blob: 2139e30a4b4490da90c2e2811d5ecc901ff07d61 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
Tejun Heo88999a82010-04-09 18:57:01 +09002 * mm/percpu.c - percpu memory allocator
Tejun Heofbf59bc2009-02-20 16:29:08 +09003 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
Tejun Heo88999a82010-04-09 18:57:01 +090010 * 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
Tejun Heo2f39e632009-07-04 08:11:00 +090013 * (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.
Tejun Heo88999a82010-04-09 18:57:01 +090016 * When a chunk is filled up, another chunk is allocated.
Tejun Heofbf59bc2009-02-20 16:29:08 +090017 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
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,
Tejun Heo2f39e632009-07-04 08:11:00 +090025 * 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.
Tejun Heofbf59bc2009-02-20 16:29:08 +090029 *
Tejun Heo2f39e632009-07-04 08:11:00 +090030 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090032 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
Namhyung Kim4785879e2010-08-11 11:24:10 +090034 * guaranteed to be equal to or larger than the maximum contiguous
Tejun Heofbf59bc2009-02-20 16:29:08 +090035 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
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.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090043 * 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.
Tejun Heofbf59bc2009-02-20 16:29:08 +090045 *
46 * To use this allocator, arch code should do the followings.
47 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090048 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090058#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090059#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090060#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090061#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090066#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090067#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090068#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090069#include <linux/workqueue.h>
Catalin Marinasf528f0b2011-09-26 17:12:53 +010070#include <linux/kmemleak.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071
72#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090073#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090074#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090075#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090076
Tejun Heofbf59bc2009-02-20 16:29:08 +090077#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 */
79
Tejun Heobbddff02010-09-03 18:22:48 +020080#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +090081/* 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) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090084 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090087#endif
88#ifndef __pcpu_ptr_to_addr
89#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090090 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090093#endif
Tejun Heobbddff02010-09-03 18:22:48 +020094#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 */
Tejun Heoe0100982009-03-10 16:27:48 +090099
Tejun Heofbf59bc2009-02-20 16:29:08 +0900100struct pcpu_chunk {
101 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +0900104 void *base_addr; /* base address of this chunk */
Al Viro723ad1d2014-03-06 21:13:18 -0500105 int map_used; /* # of map entries used before the sentry */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
Tejun Heo88999a82010-04-09 18:57:01 +0900108 void *data; /* chunk data */
Al Viro3d331ad2014-03-06 20:52:32 -0500109 int first_free; /* no free below this */
Tejun Heo8d408b42009-02-24 11:57:21 +0900110 bool immutable; /* no [de]population allowed */
Tejun Heoce3141a2009-07-04 08:11:00 +0900111 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900112};
113
Tejun Heo40150d32009-02-24 12:32:28 +0900114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900116static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900117static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900120
Tejun Heoa855b842011-11-18 10:55:35 -0800121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900124
Tejun Heofbf59bc2009-02-20 16:29:08 +0900125/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900126void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
Tejun Heofb435d52009-08-14 15:00:51 +0900129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900131
Tejun Heo65632972009-08-14 15:00:52 +0900132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static const size_t *pcpu_group_sizes __read_mostly;
136
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900137/*
138 * The first chunk which always exists. Note that unlike other
139 * chunks, this one can be allocated and mapped in several different
140 * ways and thus often doesn't live in the vmalloc area.
141 */
142static struct pcpu_chunk *pcpu_first_chunk;
143
144/*
145 * Optional reserved chunk. This chunk reserves part of the first
146 * chunk and serves it for reserved allocations. The amount of
147 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
148 * area doesn't exist, the following variables contain NULL and 0
149 * respectively.
150 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900151static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900152static int pcpu_reserved_chunk_limit;
153
Tejun Heofbf59bc2009-02-20 16:29:08 +0900154/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900155 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900156 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900157 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
Tejun Heoce3141a2009-07-04 08:11:00 +0900158 * protects allocation/reclaim paths, chunks, populated bitmap and
159 * vmalloc mapping. The latter is a spinlock and protects the index
160 * data structures - chunk slots, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900161 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900162 * During allocation, pcpu_alloc_mutex is kept locked all the time and
163 * pcpu_lock is grabbed and released as necessary. All actual memory
Jiri Kosina403a91b2009-10-29 00:25:59 +0900164 * allocations are done using GFP_KERNEL with pcpu_lock released. In
165 * general, percpu memory can't be allocated with irq off but
166 * irqsave/restore are still used in alloc path so that it can be used
167 * from early init path - sched_init() specifically.
Tejun Heoccea34b2009-03-07 00:44:13 +0900168 *
169 * Free path accesses and alters only the index data structures, so it
170 * can be safely called from atomic context. When memory needs to be
171 * returned to the system, free path schedules reclaim_work which
172 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
173 * reclaimed, release both locks and frees the chunks. Note that it's
174 * necessary to grab both locks to remove a chunk from circulation as
175 * allocation path might be referencing the chunk with only
176 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900177 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900178static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
179static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900180
Tejun Heo40150d32009-02-24 12:32:28 +0900181static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900182
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900183/* reclaim work to release fully free chunks, scheduled from free path */
184static void pcpu_reclaim(struct work_struct *work);
185static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186
Tejun Heo020ec652010-04-09 18:57:00 +0900187static bool pcpu_addr_in_first_chunk(void *addr)
188{
189 void *first_start = pcpu_first_chunk->base_addr;
190
191 return addr >= first_start && addr < first_start + pcpu_unit_size;
192}
193
194static bool pcpu_addr_in_reserved_chunk(void *addr)
195{
196 void *first_start = pcpu_first_chunk->base_addr;
197
198 return addr >= first_start &&
199 addr < first_start + pcpu_reserved_chunk_limit;
200}
201
Tejun Heod9b55ee2009-02-24 11:57:21 +0900202static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900203{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900204 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900205 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
206}
207
Tejun Heod9b55ee2009-02-24 11:57:21 +0900208static int pcpu_size_to_slot(int size)
209{
210 if (size == pcpu_unit_size)
211 return pcpu_nr_slots - 1;
212 return __pcpu_size_to_slot(size);
213}
214
Tejun Heofbf59bc2009-02-20 16:29:08 +0900215static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216{
217 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
218 return 0;
219
220 return pcpu_size_to_slot(chunk->free_size);
221}
222
Tejun Heo88999a82010-04-09 18:57:01 +0900223/* set the pointer to a chunk in a page struct */
224static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225{
226 page->index = (unsigned long)pcpu;
227}
228
229/* obtain pointer to a chunk from a page struct */
230static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231{
232 return (struct pcpu_chunk *)page->index;
233}
234
235static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900236{
Tejun Heo2f39e632009-07-04 08:11:00 +0900237 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900238}
239
Tejun Heo9983b6f02010-06-18 11:44:31 +0200240static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
241 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900242{
Tejun Heobba174f2009-08-14 15:00:51 +0900243 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900244 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900245}
246
Tejun Heo88999a82010-04-09 18:57:01 +0900247static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
248 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900249{
250 *rs = find_next_zero_bit(chunk->populated, end, *rs);
251 *re = find_next_bit(chunk->populated, end, *rs + 1);
252}
253
Tejun Heo88999a82010-04-09 18:57:01 +0900254static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
255 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900256{
257 *rs = find_next_bit(chunk->populated, end, *rs);
258 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
259}
260
261/*
262 * (Un)populated page region iterators. Iterate over (un)populated
Uwe Kleine-Königb5950762010-11-01 15:38:34 -0400263 * page regions between @start and @end in @chunk. @rs and @re should
Tejun Heoce3141a2009-07-04 08:11:00 +0900264 * be integer variables and will be set to start and end page index of
265 * the current region.
266 */
267#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 (rs) < (re); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271
272#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276
Tejun Heofbf59bc2009-02-20 16:29:08 +0900277/**
Bob Liu90459ce02011-08-04 11:02:33 +0200278 * pcpu_mem_zalloc - allocate memory
Tejun Heo1880d932009-03-07 00:44:09 +0900279 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900280 *
Tejun Heo1880d932009-03-07 00:44:09 +0900281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
Bob Liu90459ce02011-08-04 11:02:33 +0200282 * kzalloc() is used; otherwise, vzalloc() is used. The returned
Tejun Heo1880d932009-03-07 00:44:09 +0900283 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900284 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900285 * CONTEXT:
286 * Does GFP_KERNEL allocation.
