in the first chunk. This reserves
1167 * the first chunk such that it's available only through reserved
1168 * percpu allocation. This is primarily used to serve module percpu
1169 * static areas on architectures where the addressing model has
1170 * limited offset range for symbol relocations to guarantee module
1171 * percpu symbols fall inside the relocatable range.
1173 * @ai->dyn_size determines the number of bytes available for dynamic
1174 * allocation in the first chunk. The area between @ai->static_size +
1175 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1177 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1178 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1181 * @ai->atom_size is the allocation atom size and used as alignment
1184 * @ai->alloc_size is the allocation size and always multiple of
1185 * @ai->atom_size. This is larger than @ai->atom_size if
1186 * @ai->unit_size is larger than @ai->atom_size.
1188 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1189 * percpu areas. Units which should be colocated are put into the
1190 * same group. Dynamic VM areas will be allocated according to these
1191 * groupings. If @ai->nr_groups is zero, a single group containing
1192 * all units is assumed.
1194 * The caller should have mapped the first chunk at @base_addr and
1195 * copied static data to each unit.
1197 * If the first chunk ends up with both reserved and dynamic areas, it
1198 * is served by two chunks - one to serve the core static and reserved
1199 * areas and the other for the dynamic area. They share the same vm
1200 * and page map but uses different area allocation map to stay away
1201 * from each other. The latter chunk is circulated in the chunk slots
1202 * and available for dynamic allocation like any other chunks.
1205 * 0 on success, -errno on failure.
1207 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1210 static char cpus_buf[4096] __initdata;
1211 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1212 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1213 size_t dyn_size = ai->dyn_size;
1214 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1215 struct pcpu_chunk *schunk, *dchunk = NULL;
1216 unsigned long *group_offsets;
1217 size_t *group_sizes;
1218 unsigned long *unit_off;
1223 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1225 #define PCPU_SETUP_BUG_ON(cond) do { \
1226 if (unlikely(cond)) { \
1227 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1228 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1229 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1235 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1237 PCPU_SETUP_BUG_ON(!ai->static_size);
1238 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
1240 PCPU_SETUP_BUG_ON(!base_addr);
1241 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
1242 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1243 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1244 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1245 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1246 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1248 /* process group information and build config tables accordingly */
1249 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1250 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1251 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1252 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1254 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1255 unit_map[cpu] = UINT_MAX;
1257 pcpu_low_unit_cpu = NR_CPUS;
1258 pcpu_high_unit_cpu = NR_CPUS;
1260 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1261 const struct pcpu_group_info *gi = &ai->groups[group];
1263 group_offsets[group] = gi->base_offset;
1264 group_sizes[group] = gi->nr_units * ai->unit_size;
1266 for (i = 0; i < gi->nr_units; i++) {
1267 cpu = gi->cpu_map[i];
1271 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1272 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1273 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1275 unit_map[cpu] = unit + i;
1276 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1278 /* determine low/high unit_cpu */
1279 if (pcpu_low_unit_cpu == NR_CPUS ||
1280 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1281 pcpu_low_unit_cpu = cpu;
1282 if (pcpu_high_unit_cpu == NR_CPUS ||
1283 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1284 pcpu_high_unit_cpu = cpu;
1287 pcpu_nr_units = unit;
1289 for_each_possible_cpu(cpu)
1290 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1292 /* we're done parsing the input, undefine BUG macro and dump config */
1293 #undef PCPU_SETUP_BUG_ON
1294 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1296 pcpu_nr_groups = ai->nr_groups;
1297 pcpu_group_offsets = group_offsets;
1298 pcpu_group_sizes = group_sizes;
1299 pcpu_unit_map = unit_map;
1300 pcpu_unit_offsets = unit_off;
1302 /* determine basic parameters */
1303 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1304 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1305 pcpu_atom_size = ai->atom_size;
1306 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1307 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1310 * Allocate chunk slots. The additional last slot is for
1313 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1314 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1315 for (i = 0; i < pcpu_nr_slots; i++)
1316 INIT_LIST_HEAD(&pcpu_slot[i]);
1319 * Initialize static chunk. If reserved_size is zero, the
1320 * static chunk covers static area + dynamic allocation area
1321 * in the first chunk. If reserved_size is not zero, it
1322 * covers static area + reserved area (mostly used for module
1323 * static percpu allocation).
