blob: c0b2c1a76e81c280398b27ed310b1e03500e3a23 [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
11 * chunk is consisted of num_possible_cpus() units and the first chunk
12 * is used for static percpu variables in the kernel image (special
13 * boot time alloc/init handling necessary as these areas need to be
14 * brought up before allocation services are running). Unit grows as
15 * necessary and all units grow or shrink in unison. When a chunk is
16 * filled up, another chunk is allocated. ie. in vmalloc area
17 *
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{
200 return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL;
201}
202
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900203/* set the pointer to a chunk in a page struct */
204static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
205{
206 page->index = (unsigned long)pcpu;
207}
208
209/* obtain pointer to a chunk from a page struct */
210static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
211{
212 return (struct pcpu_chunk *)page->index;
213}
214
Tejun Heofbf59bc2009-02-20 16:29:08 +0900215/**
Tejun Heo1880d932009-03-07 00:44:09 +0900216 * pcpu_mem_alloc - allocate memory
217 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900218 *
Tejun Heo1880d932009-03-07 00:44:09 +0900219 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
220 * kzalloc() is used; otherwise, vmalloc() is used. The returned
221 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900222 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900223 * CONTEXT:
224 * Does GFP_KERNEL allocation.
225 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900226 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900227 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900228 */
Tejun Heo1880d932009-03-07 00:44:09 +0900229static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900230{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900231 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900232 return kzalloc(size, GFP_KERNEL);
233 else {
234 void *ptr = vmalloc(size);
235 if (ptr)
236 memset(ptr, 0, size);
237 return ptr;
238 }
239}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900240
Tejun Heo1880d932009-03-07 00:44:09 +0900241/**
242 * pcpu_mem_free - free memory
243 * @ptr: memory to free
244 * @size: size of the area
245 *
246 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
247 */
248static void pcpu_mem_free(void *ptr, size_t size)
249{
250 if (size <= PAGE_SIZE)
251 kfree(ptr);
252 else
253 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900254}
255
256/**
257 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
258 * @chunk: chunk of interest
259 * @oslot: the previous slot it was on
260 *
261 * This function is called after an allocation or free changed @chunk.
262 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900263 * moved to the slot. Note that the reserved chunk is never put on
264 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900265 *
266 * CONTEXT:
267 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900268 */
269static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
270{
271 int nslot = pcpu_chunk_slot(chunk);
272
Tejun Heoedcb4632009-03-06 14:33:59 +0900273 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900274 if (oslot < nslot)
275 list_move(&chunk->list, &pcpu_slot[nslot]);
276 else
277 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
278 }
279}
280
Tejun Heofbf59bc2009-02-20 16:29:08 +0900281/**
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900282 * pcpu_chunk_addr_search - determine chunk containing specified address
283 * @addr: address for which the chunk needs to be determined.
Tejun Heoccea34b2009-03-07 00:44:13 +0900284 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900285 * RETURNS:
286 * The address of the found chunk.
287 */
288static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
289{
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900290 void *first_start = pcpu_first_chunk->vm->addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900291
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900292 /* is it in the first chunk? */
293 if (addr >= first_start && addr < first_start + pcpu_chunk_size) {
294 /* is it in the reserved area? */
295 if (addr < first_start + pcpu_reserved_chunk_limit)
Tejun Heoedcb4632009-03-06 14:33:59 +0900296 return pcpu_reserved_chunk;
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900297 return pcpu_first_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900298 }
299
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900300 return pcpu_get_page_chunk(vmalloc_to_page(addr));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900301}
302
303/**
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900304 * pcpu_extend_area_map - extend area map for allocation
305 * @chunk: target chunk
306 *
307 * Extend area map of @chunk so that it can accomodate an allocation.
308 * A single allocation can split an area into three areas, so this
309 * function makes sure that @chunk->map has at least two extra slots.
310 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900311 * CONTEXT:
312 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
313 * if area map is extended.
314 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900315 * RETURNS:
316 * 0 if noop, 1 if successfully extended, -errno on failure.
317 */
318static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
319{
320 int new_alloc;
321 int *new;
322 size_t size;
323
324 /* has enough? */
325 if (chunk->map_alloc >= chunk->map_used + 2)
326 return 0;
327
Tejun Heoccea34b2009-03-07 00:44:13 +0900328 spin_unlock_irq(&pcpu_lock);
329
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900330 new_alloc = PCPU_DFL_MAP_ALLOC;
331 while (new_alloc < chunk->map_used + 2)
332 new_alloc *= 2;
333
334 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
Tejun Heoccea34b2009-03-07 00:44:13 +0900335 if (!new) {
336 spin_lock_irq(&pcpu_lock);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900337 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900338 }
339
340 /*
341 * Acquire pcpu_lock and switch to new area map. Only free
342 * could have happened inbetween, so map_used couldn't have
343 * grown.
344 */
345 spin_lock_irq(&pcpu_lock);
346 BUG_ON(new_alloc < chunk->map_used + 2);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900347
348 size = chunk->map_alloc * sizeof(chunk->map[0]);
349 memcpy(new, chunk->map, size);
350
351 /*
352 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
353 * one of the first chunks and still using static map.
354 */
355 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
356 pcpu_mem_free(chunk->map, size);
357
358 chunk->map_alloc = new_alloc;
359 chunk->map = new;
360 return 0;
361}
362
363/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900364 * pcpu_split_block - split a map block
365 * @chunk: chunk of interest
366 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900367 * @head: head size in bytes (can be 0)
368 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900369 *
370 * Split the @i'th map block into two or three blocks. If @head is
371 * non-zero, @head bytes block is inserted before block @i moving it
372 * to @i+1 and reducing its size by @head bytes.
373 *
374 * If @tail is non-zero, the target block, which can be @i or @i+1
375 * depending on @head, is reduced by @tail bytes and @tail byte block
376 * is inserted after the target block.
377 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900378 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900379 *
380 * CONTEXT:
381 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900382 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900383static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
384 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900385{
386 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900387
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900388 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900389
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900390 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900391 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
392 sizeof(chunk->map[0]) * (chunk->map_used - i));
393 chunk->map_used += nr_extra;
394
395 if (head) {
396 chunk->map[i + 1] = chunk->map[i] - head;
397 chunk->map[i++] = head;
398 }
399 if (tail) {
400 chunk->map[i++] -= tail;
401 chunk->map[i] = tail;
402 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900403}
404
405/**
406 * pcpu_alloc_area - allocate area from a pcpu_chunk
407 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900408 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900409 * @align: wanted align
410 *
411 * Try to allocate @size bytes area aligned at @align from @chunk.
412 * Note that this function only allocates the offset. It doesn't
413 * populate or map the area.
414 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900415 * @chunk->map must have at least two free slots.
416 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900417 * CONTEXT:
418 * pcpu_lock.
419 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900420 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900421 * Allocated offset in @chunk on success, -1 if no matching area is
422 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900423 */
424static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
425{
426 int oslot = pcpu_chunk_slot(chunk);
427 int max_contig = 0;
428 int i, off;
429
Tejun Heofbf59bc2009-02-20 16:29:08 +0900430 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
431 bool is_last = i + 1 == chunk->map_used;
432 int head, tail;
433
434 /* extra for alignment requirement */
435 head = ALIGN(off, align) - off;
436 BUG_ON(i == 0 && head != 0);
437
438 if (chunk->map[i] < 0)
439 continue;
440 if (chunk->map[i] < head + size) {
441 max_contig = max(chunk->map[i], max_contig);
442 continue;
443 }
444
445 /*
446 * If head is small or the previous block is free,
447 * merge'em. Note that 'small' is defined as smaller
448 * than sizeof(int), which is very small but isn't too
449 * uncommon for percpu allocations.
450 */
451 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
452 if (chunk->map[i - 1] > 0)
453 chunk->map[i - 1] += head;
454 else {
455 chunk->map[i - 1] -= head;
456 chunk->free_size -= head;
457 }
458 chunk->map[i] -= head;
459 off += head;
460 head = 0;
461 }
462
463 /* if tail is small, just keep it around */
464 tail = chunk->map[i] - head - size;
465 if (tail < sizeof(int))
466 tail = 0;
467
468 /* split if warranted */
469 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900470 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900471 if (head) {
472 i++;
473 off += head;
474 max_contig = max(chunk->map[i - 1], max_contig);
475 }
476 if (tail)
477 max_contig = max(chunk->map[i + 1], max_contig);
478 }
479
480 /* update hint and mark allocated */
481 if (is_last)
482 chunk->contig_hint = max_contig; /* fully scanned */
483 else
484 chunk->contig_hint = max(chunk->contig_hint,
485 max_contig);
486
487 chunk->free_size -= chunk->map[i];
488 chunk->map[i] = -chunk->map[i];
489
490 pcpu_chunk_relocate(chunk, oslot);
491 return off;
492 }
493
494 chunk->contig_hint = max_contig; /* fully scanned */
495 pcpu_chunk_relocate(chunk, oslot);
496
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900497 /* tell the upper layer that this chunk has no matching area */
498 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900499}
500
501/**
502 * pcpu_free_area - free area to a pcpu_chunk
503 * @chunk: chunk of interest
504 * @freeme: offset of area to free
505 *
506 * Free area starting from @freeme to @chunk. Note that this function
507 * only modifies the allocation map. It doesn't depopulate or unmap
508 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900509 *
510 * CONTEXT:
511 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900512 */
513static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
514{
515 int oslot = pcpu_chunk_slot(chunk);
516 int i, off;
517
518 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
519 if (off == freeme)
520 break;
521 BUG_ON(off != freeme);
522 BUG_ON(chunk->map[i] > 0);
523
524 chunk->map[i] = -chunk->map[i];
525 chunk->free_size += chunk->map[i];
526
527 /* merge with previous? */
528 if (i > 0 && chunk->map[i - 1] >= 0) {
529 chunk->map[i - 1] += chunk->map[i];
530 chunk->map_used--;
531 memmove(&chunk->map[i], &chunk->map[i + 1],
532 (chunk->map_used - i) * sizeof(chunk->map[0]));
533 i--;
534 }
535 /* merge with next? */
536 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
537 chunk->map[i] += chunk->map[i + 1];
538 chunk->map_used--;
539 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
540 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
541 }
542
543 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
544 pcpu_chunk_relocate(chunk, oslot);
545}
546
547/**
548 * pcpu_unmap - unmap pages out of a pcpu_chunk
549 * @chunk: chunk of interest
550 * @page_start: page index of the first page to unmap
551 * @page_end: page index of the last page to unmap + 1
552 * @flush: whether to flush cache and tlb or not
553 *
554 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
555 * If @flush is true, vcache is flushed before unmapping and tlb
556 * after.
557 */
558static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
559 bool flush)
560{
561 unsigned int last = num_possible_cpus() - 1;
562 unsigned int cpu;
563
Tejun Heo8d408b42009-02-24 11:57:21 +0900564 /* unmap must not be done on immutable chunk */
565 WARN_ON(chunk->immutable);
566
Tejun Heofbf59bc2009-02-20 16:29:08 +0900567 /*
568 * Each flushing trial can be very expensive, issue flush on
569 * the whole region at once rather than doing it for each cpu.
570 * This could be an overkill but is more scalable.
571 */
572 if (flush)
573 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
574 pcpu_chunk_addr(chunk, last, page_end));
575
576 for_each_possible_cpu(cpu)
577 unmap_kernel_range_noflush(
578 pcpu_chunk_addr(chunk, cpu, page_start),
579 (page_end - page_start) << PAGE_SHIFT);
580
581 /* ditto as flush_cache_vunmap() */
582 if (flush)
583 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
584 pcpu_chunk_addr(chunk, last, page_end));
585}
586
587/**
588 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
589 * @chunk: chunk to depopulate
590 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900591 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900592 * @flush: whether to flush cache and tlb or not
593 *
594 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
595 * from @chunk. If @flush is true, vcache is flushed before unmapping
596 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900597 *
598 * CONTEXT:
599 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900600 */
Tejun Heocae3aeb2009-02-21 16:56:23 +0900601static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size,
602 bool flush)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900603{
604 int page_start = PFN_DOWN(off);
605 int page_end = PFN_UP(off + size);
606 int unmap_start = -1;
607 int uninitialized_var(unmap_end);
608 unsigned int cpu;
609 int i;
610
611 for (i = page_start; i < page_end; i++) {
612 for_each_possible_cpu(cpu) {
613 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
614
615 if (!*pagep)
616 continue;
617
618 __free_page(*pagep);
619
620 /*
621 * If it's partial depopulation, it might get
622 * populated or depopulated again. Mark the
623 * page gone.
624 */
625 *pagep = NULL;
626
627 unmap_start = unmap_start < 0 ? i : unmap_start;
628 unmap_end = i + 1;
629 }
630 }
631
632 if (unmap_start >= 0)
633 pcpu_unmap(chunk, unmap_start, unmap_end, flush);
634}
635
636/**
637 * pcpu_map - map pages into a pcpu_chunk
638 * @chunk: chunk of interest
639 * @page_start: page index of the first page to map
640 * @page_end: page index of the last page to map + 1
641 *
642 * For each cpu, map pages [@page_start,@page_end) into @chunk.
643 * vcache is flushed afterwards.
644 */
645static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
646{
647 unsigned int last = num_possible_cpus() - 1;
648 unsigned int cpu;
649 int err;
650
Tejun Heo8d408b42009-02-24 11:57:21 +0900651 /* map must not be done on immutable chunk */
652 WARN_ON(chunk->immutable);
653
Tejun Heofbf59bc2009-02-20 16:29:08 +0900654 for_each_possible_cpu(cpu) {
655 err = map_kernel_range_noflush(
656 pcpu_chunk_addr(chunk, cpu, page_start),
657 (page_end - page_start) << PAGE_SHIFT,
658 PAGE_KERNEL,
659 pcpu_chunk_pagep(chunk, cpu, page_start));
660 if (err < 0)
661 return err;
662 }
663
664 /* flush at once, please read comments in pcpu_unmap() */
665 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
666 pcpu_chunk_addr(chunk, last, page_end));
667 return 0;
668}
669
670/**
671 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
672 * @chunk: chunk of interest
673 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900674 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900675 *
676 * For each cpu, populate and map pages [@page_start,@page_end) into
677 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900678 *
679 * CONTEXT:
680 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900681 */
682static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
683{
684 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
685 int page_start = PFN_DOWN(off);
686 int page_end = PFN_UP(off + size);
687 int map_start = -1;
Tejun Heo02d51fdf2009-03-01 15:42:36 +0900688 int uninitialized_var(map_end);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900689 unsigned int cpu;
690 int i;
691
692 for (i = page_start; i < page_end; i++) {
693 if (pcpu_chunk_page_occupied(chunk, i)) {
694 if (map_start >= 0) {
695 if (pcpu_map(chunk, map_start, map_end))
696 goto err;
697 map_start = -1;
698 }
699 continue;
700 }
701
702 map_start = map_start < 0 ? i : map_start;
703 map_end = i + 1;
704
705 for_each_possible_cpu(cpu) {
706 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
707
708 *pagep = alloc_pages_node(cpu_to_node(cpu),
709 alloc_mask, 0);
710 if (!*pagep)
711 goto err;
Christoph Lametere1b9aa32009-04-02 13:21:44 +0900712 pcpu_set_page_chunk(*pagep, chunk);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900713 }
714 }
715
716 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
717 goto err;
718
719 for_each_possible_cpu(cpu)
Tejun Heod9b55ee2009-02-24 11:57:21 +0900720 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0,
Tejun Heofbf59bc2009-02-20 16:29:08 +0900721 size);
722
723 return 0;
724err:
725 /* likely under heavy memory pressure, give memory back */
726 pcpu_depopulate_chunk(chunk, off, size, true);
727 return -ENOMEM;
728}
729
730static void free_pcpu_chunk(struct pcpu_chunk *chunk)
731{
732 if (!chunk)
733 return;
734 if (chunk->vm)
735 free_vm_area(chunk->vm);
Tejun Heo1880d932009-03-07 00:44:09 +0900736 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900737 kfree(chunk);
738}
739
740static struct pcpu_chunk *alloc_pcpu_chunk(void)
741{
742 struct pcpu_chunk *chunk;
743
744 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
745 if (!chunk)
746 return NULL;
747
Tejun Heo1880d932009-03-07 00:44:09 +0900748 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900749 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
750 chunk->map[chunk->map_used++] = pcpu_unit_size;
Tejun Heo3e24aa52009-03-06 14:33:59 +0900751 chunk->page = chunk->page_ar;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900752
753 chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL);
754 if (!chunk->vm) {
755 free_pcpu_chunk(chunk);
756 return NULL;
757 }
758
759 INIT_LIST_HEAD(&chunk->list);
760 chunk->free_size = pcpu_unit_size;
761 chunk->contig_hint = pcpu_unit_size;
762
763 return chunk;
764}
765
766/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900767 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900768 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900769 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900770 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900771 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900772 * Allocate percpu area of @size bytes aligned at @align.
773 *
774 * CONTEXT:
775 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900776 *
777 * RETURNS:
778 * Percpu pointer to the allocated area on success, NULL on failure.
779 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900780static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900781{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900782 struct pcpu_chunk *chunk;
783 int slot, off;
784
Tejun Heo8d408b42009-02-24 11:57:21 +0900785 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900786 WARN(true, "illegal size (%zu) or align (%zu) for "
787 "percpu allocation\n", size, align);
788 return NULL;
789 }
790
Tejun Heoccea34b2009-03-07 00:44:13 +0900791 mutex_lock(&pcpu_alloc_mutex);
792 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900793
Tejun Heoedcb4632009-03-06 14:33:59 +0900794 /* serve reserved allocations from the reserved chunk if available */
795 if (reserved && pcpu_reserved_chunk) {
796 chunk = pcpu_reserved_chunk;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900797 if (size > chunk->contig_hint ||
798 pcpu_extend_area_map(chunk) < 0)
Tejun Heoccea34b2009-03-07 00:44:13 +0900799 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900800 off = pcpu_alloc_area(chunk, size, align);
801 if (off >= 0)
802 goto area_found;
Tejun Heoccea34b2009-03-07 00:44:13 +0900803 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900804 }
805
Tejun Heoccea34b2009-03-07 00:44:13 +0900806restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900807 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900808 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
809 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
810 if (size > chunk->contig_hint)
811 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900812
813 switch (pcpu_extend_area_map(chunk)) {
814 case 0:
815 break;
816 case 1:
817 goto restart; /* pcpu_lock dropped, restart */
818 default:
819 goto fail_unlock;
820 }
821
Tejun Heofbf59bc2009-02-20 16:29:08 +0900822 off = pcpu_alloc_area(chunk, size, align);
823 if (off >= 0)
824 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900825 }
826 }
827
828 /* hmmm... no space left, create a new chunk */
Tejun Heoccea34b2009-03-07 00:44:13 +0900829 spin_unlock_irq(&pcpu_lock);
830
Tejun Heofbf59bc2009-02-20 16:29:08 +0900831 chunk = alloc_pcpu_chunk();
832 if (!chunk)
Tejun Heoccea34b2009-03-07 00:44:13 +0900833 goto fail_unlock_mutex;
834
835 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900836 pcpu_chunk_relocate(chunk, -1);
Tejun Heoccea34b2009-03-07 00:44:13 +0900837 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900838
839area_found:
Tejun Heoccea34b2009-03-07 00:44:13 +0900840 spin_unlock_irq(&pcpu_lock);
841
Tejun Heofbf59bc2009-02-20 16:29:08 +0900842 /* populate, map and clear the area */
843 if (pcpu_populate_chunk(chunk, off, size)) {
Tejun Heoccea34b2009-03-07 00:44:13 +0900844 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900845 pcpu_free_area(chunk, off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900846 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900847 }
848
Tejun Heoccea34b2009-03-07 00:44:13 +0900849 mutex_unlock(&pcpu_alloc_mutex);
850
851 return __addr_to_pcpu_ptr(chunk->vm->addr + off);
852
853fail_unlock:
854 spin_unlock_irq(&pcpu_lock);
855fail_unlock_mutex:
856 mutex_unlock(&pcpu_alloc_mutex);
857 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900858}
Tejun Heoedcb4632009-03-06 14:33:59 +0900859
860/**
861 * __alloc_percpu - allocate dynamic percpu area
862 * @size: size of area to allocate in bytes
863 * @align: alignment of area (max PAGE_SIZE)
864 *
865 * Allocate percpu area of @size bytes aligned at @align. Might
866 * sleep. Might trigger writeouts.
867 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900868 * CONTEXT:
869 * Does GFP_KERNEL allocation.
870 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900871 * RETURNS:
872 * Percpu pointer to the allocated area on success, NULL on failure.
873 */
874void *__alloc_percpu(size_t size, size_t align)
875{
876 return pcpu_alloc(size, align, false);
877}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900878EXPORT_SYMBOL_GPL(__alloc_percpu);
879
Tejun Heoedcb4632009-03-06 14:33:59 +0900880/**
881 * __alloc_reserved_percpu - allocate reserved percpu area
882 * @size: size of area to allocate in bytes
883 * @align: alignment of area (max PAGE_SIZE)
884 *
885 * Allocate percpu area of @size bytes aligned at @align from reserved
886 * percpu area if arch has set it up; otherwise, allocation is served
887 * from the same dynamic area. Might sleep. Might trigger writeouts.
888 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900889 * CONTEXT:
890 * Does GFP_KERNEL allocation.
891 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900892 * RETURNS:
893 * Percpu pointer to the allocated area on success, NULL on failure.
894 */
895void *__alloc_reserved_percpu(size_t size, size_t align)
896{
897 return pcpu_alloc(size, align, true);
898}
899
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900900/**
901 * pcpu_reclaim - reclaim fully free chunks, workqueue function
902 * @work: unused
903 *
904 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900905 *
906 * CONTEXT:
907 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900908 */
909static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900910{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900911 LIST_HEAD(todo);
912 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
913 struct pcpu_chunk *chunk, *next;
914
Tejun Heoccea34b2009-03-07 00:44:13 +0900915 mutex_lock(&pcpu_alloc_mutex);
916 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900917
918 list_for_each_entry_safe(chunk, next, head, list) {
919 WARN_ON(chunk->immutable);
920
921 /* spare the first one */
922 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
923 continue;
924
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900925 list_move(&chunk->list, &todo);
926 }
927
Tejun Heoccea34b2009-03-07 00:44:13 +0900928 spin_unlock_irq(&pcpu_lock);
929 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900930
931 list_for_each_entry_safe(chunk, next, &todo, list) {
932 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
933 free_pcpu_chunk(chunk);
934 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900935}
936
937/**
938 * free_percpu - free percpu area
939 * @ptr: pointer to area to free
940 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900941 * Free percpu area @ptr.
942 *
943 * CONTEXT:
944 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900945 */
946void free_percpu(void *ptr)
947{
948 void *addr = __pcpu_ptr_to_addr(ptr);
949 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900950 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900951 int off;
952
953 if (!ptr)
954 return;
955
Tejun Heoccea34b2009-03-07 00:44:13 +0900956 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900957
958 chunk = pcpu_chunk_addr_search(addr);
959 off = addr - chunk->vm->addr;
960
961 pcpu_free_area(chunk, off);
962
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900963 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900964 if (chunk->free_size == pcpu_unit_size) {
965 struct pcpu_chunk *pos;
966
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900967 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900968 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900969 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900970 break;
971 }
972 }
973
Tejun Heoccea34b2009-03-07 00:44:13 +0900974 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900975}
976EXPORT_SYMBOL_GPL(free_percpu);
977
978/**
Tejun Heo8d408b42009-02-24 11:57:21 +0900979 * pcpu_setup_first_chunk - initialize the first percpu chunk
980 * @get_page_fn: callback to fetch page pointer
981 * @static_size: the size of static percpu area in bytes
Tejun Heoedcb4632009-03-06 14:33:59 +0900982 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +0900983 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heo6074d5b2009-03-10 16:27:48 +0900984 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
Tejun Heo8d408b42009-02-24 11:57:21 +0900985 * @base_addr: mapped address, NULL for auto
986 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary
Tejun Heofbf59bc2009-02-20 16:29:08 +0900987 *
Tejun Heo8d408b42009-02-24 11:57:21 +0900988 * Initialize the first percpu chunk which contains the kernel static
989 * perpcu area. This function is to be called from arch percpu area
990 * setup path. The first two parameters are mandatory. The rest are
991 * optional.
992 *
993 * @get_page_fn() should return pointer to percpu page given cpu
994 * number and page number. It should at least return enough pages to
995 * cover the static area. The returned pages for static area should
996 * have been initialized with valid data. If @unit_size is specified,
997 * it can also return pages after the static area. NULL return
998 * indicates end of pages for the cpu. Note that @get_page_fn() must
999 * return the same number of pages for all cpus.
1000 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001001 * @reserved_size, if non-zero, specifies the amount of bytes to
1002 * reserve after the static area in the first chunk. This reserves
1003 * the first chunk such that it's available only through reserved
1004 * percpu allocation. This is primarily used to serve module percpu
1005 * static areas on architectures where the addressing model has
1006 * limited offset range for symbol relocations to guarantee module
1007 * percpu symbols fall inside the relocatable range.
1008 *
Tejun Heo6074d5b2009-03-10 16:27:48 +09001009 * @dyn_size, if non-negative, determines the number of bytes
1010 * available for dynamic allocation in the first chunk. Specifying
1011 * non-negative value makes percpu leave alone the area beyond
1012 * @static_size + @reserved_size + @dyn_size.
1013 *
Tejun Heocafe8812009-03-06 14:33:59 +09001014 * @unit_size, if non-negative, specifies unit size and must be
1015 * aligned to PAGE_SIZE and equal to or larger than @static_size +
Tejun Heo6074d5b2009-03-10 16:27:48 +09001016 * @reserved_size + if non-negative, @dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001017 *
1018 * Non-null @base_addr means that the caller already allocated virtual
1019 * region for the first chunk and mapped it. percpu must not mess
1020 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
1021 * @populate_pte_fn doesn't make any sense.
1022 *
1023 * @populate_pte_fn is used to populate the pagetable. NULL means the
1024 * caller already populated the pagetable.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001025 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001026 * If the first chunk ends up with both reserved and dynamic areas, it
1027 * is served by two chunks - one to serve the core static and reserved
1028 * areas and the other for the dynamic area. They share the same vm
1029 * and page map but uses different area allocation map to stay away
1030 * from each other. The latter chunk is circulated in the chunk slots
1031 * and available for dynamic allocation like any other chunks.
1032 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001033 * RETURNS:
1034 * The determined pcpu_unit_size which can be used to initialize
1035 * percpu access.
1036 */
Tejun Heo8d408b42009-02-24 11:57:21 +09001037size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
Tejun Heoedcb4632009-03-06 14:33:59 +09001038 size_t static_size, size_t reserved_size,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001039 ssize_t dyn_size, ssize_t unit_size,
Tejun Heocafe8812009-03-06 14:33:59 +09001040 void *base_addr,
Tejun Heo8d408b42009-02-24 11:57:21 +09001041 pcpu_populate_pte_fn_t populate_pte_fn)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001042{
Tejun Heo2441d152009-03-06 14:33:59 +09001043 static struct vm_struct first_vm;
Tejun Heoedcb4632009-03-06 14:33:59 +09001044 static int smap[2], dmap[2];
Tejun Heo6074d5b2009-03-10 16:27:48 +09001045 size_t size_sum = static_size + reserved_size +
1046 (dyn_size >= 0 ? dyn_size : 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001047 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001048 unsigned int cpu;
Tejun Heo8d408b42009-02-24 11:57:21 +09001049 int nr_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001050 int err, i;
1051
Tejun Heo8d408b42009-02-24 11:57:21 +09001052 /* santiy checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001053 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1054 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo8d408b42009-02-24 11:57:21 +09001055 BUG_ON(!static_size);
Tejun Heocafe8812009-03-06 14:33:59 +09001056 if (unit_size >= 0) {
Tejun Heo6074d5b2009-03-10 16:27:48 +09001057 BUG_ON(unit_size < size_sum);
Tejun Heocafe8812009-03-06 14:33:59 +09001058 BUG_ON(unit_size & ~PAGE_MASK);
Tejun Heo6074d5b2009-03-10 16:27:48 +09001059 BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE);
1060 } else
Tejun Heocafe8812009-03-06 14:33:59 +09001061 BUG_ON(base_addr);
Tejun Heo8d408b42009-02-24 11:57:21 +09001062 BUG_ON(base_addr && populate_pte_fn);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001063
Tejun Heocafe8812009-03-06 14:33:59 +09001064 if (unit_size >= 0)
Tejun Heo8d408b42009-02-24 11:57:21 +09001065 pcpu_unit_pages = unit_size >> PAGE_SHIFT;
1066 else
1067 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001068 PFN_UP(size_sum));
Tejun Heo8d408b42009-02-24 11:57:21 +09001069
Tejun Heod9b55ee2009-02-24 11:57:21 +09001070 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001071 pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001072 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
Tejun Heocb83b422009-02-24 11:57:20 +09001073 + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001074
Tejun Heocafe8812009-03-06 14:33:59 +09001075 if (dyn_size < 0)
Tejun Heoedcb4632009-03-06 14:33:59 +09001076 dyn_size = pcpu_unit_size - static_size - reserved_size;
Tejun Heocafe8812009-03-06 14:33:59 +09001077
Tejun Heod9b55ee2009-02-24 11:57:21 +09001078 /*
1079 * Allocate chunk slots. The additional last slot is for
1080 * empty chunks.
1081 */
1082 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001083 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1084 for (i = 0; i < pcpu_nr_slots; i++)
1085 INIT_LIST_HEAD(&pcpu_slot[i]);
1086
Tejun Heoedcb4632009-03-06 14:33:59 +09001087 /*
1088 * Initialize static chunk. If reserved_size is zero, the
1089 * static chunk covers static area + dynamic allocation area
1090 * in the first chunk. If reserved_size is not zero, it
1091 * covers static area + reserved area (mostly used for module
1092 * static percpu allocation).
1093 */
Tejun Heo2441d152009-03-06 14:33:59 +09001094 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1095 INIT_LIST_HEAD(&schunk->list);
1096 schunk->vm = &first_vm;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001097 schunk->map = smap;
1098 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo3e24aa52009-03-06 14:33:59 +09001099 schunk->page = schunk->page_ar;
Tejun Heoedcb4632009-03-06 14:33:59 +09001100
1101 if (reserved_size) {
1102 schunk->free_size = reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001103 pcpu_reserved_chunk = schunk;
1104 pcpu_reserved_chunk_limit = static_size + reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001105 } else {
1106 schunk->free_size = dyn_size;
1107 dyn_size = 0; /* dynamic area covered */
1108 }
Tejun Heo2441d152009-03-06 14:33:59 +09001109 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001110
Tejun Heo61ace7f2009-03-06 14:33:59 +09001111 schunk->map[schunk->map_used++] = -static_size;
1112 if (schunk->free_size)
1113 schunk->map[schunk->map_used++] = schunk->free_size;
1114
Tejun Heoedcb4632009-03-06 14:33:59 +09001115 /* init dynamic chunk if necessary */
1116 if (dyn_size) {
1117 dchunk = alloc_bootmem(sizeof(struct pcpu_chunk));
1118 INIT_LIST_HEAD(&dchunk->list);
1119 dchunk->vm = &first_vm;
1120 dchunk->map = dmap;
1121 dchunk->map_alloc = ARRAY_SIZE(dmap);
1122 dchunk->page = schunk->page_ar; /* share page map with schunk */
1123
1124 dchunk->contig_hint = dchunk->free_size = dyn_size;
1125 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1126 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1127 }
1128
Tejun Heo8d408b42009-02-24 11:57:21 +09001129 /* allocate vm address */
Tejun Heo2441d152009-03-06 14:33:59 +09001130 first_vm.flags = VM_ALLOC;
1131 first_vm.size = pcpu_chunk_size;
Tejun Heo8d408b42009-02-24 11:57:21 +09001132
1133 if (!base_addr)
Tejun Heo2441d152009-03-06 14:33:59 +09001134 vm_area_register_early(&first_vm, PAGE_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001135 else {
1136 /*
1137 * Pages already mapped. No need to remap into
Tejun Heoedcb4632009-03-06 14:33:59 +09001138 * vmalloc area. In this case the first chunks can't
1139 * be mapped or unmapped by percpu and are marked
Tejun Heo8d408b42009-02-24 11:57:21 +09001140 * immutable.
1141 */
Tejun Heo2441d152009-03-06 14:33:59 +09001142 first_vm.addr = base_addr;
1143 schunk->immutable = true;
Tejun Heoedcb4632009-03-06 14:33:59 +09001144 if (dchunk)
1145 dchunk->immutable = true;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001146 }
1147
Tejun Heo8d408b42009-02-24 11:57:21 +09001148 /* assign pages */
1149 nr_pages = -1;
1150 for_each_possible_cpu(cpu) {
1151 for (i = 0; i < pcpu_unit_pages; i++) {
1152 struct page *page = get_page_fn(cpu, i);
1153
1154 if (!page)
1155 break;
Tejun Heo2441d152009-03-06 14:33:59 +09001156 *pcpu_chunk_pagep(schunk, cpu, i) = page;
Tejun Heo8d408b42009-02-24 11:57:21 +09001157 }
1158
Tejun Heo61ace7f2009-03-06 14:33:59 +09001159 BUG_ON(i < PFN_UP(static_size));
Tejun Heo8d408b42009-02-24 11:57:21 +09001160
1161 if (nr_pages < 0)
1162 nr_pages = i;
1163 else
1164 BUG_ON(nr_pages != i);
1165 }
1166
1167 /* map them */
1168 if (populate_pte_fn) {
1169 for_each_possible_cpu(cpu)
1170 for (i = 0; i < nr_pages; i++)
Tejun Heo2441d152009-03-06 14:33:59 +09001171 populate_pte_fn(pcpu_chunk_addr(schunk,
Tejun Heo8d408b42009-02-24 11:57:21 +09001172 cpu, i));
1173
Tejun Heo2441d152009-03-06 14:33:59 +09001174 err = pcpu_map(schunk, 0, nr_pages);
Tejun Heo8d408b42009-02-24 11:57:21 +09001175 if (err)
1176 panic("failed to setup static percpu area, err=%d\n",
1177 err);
1178 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001179
Tejun Heo2441d152009-03-06 14:33:59 +09001180 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001181 pcpu_first_chunk = dchunk ?: schunk;
1182 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001183
1184 /* we're done */
Tejun Heo2441d152009-03-06 14:33:59 +09001185 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001186 return pcpu_unit_size;
1187}
Tejun Heo66c3a752009-03-10 16:27:48 +09001188
1189/*
1190 * Embedding first chunk setup helper.
1191 */
1192static void *pcpue_ptr __initdata;
1193static size_t pcpue_size __initdata;
1194static size_t pcpue_unit_size __initdata;
1195
1196static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
1197{
1198 size_t off = (size_t)pageno << PAGE_SHIFT;
1199
1200 if (off >= pcpue_size)
1201 return NULL;
1202
1203 return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off);
1204}
1205
1206/**
1207 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1208 * @static_size: the size of static percpu area in bytes
1209 * @reserved_size: the size of reserved percpu area in bytes
1210 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1211 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
1212 *
1213 * This is a helper to ease setting up embedded first percpu chunk and
1214 * can be called where pcpu_setup_first_chunk() is expected.
1215 *
1216 * If this function is used to setup the first chunk, it is allocated
1217 * as a contiguous area using bootmem allocator and used as-is without
1218 * being mapped into vmalloc area. This enables the first chunk to
1219 * piggy back on the linear physical mapping which often uses larger
1220 * page size.
1221 *
1222 * When @dyn_size is positive, dynamic area might be larger than
1223 * specified to fill page alignment. Also, when @dyn_size is auto,
1224 * @dyn_size does not fill the whole first chunk but only what's
1225 * necessary for page alignment after static and reserved areas.
1226 *
1227 * If the needed size is smaller than the minimum or specified unit
1228 * size, the leftover is returned to the bootmem allocator.
1229 *
1230 * RETURNS:
1231 * The determined pcpu_unit_size which can be used to initialize
1232 * percpu access on success, -errno on failure.
1233 */
1234ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size,
1235 ssize_t dyn_size, ssize_t unit_size)
1236{
1237 unsigned int cpu;
1238
1239 /* determine parameters and allocate */
1240 pcpue_size = PFN_ALIGN(static_size + reserved_size +
1241 (dyn_size >= 0 ? dyn_size : 0));
1242 if (dyn_size != 0)
1243 dyn_size = pcpue_size - static_size - reserved_size;
1244
1245 if (unit_size >= 0) {
1246 BUG_ON(unit_size < pcpue_size);
1247 pcpue_unit_size = unit_size;
1248 } else
1249 pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE);
1250
1251 pcpue_ptr = __alloc_bootmem_nopanic(
1252 num_possible_cpus() * pcpue_unit_size,
1253 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
1254 if (!pcpue_ptr)
1255 return -ENOMEM;
1256
1257 /* return the leftover and copy */
1258 for_each_possible_cpu(cpu) {
1259 void *ptr = pcpue_ptr + cpu * pcpue_unit_size;
1260
1261 free_bootmem(__pa(ptr + pcpue_size),
1262 pcpue_unit_size - pcpue_size);
1263 memcpy(ptr, __per_cpu_load, static_size);
1264 }
1265
1266 /* we're ready, commit */
1267 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
1268 pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size);
1269
1270 return pcpu_setup_first_chunk(pcpue_get_page, static_size,
1271 reserved_size, dyn_size,
1272 pcpue_unit_size, pcpue_ptr, NULL);
1273}