blob: 5adfc268b408936a3d18ae9014ce70e49590738d [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 Heo833af842009-11-11 15:35:18 +0900358 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
359 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900360 *
Tejun Heo833af842009-11-11 15:35:18 +0900361 * Determine whether area map of @chunk needs to be extended to
362 * accomodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900363 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900364 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900365 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900366 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900367 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900368 * New target map allocation length if extension is necessary, 0
369 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900370 */
Tejun Heo833af842009-11-11 15:35:18 +0900371static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900372{
373 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900374
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900375 if (chunk->map_alloc >= chunk->map_used + 2)
376 return 0;
377
378 new_alloc = PCPU_DFL_MAP_ALLOC;
379 while (new_alloc < chunk->map_used + 2)
380 new_alloc *= 2;
381
Tejun Heo833af842009-11-11 15:35:18 +0900382 return new_alloc;
383}
384
385/**
386 * pcpu_extend_area_map - extend area map of a chunk
387 * @chunk: chunk of interest
388 * @new_alloc: new target allocation length of the area map
389 *
390 * Extend area map of @chunk to have @new_alloc entries.
391 *
392 * CONTEXT:
393 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
394 *
395 * RETURNS:
396 * 0 on success, -errno on failure.
397 */
398static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
399{
400 int *old = NULL, *new = NULL;
401 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
402 unsigned long flags;
403
404 new = pcpu_mem_alloc(new_size);
405 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900406 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900407
Tejun Heo833af842009-11-11 15:35:18 +0900408 /* acquire pcpu_lock and switch to new area map */
409 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900410
Tejun Heo833af842009-11-11 15:35:18 +0900411 if (new_alloc <= chunk->map_alloc)
412 goto out_unlock;
413
414 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
415 memcpy(new, chunk->map, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900416
417 /*
418 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
419 * one of the first chunks and still using static map.
420 */
421 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
Tejun Heo833af842009-11-11 15:35:18 +0900422 old = chunk->map;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900423
424 chunk->map_alloc = new_alloc;
425 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900426 new = NULL;
427
428out_unlock:
429 spin_unlock_irqrestore(&pcpu_lock, flags);
430
431 /*
432 * pcpu_mem_free() might end up calling vfree() which uses
433 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
434 */
435 pcpu_mem_free(old, old_size);
436 pcpu_mem_free(new, new_size);
437
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900438 return 0;
439}
440
441/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900442 * pcpu_split_block - split a map block
443 * @chunk: chunk of interest
444 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900445 * @head: head size in bytes (can be 0)
446 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900447 *
448 * Split the @i'th map block into two or three blocks. If @head is
449 * non-zero, @head bytes block is inserted before block @i moving it
450 * to @i+1 and reducing its size by @head bytes.
451 *
452 * If @tail is non-zero, the target block, which can be @i or @i+1
453 * depending on @head, is reduced by @tail bytes and @tail byte block
454 * is inserted after the target block.
455 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900456 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900457 *
458 * CONTEXT:
459 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900460 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900461static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
462 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900463{
464 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900465
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900466 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900467
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900468 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900469 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
470 sizeof(chunk->map[0]) * (chunk->map_used - i));
471 chunk->map_used += nr_extra;
472
473 if (head) {
474 chunk->map[i + 1] = chunk->map[i] - head;
475 chunk->map[i++] = head;
476 }
477 if (tail) {
478 chunk->map[i++] -= tail;
479 chunk->map[i] = tail;
480 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900481}
482
483/**
484 * pcpu_alloc_area - allocate area from a pcpu_chunk
485 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900486 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900487 * @align: wanted align
488 *
489 * Try to allocate @size bytes area aligned at @align from @chunk.
490 * Note that this function only allocates the offset. It doesn't
491 * populate or map the area.
492 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900493 * @chunk->map must have at least two free slots.
494 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900495 * CONTEXT:
496 * pcpu_lock.
497 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900498 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900499 * Allocated offset in @chunk on success, -1 if no matching area is
500 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900501 */
502static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
503{
504 int oslot = pcpu_chunk_slot(chunk);
505 int max_contig = 0;
506 int i, off;
507
Tejun Heofbf59bc2009-02-20 16:29:08 +0900508 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
509 bool is_last = i + 1 == chunk->map_used;
510 int head, tail;
511
512 /* extra for alignment requirement */
513 head = ALIGN(off, align) - off;
514 BUG_ON(i == 0 && head != 0);
515
516 if (chunk->map[i] < 0)
517 continue;
518 if (chunk->map[i] < head + size) {
519 max_contig = max(chunk->map[i], max_contig);
520 continue;
521 }
522
523 /*
524 * If head is small or the previous block is free,
525 * merge'em. Note that 'small' is defined as smaller
526 * than sizeof(int), which is very small but isn't too
527 * uncommon for percpu allocations.
528 */
529 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
530 if (chunk->map[i - 1] > 0)
531 chunk->map[i - 1] += head;
532 else {
533 chunk->map[i - 1] -= head;
534 chunk->free_size -= head;
535 }
536 chunk->map[i] -= head;
537 off += head;
538 head = 0;
539 }
540
541 /* if tail is small, just keep it around */
542 tail = chunk->map[i] - head - size;
543 if (tail < sizeof(int))
544 tail = 0;
545
546 /* split if warranted */
547 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900548 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900549 if (head) {
550 i++;
551 off += head;
552 max_contig = max(chunk->map[i - 1], max_contig);
553 }
554 if (tail)
555 max_contig = max(chunk->map[i + 1], max_contig);
556 }
557
558 /* update hint and mark allocated */
559 if (is_last)
560 chunk->contig_hint = max_contig; /* fully scanned */
561 else
562 chunk->contig_hint = max(chunk->contig_hint,
563 max_contig);
564
565 chunk->free_size -= chunk->map[i];
566 chunk->map[i] = -chunk->map[i];
567
568 pcpu_chunk_relocate(chunk, oslot);
569 return off;
570 }
571
572 chunk->contig_hint = max_contig; /* fully scanned */
573 pcpu_chunk_relocate(chunk, oslot);
574
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900575 /* tell the upper layer that this chunk has no matching area */
576 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900577}
578
579/**
580 * pcpu_free_area - free area to a pcpu_chunk
581 * @chunk: chunk of interest
582 * @freeme: offset of area to free
583 *
584 * Free area starting from @freeme to @chunk. Note that this function
585 * only modifies the allocation map. It doesn't depopulate or unmap
586 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900587 *
588 * CONTEXT:
589 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900590 */
591static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
592{
593 int oslot = pcpu_chunk_slot(chunk);
594 int i, off;
595
596 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
597 if (off == freeme)
598 break;
599 BUG_ON(off != freeme);
600 BUG_ON(chunk->map[i] > 0);
601
602 chunk->map[i] = -chunk->map[i];
603 chunk->free_size += chunk->map[i];
604
605 /* merge with previous? */
606 if (i > 0 && chunk->map[i - 1] >= 0) {
607 chunk->map[i - 1] += chunk->map[i];
608 chunk->map_used--;
609 memmove(&chunk->map[i], &chunk->map[i + 1],
610 (chunk->map_used - i) * sizeof(chunk->map[0]));
611 i--;
612 }
613 /* merge with next? */
614 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
615 chunk->map[i] += chunk->map[i + 1];
616 chunk->map_used--;
617 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
618 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
619 }
620
621 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
622 pcpu_chunk_relocate(chunk, oslot);
623}
624
625/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900626 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900627 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900628 * @bitmapp: output parameter for bitmap
629 * @may_alloc: may allocate the array
Tejun Heofbf59bc2009-02-20 16:29:08 +0900630 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900631 * Returns pointer to array of pointers to struct page and bitmap,
632 * both of which can be indexed with pcpu_page_idx(). The returned
633 * array is cleared to zero and *@bitmapp is copied from
634 * @chunk->populated. Note that there is only one array and bitmap
635 * and access exclusion is the caller's responsibility.
636 *
637 * CONTEXT:
638 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
639 * Otherwise, don't care.
640 *
641 * RETURNS:
642 * Pointer to temp pages array on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900643 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900644static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
645 unsigned long **bitmapp,
646 bool may_alloc)
647{
648 static struct page **pages;
649 static unsigned long *bitmap;
Tejun Heo2f39e632009-07-04 08:11:00 +0900650 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
Tejun Heoce3141a2009-07-04 08:11:00 +0900651 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
652 sizeof(unsigned long);
653
654 if (!pages || !bitmap) {
655 if (may_alloc && !pages)
656 pages = pcpu_mem_alloc(pages_size);
657 if (may_alloc && !bitmap)
658 bitmap = pcpu_mem_alloc(bitmap_size);
659 if (!pages || !bitmap)
660 return NULL;
661 }
662
663 memset(pages, 0, pages_size);
664 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
665
666 *bitmapp = bitmap;
667 return pages;
668}
669
670/**
671 * pcpu_free_pages - free pages which were allocated for @chunk
672 * @chunk: chunk pages were allocated for
673 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
674 * @populated: populated bitmap
675 * @page_start: page index of the first page to be freed
676 * @page_end: page index of the last page to be freed + 1
677 *
678 * Free pages [@page_start and @page_end) in @pages for all units.
679 * The pages were allocated for @chunk.
680 */
681static void pcpu_free_pages(struct pcpu_chunk *chunk,
682 struct page **pages, unsigned long *populated,
683 int page_start, int page_end)
684{
685 unsigned int cpu;
686 int i;
687
688 for_each_possible_cpu(cpu) {
689 for (i = page_start; i < page_end; i++) {
690 struct page *page = pages[pcpu_page_idx(cpu, i)];
691
692 if (page)
693 __free_page(page);
694 }
695 }
696}
697
698/**
699 * pcpu_alloc_pages - allocates pages for @chunk
700 * @chunk: target chunk
701 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
702 * @populated: populated bitmap
703 * @page_start: page index of the first page to be allocated
704 * @page_end: page index of the last page to be allocated + 1
705 *
706 * Allocate pages [@page_start,@page_end) into @pages for all units.
707 * The allocation is for @chunk. Percpu core doesn't care about the
708 * content of @pages and will pass it verbatim to pcpu_map_pages().
709 */
710static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
711 struct page **pages, unsigned long *populated,
712 int page_start, int page_end)
713{
714 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
715 unsigned int cpu;
716 int i;
717
718 for_each_possible_cpu(cpu) {
719 for (i = page_start; i < page_end; i++) {
720 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
721
722 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
723 if (!*pagep) {
724 pcpu_free_pages(chunk, pages, populated,
725 page_start, page_end);
726 return -ENOMEM;
727 }
728 }
729 }
730 return 0;
731}
732
733/**
734 * pcpu_pre_unmap_flush - flush cache prior to unmapping
735 * @chunk: chunk the regions to be flushed belongs to
736 * @page_start: page index of the first page to be flushed
737 * @page_end: page index of the last page to be flushed + 1
738 *
739 * Pages in [@page_start,@page_end) of @chunk are about to be
740 * unmapped. Flush cache. As each flushing trial can be very
741 * expensive, issue flush on the whole region at once rather than
742 * doing it for each cpu. This could be an overkill but is more
743 * scalable.
744 */
745static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
746 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900747{
Tejun Heo2f39e632009-07-04 08:11:00 +0900748 flush_cache_vunmap(
749 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
750 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heoce3141a2009-07-04 08:11:00 +0900751}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900752
Tejun Heoce3141a2009-07-04 08:11:00 +0900753static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
754{
755 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
756}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900757
Tejun Heoce3141a2009-07-04 08:11:00 +0900758/**
759 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
760 * @chunk: chunk of interest
761 * @pages: pages array which can be used to pass information to free
762 * @populated: populated bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900763 * @page_start: page index of the first page to unmap
764 * @page_end: page index of the last page to unmap + 1
Tejun Heofbf59bc2009-02-20 16:29:08 +0900765 *
766 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
Tejun Heoce3141a2009-07-04 08:11:00 +0900767 * Corresponding elements in @pages were cleared by the caller and can
768 * be used to carry information to pcpu_free_pages() which will be
769 * called after all unmaps are finished. The caller should call
770 * proper pre/post flush functions.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900771 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900772static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
773 struct page **pages, unsigned long *populated,
774 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900775{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900776 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900777 int i;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900778
Tejun Heoce3141a2009-07-04 08:11:00 +0900779 for_each_possible_cpu(cpu) {
780 for (i = page_start; i < page_end; i++) {
781 struct page *page;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900782
Tejun Heoce3141a2009-07-04 08:11:00 +0900783 page = pcpu_chunk_page(chunk, cpu, i);
784 WARN_ON(!page);
785 pages[pcpu_page_idx(cpu, i)] = page;
786 }
787 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
788 page_end - page_start);
789 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900790
Tejun Heoce3141a2009-07-04 08:11:00 +0900791 for (i = page_start; i < page_end; i++)
792 __clear_bit(i, populated);
793}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900794
Tejun Heoce3141a2009-07-04 08:11:00 +0900795/**
796 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
797 * @chunk: pcpu_chunk the regions to be flushed belong to
798 * @page_start: page index of the first page to be flushed
799 * @page_end: page index of the last page to be flushed + 1
800 *
801 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
802 * TLB for the regions. This can be skipped if the area is to be
803 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
804 *
805 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
806 * for the whole region.
807 */
808static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
809 int page_start, int page_end)
810{
Tejun Heo2f39e632009-07-04 08:11:00 +0900811 flush_tlb_kernel_range(
812 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
813 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900814}
815
Tejun Heoc8a51be2009-07-04 08:10:59 +0900816static int __pcpu_map_pages(unsigned long addr, struct page **pages,
817 int nr_pages)
818{
819 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
820 PAGE_KERNEL, pages);
821}
822
823/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900824 * pcpu_map_pages - map pages into a pcpu_chunk
Tejun Heoc8a51be2009-07-04 08:10:59 +0900825 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900826 * @pages: pages array containing pages to be mapped
827 * @populated: populated bitmap
Tejun Heoc8a51be2009-07-04 08:10:59 +0900828 * @page_start: page index of the first page to map
829 * @page_end: page index of the last page to map + 1
830 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900831 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
832 * caller is responsible for calling pcpu_post_map_flush() after all
833 * mappings are complete.
834 *
835 * This function is responsible for setting corresponding bits in
836 * @chunk->populated bitmap and whatever is necessary for reverse
837 * lookup (addr -> chunk).
Tejun Heoc8a51be2009-07-04 08:10:59 +0900838 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900839static int pcpu_map_pages(struct pcpu_chunk *chunk,
840 struct page **pages, unsigned long *populated,
841 int page_start, int page_end)
Tejun Heoc8a51be2009-07-04 08:10:59 +0900842{
Tejun Heoce3141a2009-07-04 08:11:00 +0900843 unsigned int cpu, tcpu;
844 int i, err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900845
846 for_each_possible_cpu(cpu) {
847 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
Tejun Heoce3141a2009-07-04 08:11:00 +0900848 &pages[pcpu_page_idx(cpu, page_start)],
Tejun Heoc8a51be2009-07-04 08:10:59 +0900849 page_end - page_start);
850 if (err < 0)
Tejun Heoce3141a2009-07-04 08:11:00 +0900851 goto err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900852 }
853
Tejun Heoce3141a2009-07-04 08:11:00 +0900854 /* mapping successful, link chunk and mark populated */
855 for (i = page_start; i < page_end; i++) {
856 for_each_possible_cpu(cpu)
857 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
858 chunk);
859 __set_bit(i, populated);
860 }
861
862 return 0;
863
864err:
865 for_each_possible_cpu(tcpu) {
866 if (tcpu == cpu)
867 break;
868 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
869 page_end - page_start);
870 }
871 return err;
872}
873
874/**
875 * pcpu_post_map_flush - flush cache after mapping
876 * @chunk: pcpu_chunk the regions to be flushed belong to
877 * @page_start: page index of the first page to be flushed
878 * @page_end: page index of the last page to be flushed + 1
879 *
880 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
881 * cache.
882 *
883 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
884 * for the whole region.
885 */
886static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
887 int page_start, int page_end)
888{
Tejun Heo2f39e632009-07-04 08:11:00 +0900889 flush_cache_vmap(
890 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
891 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900892}
893
894/**
895 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
896 * @chunk: chunk to depopulate
897 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900898 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900899 * @flush: whether to flush cache and tlb or not
900 *
901 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
902 * from @chunk. If @flush is true, vcache is flushed before unmapping
903 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900904 *
905 * CONTEXT:
906 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900907 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900908static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900909{
910 int page_start = PFN_DOWN(off);
911 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900912 struct page **pages;
913 unsigned long *populated;
914 int rs, re;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900915
Tejun Heoce3141a2009-07-04 08:11:00 +0900916 /* quick path, check whether it's empty already */
917 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
918 if (rs == page_start && re == page_end)
919 return;
920 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900921 }
922
Tejun Heoce3141a2009-07-04 08:11:00 +0900923 /* immutable chunks can't be depopulated */
Tejun Heo8d408b42009-02-24 11:57:21 +0900924 WARN_ON(chunk->immutable);
925
Tejun Heoce3141a2009-07-04 08:11:00 +0900926 /*
927 * If control reaches here, there must have been at least one
928 * successful population attempt so the temp pages array must
929 * be available now.
930 */
931 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
932 BUG_ON(!pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900933
Tejun Heoce3141a2009-07-04 08:11:00 +0900934 /* unmap and free */
935 pcpu_pre_unmap_flush(chunk, page_start, page_end);
936
937 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
938 pcpu_unmap_pages(chunk, pages, populated, rs, re);
939
940 /* no need to flush tlb, vmalloc will handle it lazily */
941
942 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
943 pcpu_free_pages(chunk, pages, populated, rs, re);
944
945 /* commit new bitmap */
946 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900947}
948
949/**
950 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
951 * @chunk: chunk of interest
952 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900953 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900954 *
955 * For each cpu, populate and map pages [@page_start,@page_end) into
956 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900957 *
958 * CONTEXT:
959 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900960 */
961static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
962{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900963 int page_start = PFN_DOWN(off);
964 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900965 int free_end = page_start, unmap_end = page_start;
966 struct page **pages;
967 unsigned long *populated;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900968 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900969 int rs, re, rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900970
Tejun Heoce3141a2009-07-04 08:11:00 +0900971 /* quick path, check whether all pages are already there */
972 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
973 if (rs == page_start && re == page_end)
974 goto clear;
975 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900976 }
977
Tejun Heoce3141a2009-07-04 08:11:00 +0900978 /* need to allocate and map pages, this chunk can't be immutable */
979 WARN_ON(chunk->immutable);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900980
Tejun Heoce3141a2009-07-04 08:11:00 +0900981 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
982 if (!pages)
983 return -ENOMEM;
984
985 /* alloc and map */
986 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
987 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
988 if (rc)
989 goto err_free;
990 free_end = re;
991 }
992
993 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
994 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
995 if (rc)
996 goto err_unmap;
997 unmap_end = re;
998 }
999 pcpu_post_map_flush(chunk, page_start, page_end);
1000
1001 /* commit new bitmap */
1002 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
1003clear:
Tejun Heofbf59bc2009-02-20 16:29:08 +09001004 for_each_possible_cpu(cpu)
Tejun Heo2f39e632009-07-04 08:11:00 +09001005 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001006 return 0;
Tejun Heoce3141a2009-07-04 08:11:00 +09001007
1008err_unmap:
1009 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
1010 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
1011 pcpu_unmap_pages(chunk, pages, populated, rs, re);
1012 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
1013err_free:
1014 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
1015 pcpu_free_pages(chunk, pages, populated, rs, re);
1016 return rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001017}
1018
1019static void free_pcpu_chunk(struct pcpu_chunk *chunk)
1020{
1021 if (!chunk)
1022 return;
Tejun Heo65632972009-08-14 15:00:52 +09001023 if (chunk->vms)
1024 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
Tejun Heo1880d932009-03-07 00:44:09 +09001025 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001026 kfree(chunk);
1027}
1028
1029static struct pcpu_chunk *alloc_pcpu_chunk(void)
1030{
1031 struct pcpu_chunk *chunk;
1032
1033 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1034 if (!chunk)
1035 return NULL;
1036
Tejun Heo1880d932009-03-07 00:44:09 +09001037 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001038 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1039 chunk->map[chunk->map_used++] = pcpu_unit_size;
1040
Tejun Heo65632972009-08-14 15:00:52 +09001041 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1042 pcpu_nr_groups, pcpu_atom_size,
1043 GFP_KERNEL);
1044 if (!chunk->vms) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001045 free_pcpu_chunk(chunk);
1046 return NULL;
1047 }
1048
1049 INIT_LIST_HEAD(&chunk->list);
1050 chunk->free_size = pcpu_unit_size;
1051 chunk->contig_hint = pcpu_unit_size;
Tejun Heo65632972009-08-14 15:00:52 +09001052 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
Tejun Heofbf59bc2009-02-20 16:29:08 +09001053
1054 return chunk;
1055}
1056
1057/**
Tejun Heoedcb4632009-03-06 14:33:59 +09001058 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +09001059 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +09001060 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +09001061 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +09001062 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001063 * Allocate percpu area of @size bytes aligned at @align.
1064 *
1065 * CONTEXT:
1066 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001067 *
1068 * RETURNS:
1069 * Percpu pointer to the allocated area on success, NULL on failure.
1070 */
Tejun Heoedcb4632009-03-06 14:33:59 +09001071static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001072{
Tejun Heof2badb02009-09-29 09:17:58 +09001073 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001074 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +09001075 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +09001076 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +09001077 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001078
Tejun Heo8d408b42009-02-24 11:57:21 +09001079 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001080 WARN(true, "illegal size (%zu) or align (%zu) for "
1081 "percpu allocation\n", size, align);
1082 return NULL;
1083 }
1084
Tejun Heoccea34b2009-03-07 00:44:13 +09001085 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +09001086 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001087
Tejun Heoedcb4632009-03-06 14:33:59 +09001088 /* serve reserved allocations from the reserved chunk if available */
1089 if (reserved && pcpu_reserved_chunk) {
1090 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +09001091
1092 if (size > chunk->contig_hint) {
1093 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001094 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +09001095 }
Tejun Heo833af842009-11-11 15:35:18 +09001096
1097 while ((new_alloc = pcpu_need_to_extend(chunk))) {
1098 spin_unlock_irqrestore(&pcpu_lock, flags);
1099 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
1100 err = "failed to extend area map of reserved chunk";
1101 goto fail_unlock_mutex;
1102 }
1103 spin_lock_irqsave(&pcpu_lock, flags);
1104 }
1105
Tejun Heoedcb4632009-03-06 14:33:59 +09001106 off = pcpu_alloc_area(chunk, size, align);
1107 if (off >= 0)
1108 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +09001109
Tejun Heof2badb02009-09-29 09:17:58 +09001110 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001111 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +09001112 }
1113
Tejun Heoccea34b2009-03-07 00:44:13 +09001114restart:
Tejun Heoedcb4632009-03-06 14:33:59 +09001115 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001116 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1117 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1118 if (size > chunk->contig_hint)
1119 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +09001120
Tejun Heo833af842009-11-11 15:35:18 +09001121 new_alloc = pcpu_need_to_extend(chunk);
1122 if (new_alloc) {
1123 spin_unlock_irqrestore(&pcpu_lock, flags);
1124 if (pcpu_extend_area_map(chunk,
1125 new_alloc) < 0) {
1126 err = "failed to extend area map";
1127 goto fail_unlock_mutex;
1128 }
1129 spin_lock_irqsave(&pcpu_lock, flags);
1130 /*
1131 * pcpu_lock has been dropped, need to
1132 * restart cpu_slot list walking.
1133 */
1134 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +09001135 }
1136
Tejun Heofbf59bc2009-02-20 16:29:08 +09001137 off = pcpu_alloc_area(chunk, size, align);
1138 if (off >= 0)
1139 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001140 }
1141 }
1142
1143 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +09001144 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001145
Tejun Heofbf59bc2009-02-20 16:29:08 +09001146 chunk = alloc_pcpu_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +09001147 if (!chunk) {
1148 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +09001149 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +09001150 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001151
Jiri Kosina403a91b2009-10-29 00:25:59 +09001152 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001153 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +09001154 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001155
1156area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001157 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001158
Tejun Heofbf59bc2009-02-20 16:29:08 +09001159 /* populate, map and clear the area */
1160 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +09001161 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001162 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +09001163 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +09001164 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001165 }
1166
Tejun Heoccea34b2009-03-07 00:44:13 +09001167 mutex_unlock(&pcpu_alloc_mutex);
1168
Tejun Heobba174f2009-08-14 15:00:51 +09001169 /* return address relative to base address */
1170 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +09001171
1172fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001173 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001174fail_unlock_mutex:
1175 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +09001176 if (warn_limit) {
1177 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1178 "%s\n", size, align, err);
1179 dump_stack();
1180 if (!--warn_limit)
1181 pr_info("PERCPU: limit reached, disable warning\n");
1182 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001183 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001184}
Tejun Heoedcb4632009-03-06 14:33:59 +09001185
1186/**
1187 * __alloc_percpu - allocate dynamic percpu area
1188 * @size: size of area to allocate in bytes
1189 * @align: alignment of area (max PAGE_SIZE)
1190 *
1191 * Allocate percpu area of @size bytes aligned at @align. Might
1192 * sleep. Might trigger writeouts.
1193 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001194 * CONTEXT:
1195 * Does GFP_KERNEL allocation.
1196 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001197 * RETURNS:
1198 * Percpu pointer to the allocated area on success, NULL on failure.
1199 */
1200void *__alloc_percpu(size_t size, size_t align)
1201{
1202 return pcpu_alloc(size, align, false);
1203}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001204EXPORT_SYMBOL_GPL(__alloc_percpu);
1205
Tejun Heoedcb4632009-03-06 14:33:59 +09001206/**
1207 * __alloc_reserved_percpu - allocate reserved percpu area
1208 * @size: size of area to allocate in bytes
1209 * @align: alignment of area (max PAGE_SIZE)
1210 *
1211 * Allocate percpu area of @size bytes aligned at @align from reserved
1212 * percpu area if arch has set it up; otherwise, allocation is served
1213 * from the same dynamic area. Might sleep. Might trigger writeouts.
1214 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001215 * CONTEXT:
1216 * Does GFP_KERNEL allocation.
1217 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001218 * RETURNS:
1219 * Percpu pointer to the allocated area on success, NULL on failure.
1220 */
1221void *__alloc_reserved_percpu(size_t size, size_t align)
1222{
1223 return pcpu_alloc(size, align, true);
1224}
1225
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001226/**
1227 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1228 * @work: unused
1229 *
1230 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +09001231 *
1232 * CONTEXT:
1233 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001234 */
1235static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001236{
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001237 LIST_HEAD(todo);
1238 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1239 struct pcpu_chunk *chunk, *next;
1240
Tejun Heoccea34b2009-03-07 00:44:13 +09001241 mutex_lock(&pcpu_alloc_mutex);
1242 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001243
1244 list_for_each_entry_safe(chunk, next, head, list) {
1245 WARN_ON(chunk->immutable);
1246
1247 /* spare the first one */
1248 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1249 continue;
1250
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001251 list_move(&chunk->list, &todo);
1252 }
1253
Tejun Heoccea34b2009-03-07 00:44:13 +09001254 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001255
1256 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001257 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001258 free_pcpu_chunk(chunk);
1259 }
Tejun Heo971f3912009-08-14 15:00:49 +09001260
1261 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001262}
1263
1264/**
1265 * free_percpu - free percpu area
1266 * @ptr: pointer to area to free
1267 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001268 * Free percpu area @ptr.
1269 *
1270 * CONTEXT:
1271 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001272 */
1273void free_percpu(void *ptr)
1274{
1275 void *addr = __pcpu_ptr_to_addr(ptr);
1276 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001277 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001278 int off;
1279
1280 if (!ptr)
1281 return;
1282
Tejun Heoccea34b2009-03-07 00:44:13 +09001283 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001284
1285 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001286 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001287
1288 pcpu_free_area(chunk, off);
1289
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001290 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001291 if (chunk->free_size == pcpu_unit_size) {
1292 struct pcpu_chunk *pos;
1293
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001294 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001295 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001296 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001297 break;
1298 }
1299 }
1300
Tejun Heoccea34b2009-03-07 00:44:13 +09001301 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001302}
1303EXPORT_SYMBOL_GPL(free_percpu);
1304
Tejun Heo033e48f2009-08-14 15:00:51 +09001305static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1306 size_t reserved_size,
1307 ssize_t *dyn_sizep)
1308{
1309 size_t size_sum;
1310
1311 size_sum = PFN_ALIGN(static_size + reserved_size +
1312 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1313 if (*dyn_sizep != 0)
1314 *dyn_sizep = size_sum - static_size - reserved_size;
1315
1316 return size_sum;
1317}
1318
Tejun Heofbf59bc2009-02-20 16:29:08 +09001319/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001320 * pcpu_alloc_alloc_info - allocate percpu allocation info
1321 * @nr_groups: the number of groups
1322 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001323 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001324 * Allocate ai which is large enough for @nr_groups groups containing
1325 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1326 * cpu_map array which is long enough for @nr_units and filled with
1327 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1328 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001329 *
1330 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001331 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1332 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001333 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001334struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1335 int nr_units)
1336{
1337 struct pcpu_alloc_info *ai;
1338 size_t base_size, ai_size;
1339 void *ptr;
1340 int unit;
1341
1342 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1343 __alignof__(ai->groups[0].cpu_map[0]));
1344 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1345
1346 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1347 if (!ptr)
1348 return NULL;
1349 ai = ptr;
1350 ptr += base_size;
1351
1352 ai->groups[0].cpu_map = ptr;
1353
1354 for (unit = 0; unit < nr_units; unit++)
1355 ai->groups[0].cpu_map[unit] = NR_CPUS;
1356
1357 ai->nr_groups = nr_groups;
1358 ai->__ai_size = PFN_ALIGN(ai_size);
1359
1360 return ai;
1361}
1362
1363/**
1364 * pcpu_free_alloc_info - free percpu allocation info
1365 * @ai: pcpu_alloc_info to free
1366 *
1367 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1368 */
1369void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1370{
1371 free_bootmem(__pa(ai), ai->__ai_size);
1372}
1373
1374/**
1375 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
Tejun Heoedcb4632009-03-06 14:33:59 +09001376 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +09001377 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heofd1e8a12009-08-14 15:00:51 +09001378 * @atom_size: allocation atom size
1379 * @cpu_distance_fn: callback to determine distance between cpus, optional
1380 *
1381 * This function determines grouping of units, their mappings to cpus
1382 * and other parameters considering needed percpu size, allocation
1383 * atom size and distances between CPUs.
1384 *
1385 * Groups are always mutliples of atom size and CPUs which are of
1386 * LOCAL_DISTANCE both ways are grouped together and share space for
1387 * units in the same group. The returned configuration is guaranteed
1388 * to have CPUs on different nodes on different groups and >=75% usage
1389 * of allocated virtual address space.
1390 *
1391 * RETURNS:
1392 * On success, pointer to the new allocation_info is returned. On
1393 * failure, ERR_PTR value is returned.
1394 */
1395struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1396 size_t reserved_size, ssize_t dyn_size,
1397 size_t atom_size,
1398 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
Tejun Heo033e48f2009-08-14 15:00:51 +09001399{
1400 static int group_map[NR_CPUS] __initdata;
1401 static int group_cnt[NR_CPUS] __initdata;
1402 const size_t static_size = __per_cpu_end - __per_cpu_start;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001403 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
Tejun Heo033e48f2009-08-14 15:00:51 +09001404 size_t size_sum, min_unit_size, alloc_size;
1405 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001406 int last_allocs, group, unit;
Tejun Heo033e48f2009-08-14 15:00:51 +09001407 unsigned int cpu, tcpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001408 struct pcpu_alloc_info *ai;
1409 unsigned int *cpu_map;
Tejun Heo033e48f2009-08-14 15:00:51 +09001410
Tejun Heofb59e722009-09-24 18:50:34 +09001411 /* this function may be called multiple times */
1412 memset(group_map, 0, sizeof(group_map));
1413 memset(group_cnt, 0, sizeof(group_map));
1414
Tejun Heo033e48f2009-08-14 15:00:51 +09001415 /*
1416 * Determine min_unit_size, alloc_size and max_upa such that
Tejun Heofd1e8a12009-08-14 15:00:51 +09001417 * alloc_size is multiple of atom_size and is the smallest
Tejun Heo033e48f2009-08-14 15:00:51 +09001418 * which can accomodate 4k aligned segments which are equal to
1419 * or larger than min_unit_size.
1420 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001421 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001422 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1423
Tejun Heofd1e8a12009-08-14 15:00:51 +09001424 alloc_size = roundup(min_unit_size, atom_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001425 upa = alloc_size / min_unit_size;
1426 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1427 upa--;
1428 max_upa = upa;
1429
1430 /* group cpus according to their proximity */
1431 for_each_possible_cpu(cpu) {
1432 group = 0;
1433 next_group:
1434 for_each_possible_cpu(tcpu) {
1435 if (cpu == tcpu)
1436 break;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001437 if (group_map[tcpu] == group && cpu_distance_fn &&
Tejun Heo033e48f2009-08-14 15:00:51 +09001438 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1439 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1440 group++;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001441 nr_groups = max(nr_groups, group + 1);
Tejun Heo033e48f2009-08-14 15:00:51 +09001442 goto next_group;
1443 }
1444 }
1445 group_map[cpu] = group;
1446 group_cnt[group]++;
1447 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1448 }
1449
1450 /*
1451 * Expand unit size until address space usage goes over 75%
1452 * and then as much as possible without using more address
1453 * space.
1454 */
1455 last_allocs = INT_MAX;
1456 for (upa = max_upa; upa; upa--) {
1457 int allocs = 0, wasted = 0;
1458
1459 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1460 continue;
1461
Tejun Heofd1e8a12009-08-14 15:00:51 +09001462 for (group = 0; group < nr_groups; group++) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001463 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1464 allocs += this_allocs;
1465 wasted += this_allocs * upa - group_cnt[group];
1466 }
1467
1468 /*
1469 * Don't accept if wastage is over 25%. The
1470 * greater-than comparison ensures upa==1 always
1471 * passes the following check.
1472 */
1473 if (wasted > num_possible_cpus() / 3)
1474 continue;
1475
1476 /* and then don't consume more memory */
1477 if (allocs > last_allocs)
1478 break;
1479 last_allocs = allocs;
1480 best_upa = upa;
1481 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001482 upa = best_upa;
Tejun Heo033e48f2009-08-14 15:00:51 +09001483
Tejun Heofd1e8a12009-08-14 15:00:51 +09001484 /* allocate and fill alloc_info */
1485 for (group = 0; group < nr_groups; group++)
1486 nr_units += roundup(group_cnt[group], upa);
1487
1488 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1489 if (!ai)
1490 return ERR_PTR(-ENOMEM);
1491 cpu_map = ai->groups[0].cpu_map;
1492
1493 for (group = 0; group < nr_groups; group++) {
1494 ai->groups[group].cpu_map = cpu_map;
1495 cpu_map += roundup(group_cnt[group], upa);
Tejun Heo033e48f2009-08-14 15:00:51 +09001496 }
1497
Tejun Heofd1e8a12009-08-14 15:00:51 +09001498 ai->static_size = static_size;
1499 ai->reserved_size = reserved_size;
1500 ai->dyn_size = dyn_size;
1501 ai->unit_size = alloc_size / upa;
1502 ai->atom_size = atom_size;
1503 ai->alloc_size = alloc_size;
1504
1505 for (group = 0, unit = 0; group_cnt[group]; group++) {
1506 struct pcpu_group_info *gi = &ai->groups[group];
1507
1508 /*
1509 * Initialize base_offset as if all groups are located
1510 * back-to-back. The caller should update this to
1511 * reflect actual allocation.
1512 */
1513 gi->base_offset = unit * ai->unit_size;
1514
1515 for_each_possible_cpu(cpu)
1516 if (group_map[cpu] == group)
1517 gi->cpu_map[gi->nr_units++] = cpu;
1518 gi->nr_units = roundup(gi->nr_units, upa);
1519 unit += gi->nr_units;
1520 }
1521 BUG_ON(unit != nr_units);
1522
1523 return ai;
Tejun Heo033e48f2009-08-14 15:00:51 +09001524}
1525
Tejun Heofd1e8a12009-08-14 15:00:51 +09001526/**
1527 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1528 * @lvl: loglevel
1529 * @ai: allocation info to dump
1530 *
1531 * Print out information about @ai using loglevel @lvl.
1532 */
1533static void pcpu_dump_alloc_info(const char *lvl,
1534 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001535{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001536 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001537 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001538 int alloc = 0, alloc_end = 0;
1539 int group, v;
1540 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001541
Tejun Heofd1e8a12009-08-14 15:00:51 +09001542 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001543 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001544 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001545
Tejun Heofd1e8a12009-08-14 15:00:51 +09001546 v = num_possible_cpus();
1547 while (v /= 10)
1548 cpu_width++;
1549 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001550
Tejun Heofd1e8a12009-08-14 15:00:51 +09001551 upa = ai->alloc_size / ai->unit_size;
1552 width = upa * (cpu_width + 1) + group_width + 3;
1553 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001554
Tejun Heofd1e8a12009-08-14 15:00:51 +09001555 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1556 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1557 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1558
1559 for (group = 0; group < ai->nr_groups; group++) {
1560 const struct pcpu_group_info *gi = &ai->groups[group];
1561 int unit = 0, unit_end = 0;
1562
1563 BUG_ON(gi->nr_units % upa);
1564 for (alloc_end += gi->nr_units / upa;
1565 alloc < alloc_end; alloc++) {
1566 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001567 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001568 printk("%spcpu-alloc: ", lvl);
1569 }
1570 printk("[%0*d] ", group_width, group);
1571
1572 for (unit_end += upa; unit < unit_end; unit++)
1573 if (gi->cpu_map[unit] != NR_CPUS)
1574 printk("%0*d ", cpu_width,
1575 gi->cpu_map[unit]);
1576 else
1577 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001578 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001579 }
1580 printk("\n");
1581}
Tejun Heo033e48f2009-08-14 15:00:51 +09001582
Tejun Heofbf59bc2009-02-20 16:29:08 +09001583/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001584 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001585 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001586 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001587 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001588 * Initialize the first percpu chunk which contains the kernel static
1589 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001590 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001591 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001592 * @ai contains all information necessary to initialize the first
1593 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001594 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001595 * @ai->static_size is the size of static percpu area.
1596 *
1597 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001598 * reserve after the static area in the first chunk. This reserves
1599 * the first chunk such that it's available only through reserved
1600 * percpu allocation. This is primarily used to serve module percpu
1601 * static areas on architectures where the addressing model has
1602 * limited offset range for symbol relocations to guarantee module
1603 * percpu symbols fall inside the relocatable range.
1604 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001605 * @ai->dyn_size determines the number of bytes available for dynamic
1606 * allocation in the first chunk. The area between @ai->static_size +
1607 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001608 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001609 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1610 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1611 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001612 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001613 * @ai->atom_size is the allocation atom size and used as alignment
1614 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001615 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001616 * @ai->alloc_size is the allocation size and always multiple of
1617 * @ai->atom_size. This is larger than @ai->atom_size if
1618 * @ai->unit_size is larger than @ai->atom_size.
1619 *
1620 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1621 * percpu areas. Units which should be colocated are put into the
1622 * same group. Dynamic VM areas will be allocated according to these
1623 * groupings. If @ai->nr_groups is zero, a single group containing
1624 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001625 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001626 * The caller should have mapped the first chunk at @base_addr and
1627 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001628 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001629 * If the first chunk ends up with both reserved and dynamic areas, it
1630 * is served by two chunks - one to serve the core static and reserved
1631 * areas and the other for the dynamic area. They share the same vm
1632 * and page map but uses different area allocation map to stay away
1633 * from each other. The latter chunk is circulated in the chunk slots
1634 * and available for dynamic allocation like any other chunks.
1635 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001636 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001637 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001638 */
Tejun Heofb435d52009-08-14 15:00:51 +09001639int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1640 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001641{
Tejun Heo635b75f2009-09-24 09:43:11 +09001642 static char cpus_buf[4096] __initdata;
Tejun Heoedcb4632009-03-06 14:33:59 +09001643 static int smap[2], dmap[2];
Tejun Heofd1e8a12009-08-14 15:00:51 +09001644 size_t dyn_size = ai->dyn_size;
1645 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001646 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001647 unsigned long *group_offsets;
1648 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001649 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001650 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001651 int *unit_map;
1652 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001653
Tejun Heo635b75f2009-09-24 09:43:11 +09001654 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1655
1656#define PCPU_SETUP_BUG_ON(cond) do { \
1657 if (unlikely(cond)) { \
1658 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1659 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1660 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1661 BUG(); \
1662 } \
1663} while (0)
1664
Tejun Heo2f39e632009-07-04 08:11:00 +09001665 /* sanity checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001666 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1667 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo635b75f2009-09-24 09:43:11 +09001668 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1669 PCPU_SETUP_BUG_ON(!ai->static_size);
1670 PCPU_SETUP_BUG_ON(!base_addr);
1671 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1672 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1673 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001674
Tejun Heo65632972009-08-14 15:00:52 +09001675 /* process group information and build config tables accordingly */
1676 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1677 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001678 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001679 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001680
Tejun Heofd1e8a12009-08-14 15:00:51 +09001681 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001682 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001683 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001684
Tejun Heofd1e8a12009-08-14 15:00:51 +09001685 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1686 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001687
Tejun Heo65632972009-08-14 15:00:52 +09001688 group_offsets[group] = gi->base_offset;
1689 group_sizes[group] = gi->nr_units * ai->unit_size;
1690
Tejun Heofd1e8a12009-08-14 15:00:51 +09001691 for (i = 0; i < gi->nr_units; i++) {
1692 cpu = gi->cpu_map[i];
1693 if (cpu == NR_CPUS)
1694 continue;
1695
Tejun Heo635b75f2009-09-24 09:43:11 +09001696 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1697 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1698 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001699
1700 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001701 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1702
Tejun Heofd1e8a12009-08-14 15:00:51 +09001703 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001704 pcpu_first_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001705 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001706 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001707 pcpu_last_unit_cpu = cpu;
1708 pcpu_nr_units = unit;
1709
1710 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001711 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1712
1713 /* we're done parsing the input, undefine BUG macro and dump config */
1714#undef PCPU_SETUP_BUG_ON
1715 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001716
Tejun Heo65632972009-08-14 15:00:52 +09001717 pcpu_nr_groups = ai->nr_groups;
1718 pcpu_group_offsets = group_offsets;
1719 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001720 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001721 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001722
1723 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001724 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001725 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001726 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001727 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1728 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001729
Tejun Heod9b55ee2009-02-24 11:57:21 +09001730 /*
1731 * Allocate chunk slots. The additional last slot is for
1732 * empty chunks.
1733 */
1734 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001735 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1736 for (i = 0; i < pcpu_nr_slots; i++)
1737 INIT_LIST_HEAD(&pcpu_slot[i]);
1738
Tejun Heoedcb4632009-03-06 14:33:59 +09001739 /*
1740 * Initialize static chunk. If reserved_size is zero, the
1741 * static chunk covers static area + dynamic allocation area
1742 * in the first chunk. If reserved_size is not zero, it
1743 * covers static area + reserved area (mostly used for module
1744 * static percpu allocation).
1745 */
Tejun Heo2441d152009-03-06 14:33:59 +09001746 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1747 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001748 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001749 schunk->map = smap;
1750 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001751 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001752 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001753
Tejun Heofd1e8a12009-08-14 15:00:51 +09001754 if (ai->reserved_size) {
1755 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001756 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001757 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001758 } else {
1759 schunk->free_size = dyn_size;
1760 dyn_size = 0; /* dynamic area covered */
1761 }
Tejun Heo2441d152009-03-06 14:33:59 +09001762 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001763
Tejun Heofd1e8a12009-08-14 15:00:51 +09001764 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001765 if (schunk->free_size)
1766 schunk->map[schunk->map_used++] = schunk->free_size;
1767
Tejun Heoedcb4632009-03-06 14:33:59 +09001768 /* init dynamic chunk if necessary */
1769 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001770 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001771 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001772 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001773 dchunk->map = dmap;
1774 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001775 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001776 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001777
1778 dchunk->contig_hint = dchunk->free_size = dyn_size;
1779 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1780 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1781 }
1782
Tejun Heo2441d152009-03-06 14:33:59 +09001783 /* link the first chunk in */
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001784 pcpu_first_chunk = dchunk ?: schunk;
1785 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001786
1787 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001788 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001789 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001790}
Tejun Heo66c3a752009-03-10 16:27:48 +09001791
Tejun Heof58dc012009-08-14 15:00:50 +09001792const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1793 [PCPU_FC_AUTO] = "auto",
1794 [PCPU_FC_EMBED] = "embed",
1795 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001796};
Tejun Heo66c3a752009-03-10 16:27:48 +09001797
Tejun Heof58dc012009-08-14 15:00:50 +09001798enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1799
1800static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001801{
Tejun Heof58dc012009-08-14 15:00:50 +09001802 if (0)
1803 /* nada */;
1804#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1805 else if (!strcmp(str, "embed"))
1806 pcpu_chosen_fc = PCPU_FC_EMBED;
1807#endif
1808#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1809 else if (!strcmp(str, "page"))
1810 pcpu_chosen_fc = PCPU_FC_PAGE;
1811#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001812 else
1813 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001814
Tejun Heof58dc012009-08-14 15:00:50 +09001815 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001816}
Tejun Heof58dc012009-08-14 15:00:50 +09001817early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001818
Tejun Heo08fc4582009-08-14 15:00:49 +09001819#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1820 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo66c3a752009-03-10 16:27:48 +09001821/**
1822 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001823 * @reserved_size: the size of reserved percpu area in bytes
1824 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heoc8826dd2009-08-14 15:00:52 +09001825 * @atom_size: allocation atom size
1826 * @cpu_distance_fn: callback to determine distance between cpus, optional
1827 * @alloc_fn: function to allocate percpu page
1828 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001829 *
1830 * This is a helper to ease setting up embedded first percpu chunk and
1831 * can be called where pcpu_setup_first_chunk() is expected.
1832 *
1833 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001834 * by calling @alloc_fn and used as-is without being mapped into
1835 * vmalloc area. Allocations are always whole multiples of @atom_size
1836 * aligned to @atom_size.
1837 *
1838 * This enables the first chunk to piggy back on the linear physical
1839 * mapping which often uses larger page size. Please note that this
1840 * can result in very sparse cpu->unit mapping on NUMA machines thus
1841 * requiring large vmalloc address space. Don't use this allocator if
1842 * vmalloc space is not orders of magnitude larger than distances
1843 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001844 *
1845 * When @dyn_size is positive, dynamic area might be larger than
Tejun Heo788e5ab2009-07-04 08:10:58 +09001846 * specified to fill page alignment. When @dyn_size is auto,
1847 * @dyn_size is just big enough to fill page alignment after static
1848 * and reserved areas.
Tejun Heo66c3a752009-03-10 16:27:48 +09001849 *
1850 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001851 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001852 *
1853 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001854 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001855 */
Tejun Heoc8826dd2009-08-14 15:00:52 +09001856int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1857 size_t atom_size,
1858 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1859 pcpu_fc_alloc_fn_t alloc_fn,
1860 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001861{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001862 void *base = (void *)ULONG_MAX;
1863 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001864 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001865 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001866 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001867
Tejun Heoc8826dd2009-08-14 15:00:52 +09001868 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1869 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001870 if (IS_ERR(ai))
1871 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001872
Tejun Heofd1e8a12009-08-14 15:00:51 +09001873 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001874 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001875
Tejun Heoc8826dd2009-08-14 15:00:52 +09001876 areas = alloc_bootmem_nopanic(areas_size);
1877 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001878 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001879 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001880 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001881
Tejun Heoc8826dd2009-08-14 15:00:52 +09001882 /* allocate, copy and determine base address */
1883 for (group = 0; group < ai->nr_groups; group++) {
1884 struct pcpu_group_info *gi = &ai->groups[group];
1885 unsigned int cpu = NR_CPUS;
1886 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001887
Tejun Heoc8826dd2009-08-14 15:00:52 +09001888 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1889 cpu = gi->cpu_map[i];
1890 BUG_ON(cpu == NR_CPUS);
1891
1892 /* allocate space for the whole group */
1893 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1894 if (!ptr) {
1895 rc = -ENOMEM;
1896 goto out_free_areas;
1897 }
1898 areas[group] = ptr;
1899
1900 base = min(ptr, base);
1901
1902 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1903 if (gi->cpu_map[i] == NR_CPUS) {
1904 /* unused unit, free whole */
1905 free_fn(ptr, ai->unit_size);
1906 continue;
1907 }
1908 /* copy and return the unused part */
1909 memcpy(ptr, __per_cpu_load, ai->static_size);
1910 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1911 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001912 }
1913
Tejun Heoc8826dd2009-08-14 15:00:52 +09001914 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001915 max_distance = 0;
1916 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001917 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001918 max_distance = max_t(size_t, max_distance,
1919 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001920 }
1921 max_distance += ai->unit_size;
1922
1923 /* warn if maximum distance is further than 75% of vmalloc space */
1924 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001925 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001926 "space 0x%lx\n",
1927 max_distance, VMALLOC_END - VMALLOC_START);
1928#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1929 /* and fail if we have fallback */
1930 rc = -EINVAL;
1931 goto out_free;
1932#endif
1933 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001934
Tejun Heo004018e2009-08-14 15:00:49 +09001935 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001936 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1937 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001938
Tejun Heofb435d52009-08-14 15:00:51 +09001939 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001940 goto out_free;
1941
1942out_free_areas:
1943 for (group = 0; group < ai->nr_groups; group++)
1944 free_fn(areas[group],
1945 ai->groups[group].nr_units * ai->unit_size);
1946out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001947 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001948 if (areas)
1949 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001950 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001951}
Tejun Heo08fc4582009-08-14 15:00:49 +09001952#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1953 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heod4b95f82009-07-04 08:10:59 +09001954
Tejun Heo08fc4582009-08-14 15:00:49 +09001955#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001956/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001957 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001958 * @reserved_size: the size of reserved percpu area in bytes
1959 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1960 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1961 * @populate_pte_fn: function to populate pte
1962 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001963 * This is a helper to ease setting up page-remapped first percpu
1964 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001965 *
1966 * This is the basic allocator. Static percpu area is allocated
1967 * page-by-page into vmalloc area.
1968 *
1969 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001970 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001971 */
Tejun Heofb435d52009-08-14 15:00:51 +09001972int __init pcpu_page_first_chunk(size_t reserved_size,
1973 pcpu_fc_alloc_fn_t alloc_fn,
1974 pcpu_fc_free_fn_t free_fn,
1975 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001976{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001977 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001978 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001979 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001980 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001981 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001982 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001983 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001984
Tejun Heo00ae4062009-08-14 15:00:49 +09001985 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1986
Tejun Heofd1e8a12009-08-14 15:00:51 +09001987 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1988 if (IS_ERR(ai))
1989 return PTR_ERR(ai);
1990 BUG_ON(ai->nr_groups != 1);
1991 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1992
1993 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001994
1995 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001996 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1997 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09001998 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09001999
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002000 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09002001 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002002 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09002003 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09002004 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09002005 void *ptr;
2006
Tejun Heo3cbc8562009-08-14 15:00:50 +09002007 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09002008 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09002009 pr_warning("PERCPU: failed to allocate %s page "
2010 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09002011 goto enomem;
2012 }
Tejun Heoce3141a2009-07-04 08:11:00 +09002013 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002014 }
2015
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002016 /* allocate vm area, map the pages and copy static data */
2017 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002018 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002019 vm_area_register_early(&vm, PAGE_SIZE);
2020
Tejun Heofd1e8a12009-08-14 15:00:51 +09002021 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09002022 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09002023 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002024
Tejun Heoce3141a2009-07-04 08:11:00 +09002025 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002026 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2027
2028 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09002029 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2030 unit_pages);
2031 if (rc < 0)
2032 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002033
2034 /*
2035 * FIXME: Archs with virtual cache should flush local
2036 * cache for the linear mapping here - something
2037 * equivalent to flush_cache_vmap() on the local cpu.
2038 * flush_cache_vmap() can't be used as most supporting
2039 * data structures are not set up yet.
2040 */
2041
2042 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002043 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002044 }
2045
2046 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09002047 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002048 unit_pages, psize_str, vm.addr, ai->static_size,
2049 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002050
Tejun Heofb435d52009-08-14 15:00:51 +09002051 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002052 goto out_free_ar;
2053
2054enomem:
2055 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002056 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002057 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002058out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09002059 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002060 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002061 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002062}
Tejun Heo08fc4582009-08-14 15:00:49 +09002063#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002064
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002065/*
Tejun Heoe74e3962009-03-30 19:07:44 +09002066 * Generic percpu area setup.
2067 *
2068 * The embedding helper is used because its behavior closely resembles
2069 * the original non-dynamic generic percpu area setup. This is
2070 * important because many archs have addressing restrictions and might
2071 * fail if the percpu area is located far away from the previous
2072 * location. As an added bonus, in non-NUMA cases, embedding is
2073 * generally a good idea TLB-wise because percpu area can piggy back
2074 * on the physical linear memory mapping which uses large page
2075 * mappings on applicable archs.
2076 */
2077#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2078unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2079EXPORT_SYMBOL(__per_cpu_offset);
2080
Tejun Heoc8826dd2009-08-14 15:00:52 +09002081static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2082 size_t align)
2083{
2084 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2085}
2086
2087static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2088{
2089 free_bootmem(__pa(ptr), size);
2090}
2091
Tejun Heoe74e3962009-03-30 19:07:44 +09002092void __init setup_per_cpu_areas(void)
2093{
Tejun Heoe74e3962009-03-30 19:07:44 +09002094 unsigned long delta;
2095 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002096 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002097
2098 /*
2099 * Always reserve area for module percpu variables. That's
2100 * what the legacy allocator did.
2101 */
Tejun Heofb435d52009-08-14 15:00:51 +09002102 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002103 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2104 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002105 if (rc < 0)
Tejun Heoe74e3962009-03-30 19:07:44 +09002106 panic("Failed to initialized percpu areas.");
2107
2108 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2109 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002110 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002111}
2112#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */