blob: 442010cc91c6c82eb8489e64d21500baa52b5911 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
2 * linux/mm/percpu.c - percpu memory allocator
3 *
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
10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each
Tejun Heo2f39e632009-07-04 08:11:00 +090011 * chunk is consisted of boot-time determined number of units and the
12 * first chunk is used for static percpu variables in the kernel image
13 * (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.
16 * When a chunk is filled up, another chunk is allocated. ie. in
17 * vmalloc area
Tejun Heofbf59bc2009-02-20 16:29:08 +090018 *
19 * c0 c1 c2
20 * ------------------- ------------------- ------------
21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
22 * ------------------- ...... ------------------- .... ------------
23 *
24 * Allocation is done in offset-size areas of single unit space. Ie,
25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
Tejun Heo2f39e632009-07-04 08:11:00 +090026 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090030 *
Tejun Heo2f39e632009-07-04 08:11:00 +090031 * There are usually many small percpu allocations many of them being
32 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090033 * according to free size and tries to allocate from the fullest one.
34 * Each chunk keeps the maximum contiguous area size hint which is
35 * guaranteed to be eqaul to or larger than the maximum contiguous
36 * area in the chunk. This helps the allocator not to iterate the
37 * chunk maps unnecessarily.
38 *
39 * Allocation state in each chunk is kept using an array of integers
40 * on chunk->map. A positive value in the map represents a free
41 * region and negative allocated. Allocation inside a chunk is done
42 * by scanning this map sequentially and serving the first matching
43 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090044 * Chunks can be determined from the address using the index field
45 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090046 *
47 * To use this allocator, arch code should do the followings.
48 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090049 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090050 * regular address to percpu pointer and back if they need to be
51 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090052 *
Tejun Heo8d408b42009-02-24 11:57:21 +090053 * - use pcpu_setup_first_chunk() during percpu area initialization to
54 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090055 */
56
57#include <linux/bitmap.h>
58#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090059#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090060#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090061#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090062#include <linux/mm.h>
63#include <linux/module.h>
64#include <linux/mutex.h>
65#include <linux/percpu.h>
66#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090067#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090068#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090069#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090070#include <linux/workqueue.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071
72#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090073#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090074#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090075#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090076
Tejun Heofbf59bc2009-02-20 16:29:08 +090077#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79
Tejun Heoe0100982009-03-10 16:27:48 +090080/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
81#ifndef __addr_to_pcpu_ptr
82#define __addr_to_pcpu_ptr(addr) \
83 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
84 + (unsigned long)__per_cpu_start)
85#endif
86#ifndef __pcpu_ptr_to_addr
87#define __pcpu_ptr_to_addr(ptr) \
88 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
89 - (unsigned long)__per_cpu_start)
90#endif
91
Tejun Heofbf59bc2009-02-20 16:29:08 +090092struct pcpu_chunk {
93 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +090094 int free_size; /* free bytes in the chunk */
95 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +090096 void *base_addr; /* base address of this chunk */
Tejun Heofbf59bc2009-02-20 16:29:08 +090097 int map_used; /* # of map entries used */
98 int map_alloc; /* # of map entries allocated */
99 int *map; /* allocation map */
Tejun Heo65632972009-08-14 15:00:52 +0900100 struct vm_struct **vms; /* mapped vmalloc regions */
Tejun Heo8d408b42009-02-24 11:57:21 +0900101 bool immutable; /* no [de]population allowed */
Tejun Heoce3141a2009-07-04 08:11:00 +0900102 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900103};
104
Tejun Heo40150d32009-02-24 12:32:28 +0900105static int pcpu_unit_pages __read_mostly;
106static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900107static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900108static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900109static int pcpu_nr_slots __read_mostly;
110static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900111
Tejun Heo2f39e632009-07-04 08:11:00 +0900112/* cpus with the lowest and highest unit numbers */
113static unsigned int pcpu_first_unit_cpu __read_mostly;
114static unsigned int pcpu_last_unit_cpu __read_mostly;
115
Tejun Heofbf59bc2009-02-20 16:29:08 +0900116/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900117void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900118EXPORT_SYMBOL_GPL(pcpu_base_addr);
119
Tejun Heofb435d52009-08-14 15:00:51 +0900120static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
121const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900122
Tejun Heo65632972009-08-14 15:00:52 +0900123/* group information, used for vm allocation */
124static int pcpu_nr_groups __read_mostly;
125static const unsigned long *pcpu_group_offsets __read_mostly;
126static const size_t *pcpu_group_sizes __read_mostly;
127
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900128/*
129 * The first chunk which always exists. Note that unlike other
130 * chunks, this one can be allocated and mapped in several different
131 * ways and thus often doesn't live in the vmalloc area.
132 */
133static struct pcpu_chunk *pcpu_first_chunk;
134
135/*
136 * Optional reserved chunk. This chunk reserves part of the first
137 * chunk and serves it for reserved allocations. The amount of
138 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
139 * area doesn't exist, the following variables contain NULL and 0
140 * respectively.
141 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900142static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900143static int pcpu_reserved_chunk_limit;
144
Tejun Heofbf59bc2009-02-20 16:29:08 +0900145/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900146 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900147 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900148 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
Tejun Heoce3141a2009-07-04 08:11:00 +0900149 * protects allocation/reclaim paths, chunks, populated bitmap and
150 * vmalloc mapping. The latter is a spinlock and protects the index
151 * data structures - chunk slots, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900152 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900153 * During allocation, pcpu_alloc_mutex is kept locked all the time and
154 * pcpu_lock is grabbed and released as necessary. All actual memory
Jiri Kosina403a91b2009-10-29 00:25:59 +0900155 * allocations are done using GFP_KERNEL with pcpu_lock released. In
156 * general, percpu memory can't be allocated with irq off but
157 * irqsave/restore are still used in alloc path so that it can be used
158 * from early init path - sched_init() specifically.
Tejun Heoccea34b2009-03-07 00:44:13 +0900159 *
160 * Free path accesses and alters only the index data structures, so it
161 * can be safely called from atomic context. When memory needs to be
162 * returned to the system, free path schedules reclaim_work which
163 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
164 * reclaimed, release both locks and frees the chunks. Note that it's
165 * necessary to grab both locks to remove a chunk from circulation as
166 * allocation path might be referencing the chunk with only
167 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900168 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900169static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
170static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900171
Tejun Heo40150d32009-02-24 12:32:28 +0900172static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900173
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900174/* reclaim work to release fully free chunks, scheduled from free path */
175static void pcpu_reclaim(struct work_struct *work);
176static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
177
Tejun Heod9b55ee2009-02-24 11:57:21 +0900178static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900179{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900180 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900181 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
182}
183
Tejun Heod9b55ee2009-02-24 11:57:21 +0900184static int pcpu_size_to_slot(int size)
185{
186 if (size == pcpu_unit_size)
187 return pcpu_nr_slots - 1;
188 return __pcpu_size_to_slot(size);
189}
190
Tejun Heofbf59bc2009-02-20 16:29:08 +0900191static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
192{
193 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
194 return 0;
195
196 return pcpu_size_to_slot(chunk->free_size);
197}
198
199static int pcpu_page_idx(unsigned int cpu, int page_idx)
200{
Tejun Heo2f39e632009-07-04 08:11:00 +0900201 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900202}
203
204static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
205 unsigned int cpu, int page_idx)
206{
Tejun Heobba174f2009-08-14 15:00:51 +0900207 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900208 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900209}
210
Tejun Heoce3141a2009-07-04 08:11:00 +0900211static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
212 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900213{
Tejun Heoce3141a2009-07-04 08:11:00 +0900214 /* must not be used on pre-mapped chunk */
215 WARN_ON(chunk->immutable);
Tejun Heoc8a51be2009-07-04 08:10:59 +0900216
Tejun Heoce3141a2009-07-04 08:11:00 +0900217 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900218}
219
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900220/* set the pointer to a chunk in a page struct */
221static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
222{
223 page->index = (unsigned long)pcpu;
224}
225
226/* obtain pointer to a chunk from a page struct */
227static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
228{
229 return (struct pcpu_chunk *)page->index;
230}
231
Tejun Heoce3141a2009-07-04 08:11:00 +0900232static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
233{
234 *rs = find_next_zero_bit(chunk->populated, end, *rs);
235 *re = find_next_bit(chunk->populated, end, *rs + 1);
236}
237
238static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
239{
240 *rs = find_next_bit(chunk->populated, end, *rs);
241 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
242}
243
244/*
245 * (Un)populated page region iterators. Iterate over (un)populated
246 * page regions betwen @start and @end in @chunk. @rs and @re should
247 * be integer variables and will be set to start and end page index of
248 * the current region.
249 */
250#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
251 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
252 (rs) < (re); \
253 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
254
255#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
256 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
257 (rs) < (re); \
258 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
259
Tejun Heofbf59bc2009-02-20 16:29:08 +0900260/**
Tejun Heo1880d932009-03-07 00:44:09 +0900261 * pcpu_mem_alloc - allocate memory
262 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900263 *
Tejun Heo1880d932009-03-07 00:44:09 +0900264 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
265 * kzalloc() is used; otherwise, vmalloc() is used. The returned
266 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900267 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900268 * CONTEXT:
269 * Does GFP_KERNEL allocation.
