blob: e61dc2cc5873d7d89efc45609a4f4f2a8ae75732 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
Tejun Heo88999a82010-04-09 18:57:01 +09002 * mm/percpu.c - percpu memory allocator
Tejun Heofbf59bc2009-02-20 16:29:08 +09003 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
Tejun Heo88999a82010-04-09 18:57:01 +090010 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
Tejun Heo2f39e632009-07-04 08:11:00 +090013 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
Tejun Heo88999a82010-04-09 18:57:01 +090016 * When a chunk is filled up, another chunk is allocated.
Tejun Heofbf59bc2009-02-20 16:29:08 +090017 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
Tejun Heo2f39e632009-07-04 08:11:00 +090025 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090029 *
Tejun Heo2f39e632009-07-04 08:11:00 +090030 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090032 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be eqaul to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090043 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090045 *
46 * To use this allocator, arch code should do the followings.
47 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090048 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090058#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090059#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090060#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090061#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090066#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090067#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090068#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090069#include <linux/workqueue.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090070
71#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090072#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090073#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090074#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090075
Tejun Heofbf59bc2009-02-20 16:29:08 +090076#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
77#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
78
Tejun Heoe0100982009-03-10 16:27:48 +090079/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
80#ifndef __addr_to_pcpu_ptr
81#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090082 (void __percpu *)((unsigned long)(addr) - \
83 (unsigned long)pcpu_base_addr + \
84 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090085#endif
86#ifndef __pcpu_ptr_to_addr
87#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090088 (void __force *)((unsigned long)(ptr) + \
89 (unsigned long)pcpu_base_addr - \
90 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090091#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 Heo88999a82010-04-09 18:57:01 +0900101 void *data; /* chunk data */
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 Heo020ec652010-04-09 18:57:00 +0900179static bool pcpu_addr_in_first_chunk(void *addr)
180{
181 void *first_start = pcpu_first_chunk->base_addr;
182
183 return addr >= first_start && addr < first_start + pcpu_unit_size;
184}
185
186static bool pcpu_addr_in_reserved_chunk(void *addr)
187{
188 void *first_start = pcpu_first_chunk->base_addr;
189
190 return addr >= first_start &&
191 addr < first_start + pcpu_reserved_chunk_limit;
192}
193
Tejun Heod9b55ee2009-02-24 11:57:21 +0900194static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900195{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900196 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900197 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
198}
199
Tejun Heod9b55ee2009-02-24 11:57:21 +0900200static int pcpu_size_to_slot(int size)
201{
202 if (size == pcpu_unit_size)
203 return pcpu_nr_slots - 1;
204 return __pcpu_size_to_slot(size);
205}
206
Tejun Heofbf59bc2009-02-20 16:29:08 +0900207static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
208{
209 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
210 return 0;
211
212 return pcpu_size_to_slot(chunk->free_size);
213}
214
Tejun Heo88999a82010-04-09 18:57:01 +0900215/* set the pointer to a chunk in a page struct */
216static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
217{
218 page->index = (unsigned long)pcpu;
219}
220
221/* obtain pointer to a chunk from a page struct */
222static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
223{
224 return (struct pcpu_chunk *)page->index;
225}
226
227static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900228{
Tejun Heo2f39e632009-07-04 08:11:00 +0900229 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900230}
231
Tejun Heo9983b6f2010-06-18 11:44:31 +0200232static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
233 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900234{
Tejun Heobba174f2009-08-14 15:00:51 +0900235 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900236 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900237}
238
Tejun Heo88999a82010-04-09 18:57:01 +0900239static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
240 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900241{
242 *rs = find_next_zero_bit(chunk->populated, end, *rs);
243 *re = find_next_bit(chunk->populated, end, *rs + 1);
244}
245
Tejun Heo88999a82010-04-09 18:57:01 +0900246static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
247 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900248{
249 *rs = find_next_bit(chunk->populated, end, *rs);
250 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
251}
252
253/*
254 * (Un)populated page region iterators. Iterate over (un)populated
255 * page regions betwen @start and @end in @chunk. @rs and @re should
256 * be integer variables and will be set to start and end page index of
257 * the current region.
258 */
259#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
260 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
261 (rs) < (re); \
262 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
263
264#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
265 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
266 (rs) < (re); \
267 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
268
Tejun Heofbf59bc2009-02-20 16:29:08 +0900269/**
Tejun Heo1880d932009-03-07 00:44:09 +0900270 * pcpu_mem_alloc - allocate memory
271 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900272 *
Tejun Heo1880d932009-03-07 00:44:09 +0900273 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
274 * kzalloc() is used; otherwise, vmalloc() is used. The returned
275 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900276 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900277 * CONTEXT:
278 * Does GFP_KERNEL allocation.
279 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900280 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900281 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900282 */
Tejun Heo1880d932009-03-07 00:44:09 +0900283static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900284{
Tejun Heo099a19d2010-06-27 18:50:00 +0200285 if (WARN_ON_ONCE(!slab_is_available()))
286 return NULL;
287
Tejun Heofbf59bc2009-02-20 16:29:08 +0900288 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900289 return kzalloc(size, GFP_KERNEL);
290 else {
291 void *ptr = vmalloc(size);
292 if (ptr)
293 memset(ptr, 0, size);
294 return ptr;
295 }
296}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900297
Tejun Heo1880d932009-03-07 00:44:09 +0900298/**
299 * pcpu_mem_free - free memory
300 * @ptr: memory to free
301 * @size: size of the area
302 *
303 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
304 */
305static void pcpu_mem_free(void *ptr, size_t size)
306{
307 if (size <= PAGE_SIZE)
308 kfree(ptr);
309 else
310 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900311}
312
313/**
314 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
315 * @chunk: chunk of interest
316 * @oslot: the previous slot it was on
317 *
318 * This function is called after an allocation or free changed @chunk.
319 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900320 * moved to the slot. Note that the reserved chunk is never put on
321 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900322 *
323 * CONTEXT:
324 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900325 */
326static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
327{
328 int nslot = pcpu_chunk_slot(chunk);
329
Tejun Heoedcb4632009-03-06 14:33:59 +0900330 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900331 if (oslot < nslot)
332 list_move(&chunk->list, &pcpu_slot[nslot]);
333 else
334 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
335 }
336}
337
Tejun Heofbf59bc2009-02-20 16:29:08 +0900338/**
Tejun Heo833af842009-11-11 15:35:18 +0900339 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
340 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900341 *
Tejun Heo833af842009-11-11 15:35:18 +0900342 * Determine whether area map of @chunk needs to be extended to
343 * accomodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900344 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900345 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900346 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900347 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900348 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900349 * New target map allocation length if extension is necessary, 0
350 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900351 */
Tejun Heo833af842009-11-11 15:35:18 +0900352static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900353{
354 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900355
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900356 if (chunk->map_alloc >= chunk->map_used + 2)
357 return 0;
358
359 new_alloc = PCPU_DFL_MAP_ALLOC;
360 while (new_alloc < chunk->map_used + 2)
361 new_alloc *= 2;
362
Tejun Heo833af842009-11-11 15:35:18 +0900363 return new_alloc;
364}
365
366/**
367 * pcpu_extend_area_map - extend area map of a chunk
368 * @chunk: chunk of interest
369 * @new_alloc: new target allocation length of the area map
370 *
371 * Extend area map of @chunk to have @new_alloc entries.
372 *
373 * CONTEXT:
374 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
375 *
376 * RETURNS:
377 * 0 on success, -errno on failure.
378 */
379static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
380{
381 int *old = NULL, *new = NULL;
382 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
383 unsigned long flags;
384
385 new = pcpu_mem_alloc(new_size);
386 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900387 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900388
Tejun Heo833af842009-11-11 15:35:18 +0900389 /* acquire pcpu_lock and switch to new area map */
390 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900391
Tejun Heo833af842009-11-11 15:35:18 +0900392 if (new_alloc <= chunk->map_alloc)
393 goto out_unlock;
394
395 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
396 memcpy(new, chunk->map, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900397
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900398 chunk->map_alloc = new_alloc;
399 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900400 new = NULL;
401
402out_unlock:
403 spin_unlock_irqrestore(&pcpu_lock, flags);
404
405 /*
406 * pcpu_mem_free() might end up calling vfree() which uses
407 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
408 */
409 pcpu_mem_free(old, old_size);
410 pcpu_mem_free(new, new_size);
411
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900412 return 0;
413}
414
415/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900416 * pcpu_split_block - split a map block
417 * @chunk: chunk of interest
418 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900419 * @head: head size in bytes (can be 0)
420 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900421 *
422 * Split the @i'th map block into two or three blocks. If @head is
423 * non-zero, @head bytes block is inserted before block @i moving it
424 * to @i+1 and reducing its size by @head bytes.
425 *
426 * If @tail is non-zero, the target block, which can be @i or @i+1
427 * depending on @head, is reduced by @tail bytes and @tail byte block
428 * is inserted after the target block.
429 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900430 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900431 *
432 * CONTEXT:
433 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900434 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900435static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
436 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900437{
438 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900439
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900440 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900441
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900442 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900443 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
444 sizeof(chunk->map[0]) * (chunk->map_used - i));
445 chunk->map_used += nr_extra;
446
447 if (head) {
448 chunk->map[i + 1] = chunk->map[i] - head;
449 chunk->map[i++] = head;
450 }
451 if (tail) {
452 chunk->map[i++] -= tail;
453 chunk->map[i] = tail;
454 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900455}
456
457/**
458 * pcpu_alloc_area - allocate area from a pcpu_chunk
459 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900460 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900461 * @align: wanted align
462 *
463 * Try to allocate @size bytes area aligned at @align from @chunk.
464 * Note that this function only allocates the offset. It doesn't
465 * populate or map the area.
466 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900467 * @chunk->map must have at least two free slots.
468 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900469 * CONTEXT:
470 * pcpu_lock.
471 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900472 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900473 * Allocated offset in @chunk on success, -1 if no matching area is
474 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900475 */
476static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
477{
478 int oslot = pcpu_chunk_slot(chunk);
479 int max_contig = 0;
480 int i, off;
481
Tejun Heofbf59bc2009-02-20 16:29:08 +0900482 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
483 bool is_last = i + 1 == chunk->map_used;
484 int head, tail;
485
486 /* extra for alignment requirement */
487 head = ALIGN(off, align) - off;
488 BUG_ON(i == 0 && head != 0);
489
490 if (chunk->map[i] < 0)
491 continue;
492 if (chunk->map[i] < head + size) {
493 max_contig = max(chunk->map[i], max_contig);
494 continue;
495 }
496
497 /*
498 * If head is small or the previous block is free,
499 * merge'em. Note that 'small' is defined as smaller
500 * than sizeof(int), which is very small but isn't too
501 * uncommon for percpu allocations.
502 */
503 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
504 if (chunk->map[i - 1] > 0)
505 chunk->map[i - 1] += head;
506 else {
507 chunk->map[i - 1] -= head;
508 chunk->free_size -= head;
509 }
510 chunk->map[i] -= head;
511 off += head;
512 head = 0;
513 }
514
515 /* if tail is small, just keep it around */
516 tail = chunk->map[i] - head - size;
517 if (tail < sizeof(int))
518 tail = 0;
519
520 /* split if warranted */
521 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900522 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900523 if (head) {
524 i++;
525 off += head;
526 max_contig = max(chunk->map[i - 1], max_contig);
527 }
528 if (tail)
529 max_contig = max(chunk->map[i + 1], max_contig);
530 }
531
532 /* update hint and mark allocated */
533 if (is_last)
534 chunk->contig_hint = max_contig; /* fully scanned */
535 else
536 chunk->contig_hint = max(chunk->contig_hint,
537 max_contig);
538
539 chunk->free_size -= chunk->map[i];
540 chunk->map[i] = -chunk->map[i];
541
542 pcpu_chunk_relocate(chunk, oslot);
543 return off;
544 }
545
546 chunk->contig_hint = max_contig; /* fully scanned */
547 pcpu_chunk_relocate(chunk, oslot);
548
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900549 /* tell the upper layer that this chunk has no matching area */
550 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900551}
552
553/**
554 * pcpu_free_area - free area to a pcpu_chunk
555 * @chunk: chunk of interest
556 * @freeme: offset of area to free
557 *
558 * Free area starting from @freeme to @chunk. Note that this function
559 * only modifies the allocation map. It doesn't depopulate or unmap
560 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900561 *
562 * CONTEXT:
563 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900564 */
565static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
566{
567 int oslot = pcpu_chunk_slot(chunk);
568 int i, off;
569
570 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
571 if (off == freeme)
572 break;
573 BUG_ON(off != freeme);
574 BUG_ON(chunk->map[i] > 0);
575
576 chunk->map[i] = -chunk->map[i];
577 chunk->free_size += chunk->map[i];
578
579 /* merge with previous? */
580 if (i > 0 && chunk->map[i - 1] >= 0) {
581 chunk->map[i - 1] += chunk->map[i];
582 chunk->map_used--;
583 memmove(&chunk->map[i], &chunk->map[i + 1],
584 (chunk->map_used - i) * sizeof(chunk->map[0]));
585 i--;
586 }
587 /* merge with next? */
588 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
589 chunk->map[i] += chunk->map[i + 1];
590 chunk->map_used--;
591 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
592 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
593 }
594
595 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
596 pcpu_chunk_relocate(chunk, oslot);
597}
598
Tejun Heo60810892010-04-09 18:57:01 +0900599static struct pcpu_chunk *pcpu_alloc_chunk(void)
600{
601 struct pcpu_chunk *chunk;
602
Tejun Heo099a19d2010-06-27 18:50:00 +0200603 chunk = pcpu_mem_alloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900604 if (!chunk)
605 return NULL;
606
607 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
608 if (!chunk->map) {
609 kfree(chunk);
610 return NULL;
611 }
612
613 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
614 chunk->map[chunk->map_used++] = pcpu_unit_size;
615
616 INIT_LIST_HEAD(&chunk->list);
617 chunk->free_size = pcpu_unit_size;
618 chunk->contig_hint = pcpu_unit_size;
619
620 return chunk;
621}
622
623static void pcpu_free_chunk(struct pcpu_chunk *chunk)
624{
625 if (!chunk)
626 return;
627 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
628 kfree(chunk);
629}
630
Tejun Heo9f645532010-04-09 18:57:01 +0900631/*
632 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900633 *
Tejun Heo9f645532010-04-09 18:57:01 +0900634 * To allow different implementations, chunk alloc/free and
635 * [de]population are implemented in a separate file which is pulled
636 * into this file and compiled together. The following functions
637 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900638 *
Tejun Heo9f645532010-04-09 18:57:01 +0900639 * pcpu_populate_chunk - populate the specified range of a chunk
640 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
641 * pcpu_create_chunk - create a new chunk
642 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
643 * pcpu_addr_to_page - translate address to physical address
644 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900645 */
Tejun Heo9f645532010-04-09 18:57:01 +0900646static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
647static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
648static struct pcpu_chunk *pcpu_create_chunk(void);
649static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
650static struct page *pcpu_addr_to_page(void *addr);
651static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900652
Tejun Heob0c97782010-04-09 18:57:01 +0900653#ifdef CONFIG_NEED_PER_CPU_KM
654#include "percpu-km.c"
655#else
Tejun Heo9f645532010-04-09 18:57:01 +0900656#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900657#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900658
659/**
Tejun Heo88999a82010-04-09 18:57:01 +0900660 * pcpu_chunk_addr_search - determine chunk containing specified address
661 * @addr: address for which the chunk needs to be determined.
662 *
663 * RETURNS:
664 * The address of the found chunk.
665 */
666static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
667{
668 /* is it in the first chunk? */
669 if (pcpu_addr_in_first_chunk(addr)) {
670 /* is it in the reserved area? */
671 if (pcpu_addr_in_reserved_chunk(addr))
672 return pcpu_reserved_chunk;
673 return pcpu_first_chunk;
674 }
675
676 /*
677 * The address is relative to unit0 which might be unused and
678 * thus unmapped. Offset the address to the unit space of the
679 * current processor before looking it up in the vmalloc
680 * space. Note that any possible cpu id can be used here, so
681 * there's no need to worry about preemption or cpu hotplug.
682 */
683 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900684 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900685}
686
687/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900688 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900689 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900690 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900691 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900692 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900693 * Allocate percpu area of @size bytes aligned at @align.
694 *
695 * CONTEXT:
696 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900697 *
698 * RETURNS:
699 * Percpu pointer to the allocated area on success, NULL on failure.
700 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900701static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900702{
Tejun Heof2badb02009-09-29 09:17:58 +0900703 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900704 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900705 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +0900706 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900707 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900708
Tejun Heo8d408b42009-02-24 11:57:21 +0900709 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900710 WARN(true, "illegal size (%zu) or align (%zu) for "
711 "percpu allocation\n", size, align);
712 return NULL;
713 }
714
Tejun Heoccea34b2009-03-07 00:44:13 +0900715 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +0900716 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900717
Tejun Heoedcb4632009-03-06 14:33:59 +0900718 /* serve reserved allocations from the reserved chunk if available */
719 if (reserved && pcpu_reserved_chunk) {
720 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900721
722 if (size > chunk->contig_hint) {
723 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900724 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900725 }
Tejun Heo833af842009-11-11 15:35:18 +0900726
727 while ((new_alloc = pcpu_need_to_extend(chunk))) {
728 spin_unlock_irqrestore(&pcpu_lock, flags);
729 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
730 err = "failed to extend area map of reserved chunk";
731 goto fail_unlock_mutex;
732 }
733 spin_lock_irqsave(&pcpu_lock, flags);
734 }
735
Tejun Heoedcb4632009-03-06 14:33:59 +0900736 off = pcpu_alloc_area(chunk, size, align);
737 if (off >= 0)
738 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900739
Tejun Heof2badb02009-09-29 09:17:58 +0900740 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900741 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900742 }
743
Tejun Heoccea34b2009-03-07 00:44:13 +0900744restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900745 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900746 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
747 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
748 if (size > chunk->contig_hint)
749 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900750
Tejun Heo833af842009-11-11 15:35:18 +0900751 new_alloc = pcpu_need_to_extend(chunk);
752 if (new_alloc) {
753 spin_unlock_irqrestore(&pcpu_lock, flags);
754 if (pcpu_extend_area_map(chunk,
755 new_alloc) < 0) {
756 err = "failed to extend area map";
757 goto fail_unlock_mutex;
758 }
759 spin_lock_irqsave(&pcpu_lock, flags);
760 /*
761 * pcpu_lock has been dropped, need to
762 * restart cpu_slot list walking.
763 */
764 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900765 }
766
Tejun Heofbf59bc2009-02-20 16:29:08 +0900767 off = pcpu_alloc_area(chunk, size, align);
768 if (off >= 0)
769 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900770 }
771 }
772
773 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900774 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900775
Tejun Heo60810892010-04-09 18:57:01 +0900776 chunk = pcpu_create_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +0900777 if (!chunk) {
778 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +0900779 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +0900780 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900781
Jiri Kosina403a91b2009-10-29 00:25:59 +0900782 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900783 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900784 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900785
786area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900787 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900788
Tejun Heofbf59bc2009-02-20 16:29:08 +0900789 /* populate, map and clear the area */
790 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +0900791 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900792 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +0900793 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +0900794 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900795 }
796
Tejun Heoccea34b2009-03-07 00:44:13 +0900797 mutex_unlock(&pcpu_alloc_mutex);
798
Tejun Heobba174f2009-08-14 15:00:51 +0900799 /* return address relative to base address */
800 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900801
802fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900803 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900804fail_unlock_mutex:
805 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +0900806 if (warn_limit) {
807 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
808 "%s\n", size, align, err);
809 dump_stack();
810 if (!--warn_limit)
811 pr_info("PERCPU: limit reached, disable warning\n");
812 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900813 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900814}
Tejun Heoedcb4632009-03-06 14:33:59 +0900815
816/**
817 * __alloc_percpu - allocate dynamic percpu area
818 * @size: size of area to allocate in bytes
819 * @align: alignment of area (max PAGE_SIZE)
820 *
821 * Allocate percpu area of @size bytes aligned at @align. Might
822 * sleep. Might trigger writeouts.
823 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900824 * CONTEXT:
825 * Does GFP_KERNEL allocation.
826 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900827 * RETURNS:
828 * Percpu pointer to the allocated area on success, NULL on failure.
829 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900830void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900831{
832 return pcpu_alloc(size, align, false);
833}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900834EXPORT_SYMBOL_GPL(__alloc_percpu);
835
Tejun Heoedcb4632009-03-06 14:33:59 +0900836/**
837 * __alloc_reserved_percpu - allocate reserved percpu area
838 * @size: size of area to allocate in bytes
839 * @align: alignment of area (max PAGE_SIZE)
840 *
841 * Allocate percpu area of @size bytes aligned at @align from reserved
842 * percpu area if arch has set it up; otherwise, allocation is served
843 * from the same dynamic area. Might sleep. Might trigger writeouts.
844 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900845 * CONTEXT:
846 * Does GFP_KERNEL allocation.
847 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900848 * RETURNS:
849 * Percpu pointer to the allocated area on success, NULL on failure.
850 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900851void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900852{
853 return pcpu_alloc(size, align, true);
854}
855
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900856/**
857 * pcpu_reclaim - reclaim fully free chunks, workqueue function
858 * @work: unused
859 *
860 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900861 *
862 * CONTEXT:
863 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900864 */
865static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900866{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900867 LIST_HEAD(todo);
868 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
869 struct pcpu_chunk *chunk, *next;
870
Tejun Heoccea34b2009-03-07 00:44:13 +0900871 mutex_lock(&pcpu_alloc_mutex);
872 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900873
874 list_for_each_entry_safe(chunk, next, head, list) {
875 WARN_ON(chunk->immutable);
876
877 /* spare the first one */
878 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
879 continue;
880
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900881 list_move(&chunk->list, &todo);
882 }
883
Tejun Heoccea34b2009-03-07 00:44:13 +0900884 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900885
886 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +0900887 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heo60810892010-04-09 18:57:01 +0900888 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900889 }
Tejun Heo971f3912009-08-14 15:00:49 +0900890
891 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900892}
893
894/**
895 * free_percpu - free percpu area
896 * @ptr: pointer to area to free
897 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900898 * Free percpu area @ptr.
899 *
900 * CONTEXT:
901 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900902 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900903void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900904{
Andrew Morton129182e2010-01-08 14:42:39 -0800905 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900906 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900907 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900908 int off;
909
910 if (!ptr)
911 return;
912
Andrew Morton129182e2010-01-08 14:42:39 -0800913 addr = __pcpu_ptr_to_addr(ptr);
914
Tejun Heoccea34b2009-03-07 00:44:13 +0900915 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900916
917 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +0900918 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900919
920 pcpu_free_area(chunk, off);
921
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900922 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900923 if (chunk->free_size == pcpu_unit_size) {
924 struct pcpu_chunk *pos;
925
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900926 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900927 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900928 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900929 break;
930 }
931 }
932
Tejun Heoccea34b2009-03-07 00:44:13 +0900933 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900934}
935EXPORT_SYMBOL_GPL(free_percpu);
936
Vivek Goyal3b034b02009-11-24 15:50:03 +0900937/**
Tejun Heo10fad5e2010-03-10 18:57:54 +0900938 * is_kernel_percpu_address - test whether address is from static percpu area
939 * @addr: address to test
940 *
941 * Test whether @addr belongs to in-kernel static percpu area. Module
942 * static percpu areas are not considered. For those, use
943 * is_module_percpu_address().
944 *
945 * RETURNS:
946 * %true if @addr is from in-kernel static percpu area, %false otherwise.
947 */
948bool is_kernel_percpu_address(unsigned long addr)
949{
950 const size_t static_size = __per_cpu_end - __per_cpu_start;
951 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
952 unsigned int cpu;
953
954 for_each_possible_cpu(cpu) {
955 void *start = per_cpu_ptr(base, cpu);
956
957 if ((void *)addr >= start && (void *)addr < start + static_size)
958 return true;
959 }
960 return false;
961}
962
963/**
Vivek Goyal3b034b02009-11-24 15:50:03 +0900964 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
965 * @addr: the address to be converted to physical address
966 *
967 * Given @addr which is dereferenceable address obtained via one of
968 * percpu access macros, this function translates it into its physical
969 * address. The caller is responsible for ensuring @addr stays valid
970 * until this function finishes.
971 *
972 * RETURNS:
973 * The physical address for @addr.
974 */
975phys_addr_t per_cpu_ptr_to_phys(void *addr)
976{
Tejun Heo9983b6f2010-06-18 11:44:31 +0200977 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
978 bool in_first_chunk = false;
979 unsigned long first_start, first_end;
980 unsigned int cpu;
981
982 /*
983 * The following test on first_start/end isn't strictly
984 * necessary but will speed up lookups of addresses which
985 * aren't in the first chunk.
986 */
987 first_start = pcpu_chunk_addr(pcpu_first_chunk, pcpu_first_unit_cpu, 0);
988 first_end = pcpu_chunk_addr(pcpu_first_chunk, pcpu_last_unit_cpu,
989 pcpu_unit_pages);
990 if ((unsigned long)addr >= first_start &&
991 (unsigned long)addr < first_end) {
992 for_each_possible_cpu(cpu) {
993 void *start = per_cpu_ptr(base, cpu);
994
995 if (addr >= start && addr < start + pcpu_unit_size) {
996 in_first_chunk = true;
997 break;
998 }
999 }
1000 }
1001
1002 if (in_first_chunk) {
Tejun Heo020ec652010-04-09 18:57:00 +09001003 if ((unsigned long)addr < VMALLOC_START ||
1004 (unsigned long)addr >= VMALLOC_END)
1005 return __pa(addr);
1006 else
1007 return page_to_phys(vmalloc_to_page(addr));
1008 } else
Tejun Heo9f645532010-04-09 18:57:01 +09001009 return page_to_phys(pcpu_addr_to_page(addr));
Vivek Goyal3b034b02009-11-24 15:50:03 +09001010}
1011
Tejun Heofbf59bc2009-02-20 16:29:08 +09001012/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001013 * pcpu_alloc_alloc_info - allocate percpu allocation info
1014 * @nr_groups: the number of groups
1015 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001016 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001017 * Allocate ai which is large enough for @nr_groups groups containing
1018 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1019 * cpu_map array which is long enough for @nr_units and filled with
1020 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1021 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001022 *
1023 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001024 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1025 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001026 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001027struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1028 int nr_units)
1029{
1030 struct pcpu_alloc_info *ai;
1031 size_t base_size, ai_size;
1032 void *ptr;
1033 int unit;
1034
1035 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1036 __alignof__(ai->groups[0].cpu_map[0]));
1037 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1038
1039 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1040 if (!ptr)
1041 return NULL;
1042 ai = ptr;
1043 ptr += base_size;
1044
1045 ai->groups[0].cpu_map = ptr;
1046
1047 for (unit = 0; unit < nr_units; unit++)
1048 ai->groups[0].cpu_map[unit] = NR_CPUS;
1049
1050 ai->nr_groups = nr_groups;
1051 ai->__ai_size = PFN_ALIGN(ai_size);
1052
1053 return ai;
1054}
1055
1056/**
1057 * pcpu_free_alloc_info - free percpu allocation info
1058 * @ai: pcpu_alloc_info to free
1059 *
1060 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1061 */
1062void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1063{
1064 free_bootmem(__pa(ai), ai->__ai_size);
1065}
1066
1067/**
1068 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
Tejun Heoedcb4632009-03-06 14:33:59 +09001069 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001070 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heofd1e8a12009-08-14 15:00:51 +09001071 * @atom_size: allocation atom size
1072 * @cpu_distance_fn: callback to determine distance between cpus, optional
1073 *
1074 * This function determines grouping of units, their mappings to cpus
1075 * and other parameters considering needed percpu size, allocation
1076 * atom size and distances between CPUs.
1077 *
1078 * Groups are always mutliples of atom size and CPUs which are of
1079 * LOCAL_DISTANCE both ways are grouped together and share space for
1080 * units in the same group. The returned configuration is guaranteed
1081 * to have CPUs on different nodes on different groups and >=75% usage
1082 * of allocated virtual address space.
1083 *
1084 * RETURNS:
1085 * On success, pointer to the new allocation_info is returned. On
1086 * failure, ERR_PTR value is returned.
1087 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001088static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1089 size_t reserved_size, size_t dyn_size,
Tejun Heofd1e8a12009-08-14 15:00:51 +09001090 size_t atom_size,
1091 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
Tejun Heo033e48f2009-08-14 15:00:51 +09001092{
1093 static int group_map[NR_CPUS] __initdata;
1094 static int group_cnt[NR_CPUS] __initdata;
1095 const size_t static_size = __per_cpu_end - __per_cpu_start;
Pavel V. Panteleeva92d3ff2010-06-17 10:07:25 +02001096 int nr_groups = 1, nr_units = 0;
Tejun Heo033e48f2009-08-14 15:00:51 +09001097 size_t size_sum, min_unit_size, alloc_size;
1098 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001099 int last_allocs, group, unit;
Tejun Heo033e48f2009-08-14 15:00:51 +09001100 unsigned int cpu, tcpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001101 struct pcpu_alloc_info *ai;
1102 unsigned int *cpu_map;
Tejun Heo033e48f2009-08-14 15:00:51 +09001103
Tejun Heofb59e722009-09-24 18:50:34 +09001104 /* this function may be called multiple times */
1105 memset(group_map, 0, sizeof(group_map));
Pavel V. Panteleeva92d3ff2010-06-17 10:07:25 +02001106 memset(group_cnt, 0, sizeof(group_cnt));
Tejun Heofb59e722009-09-24 18:50:34 +09001107
Tejun Heo099a19d2010-06-27 18:50:00 +02001108 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1109 size_sum = PFN_ALIGN(static_size + reserved_size +
1110 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001111 dyn_size = size_sum - static_size - reserved_size;
1112
Tejun Heo033e48f2009-08-14 15:00:51 +09001113 /*
1114 * Determine min_unit_size, alloc_size and max_upa such that
Tejun Heofd1e8a12009-08-14 15:00:51 +09001115 * alloc_size is multiple of atom_size and is the smallest
Tejun Heo033e48f2009-08-14 15:00:51 +09001116 * which can accomodate 4k aligned segments which are equal to
1117 * or larger than min_unit_size.
1118 */
Tejun Heo033e48f2009-08-14 15:00:51 +09001119 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1120
Tejun Heofd1e8a12009-08-14 15:00:51 +09001121 alloc_size = roundup(min_unit_size, atom_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001122 upa = alloc_size / min_unit_size;
1123 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1124 upa--;
1125 max_upa = upa;
1126
1127 /* group cpus according to their proximity */
1128 for_each_possible_cpu(cpu) {
1129 group = 0;
1130 next_group:
1131 for_each_possible_cpu(tcpu) {
1132 if (cpu == tcpu)
1133 break;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001134 if (group_map[tcpu] == group && cpu_distance_fn &&
Tejun Heo033e48f2009-08-14 15:00:51 +09001135 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1136 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1137 group++;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001138 nr_groups = max(nr_groups, group + 1);
Tejun Heo033e48f2009-08-14 15:00:51 +09001139 goto next_group;
1140 }
1141 }
1142 group_map[cpu] = group;
1143 group_cnt[group]++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001144 }
1145
1146 /*
1147 * Expand unit size until address space usage goes over 75%
1148 * and then as much as possible without using more address
1149 * space.
1150 */
1151 last_allocs = INT_MAX;
1152 for (upa = max_upa; upa; upa--) {
1153 int allocs = 0, wasted = 0;
1154
1155 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1156 continue;
1157
Tejun Heofd1e8a12009-08-14 15:00:51 +09001158 for (group = 0; group < nr_groups; group++) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001159 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1160 allocs += this_allocs;
1161 wasted += this_allocs * upa - group_cnt[group];
1162 }
1163
1164 /*
1165 * Don't accept if wastage is over 25%. The
1166 * greater-than comparison ensures upa==1 always
1167 * passes the following check.
1168 */
1169 if (wasted > num_possible_cpus() / 3)
1170 continue;
1171
1172 /* and then don't consume more memory */
1173 if (allocs > last_allocs)
1174 break;
1175 last_allocs = allocs;
1176 best_upa = upa;
1177 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001178 upa = best_upa;
Tejun Heo033e48f2009-08-14 15:00:51 +09001179
Tejun Heofd1e8a12009-08-14 15:00:51 +09001180 /* allocate and fill alloc_info */
1181 for (group = 0; group < nr_groups; group++)
1182 nr_units += roundup(group_cnt[group], upa);
1183
1184 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1185 if (!ai)
1186 return ERR_PTR(-ENOMEM);
1187 cpu_map = ai->groups[0].cpu_map;
1188
1189 for (group = 0; group < nr_groups; group++) {
1190 ai->groups[group].cpu_map = cpu_map;
1191 cpu_map += roundup(group_cnt[group], upa);
Tejun Heo033e48f2009-08-14 15:00:51 +09001192 }
1193
Tejun Heofd1e8a12009-08-14 15:00:51 +09001194 ai->static_size = static_size;
1195 ai->reserved_size = reserved_size;
1196 ai->dyn_size = dyn_size;
1197 ai->unit_size = alloc_size / upa;
1198 ai->atom_size = atom_size;
1199 ai->alloc_size = alloc_size;
1200
1201 for (group = 0, unit = 0; group_cnt[group]; group++) {
1202 struct pcpu_group_info *gi = &ai->groups[group];
1203
1204 /*
1205 * Initialize base_offset as if all groups are located
1206 * back-to-back. The caller should update this to
1207 * reflect actual allocation.
1208 */
1209 gi->base_offset = unit * ai->unit_size;
1210
1211 for_each_possible_cpu(cpu)
1212 if (group_map[cpu] == group)
1213 gi->cpu_map[gi->nr_units++] = cpu;
1214 gi->nr_units = roundup(gi->nr_units, upa);
1215 unit += gi->nr_units;
1216 }
1217 BUG_ON(unit != nr_units);
1218
1219 return ai;
Tejun Heo033e48f2009-08-14 15:00:51 +09001220}
1221
Tejun Heofd1e8a12009-08-14 15:00:51 +09001222/**
1223 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1224 * @lvl: loglevel
1225 * @ai: allocation info to dump
1226 *
1227 * Print out information about @ai using loglevel @lvl.
1228 */
1229static void pcpu_dump_alloc_info(const char *lvl,
1230 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001231{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001232 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001233 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001234 int alloc = 0, alloc_end = 0;
1235 int group, v;
1236 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001237
Tejun Heofd1e8a12009-08-14 15:00:51 +09001238 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001239 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001240 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001241
Tejun Heofd1e8a12009-08-14 15:00:51 +09001242 v = num_possible_cpus();
1243 while (v /= 10)
1244 cpu_width++;
1245 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001246
Tejun Heofd1e8a12009-08-14 15:00:51 +09001247 upa = ai->alloc_size / ai->unit_size;
1248 width = upa * (cpu_width + 1) + group_width + 3;
1249 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001250
Tejun Heofd1e8a12009-08-14 15:00:51 +09001251 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1252 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1253 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1254
1255 for (group = 0; group < ai->nr_groups; group++) {
1256 const struct pcpu_group_info *gi = &ai->groups[group];
1257 int unit = 0, unit_end = 0;
1258
1259 BUG_ON(gi->nr_units % upa);
1260 for (alloc_end += gi->nr_units / upa;
1261 alloc < alloc_end; alloc++) {
1262 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001263 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001264 printk("%spcpu-alloc: ", lvl);
1265 }
1266 printk("[%0*d] ", group_width, group);
1267
1268 for (unit_end += upa; unit < unit_end; unit++)
1269 if (gi->cpu_map[unit] != NR_CPUS)
1270 printk("%0*d ", cpu_width,
1271 gi->cpu_map[unit]);
1272 else
1273 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001274 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001275 }
1276 printk("\n");
1277}
Tejun Heo033e48f2009-08-14 15:00:51 +09001278
Tejun Heofbf59bc2009-02-20 16:29:08 +09001279/**
1280 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001281 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001282 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001283 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001284 * Initialize the first percpu chunk which contains the kernel static
1285 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001286 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001287 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001288 * @ai contains all information necessary to initialize the first
1289 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001290 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001291 * @ai->static_size is the size of static percpu area.
1292 *
1293 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001294 * reserve after the static area in the first chunk. This reserves
1295 * the first chunk such that it's available only through reserved
1296 * percpu allocation. This is primarily used to serve module percpu
1297 * static areas on architectures where the addressing model has
1298 * limited offset range for symbol relocations to guarantee module
1299 * percpu symbols fall inside the relocatable range.
1300 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001301 * @ai->dyn_size determines the number of bytes available for dynamic
1302 * allocation in the first chunk. The area between @ai->static_size +
1303 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001304 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001305 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1306 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1307 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001308 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001309 * @ai->atom_size is the allocation atom size and used as alignment
1310 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001311 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001312 * @ai->alloc_size is the allocation size and always multiple of
1313 * @ai->atom_size. This is larger than @ai->atom_size if
1314 * @ai->unit_size is larger than @ai->atom_size.
1315 *
1316 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1317 * percpu areas. Units which should be colocated are put into the
1318 * same group. Dynamic VM areas will be allocated according to these
1319 * groupings. If @ai->nr_groups is zero, a single group containing
1320 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001321 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001322 * The caller should have mapped the first chunk at @base_addr and
1323 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001324 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001325 * If the first chunk ends up with both reserved and dynamic areas, it
1326 * is served by two chunks - one to serve the core static and reserved
1327 * areas and the other for the dynamic area. They share the same vm
1328 * and page map but uses different area allocation map to stay away
1329 * from each other. The latter chunk is circulated in the chunk slots
1330 * and available for dynamic allocation like any other chunks.
1331 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001332 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001333 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001334 */
Tejun Heofb435d52009-08-14 15:00:51 +09001335int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1336 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001337{
Tejun Heo635b75f2009-09-24 09:43:11 +09001338 static char cpus_buf[4096] __initdata;
Tejun Heo099a19d2010-06-27 18:50:00 +02001339 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1340 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001341 size_t dyn_size = ai->dyn_size;
1342 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001343 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001344 unsigned long *group_offsets;
1345 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001346 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001347 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001348 int *unit_map;
1349 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001350
Tejun Heo635b75f2009-09-24 09:43:11 +09001351 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1352
1353#define PCPU_SETUP_BUG_ON(cond) do { \
1354 if (unlikely(cond)) { \
1355 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1356 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1357 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1358 BUG(); \
1359 } \
1360} while (0)
1361
Tejun Heo2f39e632009-07-04 08:11:00 +09001362 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001363 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1364 PCPU_SETUP_BUG_ON(!ai->static_size);
1365 PCPU_SETUP_BUG_ON(!base_addr);
1366 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1367 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1368 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001369 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001370 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001371
Tejun Heo65632972009-08-14 15:00:52 +09001372 /* process group information and build config tables accordingly */
1373 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1374 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001375 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001376 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001377
Tejun Heofd1e8a12009-08-14 15:00:51 +09001378 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001379 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001380 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001381
Tejun Heofd1e8a12009-08-14 15:00:51 +09001382 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1383 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001384
Tejun Heo65632972009-08-14 15:00:52 +09001385 group_offsets[group] = gi->base_offset;
1386 group_sizes[group] = gi->nr_units * ai->unit_size;
1387
Tejun Heofd1e8a12009-08-14 15:00:51 +09001388 for (i = 0; i < gi->nr_units; i++) {
1389 cpu = gi->cpu_map[i];
1390 if (cpu == NR_CPUS)
1391 continue;
1392
Tejun Heo635b75f2009-09-24 09:43:11 +09001393 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1394 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1395 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001396
1397 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001398 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1399
Tejun Heofd1e8a12009-08-14 15:00:51 +09001400 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001401 pcpu_first_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001402 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001403 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001404 pcpu_last_unit_cpu = cpu;
1405 pcpu_nr_units = unit;
1406
1407 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001408 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1409
1410 /* we're done parsing the input, undefine BUG macro and dump config */
1411#undef PCPU_SETUP_BUG_ON
1412 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001413
Tejun Heo65632972009-08-14 15:00:52 +09001414 pcpu_nr_groups = ai->nr_groups;
1415 pcpu_group_offsets = group_offsets;
1416 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001417 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001418 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001419
1420 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001421 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001422 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001423 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001424 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1425 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001426
Tejun Heod9b55ee2009-02-24 11:57:21 +09001427 /*
1428 * Allocate chunk slots. The additional last slot is for
1429 * empty chunks.
1430 */
1431 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001432 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1433 for (i = 0; i < pcpu_nr_slots; i++)
1434 INIT_LIST_HEAD(&pcpu_slot[i]);
1435
Tejun Heoedcb4632009-03-06 14:33:59 +09001436 /*
1437 * Initialize static chunk. If reserved_size is zero, the
1438 * static chunk covers static area + dynamic allocation area
1439 * in the first chunk. If reserved_size is not zero, it
1440 * covers static area + reserved area (mostly used for module
1441 * static percpu allocation).
1442 */
Tejun Heo2441d152009-03-06 14:33:59 +09001443 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1444 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001445 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001446 schunk->map = smap;
1447 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001448 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001449 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001450
Tejun Heofd1e8a12009-08-14 15:00:51 +09001451 if (ai->reserved_size) {
1452 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001453 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001454 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001455 } else {
1456 schunk->free_size = dyn_size;
1457 dyn_size = 0; /* dynamic area covered */
1458 }
Tejun Heo2441d152009-03-06 14:33:59 +09001459 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001460
Tejun Heofd1e8a12009-08-14 15:00:51 +09001461 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001462 if (schunk->free_size)
1463 schunk->map[schunk->map_used++] = schunk->free_size;
1464
Tejun Heoedcb4632009-03-06 14:33:59 +09001465 /* init dynamic chunk if necessary */
1466 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001467 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001468 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001469 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001470 dchunk->map = dmap;
1471 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001472 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001473 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001474
1475 dchunk->contig_hint = dchunk->free_size = dyn_size;
1476 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1477 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1478 }
1479
Tejun Heo2441d152009-03-06 14:33:59 +09001480 /* link the first chunk in */
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001481 pcpu_first_chunk = dchunk ?: schunk;
1482 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001483
1484 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001485 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001486 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001487}
Tejun Heo66c3a752009-03-10 16:27:48 +09001488
Tejun Heof58dc012009-08-14 15:00:50 +09001489const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1490 [PCPU_FC_AUTO] = "auto",
1491 [PCPU_FC_EMBED] = "embed",
1492 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001493};
Tejun Heo66c3a752009-03-10 16:27:48 +09001494
Tejun Heof58dc012009-08-14 15:00:50 +09001495enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1496
1497static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001498{
Tejun Heof58dc012009-08-14 15:00:50 +09001499 if (0)
1500 /* nada */;
1501#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1502 else if (!strcmp(str, "embed"))
1503 pcpu_chosen_fc = PCPU_FC_EMBED;
1504#endif
1505#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1506 else if (!strcmp(str, "page"))
1507 pcpu_chosen_fc = PCPU_FC_PAGE;
1508#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001509 else
1510 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001511
Tejun Heof58dc012009-08-14 15:00:50 +09001512 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001513}
Tejun Heof58dc012009-08-14 15:00:50 +09001514early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001515
Tejun Heo08fc4582009-08-14 15:00:49 +09001516#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1517 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo66c3a752009-03-10 16:27:48 +09001518/**
1519 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001520 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001521 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001522 * @atom_size: allocation atom size
1523 * @cpu_distance_fn: callback to determine distance between cpus, optional
1524 * @alloc_fn: function to allocate percpu page
1525 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001526 *
1527 * This is a helper to ease setting up embedded first percpu chunk and
1528 * can be called where pcpu_setup_first_chunk() is expected.
1529 *
1530 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001531 * by calling @alloc_fn and used as-is without being mapped into
1532 * vmalloc area. Allocations are always whole multiples of @atom_size
1533 * aligned to @atom_size.
1534 *
1535 * This enables the first chunk to piggy back on the linear physical
1536 * mapping which often uses larger page size. Please note that this
1537 * can result in very sparse cpu->unit mapping on NUMA machines thus
1538 * requiring large vmalloc address space. Don't use this allocator if
1539 * vmalloc space is not orders of magnitude larger than distances
1540 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001541 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001542 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001543 *
1544 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001545 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001546 *
1547 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001548 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001549 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001550int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001551 size_t atom_size,
1552 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1553 pcpu_fc_alloc_fn_t alloc_fn,
1554 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001555{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001556 void *base = (void *)ULONG_MAX;
1557 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001558 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001559 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001560 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001561
Tejun Heoc8826dd2009-08-14 15:00:52 +09001562 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1563 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001564 if (IS_ERR(ai))
1565 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001566
Tejun Heofd1e8a12009-08-14 15:00:51 +09001567 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001568 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001569
Tejun Heoc8826dd2009-08-14 15:00:52 +09001570 areas = alloc_bootmem_nopanic(areas_size);
1571 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001572 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001573 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001574 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001575
Tejun Heoc8826dd2009-08-14 15:00:52 +09001576 /* allocate, copy and determine base address */
1577 for (group = 0; group < ai->nr_groups; group++) {
1578 struct pcpu_group_info *gi = &ai->groups[group];
1579 unsigned int cpu = NR_CPUS;
1580 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001581
Tejun Heoc8826dd2009-08-14 15:00:52 +09001582 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1583 cpu = gi->cpu_map[i];
1584 BUG_ON(cpu == NR_CPUS);
1585
1586 /* allocate space for the whole group */
1587 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1588 if (!ptr) {
1589 rc = -ENOMEM;
1590 goto out_free_areas;
1591 }
1592 areas[group] = ptr;
1593
1594 base = min(ptr, base);
1595
1596 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1597 if (gi->cpu_map[i] == NR_CPUS) {
1598 /* unused unit, free whole */
1599 free_fn(ptr, ai->unit_size);
1600 continue;
1601 }
1602 /* copy and return the unused part */
1603 memcpy(ptr, __per_cpu_load, ai->static_size);
1604 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1605 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001606 }
1607
Tejun Heoc8826dd2009-08-14 15:00:52 +09001608 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001609 max_distance = 0;
1610 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001611 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001612 max_distance = max_t(size_t, max_distance,
1613 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001614 }
1615 max_distance += ai->unit_size;
1616
1617 /* warn if maximum distance is further than 75% of vmalloc space */
1618 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001619 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001620 "space 0x%lx\n",
1621 max_distance, VMALLOC_END - VMALLOC_START);
1622#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1623 /* and fail if we have fallback */
1624 rc = -EINVAL;
1625 goto out_free;
1626#endif
1627 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001628
Tejun Heo004018e2009-08-14 15:00:49 +09001629 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001630 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1631 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001632
Tejun Heofb435d52009-08-14 15:00:51 +09001633 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001634 goto out_free;
1635
1636out_free_areas:
1637 for (group = 0; group < ai->nr_groups; group++)
1638 free_fn(areas[group],
1639 ai->groups[group].nr_units * ai->unit_size);
1640out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001641 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001642 if (areas)
1643 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001644 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001645}
Tejun Heo08fc4582009-08-14 15:00:49 +09001646#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1647 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heod4b95f82009-07-04 08:10:59 +09001648
Tejun Heo08fc4582009-08-14 15:00:49 +09001649#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001650/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001651 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001652 * @reserved_size: the size of reserved percpu area in bytes
1653 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1654 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1655 * @populate_pte_fn: function to populate pte
1656 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001657 * This is a helper to ease setting up page-remapped first percpu
1658 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001659 *
1660 * This is the basic allocator. Static percpu area is allocated
1661 * page-by-page into vmalloc area.
1662 *
1663 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001664 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001665 */
Tejun Heofb435d52009-08-14 15:00:51 +09001666int __init pcpu_page_first_chunk(size_t reserved_size,
1667 pcpu_fc_alloc_fn_t alloc_fn,
1668 pcpu_fc_free_fn_t free_fn,
1669 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001670{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001671 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001672 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001673 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001674 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001675 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001676 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001677 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001678
Tejun Heo00ae4062009-08-14 15:00:49 +09001679 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1680
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001681 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001682 if (IS_ERR(ai))
1683 return PTR_ERR(ai);
1684 BUG_ON(ai->nr_groups != 1);
1685 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1686
1687 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001688
1689 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001690 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1691 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09001692 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09001693
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001694 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001695 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001696 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001697 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001698 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001699 void *ptr;
1700
Tejun Heo3cbc8562009-08-14 15:00:50 +09001701 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001702 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001703 pr_warning("PERCPU: failed to allocate %s page "
1704 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001705 goto enomem;
1706 }
Tejun Heoce3141a2009-07-04 08:11:00 +09001707 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001708 }
1709
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001710 /* allocate vm area, map the pages and copy static data */
1711 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001712 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001713 vm_area_register_early(&vm, PAGE_SIZE);
1714
Tejun Heofd1e8a12009-08-14 15:00:51 +09001715 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001716 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001717 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001718
Tejun Heoce3141a2009-07-04 08:11:00 +09001719 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001720 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1721
1722 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001723 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1724 unit_pages);
1725 if (rc < 0)
1726 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001727
1728 /*
1729 * FIXME: Archs with virtual cache should flush local
1730 * cache for the linear mapping here - something
1731 * equivalent to flush_cache_vmap() on the local cpu.
1732 * flush_cache_vmap() can't be used as most supporting
1733 * data structures are not set up yet.
1734 */
1735
1736 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001737 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001738 }
1739
1740 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09001741 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001742 unit_pages, psize_str, vm.addr, ai->static_size,
1743 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001744
Tejun Heofb435d52009-08-14 15:00:51 +09001745 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001746 goto out_free_ar;
1747
1748enomem:
1749 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09001750 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09001751 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09001752out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09001753 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001754 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09001755 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001756}
Tejun Heo08fc4582009-08-14 15:00:49 +09001757#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001758
Tejun Heo8c4bfc62009-07-04 08:10:59 +09001759/*
Tejun Heoe74e3962009-03-30 19:07:44 +09001760 * Generic percpu area setup.
1761 *
1762 * The embedding helper is used because its behavior closely resembles
1763 * the original non-dynamic generic percpu area setup. This is
1764 * important because many archs have addressing restrictions and might
1765 * fail if the percpu area is located far away from the previous
1766 * location. As an added bonus, in non-NUMA cases, embedding is
1767 * generally a good idea TLB-wise because percpu area can piggy back
1768 * on the physical linear memory mapping which uses large page
1769 * mappings on applicable archs.
1770 */
1771#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1772unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1773EXPORT_SYMBOL(__per_cpu_offset);
1774
Tejun Heoc8826dd2009-08-14 15:00:52 +09001775static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1776 size_t align)
1777{
1778 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1779}
1780
1781static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1782{
1783 free_bootmem(__pa(ptr), size);
1784}
1785
Tejun Heoe74e3962009-03-30 19:07:44 +09001786void __init setup_per_cpu_areas(void)
1787{
Tejun Heoe74e3962009-03-30 19:07:44 +09001788 unsigned long delta;
1789 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09001790 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09001791
1792 /*
1793 * Always reserve area for module percpu variables. That's
1794 * what the legacy allocator did.
1795 */
Tejun Heofb435d52009-08-14 15:00:51 +09001796 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001797 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1798 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09001799 if (rc < 0)
Tejun Heoe74e3962009-03-30 19:07:44 +09001800 panic("Failed to initialized percpu areas.");
1801
1802 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1803 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09001804 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09001805}
1806#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heo099a19d2010-06-27 18:50:00 +02001807
1808/*
1809 * First and reserved chunks are initialized with temporary allocation
1810 * map in initdata so that they can be used before slab is online.
1811 * This function is called after slab is brought up and replaces those
1812 * with properly allocated maps.
1813 */
1814void __init percpu_init_late(void)
1815{
1816 struct pcpu_chunk *target_chunks[] =
1817 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1818 struct pcpu_chunk *chunk;
1819 unsigned long flags;
1820 int i;
1821
1822 for (i = 0; (chunk = target_chunks[i]); i++) {
1823 int *map;
1824 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1825
1826 BUILD_BUG_ON(size > PAGE_SIZE);
1827
1828 map = pcpu_mem_alloc(size);
1829 BUG_ON(!map);
1830
1831 spin_lock_irqsave(&pcpu_lock, flags);
1832 memcpy(map, chunk->map, size);
1833 chunk->map = map;
1834 spin_unlock_irqrestore(&pcpu_lock, flags);
1835 }
1836}