blob: efe816856a9d777b284f8bf25cc7548154f374d6 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
Tejun Heo88999a82010-04-09 18:57:01 +09002 * mm/percpu.c - percpu memory allocator
Tejun Heofbf59bc2009-02-20 16:29:08 +09003 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
Tejun Heo88999a82010-04-09 18:57:01 +090010 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
Tejun Heo2f39e632009-07-04 08:11:00 +090013 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
Tejun Heo88999a82010-04-09 18:57:01 +090016 * When a chunk is filled up, another chunk is allocated.
Tejun Heofbf59bc2009-02-20 16:29:08 +090017 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
Tejun Heo2f39e632009-07-04 08:11:00 +090025 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090029 *
Tejun Heo2f39e632009-07-04 08:11:00 +090030 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090032 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
Namhyung Kim4785879e2010-08-11 11:24:10 +090034 * guaranteed to be equal to or larger than the maximum contiguous
Tejun Heofbf59bc2009-02-20 16:29:08 +090035 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090043 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090045 *
46 * To use this allocator, arch code should do the followings.
47 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090048 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090058#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090059#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090060#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090061#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090066#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090067#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090068#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090069#include <linux/workqueue.h>
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 Heobbddff02010-09-03 18:22:48 +020079#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +090080/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
81#ifndef __addr_to_pcpu_ptr
82#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090083 (void __percpu *)((unsigned long)(addr) - \
84 (unsigned long)pcpu_base_addr + \
85 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090086#endif
87#ifndef __pcpu_ptr_to_addr
88#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090089 (void __force *)((unsigned long)(ptr) + \
90 (unsigned long)pcpu_base_addr - \
91 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090092#endif
Tejun Heobbddff02010-09-03 18:22:48 +020093#else /* CONFIG_SMP */
94/* on UP, it's always identity mapped */
95#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
96#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
97#endif /* CONFIG_SMP */
Tejun Heoe0100982009-03-10 16:27:48 +090098
Tejun Heofbf59bc2009-02-20 16:29:08 +090099struct pcpu_chunk {
100 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900101 int free_size; /* free bytes in the chunk */
102 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +0900103 void *base_addr; /* base address of this chunk */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900104 int map_used; /* # of map entries used */
105 int map_alloc; /* # of map entries allocated */
106 int *map; /* allocation map */
Tejun Heo88999a82010-04-09 18:57:01 +0900107 void *data; /* chunk data */
Tejun Heo8d408b42009-02-24 11:57:21 +0900108 bool immutable; /* no [de]population allowed */
Tejun Heoce3141a2009-07-04 08:11:00 +0900109 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900110};
111
Tejun Heo40150d32009-02-24 12:32:28 +0900112static int pcpu_unit_pages __read_mostly;
113static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900114static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900115static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900116static int pcpu_nr_slots __read_mostly;
117static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900118
Tejun Heo2f39e632009-07-04 08:11:00 +0900119/* cpus with the lowest and highest unit numbers */
120static unsigned int pcpu_first_unit_cpu __read_mostly;
121static unsigned int pcpu_last_unit_cpu __read_mostly;
122
Tejun Heofbf59bc2009-02-20 16:29:08 +0900123/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900124void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900125EXPORT_SYMBOL_GPL(pcpu_base_addr);
126
Tejun Heofb435d52009-08-14 15:00:51 +0900127static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
128const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900129
Tejun Heo65632972009-08-14 15:00:52 +0900130/* group information, used for vm allocation */
131static int pcpu_nr_groups __read_mostly;
132static const unsigned long *pcpu_group_offsets __read_mostly;
133static const size_t *pcpu_group_sizes __read_mostly;
134
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900135/*
136 * The first chunk which always exists. Note that unlike other
137 * chunks, this one can be allocated and mapped in several different
138 * ways and thus often doesn't live in the vmalloc area.
139 */
140static struct pcpu_chunk *pcpu_first_chunk;
141
142/*
143 * Optional reserved chunk. This chunk reserves part of the first
144 * chunk and serves it for reserved allocations. The amount of
145 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
146 * area doesn't exist, the following variables contain NULL and 0
147 * respectively.
148 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900149static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900150static int pcpu_reserved_chunk_limit;
151
Tejun Heofbf59bc2009-02-20 16:29:08 +0900152/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900153 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900154 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900155 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
Tejun Heoce3141a2009-07-04 08:11:00 +0900156 * protects allocation/reclaim paths, chunks, populated bitmap and
157 * vmalloc mapping. The latter is a spinlock and protects the index
158 * data structures - chunk slots, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900159 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900160 * During allocation, pcpu_alloc_mutex is kept locked all the time and
161 * pcpu_lock is grabbed and released as necessary. All actual memory
Jiri Kosina403a91b2009-10-29 00:25:59 +0900162 * allocations are done using GFP_KERNEL with pcpu_lock released. In
163 * general, percpu memory can't be allocated with irq off but
164 * irqsave/restore are still used in alloc path so that it can be used
165 * from early init path - sched_init() specifically.
Tejun Heoccea34b2009-03-07 00:44:13 +0900166 *
167 * Free path accesses and alters only the index data structures, so it
168 * can be safely called from atomic context. When memory needs to be
169 * returned to the system, free path schedules reclaim_work which
170 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
171 * reclaimed, release both locks and frees the chunks. Note that it's
172 * necessary to grab both locks to remove a chunk from circulation as
173 * allocation path might be referencing the chunk with only
174 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900175 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900176static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
177static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900178
Tejun Heo40150d32009-02-24 12:32:28 +0900179static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900180
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900181/* reclaim work to release fully free chunks, scheduled from free path */
182static void pcpu_reclaim(struct work_struct *work);
183static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
184
Tejun Heo020ec652010-04-09 18:57:00 +0900185static bool pcpu_addr_in_first_chunk(void *addr)
186{
187 void *first_start = pcpu_first_chunk->base_addr;
188
189 return addr >= first_start && addr < first_start + pcpu_unit_size;
190}
191
192static bool pcpu_addr_in_reserved_chunk(void *addr)
193{
194 void *first_start = pcpu_first_chunk->base_addr;
195
196 return addr >= first_start &&
197 addr < first_start + pcpu_reserved_chunk_limit;
198}
199
Tejun Heod9b55ee2009-02-24 11:57:21 +0900200static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900201{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900202 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900203 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
204}
205
Tejun Heod9b55ee2009-02-24 11:57:21 +0900206static int pcpu_size_to_slot(int size)
207{
208 if (size == pcpu_unit_size)
209 return pcpu_nr_slots - 1;
210 return __pcpu_size_to_slot(size);
211}
212
Tejun Heofbf59bc2009-02-20 16:29:08 +0900213static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
214{
215 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
216 return 0;
217
218 return pcpu_size_to_slot(chunk->free_size);
219}
220
Tejun Heo88999a82010-04-09 18:57:01 +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
233static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900234{
Tejun Heo2f39e632009-07-04 08:11:00 +0900235 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900236}
237
Tejun Heo9983b6f02010-06-18 11:44:31 +0200238static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
239 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900240{
Tejun Heobba174f2009-08-14 15:00:51 +0900241 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900242 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900243}
244
Tejun Heo88999a82010-04-09 18:57:01 +0900245static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
246 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900247{
248 *rs = find_next_zero_bit(chunk->populated, end, *rs);
249 *re = find_next_bit(chunk->populated, end, *rs + 1);
250}
251
Tejun Heo88999a82010-04-09 18:57:01 +0900252static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
253 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900254{
255 *rs = find_next_bit(chunk->populated, end, *rs);
256 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
257}
258
259/*
260 * (Un)populated page region iterators. Iterate over (un)populated
261 * page regions betwen @start and @end in @chunk. @rs and @re should
262 * be integer variables and will be set to start and end page index of
263 * the current region.
