blob: c52b93117dc20a0a28d545c633e90dc538a57f4c [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>
Catalin Marinasf528f0b2011-09-26 17:12:53 +010070#include <linux/kmemleak.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071
72#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090073#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090074#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090075#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090076
Tejun Heofbf59bc2009-02-20 16:29:08 +090077#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79
Tejun Heobbddff02010-09-03 18:22:48 +020080#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +090081/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82#ifndef __addr_to_pcpu_ptr
83#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090084 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090087#endif
88#ifndef __pcpu_ptr_to_addr
89#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090090 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090093#endif
Tejun Heobbddff02010-09-03 18:22:48 +020094#else /* CONFIG_SMP */
95/* on UP, it's always identity mapped */
96#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
97#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
98#endif /* CONFIG_SMP */
Tejun Heoe0100982009-03-10 16:27:48 +090099
Tejun Heofbf59bc2009-02-20 16:29:08 +0900100struct pcpu_chunk {
101 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +0900104 void *base_addr; /* base address of this chunk */
Al Viro723ad1d2014-03-06 21:13:18 -0500105 int map_used; /* # of map entries used before the sentry */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
Tejun Heo88999a82010-04-09 18:57:01 +0900108 void *data; /* chunk data */
Al Viro3d331ad2014-03-06 20:52:32 -0500109 int first_free; /* no free below this */
Tejun Heo8d408b42009-02-24 11:57:21 +0900110 bool immutable; /* no [de]population allowed */
Tejun Heoce3141a2009-07-04 08:11:00 +0900111 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900112};
113
Tejun Heo40150d32009-02-24 12:32:28 +0900114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900116static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900117static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900120
Tejun Heoa855b842011-11-18 10:55:35 -0800121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900124
Tejun Heofbf59bc2009-02-20 16:29:08 +0900125/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900126void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
Tejun Heofb435d52009-08-14 15:00:51 +0900129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900131
Tejun Heo65632972009-08-14 15:00:52 +0900132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static const size_t *pcpu_group_sizes __read_mostly;
136
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900137/*
138 * The first chunk which always exists. Note that unlike other
139 * chunks, this one can be allocated and mapped in several different
140 * ways and thus often doesn't live in the vmalloc area.
141 */
142static struct pcpu_chunk *pcpu_first_chunk;
143
144/*
145 * Optional reserved chunk. This chunk reserves part of the first
146 * chunk and serves it for reserved allocations. The amount of
147 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
148 * area doesn't exist, the following variables contain NULL and 0
149 * respectively.
150 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900151static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900152static int pcpu_reserved_chunk_limit;
153
Tejun Heob38d08f2014-09-02 14:46:02 -0400154static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
155static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900156
Tejun Heo40150d32009-02-24 12:32:28 +0900157static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900158
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900159/* reclaim work to release fully free chunks, scheduled from free path */
160static void pcpu_reclaim(struct work_struct *work);
161static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
162
Tejun Heo020ec652010-04-09 18:57:00 +0900163static bool pcpu_addr_in_first_chunk(void *addr)
164{
165 void *first_start = pcpu_first_chunk->base_addr;
166
167 return addr >= first_start && addr < first_start + pcpu_unit_size;
168}
169
170static bool pcpu_addr_in_reserved_chunk(void *addr)
171{
172 void *first_start = pcpu_first_chunk->base_addr;
173
174 return addr >= first_start &&
175 addr < first_start + pcpu_reserved_chunk_limit;
176}
177
Tejun Heod9b55ee2009-02-24 11:57:21 +0900178static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900179{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900180 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900181 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
182}
183
Tejun Heod9b55ee2009-02-24 11:57:21 +0900184static int pcpu_size_to_slot(int size)
185{
186 if (size == pcpu_unit_size)
187 return pcpu_nr_slots - 1;
188 return __pcpu_size_to_slot(size);
189}
190
Tejun Heofbf59bc2009-02-20 16:29:08 +0900191static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
192{
193 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
194 return 0;
195
196 return pcpu_size_to_slot(chunk->free_size);
197}
198
Tejun Heo88999a82010-04-09 18:57:01 +0900199/* set the pointer to a chunk in a page struct */
200static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
201{
202 page->index = (unsigned long)pcpu;
203}
204
205/* obtain pointer to a chunk from a page struct */
206static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
207{
208 return (struct pcpu_chunk *)page->index;
209}
210
211static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900212{
Tejun Heo2f39e632009-07-04 08:11:00 +0900213 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900214}
215
Tejun Heo9983b6f02010-06-18 11:44:31 +0200216static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
217 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900218{
Tejun Heobba174f2009-08-14 15:00:51 +0900219 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900220 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900221}
222
Tejun Heo88999a82010-04-09 18:57:01 +0900223static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
224 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900225{
226 *rs = find_next_zero_bit(chunk->populated, end, *rs);
227 *re = find_next_bit(chunk->populated, end, *rs + 1);
228}
229
Tejun Heo88999a82010-04-09 18:57:01 +0900230static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
231 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900232{
233 *rs = find_next_bit(chunk->populated, end, *rs);
234 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
235}
236
237/*
238 * (Un)populated page region iterators. Iterate over (un)populated
Uwe Kleine-Königb5950762010-11-01 15:38:34 -0400239 * page regions between @start and @end in @chunk. @rs and @re should
Tejun Heoce3141a2009-07-04 08:11:00 +0900240 * be integer variables and will be set to start and end page index of
241 * the current region.
242 */
243#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
244 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
245 (rs) < (re); \
246 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
247
248#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
249 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
250 (rs) < (re); \
251 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
252
Tejun Heofbf59bc2009-02-20 16:29:08 +0900253/**
Bob Liu90459ce02011-08-04 11:02:33 +0200254 * pcpu_mem_zalloc - allocate memory
Tejun Heo1880d932009-03-07 00:44:09 +0900255 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900256 *
Tejun Heo1880d932009-03-07 00:44:09 +0900257 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
Bob Liu90459ce02011-08-04 11:02:33 +0200258 * kzalloc() is used; otherwise, vzalloc() is used. The returned
Tejun Heo1880d932009-03-07 00:44:09 +0900259 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900260 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900261 * CONTEXT:
262 * Does GFP_KERNEL allocation.
263 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900264 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900265 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900266 */
Bob Liu90459ce02011-08-04 11:02:33 +0200267static void *pcpu_mem_zalloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900268{
Tejun Heo099a19d2010-06-27 18:50:00 +0200269 if (WARN_ON_ONCE(!slab_is_available()))
270 return NULL;
271
Tejun Heofbf59bc2009-02-20 16:29:08 +0900272 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900273 return kzalloc(size, GFP_KERNEL);
Jesper Juhl7af4c092010-10-30 15:56:54 +0200274 else
275 return vzalloc(size);
Tejun Heo1880d932009-03-07 00:44:09 +0900276}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900277
Tejun Heo1880d932009-03-07 00:44:09 +0900278/**
279 * pcpu_mem_free - free memory
280 * @ptr: memory to free
281 * @size: size of the area
282 *
Bob Liu90459ce02011-08-04 11:02:33 +0200283 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
Tejun Heo1880d932009-03-07 00:44:09 +0900284 */
285static void pcpu_mem_free(void *ptr, size_t size)
286{
287 if (size <= PAGE_SIZE)
288 kfree(ptr);
289 else
290 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900291}
292
293/**
294 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
295 * @chunk: chunk of interest
296 * @oslot: the previous slot it was on
297 *
298 * This function is called after an allocation or free changed @chunk.
299 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900300 * moved to the slot. Note that the reserved chunk is never put on
301 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900302 *
303 * CONTEXT:
304 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900305 */
306static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
307{
308 int nslot = pcpu_chunk_slot(chunk);
309
Tejun Heoedcb4632009-03-06 14:33:59 +0900310 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900311 if (oslot < nslot)
312 list_move(&chunk->list, &pcpu_slot[nslot]);
313 else
314 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
315 }
316}
317
Tejun Heofbf59bc2009-02-20 16:29:08 +0900318/**
Tejun Heo833af842009-11-11 15:35:18 +0900319 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
320 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900321 *
Tejun Heo833af842009-11-11 15:35:18 +0900322 * Determine whether area map of @chunk needs to be extended to
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300323 * accommodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900324 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900325 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900326 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900327 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900328 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900329 * New target map allocation length if extension is necessary, 0
330 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900331 */
Tejun Heo833af842009-11-11 15:35:18 +0900332static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900333{
334 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900335
Al Viro723ad1d2014-03-06 21:13:18 -0500336 if (chunk->map_alloc >= chunk->map_used + 3)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900337 return 0;
338
339 new_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500340 while (new_alloc < chunk->map_used + 3)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900341 new_alloc *= 2;
342
Tejun Heo833af842009-11-11 15:35:18 +0900343 return new_alloc;
344}
345
346/**
347 * pcpu_extend_area_map - extend area map of a chunk
348 * @chunk: chunk of interest
349 * @new_alloc: new target allocation length of the area map
350 *
351 * Extend area map of @chunk to have @new_alloc entries.
352 *
353 * CONTEXT:
354 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
355 *
356 * RETURNS:
357 * 0 on success, -errno on failure.
358 */
359static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
360{
361 int *old = NULL, *new = NULL;
362 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
363 unsigned long flags;
364
Bob Liu90459ce02011-08-04 11:02:33 +0200365 new = pcpu_mem_zalloc(new_size);
Tejun Heo833af842009-11-11 15:35:18 +0900366 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900367 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900368
Tejun Heo833af842009-11-11 15:35:18 +0900369 /* acquire pcpu_lock and switch to new area map */
370 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900371
Tejun Heo833af842009-11-11 15:35:18 +0900372 if (new_alloc <= chunk->map_alloc)
373 goto out_unlock;
374
375 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
Huang Shijiea002d142010-08-08 14:39:07 +0200376 old = chunk->map;
377
378 memcpy(new, old, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900379
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900380 chunk->map_alloc = new_alloc;
381 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900382 new = NULL;
383
384out_unlock:
385 spin_unlock_irqrestore(&pcpu_lock, flags);
386
387 /*
388 * pcpu_mem_free() might end up calling vfree() which uses
389 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
390 */
391 pcpu_mem_free(old, old_size);
392 pcpu_mem_free(new, new_size);
393
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900394 return 0;
395}
396
397/**
Tejun Heoa16037c2014-09-02 14:46:02 -0400398 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
399 * @chunk: chunk the candidate area belongs to
400 * @off: the offset to the start of the candidate area
401 * @this_size: the size of the candidate area
402 * @size: the size of the target allocation
403 * @align: the alignment of the target allocation
404 * @pop_only: only allocate from already populated region
405 *
406 * We're trying to allocate @size bytes aligned at @align. @chunk's area
407 * at @off sized @this_size is a candidate. This function determines
408 * whether the target allocation fits in the candidate area and returns the
409 * number of bytes to pad after @off. If the target area doesn't fit, -1
410 * is returned.
411 *
412 * If @pop_only is %true, this function only considers the already
413 * populated part of the candidate area.
414 */
415static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
416 int size, int align, bool pop_only)
417{
418 int cand_off = off;
419
420 while (true) {
421 int head = ALIGN(cand_off, align) - off;
422 int page_start, page_end, rs, re;
423
424 if (this_size < head + size)
425 return -1;
426
427 if (!pop_only)
428 return head;
429
430 /*
431 * If the first unpopulated page is beyond the end of the
432 * allocation, the whole allocation is populated;
433 * otherwise, retry from the end of the unpopulated area.
434 */
435 page_start = PFN_DOWN(head + off);
436 page_end = PFN_UP(head + off + size);
437
438 rs = page_start;
439 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
440 if (rs >= page_end)
441 return head;
442 cand_off = re * PAGE_SIZE;
443 }
444}
445
446/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900447 * pcpu_alloc_area - allocate area from a pcpu_chunk
448 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900449 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900450 * @align: wanted align
Tejun Heoa16037c2014-09-02 14:46:02 -0400451 * @pop_only: allocate only from the populated area
Tejun Heofbf59bc2009-02-20 16:29:08 +0900452 *
453 * Try to allocate @size bytes area aligned at @align from @chunk.
454 * Note that this function only allocates the offset. It doesn't
455 * populate or map the area.
456 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900457 * @chunk->map must have at least two free slots.
458 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900459 * CONTEXT:
460 * pcpu_lock.
461 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900462 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900463 * Allocated offset in @chunk on success, -1 if no matching area is
464 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900465 */
Tejun Heoa16037c2014-09-02 14:46:02 -0400466static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
467 bool pop_only)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900468{
469 int oslot = pcpu_chunk_slot(chunk);
470 int max_contig = 0;
471 int i, off;
Al Viro3d331ad2014-03-06 20:52:32 -0500472 bool seen_free = false;
Al Viro723ad1d2014-03-06 21:13:18 -0500473 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900474
Al Viro3d331ad2014-03-06 20:52:32 -0500475 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900476 int head, tail;
Al Viro723ad1d2014-03-06 21:13:18 -0500477 int this_size;
478
479 off = *p;
480 if (off & 1)
481 continue;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900482
Al Viro723ad1d2014-03-06 21:13:18 -0500483 this_size = (p[1] & ~1) - off;
Tejun Heoa16037c2014-09-02 14:46:02 -0400484
485 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
486 pop_only);
487 if (head < 0) {
Al Viro3d331ad2014-03-06 20:52:32 -0500488 if (!seen_free) {
489 chunk->first_free = i;
490 seen_free = true;
491 }
Al Viro723ad1d2014-03-06 21:13:18 -0500492 max_contig = max(this_size, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900493 continue;
494 }
495
496 /*
497 * If head is small or the previous block is free,
498 * merge'em. Note that 'small' is defined as smaller
499 * than sizeof(int), which is very small but isn't too
500 * uncommon for percpu allocations.
501 */
Al Viro723ad1d2014-03-06 21:13:18 -0500502 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800503 *p = off += head;
Al Viro723ad1d2014-03-06 21:13:18 -0500504 if (p[-1] & 1)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900505 chunk->free_size -= head;
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800506 else
507 max_contig = max(*p - p[-1], max_contig);
Al Viro723ad1d2014-03-06 21:13:18 -0500508 this_size -= head;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900509 head = 0;
510 }
511
512 /* if tail is small, just keep it around */
Al Viro723ad1d2014-03-06 21:13:18 -0500513 tail = this_size - head - size;
514 if (tail < sizeof(int)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900515 tail = 0;
Al Viro723ad1d2014-03-06 21:13:18 -0500516 size = this_size - head;
517 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900518
519 /* split if warranted */
520 if (head || tail) {
Al Viro706c16f2014-03-06 21:08:24 -0500521 int nr_extra = !!head + !!tail;
522
523 /* insert new subblocks */
Al Viro723ad1d2014-03-06 21:13:18 -0500524 memmove(p + nr_extra + 1, p + 1,
Al Viro706c16f2014-03-06 21:08:24 -0500525 sizeof(chunk->map[0]) * (chunk->map_used - i));
526 chunk->map_used += nr_extra;
527
Tejun Heofbf59bc2009-02-20 16:29:08 +0900528 if (head) {
Al Viro3d331ad2014-03-06 20:52:32 -0500529 if (!seen_free) {
530 chunk->first_free = i;
531 seen_free = true;
532 }
Al Viro723ad1d2014-03-06 21:13:18 -0500533 *++p = off += head;
534 ++i;
Al Viro706c16f2014-03-06 21:08:24 -0500535 max_contig = max(head, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900536 }
Al Viro706c16f2014-03-06 21:08:24 -0500537 if (tail) {
Al Viro723ad1d2014-03-06 21:13:18 -0500538 p[1] = off + size;
Al Viro706c16f2014-03-06 21:08:24 -0500539 max_contig = max(tail, max_contig);
540 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900541 }
542
Al Viro3d331ad2014-03-06 20:52:32 -0500543 if (!seen_free)
544 chunk->first_free = i + 1;
545
Tejun Heofbf59bc2009-02-20 16:29:08 +0900546 /* update hint and mark allocated */
Al Viro723ad1d2014-03-06 21:13:18 -0500547 if (i + 1 == chunk->map_used)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900548 chunk->contig_hint = max_contig; /* fully scanned */
549 else
550 chunk->contig_hint = max(chunk->contig_hint,
551 max_contig);
552
Al Viro723ad1d2014-03-06 21:13:18 -0500553 chunk->free_size -= size;
554 *p |= 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900555
556 pcpu_chunk_relocate(chunk, oslot);
557 return off;
558 }
559
560 chunk->contig_hint = max_contig; /* fully scanned */
561 pcpu_chunk_relocate(chunk, oslot);
562
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900563 /* tell the upper layer that this chunk has no matching area */
564 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900565}
566
567/**
568 * pcpu_free_area - free area to a pcpu_chunk
569 * @chunk: chunk of interest
570 * @freeme: offset of area to free
571 *
572 * Free area starting from @freeme to @chunk. Note that this function
573 * only modifies the allocation map. It doesn't depopulate or unmap
574 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900575 *
576 * CONTEXT:
577 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900578 */
579static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
580{
581 int oslot = pcpu_chunk_slot(chunk);
Al Viro723ad1d2014-03-06 21:13:18 -0500582 int off = 0;
583 unsigned i, j;
584 int to_free = 0;
585 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900586
Al Viro723ad1d2014-03-06 21:13:18 -0500587 freeme |= 1; /* we are searching for <given offset, in use> pair */
588
589 i = 0;
590 j = chunk->map_used;
591 while (i != j) {
592 unsigned k = (i + j) / 2;
593 off = chunk->map[k];
594 if (off < freeme)
595 i = k + 1;
596 else if (off > freeme)
597 j = k;
598 else
599 i = j = k;
600 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900601 BUG_ON(off != freeme);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900602
Al Viro3d331ad2014-03-06 20:52:32 -0500603 if (i < chunk->first_free)
604 chunk->first_free = i;
605
Al Viro723ad1d2014-03-06 21:13:18 -0500606 p = chunk->map + i;
607 *p = off &= ~1;
608 chunk->free_size += (p[1] & ~1) - off;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900609
Tejun Heofbf59bc2009-02-20 16:29:08 +0900610 /* merge with next? */
Al Viro723ad1d2014-03-06 21:13:18 -0500611 if (!(p[1] & 1))
612 to_free++;
613 /* merge with previous? */
614 if (i > 0 && !(p[-1] & 1)) {
615 to_free++;
616 i--;
617 p--;
618 }
619 if (to_free) {
620 chunk->map_used -= to_free;
621 memmove(p + 1, p + 1 + to_free,
622 (chunk->map_used - i) * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900623 }
624
Al Viro723ad1d2014-03-06 21:13:18 -0500625 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900626 pcpu_chunk_relocate(chunk, oslot);
627}
628
Tejun Heo60810892010-04-09 18:57:01 +0900629static struct pcpu_chunk *pcpu_alloc_chunk(void)
630{
631 struct pcpu_chunk *chunk;
632
Bob Liu90459ce02011-08-04 11:02:33 +0200633 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900634 if (!chunk)
635 return NULL;
636
Bob Liu90459ce02011-08-04 11:02:33 +0200637 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
638 sizeof(chunk->map[0]));
Tejun Heo60810892010-04-09 18:57:01 +0900639 if (!chunk->map) {
Jianyu Zhan5a838c32014-04-14 13:47:40 +0800640 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900641 return NULL;
642 }
643
644 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500645 chunk->map[0] = 0;
646 chunk->map[1] = pcpu_unit_size | 1;
647 chunk->map_used = 1;
Tejun Heo60810892010-04-09 18:57:01 +0900648
649 INIT_LIST_HEAD(&chunk->list);
650 chunk->free_size = pcpu_unit_size;
651 chunk->contig_hint = pcpu_unit_size;
652
653 return chunk;
654}
655
656static void pcpu_free_chunk(struct pcpu_chunk *chunk)
657{
658 if (!chunk)
659 return;
660 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Joonsoo Kimb4916cb2012-10-29 22:59:58 +0900661 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900662}
663
Tejun Heo9f645532010-04-09 18:57:01 +0900664/*
665 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900666 *
Tejun Heo9f645532010-04-09 18:57:01 +0900667 * To allow different implementations, chunk alloc/free and
668 * [de]population are implemented in a separate file which is pulled
669 * into this file and compiled together. The following functions
670 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900671 *
Tejun Heo9f645532010-04-09 18:57:01 +0900672 * pcpu_populate_chunk - populate the specified range of a chunk
673 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
674 * pcpu_create_chunk - create a new chunk
675 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
676 * pcpu_addr_to_page - translate address to physical address
677 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900678 */
Tejun Heo9f645532010-04-09 18:57:01 +0900679static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
680static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
681static struct pcpu_chunk *pcpu_create_chunk(void);
682static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
683static struct page *pcpu_addr_to_page(void *addr);
684static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900685
Tejun Heob0c97782010-04-09 18:57:01 +0900686#ifdef CONFIG_NEED_PER_CPU_KM
687#include "percpu-km.c"
688#else
Tejun Heo9f645532010-04-09 18:57:01 +0900689#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900690#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900691
692/**
Tejun Heo88999a82010-04-09 18:57:01 +0900693 * pcpu_chunk_addr_search - determine chunk containing specified address
694 * @addr: address for which the chunk needs to be determined.
695 *
696 * RETURNS:
697 * The address of the found chunk.
698 */
699static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
700{
701 /* is it in the first chunk? */
702 if (pcpu_addr_in_first_chunk(addr)) {
703 /* is it in the reserved area? */
704 if (pcpu_addr_in_reserved_chunk(addr))
705 return pcpu_reserved_chunk;
706 return pcpu_first_chunk;
707 }
708
709 /*
710 * The address is relative to unit0 which might be unused and
711 * thus unmapped. Offset the address to the unit space of the
712 * current processor before looking it up in the vmalloc
713 * space. Note that any possible cpu id can be used here, so
714 * there's no need to worry about preemption or cpu hotplug.
715 */
716 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900717 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900718}
719
720/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900721 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900722 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900723 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900724 * @reserved: allocate from the reserved chunk if available
Tejun Heo5835d962014-09-02 14:46:04 -0400725 * @gfp: allocation flags
Tejun Heofbf59bc2009-02-20 16:29:08 +0900726 *
Tejun Heo5835d962014-09-02 14:46:04 -0400727 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
728 * contain %GFP_KERNEL, the allocation is atomic.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900729 *
730 * RETURNS:
731 * Percpu pointer to the allocated area on success, NULL on failure.
732 */
Tejun Heo5835d962014-09-02 14:46:04 -0400733static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
734 gfp_t gfp)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900735{
Tejun Heof2badb02009-09-29 09:17:58 +0900736 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900737 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900738 const char *err;
Tejun Heo5835d962014-09-02 14:46:04 -0400739 bool is_atomic = !(gfp & GFP_KERNEL);
Tejun Heob38d08f2014-09-02 14:46:02 -0400740 int slot, off, new_alloc, cpu, ret;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900741 unsigned long flags;
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100742 void __percpu *ptr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900743
Al Viro723ad1d2014-03-06 21:13:18 -0500744 /*
745 * We want the lowest bit of offset available for in-use/free
Viro2f69fa82014-03-17 16:01:27 -0400746 * indicator, so force >= 16bit alignment and make size even.
Al Viro723ad1d2014-03-06 21:13:18 -0500747 */
748 if (unlikely(align < 2))
749 align = 2;
750
Christoph Lameterfb009e32014-06-19 09:59:18 -0500751 size = ALIGN(size, 2);
Viro2f69fa82014-03-17 16:01:27 -0400752
Tejun Heo8d408b42009-02-24 11:57:21 +0900753 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900754 WARN(true, "illegal size (%zu) or align (%zu) for "
755 "percpu allocation\n", size, align);
756 return NULL;
757 }
758
Jiri Kosina403a91b2009-10-29 00:25:59 +0900759 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900760
Tejun Heoedcb4632009-03-06 14:33:59 +0900761 /* serve reserved allocations from the reserved chunk if available */
762 if (reserved && pcpu_reserved_chunk) {
763 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900764
765 if (size > chunk->contig_hint) {
766 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900767 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900768 }
Tejun Heo833af842009-11-11 15:35:18 +0900769
770 while ((new_alloc = pcpu_need_to_extend(chunk))) {
771 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heo5835d962014-09-02 14:46:04 -0400772 if (is_atomic ||
773 pcpu_extend_area_map(chunk, new_alloc) < 0) {
Tejun Heo833af842009-11-11 15:35:18 +0900774 err = "failed to extend area map of reserved chunk";
Tejun Heob38d08f2014-09-02 14:46:02 -0400775 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900776 }
777 spin_lock_irqsave(&pcpu_lock, flags);
778 }
779
Tejun Heo5835d962014-09-02 14:46:04 -0400780 off = pcpu_alloc_area(chunk, size, align, is_atomic);
Tejun Heoedcb4632009-03-06 14:33:59 +0900781 if (off >= 0)
782 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900783
Tejun Heof2badb02009-09-29 09:17:58 +0900784 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900785 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900786 }
787
Tejun Heoccea34b2009-03-07 00:44:13 +0900788restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900789 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900790 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
791 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
792 if (size > chunk->contig_hint)
793 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900794
Tejun Heo833af842009-11-11 15:35:18 +0900795 new_alloc = pcpu_need_to_extend(chunk);
796 if (new_alloc) {
Tejun Heo5835d962014-09-02 14:46:04 -0400797 if (is_atomic)
798 continue;
Tejun Heo833af842009-11-11 15:35:18 +0900799 spin_unlock_irqrestore(&pcpu_lock, flags);
800 if (pcpu_extend_area_map(chunk,
801 new_alloc) < 0) {
802 err = "failed to extend area map";
Tejun Heob38d08f2014-09-02 14:46:02 -0400803 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900804 }
805 spin_lock_irqsave(&pcpu_lock, flags);
806 /*
807 * pcpu_lock has been dropped, need to
808 * restart cpu_slot list walking.
809 */
810 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900811 }
812
Tejun Heo5835d962014-09-02 14:46:04 -0400813 off = pcpu_alloc_area(chunk, size, align, is_atomic);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900814 if (off >= 0)
815 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900816 }
817 }
818
Jiri Kosina403a91b2009-10-29 00:25:59 +0900819 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900820
Tejun Heob38d08f2014-09-02 14:46:02 -0400821 /*
822 * No space left. Create a new chunk. We don't want multiple
823 * tasks to create chunks simultaneously. Serialize and create iff
824 * there's still no empty chunk after grabbing the mutex.
825 */
Tejun Heo5835d962014-09-02 14:46:04 -0400826 if (is_atomic)
827 goto fail;
828
Tejun Heob38d08f2014-09-02 14:46:02 -0400829 mutex_lock(&pcpu_alloc_mutex);
830
831 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
832 chunk = pcpu_create_chunk();
833 if (!chunk) {
834 err = "failed to allocate new chunk";
835 goto fail;
836 }
837
838 spin_lock_irqsave(&pcpu_lock, flags);
839 pcpu_chunk_relocate(chunk, -1);
840 } else {
841 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heof2badb02009-09-29 09:17:58 +0900842 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900843
Tejun Heob38d08f2014-09-02 14:46:02 -0400844 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heoccea34b2009-03-07 00:44:13 +0900845 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900846
847area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900848 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900849
Tejun Heodca49642014-09-02 14:46:01 -0400850 /* populate if not all pages are already there */
Tejun Heo5835d962014-09-02 14:46:04 -0400851 if (!is_atomic) {
Tejun Heoe04d3202014-09-02 14:46:04 -0400852 int page_start, page_end, rs, re;
Tejun Heodca49642014-09-02 14:46:01 -0400853
Tejun Heoe04d3202014-09-02 14:46:04 -0400854 mutex_lock(&pcpu_alloc_mutex);
Tejun Heodca49642014-09-02 14:46:01 -0400855
Tejun Heoe04d3202014-09-02 14:46:04 -0400856 page_start = PFN_DOWN(off);
857 page_end = PFN_UP(off + size);
Tejun Heob38d08f2014-09-02 14:46:02 -0400858
Tejun Heoe04d3202014-09-02 14:46:04 -0400859 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
860 WARN_ON(chunk->immutable);
861
862 ret = pcpu_populate_chunk(chunk, rs, re);
863
864 spin_lock_irqsave(&pcpu_lock, flags);
865 if (ret) {
866 mutex_unlock(&pcpu_alloc_mutex);
867 pcpu_free_area(chunk, off);
868 err = "failed to populate";
869 goto fail_unlock;
870 }
871 bitmap_set(chunk->populated, rs, re - rs);
872 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heodca49642014-09-02 14:46:01 -0400873 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900874
Tejun Heoe04d3202014-09-02 14:46:04 -0400875 mutex_unlock(&pcpu_alloc_mutex);
876 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900877
Tejun Heodca49642014-09-02 14:46:01 -0400878 /* clear the areas and return address relative to base address */
879 for_each_possible_cpu(cpu)
880 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
881
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100882 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
883 kmemleak_alloc_percpu(ptr, size);
884 return ptr;
Tejun Heoccea34b2009-03-07 00:44:13 +0900885
886fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900887 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heob38d08f2014-09-02 14:46:02 -0400888fail:
Tejun Heo5835d962014-09-02 14:46:04 -0400889 if (!is_atomic && warn_limit) {
890 pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n",
891 size, align, is_atomic, err);
Tejun Heof2badb02009-09-29 09:17:58 +0900892 dump_stack();
893 if (!--warn_limit)
894 pr_info("PERCPU: limit reached, disable warning\n");
895 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900896 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900897}
Tejun Heoedcb4632009-03-06 14:33:59 +0900898
899/**
Tejun Heo5835d962014-09-02 14:46:04 -0400900 * __alloc_percpu_gfp - allocate dynamic percpu area
Tejun Heoedcb4632009-03-06 14:33:59 +0900901 * @size: size of area to allocate in bytes
902 * @align: alignment of area (max PAGE_SIZE)
Tejun Heo5835d962014-09-02 14:46:04 -0400903 * @gfp: allocation flags
Tejun Heoedcb4632009-03-06 14:33:59 +0900904 *
Tejun Heo5835d962014-09-02 14:46:04 -0400905 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
906 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
907 * be called from any context but is a lot more likely to fail.
Tejun Heoccea34b2009-03-07 00:44:13 +0900908 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900909 * RETURNS:
910 * Percpu pointer to the allocated area on success, NULL on failure.
911 */
Tejun Heo5835d962014-09-02 14:46:04 -0400912void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
913{
914 return pcpu_alloc(size, align, false, gfp);
915}
916EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
917
918/**
919 * __alloc_percpu - allocate dynamic percpu area
920 * @size: size of area to allocate in bytes
921 * @align: alignment of area (max PAGE_SIZE)
922 *
923 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
924 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900925void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900926{
Tejun Heo5835d962014-09-02 14:46:04 -0400927 return pcpu_alloc(size, align, false, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +0900928}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900929EXPORT_SYMBOL_GPL(__alloc_percpu);
930
Tejun Heoedcb4632009-03-06 14:33:59 +0900931/**
932 * __alloc_reserved_percpu - allocate reserved percpu area
933 * @size: size of area to allocate in bytes
934 * @align: alignment of area (max PAGE_SIZE)
935 *
Tejun Heo9329ba92010-09-10 11:01:56 +0200936 * Allocate zero-filled percpu area of @size bytes aligned at @align
937 * from reserved percpu area if arch has set it up; otherwise,
938 * allocation is served from the same dynamic area. Might sleep.
939 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +0900940 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900941 * CONTEXT:
942 * Does GFP_KERNEL allocation.
943 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900944 * RETURNS:
945 * Percpu pointer to the allocated area on success, NULL on failure.
946 */
Tejun Heo43cf38e2010-02-02 14:38:57 +0900947void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +0900948{
Tejun Heo5835d962014-09-02 14:46:04 -0400949 return pcpu_alloc(size, align, true, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +0900950}
951
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900952/**
953 * pcpu_reclaim - reclaim fully free chunks, workqueue function
954 * @work: unused
955 *
956 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900957 *
958 * CONTEXT:
959 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900960 */
961static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900962{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900963 LIST_HEAD(todo);
964 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
965 struct pcpu_chunk *chunk, *next;
966
Tejun Heoccea34b2009-03-07 00:44:13 +0900967 mutex_lock(&pcpu_alloc_mutex);
968 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900969
970 list_for_each_entry_safe(chunk, next, head, list) {
971 WARN_ON(chunk->immutable);
972
973 /* spare the first one */
974 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
975 continue;
976
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900977 list_move(&chunk->list, &todo);
978 }
979
Tejun Heoccea34b2009-03-07 00:44:13 +0900980 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900981
982 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoa93ace42014-09-02 14:46:02 -0400983 int rs, re;
Tejun Heodca49642014-09-02 14:46:01 -0400984
Tejun Heoa93ace42014-09-02 14:46:02 -0400985 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
986 pcpu_depopulate_chunk(chunk, rs, re);
987 bitmap_clear(chunk->populated, rs, re - rs);
988 }
Tejun Heo60810892010-04-09 18:57:01 +0900989 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900990 }
Tejun Heo971f3912009-08-14 15:00:49 +0900991
992 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900993}
994
995/**
996 * free_percpu - free percpu area
997 * @ptr: pointer to area to free
998 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900999 * Free percpu area @ptr.
1000 *
1001 * CONTEXT:
1002 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001003 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001004void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001005{
Andrew Morton129182e2010-01-08 14:42:39 -08001006 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001007 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001008 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001009 int off;
1010
1011 if (!ptr)
1012 return;
1013
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001014 kmemleak_free_percpu(ptr);
1015
Andrew Morton129182e2010-01-08 14:42:39 -08001016 addr = __pcpu_ptr_to_addr(ptr);
1017
Tejun Heoccea34b2009-03-07 00:44:13 +09001018 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001019
1020 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001021 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001022
1023 pcpu_free_area(chunk, off);
1024
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001025 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001026 if (chunk->free_size == pcpu_unit_size) {
1027 struct pcpu_chunk *pos;
1028
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001029 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001030 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001031 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001032 break;
1033 }
1034 }
1035
Tejun Heoccea34b2009-03-07 00:44:13 +09001036 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001037}
1038EXPORT_SYMBOL_GPL(free_percpu);
1039
Vivek Goyal3b034b02009-11-24 15:50:03 +09001040/**
Tejun Heo10fad5e2010-03-10 18:57:54 +09001041 * is_kernel_percpu_address - test whether address is from static percpu area
1042 * @addr: address to test
1043 *
1044 * Test whether @addr belongs to in-kernel static percpu area. Module
1045 * static percpu areas are not considered. For those, use
1046 * is_module_percpu_address().
1047 *
1048 * RETURNS:
1049 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1050 */
1051bool is_kernel_percpu_address(unsigned long addr)
1052{
Tejun Heobbddff02010-09-03 18:22:48 +02001053#ifdef CONFIG_SMP
Tejun Heo10fad5e2010-03-10 18:57:54 +09001054 const size_t static_size = __per_cpu_end - __per_cpu_start;
1055 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1056 unsigned int cpu;
1057
1058 for_each_possible_cpu(cpu) {
1059 void *start = per_cpu_ptr(base, cpu);
1060
1061 if ((void *)addr >= start && (void *)addr < start + static_size)
1062 return true;
1063 }
Tejun Heobbddff02010-09-03 18:22:48 +02001064#endif
1065 /* on UP, can't distinguish from other static vars, always false */
Tejun Heo10fad5e2010-03-10 18:57:54 +09001066 return false;
1067}
1068
1069/**
Vivek Goyal3b034b02009-11-24 15:50:03 +09001070 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1071 * @addr: the address to be converted to physical address
1072 *
1073 * Given @addr which is dereferenceable address obtained via one of
1074 * percpu access macros, this function translates it into its physical
1075 * address. The caller is responsible for ensuring @addr stays valid
1076 * until this function finishes.
1077 *
Dave Young67589c712011-11-23 08:20:53 -08001078 * percpu allocator has special setup for the first chunk, which currently
1079 * supports either embedding in linear address space or vmalloc mapping,
1080 * and, from the second one, the backing allocator (currently either vm or
1081 * km) provides translation.
1082 *
1083 * The addr can be tranlated simply without checking if it falls into the
1084 * first chunk. But the current code reflects better how percpu allocator
1085 * actually works, and the verification can discover both bugs in percpu
1086 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1087 * code.
1088 *
Vivek Goyal3b034b02009-11-24 15:50:03 +09001089 * RETURNS:
1090 * The physical address for @addr.
1091 */
1092phys_addr_t per_cpu_ptr_to_phys(void *addr)
1093{
Tejun Heo9983b6f02010-06-18 11:44:31 +02001094 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1095 bool in_first_chunk = false;
Tejun Heoa855b842011-11-18 10:55:35 -08001096 unsigned long first_low, first_high;
Tejun Heo9983b6f02010-06-18 11:44:31 +02001097 unsigned int cpu;
1098
1099 /*
Tejun Heoa855b842011-11-18 10:55:35 -08001100 * The following test on unit_low/high isn't strictly
Tejun Heo9983b6f02010-06-18 11:44:31 +02001101 * necessary but will speed up lookups of addresses which
1102 * aren't in the first chunk.
1103 */
Tejun Heoa855b842011-11-18 10:55:35 -08001104 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1105 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1106 pcpu_unit_pages);
1107 if ((unsigned long)addr >= first_low &&
1108 (unsigned long)addr < first_high) {
Tejun Heo9983b6f02010-06-18 11:44:31 +02001109 for_each_possible_cpu(cpu) {
1110 void *start = per_cpu_ptr(base, cpu);
1111
1112 if (addr >= start && addr < start + pcpu_unit_size) {
1113 in_first_chunk = true;
1114 break;
1115 }
1116 }
1117 }
1118
1119 if (in_first_chunk) {
David Howellseac522e2011-03-28 12:53:29 +01001120 if (!is_vmalloc_addr(addr))
Tejun Heo020ec652010-04-09 18:57:00 +09001121 return __pa(addr);
1122 else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001123 return page_to_phys(vmalloc_to_page(addr)) +
1124 offset_in_page(addr);
Tejun Heo020ec652010-04-09 18:57:00 +09001125 } else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001126 return page_to_phys(pcpu_addr_to_page(addr)) +
1127 offset_in_page(addr);
Vivek Goyal3b034b02009-11-24 15:50:03 +09001128}
1129
Tejun Heofbf59bc2009-02-20 16:29:08 +09001130/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001131 * pcpu_alloc_alloc_info - allocate percpu allocation info
1132 * @nr_groups: the number of groups
1133 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001134 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001135 * Allocate ai which is large enough for @nr_groups groups containing
1136 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1137 * cpu_map array which is long enough for @nr_units and filled with
1138 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1139 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001140 *
1141 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001142 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1143 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001144 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001145struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1146 int nr_units)
1147{
1148 struct pcpu_alloc_info *ai;
1149 size_t base_size, ai_size;
1150 void *ptr;
1151 int unit;
1152
1153 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1154 __alignof__(ai->groups[0].cpu_map[0]));
1155 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1156
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001157 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001158 if (!ptr)
1159 return NULL;
1160 ai = ptr;
1161 ptr += base_size;
1162
1163 ai->groups[0].cpu_map = ptr;
1164
1165 for (unit = 0; unit < nr_units; unit++)
1166 ai->groups[0].cpu_map[unit] = NR_CPUS;
1167
1168 ai->nr_groups = nr_groups;
1169 ai->__ai_size = PFN_ALIGN(ai_size);
1170
1171 return ai;
1172}
1173
1174/**
1175 * pcpu_free_alloc_info - free percpu allocation info
1176 * @ai: pcpu_alloc_info to free
1177 *
1178 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1179 */
1180void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1181{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001182 memblock_free_early(__pa(ai), ai->__ai_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001183}
1184
1185/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001186 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1187 * @lvl: loglevel
1188 * @ai: allocation info to dump
1189 *
1190 * Print out information about @ai using loglevel @lvl.
1191 */
1192static void pcpu_dump_alloc_info(const char *lvl,
1193 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001194{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001195 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001196 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001197 int alloc = 0, alloc_end = 0;
1198 int group, v;
1199 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001200
Tejun Heofd1e8a12009-08-14 15:00:51 +09001201 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001202 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001203 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001204
Tejun Heofd1e8a12009-08-14 15:00:51 +09001205 v = num_possible_cpus();
1206 while (v /= 10)
1207 cpu_width++;
1208 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001209
Tejun Heofd1e8a12009-08-14 15:00:51 +09001210 upa = ai->alloc_size / ai->unit_size;
1211 width = upa * (cpu_width + 1) + group_width + 3;
1212 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001213
Tejun Heofd1e8a12009-08-14 15:00:51 +09001214 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1215 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1216 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1217
1218 for (group = 0; group < ai->nr_groups; group++) {
1219 const struct pcpu_group_info *gi = &ai->groups[group];
1220 int unit = 0, unit_end = 0;
1221
1222 BUG_ON(gi->nr_units % upa);
1223 for (alloc_end += gi->nr_units / upa;
1224 alloc < alloc_end; alloc++) {
1225 if (!(alloc % apl)) {
Tejun Heocb129822012-03-29 09:45:58 -07001226 printk(KERN_CONT "\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001227 printk("%spcpu-alloc: ", lvl);
1228 }
Tejun Heocb129822012-03-29 09:45:58 -07001229 printk(KERN_CONT "[%0*d] ", group_width, group);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001230
1231 for (unit_end += upa; unit < unit_end; unit++)
1232 if (gi->cpu_map[unit] != NR_CPUS)
Tejun Heocb129822012-03-29 09:45:58 -07001233 printk(KERN_CONT "%0*d ", cpu_width,
Tejun Heofd1e8a12009-08-14 15:00:51 +09001234 gi->cpu_map[unit]);
1235 else
Tejun Heocb129822012-03-29 09:45:58 -07001236 printk(KERN_CONT "%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001237 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001238 }
Tejun Heocb129822012-03-29 09:45:58 -07001239 printk(KERN_CONT "\n");
Tejun Heo033e48f2009-08-14 15:00:51 +09001240}
Tejun Heo033e48f2009-08-14 15:00:51 +09001241
Tejun Heofbf59bc2009-02-20 16:29:08 +09001242/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001243 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001244 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001245 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001246 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001247 * Initialize the first percpu chunk which contains the kernel static
1248 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001249 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001250 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001251 * @ai contains all information necessary to initialize the first
1252 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001253 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001254 * @ai->static_size is the size of static percpu area.
1255 *
1256 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001257 * reserve after the static area in the first chunk. This reserves
1258 * the first chunk such that it's available only through reserved
1259 * percpu allocation. This is primarily used to serve module percpu
1260 * static areas on architectures where the addressing model has
1261 * limited offset range for symbol relocations to guarantee module
1262 * percpu symbols fall inside the relocatable range.
1263 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001264 * @ai->dyn_size determines the number of bytes available for dynamic
1265 * allocation in the first chunk. The area between @ai->static_size +
1266 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001267 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001268 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1269 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1270 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001271 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001272 * @ai->atom_size is the allocation atom size and used as alignment
1273 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001274 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001275 * @ai->alloc_size is the allocation size and always multiple of
1276 * @ai->atom_size. This is larger than @ai->atom_size if
1277 * @ai->unit_size is larger than @ai->atom_size.
1278 *
1279 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1280 * percpu areas. Units which should be colocated are put into the
1281 * same group. Dynamic VM areas will be allocated according to these
1282 * groupings. If @ai->nr_groups is zero, a single group containing
1283 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001284 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001285 * The caller should have mapped the first chunk at @base_addr and
1286 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001287 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001288 * If the first chunk ends up with both reserved and dynamic areas, it
1289 * is served by two chunks - one to serve the core static and reserved
1290 * areas and the other for the dynamic area. They share the same vm
1291 * and page map but uses different area allocation map to stay away
1292 * from each other. The latter chunk is circulated in the chunk slots
1293 * and available for dynamic allocation like any other chunks.
1294 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001295 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001296 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001297 */
Tejun Heofb435d52009-08-14 15:00:51 +09001298int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1299 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001300{
Tejun Heo635b75f2009-09-24 09:43:11 +09001301 static char cpus_buf[4096] __initdata;
Tejun Heo099a19d2010-06-27 18:50:00 +02001302 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1303 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001304 size_t dyn_size = ai->dyn_size;
1305 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001306 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001307 unsigned long *group_offsets;
1308 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001309 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001310 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001311 int *unit_map;
1312 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001313
Tejun Heo635b75f2009-09-24 09:43:11 +09001314 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1315
1316#define PCPU_SETUP_BUG_ON(cond) do { \
1317 if (unlikely(cond)) { \
1318 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1319 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1320 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1321 BUG(); \
1322 } \
1323} while (0)
1324
Tejun Heo2f39e632009-07-04 08:11:00 +09001325 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001326 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02001327#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09001328 PCPU_SETUP_BUG_ON(!ai->static_size);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001329 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
Tejun Heobbddff02010-09-03 18:22:48 +02001330#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09001331 PCPU_SETUP_BUG_ON(!base_addr);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001332 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
Tejun Heo635b75f2009-09-24 09:43:11 +09001333 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1334 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1335 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001336 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001337 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001338
Tejun Heo65632972009-08-14 15:00:52 +09001339 /* process group information and build config tables accordingly */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001340 group_offsets = memblock_virt_alloc(ai->nr_groups *
1341 sizeof(group_offsets[0]), 0);
1342 group_sizes = memblock_virt_alloc(ai->nr_groups *
1343 sizeof(group_sizes[0]), 0);
1344 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1345 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
Tejun Heo2f39e632009-07-04 08:11:00 +09001346
Tejun Heofd1e8a12009-08-14 15:00:51 +09001347 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001348 unit_map[cpu] = UINT_MAX;
Tejun Heoa855b842011-11-18 10:55:35 -08001349
1350 pcpu_low_unit_cpu = NR_CPUS;
1351 pcpu_high_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001352
Tejun Heofd1e8a12009-08-14 15:00:51 +09001353 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1354 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001355
Tejun Heo65632972009-08-14 15:00:52 +09001356 group_offsets[group] = gi->base_offset;
1357 group_sizes[group] = gi->nr_units * ai->unit_size;
1358
Tejun Heofd1e8a12009-08-14 15:00:51 +09001359 for (i = 0; i < gi->nr_units; i++) {
1360 cpu = gi->cpu_map[i];
1361 if (cpu == NR_CPUS)
1362 continue;
1363
Tejun Heo635b75f2009-09-24 09:43:11 +09001364 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1365 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1366 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001367
1368 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001369 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1370
Tejun Heoa855b842011-11-18 10:55:35 -08001371 /* determine low/high unit_cpu */
1372 if (pcpu_low_unit_cpu == NR_CPUS ||
1373 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1374 pcpu_low_unit_cpu = cpu;
1375 if (pcpu_high_unit_cpu == NR_CPUS ||
1376 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1377 pcpu_high_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001378 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001379 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001380 pcpu_nr_units = unit;
1381
1382 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001383 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1384
1385 /* we're done parsing the input, undefine BUG macro and dump config */
1386#undef PCPU_SETUP_BUG_ON
Tejun Heobcbea792010-12-22 14:19:14 +01001387 pcpu_dump_alloc_info(KERN_DEBUG, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001388
Tejun Heo65632972009-08-14 15:00:52 +09001389 pcpu_nr_groups = ai->nr_groups;
1390 pcpu_group_offsets = group_offsets;
1391 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001392 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001393 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001394
1395 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001396 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001397 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001398 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001399 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1400 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001401
Tejun Heod9b55ee2009-02-24 11:57:21 +09001402 /*
1403 * Allocate chunk slots. The additional last slot is for
1404 * empty chunks.
1405 */
1406 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001407 pcpu_slot = memblock_virt_alloc(
1408 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001409 for (i = 0; i < pcpu_nr_slots; i++)
1410 INIT_LIST_HEAD(&pcpu_slot[i]);
1411
Tejun Heoedcb4632009-03-06 14:33:59 +09001412 /*
1413 * Initialize static chunk. If reserved_size is zero, the
1414 * static chunk covers static area + dynamic allocation area
1415 * in the first chunk. If reserved_size is not zero, it
1416 * covers static area + reserved area (mostly used for module
1417 * static percpu allocation).
1418 */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001419 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heo2441d152009-03-06 14:33:59 +09001420 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001421 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001422 schunk->map = smap;
1423 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001424 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001425 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001426
Tejun Heofd1e8a12009-08-14 15:00:51 +09001427 if (ai->reserved_size) {
1428 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001429 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001430 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001431 } else {
1432 schunk->free_size = dyn_size;
1433 dyn_size = 0; /* dynamic area covered */
1434 }
Tejun Heo2441d152009-03-06 14:33:59 +09001435 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001436
Al Viro723ad1d2014-03-06 21:13:18 -05001437 schunk->map[0] = 1;
1438 schunk->map[1] = ai->static_size;
1439 schunk->map_used = 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001440 if (schunk->free_size)
Al Viro723ad1d2014-03-06 21:13:18 -05001441 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1442 else
1443 schunk->map[1] |= 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001444
Tejun Heoedcb4632009-03-06 14:33:59 +09001445 /* init dynamic chunk if necessary */
1446 if (dyn_size) {
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001447 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001448 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001449 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001450 dchunk->map = dmap;
1451 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001452 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001453 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001454
1455 dchunk->contig_hint = dchunk->free_size = dyn_size;
Al Viro723ad1d2014-03-06 21:13:18 -05001456 dchunk->map[0] = 1;
1457 dchunk->map[1] = pcpu_reserved_chunk_limit;
1458 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1459 dchunk->map_used = 2;
Tejun Heoedcb4632009-03-06 14:33:59 +09001460 }
1461
Tejun Heo2441d152009-03-06 14:33:59 +09001462 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001463 pcpu_first_chunk = dchunk ?: schunk;
1464 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001465
1466 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001467 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001468 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001469}
Tejun Heo66c3a752009-03-10 16:27:48 +09001470
Tejun Heobbddff02010-09-03 18:22:48 +02001471#ifdef CONFIG_SMP
1472
Andi Kleen17f36092012-10-04 17:12:07 -07001473const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
Tejun Heof58dc012009-08-14 15:00:50 +09001474 [PCPU_FC_AUTO] = "auto",
1475 [PCPU_FC_EMBED] = "embed",
1476 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001477};
Tejun Heo66c3a752009-03-10 16:27:48 +09001478
Tejun Heof58dc012009-08-14 15:00:50 +09001479enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1480
1481static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001482{
Cyrill Gorcunov5479c782012-11-25 01:17:13 +04001483 if (!str)
1484 return -EINVAL;
1485
Tejun Heof58dc012009-08-14 15:00:50 +09001486 if (0)
1487 /* nada */;
1488#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1489 else if (!strcmp(str, "embed"))
1490 pcpu_chosen_fc = PCPU_FC_EMBED;
1491#endif
1492#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1493 else if (!strcmp(str, "page"))
1494 pcpu_chosen_fc = PCPU_FC_PAGE;
1495#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001496 else
1497 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001498
Tejun Heof58dc012009-08-14 15:00:50 +09001499 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001500}
Tejun Heof58dc012009-08-14 15:00:50 +09001501early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001502
Tejun Heo3c9a0242010-09-09 18:00:15 +02001503/*
1504 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1505 * Build it if needed by the arch config or the generic setup is going
1506 * to be used.
1507 */
Tejun Heo08fc4582009-08-14 15:00:49 +09001508#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1509 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02001510#define BUILD_EMBED_FIRST_CHUNK
1511#endif
1512
1513/* build pcpu_page_first_chunk() iff needed by the arch config */
1514#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1515#define BUILD_PAGE_FIRST_CHUNK
1516#endif
1517
1518/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1519#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1520/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09001521 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1522 * @reserved_size: the size of reserved percpu area in bytes
1523 * @dyn_size: minimum free size for dynamic allocation in bytes
1524 * @atom_size: allocation atom size
1525 * @cpu_distance_fn: callback to determine distance between cpus, optional
1526 *
1527 * This function determines grouping of units, their mappings to cpus
1528 * and other parameters considering needed percpu size, allocation
1529 * atom size and distances between CPUs.
1530 *
1531 * Groups are always mutliples of atom size and CPUs which are of
1532 * LOCAL_DISTANCE both ways are grouped together and share space for
1533 * units in the same group. The returned configuration is guaranteed
1534 * to have CPUs on different nodes on different groups and >=75% usage
1535 * of allocated virtual address space.
1536 *
1537 * RETURNS:
1538 * On success, pointer to the new allocation_info is returned. On
1539 * failure, ERR_PTR value is returned.
1540 */
1541static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1542 size_t reserved_size, size_t dyn_size,
1543 size_t atom_size,
1544 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1545{
1546 static int group_map[NR_CPUS] __initdata;
1547 static int group_cnt[NR_CPUS] __initdata;
1548 const size_t static_size = __per_cpu_end - __per_cpu_start;
1549 int nr_groups = 1, nr_units = 0;
1550 size_t size_sum, min_unit_size, alloc_size;
1551 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1552 int last_allocs, group, unit;
1553 unsigned int cpu, tcpu;
1554 struct pcpu_alloc_info *ai;
1555 unsigned int *cpu_map;
1556
1557 /* this function may be called multiple times */
1558 memset(group_map, 0, sizeof(group_map));
1559 memset(group_cnt, 0, sizeof(group_cnt));
1560
1561 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1562 size_sum = PFN_ALIGN(static_size + reserved_size +
1563 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1564 dyn_size = size_sum - static_size - reserved_size;
1565
1566 /*
1567 * Determine min_unit_size, alloc_size and max_upa such that
1568 * alloc_size is multiple of atom_size and is the smallest
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001569 * which can accommodate 4k aligned segments which are equal to
Tejun Heofbf59bc2009-02-20 16:29:08 +09001570 * or larger than min_unit_size.
1571 */
1572 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1573
1574 alloc_size = roundup(min_unit_size, atom_size);
1575 upa = alloc_size / min_unit_size;
1576 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1577 upa--;
1578 max_upa = upa;
1579
1580 /* group cpus according to their proximity */
1581 for_each_possible_cpu(cpu) {
1582 group = 0;
1583 next_group:
1584 for_each_possible_cpu(tcpu) {
1585 if (cpu == tcpu)
1586 break;
1587 if (group_map[tcpu] == group && cpu_distance_fn &&
1588 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1589 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1590 group++;
1591 nr_groups = max(nr_groups, group + 1);
1592 goto next_group;
1593 }
1594 }
1595 group_map[cpu] = group;
1596 group_cnt[group]++;
1597 }
1598
1599 /*
1600 * Expand unit size until address space usage goes over 75%
1601 * and then as much as possible without using more address
1602 * space.
1603 */
1604 last_allocs = INT_MAX;
1605 for (upa = max_upa; upa; upa--) {
1606 int allocs = 0, wasted = 0;
1607
1608 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1609 continue;
1610
1611 for (group = 0; group < nr_groups; group++) {
1612 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1613 allocs += this_allocs;
1614 wasted += this_allocs * upa - group_cnt[group];
1615 }
1616
1617 /*
1618 * Don't accept if wastage is over 1/3. The
1619 * greater-than comparison ensures upa==1 always
1620 * passes the following check.
1621 */
1622 if (wasted > num_possible_cpus() / 3)
1623 continue;
1624
1625 /* and then don't consume more memory */
1626 if (allocs > last_allocs)
1627 break;
1628 last_allocs = allocs;
1629 best_upa = upa;
1630 }
1631 upa = best_upa;
1632
1633 /* allocate and fill alloc_info */
1634 for (group = 0; group < nr_groups; group++)
1635 nr_units += roundup(group_cnt[group], upa);
1636
1637 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1638 if (!ai)
1639 return ERR_PTR(-ENOMEM);
1640 cpu_map = ai->groups[0].cpu_map;
1641
1642 for (group = 0; group < nr_groups; group++) {
1643 ai->groups[group].cpu_map = cpu_map;
1644 cpu_map += roundup(group_cnt[group], upa);
1645 }
1646
1647 ai->static_size = static_size;
1648 ai->reserved_size = reserved_size;
1649 ai->dyn_size = dyn_size;
1650 ai->unit_size = alloc_size / upa;
1651 ai->atom_size = atom_size;
1652 ai->alloc_size = alloc_size;
1653
1654 for (group = 0, unit = 0; group_cnt[group]; group++) {
1655 struct pcpu_group_info *gi = &ai->groups[group];
1656
1657 /*
1658 * Initialize base_offset as if all groups are located
1659 * back-to-back. The caller should update this to
1660 * reflect actual allocation.
1661 */
1662 gi->base_offset = unit * ai->unit_size;
1663
1664 for_each_possible_cpu(cpu)
1665 if (group_map[cpu] == group)
1666 gi->cpu_map[gi->nr_units++] = cpu;
1667 gi->nr_units = roundup(gi->nr_units, upa);
1668 unit += gi->nr_units;
1669 }
1670 BUG_ON(unit != nr_units);
1671
1672 return ai;
1673}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001674#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001675
Tejun Heo3c9a0242010-09-09 18:00:15 +02001676#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09001677/**
1678 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001679 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001680 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001681 * @atom_size: allocation atom size
1682 * @cpu_distance_fn: callback to determine distance between cpus, optional
1683 * @alloc_fn: function to allocate percpu page
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001684 * @free_fn: function to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001685 *
1686 * This is a helper to ease setting up embedded first percpu chunk and
1687 * can be called where pcpu_setup_first_chunk() is expected.
1688 *
1689 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001690 * by calling @alloc_fn and used as-is without being mapped into
1691 * vmalloc area. Allocations are always whole multiples of @atom_size
1692 * aligned to @atom_size.
1693 *
1694 * This enables the first chunk to piggy back on the linear physical
1695 * mapping which often uses larger page size. Please note that this
1696 * can result in very sparse cpu->unit mapping on NUMA machines thus
1697 * requiring large vmalloc address space. Don't use this allocator if
1698 * vmalloc space is not orders of magnitude larger than distances
1699 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001700 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001701 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001702 *
1703 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001704 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001705 *
1706 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001707 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001708 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001709int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001710 size_t atom_size,
1711 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1712 pcpu_fc_alloc_fn_t alloc_fn,
1713 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001714{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001715 void *base = (void *)ULONG_MAX;
1716 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001717 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001718 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001719 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001720
Tejun Heoc8826dd2009-08-14 15:00:52 +09001721 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1722 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001723 if (IS_ERR(ai))
1724 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001725
Tejun Heofd1e8a12009-08-14 15:00:51 +09001726 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001727 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001728
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001729 areas = memblock_virt_alloc_nopanic(areas_size, 0);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001730 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001731 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001732 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001733 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001734
Tejun Heoc8826dd2009-08-14 15:00:52 +09001735 /* allocate, copy and determine base address */
1736 for (group = 0; group < ai->nr_groups; group++) {
1737 struct pcpu_group_info *gi = &ai->groups[group];
1738 unsigned int cpu = NR_CPUS;
1739 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001740
Tejun Heoc8826dd2009-08-14 15:00:52 +09001741 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1742 cpu = gi->cpu_map[i];
1743 BUG_ON(cpu == NR_CPUS);
1744
1745 /* allocate space for the whole group */
1746 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1747 if (!ptr) {
1748 rc = -ENOMEM;
1749 goto out_free_areas;
1750 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001751 /* kmemleak tracks the percpu allocations separately */
1752 kmemleak_free(ptr);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001753 areas[group] = ptr;
1754
1755 base = min(ptr, base);
Tejun Heo42b64282012-04-27 08:42:53 -07001756 }
1757
1758 /*
1759 * Copy data and free unused parts. This should happen after all
1760 * allocations are complete; otherwise, we may end up with
1761 * overlapping groups.
1762 */
1763 for (group = 0; group < ai->nr_groups; group++) {
1764 struct pcpu_group_info *gi = &ai->groups[group];
1765 void *ptr = areas[group];
Tejun Heoc8826dd2009-08-14 15:00:52 +09001766
1767 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1768 if (gi->cpu_map[i] == NR_CPUS) {
1769 /* unused unit, free whole */
1770 free_fn(ptr, ai->unit_size);
1771 continue;
1772 }
1773 /* copy and return the unused part */
1774 memcpy(ptr, __per_cpu_load, ai->static_size);
1775 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1776 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001777 }
1778
Tejun Heoc8826dd2009-08-14 15:00:52 +09001779 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001780 max_distance = 0;
1781 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001782 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001783 max_distance = max_t(size_t, max_distance,
1784 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001785 }
1786 max_distance += ai->unit_size;
1787
1788 /* warn if maximum distance is further than 75% of vmalloc space */
Laura Abbott8a092172014-01-02 13:53:21 -08001789 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001790 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Mike Frysinger787e5b02011-03-23 08:23:52 +01001791 "space 0x%lx\n", max_distance,
Laura Abbott8a092172014-01-02 13:53:21 -08001792 VMALLOC_TOTAL);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001793#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1794 /* and fail if we have fallback */
1795 rc = -EINVAL;
1796 goto out_free;
1797#endif
1798 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001799
Tejun Heo004018e2009-08-14 15:00:49 +09001800 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001801 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1802 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001803
Tejun Heofb435d52009-08-14 15:00:51 +09001804 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001805 goto out_free;
1806
1807out_free_areas:
1808 for (group = 0; group < ai->nr_groups; group++)
Michael Holzheuf851c8d2013-09-17 16:57:34 +02001809 if (areas[group])
1810 free_fn(areas[group],
1811 ai->groups[group].nr_units * ai->unit_size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001812out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001813 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001814 if (areas)
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001815 memblock_free_early(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001816 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001817}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001818#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001819
Tejun Heo3c9a0242010-09-09 18:00:15 +02001820#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001821/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001822 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001823 * @reserved_size: the size of reserved percpu area in bytes
1824 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001825 * @free_fn: function to free percpu page, always called with PAGE_SIZE
Tejun Heod4b95f82009-07-04 08:10:59 +09001826 * @populate_pte_fn: function to populate pte
1827 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001828 * This is a helper to ease setting up page-remapped first percpu
1829 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001830 *
1831 * This is the basic allocator. Static percpu area is allocated
1832 * page-by-page into vmalloc area.
1833 *
1834 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001835 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001836 */
Tejun Heofb435d52009-08-14 15:00:51 +09001837int __init pcpu_page_first_chunk(size_t reserved_size,
1838 pcpu_fc_alloc_fn_t alloc_fn,
1839 pcpu_fc_free_fn_t free_fn,
1840 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001841{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001842 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001843 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001844 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09001845 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09001846 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001847 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09001848 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001849
Tejun Heo00ae4062009-08-14 15:00:49 +09001850 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1851
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001852 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001853 if (IS_ERR(ai))
1854 return PTR_ERR(ai);
1855 BUG_ON(ai->nr_groups != 1);
1856 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1857
1858 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09001859
1860 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001861 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1862 sizeof(pages[0]));
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001863 pages = memblock_virt_alloc(pages_size, 0);
Tejun Heod4b95f82009-07-04 08:10:59 +09001864
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001865 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09001866 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001867 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09001868 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09001869 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09001870 void *ptr;
1871
Tejun Heo3cbc8562009-08-14 15:00:50 +09001872 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09001873 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09001874 pr_warning("PERCPU: failed to allocate %s page "
1875 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09001876 goto enomem;
1877 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001878 /* kmemleak tracks the percpu allocations separately */
1879 kmemleak_free(ptr);
Tejun Heoce3141a2009-07-04 08:11:00 +09001880 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001881 }
1882
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001883 /* allocate vm area, map the pages and copy static data */
1884 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001885 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001886 vm_area_register_early(&vm, PAGE_SIZE);
1887
Tejun Heofd1e8a12009-08-14 15:00:51 +09001888 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09001889 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09001890 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001891
Tejun Heoce3141a2009-07-04 08:11:00 +09001892 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001893 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1894
1895 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09001896 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1897 unit_pages);
1898 if (rc < 0)
1899 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001900
1901 /*
1902 * FIXME: Archs with virtual cache should flush local
1903 * cache for the linear mapping here - something
1904 * equivalent to flush_cache_vmap() on the local cpu.
1905 * flush_cache_vmap() can't be used as most supporting
1906 * data structures are not set up yet.
1907 */
1908
1909 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001910 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001911 }
1912
1913 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09001914 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001915 unit_pages, psize_str, vm.addr, ai->static_size,
1916 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001917
Tejun Heofb435d52009-08-14 15:00:51 +09001918 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09001919 goto out_free_ar;
1920
1921enomem:
1922 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09001923 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09001924 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09001925out_free_ar:
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001926 memblock_free_early(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001927 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09001928 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001929}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001930#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09001931
Tejun Heobbddff02010-09-03 18:22:48 +02001932#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09001933/*
Tejun Heobbddff02010-09-03 18:22:48 +02001934 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09001935 *
1936 * The embedding helper is used because its behavior closely resembles
1937 * the original non-dynamic generic percpu area setup. This is
1938 * important because many archs have addressing restrictions and might
1939 * fail if the percpu area is located far away from the previous
1940 * location. As an added bonus, in non-NUMA cases, embedding is
1941 * generally a good idea TLB-wise because percpu area can piggy back
1942 * on the physical linear memory mapping which uses large page
1943 * mappings on applicable archs.
1944 */
Tejun Heoe74e3962009-03-30 19:07:44 +09001945unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1946EXPORT_SYMBOL(__per_cpu_offset);
1947
Tejun Heoc8826dd2009-08-14 15:00:52 +09001948static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1949 size_t align)
1950{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001951 return memblock_virt_alloc_from_nopanic(
1952 size, align, __pa(MAX_DMA_ADDRESS));
Tejun Heoc8826dd2009-08-14 15:00:52 +09001953}
1954
1955static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1956{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001957 memblock_free_early(__pa(ptr), size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001958}
1959
Tejun Heoe74e3962009-03-30 19:07:44 +09001960void __init setup_per_cpu_areas(void)
1961{
Tejun Heoe74e3962009-03-30 19:07:44 +09001962 unsigned long delta;
1963 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09001964 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09001965
1966 /*
1967 * Always reserve area for module percpu variables. That's
1968 * what the legacy allocator did.
1969 */
Tejun Heofb435d52009-08-14 15:00:51 +09001970 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001971 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1972 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09001973 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02001974 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09001975
1976 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1977 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09001978 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09001979}
Tejun Heobbddff02010-09-03 18:22:48 +02001980#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1981
1982#else /* CONFIG_SMP */
1983
1984/*
1985 * UP percpu area setup.
1986 *
1987 * UP always uses km-based percpu allocator with identity mapping.
1988 * Static percpu variables are indistinguishable from the usual static
1989 * variables and don't require any special preparation.
1990 */
1991void __init setup_per_cpu_areas(void)
1992{
1993 const size_t unit_size =
1994 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1995 PERCPU_DYNAMIC_RESERVE));
1996 struct pcpu_alloc_info *ai;
1997 void *fc;
1998
1999 ai = pcpu_alloc_alloc_info(1, 1);
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002000 fc = memblock_virt_alloc_from_nopanic(unit_size,
2001 PAGE_SIZE,
2002 __pa(MAX_DMA_ADDRESS));
Tejun Heobbddff02010-09-03 18:22:48 +02002003 if (!ai || !fc)
2004 panic("Failed to allocate memory for percpu areas.");
Catalin Marinas100d13c2012-05-09 16:55:19 +01002005 /* kmemleak tracks the percpu allocations separately */
2006 kmemleak_free(fc);
Tejun Heobbddff02010-09-03 18:22:48 +02002007
2008 ai->dyn_size = unit_size;
2009 ai->unit_size = unit_size;
2010 ai->atom_size = unit_size;
2011 ai->alloc_size = unit_size;
2012 ai->groups[0].nr_units = 1;
2013 ai->groups[0].cpu_map[0] = 0;
2014
2015 if (pcpu_setup_first_chunk(ai, fc) < 0)
2016 panic("Failed to initialize percpu areas.");
Honggang Li3189edd2014-08-12 21:36:15 +08002017
2018 pcpu_free_alloc_info(ai);
Tejun Heobbddff02010-09-03 18:22:48 +02002019}
2020
2021#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02002022
2023/*
2024 * First and reserved chunks are initialized with temporary allocation
2025 * map in initdata so that they can be used before slab is online.
2026 * This function is called after slab is brought up and replaces those
2027 * with properly allocated maps.
2028 */
2029void __init percpu_init_late(void)
2030{
2031 struct pcpu_chunk *target_chunks[] =
2032 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2033 struct pcpu_chunk *chunk;
2034 unsigned long flags;
2035 int i;
2036
2037 for (i = 0; (chunk = target_chunks[i]); i++) {
2038 int *map;
2039 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2040
2041 BUILD_BUG_ON(size > PAGE_SIZE);
2042
Bob Liu90459ce02011-08-04 11:02:33 +02002043 map = pcpu_mem_zalloc(size);
Tejun Heo099a19d2010-06-27 18:50:00 +02002044 BUG_ON(!map);
2045
2046 spin_lock_irqsave(&pcpu_lock, flags);
2047 memcpy(map, chunk->map, size);
2048 chunk->map = map;
2049 spin_unlock_irqrestore(&pcpu_lock, flags);
2050 }
2051}