270 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900271 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900272 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900273 */
Tejun Heo1880d932009-03-07 00:44:09 +0900274static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900275{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900276 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900277 return kzalloc(size, GFP_KERNEL);
278 else {
279 void *ptr = vmalloc(size);
280 if (ptr)
281 memset(ptr, 0, size);
282 return ptr;
283 }
284}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900285
Tejun Heo1880d932009-03-07 00:44:09 +0900286/**
287 * pcpu_mem_free - free memory
288 * @ptr: memory to free
289 * @size: size of the area
290 *
291 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
292 */
293static void pcpu_mem_free(void *ptr, size_t size)
294{
295 if (size <= PAGE_SIZE)
296 kfree(ptr);
297 else
298 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900299}
300
301/**
302 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
303 * @chunk: chunk of interest
304 * @oslot: the previous slot it was on
305 *
306 * This function is called after an allocation or free changed @chunk.
307 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900308 * moved to the slot. Note that the reserved chunk is never put on
309 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900310 *
311 * CONTEXT:
312 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900313 */
314static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
315{
316 int nslot = pcpu_chunk_slot(chunk);
317
Tejun Heoedcb4632009-03-06 14:33:59 +0900318 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900319 if (oslot < nslot)
320 list_move(&chunk->list, &pcpu_slot[nslot]);
321 else
322 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
323 }
324}
325
Tejun Heofbf59bc2009-02-20 16:29:08 +0900326/**
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900327 * pcpu_chunk_addr_search - determine chunk containing specified address
328 * @addr: address for which the chunk needs to be determined.
Tejun Heoccea34b2009-03-07 00:44:13 +0900329 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900330 * RETURNS:
331 * The address of the found chunk.
332 */
333static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
334{
Tejun Heobba174f2009-08-14 15:00:51 +0900335 void *first_start = pcpu_first_chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900336
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900337 /* is it in the first chunk? */
Tejun Heo79ba6ac2009-07-04 08:10:58 +0900338 if (addr >= first_start && addr < first_start + pcpu_unit_size) {
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900339 /* is it in the reserved area? */
340 if (addr < first_start + pcpu_reserved_chunk_limit)
Tejun Heoedcb4632009-03-06 14:33:59 +0900341 return pcpu_reserved_chunk;
Tejun Heoae9e6bc2009-04-02 13:19:54 +0900342 return pcpu_first_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900343 }
344
Tejun Heo04a13c72009-09-01 21:12:28 +0900345 /*
346 * The address is relative to unit0 which might be unused and
347 * thus unmapped. Offset the address to the unit space of the
348 * current processor before looking it up in the vmalloc
349 * space. Note that any possible cpu id can be used here, so
350 * there's no need to worry about preemption or cpu hotplug.
351 */
Tejun Heo5579fd72009-09-15 09:57:19 +0900352 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900353 return pcpu_get_page_chunk(vmalloc_to_page(addr));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900354}
355
356/**
Tejun Heo833af842009-11-11 15:35:18 +0900357 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
358 * @chunk: chunk of interest
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900359 *
Tejun Heo833af842009-11-11 15:35:18 +0900360 * Determine whether area map of @chunk needs to be extended to
361 * accomodate a new allocation.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900362 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900363 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900364 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900365 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900366 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900367 * New target map allocation length if extension is necessary, 0
368 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900369 */
Tejun Heo833af842009-11-11 15:35:18 +0900370static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900371{
372 int new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900373
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900374 if (chunk->map_alloc >= chunk->map_used + 2)
375 return 0;
376
377 new_alloc = PCPU_DFL_MAP_ALLOC;
378 while (new_alloc < chunk->map_used + 2)
379 new_alloc *= 2;
380
Tejun Heo833af842009-11-11 15:35:18 +0900381 return new_alloc;
382}
383
384/**
385 * pcpu_extend_area_map - extend area map of a chunk
386 * @chunk: chunk of interest
387 * @new_alloc: new target allocation length of the area map
388 *
389 * Extend area map of @chunk to have @new_alloc entries.
390 *
391 * CONTEXT:
392 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
393 *
394 * RETURNS:
395 * 0 on success, -errno on failure.
396 */
397static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
398{
399 int *old = NULL, *new = NULL;
400 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
401 unsigned long flags;
402
403 new = pcpu_mem_alloc(new_size);
404 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900405 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900406
Tejun Heo833af842009-11-11 15:35:18 +0900407 /* acquire pcpu_lock and switch to new area map */
408 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900409
Tejun Heo833af842009-11-11 15:35:18 +0900410 if (new_alloc <= chunk->map_alloc)
411 goto out_unlock;
412
413 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
414 memcpy(new, chunk->map, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900415
416 /*
417 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
418 * one of the first chunks and still using static map.
419 */
420 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
Tejun Heo833af842009-11-11 15:35:18 +0900421 old = chunk->map;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900422
423 chunk->map_alloc = new_alloc;
424 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900425 new = NULL;
426
427out_unlock:
428 spin_unlock_irqrestore(&pcpu_lock, flags);
429
430 /*
431 * pcpu_mem_free() might end up calling vfree() which uses
432 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
433 */
434 pcpu_mem_free(old, old_size);
435 pcpu_mem_free(new, new_size);
436
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900437 return 0;
438}
439
440/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900441 * pcpu_split_block - split a map block
442 * @chunk: chunk of interest
443 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900444 * @head: head size in bytes (can be 0)
445 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900446 *
447 * Split the @i'th map block into two or three blocks. If @head is
448 * non-zero, @head bytes block is inserted before block @i moving it
449 * to @i+1 and reducing its size by @head bytes.
450 *
451 * If @tail is non-zero, the target block, which can be @i or @i+1
452 * depending on @head, is reduced by @tail bytes and @tail byte block
453 * is inserted after the target block.
454 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900455 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900456 *
457 * CONTEXT:
458 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900459 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900460static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
461 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900462{
463 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900464
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900465 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900466
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900467 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900468 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
469 sizeof(chunk->map[0]) * (chunk->map_used - i));
470 chunk->map_used += nr_extra;
471
472 if (head) {
473 chunk->map[i + 1] = chunk->map[i] - head;
474 chunk->map[i++] = head;
475 }
476 if (tail) {
477 chunk->map[i++] -= tail;
478 chunk->map[i] = tail;
479 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900480}
481
482/**
483 * pcpu_alloc_area - allocate area from a pcpu_chunk
484 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900485 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900486 * @align: wanted align
487 *
488 * Try to allocate @size bytes area aligned at @align from @chunk.
489 * Note that this function only allocates the offset. It doesn't
490 * populate or map the area.
491 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900492 * @chunk->map must have at least two free slots.
493 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900494 * CONTEXT:
495 * pcpu_lock.
496 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900497 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900498 * Allocated offset in @chunk on success, -1 if no matching area is
499 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900500 */
501static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
502{
503 int oslot = pcpu_chunk_slot(chunk);
504 int max_contig = 0;
505 int i, off;
506
Tejun Heofbf59bc2009-02-20 16:29:08 +0900507 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
508 bool is_last = i + 1 == chunk->map_used;
509 int head, tail;
510
511 /* extra for alignment requirement */
512 head = ALIGN(off, align) - off;
513 BUG_ON(i == 0 && head != 0);
514
515 if (chunk->map[i] < 0)
516 continue;
517 if (chunk->map[i] < head + size) {
518 max_contig = max(chunk->map[i], max_contig);
519 continue;
520 }
521
522 /*
523 * If head is small or the previous block is free,
524 * merge'em. Note that 'small' is defined as smaller
525 * than sizeof(int), which is very small but isn't too
526 * uncommon for percpu allocations.
527 */
528 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
529 if (chunk->map[i - 1] > 0)
530 chunk->map[i - 1] += head;
531 else {
532 chunk->map[i - 1] -= head;
533 chunk->free_size -= head;
534 }
535 chunk->map[i] -= head;
536 off += head;
537 head = 0;
538 }
539
540 /* if tail is small, just keep it around */
541 tail = chunk->map[i] - head - size;
542 if (tail < sizeof(int))
543 tail = 0;
544
545 /* split if warranted */
546 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900547 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900548 if (head) {
549 i++;
550 off += head;
551 max_contig = max(chunk->map[i - 1], max_contig);
552 }
553 if (tail)
554 max_contig = max(chunk->map[i + 1], max_contig);
555 }
556
557 /* update hint and mark allocated */
558 if (is_last)
559 chunk->contig_hint = max_contig; /* fully scanned */
560 else
561 chunk->contig_hint = max(chunk->contig_hint,
562 max_contig);
563
564 chunk->free_size -= chunk->map[i];
565 chunk->map[i] = -chunk->map[i];
566
567 pcpu_chunk_relocate(chunk, oslot);
568 return off;
569 }
570
571 chunk->contig_hint = max_contig; /* fully scanned */
572 pcpu_chunk_relocate(chunk, oslot);
573
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900574 /* tell the upper layer that this chunk has no matching area */
575 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900576}
577
578/**
579 * pcpu_free_area - free area to a pcpu_chunk
580 * @chunk: chunk of interest
581 * @freeme: offset of area to free
582 *
583 * Free area starting from @freeme to @chunk. Note that this function
584 * only modifies the allocation map. It doesn't depopulate or unmap
585 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900586 *
587 * CONTEXT:
588 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900589 */
590static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
591{
592 int oslot = pcpu_chunk_slot(chunk);
593 int i, off;
594
595 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
596 if (off == freeme)
597 break;
598 BUG_ON(off != freeme);
599 BUG_ON(chunk->map[i] > 0);
600
601 chunk->map[i] = -chunk->map[i];
602 chunk->free_size += chunk->map[i];
603
604 /* merge with previous? */
605 if (i > 0 && chunk->map[i - 1] >= 0) {
606 chunk->map[i - 1] += chunk->map[i];
607 chunk->map_used--;
608 memmove(&chunk->map[i], &chunk->map[i + 1],
609 (chunk->map_used - i) * sizeof(chunk->map[0]));
610 i--;
611 }
612 /* merge with next? */
613 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
614 chunk->map[i] += chunk->map[i + 1];
615 chunk->map_used--;
616 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
617 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
618 }
619
620 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
621 pcpu_chunk_relocate(chunk, oslot);
622}
623
624/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900625 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900626 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900627 * @bitmapp: output parameter for bitmap
628 * @may_alloc: may allocate the array
Tejun Heofbf59bc2009-02-20 16:29:08 +0900629 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900630 * Returns pointer to array of pointers to struct page and bitmap,
631 * both of which can be indexed with pcpu_page_idx(). The returned
632 * array is cleared to zero and *@bitmapp is copied from
633 * @chunk->populated. Note that there is only one array and bitmap
634 * and access exclusion is the caller's responsibility.
635 *
636 * CONTEXT:
637 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
638 * Otherwise, don't care.
639 *
640 * RETURNS:
641 * Pointer to temp pages array on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900642 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900643static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
644 unsigned long **bitmapp,
645 bool may_alloc)
646{
647 static struct page **pages;
648 static unsigned long *bitmap;
Tejun Heo2f39e632009-07-04 08:11:00 +0900649 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
Tejun Heoce3141a2009-07-04 08:11:00 +0900650 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
651 sizeof(unsigned long);
652
653 if (!pages || !bitmap) {
654 if (may_alloc && !pages)
655 pages = pcpu_mem_alloc(pages_size);
656 if (may_alloc && !bitmap)
657 bitmap = pcpu_mem_alloc(bitmap_size);
658 if (!pages || !bitmap)
659 return NULL;
660 }
661
662 memset(pages, 0, pages_size);
663 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
664
665 *bitmapp = bitmap;
666 return pages;
667}
668
669/**
670 * pcpu_free_pages - free pages which were allocated for @chunk
671 * @chunk: chunk pages were allocated for
672 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
673 * @populated: populated bitmap
674 * @page_start: page index of the first page to be freed
675 * @page_end: page index of the last page to be freed + 1
676 *
677 * Free pages [@page_start and @page_end) in @pages for all units.
678 * The pages were allocated for @chunk.
679 */
680static void pcpu_free_pages(struct pcpu_chunk *chunk,
681 struct page **pages, unsigned long *populated,
682 int page_start, int page_end)
683{
684 unsigned int cpu;
685 int i;
686
687 for_each_possible_cpu(cpu) {
688 for (i = page_start; i < page_end; i++) {
689 struct page *page = pages[pcpu_page_idx(cpu, i)];
690
691 if (page)
692 __free_page(page);
693 }
694 }
695}
696
697/**
698 * pcpu_alloc_pages - allocates pages for @chunk
699 * @chunk: target chunk
700 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
701 * @populated: populated bitmap
702 * @page_start: page index of the first page to be allocated
703 * @page_end: page index of the last page to be allocated + 1
704 *
705 * Allocate pages [@page_start,@page_end) into @pages for all units.
706 * The allocation is for @chunk. Percpu core doesn't care about the
707 * content of @pages and will pass it verbatim to pcpu_map_pages().
708 */
709static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
710 struct page **pages, unsigned long *populated,
711 int page_start, int page_end)
712{
713 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
714 unsigned int cpu;
715 int i;
716
717 for_each_possible_cpu(cpu) {
718 for (i = page_start; i < page_end; i++) {
719 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
720
721 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
722 if (!*pagep) {
723 pcpu_free_pages(chunk, pages, populated,
724 page_start, page_end);
725 return -ENOMEM;
726 }
727 }
728 }
729 return 0;
730}
731
732/**
733 * pcpu_pre_unmap_flush - flush cache prior to unmapping
734 * @chunk: chunk the regions to be flushed belongs to
735 * @page_start: page index of the first page to be flushed
736 * @page_end: page index of the last page to be flushed + 1
737 *
738 * Pages in [@page_start,@page_end) of @chunk are about to be
739 * unmapped. Flush cache. As each flushing trial can be very
740 * expensive, issue flush on the whole region at once rather than
741 * doing it for each cpu. This could be an overkill but is more
742 * scalable.
743 */
744static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
745 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900746{
Tejun Heo2f39e632009-07-04 08:11:00 +0900747 flush_cache_vunmap(
748 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
749 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heoce3141a2009-07-04 08:11:00 +0900750}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900751
Tejun Heoce3141a2009-07-04 08:11:00 +0900752static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
753{
754 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
755}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900756
Tejun Heoce3141a2009-07-04 08:11:00 +0900757/**
758 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
759 * @chunk: chunk of interest
760 * @pages: pages array which can be used to pass information to free
761 * @populated: populated bitmap
Tejun Heofbf59bc2009-02-20 16:29:08 +0900762 * @page_start: page index of the first page to unmap
763 * @page_end: page index of the last page to unmap + 1
Tejun Heofbf59bc2009-02-20 16:29:08 +0900764 *
765 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
Tejun Heoce3141a2009-07-04 08:11:00 +0900766 * Corresponding elements in @pages were cleared by the caller and can
767 * be used to carry information to pcpu_free_pages() which will be
768 * called after all unmaps are finished. The caller should call
769 * proper pre/post flush functions.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900770 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900771static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
772 struct page **pages, unsigned long *populated,
773 int page_start, int page_end)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900774{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900775 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900776 int i;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900777
Tejun Heoce3141a2009-07-04 08:11:00 +0900778 for_each_possible_cpu(cpu) {
779 for (i = page_start; i < page_end; i++) {
780 struct page *page;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900781
Tejun Heoce3141a2009-07-04 08:11:00 +0900782 page = pcpu_chunk_page(chunk, cpu, i);
783 WARN_ON(!page);
784 pages[pcpu_page_idx(cpu, i)] = page;
785 }
786 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
787 page_end - page_start);
788 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900789
Tejun Heoce3141a2009-07-04 08:11:00 +0900790 for (i = page_start; i < page_end; i++)
791 __clear_bit(i, populated);
792}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900793
Tejun Heoce3141a2009-07-04 08:11:00 +0900794/**
795 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
796 * @chunk: pcpu_chunk the regions to be flushed belong to
797 * @page_start: page index of the first page to be flushed
798 * @page_end: page index of the last page to be flushed + 1
799 *
800 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
801 * TLB for the regions. This can be skipped if the area is to be
802 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
803 *
804 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
805 * for the whole region.
806 */
807static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
808 int page_start, int page_end)
809{
Tejun Heo2f39e632009-07-04 08:11:00 +0900810 flush_tlb_kernel_range(
811 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
812 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900813}
814
Tejun Heoc8a51be2009-07-04 08:10:59 +0900815static int __pcpu_map_pages(unsigned long addr, struct page **pages,
816 int nr_pages)
817{
818 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
819 PAGE_KERNEL, pages);
820}
821
822/**
Tejun Heoce3141a2009-07-04 08:11:00 +0900823 * pcpu_map_pages - map pages into a pcpu_chunk
Tejun Heoc8a51be2009-07-04 08:10:59 +0900824 * @chunk: chunk of interest
Tejun Heoce3141a2009-07-04 08:11:00 +0900825 * @pages: pages array containing pages to be mapped
826 * @populated: populated bitmap
Tejun Heoc8a51be2009-07-04 08:10:59 +0900827 * @page_start: page index of the first page to map
828 * @page_end: page index of the last page to map + 1
829 *
Tejun Heoce3141a2009-07-04 08:11:00 +0900830 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
831 * caller is responsible for calling pcpu_post_map_flush() after all
832 * mappings are complete.
833 *
834 * This function is responsible for setting corresponding bits in
835 * @chunk->populated bitmap and whatever is necessary for reverse
836 * lookup (addr -> chunk).
