blob: 39f7dfd59585a118f4a8c6be480bf3b866a50f25 [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 Heoae9e6bc92009-04-02 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 Heo88999a82010-04-09 18:57:01 +0900232static unsigned long __maybe_unused 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 Heofbf59bc2009-02-20 16:29:08 +0900285 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900286 return kzalloc(size, GFP_KERNEL);
287 else {
288 void *ptr = vmalloc(size);
289 if (ptr)
290 memset(ptr, 0, size);
291 return ptr;
292 }
293}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900294
Tejun Heo1880d932009-03-07 00:44:09 +0900295/**
296 * pcpu_mem_free - free memory
297 * @ptr: memory to free
298 * @size: size of the area
299 *
300 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
301 */
302static void pcpu_mem_free(void *ptr, size_t size)
303{
304 if (size <= PAGE_SIZE)
305 kfree(ptr);
306 else
307 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900308}
309
310/**
311 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
312 * @chunk: chunk of interest
313 * @oslot: the previous slot it was on
314 *
315 * This function is called after an allocation or free changed @chunk.
316 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900317 * moved to the slot. Note that the reserved chunk is never put on
318 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900319 *
320 * CONTEXT:
321 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900322 */
323static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
324{
325 int nslot = pcpu_chunk_slot(chunk);
326
Tejun Heoedcb4632009-03-06 14:33:59 +0900327 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900328 if (oslot < nslot)
329 list_move(&chunk->list, &pcpu_slot[nslot]);
330 else
331 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
332 }
333}
334
Tejun Heofbf59bc2009-02-20 16:29:08 +0900335/**
Tejun Heo833af842009-11-11 15:35:18 +0900336 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
337 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900338 *
Tejun Heo833af842009-11-11 15:35:18 +0900339 * Determine whether area map of @chunk needs to be extended to
340 * accomodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900341 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900342 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900343 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900344 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900345 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900346 * New target map allocation length if extension is necessary, 0
347 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900348 */
Tejun Heo833af842009-11-11 15:35:18 +0900349static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900350{
351 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900352
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900353 if (chunk->map_alloc >= chunk->map_used + 2)
354 return 0;
355
356 new_alloc = PCPU_DFL_MAP_ALLOC;
357 while (new_alloc < chunk->map_used + 2)
358 new_alloc *= 2;
359
Tejun Heo833af842009-11-11 15:35:18 +0900360 return new_alloc;
361}
362
363/**
364 * pcpu_extend_area_map - extend area map of a chunk
365 * @chunk: chunk of interest
366 * @new_alloc: new target allocation length of the area map
367 *
368 * Extend area map of @chunk to have @new_alloc entries.
369 *
370 * CONTEXT:
371 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
372 *
373 * RETURNS:
374 * 0 on success, -errno on failure.
375 */
376static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
377{
378 int *old = NULL, *new = NULL;
379 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
380 unsigned long flags;
381
382 new = pcpu_mem_alloc(new_size);
383 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900384 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900385
Tejun Heo833af842009-11-11 15:35:18 +0900386 /* acquire pcpu_lock and switch to new area map */
387 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900388
Tejun Heo833af842009-11-11 15:35:18 +0900389 if (new_alloc <= chunk->map_alloc)
390 goto out_unlock;
391
392 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
393 memcpy(new, chunk->map, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900394
395 /*
396 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
397 * one of the first chunks and still using static map.
398 */
399 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
Tejun Heo833af842009-11-11 15:35:18 +0900400 old = chunk->map;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900401
402 chunk->map_alloc = new_alloc;
403 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900404 new = NULL;
405
406out_unlock:
407 spin_unlock_irqrestore(&pcpu_lock, flags);
408
409 /*
410 * pcpu_mem_free() might end up calling vfree() which uses
411 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
412 */
413 pcpu_mem_free(old, old_size);
414 pcpu_mem_free(new, new_size);
415
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900416 return 0;
417}
418
419/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900420 * pcpu_split_block - split a map block
421 * @chunk: chunk of interest
422 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900423 * @head: head size in bytes (can be 0)
424 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900425 *
426 * Split the @i'th map block into two or three blocks. If @head is
427 * non-zero, @head bytes block is inserted before block @i moving it
428 * to @i+1 and reducing its size by @head bytes.
429 *
430 * If @tail is non-zero, the target block, which can be @i or @i+1
431 * depending on @head, is reduced by @tail bytes and @tail byte block
432 * is inserted after the target block.
433 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900434 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900435 *
436 * CONTEXT:
437 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900438 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900439static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
440 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900441{
442 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900443
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900444 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900445
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900446 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900447 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
448 sizeof(chunk->map[0]) * (chunk->map_used - i));
449 chunk->map_used += nr_extra;
450
451 if (head) {
452 chunk->map[i + 1] = chunk->map[i] - head;
453 chunk->map[i++] = head;
454 }
455 if (tail) {
456 chunk->map[i++] -= tail;
457 chunk->map[i] = tail;
458 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900459}
460
461/**
462 * pcpu_alloc_area - allocate area from a pcpu_chunk
463 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900464 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900465 * @align: wanted align
466 *
467 * Try to allocate @size bytes area aligned at @align from @chunk.
468 * Note that this function only allocates the offset. It doesn't
469 * populate or map the area.
470 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900471 * @chunk->map must have at least two free slots.
472 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900473 * CONTEXT:
474 * pcpu_lock.
475 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900476 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900477 * Allocated offset in @chunk on success, -1 if no matching area is
478 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900479 */
480static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
481{
482 int oslot = pcpu_chunk_slot(chunk);
483 int max_contig = 0;
484 int i, off;
485
Tejun Heofbf59bc2009-02-20 16:29:08 +0900486 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
487 bool is_last = i + 1 == chunk->map_used;
488 int head, tail;
489
490 /* extra for alignment requirement */
491 head = ALIGN(off, align) - off;
492 BUG_ON(i == 0 && head != 0);
493
494 if (chunk->map[i] < 0)
495 continue;
496 if (chunk->map[i] < head + size) {
497 max_contig = max(chunk->map[i], max_contig);
498 continue;
499 }
500
501 /*
502 * If head is small or the previous block is free,
503 * merge'em. Note that 'small' is defined as smaller
504 * than sizeof(int), which is very small but isn't too
505 * uncommon for percpu allocations.
506 */
507 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
508 if (chunk->map[i - 1] > 0)
509 chunk->map[i - 1] += head;
510 else {
511 chunk->map[i - 1] -= head;
512 chunk->free_size -= head;
513 }
514 chunk->map[i] -= head;
515 off += head;
516 head = 0;
517 }
518
519 /* if tail is small, just keep it around */
520 tail = chunk->map[i] - head - size;
521 if (tail < sizeof(int))
522 tail = 0;
523
524 /* split if warranted */
525 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900526 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900527 if (head) {
528 i++;
529 off += head;
530 max_contig = max(chunk->map[i - 1], max_contig);
531 }
532 if (tail)
533 max_contig = max(chunk->map[i + 1], max_contig);
534 }
535
536 /* update hint and mark allocated */
537 if (is_last)
538 chunk->contig_hint = max_contig; /* fully scanned */
539 else
540 chunk->contig_hint = max(chunk->contig_hint,
541 max_contig);
542
543 chunk->free_size -= chunk->map[i];
544 chunk->map[i] = -chunk->map[i];
545
546 pcpu_chunk_relocate(chunk, oslot);
547 return off;
548 }
549
550 chunk->contig_hint = max_contig; /* fully scanned */
551 pcpu_chunk_relocate(chunk, oslot);
552
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900553 /* tell the upper layer that this chunk has no matching area */
554 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900555}
556
557/**
558 * pcpu_free_area - free area to a pcpu_chunk
559 * @chunk: chunk of interest
560 * @freeme: offset of area to free
561 *
562 * Free area starting from @freeme to @chunk. Note that this function
563 * only modifies the allocation map. It doesn't depopulate or unmap
564 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900565 *
566 * CONTEXT:
567 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900568 */
569static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
570{
571 int oslot = pcpu_chunk_slot(chunk);
572 int i, off;
573
574 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
575 if (off == freeme)
576 break;
577 BUG_ON(off != freeme);
578 BUG_ON(chunk->map[i] > 0);
579
580 chunk->map[i] = -chunk->map[i];
581 chunk->free_size += chunk->map[i];
582
583 /* merge with previous? */
584 if (i > 0 && chunk->map[i - 1] >= 0) {
585 chunk->map[i - 1] += chunk->map[i];
586 chunk->map_used--;
587 memmove(&chunk->map[i], &chunk->map[i + 1],
588 (chunk->map_used - i) * sizeof(chunk->map[0]));
589 i--;
590 }
591 /* merge with next? */
592 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
593 chunk->map[i] += chunk->map[i + 1];
594 chunk->map_used--;
595 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
596 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
597 }
598
599 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
600 pcpu_chunk_relocate(chunk, oslot);
601}
602
Tejun Heo60810892010-04-09 18:57:01 +0900603static struct pcpu_chunk *pcpu_alloc_chunk(void)
604{
605 struct pcpu_chunk *chunk;
606
607 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
608 if (!chunk)
609 return NULL;
610
611 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
612 if (!chunk->map) {
613 kfree(chunk);
614 return NULL;
615 }
616
617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618 chunk->map[chunk->map_used++] = pcpu_unit_size;
619
620 INIT_LIST_HEAD(&chunk->list);
621 chunk->free_size = pcpu_unit_size;
622 chunk->contig_hint = pcpu_unit_size;
623
624 return chunk;
625}
626
627static void pcpu_free_chunk(struct pcpu_chunk *chunk)
628{
629 if (!chunk)
630 return;
631 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
632 kfree(chunk);
633}
634
Tejun Heo9f645532010-04-09 18:57:01 +0900635/*
636 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900637 *
Tejun Heo9f645532010-04-09 18:57:01 +0900638 * To allow different implementations, chunk alloc/free and
639 * [de]population are implemented in a separate file which is pulled
640 * into this file and compiled together. The following functions
641 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900642 *
Tejun Heo9f645532010-04-09 18:57:01 +0900643 * pcpu_populate_chunk - populate the specified range of a chunk
644 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
645 * pcpu_create_chunk - create a new chunk
646 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
647 * pcpu_addr_to_page - translate address to physical address
648 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900649 */
Tejun Heo9f645532010-04-09 18:57:01 +0900650static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
651static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
652static struct pcpu_chunk *pcpu_create_chunk(void);
653static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
654static struct page *pcpu_addr_to_page(void *addr);
655static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900656
Tejun Heob0c97782010-04-09 18:57:01 +0900657#ifdef CONFIG_NEED_PER_CPU_KM
658#include "percpu-km.c"
659#else
Tejun Heo9f645532010-04-09 18:57:01 +0900660#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900661#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900662
663/**
Tejun Heo88999a82010-04-09 18:57:01 +0900664 * pcpu_chunk_addr_search - determine chunk containing specified address
665 * @addr: address for which the chunk needs to be determined.
666 *
667 * RETURNS:
668 * The address of the found chunk.
669 */
670static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
671{
672 /* is it in the first chunk? */
673 if (pcpu_addr_in_first_chunk(addr)) {
674 /* is it in the reserved area? */
675 if (pcpu_addr_in_reserved_chunk(addr))
676 return pcpu_reserved_chunk;
677 return pcpu_first_chunk;
678 }
679
680 /*
681 * The address is relative to unit0 which might be unused and
682 * thus unmapped. Offset the address to the unit space of the
683 * current processor before looking it up in the vmalloc
684 * space. Note that any possible cpu id can be used here, so
685 * there's no need to worry about preemption or cpu hotplug.
686 */
687 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900688 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900689}
690
691/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900692 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900693 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900694 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900695 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900696 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900697 * Allocate percpu area of @size bytes aligned at @align.
698 *
699 * CONTEXT:
700 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900701 *
702 * RETURNS:
703 * Percpu pointer to the allocated area on success, NULL on failure.
704 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900705static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900706{
Tejun Heof2badb02009-09-29 09:17:58 +0900707 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900708 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900709 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +0900710 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900711 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900712
Tejun Heo8d408b42009-02-24 11:57:21 +0900713 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900714 WARN(true, "illegal size (%zu) or align (%zu) for "
715 "percpu allocation\n", size, align);
716 return NULL;
717 }
718
Tejun Heoccea34b2009-03-07 00:44:13 +0900719 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +0900720 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900721
Tejun Heoedcb4632009-03-06 14:33:59 +0900722 /* serve reserved allocations from the reserved chunk if available */
723 if (reserved && pcpu_reserved_chunk) {
724 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900725
726 if (size > chunk->contig_hint) {
727 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900728 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900729 }
Tejun Heo833af842009-11-11 15:35:18 +0900730
731 while ((new_alloc = pcpu_need_to_extend(chunk))) {
732 spin_unlock_irqrestore(&pcpu_lock, flags);
733 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
734 err = "failed to extend area map of reserved chunk";
735 goto fail_unlock_mutex;
736 }
737 spin_lock_irqsave(&pcpu_lock, flags);
738 }
739
Tejun Heoedcb4632009-03-06 14:33:59 +0900740 off = pcpu_alloc_area(chunk, size, align);
741 if (off >= 0)
742 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900743
Tejun Heof2badb02009-09-29 09:17:58 +0900744 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900745 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900746 }
747
Tejun Heoccea34b2009-03-07 00:44:13 +0900748restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900749 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900750 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
751 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
752 if (size > chunk->contig_hint)
753 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900754
Tejun Heo833af842009-11-11 15:35:18 +0900755 new_alloc = pcpu_need_to_extend(chunk);
756 if (new_alloc) {
757 spin_unlock_irqrestore(&pcpu_lock, flags);
758 if (pcpu_extend_area_map(chunk,
759 new_alloc) < 0) {
760 err = "failed to extend area map";
761 goto fail_unlock_mutex;
762 }
763 spin_lock_irqsave(&pcpu_lock, flags);
764 /*
765 * pcpu_lock has been dropped, need to
766 * restart cpu_slot list walking.
767 */
768 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900769 }
770
Tejun Heofbf59bc2009-02-20 16:29:08 +0900771 off = pcpu_alloc_area(chunk, size, align);
772 if (off >= 0)
773 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900774 }
775 }
776
777 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +0900778 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900779
Tejun Heo60810892010-04-09 18:57:01 +0900780 chunk = pcpu_create_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +0900781 if (!chunk) {
782 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +0900783 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +0900784 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900785
Jiri Kosina403a91b2009-10-29 00:25:59 +0900786 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900787 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900788 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900789
790area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900791 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900792
Tejun Heofbf59bc2009-02-20 16:29:08 +0900793 /* populate, map and clear the area */
794 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +0900795 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900796 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +0900797 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +0900798 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900799 }
800
Tejun Heoccea34b2009-03-07 00:44:13 +0900801 mutex_unlock(&pcpu_alloc_mutex);
802
Tejun Heobba174f2009-08-14 15:00:51 +0900803 /* return address relative to base address */
804 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900805
806fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900807 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900808fail_unlock_mutex:
809 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +0900810 if (warn_limit) {
811 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
812 "%s\n", size, align, err);
813 dump_stack();
814 if (!--warn_limit)
815 pr_info("PERCPU: limit reached, disable warning\n");
816 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900817 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900818}
Tejun Heoedcb4632009-03-06 14:33:59 +0900819
820/**
821 * __alloc_percpu - allocate dynamic percpu area
822 * @size: size of area to allocate in bytes
823 * @align: alignment of area (max PAGE_SIZE)
824 *
825 * Allocate percpu area of @size bytes aligned at @align. Might
826 * sleep. Might trigger writeouts.
827 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900828 * CONTEXT:
829 * Does GFP_KERNEL allocation.
830 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900831 * RETURNS:
832 * Percpu pointer to the allocated area on success, NULL on failure.
833 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900834void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900835{
836 return pcpu_alloc(size, align, false);
837}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900838EXPORT_SYMBOL_GPL(__alloc_percpu);
839
Tejun Heoedcb4632009-03-06 14:33:59 +0900840/**
841 * __alloc_reserved_percpu - allocate reserved percpu area
842 * @size: size of area to allocate in bytes
843 * @align: alignment of area (max PAGE_SIZE)
844 *
845 * Allocate percpu area of @size bytes aligned at @align from reserved
846 * percpu area if arch has set it up; otherwise, allocation is served
847 * from the same dynamic area. Might sleep. Might trigger writeouts.
848 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900849 * CONTEXT:
850 * Does GFP_KERNEL allocation.
851 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900852 * RETURNS:
853 * Percpu pointer to the allocated area on success, NULL on failure.
854 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900855void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900856{
857 return pcpu_alloc(size, align, true);
858}
859
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900860/**
861 * pcpu_reclaim - reclaim fully free chunks, workqueue function
862 * @work: unused
863 *
864 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900865 *
866 * CONTEXT:
867 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900868 */
869static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900870{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900871 LIST_HEAD(todo);
872 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
873 struct pcpu_chunk *chunk, *next;
874
Tejun Heoccea34b2009-03-07 00:44:13 +0900875 mutex_lock(&pcpu_alloc_mutex);
876 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900877
878 list_for_each_entry_safe(chunk, next, head, list) {
879 WARN_ON(chunk->immutable);
880
881 /* spare the first one */
882 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
883 continue;
884
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900885 list_move(&chunk->list, &todo);
886 }
887
Tejun Heoccea34b2009-03-07 00:44:13 +0900888 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900889
890 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +0900891 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heo60810892010-04-09 18:57:01 +0900892 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900893 }
Tejun Heo971f3912009-08-14 15:00:49 +0900894
895 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900896}
897
898/**
899 * free_percpu - free percpu area
900 * @ptr: pointer to area to free
901 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900902 * Free percpu area @ptr.