1325 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1326 INIT_LIST_HEAD(&schunk->list);
1327 schunk->base_addr = base_addr;
1329 schunk->map_alloc = ARRAY_SIZE(smap);
1330 schunk->immutable = true;
1331 bitmap_fill(schunk->populated, pcpu_unit_pages);
1333 if (ai->reserved_size) {
1334 schunk->free_size = ai->reserved_size;
1335 pcpu_reserved_chunk = schunk;
1336 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1338 schunk->free_size = dyn_size;
1339 dyn_size = 0; /* dynamic area covered */
1341 schunk->contig_hint = schunk->free_size;
1343 schunk->map[schunk->map_used++] = -ai->static_size;
1344 if (schunk->free_size)
1345 schunk->map[schunk->map_used++] = schunk->free_size;
1347 /* init dynamic chunk if necessary */
1349 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1350 INIT_LIST_HEAD(&dchunk->list);
1351 dchunk->base_addr = base_addr;
1353 dchunk->map_alloc = ARRAY_SIZE(dmap);
1354 dchunk->immutable = true;
1355 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1357 dchunk->contig_hint = dchunk->free_size = dyn_size;
1358 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1359 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1362 /* link the first chunk in */
1363 pcpu_first_chunk = dchunk ?: schunk;
1364 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1367 pcpu_base_addr = base_addr;
1373 const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1374 [PCPU_FC_AUTO] = "auto",
1375 [PCPU_FC_EMBED] = "embed",
1376 [PCPU_FC_PAGE] = "page",
1379 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1381 static int __init percpu_alloc_setup(char *str)
1385 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1386 else if (!strcmp(str, "embed"))
1387 pcpu_chosen_fc = PCPU_FC_EMBED;
1389 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1390 else if (!strcmp(str, "page"))
1391 pcpu_chosen_fc = PCPU_FC_PAGE;
1394 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1398 early_param("percpu_alloc", percpu_alloc_setup);
1401 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1402 * Build it if needed by the arch config or the generic setup is going
1405 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1406 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1407 #define BUILD_EMBED_FIRST_CHUNK
1410 /* build pcpu_page_first_chunk() iff needed by the arch config */
1411 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1412 #define BUILD_PAGE_FIRST_CHUNK
1415 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1416 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1418 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1419 * @reserved_size: the size of reserved percpu area in bytes
1420 * @dyn_size: minimum free size for dynamic allocation in bytes
1421 * @atom_size: allocation atom size
1422 * @cpu_distance_fn: callback to determine distance between cpus, optional
1424 * This function determines grouping of units, their mappings to cpus
1425 * and other parameters considering needed percpu size, allocation
1426 * atom size and distances between CPUs.
1428 * Groups are always mutliples of atom size and CPUs which are of
1429 * LOCAL_DISTANCE both ways are grouped together and share space for
1430 * units in the same group. The returned configuration is guaranteed
1431 * to have CPUs on different nodes on different groups and >=75% usage
1432 * of allocated virtual address space.
1435 * On success, pointer to the new allocation_info is returned. On
1436 * failure, ERR_PTR value is returned.
1438 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1439 size_t reserved_size, size_t dyn_size,
1441 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1443 static int group_map[NR_CPUS] __initdata;
1444 static int group_cnt[NR_CPUS] __initdata;
1445 const size_t static_size = __per_cpu_end - __per_cpu_start;
1446 int nr_groups = 1, nr_units = 0;
1447 size_t size_sum, min_unit_size, alloc_size;
1448 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1449 int last_allocs, group, unit;
1450 unsigned int cpu, tcpu;
1451 struct pcpu_alloc_info *ai;
1452 unsigned int *cpu_map;
1454 /* this function may be called multiple times */
1455 memset(group_map, 0, sizeof(group_map));
1456 memset(group_cnt, 0, sizeof(group_cnt));
1458 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1459 size_sum = PFN_ALIGN(static_size + reserved_size +
1460 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1461 dyn_size = size_sum - static_size - reserved_size;
1464 * Determine min_unit_size, alloc_size and max_upa such that
1465 * alloc_size is multiple of atom_size and is the smallest
1466 * which can accommodate 4k aligned segments which are equal to
1467 * or larger than min_unit_size.
1469 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1471 alloc_size = roundup(min_unit_size, atom_size);
1472 upa = alloc_size / min_unit_size;
1473 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1477 /* group cpus according to their proximity */
1478 for_each_possible_cpu(cpu) {
1481 for_each_possible_cpu(tcpu) {
1484 if (group_map[tcpu] == group && cpu_distance_fn &&
1485 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1486 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1488 nr_groups = max(nr_groups, group + 1);
1492 group_map[cpu] = group;
1497 * Expand unit size until address space usage goes over 75%
1498 * and then as much as possible without using more address
1501 last_allocs = INT_MAX;
1502 for (upa = max_upa; upa; upa--) {
1503 int allocs = 0, wasted = 0;
1505 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1508 for (group = 0; group < nr_groups; group++) {
1509 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1510 allocs += this_allocs;
1511 wasted += this_allocs * upa - group_cnt[group];
1515 * Don't accept if wastage is over 1/3. The
1516 * greater-than comparison ensures upa==1 always
1517 * passes the following check.
1519 if (wasted > num_possible_cpus() / 3)
1522 /* and then don't consume more memory */
1523 if (allocs > last_allocs)
1525 last_allocs = allocs;
1530 /* allocate and fill alloc_info */
1531 for (group = 0; group < nr_groups; group++)
1532 nr_units += roundup(group_cnt[group], upa);
1534 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1536 return ERR_PTR(-ENOMEM);
1537 cpu_map = ai->groups[0].cpu_map;
1539 for (group = 0; group < nr_groups; group++) {
1540 ai->groups[group].cpu_map = cpu_map;
1541 cpu_map += roundup(group_cnt[group], upa);
1544 ai->static_size = static_size;
1545 ai->reserved_size = reserved_size;
1546 ai->dyn_size = dyn_size;
1547 ai->unit_size = alloc_size / upa;
1548 ai->atom_size = atom_size;
1549 ai->alloc_size = alloc_size;
1551 for (group = 0, unit = 0; group_cnt[group]; group++) {
1552 struct pcpu_group_info *gi = &ai->groups[group];
1555 * Initialize base_offset as if all groups are located
1556 * back-to-back. The caller should update this to
1557 * reflect actual allocation.
1559 gi->base_offset = unit * ai->unit_size;
1561 for_each_possible_cpu(cpu)
1562 if (group_map[cpu] == group)
1563 gi->cpu_map[gi->nr_units++] = cpu;
1564 gi->nr_units = roundup(gi->nr_units, upa);
1565 unit += gi->nr_units;
1567 BUG_ON(unit != nr_units);
1571 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1573 #if defined(BUILD_EMBED_FIRST_CHUNK)
1575 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1576 * @reserved_size: the size of reserved percpu area in bytes
1577 * @dyn_size: minimum free size for dynamic allocation in bytes
1578 * @atom_size: allocation atom size
1579 * @cpu_distance_fn: callback to determine distance between cpus, optional
1580 * @alloc_fn: function to allocate percpu page
1581 * @free_fn: function to free percpu page
1583 * This is a helper to ease setting up embedded first percpu chunk and
1584 * can be called where pcpu_setup_first_chunk() is expected.
1586 * If this function is used to setup the first chunk, it is allocated
1587 * by calling @alloc_fn and used as-is without being mapped into
1588 * vmalloc area. Allocations are always whole multiples of @atom_size
1589 * aligned to @atom_size.
1591 * This enables the first chunk to piggy back on the linear physical
1592 * mapping which often uses larger page size. Please note that this
1593 * can result in very sparse cpu->unit mapping on NUMA machines thus
1594 * requiring large vmalloc address space. Don't use this allocator if
1595 * vmalloc space is not orders of magnitude larger than distances
1596 * between node memory addresses (ie. 32bit NUMA machines).
1598 * @dyn_size specifies the minimum dynamic area size.
1600 * If the needed size is smaller than the minimum or specified unit
1601 * size, the leftover is returned using @free_fn.
1604 * 0 on success, -errno on failure.
1606 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1608 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1609 pcpu_fc_alloc_fn_t alloc_fn,
1610 pcpu_fc_free_fn_t free_fn)
1612 void *base = (void *)ULONG_MAX;
1613 void **areas = NULL;
1614 struct pcpu_alloc_info *ai;
1615 size_t size_sum, areas_size, max_distance;
1618 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1623 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1624 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1626 areas = alloc_bootmem_nopanic(areas_size);
1632 /* allocate, copy and determine base address */
1633 for (group = 0; group < ai->nr_groups; group++) {
1634 struct pcpu_group_info *gi = &ai->groups[group];
1635 unsigned int cpu = NR_CPUS;
1638 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1639 cpu = gi->cpu_map[i];
1640 BUG_ON(cpu == NR_CPUS);
1642 /* allocate space for the whole group */
1643 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1646 goto out_free_areas;
1648 /* kmemleak tracks the percpu allocations separately */
1652 base = min(ptr, base);
1656 * Copy data and free unused parts. This should happen after all
1657 * allocations are complete; otherwise, we may end up with
1658 * overlapping groups.
1660 for (group = 0; group < ai->nr_groups; group++) {
1661 struct pcpu_group_info *gi = &ai->groups[group];
1662 void *ptr = areas[group];
1664 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1665 if (gi->cpu_map[i] == NR_CPUS) {
1666 /* unused unit, free whole */
1667 free_fn(ptr, ai->unit_size);
1670 /* copy and return the unused part */
1671 memcpy(ptr, __per_cpu_load, ai->static_size);
1672 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1676 /* base address is now known, determine group base offsets */
1678 for (group = 0; group < ai->nr_groups; group++) {
1679 ai->groups[group].base_offset = areas[group] - base;
1680 max_distance = max_t(size_t, max_distance,
1681 ai->groups[group].base_offset);
1683 max_distance += ai->unit_size;
1685 /* warn if maximum distance is further than 75% of vmalloc space */
1686 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1687 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1688 "space 0x%lx\n", max_distance,
1689 (unsigned long)(VMALLOC_END - VMALLOC_START));
1690 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1691 /* and fail if we have fallback */
1697 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1698 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1699 ai->dyn_size, ai->unit_size);
1701 rc = pcpu_setup_first_chunk(ai, base);
1705 for (group = 0; group < ai->nr_groups; group++)
1706 free_fn(areas[group],
1707 ai->groups[group].nr_units * ai->unit_size);
1709 pcpu_free_alloc_info(ai);
1711 free_bootmem(__pa(areas), areas_size);
1714 #endif /* BUILD_EMBED_FIRST_CHUNK */
1716 #ifdef BUILD_PAGE_FIRST_CHUNK
1718 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1719 * @reserved_size: the size of reserved percpu area in bytes
1720 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1721 * @free_fn: function to free percpu page, always called with PAGE_SIZE
1722 * @populate_pte_fn: function to populate pte
1724 * This is a helper to ease setting up page-remapped first percpu
1725 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1727 * This is the basic allocator. Static percpu area is allocated
1728 * page-by-page into vmalloc area.
1731 * 0 on success, -errno on failure.
1733 int __init pcpu_page_first_chunk(size_t reserved_size,
1734 pcpu_fc_alloc_fn_t alloc_fn,
1735 pcpu_fc_free_fn_t free_fn,
1736 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1738 static struct vm_struct vm;
1739 struct pcpu_alloc_info *ai;
1743 struct page **pages;
1746 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1748 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
1751 BUG_ON(ai->nr_groups != 1);
1752 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1754 unit_pages = ai->unit_size >> PAGE_SHIFT;
1756 /* unaligned allocations can't be freed, round up to page size */
1757 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1759 pages = alloc_bootmem(pages_size);
1761 /* allocate pages */
1763 for (unit = 0; unit < num_possible_cpus(); unit++)
1764 for (i = 0; i < unit_pages; i++) {
1765 unsigned int cpu = ai->groups[0].cpu_map[unit];
1768 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1770 pr_warning("PERCPU: failed to allocate %s page "
1771 "for cpu%u\n", psize_str, cpu);
1774 /* kmemleak tracks the percpu allocations separately */
1776 pages[j++] = virt_to_page(ptr);
1779 /* allocate vm area, map the pages and copy static data */
1780 vm.flags = VM_ALLOC;
1781 vm.size = num_possible_cpus() * ai->unit_size;
1782 vm_area_register_early(&vm, PAGE_SIZE);
1784 for (unit = 0; unit < num_possible_cpus(); unit++) {
1785 unsigned long unit_addr =
1786 (unsigned long)vm.addr + unit * ai->unit_size;
1788 for (i = 0; i < unit_pages; i++)
1789 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1791 /* pte already populated, the following shouldn't fail */
1792 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1795 panic("failed to map percpu area, err=%d\n", rc);
1798 * FIXME: Archs with virtual cache should flush local
1799 * cache for the linear mapping here - something
1800 * equivalent to flush_cache_vmap() on the local cpu.
1801 * flush_cache_vmap() can't be used as most supporting
1802 * data structures are not set up yet.
1805 /* copy static data */
1806 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1809 /* we're ready, commit */
1810 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1811 unit_pages, psize_str, vm.addr, ai->static_size,
1812 ai->reserved_size, ai->dyn_size);
1814 rc = pcpu_setup_first_chunk(ai, vm.addr);
1819 free_fn(page_address(pages[j]), PAGE_SIZE);
1822 free_bootmem(__pa(pages), pages_size);
1823 pcpu_free_alloc_info(ai);
1826 #endif /* BUILD_PAGE_FIRST_CHUNK */
1828 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1830 * Generic SMP percpu area setup.
1832 * The embedding helper is used because its behavior closely resembles
1833 * the original non-dynamic generic percpu area setup. This is
1834 * important because many archs have addressing restrictions and might
1835 * fail if the percpu area is located far away from the previous
1836 * location. As an added bonus, in non-NUMA cases, embedding is
1837 * generally a good idea TLB-wise because percpu area can piggy back
1838 * on the physical linear memory mapping which uses large page
1839 * mappings on applicable archs.
1841 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1842 EXPORT_SYMBOL(__per_cpu_offset);
1844 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1847 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1850 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1852 free_bootmem(__pa(ptr), size);
1855 void __init setup_per_cpu_areas(void)
1857 unsigned long delta;
1862 * Always reserve area for module percpu variables. That's
1863 * what the legacy allocator did.
1865 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1866 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1867 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1869 panic("Failed to initialize percpu areas.");
1871 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1872 for_each_possible_cpu(cpu)
1873 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1875 #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1877 #else /* CONFIG_SMP */
1880 * UP percpu area setup.
1882 * UP always uses km-based percpu allocator with identity mapping.
1883 * Static percpu variables are indistinguishable from the usual static
1884 * variables and don't require any special preparation.
1886 void __init setup_per_cpu_areas(void)
1888 const size_t unit_size =
1889 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1890 PERCPU_DYNAMIC_RESERVE));
1891 struct pcpu_alloc_info *ai;
1894 ai = pcpu_alloc_alloc_info(1, 1);
1895 fc = __alloc_bootmem(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
1897 panic("Failed to allocate memory for percpu areas.");
1898 /* kmemleak tracks the percpu allocations separately */
1901 ai->dyn_size = unit_size;
1902 ai->unit_size = unit_size;
1903 ai->atom_size = unit_size;
1904 ai->alloc_size = unit_size;
1905 ai->groups[0].nr_units = 1;
1906 ai->groups[0].cpu_map[0] = 0;
1908 if (pcpu_setup_first_chunk(ai, fc) < 0)
1909 panic("Failed to initialize percpu areas.");
1912 #endif /* CONFIG_SMP */
1915 * First and reserved chunks are initialized with temporary allocation
1916 * map in initdata so that they can be used before slab is online.
1917 * This function is called after slab is brought up and replaces those
1918 * with properly allocated maps.
1920 void __init percpu_init_late(void)
1922 struct pcpu_chunk *target_chunks[] =
1923 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1924 struct pcpu_chunk *chunk;
1925 unsigned long flags;
1928 for (i = 0; (chunk = target_chunks[i]); i++) {
1930 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1932 BUILD_BUG_ON(size > PAGE_SIZE);
1934 map = pcpu_mem_zalloc(size);
1937 spin_lock_irqsave(&pcpu_lock, flags);
1938 memcpy(map, chunk->map, size);
1940 spin_unlock_irqrestore(&pcpu_lock, flags);