Tejun Heoc8a51be2009-07-04 08:10:59 +0900837 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900838static int pcpu_map_pages(struct pcpu_chunk *chunk,
839 struct page **pages, unsigned long *populated,
840 int page_start, int page_end)
Tejun Heoc8a51be2009-07-04 08:10:59 +0900841{
Tejun Heoce3141a2009-07-04 08:11:00 +0900842 unsigned int cpu, tcpu;
843 int i, err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900844
845 for_each_possible_cpu(cpu) {
846 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
Tejun Heoce3141a2009-07-04 08:11:00 +0900847 &pages[pcpu_page_idx(cpu, page_start)],
Tejun Heoc8a51be2009-07-04 08:10:59 +0900848 page_end - page_start);
849 if (err < 0)
Tejun Heoce3141a2009-07-04 08:11:00 +0900850 goto err;
Tejun Heoc8a51be2009-07-04 08:10:59 +0900851 }
852
Tejun Heoce3141a2009-07-04 08:11:00 +0900853 /* mapping successful, link chunk and mark populated */
854 for (i = page_start; i < page_end; i++) {
855 for_each_possible_cpu(cpu)
856 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
857 chunk);
858 __set_bit(i, populated);
859 }
860
861 return 0;
862
863err:
864 for_each_possible_cpu(tcpu) {
865 if (tcpu == cpu)
866 break;
867 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
868 page_end - page_start);
869 }
870 return err;
871}
872
873/**
874 * pcpu_post_map_flush - flush cache after mapping
875 * @chunk: pcpu_chunk the regions to be flushed belong to
876 * @page_start: page index of the first page to be flushed
877 * @page_end: page index of the last page to be flushed + 1
878 *
879 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
880 * cache.
881 *
882 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
883 * for the whole region.
884 */
885static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
886 int page_start, int page_end)
887{
Tejun Heo2f39e632009-07-04 08:11:00 +0900888 flush_cache_vmap(
889 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
890 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900891}
892
893/**
894 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
895 * @chunk: chunk to depopulate
896 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900897 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900898 * @flush: whether to flush cache and tlb or not
899 *
900 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
901 * from @chunk. If @flush is true, vcache is flushed before unmapping
902 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900903 *
904 * CONTEXT:
905 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900906 */
Tejun Heoce3141a2009-07-04 08:11:00 +0900907static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900908{
909 int page_start = PFN_DOWN(off);
910 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900911 struct page **pages;
912 unsigned long *populated;
913 int rs, re;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900914
Tejun Heoce3141a2009-07-04 08:11:00 +0900915 /* quick path, check whether it's empty already */
916 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
917 if (rs == page_start && re == page_end)
918 return;
919 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900920 }
921
Tejun Heoce3141a2009-07-04 08:11:00 +0900922 /* immutable chunks can't be depopulated */
Tejun Heo8d408b42009-02-24 11:57:21 +0900923 WARN_ON(chunk->immutable);
924
Tejun Heoce3141a2009-07-04 08:11:00 +0900925 /*
926 * If control reaches here, there must have been at least one
927 * successful population attempt so the temp pages array must
928 * be available now.
929 */
930 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
931 BUG_ON(!pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900932
Tejun Heoce3141a2009-07-04 08:11:00 +0900933 /* unmap and free */
934 pcpu_pre_unmap_flush(chunk, page_start, page_end);
935
936 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
937 pcpu_unmap_pages(chunk, pages, populated, rs, re);
938
939 /* no need to flush tlb, vmalloc will handle it lazily */
940
941 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
942 pcpu_free_pages(chunk, pages, populated, rs, re);
943
944 /* commit new bitmap */
945 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900946}
947
948/**
949 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
950 * @chunk: chunk of interest
951 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900952 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900953 *
954 * For each cpu, populate and map pages [@page_start,@page_end) into
955 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900956 *
957 * CONTEXT:
958 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900959 */
960static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
961{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900962 int page_start = PFN_DOWN(off);
963 int page_end = PFN_UP(off + size);
Tejun Heoce3141a2009-07-04 08:11:00 +0900964 int free_end = page_start, unmap_end = page_start;
965 struct page **pages;
966 unsigned long *populated;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900967 unsigned int cpu;
Tejun Heoce3141a2009-07-04 08:11:00 +0900968 int rs, re, rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900969
Tejun Heoce3141a2009-07-04 08:11:00 +0900970 /* quick path, check whether all pages are already there */
971 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
972 if (rs == page_start && re == page_end)
973 goto clear;
974 break;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900975 }
976
Tejun Heoce3141a2009-07-04 08:11:00 +0900977 /* need to allocate and map pages, this chunk can't be immutable */
978 WARN_ON(chunk->immutable);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900979
Tejun Heoce3141a2009-07-04 08:11:00 +0900980 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
981 if (!pages)
982 return -ENOMEM;
983
984 /* alloc and map */
985 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
986 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
987 if (rc)
988 goto err_free;
989 free_end = re;
990 }
991
992 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
993 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
994 if (rc)
995 goto err_unmap;
996 unmap_end = re;
997 }
998 pcpu_post_map_flush(chunk, page_start, page_end);
999
1000 /* commit new bitmap */
1001 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
1002clear:
Tejun Heofbf59bc2009-02-20 16:29:08 +09001003 for_each_possible_cpu(cpu)
Tejun Heo2f39e632009-07-04 08:11:00 +09001004 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001005 return 0;
Tejun Heoce3141a2009-07-04 08:11:00 +09001006
1007err_unmap:
1008 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
1009 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
1010 pcpu_unmap_pages(chunk, pages, populated, rs, re);
1011 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
1012err_free:
1013 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
1014 pcpu_free_pages(chunk, pages, populated, rs, re);
1015 return rc;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001016}
1017
1018static void free_pcpu_chunk(struct pcpu_chunk *chunk)
1019{
1020 if (!chunk)
1021 return;
Tejun Heo65632972009-08-14 15:00:52 +09001022 if (chunk->vms)
1023 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
Tejun Heo1880d932009-03-07 00:44:09 +09001024 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001025 kfree(chunk);
1026}
1027
1028static struct pcpu_chunk *alloc_pcpu_chunk(void)
1029{
1030 struct pcpu_chunk *chunk;
1031
1032 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1033 if (!chunk)
1034 return NULL;
1035
Tejun Heo1880d932009-03-07 00:44:09 +09001036 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +09001037 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1038 chunk->map[chunk->map_used++] = pcpu_unit_size;
1039
Tejun Heo65632972009-08-14 15:00:52 +09001040 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1041 pcpu_nr_groups, pcpu_atom_size,
1042 GFP_KERNEL);
1043 if (!chunk->vms) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001044 free_pcpu_chunk(chunk);
1045 return NULL;
1046 }
1047
1048 INIT_LIST_HEAD(&chunk->list);
1049 chunk->free_size = pcpu_unit_size;
1050 chunk->contig_hint = pcpu_unit_size;
Tejun Heo65632972009-08-14 15:00:52 +09001051 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
Tejun Heofbf59bc2009-02-20 16:29:08 +09001052
1053 return chunk;
1054}
1055
1056/**
Tejun Heoedcb4632009-03-06 14:33:59 +09001057 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +09001058 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +09001059 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +09001060 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +09001061 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001062 * Allocate percpu area of @size bytes aligned at @align.
1063 *
1064 * CONTEXT:
1065 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001066 *
1067 * RETURNS:
1068 * Percpu pointer to the allocated area on success, NULL on failure.
1069 */
Tejun Heoedcb4632009-03-06 14:33:59 +09001070static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001071{
Tejun Heof2badb02009-09-29 09:17:58 +09001072 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001073 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +09001074 const char *err;
Tejun Heo833af842009-11-11 15:35:18 +09001075 int slot, off, new_alloc;
Jiri Kosina403a91b2009-10-29 00:25:59 +09001076 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001077
Tejun Heo8d408b42009-02-24 11:57:21 +09001078 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +09001079 WARN(true, "illegal size (%zu) or align (%zu) for "
1080 "percpu allocation\n", size, align);
1081 return NULL;
1082 }
1083
Tejun Heoccea34b2009-03-07 00:44:13 +09001084 mutex_lock(&pcpu_alloc_mutex);
Jiri Kosina403a91b2009-10-29 00:25:59 +09001085 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001086
Tejun Heoedcb4632009-03-06 14:33:59 +09001087 /* serve reserved allocations from the reserved chunk if available */
1088 if (reserved && pcpu_reserved_chunk) {
1089 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +09001090
1091 if (size > chunk->contig_hint) {
1092 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001093 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +09001094 }
Tejun Heo833af842009-11-11 15:35:18 +09001095
1096 while ((new_alloc = pcpu_need_to_extend(chunk))) {
1097 spin_unlock_irqrestore(&pcpu_lock, flags);
1098 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
1099 err = "failed to extend area map of reserved chunk";
1100 goto fail_unlock_mutex;
1101 }
1102 spin_lock_irqsave(&pcpu_lock, flags);
1103 }
1104
Tejun Heoedcb4632009-03-06 14:33:59 +09001105 off = pcpu_alloc_area(chunk, size, align);
1106 if (off >= 0)
1107 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +09001108
Tejun Heof2badb02009-09-29 09:17:58 +09001109 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +09001110 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +09001111 }
1112
Tejun Heoccea34b2009-03-07 00:44:13 +09001113restart:
Tejun Heoedcb4632009-03-06 14:33:59 +09001114 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001115 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1116 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1117 if (size > chunk->contig_hint)
1118 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +09001119
Tejun Heo833af842009-11-11 15:35:18 +09001120 new_alloc = pcpu_need_to_extend(chunk);
1121 if (new_alloc) {
1122 spin_unlock_irqrestore(&pcpu_lock, flags);
1123 if (pcpu_extend_area_map(chunk,
1124 new_alloc) < 0) {
1125 err = "failed to extend area map";
1126 goto fail_unlock_mutex;
1127 }
1128 spin_lock_irqsave(&pcpu_lock, flags);
1129 /*
1130 * pcpu_lock has been dropped, need to
1131 * restart cpu_slot list walking.
1132 */
1133 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +09001134 }
1135
Tejun Heofbf59bc2009-02-20 16:29:08 +09001136 off = pcpu_alloc_area(chunk, size, align);
1137 if (off >= 0)
1138 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001139 }
1140 }
1141
1142 /* hmmm... no space left, create a new chunk */
Jiri Kosina403a91b2009-10-29 00:25:59 +09001143 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001144
Tejun Heofbf59bc2009-02-20 16:29:08 +09001145 chunk = alloc_pcpu_chunk();
Tejun Heof2badb02009-09-29 09:17:58 +09001146 if (!chunk) {
1147 err = "failed to allocate new chunk";
Tejun Heoccea34b2009-03-07 00:44:13 +09001148 goto fail_unlock_mutex;
Tejun Heof2badb02009-09-29 09:17:58 +09001149 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001150
Jiri Kosina403a91b2009-10-29 00:25:59 +09001151 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001152 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +09001153 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001154
1155area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001156 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001157
Tejun Heofbf59bc2009-02-20 16:29:08 +09001158 /* populate, map and clear the area */
1159 if (pcpu_populate_chunk(chunk, off, size)) {
Jiri Kosina403a91b2009-10-29 00:25:59 +09001160 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001161 pcpu_free_area(chunk, off);
Tejun Heof2badb02009-09-29 09:17:58 +09001162 err = "failed to populate";
Tejun Heoccea34b2009-03-07 00:44:13 +09001163 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001164 }
1165
Tejun Heoccea34b2009-03-07 00:44:13 +09001166 mutex_unlock(&pcpu_alloc_mutex);
1167
Tejun Heobba174f2009-08-14 15:00:51 +09001168 /* return address relative to base address */
1169 return __addr_to_pcpu_ptr(chunk->base_addr + off);
Tejun Heoccea34b2009-03-07 00:44:13 +09001170
1171fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001172 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +09001173fail_unlock_mutex:
1174 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heof2badb02009-09-29 09:17:58 +09001175 if (warn_limit) {
1176 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1177 "%s\n", size, align, err);
1178 dump_stack();
1179 if (!--warn_limit)
1180 pr_info("PERCPU: limit reached, disable warning\n");
1181 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001182 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001183}
Tejun Heoedcb4632009-03-06 14:33:59 +09001184
1185/**
1186 * __alloc_percpu - allocate dynamic percpu area
1187 * @size: size of area to allocate in bytes
1188 * @align: alignment of area (max PAGE_SIZE)
1189 *
1190 * Allocate percpu area of @size bytes aligned at @align. Might
1191 * sleep. Might trigger writeouts.
1192 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001193 * CONTEXT:
1194 * Does GFP_KERNEL allocation.
1195 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001196 * RETURNS:
1197 * Percpu pointer to the allocated area on success, NULL on failure.
1198 */
1199void *__alloc_percpu(size_t size, size_t align)
1200{
1201 return pcpu_alloc(size, align, false);
1202}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001203EXPORT_SYMBOL_GPL(__alloc_percpu);
1204
Tejun Heoedcb4632009-03-06 14:33:59 +09001205/**
1206 * __alloc_reserved_percpu - allocate reserved percpu area
1207 * @size: size of area to allocate in bytes
1208 * @align: alignment of area (max PAGE_SIZE)
1209 *
1210 * Allocate percpu area of @size bytes aligned at @align from reserved
1211 * percpu area if arch has set it up; otherwise, allocation is served
1212 * from the same dynamic area. Might sleep. Might trigger writeouts.
1213 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001214 * CONTEXT:
1215 * Does GFP_KERNEL allocation.
1216 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001217 * RETURNS:
1218 * Percpu pointer to the allocated area on success, NULL on failure.
1219 */
1220void *__alloc_reserved_percpu(size_t size, size_t align)
1221{
1222 return pcpu_alloc(size, align, true);
1223}
1224
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001225/**
1226 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1227 * @work: unused
1228 *
1229 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +09001230 *
1231 * CONTEXT:
1232 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001233 */
1234static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001235{
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001236 LIST_HEAD(todo);
1237 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1238 struct pcpu_chunk *chunk, *next;
1239
Tejun Heoccea34b2009-03-07 00:44:13 +09001240 mutex_lock(&pcpu_alloc_mutex);
1241 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001242
1243 list_for_each_entry_safe(chunk, next, head, list) {
1244 WARN_ON(chunk->immutable);
1245
1246 /* spare the first one */
1247 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1248 continue;
1249
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001250 list_move(&chunk->list, &todo);
1251 }
1252
Tejun Heoccea34b2009-03-07 00:44:13 +09001253 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001254
1255 list_for_each_entry_safe(chunk, next, &todo, list) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001256 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001257 free_pcpu_chunk(chunk);
1258 }
Tejun Heo971f3912009-08-14 15:00:49 +09001259
1260 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001261}
1262
1263/**
1264 * free_percpu - free percpu area
1265 * @ptr: pointer to area to free
1266 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001267 * Free percpu area @ptr.
1268 *
1269 * CONTEXT:
1270 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001271 */
1272void free_percpu(void *ptr)
1273{
1274 void *addr = __pcpu_ptr_to_addr(ptr);
1275 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001276 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001277 int off;
1278
1279 if (!ptr)
1280 return;
1281
Tejun Heoccea34b2009-03-07 00:44:13 +09001282 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001283
1284 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001285 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001286
1287 pcpu_free_area(chunk, off);
1288
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001289 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001290 if (chunk->free_size == pcpu_unit_size) {
1291 struct pcpu_chunk *pos;
1292
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001293 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001294 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001295 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001296 break;
1297 }
1298 }
1299
Tejun Heoccea34b2009-03-07 00:44:13 +09001300 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001301}
1302EXPORT_SYMBOL_GPL(free_percpu);
1303
Vivek Goyal3b034b02009-11-24 15:50:03 +09001304/**
1305 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1306 * @addr: the address to be converted to physical address
1307 *
1308 * Given @addr which is dereferenceable address obtained via one of
1309 * percpu access macros, this function translates it into its physical
1310 * address. The caller is responsible for ensuring @addr stays valid
1311 * until this function finishes.
1312 *
1313 * RETURNS:
1314 * The physical address for @addr.
1315 */
1316phys_addr_t per_cpu_ptr_to_phys(void *addr)
1317{
1318 if ((unsigned long)addr < VMALLOC_START ||
1319 (unsigned long)addr >= VMALLOC_END)
1320 return __pa(addr);
1321 else
1322 return page_to_phys(vmalloc_to_page(addr));
1323}
1324
Tejun Heo033e48f2009-08-14 15:00:51 +09001325static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1326 size_t reserved_size,
1327 ssize_t *dyn_sizep)
1328{
1329 size_t size_sum;
1330
1331 size_sum = PFN_ALIGN(static_size + reserved_size +
1332 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1333 if (*dyn_sizep != 0)
1334 *dyn_sizep = size_sum - static_size - reserved_size;
1335
1336 return size_sum;
1337}
1338
Tejun Heofbf59bc2009-02-20 16:29:08 +09001339/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001340 * pcpu_alloc_alloc_info - allocate percpu allocation info
1341 * @nr_groups: the number of groups
1342 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001343 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001344 * Allocate ai which is large enough for @nr_groups groups containing
1345 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1346 * cpu_map array which is long enough for @nr_units and filled with
1347 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1348 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001349 *
1350 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001351 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1352 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001353 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001354struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1355 int nr_units)
1356{
1357 struct pcpu_alloc_info *ai;
1358 size_t base_size, ai_size;
1359 void *ptr;
1360 int unit;
1361
1362 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1363 __alignof__(ai->groups[0].cpu_map[0]));
1364 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1365
1366 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1367 if (!ptr)
1368 return NULL;
1369 ai = ptr;
1370 ptr += base_size;
1371
1372 ai->groups[0].cpu_map = ptr;
1373
1374 for (unit = 0; unit < nr_units; unit++)
1375 ai->groups[0].cpu_map[unit] = NR_CPUS;
1376
1377 ai->nr_groups = nr_groups;
1378 ai->__ai_size = PFN_ALIGN(ai_size);
1379
1380 return ai;
1381}
1382
1383/**
1384 * pcpu_free_alloc_info - free percpu allocation info
1385 * @ai: pcpu_alloc_info to free
1386 *
1387 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1388 */
1389void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1390{
1391 free_bootmem(__pa(ai), ai->__ai_size);
1392}
1393
1394/**
1395 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
Tejun Heoedcb4632009-03-06 14:33:59 +09001396 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +09001397 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heofd1e8a12009-08-14 15:00:51 +09001398 * @atom_size: allocation atom size
1399 * @cpu_distance_fn: callback to determine distance between cpus, optional
1400 *
1401 * This function determines grouping of units, their mappings to cpus
1402 * and other parameters considering needed percpu size, allocation
1403 * atom size and distances between CPUs.
1404 *
1405 * Groups are always mutliples of atom size and CPUs which are of
1406 * LOCAL_DISTANCE both ways are grouped together and share space for
1407 * units in the same group. The returned configuration is guaranteed
1408 * to have CPUs on different nodes on different groups and >=75% usage
1409 * of allocated virtual address space.
1410 *
1411 * RETURNS:
1412 * On success, pointer to the new allocation_info is returned. On
1413 * failure, ERR_PTR value is returned.
1414 */
1415struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1416 size_t reserved_size, ssize_t dyn_size,
1417 size_t atom_size,
1418 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
Tejun Heo033e48f2009-08-14 15:00:51 +09001419{
1420 static int group_map[NR_CPUS] __initdata;
1421 static int group_cnt[NR_CPUS] __initdata;
1422 const size_t static_size = __per_cpu_end - __per_cpu_start;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001423 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
Tejun Heo033e48f2009-08-14 15:00:51 +09001424 size_t size_sum, min_unit_size, alloc_size;
1425 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001426 int last_allocs, group, unit;
Tejun Heo033e48f2009-08-14 15:00:51 +09001427 unsigned int cpu, tcpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001428 struct pcpu_alloc_info *ai;
1429 unsigned int *cpu_map;
Tejun Heo033e48f2009-08-14 15:00:51 +09001430
Tejun Heofb59e722009-09-24 18:50:34 +09001431 /* this function may be called multiple times */
1432 memset(group_map, 0, sizeof(group_map));
1433 memset(group_cnt, 0, sizeof(group_map));
1434
Tejun Heo033e48f2009-08-14 15:00:51 +09001435 /*
1436 * Determine min_unit_size, alloc_size and max_upa such that
Tejun Heofd1e8a12009-08-14 15:00:51 +09001437 * alloc_size is multiple of atom_size and is the smallest
Tejun Heo033e48f2009-08-14 15:00:51 +09001438 * which can accomodate 4k aligned segments which are equal to
1439 * or larger than min_unit_size.
1440 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001441 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001442 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1443
Tejun Heofd1e8a12009-08-14 15:00:51 +09001444 alloc_size = roundup(min_unit_size, atom_size);
Tejun Heo033e48f2009-08-14 15:00:51 +09001445 upa = alloc_size / min_unit_size;
1446 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1447 upa--;
1448 max_upa = upa;
1449
1450 /* group cpus according to their proximity */
1451 for_each_possible_cpu(cpu) {
1452 group = 0;
1453 next_group:
1454 for_each_possible_cpu(tcpu) {
1455 if (cpu == tcpu)
1456 break;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001457 if (group_map[tcpu] == group && cpu_distance_fn &&
Tejun Heo033e48f2009-08-14 15:00:51 +09001458 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1459 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1460 group++;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001461 nr_groups = max(nr_groups, group + 1);
Tejun Heo033e48f2009-08-14 15:00:51 +09001462 goto next_group;
1463 }
1464 }
1465 group_map[cpu] = group;
1466 group_cnt[group]++;
1467 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1468 }
1469
1470 /*
1471 * Expand unit size until address space usage goes over 75%
1472 * and then as much as possible without using more address
1473 * space.
1474 */
1475 last_allocs = INT_MAX;
1476 for (upa = max_upa; upa; upa--) {
1477 int allocs = 0, wasted = 0;
1478
1479 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1480 continue;
1481
Tejun Heofd1e8a12009-08-14 15:00:51 +09001482 for (group = 0; group < nr_groups; group++) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001483 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1484 allocs += this_allocs;
1485 wasted += this_allocs * upa - group_cnt[group];
1486 }
1487
1488 /*
1489 * Don't accept if wastage is over 25%. The
1490 * greater-than comparison ensures upa==1 always
1491 * passes the following check.
1492 */
1493 if (wasted > num_possible_cpus() / 3)
1494 continue;
1495
1496 /* and then don't consume more memory */
1497 if (allocs > last_allocs)
1498 break;
1499 last_allocs = allocs;
1500 best_upa = upa;
1501 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001502 upa = best_upa;
Tejun Heo033e48f2009-08-14 15:00:51 +09001503
Tejun Heofd1e8a12009-08-14 15:00:51 +09001504 /* allocate and fill alloc_info */
1505 for (group = 0; group < nr_groups; group++)
1506 nr_units += roundup(group_cnt[group], upa);
1507
1508 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1509 if (!ai)
1510 return ERR_PTR(-ENOMEM);
1511 cpu_map = ai->groups[0].cpu_map;
1512
1513 for (group = 0; group < nr_groups; group++) {
1514 ai->groups[group].cpu_map = cpu_map;
1515 cpu_map += roundup(group_cnt[group], upa);
Tejun Heo033e48f2009-08-14 15:00:51 +09001516 }
1517
Tejun Heofd1e8a12009-08-14 15:00:51 +09001518 ai->static_size = static_size;
1519 ai->reserved_size = reserved_size;
1520 ai->dyn_size = dyn_size;
1521 ai->unit_size = alloc_size / upa;
1522 ai->atom_size = atom_size;
1523 ai->alloc_size = alloc_size;
1524
1525 for (group = 0, unit = 0; group_cnt[group]; group++) {
1526 struct pcpu_group_info *gi = &ai->groups[group];
1527
1528 /*
1529 * Initialize base_offset as if all groups are located
1530 * back-to-back. The caller should update this to
1531 * reflect actual allocation.
1532 */
1533 gi->base_offset = unit * ai->unit_size;
1534
1535 for_each_possible_cpu(cpu)
1536 if (group_map[cpu] == group)
1537 gi->cpu_map[gi->nr_units++] = cpu;
1538 gi->nr_units = roundup(gi->nr_units, upa);
1539 unit += gi->nr_units;
1540 }
1541 BUG_ON(unit != nr_units);
1542
1543 return ai;
Tejun Heo033e48f2009-08-14 15:00:51 +09001544}
1545
Tejun Heofd1e8a12009-08-14 15:00:51 +09001546/**
1547 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1548 * @lvl: loglevel
1549 * @ai: allocation info to dump
1550 *
1551 * Print out information about @ai using loglevel @lvl.
1552 */
1553static void pcpu_dump_alloc_info(const char *lvl,
1554 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001555{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001556 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001557 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001558 int alloc = 0, alloc_end = 0;
1559 int group, v;
1560 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001561
Tejun Heofd1e8a12009-08-14 15:00:51 +09001562 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001563 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001564 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001565
Tejun Heofd1e8a12009-08-14 15:00:51 +09001566 v = num_possible_cpus();
1567 while (v /= 10)
1568 cpu_width++;
1569 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001570
Tejun Heofd1e8a12009-08-14 15:00:51 +09001571 upa = ai->alloc_size / ai->unit_size;
1572 width = upa * (cpu_width + 1) + group_width + 3;
1573 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001574
Tejun Heofd1e8a12009-08-14 15:00:51 +09001575 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1576 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1577 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1578
1579 for (group = 0; group < ai->nr_groups; group++) {
1580 const struct pcpu_group_info *gi = &ai->groups[group];
1581 int unit = 0, unit_end = 0;
1582
1583 BUG_ON(gi->nr_units % upa);
1584 for (alloc_end += gi->nr_units / upa;
1585 alloc < alloc_end; alloc++) {
1586 if (!(alloc % apl)) {
Tejun Heo033e48f2009-08-14 15:00:51 +09001587 printk("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001588 printk("%spcpu-alloc: ", lvl);
1589 }
1590 printk("[%0*d] ", group_width, group);
1591
1592 for (unit_end += upa; unit < unit_end; unit++)
1593 if (gi->cpu_map[unit] != NR_CPUS)
1594 printk("%0*d ", cpu_width,
1595 gi->cpu_map[unit]);
1596 else
1597 printk("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001598 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001599 }
1600 printk("\n");
1601}
Tejun Heo033e48f2009-08-14 15:00:51 +09001602
Tejun Heofbf59bc2009-02-20 16:29:08 +09001603/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001604 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001605 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001606 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001607 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001608 * Initialize the first percpu chunk which contains the kernel static
1609 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001610 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001611 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001612 * @ai contains all information necessary to initialize the first
1613 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001614 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001615 * @ai->static_size is the size of static percpu area.
1616 *
1617 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001618 * reserve after the static area in the first chunk. This reserves
1619 * the first chunk such that it's available only through reserved
1620 * percpu allocation. This is primarily used to serve module percpu
1621 * static areas on architectures where the addressing model has
1622 * limited offset range for symbol relocations to guarantee module
1623 * percpu symbols fall inside the relocatable range.
1624 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001625 * @ai->dyn_size determines the number of bytes available for dynamic
1626 * allocation in the first chunk. The area between @ai->static_size +
1627 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001628 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001629 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1630 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1631 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001632 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001633 * @ai->atom_size is the allocation atom size and used as alignment
1634 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001635 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001636 * @ai->alloc_size is the allocation size and always multiple of
1637 * @ai->atom_size. This is larger than @ai->atom_size if
1638 * @ai->unit_size is larger than @ai->atom_size.
1639 *
1640 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1641 * percpu areas. Units which should be colocated are put into the
1642 * same group. Dynamic VM areas will be allocated according to these
1643 * groupings. If @ai->nr_groups is zero, a single group containing
1644 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001645 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001646 * The caller should have mapped the first chunk at @base_addr and
1647 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001648 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001649 * If the first chunk ends up with both reserved and dynamic areas, it
1650 * is served by two chunks - one to serve the core static and reserved
1651 * areas and the other for the dynamic area. They share the same vm
1652 * and page map but uses different area allocation map to stay away
1653 * from each other. The latter chunk is circulated in the chunk slots
1654 * and available for dynamic allocation like any other chunks.
1655 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001656 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001657 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001658 */
Tejun Heofb435d52009-08-14 15:00:51 +09001659int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1660 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001661{
Tejun Heo635b75f2009-09-24 09:43:11 +09001662 static char cpus_buf[4096] __initdata;
Tejun Heoedcb4632009-03-06 14:33:59 +09001663 static int smap[2], dmap[2];
Tejun Heofd1e8a12009-08-14 15:00:51 +09001664 size_t dyn_size = ai->dyn_size;
1665 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001666 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001667 unsigned long *group_offsets;
1668 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001669 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001670 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001671 int *unit_map;
1672 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001673
Tejun Heo635b75f2009-09-24 09:43:11 +09001674 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1675
1676#define PCPU_SETUP_BUG_ON(cond) do { \
1677 if (unlikely(cond)) { \
1678 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1679 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1680 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1681 BUG(); \
1682 } \
1683} while (0)
1684
Tejun Heo2f39e632009-07-04 08:11:00 +09001685 /* sanity checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001686 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1687 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo635b75f2009-09-24 09:43:11 +09001688 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1689 PCPU_SETUP_BUG_ON(!ai->static_size);
1690 PCPU_SETUP_BUG_ON(!base_addr);
1691 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1692 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1693 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001694
Tejun Heo65632972009-08-14 15:00:52 +09001695 /* process group information and build config tables accordingly */
1696 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1697 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
Tejun Heofd1e8a12009-08-14 15:00:51 +09001698 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
Tejun Heofb435d52009-08-14 15:00:51 +09001699 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
Tejun Heo2f39e632009-07-04 08:11:00 +09001700
Tejun Heofd1e8a12009-08-14 15:00:51 +09001701 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001702 unit_map[cpu] = UINT_MAX;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001703 pcpu_first_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001704
Tejun Heofd1e8a12009-08-14 15:00:51 +09001705 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1706 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001707
Tejun Heo65632972009-08-14 15:00:52 +09001708 group_offsets[group] = gi->base_offset;
1709 group_sizes[group] = gi->nr_units * ai->unit_size;
1710
Tejun Heofd1e8a12009-08-14 15:00:51 +09001711 for (i = 0; i < gi->nr_units; i++) {
1712 cpu = gi->cpu_map[i];
1713 if (cpu == NR_CPUS)
1714 continue;
1715
Tejun Heo635b75f2009-09-24 09:43:11 +09001716 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1717 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1718 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001719
1720 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001721 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1722
Tejun Heofd1e8a12009-08-14 15:00:51 +09001723 if (pcpu_first_unit_cpu == NR_CPUS)
Tejun Heo2f39e632009-07-04 08:11:00 +09001724 pcpu_first_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001725 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001726 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001727 pcpu_last_unit_cpu = cpu;
1728 pcpu_nr_units = unit;
1729
1730 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001731 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1732
1733 /* we're done parsing the input, undefine BUG macro and dump config */
1734#undef PCPU_SETUP_BUG_ON
1735 pcpu_dump_alloc_info(KERN_INFO, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001736
Tejun Heo65632972009-08-14 15:00:52 +09001737 pcpu_nr_groups = ai->nr_groups;
1738 pcpu_group_offsets = group_offsets;
1739 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001740 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001741 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001742
1743 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001744 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001745 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001746 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001747 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1748 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001749
Tejun Heod9b55ee2009-02-24 11:57:21 +09001750 /*
1751 * Allocate chunk slots. The additional last slot is for
1752 * empty chunks.
1753 */
1754 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001755 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1756 for (i = 0; i < pcpu_nr_slots; i++)
1757 INIT_LIST_HEAD(&pcpu_slot[i]);
1758
Tejun Heoedcb4632009-03-06 14:33:59 +09001759 /*
1760 * Initialize static chunk. If reserved_size is zero, the
1761 * static chunk covers static area + dynamic allocation area
1762 * in the first chunk. If reserved_size is not zero, it
1763 * covers static area + reserved area (mostly used for module
1764 * static percpu allocation).
1765 */
Tejun Heo2441d152009-03-06 14:33:59 +09001766 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1767 INIT_LIST_HEAD(&schunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001768 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001769 schunk->map = smap;
1770 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001771 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001772 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001773
Tejun Heofd1e8a12009-08-14 15:00:51 +09001774 if (ai->reserved_size) {
1775 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001776 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001777 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001778 } else {
1779 schunk->free_size = dyn_size;
1780 dyn_size = 0; /* dynamic area covered */
1781 }
Tejun Heo2441d152009-03-06 14:33:59 +09001782 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001783
Tejun Heofd1e8a12009-08-14 15:00:51 +09001784 schunk->map[schunk->map_used++] = -ai->static_size;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001785 if (schunk->free_size)
1786 schunk->map[schunk->map_used++] = schunk->free_size;
1787
Tejun Heoedcb4632009-03-06 14:33:59 +09001788 /* init dynamic chunk if necessary */
1789 if (dyn_size) {
Tejun Heoce3141a2009-07-04 08:11:00 +09001790 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
Tejun Heoedcb4632009-03-06 14:33:59 +09001791 INIT_LIST_HEAD(&dchunk->list);
Tejun Heobba174f2009-08-14 15:00:51 +09001792 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001793 dchunk->map = dmap;
1794 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001795 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001796 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +09001797
1798 dchunk->contig_hint = dchunk->free_size = dyn_size;
1799 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1800 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1801 }
1802
Tejun Heo2441d152009-03-06 14:33:59 +09001803 /* link the first chunk in */
Tejun Heoae9e6bc2009-04-02 13:19:54 +09001804 pcpu_first_chunk = dchunk ?: schunk;
1805 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001806
1807 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001808 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001809 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001810}
Tejun Heo66c3a752009-03-10 16:27:48 +09001811
Tejun Heof58dc012009-08-14 15:00:50 +09001812const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1813 [PCPU_FC_AUTO] = "auto",
1814 [PCPU_FC_EMBED] = "embed",
1815 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001816};
Tejun Heo66c3a752009-03-10 16:27:48 +09001817
Tejun Heof58dc012009-08-14 15:00:50 +09001818enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1819
1820static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001821{
Tejun Heof58dc012009-08-14 15:00:50 +09001822 if (0)
1823 /* nada */;
1824#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1825 else if (!strcmp(str, "embed"))
1826 pcpu_chosen_fc = PCPU_FC_EMBED;
1827#endif
1828#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1829 else if (!strcmp(str, "page"))
1830 pcpu_chosen_fc = PCPU_FC_PAGE;
1831#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001832 else
1833 pr_warning("PERCPU: unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001834
Tejun Heof58dc012009-08-14 15:00:50 +09001835 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001836}
Tejun Heof58dc012009-08-14 15:00:50 +09001837early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001838
Tejun Heo08fc4582009-08-14 15:00:49 +09001839#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1840 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo66c3a752009-03-10 16:27:48 +09001841/**
1842 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001843 * @reserved_size: the size of reserved percpu area in bytes
1844 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heoc8826dd2009-08-14 15:00:52 +09001845 * @atom_size: allocation atom size
1846 * @cpu_distance_fn: callback to determine distance between cpus, optional
1847 * @alloc_fn: function to allocate percpu page
1848 * @free_fn: funtion to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001849 *
1850 * This is a helper to ease setting up embedded first percpu chunk and
1851 * can be called where pcpu_setup_first_chunk() is expected.
1852 *
1853 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001854 * by calling @alloc_fn and used as-is without being mapped into
1855 * vmalloc area. Allocations are always whole multiples of @atom_size
1856 * aligned to @atom_size.
1857 *
1858 * This enables the first chunk to piggy back on the linear physical
1859 * mapping which often uses larger page size. Please note that this
1860 * can result in very sparse cpu->unit mapping on NUMA machines thus
1861 * requiring large vmalloc address space. Don't use this allocator if
1862 * vmalloc space is not orders of magnitude larger than distances
1863 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001864 *
1865 * When @dyn_size is positive, dynamic area might be larger than
Tejun Heo788e5ab2009-07-04 08:10:58 +09001866 * specified to fill page alignment. When @dyn_size is auto,
1867 * @dyn_size is just big enough to fill page alignment after static
1868 * and reserved areas.
Tejun Heo66c3a752009-03-10 16:27:48 +09001869 *
1870 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001871 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001872 *
1873 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001874 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001875 */
Tejun Heoc8826dd2009-08-14 15:00:52 +09001876int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1877 size_t atom_size,
1878 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1879 pcpu_fc_alloc_fn_t alloc_fn,
1880 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001881{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001882 void *base = (void *)ULONG_MAX;
1883 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001884 struct pcpu_alloc_info *ai;
Tejun Heo6ea529a2009-09-24 18:46:01 +09001885 size_t size_sum, areas_size, max_distance;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001886 int group, i, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001887
Tejun Heoc8826dd2009-08-14 15:00:52 +09001888 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1889 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001890 if (IS_ERR(ai))
1891 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001892
Tejun Heofd1e8a12009-08-14 15:00:51 +09001893 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001894 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001895
Tejun Heoc8826dd2009-08-14 15:00:52 +09001896 areas = alloc_bootmem_nopanic(areas_size);
1897 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001898 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001899 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001900 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001901
Tejun Heoc8826dd2009-08-14 15:00:52 +09001902 /* allocate, copy and determine base address */
1903 for (group = 0; group < ai->nr_groups; group++) {
1904 struct pcpu_group_info *gi = &ai->groups[group];
1905 unsigned int cpu = NR_CPUS;
1906 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001907
Tejun Heoc8826dd2009-08-14 15:00:52 +09001908 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1909 cpu = gi->cpu_map[i];
1910 BUG_ON(cpu == NR_CPUS);
1911
1912 /* allocate space for the whole group */
1913 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1914 if (!ptr) {
1915 rc = -ENOMEM;
1916 goto out_free_areas;
1917 }
1918 areas[group] = ptr;
1919
1920 base = min(ptr, base);
1921
1922 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1923 if (gi->cpu_map[i] == NR_CPUS) {
1924 /* unused unit, free whole */
1925 free_fn(ptr, ai->unit_size);
1926 continue;
1927 }
1928 /* copy and return the unused part */
1929 memcpy(ptr, __per_cpu_load, ai->static_size);
1930 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1931 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001932 }
1933
Tejun Heoc8826dd2009-08-14 15:00:52 +09001934 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09001935 max_distance = 0;
1936 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09001937 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo1a0c3292009-10-04 09:31:05 +09001938 max_distance = max_t(size_t, max_distance,
1939 ai->groups[group].base_offset);
Tejun Heo6ea529a2009-09-24 18:46:01 +09001940 }
1941 max_distance += ai->unit_size;
1942
1943 /* warn if maximum distance is further than 75% of vmalloc space */
1944 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
Tejun Heo1a0c3292009-10-04 09:31:05 +09001945 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
Tejun Heo6ea529a2009-09-24 18:46:01 +09001946 "space 0x%lx\n",
1947 max_distance, VMALLOC_END - VMALLOC_START);
1948#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1949 /* and fail if we have fallback */
1950 rc = -EINVAL;
1951 goto out_free;
1952#endif
1953 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09001954
Tejun Heo004018e2009-08-14 15:00:49 +09001955 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09001956 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1957 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001958
Tejun Heofb435d52009-08-14 15:00:51 +09001959 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001960 goto out_free;
1961
1962out_free_areas:
1963 for (group = 0; group < ai->nr_groups; group++)
1964 free_fn(areas[group],
1965 ai->groups[group].nr_units * ai->unit_size);
1966out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001967 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001968 if (areas)
1969 free_bootmem(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09001970 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09001971}
Tejun Heo08fc4582009-08-14 15:00:49 +09001972#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1973 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
Tejun Heod4b95f82009-07-04 08:10:59 +09001974
Tejun Heo08fc4582009-08-14 15:00:49 +09001975#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09001976/**
Tejun Heo00ae4062009-08-14 15:00:49 +09001977 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09001978 * @reserved_size: the size of reserved percpu area in bytes
1979 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1980 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1981 * @populate_pte_fn: function to populate pte
1982 *
Tejun Heo00ae4062009-08-14 15:00:49 +09001983 * This is a helper to ease setting up page-remapped first percpu
1984 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09001985 *
1986 * This is the basic allocator. Static percpu area is allocated
1987 * page-by-page into vmalloc area.
1988 *
1989 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001990 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09001991 */
Tejun Heofb435d52009-08-14 15:00:51 +09001992int __init pcpu_page_first_chunk(size_t reserved_size,
1993 pcpu_fc_alloc_fn_t alloc_fn,
1994 pcpu_fc_free_fn_t free_fn,
1995 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09001996{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09001997 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001998 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09001999 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09002000 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09002001 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09002002 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09002003 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002004
Tejun Heo00ae4062009-08-14 15:00:49 +09002005 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2006
Tejun Heofd1e8a12009-08-14 15:00:51 +09002007 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
2008 if (IS_ERR(ai))
2009 return PTR_ERR(ai);
2010 BUG_ON(ai->nr_groups != 1);
2011 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2012
2013 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09002014
2015 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002016 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2017 sizeof(pages[0]));
Tejun Heoce3141a2009-07-04 08:11:00 +09002018 pages = alloc_bootmem(pages_size);
Tejun Heod4b95f82009-07-04 08:10:59 +09002019
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002020 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09002021 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002022 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09002023 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09002024 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09002025 void *ptr;
2026
Tejun Heo3cbc8562009-08-14 15:00:50 +09002027 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09002028 if (!ptr) {
Tejun Heo00ae4062009-08-14 15:00:49 +09002029 pr_warning("PERCPU: failed to allocate %s page "
2030 "for cpu%u\n", psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09002031 goto enomem;
2032 }
Tejun Heoce3141a2009-07-04 08:11:00 +09002033 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002034 }
2035
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002036 /* allocate vm area, map the pages and copy static data */
2037 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002038 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002039 vm_area_register_early(&vm, PAGE_SIZE);
2040
Tejun Heofd1e8a12009-08-14 15:00:51 +09002041 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09002042 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09002043 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002044
Tejun Heoce3141a2009-07-04 08:11:00 +09002045 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002046 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2047
2048 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09002049 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2050 unit_pages);
2051 if (rc < 0)
2052 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002053
2054 /*
2055 * FIXME: Archs with virtual cache should flush local
2056 * cache for the linear mapping here - something
2057 * equivalent to flush_cache_vmap() on the local cpu.
2058 * flush_cache_vmap() can't be used as most supporting
2059 * data structures are not set up yet.
2060 */
2061
2062 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002063 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002064 }
2065
2066 /* we're ready, commit */
Tejun Heo1d9d3252009-08-14 15:00:50 +09002067 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002068 unit_pages, psize_str, vm.addr, ai->static_size,
2069 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002070
Tejun Heofb435d52009-08-14 15:00:51 +09002071 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002072 goto out_free_ar;
2073
2074enomem:
2075 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002076 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002077 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002078out_free_ar:
Tejun Heoce3141a2009-07-04 08:11:00 +09002079 free_bootmem(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002080 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002081 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002082}
Tejun Heo08fc4582009-08-14 15:00:49 +09002083#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002084
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002085/*
Tejun Heoe74e3962009-03-30 19:07:44 +09002086 * Generic percpu area setup.
2087 *
2088 * The embedding helper is used because its behavior closely resembles
2089 * the original non-dynamic generic percpu area setup. This is
2090 * important because many archs have addressing restrictions and might
2091 * fail if the percpu area is located far away from the previous
2092 * location. As an added bonus, in non-NUMA cases, embedding is
2093 * generally a good idea TLB-wise because percpu area can piggy back
2094 * on the physical linear memory mapping which uses large page
2095 * mappings on applicable archs.
2096 */
2097#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2098unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2099EXPORT_SYMBOL(__per_cpu_offset);
2100
Tejun Heoc8826dd2009-08-14 15:00:52 +09002101static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2102 size_t align)
2103{
2104 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2105}
2106
2107static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2108{
2109 free_bootmem(__pa(ptr), size);
2110}
2111
Tejun Heoe74e3962009-03-30 19:07:44 +09002112void __init setup_per_cpu_areas(void)
2113{
Tejun Heoe74e3962009-03-30 19:07:44 +09002114 unsigned long delta;
2115 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002116 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002117
2118 /*
2119 * Always reserve area for module percpu variables. That's
2120 * what the legacy allocator did.
2121 */
Tejun Heofb435d52009-08-14 15:00:51 +09002122 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002123 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2124 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002125 if (rc < 0)
Tejun Heoe74e3962009-03-30 19:07:44 +09002126 panic("Failed to initialized percpu areas.");
2127
2128 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2129 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002130 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002131}
2132#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */