blob: 3311c8919f375fa062faad78b2439176f436301a [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 Heo74d46d62009-07-21 17:11:50 +090011 * chunk is consisted of nr_cpu_ids units and the first chunk is used
12 * for static percpu variables in the kernel image (special boot time
13 * alloc/init handling necessary as these areas need to be brought up
14 * before allocation services are running). Unit grows as necessary
15 * and all units grow or shrink in unison. When a chunk is filled up,
16 * another chunk is allocated. ie. in vmalloc area
Tejun Heofbf59bc2009-02-20 16:29:08 +090017 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. Percpu access can be done by configuring
Christoph Lametere1b9aa32009-04-02 13:21:44 +090026 * percpu base registers pcpu_unit_size apart.
Tejun Heofbf59bc2009-02-20 16:29:08 +090027 *
28 * There are usually many small percpu allocations many of them as
29 * small as 4 bytes. The allocator organizes chunks into lists
30 * according to free size and tries to allocate from the fullest one.
31 * Each chunk keeps the maximum contiguous area size hint which is
32 * guaranteed to be eqaul to or larger than the maximum contiguous
33 * area in the chunk. This helps the allocator not to iterate the
34 * chunk maps unnecessarily.
35 *
36 * Allocation state in each chunk is kept using an array of integers
37 * on chunk->map. A positive value in the map represents a free
38 * region and negative allocated. Allocation inside a chunk is done
39 * by scanning this map sequentially and serving the first matching
40 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090041 * Chunks can be determined from the address using the index field
42 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090043 *
44 * To use this allocator, arch code should do the followings.
45 *
46 * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/list.h>
59#include <linux/mm.h>
60#include <linux/module.h>
61#include <linux/mutex.h>
62#include <linux/percpu.h>
63#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090064#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090065#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090066#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090067#include <linux/workqueue.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090068
69#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090070#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090071#include <asm/tlbflush.h>
72
Tejun Heofbf59bc2009-02-20 16:29:08 +090073#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
74#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
75
Tejun Heoe0100982009-03-10 16:27:48 +090076/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
77#ifndef __addr_to_pcpu_ptr
78#define __addr_to_pcpu_ptr(addr) \
79 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
80 + (unsigned long)__per_cpu_start)
81#endif
82#ifndef __pcpu_ptr_to_addr
83#define __pcpu_ptr_to_addr(ptr) \
84 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
85 - (unsigned long)__per_cpu_start)
86#endif
87
Tejun Heofbf59bc2009-02-20 16:29:08 +090088struct pcpu_chunk {
89 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +090090 int free_size; /* free bytes in the chunk */
91 int contig_hint; /* max contiguous size hint */
92 struct vm_struct *vm; /* mapped vmalloc region */
93 int map_used; /* # of map entries used */
94 int map_alloc; /* # of map entries allocated */
95 int *map; /* allocation map */
Tejun Heo8d408b42009-02-24 11:57:21 +090096 bool immutable; /* no [de]population allowed */
Tejun Heo3e24aa52009-03-06 14:33:59 +090097 struct page **page; /* points to page array */
98 struct page *page_ar[]; /* #cpus * UNIT_PAGES */
Tejun Heofbf59bc2009-02-20 16:29:08 +090099};
100
Tejun Heo40150d32009-02-24 12:32:28 +0900101static int pcpu_unit_pages __read_mostly;
102static int pcpu_unit_size __read_mostly;
103static int pcpu_chunk_size __read_mostly;
104static int pcpu_nr_slots __read_mostly;
105static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900106
107/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900108void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900109EXPORT_SYMBOL_GPL(pcpu_base_addr);
110
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900111/*
112 * The first chunk which always exists. Note that unlike other
113 * chunks, this one can be allocated and mapped in several different
114 * ways and thus often doesn't live in the vmalloc area.
115 */
116static struct pcpu_chunk *pcpu_first_chunk;
117
118/*
119 * Optional reserved chunk. This chunk reserves part of the first
120 * chunk and serves it for reserved allocations. The amount of
121 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
122 * area doesn't exist, the following variables contain NULL and 0
123 * respectively.
124 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900125static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900126static int pcpu_reserved_chunk_limit;
127
Tejun Heofbf59bc2009-02-20 16:29:08 +0900128/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900129 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900130 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900131 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
132 * protects allocation/reclaim paths, chunks and chunk->page arrays.
133 * The latter is a spinlock and protects the index data structures -
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900134 * chunk slots, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900135 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900136 * During allocation, pcpu_alloc_mutex is kept locked all the time and
137 * pcpu_lock is grabbed and released as necessary. All actual memory
138 * allocations are done using GFP_KERNEL with pcpu_lock released.
139 *
140 * Free path accesses and alters only the index data structures, so it
141 * can be safely called from atomic context. When memory needs to be
142 * returned to the system, free path schedules reclaim_work which
143 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
144 * reclaimed, release both locks and frees the chunks. Note that it's
145 * necessary to grab both locks to remove a chunk from circulation as
146 * allocation path might be referencing the chunk with only
147 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900148 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900149static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
150static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900151
Tejun Heo40150d32009-02-24 12:32:28 +0900152static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900153
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900154/* reclaim work to release fully free chunks, scheduled from free path */
155static void pcpu_reclaim(struct work_struct *work);
156static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
157
Tejun Heod9b55ee2009-02-24 11:57:21 +0900158static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900159{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900160 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900161 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
162}
163
Tejun Heod9b55ee2009-02-24 11:57:21 +0900164static int pcpu_size_to_slot(int size)
165{
166 if (size == pcpu_unit_size)
167 return pcpu_nr_slots - 1;
168 return __pcpu_size_to_slot(size);
169}
170
Tejun Heofbf59bc2009-02-20 16:29:08 +0900171static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
172{
173 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
174 return 0;
175
176 return pcpu_size_to_slot(chunk->free_size);
177}
178
179static int pcpu_page_idx(unsigned int cpu, int page_idx)
180{
Tejun Heod9b55ee2009-02-24 11:57:21 +0900181 return cpu * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900182}
183
184static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk,
185 unsigned int cpu, int page_idx)
186{
187 return &chunk->page[pcpu_page_idx(cpu, page_idx)];
188}
189
190static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
191 unsigned int cpu, int page_idx)
192{
193 return (unsigned long)chunk->vm->addr +
194 (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT);
195}
196
197static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk,
198 int page_idx)
199{
Tejun Heo04a13c72009-09-01 21:12:28 +0900200 /*
201 * Any possible cpu id can be used here, so there's no need to
202 * worry about preemption or cpu hotplug.
203 */
204 return *pcpu_chunk_pagep(chunk, raw_smp_processor_id(),
205 page_idx) != NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900206}
207
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900208/* set the pointer to a chunk in a page struct */
209static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
210{
211 page->index = (unsigned long)pcpu;
212}
213
214/* obtain pointer to a chunk from a page struct */
215static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
216{
217 return (struct pcpu_chunk *)page->index;
218}
219
Tejun Heofbf59bc2009-02-20 16:29:08 +0900220/**
Tejun Heo1880d932009-03-07 00:44:09 +0900221 * pcpu_mem_alloc - allocate memory
222 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900223 *
Tejun Heo1880d932009-03-07 00:44:09 +0900224 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
225 * kzalloc() is used; otherwise, vmalloc() is used. The returned
226 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900227 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900228 * CONTEXT:
229 * Does GFP_KERNEL allocation.
230 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900231 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900232 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900233 */
Tejun Heo1880d932009-03-07 00:44:09 +0900234static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900235{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900236 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900237 return kzalloc(size, GFP_KERNEL);
238 else {
239 void *ptr = vmalloc(size);
240 if (ptr)
241 memset(ptr, 0, size);
242 return ptr;
243 }
244}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900245
Tejun Heo1880d932009-03-07 00:44:09 +0900246/**
247 * pcpu_mem_free - free memory
248 * @ptr: memory to free
249 * @size: size of the area
250 *
251 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
252 */
253static void pcpu_mem_free(void *ptr, size_t size)
254{
255 if (size <= PAGE_SIZE)
256 kfree(ptr);
257 else
258 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900259}
260
261/**
262 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
263 * @chunk: chunk of interest
264 * @oslot: the previous slot it was on
265 *
266 * This function is called after an allocation or free changed @chunk.
267 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900268 * moved to the slot. Note that the reserved chunk is never put on
269 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900270 *
271 * CONTEXT:
272 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900273 */
274static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
275{
276 int nslot = pcpu_chunk_slot(chunk);
277
Tejun Heoedcb4632009-03-06 14:33:59 +0900278 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900279 if (oslot < nslot)
280 list_move(&chunk->list, &pcpu_slot[nslot]);
281 else
282 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
283 }
284}
285
Tejun Heofbf59bc2009-02-20 16:29:08 +0900286/**
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900287 * pcpu_chunk_addr_search - determine chunk containing specified address
288 * @addr: address for which the chunk needs to be determined.
Tejun Heoccea34b2009-03-07 00:44:13 +0900289 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900290 * RETURNS:
291 * The address of the found chunk.
292 */
293static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
294{
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900295 void *first_start = pcpu_first_chunk->vm->addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900296
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900297 /* is it in the first chunk? */
298 if (addr >= first_start && addr < first_start + pcpu_chunk_size) {
299 /* is it in the reserved area? */
300 if (addr < first_start + pcpu_reserved_chunk_limit)
Tejun Heoedcb4632009-03-06 14:33:59 +0900301 return pcpu_reserved_chunk;
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900302 return pcpu_first_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900303 }
304
Tejun Heo04a13c72009-09-01 21:12:28 +0900305 /*
306 * The address is relative to unit0 which might be unused and
307 * thus unmapped. Offset the address to the unit space of the
308 * current processor before looking it up in the vmalloc
309 * space. Note that any possible cpu id can be used here, so
310 * there's no need to worry about preemption or cpu hotplug.
311 */
312 addr += raw_smp_processor_id() * pcpu_unit_size;
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900313 return pcpu_get_page_chunk(vmalloc_to_page(addr));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900314}
315
316/**
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900317 * pcpu_extend_area_map - extend area map for allocation
318 * @chunk: target chunk
319 *
320 * Extend area map of @chunk so that it can accomodate an allocation.
321 * A single allocation can split an area into three areas, so this
322 * function makes sure that @chunk->map has at least two extra slots.
323 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900324 * CONTEXT:
325 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
326 * if area map is extended.
327 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900328 * RETURNS:
329 * 0 if noop, 1 if successfully extended, -errno on failure.
330 */
331static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
332{
333 int new_alloc;
334 int *new;
335 size_t size;
336
337 /* has enough? */
338 if (chunk->map_alloc >= chunk->map_used + 2)
339 return 0;
340
Tejun Heoccea34b2009-03-07 00:44:13 +0900341 spin_unlock_irq(&pcpu_lock);
342
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900343 new_alloc = PCPU_DFL_MAP_ALLOC;
344 while (new_alloc < chunk->map_used + 2)
345 new_alloc *= 2;
346
347 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
Tejun Heoccea34b2009-03-07 00:44:13 +0900348 if (!new) {
349 spin_lock_irq(&pcpu_lock);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900350 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900351 }
352
353 /*
354 * Acquire pcpu_lock and switch to new area map. Only free
355 * could have happened inbetween, so map_used couldn't have
356 * grown.
357 */
358 spin_lock_irq(&pcpu_lock);
359 BUG_ON(new_alloc < chunk->map_used + 2);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900360
361 size = chunk->map_alloc * sizeof(chunk->map[0]);
362 memcpy(new, chunk->map, size);
363
364 /*
365 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
366 * one of the first chunks and still using static map.
367 */
368 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
369 pcpu_mem_free(chunk->map, size);
370
371 chunk->map_alloc = new_alloc;
372 chunk->map = new;
373 return 0;
374}
375
376/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900377 * pcpu_split_block - split a map block
378 * @chunk: chunk of interest
379 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900380 * @head: head size in bytes (can be 0)
381 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900382 *
383 * Split the @i'th map block into two or three blocks. If @head is
384 * non-zero, @head bytes block is inserted before block @i moving it
385 * to @i+1 and reducing its size by @head bytes.
386 *
387 * If @tail is non-zero, the target block, which can be @i or @i+1
388 * depending on @head, is reduced by @tail bytes and @tail byte block
389 * is inserted after the target block.
390 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900391 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900392 *
393 * CONTEXT:
394 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900395 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900396static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
397 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900398{
399 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900400
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900401 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900402
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900403 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900404 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
405 sizeof(chunk->map[0]) * (chunk->map_used - i));
406 chunk->map_used += nr_extra;
407
408 if (head) {
409 chunk->map[i + 1] = chunk->map[i] - head;
410 chunk->map[i++] = head;
411 }
412 if (tail) {
413 chunk->map[i++] -= tail;
414 chunk->map[i] = tail;
415 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900416}
417
418/**
419 * pcpu_alloc_area - allocate area from a pcpu_chunk
420 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900421 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900422 * @align: wanted align
423 *
424 * Try to allocate @size bytes area aligned at @align from @chunk.
425 * Note that this function only allocates the offset. It doesn't
426 * populate or map the area.
427 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900428 * @chunk->map must have at least two free slots.
429 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900430 * CONTEXT:
431 * pcpu_lock.
432 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900433 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900434 * Allocated offset in @chunk on success, -1 if no matching area is
435 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900436 */
437static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
438{
439 int oslot = pcpu_chunk_slot(chunk);
440 int max_contig = 0;
441 int i, off;
442
Tejun Heofbf59bc2009-02-20 16:29:08 +0900443 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
444 bool is_last = i + 1 == chunk->map_used;
445 int head, tail;
446
447 /* extra for alignment requirement */
448 head = ALIGN(off, align) - off;
449 BUG_ON(i == 0 && head != 0);
450
451 if (chunk->map[i] < 0)
452 continue;
453 if (chunk->map[i] < head + size) {
454 max_contig = max(chunk->map[i], max_contig);
455 continue;
456 }
457
458 /*
459 * If head is small or the previous block is free,
460 * merge'em. Note that 'small' is defined as smaller
461 * than sizeof(int), which is very small but isn't too
462 * uncommon for percpu allocations.
463 */
464 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
465 if (chunk->map[i - 1] > 0)
466 chunk->map[i - 1] += head;
467 else {
468 chunk->map[i - 1] -= head;
469 chunk->free_size -= head;
470 }
471 chunk->map[i] -= head;
472 off += head;
473 head = 0;
474 }
475
476 /* if tail is small, just keep it around */
477 tail = chunk->map[i] - head - size;
478 if (tail < sizeof(int))
479 tail = 0;
480
481 /* split if warranted */
482 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900483 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900484 if (head) {
485 i++;
486 off += head;
487 max_contig = max(chunk->map[i - 1], max_contig);
488 }
489 if (tail)
490 max_contig = max(chunk->map[i + 1], max_contig);
491 }
492
493 /* update hint and mark allocated */
494 if (is_last)
495 chunk->contig_hint = max_contig; /* fully scanned */
496 else
497 chunk->contig_hint = max(chunk->contig_hint,
498 max_contig);
499
500 chunk->free_size -= chunk->map[i];
501 chunk->map[i] = -chunk->map[i];
502
503 pcpu_chunk_relocate(chunk, oslot);
504 return off;
505 }
506
507 chunk->contig_hint = max_contig; /* fully scanned */
508 pcpu_chunk_relocate(chunk, oslot);
509
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900510 /* tell the upper layer that this chunk has no matching area */
511 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900512}
513
514/**
515 * pcpu_free_area - free area to a pcpu_chunk
516 * @chunk: chunk of interest
517 * @freeme: offset of area to free
518 *
519 * Free area starting from @freeme to @chunk. Note that this function
520 * only modifies the allocation map. It doesn't depopulate or unmap
521 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900522 *
523 * CONTEXT:
524 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900525 */
526static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
527{
528 int oslot = pcpu_chunk_slot(chunk);
529 int i, off;
530
531 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
532 if (off == freeme)
533 break;
534 BUG_ON(off != freeme);
535 BUG_ON(chunk->map[i] > 0);
536
537 chunk->map[i] = -chunk->map[i];
538 chunk->free_size += chunk->map[i];
539
540 /* merge with previous? */
541 if (i > 0 && chunk->map[i - 1] >= 0) {
542 chunk->map[i - 1] += chunk->map[i];
543 chunk->map_used--;
544 memmove(&chunk->map[i], &chunk->map[i + 1],
545 (chunk->map_used - i) * sizeof(chunk->map[0]));
546 i--;
547 }
548 /* merge with next? */
549 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
550 chunk->map[i] += chunk->map[i + 1];
551 chunk->map_used--;
552 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
553 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
554 }
555
556 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
557 pcpu_chunk_relocate(chunk, oslot);
558}
559
560/**
561 * pcpu_unmap - unmap pages out of a pcpu_chunk
562 * @chunk: chunk of interest
563 * @page_start: page index of the first page to unmap
564 * @page_end: page index of the last page to unmap + 1
Tejun Heo85ae87c2009-06-22 11:56:23 +0900565 * @flush_tlb: whether to flush tlb or not
Tejun Heofbf59bc2009-02-20 16:29:08 +0900566 *
567 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
568 * If @flush is true, vcache is flushed before unmapping and tlb
569 * after.
570 */
571static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
Tejun Heo85ae87c2009-06-22 11:56:23 +0900572 bool flush_tlb)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900573{
Tejun Heo74d46d62009-07-21 17:11:50 +0900574 unsigned int last = nr_cpu_ids - 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900575 unsigned int cpu;
576
Tejun Heo8d408b42009-02-24 11:57:21 +0900577 /* unmap must not be done on immutable chunk */
578 WARN_ON(chunk->immutable);
579
Tejun Heofbf59bc2009-02-20 16:29:08 +0900580 /*
581 * Each flushing trial can be very expensive, issue flush on
582 * the whole region at once rather than doing it for each cpu.
583 * This could be an overkill but is more scalable.
584 */
Tejun Heo85ae87c2009-06-22 11:56:23 +0900585 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
586 pcpu_chunk_addr(chunk, last, page_end));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900587
588 for_each_possible_cpu(cpu)
589 unmap_kernel_range_noflush(
590 pcpu_chunk_addr(chunk, cpu, page_start),
591 (page_end - page_start) << PAGE_SHIFT);
592
593 /* ditto as flush_cache_vunmap() */
Tejun Heo85ae87c2009-06-22 11:56:23 +0900594 if (flush_tlb)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900595 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
596 pcpu_chunk_addr(chunk, last, page_end));
597}
598
599/**
600 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
601 * @chunk: chunk to depopulate
602 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900603 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900604 * @flush: whether to flush cache and tlb or not
605 *
606 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
607 * from @chunk. If @flush is true, vcache is flushed before unmapping
608 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900609 *
610 * CONTEXT:
611 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900612 */
Tejun Heocae3aeb2009-02-21 16:56:23 +0900613static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size,
614 bool flush)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900615{
616 int page_start = PFN_DOWN(off);
617 int page_end = PFN_UP(off + size);
618 int unmap_start = -1;
619 int uninitialized_var(unmap_end);
620 unsigned int cpu;
621 int i;
622
623 for (i = page_start; i < page_end; i++) {
624 for_each_possible_cpu(cpu) {
625 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
626
627 if (!*pagep)
628 continue;
629
630 __free_page(*pagep);
631
632 /*
633 * If it's partial depopulation, it might get
634 * populated or depopulated again. Mark the
635 * page gone.
636 */
637 *pagep = NULL;
638
639 unmap_start = unmap_start < 0 ? i : unmap_start;
640 unmap_end = i + 1;
641 }
642 }
643
644 if (unmap_start >= 0)
645 pcpu_unmap(chunk, unmap_start, unmap_end, flush);
646}
647
648/**
649 * pcpu_map - map pages into a pcpu_chunk
650 * @chunk: chunk of interest
651 * @page_start: page index of the first page to map
652 * @page_end: page index of the last page to map + 1
653 *
654 * For each cpu, map pages [@page_start,@page_end) into @chunk.
655 * vcache is flushed afterwards.
656 */
657static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
658{
Tejun Heo74d46d62009-07-21 17:11:50 +0900659 unsigned int last = nr_cpu_ids - 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900660 unsigned int cpu;
661 int err;
662
Tejun Heo8d408b42009-02-24 11:57:21 +0900663 /* map must not be done on immutable chunk */
664 WARN_ON(chunk->immutable);
665
Tejun Heofbf59bc2009-02-20 16:29:08 +0900666 for_each_possible_cpu(cpu) {
667 err = map_kernel_range_noflush(
668 pcpu_chunk_addr(chunk, cpu, page_start),
669 (page_end - page_start) << PAGE_SHIFT,
670 PAGE_KERNEL,
671 pcpu_chunk_pagep(chunk, cpu, page_start));
672 if (err < 0)
673 return err;
674 }
675
676 /* flush at once, please read comments in pcpu_unmap() */
677 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
678 pcpu_chunk_addr(chunk, last, page_end));
679 return 0;
680}
681
682/**
683 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
684 * @chunk: chunk of interest
685 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900686 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900687 *
688 * For each cpu, populate and map pages [@page_start,@page_end) into
689 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900690 *
691 * CONTEXT:
692 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900693 */
694static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
695{
696 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
697 int page_start = PFN_DOWN(off);
698 int page_end = PFN_UP(off + size);
699 int map_start = -1;
Tejun Heo02d51fdf2009-03-01 15:42:36 +0900700 int uninitialized_var(map_end);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900701 unsigned int cpu;
702 int i;
703
704 for (i = page_start; i < page_end; i++) {
705 if (pcpu_chunk_page_occupied(chunk, i)) {
706 if (map_start >= 0) {
707 if (pcpu_map(chunk, map_start, map_end))
708 goto err;
709 map_start = -1;
710 }
711 continue;
712 }
713
714 map_start = map_start < 0 ? i : map_start;
715 map_end = i + 1;
716
717 for_each_possible_cpu(cpu) {
718 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
719
720 *pagep = alloc_pages_node(cpu_to_node(cpu),
721 alloc_mask, 0);
722 if (!*pagep)
723 goto err;
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900724 pcpu_set_page_chunk(*pagep, chunk);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900725 }
726 }
727
728 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
729 goto err;
730
731 for_each_possible_cpu(cpu)
Tejun Heod9b55ee2009-02-24 11:57:21 +0900732 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0,
Tejun Heofbf59bc2009-02-20 16:29:08 +0900733 size);
734
735 return 0;
736err:
737 /* likely under heavy memory pressure, give memory back */
738 pcpu_depopulate_chunk(chunk, off, size, true);
739 return -ENOMEM;
740}
741
742static void free_pcpu_chunk(struct pcpu_chunk *chunk)
743{
744 if (!chunk)
745 return;
746 if (chunk->vm)
747 free_vm_area(chunk->vm);
Tejun Heo1880d932009-03-07 00:44:09 +0900748 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900749 kfree(chunk);
750}
751
752static struct pcpu_chunk *alloc_pcpu_chunk(void)
753{
754 struct pcpu_chunk *chunk;
755
756 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
757 if (!chunk)
758 return NULL;
759
Tejun Heo1880d932009-03-07 00:44:09 +0900760 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900761 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
762 chunk->map[chunk->map_used++] = pcpu_unit_size;
Tejun Heo3e24aa52009-03-06 14:33:59 +0900763 chunk->page = chunk->page_ar;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900764
Amerigo Wang142d44b2009-08-13 02:00:13 -0400765 chunk->vm = get_vm_area(pcpu_chunk_size, VM_ALLOC);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900766 if (!chunk->vm) {
767 free_pcpu_chunk(chunk);
768 return NULL;
769 }
770
771 INIT_LIST_HEAD(&chunk->list);
772 chunk->free_size = pcpu_unit_size;
773 chunk->contig_hint = pcpu_unit_size;
774
775 return chunk;
776}
777
778/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900779 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900780 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900781 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900782 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900783 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900784 * Allocate percpu area of @size bytes aligned at @align.
785 *
786 * CONTEXT:
787 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900788 *
789 * RETURNS:
790 * Percpu pointer to the allocated area on success, NULL on failure.
791 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900792static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900793{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900794 struct pcpu_chunk *chunk;
795 int slot, off;
796
Tejun Heo8d408b42009-02-24 11:57:21 +0900797 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900798 WARN(true, "illegal size (%zu) or align (%zu) for "
799 "percpu allocation\n", size, align);
800 return NULL;
801 }
802
Tejun Heoccea34b2009-03-07 00:44:13 +0900803 mutex_lock(&pcpu_alloc_mutex);
804 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900805
Tejun Heoedcb4632009-03-06 14:33:59 +0900806 /* serve reserved allocations from the reserved chunk if available */
807 if (reserved && pcpu_reserved_chunk) {
808 chunk = pcpu_reserved_chunk;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900809 if (size > chunk->contig_hint ||
810 pcpu_extend_area_map(chunk) < 0)
Tejun Heoccea34b2009-03-07 00:44:13 +0900811 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900812 off = pcpu_alloc_area(chunk, size, align);
813 if (off >= 0)
814 goto area_found;
Tejun Heoccea34b2009-03-07 00:44:13 +0900815 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900816 }
817
Tejun Heoccea34b2009-03-07 00:44:13 +0900818restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900819 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900820 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
821 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
822 if (size > chunk->contig_hint)
823 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900824
825 switch (pcpu_extend_area_map(chunk)) {
826 case 0:
827 break;
828 case 1:
829 goto restart; /* pcpu_lock dropped, restart */
830 default:
831 goto fail_unlock;
832 }
833
Tejun Heofbf59bc2009-02-20 16:29:08 +0900834 off = pcpu_alloc_area(chunk, size, align);
835 if (off >= 0)
836 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900837 }
838 }
839
840 /* hmmm... no space left, create a new chunk */
Tejun Heoccea34b2009-03-07 00:44:13 +0900841 spin_unlock_irq(&pcpu_lock);
842
Tejun Heofbf59bc2009-02-20 16:29:08 +0900843 chunk = alloc_pcpu_chunk();
844 if (!chunk)
Tejun Heoccea34b2009-03-07 00:44:13 +0900845 goto fail_unlock_mutex;
846
847 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900848 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900849 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900850
851area_found:
Tejun Heoccea34b2009-03-07 00:44:13 +0900852 spin_unlock_irq(&pcpu_lock);
853
Tejun Heofbf59bc2009-02-20 16:29:08 +0900854 /* populate, map and clear the area */
855 if (pcpu_populate_chunk(chunk, off, size)) {
Tejun Heoccea34b2009-03-07 00:44:13 +0900856 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900857 pcpu_free_area(chunk, off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900858 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900859 }
860
Tejun Heoccea34b2009-03-07 00:44:13 +0900861 mutex_unlock(&pcpu_alloc_mutex);
862
863 return __addr_to_pcpu_ptr(chunk->vm->addr + off);
864
865fail_unlock:
866 spin_unlock_irq(&pcpu_lock);
867fail_unlock_mutex:
868 mutex_unlock(&pcpu_alloc_mutex);
869 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900870}
Tejun Heoedcb4632009-03-06 14:33:59 +0900871
872/**
873 * __alloc_percpu - allocate dynamic percpu area
874 * @size: size of area to allocate in bytes
875 * @align: alignment of area (max PAGE_SIZE)
876 *
877 * Allocate percpu area of @size bytes aligned at @align. Might
878 * sleep. Might trigger writeouts.
879 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900880 * CONTEXT:
881 * Does GFP_KERNEL allocation.
882 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900883 * RETURNS:
884 * Percpu pointer to the allocated area on success, NULL on failure.
885 */
886void *__alloc_percpu(size_t size, size_t align)
887{
888 return pcpu_alloc(size, align, false);
889}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900890EXPORT_SYMBOL_GPL(__alloc_percpu);
891
Tejun Heoedcb4632009-03-06 14:33:59 +0900892/**
893 * __alloc_reserved_percpu - allocate reserved percpu area
894 * @size: size of area to allocate in bytes
895 * @align: alignment of area (max PAGE_SIZE)
896 *
897 * Allocate percpu area of @size bytes aligned at @align from reserved
898 * percpu area if arch has set it up; otherwise, allocation is served
899 * from the same dynamic area. Might sleep. Might trigger writeouts.
900 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900901 * CONTEXT:
902 * Does GFP_KERNEL allocation.
903 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900904 * RETURNS:
905 * Percpu pointer to the allocated area on success, NULL on failure.
906 */
907void *__alloc_reserved_percpu(size_t size, size_t align)
908{
909 return pcpu_alloc(size, align, true);
910}
911
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900912/**
913 * pcpu_reclaim - reclaim fully free chunks, workqueue function
914 * @work: unused
915 *
916 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900917 *
918 * CONTEXT:
919 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900920 */
921static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900922{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900923 LIST_HEAD(todo);
924 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
925 struct pcpu_chunk *chunk, *next;
926
Tejun Heoccea34b2009-03-07 00:44:13 +0900927 mutex_lock(&pcpu_alloc_mutex);
928 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900929
930 list_for_each_entry_safe(chunk, next, head, list) {
931 WARN_ON(chunk->immutable);
932
933 /* spare the first one */
934 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
935 continue;
936
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900937 list_move(&chunk->list, &todo);
938 }
939
Tejun Heoccea34b2009-03-07 00:44:13 +0900940 spin_unlock_irq(&pcpu_lock);
941 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900942
943 list_for_each_entry_safe(chunk, next, &todo, list) {
944 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
945 free_pcpu_chunk(chunk);
946 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900947}
948
949/**
950 * free_percpu - free percpu area
951 * @ptr: pointer to area to free
952 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900953 * Free percpu area @ptr.
954 *
955 * CONTEXT:
956 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900957 */
958void free_percpu(void *ptr)
959{
960 void *addr = __pcpu_ptr_to_addr(ptr);
961 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900962 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900963 int off;
964
965 if (!ptr)
966 return;
967
Tejun Heoccea34b2009-03-07 00:44:13 +0900968 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900969
970 chunk = pcpu_chunk_addr_search(addr);
971 off = addr - chunk->vm->addr;
972
973 pcpu_free_area(chunk, off);
974
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900975 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900976 if (chunk->free_size == pcpu_unit_size) {
977 struct pcpu_chunk *pos;
978
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900979 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900980 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900981 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900982 break;
983 }
984 }
985
Tejun Heoccea34b2009-03-07 00:44:13 +0900986 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900987}
988EXPORT_SYMBOL_GPL(free_percpu);
989
990/**
Tejun Heo8d408b42009-02-24 11:57:21 +0900991 * pcpu_setup_first_chunk - initialize the first percpu chunk
992 * @get_page_fn: callback to fetch page pointer
993 * @static_size: the size of static percpu area in bytes
Tejun Heoedcb4632009-03-06 14:33:59 +0900994 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +0900995 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heo6074d5b2009-03-10 16:27:48 +0900996 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
Tejun Heo8d408b42009-02-24 11:57:21 +0900997 * @base_addr: mapped address, NULL for auto
998 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary
Tejun Heofbf59bc2009-02-20 16:29:08 +0900999 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001000 * Initialize the first percpu chunk which contains the kernel static
1001 * perpcu area. This function is to be called from arch percpu area
1002 * setup path. The first two parameters are mandatory. The rest are
1003 * optional.
1004 *
1005 * @get_page_fn() should return pointer to percpu page given cpu
1006 * number and page number. It should at least return enough pages to
1007 * cover the static area. The returned pages for static area should
1008 * have been initialized with valid data. If @unit_size is specified,
1009 * it can also return pages after the static area. NULL return
1010 * indicates end of pages for the cpu. Note that @get_page_fn() must
1011 * return the same number of pages for all cpus.
1012 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001013 * @reserved_size, if non-zero, specifies the amount of bytes to
1014 * reserve after the static area in the first chunk. This reserves
1015 * the first chunk such that it's available only through reserved
1016 * percpu allocation. This is primarily used to serve module percpu
1017 * static areas on architectures where the addressing model has
1018 * limited offset range for symbol relocations to guarantee module
1019 * percpu symbols fall inside the relocatable range.
1020 *
Tejun Heo6074d5b2009-03-10 16:27:48 +09001021 * @dyn_size, if non-negative, determines the number of bytes
1022 * available for dynamic allocation in the first chunk. Specifying
1023 * non-negative value makes percpu leave alone the area beyond
1024 * @static_size + @reserved_size + @dyn_size.
1025 *
Tejun Heocafe8812009-03-06 14:33:59 +09001026 * @unit_size, if non-negative, specifies unit size and must be
1027 * aligned to PAGE_SIZE and equal to or larger than @static_size +
Tejun Heo6074d5b2009-03-10 16:27:48 +09001028 * @reserved_size + if non-negative, @dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001029 *
1030 * Non-null @base_addr means that the caller already allocated virtual
1031 * region for the first chunk and mapped it. percpu must not mess
1032 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
1033 * @populate_pte_fn doesn't make any sense.
1034 *
1035 * @populate_pte_fn is used to populate the pagetable. NULL means the
1036 * caller already populated the pagetable.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001037 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001038 * If the first chunk ends up with both reserved and dynamic areas, it
1039 * is served by two chunks - one to serve the core static and reserved
1040 * areas and the other for the dynamic area. They share the same vm
1041 * and page map but uses different area allocation map to stay away
1042 * from each other. The latter chunk is circulated in the chunk slots
1043 * and available for dynamic allocation like any other chunks.
1044 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001045 * RETURNS:
1046 * The determined pcpu_unit_size which can be used to initialize
1047 * percpu access.
1048 */
Tejun Heo8d408b42009-02-24 11:57:21 +09001049size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
Tejun Heoedcb4632009-03-06 14:33:59 +09001050 size_t static_size, size_t reserved_size,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001051 ssize_t dyn_size, ssize_t unit_size,
Tejun Heocafe8812009-03-06 14:33:59 +09001052 void *base_addr,
Tejun Heo8d408b42009-02-24 11:57:21 +09001053 pcpu_populate_pte_fn_t populate_pte_fn)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001054{
Tejun Heo2441d152009-03-06 14:33:59 +09001055 static struct vm_struct first_vm;
Tejun Heoedcb4632009-03-06 14:33:59 +09001056 static int smap[2], dmap[2];
Tejun Heo6074d5b2009-03-10 16:27:48 +09001057 size_t size_sum = static_size + reserved_size +
1058 (dyn_size >= 0 ? dyn_size : 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001059 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001060 unsigned int cpu;
Tejun Heo8d408b42009-02-24 11:57:21 +09001061 int nr_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001062 int err, i;
1063
Tejun Heo8d408b42009-02-24 11:57:21 +09001064 /* santiy checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001065 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1066 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo8d408b42009-02-24 11:57:21 +09001067 BUG_ON(!static_size);
Tejun Heocafe8812009-03-06 14:33:59 +09001068 if (unit_size >= 0) {
Tejun Heo6074d5b2009-03-10 16:27:48 +09001069 BUG_ON(unit_size < size_sum);
Tejun Heocafe8812009-03-06 14:33:59 +09001070 BUG_ON(unit_size & ~PAGE_MASK);
Tejun Heo6074d5b2009-03-10 16:27:48 +09001071 BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE);
1072 } else
Tejun Heocafe8812009-03-06 14:33:59 +09001073 BUG_ON(base_addr);
Tejun Heo8d408b42009-02-24 11:57:21 +09001074 BUG_ON(base_addr && populate_pte_fn);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001075
Tejun Heocafe8812009-03-06 14:33:59 +09001076 if (unit_size >= 0)
Tejun Heo8d408b42009-02-24 11:57:21 +09001077 pcpu_unit_pages = unit_size >> PAGE_SHIFT;
1078 else
1079 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001080 PFN_UP(size_sum));
Tejun Heo8d408b42009-02-24 11:57:21 +09001081
Tejun Heod9b55ee2009-02-24 11:57:21 +09001082 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo74d46d62009-07-21 17:11:50 +09001083 pcpu_chunk_size = nr_cpu_ids * pcpu_unit_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001084 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
Tejun Heo74d46d62009-07-21 17:11:50 +09001085 + nr_cpu_ids * pcpu_unit_pages * sizeof(struct page *);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001086
Tejun Heocafe8812009-03-06 14:33:59 +09001087 if (dyn_size < 0)
Tejun Heoedcb4632009-03-06 14:33:59 +09001088 dyn_size = pcpu_unit_size - static_size - reserved_size;
Tejun Heocafe8812009-03-06 14:33:59 +09001089
Tejun Heod9b55ee2009-02-24 11:57:21 +09001090 /*
1091 * Allocate chunk slots. The additional last slot is for
1092 * empty chunks.
1093 */
1094 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001095 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1096 for (i = 0; i < pcpu_nr_slots; i++)
1097 INIT_LIST_HEAD(&pcpu_slot[i]);
1098
Tejun Heoedcb4632009-03-06 14:33:59 +09001099 /*
1100 * Initialize static chunk. If reserved_size is zero, the
1101 * static chunk covers static area + dynamic allocation area
1102 * in the first chunk. If reserved_size is not zero, it
1103 * covers static area + reserved area (mostly used for module
1104 * static percpu allocation).
1105 */
Tejun Heo2441d152009-03-06 14:33:59 +09001106 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1107 INIT_LIST_HEAD(&schunk->list);
1108 schunk->vm = &first_vm;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001109 schunk->map = smap;
1110 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo3e24aa52009-03-06 14:33:59 +09001111 schunk->page = schunk->page_ar;
Tejun Heoedcb4632009-03-06 14:33:59 +09001112
1113 if (reserved_size) {
1114 schunk->free_size = reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001115 pcpu_reserved_chunk = schunk;
1116 pcpu_reserved_chunk_limit = static_size + reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001117 } else {
1118 schunk->free_size = dyn_size;
1119 dyn_size = 0; /* dynamic area covered */
1120 }
Tejun Heo2441d152009-03-06 14:33:59 +09001121 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001122
Tejun Heo61ace7f2009-03-06 14:33:59 +09001123 schunk->map[schunk->map_used++] = -static_size;
1124 if (schunk->free_size)
1125 schunk->map[schunk->map_used++] = schunk->free_size;
1126
Tejun Heoedcb4632009-03-06 14:33:59 +09001127 /* init dynamic chunk if necessary */
1128 if (dyn_size) {
1129 dchunk = alloc_bootmem(sizeof(struct pcpu_chunk));
1130 INIT_LIST_HEAD(&dchunk->list);
1131 dchunk->vm = &first_vm;
1132 dchunk->map = dmap;
1133 dchunk->map_alloc = ARRAY_SIZE(dmap);
1134 dchunk->page = schunk->page_ar; /* share page map with schunk */
1135
1136 dchunk->contig_hint = dchunk->free_size = dyn_size;
1137 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1138 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1139 }
1140
Tejun Heo8d408b42009-02-24 11:57:21 +09001141 /* allocate vm address */
Tejun Heo2441d152009-03-06 14:33:59 +09001142 first_vm.flags = VM_ALLOC;
1143 first_vm.size = pcpu_chunk_size;
Tejun Heo8d408b42009-02-24 11:57:21 +09001144
1145 if (!base_addr)
Tejun Heo2441d152009-03-06 14:33:59 +09001146 vm_area_register_early(&first_vm, PAGE_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001147 else {
1148 /*
1149 * Pages already mapped. No need to remap into
Tejun Heoedcb4632009-03-06 14:33:59 +09001150 * vmalloc area. In this case the first chunks can't
1151 * be mapped or unmapped by percpu and are marked
Tejun Heo8d408b42009-02-24 11:57:21 +09001152 * immutable.
1153 */
Tejun Heo2441d152009-03-06 14:33:59 +09001154 first_vm.addr = base_addr;
1155 schunk->immutable = true;
Tejun Heoedcb4632009-03-06 14:33:59 +09001156 if (dchunk)
1157 dchunk->immutable = true;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001158 }
1159
Tejun Heo8d408b42009-02-24 11:57:21 +09001160 /* assign pages */
1161 nr_pages = -1;
1162 for_each_possible_cpu(cpu) {
1163 for (i = 0; i < pcpu_unit_pages; i++) {
1164 struct page *page = get_page_fn(cpu, i);
1165
1166 if (!page)
1167 break;
Tejun Heo2441d152009-03-06 14:33:59 +09001168 *pcpu_chunk_pagep(schunk, cpu, i) = page;
Tejun Heo8d408b42009-02-24 11:57:21 +09001169 }
1170
Tejun Heo61ace7f2009-03-06 14:33:59 +09001171 BUG_ON(i < PFN_UP(static_size));
Tejun Heo8d408b42009-02-24 11:57:21 +09001172
1173 if (nr_pages < 0)
1174 nr_pages = i;
1175 else
1176 BUG_ON(nr_pages != i);
1177 }
1178
1179 /* map them */
1180 if (populate_pte_fn) {
1181 for_each_possible_cpu(cpu)
1182 for (i = 0; i < nr_pages; i++)
Tejun Heo2441d152009-03-06 14:33:59 +09001183 populate_pte_fn(pcpu_chunk_addr(schunk,
Tejun Heo8d408b42009-02-24 11:57:21 +09001184 cpu, i));
1185
Tejun Heo2441d152009-03-06 14:33:59 +09001186 err = pcpu_map(schunk, 0, nr_pages);
Tejun Heo8d408b42009-02-24 11:57:21 +09001187 if (err)
1188 panic("failed to setup static percpu area, err=%d\n",
1189 err);
1190 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001191
Tejun Heo2441d152009-03-06 14:33:59 +09001192 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001193 pcpu_first_chunk = dchunk ?: schunk;
1194 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001195
1196 /* we're done */
Tejun Heo2441d152009-03-06 14:33:59 +09001197 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001198 return pcpu_unit_size;
1199}
Tejun Heo66c3a752009-03-10 16:27:48 +09001200
1201/*
1202 * Embedding first chunk setup helper.
1203 */
1204static void *pcpue_ptr __initdata;
1205static size_t pcpue_size __initdata;
1206static size_t pcpue_unit_size __initdata;
1207
1208static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
1209{
1210 size_t off = (size_t)pageno << PAGE_SHIFT;
1211
1212 if (off >= pcpue_size)
1213 return NULL;
1214
1215 return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off);
1216}
1217
1218/**
1219 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1220 * @static_size: the size of static percpu area in bytes
1221 * @reserved_size: the size of reserved percpu area in bytes
1222 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1223 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
1224 *
1225 * This is a helper to ease setting up embedded first percpu chunk and
1226 * can be called where pcpu_setup_first_chunk() is expected.
1227 *
1228 * If this function is used to setup the first chunk, it is allocated
1229 * as a contiguous area using bootmem allocator and used as-is without
1230 * being mapped into vmalloc area. This enables the first chunk to
1231 * piggy back on the linear physical mapping which often uses larger
1232 * page size.
1233 *
1234 * When @dyn_size is positive, dynamic area might be larger than
1235 * specified to fill page alignment. Also, when @dyn_size is auto,
1236 * @dyn_size does not fill the whole first chunk but only what's
1237 * necessary for page alignment after static and reserved areas.
1238 *
1239 * If the needed size is smaller than the minimum or specified unit
1240 * size, the leftover is returned to the bootmem allocator.
1241 *
1242 * RETURNS:
1243 * The determined pcpu_unit_size which can be used to initialize
1244 * percpu access on success, -errno on failure.
1245 */
1246ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size,
1247 ssize_t dyn_size, ssize_t unit_size)
1248{
Tejun Heofa8a7092009-06-22 11:56:24 +09001249 size_t chunk_size;
Tejun Heo66c3a752009-03-10 16:27:48 +09001250 unsigned int cpu;
1251
1252 /* determine parameters and allocate */
1253 pcpue_size = PFN_ALIGN(static_size + reserved_size +
1254 (dyn_size >= 0 ? dyn_size : 0));
1255 if (dyn_size != 0)
1256 dyn_size = pcpue_size - static_size - reserved_size;
1257
1258 if (unit_size >= 0) {
1259 BUG_ON(unit_size < pcpue_size);
1260 pcpue_unit_size = unit_size;
1261 } else
1262 pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE);
1263
Tejun Heo74d46d62009-07-21 17:11:50 +09001264 chunk_size = pcpue_unit_size * nr_cpu_ids;
Tejun Heofa8a7092009-06-22 11:56:24 +09001265
1266 pcpue_ptr = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE,
1267 __pa(MAX_DMA_ADDRESS));
1268 if (!pcpue_ptr) {
1269 pr_warning("PERCPU: failed to allocate %zu bytes for "
1270 "embedding\n", chunk_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001271 return -ENOMEM;
Tejun Heofa8a7092009-06-22 11:56:24 +09001272 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001273
1274 /* return the leftover and copy */
Tejun Heo74d46d62009-07-21 17:11:50 +09001275 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
Tejun Heo66c3a752009-03-10 16:27:48 +09001276 void *ptr = pcpue_ptr + cpu * pcpue_unit_size;
1277
Tejun Heo74d46d62009-07-21 17:11:50 +09001278 if (cpu_possible(cpu)) {
1279 free_bootmem(__pa(ptr + pcpue_size),
1280 pcpue_unit_size - pcpue_size);
1281 memcpy(ptr, __per_cpu_load, static_size);
1282 } else
1283 free_bootmem(__pa(ptr), pcpue_unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09001284 }
1285
1286 /* we're ready, commit */
1287 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
1288 pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size);
1289
1290 return pcpu_setup_first_chunk(pcpue_get_page, static_size,
1291 reserved_size, dyn_size,
1292 pcpue_unit_size, pcpue_ptr, NULL);
1293}