903 *
904 * CONTEXT:
905 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900906 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900907void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900908{
Andrew Morton129182e2010-01-08 14:42:39 -0800909 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900910 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900911 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900912 int off;
913
914 if (!ptr)
915 return;
916
Andrew Morton129182e2010-01-08 14:42:39 -0800917 addr = __pcpu_ptr_to_addr(ptr);
918
Tejun Heoccea34b2009-03-07 00:44:13 +0900919 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900920
921 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +0900922 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900923
924 pcpu_free_area(chunk, off);
925
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900926 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900927 if (chunk->free_size == pcpu_unit_size) {
928 struct pcpu_chunk *pos;
929
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900930 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900931 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900932 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900933 break;
934 }
935 }
936
Tejun Heoccea34b2009-03-07 00:44:13 +0900937 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900938}
939EXPORT_SYMBOL_GPL(free_percpu);
940
Vivek Goyal3b034b02009-11-24 15:50:03 +0900941/**
Tejun Heo10fad5e2010-03-10 18:57:54 +0900942 * is_kernel_percpu_address - test whether address is from static percpu area
943 * @addr: address to test
944 *
945 * Test whether @addr belongs to in-kernel static percpu area. Module
946 * static percpu areas are not considered. For those, use
947 * is_module_percpu_address().
948 *
949 * RETURNS:
950 * %true if @addr is from in-kernel static percpu area, %false otherwise.
951 */
952bool is_kernel_percpu_address(unsigned long addr)
953{
954 const size_t static_size = __per_cpu_end - __per_cpu_start;
955 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
956 unsigned int cpu;
957
958 for_each_possible_cpu(cpu) {
959 void *start = per_cpu_ptr(base, cpu);
960
961 if ((void *)addr >= start && (void *)addr < start + static_size)
962 return true;
963 }
964 return false;
965}
966
967/**
Vivek Goyal3b034b02009-11-24 15:50:03 +0900968 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
969 * @addr: the address to be converted to physical address
970 *
971 * Given @addr which is dereferenceable address obtained via one of
972 * percpu access macros, this function translates it into its physical
973 * address. The caller is responsible for ensuring @addr stays valid
974 * until this function finishes.
975 *
976 * RETURNS:
977 * The physical address for @addr.
978 */
979phys_addr_t per_cpu_ptr_to_phys(void *addr)
980{
Tejun Heo020ec652010-04-09 18:57:00 +0900981 if (pcpu_addr_in_first_chunk(addr)) {
982 if ((unsigned long)addr < VMALLOC_START ||
983 (unsigned long)addr >= VMALLOC_END)
984 return __pa(addr);
985 else
986 return page_to_phys(vmalloc_to_page(addr));
987 } else
Tejun Heo9f645532010-04-09 18:57:01 +0900988 return page_to_phys(pcpu_addr_to_page(addr));
Vivek Goyal3b034b02009-11-24 15:50:03 +0900989}
990
Tejun Heo033e48f2009-08-14 15:00:51 +0900991static inline size_t pcpu_calc_fc_sizes(size_t static_size,
992 size_t reserved_size,
993 ssize_t *dyn_sizep)
994{
995 size_t size_sum;
996
997 size_sum = PFN_ALIGN(static_size + reserved_size +
998 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
999 if (*dyn_sizep != 0)
1000 *dyn_sizep = size_sum - static_size - reserved_size;
1001
1002 return size_sum;
1003}
1004
Tejun Heofbf59bc2009-02-20 16:29:08 +09001005/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001006 * pcpu_alloc_alloc_info - allocate percpu allocation info
1007 * @nr_groups: the number of groups
1008 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001009 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001010 * Allocate ai which is large enough for @nr_groups groups containing
1011 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1012 * cpu_map array which is long enough for @nr_units and filled with
1013 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1014 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001015 *
1016 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001017 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1018 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001019 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001020struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1021 int nr_units)
1022{
1023 struct pcpu_alloc_info *ai;
1024 size_t base_size, ai_size;
1025 void *ptr;
1026 int unit;
1027
1028 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1029 __alignof__(ai->groups[0].cpu_map[0]));
1030 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1031
1032 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1033 if (!ptr)
1034 return NULL;
1035 ai = ptr;
1036 ptr += base_size;
1037
1038 ai->groups[0].cpu_map = ptr;
1039
1040 for (unit = 0; unit < nr_units; unit++)
1041 ai->groups[0].cpu_map[unit] = NR_CPUS;
1042
1043 ai->nr_groups = nr_groups;
1044 ai->__ai_size = PFN_ALIGN(ai_size);
1045
1046 return ai;
1047}
1048
1049/**
1050 * pcpu_free_alloc_info - free percpu allocation info
1051 * @ai: pcpu_alloc_info to free
1052 *
1053 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1054 */
1055void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1056{
1057 free_bootmem(__pa(ai), ai->__ai_size);
1058}
1059
1060/**
1061 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
Tejun Heoedcb4632009-03-06 14:33:59 +09001062 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +09001063 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heofd1e8a12009-08-14 15:00:51 +09001064 * @atom_size: allocation atom size
1065 * @cpu_distance_fn: callback to determine distance between cpus, optional
1066 *
1067 * This function determines grouping of units, their mappings to cpus
1068 * and other parameters considering needed percpu size, allocation
1069 * atom size and distances between CPUs.
1070 *
1071 * Groups are always mutliples of atom size and CPUs which are of
1072 * LOCAL_DISTANCE both ways are grouped together and share space for
1073 * units in the same group. The returned configuration is guaranteed
1074 * to have CPUs on different nodes on different groups and >=75% usage
1075 * of allocated virtual address space.
1076 *
1077 * RETURNS:
1078 * On success, pointer to the new allocation_info is returned. On
1079 * failure, ERR_PTR value is returned.
1080 */
1081struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1082 size_t reserved_size, ssize_t dyn_size,
1083 size_t atom_size,
1084 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
Tejun Heo033e48f2009-08-14 15:00:51 +09001085{
1086 static int group_map[NR_CPUS] __initdata;
1087 static int group_cnt[NR_CPUS] __initdata;
1088 const size_t static_size = __per_cpu_end - __per_cpu_start;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001089 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
Tejun Heo033e48f2009-08-14 15:00:51 +09001090 size_t size_sum, min_unit_size, alloc_size;
1091 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001092 int last_allocs, group, unit;
Tejun Heo033e48f2009-08-14 15:00:51 +09001093 unsigned int cpu, tcpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001094 struct pcpu_alloc_info *ai;
1095 unsigned int *cpu_map;
Tejun Heo033e48f2009-08-14 15:00:51 +09001096
Tejun Heofb59e722009-09-24 18:50:34 +09001097 /* this function may be called multiple times */
1098 memset(group_map, 0, sizeof(group_map));
1099 memset(group_cnt, 0, sizeof(group_map));
1100
Tejun Heo033e48f2009-08-14 15:00:51 +09001101 /*
1102 * Determine min_unit_size, alloc_size and max_upa such that
Tejun Heofd1e8a12009-08-14 15:00:51 +09001103 * alloc_size is multiple of atom_size and is the smallest
Tejun Heo033e48f2009-08-14 15:00:51 +09001104 * which can accomodate 4k aligned segments which are equal to
1105 * or larger than min_unit_size.
1106 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001107 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001108 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1109
Tejun Heofd1e8a12009-08-14 15:00:51 +09001110 alloc_size = roundup(min_unit_size, atom_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001111 upa = alloc_size / min_unit_size;
1112 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1113 upa--;
1114 max_upa = upa;
1115
1116 /* group cpus according to their proximity */
1117 for_each_possible_cpu(cpu) {
1118 group = 0;
1119 next_group:
1120 for_each_possible_cpu(tcpu) {
1121 if (cpu == tcpu)
1122 break;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001123 if (group_map[tcpu] == group && cpu_distance_fn &&
Tejun Heo033e48f2009-08-14 15:00:51 +09001124 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1125 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1126 group++;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001127 nr_groups = max(nr_groups, group + 1);
Tejun Heo033e48f2009-08-14 15:00:51 +09001128 goto next_group;
1129 }
1130 }
1131 group_map[cpu] = group;
1132 group_cnt[group]++;
1133 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1134 }
1135
1136 /*
1137 * Expand unit size until address space usage goes over 75%
1138 * and then as much as possible without using more address
1139 * space.
1140 */
1141 last_allocs = INT_MAX;
1142 for (upa = max_upa; upa; upa--) {
1143 int allocs = 0, wasted = 0;
1144
1145 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1146 continue;
1147
Tejun Heofd1e8a12009-08-14 15:00:51 +09001148 for (group = 0; group < nr_groups; group++) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001149 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1150 allocs += this_allocs;
1151 wasted += this_allocs * upa - group_cnt[group];
1152 }
1153
1154 /*
1155 * Don't accept if wastage is over 25%. The
1156 * greater-than comparison ensures upa==1 always
1157 * passes the following check.
1158 */
1159 if (wasted > num_possible_cpus() / 3)
1160 continue;
1161
1162 /* and then don't consume more memory */
1163 if (allocs > last_allocs)
1164 break;
1165 last_allocs = allocs;
1166 best_upa = upa;
1167 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001168 upa = best_upa;
Tejun Heo033e48f2009-08-14 15:00:51 +09001169
Tejun Heofd1e8a12009-08-14 15:00:51 +09001170 /* allocate and fill alloc_info */
1171 for (group = 0; group < nr_groups; group++)
1172 nr_units += roundup(group_cnt[group], upa);
1173
1174 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1175 if (!ai)
1176 return ERR_PTR(-ENOMEM);
1177 cpu_map = ai->groups[0].cpu_map;
1178
1179 for (group = 0; group < nr_groups; group++) {
1180 ai->groups[group].cpu_map = cpu_map;
1181 cpu_map += roundup(group_cnt[group], upa);
Tejun Heo033e48f2009-08-14 15:00:51 +09001182 }
1183
Tejun Heofd1e8a12009-08-14 15:00:51 +09001184 ai->static_size = static_size;
1185 ai->reserved_size = reserved_size;
1186 ai->dyn_size = dyn_size;
1187 ai->unit_size = alloc_size / upa;
1188 ai->atom_size = atom_size;
1189 ai->alloc_size = alloc_size;
1190
1191 for (group = 0, unit = 0; group_cnt[group]; group++) {
1192 struct pcpu_group_info *gi = &ai->groups[group];
1193
1194 /*
1195 * Initialize base_offset as if all groups are located
1196 * back-to-back. The caller should update this to
1197 * reflect actual allocation.
1198 */
1199 gi->base_offset = unit * ai->unit_size;
1200
1201 for_each_possible_cpu(cpu)
1202 if (group_map[cpu] == group)
1203 gi->cpu_map[gi->nr_units++] = cpu;
1204 gi->nr_units = roundup(gi->nr_units, upa);
1205 unit += gi->nr_units;
1206 }
1207 BUG_ON(unit != nr_units);
1208
1209 return ai;
Tejun Heo033e48f2009-08-14 15:00:51 +09001210}
1211
Tejun Heofd1e8a12009-08-14 15:00:51 +09001212/**
1213 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1214 * @lvl: loglevel
1215 * @ai: allocation info to dump
1216 *
1217 * Print out information about @ai using loglevel @lvl.
