blob: d39e2f4e335c36497520b02859c02eb7f46f220f [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 Kim47858792010-08-11 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 */
Tejun Heo9c824b62014-09-02 14:46:05 -040079#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
80#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
Tejun Heo1a4d7602014-09-02 14:46:05 -040081#define PCPU_EMPTY_POP_PAGES_LOW 2
82#define PCPU_EMPTY_POP_PAGES_HIGH 4
Tejun Heofbf59bc2009-02-20 16:29:08 +090083
Tejun Heobbddff02010-09-03 18:22:48 +020084#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +090085/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
86#ifndef __addr_to_pcpu_ptr
87#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090088 (void __percpu *)((unsigned long)(addr) - \
89 (unsigned long)pcpu_base_addr + \
90 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090091#endif
92#ifndef __pcpu_ptr_to_addr
93#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090094 (void __force *)((unsigned long)(ptr) + \
95 (unsigned long)pcpu_base_addr - \
96 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090097#endif
Tejun Heobbddff02010-09-03 18:22:48 +020098#else /* CONFIG_SMP */
99/* on UP, it's always identity mapped */
100#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
101#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
102#endif /* CONFIG_SMP */
Tejun Heoe0100982009-03-10 16:27:48 +0900103
Tejun Heofbf59bc2009-02-20 16:29:08 +0900104struct pcpu_chunk {
105 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900106 int free_size; /* free bytes in the chunk */
107 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +0900108 void *base_addr; /* base address of this chunk */
Tejun Heo9c824b62014-09-02 14:46:05 -0400109
Al Viro723ad1d2014-03-06 21:13:18 -0500110 int map_used; /* # of map entries used before the sentry */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900111 int map_alloc; /* # of map entries allocated */
112 int *map; /* allocation map */
Tejun Heo9c824b62014-09-02 14:46:05 -0400113 struct work_struct map_extend_work;/* async ->map[] extension */
114
Tejun Heo88999a82010-04-09 18:57:01 +0900115 void *data; /* chunk data */
Al Viro3d331ad2014-03-06 20:52:32 -0500116 int first_free; /* no free below this */
Tejun Heo8d408b42009-02-24 11:57:21 +0900117 bool immutable; /* no [de]population allowed */
Tejun Heob539b872014-09-02 14:46:05 -0400118 int nr_populated; /* # of populated pages */
Tejun Heoce3141a2009-07-04 08:11:00 +0900119 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900120};
121
Tejun Heo40150d32009-02-24 12:32:28 +0900122static int pcpu_unit_pages __read_mostly;
123static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900124static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900125static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900126static int pcpu_nr_slots __read_mostly;
127static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900128
Tejun Heoa855b842011-11-18 10:55:35 -0800129/* cpus with the lowest and highest unit addresses */
130static unsigned int pcpu_low_unit_cpu __read_mostly;
131static unsigned int pcpu_high_unit_cpu __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900132
Tejun Heofbf59bc2009-02-20 16:29:08 +0900133/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900134void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900135EXPORT_SYMBOL_GPL(pcpu_base_addr);
136
Tejun Heofb435d52009-08-14 15:00:51 +0900137static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
138const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900139
Tejun Heo65632972009-08-14 15:00:52 +0900140/* group information, used for vm allocation */
141static int pcpu_nr_groups __read_mostly;
142static const unsigned long *pcpu_group_offsets __read_mostly;
143static const size_t *pcpu_group_sizes __read_mostly;
144
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900145/*
146 * The first chunk which always exists. Note that unlike other
147 * chunks, this one can be allocated and mapped in several different
148 * ways and thus often doesn't live in the vmalloc area.
149 */
150static struct pcpu_chunk *pcpu_first_chunk;
151
152/*
153 * Optional reserved chunk. This chunk reserves part of the first
154 * chunk and serves it for reserved allocations. The amount of
155 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
156 * area doesn't exist, the following variables contain NULL and 0
157 * respectively.
158 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900159static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900160static int pcpu_reserved_chunk_limit;
161
Tejun Heob38d08f2014-09-02 14:46:02 -0400162static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
163static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900164
Tejun Heo40150d32009-02-24 12:32:28 +0900165static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900166
Tejun Heob539b872014-09-02 14:46:05 -0400167/*
168 * The number of empty populated pages, protected by pcpu_lock. The
169 * reserved chunk doesn't contribute to the count.
170 */
171static int pcpu_nr_empty_pop_pages;
172
Tejun Heo1a4d7602014-09-02 14:46:05 -0400173/*
174 * Balance work is used to populate or destroy chunks asynchronously. We
175 * try to keep the number of populated free pages between
176 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
177 * empty chunk.
178 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -0400179static void pcpu_balance_workfn(struct work_struct *work);
180static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
Tejun Heo1a4d7602014-09-02 14:46:05 -0400181static bool pcpu_async_enabled __read_mostly;
182static bool pcpu_atomic_alloc_failed;
183
184static void pcpu_schedule_balance_work(void)
185{
186 if (pcpu_async_enabled)
187 schedule_work(&pcpu_balance_work);
188}
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900189
Tejun Heo020ec652010-04-09 18:57:00 +0900190static bool pcpu_addr_in_first_chunk(void *addr)
191{
192 void *first_start = pcpu_first_chunk->base_addr;
193
194 return addr >= first_start && addr < first_start + pcpu_unit_size;
195}
196
197static bool pcpu_addr_in_reserved_chunk(void *addr)
198{
199 void *first_start = pcpu_first_chunk->base_addr;
200
201 return addr >= first_start &&
202 addr < first_start + pcpu_reserved_chunk_limit;
203}
204
Tejun Heod9b55ee2009-02-24 11:57:21 +0900205static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900206{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900207 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900208 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
209}
210
Tejun Heod9b55ee2009-02-24 11:57:21 +0900211static int pcpu_size_to_slot(int size)
212{
213 if (size == pcpu_unit_size)
214 return pcpu_nr_slots - 1;
215 return __pcpu_size_to_slot(size);
216}
217
Tejun Heofbf59bc2009-02-20 16:29:08 +0900218static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
219{
220 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
221 return 0;
222
223 return pcpu_size_to_slot(chunk->free_size);
224}
225
Tejun Heo88999a82010-04-09 18:57:01 +0900226/* set the pointer to a chunk in a page struct */
227static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
228{
229 page->index = (unsigned long)pcpu;
230}
231
232/* obtain pointer to a chunk from a page struct */
233static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
234{
235 return (struct pcpu_chunk *)page->index;
236}
237
238static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900239{
Tejun Heo2f39e632009-07-04 08:11:00 +0900240 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900241}
242
Tejun Heo9983b6f02010-06-18 11:44:31 +0200243static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
244 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900245{
Tejun Heobba174f2009-08-14 15:00:51 +0900246 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900247 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900248}
249
Tejun Heo88999a82010-04-09 18:57:01 +0900250static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
251 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900252{
253 *rs = find_next_zero_bit(chunk->populated, end, *rs);
254 *re = find_next_bit(chunk->populated, end, *rs + 1);
255}
256
Tejun Heo88999a82010-04-09 18:57:01 +0900257static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
258 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900259{
260 *rs = find_next_bit(chunk->populated, end, *rs);
261 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
262}
263
264/*
265 * (Un)populated page region iterators. Iterate over (un)populated
Uwe Kleine-Königb5950762010-11-01 15:38:34 -0400266 * page regions between @start and @end in @chunk. @rs and @re should
Tejun Heoce3141a2009-07-04 08:11:00 +0900267 * be integer variables and will be set to start and end page index of
268 * the current region.
269 */
270#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
271 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
272 (rs) < (re); \
273 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
274
275#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
276 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
277 (rs) < (re); \
278 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
279
Tejun Heofbf59bc2009-02-20 16:29:08 +0900280/**
Bob Liu90459ce02011-08-04 11:02:33 +0200281 * pcpu_mem_zalloc - allocate memory
Tejun Heo1880d932009-03-07 00:44:09 +0900282 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900283 *
Tejun Heo1880d932009-03-07 00:44:09 +0900284 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
Bob Liu90459ce02011-08-04 11:02:33 +0200285 * kzalloc() is used; otherwise, vzalloc() is used. The returned
Tejun Heo1880d932009-03-07 00:44:09 +0900286 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900287 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900288 * CONTEXT:
289 * Does GFP_KERNEL allocation.
290 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900291 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900292 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900293 */
Bob Liu90459ce02011-08-04 11:02:33 +0200294static void *pcpu_mem_zalloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900295{
Tejun Heo099a19d2010-06-27 18:50:00 +0200296 if (WARN_ON_ONCE(!slab_is_available()))
297 return NULL;
298
Tejun Heofbf59bc2009-02-20 16:29:08 +0900299 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900300 return kzalloc(size, GFP_KERNEL);
Jesper Juhl7af4c092010-10-30 15:56:54 +0200301 else
302 return vzalloc(size);
Tejun Heo1880d932009-03-07 00:44:09 +0900303}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900304
Tejun Heo1880d932009-03-07 00:44:09 +0900305/**
306 * pcpu_mem_free - free memory
307 * @ptr: memory to free
308 * @size: size of the area
309 *
Bob Liu90459ce02011-08-04 11:02:33 +0200310 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
Tejun Heo1880d932009-03-07 00:44:09 +0900311 */
312static void pcpu_mem_free(void *ptr, size_t size)
313{
314 if (size <= PAGE_SIZE)
315 kfree(ptr);
316 else
317 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900318}
319
320/**
Tejun Heob539b872014-09-02 14:46:05 -0400321 * pcpu_count_occupied_pages - count the number of pages an area occupies
322 * @chunk: chunk of interest
323 * @i: index of the area in question
324 *
325 * Count the number of pages chunk's @i'th area occupies. When the area's
326 * start and/or end address isn't aligned to page boundary, the straddled
327 * page is included in the count iff the rest of the page is free.
328 */
329static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
330{
331 int off = chunk->map[i] & ~1;
332 int end = chunk->map[i + 1] & ~1;
333
334 if (!PAGE_ALIGNED(off) && i > 0) {
335 int prev = chunk->map[i - 1];
336
337 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
338 off = round_down(off, PAGE_SIZE);
339 }
340
341 if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
342 int next = chunk->map[i + 1];
343 int nend = chunk->map[i + 2] & ~1;
344
345 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
346 end = round_up(end, PAGE_SIZE);
347 }
348
349 return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
350}
351
352/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900353 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
354 * @chunk: chunk of interest
355 * @oslot: the previous slot it was on
356 *
357 * This function is called after an allocation or free changed @chunk.
358 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900359 * moved to the slot. Note that the reserved chunk is never put on
360 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900361 *
362 * CONTEXT:
363 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900364 */
365static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
366{
367 int nslot = pcpu_chunk_slot(chunk);
368
Tejun Heoedcb4632009-03-06 14:33:59 +0900369 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900370 if (oslot < nslot)
371 list_move(&chunk->list, &pcpu_slot[nslot]);
372 else
373 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
374 }
375}
376
Tejun Heofbf59bc2009-02-20 16:29:08 +0900377/**
Tejun Heo833af842009-11-11 15:35:18 +0900378 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
379 * @chunk: chunk of interest
Tejun Heo9c824b62014-09-02 14:46:05 -0400380 * @is_atomic: the allocation context
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900381 *
Tejun Heo9c824b62014-09-02 14:46:05 -0400382 * Determine whether area map of @chunk needs to be extended. If
383 * @is_atomic, only the amount necessary for a new allocation is
384 * considered; however, async extension is scheduled if the left amount is
385 * low. If !@is_atomic, it aims for more empty space. Combined, this
386 * ensures that the map is likely to have enough available space to
387 * accomodate atomic allocations which can't extend maps directly.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900388 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900389 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900390 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900391 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900392 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900393 * New target map allocation length if extension is necessary, 0
394 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900395 */
Tejun Heo9c824b62014-09-02 14:46:05 -0400396static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900397{
Tejun Heo9c824b62014-09-02 14:46:05 -0400398 int margin, new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900399
Tejun Heo9c824b62014-09-02 14:46:05 -0400400 if (is_atomic) {
401 margin = 3;
402
403 if (chunk->map_alloc <
Tejun Heo1a4d7602014-09-02 14:46:05 -0400404 chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
405 pcpu_async_enabled)
Tejun Heo9c824b62014-09-02 14:46:05 -0400406 schedule_work(&chunk->map_extend_work);
407 } else {
408 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
409 }
410
411 if (chunk->map_alloc >= chunk->map_used + margin)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900412 return 0;
413
414 new_alloc = PCPU_DFL_MAP_ALLOC;
Tejun Heo9c824b62014-09-02 14:46:05 -0400415 while (new_alloc < chunk->map_used + margin)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900416 new_alloc *= 2;
417
Tejun Heo833af842009-11-11 15:35:18 +0900418 return new_alloc;
419}
420
421/**
422 * pcpu_extend_area_map - extend area map of a chunk
423 * @chunk: chunk of interest
424 * @new_alloc: new target allocation length of the area map
425 *
426 * Extend area map of @chunk to have @new_alloc entries.
427 *
428 * CONTEXT:
429 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
430 *
431 * RETURNS:
432 * 0 on success, -errno on failure.
433 */
434static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
435{
436 int *old = NULL, *new = NULL;
437 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
438 unsigned long flags;
439
Bob Liu90459ce02011-08-04 11:02:33 +0200440 new = pcpu_mem_zalloc(new_size);
Tejun Heo833af842009-11-11 15:35:18 +0900441 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900442 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900443
Tejun Heo833af842009-11-11 15:35:18 +0900444 /* acquire pcpu_lock and switch to new area map */
445 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900446
Tejun Heo833af842009-11-11 15:35:18 +0900447 if (new_alloc <= chunk->map_alloc)
448 goto out_unlock;
449
450 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
Huang Shijiea002d142010-08-08 14:39:07 +0200451 old = chunk->map;
452
453 memcpy(new, old, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900454
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900455 chunk->map_alloc = new_alloc;
456 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900457 new = NULL;
458
459out_unlock:
460 spin_unlock_irqrestore(&pcpu_lock, flags);
461
462 /*
463 * pcpu_mem_free() might end up calling vfree() which uses
464 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
465 */
466 pcpu_mem_free(old, old_size);
467 pcpu_mem_free(new, new_size);
468
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900469 return 0;
470}
471
Tejun Heo9c824b62014-09-02 14:46:05 -0400472static void pcpu_map_extend_workfn(struct work_struct *work)
473{
474 struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk,
475 map_extend_work);
476 int new_alloc;
477
478 spin_lock_irq(&pcpu_lock);
479 new_alloc = pcpu_need_to_extend(chunk, false);
480 spin_unlock_irq(&pcpu_lock);
481
482 if (new_alloc)
483 pcpu_extend_area_map(chunk, new_alloc);
484}
485
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900486/**
Tejun Heoa16037c2014-09-02 14:46:02 -0400487 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
488 * @chunk: chunk the candidate area belongs to
489 * @off: the offset to the start of the candidate area
490 * @this_size: the size of the candidate area
491 * @size: the size of the target allocation
492 * @align: the alignment of the target allocation
493 * @pop_only: only allocate from already populated region
494 *
495 * We're trying to allocate @size bytes aligned at @align. @chunk's area
496 * at @off sized @this_size is a candidate. This function determines
497 * whether the target allocation fits in the candidate area and returns the
498 * number of bytes to pad after @off. If the target area doesn't fit, -1
499 * is returned.
500 *
501 * If @pop_only is %true, this function only considers the already
502 * populated part of the candidate area.
503 */
504static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
505 int size, int align, bool pop_only)
506{
507 int cand_off = off;
508
509 while (true) {
510 int head = ALIGN(cand_off, align) - off;
511 int page_start, page_end, rs, re;
512
513 if (this_size < head + size)
514 return -1;
515
516 if (!pop_only)
517 return head;
518
519 /*
520 * If the first unpopulated page is beyond the end of the
521 * allocation, the whole allocation is populated;
522 * otherwise, retry from the end of the unpopulated area.
523 */
524 page_start = PFN_DOWN(head + off);
525 page_end = PFN_UP(head + off + size);
526
527 rs = page_start;
528 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
529 if (rs >= page_end)
530 return head;
531 cand_off = re * PAGE_SIZE;
532 }
533}
534
535/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900536 * pcpu_alloc_area - allocate area from a pcpu_chunk
537 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900538 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900539 * @align: wanted align
Tejun Heoa16037c2014-09-02 14:46:02 -0400540 * @pop_only: allocate only from the populated area
Tejun Heob539b872014-09-02 14:46:05 -0400541 * @occ_pages_p: out param for the number of pages the area occupies
Tejun Heofbf59bc2009-02-20 16:29:08 +0900542 *
543 * Try to allocate @size bytes area aligned at @align from @chunk.
544 * Note that this function only allocates the offset. It doesn't
545 * populate or map the area.
546 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900547 * @chunk->map must have at least two free slots.
548 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900549 * CONTEXT:
550 * pcpu_lock.
551 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900552 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900553 * Allocated offset in @chunk on success, -1 if no matching area is
554 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900555 */
Tejun Heoa16037c2014-09-02 14:46:02 -0400556static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
Tejun Heob539b872014-09-02 14:46:05 -0400557 bool pop_only, int *occ_pages_p)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900558{
559 int oslot = pcpu_chunk_slot(chunk);
560 int max_contig = 0;
561 int i, off;
Al Viro3d331ad2014-03-06 20:52:32 -0500562 bool seen_free = false;
Al Viro723ad1d2014-03-06 21:13:18 -0500563 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900564
Al Viro3d331ad2014-03-06 20:52:32 -0500565 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900566 int head, tail;
Al Viro723ad1d2014-03-06 21:13:18 -0500567 int this_size;
568
569 off = *p;
570 if (off & 1)
571 continue;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900572
Al Viro723ad1d2014-03-06 21:13:18 -0500573 this_size = (p[1] & ~1) - off;
Tejun Heoa16037c2014-09-02 14:46:02 -0400574
575 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
576 pop_only);
577 if (head < 0) {
Al Viro3d331ad2014-03-06 20:52:32 -0500578 if (!seen_free) {
579 chunk->first_free = i;
580 seen_free = true;
581 }
Al Viro723ad1d2014-03-06 21:13:18 -0500582 max_contig = max(this_size, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900583 continue;
584 }
585
586 /*
587 * If head is small or the previous block is free,
588 * merge'em. Note that 'small' is defined as smaller
589 * than sizeof(int), which is very small but isn't too
590 * uncommon for percpu allocations.
591 */
Al Viro723ad1d2014-03-06 21:13:18 -0500592 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800593 *p = off += head;
Al Viro723ad1d2014-03-06 21:13:18 -0500594 if (p[-1] & 1)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900595 chunk->free_size -= head;
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800596 else
597 max_contig = max(*p - p[-1], max_contig);
Al Viro723ad1d2014-03-06 21:13:18 -0500598 this_size -= head;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900599 head = 0;
600 }
601
602 /* if tail is small, just keep it around */
Al Viro723ad1d2014-03-06 21:13:18 -0500603 tail = this_size - head - size;
604 if (tail < sizeof(int)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900605 tail = 0;
Al Viro723ad1d2014-03-06 21:13:18 -0500606 size = this_size - head;
607 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900608
609 /* split if warranted */
610 if (head || tail) {
Al Viro706c16f2014-03-06 21:08:24 -0500611 int nr_extra = !!head + !!tail;
612
613 /* insert new subblocks */
Al Viro723ad1d2014-03-06 21:13:18 -0500614 memmove(p + nr_extra + 1, p + 1,
Al Viro706c16f2014-03-06 21:08:24 -0500615 sizeof(chunk->map[0]) * (chunk->map_used - i));
616 chunk->map_used += nr_extra;
617
Tejun Heofbf59bc2009-02-20 16:29:08 +0900618 if (head) {
Al Viro3d331ad2014-03-06 20:52:32 -0500619 if (!seen_free) {
620 chunk->first_free = i;
621 seen_free = true;
622 }
Al Viro723ad1d2014-03-06 21:13:18 -0500623 *++p = off += head;
624 ++i;
Al Viro706c16f2014-03-06 21:08:24 -0500625 max_contig = max(head, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900626 }
Al Viro706c16f2014-03-06 21:08:24 -0500627 if (tail) {
Al Viro723ad1d2014-03-06 21:13:18 -0500628 p[1] = off + size;
Al Viro706c16f2014-03-06 21:08:24 -0500629 max_contig = max(tail, max_contig);
630 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900631 }
632
Al Viro3d331ad2014-03-06 20:52:32 -0500633 if (!seen_free)
634 chunk->first_free = i + 1;
635
Tejun Heofbf59bc2009-02-20 16:29:08 +0900636 /* update hint and mark allocated */
Al Viro723ad1d2014-03-06 21:13:18 -0500637 if (i + 1 == chunk->map_used)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900638 chunk->contig_hint = max_contig; /* fully scanned */
639 else
640 chunk->contig_hint = max(chunk->contig_hint,
641 max_contig);
642
Al Viro723ad1d2014-03-06 21:13:18 -0500643 chunk->free_size -= size;
644 *p |= 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900645
Tejun Heob539b872014-09-02 14:46:05 -0400646 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900647 pcpu_chunk_relocate(chunk, oslot);
648 return off;
649 }
650
651 chunk->contig_hint = max_contig; /* fully scanned */
652 pcpu_chunk_relocate(chunk, oslot);
653
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900654 /* tell the upper layer that this chunk has no matching area */
655 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900656}
657
658/**
659 * pcpu_free_area - free area to a pcpu_chunk
660 * @chunk: chunk of interest
661 * @freeme: offset of area to free
Tejun Heob539b872014-09-02 14:46:05 -0400662 * @occ_pages_p: out param for the number of pages the area occupies
Tejun Heofbf59bc2009-02-20 16:29:08 +0900663 *
664 * Free area starting from @freeme to @chunk. Note that this function
665 * only modifies the allocation map. It doesn't depopulate or unmap
666 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900667 *
668 * CONTEXT:
669 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900670 */
Tejun Heob539b872014-09-02 14:46:05 -0400671static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
672 int *occ_pages_p)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900673{
674 int oslot = pcpu_chunk_slot(chunk);
Al Viro723ad1d2014-03-06 21:13:18 -0500675 int off = 0;
676 unsigned i, j;
677 int to_free = 0;
678 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900679
Al Viro723ad1d2014-03-06 21:13:18 -0500680 freeme |= 1; /* we are searching for <given offset, in use> pair */
681
682 i = 0;
683 j = chunk->map_used;
684 while (i != j) {
685 unsigned k = (i + j) / 2;
686 off = chunk->map[k];
687 if (off < freeme)
688 i = k + 1;
689 else if (off > freeme)
690 j = k;
691 else
692 i = j = k;
693 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900694 BUG_ON(off != freeme);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900695
Al Viro3d331ad2014-03-06 20:52:32 -0500696 if (i < chunk->first_free)
697 chunk->first_free = i;
698
Al Viro723ad1d2014-03-06 21:13:18 -0500699 p = chunk->map + i;
700 *p = off &= ~1;
701 chunk->free_size += (p[1] & ~1) - off;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900702
Tejun Heob539b872014-09-02 14:46:05 -0400703 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
704
Tejun Heofbf59bc2009-02-20 16:29:08 +0900705 /* merge with next? */
Al Viro723ad1d2014-03-06 21:13:18 -0500706 if (!(p[1] & 1))
707 to_free++;
708 /* merge with previous? */
709 if (i > 0 && !(p[-1] & 1)) {
710 to_free++;
711 i--;
712 p--;
713 }
714 if (to_free) {
715 chunk->map_used -= to_free;
716 memmove(p + 1, p + 1 + to_free,
717 (chunk->map_used - i) * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900718 }
719
Al Viro723ad1d2014-03-06 21:13:18 -0500720 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900721 pcpu_chunk_relocate(chunk, oslot);
722}
723
Tejun Heo60810892010-04-09 18:57:01 +0900724static struct pcpu_chunk *pcpu_alloc_chunk(void)
725{
726 struct pcpu_chunk *chunk;
727
Bob Liu90459ce02011-08-04 11:02:33 +0200728 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900729 if (!chunk)
730 return NULL;
731
Bob Liu90459ce02011-08-04 11:02:33 +0200732 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
733 sizeof(chunk->map[0]));
Tejun Heo60810892010-04-09 18:57:01 +0900734 if (!chunk->map) {
Jianyu Zhan5a838c32014-04-14 13:47:40 +0800735 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900736 return NULL;
737 }
738
739 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500740 chunk->map[0] = 0;
741 chunk->map[1] = pcpu_unit_size | 1;
742 chunk->map_used = 1;
Tejun Heo60810892010-04-09 18:57:01 +0900743
744 INIT_LIST_HEAD(&chunk->list);
Tejun Heo9c824b62014-09-02 14:46:05 -0400745 INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn);
Tejun Heo60810892010-04-09 18:57:01 +0900746 chunk->free_size = pcpu_unit_size;
747 chunk->contig_hint = pcpu_unit_size;
748
749 return chunk;
750}
751
752static void pcpu_free_chunk(struct pcpu_chunk *chunk)
753{
754 if (!chunk)
755 return;
756 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Joonsoo Kimb4916cb2012-10-29 22:59:58 +0900757 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900758}
759
Tejun Heob539b872014-09-02 14:46:05 -0400760/**
761 * pcpu_chunk_populated - post-population bookkeeping
762 * @chunk: pcpu_chunk which got populated
763 * @page_start: the start page
764 * @page_end: the end page
765 *
766 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
767 * the bookkeeping information accordingly. Must be called after each
768 * successful population.
769 */
770static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
771 int page_start, int page_end)
772{
773 int nr = page_end - page_start;
774
775 lockdep_assert_held(&pcpu_lock);
776
777 bitmap_set(chunk->populated, page_start, nr);
778 chunk->nr_populated += nr;
779 pcpu_nr_empty_pop_pages += nr;
780}
781
782/**
783 * pcpu_chunk_depopulated - post-depopulation bookkeeping
784 * @chunk: pcpu_chunk which got depopulated
785 * @page_start: the start page
786 * @page_end: the end page
787 *
788 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
789 * Update the bookkeeping information accordingly. Must be called after
790 * each successful depopulation.
791 */
792static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
793 int page_start, int page_end)
794{
795 int nr = page_end - page_start;
796
797 lockdep_assert_held(&pcpu_lock);
798
799 bitmap_clear(chunk->populated, page_start, nr);
800 chunk->nr_populated -= nr;
801 pcpu_nr_empty_pop_pages -= nr;
802}
803
Tejun Heo9f645532010-04-09 18:57:01 +0900804/*
805 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900806 *
Tejun Heo9f645532010-04-09 18:57:01 +0900807 * To allow different implementations, chunk alloc/free and
808 * [de]population are implemented in a separate file which is pulled
809 * into this file and compiled together. The following functions
810 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900811 *
Tejun Heo9f645532010-04-09 18:57:01 +0900812 * pcpu_populate_chunk - populate the specified range of a chunk
813 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
814 * pcpu_create_chunk - create a new chunk
815 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
816 * pcpu_addr_to_page - translate address to physical address
817 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900818 */
Tejun Heo9f645532010-04-09 18:57:01 +0900819static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
820static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
821static struct pcpu_chunk *pcpu_create_chunk(void);
822static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
823static struct page *pcpu_addr_to_page(void *addr);
824static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900825
Tejun Heob0c97782010-04-09 18:57:01 +0900826#ifdef CONFIG_NEED_PER_CPU_KM
827#include "percpu-km.c"
828#else
Tejun Heo9f645532010-04-09 18:57:01 +0900829#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900830#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900831
832/**
Tejun Heo88999a82010-04-09 18:57:01 +0900833 * pcpu_chunk_addr_search - determine chunk containing specified address
834 * @addr: address for which the chunk needs to be determined.
835 *
836 * RETURNS:
837 * The address of the found chunk.
838 */
839static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
840{
841 /* is it in the first chunk? */
842 if (pcpu_addr_in_first_chunk(addr)) {
843 /* is it in the reserved area? */
844 if (pcpu_addr_in_reserved_chunk(addr))
845 return pcpu_reserved_chunk;
846 return pcpu_first_chunk;
847 }
848
849 /*
850 * The address is relative to unit0 which might be unused and
851 * thus unmapped. Offset the address to the unit space of the
852 * current processor before looking it up in the vmalloc
853 * space. Note that any possible cpu id can be used here, so
854 * there's no need to worry about preemption or cpu hotplug.
855 */
856 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900857 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900858}
859
860/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900861 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900862 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900863 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900864 * @reserved: allocate from the reserved chunk if available
Tejun Heo5835d962014-09-02 14:46:04 -0400865 * @gfp: allocation flags
Tejun Heofbf59bc2009-02-20 16:29:08 +0900866 *
Tejun Heo5835d962014-09-02 14:46:04 -0400867 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
868 * contain %GFP_KERNEL, the allocation is atomic.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900869 *
870 * RETURNS:
871 * Percpu pointer to the allocated area on success, NULL on failure.
872 */
Tejun Heo5835d962014-09-02 14:46:04 -0400873static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
874 gfp_t gfp)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900875{
Tejun Heof2badb02009-09-29 09:17:58 +0900876 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900877 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900878 const char *err;
Tejun Heo6ae833c2014-10-08 12:01:52 -0400879 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
Tejun Heob539b872014-09-02 14:46:05 -0400880 int occ_pages = 0;
Tejun Heob38d08f2014-09-02 14:46:02 -0400881 int slot, off, new_alloc, cpu, ret;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900882 unsigned long flags;
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100883 void __percpu *ptr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900884
Al Viro723ad1d2014-03-06 21:13:18 -0500885 /*
886 * We want the lowest bit of offset available for in-use/free
Viro2f69fa82014-03-17 16:01:27 -0400887 * indicator, so force >= 16bit alignment and make size even.
Al Viro723ad1d2014-03-06 21:13:18 -0500888 */
889 if (unlikely(align < 2))
890 align = 2;
891
Christoph Lameterfb009e32014-06-19 09:59:18 -0500892 size = ALIGN(size, 2);
Viro2f69fa82014-03-17 16:01:27 -0400893
Tejun Heo8d408b42009-02-24 11:57:21 +0900894 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900895 WARN(true, "illegal size (%zu) or align (%zu) for "
896 "percpu allocation\n", size, align);
897 return NULL;
898 }
899
Jiri Kosina403a91b2009-10-29 00:25:59 +0900900 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900901
Tejun Heoedcb4632009-03-06 14:33:59 +0900902 /* serve reserved allocations from the reserved chunk if available */
903 if (reserved && pcpu_reserved_chunk) {
904 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900905
906 if (size > chunk->contig_hint) {
907 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900908 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900909 }
Tejun Heo833af842009-11-11 15:35:18 +0900910
Tejun Heo9c824b62014-09-02 14:46:05 -0400911 while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
Tejun Heo833af842009-11-11 15:35:18 +0900912 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heo5835d962014-09-02 14:46:04 -0400913 if (is_atomic ||
914 pcpu_extend_area_map(chunk, new_alloc) < 0) {
Tejun Heo833af842009-11-11 15:35:18 +0900915 err = "failed to extend area map of reserved chunk";
Tejun Heob38d08f2014-09-02 14:46:02 -0400916 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900917 }
918 spin_lock_irqsave(&pcpu_lock, flags);
919 }
920
Tejun Heob539b872014-09-02 14:46:05 -0400921 off = pcpu_alloc_area(chunk, size, align, is_atomic,
922 &occ_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +0900923 if (off >= 0)
924 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900925
Tejun Heof2badb02009-09-29 09:17:58 +0900926 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900927 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900928 }
929
Tejun Heoccea34b2009-03-07 00:44:13 +0900930restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900931 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900932 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
933 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
934 if (size > chunk->contig_hint)
935 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900936
Tejun Heo9c824b62014-09-02 14:46:05 -0400937 new_alloc = pcpu_need_to_extend(chunk, is_atomic);
Tejun Heo833af842009-11-11 15:35:18 +0900938 if (new_alloc) {
Tejun Heo5835d962014-09-02 14:46:04 -0400939 if (is_atomic)
940 continue;
Tejun Heo833af842009-11-11 15:35:18 +0900941 spin_unlock_irqrestore(&pcpu_lock, flags);
942 if (pcpu_extend_area_map(chunk,
943 new_alloc) < 0) {
944 err = "failed to extend area map";
Tejun Heob38d08f2014-09-02 14:46:02 -0400945 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900946 }
947 spin_lock_irqsave(&pcpu_lock, flags);
948 /*
949 * pcpu_lock has been dropped, need to
950 * restart cpu_slot list walking.
951 */
952 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900953 }
954
Tejun Heob539b872014-09-02 14:46:05 -0400955 off = pcpu_alloc_area(chunk, size, align, is_atomic,
956 &occ_pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900957 if (off >= 0)
958 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900959 }
960 }
961
Jiri Kosina403a91b2009-10-29 00:25:59 +0900962 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900963
Tejun Heob38d08f2014-09-02 14:46:02 -0400964 /*
965 * No space left. Create a new chunk. We don't want multiple
966 * tasks to create chunks simultaneously. Serialize and create iff
967 * there's still no empty chunk after grabbing the mutex.
968 */
Tejun Heo5835d962014-09-02 14:46:04 -0400969 if (is_atomic)
970 goto fail;
971
Tejun Heob38d08f2014-09-02 14:46:02 -0400972 mutex_lock(&pcpu_alloc_mutex);
973
974 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
975 chunk = pcpu_create_chunk();
976 if (!chunk) {
Tejun Heo23cb8982014-09-09 08:02:45 +0900977 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heob38d08f2014-09-02 14:46:02 -0400978 err = "failed to allocate new chunk";
979 goto fail;
980 }
981
982 spin_lock_irqsave(&pcpu_lock, flags);
983 pcpu_chunk_relocate(chunk, -1);
984 } else {
985 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heof2badb02009-09-29 09:17:58 +0900986 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900987
Tejun Heob38d08f2014-09-02 14:46:02 -0400988 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heoccea34b2009-03-07 00:44:13 +0900989 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900990
991area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900992 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900993
Tejun Heodca49642014-09-02 14:46:01 -0400994 /* populate if not all pages are already there */
Tejun Heo5835d962014-09-02 14:46:04 -0400995 if (!is_atomic) {
Tejun Heoe04d3202014-09-02 14:46:04 -0400996 int page_start, page_end, rs, re;
Tejun Heodca49642014-09-02 14:46:01 -0400997
Tejun Heoe04d3202014-09-02 14:46:04 -0400998 mutex_lock(&pcpu_alloc_mutex);
Tejun Heodca49642014-09-02 14:46:01 -0400999
Tejun Heoe04d3202014-09-02 14:46:04 -04001000 page_start = PFN_DOWN(off);
1001 page_end = PFN_UP(off + size);
Tejun Heob38d08f2014-09-02 14:46:02 -04001002
Tejun Heoe04d3202014-09-02 14:46:04 -04001003 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
1004 WARN_ON(chunk->immutable);
1005
1006 ret = pcpu_populate_chunk(chunk, rs, re);
1007
1008 spin_lock_irqsave(&pcpu_lock, flags);
1009 if (ret) {
1010 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heob539b872014-09-02 14:46:05 -04001011 pcpu_free_area(chunk, off, &occ_pages);
Tejun Heoe04d3202014-09-02 14:46:04 -04001012 err = "failed to populate";
1013 goto fail_unlock;
1014 }
Tejun Heob539b872014-09-02 14:46:05 -04001015 pcpu_chunk_populated(chunk, rs, re);
Tejun Heoe04d3202014-09-02 14:46:04 -04001016 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heodca49642014-09-02 14:46:01 -04001017 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001018
Tejun Heoe04d3202014-09-02 14:46:04 -04001019 mutex_unlock(&pcpu_alloc_mutex);
1020 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001021
Tejun Heob539b872014-09-02 14:46:05 -04001022 if (chunk != pcpu_reserved_chunk)
1023 pcpu_nr_empty_pop_pages -= occ_pages;
1024
Tejun Heo1a4d7602014-09-02 14:46:05 -04001025 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1026 pcpu_schedule_balance_work();
1027
Tejun Heodca49642014-09-02 14:46:01 -04001028 /* clear the areas and return address relative to base address */
1029 for_each_possible_cpu(cpu)
1030 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1031
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001032 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1033 kmemleak_alloc_percpu(ptr, size);
1034 return ptr;
Tejun Heoccea34b2009-03-07 00:44:13 +09001035
1036fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001037 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heob38d08f2014-09-02 14:46:02 -04001038fail:
Tejun Heo5835d962014-09-02 14:46:04 -04001039 if (!is_atomic && warn_limit) {
1040 pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1041 size, align, is_atomic, err);
Tejun Heof2badb02009-09-29 09:17:58 +09001042 dump_stack();
1043 if (!--warn_limit)
1044 pr_info("PERCPU: limit reached, disable warning\n");
1045 }
Tejun Heo1a4d7602014-09-02 14:46:05 -04001046 if (is_atomic) {
1047 /* see the flag handling in pcpu_blance_workfn() */
1048 pcpu_atomic_alloc_failed = true;
1049 pcpu_schedule_balance_work();
1050 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001051 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001052}
Tejun Heoedcb4632009-03-06 14:33:59 +09001053
1054/**
Tejun Heo5835d962014-09-02 14:46:04 -04001055 * __alloc_percpu_gfp - allocate dynamic percpu area
Tejun Heoedcb4632009-03-06 14:33:59 +09001056 * @size: size of area to allocate in bytes
1057 * @align: alignment of area (max PAGE_SIZE)
Tejun Heo5835d962014-09-02 14:46:04 -04001058 * @gfp: allocation flags
Tejun Heoedcb4632009-03-06 14:33:59 +09001059 *
Tejun Heo5835d962014-09-02 14:46:04 -04001060 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1061 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1062 * be called from any context but is a lot more likely to fail.
Tejun Heoccea34b2009-03-07 00:44:13 +09001063 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001064 * RETURNS:
1065 * Percpu pointer to the allocated area on success, NULL on failure.
1066 */
Tejun Heo5835d962014-09-02 14:46:04 -04001067void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1068{
1069 return pcpu_alloc(size, align, false, gfp);
1070}
1071EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1072
1073/**
1074 * __alloc_percpu - allocate dynamic percpu area
1075 * @size: size of area to allocate in bytes
1076 * @align: alignment of area (max PAGE_SIZE)
1077 *
1078 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1079 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001080void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001081{
Tejun Heo5835d962014-09-02 14:46:04 -04001082 return pcpu_alloc(size, align, false, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001083}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001084EXPORT_SYMBOL_GPL(__alloc_percpu);
1085
Tejun Heoedcb4632009-03-06 14:33:59 +09001086/**
1087 * __alloc_reserved_percpu - allocate reserved percpu area
1088 * @size: size of area to allocate in bytes
1089 * @align: alignment of area (max PAGE_SIZE)
1090 *
Tejun Heo9329ba92010-09-10 11:01:56 +02001091 * Allocate zero-filled percpu area of @size bytes aligned at @align
1092 * from reserved percpu area if arch has set it up; otherwise,
1093 * allocation is served from the same dynamic area. Might sleep.
1094 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +09001095 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001096 * CONTEXT:
1097 * Does GFP_KERNEL allocation.
1098 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001099 * RETURNS:
1100 * Percpu pointer to the allocated area on success, NULL on failure.
1101 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001102void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001103{
Tejun Heo5835d962014-09-02 14:46:04 -04001104 return pcpu_alloc(size, align, true, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001105}
1106
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001107/**
Tejun Heo1a4d7602014-09-02 14:46:05 -04001108 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001109 * @work: unused
1110 *
1111 * Reclaim all fully free chunks except for the first one.
1112 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001113static void pcpu_balance_workfn(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001114{
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001115 LIST_HEAD(to_free);
1116 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001117 struct pcpu_chunk *chunk, *next;
Tejun Heo1a4d7602014-09-02 14:46:05 -04001118 int slot, nr_to_pop, ret;
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001119
Tejun Heo1a4d7602014-09-02 14:46:05 -04001120 /*
1121 * There's no reason to keep around multiple unused chunks and VM
1122 * areas can be scarce. Destroy all free chunks except for one.
1123 */
Tejun Heoccea34b2009-03-07 00:44:13 +09001124 mutex_lock(&pcpu_alloc_mutex);
1125 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001126
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001127 list_for_each_entry_safe(chunk, next, free_head, list) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001128 WARN_ON(chunk->immutable);
1129
1130 /* spare the first one */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001131 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001132 continue;
1133
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001134 list_move(&chunk->list, &to_free);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001135 }
1136
Tejun Heoccea34b2009-03-07 00:44:13 +09001137 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001138
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001139 list_for_each_entry_safe(chunk, next, &to_free, list) {
Tejun Heoa93ace42014-09-02 14:46:02 -04001140 int rs, re;
Tejun Heodca49642014-09-02 14:46:01 -04001141
Tejun Heoa93ace42014-09-02 14:46:02 -04001142 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1143 pcpu_depopulate_chunk(chunk, rs, re);
Tejun Heob539b872014-09-02 14:46:05 -04001144 spin_lock_irq(&pcpu_lock);
1145 pcpu_chunk_depopulated(chunk, rs, re);
1146 spin_unlock_irq(&pcpu_lock);
Tejun Heoa93ace42014-09-02 14:46:02 -04001147 }
Tejun Heo60810892010-04-09 18:57:01 +09001148 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001149 }
Tejun Heo971f3912009-08-14 15:00:49 +09001150
Tejun Heo1a4d7602014-09-02 14:46:05 -04001151 /*
1152 * Ensure there are certain number of free populated pages for
1153 * atomic allocs. Fill up from the most packed so that atomic
1154 * allocs don't increase fragmentation. If atomic allocation
1155 * failed previously, always populate the maximum amount. This
1156 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1157 * failing indefinitely; however, large atomic allocs are not
1158 * something we support properly and can be highly unreliable and
1159 * inefficient.
1160 */
1161retry_pop:
1162 if (pcpu_atomic_alloc_failed) {
1163 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1164 /* best effort anyway, don't worry about synchronization */
1165 pcpu_atomic_alloc_failed = false;
1166 } else {
1167 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1168 pcpu_nr_empty_pop_pages,
1169 0, PCPU_EMPTY_POP_PAGES_HIGH);
1170 }
1171
1172 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1173 int nr_unpop = 0, rs, re;
1174
1175 if (!nr_to_pop)
1176 break;
1177
1178 spin_lock_irq(&pcpu_lock);
1179 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1180 nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1181 if (nr_unpop)
1182 break;
1183 }
1184 spin_unlock_irq(&pcpu_lock);
1185
1186 if (!nr_unpop)
1187 continue;
1188
1189 /* @chunk can't go away while pcpu_alloc_mutex is held */
1190 pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1191 int nr = min(re - rs, nr_to_pop);
1192
1193 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1194 if (!ret) {
1195 nr_to_pop -= nr;
1196 spin_lock_irq(&pcpu_lock);
1197 pcpu_chunk_populated(chunk, rs, rs + nr);
1198 spin_unlock_irq(&pcpu_lock);
1199 } else {
1200 nr_to_pop = 0;
1201 }
1202
1203 if (!nr_to_pop)
1204 break;
1205 }
1206 }
1207
1208 if (nr_to_pop) {
1209 /* ran out of chunks to populate, create a new one and retry */
1210 chunk = pcpu_create_chunk();
1211 if (chunk) {
1212 spin_lock_irq(&pcpu_lock);
1213 pcpu_chunk_relocate(chunk, -1);
1214 spin_unlock_irq(&pcpu_lock);
1215 goto retry_pop;
1216 }
1217 }
1218
Tejun Heo971f3912009-08-14 15:00:49 +09001219 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001220}
1221
1222/**
1223 * free_percpu - free percpu area
1224 * @ptr: pointer to area to free
1225 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001226 * Free percpu area @ptr.
1227 *
1228 * CONTEXT:
1229 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001230 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001231void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001232{
Andrew Morton129182e2010-01-08 14:42:39 -08001233 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001234 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001235 unsigned long flags;
Tejun Heob539b872014-09-02 14:46:05 -04001236 int off, occ_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001237
1238 if (!ptr)
1239 return;
1240
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001241 kmemleak_free_percpu(ptr);
1242
Andrew Morton129182e2010-01-08 14:42:39 -08001243 addr = __pcpu_ptr_to_addr(ptr);
1244
Tejun Heoccea34b2009-03-07 00:44:13 +09001245 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001246
1247 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001248 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001249
Tejun Heob539b872014-09-02 14:46:05 -04001250 pcpu_free_area(chunk, off, &occ_pages);
1251
1252 if (chunk != pcpu_reserved_chunk)
1253 pcpu_nr_empty_pop_pages += occ_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001254
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001255 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001256 if (chunk->free_size == pcpu_unit_size) {
1257 struct pcpu_chunk *pos;
1258
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001259 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001260 if (pos != chunk) {
Tejun Heo1a4d7602014-09-02 14:46:05 -04001261 pcpu_schedule_balance_work();
Tejun Heofbf59bc2009-02-20 16:29:08 +09001262 break;
1263 }
1264 }
1265
Tejun Heoccea34b2009-03-07 00:44:13 +09001266 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001267}
1268EXPORT_SYMBOL_GPL(free_percpu);
1269
Vivek Goyal3b034b02009-11-24 15:50:03 +09001270/**
Tejun Heo10fad5e2010-03-10 18:57:54 +09001271 * is_kernel_percpu_address - test whether address is from static percpu area
1272 * @addr: address to test
1273 *
1274 * Test whether @addr belongs to in-kernel static percpu area. Module
1275 * static percpu areas are not considered. For those, use
1276 * is_module_percpu_address().
1277 *
1278 * RETURNS:
1279 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1280 */
1281bool is_kernel_percpu_address(unsigned long addr)
1282{
Tejun Heobbddff02010-09-03 18:22:48 +02001283#ifdef CONFIG_SMP
Tejun Heo10fad5e2010-03-10 18:57:54 +09001284 const size_t static_size = __per_cpu_end - __per_cpu_start;
1285 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1286 unsigned int cpu;
1287
1288 for_each_possible_cpu(cpu) {
1289 void *start = per_cpu_ptr(base, cpu);
1290
1291 if ((void *)addr >= start && (void *)addr < start + static_size)
1292 return true;
1293 }
Tejun Heobbddff02010-09-03 18:22:48 +02001294#endif
1295 /* on UP, can't distinguish from other static vars, always false */
Tejun Heo10fad5e2010-03-10 18:57:54 +09001296 return false;
1297}
1298
1299/**
Vivek Goyal3b034b02009-11-24 15:50:03 +09001300 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1301 * @addr: the address to be converted to physical address
1302 *
1303 * Given @addr which is dereferenceable address obtained via one of
1304 * percpu access macros, this function translates it into its physical
1305 * address. The caller is responsible for ensuring @addr stays valid
1306 * until this function finishes.
1307 *
Dave Young67589c72011-11-23 08:20:53 -08001308 * percpu allocator has special setup for the first chunk, which currently
1309 * supports either embedding in linear address space or vmalloc mapping,
1310 * and, from the second one, the backing allocator (currently either vm or
1311 * km) provides translation.
1312 *
1313 * The addr can be tranlated simply without checking if it falls into the
1314 * first chunk. But the current code reflects better how percpu allocator
1315 * actually works, and the verification can discover both bugs in percpu
1316 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1317 * code.
1318 *
Vivek Goyal3b034b02009-11-24 15:50:03 +09001319 * RETURNS:
1320 * The physical address for @addr.
1321 */
1322phys_addr_t per_cpu_ptr_to_phys(void *addr)
1323{
Tejun Heo9983b6f02010-06-18 11:44:31 +02001324 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1325 bool in_first_chunk = false;
Tejun Heoa855b842011-11-18 10:55:35 -08001326 unsigned long first_low, first_high;
Tejun Heo9983b6f02010-06-18 11:44:31 +02001327 unsigned int cpu;
1328
1329 /*
Tejun Heoa855b842011-11-18 10:55:35 -08001330 * The following test on unit_low/high isn't strictly
Tejun Heo9983b6f02010-06-18 11:44:31 +02001331 * necessary but will speed up lookups of addresses which
1332 * aren't in the first chunk.
1333 */
Tejun Heoa855b842011-11-18 10:55:35 -08001334 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1335 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1336 pcpu_unit_pages);
1337 if ((unsigned long)addr >= first_low &&
1338 (unsigned long)addr < first_high) {
Tejun Heo9983b6f02010-06-18 11:44:31 +02001339 for_each_possible_cpu(cpu) {
1340 void *start = per_cpu_ptr(base, cpu);
1341
1342 if (addr >= start && addr < start + pcpu_unit_size) {
1343 in_first_chunk = true;
1344 break;
1345 }
1346 }
1347 }
1348
1349 if (in_first_chunk) {
David Howellseac522e2011-03-28 12:53:29 +01001350 if (!is_vmalloc_addr(addr))
Tejun Heo020ec652010-04-09 18:57:00 +09001351 return __pa(addr);
1352 else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001353 return page_to_phys(vmalloc_to_page(addr)) +
1354 offset_in_page(addr);
Tejun Heo020ec652010-04-09 18:57:00 +09001355 } else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001356 return page_to_phys(pcpu_addr_to_page(addr)) +
1357 offset_in_page(addr);
Vivek Goyal3b034b02009-11-24 15:50:03 +09001358}
1359
Tejun Heofbf59bc2009-02-20 16:29:08 +09001360/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001361 * pcpu_alloc_alloc_info - allocate percpu allocation info
1362 * @nr_groups: the number of groups
1363 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001364 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001365 * Allocate ai which is large enough for @nr_groups groups containing
1366 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1367 * cpu_map array which is long enough for @nr_units and filled with
1368 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1369 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001370 *
1371 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001372 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1373 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001374 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001375struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1376 int nr_units)
1377{
1378 struct pcpu_alloc_info *ai;
1379 size_t base_size, ai_size;
1380 void *ptr;
1381 int unit;
1382
1383 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1384 __alignof__(ai->groups[0].cpu_map[0]));
1385 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1386
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001387 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001388 if (!ptr)
1389 return NULL;
1390 ai = ptr;
1391 ptr += base_size;
1392
1393 ai->groups[0].cpu_map = ptr;
1394
1395 for (unit = 0; unit < nr_units; unit++)
1396 ai->groups[0].cpu_map[unit] = NR_CPUS;
1397
1398 ai->nr_groups = nr_groups;
1399 ai->__ai_size = PFN_ALIGN(ai_size);
1400
1401 return ai;
1402}
1403
1404/**
1405 * pcpu_free_alloc_info - free percpu allocation info
1406 * @ai: pcpu_alloc_info to free
1407 *
1408 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1409 */
1410void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1411{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001412 memblock_free_early(__pa(ai), ai->__ai_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001413}
1414
1415/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001416 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1417 * @lvl: loglevel
1418 * @ai: allocation info to dump
1419 *
1420 * Print out information about @ai using loglevel @lvl.
1421 */
1422static void pcpu_dump_alloc_info(const char *lvl,
1423 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001424{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001425 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001426 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001427 int alloc = 0, alloc_end = 0;
1428 int group, v;
1429 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001430
Tejun Heofd1e8a12009-08-14 15:00:51 +09001431 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001432 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001433 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001434
Tejun Heofd1e8a12009-08-14 15:00:51 +09001435 v = num_possible_cpus();
1436 while (v /= 10)
1437 cpu_width++;
1438 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001439
Tejun Heofd1e8a12009-08-14 15:00:51 +09001440 upa = ai->alloc_size / ai->unit_size;
1441 width = upa * (cpu_width + 1) + group_width + 3;
1442 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001443
Tejun Heofd1e8a12009-08-14 15:00:51 +09001444 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1445 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1446 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1447
1448 for (group = 0; group < ai->nr_groups; group++) {
1449 const struct pcpu_group_info *gi = &ai->groups[group];
1450 int unit = 0, unit_end = 0;
1451
1452 BUG_ON(gi->nr_units % upa);
1453 for (alloc_end += gi->nr_units / upa;
1454 alloc < alloc_end; alloc++) {
1455 if (!(alloc % apl)) {
Tejun Heocb129822012-03-29 09:45:58 -07001456 printk(KERN_CONT "\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001457 printk("%spcpu-alloc: ", lvl);
1458 }
Tejun Heocb129822012-03-29 09:45:58 -07001459 printk(KERN_CONT "[%0*d] ", group_width, group);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001460
1461 for (unit_end += upa; unit < unit_end; unit++)
1462 if (gi->cpu_map[unit] != NR_CPUS)
Tejun Heocb129822012-03-29 09:45:58 -07001463 printk(KERN_CONT "%0*d ", cpu_width,
Tejun Heofd1e8a12009-08-14 15:00:51 +09001464 gi->cpu_map[unit]);
1465 else
Tejun Heocb129822012-03-29 09:45:58 -07001466 printk(KERN_CONT "%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001467 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001468 }
Tejun Heocb129822012-03-29 09:45:58 -07001469 printk(KERN_CONT "\n");
Tejun Heo033e48f2009-08-14 15:00:51 +09001470}
Tejun Heo033e48f2009-08-14 15:00:51 +09001471
Tejun Heofbf59bc2009-02-20 16:29:08 +09001472/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001473 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001474 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001475 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001476 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001477 * Initialize the first percpu chunk which contains the kernel static
1478 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001479 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001480 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001481 * @ai contains all information necessary to initialize the first
1482 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001483 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001484 * @ai->static_size is the size of static percpu area.
1485 *
1486 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001487 * reserve after the static area in the first chunk. This reserves
1488 * the first chunk such that it's available only through reserved
1489 * percpu allocation. This is primarily used to serve module percpu
1490 * static areas on architectures where the addressing model has
1491 * limited offset range for symbol relocations to guarantee module
1492 * percpu symbols fall inside the relocatable range.
1493 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001494 * @ai->dyn_size determines the number of bytes available for dynamic
1495 * allocation in the first chunk. The area between @ai->static_size +
1496 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001497 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001498 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1499 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1500 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001501 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001502 * @ai->atom_size is the allocation atom size and used as alignment
1503 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001504 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001505 * @ai->alloc_size is the allocation size and always multiple of
1506 * @ai->atom_size. This is larger than @ai->atom_size if
1507 * @ai->unit_size is larger than @ai->atom_size.
1508 *
1509 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1510 * percpu areas. Units which should be colocated are put into the
1511 * same group. Dynamic VM areas will be allocated according to these
1512 * groupings. If @ai->nr_groups is zero, a single group containing
1513 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001514 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001515 * The caller should have mapped the first chunk at @base_addr and
1516 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001517 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001518 * If the first chunk ends up with both reserved and dynamic areas, it
1519 * is served by two chunks - one to serve the core static and reserved
1520 * areas and the other for the dynamic area. They share the same vm
1521 * and page map but uses different area allocation map to stay away
1522 * from each other. The latter chunk is circulated in the chunk slots
1523 * and available for dynamic allocation like any other chunks.
1524 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001525 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001526 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001527 */
Tejun Heofb435d52009-08-14 15:00:51 +09001528int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1529 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001530{
Tejun Heo635b75f2009-09-24 09:43:11 +09001531 static char cpus_buf[4096] __initdata;
Tejun Heo099a19d2010-06-27 18:50:00 +02001532 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1533 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001534 size_t dyn_size = ai->dyn_size;
1535 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001536 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001537 unsigned long *group_offsets;
1538 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001539 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001540 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001541 int *unit_map;
1542 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001543
Tejun Heo635b75f2009-09-24 09:43:11 +09001544 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1545
1546#define PCPU_SETUP_BUG_ON(cond) do { \
1547 if (unlikely(cond)) { \
1548 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1549 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1550 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1551 BUG(); \
1552 } \
1553} while (0)
1554
Tejun Heo2f39e632009-07-04 08:11:00 +09001555 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001556 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02001557#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09001558 PCPU_SETUP_BUG_ON(!ai->static_size);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001559 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
Tejun Heobbddff02010-09-03 18:22:48 +02001560#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09001561 PCPU_SETUP_BUG_ON(!base_addr);
Tejun Heo0415b00d12011-03-24 18:50:09 +01001562 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
Tejun Heo635b75f2009-09-24 09:43:11 +09001563 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1564 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1565 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001566 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001567 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001568
Tejun Heo65632972009-08-14 15:00:52 +09001569 /* process group information and build config tables accordingly */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001570 group_offsets = memblock_virt_alloc(ai->nr_groups *
1571 sizeof(group_offsets[0]), 0);
1572 group_sizes = memblock_virt_alloc(ai->nr_groups *
1573 sizeof(group_sizes[0]), 0);
1574 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1575 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
Tejun Heo2f39e632009-07-04 08:11:00 +09001576
Tejun Heofd1e8a12009-08-14 15:00:51 +09001577 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001578 unit_map[cpu] = UINT_MAX;
Tejun Heoa855b842011-11-18 10:55:35 -08001579
1580 pcpu_low_unit_cpu = NR_CPUS;
1581 pcpu_high_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001582
Tejun Heofd1e8a12009-08-14 15:00:51 +09001583 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1584 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001585
Tejun Heo65632972009-08-14 15:00:52 +09001586 group_offsets[group] = gi->base_offset;
1587 group_sizes[group] = gi->nr_units * ai->unit_size;
1588
Tejun Heofd1e8a12009-08-14 15:00:51 +09001589 for (i = 0; i < gi->nr_units; i++) {
1590 cpu = gi->cpu_map[i];
1591 if (cpu == NR_CPUS)
1592 continue;
1593
Dan Carpenter9f295662014-10-29 11:45:04 +03001594 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
Tejun Heo635b75f2009-09-24 09:43:11 +09001595 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1596 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001597
1598 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001599 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1600
Tejun Heoa855b842011-11-18 10:55:35 -08001601 /* determine low/high unit_cpu */
1602 if (pcpu_low_unit_cpu == NR_CPUS ||
1603 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1604 pcpu_low_unit_cpu = cpu;
1605 if (pcpu_high_unit_cpu == NR_CPUS ||
1606 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1607 pcpu_high_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001608 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001609 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001610 pcpu_nr_units = unit;
1611
1612 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001613 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1614
1615 /* we're done parsing the input, undefine BUG macro and dump config */
1616#undef PCPU_SETUP_BUG_ON
Tejun Heobcbea792010-12-22 14:19:14 +01001617 pcpu_dump_alloc_info(KERN_DEBUG, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001618
Tejun Heo65632972009-08-14 15:00:52 +09001619 pcpu_nr_groups = ai->nr_groups;
1620 pcpu_group_offsets = group_offsets;
1621 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001622 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001623 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001624
1625 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001626 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001627 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001628 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001629 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1630 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001631
Tejun Heod9b55ee2009-02-24 11:57:21 +09001632 /*
1633 * Allocate chunk slots. The additional last slot is for
1634 * empty chunks.
1635 */
1636 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001637 pcpu_slot = memblock_virt_alloc(
1638 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001639 for (i = 0; i < pcpu_nr_slots; i++)
1640 INIT_LIST_HEAD(&pcpu_slot[i]);
1641
Tejun Heoedcb4632009-03-06 14:33:59 +09001642 /*
1643 * Initialize static chunk. If reserved_size is zero, the
1644 * static chunk covers static area + dynamic allocation area
1645 * in the first chunk. If reserved_size is not zero, it
1646 * covers static area + reserved area (mostly used for module
1647 * static percpu allocation).
1648 */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001649 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heo2441d152009-03-06 14:33:59 +09001650 INIT_LIST_HEAD(&schunk->list);
Tejun Heo9c824b62014-09-02 14:46:05 -04001651 INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn);
Tejun Heobba174f2009-08-14 15:00:51 +09001652 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001653 schunk->map = smap;
1654 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001655 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001656 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heob539b872014-09-02 14:46:05 -04001657 schunk->nr_populated = pcpu_unit_pages;
Tejun Heoedcb4632009-03-06 14:33:59 +09001658
Tejun Heofd1e8a12009-08-14 15:00:51 +09001659 if (ai->reserved_size) {
1660 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001661 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001662 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001663 } else {
1664 schunk->free_size = dyn_size;
1665 dyn_size = 0; /* dynamic area covered */
1666 }
Tejun Heo2441d152009-03-06 14:33:59 +09001667 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001668
Al Viro723ad1d2014-03-06 21:13:18 -05001669 schunk->map[0] = 1;
1670 schunk->map[1] = ai->static_size;
1671 schunk->map_used = 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001672 if (schunk->free_size)
Al Viro723ad1d2014-03-06 21:13:18 -05001673 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1674 else
1675 schunk->map[1] |= 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001676
Tejun Heoedcb4632009-03-06 14:33:59 +09001677 /* init dynamic chunk if necessary */
1678 if (dyn_size) {
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001679 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001680 INIT_LIST_HEAD(&dchunk->list);
Tejun Heo9c824b62014-09-02 14:46:05 -04001681 INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn);
Tejun Heobba174f2009-08-14 15:00:51 +09001682 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001683 dchunk->map = dmap;
1684 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001685 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001686 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heob539b872014-09-02 14:46:05 -04001687 dchunk->nr_populated = pcpu_unit_pages;
Tejun Heoedcb4632009-03-06 14:33:59 +09001688
1689 dchunk->contig_hint = dchunk->free_size = dyn_size;
Al Viro723ad1d2014-03-06 21:13:18 -05001690 dchunk->map[0] = 1;
1691 dchunk->map[1] = pcpu_reserved_chunk_limit;
1692 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1693 dchunk->map_used = 2;
Tejun Heoedcb4632009-03-06 14:33:59 +09001694 }
1695
Tejun Heo2441d152009-03-06 14:33:59 +09001696 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001697 pcpu_first_chunk = dchunk ?: schunk;
Tejun Heob539b872014-09-02 14:46:05 -04001698 pcpu_nr_empty_pop_pages +=
1699 pcpu_count_occupied_pages(pcpu_first_chunk, 1);
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001700 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001701
1702 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001703 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001704 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001705}
Tejun Heo66c3a752009-03-10 16:27:48 +09001706
Tejun Heobbddff02010-09-03 18:22:48 +02001707#ifdef CONFIG_SMP
1708
Andi Kleen17f36092012-10-04 17:12:07 -07001709const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
Tejun Heof58dc012009-08-14 15:00:50 +09001710 [PCPU_FC_AUTO] = "auto",
1711 [PCPU_FC_EMBED] = "embed",
1712 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001713};
Tejun Heo66c3a752009-03-10 16:27:48 +09001714
Tejun Heof58dc012009-08-14 15:00:50 +09001715enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1716
1717static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001718{
Cyrill Gorcunov5479c782012-11-25 01:17:13 +04001719 if (!str)
1720 return -EINVAL;
1721
Tejun Heof58dc012009-08-14 15:00:50 +09001722 if (0)
1723 /* nada */;
1724#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1725 else if (!strcmp(str, "embed"))
1726 pcpu_chosen_fc = PCPU_FC_EMBED;
1727#endif
1728#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1729 else if (!strcmp(str, "page"))
1730 pcpu_chosen_fc = PCPU_FC_PAGE;
1731#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001732 else
1733 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001734
Tejun Heof58dc012009-08-14 15:00:50 +09001735 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001736}
Tejun Heof58dc012009-08-14 15:00:50 +09001737early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001738
Tejun Heo3c9a0242010-09-09 18:00:15 +02001739/*
1740 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1741 * Build it if needed by the arch config or the generic setup is going
1742 * to be used.
1743 */
Tejun Heo08fc4582009-08-14 15:00:49 +09001744#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1745 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02001746#define BUILD_EMBED_FIRST_CHUNK
1747#endif
1748
1749/* build pcpu_page_first_chunk() iff needed by the arch config */
1750#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1751#define BUILD_PAGE_FIRST_CHUNK
1752#endif
1753
1754/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1755#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1756/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09001757 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1758 * @reserved_size: the size of reserved percpu area in bytes
1759 * @dyn_size: minimum free size for dynamic allocation in bytes
1760 * @atom_size: allocation atom size
1761 * @cpu_distance_fn: callback to determine distance between cpus, optional
1762 *
1763 * This function determines grouping of units, their mappings to cpus
1764 * and other parameters considering needed percpu size, allocation
1765 * atom size and distances between CPUs.
1766 *
1767 * Groups are always mutliples of atom size and CPUs which are of
1768 * LOCAL_DISTANCE both ways are grouped together and share space for
1769 * units in the same group. The returned configuration is guaranteed
1770 * to have CPUs on different nodes on different groups and >=75% usage
1771 * of allocated virtual address space.
1772 *
1773 * RETURNS:
1774 * On success, pointer to the new allocation_info is returned. On
1775 * failure, ERR_PTR value is returned.
1776 */
1777static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1778 size_t reserved_size, size_t dyn_size,
1779 size_t atom_size,
1780 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1781{
1782 static int group_map[NR_CPUS] __initdata;
1783 static int group_cnt[NR_CPUS] __initdata;
1784 const size_t static_size = __per_cpu_end - __per_cpu_start;
1785 int nr_groups = 1, nr_units = 0;
1786 size_t size_sum, min_unit_size, alloc_size;
1787 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1788 int last_allocs, group, unit;
1789 unsigned int cpu, tcpu;
1790 struct pcpu_alloc_info *ai;
1791 unsigned int *cpu_map;
1792
1793 /* this function may be called multiple times */
1794 memset(group_map, 0, sizeof(group_map));
1795 memset(group_cnt, 0, sizeof(group_cnt));
1796
1797 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1798 size_sum = PFN_ALIGN(static_size + reserved_size +
1799 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1800 dyn_size = size_sum - static_size - reserved_size;
1801
1802 /*
1803 * Determine min_unit_size, alloc_size and max_upa such that
1804 * alloc_size is multiple of atom_size and is the smallest
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001805 * which can accommodate 4k aligned segments which are equal to
Tejun Heofbf59bc2009-02-20 16:29:08 +09001806 * or larger than min_unit_size.
1807 */
1808 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1809
1810 alloc_size = roundup(min_unit_size, atom_size);
1811 upa = alloc_size / min_unit_size;
1812 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1813 upa--;
1814 max_upa = upa;
1815
1816 /* group cpus according to their proximity */
1817 for_each_possible_cpu(cpu) {
1818 group = 0;
1819 next_group:
1820 for_each_possible_cpu(tcpu) {
1821 if (cpu == tcpu)
1822 break;
1823 if (group_map[tcpu] == group && cpu_distance_fn &&
1824 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1825 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1826 group++;
1827 nr_groups = max(nr_groups, group + 1);
1828 goto next_group;
1829 }
1830 }
1831 group_map[cpu] = group;
1832 group_cnt[group]++;
1833 }
1834
1835 /*
1836 * Expand unit size until address space usage goes over 75%
1837 * and then as much as possible without using more address
1838 * space.
1839 */
1840 last_allocs = INT_MAX;
1841 for (upa = max_upa; upa; upa--) {
1842 int allocs = 0, wasted = 0;
1843
1844 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1845 continue;
1846
1847 for (group = 0; group < nr_groups; group++) {
1848 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1849 allocs += this_allocs;
1850 wasted += this_allocs * upa - group_cnt[group];
1851 }
1852
1853 /*
1854 * Don't accept if wastage is over 1/3. The
1855 * greater-than comparison ensures upa==1 always
1856 * passes the following check.
1857 */
1858 if (wasted > num_possible_cpus() / 3)
1859 continue;
1860
1861 /* and then don't consume more memory */
1862 if (allocs > last_allocs)
1863 break;
1864 last_allocs = allocs;
1865 best_upa = upa;
1866 }
1867 upa = best_upa;
1868
1869 /* allocate and fill alloc_info */
1870 for (group = 0; group < nr_groups; group++)
1871 nr_units += roundup(group_cnt[group], upa);
1872
1873 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1874 if (!ai)
1875 return ERR_PTR(-ENOMEM);
1876 cpu_map = ai->groups[0].cpu_map;
1877
1878 for (group = 0; group < nr_groups; group++) {
1879 ai->groups[group].cpu_map = cpu_map;
1880 cpu_map += roundup(group_cnt[group], upa);
1881 }
1882
1883 ai->static_size = static_size;
1884 ai->reserved_size = reserved_size;
1885 ai->dyn_size = dyn_size;
1886 ai->unit_size = alloc_size / upa;
1887 ai->atom_size = atom_size;
1888 ai->alloc_size = alloc_size;
1889
1890 for (group = 0, unit = 0; group_cnt[group]; group++) {
1891 struct pcpu_group_info *gi = &ai->groups[group];
1892
1893 /*
1894 * Initialize base_offset as if all groups are located
1895 * back-to-back. The caller should update this to
1896 * reflect actual allocation.
1897 */
1898 gi->base_offset = unit * ai->unit_size;
1899
1900 for_each_possible_cpu(cpu)
1901 if (group_map[cpu] == group)
1902 gi->cpu_map[gi->nr_units++] = cpu;
1903 gi->nr_units = roundup(gi->nr_units, upa);
1904 unit += gi->nr_units;
1905 }
1906 BUG_ON(unit != nr_units);
1907
1908 return ai;
1909}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001910#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001911
Tejun Heo3c9a0242010-09-09 18:00:15 +02001912#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09001913/**
1914 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001915 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001916 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001917 * @atom_size: allocation atom size
1918 * @cpu_distance_fn: callback to determine distance between cpus, optional
1919 * @alloc_fn: function to allocate percpu page
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001920 * @free_fn: function to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001921 *
1922 * This is a helper to ease setting up embedded first percpu chunk and
1923 * can be called where pcpu_setup_first_chunk() is expected.
1924 *
1925 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001926 * by calling @alloc_fn and used as-is without being mapped into
1927 * vmalloc area. Allocations are always whole multiples of @atom_size
1928 * aligned to @atom_size.
1929 *
1930 * This enables the first chunk to piggy back on the linear physical
1931 * mapping which often uses larger page size. Please note that this
1932 * can result in very sparse cpu->unit mapping on NUMA machines thus
1933 * requiring large vmalloc address space. Don't use this allocator if
1934 * vmalloc space is not orders of magnitude larger than distances
1935 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001936 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001937 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001938 *
1939 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001940 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001941 *
1942 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001943 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001944 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001945int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001946 size_t atom_size,
1947 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1948 pcpu_fc_alloc_fn_t alloc_fn,
1949 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001950{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001951 void *base = (void *)ULONG_MAX;
1952 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001953 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001954 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001955 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001956
Tejun Heoc8826dd2009-08-14 15:00:52 +09001957 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1958 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001959 if (IS_ERR(ai))
1960 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001961
Tejun Heofd1e8a12009-08-14 15:00:51 +09001962 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001963 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001964
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001965 areas = memblock_virt_alloc_nopanic(areas_size, 0);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001966 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001967 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001968 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001969 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001970
Tejun Heoc8826dd2009-08-14 15:00:52 +09001971 /* allocate, copy and determine base address */
1972 for (group = 0; group < ai->nr_groups; group++) {
1973 struct pcpu_group_info *gi = &ai->groups[group];
1974 unsigned int cpu = NR_CPUS;
1975 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001976
Tejun Heoc8826dd2009-08-14 15:00:52 +09001977 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1978 cpu = gi->cpu_map[i];
1979 BUG_ON(cpu == NR_CPUS);
1980
1981 /* allocate space for the whole group */
1982 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1983 if (!ptr) {
1984 rc = -ENOMEM;
1985 goto out_free_areas;
1986 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001987 /* kmemleak tracks the percpu allocations separately */
1988 kmemleak_free(ptr);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001989 areas[group] = ptr;
1990
1991 base = min(ptr, base);
Tejun Heo42b64282012-04-27 08:42:53 -07001992 }
1993
1994 /*
1995 * Copy data and free unused parts. This should happen after all
1996 * allocations are complete; otherwise, we may end up with
1997 * overlapping groups.
1998 */
1999 for (group = 0; group < ai->nr_groups; group++) {
2000 struct pcpu_group_info *gi = &ai->groups[group];
2001 void *ptr = areas[group];
Tejun Heoc8826dd2009-08-14 15:00:52 +09002002
2003 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2004 if (gi->cpu_map[i] == NR_CPUS) {
2005 /* unused unit, free whole */
2006 free_fn(ptr, ai->unit_size);
2007 continue;
2008 }
2009 /* copy and return the unused part */
2010 memcpy(ptr, __per_cpu_load, ai->static_size);
2011 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2012 }
Tejun Heo66c3a752009-03-10 16:27:48 +09002013 }
2014
Tejun Heoc8826dd2009-08-14 15:00:52 +09002015 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09002016 max_distance = 0;
2017 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09002018 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09002019 max_distance = max_t(size_t, max_distance,
2020 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09002021 }
2022 max_distance += ai->unit_size;
2023
2024 /* warn if maximum distance is further than 75% of vmalloc space */
Laura Abbott8a092172014-01-02 13:53:21 -08002025 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09002026 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Mike Frysinger787e5b02011-03-23 08:23:52 +01002027 "space 0x%lx\n", max_distance,
Laura Abbott8a092172014-01-02 13:53:21 -08002028 VMALLOC_TOTAL);
Tejun Heo6ea529a2009-09-24 18:46:01 +09002029#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2030 /* and fail if we have fallback */
2031 rc = -EINVAL;
2032 goto out_free;
2033#endif
2034 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09002035
Tejun Heo004018e2009-08-14 15:00:49 +09002036 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002037 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2038 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002039
Tejun Heofb435d52009-08-14 15:00:51 +09002040 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002041 goto out_free;
2042
2043out_free_areas:
2044 for (group = 0; group < ai->nr_groups; group++)
Michael Holzheuf851c8d2013-09-17 16:57:34 +02002045 if (areas[group])
2046 free_fn(areas[group],
2047 ai->groups[group].nr_units * ai->unit_size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002048out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09002049 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002050 if (areas)
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002051 memblock_free_early(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09002052 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002053}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002054#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002055
Tejun Heo3c9a0242010-09-09 18:00:15 +02002056#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09002057/**
Tejun Heo00ae4062009-08-14 15:00:49 +09002058 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09002059 * @reserved_size: the size of reserved percpu area in bytes
2060 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002061 * @free_fn: function to free percpu page, always called with PAGE_SIZE
Tejun Heod4b95f82009-07-04 08:10:59 +09002062 * @populate_pte_fn: function to populate pte
2063 *
Tejun Heo00ae4062009-08-14 15:00:49 +09002064 * This is a helper to ease setting up page-remapped first percpu
2065 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09002066 *
2067 * This is the basic allocator. Static percpu area is allocated
2068 * page-by-page into vmalloc area.
2069 *
2070 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09002071 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09002072 */
Tejun Heofb435d52009-08-14 15:00:51 +09002073int __init pcpu_page_first_chunk(size_t reserved_size,
2074 pcpu_fc_alloc_fn_t alloc_fn,
2075 pcpu_fc_free_fn_t free_fn,
2076 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09002077{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002078 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002079 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09002080 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09002081 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09002082 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09002083 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09002084 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002085
Tejun Heo00ae4062009-08-14 15:00:49 +09002086 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2087
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002088 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002089 if (IS_ERR(ai))
2090 return PTR_ERR(ai);
2091 BUG_ON(ai->nr_groups != 1);
2092 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2093
2094 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09002095
2096 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002097 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2098 sizeof(pages[0]));
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002099 pages = memblock_virt_alloc(pages_size, 0);
Tejun Heod4b95f82009-07-04 08:10:59 +09002100
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002101 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09002102 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002103 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09002104 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09002105 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09002106 void *ptr;
2107
Tejun Heo3cbc8562009-08-14 15:00:50 +09002108 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09002109 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09002110 pr_warning("PERCPU: failed to allocate %s page "
2111 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09002112 goto enomem;
2113 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01002114 /* kmemleak tracks the percpu allocations separately */
2115 kmemleak_free(ptr);
Tejun Heoce3141a2009-07-04 08:11:00 +09002116 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002117 }
2118
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002119 /* allocate vm area, map the pages and copy static data */
2120 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002121 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002122 vm_area_register_early(&vm, PAGE_SIZE);
2123
Tejun Heofd1e8a12009-08-14 15:00:51 +09002124 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09002125 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09002126 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002127
Tejun Heoce3141a2009-07-04 08:11:00 +09002128 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002129 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2130
2131 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09002132 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2133 unit_pages);
2134 if (rc < 0)
2135 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002136
2137 /*
2138 * FIXME: Archs with virtual cache should flush local
2139 * cache for the linear mapping here - something
2140 * equivalent to flush_cache_vmap() on the local cpu.
2141 * flush_cache_vmap() can't be used as most supporting
2142 * data structures are not set up yet.
2143 */
2144
2145 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002146 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002147 }
2148
2149 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09002150 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002151 unit_pages, psize_str, vm.addr, ai->static_size,
2152 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002153
Tejun Heofb435d52009-08-14 15:00:51 +09002154 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002155 goto out_free_ar;
2156
2157enomem:
2158 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002159 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002160 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002161out_free_ar:
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002162 memblock_free_early(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002163 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002164 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002165}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002166#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002167
Tejun Heobbddff02010-09-03 18:22:48 +02002168#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002169/*
Tejun Heobbddff02010-09-03 18:22:48 +02002170 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09002171 *
2172 * The embedding helper is used because its behavior closely resembles
2173 * the original non-dynamic generic percpu area setup. This is
2174 * important because many archs have addressing restrictions and might
2175 * fail if the percpu area is located far away from the previous
2176 * location. As an added bonus, in non-NUMA cases, embedding is
2177 * generally a good idea TLB-wise because percpu area can piggy back
2178 * on the physical linear memory mapping which uses large page
2179 * mappings on applicable archs.
2180 */
Tejun Heoe74e3962009-03-30 19:07:44 +09002181unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2182EXPORT_SYMBOL(__per_cpu_offset);
2183
Tejun Heoc8826dd2009-08-14 15:00:52 +09002184static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2185 size_t align)
2186{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002187 return memblock_virt_alloc_from_nopanic(
2188 size, align, __pa(MAX_DMA_ADDRESS));
Tejun Heoc8826dd2009-08-14 15:00:52 +09002189}
2190
2191static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2192{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002193 memblock_free_early(__pa(ptr), size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002194}
2195
Tejun Heoe74e3962009-03-30 19:07:44 +09002196void __init setup_per_cpu_areas(void)
2197{
Tejun Heoe74e3962009-03-30 19:07:44 +09002198 unsigned long delta;
2199 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002200 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002201
2202 /*
2203 * Always reserve area for module percpu variables. That's
2204 * what the legacy allocator did.
2205 */
Tejun Heofb435d52009-08-14 15:00:51 +09002206 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002207 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2208 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002209 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02002210 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09002211
2212 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2213 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002214 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002215}
Tejun Heobbddff02010-09-03 18:22:48 +02002216#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2217
2218#else /* CONFIG_SMP */
2219
2220/*
2221 * UP percpu area setup.
2222 *
2223 * UP always uses km-based percpu allocator with identity mapping.
2224 * Static percpu variables are indistinguishable from the usual static
2225 * variables and don't require any special preparation.
2226 */
2227void __init setup_per_cpu_areas(void)
2228{
2229 const size_t unit_size =
2230 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2231 PERCPU_DYNAMIC_RESERVE));
2232 struct pcpu_alloc_info *ai;
2233 void *fc;
2234
2235 ai = pcpu_alloc_alloc_info(1, 1);
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002236 fc = memblock_virt_alloc_from_nopanic(unit_size,
2237 PAGE_SIZE,
2238 __pa(MAX_DMA_ADDRESS));
Tejun Heobbddff02010-09-03 18:22:48 +02002239 if (!ai || !fc)
2240 panic("Failed to allocate memory for percpu areas.");
Catalin Marinas100d13c2012-05-09 16:55:19 +01002241 /* kmemleak tracks the percpu allocations separately */
2242 kmemleak_free(fc);
Tejun Heobbddff02010-09-03 18:22:48 +02002243
2244 ai->dyn_size = unit_size;
2245 ai->unit_size = unit_size;
2246 ai->atom_size = unit_size;
2247 ai->alloc_size = unit_size;
2248 ai->groups[0].nr_units = 1;
2249 ai->groups[0].cpu_map[0] = 0;
2250
2251 if (pcpu_setup_first_chunk(ai, fc) < 0)
2252 panic("Failed to initialize percpu areas.");
2253}
2254
2255#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02002256
2257/*
2258 * First and reserved chunks are initialized with temporary allocation
2259 * map in initdata so that they can be used before slab is online.
2260 * This function is called after slab is brought up and replaces those
2261 * with properly allocated maps.
2262 */
2263void __init percpu_init_late(void)
2264{
2265 struct pcpu_chunk *target_chunks[] =
2266 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2267 struct pcpu_chunk *chunk;
2268 unsigned long flags;
2269 int i;
2270
2271 for (i = 0; (chunk = target_chunks[i]); i++) {
2272 int *map;
2273 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2274
2275 BUILD_BUG_ON(size > PAGE_SIZE);
2276
Bob Liu90459ce02011-08-04 11:02:33 +02002277 map = pcpu_mem_zalloc(size);
Tejun Heo099a19d2010-06-27 18:50:00 +02002278 BUG_ON(!map);
2279
2280 spin_lock_irqsave(&pcpu_lock, flags);
2281 memcpy(map, chunk->map, size);
2282 chunk->map = map;
2283 spin_unlock_irqrestore(&pcpu_lock, flags);
2284 }
2285}
Tejun Heo1a4d7602014-09-02 14:46:05 -04002286
2287/*
2288 * Percpu allocator is initialized early during boot when neither slab or
2289 * workqueue is available. Plug async management until everything is up
2290 * and running.
2291 */
2292static int __init percpu_enable_async(void)
2293{
2294 pcpu_async_enabled = true;
2295 return 0;
2296}
2297subsys_initcall(percpu_enable_async);