264 */
265#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
266 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
267 (rs) < (re); \
268 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
269
270#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
271 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
272 (rs) < (re); \
273 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
274
Tejun Heofbf59bc2009-02-20 16:29:08 +0900275/**
Tejun Heo1880d932009-03-07 00:44:09 +0900276 * pcpu_mem_alloc - allocate memory
277 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900278 *
Tejun Heo1880d932009-03-07 00:44:09 +0900279 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
280 * kzalloc() is used; otherwise, vmalloc() is used. The returned
281 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900282 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900283 * CONTEXT:
284 * Does GFP_KERNEL allocation.
285 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900286 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900287 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900288 */
Tejun Heo1880d932009-03-07 00:44:09 +0900289static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900290{
Tejun Heo099a19d2010-06-27 18:50:00 +0200291 if (WARN_ON_ONCE(!slab_is_available()))
292 return NULL;
293
Tejun Heofbf59bc2009-02-20 16:29:08 +0900294 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900295 return kzalloc(size, GFP_KERNEL);
296 else {
297 void *ptr = vmalloc(size);
298 if (ptr)
299 memset(ptr, 0, size);
300 return ptr;
301 }
302}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900303
Tejun Heo1880d932009-03-07 00:44:09 +0900304/**
305 * pcpu_mem_free - free memory
306 * @ptr: memory to free
307 * @size: size of the area
308 *
309 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
310 */
311static void pcpu_mem_free(void *ptr, size_t size)
312{
313 if (size <= PAGE_SIZE)
314 kfree(ptr);
315 else
316 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900317}
318
319/**
320 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
321 * @chunk: chunk of interest
322 * @oslot: the previous slot it was on
323 *
324 * This function is called after an allocation or free changed @chunk.
325 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900326 * moved to the slot. Note that the reserved chunk is never put on
327 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900328 *
329 * CONTEXT:
330 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900331 */
332static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
333{
334 int nslot = pcpu_chunk_slot(chunk);
335
Tejun Heoedcb4632009-03-06 14:33:59 +0900336 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900337 if (oslot < nslot)
338 list_move(&chunk->list, &pcpu_slot[nslot]);
339 else
340 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
341 }
342}
343
Tejun Heofbf59bc2009-02-20 16:29:08 +0900344/**
Tejun Heo833af842009-11-11 15:35:18 +0900345 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
346 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900347 *
Tejun Heo833af842009-11-11 15:35:18 +0900348 * Determine whether area map of @chunk needs to be extended to
349 * accomodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900350 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900351 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900352 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900353 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900354 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900355 * New target map allocation length if extension is necessary, 0
356 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900357 */
Tejun Heo833af842009-11-11 15:35:18 +0900358static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900359{
360 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900361
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900362 if (chunk->map_alloc >= chunk->map_used + 2)
363 return 0;
364
365 new_alloc = PCPU_DFL_MAP_ALLOC;
366 while (new_alloc < chunk->map_used + 2)
367 new_alloc *= 2;
368
Tejun Heo833af842009-11-11 15:35:18 +0900369 return new_alloc;
370}
371
372/**
373 * pcpu_extend_area_map - extend area map of a chunk
374 * @chunk: chunk of interest
375 * @new_alloc: new target allocation length of the area map
376 *
377 * Extend area map of @chunk to have @new_alloc entries.
378 *
379 * CONTEXT:
380 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
381 *
382 * RETURNS:
383 * 0 on success, -errno on failure.
384 */
385static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
386{
387 int *old = NULL, *new = NULL;
388 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
389 unsigned long flags;
390
391 new = pcpu_mem_alloc(new_size);
392 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900393 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900394
Tejun Heo833af842009-11-11 15:35:18 +0900395 /* acquire pcpu_lock and switch to new area map */
396 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900397
Tejun Heo833af842009-11-11 15:35:18 +0900398 if (new_alloc <= chunk->map_alloc)
399 goto out_unlock;
400
401 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
Huang Shijiea002d142010-08-08 14:39:07 +0200402 old = chunk->map;
403
404 memcpy(new, old, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900405
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900406 chunk->map_alloc = new_alloc;
407 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900408 new = NULL;
409
410out_unlock:
411 spin_unlock_irqrestore(&pcpu_lock, flags);
412
413 /*
414 * pcpu_mem_free() might end up calling vfree() which uses
415 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
416 */
417 pcpu_mem_free(old, old_size);
418 pcpu_mem_free(new, new_size);
419
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900420 return 0;
421}
422
423/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900424 * pcpu_split_block - split a map block
425 * @chunk: chunk of interest
426 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900427 * @head: head size in bytes (can be 0)
428 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900429 *
430 * Split the @i'th map block into two or three blocks. If @head is
431 * non-zero, @head bytes block is inserted before block @i moving it
432 * to @i+1 and reducing its size by @head bytes.
433 *
434 * If @tail is non-zero, the target block, which can be @i or @i+1
435 * depending on @head, is reduced by @tail bytes and @tail byte block
436 * is inserted after the target block.
437 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900438 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900439 *
440 * CONTEXT:
441 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900442 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900443static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
444 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900445{
446 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900447
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900448 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900449
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900450 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900451 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
452 sizeof(chunk->map[0]) * (chunk->map_used - i));
453 chunk->map_used += nr_extra;
454
455 if (head) {
456 chunk->map[i + 1] = chunk->map[i] - head;
457 chunk->map[i++] = head;
458 }
459 if (tail) {
460 chunk->map[i++] -= tail;
461 chunk->map[i] = tail;
462 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900463}
464
465/**
466 * pcpu_alloc_area - allocate area from a pcpu_chunk
467 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900468 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900469 * @align: wanted align
470 *
471 * Try to allocate @size bytes area aligned at @align from @chunk.
472 * Note that this function only allocates the offset. It doesn't
473 * populate or map the area.
474 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900475 * @chunk->map must have at least two free slots.
476 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900477 * CONTEXT:
478 * pcpu_lock.
479 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900480 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900481 * Allocated offset in @chunk on success, -1 if no matching area is
482 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900483 */
484static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
485{
486 int oslot = pcpu_chunk_slot(chunk);
487 int max_contig = 0;
488 int i, off;
489
Tejun Heofbf59bc2009-02-20 16:29:08 +0900490 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
491 bool is_last = i + 1 == chunk->map_used;
492 int head, tail;
493
494 /* extra for alignment requirement */
495 head = ALIGN(off, align) - off;
496 BUG_ON(i == 0 && head != 0);
497
498 if (chunk->map[i] < 0)
499 continue;
500 if (chunk->map[i] < head + size) {
501 max_contig = max(chunk->map[i], max_contig);
502 continue;
503 }
504
505 /*
506 * If head is small or the previous block is free,
507 * merge'em. Note that 'small' is defined as smaller
508 * than sizeof(int), which is very small but isn't too
509 * uncommon for percpu allocations.
510 */
511 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
512 if (chunk->map[i - 1] > 0)
513 chunk->map[i - 1] += head;
514 else {
515 chunk->map[i - 1] -= head;
516 chunk->free_size -= head;
517 }
518 chunk->map[i] -= head;
519 off += head;
520 head = 0;
521 }
522
523 /* if tail is small, just keep it around */
524 tail = chunk->map[i] - head - size;
525 if (tail < sizeof(int))
526 tail = 0;
527
528 /* split if warranted */
529 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900530 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900531 if (head) {
532 i++;
533 off += head;
534 max_contig = max(chunk->map[i - 1], max_contig);
535 }
536 if (tail)
537 max_contig = max(chunk->map[i + 1], max_contig);
538 }
539
540 /* update hint and mark allocated */
541 if (is_last)
542 chunk->contig_hint = max_contig; /* fully scanned */
543 else
544 chunk->contig_hint = max(chunk->contig_hint,
545 max_contig);
546
547 chunk->free_size -= chunk->map[i];
548 chunk->map[i] = -chunk->map[i];
549
550 pcpu_chunk_relocate(chunk, oslot);
551 return off;
552 }
553
554 chunk->contig_hint = max_contig; /* fully scanned */
555 pcpu_chunk_relocate(chunk, oslot);
556
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900557 /* tell the upper layer that this chunk has no matching area */
558 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900559}
560
561/**
562 * pcpu_free_area - free area to a pcpu_chunk
563 * @chunk: chunk of interest
564 * @freeme: offset of area to free
565 *
566 * Free area starting from @freeme to @chunk. Note that this function
567 * only modifies the allocation map. It doesn't depopulate or unmap
568 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900569 *
570 * CONTEXT:
571 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900572 */
573static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
574{
575 int oslot = pcpu_chunk_slot(chunk);
576 int i, off;
577
578 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
579 if (off == freeme)
580 break;
581 BUG_ON(off != freeme);
582 BUG_ON(chunk->map[i] > 0);
583
584 chunk->map[i] = -chunk->map[i];
585 chunk->free_size += chunk->map[i];
586
587 /* merge with previous? */
588 if (i > 0 && chunk->map[i - 1] >= 0) {
589 chunk->map[i - 1] += chunk->map[i];
590 chunk->map_used--;
591 memmove(&chunk->map[i], &chunk->map[i + 1],
592 (chunk->map_used - i) * sizeof(chunk->map[0]));
593 i--;
594 }
595 /* merge with next? */
596 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
597 chunk->map[i] += chunk->map[i + 1];
598 chunk->map_used--;
599 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
600 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
601 }
602
603 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
604 pcpu_chunk_relocate(chunk, oslot);
605}
606
Tejun Heo60810892010-04-09 18:57:01 +0900607static struct pcpu_chunk *pcpu_alloc_chunk(void)
608{
609 struct pcpu_chunk *chunk;
610
Tejun Heo099a19d2010-06-27 18:50:00 +0200611 chunk = pcpu_mem_alloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900612 if (!chunk)
613 return NULL;
614
615 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
616 if (!chunk->map) {
617 kfree(chunk);
618 return NULL;
619 }
620
621 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
622 chunk->map[chunk->map_used++] = pcpu_unit_size;
623
624 INIT_LIST_HEAD(&chunk->list);
625 chunk->free_size = pcpu_unit_size;
626 chunk->contig_hint = pcpu_unit_size;
627
628 return chunk;
629}
630
631static void pcpu_free_chunk(struct pcpu_chunk *chunk)
632{
633 if (!chunk)
634 return;
635 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
636 kfree(chunk);
637}
638
Tejun Heo9f645532010-04-09 18:57:01 +0900639/*
640 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900641 *
Tejun Heo9f645532010-04-09 18:57:01 +0900642 * To allow different implementations, chunk alloc/free and
643 * [de]population are implemented in a separate file which is pulled
644 * into this file and compiled together. The following functions
645 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900646 *
Tejun Heo9f645532010-04-09 18:57:01 +0900647 * pcpu_populate_chunk - populate the specified range of a chunk
648 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
649 * pcpu_create_chunk - create a new chunk
650 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
651 * pcpu_addr_to_page - translate address to physical address
652 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900653 */
Tejun Heo9f645532010-04-09 18:57:01 +0900654static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
655static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
656static struct pcpu_chunk *pcpu_create_chunk(void);
657static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
658static struct page *pcpu_addr_to_page(void *addr);
659static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900660
Tejun Heob0c97782010-04-09 18:57:01 +0900661#ifdef CONFIG_NEED_PER_CPU_KM
662#include "percpu-km.c"
663#else
Tejun Heo9f645532010-04-09 18:57:01 +0900664#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900665#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900666
667/**
Tejun Heo88999a82010-04-09 18:57:01 +0900668 * pcpu_chunk_addr_search - determine chunk containing specified address
669 * @addr: address for which the chunk needs to be determined.
670 *
671 * RETURNS:
672 * The address of the found chunk.
673 */
674static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
675{
676 /* is it in the first chunk? */
677 if (pcpu_addr_in_first_chunk(addr)) {
678 /* is it in the reserved area? */
679 if (pcpu_addr_in_reserved_chunk(addr))
680 return pcpu_reserved_chunk;
681 return pcpu_first_chunk;
682 }
683
684 /*
685 * The address is relative to unit0 which might be unused and
686 * thus unmapped. Offset the address to the unit space of the
687 * current processor before looking it up in the vmalloc
688 * space. Note that any possible cpu id can be used here, so
689 * there's no need to worry about preemption or cpu hotplug.
690 */
691 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900692 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900693}
694
695/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900696 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900697 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900698 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900699 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900700 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900701 * Allocate percpu area of @size bytes aligned at @align.
702 *
703 * CONTEXT:
704 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900705 *
706 * RETURNS:
707 * Percpu pointer to the allocated area on success, NULL on failure.
708 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900709static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900710{
Tejun Heof2badb02009-09-29 09:17:58 +0900711 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900712 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900713 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +0900714 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900715 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900716
Tejun Heo8d408b42009-02-24 11:57:21 +0900717 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900718 WARN(true, "illegal size (%zu) or align (%zu) for "
719 "percpu allocation\n", size, align);
720 return NULL;
721 }
722
Tejun Heoccea34b2009-03-07 00:44:13 +0900723 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +0900724 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900725
Tejun Heoedcb4632009-03-06 14:33:59 +0900726 /* serve reserved allocations from the reserved chunk if available */
727 if (reserved && pcpu_reserved_chunk) {
728 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900729
730 if (size > chunk->contig_hint) {
731 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900732 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900733 }
Tejun Heo833af842009-11-11 15:35:18 +0900734
735 while ((new_alloc = pcpu_need_to_extend(chunk))) {
736 spin_unlock_irqrestore(&pcpu_lock, flags);
737 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
738 err = "failed to extend area map of reserved chunk";
739 goto fail_unlock_mutex;
740 }
741 spin_lock_irqsave(&pcpu_lock, flags);
742 }
743
Tejun Heoedcb4632009-03-06 14:33:59 +0900744 off = pcpu_alloc_area(chunk, size, align);
745 if (off >= 0)
746 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900747
Tejun Heof2badb02009-09-29 09:17:58 +0900748 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900749 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900750 }
751
Tejun Heoccea34b2009-03-07 00:44:13 +0900752restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900753 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900754 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
755 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
756 if (size > chunk->contig_hint)
757 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900758
Tejun Heo833af842009-11-11 15:35:18 +0900759 new_alloc = pcpu_need_to_extend(chunk);
760 if (new_alloc) {
761 spin_unlock_irqrestore(&pcpu_lock, flags);
762 if (pcpu_extend_area_map(chunk,
763 new_alloc) < 0) {
764 err = "failed to extend area map";
765 goto fail_unlock_mutex;
766 }
767 spin_lock_irqsave(&pcpu_lock, flags);
768 /*
769 * pcpu_lock has been dropped, need to
770 * restart cpu_slot list walking.
771 */
772 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900773 }
774
Tejun Heofbf59bc2009-02-20 16:29:08 +0900775 off = pcpu_alloc_area(chunk, size, align);
776 if (off >= 0)
777 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900778 }
779 }
780
781 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900782 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900783
Tejun Heo60810892010-04-09 18:57:01 +0900784 chunk = pcpu_create_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +0900785 if (!chunk) {
786 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +0900787 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +0900788 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900789
Jiri Kosina403a91b2009-10-29 00:25:59 +0900790 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900791 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900792 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900793
794area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900795 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900796
Tejun Heofbf59bc2009-02-20 16:29:08 +0900797 /* populate, map and clear the area */
798 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +0900799 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900800 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +0900801 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +0900802 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900803 }
804
Tejun Heoccea34b2009-03-07 00:44:13 +0900805 mutex_unlock(&pcpu_alloc_mutex);
806
Tejun Heobba174f2009-08-14 15:00:51 +0900807 /* return address relative to base address */
808 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900809
810fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900811 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900812fail_unlock_mutex:
813 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +0900814 if (warn_limit) {
815 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
816 "%s\n", size, align, err);
817 dump_stack();
818 if (!--warn_limit)
819 pr_info("PERCPU: limit reached, disable warning\n");
820 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900821 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900822}
Tejun Heoedcb4632009-03-06 14:33:59 +0900823
824/**
825 * __alloc_percpu - allocate dynamic percpu area
826 * @size: size of area to allocate in bytes
827 * @align: alignment of area (max PAGE_SIZE)
828 *
Tejun Heo9329ba92010-09-10 11:01:56 +0200829 * Allocate zero-filled percpu area of @size bytes aligned at @align.
830 * Might sleep. Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +0900831 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900832 * CONTEXT:
833 * Does GFP_KERNEL allocation.
834 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900835 * RETURNS:
836 * Percpu pointer to the allocated area on success, NULL on failure.
837 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900838void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900839{
840 return pcpu_alloc(size, align, false);
841}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900842EXPORT_SYMBOL_GPL(__alloc_percpu);
843
Tejun Heoedcb4632009-03-06 14:33:59 +0900844/**
845 * __alloc_reserved_percpu - allocate reserved percpu area
846 * @size: size of area to allocate in bytes
847 * @align: alignment of area (max PAGE_SIZE)
848 *
Tejun Heo9329ba92010-09-10 11:01:56 +0200849 * Allocate zero-filled percpu area of @size bytes aligned at @align
850 * from reserved percpu area if arch has set it up; otherwise,
851 * allocation is served from the same dynamic area. Might sleep.
852 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +0900853 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900854 * CONTEXT:
855 * Does GFP_KERNEL allocation.
856 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900857 * RETURNS:
858 * Percpu pointer to the allocated area on success, NULL on failure.
859 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900860void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900861{
862 return pcpu_alloc(size, align, true);
863}
864
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900865/**
866 * pcpu_reclaim - reclaim fully free chunks, workqueue function
867 * @work: unused
868 *
869 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900870 *
871 * CONTEXT:
872 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900873 */
874static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900875{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900876 LIST_HEAD(todo);
877 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
878 struct pcpu_chunk *chunk, *next;
879
Tejun Heoccea34b2009-03-07 00:44:13 +0900880 mutex_lock(&pcpu_alloc_mutex);
881 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900882
883 list_for_each_entry_safe(chunk, next, head, list) {
884 WARN_ON(chunk->immutable);
885
886 /* spare the first one */
887 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
888 continue;
889
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900890 list_move(&chunk->list, &todo);
891 }
892
Tejun Heoccea34b2009-03-07 00:44:13 +0900893 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900894
895 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +0900896 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heo60810892010-04-09 18:57:01 +0900897 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900898 }
Tejun Heo971f3912009-08-14 15:00:49 +0900899
900 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900901}
902
903/**
904 * free_percpu - free percpu area
905 * @ptr: pointer to area to free
906 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900907 * Free percpu area @ptr.
908 *
909 * CONTEXT:
910 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900911 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900912void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900913{
Andrew Morton129182e2010-01-08 14:42:39 -0800914 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900915 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900916 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900917 int off;
918
919 if (!ptr)
920 return;
921
Andrew Morton129182e2010-01-08 14:42:39 -0800922 addr = __pcpu_ptr_to_addr(ptr);
923
Tejun Heoccea34b2009-03-07 00:44:13 +0900924 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900925
926 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +0900927 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900928
929 pcpu_free_area(chunk, off);
930
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900931 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900932 if (chunk->free_size == pcpu_unit_size) {
933 struct pcpu_chunk *pos;
934
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900935 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900936 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900937 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900938 break;
939 }
940 }
941
Tejun Heoccea34b2009-03-07 00:44:13 +0900942 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900943}
944EXPORT_SYMBOL_GPL(free_percpu);
945
Vivek Goyal3b034b02009-11-24 15:50:03 +0900946/**
Tejun Heo10fad5e2010-03-10 18:57:54 +0900947 * is_kernel_percpu_address - test whether address is from static percpu area
948 * @addr: address to test
949 *
950 * Test whether @addr belongs to in-kernel static percpu area. Module
951 * static percpu areas are not considered. For those, use
952 * is_module_percpu_address().
953 *
954 * RETURNS:
955 * %true if @addr is from in-kernel static percpu area, %false otherwise.
956 */
957bool is_kernel_percpu_address(unsigned long addr)
958{
Tejun Heobbddff02010-09-03 18:22:48 +0200959#ifdef CONFIG_SMP
Tejun Heo10fad5e2010-03-10 18:57:54 +0900960 const size_t static_size = __per_cpu_end - __per_cpu_start;
961 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
962 unsigned int cpu;
963
964 for_each_possible_cpu(cpu) {
965 void *start = per_cpu_ptr(base, cpu);
966
967 if ((void *)addr >= start && (void *)addr < start + static_size)
968 return true;
969 }
Tejun Heobbddff02010-09-03 18:22:48 +0200970#endif
971 /* on UP, can't distinguish from other static vars, always false */
Tejun Heo10fad5e2010-03-10 18:57:54 +0900972 return false;
973}
974
975/**
Vivek Goyal3b034b02009-11-24 15:50:03 +0900976 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
977 * @addr: the address to be converted to physical address
978 *
979 * Given @addr which is dereferenceable address obtained via one of
980 * percpu access macros, this function translates it into its physical
981 * address. The caller is responsible for ensuring @addr stays valid
982 * until this function finishes.
983 *
984 * RETURNS:
985 * The physical address for @addr.
986 */
987phys_addr_t per_cpu_ptr_to_phys(void *addr)
988{
Tejun Heo9983b6f02010-06-18 11:44:31 +0200989 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
990 bool in_first_chunk = false;
991 unsigned long first_start, first_end;
992 unsigned int cpu;
993
994 /*
995 * The following test on first_start/end isn't strictly
996 * necessary but will speed up lookups of addresses which
997 * aren't in the first chunk.
998 */
999 first_start = pcpu_chunk_addr(pcpu_first_chunk, pcpu_first_unit_cpu, 0);
1000 first_end = pcpu_chunk_addr(pcpu_first_chunk, pcpu_last_unit_cpu,
1001 pcpu_unit_pages);
1002 if ((unsigned long)addr >= first_start &&
1003 (unsigned long)addr < first_end) {
1004 for_each_possible_cpu(cpu) {
1005 void *start = per_cpu_ptr(base, cpu);
1006
1007 if (addr >= start && addr < start + pcpu_unit_size) {
1008 in_first_chunk = true;
1009 break;
1010 }
1011 }
1012 }
1013
1014 if (in_first_chunk) {
Tejun Heo020ec652010-04-09 18:57:00 +09001015 if ((unsigned long)addr < VMALLOC_START ||
1016 (unsigned long)addr >= VMALLOC_END)
1017 return __pa(addr);
1018 else
1019 return page_to_phys(vmalloc_to_page(addr));
1020 } else
Tejun Heo9f645532010-04-09 18:57:01 +09001021 return page_to_phys(pcpu_addr_to_page(addr));
Vivek Goyal3b034b02009-11-24 15:50:03 +09001022}
1023
Tejun Heofbf59bc2009-02-20 16:29:08 +09001024/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001025 * pcpu_alloc_alloc_info - allocate percpu allocation info
1026 * @nr_groups: the number of groups
1027 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001028 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001029 * Allocate ai which is large enough for @nr_groups groups containing
1030 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1031 * cpu_map array which is long enough for @nr_units and filled with
1032 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1033 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001034 *
1035 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001036 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1037 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001038 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001039struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1040 int nr_units)
1041{
1042 struct pcpu_alloc_info *ai;
1043 size_t base_size, ai_size;
1044 void *ptr;
1045 int unit;
1046
1047 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1048 __alignof__(ai->groups[0].cpu_map[0]));
1049 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1050
1051 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1052 if (!ptr)
1053 return NULL;
1054 ai = ptr;
1055 ptr += base_size;
1056
1057 ai->groups[0].cpu_map = ptr;
1058
1059 for (unit = 0; unit < nr_units; unit++)
1060 ai->groups[0].cpu_map[unit] = NR_CPUS;
1061
1062 ai->nr_groups = nr_groups;
1063 ai->__ai_size = PFN_ALIGN(ai_size);
1064
1065 return ai;
1066}
1067
1068/**
1069 * pcpu_free_alloc_info - free percpu allocation info
1070 * @ai: pcpu_alloc_info to free
1071 *
1072 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1073 */
1074void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1075{
1076 free_bootmem(__pa(ai), ai->__ai_size);
1077}
1078
1079/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001080 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1081 * @lvl: loglevel
1082 * @ai: allocation info to dump
1083 *
1084 * Print out information about @ai using loglevel @lvl.
1085 */
1086static void pcpu_dump_alloc_info(const char *lvl,
1087 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001088{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001089 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001090 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001091 int alloc = 0, alloc_end = 0;
1092 int group, v;
1093 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001094
Tejun Heofd1e8a12009-08-14 15:00:51 +09001095 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001096 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001097 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001098
Tejun Heofd1e8a12009-08-14 15:00:51 +09001099 v = num_possible_cpus();
1100 while (v /= 10)
1101 cpu_width++;
1102 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001103
Tejun Heofd1e8a12009-08-14 15:00:51 +09001104 upa = ai->alloc_size / ai->unit_size;
1105 width = upa * (cpu_width + 1) + group_width + 3;
1106 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001107
Tejun Heofd1e8a12009-08-14 15:00:51 +09001108 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1109 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1110 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1111
1112 for (group = 0; group < ai->nr_groups; group++) {
1113 const struct pcpu_group_info *gi = &ai->groups[group];
1114 int unit = 0, unit_end = 0;
1115
1116 BUG_ON(gi->nr_units % upa);
1117 for (alloc_end += gi->nr_units / upa;
1118 alloc < alloc_end; alloc++) {
1119 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001120 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001121 printk("%spcpu-alloc: ", lvl);
1122 }
1123 printk("[%0*d] ", group_width, group);
1124
1125 for (unit_end += upa; unit < unit_end; unit++)
1126 if (gi->cpu_map[unit] != NR_CPUS)
1127 printk("%0*d ", cpu_width,
1128 gi->cpu_map[unit]);
1129 else
1130 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001131 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001132 }
1133 printk("\n");
1134}
Tejun Heo033e48f2009-08-14 15:00:51 +09001135
Tejun Heofbf59bc2009-02-20 16:29:08 +09001136/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001137 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001138 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001139 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001140 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001141 * Initialize the first percpu chunk which contains the kernel static
1142 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001143 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001144 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001145 * @ai contains all information necessary to initialize the first
1146 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001147 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001148 * @ai->static_size is the size of static percpu area.
1149 *
1150 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001151 * reserve after the static area in the first chunk. This reserves
1152 * the first chunk such that it's available only through reserved
1153 * percpu allocation. This is primarily used to serve module percpu
1154 * static areas on architectures where the addressing model has
1155 * limited offset range for symbol relocations to guarantee module
1156 * percpu symbols fall inside the relocatable range.
1157 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001158 * @ai->dyn_size determines the number of bytes available for dynamic
1159 * allocation in the first chunk. The area between @ai->static_size +
1160 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001161 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001162 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1163 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1164 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001165 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001166 * @ai->atom_size is the allocation atom size and used as alignment
1167 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001168 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001169 * @ai->alloc_size is the allocation size and always multiple of
1170 * @ai->atom_size. This is larger than @ai->atom_size if
1171 * @ai->unit_size is larger than @ai->atom_size.
1172 *
1173 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1174 * percpu areas. Units which should be colocated are put into the
1175 * same group. Dynamic VM areas will be allocated according to these
1176 * groupings. If @ai->nr_groups is zero, a single group containing
1177 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001178 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001179 * The caller should have mapped the first chunk at @base_addr and
1180 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001181 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001182 * If the first chunk ends up with both reserved and dynamic areas, it
1183 * is served by two chunks - one to serve the core static and reserved
1184 * areas and the other for the dynamic area. They share the same vm
1185 * and page map but uses different area allocation map to stay away
1186 * from each other. The latter chunk is circulated in the chunk slots
1187 * and available for dynamic allocation like any other chunks.
1188 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001189 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001190 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001191 */
Tejun Heofb435d52009-08-14 15:00:51 +09001192int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1193 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001194{
Tejun Heo635b75f2009-09-24 09:43:11 +09001195 static char cpus_buf[4096] __initdata;
Tejun Heo099a19d2010-06-27 18:50:00 +02001196 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1197 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001198 size_t dyn_size = ai->dyn_size;
1199 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001200 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001201 unsigned long *group_offsets;
1202 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001203 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001204 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001205 int *unit_map;
1206 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001207
Tejun Heo635b75f2009-09-24 09:43:11 +09001208 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1209
1210#define PCPU_SETUP_BUG_ON(cond) do { \
1211 if (unlikely(cond)) { \
1212 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1213 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1214 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1215 BUG(); \
1216 } \
1217} while (0)
1218
Tejun Heo2f39e632009-07-04 08:11:00 +09001219 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001220 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02001221#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09001222 PCPU_SETUP_BUG_ON(!ai->static_size);
Tejun Heobbddff02010-09-03 18:22:48 +02001223#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09001224 PCPU_SETUP_BUG_ON(!base_addr);
1225 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1226 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1227 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001228 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001229 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001230
Tejun Heo65632972009-08-14 15:00:52 +09001231 /* process group information and build config tables accordingly */
1232 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1233 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001234 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001235 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001236
Tejun Heofd1e8a12009-08-14 15:00:51 +09001237 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001238 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001239 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001240
Tejun Heofd1e8a12009-08-14 15:00:51 +09001241 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1242 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001243
Tejun Heo65632972009-08-14 15:00:52 +09001244 group_offsets[group] = gi->base_offset;
1245 group_sizes[group] = gi->nr_units * ai->unit_size;
1246
Tejun Heofd1e8a12009-08-14 15:00:51 +09001247 for (i = 0; i < gi->nr_units; i++) {
1248 cpu = gi->cpu_map[i];
1249 if (cpu == NR_CPUS)
1250 continue;
1251
Tejun Heo635b75f2009-09-24 09:43:11 +09001252 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1253 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1254 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001255
1256 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001257 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1258
Tejun Heofd1e8a12009-08-14 15:00:51 +09001259 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001260 pcpu_first_unit_cpu = cpu;
Linus Torvalds0fc05312010-10-22 17:31:36 -07001261 pcpu_last_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001262 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001263 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001264 pcpu_nr_units = unit;
1265
1266 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001267 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1268
1269 /* we're done parsing the input, undefine BUG macro and dump config */
1270#undef PCPU_SETUP_BUG_ON
1271 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001272
Tejun Heo65632972009-08-14 15:00:52 +09001273 pcpu_nr_groups = ai->nr_groups;
1274 pcpu_group_offsets = group_offsets;
1275 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001276 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001277 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001278
1279 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001280 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001281 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001282 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001283 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1284 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001285
Tejun Heod9b55ee2009-02-24 11:57:21 +09001286 /*
1287 * Allocate chunk slots. The additional last slot is for
1288 * empty chunks.
1289 */
1290 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001291 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1292 for (i = 0; i < pcpu_nr_slots; i++)
1293 INIT_LIST_HEAD(&pcpu_slot[i]);
1294
Tejun Heoedcb4632009-03-06 14:33:59 +09001295 /*
1296 * Initialize static chunk. If reserved_size is zero, the
1297 * static chunk covers static area + dynamic allocation area
1298 * in the first chunk. If reserved_size is not zero, it
1299 * covers static area + reserved area (mostly used for module
1300 * static percpu allocation).
1301 */
Tejun Heo2441d152009-03-06 14:33:59 +09001302 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1303 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001304 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001305 schunk->map = smap;
1306 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001307 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001308 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001309
Tejun Heofd1e8a12009-08-14 15:00:51 +09001310 if (ai->reserved_size) {
1311 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001312 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001313 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001314 } else {
1315 schunk->free_size = dyn_size;
1316 dyn_size = 0; /* dynamic area covered */
1317 }
Tejun Heo2441d152009-03-06 14:33:59 +09001318 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001319
Tejun Heofd1e8a12009-08-14 15:00:51 +09001320 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001321 if (schunk->free_size)
1322 schunk->map[schunk->map_used++] = schunk->free_size;
1323
Tejun Heoedcb4632009-03-06 14:33:59 +09001324 /* init dynamic chunk if necessary */
1325 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001326 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001327 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001328 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001329 dchunk->map = dmap;
1330 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001331 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001332 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001333
1334 dchunk->contig_hint = dchunk->free_size = dyn_size;
1335 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1336 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1337 }
1338
Tejun Heo2441d152009-03-06 14:33:59 +09001339 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001340 pcpu_first_chunk = dchunk ?: schunk;
1341 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001342
1343 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001344 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001345 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001346}
Tejun Heo66c3a752009-03-10 16:27:48 +09001347
Tejun Heobbddff02010-09-03 18:22:48 +02001348#ifdef CONFIG_SMP
1349
Tejun Heof58dc012009-08-14 15:00:50 +09001350const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1351 [PCPU_FC_AUTO] = "auto",
1352 [PCPU_FC_EMBED] = "embed",
1353 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001354};
Tejun Heo66c3a752009-03-10 16:27:48 +09001355
Tejun Heof58dc012009-08-14 15:00:50 +09001356enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1357
1358static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001359{
Tejun Heof58dc012009-08-14 15:00:50 +09001360 if (0)
1361 /* nada */;
1362#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1363 else if (!strcmp(str, "embed"))
1364 pcpu_chosen_fc = PCPU_FC_EMBED;
1365#endif
1366#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1367 else if (!strcmp(str, "page"))
1368 pcpu_chosen_fc = PCPU_FC_PAGE;
1369#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001370 else
1371 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001372
Tejun Heof58dc012009-08-14 15:00:50 +09001373 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001374}
Tejun Heof58dc012009-08-14 15:00:50 +09001375early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001376
Tejun Heo3c9a0242010-09-09 18:00:15 +02001377/*
1378 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1379 * Build it if needed by the arch config or the generic setup is going
1380 * to be used.
1381 */
Tejun Heo08fc4582009-08-14 15:00:49 +09001382#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1383 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02001384#define BUILD_EMBED_FIRST_CHUNK
1385#endif
1386
1387/* build pcpu_page_first_chunk() iff needed by the arch config */
1388#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1389#define BUILD_PAGE_FIRST_CHUNK
1390#endif
1391
1392/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1393#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1394/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09001395 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1396 * @reserved_size: the size of reserved percpu area in bytes
1397 * @dyn_size: minimum free size for dynamic allocation in bytes
1398 * @atom_size: allocation atom size
1399 * @cpu_distance_fn: callback to determine distance between cpus, optional
1400 *
1401 * This function determines grouping of units, their mappings to cpus
1402 * and other parameters considering needed percpu size, allocation
1403 * atom size and distances between CPUs.
1404 *
1405 * Groups are always mutliples of atom size and CPUs which are of
1406 * LOCAL_DISTANCE both ways are grouped together and share space for
1407 * units in the same group. The returned configuration is guaranteed
1408 * to have CPUs on different nodes on different groups and >=75% usage
1409 * of allocated virtual address space.
1410 *
1411 * RETURNS:
1412 * On success, pointer to the new allocation_info is returned. On
1413 * failure, ERR_PTR value is returned.
1414 */
1415static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1416 size_t reserved_size, size_t dyn_size,
1417 size_t atom_size,
1418 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1419{
1420 static int group_map[NR_CPUS] __initdata;
1421 static int group_cnt[NR_CPUS] __initdata;
1422 const size_t static_size = __per_cpu_end - __per_cpu_start;
1423 int nr_groups = 1, nr_units = 0;
1424 size_t size_sum, min_unit_size, alloc_size;
1425 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1426 int last_allocs, group, unit;
1427 unsigned int cpu, tcpu;
1428 struct pcpu_alloc_info *ai;
1429 unsigned int *cpu_map;
1430
1431 /* this function may be called multiple times */
1432 memset(group_map, 0, sizeof(group_map));
1433 memset(group_cnt, 0, sizeof(group_cnt));
1434
1435 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1436 size_sum = PFN_ALIGN(static_size + reserved_size +
1437 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1438 dyn_size = size_sum - static_size - reserved_size;
1439
1440 /*
1441 * Determine min_unit_size, alloc_size and max_upa such that
1442 * alloc_size is multiple of atom_size and is the smallest
1443 * which can accomodate 4k aligned segments which are equal to
1444 * or larger than min_unit_size.
1445 */
1446 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1447
1448 alloc_size = roundup(min_unit_size, atom_size);
1449 upa = alloc_size / min_unit_size;
1450 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1451 upa--;
1452 max_upa = upa;
1453
1454 /* group cpus according to their proximity */
1455 for_each_possible_cpu(cpu) {
1456 group = 0;
1457 next_group:
1458 for_each_possible_cpu(tcpu) {
1459 if (cpu == tcpu)
1460 break;
1461 if (group_map[tcpu] == group && cpu_distance_fn &&
1462 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1463 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1464 group++;
1465 nr_groups = max(nr_groups, group + 1);
1466 goto next_group;
1467 }
1468 }
1469 group_map[cpu] = group;
1470 group_cnt[group]++;
1471 }
1472
1473 /*
1474 * Expand unit size until address space usage goes over 75%
1475 * and then as much as possible without using more address
1476 * space.
1477 */
1478 last_allocs = INT_MAX;
1479 for (upa = max_upa; upa; upa--) {
1480 int allocs = 0, wasted = 0;
1481
1482 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1483 continue;
1484
1485 for (group = 0; group < nr_groups; group++) {
1486 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1487 allocs += this_allocs;
1488 wasted += this_allocs * upa - group_cnt[group];
1489 }
1490
1491 /*
1492 * Don't accept if wastage is over 1/3. The
1493 * greater-than comparison ensures upa==1 always
1494 * passes the following check.
1495 */
1496 if (wasted > num_possible_cpus() / 3)
1497 continue;
1498
1499 /* and then don't consume more memory */
1500 if (allocs > last_allocs)
1501 break;
1502 last_allocs = allocs;
1503 best_upa = upa;
1504 }
1505 upa = best_upa;
1506
1507 /* allocate and fill alloc_info */
1508 for (group = 0; group < nr_groups; group++)
1509 nr_units += roundup(group_cnt[group], upa);
1510
1511 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1512 if (!ai)
1513 return ERR_PTR(-ENOMEM);
1514 cpu_map = ai->groups[0].cpu_map;
1515
1516 for (group = 0; group < nr_groups; group++) {
1517 ai->groups[group].cpu_map = cpu_map;
1518 cpu_map += roundup(group_cnt[group], upa);
1519 }
1520
1521 ai->static_size = static_size;
1522 ai->reserved_size = reserved_size;
1523 ai->dyn_size = dyn_size;
1524 ai->unit_size = alloc_size / upa;
1525 ai->atom_size = atom_size;
1526 ai->alloc_size = alloc_size;
1527
1528 for (group = 0, unit = 0; group_cnt[group]; group++) {
1529 struct pcpu_group_info *gi = &ai->groups[group];
1530
1531 /*
1532 * Initialize base_offset as if all groups are located
1533 * back-to-back. The caller should update this to
1534 * reflect actual allocation.
1535 */
1536 gi->base_offset = unit * ai->unit_size;
1537
1538 for_each_possible_cpu(cpu)
1539 if (group_map[cpu] == group)
1540 gi->cpu_map[gi->nr_units++] = cpu;
1541 gi->nr_units = roundup(gi->nr_units, upa);
1542 unit += gi->nr_units;
1543 }
1544 BUG_ON(unit != nr_units);
1545
1546 return ai;
1547}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001548#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001549
Tejun Heo3c9a0242010-09-09 18:00:15 +02001550#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09001551/**
1552 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001553 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001554 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001555 * @atom_size: allocation atom size
1556 * @cpu_distance_fn: callback to determine distance between cpus, optional
1557 * @alloc_fn: function to allocate percpu page
1558 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001559 *
1560 * This is a helper to ease setting up embedded first percpu chunk and
1561 * can be called where pcpu_setup_first_chunk() is expected.
1562 *
1563 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001564 * by calling @alloc_fn and used as-is without being mapped into
1565 * vmalloc area. Allocations are always whole multiples of @atom_size
1566 * aligned to @atom_size.
1567 *
1568 * This enables the first chunk to piggy back on the linear physical
1569 * mapping which often uses larger page size. Please note that this
1570 * can result in very sparse cpu->unit mapping on NUMA machines thus
1571 * requiring large vmalloc address space. Don't use this allocator if
1572 * vmalloc space is not orders of magnitude larger than distances
1573 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001574 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001575 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001576 *
1577 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001578 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001579 *
1580 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001581 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001582 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001583int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001584 size_t atom_size,
1585 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1586 pcpu_fc_alloc_fn_t alloc_fn,
1587 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001588{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001589 void *base = (void *)ULONG_MAX;
1590 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001591 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001592 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001593 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001594
Tejun Heoc8826dd2009-08-14 15:00:52 +09001595 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1596 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001597 if (IS_ERR(ai))
1598 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001599
Tejun Heofd1e8a12009-08-14 15:00:51 +09001600 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001601 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001602
Tejun Heoc8826dd2009-08-14 15:00:52 +09001603 areas = alloc_bootmem_nopanic(areas_size);
1604 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001605 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001606 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001607 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001608
Tejun Heoc8826dd2009-08-14 15:00:52 +09001609 /* allocate, copy and determine base address */
1610 for (group = 0; group < ai->nr_groups; group++) {
1611 struct pcpu_group_info *gi = &ai->groups[group];
1612 unsigned int cpu = NR_CPUS;
1613 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001614
Tejun Heoc8826dd2009-08-14 15:00:52 +09001615 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1616 cpu = gi->cpu_map[i];
1617 BUG_ON(cpu == NR_CPUS);
1618
1619 /* allocate space for the whole group */
1620 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1621 if (!ptr) {
1622 rc = -ENOMEM;
1623 goto out_free_areas;
1624 }
1625 areas[group] = ptr;
1626
1627 base = min(ptr, base);
1628
1629 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1630 if (gi->cpu_map[i] == NR_CPUS) {
1631 /* unused unit, free whole */
1632 free_fn(ptr, ai->unit_size);
1633 continue;
1634 }
1635 /* copy and return the unused part */
1636 memcpy(ptr, __per_cpu_load, ai->static_size);
1637 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1638 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001639 }
1640
Tejun Heoc8826dd2009-08-14 15:00:52 +09001641 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001642 max_distance = 0;
1643 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001644 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001645 max_distance = max_t(size_t, max_distance,
1646 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001647 }
1648 max_distance += ai->unit_size;
1649
1650 /* warn if maximum distance is further than 75% of vmalloc space */
1651 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001652 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001653 "space 0x%lx\n",
1654 max_distance, VMALLOC_END - VMALLOC_START);
1655#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1656 /* and fail if we have fallback */
1657 rc = -EINVAL;
1658 goto out_free;
1659#endif
1660 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001661
Tejun Heo004018e2009-08-14 15:00:49 +09001662 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001663 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1664 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001665
Tejun Heofb435d52009-08-14 15:00:51 +09001666 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001667 goto out_free;
1668
1669out_free_areas:
1670 for (group = 0; group < ai->nr_groups; group++)
1671 free_fn(areas[group],
1672 ai->groups[group].nr_units * ai->unit_size);
1673out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001674 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001675 if (areas)
1676 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001677 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001678}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001679#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001680
Tejun Heo3c9a0242010-09-09 18:00:15 +02001681#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001682/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001683 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001684 * @reserved_size: the size of reserved percpu area in bytes
1685 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1686 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1687 * @populate_pte_fn: function to populate pte
1688 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001689 * This is a helper to ease setting up page-remapped first percpu
1690 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001691 *
1692 * This is the basic allocator. Static percpu area is allocated
1693 * page-by-page into vmalloc area.
1694 *
1695 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001696 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001697 */
Tejun Heofb435d52009-08-14 15:00:51 +09001698int __init pcpu_page_first_chunk(size_t reserved_size,
1699 pcpu_fc_alloc_fn_t alloc_fn,
1700 pcpu_fc_free_fn_t free_fn,
1701 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001702{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001703 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001704 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001705 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001706 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001707 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001708 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001709 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001710
Tejun Heo00ae4062009-08-14 15:00:49 +09001711 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1712
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001713 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001714 if (IS_ERR(ai))
1715 return PTR_ERR(ai);
1716 BUG_ON(ai->nr_groups != 1);
1717 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1718
1719 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001720
1721 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001722 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1723 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09001724 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09001725
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001726 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001727 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001728 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001729 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001730 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001731 void *ptr;
1732
Tejun Heo3cbc8562009-08-14 15:00:50 +09001733 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001734 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001735 pr_warning("PERCPU: failed to allocate %s page "
1736 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001737 goto enomem;
1738 }
Tejun Heoce3141a2009-07-04 08:11:00 +09001739 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001740 }
1741
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001742 /* allocate vm area, map the pages and copy static data */
1743 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001744 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001745 vm_area_register_early(&vm, PAGE_SIZE);
1746
Tejun Heofd1e8a12009-08-14 15:00:51 +09001747 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001748 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001749 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001750
Tejun Heoce3141a2009-07-04 08:11:00 +09001751 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001752 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1753
1754 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001755 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1756 unit_pages);
1757 if (rc < 0)
1758 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001759
1760 /*
1761 * FIXME: Archs with virtual cache should flush local
1762 * cache for the linear mapping here - something
1763 * equivalent to flush_cache_vmap() on the local cpu.
1764 * flush_cache_vmap() can't be used as most supporting
1765 * data structures are not set up yet.
1766 */
1767
1768 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001769 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001770 }
1771
1772 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09001773 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001774 unit_pages, psize_str, vm.addr, ai->static_size,
1775 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001776
Tejun Heofb435d52009-08-14 15:00:51 +09001777 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001778 goto out_free_ar;
1779
1780enomem:
1781 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09001782 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09001783 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09001784out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09001785 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001786 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09001787 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001788}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001789#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001790
Tejun Heobbddff02010-09-03 18:22:48 +02001791#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09001792/*
Tejun Heobbddff02010-09-03 18:22:48 +02001793 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09001794 *
1795 * The embedding helper is used because its behavior closely resembles
1796 * the original non-dynamic generic percpu area setup. This is
1797 * important because many archs have addressing restrictions and might
1798 * fail if the percpu area is located far away from the previous
1799 * location. As an added bonus, in non-NUMA cases, embedding is
1800 * generally a good idea TLB-wise because percpu area can piggy back
1801 * on the physical linear memory mapping which uses large page
1802 * mappings on applicable archs.
1803 */
Tejun Heoe74e3962009-03-30 19:07:44 +09001804unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1805EXPORT_SYMBOL(__per_cpu_offset);
1806
Tejun Heoc8826dd2009-08-14 15:00:52 +09001807static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1808 size_t align)
1809{
1810 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1811}
1812
1813static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1814{
1815 free_bootmem(__pa(ptr), size);
1816}
1817
Tejun Heoe74e3962009-03-30 19:07:44 +09001818void __init setup_per_cpu_areas(void)
1819{
Tejun Heoe74e3962009-03-30 19:07:44 +09001820 unsigned long delta;
1821 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09001822 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09001823
1824 /*
1825 * Always reserve area for module percpu variables. That's
1826 * what the legacy allocator did.
1827 */
Tejun Heofb435d52009-08-14 15:00:51 +09001828 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001829 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1830 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09001831 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02001832 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09001833
1834 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1835 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09001836 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09001837}
Tejun Heobbddff02010-09-03 18:22:48 +02001838#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1839
1840#else /* CONFIG_SMP */
1841
1842/*
1843 * UP percpu area setup.
1844 *
1845 * UP always uses km-based percpu allocator with identity mapping.
1846 * Static percpu variables are indistinguishable from the usual static
1847 * variables and don't require any special preparation.
1848 */
1849void __init setup_per_cpu_areas(void)
1850{
1851 const size_t unit_size =
1852 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1853 PERCPU_DYNAMIC_RESERVE));
1854 struct pcpu_alloc_info *ai;
1855 void *fc;
1856
1857 ai = pcpu_alloc_alloc_info(1, 1);
1858 fc = __alloc_bootmem(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
1859 if (!ai || !fc)
1860 panic("Failed to allocate memory for percpu areas.");
1861
1862 ai->dyn_size = unit_size;
1863 ai->unit_size = unit_size;
1864 ai->atom_size = unit_size;
1865 ai->alloc_size = unit_size;
1866 ai->groups[0].nr_units = 1;
1867 ai->groups[0].cpu_map[0] = 0;
1868
1869 if (pcpu_setup_first_chunk(ai, fc) < 0)
1870 panic("Failed to initialize percpu areas.");
1871}
1872
1873#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02001874
1875/*
1876 * First and reserved chunks are initialized with temporary allocation
1877 * map in initdata so that they can be used before slab is online.
1878 * This function is called after slab is brought up and replaces those
1879 * with properly allocated maps.
1880 */
1881void __init percpu_init_late(void)
1882{
1883 struct pcpu_chunk *target_chunks[] =
1884 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1885 struct pcpu_chunk *chunk;
1886 unsigned long flags;
1887 int i;
1888
1889 for (i = 0; (chunk = target_chunks[i]); i++) {
1890 int *map;
1891 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1892
1893 BUILD_BUG_ON(size > PAGE_SIZE);
1894
1895 map = pcpu_mem_alloc(size);
1896 BUG_ON(!map);
1897
1898 spin_lock_irqsave(&pcpu_lock, flags);
1899 memcpy(map, chunk->map, size);
1900 chunk->map = map;
1901 spin_unlock_irqrestore(&pcpu_lock, flags);
1902 }
1903}