1218 */
1219static void pcpu_dump_alloc_info(const char *lvl,
1220 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001221{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001222 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001223 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001224 int alloc = 0, alloc_end = 0;
1225 int group, v;
1226 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001227
Tejun Heofd1e8a12009-08-14 15:00:51 +09001228 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001229 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001230 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001231
Tejun Heofd1e8a12009-08-14 15:00:51 +09001232 v = num_possible_cpus();
1233 while (v /= 10)
1234 cpu_width++;
1235 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001236
Tejun Heofd1e8a12009-08-14 15:00:51 +09001237 upa = ai->alloc_size / ai->unit_size;
1238 width = upa * (cpu_width + 1) + group_width + 3;
1239 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001240
Tejun Heofd1e8a12009-08-14 15:00:51 +09001241 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1242 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1243 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1244
1245 for (group = 0; group < ai->nr_groups; group++) {
1246 const struct pcpu_group_info *gi = &ai->groups[group];
1247 int unit = 0, unit_end = 0;
1248
1249 BUG_ON(gi->nr_units % upa);
1250 for (alloc_end += gi->nr_units / upa;
1251 alloc < alloc_end; alloc++) {
1252 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001253 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001254 printk("%spcpu-alloc: ", lvl);
1255 }
1256 printk("[%0*d] ", group_width, group);
1257
1258 for (unit_end += upa; unit < unit_end; unit++)
1259 if (gi->cpu_map[unit] != NR_CPUS)
1260 printk("%0*d ", cpu_width,
1261 gi->cpu_map[unit]);
1262 else
1263 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001264 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001265 }
1266 printk("\n");
1267}
Tejun Heo033e48f2009-08-14 15:00:51 +09001268
Tejun Heofbf59bc2009-02-20 16:29:08 +09001269/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001270 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001271 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001272 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001273 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001274 * Initialize the first percpu chunk which contains the kernel static
1275 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001276 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001277 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001278 * @ai contains all information necessary to initialize the first
1279 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001280 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001281 * @ai->static_size is the size of static percpu area.
1282 *
1283 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001284 * reserve after the static area in the first chunk. This reserves
1285 * the first chunk such that it's available only through reserved
1286 * percpu allocation. This is primarily used to serve module percpu
1287 * static areas on architectures where the addressing model has
1288 * limited offset range for symbol relocations to guarantee module
1289 * percpu symbols fall inside the relocatable range.
1290 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001291 * @ai->dyn_size determines the number of bytes available for dynamic
1292 * allocation in the first chunk. The area between @ai->static_size +
1293 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001294 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001295 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1296 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1297 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001298 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001299 * @ai->atom_size is the allocation atom size and used as alignment
1300 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001301 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001302 * @ai->alloc_size is the allocation size and always multiple of
1303 * @ai->atom_size. This is larger than @ai->atom_size if
1304 * @ai->unit_size is larger than @ai->atom_size.
1305 *
1306 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1307 * percpu areas. Units which should be colocated are put into the
1308 * same group. Dynamic VM areas will be allocated according to these
1309 * groupings. If @ai->nr_groups is zero, a single group containing
1310 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001311 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001312 * The caller should have mapped the first chunk at @base_addr and
1313 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001314 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001315 * If the first chunk ends up with both reserved and dynamic areas, it
1316 * is served by two chunks - one to serve the core static and reserved
1317 * areas and the other for the dynamic area. They share the same vm
1318 * and page map but uses different area allocation map to stay away
1319 * from each other. The latter chunk is circulated in the chunk slots
1320 * and available for dynamic allocation like any other chunks.
1321 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001322 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001323 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001324 */
Tejun Heofb435d52009-08-14 15:00:51 +09001325int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1326 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001327{
Tejun Heo635b75f2009-09-24 09:43:11 +09001328 static char cpus_buf[4096] __initdata;
Tejun Heoedcb4632009-03-06 14:33:59 +09001329 static int smap[2], dmap[2];
Tejun Heofd1e8a12009-08-14 15:00:51 +09001330 size_t dyn_size = ai->dyn_size;
1331 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001332 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001333 unsigned long *group_offsets;
1334 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001335 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001336 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001337 int *unit_map;
1338 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001339
Tejun Heo635b75f2009-09-24 09:43:11 +09001340 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1341
1342#define PCPU_SETUP_BUG_ON(cond) do { \
1343 if (unlikely(cond)) { \
1344 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1345 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1346 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1347 BUG(); \
1348 } \
1349} while (0)
1350
Tejun Heo2f39e632009-07-04 08:11:00 +09001351 /* sanity checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001352 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1353 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo635b75f2009-09-24 09:43:11 +09001354 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1355 PCPU_SETUP_BUG_ON(!ai->static_size);
1356 PCPU_SETUP_BUG_ON(!base_addr);
1357 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1358 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1359 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001360 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001361
Tejun Heo65632972009-08-14 15:00:52 +09001362 /* process group information and build config tables accordingly */
1363 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1364 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001365 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001366 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001367
Tejun Heofd1e8a12009-08-14 15:00:51 +09001368 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001369 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001370 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001371
Tejun Heofd1e8a12009-08-14 15:00:51 +09001372 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1373 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001374
Tejun Heo65632972009-08-14 15:00:52 +09001375 group_offsets[group] = gi->base_offset;
1376 group_sizes[group] = gi->nr_units * ai->unit_size;
1377
Tejun Heofd1e8a12009-08-14 15:00:51 +09001378 for (i = 0; i < gi->nr_units; i++) {
1379 cpu = gi->cpu_map[i];
1380 if (cpu == NR_CPUS)
1381 continue;
1382
Tejun Heo635b75f2009-09-24 09:43:11 +09001383 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1384 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1385 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001386
1387 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001388 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1389
Tejun Heofd1e8a12009-08-14 15:00:51 +09001390 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001391 pcpu_first_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001392 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001393 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001394 pcpu_last_unit_cpu = cpu;
1395 pcpu_nr_units = unit;
1396
1397 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001398 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1399
1400 /* we're done parsing the input, undefine BUG macro and dump config */
1401#undef PCPU_SETUP_BUG_ON
1402 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001403
Tejun Heo65632972009-08-14 15:00:52 +09001404 pcpu_nr_groups = ai->nr_groups;
1405 pcpu_group_offsets = group_offsets;
1406 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001407 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001408 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001409
1410 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001411 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001412 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001413 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001414 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1415 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001416
Tejun Heod9b55ee2009-02-24 11:57:21 +09001417 /*
1418 * Allocate chunk slots. The additional last slot is for
1419 * empty chunks.
1420 */
1421 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001422 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1423 for (i = 0; i < pcpu_nr_slots; i++)
1424 INIT_LIST_HEAD(&pcpu_slot[i]);
1425
Tejun Heoedcb4632009-03-06 14:33:59 +09001426 /*
1427 * Initialize static chunk. If reserved_size is zero, the
1428 * static chunk covers static area + dynamic allocation area
1429 * in the first chunk. If reserved_size is not zero, it
1430 * covers static area + reserved area (mostly used for module
1431 * static percpu allocation).
1432 */
Tejun Heo2441d152009-03-06 14:33:59 +09001433 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1434 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001435 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001436 schunk->map = smap;
1437 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001438 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001439 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001440
Tejun Heofd1e8a12009-08-14 15:00:51 +09001441 if (ai->reserved_size) {
1442 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001443 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001444 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001445 } else {
1446 schunk->free_size = dyn_size;
1447 dyn_size = 0; /* dynamic area covered */
1448 }
Tejun Heo2441d152009-03-06 14:33:59 +09001449 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001450
Tejun Heofd1e8a12009-08-14 15:00:51 +09001451 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001452 if (schunk->free_size)
1453 schunk->map[schunk->map_used++] = schunk->free_size;
1454
Tejun Heoedcb4632009-03-06 14:33:59 +09001455 /* init dynamic chunk if necessary */
1456 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001457 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001458 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001459 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001460 dchunk->map = dmap;
1461 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001462 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001463 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001464
1465 dchunk->contig_hint = dchunk->free_size = dyn_size;
1466 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1467 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1468 }
1469
Tejun Heo2441d152009-03-06 14:33:59 +09001470 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001471 pcpu_first_chunk = dchunk ?: schunk;
1472 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001473
1474 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001475 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001476 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001477}
Tejun Heo66c3a752009-03-10 16:27:48 +09001478
Tejun Heof58dc012009-08-14 15:00:50 +09001479const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1480 [PCPU_FC_AUTO] = "auto",
1481 [PCPU_FC_EMBED] = "embed",
1482 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001483};
Tejun Heo66c3a752009-03-10 16:27:48 +09001484
Tejun Heof58dc012009-08-14 15:00:50 +09001485enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1486
1487static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001488{
Tejun Heof58dc012009-08-14 15:00:50 +09001489 if (0)
1490 /* nada */;
1491#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1492 else if (!strcmp(str, "embed"))
1493 pcpu_chosen_fc = PCPU_FC_EMBED;
1494#endif
1495#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1496 else if (!strcmp(str, "page"))
1497 pcpu_chosen_fc = PCPU_FC_PAGE;
1498#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001499 else
1500 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001501
Tejun Heof58dc012009-08-14 15:00:50 +09001502 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001503}
Tejun Heof58dc012009-08-14 15:00:50 +09001504early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001505
Tejun Heo08fc4582009-08-14 15:00:49 +09001506#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1507 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo66c3a752009-03-10 16:27:48 +09001508/**
1509 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001510 * @reserved_size: the size of reserved percpu area in bytes
1511 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heoc8826dd2009-08-14 15:00:52 +09001512 * @atom_size: allocation atom size
1513 * @cpu_distance_fn: callback to determine distance between cpus, optional
1514 * @alloc_fn: function to allocate percpu page
1515 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001516 *
1517 * This is a helper to ease setting up embedded first percpu chunk and
1518 * can be called where pcpu_setup_first_chunk() is expected.
1519 *
1520 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001521 * by calling @alloc_fn and used as-is without being mapped into
1522 * vmalloc area. Allocations are always whole multiples of @atom_size
1523 * aligned to @atom_size.
1524 *
1525 * This enables the first chunk to piggy back on the linear physical
1526 * mapping which often uses larger page size. Please note that this
1527 * can result in very sparse cpu->unit mapping on NUMA machines thus
1528 * requiring large vmalloc address space. Don't use this allocator if
1529 * vmalloc space is not orders of magnitude larger than distances
1530 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001531 *
1532 * When @dyn_size is positive, dynamic area might be larger than
Tejun Heo788e5ab2009-07-04 08:10:58 +09001533 * specified to fill page alignment. When @dyn_size is auto,
1534 * @dyn_size is just big enough to fill page alignment after static
1535 * and reserved areas.
Tejun Heo66c3a752009-03-10 16:27:48 +09001536 *
1537 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001538 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001539 *
1540 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001541 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001542 */
Tejun Heoc8826dd2009-08-14 15:00:52 +09001543int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1544 size_t atom_size,
1545 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1546 pcpu_fc_alloc_fn_t alloc_fn,
1547 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001548{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001549 void *base = (void *)ULONG_MAX;
1550 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001551 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001552 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001553 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001554
Tejun Heoc8826dd2009-08-14 15:00:52 +09001555 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1556 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001557 if (IS_ERR(ai))
1558 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001559
Tejun Heofd1e8a12009-08-14 15:00:51 +09001560 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001561 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001562
Tejun Heoc8826dd2009-08-14 15:00:52 +09001563 areas = alloc_bootmem_nopanic(areas_size);
1564 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001565 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001566 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001567 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001568
Tejun Heoc8826dd2009-08-14 15:00:52 +09001569 /* allocate, copy and determine base address */
1570 for (group = 0; group < ai->nr_groups; group++) {
1571 struct pcpu_group_info *gi = &ai->groups[group];
1572 unsigned int cpu = NR_CPUS;
1573 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001574
Tejun Heoc8826dd2009-08-14 15:00:52 +09001575 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1576 cpu = gi->cpu_map[i];
1577 BUG_ON(cpu == NR_CPUS);
1578
1579 /* allocate space for the whole group */
1580 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1581 if (!ptr) {
1582 rc = -ENOMEM;
1583 goto out_free_areas;
1584 }
1585 areas[group] = ptr;
1586
1587 base = min(ptr, base);
1588
1589 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1590 if (gi->cpu_map[i] == NR_CPUS) {
1591 /* unused unit, free whole */
1592 free_fn(ptr, ai->unit_size);
1593 continue;
1594 }
1595 /* copy and return the unused part */
1596 memcpy(ptr, __per_cpu_load, ai->static_size);
1597 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1598 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001599 }
1600
Tejun Heoc8826dd2009-08-14 15:00:52 +09001601 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001602 max_distance = 0;
1603 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001604 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001605 max_distance = max_t(size_t, max_distance,
1606 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001607 }
1608 max_distance += ai->unit_size;
1609
1610 /* warn if maximum distance is further than 75% of vmalloc space */
1611 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001612 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001613 "space 0x%lx\n",
1614 max_distance, VMALLOC_END - VMALLOC_START);
1615#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1616 /* and fail if we have fallback */
1617 rc = -EINVAL;
1618 goto out_free;
1619#endif
1620 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001621
Tejun Heo004018e2009-08-14 15:00:49 +09001622 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001623 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1624 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001625
Tejun Heofb435d52009-08-14 15:00:51 +09001626 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001627 goto out_free;
1628
1629out_free_areas:
1630 for (group = 0; group < ai->nr_groups; group++)
1631 free_fn(areas[group],
1632 ai->groups[group].nr_units * ai->unit_size);
1633out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001634 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001635 if (areas)
1636 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001637 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001638}
Tejun Heo08fc4582009-08-14 15:00:49 +09001639#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1640 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heod4b95f82009-07-04 08:10:59 +09001641
Tejun Heo08fc4582009-08-14 15:00:49 +09001642#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001643/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001644 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001645 * @reserved_size: the size of reserved percpu area in bytes
1646 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1647 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1648 * @populate_pte_fn: function to populate pte
1649 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001650 * This is a helper to ease setting up page-remapped first percpu
1651 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001652 *
1653 * This is the basic allocator. Static percpu area is allocated
1654 * page-by-page into vmalloc area.
1655 *
1656 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001657 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001658 */
Tejun Heofb435d52009-08-14 15:00:51 +09001659int __init pcpu_page_first_chunk(size_t reserved_size,
1660 pcpu_fc_alloc_fn_t alloc_fn,
1661 pcpu_fc_free_fn_t free_fn,
1662 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001663{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001664 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001665 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001666 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001667 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001668 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001669 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001670 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001671
Tejun Heo00ae4062009-08-14 15:00:49 +09001672 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1673
Tejun Heofd1e8a12009-08-14 15:00:51 +09001674 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1675 if (IS_ERR(ai))
1676 return PTR_ERR(ai);
1677 BUG_ON(ai->nr_groups != 1);
1678 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1679
1680 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001681
1682 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001683 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1684 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09001685 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09001686
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001687 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001688 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001689 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001690 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001691 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001692 void *ptr;
1693
Tejun Heo3cbc8562009-08-14 15:00:50 +09001694 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001695 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001696 pr_warning("PERCPU: failed to allocate %s page "
1697 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001698 goto enomem;
1699 }
Tejun Heoce3141a2009-07-04 08:11:00 +09001700 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001701 }
1702
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001703 /* allocate vm area, map the pages and copy static data */
1704 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001705 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001706 vm_area_register_early(&vm, PAGE_SIZE);
1707
Tejun Heofd1e8a12009-08-14 15:00:51 +09001708 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001709 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001710 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001711
Tejun Heoce3141a2009-07-04 08:11:00 +09001712 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001713 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1714
1715 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001716 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1717 unit_pages);
1718 if (rc < 0)
1719 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001720
1721 /*
1722 * FIXME: Archs with virtual cache should flush local
1723 * cache for the linear mapping here - something
1724 * equivalent to flush_cache_vmap() on the local cpu.
1725 * flush_cache_vmap() can't be used as most supporting
1726 * data structures are not set up yet.
1727 */
1728
1729 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001730 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001731 }
1732
1733 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09001734 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001735 unit_pages, psize_str, vm.addr, ai->static_size,
1736 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001737
Tejun Heofb435d52009-08-14 15:00:51 +09001738 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001739 goto out_free_ar;
1740
1741enomem:
1742 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09001743 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09001744 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09001745out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09001746 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001747 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09001748 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001749}
Tejun Heo08fc4582009-08-14 15:00:49 +09001750#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001751
Tejun Heo8c4bfc62009-07-04 08:10:59 +09001752/*
Tejun Heoe74e3962009-03-30 19:07:44 +09001753 * Generic percpu area setup.
1754 *
1755 * The embedding helper is used because its behavior closely resembles
1756 * the original non-dynamic generic percpu area setup. This is
1757 * important because many archs have addressing restrictions and might
1758 * fail if the percpu area is located far away from the previous
1759 * location. As an added bonus, in non-NUMA cases, embedding is
1760 * generally a good idea TLB-wise because percpu area can piggy back
1761 * on the physical linear memory mapping which uses large page
1762 * mappings on applicable archs.
1763 */
1764#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1765unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1766EXPORT_SYMBOL(__per_cpu_offset);
1767
Tejun Heoc8826dd2009-08-14 15:00:52 +09001768static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1769 size_t align)
1770{
1771 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1772}
1773
1774static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1775{
1776 free_bootmem(__pa(ptr), size);
1777}
1778
Tejun Heoe74e3962009-03-30 19:07:44 +09001779void __init setup_per_cpu_areas(void)
1780{
Tejun Heoe74e3962009-03-30 19:07:44 +09001781 unsigned long delta;
1782 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09001783 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09001784
1785 /*
1786 * Always reserve area for module percpu variables. That's
1787 * what the legacy allocator did.
1788 */
Tejun Heofb435d52009-08-14 15:00:51 +09001789 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001790 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1791 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09001792 if (rc < 0)
Tejun Heoe74e3962009-03-30 19:07:44 +09001793 panic("Failed to initialized percpu areas.");
1794
1795 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1796 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09001797 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09001798}
1799#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */