blob: 255714302394137e37e4d8bd5cf87d685732d9d6 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
Tejun Heo88999a82010-04-09 18:57:01 +09002 * mm/percpu.c - percpu memory allocator
Tejun Heofbf59bc2009-02-20 16:29:08 +09003 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
Tejun Heo88999a82010-04-09 18:57:01 +090010 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
Tejun Heo2f39e632009-07-04 08:11:00 +090013 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
Tejun Heo88999a82010-04-09 18:57:01 +090016 * When a chunk is filled up, another chunk is allocated.
Tejun Heofbf59bc2009-02-20 16:29:08 +090017 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
Tejun Heo2f39e632009-07-04 08:11:00 +090025 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
Tejun Heofbf59bc2009-02-20 16:29:08 +090029 *
Tejun Heo2f39e632009-07-04 08:11:00 +090030 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
Tejun Heofbf59bc2009-02-20 16:29:08 +090032 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
Namhyung Kim47858792010-08-11 11:24:10 +090034 * guaranteed to be equal to or larger than the maximum contiguous
Tejun Heofbf59bc2009-02-20 16:29:08 +090035 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
Christoph Lametere1b9aa32009-04-02 13:21:44 +090043 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
Tejun Heofbf59bc2009-02-20 16:29:08 +090045 *
46 * To use this allocator, arch code should do the followings.
47 *
Tejun Heofbf59bc2009-02-20 16:29:08 +090048 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
Joe Perches870d4b12016-03-17 14:19:53 -070056#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
57
Tejun Heofbf59bc2009-02-20 16:29:08 +090058#include <linux/bitmap.h>
59#include <linux/bootmem.h>
Tejun Heofd1e8a12009-08-14 15:00:51 +090060#include <linux/err.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090061#include <linux/list.h>
Tejun Heoa530b792009-07-04 08:11:00 +090062#include <linux/log2.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090063#include <linux/mm.h>
64#include <linux/module.h>
65#include <linux/mutex.h>
66#include <linux/percpu.h>
67#include <linux/pfn.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090068#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090069#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090070#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090071#include <linux/workqueue.h>
Catalin Marinasf528f0b2011-09-26 17:12:53 +010072#include <linux/kmemleak.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090073
74#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090075#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090076#include <asm/tlbflush.h>
Vivek Goyal3b034b02009-11-24 15:50:03 +090077#include <asm/io.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090078
Tejun Heofbf59bc2009-02-20 16:29:08 +090079#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
80#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
Tejun Heo9c824b62014-09-02 14:46:05 -040081#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
82#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
Tejun Heo1a4d7602014-09-02 14:46:05 -040083#define PCPU_EMPTY_POP_PAGES_LOW 2
84#define PCPU_EMPTY_POP_PAGES_HIGH 4
Tejun Heofbf59bc2009-02-20 16:29:08 +090085
Tejun Heobbddff02010-09-03 18:22:48 +020086#ifdef CONFIG_SMP
Tejun Heoe0100982009-03-10 16:27:48 +090087/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
88#ifndef __addr_to_pcpu_ptr
89#define __addr_to_pcpu_ptr(addr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090090 (void __percpu *)((unsigned long)(addr) - \
91 (unsigned long)pcpu_base_addr + \
92 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090093#endif
94#ifndef __pcpu_ptr_to_addr
95#define __pcpu_ptr_to_addr(ptr) \
Tejun Heo43cf38e2010-02-02 14:38:57 +090096 (void __force *)((unsigned long)(ptr) + \
97 (unsigned long)pcpu_base_addr - \
98 (unsigned long)__per_cpu_start)
Tejun Heoe0100982009-03-10 16:27:48 +090099#endif
Tejun Heobbddff02010-09-03 18:22:48 +0200100#else /* CONFIG_SMP */
101/* on UP, it's always identity mapped */
102#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
103#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
104#endif /* CONFIG_SMP */
Tejun Heoe0100982009-03-10 16:27:48 +0900105
Tejun Heofbf59bc2009-02-20 16:29:08 +0900106struct pcpu_chunk {
107 struct list_head list; /* linked to pcpu_slot lists */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900108 int free_size; /* free bytes in the chunk */
109 int contig_hint; /* max contiguous size hint */
Tejun Heobba174f2009-08-14 15:00:51 +0900110 void *base_addr; /* base address of this chunk */
Tejun Heo9c824b62014-09-02 14:46:05 -0400111
Al Viro723ad1d2014-03-06 21:13:18 -0500112 int map_used; /* # of map entries used before the sentry */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900113 int map_alloc; /* # of map entries allocated */
114 int *map; /* allocation map */
Tejun Heo4f996e22016-05-25 11:48:25 -0400115 struct list_head map_extend_list;/* on pcpu_map_extend_chunks */
Tejun Heo9c824b62014-09-02 14:46:05 -0400116
Tejun Heo88999a82010-04-09 18:57:01 +0900117 void *data; /* chunk data */
Al Viro3d331ad2014-03-06 20:52:32 -0500118 int first_free; /* no free below this */
Tejun Heo8d408b42009-02-24 11:57:21 +0900119 bool immutable; /* no [de]population allowed */
Tejun Heob539b872014-09-02 14:46:05 -0400120 int nr_populated; /* # of populated pages */
Tejun Heoce3141a2009-07-04 08:11:00 +0900121 unsigned long populated[]; /* populated bitmap */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900122};
123
Tejun Heo40150d32009-02-24 12:32:28 +0900124static int pcpu_unit_pages __read_mostly;
125static int pcpu_unit_size __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900126static int pcpu_nr_units __read_mostly;
Tejun Heo65632972009-08-14 15:00:52 +0900127static int pcpu_atom_size __read_mostly;
Tejun Heo40150d32009-02-24 12:32:28 +0900128static int pcpu_nr_slots __read_mostly;
129static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900130
Tejun Heoa855b842011-11-18 10:55:35 -0800131/* cpus with the lowest and highest unit addresses */
132static unsigned int pcpu_low_unit_cpu __read_mostly;
133static unsigned int pcpu_high_unit_cpu __read_mostly;
Tejun Heo2f39e632009-07-04 08:11:00 +0900134
Tejun Heofbf59bc2009-02-20 16:29:08 +0900135/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900136void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900137EXPORT_SYMBOL_GPL(pcpu_base_addr);
138
Tejun Heofb435d52009-08-14 15:00:51 +0900139static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
140const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
Tejun Heo2f39e632009-07-04 08:11:00 +0900141
Tejun Heo65632972009-08-14 15:00:52 +0900142/* group information, used for vm allocation */
143static int pcpu_nr_groups __read_mostly;
144static const unsigned long *pcpu_group_offsets __read_mostly;
145static const size_t *pcpu_group_sizes __read_mostly;
146
Tejun Heoae9e6bc92009-04-02 13:19:54 +0900147/*
148 * The first chunk which always exists. Note that unlike other
149 * chunks, this one can be allocated and mapped in several different
150 * ways and thus often doesn't live in the vmalloc area.
151 */
152static struct pcpu_chunk *pcpu_first_chunk;
153
154/*
155 * Optional reserved chunk. This chunk reserves part of the first
156 * chunk and serves it for reserved allocations. The amount of
157 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
158 * area doesn't exist, the following variables contain NULL and 0
159 * respectively.
160 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900161static struct pcpu_chunk *pcpu_reserved_chunk;
Tejun Heoedcb4632009-03-06 14:33:59 +0900162static int pcpu_reserved_chunk_limit;
163
Tejun Heob38d08f2014-09-02 14:46:02 -0400164static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
Tejun Heo6710e592016-05-25 11:48:25 -0400165static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900166
Tejun Heo40150d32009-02-24 12:32:28 +0900167static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900168
Tejun Heo4f996e22016-05-25 11:48:25 -0400169/* chunks which need their map areas extended, protected by pcpu_lock */
170static LIST_HEAD(pcpu_map_extend_chunks);
171
Tejun Heob539b872014-09-02 14:46:05 -0400172/*
173 * The number of empty populated pages, protected by pcpu_lock. The
174 * reserved chunk doesn't contribute to the count.
175 */
176static int pcpu_nr_empty_pop_pages;
177
Tejun Heo1a4d7602014-09-02 14:46:05 -0400178/*
179 * Balance work is used to populate or destroy chunks asynchronously. We
180 * try to keep the number of populated free pages between
181 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
182 * empty chunk.
183 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -0400184static void pcpu_balance_workfn(struct work_struct *work);
185static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
Tejun Heo1a4d7602014-09-02 14:46:05 -0400186static bool pcpu_async_enabled __read_mostly;
187static bool pcpu_atomic_alloc_failed;
188
189static void pcpu_schedule_balance_work(void)
190{
191 if (pcpu_async_enabled)
192 schedule_work(&pcpu_balance_work);
193}
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900194
Tejun Heo020ec652010-04-09 18:57:00 +0900195static bool pcpu_addr_in_first_chunk(void *addr)
196{
197 void *first_start = pcpu_first_chunk->base_addr;
198
199 return addr >= first_start && addr < first_start + pcpu_unit_size;
200}
201
202static bool pcpu_addr_in_reserved_chunk(void *addr)
203{
204 void *first_start = pcpu_first_chunk->base_addr;
205
206 return addr >= first_start &&
207 addr < first_start + pcpu_reserved_chunk_limit;
208}
209
Tejun Heod9b55ee2009-02-24 11:57:21 +0900210static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900211{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900212 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900213 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
214}
215
Tejun Heod9b55ee2009-02-24 11:57:21 +0900216static int pcpu_size_to_slot(int size)
217{
218 if (size == pcpu_unit_size)
219 return pcpu_nr_slots - 1;
220 return __pcpu_size_to_slot(size);
221}
222
Tejun Heofbf59bc2009-02-20 16:29:08 +0900223static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
224{
225 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
226 return 0;
227
228 return pcpu_size_to_slot(chunk->free_size);
229}
230
Tejun Heo88999a82010-04-09 18:57:01 +0900231/* set the pointer to a chunk in a page struct */
232static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
233{
234 page->index = (unsigned long)pcpu;
235}
236
237/* obtain pointer to a chunk from a page struct */
238static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
239{
240 return (struct pcpu_chunk *)page->index;
241}
242
243static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900244{
Tejun Heo2f39e632009-07-04 08:11:00 +0900245 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900246}
247
Tejun Heo9983b6f02010-06-18 11:44:31 +0200248static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
249 unsigned int cpu, int page_idx)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900250{
Tejun Heobba174f2009-08-14 15:00:51 +0900251 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
Tejun Heofb435d52009-08-14 15:00:51 +0900252 (page_idx << PAGE_SHIFT);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900253}
254
Tejun Heo88999a82010-04-09 18:57:01 +0900255static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
256 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900257{
258 *rs = find_next_zero_bit(chunk->populated, end, *rs);
259 *re = find_next_bit(chunk->populated, end, *rs + 1);
260}
261
Tejun Heo88999a82010-04-09 18:57:01 +0900262static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
263 int *rs, int *re, int end)
Tejun Heoce3141a2009-07-04 08:11:00 +0900264{
265 *rs = find_next_bit(chunk->populated, end, *rs);
266 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
267}
268
269/*
270 * (Un)populated page region iterators. Iterate over (un)populated
Uwe Kleine-Königb5950762010-11-01 15:38:34 -0400271 * page regions between @start and @end in @chunk. @rs and @re should
Tejun Heoce3141a2009-07-04 08:11:00 +0900272 * be integer variables and will be set to start and end page index of
273 * the current region.
274 */
275#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
276 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
277 (rs) < (re); \
278 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
279
280#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
281 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
282 (rs) < (re); \
283 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
284
Tejun Heofbf59bc2009-02-20 16:29:08 +0900285/**
Bob Liu90459ce02011-08-04 11:02:33 +0200286 * pcpu_mem_zalloc - allocate memory
Tejun Heo1880d932009-03-07 00:44:09 +0900287 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900288 *
Tejun Heo1880d932009-03-07 00:44:09 +0900289 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
Bob Liu90459ce02011-08-04 11:02:33 +0200290 * kzalloc() is used; otherwise, vzalloc() is used. The returned
Tejun Heo1880d932009-03-07 00:44:09 +0900291 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900292 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900293 * CONTEXT:
294 * Does GFP_KERNEL allocation.
295 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900296 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900297 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900298 */
Bob Liu90459ce02011-08-04 11:02:33 +0200299static void *pcpu_mem_zalloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900300{
Tejun Heo099a19d2010-06-27 18:50:00 +0200301 if (WARN_ON_ONCE(!slab_is_available()))
302 return NULL;
303
Tejun Heofbf59bc2009-02-20 16:29:08 +0900304 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900305 return kzalloc(size, GFP_KERNEL);
Jesper Juhl7af4c092010-10-30 15:56:54 +0200306 else
307 return vzalloc(size);
Tejun Heo1880d932009-03-07 00:44:09 +0900308}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900309
Tejun Heo1880d932009-03-07 00:44:09 +0900310/**
311 * pcpu_mem_free - free memory
312 * @ptr: memory to free
Tejun Heo1880d932009-03-07 00:44:09 +0900313 *
Bob Liu90459ce02011-08-04 11:02:33 +0200314 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
Tejun Heo1880d932009-03-07 00:44:09 +0900315 */
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800316static void pcpu_mem_free(void *ptr)
Tejun Heo1880d932009-03-07 00:44:09 +0900317{
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800318 kvfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900319}
320
321/**
Tejun Heob539b872014-09-02 14:46:05 -0400322 * pcpu_count_occupied_pages - count the number of pages an area occupies
323 * @chunk: chunk of interest
324 * @i: index of the area in question
325 *
326 * Count the number of pages chunk's @i'th area occupies. When the area's
327 * start and/or end address isn't aligned to page boundary, the straddled
328 * page is included in the count iff the rest of the page is free.
329 */
330static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
331{
332 int off = chunk->map[i] & ~1;
333 int end = chunk->map[i + 1] & ~1;
334
335 if (!PAGE_ALIGNED(off) && i > 0) {
336 int prev = chunk->map[i - 1];
337
338 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
339 off = round_down(off, PAGE_SIZE);
340 }
341
342 if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
343 int next = chunk->map[i + 1];
344 int nend = chunk->map[i + 2] & ~1;
345
346 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
347 end = round_up(end, PAGE_SIZE);
348 }
349
350 return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
351}
352
353/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900354 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
355 * @chunk: chunk of interest
356 * @oslot: the previous slot it was on
357 *
358 * This function is called after an allocation or free changed @chunk.
359 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900360 * moved to the slot. Note that the reserved chunk is never put on
361 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900362 *
363 * CONTEXT:
364 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900365 */
366static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
367{
368 int nslot = pcpu_chunk_slot(chunk);
369
Tejun Heoedcb4632009-03-06 14:33:59 +0900370 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900371 if (oslot < nslot)
372 list_move(&chunk->list, &pcpu_slot[nslot]);
373 else
374 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
375 }
376}
377
Tejun Heofbf59bc2009-02-20 16:29:08 +0900378/**
Tejun Heo833af842009-11-11 15:35:18 +0900379 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
380 * @chunk: chunk of interest
Tejun Heo9c824b62014-09-02 14:46:05 -0400381 * @is_atomic: the allocation context
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900382 *
Tejun Heo9c824b62014-09-02 14:46:05 -0400383 * Determine whether area map of @chunk needs to be extended. If
384 * @is_atomic, only the amount necessary for a new allocation is
385 * considered; however, async extension is scheduled if the left amount is
386 * low. If !@is_atomic, it aims for more empty space. Combined, this
387 * ensures that the map is likely to have enough available space to
388 * accomodate atomic allocations which can't extend maps directly.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900389 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900390 * CONTEXT:
Tejun Heo833af842009-11-11 15:35:18 +0900391 * pcpu_lock.
Tejun Heoccea34b2009-03-07 00:44:13 +0900392 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900393 * RETURNS:
Tejun Heo833af842009-11-11 15:35:18 +0900394 * New target map allocation length if extension is necessary, 0
395 * otherwise.
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900396 */
Tejun Heo9c824b62014-09-02 14:46:05 -0400397static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900398{
Tejun Heo9c824b62014-09-02 14:46:05 -0400399 int margin, new_alloc;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900400
Tejun Heo4f996e22016-05-25 11:48:25 -0400401 lockdep_assert_held(&pcpu_lock);
402
Tejun Heo9c824b62014-09-02 14:46:05 -0400403 if (is_atomic) {
404 margin = 3;
405
406 if (chunk->map_alloc <
Tejun Heo4f996e22016-05-25 11:48:25 -0400407 chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) {
408 if (list_empty(&chunk->map_extend_list)) {
409 list_add_tail(&chunk->map_extend_list,
410 &pcpu_map_extend_chunks);
411 pcpu_schedule_balance_work();
412 }
413 }
Tejun Heo9c824b62014-09-02 14:46:05 -0400414 } else {
415 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
416 }
417
418 if (chunk->map_alloc >= chunk->map_used + margin)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900419 return 0;
420
421 new_alloc = PCPU_DFL_MAP_ALLOC;
Tejun Heo9c824b62014-09-02 14:46:05 -0400422 while (new_alloc < chunk->map_used + margin)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900423 new_alloc *= 2;
424
Tejun Heo833af842009-11-11 15:35:18 +0900425 return new_alloc;
426}
427
428/**
429 * pcpu_extend_area_map - extend area map of a chunk
430 * @chunk: chunk of interest
431 * @new_alloc: new target allocation length of the area map
432 *
433 * Extend area map of @chunk to have @new_alloc entries.
434 *
435 * CONTEXT:
436 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
437 *
438 * RETURNS:
439 * 0 on success, -errno on failure.
440 */
441static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
442{
443 int *old = NULL, *new = NULL;
444 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
445 unsigned long flags;
446
Tejun Heo6710e592016-05-25 11:48:25 -0400447 lockdep_assert_held(&pcpu_alloc_mutex);
448
Bob Liu90459ce02011-08-04 11:02:33 +0200449 new = pcpu_mem_zalloc(new_size);
Tejun Heo833af842009-11-11 15:35:18 +0900450 if (!new)
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900451 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900452
Tejun Heo833af842009-11-11 15:35:18 +0900453 /* acquire pcpu_lock and switch to new area map */
454 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900455
Tejun Heo833af842009-11-11 15:35:18 +0900456 if (new_alloc <= chunk->map_alloc)
457 goto out_unlock;
458
459 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
Huang Shijiea002d142010-08-08 14:39:07 +0200460 old = chunk->map;
461
462 memcpy(new, old, old_size);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900463
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900464 chunk->map_alloc = new_alloc;
465 chunk->map = new;
Tejun Heo833af842009-11-11 15:35:18 +0900466 new = NULL;
467
468out_unlock:
469 spin_unlock_irqrestore(&pcpu_lock, flags);
470
471 /*
472 * pcpu_mem_free() might end up calling vfree() which uses
473 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
474 */
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800475 pcpu_mem_free(old);
476 pcpu_mem_free(new);
Tejun Heo833af842009-11-11 15:35:18 +0900477
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900478 return 0;
479}
480
481/**
Tejun Heoa16037c2014-09-02 14:46:02 -0400482 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
483 * @chunk: chunk the candidate area belongs to
484 * @off: the offset to the start of the candidate area
485 * @this_size: the size of the candidate area
486 * @size: the size of the target allocation
487 * @align: the alignment of the target allocation
488 * @pop_only: only allocate from already populated region
489 *
490 * We're trying to allocate @size bytes aligned at @align. @chunk's area
491 * at @off sized @this_size is a candidate. This function determines
492 * whether the target allocation fits in the candidate area and returns the
493 * number of bytes to pad after @off. If the target area doesn't fit, -1
494 * is returned.
495 *
496 * If @pop_only is %true, this function only considers the already
497 * populated part of the candidate area.
498 */
499static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
500 int size, int align, bool pop_only)
501{
502 int cand_off = off;
503
504 while (true) {
505 int head = ALIGN(cand_off, align) - off;
506 int page_start, page_end, rs, re;
507
508 if (this_size < head + size)
509 return -1;
510
511 if (!pop_only)
512 return head;
513
514 /*
515 * If the first unpopulated page is beyond the end of the
516 * allocation, the whole allocation is populated;
517 * otherwise, retry from the end of the unpopulated area.
518 */
519 page_start = PFN_DOWN(head + off);
520 page_end = PFN_UP(head + off + size);
521
522 rs = page_start;
523 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
524 if (rs >= page_end)
525 return head;
526 cand_off = re * PAGE_SIZE;
527 }
528}
529
530/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900531 * pcpu_alloc_area - allocate area from a pcpu_chunk
532 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900533 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900534 * @align: wanted align
Tejun Heoa16037c2014-09-02 14:46:02 -0400535 * @pop_only: allocate only from the populated area
Tejun Heob539b872014-09-02 14:46:05 -0400536 * @occ_pages_p: out param for the number of pages the area occupies
Tejun Heofbf59bc2009-02-20 16:29:08 +0900537 *
538 * Try to allocate @size bytes area aligned at @align from @chunk.
539 * Note that this function only allocates the offset. It doesn't
540 * populate or map the area.
541 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900542 * @chunk->map must have at least two free slots.
543 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900544 * CONTEXT:
545 * pcpu_lock.
546 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900547 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900548 * Allocated offset in @chunk on success, -1 if no matching area is
549 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900550 */
Tejun Heoa16037c2014-09-02 14:46:02 -0400551static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
Tejun Heob539b872014-09-02 14:46:05 -0400552 bool pop_only, int *occ_pages_p)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900553{
554 int oslot = pcpu_chunk_slot(chunk);
555 int max_contig = 0;
556 int i, off;
Al Viro3d331ad2014-03-06 20:52:32 -0500557 bool seen_free = false;
Al Viro723ad1d2014-03-06 21:13:18 -0500558 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900559
Al Viro3d331ad2014-03-06 20:52:32 -0500560 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900561 int head, tail;
Al Viro723ad1d2014-03-06 21:13:18 -0500562 int this_size;
563
564 off = *p;
565 if (off & 1)
566 continue;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900567
Al Viro723ad1d2014-03-06 21:13:18 -0500568 this_size = (p[1] & ~1) - off;
Tejun Heoa16037c2014-09-02 14:46:02 -0400569
570 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
571 pop_only);
572 if (head < 0) {
Al Viro3d331ad2014-03-06 20:52:32 -0500573 if (!seen_free) {
574 chunk->first_free = i;
575 seen_free = true;
576 }
Al Viro723ad1d2014-03-06 21:13:18 -0500577 max_contig = max(this_size, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900578 continue;
579 }
580
581 /*
582 * If head is small or the previous block is free,
583 * merge'em. Note that 'small' is defined as smaller
584 * than sizeof(int), which is very small but isn't too
585 * uncommon for percpu allocations.
586 */
Al Viro723ad1d2014-03-06 21:13:18 -0500587 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800588 *p = off += head;
Al Viro723ad1d2014-03-06 21:13:18 -0500589 if (p[-1] & 1)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900590 chunk->free_size -= head;
Jianyu Zhan21ddfd32014-03-28 20:55:21 +0800591 else
592 max_contig = max(*p - p[-1], max_contig);
Al Viro723ad1d2014-03-06 21:13:18 -0500593 this_size -= head;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900594 head = 0;
595 }
596
597 /* if tail is small, just keep it around */
Al Viro723ad1d2014-03-06 21:13:18 -0500598 tail = this_size - head - size;
599 if (tail < sizeof(int)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900600 tail = 0;
Al Viro723ad1d2014-03-06 21:13:18 -0500601 size = this_size - head;
602 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900603
604 /* split if warranted */
605 if (head || tail) {
Al Viro706c16f2014-03-06 21:08:24 -0500606 int nr_extra = !!head + !!tail;
607
608 /* insert new subblocks */
Al Viro723ad1d2014-03-06 21:13:18 -0500609 memmove(p + nr_extra + 1, p + 1,
Al Viro706c16f2014-03-06 21:08:24 -0500610 sizeof(chunk->map[0]) * (chunk->map_used - i));
611 chunk->map_used += nr_extra;
612
Tejun Heofbf59bc2009-02-20 16:29:08 +0900613 if (head) {
Al Viro3d331ad2014-03-06 20:52:32 -0500614 if (!seen_free) {
615 chunk->first_free = i;
616 seen_free = true;
617 }
Al Viro723ad1d2014-03-06 21:13:18 -0500618 *++p = off += head;
619 ++i;
Al Viro706c16f2014-03-06 21:08:24 -0500620 max_contig = max(head, max_contig);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900621 }
Al Viro706c16f2014-03-06 21:08:24 -0500622 if (tail) {
Al Viro723ad1d2014-03-06 21:13:18 -0500623 p[1] = off + size;
Al Viro706c16f2014-03-06 21:08:24 -0500624 max_contig = max(tail, max_contig);
625 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900626 }
627
Al Viro3d331ad2014-03-06 20:52:32 -0500628 if (!seen_free)
629 chunk->first_free = i + 1;
630
Tejun Heofbf59bc2009-02-20 16:29:08 +0900631 /* update hint and mark allocated */
Al Viro723ad1d2014-03-06 21:13:18 -0500632 if (i + 1 == chunk->map_used)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900633 chunk->contig_hint = max_contig; /* fully scanned */
634 else
635 chunk->contig_hint = max(chunk->contig_hint,
636 max_contig);
637
Al Viro723ad1d2014-03-06 21:13:18 -0500638 chunk->free_size -= size;
639 *p |= 1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900640
Tejun Heob539b872014-09-02 14:46:05 -0400641 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900642 pcpu_chunk_relocate(chunk, oslot);
643 return off;
644 }
645
646 chunk->contig_hint = max_contig; /* fully scanned */
647 pcpu_chunk_relocate(chunk, oslot);
648
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900649 /* tell the upper layer that this chunk has no matching area */
650 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900651}
652
653/**
654 * pcpu_free_area - free area to a pcpu_chunk
655 * @chunk: chunk of interest
656 * @freeme: offset of area to free
Tejun Heob539b872014-09-02 14:46:05 -0400657 * @occ_pages_p: out param for the number of pages the area occupies
Tejun Heofbf59bc2009-02-20 16:29:08 +0900658 *
659 * Free area starting from @freeme to @chunk. Note that this function
660 * only modifies the allocation map. It doesn't depopulate or unmap
661 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900662 *
663 * CONTEXT:
664 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900665 */
Tejun Heob539b872014-09-02 14:46:05 -0400666static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
667 int *occ_pages_p)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900668{
669 int oslot = pcpu_chunk_slot(chunk);
Al Viro723ad1d2014-03-06 21:13:18 -0500670 int off = 0;
671 unsigned i, j;
672 int to_free = 0;
673 int *p;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900674
Al Viro723ad1d2014-03-06 21:13:18 -0500675 freeme |= 1; /* we are searching for <given offset, in use> pair */
676
677 i = 0;
678 j = chunk->map_used;
679 while (i != j) {
680 unsigned k = (i + j) / 2;
681 off = chunk->map[k];
682 if (off < freeme)
683 i = k + 1;
684 else if (off > freeme)
685 j = k;
686 else
687 i = j = k;
688 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900689 BUG_ON(off != freeme);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900690
Al Viro3d331ad2014-03-06 20:52:32 -0500691 if (i < chunk->first_free)
692 chunk->first_free = i;
693
Al Viro723ad1d2014-03-06 21:13:18 -0500694 p = chunk->map + i;
695 *p = off &= ~1;
696 chunk->free_size += (p[1] & ~1) - off;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900697
Tejun Heob539b872014-09-02 14:46:05 -0400698 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
699
Tejun Heofbf59bc2009-02-20 16:29:08 +0900700 /* merge with next? */
Al Viro723ad1d2014-03-06 21:13:18 -0500701 if (!(p[1] & 1))
702 to_free++;
703 /* merge with previous? */
704 if (i > 0 && !(p[-1] & 1)) {
705 to_free++;
706 i--;
707 p--;
708 }
709 if (to_free) {
710 chunk->map_used -= to_free;
711 memmove(p + 1, p + 1 + to_free,
712 (chunk->map_used - i) * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900713 }
714
Al Viro723ad1d2014-03-06 21:13:18 -0500715 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900716 pcpu_chunk_relocate(chunk, oslot);
717}
718
Tejun Heo60810892010-04-09 18:57:01 +0900719static struct pcpu_chunk *pcpu_alloc_chunk(void)
720{
721 struct pcpu_chunk *chunk;
722
Bob Liu90459ce02011-08-04 11:02:33 +0200723 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
Tejun Heo60810892010-04-09 18:57:01 +0900724 if (!chunk)
725 return NULL;
726
Bob Liu90459ce02011-08-04 11:02:33 +0200727 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
728 sizeof(chunk->map[0]));
Tejun Heo60810892010-04-09 18:57:01 +0900729 if (!chunk->map) {
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800730 pcpu_mem_free(chunk);
Tejun Heo60810892010-04-09 18:57:01 +0900731 return NULL;
732 }
733
734 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
Al Viro723ad1d2014-03-06 21:13:18 -0500735 chunk->map[0] = 0;
736 chunk->map[1] = pcpu_unit_size | 1;
737 chunk->map_used = 1;
Tejun Heo60810892010-04-09 18:57:01 +0900738
739 INIT_LIST_HEAD(&chunk->list);
Tejun Heo4f996e22016-05-25 11:48:25 -0400740 INIT_LIST_HEAD(&chunk->map_extend_list);
Tejun Heo60810892010-04-09 18:57:01 +0900741 chunk->free_size = pcpu_unit_size;
742 chunk->contig_hint = pcpu_unit_size;
743
744 return chunk;
745}
746
747static void pcpu_free_chunk(struct pcpu_chunk *chunk)
748{
749 if (!chunk)
750 return;
Tetsuo Handa1d5cfdb2016-01-22 15:11:02 -0800751 pcpu_mem_free(chunk->map);
752 pcpu_mem_free(chunk);
Tejun Heo60810892010-04-09 18:57:01 +0900753}
754
Tejun Heob539b872014-09-02 14:46:05 -0400755/**
756 * pcpu_chunk_populated - post-population bookkeeping
757 * @chunk: pcpu_chunk which got populated
758 * @page_start: the start page
759 * @page_end: the end page
760 *
761 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
762 * the bookkeeping information accordingly. Must be called after each
763 * successful population.
764 */
765static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
766 int page_start, int page_end)
767{
768 int nr = page_end - page_start;
769
770 lockdep_assert_held(&pcpu_lock);
771
772 bitmap_set(chunk->populated, page_start, nr);
773 chunk->nr_populated += nr;
774 pcpu_nr_empty_pop_pages += nr;
775}
776
777/**
778 * pcpu_chunk_depopulated - post-depopulation bookkeeping
779 * @chunk: pcpu_chunk which got depopulated
780 * @page_start: the start page
781 * @page_end: the end page
782 *
783 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
784 * Update the bookkeeping information accordingly. Must be called after
785 * each successful depopulation.
786 */
787static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
788 int page_start, int page_end)
789{
790 int nr = page_end - page_start;
791
792 lockdep_assert_held(&pcpu_lock);
793
794 bitmap_clear(chunk->populated, page_start, nr);
795 chunk->nr_populated -= nr;
796 pcpu_nr_empty_pop_pages -= nr;
797}
798
Tejun Heo9f645532010-04-09 18:57:01 +0900799/*
800 * Chunk management implementation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900801 *
Tejun Heo9f645532010-04-09 18:57:01 +0900802 * To allow different implementations, chunk alloc/free and
803 * [de]population are implemented in a separate file which is pulled
804 * into this file and compiled together. The following functions
805 * should be implemented.
Tejun Heoce3141a2009-07-04 08:11:00 +0900806 *
Tejun Heo9f645532010-04-09 18:57:01 +0900807 * pcpu_populate_chunk - populate the specified range of a chunk
808 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
809 * pcpu_create_chunk - create a new chunk
810 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
811 * pcpu_addr_to_page - translate address to physical address
812 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
Tejun Heofbf59bc2009-02-20 16:29:08 +0900813 */
Tejun Heo9f645532010-04-09 18:57:01 +0900814static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
815static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
816static struct pcpu_chunk *pcpu_create_chunk(void);
817static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
818static struct page *pcpu_addr_to_page(void *addr);
819static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
Tejun Heoce3141a2009-07-04 08:11:00 +0900820
Tejun Heob0c97782010-04-09 18:57:01 +0900821#ifdef CONFIG_NEED_PER_CPU_KM
822#include "percpu-km.c"
823#else
Tejun Heo9f645532010-04-09 18:57:01 +0900824#include "percpu-vm.c"
Tejun Heob0c97782010-04-09 18:57:01 +0900825#endif
Tejun Heofbf59bc2009-02-20 16:29:08 +0900826
827/**
Tejun Heo88999a82010-04-09 18:57:01 +0900828 * pcpu_chunk_addr_search - determine chunk containing specified address
829 * @addr: address for which the chunk needs to be determined.
830 *
831 * RETURNS:
832 * The address of the found chunk.
833 */
834static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
835{
836 /* is it in the first chunk? */
837 if (pcpu_addr_in_first_chunk(addr)) {
838 /* is it in the reserved area? */
839 if (pcpu_addr_in_reserved_chunk(addr))
840 return pcpu_reserved_chunk;
841 return pcpu_first_chunk;
842 }
843
844 /*
845 * The address is relative to unit0 which might be unused and
846 * thus unmapped. Offset the address to the unit space of the
847 * current processor before looking it up in the vmalloc
848 * space. Note that any possible cpu id can be used here, so
849 * there's no need to worry about preemption or cpu hotplug.
850 */
851 addr += pcpu_unit_offsets[raw_smp_processor_id()];
Tejun Heo9f645532010-04-09 18:57:01 +0900852 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
Tejun Heo88999a82010-04-09 18:57:01 +0900853}
854
855/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900856 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900857 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900858 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900859 * @reserved: allocate from the reserved chunk if available
Tejun Heo5835d962014-09-02 14:46:04 -0400860 * @gfp: allocation flags
Tejun Heofbf59bc2009-02-20 16:29:08 +0900861 *
Tejun Heo5835d962014-09-02 14:46:04 -0400862 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
863 * contain %GFP_KERNEL, the allocation is atomic.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900864 *
865 * RETURNS:
866 * Percpu pointer to the allocated area on success, NULL on failure.
867 */
Tejun Heo5835d962014-09-02 14:46:04 -0400868static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
869 gfp_t gfp)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900870{
Tejun Heof2badb02009-09-29 09:17:58 +0900871 static int warn_limit = 10;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900872 struct pcpu_chunk *chunk;
Tejun Heof2badb02009-09-29 09:17:58 +0900873 const char *err;
Tejun Heo6ae833c2014-10-08 12:01:52 -0400874 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
Tejun Heob539b872014-09-02 14:46:05 -0400875 int occ_pages = 0;
Tejun Heob38d08f2014-09-02 14:46:02 -0400876 int slot, off, new_alloc, cpu, ret;
Jiri Kosina403a91b2009-10-29 00:25:59 +0900877 unsigned long flags;
Catalin Marinasf528f0b2011-09-26 17:12:53 +0100878 void __percpu *ptr;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900879
Al Viro723ad1d2014-03-06 21:13:18 -0500880 /*
881 * We want the lowest bit of offset available for in-use/free
Viro2f69fa82014-03-17 16:01:27 -0400882 * indicator, so force >= 16bit alignment and make size even.
Al Viro723ad1d2014-03-06 21:13:18 -0500883 */
884 if (unlikely(align < 2))
885 align = 2;
886
Christoph Lameterfb009e32014-06-19 09:59:18 -0500887 size = ALIGN(size, 2);
Viro2f69fa82014-03-17 16:01:27 -0400888
Tejun Heo8d408b42009-02-24 11:57:21 +0900889 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Joe Perches756a025f02016-03-17 14:19:47 -0700890 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
891 size, align);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900892 return NULL;
893 }
894
Tejun Heo6710e592016-05-25 11:48:25 -0400895 if (!is_atomic)
896 mutex_lock(&pcpu_alloc_mutex);
897
Jiri Kosina403a91b2009-10-29 00:25:59 +0900898 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900899
Tejun Heoedcb4632009-03-06 14:33:59 +0900900 /* serve reserved allocations from the reserved chunk if available */
901 if (reserved && pcpu_reserved_chunk) {
902 chunk = pcpu_reserved_chunk;
Tejun Heo833af842009-11-11 15:35:18 +0900903
904 if (size > chunk->contig_hint) {
905 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900906 goto fail_unlock;
Tejun Heof2badb02009-09-29 09:17:58 +0900907 }
Tejun Heo833af842009-11-11 15:35:18 +0900908
Tejun Heo9c824b62014-09-02 14:46:05 -0400909 while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
Tejun Heo833af842009-11-11 15:35:18 +0900910 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heo5835d962014-09-02 14:46:04 -0400911 if (is_atomic ||
912 pcpu_extend_area_map(chunk, new_alloc) < 0) {
Tejun Heo833af842009-11-11 15:35:18 +0900913 err = "failed to extend area map of reserved chunk";
Tejun Heob38d08f2014-09-02 14:46:02 -0400914 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900915 }
916 spin_lock_irqsave(&pcpu_lock, flags);
917 }
918
Tejun Heob539b872014-09-02 14:46:05 -0400919 off = pcpu_alloc_area(chunk, size, align, is_atomic,
920 &occ_pages);
Tejun Heoedcb4632009-03-06 14:33:59 +0900921 if (off >= 0)
922 goto area_found;
Tejun Heo833af842009-11-11 15:35:18 +0900923
Tejun Heof2badb02009-09-29 09:17:58 +0900924 err = "alloc from reserved chunk failed";
Tejun Heoccea34b2009-03-07 00:44:13 +0900925 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900926 }
927
Tejun Heoccea34b2009-03-07 00:44:13 +0900928restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900929 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900930 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
931 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
932 if (size > chunk->contig_hint)
933 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900934
Tejun Heo9c824b62014-09-02 14:46:05 -0400935 new_alloc = pcpu_need_to_extend(chunk, is_atomic);
Tejun Heo833af842009-11-11 15:35:18 +0900936 if (new_alloc) {
Tejun Heo5835d962014-09-02 14:46:04 -0400937 if (is_atomic)
938 continue;
Tejun Heo833af842009-11-11 15:35:18 +0900939 spin_unlock_irqrestore(&pcpu_lock, flags);
940 if (pcpu_extend_area_map(chunk,
941 new_alloc) < 0) {
942 err = "failed to extend area map";
Tejun Heob38d08f2014-09-02 14:46:02 -0400943 goto fail;
Tejun Heo833af842009-11-11 15:35:18 +0900944 }
945 spin_lock_irqsave(&pcpu_lock, flags);
946 /*
947 * pcpu_lock has been dropped, need to
948 * restart cpu_slot list walking.
949 */
950 goto restart;
Tejun Heoccea34b2009-03-07 00:44:13 +0900951 }
952
Tejun Heob539b872014-09-02 14:46:05 -0400953 off = pcpu_alloc_area(chunk, size, align, is_atomic,
954 &occ_pages);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900955 if (off >= 0)
956 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900957 }
958 }
959
Jiri Kosina403a91b2009-10-29 00:25:59 +0900960 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900961
Tejun Heob38d08f2014-09-02 14:46:02 -0400962 /*
963 * No space left. Create a new chunk. We don't want multiple
964 * tasks to create chunks simultaneously. Serialize and create iff
965 * there's still no empty chunk after grabbing the mutex.
966 */
Tejun Heo5835d962014-09-02 14:46:04 -0400967 if (is_atomic)
968 goto fail;
969
Tejun Heob38d08f2014-09-02 14:46:02 -0400970 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
971 chunk = pcpu_create_chunk();
972 if (!chunk) {
973 err = "failed to allocate new chunk";
974 goto fail;
975 }
976
977 spin_lock_irqsave(&pcpu_lock, flags);
978 pcpu_chunk_relocate(chunk, -1);
979 } else {
980 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heof2badb02009-09-29 09:17:58 +0900981 }
Tejun Heoccea34b2009-03-07 00:44:13 +0900982
Tejun Heoccea34b2009-03-07 00:44:13 +0900983 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900984
985area_found:
Jiri Kosina403a91b2009-10-29 00:25:59 +0900986 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heoccea34b2009-03-07 00:44:13 +0900987
Tejun Heodca49642014-09-02 14:46:01 -0400988 /* populate if not all pages are already there */
Tejun Heo5835d962014-09-02 14:46:04 -0400989 if (!is_atomic) {
Tejun Heoe04d3202014-09-02 14:46:04 -0400990 int page_start, page_end, rs, re;
Tejun Heodca49642014-09-02 14:46:01 -0400991
Tejun Heoe04d3202014-09-02 14:46:04 -0400992 page_start = PFN_DOWN(off);
993 page_end = PFN_UP(off + size);
Tejun Heob38d08f2014-09-02 14:46:02 -0400994
Tejun Heoe04d3202014-09-02 14:46:04 -0400995 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
996 WARN_ON(chunk->immutable);
997
998 ret = pcpu_populate_chunk(chunk, rs, re);
999
1000 spin_lock_irqsave(&pcpu_lock, flags);
1001 if (ret) {
Tejun Heob539b872014-09-02 14:46:05 -04001002 pcpu_free_area(chunk, off, &occ_pages);
Tejun Heoe04d3202014-09-02 14:46:04 -04001003 err = "failed to populate";
1004 goto fail_unlock;
1005 }
Tejun Heob539b872014-09-02 14:46:05 -04001006 pcpu_chunk_populated(chunk, rs, re);
Tejun Heoe04d3202014-09-02 14:46:04 -04001007 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heodca49642014-09-02 14:46:01 -04001008 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001009
Tejun Heoe04d3202014-09-02 14:46:04 -04001010 mutex_unlock(&pcpu_alloc_mutex);
1011 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001012
Tejun Heob539b872014-09-02 14:46:05 -04001013 if (chunk != pcpu_reserved_chunk)
1014 pcpu_nr_empty_pop_pages -= occ_pages;
1015
Tejun Heo1a4d7602014-09-02 14:46:05 -04001016 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1017 pcpu_schedule_balance_work();
1018
Tejun Heodca49642014-09-02 14:46:01 -04001019 /* clear the areas and return address relative to base address */
1020 for_each_possible_cpu(cpu)
1021 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1022
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001023 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
Larry Finger8a8c35f2015-06-24 16:58:51 -07001024 kmemleak_alloc_percpu(ptr, size, gfp);
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001025 return ptr;
Tejun Heoccea34b2009-03-07 00:44:13 +09001026
1027fail_unlock:
Jiri Kosina403a91b2009-10-29 00:25:59 +09001028 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heob38d08f2014-09-02 14:46:02 -04001029fail:
Tejun Heo5835d962014-09-02 14:46:04 -04001030 if (!is_atomic && warn_limit) {
Joe Perches870d4b12016-03-17 14:19:53 -07001031 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
Joe Perches598d8092016-03-17 14:19:44 -07001032 size, align, is_atomic, err);
Tejun Heof2badb02009-09-29 09:17:58 +09001033 dump_stack();
1034 if (!--warn_limit)
Joe Perches870d4b12016-03-17 14:19:53 -07001035 pr_info("limit reached, disable warning\n");
Tejun Heof2badb02009-09-29 09:17:58 +09001036 }
Tejun Heo1a4d7602014-09-02 14:46:05 -04001037 if (is_atomic) {
1038 /* see the flag handling in pcpu_blance_workfn() */
1039 pcpu_atomic_alloc_failed = true;
1040 pcpu_schedule_balance_work();
Tejun Heo6710e592016-05-25 11:48:25 -04001041 } else {
1042 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heo1a4d7602014-09-02 14:46:05 -04001043 }
Tejun Heoccea34b2009-03-07 00:44:13 +09001044 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001045}
Tejun Heoedcb4632009-03-06 14:33:59 +09001046
1047/**
Tejun Heo5835d962014-09-02 14:46:04 -04001048 * __alloc_percpu_gfp - allocate dynamic percpu area
Tejun Heoedcb4632009-03-06 14:33:59 +09001049 * @size: size of area to allocate in bytes
1050 * @align: alignment of area (max PAGE_SIZE)
Tejun Heo5835d962014-09-02 14:46:04 -04001051 * @gfp: allocation flags
Tejun Heoedcb4632009-03-06 14:33:59 +09001052 *
Tejun Heo5835d962014-09-02 14:46:04 -04001053 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1054 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1055 * be called from any context but is a lot more likely to fail.
Tejun Heoccea34b2009-03-07 00:44:13 +09001056 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001057 * RETURNS:
1058 * Percpu pointer to the allocated area on success, NULL on failure.
1059 */
Tejun Heo5835d962014-09-02 14:46:04 -04001060void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1061{
1062 return pcpu_alloc(size, align, false, gfp);
1063}
1064EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1065
1066/**
1067 * __alloc_percpu - allocate dynamic percpu area
1068 * @size: size of area to allocate in bytes
1069 * @align: alignment of area (max PAGE_SIZE)
1070 *
1071 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1072 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001073void __percpu *__alloc_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001074{
Tejun Heo5835d962014-09-02 14:46:04 -04001075 return pcpu_alloc(size, align, false, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001076}
Tejun Heofbf59bc2009-02-20 16:29:08 +09001077EXPORT_SYMBOL_GPL(__alloc_percpu);
1078
Tejun Heoedcb4632009-03-06 14:33:59 +09001079/**
1080 * __alloc_reserved_percpu - allocate reserved percpu area
1081 * @size: size of area to allocate in bytes
1082 * @align: alignment of area (max PAGE_SIZE)
1083 *
Tejun Heo9329ba92010-09-10 11:01:56 +02001084 * Allocate zero-filled percpu area of @size bytes aligned at @align
1085 * from reserved percpu area if arch has set it up; otherwise,
1086 * allocation is served from the same dynamic area. Might sleep.
1087 * Might trigger writeouts.
Tejun Heoedcb4632009-03-06 14:33:59 +09001088 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001089 * CONTEXT:
1090 * Does GFP_KERNEL allocation.
1091 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001092 * RETURNS:
1093 * Percpu pointer to the allocated area on success, NULL on failure.
1094 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001095void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
Tejun Heoedcb4632009-03-06 14:33:59 +09001096{
Tejun Heo5835d962014-09-02 14:46:04 -04001097 return pcpu_alloc(size, align, true, GFP_KERNEL);
Tejun Heoedcb4632009-03-06 14:33:59 +09001098}
1099
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001100/**
Tejun Heo1a4d7602014-09-02 14:46:05 -04001101 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001102 * @work: unused
1103 *
1104 * Reclaim all fully free chunks except for the first one.
1105 */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001106static void pcpu_balance_workfn(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001107{
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001108 LIST_HEAD(to_free);
1109 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001110 struct pcpu_chunk *chunk, *next;
Tejun Heo1a4d7602014-09-02 14:46:05 -04001111 int slot, nr_to_pop, ret;
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001112
Tejun Heo1a4d7602014-09-02 14:46:05 -04001113 /*
1114 * There's no reason to keep around multiple unused chunks and VM
1115 * areas can be scarce. Destroy all free chunks except for one.
1116 */
Tejun Heoccea34b2009-03-07 00:44:13 +09001117 mutex_lock(&pcpu_alloc_mutex);
1118 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001119
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001120 list_for_each_entry_safe(chunk, next, free_head, list) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001121 WARN_ON(chunk->immutable);
1122
1123 /* spare the first one */
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001124 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001125 continue;
1126
Tejun Heo4f996e22016-05-25 11:48:25 -04001127 list_del_init(&chunk->map_extend_list);
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001128 list_move(&chunk->list, &to_free);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001129 }
1130
Tejun Heoccea34b2009-03-07 00:44:13 +09001131 spin_unlock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001132
Tejun Heofe6bd8c2014-09-02 14:46:05 -04001133 list_for_each_entry_safe(chunk, next, &to_free, list) {
Tejun Heoa93ace42014-09-02 14:46:02 -04001134 int rs, re;
Tejun Heodca49642014-09-02 14:46:01 -04001135
Tejun Heoa93ace42014-09-02 14:46:02 -04001136 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1137 pcpu_depopulate_chunk(chunk, rs, re);
Tejun Heob539b872014-09-02 14:46:05 -04001138 spin_lock_irq(&pcpu_lock);
1139 pcpu_chunk_depopulated(chunk, rs, re);
1140 spin_unlock_irq(&pcpu_lock);
Tejun Heoa93ace42014-09-02 14:46:02 -04001141 }
Tejun Heo60810892010-04-09 18:57:01 +09001142 pcpu_destroy_chunk(chunk);
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001143 }
Tejun Heo971f3912009-08-14 15:00:49 +09001144
Tejun Heo4f996e22016-05-25 11:48:25 -04001145 /* service chunks which requested async area map extension */
1146 do {
1147 int new_alloc = 0;
1148
1149 spin_lock_irq(&pcpu_lock);
1150
1151 chunk = list_first_entry_or_null(&pcpu_map_extend_chunks,
1152 struct pcpu_chunk, map_extend_list);
1153 if (chunk) {
1154 list_del_init(&chunk->map_extend_list);
1155 new_alloc = pcpu_need_to_extend(chunk, false);
1156 }
1157
1158 spin_unlock_irq(&pcpu_lock);
1159
1160 if (new_alloc)
1161 pcpu_extend_area_map(chunk, new_alloc);
1162 } while (chunk);
1163
Tejun Heo1a4d7602014-09-02 14:46:05 -04001164 /*
1165 * Ensure there are certain number of free populated pages for
1166 * atomic allocs. Fill up from the most packed so that atomic
1167 * allocs don't increase fragmentation. If atomic allocation
1168 * failed previously, always populate the maximum amount. This
1169 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1170 * failing indefinitely; however, large atomic allocs are not
1171 * something we support properly and can be highly unreliable and
1172 * inefficient.
1173 */
1174retry_pop:
1175 if (pcpu_atomic_alloc_failed) {
1176 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1177 /* best effort anyway, don't worry about synchronization */
1178 pcpu_atomic_alloc_failed = false;
1179 } else {
1180 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1181 pcpu_nr_empty_pop_pages,
1182 0, PCPU_EMPTY_POP_PAGES_HIGH);
1183 }
1184
1185 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1186 int nr_unpop = 0, rs, re;
1187
1188 if (!nr_to_pop)
1189 break;
1190
1191 spin_lock_irq(&pcpu_lock);
1192 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1193 nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1194 if (nr_unpop)
1195 break;
1196 }
1197 spin_unlock_irq(&pcpu_lock);
1198
1199 if (!nr_unpop)
1200 continue;
1201
1202 /* @chunk can't go away while pcpu_alloc_mutex is held */
1203 pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1204 int nr = min(re - rs, nr_to_pop);
1205
1206 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1207 if (!ret) {
1208 nr_to_pop -= nr;
1209 spin_lock_irq(&pcpu_lock);
1210 pcpu_chunk_populated(chunk, rs, rs + nr);
1211 spin_unlock_irq(&pcpu_lock);
1212 } else {
1213 nr_to_pop = 0;
1214 }
1215
1216 if (!nr_to_pop)
1217 break;
1218 }
1219 }
1220
1221 if (nr_to_pop) {
1222 /* ran out of chunks to populate, create a new one and retry */
1223 chunk = pcpu_create_chunk();
1224 if (chunk) {
1225 spin_lock_irq(&pcpu_lock);
1226 pcpu_chunk_relocate(chunk, -1);
1227 spin_unlock_irq(&pcpu_lock);
1228 goto retry_pop;
1229 }
1230 }
1231
Tejun Heo971f3912009-08-14 15:00:49 +09001232 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001233}
1234
1235/**
1236 * free_percpu - free percpu area
1237 * @ptr: pointer to area to free
1238 *
Tejun Heoccea34b2009-03-07 00:44:13 +09001239 * Free percpu area @ptr.
1240 *
1241 * CONTEXT:
1242 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001243 */
Tejun Heo43cf38e2010-02-02 14:38:57 +09001244void free_percpu(void __percpu *ptr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001245{
Andrew Morton129182e2010-01-08 14:42:39 -08001246 void *addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001247 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +09001248 unsigned long flags;
Tejun Heob539b872014-09-02 14:46:05 -04001249 int off, occ_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001250
1251 if (!ptr)
1252 return;
1253
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001254 kmemleak_free_percpu(ptr);
1255
Andrew Morton129182e2010-01-08 14:42:39 -08001256 addr = __pcpu_ptr_to_addr(ptr);
1257
Tejun Heoccea34b2009-03-07 00:44:13 +09001258 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001259
1260 chunk = pcpu_chunk_addr_search(addr);
Tejun Heobba174f2009-08-14 15:00:51 +09001261 off = addr - chunk->base_addr;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001262
Tejun Heob539b872014-09-02 14:46:05 -04001263 pcpu_free_area(chunk, off, &occ_pages);
1264
1265 if (chunk != pcpu_reserved_chunk)
1266 pcpu_nr_empty_pop_pages += occ_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001267
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001268 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001269 if (chunk->free_size == pcpu_unit_size) {
1270 struct pcpu_chunk *pos;
1271
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001272 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001273 if (pos != chunk) {
Tejun Heo1a4d7602014-09-02 14:46:05 -04001274 pcpu_schedule_balance_work();
Tejun Heofbf59bc2009-02-20 16:29:08 +09001275 break;
1276 }
1277 }
1278
Tejun Heoccea34b2009-03-07 00:44:13 +09001279 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001280}
1281EXPORT_SYMBOL_GPL(free_percpu);
1282
Vivek Goyal3b034b02009-11-24 15:50:03 +09001283/**
Tejun Heo10fad5e2010-03-10 18:57:54 +09001284 * is_kernel_percpu_address - test whether address is from static percpu area
1285 * @addr: address to test
1286 *
1287 * Test whether @addr belongs to in-kernel static percpu area. Module
1288 * static percpu areas are not considered. For those, use
1289 * is_module_percpu_address().
1290 *
1291 * RETURNS:
1292 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1293 */
1294bool is_kernel_percpu_address(unsigned long addr)
1295{
Tejun Heobbddff02010-09-03 18:22:48 +02001296#ifdef CONFIG_SMP
Tejun Heo10fad5e2010-03-10 18:57:54 +09001297 const size_t static_size = __per_cpu_end - __per_cpu_start;
1298 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1299 unsigned int cpu;
1300
1301 for_each_possible_cpu(cpu) {
1302 void *start = per_cpu_ptr(base, cpu);
1303
1304 if ((void *)addr >= start && (void *)addr < start + static_size)
1305 return true;
1306 }
Tejun Heobbddff02010-09-03 18:22:48 +02001307#endif
1308 /* on UP, can't distinguish from other static vars, always false */
Tejun Heo10fad5e2010-03-10 18:57:54 +09001309 return false;
1310}
1311
1312/**
Vivek Goyal3b034b02009-11-24 15:50:03 +09001313 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1314 * @addr: the address to be converted to physical address
1315 *
1316 * Given @addr which is dereferenceable address obtained via one of
1317 * percpu access macros, this function translates it into its physical
1318 * address. The caller is responsible for ensuring @addr stays valid
1319 * until this function finishes.
1320 *
Dave Young67589c72011-11-23 08:20:53 -08001321 * percpu allocator has special setup for the first chunk, which currently
1322 * supports either embedding in linear address space or vmalloc mapping,
1323 * and, from the second one, the backing allocator (currently either vm or
1324 * km) provides translation.
1325 *
Yannick Guerrinibffc4372015-03-06 23:30:42 +01001326 * The addr can be translated simply without checking if it falls into the
Dave Young67589c72011-11-23 08:20:53 -08001327 * first chunk. But the current code reflects better how percpu allocator
1328 * actually works, and the verification can discover both bugs in percpu
1329 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1330 * code.
1331 *
Vivek Goyal3b034b02009-11-24 15:50:03 +09001332 * RETURNS:
1333 * The physical address for @addr.
1334 */
1335phys_addr_t per_cpu_ptr_to_phys(void *addr)
1336{
Tejun Heo9983b6f02010-06-18 11:44:31 +02001337 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1338 bool in_first_chunk = false;
Tejun Heoa855b842011-11-18 10:55:35 -08001339 unsigned long first_low, first_high;
Tejun Heo9983b6f02010-06-18 11:44:31 +02001340 unsigned int cpu;
1341
1342 /*
Tejun Heoa855b842011-11-18 10:55:35 -08001343 * The following test on unit_low/high isn't strictly
Tejun Heo9983b6f02010-06-18 11:44:31 +02001344 * necessary but will speed up lookups of addresses which
1345 * aren't in the first chunk.
1346 */
Tejun Heoa855b842011-11-18 10:55:35 -08001347 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1348 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1349 pcpu_unit_pages);
1350 if ((unsigned long)addr >= first_low &&
1351 (unsigned long)addr < first_high) {
Tejun Heo9983b6f02010-06-18 11:44:31 +02001352 for_each_possible_cpu(cpu) {
1353 void *start = per_cpu_ptr(base, cpu);
1354
1355 if (addr >= start && addr < start + pcpu_unit_size) {
1356 in_first_chunk = true;
1357 break;
1358 }
1359 }
1360 }
1361
1362 if (in_first_chunk) {
David Howellseac522e2011-03-28 12:53:29 +01001363 if (!is_vmalloc_addr(addr))
Tejun Heo020ec652010-04-09 18:57:00 +09001364 return __pa(addr);
1365 else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001366 return page_to_phys(vmalloc_to_page(addr)) +
1367 offset_in_page(addr);
Tejun Heo020ec652010-04-09 18:57:00 +09001368 } else
Eugene Surovegin9f57bd42011-12-15 11:25:59 -08001369 return page_to_phys(pcpu_addr_to_page(addr)) +
1370 offset_in_page(addr);
Vivek Goyal3b034b02009-11-24 15:50:03 +09001371}
1372
Tejun Heofbf59bc2009-02-20 16:29:08 +09001373/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001374 * pcpu_alloc_alloc_info - allocate percpu allocation info
1375 * @nr_groups: the number of groups
1376 * @nr_units: the number of units
Tejun Heo033e48f2009-08-14 15:00:51 +09001377 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001378 * Allocate ai which is large enough for @nr_groups groups containing
1379 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1380 * cpu_map array which is long enough for @nr_units and filled with
1381 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1382 * pointer of other groups.
Tejun Heo033e48f2009-08-14 15:00:51 +09001383 *
1384 * RETURNS:
Tejun Heofd1e8a12009-08-14 15:00:51 +09001385 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1386 * failure.
Tejun Heo033e48f2009-08-14 15:00:51 +09001387 */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001388struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1389 int nr_units)
1390{
1391 struct pcpu_alloc_info *ai;
1392 size_t base_size, ai_size;
1393 void *ptr;
1394 int unit;
1395
1396 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1397 __alignof__(ai->groups[0].cpu_map[0]));
1398 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1399
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001400 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001401 if (!ptr)
1402 return NULL;
1403 ai = ptr;
1404 ptr += base_size;
1405
1406 ai->groups[0].cpu_map = ptr;
1407
1408 for (unit = 0; unit < nr_units; unit++)
1409 ai->groups[0].cpu_map[unit] = NR_CPUS;
1410
1411 ai->nr_groups = nr_groups;
1412 ai->__ai_size = PFN_ALIGN(ai_size);
1413
1414 return ai;
1415}
1416
1417/**
1418 * pcpu_free_alloc_info - free percpu allocation info
1419 * @ai: pcpu_alloc_info to free
1420 *
1421 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1422 */
1423void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1424{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001425 memblock_free_early(__pa(ai), ai->__ai_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001426}
1427
1428/**
Tejun Heofd1e8a12009-08-14 15:00:51 +09001429 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1430 * @lvl: loglevel
1431 * @ai: allocation info to dump
1432 *
1433 * Print out information about @ai using loglevel @lvl.
1434 */
1435static void pcpu_dump_alloc_info(const char *lvl,
1436 const struct pcpu_alloc_info *ai)
Tejun Heo033e48f2009-08-14 15:00:51 +09001437{
Tejun Heofd1e8a12009-08-14 15:00:51 +09001438 int group_width = 1, cpu_width = 1, width;
Tejun Heo033e48f2009-08-14 15:00:51 +09001439 char empty_str[] = "--------";
Tejun Heofd1e8a12009-08-14 15:00:51 +09001440 int alloc = 0, alloc_end = 0;
1441 int group, v;
1442 int upa, apl; /* units per alloc, allocs per line */
Tejun Heo033e48f2009-08-14 15:00:51 +09001443
Tejun Heofd1e8a12009-08-14 15:00:51 +09001444 v = ai->nr_groups;
Tejun Heo033e48f2009-08-14 15:00:51 +09001445 while (v /= 10)
Tejun Heofd1e8a12009-08-14 15:00:51 +09001446 group_width++;
Tejun Heo033e48f2009-08-14 15:00:51 +09001447
Tejun Heofd1e8a12009-08-14 15:00:51 +09001448 v = num_possible_cpus();
1449 while (v /= 10)
1450 cpu_width++;
1451 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
Tejun Heo033e48f2009-08-14 15:00:51 +09001452
Tejun Heofd1e8a12009-08-14 15:00:51 +09001453 upa = ai->alloc_size / ai->unit_size;
1454 width = upa * (cpu_width + 1) + group_width + 3;
1455 apl = rounddown_pow_of_two(max(60 / width, 1));
Tejun Heo033e48f2009-08-14 15:00:51 +09001456
Tejun Heofd1e8a12009-08-14 15:00:51 +09001457 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1458 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1459 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1460
1461 for (group = 0; group < ai->nr_groups; group++) {
1462 const struct pcpu_group_info *gi = &ai->groups[group];
1463 int unit = 0, unit_end = 0;
1464
1465 BUG_ON(gi->nr_units % upa);
1466 for (alloc_end += gi->nr_units / upa;
1467 alloc < alloc_end; alloc++) {
1468 if (!(alloc % apl)) {
Joe Perches11705322016-03-17 14:19:50 -07001469 pr_cont("\n");
Tejun Heofd1e8a12009-08-14 15:00:51 +09001470 printk("%spcpu-alloc: ", lvl);
1471 }
Joe Perches11705322016-03-17 14:19:50 -07001472 pr_cont("[%0*d] ", group_width, group);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001473
1474 for (unit_end += upa; unit < unit_end; unit++)
1475 if (gi->cpu_map[unit] != NR_CPUS)
Joe Perches11705322016-03-17 14:19:50 -07001476 pr_cont("%0*d ",
1477 cpu_width, gi->cpu_map[unit]);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001478 else
Joe Perches11705322016-03-17 14:19:50 -07001479 pr_cont("%s ", empty_str);
Tejun Heo033e48f2009-08-14 15:00:51 +09001480 }
Tejun Heo033e48f2009-08-14 15:00:51 +09001481 }
Joe Perches11705322016-03-17 14:19:50 -07001482 pr_cont("\n");
Tejun Heo033e48f2009-08-14 15:00:51 +09001483}
Tejun Heo033e48f2009-08-14 15:00:51 +09001484
Tejun Heofbf59bc2009-02-20 16:29:08 +09001485/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001486 * pcpu_setup_first_chunk - initialize the first percpu chunk
Tejun Heofd1e8a12009-08-14 15:00:51 +09001487 * @ai: pcpu_alloc_info describing how to percpu area is shaped
Tejun Heo38a6be52009-07-04 08:10:59 +09001488 * @base_addr: mapped address
Tejun Heofbf59bc2009-02-20 16:29:08 +09001489 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001490 * Initialize the first percpu chunk which contains the kernel static
1491 * perpcu area. This function is to be called from arch percpu area
Tejun Heo38a6be52009-07-04 08:10:59 +09001492 * setup path.
Tejun Heo8d408b42009-02-24 11:57:21 +09001493 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001494 * @ai contains all information necessary to initialize the first
1495 * chunk and prime the dynamic percpu allocator.
Tejun Heo8d408b42009-02-24 11:57:21 +09001496 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001497 * @ai->static_size is the size of static percpu area.
1498 *
1499 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
Tejun Heoedcb4632009-03-06 14:33:59 +09001500 * reserve after the static area in the first chunk. This reserves
1501 * the first chunk such that it's available only through reserved
1502 * percpu allocation. This is primarily used to serve module percpu
1503 * static areas on architectures where the addressing model has
1504 * limited offset range for symbol relocations to guarantee module
1505 * percpu symbols fall inside the relocatable range.
1506 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001507 * @ai->dyn_size determines the number of bytes available for dynamic
1508 * allocation in the first chunk. The area between @ai->static_size +
1509 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
Tejun Heo6074d5b2009-03-10 16:27:48 +09001510 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001511 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1512 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1513 * @ai->dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001514 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001515 * @ai->atom_size is the allocation atom size and used as alignment
1516 * for vm areas.
Tejun Heo8d408b42009-02-24 11:57:21 +09001517 *
Tejun Heofd1e8a12009-08-14 15:00:51 +09001518 * @ai->alloc_size is the allocation size and always multiple of
1519 * @ai->atom_size. This is larger than @ai->atom_size if
1520 * @ai->unit_size is larger than @ai->atom_size.
1521 *
1522 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1523 * percpu areas. Units which should be colocated are put into the
1524 * same group. Dynamic VM areas will be allocated according to these
1525 * groupings. If @ai->nr_groups is zero, a single group containing
1526 * all units is assumed.
Tejun Heo8d408b42009-02-24 11:57:21 +09001527 *
Tejun Heo38a6be52009-07-04 08:10:59 +09001528 * The caller should have mapped the first chunk at @base_addr and
1529 * copied static data to each unit.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001530 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001531 * If the first chunk ends up with both reserved and dynamic areas, it
1532 * is served by two chunks - one to serve the core static and reserved
1533 * areas and the other for the dynamic area. They share the same vm
1534 * and page map but uses different area allocation map to stay away
1535 * from each other. The latter chunk is circulated in the chunk slots
1536 * and available for dynamic allocation like any other chunks.
1537 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001538 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001539 * 0 on success, -errno on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001540 */
Tejun Heofb435d52009-08-14 15:00:51 +09001541int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1542 void *base_addr)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001543{
Tejun Heo099a19d2010-06-27 18:50:00 +02001544 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1545 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001546 size_t dyn_size = ai->dyn_size;
1547 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001548 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heo65632972009-08-14 15:00:52 +09001549 unsigned long *group_offsets;
1550 size_t *group_sizes;
Tejun Heofb435d52009-08-14 15:00:51 +09001551 unsigned long *unit_off;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001552 unsigned int cpu;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001553 int *unit_map;
1554 int group, unit, i;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001555
Tejun Heo635b75f2009-09-24 09:43:11 +09001556#define PCPU_SETUP_BUG_ON(cond) do { \
1557 if (unlikely(cond)) { \
Joe Perches870d4b12016-03-17 14:19:53 -07001558 pr_emerg("failed to initialize, %s\n", #cond); \
1559 pr_emerg("cpu_possible_mask=%*pb\n", \
Tejun Heo807de072015-02-13 14:37:34 -08001560 cpumask_pr_args(cpu_possible_mask)); \
Tejun Heo635b75f2009-09-24 09:43:11 +09001561 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1562 BUG(); \
1563 } \
1564} while (0)
1565
Tejun Heo2f39e632009-07-04 08:11:00 +09001566 /* sanity checks */
Tejun Heo635b75f2009-09-24 09:43:11 +09001567 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
Tejun Heobbddff02010-09-03 18:22:48 +02001568#ifdef CONFIG_SMP
Tejun Heo635b75f2009-09-24 09:43:11 +09001569 PCPU_SETUP_BUG_ON(!ai->static_size);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08001570 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
Tejun Heobbddff02010-09-03 18:22:48 +02001571#endif
Tejun Heo635b75f2009-09-24 09:43:11 +09001572 PCPU_SETUP_BUG_ON(!base_addr);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08001573 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
Tejun Heo635b75f2009-09-24 09:43:11 +09001574 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08001575 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
Tejun Heo635b75f2009-09-24 09:43:11 +09001576 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
Tejun Heo099a19d2010-06-27 18:50:00 +02001577 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
Tejun Heo9f645532010-04-09 18:57:01 +09001578 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
Tejun Heo8d408b42009-02-24 11:57:21 +09001579
Tejun Heo65632972009-08-14 15:00:52 +09001580 /* process group information and build config tables accordingly */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001581 group_offsets = memblock_virt_alloc(ai->nr_groups *
1582 sizeof(group_offsets[0]), 0);
1583 group_sizes = memblock_virt_alloc(ai->nr_groups *
1584 sizeof(group_sizes[0]), 0);
1585 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1586 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
Tejun Heo2f39e632009-07-04 08:11:00 +09001587
Tejun Heofd1e8a12009-08-14 15:00:51 +09001588 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
Tejun Heoffe0d5a2009-09-29 09:17:56 +09001589 unit_map[cpu] = UINT_MAX;
Tejun Heoa855b842011-11-18 10:55:35 -08001590
1591 pcpu_low_unit_cpu = NR_CPUS;
1592 pcpu_high_unit_cpu = NR_CPUS;
Tejun Heo2f39e632009-07-04 08:11:00 +09001593
Tejun Heofd1e8a12009-08-14 15:00:51 +09001594 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1595 const struct pcpu_group_info *gi = &ai->groups[group];
Tejun Heo2f39e632009-07-04 08:11:00 +09001596
Tejun Heo65632972009-08-14 15:00:52 +09001597 group_offsets[group] = gi->base_offset;
1598 group_sizes[group] = gi->nr_units * ai->unit_size;
1599
Tejun Heofd1e8a12009-08-14 15:00:51 +09001600 for (i = 0; i < gi->nr_units; i++) {
1601 cpu = gi->cpu_map[i];
1602 if (cpu == NR_CPUS)
1603 continue;
1604
Dan Carpenter9f295662014-10-29 11:45:04 +03001605 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
Tejun Heo635b75f2009-09-24 09:43:11 +09001606 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1607 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001608
1609 unit_map[cpu] = unit + i;
Tejun Heofb435d52009-08-14 15:00:51 +09001610 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1611
Tejun Heoa855b842011-11-18 10:55:35 -08001612 /* determine low/high unit_cpu */
1613 if (pcpu_low_unit_cpu == NR_CPUS ||
1614 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1615 pcpu_low_unit_cpu = cpu;
1616 if (pcpu_high_unit_cpu == NR_CPUS ||
1617 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1618 pcpu_high_unit_cpu = cpu;
Tejun Heo2f39e632009-07-04 08:11:00 +09001619 }
Tejun Heo2f39e632009-07-04 08:11:00 +09001620 }
Tejun Heofd1e8a12009-08-14 15:00:51 +09001621 pcpu_nr_units = unit;
1622
1623 for_each_possible_cpu(cpu)
Tejun Heo635b75f2009-09-24 09:43:11 +09001624 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1625
1626 /* we're done parsing the input, undefine BUG macro and dump config */
1627#undef PCPU_SETUP_BUG_ON
Tejun Heobcbea792010-12-22 14:19:14 +01001628 pcpu_dump_alloc_info(KERN_DEBUG, ai);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001629
Tejun Heo65632972009-08-14 15:00:52 +09001630 pcpu_nr_groups = ai->nr_groups;
1631 pcpu_group_offsets = group_offsets;
1632 pcpu_group_sizes = group_sizes;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001633 pcpu_unit_map = unit_map;
Tejun Heofb435d52009-08-14 15:00:51 +09001634 pcpu_unit_offsets = unit_off;
Tejun Heo2f39e632009-07-04 08:11:00 +09001635
1636 /* determine basic parameters */
Tejun Heofd1e8a12009-08-14 15:00:51 +09001637 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod9b55ee2009-02-24 11:57:21 +09001638 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heo65632972009-08-14 15:00:52 +09001639 pcpu_atom_size = ai->atom_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09001640 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1641 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
Tejun Heocafe8812009-03-06 14:33:59 +09001642
Tejun Heod9b55ee2009-02-24 11:57:21 +09001643 /*
1644 * Allocate chunk slots. The additional last slot is for
1645 * empty chunks.
1646 */
1647 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001648 pcpu_slot = memblock_virt_alloc(
1649 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001650 for (i = 0; i < pcpu_nr_slots; i++)
1651 INIT_LIST_HEAD(&pcpu_slot[i]);
1652
Tejun Heoedcb4632009-03-06 14:33:59 +09001653 /*
1654 * Initialize static chunk. If reserved_size is zero, the
1655 * static chunk covers static area + dynamic allocation area
1656 * in the first chunk. If reserved_size is not zero, it
1657 * covers static area + reserved area (mostly used for module
1658 * static percpu allocation).
1659 */
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001660 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heo2441d152009-03-06 14:33:59 +09001661 INIT_LIST_HEAD(&schunk->list);
Tejun Heo4f996e22016-05-25 11:48:25 -04001662 INIT_LIST_HEAD(&schunk->map_extend_list);
Tejun Heobba174f2009-08-14 15:00:51 +09001663 schunk->base_addr = base_addr;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001664 schunk->map = smap;
1665 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001666 schunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001667 bitmap_fill(schunk->populated, pcpu_unit_pages);
Tejun Heob539b872014-09-02 14:46:05 -04001668 schunk->nr_populated = pcpu_unit_pages;
Tejun Heoedcb4632009-03-06 14:33:59 +09001669
Tejun Heofd1e8a12009-08-14 15:00:51 +09001670 if (ai->reserved_size) {
1671 schunk->free_size = ai->reserved_size;
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001672 pcpu_reserved_chunk = schunk;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001673 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
Tejun Heoedcb4632009-03-06 14:33:59 +09001674 } else {
1675 schunk->free_size = dyn_size;
1676 dyn_size = 0; /* dynamic area covered */
1677 }
Tejun Heo2441d152009-03-06 14:33:59 +09001678 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001679
Al Viro723ad1d2014-03-06 21:13:18 -05001680 schunk->map[0] = 1;
1681 schunk->map[1] = ai->static_size;
1682 schunk->map_used = 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001683 if (schunk->free_size)
Baoquan He292c24a2015-07-20 22:55:28 +08001684 schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1685 schunk->map[schunk->map_used] |= 1;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001686
Tejun Heoedcb4632009-03-06 14:33:59 +09001687 /* init dynamic chunk if necessary */
1688 if (dyn_size) {
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001689 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001690 INIT_LIST_HEAD(&dchunk->list);
Tejun Heo4f996e22016-05-25 11:48:25 -04001691 INIT_LIST_HEAD(&dchunk->map_extend_list);
Tejun Heobba174f2009-08-14 15:00:51 +09001692 dchunk->base_addr = base_addr;
Tejun Heoedcb4632009-03-06 14:33:59 +09001693 dchunk->map = dmap;
1694 dchunk->map_alloc = ARRAY_SIZE(dmap);
Tejun Heo38a6be52009-07-04 08:10:59 +09001695 dchunk->immutable = true;
Tejun Heoce3141a2009-07-04 08:11:00 +09001696 bitmap_fill(dchunk->populated, pcpu_unit_pages);
Tejun Heob539b872014-09-02 14:46:05 -04001697 dchunk->nr_populated = pcpu_unit_pages;
Tejun Heoedcb4632009-03-06 14:33:59 +09001698
1699 dchunk->contig_hint = dchunk->free_size = dyn_size;
Al Viro723ad1d2014-03-06 21:13:18 -05001700 dchunk->map[0] = 1;
1701 dchunk->map[1] = pcpu_reserved_chunk_limit;
1702 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1703 dchunk->map_used = 2;
Tejun Heoedcb4632009-03-06 14:33:59 +09001704 }
1705
Tejun Heo2441d152009-03-06 14:33:59 +09001706 /* link the first chunk in */
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001707 pcpu_first_chunk = dchunk ?: schunk;
Tejun Heob539b872014-09-02 14:46:05 -04001708 pcpu_nr_empty_pop_pages +=
1709 pcpu_count_occupied_pages(pcpu_first_chunk, 1);
Tejun Heoae9e6bc92009-04-02 13:19:54 +09001710 pcpu_chunk_relocate(pcpu_first_chunk, -1);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001711
1712 /* we're done */
Tejun Heobba174f2009-08-14 15:00:51 +09001713 pcpu_base_addr = base_addr;
Tejun Heofb435d52009-08-14 15:00:51 +09001714 return 0;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001715}
Tejun Heo66c3a752009-03-10 16:27:48 +09001716
Tejun Heobbddff02010-09-03 18:22:48 +02001717#ifdef CONFIG_SMP
1718
Andi Kleen17f36092012-10-04 17:12:07 -07001719const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
Tejun Heof58dc012009-08-14 15:00:50 +09001720 [PCPU_FC_AUTO] = "auto",
1721 [PCPU_FC_EMBED] = "embed",
1722 [PCPU_FC_PAGE] = "page",
Tejun Heof58dc012009-08-14 15:00:50 +09001723};
Tejun Heo66c3a752009-03-10 16:27:48 +09001724
Tejun Heof58dc012009-08-14 15:00:50 +09001725enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1726
1727static int __init percpu_alloc_setup(char *str)
Tejun Heo66c3a752009-03-10 16:27:48 +09001728{
Cyrill Gorcunov5479c782012-11-25 01:17:13 +04001729 if (!str)
1730 return -EINVAL;
1731
Tejun Heof58dc012009-08-14 15:00:50 +09001732 if (0)
1733 /* nada */;
1734#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1735 else if (!strcmp(str, "embed"))
1736 pcpu_chosen_fc = PCPU_FC_EMBED;
1737#endif
1738#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1739 else if (!strcmp(str, "page"))
1740 pcpu_chosen_fc = PCPU_FC_PAGE;
1741#endif
Tejun Heof58dc012009-08-14 15:00:50 +09001742 else
Joe Perches870d4b12016-03-17 14:19:53 -07001743 pr_warn("unknown allocator %s specified\n", str);
Tejun Heo66c3a752009-03-10 16:27:48 +09001744
Tejun Heof58dc012009-08-14 15:00:50 +09001745 return 0;
Tejun Heo66c3a752009-03-10 16:27:48 +09001746}
Tejun Heof58dc012009-08-14 15:00:50 +09001747early_param("percpu_alloc", percpu_alloc_setup);
Tejun Heo66c3a752009-03-10 16:27:48 +09001748
Tejun Heo3c9a0242010-09-09 18:00:15 +02001749/*
1750 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1751 * Build it if needed by the arch config or the generic setup is going
1752 * to be used.
1753 */
Tejun Heo08fc4582009-08-14 15:00:49 +09001754#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1755 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
Tejun Heo3c9a0242010-09-09 18:00:15 +02001756#define BUILD_EMBED_FIRST_CHUNK
1757#endif
1758
1759/* build pcpu_page_first_chunk() iff needed by the arch config */
1760#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1761#define BUILD_PAGE_FIRST_CHUNK
1762#endif
1763
1764/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1765#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1766/**
Tejun Heofbf59bc2009-02-20 16:29:08 +09001767 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1768 * @reserved_size: the size of reserved percpu area in bytes
1769 * @dyn_size: minimum free size for dynamic allocation in bytes
1770 * @atom_size: allocation atom size
1771 * @cpu_distance_fn: callback to determine distance between cpus, optional
1772 *
1773 * This function determines grouping of units, their mappings to cpus
1774 * and other parameters considering needed percpu size, allocation
1775 * atom size and distances between CPUs.
1776 *
Yannick Guerrinibffc4372015-03-06 23:30:42 +01001777 * Groups are always multiples of atom size and CPUs which are of
Tejun Heofbf59bc2009-02-20 16:29:08 +09001778 * LOCAL_DISTANCE both ways are grouped together and share space for
1779 * units in the same group. The returned configuration is guaranteed
1780 * to have CPUs on different nodes on different groups and >=75% usage
1781 * of allocated virtual address space.
1782 *
1783 * RETURNS:
1784 * On success, pointer to the new allocation_info is returned. On
1785 * failure, ERR_PTR value is returned.
1786 */
1787static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1788 size_t reserved_size, size_t dyn_size,
1789 size_t atom_size,
1790 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1791{
1792 static int group_map[NR_CPUS] __initdata;
1793 static int group_cnt[NR_CPUS] __initdata;
1794 const size_t static_size = __per_cpu_end - __per_cpu_start;
1795 int nr_groups = 1, nr_units = 0;
1796 size_t size_sum, min_unit_size, alloc_size;
1797 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1798 int last_allocs, group, unit;
1799 unsigned int cpu, tcpu;
1800 struct pcpu_alloc_info *ai;
1801 unsigned int *cpu_map;
1802
1803 /* this function may be called multiple times */
1804 memset(group_map, 0, sizeof(group_map));
1805 memset(group_cnt, 0, sizeof(group_cnt));
1806
1807 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1808 size_sum = PFN_ALIGN(static_size + reserved_size +
1809 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1810 dyn_size = size_sum - static_size - reserved_size;
1811
1812 /*
1813 * Determine min_unit_size, alloc_size and max_upa such that
1814 * alloc_size is multiple of atom_size and is the smallest
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001815 * which can accommodate 4k aligned segments which are equal to
Tejun Heofbf59bc2009-02-20 16:29:08 +09001816 * or larger than min_unit_size.
1817 */
1818 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1819
1820 alloc_size = roundup(min_unit_size, atom_size);
1821 upa = alloc_size / min_unit_size;
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08001822 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
Tejun Heofbf59bc2009-02-20 16:29:08 +09001823 upa--;
1824 max_upa = upa;
1825
1826 /* group cpus according to their proximity */
1827 for_each_possible_cpu(cpu) {
1828 group = 0;
1829 next_group:
1830 for_each_possible_cpu(tcpu) {
1831 if (cpu == tcpu)
1832 break;
1833 if (group_map[tcpu] == group && cpu_distance_fn &&
1834 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1835 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1836 group++;
1837 nr_groups = max(nr_groups, group + 1);
1838 goto next_group;
1839 }
1840 }
1841 group_map[cpu] = group;
1842 group_cnt[group]++;
1843 }
1844
1845 /*
1846 * Expand unit size until address space usage goes over 75%
1847 * and then as much as possible without using more address
1848 * space.
1849 */
1850 last_allocs = INT_MAX;
1851 for (upa = max_upa; upa; upa--) {
1852 int allocs = 0, wasted = 0;
1853
Alexander Kuleshovf09f1242015-11-05 18:46:43 -08001854 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
Tejun Heofbf59bc2009-02-20 16:29:08 +09001855 continue;
1856
1857 for (group = 0; group < nr_groups; group++) {
1858 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1859 allocs += this_allocs;
1860 wasted += this_allocs * upa - group_cnt[group];
1861 }
1862
1863 /*
1864 * Don't accept if wastage is over 1/3. The
1865 * greater-than comparison ensures upa==1 always
1866 * passes the following check.
1867 */
1868 if (wasted > num_possible_cpus() / 3)
1869 continue;
1870
1871 /* and then don't consume more memory */
1872 if (allocs > last_allocs)
1873 break;
1874 last_allocs = allocs;
1875 best_upa = upa;
1876 }
1877 upa = best_upa;
1878
1879 /* allocate and fill alloc_info */
1880 for (group = 0; group < nr_groups; group++)
1881 nr_units += roundup(group_cnt[group], upa);
1882
1883 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1884 if (!ai)
1885 return ERR_PTR(-ENOMEM);
1886 cpu_map = ai->groups[0].cpu_map;
1887
1888 for (group = 0; group < nr_groups; group++) {
1889 ai->groups[group].cpu_map = cpu_map;
1890 cpu_map += roundup(group_cnt[group], upa);
1891 }
1892
1893 ai->static_size = static_size;
1894 ai->reserved_size = reserved_size;
1895 ai->dyn_size = dyn_size;
1896 ai->unit_size = alloc_size / upa;
1897 ai->atom_size = atom_size;
1898 ai->alloc_size = alloc_size;
1899
1900 for (group = 0, unit = 0; group_cnt[group]; group++) {
1901 struct pcpu_group_info *gi = &ai->groups[group];
1902
1903 /*
1904 * Initialize base_offset as if all groups are located
1905 * back-to-back. The caller should update this to
1906 * reflect actual allocation.
1907 */
1908 gi->base_offset = unit * ai->unit_size;
1909
1910 for_each_possible_cpu(cpu)
1911 if (group_map[cpu] == group)
1912 gi->cpu_map[gi->nr_units++] = cpu;
1913 gi->nr_units = roundup(gi->nr_units, upa);
1914 unit += gi->nr_units;
1915 }
1916 BUG_ON(unit != nr_units);
1917
1918 return ai;
1919}
Tejun Heo3c9a0242010-09-09 18:00:15 +02001920#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001921
Tejun Heo3c9a0242010-09-09 18:00:15 +02001922#if defined(BUILD_EMBED_FIRST_CHUNK)
Tejun Heo66c3a752009-03-10 16:27:48 +09001923/**
1924 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
Tejun Heo66c3a752009-03-10 16:27:48 +09001925 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001926 * @dyn_size: minimum free size for dynamic allocation in bytes
Tejun Heoc8826dd2009-08-14 15:00:52 +09001927 * @atom_size: allocation atom size
1928 * @cpu_distance_fn: callback to determine distance between cpus, optional
1929 * @alloc_fn: function to allocate percpu page
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001930 * @free_fn: function to free percpu page
Tejun Heo66c3a752009-03-10 16:27:48 +09001931 *
1932 * This is a helper to ease setting up embedded first percpu chunk and
1933 * can be called where pcpu_setup_first_chunk() is expected.
1934 *
1935 * If this function is used to setup the first chunk, it is allocated
Tejun Heoc8826dd2009-08-14 15:00:52 +09001936 * by calling @alloc_fn and used as-is without being mapped into
1937 * vmalloc area. Allocations are always whole multiples of @atom_size
1938 * aligned to @atom_size.
1939 *
1940 * This enables the first chunk to piggy back on the linear physical
1941 * mapping which often uses larger page size. Please note that this
1942 * can result in very sparse cpu->unit mapping on NUMA machines thus
1943 * requiring large vmalloc address space. Don't use this allocator if
1944 * vmalloc space is not orders of magnitude larger than distances
1945 * between node memory addresses (ie. 32bit NUMA machines).
Tejun Heo66c3a752009-03-10 16:27:48 +09001946 *
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001947 * @dyn_size specifies the minimum dynamic area size.
Tejun Heo66c3a752009-03-10 16:27:48 +09001948 *
1949 * If the needed size is smaller than the minimum or specified unit
Tejun Heoc8826dd2009-08-14 15:00:52 +09001950 * size, the leftover is returned using @free_fn.
Tejun Heo66c3a752009-03-10 16:27:48 +09001951 *
1952 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09001953 * 0 on success, -errno on failure.
Tejun Heo66c3a752009-03-10 16:27:48 +09001954 */
Tejun Heo4ba6ce22010-06-27 18:49:59 +02001955int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
Tejun Heoc8826dd2009-08-14 15:00:52 +09001956 size_t atom_size,
1957 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1958 pcpu_fc_alloc_fn_t alloc_fn,
1959 pcpu_fc_free_fn_t free_fn)
Tejun Heo66c3a752009-03-10 16:27:48 +09001960{
Tejun Heoc8826dd2009-08-14 15:00:52 +09001961 void *base = (void *)ULONG_MAX;
1962 void **areas = NULL;
Tejun Heofd1e8a12009-08-14 15:00:51 +09001963 struct pcpu_alloc_info *ai;
zijun_hu93c76b6b2016-10-05 21:19:11 +08001964 size_t size_sum, areas_size;
1965 unsigned long max_distance;
zijun_hu9b739662016-10-05 21:30:24 +08001966 int group, i, highest_group, rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09001967
Tejun Heoc8826dd2009-08-14 15:00:52 +09001968 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1969 cpu_distance_fn);
Tejun Heofd1e8a12009-08-14 15:00:51 +09001970 if (IS_ERR(ai))
1971 return PTR_ERR(ai);
Tejun Heo66c3a752009-03-10 16:27:48 +09001972
Tejun Heofd1e8a12009-08-14 15:00:51 +09001973 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001974 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
Tejun Heo66c3a752009-03-10 16:27:48 +09001975
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08001976 areas = memblock_virt_alloc_nopanic(areas_size, 0);
Tejun Heoc8826dd2009-08-14 15:00:52 +09001977 if (!areas) {
Tejun Heofb435d52009-08-14 15:00:51 +09001978 rc = -ENOMEM;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001979 goto out_free;
Tejun Heofa8a7092009-06-22 11:56:24 +09001980 }
Tejun Heo66c3a752009-03-10 16:27:48 +09001981
zijun_hu9b739662016-10-05 21:30:24 +08001982 /* allocate, copy and determine base address & max_distance */
1983 highest_group = 0;
Tejun Heoc8826dd2009-08-14 15:00:52 +09001984 for (group = 0; group < ai->nr_groups; group++) {
1985 struct pcpu_group_info *gi = &ai->groups[group];
1986 unsigned int cpu = NR_CPUS;
1987 void *ptr;
Tejun Heo66c3a752009-03-10 16:27:48 +09001988
Tejun Heoc8826dd2009-08-14 15:00:52 +09001989 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1990 cpu = gi->cpu_map[i];
1991 BUG_ON(cpu == NR_CPUS);
1992
1993 /* allocate space for the whole group */
1994 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1995 if (!ptr) {
1996 rc = -ENOMEM;
1997 goto out_free_areas;
1998 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01001999 /* kmemleak tracks the percpu allocations separately */
2000 kmemleak_free(ptr);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002001 areas[group] = ptr;
2002
2003 base = min(ptr, base);
zijun_hu9b739662016-10-05 21:30:24 +08002004 if (ptr > areas[highest_group])
2005 highest_group = group;
2006 }
2007 max_distance = areas[highest_group] - base;
2008 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2009
2010 /* warn if maximum distance is further than 75% of vmalloc space */
2011 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2012 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2013 max_distance, VMALLOC_TOTAL);
2014#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2015 /* and fail if we have fallback */
2016 rc = -EINVAL;
2017 goto out_free_areas;
2018#endif
Tejun Heo42b64282012-04-27 08:42:53 -07002019 }
2020
2021 /*
2022 * Copy data and free unused parts. This should happen after all
2023 * allocations are complete; otherwise, we may end up with
2024 * overlapping groups.
2025 */
2026 for (group = 0; group < ai->nr_groups; group++) {
2027 struct pcpu_group_info *gi = &ai->groups[group];
2028 void *ptr = areas[group];
Tejun Heoc8826dd2009-08-14 15:00:52 +09002029
2030 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2031 if (gi->cpu_map[i] == NR_CPUS) {
2032 /* unused unit, free whole */
2033 free_fn(ptr, ai->unit_size);
2034 continue;
2035 }
2036 /* copy and return the unused part */
2037 memcpy(ptr, __per_cpu_load, ai->static_size);
2038 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2039 }
Tejun Heo66c3a752009-03-10 16:27:48 +09002040 }
2041
Tejun Heoc8826dd2009-08-14 15:00:52 +09002042 /* base address is now known, determine group base offsets */
Tejun Heo6ea529a2009-09-24 18:46:01 +09002043 for (group = 0; group < ai->nr_groups; group++) {
Tejun Heoc8826dd2009-08-14 15:00:52 +09002044 ai->groups[group].base_offset = areas[group] - base;
Tejun Heo6ea529a2009-09-24 18:46:01 +09002045 }
Tejun Heoc8826dd2009-08-14 15:00:52 +09002046
Joe Perches870d4b12016-03-17 14:19:53 -07002047 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002048 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2049 ai->dyn_size, ai->unit_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002050
Tejun Heofb435d52009-08-14 15:00:51 +09002051 rc = pcpu_setup_first_chunk(ai, base);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002052 goto out_free;
2053
2054out_free_areas:
2055 for (group = 0; group < ai->nr_groups; group++)
Michael Holzheuf851c8d2013-09-17 16:57:34 +02002056 if (areas[group])
2057 free_fn(areas[group],
2058 ai->groups[group].nr_units * ai->unit_size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002059out_free:
Tejun Heofd1e8a12009-08-14 15:00:51 +09002060 pcpu_free_alloc_info(ai);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002061 if (areas)
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002062 memblock_free_early(__pa(areas), areas_size);
Tejun Heofb435d52009-08-14 15:00:51 +09002063 return rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002064}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002065#endif /* BUILD_EMBED_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002066
Tejun Heo3c9a0242010-09-09 18:00:15 +02002067#ifdef BUILD_PAGE_FIRST_CHUNK
Tejun Heod4b95f82009-07-04 08:10:59 +09002068/**
Tejun Heo00ae4062009-08-14 15:00:49 +09002069 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
Tejun Heod4b95f82009-07-04 08:10:59 +09002070 * @reserved_size: the size of reserved percpu area in bytes
2071 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002072 * @free_fn: function to free percpu page, always called with PAGE_SIZE
Tejun Heod4b95f82009-07-04 08:10:59 +09002073 * @populate_pte_fn: function to populate pte
2074 *
Tejun Heo00ae4062009-08-14 15:00:49 +09002075 * This is a helper to ease setting up page-remapped first percpu
2076 * chunk and can be called where pcpu_setup_first_chunk() is expected.
Tejun Heod4b95f82009-07-04 08:10:59 +09002077 *
2078 * This is the basic allocator. Static percpu area is allocated
2079 * page-by-page into vmalloc area.
2080 *
2081 * RETURNS:
Tejun Heofb435d52009-08-14 15:00:51 +09002082 * 0 on success, -errno on failure.
Tejun Heod4b95f82009-07-04 08:10:59 +09002083 */
Tejun Heofb435d52009-08-14 15:00:51 +09002084int __init pcpu_page_first_chunk(size_t reserved_size,
2085 pcpu_fc_alloc_fn_t alloc_fn,
2086 pcpu_fc_free_fn_t free_fn,
2087 pcpu_fc_populate_pte_fn_t populate_pte_fn)
Tejun Heod4b95f82009-07-04 08:10:59 +09002088{
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002089 static struct vm_struct vm;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002090 struct pcpu_alloc_info *ai;
Tejun Heo00ae4062009-08-14 15:00:49 +09002091 char psize_str[16];
Tejun Heoce3141a2009-07-04 08:11:00 +09002092 int unit_pages;
Tejun Heod4b95f82009-07-04 08:10:59 +09002093 size_t pages_size;
Tejun Heoce3141a2009-07-04 08:11:00 +09002094 struct page **pages;
Tejun Heofb435d52009-08-14 15:00:51 +09002095 int unit, i, j, rc;
Tejun Heod4b95f82009-07-04 08:10:59 +09002096
Tejun Heo00ae4062009-08-14 15:00:49 +09002097 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2098
Tejun Heo4ba6ce22010-06-27 18:49:59 +02002099 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002100 if (IS_ERR(ai))
2101 return PTR_ERR(ai);
2102 BUG_ON(ai->nr_groups != 1);
2103 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2104
2105 unit_pages = ai->unit_size >> PAGE_SHIFT;
Tejun Heod4b95f82009-07-04 08:10:59 +09002106
2107 /* unaligned allocations can't be freed, round up to page size */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002108 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2109 sizeof(pages[0]));
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002110 pages = memblock_virt_alloc(pages_size, 0);
Tejun Heod4b95f82009-07-04 08:10:59 +09002111
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002112 /* allocate pages */
Tejun Heod4b95f82009-07-04 08:10:59 +09002113 j = 0;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002114 for (unit = 0; unit < num_possible_cpus(); unit++)
Tejun Heoce3141a2009-07-04 08:11:00 +09002115 for (i = 0; i < unit_pages; i++) {
Tejun Heofd1e8a12009-08-14 15:00:51 +09002116 unsigned int cpu = ai->groups[0].cpu_map[unit];
Tejun Heod4b95f82009-07-04 08:10:59 +09002117 void *ptr;
2118
Tejun Heo3cbc8562009-08-14 15:00:50 +09002119 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
Tejun Heod4b95f82009-07-04 08:10:59 +09002120 if (!ptr) {
Joe Perches870d4b12016-03-17 14:19:53 -07002121 pr_warn("failed to allocate %s page for cpu%u\n",
Joe Perches598d8092016-03-17 14:19:44 -07002122 psize_str, cpu);
Tejun Heod4b95f82009-07-04 08:10:59 +09002123 goto enomem;
2124 }
Catalin Marinasf528f0b2011-09-26 17:12:53 +01002125 /* kmemleak tracks the percpu allocations separately */
2126 kmemleak_free(ptr);
Tejun Heoce3141a2009-07-04 08:11:00 +09002127 pages[j++] = virt_to_page(ptr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002128 }
2129
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002130 /* allocate vm area, map the pages and copy static data */
2131 vm.flags = VM_ALLOC;
Tejun Heofd1e8a12009-08-14 15:00:51 +09002132 vm.size = num_possible_cpus() * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002133 vm_area_register_early(&vm, PAGE_SIZE);
2134
Tejun Heofd1e8a12009-08-14 15:00:51 +09002135 for (unit = 0; unit < num_possible_cpus(); unit++) {
Tejun Heo1d9d3252009-08-14 15:00:50 +09002136 unsigned long unit_addr =
Tejun Heofd1e8a12009-08-14 15:00:51 +09002137 (unsigned long)vm.addr + unit * ai->unit_size;
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002138
Tejun Heoce3141a2009-07-04 08:11:00 +09002139 for (i = 0; i < unit_pages; i++)
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002140 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2141
2142 /* pte already populated, the following shouldn't fail */
Tejun Heofb435d52009-08-14 15:00:51 +09002143 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2144 unit_pages);
2145 if (rc < 0)
2146 panic("failed to map percpu area, err=%d\n", rc);
Tejun Heo8f05a6a2009-07-04 08:10:59 +09002147
2148 /*
2149 * FIXME: Archs with virtual cache should flush local
2150 * cache for the linear mapping here - something
2151 * equivalent to flush_cache_vmap() on the local cpu.
2152 * flush_cache_vmap() can't be used as most supporting
2153 * data structures are not set up yet.
2154 */
2155
2156 /* copy static data */
Tejun Heofd1e8a12009-08-14 15:00:51 +09002157 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002158 }
2159
2160 /* we're ready, commit */
Joe Perches870d4b12016-03-17 14:19:53 -07002161 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
Tejun Heofd1e8a12009-08-14 15:00:51 +09002162 unit_pages, psize_str, vm.addr, ai->static_size,
2163 ai->reserved_size, ai->dyn_size);
Tejun Heo66c3a752009-03-10 16:27:48 +09002164
Tejun Heofb435d52009-08-14 15:00:51 +09002165 rc = pcpu_setup_first_chunk(ai, vm.addr);
Tejun Heod4b95f82009-07-04 08:10:59 +09002166 goto out_free_ar;
2167
2168enomem:
2169 while (--j >= 0)
Tejun Heoce3141a2009-07-04 08:11:00 +09002170 free_fn(page_address(pages[j]), PAGE_SIZE);
Tejun Heofb435d52009-08-14 15:00:51 +09002171 rc = -ENOMEM;
Tejun Heod4b95f82009-07-04 08:10:59 +09002172out_free_ar:
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002173 memblock_free_early(__pa(pages), pages_size);
Tejun Heofd1e8a12009-08-14 15:00:51 +09002174 pcpu_free_alloc_info(ai);
Tejun Heofb435d52009-08-14 15:00:51 +09002175 return rc;
Tejun Heo66c3a752009-03-10 16:27:48 +09002176}
Tejun Heo3c9a0242010-09-09 18:00:15 +02002177#endif /* BUILD_PAGE_FIRST_CHUNK */
Tejun Heod4b95f82009-07-04 08:10:59 +09002178
Tejun Heobbddff02010-09-03 18:22:48 +02002179#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
Tejun Heo8c4bfc62009-07-04 08:10:59 +09002180/*
Tejun Heobbddff02010-09-03 18:22:48 +02002181 * Generic SMP percpu area setup.
Tejun Heoe74e3962009-03-30 19:07:44 +09002182 *
2183 * The embedding helper is used because its behavior closely resembles
2184 * the original non-dynamic generic percpu area setup. This is
2185 * important because many archs have addressing restrictions and might
2186 * fail if the percpu area is located far away from the previous
2187 * location. As an added bonus, in non-NUMA cases, embedding is
2188 * generally a good idea TLB-wise because percpu area can piggy back
2189 * on the physical linear memory mapping which uses large page
2190 * mappings on applicable archs.
2191 */
Tejun Heoe74e3962009-03-30 19:07:44 +09002192unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2193EXPORT_SYMBOL(__per_cpu_offset);
2194
Tejun Heoc8826dd2009-08-14 15:00:52 +09002195static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2196 size_t align)
2197{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002198 return memblock_virt_alloc_from_nopanic(
2199 size, align, __pa(MAX_DMA_ADDRESS));
Tejun Heoc8826dd2009-08-14 15:00:52 +09002200}
2201
2202static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2203{
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002204 memblock_free_early(__pa(ptr), size);
Tejun Heoc8826dd2009-08-14 15:00:52 +09002205}
2206
Tejun Heoe74e3962009-03-30 19:07:44 +09002207void __init setup_per_cpu_areas(void)
2208{
Tejun Heoe74e3962009-03-30 19:07:44 +09002209 unsigned long delta;
2210 unsigned int cpu;
Tejun Heofb435d52009-08-14 15:00:51 +09002211 int rc;
Tejun Heoe74e3962009-03-30 19:07:44 +09002212
2213 /*
2214 * Always reserve area for module percpu variables. That's
2215 * what the legacy allocator did.
2216 */
Tejun Heofb435d52009-08-14 15:00:51 +09002217 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
Tejun Heoc8826dd2009-08-14 15:00:52 +09002218 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2219 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
Tejun Heofb435d52009-08-14 15:00:51 +09002220 if (rc < 0)
Tejun Heobbddff02010-09-03 18:22:48 +02002221 panic("Failed to initialize percpu areas.");
Tejun Heoe74e3962009-03-30 19:07:44 +09002222
2223 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2224 for_each_possible_cpu(cpu)
Tejun Heofb435d52009-08-14 15:00:51 +09002225 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
Tejun Heoe74e3962009-03-30 19:07:44 +09002226}
Tejun Heobbddff02010-09-03 18:22:48 +02002227#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2228
2229#else /* CONFIG_SMP */
2230
2231/*
2232 * UP percpu area setup.
2233 *
2234 * UP always uses km-based percpu allocator with identity mapping.
2235 * Static percpu variables are indistinguishable from the usual static
2236 * variables and don't require any special preparation.
2237 */
2238void __init setup_per_cpu_areas(void)
2239{
2240 const size_t unit_size =
2241 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2242 PERCPU_DYNAMIC_RESERVE));
2243 struct pcpu_alloc_info *ai;
2244 void *fc;
2245
2246 ai = pcpu_alloc_alloc_info(1, 1);
Santosh Shilimkar999c17e2014-01-21 15:50:40 -08002247 fc = memblock_virt_alloc_from_nopanic(unit_size,
2248 PAGE_SIZE,
2249 __pa(MAX_DMA_ADDRESS));
Tejun Heobbddff02010-09-03 18:22:48 +02002250 if (!ai || !fc)
2251 panic("Failed to allocate memory for percpu areas.");
Catalin Marinas100d13c2012-05-09 16:55:19 +01002252 /* kmemleak tracks the percpu allocations separately */
2253 kmemleak_free(fc);
Tejun Heobbddff02010-09-03 18:22:48 +02002254
2255 ai->dyn_size = unit_size;
2256 ai->unit_size = unit_size;
2257 ai->atom_size = unit_size;
2258 ai->alloc_size = unit_size;
2259 ai->groups[0].nr_units = 1;
2260 ai->groups[0].cpu_map[0] = 0;
2261
2262 if (pcpu_setup_first_chunk(ai, fc) < 0)
2263 panic("Failed to initialize percpu areas.");
2264}
2265
2266#endif /* CONFIG_SMP */
Tejun Heo099a19d2010-06-27 18:50:00 +02002267
2268/*
2269 * First and reserved chunks are initialized with temporary allocation
2270 * map in initdata so that they can be used before slab is online.
2271 * This function is called after slab is brought up and replaces those
2272 * with properly allocated maps.
2273 */
2274void __init percpu_init_late(void)
2275{
2276 struct pcpu_chunk *target_chunks[] =
2277 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2278 struct pcpu_chunk *chunk;
2279 unsigned long flags;
2280 int i;
2281
2282 for (i = 0; (chunk = target_chunks[i]); i++) {
2283 int *map;
2284 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2285
2286 BUILD_BUG_ON(size > PAGE_SIZE);
2287
Bob Liu90459ce02011-08-04 11:02:33 +02002288 map = pcpu_mem_zalloc(size);
Tejun Heo099a19d2010-06-27 18:50:00 +02002289 BUG_ON(!map);
2290
2291 spin_lock_irqsave(&pcpu_lock, flags);
2292 memcpy(map, chunk->map, size);
2293 chunk->map = map;
2294 spin_unlock_irqrestore(&pcpu_lock, flags);
2295 }
2296}
Tejun Heo1a4d7602014-09-02 14:46:05 -04002297
2298/*
2299 * Percpu allocator is initialized early during boot when neither slab or
2300 * workqueue is available. Plug async management until everything is up
2301 * and running.
2302 */
2303static int __init percpu_enable_async(void)
2304{
2305 pcpu_async_enabled = true;
2306 return 0;
2307}
2308subsys_initcall(percpu_enable_async);