287 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900288 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900289 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900290 */
Bob Liu90459ce02011-08-04 11:02:33 +0200291static void *pcpu_mem_zalloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900292{
Tejun Heo099a19d2010-06-27 18:50:00 +0200293 if (WARN_ON_ONCE(!slab_is_available()))
294 return NULL;
295
Tejun Heofbf59bc2009-02-20 16:29:08 +0900296 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900297 return kzalloc(size, GFP_KERNEL);
Jesper Juhl7af4c092010-10-30 15:56:54 +0200298 else
299 return vzalloc(size);
Tejun Heo1880d932009-03-07 00:44:09 +0900300}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900301
Tejun Heo1880d932009-03-07 00:44:09 +0900302/**
303 * pcpu_mem_free - free memory
304 * @ptr: memory to free
305 * @size: size of the area
306 *
Bob Liu90459ce02011-08-04 11:02:33 +0200307 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
Tejun Heo1880d932009-03-07 00:44:09 +0900308 */
309static void pcpu_mem_free(void *ptr, size_t size)
310{
311 if (size <= PAGE_SIZE)
312 kfree(ptr);
313 else
314 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900315}
316
317/**
318 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
319 * @chunk: chunk of interest
320 * @oslot: the previous slot it was on
321 *
322 * This function is called after an allocation or free changed @chunk.
323 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900324 * moved to the slot. Note that the reserved chunk is never put on
325 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900326 *
327 * CONTEXT:
328 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900329 */
330static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331{
332 int nslot = pcpu_chunk_slot(chunk);
333
Tejun Heoedcb4632009-03-06 14:33:59 +0900334 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900335 if (oslot < nslot)
336 list_move(&chunk->list, &pcpu_slot[nslot]);
337 else
338 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
339 }
340}
341
Tejun Heofbf59bc2009-02-20 16:29:08 +0900342/**
Tejun Heo833af842009-11-11 15:35:18 +0900343 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
344 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900345 *
Tejun Heo833af842009-11-11 15:35:18 +0900346 * Determine whether area map of @chunk needs to be extended to
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300347 * accommodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900348 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900349 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900350 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900351 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900352 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900353 * New target map allocation length if extension is necessary, 0
354 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900355 */
Tejun Heo833af842009-11-11 15:35:18 +0900356static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900357{
358 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900359
Al Viro723ad1d2014-03-06 21:13:18 -0500360 if (chunk->map_alloc >= chunk->map_used + 3)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900361 return 0;
362
363 new_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500364 while (new_alloc < chunk->map_used + 3)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900365 new_alloc *= 2;
366
Tejun Heo833af842009-11-11 15:35:18 +0900367 return new_alloc;
368}
369
370/**
371 * pcpu_extend_area_map - extend area map of a chunk
372 * @chunk: chunk of interest
373 * @new_alloc: new target allocation length of the area map
374 *
375 * Extend area map of @chunk to have @new_alloc entries.
376 *
377 * CONTEXT:
378 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
379 *
380 * RETURNS:
381 * 0 on success, -errno on failure.
382 */
383static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384{
385 int *old = NULL, *new = NULL;
386 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
387 unsigned long flags;
388
Bob Liu90459ce02011-08-04 11:02:33 +0200389 new = pcpu_mem_zalloc(new_size);
Tejun Heo833af842009-11-11 15:35:18 +0900390 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900391 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900392
Tejun Heo833af842009-11-11 15:35:18 +0900393 /* acquire pcpu_lock and switch to new area map */
394 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900395
Tejun Heo833af842009-11-11 15:35:18 +0900396 if (new_alloc <= chunk->map_alloc)
397 goto out_unlock;
398
399 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
Huang Shijiea002d142010-08-08 14:39:07 +0200400 old = chunk->map;
401
402 memcpy(new, old, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900403
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900404 chunk->map_alloc = new_alloc;
405 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900406 new = NULL;
407
408out_unlock:
409 spin_unlock_irqrestore(&pcpu_lock, flags);
410
411 /*
412 * pcpu_mem_free() might end up calling vfree() which uses
413 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 */
415 pcpu_mem_free(old, old_size);
416 pcpu_mem_free(new, new_size);
417
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900418 return 0;
419}
420
421/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900422 * pcpu_alloc_area - allocate area from a pcpu_chunk
423 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900424 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900425 * @align: wanted align
426 *
427 * Try to allocate @size bytes area aligned at @align from @chunk.
428 * Note that this function only allocates the offset. It doesn't
429 * populate or map the area.
430 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900431 * @chunk->map must have at least two free slots.
432 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900433 * CONTEXT:
434 * pcpu_lock.
435 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900436 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900437 * Allocated offset in @chunk on success, -1 if no matching area is
438 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900439 */
440static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
441{
442 int oslot = pcpu_chunk_slot(chunk);
443 int max_contig = 0;
444 int i, off;
Al Viro3d331ad2014-03-06 20:52:32 -0500445 bool seen_free = false;
Al Viro723ad1d2014-03-06 21:13:18 -0500446 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900447
Al Viro3d331ad2014-03-06 20:52:32 -0500448 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900449 int head, tail;
Al Viro723ad1d2014-03-06 21:13:18 -0500450 int this_size;
451
452 off = *p;
453 if (off & 1)
454 continue;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900455
456 /* extra for alignment requirement */
457 head = ALIGN(off, align) - off;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900458
Al Viro723ad1d2014-03-06 21:13:18 -0500459 this_size = (p[1] & ~1) - off;
460 if (this_size < head + size) {
Al Viro3d331ad2014-03-06 20:52:32 -0500461 if (!seen_free) {
462 chunk->first_free = i;
463 seen_free = true;
464 }
Al Viro723ad1d2014-03-06 21:13:18 -0500465 max_contig = max(this_size, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900466 continue;
467 }
468
469 /*
470 * If head is small or the previous block is free,
471 * merge'em. Note that 'small' is defined as smaller
472 * than sizeof(int), which is very small but isn't too
473 * uncommon for percpu allocations.
474 */
Al Viro723ad1d2014-03-06 21:13:18 -0500475 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800476 *p = off += head;
Al Viro723ad1d2014-03-06 21:13:18 -0500477 if (p[-1] & 1)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900478 chunk->free_size -= head;
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800479 else
480 max_contig = max(*p - p[-1], max_contig);
Al Viro723ad1d2014-03-06 21:13:18 -0500481 this_size -= head;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900482 head = 0;
483 }
484
485 /* if tail is small, just keep it around */
Al Viro723ad1d2014-03-06 21:13:18 -0500486 tail = this_size - head - size;
487 if (tail < sizeof(int)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900488 tail = 0;
Al Viro723ad1d2014-03-06 21:13:18 -0500489 size = this_size - head;
490 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900491
492 /* split if warranted */
493 if (head || tail) {
Al Viro706c16f2014-03-06 21:08:24 -0500494 int nr_extra = !!head + !!tail;
495
496 /* insert new subblocks */
Al Viro723ad1d2014-03-06 21:13:18 -0500497 memmove(p + nr_extra + 1, p + 1,
Al Viro706c16f2014-03-06 21:08:24 -0500498 sizeof(chunk->map[0]) * (chunk->map_used - i));
499 chunk->map_used += nr_extra;
500
Tejun Heofbf59bc2009-02-20 16:29:08 +0900501 if (head) {
Al Viro3d331ad2014-03-06 20:52:32 -0500502 if (!seen_free) {
503 chunk->first_free = i;
504 seen_free = true;
505 }
Al Viro723ad1d2014-03-06 21:13:18 -0500506 *++p = off += head;
507 ++i;
Al Viro706c16f2014-03-06 21:08:24 -0500508 max_contig = max(head, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900509 }
Al Viro706c16f2014-03-06 21:08:24 -0500510 if (tail) {
Al Viro723ad1d2014-03-06 21:13:18 -0500511 p[1] = off + size;
Al Viro706c16f2014-03-06 21:08:24 -0500512 max_contig = max(tail, max_contig);
513 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900514 }
515
Al Viro3d331ad2014-03-06 20:52:32 -0500516 if (!seen_free)
517 chunk->first_free = i + 1;
518
Tejun Heofbf59bc2009-02-20 16:29:08 +0900519 /* update hint and mark allocated */
Al Viro723ad1d2014-03-06 21:13:18 -0500520 if (i + 1 == chunk->map_used)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900521 chunk->contig_hint = max_contig; /* fully scanned */
522 else
523 chunk->contig_hint = max(chunk->contig_hint,
524 max_contig);
525
Al Viro723ad1d2014-03-06 21:13:18 -0500526 chunk->free_size -= size;
527 *p |= 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900528
529 pcpu_chunk_relocate(chunk, oslot);
530 return off;
531 }
532
533 chunk->contig_hint = max_contig; /* fully scanned */
534 pcpu_chunk_relocate(chunk, oslot);
535
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900536 /* tell the upper layer that this chunk has no matching area */
537 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900538}
539
540/**
541 * pcpu_free_area - free area to a pcpu_chunk
542 * @chunk: chunk of interest
543 * @freeme: offset of area to free
544 *
545 * Free area starting from @freeme to @chunk. Note that this function
546 * only modifies the allocation map. It doesn't depopulate or unmap
547 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900548 *
549 * CONTEXT:
550 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900551 */
552static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
553{
554 int oslot = pcpu_chunk_slot(chunk);
Al Viro723ad1d2014-03-06 21:13:18 -0500555 int off = 0;
556 unsigned i, j;
557 int to_free = 0;
558 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900559
Al Viro723ad1d2014-03-06 21:13:18 -0500560 freeme |= 1; /* we are searching for <given offset, in use> pair */
561
562 i = 0;
563 j = chunk->map_used;
564 while (i != j) {
565 unsigned k = (i + j) / 2;
566 off = chunk->map[k];
567 if (off < freeme)
568 i = k + 1;
569 else if (off > freeme)
570 j = k;
571 else
572 i = j = k;
573 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900574 BUG_ON(off != freeme);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900575
Al Viro3d331ad2014-03-06 20:52:32 -0500576 if (i < chunk->first_free)
577 chunk->first_free = i;
578
Al Viro723ad1d2014-03-06 21:13:18 -0500579 p = chunk->map + i;
580 *p = off &= ~1;
581 chunk->free_size += (p[1] & ~1) - off;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900582
Tejun Heofbf59bc2009-02-20 16:29:08 +0900583 /* merge with next? */
Al Viro723ad1d2014-03-06 21:13:18 -0500584 if (!(p[1] & 1))
585 to_free++;
586 /* merge with previous? */
587 if (i > 0 && !(p[-1] & 1)) {
588 to_free++;
589 i--;
590 p--;
591 }
592 if (to_free) {
593 chunk->map_used -= to_free;
594 memmove(p + 1, p + 1 + to_free,
595 (chunk->map_used - i) * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900596 }
597
Al Viro723ad1d2014-03-06 21:13:18 -0500598 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900599 pcpu_chunk_relocate(chunk, oslot);
600}
601
Tejun Heo60810892010-04-09 18:57:01 +0900602static struct pcpu_chunk *pcpu_alloc_chunk(void)
603{
604 struct pcpu_chunk *chunk;
605
Bob Liu90459ce02011-08-04 11:02:33 +0200606 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900607 if (!chunk)
608 return NULL;
609
Bob Liu90459ce02011-08-04 11:02:33 +0200610 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
611 sizeof(chunk->map[0]));
Tejun Heo60810892010-04-09 18:57:01 +0900612 if (!chunk->map) {
Jianyu Zhan5a838c32014-04-14 13:47:40 +0800613 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900614 return NULL;
615 }
616
617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500618 chunk->map[0] = 0;
619 chunk->map[1] = pcpu_unit_size | 1;
620 chunk->map_used = 1;
Tejun Heo60810892010-04-09 18:57:01 +0900621
622 INIT_LIST_HEAD(&chunk->list);
623 chunk->free_size = pcpu_unit_size;
624 chunk->contig_hint = pcpu_unit_size;
625
626 return chunk;
627}
628
629static void pcpu_free_chunk(struct pcpu_chunk *chunk)
630{
631 if (!chunk)
632 return;
633 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Joonsoo Kimb4916cb2012-10-29 22:59:58 +0900634 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900635}
636
Tejun Heo9f645532010-04-09 18:57:01 +0900637/*
638 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900639 *
Tejun Heo9f645532010-04-09 18:57:01 +0900640 * 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.
Tejun Heoce3141a2009-07-04 08:11:00 +0900644 *
Tejun Heo9f645532010-04-09 18:57:01 +0900645 * 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
Tejun Heofbf59bc2009-02-20 16:29:08 +0900651 */
Tejun Heo9f645532010-04-09 18:57:01 +0900652static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654static struct pcpu_chunk *pcpu_create_chunk(void);
655static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656static struct page *pcpu_addr_to_page(void *addr);
657static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900658
Tejun Heob0c97782010-04-09 18:57:01 +0900659#ifdef CONFIG_NEED_PER_CPU_KM
660#include "percpu-km.c"
661#else
Tejun Heo9f645532010-04-09 18:57:01 +0900662#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900663#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900664
665/**
Tejun Heo88999a82010-04-09 18:57:01 +0900666 * pcpu_chunk_addr_search - determine chunk containing specified address
667 * @addr: address for which the chunk needs to be determined.
668 *
669 * RETURNS:
670 * The address of the found chunk.
671 */
672static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
673{
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;
680 }
681
682 /*
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.
688 */
689 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900690 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900691}
692
693/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900694 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900695 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900696 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900697 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900698 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900699 * Allocate percpu area of @size bytes aligned at @align.
700 *
701 * CONTEXT:
702 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900703 *
704 * RETURNS:
705 * Percpu pointer to the allocated area on success, NULL on failure.
706 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900707static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900708{
Tejun Heof2badb02009-09-29 09:17:58 +0900709 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900710 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900711 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +0900712 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900713 unsigned long flags;
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100714 void __percpu *ptr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900715
Al Viro723ad1d2014-03-06 21:13:18 -0500716 /*
717 * We want the lowest bit of offset available for in-use/free
Viro2f69fa82014-03-17 16:01:27 -0400718 * indicator, so force >= 16bit alignment and make size even.
Al Viro723ad1d2014-03-06 21:13:18 -0500719 */
720 if (unlikely(align < 2))
721 align = 2;
722
Christoph Lameterfb009e32014-06-19 09:59:18 -0500723 size = ALIGN(size, 2);
Viro2f69fa82014-03-17 16:01:27 -0400724
Tejun Heo8d408b42009-02-24 11:57:21 +0900725 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900726 WARN(true, "illegal size (%zu) or align (%zu) for "
727 "percpu allocation\n", size, align);
728 return NULL;
729 }
730
Tejun Heoccea34b2009-03-07 00:44:13 +0900731 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +0900732 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900733
Tejun Heoedcb4632009-03-06 14:33:59 +0900734 /* serve reserved allocations from the reserved chunk if available */
735 if (reserved && pcpu_reserved_chunk) {
736 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900737
738 if (size > chunk->contig_hint) {
739 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900740 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900741 }
Tejun Heo833af842009-11-11 15:35:18 +0900742
743 while ((new_alloc = pcpu_need_to_extend(chunk))) {
744 spin_unlock_irqrestore(&pcpu_lock, flags);
745 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
746 err = "failed to extend area map of reserved chunk";
747 goto fail_unlock_mutex;
748 }
749 spin_lock_irqsave(&pcpu_lock, flags);
750 }
751
Tejun Heoedcb4632009-03-06 14:33:59 +0900752 off = pcpu_alloc_area(chunk, size, align);
753 if (off >= 0)
754 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900755
Tejun Heof2badb02009-09-29 09:17:58 +0900756 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900757 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900758 }
759
Tejun Heoccea34b2009-03-07 00:44:13 +0900760restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900761 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900762 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
763 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
764 if (size > chunk->contig_hint)
765 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900766
Tejun Heo833af842009-11-11 15:35:18 +0900767 new_alloc = pcpu_need_to_extend(chunk);
768 if (new_alloc) {
769 spin_unlock_irqrestore(&pcpu_lock, flags);
770 if (pcpu_extend_area_map(chunk,
771 new_alloc) < 0) {
772 err = "failed to extend area map";
773 goto fail_unlock_mutex;
774 }
775 spin_lock_irqsave(&pcpu_lock, flags);
776 /*
777 * pcpu_lock has been dropped, need to
778 * restart cpu_slot list walking.
779 */
780 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900781 }
782
Tejun Heofbf59bc2009-02-20 16:29:08 +0900783 off = pcpu_alloc_area(chunk, size, align);
784 if (off >= 0)
785 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900786 }
787 }
788
789 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900790 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900791
Tejun Heo60810892010-04-09 18:57:01 +0900792 chunk = pcpu_create_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +0900793 if (!chunk) {
794 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +0900795 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +0900796 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900797
Jiri Kosina403a91b2009-10-29 00:25:59 +0900798 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900799 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900800 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900801
802area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900803 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900804
Tejun Heofbf59bc2009-02-20 16:29:08 +0900805 /* populate, map and clear the area */
806 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +0900807 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900808 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +0900809 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +0900810 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900811 }
812
Tejun Heoccea34b2009-03-07 00:44:13 +0900813 mutex_unlock(&pcpu_alloc_mutex);
814
Tejun Heobba174f2009-08-14 15:00:51 +0900815 /* return address relative to base address */
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100816 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
817 kmemleak_alloc_percpu(ptr, size);
818 return ptr;
Tejun Heoccea34b2009-03-07 00:44:13 +0900819
820fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900821 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900822fail_unlock_mutex:
823 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +0900824 if (warn_limit) {
825 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
826 "%s\n", size, align, err);
827 dump_stack();
828 if (!--warn_limit)
829 pr_info("PERCPU: limit reached, disable warning\n");
830 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900831 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900832}
Tejun Heoedcb4632009-03-06 14:33:59 +0900833
834/**
835 * __alloc_percpu - allocate dynamic percpu area
836 * @size: size of area to allocate in bytes
837 * @align: alignment of area (max PAGE_SIZE)
838 *
Tejun Heo9329ba92010-09-10 11:01:56 +0200839 * Allocate zero-filled percpu area of @size bytes aligned at @align.
840 * Might sleep. Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +0900841 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900842 * CONTEXT:
843 * Does GFP_KERNEL allocation.
844 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900845 * RETURNS:
846 * Percpu pointer to the allocated area on success, NULL on failure.
847 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900848void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900849{
850 return pcpu_alloc(size, align, false);
851}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900852EXPORT_SYMBOL_GPL(__alloc_percpu);
853
Tejun Heoedcb4632009-03-06 14:33:59 +0900854/**
855 * __alloc_reserved_percpu - allocate reserved percpu area
856 * @size: size of area to allocate in bytes
857 * @align: alignment of area (max PAGE_SIZE)
858 *
Tejun Heo9329ba92010-09-10 11:01:56 +0200859 * Allocate zero-filled percpu area of @size bytes aligned at @align
860 * from reserved percpu area if arch has set it up; otherwise,
861 * allocation is served from the same dynamic area. Might sleep.
862 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +0900863 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900864 * CONTEXT:
865 * Does GFP_KERNEL allocation.
866 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900867 * RETURNS:
868 * Percpu pointer to the allocated area on success, NULL on failure.
869 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900870void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900871{
872 return pcpu_alloc(size, align, true);
873}
874
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900875/**
876 * pcpu_reclaim - reclaim fully free chunks, workqueue function
877 * @work: unused
878 *
879 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900880 *
881 * CONTEXT:
882 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900883 */
884static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900885{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900886 LIST_HEAD(todo);
887 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
888 struct pcpu_chunk *chunk, *next;
889
Tejun Heoccea34b2009-03-07 00:44:13 +0900890 mutex_lock(&pcpu_alloc_mutex);
891 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900892
893 list_for_each_entry_safe(chunk, next, head, list) {
894 WARN_ON(chunk->immutable);
895
896 /* spare the first one */
897 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
898 continue;
899
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900900 list_move(&chunk->list, &todo);
901 }
902
Tejun Heoccea34b2009-03-07 00:44:13 +0900903 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900904
905 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +0900906 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heo60810892010-04-09 18:57:01 +0900907 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900908 }
Tejun Heo971f3912009-08-14 15:00:49 +0900909
910 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900911}
912
913/**
914 * free_percpu - free percpu area
915 * @ptr: pointer to area to free
916 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900917 * Free percpu area @ptr.
918 *
919 * CONTEXT:
920 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900921 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900922void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900923{
Andrew Morton129182e2010-01-08 14:42:39 -0800924 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900925 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900926 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900927 int off;
928
929 if (!ptr)
930 return;
931
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100932 kmemleak_free_percpu(ptr);
933
Andrew Morton129182e2010-01-08 14:42:39 -0800934 addr = __pcpu_ptr_to_addr(ptr);
935
Tejun Heoccea34b2009-03-07 00:44:13 +0900936 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900937
938 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +0900939 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900940
941 pcpu_free_area(chunk, off);
942
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900943 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900944 if (chunk->free_size == pcpu_unit_size) {
945 struct pcpu_chunk *pos;
946
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900947 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900948 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900949 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900950 break;
951 }
952 }
953
Tejun Heoccea34b2009-03-07 00:44:13 +0900954 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900955}
956EXPORT_SYMBOL_GPL(free_percpu);
957
Vivek Goyal3b034b02009-11-24 15:50:03 +0900958/**
Tejun Heo10fad5e2010-03-10 18:57:54 +0900959 * is_kernel_percpu_address - test whether address is from static percpu area
960 * @addr: address to test
961 *
962 * Test whether @addr belongs to in-kernel static percpu area. Module
963 * static percpu areas are not considered. For those, use
964 * is_module_percpu_address().
965 *
966 * RETURNS:
967 * %true if @addr is from in-kernel static percpu area, %false otherwise.
968 */
969bool is_kernel_percpu_address(unsigned long addr)
970{
Tejun Heobbddff02010-09-03 18:22:48 +0200971#ifdef CONFIG_SMP
Tejun Heo10fad5e2010-03-10 18:57:54 +0900972 const size_t static_size = __per_cpu_end - __per_cpu_start;
973 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
974 unsigned int cpu;
975
976 for_each_possible_cpu(cpu) {
977 void *start = per_cpu_ptr(base, cpu);
978
979 if ((void *)addr >= start && (void *)addr < start + static_size)
980 return true;
981 }
Tejun Heobbddff02010-09-03 18:22:48 +0200982#endif
983 /* on UP, can't distinguish from other static vars, always false */
Tejun Heo10fad5e2010-03-10 18:57:54 +0900984 return false;
985}
986
987/**
Vivek Goyal3b034b02009-11-24 15:50:03 +0900988 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
989 * @addr: the address to be converted to physical address
990 *
991 * Given @addr which is dereferenceable address obtained via one of
992 * percpu access macros, this function translates it into its physical
993 * address. The caller is responsible for ensuring @addr stays valid
994 * until this function finishes.
995 *
Dave Young67589c712011-11-23 08:20:53 -0800996 * percpu allocator has special setup for the first chunk, which currently
997 * supports either embedding in linear address space or vmalloc mapping,
998 * and, from the second one, the backing allocator (currently either vm or
999 * km) provides translation.
1000 *
1001 * The addr can be tranlated simply without checking if it falls into the
1002 * first chunk. But the current code reflects better how percpu allocator
1003 * actually works, and the verification can discover both bugs in percpu
1004 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1005 * code.
1006 *
Vivek Goyal3b034b02009-11-24 15:50:03 +09001007 * RETURNS:
1008 * The physical address for @addr.
1009 */
1010phys_addr_t per_cpu_ptr_to_phys(void *addr)
1011{
Tejun Heo9983b6f02010-06-18 11:44:31 +02001012 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1013 bool in_first_chunk = false;
Tejun Heoa855b842011-11-18 10:55:35 -08001014 unsigned long first_low, first_high;
Tejun Heo9983b6f02010-06-18 11:44:31 +02001015 unsigned int cpu;
1016
1017 /*
Tejun Heoa855b842011-11-18 10:55:35 -08001018 * The following test on unit_low/high isn't strictly
Tejun Heo9983b6f02010-06-18 11:44:31 +02001019 * necessary but will speed up lookups of addresses which
1020 * aren't in the first chunk.
1021 */
Tejun Heoa855b842011-11-18 10:55:35 -08001022 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1023 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1024 pcpu_unit_pages);
1025 if ((unsigned long)addr >= first_low &&
1026 (unsigned long)addr < first_high) {
Tejun Heo9983b6f02010-06-18 11:44:31 +02001027 for_each_possible_cpu(cpu) {
1028 void *start = per_cpu_ptr(base, cpu);
1029
1030 if (addr >= start && addr < start + pcpu_unit_size) {
1031 in_first_chunk = true;
1032 break;
1033 }
1034 }
1035 }
1036
1037 if (in_first_chunk) {
David Howellseac522e2011-03-28 12:53:29 +01001038 if (!is_vmalloc_addr(addr))
Tejun Heo020ec652010-04-09 18:57:00 +09001039 return __pa(addr);
1040 else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001041 return page_to_phys(vmalloc_to_page(addr)) +
1042 offset_in_page(addr);
Tejun Heo020ec652010-04-09 18:57:00 +09001043 } else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001044 return page_to_phys(pcpu_addr_to_page(addr)) +
1045 offset_in_page(addr);
Vivek Goyal3b034b02009-11-24 15:50:03 +09001046}
1047
Tejun Heofbf59bc2009-02-20 16:29:08 +09001048/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001049 * pcpu_alloc_alloc_info - allocate percpu allocation info
1050 * @nr_groups: the number of groups
1051 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001052 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001053 * Allocate ai which is large enough for @nr_groups groups containing
1054 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1055 * cpu_map array which is long enough for @nr_units and filled with
1056 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1057 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001058 *
1059 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001060 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1061 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001062 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001063struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1064 int nr_units)
1065{
1066 struct pcpu_alloc_info *ai;
1067 size_t base_size, ai_size;
1068 void *ptr;
1069 int unit;
1070
1071 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1072 __alignof__(ai->groups[0].cpu_map[0]));
1073 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1074
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001075 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001076 if (!ptr)
1077 return NULL;
1078 ai = ptr;
1079 ptr += base_size;
1080
1081 ai->groups[0].cpu_map = ptr;
1082
1083 for (unit = 0; unit < nr_units; unit++)
1084 ai->groups[0].cpu_map[unit] = NR_CPUS;
1085
1086 ai->nr_groups = nr_groups;
1087 ai->__ai_size = PFN_ALIGN(ai_size);
1088
1089 return ai;
1090}
1091
1092/**
1093 * pcpu_free_alloc_info - free percpu allocation info
1094 * @ai: pcpu_alloc_info to free
1095 *
1096 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1097 */
1098void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1099{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001100 memblock_free_early(__pa(ai), ai->__ai_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001101}
1102
1103/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001104 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1105 * @lvl: loglevel
1106 * @ai: allocation info to dump
1107 *
1108 * Print out information about @ai using loglevel @lvl.
1109 */
1110static void pcpu_dump_alloc_info(const char *lvl,
1111 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001112{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001113 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001114 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001115 int alloc = 0, alloc_end = 0;
1116 int group, v;
1117 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001118
Tejun Heofd1e8a12009-08-14 15:00:51 +09001119 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001120 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001121 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001122
Tejun Heofd1e8a12009-08-14 15:00:51 +09001123 v = num_possible_cpus();
1124 while (v /= 10)
1125 cpu_width++;
1126 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001127
Tejun Heofd1e8a12009-08-14 15:00:51 +09001128 upa = ai->alloc_size / ai->unit_size;
1129 width = upa * (cpu_width + 1) + group_width + 3;
1130 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001131
Tejun Heofd1e8a12009-08-14 15:00:51 +09001132 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1133 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1134 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1135
1136 for (group = 0; group < ai->nr_groups; group++) {
1137 const struct pcpu_group_info *gi = &ai->groups[group];
1138 int unit = 0, unit_end = 0;
1139
1140 BUG_ON(gi->nr_units % upa);
1141 for (alloc_end += gi->nr_units / upa;
1142 alloc < alloc_end; alloc++) {
1143 if (!(alloc % apl)) {
Tejun Heocb129822012-03-29 09:45:58 -07001144 printk(KERN_CONT "\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001145 printk("%spcpu-alloc: ", lvl);
1146 }
Tejun Heocb129822012-03-29 09:45:58 -07001147 printk(KERN_CONT "[%0*d] ", group_width, group);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001148
1149 for (unit_end += upa; unit < unit_end; unit++)
1150 if (gi->cpu_map[unit] != NR_CPUS)
Tejun Heocb129822012-03-29 09:45:58 -07001151 printk(KERN_CONT "%0*d ", cpu_width,
Tejun Heofd1e8a12009-08-14 15:00:51 +09001152 gi->cpu_map[unit]);
1153 else
Tejun Heocb129822012-03-29 09:45:58 -07001154 printk(KERN_CONT "%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001155 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001156 }
Tejun Heocb129822012-03-29 09:45:58 -07001157 printk(KERN_CONT "\n");
Tejun Heo033e48f2009-08-14 15:00:51 +09001158}
Tejun Heo033e48f2009-08-14 15:00:51 +09001159
Tejun Heofbf59bc2009-02-20 16:29:08 +09001160/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001161 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001162 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001163 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001164 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001165 * Initialize the first percpu chunk which contains the kernel static
1166 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001167 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001168 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001169 * @ai contains all information necessary to initialize the first
1170 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001171 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001172 * @ai->static_size is the size of static percpu area.
1173 *
1174 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001175 * reserve after the static area in the first chunk. This reserves
1176 * the first chunk such that it's available only through reserved
1177 * percpu allocation. This is primarily used to serve module percpu
1178 * static areas on architectures where the addressing model has
1179 * limited offset range for symbol relocations to guarantee module
1180 * percpu symbols fall inside the relocatable range.
1181 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001182 * @ai->dyn_size determines the number of bytes available for dynamic
1183 * allocation in the first chunk. The area between @ai->static_size +
1184 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001185 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001186 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1187 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1188 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001189 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001190 * @ai->atom_size is the allocation atom size and used as alignment
1191 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001192 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001193 * @ai->alloc_size is the allocation size and always multiple of
1194 * @ai->atom_size. This is larger than @ai->atom_size if
1195 * @ai->unit_size is larger than @ai->atom_size.
1196 *
1197 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1198 * percpu areas. Units which should be colocated are put into the
1199 * same group. Dynamic VM areas will be allocated according to these
1200 * groupings. If @ai->nr_groups is zero, a single group containing
1201 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001202 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001203 * The caller should have mapped the first chunk at @base_addr and
1204 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001205 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001206 * If the first chunk ends up with both reserved and dynamic areas, it
1207 * is served by two chunks - one to serve the core static and reserved
1208 * areas and the other for the dynamic area. They share the same vm
1209 * and page map but uses different area allocation map to stay away
1210 * from each other. The latter chunk is circulated in the chunk slots
1211 * and available for dynamic allocation like any other chunks.
1212 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001213 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001214 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001215 */
Tejun Heofb435d52009-08-14 15:00:51 +09001216int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1217 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001218{
Tejun Heo635b75f2009-09-24 09:43:11 +09001219 static char cpus_buf[4096] __initdata;
Tejun Heo099a19d2010-06-27 18:50:00 +02001220 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1221 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001222 size_t dyn_size = ai->dyn_size;
1223 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001224 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001225 unsigned long *group_offsets;
1226 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001227 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001228 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001229 int *unit_map;
1230 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001231
Tejun Heo635b75f2009-09-24 09:43:11 +09001232 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1233
1234#define PCPU_SETUP_BUG_ON(cond) do { \
1235 if (unlikely(cond)) { \
1236 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1237 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1238 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1239 BUG(); \
1240 } \
1241} while (0)
1242
Tejun Heo2f39e632009-07-04 08:11:00 +09001243 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001244 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02001245#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09001246 PCPU_SETUP_BUG_ON(!ai->static_size);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001247 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
Tejun Heobbddff02010-09-03 18:22:48 +02001248#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09001249 PCPU_SETUP_BUG_ON(!base_addr);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001250 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
Tejun Heo635b75f2009-09-24 09:43:11 +09001251 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1252 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1253 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001254 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001255 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001256
Tejun Heo65632972009-08-14 15:00:52 +09001257 /* process group information and build config tables accordingly */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001258 group_offsets = memblock_virt_alloc(ai->nr_groups *
1259 sizeof(group_offsets[0]), 0);
1260 group_sizes = memblock_virt_alloc(ai->nr_groups *
1261 sizeof(group_sizes[0]), 0);
1262 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1263 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
Tejun Heo2f39e632009-07-04 08:11:00 +09001264
Tejun Heofd1e8a12009-08-14 15:00:51 +09001265 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001266 unit_map[cpu] = UINT_MAX;
Tejun Heoa855b842011-11-18 10:55:35 -08001267
1268 pcpu_low_unit_cpu = NR_CPUS;
1269 pcpu_high_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001270
Tejun Heofd1e8a12009-08-14 15:00:51 +09001271 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1272 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001273
Tejun Heo65632972009-08-14 15:00:52 +09001274 group_offsets[group] = gi->base_offset;
1275 group_sizes[group] = gi->nr_units * ai->unit_size;
1276
Tejun Heofd1e8a12009-08-14 15:00:51 +09001277 for (i = 0; i < gi->nr_units; i++) {
1278 cpu = gi->cpu_map[i];
1279 if (cpu == NR_CPUS)
1280 continue;
1281
Tejun Heo635b75f2009-09-24 09:43:11 +09001282 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1283 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1284 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001285
1286 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001287 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1288
Tejun Heoa855b842011-11-18 10:55:35 -08001289 /* determine low/high unit_cpu */
1290 if (pcpu_low_unit_cpu == NR_CPUS ||
1291 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1292 pcpu_low_unit_cpu = cpu;
1293 if (pcpu_high_unit_cpu == NR_CPUS ||
1294 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1295 pcpu_high_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001296 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001297 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001298 pcpu_nr_units = unit;
1299
1300 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001301 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1302
1303 /* we're done parsing the input, undefine BUG macro and dump config */
1304#undef PCPU_SETUP_BUG_ON
Tejun Heobcbea792010-12-22 14:19:14 +01001305 pcpu_dump_alloc_info(KERN_DEBUG, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001306
Tejun Heo65632972009-08-14 15:00:52 +09001307 pcpu_nr_groups = ai->nr_groups;
1308 pcpu_group_offsets = group_offsets;
1309 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001310 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001311 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001312
1313 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001314 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001315 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001316 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001317 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1318 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001319
Tejun Heod9b55ee2009-02-24 11:57:21 +09001320 /*
1321 * Allocate chunk slots. The additional last slot is for
1322 * empty chunks.
1323 */
1324 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001325 pcpu_slot = memblock_virt_alloc(
1326 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001327 for (i = 0; i < pcpu_nr_slots; i++)
1328 INIT_LIST_HEAD(&pcpu_slot[i]);
1329
Tejun Heoedcb4632009-03-06 14:33:59 +09001330 /*
1331 * Initialize static chunk. If reserved_size is zero, the
1332 * static chunk covers static area + dynamic allocation area
1333 * in the first chunk. If reserved_size is not zero, it
1334 * covers static area + reserved area (mostly used for module
1335 * static percpu allocation).
1336 */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001337 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heo2441d152009-03-06 14:33:59 +09001338 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001339 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001340 schunk->map = smap;
1341 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001342 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001343 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001344
Tejun Heofd1e8a12009-08-14 15:00:51 +09001345 if (ai->reserved_size) {
1346 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001347 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001348 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001349 } else {
1350 schunk->free_size = dyn_size;
1351 dyn_size = 0; /* dynamic area covered */
1352 }
Tejun Heo2441d152009-03-06 14:33:59 +09001353 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001354
Al Viro723ad1d2014-03-06 21:13:18 -05001355 schunk->map[0] = 1;
1356 schunk->map[1] = ai->static_size;
1357 schunk->map_used = 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001358 if (schunk->free_size)
Al Viro723ad1d2014-03-06 21:13:18 -05001359 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1360 else
1361 schunk->map[1] |= 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001362
Tejun Heoedcb4632009-03-06 14:33:59 +09001363 /* init dynamic chunk if necessary */
1364 if (dyn_size) {
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001365 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001366 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001367 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001368 dchunk->map = dmap;
1369 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001370 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001371 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001372
1373 dchunk->contig_hint = dchunk->free_size = dyn_size;
Al Viro723ad1d2014-03-06 21:13:18 -05001374 dchunk->map[0] = 1;
1375 dchunk->map[1] = pcpu_reserved_chunk_limit;
1376 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1377 dchunk->map_used = 2;
Tejun Heoedcb4632009-03-06 14:33:59 +09001378 }
1379
Tejun Heo2441d152009-03-06 14:33:59 +09001380 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001381 pcpu_first_chunk = dchunk ?: schunk;
1382 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001383
1384 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001385 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001386 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001387}
Tejun Heo66c3a752009-03-10 16:27:48 +09001388
Tejun Heobbddff02010-09-03 18:22:48 +02001389#ifdef CONFIG_SMP
1390
Andi Kleen17f36092012-10-04 17:12:07 -07001391const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
Tejun Heof58dc012009-08-14 15:00:50 +09001392 [PCPU_FC_AUTO] = "auto",
1393 [PCPU_FC_EMBED] = "embed",
1394 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001395};
Tejun Heo66c3a752009-03-10 16:27:48 +09001396
Tejun Heof58dc012009-08-14 15:00:50 +09001397enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1398
1399static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001400{
Cyrill Gorcunov5479c782012-11-25 01:17:13 +04001401 if (!str)
1402 return -EINVAL;
1403
Tejun Heof58dc012009-08-14 15:00:50 +09001404 if (0)
1405 /* nada */;
1406#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1407 else if (!strcmp(str, "embed"))
1408 pcpu_chosen_fc = PCPU_FC_EMBED;
1409#endif
1410#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1411 else if (!strcmp(str, "page"))
1412 pcpu_chosen_fc = PCPU_FC_PAGE;
1413#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001414 else
1415 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001416
Tejun Heof58dc012009-08-14 15:00:50 +09001417 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001418}
Tejun Heof58dc012009-08-14 15:00:50 +09001419early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001420
Tejun Heo3c9a0242010-09-09 18:00:15 +02001421/*
1422 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1423 * Build it if needed by the arch config or the generic setup is going
1424 * to be used.
1425 */
Tejun Heo08fc4582009-08-14 15:00:49 +09001426#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1427 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02001428#define BUILD_EMBED_FIRST_CHUNK
1429#endif
1430
1431/* build pcpu_page_first_chunk() iff needed by the arch config */
1432#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1433#define BUILD_PAGE_FIRST_CHUNK
1434#endif
1435
1436/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1437#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1438/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09001439 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1440 * @reserved_size: the size of reserved percpu area in bytes
1441 * @dyn_size: minimum free size for dynamic allocation in bytes
1442 * @atom_size: allocation atom size
1443 * @cpu_distance_fn: callback to determine distance between cpus, optional
1444 *
1445 * This function determines grouping of units, their mappings to cpus
1446 * and other parameters considering needed percpu size, allocation
1447 * atom size and distances between CPUs.
1448 *
1449 * Groups are always mutliples of atom size and CPUs which are of
1450 * LOCAL_DISTANCE both ways are grouped together and share space for
1451 * units in the same group. The returned configuration is guaranteed
1452 * to have CPUs on different nodes on different groups and >=75% usage
1453 * of allocated virtual address space.
1454 *
1455 * RETURNS:
1456 * On success, pointer to the new allocation_info is returned. On
1457 * failure, ERR_PTR value is returned.
1458 */
1459static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1460 size_t reserved_size, size_t dyn_size,
1461 size_t atom_size,
1462 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1463{
1464 static int group_map[NR_CPUS] __initdata;
1465 static int group_cnt[NR_CPUS] __initdata;
1466 const size_t static_size = __per_cpu_end - __per_cpu_start;
1467 int nr_groups = 1, nr_units = 0;
1468 size_t size_sum, min_unit_size, alloc_size;
1469 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1470 int last_allocs, group, unit;
1471 unsigned int cpu, tcpu;
1472 struct pcpu_alloc_info *ai;
1473 unsigned int *cpu_map;
1474
1475 /* this function may be called multiple times */
1476 memset(group_map, 0, sizeof(group_map));
1477 memset(group_cnt, 0, sizeof(group_cnt));
1478
1479 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1480 size_sum = PFN_ALIGN(static_size + reserved_size +
1481 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1482 dyn_size = size_sum - static_size - reserved_size;
1483
1484 /*
1485 * Determine min_unit_size, alloc_size and max_upa such that
1486 * alloc_size is multiple of atom_size and is the smallest
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001487 * which can accommodate 4k aligned segments which are equal to
Tejun Heofbf59bc2009-02-20 16:29:08 +09001488 * or larger than min_unit_size.
1489 */
1490 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1491
1492 alloc_size = roundup(min_unit_size, atom_size);
1493 upa = alloc_size / min_unit_size;
1494 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1495 upa--;
1496 max_upa = upa;
1497
1498 /* group cpus according to their proximity */
1499 for_each_possible_cpu(cpu) {
1500 group = 0;
1501 next_group:
1502 for_each_possible_cpu(tcpu) {
1503 if (cpu == tcpu)
1504 break;
1505 if (group_map[tcpu] == group && cpu_distance_fn &&
1506 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1507 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1508 group++;
1509 nr_groups = max(nr_groups, group + 1);
1510 goto next_group;
1511 }
1512 }
1513 group_map[cpu] = group;
1514 group_cnt[group]++;
1515 }
1516
1517 /*
1518 * Expand unit size until address space usage goes over 75%
1519 * and then as much as possible without using more address
1520 * space.
1521 */
1522 last_allocs = INT_MAX;
1523 for (upa = max_upa; upa; upa--) {
1524 int allocs = 0, wasted = 0;
1525
1526 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1527 continue;
1528
1529 for (group = 0; group < nr_groups; group++) {
1530 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1531 allocs += this_allocs;
1532 wasted += this_allocs * upa - group_cnt[group];
1533 }
1534
1535 /*
1536 * Don't accept if wastage is over 1/3. The
1537 * greater-than comparison ensures upa==1 always
1538 * passes the following check.
1539 */
1540 if (wasted > num_possible_cpus() / 3)
1541 continue;
1542
1543 /* and then don't consume more memory */
1544 if (allocs > last_allocs)
1545 break;
1546 last_allocs = allocs;
1547 best_upa = upa;
1548 }
1549 upa = best_upa;
1550
1551 /* allocate and fill alloc_info */
1552 for (group = 0; group < nr_groups; group++)
1553 nr_units += roundup(group_cnt[group], upa);
1554
1555 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1556 if (!ai)
1557 return ERR_PTR(-ENOMEM);
1558 cpu_map = ai->groups[0].cpu_map;
1559
1560 for (group = 0; group < nr_groups; group++) {
1561 ai->groups[group].cpu_map = cpu_map;
1562 cpu_map += roundup(group_cnt[group], upa);
1563 }
1564
1565 ai->static_size = static_size;
1566 ai->reserved_size = reserved_size;
1567 ai->dyn_size = dyn_size;
1568 ai->unit_size = alloc_size / upa;
1569 ai->atom_size = atom_size;
1570 ai->alloc_size = alloc_size;
1571
1572 for (group = 0, unit = 0; group_cnt[group]; group++) {
1573 struct pcpu_group_info *gi = &ai->groups[group];
1574
1575 /*
1576 * Initialize base_offset as if all groups are located
1577 * back-to-back. The caller should update this to
1578 * reflect actual allocation.
1579 */
1580 gi->base_offset = unit * ai->unit_size;
1581
1582 for_each_possible_cpu(cpu)
1583 if (group_map[cpu] == group)
1584 gi->cpu_map[gi->nr_units++] = cpu;
1585 gi->nr_units = roundup(gi->nr_units, upa);
1586 unit += gi->nr_units;
1587 }
1588 BUG_ON(unit != nr_units);
1589
1590 return ai;
1591}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001592#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001593
Tejun Heo3c9a0242010-09-09 18:00:15 +02001594#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09001595/**
1596 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001597 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001598 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001599 * @atom_size: allocation atom size
1600 * @cpu_distance_fn: callback to determine distance between cpus, optional
1601 * @alloc_fn: function to allocate percpu page
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001602 * @free_fn: function to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001603 *
1604 * This is a helper to ease setting up embedded first percpu chunk and
1605 * can be called where pcpu_setup_first_chunk() is expected.
1606 *
1607 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001608 * by calling @alloc_fn and used as-is without being mapped into
1609 * vmalloc area. Allocations are always whole multiples of @atom_size
1610 * aligned to @atom_size.
1611 *
1612 * This enables the first chunk to piggy back on the linear physical
1613 * mapping which often uses larger page size. Please note that this
1614 * can result in very sparse cpu->unit mapping on NUMA machines thus
1615 * requiring large vmalloc address space. Don't use this allocator if
1616 * vmalloc space is not orders of magnitude larger than distances
1617 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001618 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001619 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001620 *
1621 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001622 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001623 *
1624 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001625 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001626 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001627int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001628 size_t atom_size,
1629 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1630 pcpu_fc_alloc_fn_t alloc_fn,
1631 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001632{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001633 void *base = (void *)ULONG_MAX;
1634 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001635 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001636 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001637 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001638
Tejun Heoc8826dd2009-08-14 15:00:52 +09001639 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1640 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001641 if (IS_ERR(ai))
1642 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001643
Tejun Heofd1e8a12009-08-14 15:00:51 +09001644 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001645 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001646
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001647 areas = memblock_virt_alloc_nopanic(areas_size, 0);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001648 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001649 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001650 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001651 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001652
Tejun Heoc8826dd2009-08-14 15:00:52 +09001653 /* allocate, copy and determine base address */
1654 for (group = 0; group < ai->nr_groups; group++) {
1655 struct pcpu_group_info *gi = &ai->groups[group];
1656 unsigned int cpu = NR_CPUS;
1657 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001658
Tejun Heoc8826dd2009-08-14 15:00:52 +09001659 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1660 cpu = gi->cpu_map[i];
1661 BUG_ON(cpu == NR_CPUS);
1662
1663 /* allocate space for the whole group */
1664 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1665 if (!ptr) {
1666 rc = -ENOMEM;
1667 goto out_free_areas;
1668 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001669 /* kmemleak tracks the percpu allocations separately */
1670 kmemleak_free(ptr);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001671 areas[group] = ptr;
1672
1673 base = min(ptr, base);
Tejun Heo42b64282012-04-27 08:42:53 -07001674 }
1675
1676 /*
1677 * Copy data and free unused parts. This should happen after all
1678 * allocations are complete; otherwise, we may end up with
1679 * overlapping groups.
1680 */
1681 for (group = 0; group < ai->nr_groups; group++) {
1682 struct pcpu_group_info *gi = &ai->groups[group];
1683 void *ptr = areas[group];
Tejun Heoc8826dd2009-08-14 15:00:52 +09001684
1685 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1686 if (gi->cpu_map[i] == NR_CPUS) {
1687 /* unused unit, free whole */
1688 free_fn(ptr, ai->unit_size);
1689 continue;
1690 }
1691 /* copy and return the unused part */
1692 memcpy(ptr, __per_cpu_load, ai->static_size);
1693 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1694 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001695 }
1696
Tejun Heoc8826dd2009-08-14 15:00:52 +09001697 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001698 max_distance = 0;
1699 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001700 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001701 max_distance = max_t(size_t, max_distance,
1702 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001703 }
1704 max_distance += ai->unit_size;
1705
1706 /* warn if maximum distance is further than 75% of vmalloc space */
Laura Abbott8a092172014-01-02 13:53:21 -08001707 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001708 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Mike Frysinger787e5b02011-03-23 08:23:52 +01001709 "space 0x%lx\n", max_distance,
Laura Abbott8a092172014-01-02 13:53:21 -08001710 VMALLOC_TOTAL);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001711#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1712 /* and fail if we have fallback */
1713 rc = -EINVAL;
1714 goto out_free;
1715#endif
1716 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001717
Tejun Heo004018e2009-08-14 15:00:49 +09001718 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001719 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1720 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001721
Tejun Heofb435d52009-08-14 15:00:51 +09001722 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001723 goto out_free;
1724
1725out_free_areas:
1726 for (group = 0; group < ai->nr_groups; group++)
Michael Holzheuf851c8d2013-09-17 16:57:34 +02001727 if (areas[group])
1728 free_fn(areas[group],
1729 ai->groups[group].nr_units * ai->unit_size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001730out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001731 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001732 if (areas)
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001733 memblock_free_early(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001734 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001735}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001736#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001737
Tejun Heo3c9a0242010-09-09 18:00:15 +02001738#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001739/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001740 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001741 * @reserved_size: the size of reserved percpu area in bytes
1742 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001743 * @free_fn: function to free percpu page, always called with PAGE_SIZE
Tejun Heod4b95f82009-07-04 08:10:59 +09001744 * @populate_pte_fn: function to populate pte
1745 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001746 * This is a helper to ease setting up page-remapped first percpu
1747 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001748 *
1749 * This is the basic allocator. Static percpu area is allocated
1750 * page-by-page into vmalloc area.
1751 *
1752 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001753 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001754 */
Tejun Heofb435d52009-08-14 15:00:51 +09001755int __init pcpu_page_first_chunk(size_t reserved_size,
1756 pcpu_fc_alloc_fn_t alloc_fn,
1757 pcpu_fc_free_fn_t free_fn,
1758 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001759{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001760 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001761 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001762 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001763 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001764 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001765 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001766 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001767
Tejun Heo00ae4062009-08-14 15:00:49 +09001768 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1769
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001770 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001771 if (IS_ERR(ai))
1772 return PTR_ERR(ai);
1773 BUG_ON(ai->nr_groups != 1);
1774 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1775
1776 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001777
1778 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001779 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1780 sizeof(pages[0]));
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001781 pages = memblock_virt_alloc(pages_size, 0);
Tejun Heod4b95f82009-07-04 08:10:59 +09001782
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001783 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001784 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001785 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001786 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001787 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001788 void *ptr;
1789
Tejun Heo3cbc8562009-08-14 15:00:50 +09001790 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001791 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001792 pr_warning("PERCPU: failed to allocate %s page "
1793 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001794 goto enomem;
1795 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001796 /* kmemleak tracks the percpu allocations separately */
1797 kmemleak_free(ptr);
Tejun Heoce3141a2009-07-04 08:11:00 +09001798 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001799 }
1800
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001801 /* allocate vm area, map the pages and copy static data */
1802 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001803 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001804 vm_area_register_early(&vm, PAGE_SIZE);
1805
Tejun Heofd1e8a12009-08-14 15:00:51 +09001806 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001807 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001808 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001809
Tejun Heoce3141a2009-07-04 08:11:00 +09001810 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001811 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1812
1813 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001814 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1815 unit_pages);
1816 if (rc < 0)
1817 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001818
1819 /*
1820 * FIXME: Archs with virtual cache should flush local
1821 * cache for the linear mapping here - something
1822 * equivalent to flush_cache_vmap() on the local cpu.
1823 * flush_cache_vmap() can't be used as most supporting
1824 * data structures are not set up yet.
1825 */
1826
1827 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001828 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001829 }
1830
1831 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09001832 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001833 unit_pages, psize_str, vm.addr, ai->static_size,
1834 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001835
Tejun Heofb435d52009-08-14 15:00:51 +09001836 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001837 goto out_free_ar;
1838
1839enomem:
1840 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09001841 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09001842 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09001843out_free_ar:
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001844 memblock_free_early(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001845 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09001846 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001847}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001848#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001849
Tejun Heobbddff02010-09-03 18:22:48 +02001850#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09001851/*
Tejun Heobbddff02010-09-03 18:22:48 +02001852 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09001853 *
1854 * The embedding helper is used because its behavior closely resembles
1855 * the original non-dynamic generic percpu area setup. This is
1856 * important because many archs have addressing restrictions and might
1857 * fail if the percpu area is located far away from the previous
1858 * location. As an added bonus, in non-NUMA cases, embedding is
1859 * generally a good idea TLB-wise because percpu area can piggy back
1860 * on the physical linear memory mapping which uses large page
1861 * mappings on applicable archs.
1862 */
Tejun Heoe74e3962009-03-30 19:07:44 +09001863unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1864EXPORT_SYMBOL(__per_cpu_offset);
1865
Tejun Heoc8826dd2009-08-14 15:00:52 +09001866static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1867 size_t align)
1868{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001869 return memblock_virt_alloc_from_nopanic(
1870 size, align, __pa(MAX_DMA_ADDRESS));
Tejun Heoc8826dd2009-08-14 15:00:52 +09001871}
1872
1873static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1874{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001875 memblock_free_early(__pa(ptr), size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001876}
1877
Tejun Heoe74e3962009-03-30 19:07:44 +09001878void __init setup_per_cpu_areas(void)
1879{
Tejun Heoe74e3962009-03-30 19:07:44 +09001880 unsigned long delta;
1881 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09001882 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09001883
1884 /*
1885 * Always reserve area for module percpu variables. That's
1886 * what the legacy allocator did.
1887 */
Tejun Heofb435d52009-08-14 15:00:51 +09001888 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001889 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1890 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09001891 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02001892 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09001893
1894 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1895 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09001896 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09001897}
Tejun Heobbddff02010-09-03 18:22:48 +02001898#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1899
1900#else /* CONFIG_SMP */
1901
1902/*
1903 * UP percpu area setup.
1904 *
1905 * UP always uses km-based percpu allocator with identity mapping.
1906 * Static percpu variables are indistinguishable from the usual static
1907 * variables and don't require any special preparation.
1908 */
1909void __init setup_per_cpu_areas(void)
1910{
1911 const size_t unit_size =
1912 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1913 PERCPU_DYNAMIC_RESERVE));
1914 struct pcpu_alloc_info *ai;
1915 void *fc;
1916
1917 ai = pcpu_alloc_alloc_info(1, 1);
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001918 fc = memblock_virt_alloc_from_nopanic(unit_size,
1919 PAGE_SIZE,
1920 __pa(MAX_DMA_ADDRESS));
Tejun Heobbddff02010-09-03 18:22:48 +02001921 if (!ai || !fc)
1922 panic("Failed to allocate memory for percpu areas.");
Catalin Marinas100d13c2012-05-09 16:55:19 +01001923 /* kmemleak tracks the percpu allocations separately */
1924 kmemleak_free(fc);
Tejun Heobbddff02010-09-03 18:22:48 +02001925
1926 ai->dyn_size = unit_size;
1927 ai->unit_size = unit_size;
1928 ai->atom_size = unit_size;
1929 ai->alloc_size = unit_size;
1930 ai->groups[0].nr_units = 1;
1931 ai->groups[0].cpu_map[0] = 0;
1932
1933 if (pcpu_setup_first_chunk(ai, fc) < 0)
1934 panic("Failed to initialize percpu areas.");
1935}
1936
1937#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02001938
1939/*
1940 * First and reserved chunks are initialized with temporary allocation
1941 * map in initdata so that they can be used before slab is online.
1942 * This function is called after slab is brought up and replaces those
1943 * with properly allocated maps.
1944 */
1945void __init percpu_init_late(void)
1946{
1947 struct pcpu_chunk *target_chunks[] =
1948 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1949 struct pcpu_chunk *chunk;
1950 unsigned long flags;
1951 int i;
1952
1953 for (i = 0; (chunk = target_chunks[i]); i++) {
1954 int *map;
1955 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1956
1957 BUILD_BUG_ON(size > PAGE_SIZE);
1958
Bob Liu90459ce02011-08-04 11:02:33 +02001959 map = pcpu_mem_zalloc(size);
Tejun Heo099a19d2010-06-27 18:50:00 +02001960 BUG_ON(!map);
1961
1962 spin_lock_irqsave(&pcpu_lock, flags);
1963 memcpy(map, chunk->map, size);
1964 chunk->map = map;
1965 spin_unlock_irqrestore(&pcpu_lock, flags);
1966 }
1967}