blob: d90797160c2a7fff7be950b04ac9817292208ee0 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
2 * linux/mm/percpu.c - percpu memory allocator
3 *
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
10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each
Tejun Heo2f39e632009-07-04 08:11:00 +090011 * chunk is consisted of boot-time determined number of units and the
12 * first chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated. ie. in
17 * vmalloc area
Tejun Heofbf59bc2009-02-20 16:29:08 +090018 *
19 * c0 c1 c2
20 * ------------------- ------------------- ------------
21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
22 * ------------------- ...... ------------------- .... ------------
23 *
24 * Allocation is done in offset-size areas of single unit space. Ie,
25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
Tejun Heo2f39e632009-07-04 08:11:00 +090026 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090030 *
Tejun Heo2f39e632009-07-04 08:11:00 +090031 * There are usually many small percpu allocations many of them being
32 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090033 * according to free size and tries to allocate from the fullest one.
34 * Each chunk keeps the maximum contiguous area size hint which is
35 * guaranteed to be eqaul to or larger than the maximum contiguous
36 * area in the chunk. This helps the allocator not to iterate the
37 * chunk maps unnecessarily.
38 *
39 * Allocation state in each chunk is kept using an array of integers
40 * on chunk->map. A positive value in the map represents a free
41 * region and negative allocated. Allocation inside a chunk is done
42 * by scanning this map sequentially and serving the first matching
43 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090044 * Chunks can be determined from the address using the index field
45 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090046 *
47 * To use this allocator, arch code should do the followings.
48 *
Tejun Heoe74e3962009-03-30 19:07:44 +090049 * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA
Tejun Heofbf59bc2009-02-20 16:29:08 +090050 *
51 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090052 * regular address to percpu pointer and back if they need to be
53 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 *
Tejun Heo8d408b42009-02-24 11:57:21 +090055 * - use pcpu_setup_first_chunk() during percpu area initialization to
56 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090057 */
58
59#include <linux/bitmap.h>
60#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090061#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090062#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090063#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090064#include <linux/mm.h>
65#include <linux/module.h>
66#include <linux/mutex.h>
67#include <linux/percpu.h>
68#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090069#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090070#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090072#include <linux/workqueue.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090073
74#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090075#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090076#include <asm/tlbflush.h>
77
Tejun Heofbf59bc2009-02-20 16:29:08 +090078#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
79#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
80
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) \
84 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
85 + (unsigned long)__per_cpu_start)
86#endif
87#ifndef __pcpu_ptr_to_addr
88#define __pcpu_ptr_to_addr(ptr) \
89 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
90 - (unsigned long)__per_cpu_start)
91#endif
92
Tejun Heofbf59bc2009-02-20 16:29:08 +090093struct pcpu_chunk {
94 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +090095 int free_size; /* free bytes in the chunk */
96 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +090097 void *base_addr; /* base address of this chunk */
Tejun Heofbf59bc2009-02-20 16:29:08 +090098 int map_used; /* # of map entries used */
99 int map_alloc; /* # of map entries allocated */
100 int *map; /* allocation map */
Tejun Heo65632972009-08-14 15:00:52 +0900101 struct vm_struct **vms; /* mapped vmalloc regions */
Tejun Heo8d408b42009-02-24 11:57:21 +0900102 bool immutable; /* no [de]population allowed */
Tejun Heoce3141a2009-07-04 08:11:00 +0900103 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900104};
105
Tejun Heo40150d32009-02-24 12:32:28 +0900106static int pcpu_unit_pages __read_mostly;
107static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900108static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900109static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900110static int pcpu_nr_slots __read_mostly;
111static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900112
Tejun Heo2f39e632009-07-04 08:11:00 +0900113/* cpus with the lowest and highest unit numbers */
114static unsigned int pcpu_first_unit_cpu __read_mostly;
115static unsigned int pcpu_last_unit_cpu __read_mostly;
116
Tejun Heofbf59bc2009-02-20 16:29:08 +0900117/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900118void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900119EXPORT_SYMBOL_GPL(pcpu_base_addr);
120
Tejun Heofb435d52009-08-14 15:00:51 +0900121static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
122const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900123
Tejun Heo65632972009-08-14 15:00:52 +0900124/* group information, used for vm allocation */
125static int pcpu_nr_groups __read_mostly;
126static const unsigned long *pcpu_group_offsets __read_mostly;
127static const size_t *pcpu_group_sizes __read_mostly;
128
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900129/*
130 * The first chunk which always exists. Note that unlike other
131 * chunks, this one can be allocated and mapped in several different
132 * ways and thus often doesn't live in the vmalloc area.
133 */
134static struct pcpu_chunk *pcpu_first_chunk;
135
136/*
137 * Optional reserved chunk. This chunk reserves part of the first
138 * chunk and serves it for reserved allocations. The amount of
139 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
140 * area doesn't exist, the following variables contain NULL and 0
141 * respectively.
142 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900143static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900144static int pcpu_reserved_chunk_limit;
145
Tejun Heofbf59bc2009-02-20 16:29:08 +0900146/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900147 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900148 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900149 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
Tejun Heoce3141a2009-07-04 08:11:00 +0900150 * protects allocation/reclaim paths, chunks, populated bitmap and
151 * vmalloc mapping. The latter is a spinlock and protects the index
152 * data structures - chunk slots, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900153 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900154 * During allocation, pcpu_alloc_mutex is kept locked all the time and
155 * pcpu_lock is grabbed and released as necessary. All actual memory
Jiri Kosina403a91b2009-10-29 00:25:59 +0900156 * allocations are done using GFP_KERNEL with pcpu_lock released. In
157 * general, percpu memory can't be allocated with irq off but
158 * irqsave/restore are still used in alloc path so that it can be used
159 * from early init path - sched_init() specifically.
Tejun Heoccea34b2009-03-07 00:44:13 +0900160 *
161 * Free path accesses and alters only the index data structures, so it
162 * can be safely called from atomic context. When memory needs to be
163 * returned to the system, free path schedules reclaim_work which
164 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
165 * reclaimed, release both locks and frees the chunks. Note that it's
166 * necessary to grab both locks to remove a chunk from circulation as
167 * allocation path might be referencing the chunk with only
168 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900169 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900170static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
171static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900172
Tejun Heo40150d32009-02-24 12:32:28 +0900173static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900174
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900175/* reclaim work to release fully free chunks, scheduled from free path */
176static void pcpu_reclaim(struct work_struct *work);
177static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
178
Tejun Heod9b55ee2009-02-24 11:57:21 +0900179static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900180{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900181 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900182 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
183}
184
Tejun Heod9b55ee2009-02-24 11:57:21 +0900185static int pcpu_size_to_slot(int size)
186{
187 if (size == pcpu_unit_size)
188 return pcpu_nr_slots - 1;
189 return __pcpu_size_to_slot(size);
190}
191
Tejun Heofbf59bc2009-02-20 16:29:08 +0900192static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
193{
194 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
195 return 0;
196
197 return pcpu_size_to_slot(chunk->free_size);
198}
199
200static int pcpu_page_idx(unsigned int cpu, int page_idx)
201{
Tejun Heo2f39e632009-07-04 08:11:00 +0900202 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900203}
204
205static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
206 unsigned int cpu, int page_idx)
207{
Tejun Heobba174f2009-08-14 15:00:51 +0900208 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900209 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900210}
211
Tejun Heoce3141a2009-07-04 08:11:00 +0900212static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
213 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900214{
Tejun Heoce3141a2009-07-04 08:11:00 +0900215 /* must not be used on pre-mapped chunk */
216 WARN_ON(chunk->immutable);
Tejun Heoc8a51be2009-07-04 08:10:59 +0900217
Tejun Heoce3141a2009-07-04 08:11:00 +0900218 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900219}
220
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900221/* set the pointer to a chunk in a page struct */
222static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
223{
224 page->index = (unsigned long)pcpu;
225}
226
227/* obtain pointer to a chunk from a page struct */
228static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
229{
230 return (struct pcpu_chunk *)page->index;
231}
232
Tejun Heoce3141a2009-07-04 08:11:00 +0900233static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
234{
235 *rs = find_next_zero_bit(chunk->populated, end, *rs);
236 *re = find_next_bit(chunk->populated, end, *rs + 1);
237}
238
239static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
240{
241 *rs = find_next_bit(chunk->populated, end, *rs);
242 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
243}
244
245/*
246 * (Un)populated page region iterators. Iterate over (un)populated
247 * page regions betwen @start and @end in @chunk. @rs and @re should
248 * be integer variables and will be set to start and end page index of
249 * the current region.
250 */
251#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
252 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
253 (rs) < (re); \
254 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
255
256#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
257 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
258 (rs) < (re); \
259 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
260
Tejun Heofbf59bc2009-02-20 16:29:08 +0900261/**
Tejun Heo1880d932009-03-07 00:44:09 +0900262 * pcpu_mem_alloc - allocate memory
263 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900264 *
Tejun Heo1880d932009-03-07 00:44:09 +0900265 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
266 * kzalloc() is used; otherwise, vmalloc() is used. The returned
267 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900268 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900269 * CONTEXT:
270 * Does GFP_KERNEL allocation.
271 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900272 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900273 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900274 */
Tejun Heo1880d932009-03-07 00:44:09 +0900275static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900276{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900277 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900278 return kzalloc(size, GFP_KERNEL);
279 else {
280 void *ptr = vmalloc(size);
281 if (ptr)
282 memset(ptr, 0, size);
283 return ptr;
284 }
285}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900286
Tejun Heo1880d932009-03-07 00:44:09 +0900287/**
288 * pcpu_mem_free - free memory
289 * @ptr: memory to free
290 * @size: size of the area
291 *
292 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
293 */
294static void pcpu_mem_free(void *ptr, size_t size)
295{
296 if (size <= PAGE_SIZE)
297 kfree(ptr);
298 else
299 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900300}
301
302/**
303 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
304 * @chunk: chunk of interest
305 * @oslot: the previous slot it was on
306 *
307 * This function is called after an allocation or free changed @chunk.
308 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900309 * moved to the slot. Note that the reserved chunk is never put on
310 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900311 *
312 * CONTEXT:
313 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900314 */
315static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
316{
317 int nslot = pcpu_chunk_slot(chunk);
318
Tejun Heoedcb4632009-03-06 14:33:59 +0900319 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900320 if (oslot < nslot)
321 list_move(&chunk->list, &pcpu_slot[nslot]);
322 else
323 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
324 }
325}
326
Tejun Heofbf59bc2009-02-20 16:29:08 +0900327/**
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900328 * pcpu_chunk_addr_search - determine chunk containing specified address
329 * @addr: address for which the chunk needs to be determined.
Tejun Heoccea34b2009-03-07 00:44:13 +0900330 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900331 * RETURNS:
332 * The address of the found chunk.
333 */
334static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
335{
Tejun Heobba174f2009-08-14 15:00:51 +0900336 void *first_start = pcpu_first_chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900337
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900338 /* is it in the first chunk? */
Tejun Heo79ba6ac2009-07-04 08:10:58 +0900339 if (addr >= first_start && addr < first_start + pcpu_unit_size) {
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900340 /* is it in the reserved area? */
341 if (addr < first_start + pcpu_reserved_chunk_limit)
Tejun Heoedcb4632009-03-06 14:33:59 +0900342 return pcpu_reserved_chunk;
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900343 return pcpu_first_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900344 }
345
Tejun Heo04a13c72009-09-01 21:12:28 +0900346 /*
347 * The address is relative to unit0 which might be unused and
348 * thus unmapped. Offset the address to the unit space of the
349 * current processor before looking it up in the vmalloc
350 * space. Note that any possible cpu id can be used here, so
351 * there's no need to worry about preemption or cpu hotplug.
352 */
Tejun Heo5579fd72009-09-15 09:57:19 +0900353 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900354 return pcpu_get_page_chunk(vmalloc_to_page(addr));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900355}
356
357/**
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900358 * pcpu_extend_area_map - extend area map for allocation
359 * @chunk: target chunk
360 *
361 * Extend area map of @chunk so that it can accomodate an allocation.
362 * A single allocation can split an area into three areas, so this
363 * function makes sure that @chunk->map has at least two extra slots.
364 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900365 * CONTEXT:
366 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
367 * if area map is extended.
368 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900369 * RETURNS:
370 * 0 if noop, 1 if successfully extended, -errno on failure.
371 */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900372static int pcpu_extend_area_map(struct pcpu_chunk *chunk, unsigned long *flags)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900373{
374 int new_alloc;
375 int *new;
376 size_t size;
377
378 /* has enough? */
379 if (chunk->map_alloc >= chunk->map_used + 2)
380 return 0;
381
Jiri Kosina403a91b2009-10-29 00:25:59 +0900382 spin_unlock_irqrestore(&pcpu_lock, *flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900383
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900384 new_alloc = PCPU_DFL_MAP_ALLOC;
385 while (new_alloc < chunk->map_used + 2)
386 new_alloc *= 2;
387
388 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
Tejun Heoccea34b2009-03-07 00:44:13 +0900389 if (!new) {
Jiri Kosina403a91b2009-10-29 00:25:59 +0900390 spin_lock_irqsave(&pcpu_lock, *flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900391 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900392 }
393
394 /*
395 * Acquire pcpu_lock and switch to new area map. Only free
396 * could have happened inbetween, so map_used couldn't have
397 * grown.
398 */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900399 spin_lock_irqsave(&pcpu_lock, *flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900400 BUG_ON(new_alloc < chunk->map_used + 2);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900401
402 size = chunk->map_alloc * sizeof(chunk->map[0]);
403 memcpy(new, chunk->map, size);
404
405 /*
406 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
407 * one of the first chunks and still using static map.
408 */
409 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
410 pcpu_mem_free(chunk->map, size);
411
412 chunk->map_alloc = new_alloc;
413 chunk->map = new;
414 return 0;
415}
416
417/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900418 * pcpu_split_block - split a map block
419 * @chunk: chunk of interest
420 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900421 * @head: head size in bytes (can be 0)
422 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900423 *
424 * Split the @i'th map block into two or three blocks. If @head is
425 * non-zero, @head bytes block is inserted before block @i moving it
426 * to @i+1 and reducing its size by @head bytes.
427 *
428 * If @tail is non-zero, the target block, which can be @i or @i+1
429 * depending on @head, is reduced by @tail bytes and @tail byte block
430 * is inserted after the target block.
431 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900432 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900433 *
434 * CONTEXT:
435 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900436 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900437static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
438 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900439{
440 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900441
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900442 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900443
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900444 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900445 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
446 sizeof(chunk->map[0]) * (chunk->map_used - i));
447 chunk->map_used += nr_extra;
448
449 if (head) {
450 chunk->map[i + 1] = chunk->map[i] - head;
451 chunk->map[i++] = head;
452 }
453 if (tail) {
454 chunk->map[i++] -= tail;
455 chunk->map[i] = tail;
456 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900457}
458
459/**
460 * pcpu_alloc_area - allocate area from a pcpu_chunk
461 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900462 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900463 * @align: wanted align
464 *
465 * Try to allocate @size bytes area aligned at @align from @chunk.
466 * Note that this function only allocates the offset. It doesn't
467 * populate or map the area.
468 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900469 * @chunk->map must have at least two free slots.
470 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900471 * CONTEXT:
472 * pcpu_lock.
473 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900474 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900475 * Allocated offset in @chunk on success, -1 if no matching area is
476 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900477 */
478static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
479{
480 int oslot = pcpu_chunk_slot(chunk);
481 int max_contig = 0;
482 int i, off;
483
Tejun Heofbf59bc2009-02-20 16:29:08 +0900484 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
485 bool is_last = i + 1 == chunk->map_used;
486 int head, tail;
487
488 /* extra for alignment requirement */
489 head = ALIGN(off, align) - off;
490 BUG_ON(i == 0 && head != 0);
491
492 if (chunk->map[i] < 0)
493 continue;
494 if (chunk->map[i] < head + size) {
495 max_contig = max(chunk->map[i], max_contig);
496 continue;
497 }
498
499 /*
500 * If head is small or the previous block is free,
501 * merge'em. Note that 'small' is defined as smaller
502 * than sizeof(int), which is very small but isn't too
503 * uncommon for percpu allocations.
504 */
505 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
506 if (chunk->map[i - 1] > 0)
507 chunk->map[i - 1] += head;
508 else {
509 chunk->map[i - 1] -= head;
510 chunk->free_size -= head;
511 }
512 chunk->map[i] -= head;
513 off += head;
514 head = 0;
515 }
516
517 /* if tail is small, just keep it around */
518 tail = chunk->map[i] - head - size;
519 if (tail < sizeof(int))
520 tail = 0;
521
522 /* split if warranted */
523 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900524 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900525 if (head) {
526 i++;
527 off += head;
528 max_contig = max(chunk->map[i - 1], max_contig);
529 }
530 if (tail)
531 max_contig = max(chunk->map[i + 1], max_contig);
532 }
533
534 /* update hint and mark allocated */
535 if (is_last)
536 chunk->contig_hint = max_contig; /* fully scanned */
537 else
538 chunk->contig_hint = max(chunk->contig_hint,
539 max_contig);
540
541 chunk->free_size -= chunk->map[i];
542 chunk->map[i] = -chunk->map[i];
543
544 pcpu_chunk_relocate(chunk, oslot);
545 return off;
546 }
547
548 chunk->contig_hint = max_contig; /* fully scanned */
549 pcpu_chunk_relocate(chunk, oslot);
550
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900551 /* tell the upper layer that this chunk has no matching area */
552 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900553}
554
555/**
556 * pcpu_free_area - free area to a pcpu_chunk
557 * @chunk: chunk of interest
558 * @freeme: offset of area to free
559 *
560 * Free area starting from @freeme to @chunk. Note that this function
561 * only modifies the allocation map. It doesn't depopulate or unmap
562 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900563 *
564 * CONTEXT:
565 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900566 */
567static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
568{
569 int oslot = pcpu_chunk_slot(chunk);
570 int i, off;
571
572 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
573 if (off == freeme)
574 break;
575 BUG_ON(off != freeme);
576 BUG_ON(chunk->map[i] > 0);
577
578 chunk->map[i] = -chunk->map[i];
579 chunk->free_size += chunk->map[i];
580
581 /* merge with previous? */
582 if (i > 0 && chunk->map[i - 1] >= 0) {
583 chunk->map[i - 1] += chunk->map[i];
584 chunk->map_used--;
585 memmove(&chunk->map[i], &chunk->map[i + 1],
586 (chunk->map_used - i) * sizeof(chunk->map[0]));
587 i--;
588 }
589 /* merge with next? */
590 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
591 chunk->map[i] += chunk->map[i + 1];
592 chunk->map_used--;
593 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
594 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
595 }
596
597 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
598 pcpu_chunk_relocate(chunk, oslot);
599}
600
601/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900602 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900603 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900604 * @bitmapp: output parameter for bitmap
605 * @may_alloc: may allocate the array
Tejun Heofbf59bc2009-02-20 16:29:08 +0900606 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900607 * Returns pointer to array of pointers to struct page and bitmap,
608 * both of which can be indexed with pcpu_page_idx(). The returned
609 * array is cleared to zero and *@bitmapp is copied from
610 * @chunk->populated. Note that there is only one array and bitmap
611 * and access exclusion is the caller's responsibility.
612 *
613 * CONTEXT:
614 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
615 * Otherwise, don't care.
616 *
617 * RETURNS:
618 * Pointer to temp pages array on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900619 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900620static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
621 unsigned long **bitmapp,
622 bool may_alloc)
623{
624 static struct page **pages;
625 static unsigned long *bitmap;
Tejun Heo2f39e632009-07-04 08:11:00 +0900626 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
Tejun Heoce3141a2009-07-04 08:11:00 +0900627 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
628 sizeof(unsigned long);
629
630 if (!pages || !bitmap) {
631 if (may_alloc && !pages)
632 pages = pcpu_mem_alloc(pages_size);
633 if (may_alloc && !bitmap)
634 bitmap = pcpu_mem_alloc(bitmap_size);
635 if (!pages || !bitmap)
636 return NULL;
637 }
638
639 memset(pages, 0, pages_size);
640 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
641
642 *bitmapp = bitmap;
643 return pages;
644}
645
646/**
647 * pcpu_free_pages - free pages which were allocated for @chunk
648 * @chunk: chunk pages were allocated for
649 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
650 * @populated: populated bitmap
651 * @page_start: page index of the first page to be freed
652 * @page_end: page index of the last page to be freed + 1
653 *
654 * Free pages [@page_start and @page_end) in @pages for all units.
655 * The pages were allocated for @chunk.
656 */
657static void pcpu_free_pages(struct pcpu_chunk *chunk,
658 struct page **pages, unsigned long *populated,
659 int page_start, int page_end)
660{
661 unsigned int cpu;
662 int i;
663
664 for_each_possible_cpu(cpu) {
665 for (i = page_start; i < page_end; i++) {
666 struct page *page = pages[pcpu_page_idx(cpu, i)];
667
668 if (page)
669 __free_page(page);
670 }
671 }
672}
673
674/**
675 * pcpu_alloc_pages - allocates pages for @chunk
676 * @chunk: target chunk
677 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
678 * @populated: populated bitmap
679 * @page_start: page index of the first page to be allocated
680 * @page_end: page index of the last page to be allocated + 1
681 *
682 * Allocate pages [@page_start,@page_end) into @pages for all units.
683 * The allocation is for @chunk. Percpu core doesn't care about the
684 * content of @pages and will pass it verbatim to pcpu_map_pages().
685 */
686static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
687 struct page **pages, unsigned long *populated,
688 int page_start, int page_end)
689{
690 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
691 unsigned int cpu;
692 int i;
693
694 for_each_possible_cpu(cpu) {
695 for (i = page_start; i < page_end; i++) {
696 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
697
698 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
699 if (!*pagep) {
700 pcpu_free_pages(chunk, pages, populated,
701 page_start, page_end);
702 return -ENOMEM;
703 }
704 }
705 }
706 return 0;
707}
708
709/**
710 * pcpu_pre_unmap_flush - flush cache prior to unmapping
711 * @chunk: chunk the regions to be flushed belongs to
712 * @page_start: page index of the first page to be flushed
713 * @page_end: page index of the last page to be flushed + 1
714 *
715 * Pages in [@page_start,@page_end) of @chunk are about to be
716 * unmapped. Flush cache. As each flushing trial can be very
717 * expensive, issue flush on the whole region at once rather than
718 * doing it for each cpu. This could be an overkill but is more
719 * scalable.
720 */
721static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
722 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900723{
Tejun Heo2f39e632009-07-04 08:11:00 +0900724 flush_cache_vunmap(
725 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
726 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heoce3141a2009-07-04 08:11:00 +0900727}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900728
Tejun Heoce3141a2009-07-04 08:11:00 +0900729static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
730{
731 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
732}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900733
Tejun Heoce3141a2009-07-04 08:11:00 +0900734/**
735 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
736 * @chunk: chunk of interest
737 * @pages: pages array which can be used to pass information to free
738 * @populated: populated bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900739 * @page_start: page index of the first page to unmap
740 * @page_end: page index of the last page to unmap + 1
Tejun Heofbf59bc2009-02-20 16:29:08 +0900741 *
742 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
Tejun Heoce3141a2009-07-04 08:11:00 +0900743 * Corresponding elements in @pages were cleared by the caller and can
744 * be used to carry information to pcpu_free_pages() which will be
745 * called after all unmaps are finished. The caller should call
746 * proper pre/post flush functions.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900747 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900748static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
749 struct page **pages, unsigned long *populated,
750 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900751{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900752 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900753 int i;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900754
Tejun Heoce3141a2009-07-04 08:11:00 +0900755 for_each_possible_cpu(cpu) {
756 for (i = page_start; i < page_end; i++) {
757 struct page *page;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900758
Tejun Heoce3141a2009-07-04 08:11:00 +0900759 page = pcpu_chunk_page(chunk, cpu, i);
760 WARN_ON(!page);
761 pages[pcpu_page_idx(cpu, i)] = page;
762 }
763 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
764 page_end - page_start);
765 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900766
Tejun Heoce3141a2009-07-04 08:11:00 +0900767 for (i = page_start; i < page_end; i++)
768 __clear_bit(i, populated);
769}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900770
Tejun Heoce3141a2009-07-04 08:11:00 +0900771/**
772 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
773 * @chunk: pcpu_chunk the regions to be flushed belong to
774 * @page_start: page index of the first page to be flushed
775 * @page_end: page index of the last page to be flushed + 1
776 *
777 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
778 * TLB for the regions. This can be skipped if the area is to be
779 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
780 *
781 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
782 * for the whole region.
783 */
784static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
785 int page_start, int page_end)
786{
Tejun Heo2f39e632009-07-04 08:11:00 +0900787 flush_tlb_kernel_range(
788 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
789 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900790}
791
Tejun Heoc8a51be2009-07-04 08:10:59 +0900792static int __pcpu_map_pages(unsigned long addr, struct page **pages,
793 int nr_pages)
794{
795 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
796 PAGE_KERNEL, pages);
797}
798
799/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900800 * pcpu_map_pages - map pages into a pcpu_chunk
Tejun Heoc8a51be2009-07-04 08:10:59 +0900801 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900802 * @pages: pages array containing pages to be mapped
803 * @populated: populated bitmap
Tejun Heoc8a51be2009-07-04 08:10:59 +0900804 * @page_start: page index of the first page to map
805 * @page_end: page index of the last page to map + 1
806 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900807 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
808 * caller is responsible for calling pcpu_post_map_flush() after all
809 * mappings are complete.
810 *
811 * This function is responsible for setting corresponding bits in
812 * @chunk->populated bitmap and whatever is necessary for reverse
813 * lookup (addr -> chunk).
Tejun Heoc8a51be2009-07-04 08:10:59 +0900814 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900815static int pcpu_map_pages(struct pcpu_chunk *chunk,
816 struct page **pages, unsigned long *populated,
817 int page_start, int page_end)
Tejun Heoc8a51be2009-07-04 08:10:59 +0900818{
Tejun Heoce3141a2009-07-04 08:11:00 +0900819 unsigned int cpu, tcpu;
820 int i, err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900821
822 for_each_possible_cpu(cpu) {
823 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
Tejun Heoce3141a2009-07-04 08:11:00 +0900824 &pages[pcpu_page_idx(cpu, page_start)],
Tejun Heoc8a51be2009-07-04 08:10:59 +0900825 page_end - page_start);
826 if (err < 0)
Tejun Heoce3141a2009-07-04 08:11:00 +0900827 goto err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900828 }
829
Tejun Heoce3141a2009-07-04 08:11:00 +0900830 /* mapping successful, link chunk and mark populated */
831 for (i = page_start; i < page_end; i++) {
832 for_each_possible_cpu(cpu)
833 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
834 chunk);
835 __set_bit(i, populated);
836 }
837
838 return 0;
839
840err:
841 for_each_possible_cpu(tcpu) {
842 if (tcpu == cpu)
843 break;
844 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
845 page_end - page_start);
846 }
847 return err;
848}
849
850/**
851 * pcpu_post_map_flush - flush cache after mapping
852 * @chunk: pcpu_chunk the regions to be flushed belong to
853 * @page_start: page index of the first page to be flushed
854 * @page_end: page index of the last page to be flushed + 1
855 *
856 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
857 * cache.
858 *
859 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
860 * for the whole region.
861 */
862static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
863 int page_start, int page_end)
864{
Tejun Heo2f39e632009-07-04 08:11:00 +0900865 flush_cache_vmap(
866 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
867 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900868}
869
870/**
871 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
872 * @chunk: chunk to depopulate
873 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900874 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900875 * @flush: whether to flush cache and tlb or not
876 *
877 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
878 * from @chunk. If @flush is true, vcache is flushed before unmapping
879 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900880 *
881 * CONTEXT:
882 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900883 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900884static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900885{
886 int page_start = PFN_DOWN(off);
887 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900888 struct page **pages;
889 unsigned long *populated;
890 int rs, re;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900891
Tejun Heoce3141a2009-07-04 08:11:00 +0900892 /* quick path, check whether it's empty already */
893 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
894 if (rs == page_start && re == page_end)
895 return;
896 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900897 }
898
Tejun Heoce3141a2009-07-04 08:11:00 +0900899 /* immutable chunks can't be depopulated */
Tejun Heo8d408b42009-02-24 11:57:21 +0900900 WARN_ON(chunk->immutable);
901
Tejun Heoce3141a2009-07-04 08:11:00 +0900902 /*
903 * If control reaches here, there must have been at least one
904 * successful population attempt so the temp pages array must
905 * be available now.
906 */
907 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
908 BUG_ON(!pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900909
Tejun Heoce3141a2009-07-04 08:11:00 +0900910 /* unmap and free */
911 pcpu_pre_unmap_flush(chunk, page_start, page_end);
912
913 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
914 pcpu_unmap_pages(chunk, pages, populated, rs, re);
915
916 /* no need to flush tlb, vmalloc will handle it lazily */
917
918 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
919 pcpu_free_pages(chunk, pages, populated, rs, re);
920
921 /* commit new bitmap */
922 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900923}
924
925/**
926 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
927 * @chunk: chunk of interest
928 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900929 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900930 *
931 * For each cpu, populate and map pages [@page_start,@page_end) into
932 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900933 *
934 * CONTEXT:
935 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900936 */
937static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
938{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900939 int page_start = PFN_DOWN(off);
940 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900941 int free_end = page_start, unmap_end = page_start;
942 struct page **pages;
943 unsigned long *populated;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900944 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900945 int rs, re, rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900946
Tejun Heoce3141a2009-07-04 08:11:00 +0900947 /* quick path, check whether all pages are already there */
948 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
949 if (rs == page_start && re == page_end)
950 goto clear;
951 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900952 }
953
Tejun Heoce3141a2009-07-04 08:11:00 +0900954 /* need to allocate and map pages, this chunk can't be immutable */
955 WARN_ON(chunk->immutable);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900956
Tejun Heoce3141a2009-07-04 08:11:00 +0900957 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
958 if (!pages)
959 return -ENOMEM;
960
961 /* alloc and map */
962 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
963 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
964 if (rc)
965 goto err_free;
966 free_end = re;
967 }
968
969 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
970 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
971 if (rc)
972 goto err_unmap;
973 unmap_end = re;
974 }
975 pcpu_post_map_flush(chunk, page_start, page_end);
976
977 /* commit new bitmap */
978 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
979clear:
Tejun Heofbf59bc2009-02-20 16:29:08 +0900980 for_each_possible_cpu(cpu)
Tejun Heo2f39e632009-07-04 08:11:00 +0900981 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900982 return 0;
Tejun Heoce3141a2009-07-04 08:11:00 +0900983
984err_unmap:
985 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
986 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
987 pcpu_unmap_pages(chunk, pages, populated, rs, re);
988 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
989err_free:
990 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
991 pcpu_free_pages(chunk, pages, populated, rs, re);
992 return rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900993}
994
995static void free_pcpu_chunk(struct pcpu_chunk *chunk)
996{
997 if (!chunk)
998 return;
Tejun Heo65632972009-08-14 15:00:52 +0900999 if (chunk->vms)
1000 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
Tejun Heo1880d932009-03-07 00:44:09 +09001001 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001002 kfree(chunk);
1003}
1004
1005static struct pcpu_chunk *alloc_pcpu_chunk(void)
1006{
1007 struct pcpu_chunk *chunk;
1008
1009 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1010 if (!chunk)
1011 return NULL;
1012
Tejun Heo1880d932009-03-07 00:44:09 +09001013 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001014 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1015 chunk->map[chunk->map_used++] = pcpu_unit_size;
1016
Tejun Heo65632972009-08-14 15:00:52 +09001017 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1018 pcpu_nr_groups, pcpu_atom_size,
1019 GFP_KERNEL);
1020 if (!chunk->vms) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001021 free_pcpu_chunk(chunk);
1022 return NULL;
1023 }
1024
1025 INIT_LIST_HEAD(&chunk->list);
1026 chunk->free_size = pcpu_unit_size;
1027 chunk->contig_hint = pcpu_unit_size;
Tejun Heo65632972009-08-14 15:00:52 +09001028 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
Tejun Heofbf59bc2009-02-20 16:29:08 +09001029
1030 return chunk;
1031}
1032
1033/**
Tejun Heoedcb4632009-03-06 14:33:59 +09001034 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +09001035 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +09001036 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +09001037 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +09001038 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001039 * Allocate percpu area of @size bytes aligned at @align.
1040 *
1041 * CONTEXT:
1042 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001043 *
1044 * RETURNS:
1045 * Percpu pointer to the allocated area on success, NULL on failure.
1046 */
Tejun Heoedcb4632009-03-06 14:33:59 +09001047static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001048{
Tejun Heof2badb02009-09-29 09:17:58 +09001049 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001050 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +09001051 const char *err;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001052 int slot, off;
Jiri Kosina403a91b2009-10-29 00:25:59 +09001053 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001054
Tejun Heo8d408b42009-02-24 11:57:21 +09001055 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001056 WARN(true, "illegal size (%zu) or align (%zu) for "
1057 "percpu allocation\n", size, align);
1058 return NULL;
1059 }
1060
Tejun Heoccea34b2009-03-07 00:44:13 +09001061 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +09001062 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001063
Tejun Heoedcb4632009-03-06 14:33:59 +09001064 /* serve reserved allocations from the reserved chunk if available */
1065 if (reserved && pcpu_reserved_chunk) {
1066 chunk = pcpu_reserved_chunk;
Tejun Heo9f7dcf22009-03-07 00:44:09 +09001067 if (size > chunk->contig_hint ||
Jiri Kosina403a91b2009-10-29 00:25:59 +09001068 pcpu_extend_area_map(chunk, &flags) < 0) {
Tejun Heof2badb02009-09-29 09:17:58 +09001069 err = "failed to extend area map of reserved chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +09001070 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +09001071 }
Tejun Heoedcb4632009-03-06 14:33:59 +09001072 off = pcpu_alloc_area(chunk, size, align);
1073 if (off >= 0)
1074 goto area_found;
Tejun Heof2badb02009-09-29 09:17:58 +09001075 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001076 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +09001077 }
1078
Tejun Heoccea34b2009-03-07 00:44:13 +09001079restart:
Tejun Heoedcb4632009-03-06 14:33:59 +09001080 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001081 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1082 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1083 if (size > chunk->contig_hint)
1084 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +09001085
Jiri Kosina403a91b2009-10-29 00:25:59 +09001086 switch (pcpu_extend_area_map(chunk, &flags)) {
Tejun Heoccea34b2009-03-07 00:44:13 +09001087 case 0:
1088 break;
1089 case 1:
1090 goto restart; /* pcpu_lock dropped, restart */
1091 default:
Tejun Heof2badb02009-09-29 09:17:58 +09001092 err = "failed to extend area map";
Tejun Heoccea34b2009-03-07 00:44:13 +09001093 goto fail_unlock;
1094 }
1095
Tejun Heofbf59bc2009-02-20 16:29:08 +09001096 off = pcpu_alloc_area(chunk, size, align);
1097 if (off >= 0)
1098 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001099 }
1100 }
1101
1102 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +09001103 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001104
Tejun Heofbf59bc2009-02-20 16:29:08 +09001105 chunk = alloc_pcpu_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +09001106 if (!chunk) {
1107 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +09001108 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +09001109 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001110
Jiri Kosina403a91b2009-10-29 00:25:59 +09001111 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001112 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +09001113 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001114
1115area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001116 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001117
Tejun Heofbf59bc2009-02-20 16:29:08 +09001118 /* populate, map and clear the area */
1119 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +09001120 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001121 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +09001122 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +09001123 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001124 }
1125
Tejun Heoccea34b2009-03-07 00:44:13 +09001126 mutex_unlock(&pcpu_alloc_mutex);
1127
Tejun Heobba174f2009-08-14 15:00:51 +09001128 /* return address relative to base address */
1129 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +09001130
1131fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001132 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001133fail_unlock_mutex:
1134 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +09001135 if (warn_limit) {
1136 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1137 "%s\n", size, align, err);
1138 dump_stack();
1139 if (!--warn_limit)
1140 pr_info("PERCPU: limit reached, disable warning\n");
1141 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001142 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001143}
Tejun Heoedcb4632009-03-06 14:33:59 +09001144
1145/**
1146 * __alloc_percpu - allocate dynamic percpu area
1147 * @size: size of area to allocate in bytes
1148 * @align: alignment of area (max PAGE_SIZE)
1149 *
1150 * Allocate percpu area of @size bytes aligned at @align. Might
1151 * sleep. Might trigger writeouts.
1152 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001153 * CONTEXT:
1154 * Does GFP_KERNEL allocation.
1155 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001156 * RETURNS:
1157 * Percpu pointer to the allocated area on success, NULL on failure.
1158 */
1159void *__alloc_percpu(size_t size, size_t align)
1160{
1161 return pcpu_alloc(size, align, false);
1162}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001163EXPORT_SYMBOL_GPL(__alloc_percpu);
1164
Tejun Heoedcb4632009-03-06 14:33:59 +09001165/**
1166 * __alloc_reserved_percpu - allocate reserved percpu area
1167 * @size: size of area to allocate in bytes
1168 * @align: alignment of area (max PAGE_SIZE)
1169 *
1170 * Allocate percpu area of @size bytes aligned at @align from reserved
1171 * percpu area if arch has set it up; otherwise, allocation is served
1172 * from the same dynamic area. Might sleep. Might trigger writeouts.
1173 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001174 * CONTEXT:
1175 * Does GFP_KERNEL allocation.
1176 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001177 * RETURNS:
1178 * Percpu pointer to the allocated area on success, NULL on failure.
1179 */
1180void *__alloc_reserved_percpu(size_t size, size_t align)
1181{
1182 return pcpu_alloc(size, align, true);
1183}
1184
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001185/**
1186 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1187 * @work: unused
1188 *
1189 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +09001190 *
1191 * CONTEXT:
1192 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001193 */
1194static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001195{
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001196 LIST_HEAD(todo);
1197 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1198 struct pcpu_chunk *chunk, *next;
1199
Tejun Heoccea34b2009-03-07 00:44:13 +09001200 mutex_lock(&pcpu_alloc_mutex);
1201 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001202
1203 list_for_each_entry_safe(chunk, next, head, list) {
1204 WARN_ON(chunk->immutable);
1205
1206 /* spare the first one */
1207 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1208 continue;
1209
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001210 list_move(&chunk->list, &todo);
1211 }
1212
Tejun Heoccea34b2009-03-07 00:44:13 +09001213 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001214
1215 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001216 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001217 free_pcpu_chunk(chunk);
1218 }
Tejun Heo971f3912009-08-14 15:00:49 +09001219
1220 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001221}
1222
1223/**
1224 * free_percpu - free percpu area
1225 * @ptr: pointer to area to free
1226 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001227 * Free percpu area @ptr.
1228 *
1229 * CONTEXT:
1230 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001231 */
1232void free_percpu(void *ptr)
1233{
1234 void *addr = __pcpu_ptr_to_addr(ptr);
1235 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001236 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001237 int off;
1238
1239 if (!ptr)
1240 return;
1241
Tejun Heoccea34b2009-03-07 00:44:13 +09001242 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001243
1244 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001245 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001246
1247 pcpu_free_area(chunk, off);
1248
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001249 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001250 if (chunk->free_size == pcpu_unit_size) {
1251 struct pcpu_chunk *pos;
1252
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001253 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001254 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001255 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001256 break;
1257 }
1258 }
1259
Tejun Heoccea34b2009-03-07 00:44:13 +09001260 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001261}
1262EXPORT_SYMBOL_GPL(free_percpu);
1263
Tejun Heo033e48f2009-08-14 15:00:51 +09001264static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1265 size_t reserved_size,
1266 ssize_t *dyn_sizep)
1267{
1268 size_t size_sum;
1269
1270 size_sum = PFN_ALIGN(static_size + reserved_size +
1271 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1272 if (*dyn_sizep != 0)
1273 *dyn_sizep = size_sum - static_size - reserved_size;
1274
1275 return size_sum;
1276}
1277
Tejun Heofbf59bc2009-02-20 16:29:08 +09001278/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001279 * pcpu_alloc_alloc_info - allocate percpu allocation info
1280 * @nr_groups: the number of groups
1281 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001282 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001283 * Allocate ai which is large enough for @nr_groups groups containing
1284 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1285 * cpu_map array which is long enough for @nr_units and filled with
1286 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1287 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001288 *
1289 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001290 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1291 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001292 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001293struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1294 int nr_units)
1295{
1296 struct pcpu_alloc_info *ai;
1297 size_t base_size, ai_size;
1298 void *ptr;
1299 int unit;
1300
1301 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1302 __alignof__(ai->groups[0].cpu_map[0]));
1303 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1304
1305 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1306 if (!ptr)
1307 return NULL;
1308 ai = ptr;
1309 ptr += base_size;
1310
1311 ai->groups[0].cpu_map = ptr;
1312
1313 for (unit = 0; unit < nr_units; unit++)
1314 ai->groups[0].cpu_map[unit] = NR_CPUS;
1315
1316 ai->nr_groups = nr_groups;
1317 ai->__ai_size = PFN_ALIGN(ai_size);
1318
1319 return ai;
1320}
1321
1322/**
1323 * pcpu_free_alloc_info - free percpu allocation info
1324 * @ai: pcpu_alloc_info to free
1325 *
1326 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1327 */
1328void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1329{
1330 free_bootmem(__pa(ai), ai->__ai_size);
1331}
1332
1333/**
1334 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
Tejun Heoedcb4632009-03-06 14:33:59 +09001335 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +09001336 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heofd1e8a12009-08-14 15:00:51 +09001337 * @atom_size: allocation atom size
1338 * @cpu_distance_fn: callback to determine distance between cpus, optional
1339 *
1340 * This function determines grouping of units, their mappings to cpus
1341 * and other parameters considering needed percpu size, allocation
1342 * atom size and distances between CPUs.
1343 *
1344 * Groups are always mutliples of atom size and CPUs which are of
1345 * LOCAL_DISTANCE both ways are grouped together and share space for
1346 * units in the same group. The returned configuration is guaranteed
1347 * to have CPUs on different nodes on different groups and >=75% usage
1348 * of allocated virtual address space.
1349 *
1350 * RETURNS:
1351 * On success, pointer to the new allocation_info is returned. On
1352 * failure, ERR_PTR value is returned.
1353 */
1354struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1355 size_t reserved_size, ssize_t dyn_size,
1356 size_t atom_size,
1357 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
Tejun Heo033e48f2009-08-14 15:00:51 +09001358{
1359 static int group_map[NR_CPUS] __initdata;
1360 static int group_cnt[NR_CPUS] __initdata;
1361 const size_t static_size = __per_cpu_end - __per_cpu_start;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001362 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
Tejun Heo033e48f2009-08-14 15:00:51 +09001363 size_t size_sum, min_unit_size, alloc_size;
1364 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001365 int last_allocs, group, unit;
Tejun Heo033e48f2009-08-14 15:00:51 +09001366 unsigned int cpu, tcpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001367 struct pcpu_alloc_info *ai;
1368 unsigned int *cpu_map;
Tejun Heo033e48f2009-08-14 15:00:51 +09001369
Tejun Heofb59e722009-09-24 18:50:34 +09001370 /* this function may be called multiple times */
1371 memset(group_map, 0, sizeof(group_map));
1372 memset(group_cnt, 0, sizeof(group_map));
1373
Tejun Heo033e48f2009-08-14 15:00:51 +09001374 /*
1375 * Determine min_unit_size, alloc_size and max_upa such that
Tejun Heofd1e8a12009-08-14 15:00:51 +09001376 * alloc_size is multiple of atom_size and is the smallest
Tejun Heo033e48f2009-08-14 15:00:51 +09001377 * which can accomodate 4k aligned segments which are equal to
1378 * or larger than min_unit_size.
1379 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001380 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001381 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1382
Tejun Heofd1e8a12009-08-14 15:00:51 +09001383 alloc_size = roundup(min_unit_size, atom_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001384 upa = alloc_size / min_unit_size;
1385 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1386 upa--;
1387 max_upa = upa;
1388
1389 /* group cpus according to their proximity */
1390 for_each_possible_cpu(cpu) {
1391 group = 0;
1392 next_group:
1393 for_each_possible_cpu(tcpu) {
1394 if (cpu == tcpu)
1395 break;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001396 if (group_map[tcpu] == group && cpu_distance_fn &&
Tejun Heo033e48f2009-08-14 15:00:51 +09001397 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1398 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1399 group++;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001400 nr_groups = max(nr_groups, group + 1);
Tejun Heo033e48f2009-08-14 15:00:51 +09001401 goto next_group;
1402 }
1403 }
1404 group_map[cpu] = group;
1405 group_cnt[group]++;
1406 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1407 }
1408
1409 /*
1410 * Expand unit size until address space usage goes over 75%
1411 * and then as much as possible without using more address
1412 * space.
1413 */
1414 last_allocs = INT_MAX;
1415 for (upa = max_upa; upa; upa--) {
1416 int allocs = 0, wasted = 0;
1417
1418 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1419 continue;
1420
Tejun Heofd1e8a12009-08-14 15:00:51 +09001421 for (group = 0; group < nr_groups; group++) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001422 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1423 allocs += this_allocs;
1424 wasted += this_allocs * upa - group_cnt[group];
1425 }
1426
1427 /*
1428 * Don't accept if wastage is over 25%. The
1429 * greater-than comparison ensures upa==1 always
1430 * passes the following check.
1431 */
1432 if (wasted > num_possible_cpus() / 3)
1433 continue;
1434
1435 /* and then don't consume more memory */
1436 if (allocs > last_allocs)
1437 break;
1438 last_allocs = allocs;
1439 best_upa = upa;
1440 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001441 upa = best_upa;
Tejun Heo033e48f2009-08-14 15:00:51 +09001442
Tejun Heofd1e8a12009-08-14 15:00:51 +09001443 /* allocate and fill alloc_info */
1444 for (group = 0; group < nr_groups; group++)
1445 nr_units += roundup(group_cnt[group], upa);
1446
1447 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1448 if (!ai)
1449 return ERR_PTR(-ENOMEM);
1450 cpu_map = ai->groups[0].cpu_map;
1451
1452 for (group = 0; group < nr_groups; group++) {
1453 ai->groups[group].cpu_map = cpu_map;
1454 cpu_map += roundup(group_cnt[group], upa);
Tejun Heo033e48f2009-08-14 15:00:51 +09001455 }
1456
Tejun Heofd1e8a12009-08-14 15:00:51 +09001457 ai->static_size = static_size;
1458 ai->reserved_size = reserved_size;
1459 ai->dyn_size = dyn_size;
1460 ai->unit_size = alloc_size / upa;
1461 ai->atom_size = atom_size;
1462 ai->alloc_size = alloc_size;
1463
1464 for (group = 0, unit = 0; group_cnt[group]; group++) {
1465 struct pcpu_group_info *gi = &ai->groups[group];
1466
1467 /*
1468 * Initialize base_offset as if all groups are located
1469 * back-to-back. The caller should update this to
1470 * reflect actual allocation.
1471 */
1472 gi->base_offset = unit * ai->unit_size;
1473
1474 for_each_possible_cpu(cpu)
1475 if (group_map[cpu] == group)
1476 gi->cpu_map[gi->nr_units++] = cpu;
1477 gi->nr_units = roundup(gi->nr_units, upa);
1478 unit += gi->nr_units;
1479 }
1480 BUG_ON(unit != nr_units);
1481
1482 return ai;
Tejun Heo033e48f2009-08-14 15:00:51 +09001483}
1484
Tejun Heofd1e8a12009-08-14 15:00:51 +09001485/**
1486 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1487 * @lvl: loglevel
1488 * @ai: allocation info to dump
1489 *
1490 * Print out information about @ai using loglevel @lvl.
1491 */
1492static void pcpu_dump_alloc_info(const char *lvl,
1493 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001494{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001495 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001496 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001497 int alloc = 0, alloc_end = 0;
1498 int group, v;
1499 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001500
Tejun Heofd1e8a12009-08-14 15:00:51 +09001501 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001502 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001503 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001504
Tejun Heofd1e8a12009-08-14 15:00:51 +09001505 v = num_possible_cpus();
1506 while (v /= 10)
1507 cpu_width++;
1508 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001509
Tejun Heofd1e8a12009-08-14 15:00:51 +09001510 upa = ai->alloc_size / ai->unit_size;
1511 width = upa * (cpu_width + 1) + group_width + 3;
1512 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001513
Tejun Heofd1e8a12009-08-14 15:00:51 +09001514 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1515 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1516 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1517
1518 for (group = 0; group < ai->nr_groups; group++) {
1519 const struct pcpu_group_info *gi = &ai->groups[group];
1520 int unit = 0, unit_end = 0;
1521
1522 BUG_ON(gi->nr_units % upa);
1523 for (alloc_end += gi->nr_units / upa;
1524 alloc < alloc_end; alloc++) {
1525 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001526 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001527 printk("%spcpu-alloc: ", lvl);
1528 }
1529 printk("[%0*d] ", group_width, group);
1530
1531 for (unit_end += upa; unit < unit_end; unit++)
1532 if (gi->cpu_map[unit] != NR_CPUS)
1533 printk("%0*d ", cpu_width,
1534 gi->cpu_map[unit]);
1535 else
1536 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001537 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001538 }
1539 printk("\n");
1540}
Tejun Heo033e48f2009-08-14 15:00:51 +09001541
Tejun Heofbf59bc2009-02-20 16:29:08 +09001542/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001543 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001544 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001545 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001546 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001547 * Initialize the first percpu chunk which contains the kernel static
1548 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001549 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001550 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001551 * @ai contains all information necessary to initialize the first
1552 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001553 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001554 * @ai->static_size is the size of static percpu area.
1555 *
1556 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001557 * reserve after the static area in the first chunk. This reserves
1558 * the first chunk such that it's available only through reserved
1559 * percpu allocation. This is primarily used to serve module percpu
1560 * static areas on architectures where the addressing model has
1561 * limited offset range for symbol relocations to guarantee module
1562 * percpu symbols fall inside the relocatable range.
1563 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001564 * @ai->dyn_size determines the number of bytes available for dynamic
1565 * allocation in the first chunk. The area between @ai->static_size +
1566 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001567 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001568 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1569 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1570 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001571 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001572 * @ai->atom_size is the allocation atom size and used as alignment
1573 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001574 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001575 * @ai->alloc_size is the allocation size and always multiple of
1576 * @ai->atom_size. This is larger than @ai->atom_size if
1577 * @ai->unit_size is larger than @ai->atom_size.
1578 *
1579 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1580 * percpu areas. Units which should be colocated are put into the
1581 * same group. Dynamic VM areas will be allocated according to these
1582 * groupings. If @ai->nr_groups is zero, a single group containing
1583 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001584 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001585 * The caller should have mapped the first chunk at @base_addr and
1586 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001587 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001588 * If the first chunk ends up with both reserved and dynamic areas, it
1589 * is served by two chunks - one to serve the core static and reserved
1590 * areas and the other for the dynamic area. They share the same vm
1591 * and page map but uses different area allocation map to stay away
1592 * from each other. The latter chunk is circulated in the chunk slots
1593 * and available for dynamic allocation like any other chunks.
1594 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001595 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001596 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001597 */
Tejun Heofb435d52009-08-14 15:00:51 +09001598int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1599 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001600{
Tejun Heo635b75f2009-09-24 09:43:11 +09001601 static char cpus_buf[4096] __initdata;
Tejun Heoedcb4632009-03-06 14:33:59 +09001602 static int smap[2], dmap[2];
Tejun Heofd1e8a12009-08-14 15:00:51 +09001603 size_t dyn_size = ai->dyn_size;
1604 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001605 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001606 unsigned long *group_offsets;
1607 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001608 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001609 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001610 int *unit_map;
1611 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001612
Tejun Heo635b75f2009-09-24 09:43:11 +09001613 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1614
1615#define PCPU_SETUP_BUG_ON(cond) do { \
1616 if (unlikely(cond)) { \
1617 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1618 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1619 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1620 BUG(); \
1621 } \
1622} while (0)
1623
Tejun Heo2f39e632009-07-04 08:11:00 +09001624 /* sanity checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001625 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1626 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo635b75f2009-09-24 09:43:11 +09001627 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1628 PCPU_SETUP_BUG_ON(!ai->static_size);
1629 PCPU_SETUP_BUG_ON(!base_addr);
1630 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1631 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1632 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001633
Tejun Heo65632972009-08-14 15:00:52 +09001634 /* process group information and build config tables accordingly */
1635 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1636 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001637 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001638 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001639
Tejun Heofd1e8a12009-08-14 15:00:51 +09001640 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001641 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001642 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001643
Tejun Heofd1e8a12009-08-14 15:00:51 +09001644 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1645 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001646
Tejun Heo65632972009-08-14 15:00:52 +09001647 group_offsets[group] = gi->base_offset;
1648 group_sizes[group] = gi->nr_units * ai->unit_size;
1649
Tejun Heofd1e8a12009-08-14 15:00:51 +09001650 for (i = 0; i < gi->nr_units; i++) {
1651 cpu = gi->cpu_map[i];
1652 if (cpu == NR_CPUS)
1653 continue;
1654
Tejun Heo635b75f2009-09-24 09:43:11 +09001655 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1656 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1657 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001658
1659 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001660 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1661
Tejun Heofd1e8a12009-08-14 15:00:51 +09001662 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001663 pcpu_first_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001664 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001665 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001666 pcpu_last_unit_cpu = cpu;
1667 pcpu_nr_units = unit;
1668
1669 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001670 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1671
1672 /* we're done parsing the input, undefine BUG macro and dump config */
1673#undef PCPU_SETUP_BUG_ON
1674 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001675
Tejun Heo65632972009-08-14 15:00:52 +09001676 pcpu_nr_groups = ai->nr_groups;
1677 pcpu_group_offsets = group_offsets;
1678 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001679 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001680 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001681
1682 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001683 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001684 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001685 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001686 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1687 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001688
Tejun Heod9b55ee2009-02-24 11:57:21 +09001689 /*
1690 * Allocate chunk slots. The additional last slot is for
1691 * empty chunks.
1692 */
1693 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001694 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1695 for (i = 0; i < pcpu_nr_slots; i++)
1696 INIT_LIST_HEAD(&pcpu_slot[i]);
1697
Tejun Heoedcb4632009-03-06 14:33:59 +09001698 /*
1699 * Initialize static chunk. If reserved_size is zero, the
1700 * static chunk covers static area + dynamic allocation area
1701 * in the first chunk. If reserved_size is not zero, it
1702 * covers static area + reserved area (mostly used for module
1703 * static percpu allocation).
1704 */
Tejun Heo2441d152009-03-06 14:33:59 +09001705 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1706 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001707 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001708 schunk->map = smap;
1709 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001710 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001711 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001712
Tejun Heofd1e8a12009-08-14 15:00:51 +09001713 if (ai->reserved_size) {
1714 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001715 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001716 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001717 } else {
1718 schunk->free_size = dyn_size;
1719 dyn_size = 0; /* dynamic area covered */
1720 }
Tejun Heo2441d152009-03-06 14:33:59 +09001721 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001722
Tejun Heofd1e8a12009-08-14 15:00:51 +09001723 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001724 if (schunk->free_size)
1725 schunk->map[schunk->map_used++] = schunk->free_size;
1726
Tejun Heoedcb4632009-03-06 14:33:59 +09001727 /* init dynamic chunk if necessary */
1728 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001729 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001730 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001731 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001732 dchunk->map = dmap;
1733 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001734 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001735 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001736
1737 dchunk->contig_hint = dchunk->free_size = dyn_size;
1738 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1739 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1740 }
1741
Tejun Heo2441d152009-03-06 14:33:59 +09001742 /* link the first chunk in */
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001743 pcpu_first_chunk = dchunk ?: schunk;
1744 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001745
1746 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001747 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001748 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001749}
Tejun Heo66c3a752009-03-10 16:27:48 +09001750
Tejun Heof58dc012009-08-14 15:00:50 +09001751const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1752 [PCPU_FC_AUTO] = "auto",
1753 [PCPU_FC_EMBED] = "embed",
1754 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001755};
Tejun Heo66c3a752009-03-10 16:27:48 +09001756
Tejun Heof58dc012009-08-14 15:00:50 +09001757enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1758
1759static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001760{
Tejun Heof58dc012009-08-14 15:00:50 +09001761 if (0)
1762 /* nada */;
1763#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1764 else if (!strcmp(str, "embed"))
1765 pcpu_chosen_fc = PCPU_FC_EMBED;
1766#endif
1767#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1768 else if (!strcmp(str, "page"))
1769 pcpu_chosen_fc = PCPU_FC_PAGE;
1770#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001771 else
1772 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001773
Tejun Heof58dc012009-08-14 15:00:50 +09001774 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001775}
Tejun Heof58dc012009-08-14 15:00:50 +09001776early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001777
Tejun Heo08fc4582009-08-14 15:00:49 +09001778#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1779 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo66c3a752009-03-10 16:27:48 +09001780/**
1781 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001782 * @reserved_size: the size of reserved percpu area in bytes
1783 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heoc8826dd2009-08-14 15:00:52 +09001784 * @atom_size: allocation atom size
1785 * @cpu_distance_fn: callback to determine distance between cpus, optional
1786 * @alloc_fn: function to allocate percpu page
1787 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001788 *
1789 * This is a helper to ease setting up embedded first percpu chunk and
1790 * can be called where pcpu_setup_first_chunk() is expected.
1791 *
1792 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001793 * by calling @alloc_fn and used as-is without being mapped into
1794 * vmalloc area. Allocations are always whole multiples of @atom_size
1795 * aligned to @atom_size.
1796 *
1797 * This enables the first chunk to piggy back on the linear physical
1798 * mapping which often uses larger page size. Please note that this
1799 * can result in very sparse cpu->unit mapping on NUMA machines thus
1800 * requiring large vmalloc address space. Don't use this allocator if
1801 * vmalloc space is not orders of magnitude larger than distances
1802 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001803 *
1804 * When @dyn_size is positive, dynamic area might be larger than
Tejun Heo788e5ab2009-07-04 08:10:58 +09001805 * specified to fill page alignment. When @dyn_size is auto,
1806 * @dyn_size is just big enough to fill page alignment after static
1807 * and reserved areas.
Tejun Heo66c3a752009-03-10 16:27:48 +09001808 *
1809 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001810 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001811 *
1812 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001813 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001814 */
Tejun Heoc8826dd2009-08-14 15:00:52 +09001815int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1816 size_t atom_size,
1817 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1818 pcpu_fc_alloc_fn_t alloc_fn,
1819 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001820{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001821 void *base = (void *)ULONG_MAX;
1822 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001823 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001824 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001825 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001826
Tejun Heoc8826dd2009-08-14 15:00:52 +09001827 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1828 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001829 if (IS_ERR(ai))
1830 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001831
Tejun Heofd1e8a12009-08-14 15:00:51 +09001832 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001833 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001834
Tejun Heoc8826dd2009-08-14 15:00:52 +09001835 areas = alloc_bootmem_nopanic(areas_size);
1836 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001837 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001838 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001839 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001840
Tejun Heoc8826dd2009-08-14 15:00:52 +09001841 /* allocate, copy and determine base address */
1842 for (group = 0; group < ai->nr_groups; group++) {
1843 struct pcpu_group_info *gi = &ai->groups[group];
1844 unsigned int cpu = NR_CPUS;
1845 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001846
Tejun Heoc8826dd2009-08-14 15:00:52 +09001847 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1848 cpu = gi->cpu_map[i];
1849 BUG_ON(cpu == NR_CPUS);
1850
1851 /* allocate space for the whole group */
1852 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1853 if (!ptr) {
1854 rc = -ENOMEM;
1855 goto out_free_areas;
1856 }
1857 areas[group] = ptr;
1858
1859 base = min(ptr, base);
1860
1861 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1862 if (gi->cpu_map[i] == NR_CPUS) {
1863 /* unused unit, free whole */
1864 free_fn(ptr, ai->unit_size);
1865 continue;
1866 }
1867 /* copy and return the unused part */
1868 memcpy(ptr, __per_cpu_load, ai->static_size);
1869 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1870 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001871 }
1872
Tejun Heoc8826dd2009-08-14 15:00:52 +09001873 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001874 max_distance = 0;
1875 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001876 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001877 max_distance = max_t(size_t, max_distance,
1878 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001879 }
1880 max_distance += ai->unit_size;
1881
1882 /* warn if maximum distance is further than 75% of vmalloc space */
1883 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001884 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001885 "space 0x%lx\n",
1886 max_distance, VMALLOC_END - VMALLOC_START);
1887#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1888 /* and fail if we have fallback */
1889 rc = -EINVAL;
1890 goto out_free;
1891#endif
1892 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001893
Tejun Heo004018e2009-08-14 15:00:49 +09001894 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001895 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1896 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001897
Tejun Heofb435d52009-08-14 15:00:51 +09001898 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001899 goto out_free;
1900
1901out_free_areas:
1902 for (group = 0; group < ai->nr_groups; group++)
1903 free_fn(areas[group],
1904 ai->groups[group].nr_units * ai->unit_size);
1905out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001906 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001907 if (areas)
1908 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001909 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001910}
Tejun Heo08fc4582009-08-14 15:00:49 +09001911#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1912 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heod4b95f82009-07-04 08:10:59 +09001913
Tejun Heo08fc4582009-08-14 15:00:49 +09001914#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001915/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001916 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001917 * @reserved_size: the size of reserved percpu area in bytes
1918 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1919 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1920 * @populate_pte_fn: function to populate pte
1921 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001922 * This is a helper to ease setting up page-remapped first percpu
1923 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001924 *
1925 * This is the basic allocator. Static percpu area is allocated
1926 * page-by-page into vmalloc area.
1927 *
1928 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001929 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001930 */
Tejun Heofb435d52009-08-14 15:00:51 +09001931int __init pcpu_page_first_chunk(size_t reserved_size,
1932 pcpu_fc_alloc_fn_t alloc_fn,
1933 pcpu_fc_free_fn_t free_fn,
1934 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001935{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001936 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001937 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001938 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001939 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001940 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001941 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001942 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001943
Tejun Heo00ae4062009-08-14 15:00:49 +09001944 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1945
Tejun Heofd1e8a12009-08-14 15:00:51 +09001946 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1947 if (IS_ERR(ai))
1948 return PTR_ERR(ai);
1949 BUG_ON(ai->nr_groups != 1);
1950 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1951
1952 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001953
1954 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001955 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1956 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09001957 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09001958
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001959 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001960 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001961 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001962 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001963 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001964 void *ptr;
1965
Tejun Heo3cbc8562009-08-14 15:00:50 +09001966 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001967 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001968 pr_warning("PERCPU: failed to allocate %s page "
1969 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001970 goto enomem;
1971 }
Tejun Heoce3141a2009-07-04 08:11:00 +09001972 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001973 }
1974
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001975 /* allocate vm area, map the pages and copy static data */
1976 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001977 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001978 vm_area_register_early(&vm, PAGE_SIZE);
1979
Tejun Heofd1e8a12009-08-14 15:00:51 +09001980 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001981 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001982 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001983
Tejun Heoce3141a2009-07-04 08:11:00 +09001984 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001985 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1986
1987 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001988 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1989 unit_pages);
1990 if (rc < 0)
1991 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001992
1993 /*
1994 * FIXME: Archs with virtual cache should flush local
1995 * cache for the linear mapping here - something
1996 * equivalent to flush_cache_vmap() on the local cpu.
1997 * flush_cache_vmap() can't be used as most supporting
1998 * data structures are not set up yet.
1999 */
2000
2001 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002002 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002003 }
2004
2005 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09002006 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002007 unit_pages, psize_str, vm.addr, ai->static_size,
2008 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002009
Tejun Heofb435d52009-08-14 15:00:51 +09002010 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002011 goto out_free_ar;
2012
2013enomem:
2014 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002015 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002016 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002017out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09002018 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002019 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002020 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002021}
Tejun Heo08fc4582009-08-14 15:00:49 +09002022#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002023
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002024/*
Tejun Heoe74e3962009-03-30 19:07:44 +09002025 * Generic percpu area setup.
2026 *
2027 * The embedding helper is used because its behavior closely resembles
2028 * the original non-dynamic generic percpu area setup. This is
2029 * important because many archs have addressing restrictions and might
2030 * fail if the percpu area is located far away from the previous
2031 * location. As an added bonus, in non-NUMA cases, embedding is
2032 * generally a good idea TLB-wise because percpu area can piggy back
2033 * on the physical linear memory mapping which uses large page
2034 * mappings on applicable archs.
2035 */
2036#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2037unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2038EXPORT_SYMBOL(__per_cpu_offset);
2039
Tejun Heoc8826dd2009-08-14 15:00:52 +09002040static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2041 size_t align)
2042{
2043 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2044}
2045
2046static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2047{
2048 free_bootmem(__pa(ptr), size);
2049}
2050
Tejun Heoe74e3962009-03-30 19:07:44 +09002051void __init setup_per_cpu_areas(void)
2052{
Tejun Heoe74e3962009-03-30 19:07:44 +09002053 unsigned long delta;
2054 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002055 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002056
2057 /*
2058 * Always reserve area for module percpu variables. That's
2059 * what the legacy allocator did.
2060 */
Tejun Heofb435d52009-08-14 15:00:51 +09002061 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002062 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2063 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002064 if (rc < 0)
Tejun Heoe74e3962009-03-30 19:07:44 +09002065 panic("Failed to initialized percpu areas.");
2066
2067 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2068 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002069 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002070}
2071#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */