Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1 | /* |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 2 | * mm/percpu.c - percpu memory allocator |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 3 | * |
| 4 | * Copyright (C) 2009 SUSE Linux Products GmbH |
| 5 | * Copyright (C) 2009 Tejun Heo <tj@kernel.org> |
| 6 | * |
| 7 | * This file is released under the GPLv2. |
| 8 | * |
| 9 | * This is percpu allocator which can handle both static and dynamic |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 10 | * 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 Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 13 | * (special boot time alloc/init handling necessary as these areas |
| 14 | * need to be brought up before allocation services are running). |
| 15 | * Unit grows as necessary and all units grow or shrink in unison. |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 16 | * When a chunk is filled up, another chunk is allocated. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 17 | * |
| 18 | * c0 c1 c2 |
| 19 | * ------------------- ------------------- ------------ |
| 20 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u |
| 21 | * ------------------- ...... ------------------- .... ------------ |
| 22 | * |
| 23 | * Allocation is done in offset-size areas of single unit space. Ie, |
| 24 | * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 25 | * 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 Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 29 | * |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 30 | * There are usually many small percpu allocations many of them being |
| 31 | * as small as 4 bytes. The allocator organizes chunks into lists |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 32 | * 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 Kim | 4785879 | 2010-08-11 11:24:10 +0900 | [diff] [blame] | 34 | * guaranteed to be equal to or larger than the maximum contiguous |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 35 | * 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 Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 43 | * 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 Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 45 | * |
| 46 | * To use this allocator, arch code should do the followings. |
| 47 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 48 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 49 | * regular address to percpu pointer and back if they need to be |
| 50 | * different from the default |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 51 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 52 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
| 53 | * setup the first chunk containing the kernel static percpu area |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 54 | */ |
| 55 | |
| 56 | #include <linux/bitmap.h> |
| 57 | #include <linux/bootmem.h> |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 58 | #include <linux/err.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 59 | #include <linux/list.h> |
Tejun Heo | a530b79 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 60 | #include <linux/log2.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 61 | #include <linux/mm.h> |
| 62 | #include <linux/module.h> |
| 63 | #include <linux/mutex.h> |
| 64 | #include <linux/percpu.h> |
| 65 | #include <linux/pfn.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 66 | #include <linux/slab.h> |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 67 | #include <linux/spinlock.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 68 | #include <linux/vmalloc.h> |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 69 | #include <linux/workqueue.h> |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 70 | #include <linux/kmemleak.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 71 | |
| 72 | #include <asm/cacheflush.h> |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 73 | #include <asm/sections.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 74 | #include <asm/tlbflush.h> |
Vivek Goyal | 3b034b0 | 2009-11-24 15:50:03 +0900 | [diff] [blame] | 75 | #include <asm/io.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 76 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 77 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
| 78 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 79 | #define PCPU_ATOMIC_MAP_MARGIN_LOW 32 |
| 80 | #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64 |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 81 | #define PCPU_EMPTY_POP_PAGES_LOW 2 |
| 82 | #define PCPU_EMPTY_POP_PAGES_HIGH 4 |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 83 | |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 84 | #ifdef CONFIG_SMP |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 85 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
| 86 | #ifndef __addr_to_pcpu_ptr |
| 87 | #define __addr_to_pcpu_ptr(addr) \ |
Tejun Heo | 43cf38e | 2010-02-02 14:38:57 +0900 | [diff] [blame] | 88 | (void __percpu *)((unsigned long)(addr) - \ |
| 89 | (unsigned long)pcpu_base_addr + \ |
| 90 | (unsigned long)__per_cpu_start) |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 91 | #endif |
| 92 | #ifndef __pcpu_ptr_to_addr |
| 93 | #define __pcpu_ptr_to_addr(ptr) \ |
Tejun Heo | 43cf38e | 2010-02-02 14:38:57 +0900 | [diff] [blame] | 94 | (void __force *)((unsigned long)(ptr) + \ |
| 95 | (unsigned long)pcpu_base_addr - \ |
| 96 | (unsigned long)__per_cpu_start) |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 97 | #endif |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 98 | #else /* CONFIG_SMP */ |
| 99 | /* on UP, it's always identity mapped */ |
| 100 | #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) |
| 101 | #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) |
| 102 | #endif /* CONFIG_SMP */ |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 103 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 104 | struct pcpu_chunk { |
| 105 | struct list_head list; /* linked to pcpu_slot lists */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 106 | int free_size; /* free bytes in the chunk */ |
| 107 | int contig_hint; /* max contiguous size hint */ |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 108 | void *base_addr; /* base address of this chunk */ |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 109 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 110 | int map_used; /* # of map entries used before the sentry */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 111 | int map_alloc; /* # of map entries allocated */ |
| 112 | int *map; /* allocation map */ |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 113 | struct work_struct map_extend_work;/* async ->map[] extension */ |
| 114 | |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 115 | void *data; /* chunk data */ |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 116 | int first_free; /* no free below this */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 117 | bool immutable; /* no [de]population allowed */ |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 118 | int nr_populated; /* # of populated pages */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 119 | unsigned long populated[]; /* populated bitmap */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 120 | }; |
| 121 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 122 | static int pcpu_unit_pages __read_mostly; |
| 123 | static int pcpu_unit_size __read_mostly; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 124 | static int pcpu_nr_units __read_mostly; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 125 | static int pcpu_atom_size __read_mostly; |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 126 | static int pcpu_nr_slots __read_mostly; |
| 127 | static size_t pcpu_chunk_struct_size __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 128 | |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 129 | /* cpus with the lowest and highest unit addresses */ |
| 130 | static unsigned int pcpu_low_unit_cpu __read_mostly; |
| 131 | static unsigned int pcpu_high_unit_cpu __read_mostly; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 132 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 133 | /* the address of the first chunk which starts with the kernel static area */ |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 134 | void *pcpu_base_addr __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 135 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
| 136 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 137 | static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ |
| 138 | const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 139 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 140 | /* group information, used for vm allocation */ |
| 141 | static int pcpu_nr_groups __read_mostly; |
| 142 | static const unsigned long *pcpu_group_offsets __read_mostly; |
| 143 | static const size_t *pcpu_group_sizes __read_mostly; |
| 144 | |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 145 | /* |
| 146 | * The first chunk which always exists. Note that unlike other |
| 147 | * chunks, this one can be allocated and mapped in several different |
| 148 | * ways and thus often doesn't live in the vmalloc area. |
| 149 | */ |
| 150 | static struct pcpu_chunk *pcpu_first_chunk; |
| 151 | |
| 152 | /* |
| 153 | * Optional reserved chunk. This chunk reserves part of the first |
| 154 | * chunk and serves it for reserved allocations. The amount of |
| 155 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved |
| 156 | * area doesn't exist, the following variables contain NULL and 0 |
| 157 | * respectively. |
| 158 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 159 | static struct pcpu_chunk *pcpu_reserved_chunk; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 160 | static int pcpu_reserved_chunk_limit; |
| 161 | |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 162 | static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ |
| 163 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 164 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 165 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 166 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 167 | /* |
| 168 | * The number of empty populated pages, protected by pcpu_lock. The |
| 169 | * reserved chunk doesn't contribute to the count. |
| 170 | */ |
| 171 | static int pcpu_nr_empty_pop_pages; |
| 172 | |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 173 | /* |
| 174 | * Balance work is used to populate or destroy chunks asynchronously. We |
| 175 | * try to keep the number of populated free pages between |
| 176 | * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one |
| 177 | * empty chunk. |
| 178 | */ |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 179 | static void pcpu_balance_workfn(struct work_struct *work); |
| 180 | static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 181 | static bool pcpu_async_enabled __read_mostly; |
| 182 | static bool pcpu_atomic_alloc_failed; |
| 183 | |
| 184 | static void pcpu_schedule_balance_work(void) |
| 185 | { |
| 186 | if (pcpu_async_enabled) |
| 187 | schedule_work(&pcpu_balance_work); |
| 188 | } |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 189 | |
Tejun Heo | 020ec65 | 2010-04-09 18:57:00 +0900 | [diff] [blame] | 190 | static bool pcpu_addr_in_first_chunk(void *addr) |
| 191 | { |
| 192 | void *first_start = pcpu_first_chunk->base_addr; |
| 193 | |
| 194 | return addr >= first_start && addr < first_start + pcpu_unit_size; |
| 195 | } |
| 196 | |
| 197 | static bool pcpu_addr_in_reserved_chunk(void *addr) |
| 198 | { |
| 199 | void *first_start = pcpu_first_chunk->base_addr; |
| 200 | |
| 201 | return addr >= first_start && |
| 202 | addr < first_start + pcpu_reserved_chunk_limit; |
| 203 | } |
| 204 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 205 | static int __pcpu_size_to_slot(int size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 206 | { |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 207 | int highbit = fls(size); /* size is in bytes */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 208 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
| 209 | } |
| 210 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 211 | static int pcpu_size_to_slot(int size) |
| 212 | { |
| 213 | if (size == pcpu_unit_size) |
| 214 | return pcpu_nr_slots - 1; |
| 215 | return __pcpu_size_to_slot(size); |
| 216 | } |
| 217 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 218 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
| 219 | { |
| 220 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) |
| 221 | return 0; |
| 222 | |
| 223 | return pcpu_size_to_slot(chunk->free_size); |
| 224 | } |
| 225 | |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 226 | /* set the pointer to a chunk in a page struct */ |
| 227 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) |
| 228 | { |
| 229 | page->index = (unsigned long)pcpu; |
| 230 | } |
| 231 | |
| 232 | /* obtain pointer to a chunk from a page struct */ |
| 233 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) |
| 234 | { |
| 235 | return (struct pcpu_chunk *)page->index; |
| 236 | } |
| 237 | |
| 238 | static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 239 | { |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 240 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 241 | } |
| 242 | |
Tejun Heo | 9983b6f0 | 2010-06-18 11:44:31 +0200 | [diff] [blame] | 243 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
| 244 | unsigned int cpu, int page_idx) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 245 | { |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 246 | return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 247 | (page_idx << PAGE_SHIFT); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 248 | } |
| 249 | |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 250 | static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, |
| 251 | int *rs, int *re, int end) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 252 | { |
| 253 | *rs = find_next_zero_bit(chunk->populated, end, *rs); |
| 254 | *re = find_next_bit(chunk->populated, end, *rs + 1); |
| 255 | } |
| 256 | |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 257 | static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, |
| 258 | int *rs, int *re, int end) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 259 | { |
| 260 | *rs = find_next_bit(chunk->populated, end, *rs); |
| 261 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); |
| 262 | } |
| 263 | |
| 264 | /* |
| 265 | * (Un)populated page region iterators. Iterate over (un)populated |
Uwe Kleine-König | b595076 | 2010-11-01 15:38:34 -0400 | [diff] [blame] | 266 | * page regions between @start and @end in @chunk. @rs and @re should |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 267 | * be integer variables and will be set to start and end page index of |
| 268 | * the current region. |
| 269 | */ |
| 270 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ |
| 271 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ |
| 272 | (rs) < (re); \ |
| 273 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) |
| 274 | |
| 275 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ |
| 276 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ |
| 277 | (rs) < (re); \ |
| 278 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) |
| 279 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 280 | /** |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 281 | * pcpu_mem_zalloc - allocate memory |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 282 | * @size: bytes to allocate |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 283 | * |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 284 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 285 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 286 | * memory is always zeroed. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 287 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 288 | * CONTEXT: |
| 289 | * Does GFP_KERNEL allocation. |
| 290 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 291 | * RETURNS: |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 292 | * Pointer to the allocated area on success, NULL on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 293 | */ |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 294 | static void *pcpu_mem_zalloc(size_t size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 295 | { |
Tejun Heo | 099a19d | 2010-06-27 18:50:00 +0200 | [diff] [blame] | 296 | if (WARN_ON_ONCE(!slab_is_available())) |
| 297 | return NULL; |
| 298 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 299 | if (size <= PAGE_SIZE) |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 300 | return kzalloc(size, GFP_KERNEL); |
Jesper Juhl | 7af4c09 | 2010-10-30 15:56:54 +0200 | [diff] [blame] | 301 | else |
| 302 | return vzalloc(size); |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 303 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 304 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 305 | /** |
| 306 | * pcpu_mem_free - free memory |
| 307 | * @ptr: memory to free |
| 308 | * @size: size of the area |
| 309 | * |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 310 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 311 | */ |
| 312 | static void pcpu_mem_free(void *ptr, size_t size) |
| 313 | { |
| 314 | if (size <= PAGE_SIZE) |
| 315 | kfree(ptr); |
| 316 | else |
| 317 | vfree(ptr); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 318 | } |
| 319 | |
| 320 | /** |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 321 | * pcpu_count_occupied_pages - count the number of pages an area occupies |
| 322 | * @chunk: chunk of interest |
| 323 | * @i: index of the area in question |
| 324 | * |
| 325 | * Count the number of pages chunk's @i'th area occupies. When the area's |
| 326 | * start and/or end address isn't aligned to page boundary, the straddled |
| 327 | * page is included in the count iff the rest of the page is free. |
| 328 | */ |
| 329 | static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i) |
| 330 | { |
| 331 | int off = chunk->map[i] & ~1; |
| 332 | int end = chunk->map[i + 1] & ~1; |
| 333 | |
| 334 | if (!PAGE_ALIGNED(off) && i > 0) { |
| 335 | int prev = chunk->map[i - 1]; |
| 336 | |
| 337 | if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE)) |
| 338 | off = round_down(off, PAGE_SIZE); |
| 339 | } |
| 340 | |
| 341 | if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) { |
| 342 | int next = chunk->map[i + 1]; |
| 343 | int nend = chunk->map[i + 2] & ~1; |
| 344 | |
| 345 | if (!(next & 1) && nend >= round_up(end, PAGE_SIZE)) |
| 346 | end = round_up(end, PAGE_SIZE); |
| 347 | } |
| 348 | |
| 349 | return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0); |
| 350 | } |
| 351 | |
| 352 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 353 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot |
| 354 | * @chunk: chunk of interest |
| 355 | * @oslot: the previous slot it was on |
| 356 | * |
| 357 | * This function is called after an allocation or free changed @chunk. |
| 358 | * New slot according to the changed state is determined and @chunk is |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 359 | * moved to the slot. Note that the reserved chunk is never put on |
| 360 | * chunk slots. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 361 | * |
| 362 | * CONTEXT: |
| 363 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 364 | */ |
| 365 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) |
| 366 | { |
| 367 | int nslot = pcpu_chunk_slot(chunk); |
| 368 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 369 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 370 | if (oslot < nslot) |
| 371 | list_move(&chunk->list, &pcpu_slot[nslot]); |
| 372 | else |
| 373 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); |
| 374 | } |
| 375 | } |
| 376 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 377 | /** |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 378 | * pcpu_need_to_extend - determine whether chunk area map needs to be extended |
| 379 | * @chunk: chunk of interest |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 380 | * @is_atomic: the allocation context |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 381 | * |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 382 | * Determine whether area map of @chunk needs to be extended. If |
| 383 | * @is_atomic, only the amount necessary for a new allocation is |
| 384 | * considered; however, async extension is scheduled if the left amount is |
| 385 | * low. If !@is_atomic, it aims for more empty space. Combined, this |
| 386 | * ensures that the map is likely to have enough available space to |
| 387 | * accomodate atomic allocations which can't extend maps directly. |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 388 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 389 | * CONTEXT: |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 390 | * pcpu_lock. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 391 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 392 | * RETURNS: |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 393 | * New target map allocation length if extension is necessary, 0 |
| 394 | * otherwise. |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 395 | */ |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 396 | static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic) |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 397 | { |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 398 | int margin, new_alloc; |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 399 | |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 400 | if (is_atomic) { |
| 401 | margin = 3; |
| 402 | |
| 403 | if (chunk->map_alloc < |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 404 | chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW && |
| 405 | pcpu_async_enabled) |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 406 | schedule_work(&chunk->map_extend_work); |
| 407 | } else { |
| 408 | margin = PCPU_ATOMIC_MAP_MARGIN_HIGH; |
| 409 | } |
| 410 | |
| 411 | if (chunk->map_alloc >= chunk->map_used + margin) |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 412 | return 0; |
| 413 | |
| 414 | new_alloc = PCPU_DFL_MAP_ALLOC; |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 415 | while (new_alloc < chunk->map_used + margin) |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 416 | new_alloc *= 2; |
| 417 | |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 418 | return new_alloc; |
| 419 | } |
| 420 | |
| 421 | /** |
| 422 | * pcpu_extend_area_map - extend area map of a chunk |
| 423 | * @chunk: chunk of interest |
| 424 | * @new_alloc: new target allocation length of the area map |
| 425 | * |
| 426 | * Extend area map of @chunk to have @new_alloc entries. |
| 427 | * |
| 428 | * CONTEXT: |
| 429 | * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. |
| 430 | * |
| 431 | * RETURNS: |
| 432 | * 0 on success, -errno on failure. |
| 433 | */ |
| 434 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) |
| 435 | { |
| 436 | int *old = NULL, *new = NULL; |
| 437 | size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); |
| 438 | unsigned long flags; |
| 439 | |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 440 | new = pcpu_mem_zalloc(new_size); |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 441 | if (!new) |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 442 | return -ENOMEM; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 443 | |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 444 | /* acquire pcpu_lock and switch to new area map */ |
| 445 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 446 | |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 447 | if (new_alloc <= chunk->map_alloc) |
| 448 | goto out_unlock; |
| 449 | |
| 450 | old_size = chunk->map_alloc * sizeof(chunk->map[0]); |
Huang Shijie | a002d14 | 2010-08-08 14:39:07 +0200 | [diff] [blame] | 451 | old = chunk->map; |
| 452 | |
| 453 | memcpy(new, old, old_size); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 454 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 455 | chunk->map_alloc = new_alloc; |
| 456 | chunk->map = new; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 457 | new = NULL; |
| 458 | |
| 459 | out_unlock: |
| 460 | spin_unlock_irqrestore(&pcpu_lock, flags); |
| 461 | |
| 462 | /* |
| 463 | * pcpu_mem_free() might end up calling vfree() which uses |
| 464 | * IRQ-unsafe lock and thus can't be called under pcpu_lock. |
| 465 | */ |
| 466 | pcpu_mem_free(old, old_size); |
| 467 | pcpu_mem_free(new, new_size); |
| 468 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 469 | return 0; |
| 470 | } |
| 471 | |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 472 | static void pcpu_map_extend_workfn(struct work_struct *work) |
| 473 | { |
| 474 | struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk, |
| 475 | map_extend_work); |
| 476 | int new_alloc; |
| 477 | |
| 478 | spin_lock_irq(&pcpu_lock); |
| 479 | new_alloc = pcpu_need_to_extend(chunk, false); |
| 480 | spin_unlock_irq(&pcpu_lock); |
| 481 | |
| 482 | if (new_alloc) |
| 483 | pcpu_extend_area_map(chunk, new_alloc); |
| 484 | } |
| 485 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 486 | /** |
Tejun Heo | a16037c | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 487 | * pcpu_fit_in_area - try to fit the requested allocation in a candidate area |
| 488 | * @chunk: chunk the candidate area belongs to |
| 489 | * @off: the offset to the start of the candidate area |
| 490 | * @this_size: the size of the candidate area |
| 491 | * @size: the size of the target allocation |
| 492 | * @align: the alignment of the target allocation |
| 493 | * @pop_only: only allocate from already populated region |
| 494 | * |
| 495 | * We're trying to allocate @size bytes aligned at @align. @chunk's area |
| 496 | * at @off sized @this_size is a candidate. This function determines |
| 497 | * whether the target allocation fits in the candidate area and returns the |
| 498 | * number of bytes to pad after @off. If the target area doesn't fit, -1 |
| 499 | * is returned. |
| 500 | * |
| 501 | * If @pop_only is %true, this function only considers the already |
| 502 | * populated part of the candidate area. |
| 503 | */ |
| 504 | static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size, |
| 505 | int size, int align, bool pop_only) |
| 506 | { |
| 507 | int cand_off = off; |
| 508 | |
| 509 | while (true) { |
| 510 | int head = ALIGN(cand_off, align) - off; |
| 511 | int page_start, page_end, rs, re; |
| 512 | |
| 513 | if (this_size < head + size) |
| 514 | return -1; |
| 515 | |
| 516 | if (!pop_only) |
| 517 | return head; |
| 518 | |
| 519 | /* |
| 520 | * If the first unpopulated page is beyond the end of the |
| 521 | * allocation, the whole allocation is populated; |
| 522 | * otherwise, retry from the end of the unpopulated area. |
| 523 | */ |
| 524 | page_start = PFN_DOWN(head + off); |
| 525 | page_end = PFN_UP(head + off + size); |
| 526 | |
| 527 | rs = page_start; |
| 528 | pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size)); |
| 529 | if (rs >= page_end) |
| 530 | return head; |
| 531 | cand_off = re * PAGE_SIZE; |
| 532 | } |
| 533 | } |
| 534 | |
| 535 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 536 | * pcpu_alloc_area - allocate area from a pcpu_chunk |
| 537 | * @chunk: chunk of interest |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 538 | * @size: wanted size in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 539 | * @align: wanted align |
Tejun Heo | a16037c | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 540 | * @pop_only: allocate only from the populated area |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 541 | * @occ_pages_p: out param for the number of pages the area occupies |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 542 | * |
| 543 | * Try to allocate @size bytes area aligned at @align from @chunk. |
| 544 | * Note that this function only allocates the offset. It doesn't |
| 545 | * populate or map the area. |
| 546 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 547 | * @chunk->map must have at least two free slots. |
| 548 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 549 | * CONTEXT: |
| 550 | * pcpu_lock. |
| 551 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 552 | * RETURNS: |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 553 | * Allocated offset in @chunk on success, -1 if no matching area is |
| 554 | * found. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 555 | */ |
Tejun Heo | a16037c | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 556 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align, |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 557 | bool pop_only, int *occ_pages_p) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 558 | { |
| 559 | int oslot = pcpu_chunk_slot(chunk); |
| 560 | int max_contig = 0; |
| 561 | int i, off; |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 562 | bool seen_free = false; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 563 | int *p; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 564 | |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 565 | for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 566 | int head, tail; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 567 | int this_size; |
| 568 | |
| 569 | off = *p; |
| 570 | if (off & 1) |
| 571 | continue; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 572 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 573 | this_size = (p[1] & ~1) - off; |
Tejun Heo | a16037c | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 574 | |
| 575 | head = pcpu_fit_in_area(chunk, off, this_size, size, align, |
| 576 | pop_only); |
| 577 | if (head < 0) { |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 578 | if (!seen_free) { |
| 579 | chunk->first_free = i; |
| 580 | seen_free = true; |
| 581 | } |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 582 | max_contig = max(this_size, max_contig); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 583 | continue; |
| 584 | } |
| 585 | |
| 586 | /* |
| 587 | * If head is small or the previous block is free, |
| 588 | * merge'em. Note that 'small' is defined as smaller |
| 589 | * than sizeof(int), which is very small but isn't too |
| 590 | * uncommon for percpu allocations. |
| 591 | */ |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 592 | if (head && (head < sizeof(int) || !(p[-1] & 1))) { |
Jianyu Zhan | 21ddfd3 | 2014-03-28 20:55:21 +0800 | [diff] [blame] | 593 | *p = off += head; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 594 | if (p[-1] & 1) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 595 | chunk->free_size -= head; |
Jianyu Zhan | 21ddfd3 | 2014-03-28 20:55:21 +0800 | [diff] [blame] | 596 | else |
| 597 | max_contig = max(*p - p[-1], max_contig); |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 598 | this_size -= head; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 599 | head = 0; |
| 600 | } |
| 601 | |
| 602 | /* if tail is small, just keep it around */ |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 603 | tail = this_size - head - size; |
| 604 | if (tail < sizeof(int)) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 605 | tail = 0; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 606 | size = this_size - head; |
| 607 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 608 | |
| 609 | /* split if warranted */ |
| 610 | if (head || tail) { |
Al Viro | 706c16f | 2014-03-06 21:08:24 -0500 | [diff] [blame] | 611 | int nr_extra = !!head + !!tail; |
| 612 | |
| 613 | /* insert new subblocks */ |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 614 | memmove(p + nr_extra + 1, p + 1, |
Al Viro | 706c16f | 2014-03-06 21:08:24 -0500 | [diff] [blame] | 615 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
| 616 | chunk->map_used += nr_extra; |
| 617 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 618 | if (head) { |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 619 | if (!seen_free) { |
| 620 | chunk->first_free = i; |
| 621 | seen_free = true; |
| 622 | } |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 623 | *++p = off += head; |
| 624 | ++i; |
Al Viro | 706c16f | 2014-03-06 21:08:24 -0500 | [diff] [blame] | 625 | max_contig = max(head, max_contig); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 626 | } |
Al Viro | 706c16f | 2014-03-06 21:08:24 -0500 | [diff] [blame] | 627 | if (tail) { |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 628 | p[1] = off + size; |
Al Viro | 706c16f | 2014-03-06 21:08:24 -0500 | [diff] [blame] | 629 | max_contig = max(tail, max_contig); |
| 630 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 631 | } |
| 632 | |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 633 | if (!seen_free) |
| 634 | chunk->first_free = i + 1; |
| 635 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 636 | /* update hint and mark allocated */ |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 637 | if (i + 1 == chunk->map_used) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 638 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 639 | else |
| 640 | chunk->contig_hint = max(chunk->contig_hint, |
| 641 | max_contig); |
| 642 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 643 | chunk->free_size -= size; |
| 644 | *p |= 1; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 645 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 646 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 647 | pcpu_chunk_relocate(chunk, oslot); |
| 648 | return off; |
| 649 | } |
| 650 | |
| 651 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 652 | pcpu_chunk_relocate(chunk, oslot); |
| 653 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 654 | /* tell the upper layer that this chunk has no matching area */ |
| 655 | return -1; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 656 | } |
| 657 | |
| 658 | /** |
| 659 | * pcpu_free_area - free area to a pcpu_chunk |
| 660 | * @chunk: chunk of interest |
| 661 | * @freeme: offset of area to free |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 662 | * @occ_pages_p: out param for the number of pages the area occupies |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 663 | * |
| 664 | * Free area starting from @freeme to @chunk. Note that this function |
| 665 | * only modifies the allocation map. It doesn't depopulate or unmap |
| 666 | * the area. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 667 | * |
| 668 | * CONTEXT: |
| 669 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 670 | */ |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 671 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme, |
| 672 | int *occ_pages_p) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 673 | { |
| 674 | int oslot = pcpu_chunk_slot(chunk); |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 675 | int off = 0; |
| 676 | unsigned i, j; |
| 677 | int to_free = 0; |
| 678 | int *p; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 679 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 680 | freeme |= 1; /* we are searching for <given offset, in use> pair */ |
| 681 | |
| 682 | i = 0; |
| 683 | j = chunk->map_used; |
| 684 | while (i != j) { |
| 685 | unsigned k = (i + j) / 2; |
| 686 | off = chunk->map[k]; |
| 687 | if (off < freeme) |
| 688 | i = k + 1; |
| 689 | else if (off > freeme) |
| 690 | j = k; |
| 691 | else |
| 692 | i = j = k; |
| 693 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 694 | BUG_ON(off != freeme); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 695 | |
Al Viro | 3d331ad | 2014-03-06 20:52:32 -0500 | [diff] [blame] | 696 | if (i < chunk->first_free) |
| 697 | chunk->first_free = i; |
| 698 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 699 | p = chunk->map + i; |
| 700 | *p = off &= ~1; |
| 701 | chunk->free_size += (p[1] & ~1) - off; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 702 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 703 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
| 704 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 705 | /* merge with next? */ |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 706 | if (!(p[1] & 1)) |
| 707 | to_free++; |
| 708 | /* merge with previous? */ |
| 709 | if (i > 0 && !(p[-1] & 1)) { |
| 710 | to_free++; |
| 711 | i--; |
| 712 | p--; |
| 713 | } |
| 714 | if (to_free) { |
| 715 | chunk->map_used -= to_free; |
| 716 | memmove(p + 1, p + 1 + to_free, |
| 717 | (chunk->map_used - i) * sizeof(chunk->map[0])); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 718 | } |
| 719 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 720 | chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 721 | pcpu_chunk_relocate(chunk, oslot); |
| 722 | } |
| 723 | |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 724 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
| 725 | { |
| 726 | struct pcpu_chunk *chunk; |
| 727 | |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 728 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 729 | if (!chunk) |
| 730 | return NULL; |
| 731 | |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 732 | chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * |
| 733 | sizeof(chunk->map[0])); |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 734 | if (!chunk->map) { |
Jianyu Zhan | 5a838c3 | 2014-04-14 13:47:40 +0800 | [diff] [blame] | 735 | pcpu_mem_free(chunk, pcpu_chunk_struct_size); |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 736 | return NULL; |
| 737 | } |
| 738 | |
| 739 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 740 | chunk->map[0] = 0; |
| 741 | chunk->map[1] = pcpu_unit_size | 1; |
| 742 | chunk->map_used = 1; |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 743 | |
| 744 | INIT_LIST_HEAD(&chunk->list); |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 745 | INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn); |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 746 | chunk->free_size = pcpu_unit_size; |
| 747 | chunk->contig_hint = pcpu_unit_size; |
| 748 | |
| 749 | return chunk; |
| 750 | } |
| 751 | |
| 752 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) |
| 753 | { |
| 754 | if (!chunk) |
| 755 | return; |
| 756 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
Joonsoo Kim | b4916cb | 2012-10-29 22:59:58 +0900 | [diff] [blame] | 757 | pcpu_mem_free(chunk, pcpu_chunk_struct_size); |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 758 | } |
| 759 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 760 | /** |
| 761 | * pcpu_chunk_populated - post-population bookkeeping |
| 762 | * @chunk: pcpu_chunk which got populated |
| 763 | * @page_start: the start page |
| 764 | * @page_end: the end page |
| 765 | * |
| 766 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update |
| 767 | * the bookkeeping information accordingly. Must be called after each |
| 768 | * successful population. |
| 769 | */ |
| 770 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, |
| 771 | int page_start, int page_end) |
| 772 | { |
| 773 | int nr = page_end - page_start; |
| 774 | |
| 775 | lockdep_assert_held(&pcpu_lock); |
| 776 | |
| 777 | bitmap_set(chunk->populated, page_start, nr); |
| 778 | chunk->nr_populated += nr; |
| 779 | pcpu_nr_empty_pop_pages += nr; |
| 780 | } |
| 781 | |
| 782 | /** |
| 783 | * pcpu_chunk_depopulated - post-depopulation bookkeeping |
| 784 | * @chunk: pcpu_chunk which got depopulated |
| 785 | * @page_start: the start page |
| 786 | * @page_end: the end page |
| 787 | * |
| 788 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. |
| 789 | * Update the bookkeeping information accordingly. Must be called after |
| 790 | * each successful depopulation. |
| 791 | */ |
| 792 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, |
| 793 | int page_start, int page_end) |
| 794 | { |
| 795 | int nr = page_end - page_start; |
| 796 | |
| 797 | lockdep_assert_held(&pcpu_lock); |
| 798 | |
| 799 | bitmap_clear(chunk->populated, page_start, nr); |
| 800 | chunk->nr_populated -= nr; |
| 801 | pcpu_nr_empty_pop_pages -= nr; |
| 802 | } |
| 803 | |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 804 | /* |
| 805 | * Chunk management implementation. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 806 | * |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 807 | * To allow different implementations, chunk alloc/free and |
| 808 | * [de]population are implemented in a separate file which is pulled |
| 809 | * into this file and compiled together. The following functions |
| 810 | * should be implemented. |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 811 | * |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 812 | * pcpu_populate_chunk - populate the specified range of a chunk |
| 813 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk |
| 814 | * pcpu_create_chunk - create a new chunk |
| 815 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop |
| 816 | * pcpu_addr_to_page - translate address to physical address |
| 817 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 818 | */ |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 819 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
| 820 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); |
| 821 | static struct pcpu_chunk *pcpu_create_chunk(void); |
| 822 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); |
| 823 | static struct page *pcpu_addr_to_page(void *addr); |
| 824 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 825 | |
Tejun Heo | b0c9778 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 826 | #ifdef CONFIG_NEED_PER_CPU_KM |
| 827 | #include "percpu-km.c" |
| 828 | #else |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 829 | #include "percpu-vm.c" |
Tejun Heo | b0c9778 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 830 | #endif |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 831 | |
| 832 | /** |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 833 | * pcpu_chunk_addr_search - determine chunk containing specified address |
| 834 | * @addr: address for which the chunk needs to be determined. |
| 835 | * |
| 836 | * RETURNS: |
| 837 | * The address of the found chunk. |
| 838 | */ |
| 839 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) |
| 840 | { |
| 841 | /* is it in the first chunk? */ |
| 842 | if (pcpu_addr_in_first_chunk(addr)) { |
| 843 | /* is it in the reserved area? */ |
| 844 | if (pcpu_addr_in_reserved_chunk(addr)) |
| 845 | return pcpu_reserved_chunk; |
| 846 | return pcpu_first_chunk; |
| 847 | } |
| 848 | |
| 849 | /* |
| 850 | * The address is relative to unit0 which might be unused and |
| 851 | * thus unmapped. Offset the address to the unit space of the |
| 852 | * current processor before looking it up in the vmalloc |
| 853 | * space. Note that any possible cpu id can be used here, so |
| 854 | * there's no need to worry about preemption or cpu hotplug. |
| 855 | */ |
| 856 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 857 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
Tejun Heo | 88999a8 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 858 | } |
| 859 | |
| 860 | /** |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 861 | * pcpu_alloc - the percpu allocator |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 862 | * @size: size of area to allocate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 863 | * @align: alignment of area (max PAGE_SIZE) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 864 | * @reserved: allocate from the reserved chunk if available |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 865 | * @gfp: allocation flags |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 866 | * |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 867 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
| 868 | * contain %GFP_KERNEL, the allocation is atomic. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 869 | * |
| 870 | * RETURNS: |
| 871 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 872 | */ |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 873 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
| 874 | gfp_t gfp) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 875 | { |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 876 | static int warn_limit = 10; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 877 | struct pcpu_chunk *chunk; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 878 | const char *err; |
Tejun Heo | 6ae833c | 2014-10-08 12:01:52 -0400 | [diff] [blame] | 879 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 880 | int occ_pages = 0; |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 881 | int slot, off, new_alloc, cpu, ret; |
Jiri Kosina | 403a91b | 2009-10-29 00:25:59 +0900 | [diff] [blame] | 882 | unsigned long flags; |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 883 | void __percpu *ptr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 884 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 885 | /* |
| 886 | * We want the lowest bit of offset available for in-use/free |
Viro | 2f69fa8 | 2014-03-17 16:01:27 -0400 | [diff] [blame] | 887 | * indicator, so force >= 16bit alignment and make size even. |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 888 | */ |
| 889 | if (unlikely(align < 2)) |
| 890 | align = 2; |
| 891 | |
Christoph Lameter | fb009e3 | 2014-06-19 09:59:18 -0500 | [diff] [blame] | 892 | size = ALIGN(size, 2); |
Viro | 2f69fa8 | 2014-03-17 16:01:27 -0400 | [diff] [blame] | 893 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 894 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 895 | WARN(true, "illegal size (%zu) or align (%zu) for " |
| 896 | "percpu allocation\n", size, align); |
| 897 | return NULL; |
| 898 | } |
| 899 | |
Jiri Kosina | 403a91b | 2009-10-29 00:25:59 +0900 | [diff] [blame] | 900 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 901 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 902 | /* serve reserved allocations from the reserved chunk if available */ |
| 903 | if (reserved && pcpu_reserved_chunk) { |
| 904 | chunk = pcpu_reserved_chunk; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 905 | |
| 906 | if (size > chunk->contig_hint) { |
| 907 | err = "alloc from reserved chunk failed"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 908 | goto fail_unlock; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 909 | } |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 910 | |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 911 | while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) { |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 912 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 913 | if (is_atomic || |
| 914 | pcpu_extend_area_map(chunk, new_alloc) < 0) { |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 915 | err = "failed to extend area map of reserved chunk"; |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 916 | goto fail; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 917 | } |
| 918 | spin_lock_irqsave(&pcpu_lock, flags); |
| 919 | } |
| 920 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 921 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
| 922 | &occ_pages); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 923 | if (off >= 0) |
| 924 | goto area_found; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 925 | |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 926 | err = "alloc from reserved chunk failed"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 927 | goto fail_unlock; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 928 | } |
| 929 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 930 | restart: |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 931 | /* search through normal chunks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 932 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
| 933 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 934 | if (size > chunk->contig_hint) |
| 935 | continue; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 936 | |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 937 | new_alloc = pcpu_need_to_extend(chunk, is_atomic); |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 938 | if (new_alloc) { |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 939 | if (is_atomic) |
| 940 | continue; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 941 | spin_unlock_irqrestore(&pcpu_lock, flags); |
| 942 | if (pcpu_extend_area_map(chunk, |
| 943 | new_alloc) < 0) { |
| 944 | err = "failed to extend area map"; |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 945 | goto fail; |
Tejun Heo | 833af84 | 2009-11-11 15:35:18 +0900 | [diff] [blame] | 946 | } |
| 947 | spin_lock_irqsave(&pcpu_lock, flags); |
| 948 | /* |
| 949 | * pcpu_lock has been dropped, need to |
| 950 | * restart cpu_slot list walking. |
| 951 | */ |
| 952 | goto restart; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 953 | } |
| 954 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 955 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
| 956 | &occ_pages); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 957 | if (off >= 0) |
| 958 | goto area_found; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 959 | } |
| 960 | } |
| 961 | |
Jiri Kosina | 403a91b | 2009-10-29 00:25:59 +0900 | [diff] [blame] | 962 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 963 | |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 964 | /* |
| 965 | * No space left. Create a new chunk. We don't want multiple |
| 966 | * tasks to create chunks simultaneously. Serialize and create iff |
| 967 | * there's still no empty chunk after grabbing the mutex. |
| 968 | */ |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 969 | if (is_atomic) |
| 970 | goto fail; |
| 971 | |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 972 | mutex_lock(&pcpu_alloc_mutex); |
| 973 | |
| 974 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
| 975 | chunk = pcpu_create_chunk(); |
| 976 | if (!chunk) { |
Tejun Heo | 23cb898 | 2014-09-09 08:02:45 +0900 | [diff] [blame] | 977 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 978 | err = "failed to allocate new chunk"; |
| 979 | goto fail; |
| 980 | } |
| 981 | |
| 982 | spin_lock_irqsave(&pcpu_lock, flags); |
| 983 | pcpu_chunk_relocate(chunk, -1); |
| 984 | } else { |
| 985 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 986 | } |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 987 | |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 988 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 989 | goto restart; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 990 | |
| 991 | area_found: |
Jiri Kosina | 403a91b | 2009-10-29 00:25:59 +0900 | [diff] [blame] | 992 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 993 | |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 994 | /* populate if not all pages are already there */ |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 995 | if (!is_atomic) { |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 996 | int page_start, page_end, rs, re; |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 997 | |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 998 | mutex_lock(&pcpu_alloc_mutex); |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 999 | |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1000 | page_start = PFN_DOWN(off); |
| 1001 | page_end = PFN_UP(off + size); |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 1002 | |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1003 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
| 1004 | WARN_ON(chunk->immutable); |
| 1005 | |
| 1006 | ret = pcpu_populate_chunk(chunk, rs, re); |
| 1007 | |
| 1008 | spin_lock_irqsave(&pcpu_lock, flags); |
| 1009 | if (ret) { |
| 1010 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1011 | pcpu_free_area(chunk, off, &occ_pages); |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1012 | err = "failed to populate"; |
| 1013 | goto fail_unlock; |
| 1014 | } |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1015 | pcpu_chunk_populated(chunk, rs, re); |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1016 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 1017 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1018 | |
Tejun Heo | e04d320 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1019 | mutex_unlock(&pcpu_alloc_mutex); |
| 1020 | } |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1021 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1022 | if (chunk != pcpu_reserved_chunk) |
| 1023 | pcpu_nr_empty_pop_pages -= occ_pages; |
| 1024 | |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1025 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
| 1026 | pcpu_schedule_balance_work(); |
| 1027 | |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 1028 | /* clear the areas and return address relative to base address */ |
| 1029 | for_each_possible_cpu(cpu) |
| 1030 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); |
| 1031 | |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 1032 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
Larry Finger | 8a8c35f | 2015-06-24 16:58:51 -0700 | [diff] [blame] | 1033 | kmemleak_alloc_percpu(ptr, size, gfp); |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 1034 | return ptr; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1035 | |
| 1036 | fail_unlock: |
Jiri Kosina | 403a91b | 2009-10-29 00:25:59 +0900 | [diff] [blame] | 1037 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | b38d08f | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 1038 | fail: |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1039 | if (!is_atomic && warn_limit) { |
| 1040 | pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
| 1041 | size, align, is_atomic, err); |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1042 | dump_stack(); |
| 1043 | if (!--warn_limit) |
| 1044 | pr_info("PERCPU: limit reached, disable warning\n"); |
| 1045 | } |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1046 | if (is_atomic) { |
| 1047 | /* see the flag handling in pcpu_blance_workfn() */ |
| 1048 | pcpu_atomic_alloc_failed = true; |
| 1049 | pcpu_schedule_balance_work(); |
| 1050 | } |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1051 | return NULL; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1052 | } |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1053 | |
| 1054 | /** |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1055 | * __alloc_percpu_gfp - allocate dynamic percpu area |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1056 | * @size: size of area to allocate in bytes |
| 1057 | * @align: alignment of area (max PAGE_SIZE) |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1058 | * @gfp: allocation flags |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1059 | * |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1060 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
| 1061 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can |
| 1062 | * be called from any context but is a lot more likely to fail. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1063 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1064 | * RETURNS: |
| 1065 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 1066 | */ |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1067 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
| 1068 | { |
| 1069 | return pcpu_alloc(size, align, false, gfp); |
| 1070 | } |
| 1071 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); |
| 1072 | |
| 1073 | /** |
| 1074 | * __alloc_percpu - allocate dynamic percpu area |
| 1075 | * @size: size of area to allocate in bytes |
| 1076 | * @align: alignment of area (max PAGE_SIZE) |
| 1077 | * |
| 1078 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). |
| 1079 | */ |
Tejun Heo | 43cf38e | 2010-02-02 14:38:57 +0900 | [diff] [blame] | 1080 | void __percpu *__alloc_percpu(size_t size, size_t align) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1081 | { |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1082 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1083 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1084 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
| 1085 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1086 | /** |
| 1087 | * __alloc_reserved_percpu - allocate reserved percpu area |
| 1088 | * @size: size of area to allocate in bytes |
| 1089 | * @align: alignment of area (max PAGE_SIZE) |
| 1090 | * |
Tejun Heo | 9329ba9 | 2010-09-10 11:01:56 +0200 | [diff] [blame] | 1091 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
| 1092 | * from reserved percpu area if arch has set it up; otherwise, |
| 1093 | * allocation is served from the same dynamic area. Might sleep. |
| 1094 | * Might trigger writeouts. |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1095 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1096 | * CONTEXT: |
| 1097 | * Does GFP_KERNEL allocation. |
| 1098 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1099 | * RETURNS: |
| 1100 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 1101 | */ |
Tejun Heo | 43cf38e | 2010-02-02 14:38:57 +0900 | [diff] [blame] | 1102 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1103 | { |
Tejun Heo | 5835d96 | 2014-09-02 14:46:04 -0400 | [diff] [blame] | 1104 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1105 | } |
| 1106 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1107 | /** |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1108 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1109 | * @work: unused |
| 1110 | * |
| 1111 | * Reclaim all fully free chunks except for the first one. |
| 1112 | */ |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1113 | static void pcpu_balance_workfn(struct work_struct *work) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1114 | { |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1115 | LIST_HEAD(to_free); |
| 1116 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1117 | struct pcpu_chunk *chunk, *next; |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1118 | int slot, nr_to_pop, ret; |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1119 | |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1120 | /* |
| 1121 | * There's no reason to keep around multiple unused chunks and VM |
| 1122 | * areas can be scarce. Destroy all free chunks except for one. |
| 1123 | */ |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1124 | mutex_lock(&pcpu_alloc_mutex); |
| 1125 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1126 | |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1127 | list_for_each_entry_safe(chunk, next, free_head, list) { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1128 | WARN_ON(chunk->immutable); |
| 1129 | |
| 1130 | /* spare the first one */ |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1131 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1132 | continue; |
| 1133 | |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1134 | list_move(&chunk->list, &to_free); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1135 | } |
| 1136 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1137 | spin_unlock_irq(&pcpu_lock); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1138 | |
Tejun Heo | fe6bd8c | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1139 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
Tejun Heo | a93ace4 | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 1140 | int rs, re; |
Tejun Heo | dca4964 | 2014-09-02 14:46:01 -0400 | [diff] [blame] | 1141 | |
Tejun Heo | a93ace4 | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 1142 | pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) { |
| 1143 | pcpu_depopulate_chunk(chunk, rs, re); |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1144 | spin_lock_irq(&pcpu_lock); |
| 1145 | pcpu_chunk_depopulated(chunk, rs, re); |
| 1146 | spin_unlock_irq(&pcpu_lock); |
Tejun Heo | a93ace4 | 2014-09-02 14:46:02 -0400 | [diff] [blame] | 1147 | } |
Tejun Heo | 6081089 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 1148 | pcpu_destroy_chunk(chunk); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1149 | } |
Tejun Heo | 971f391 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1150 | |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1151 | /* |
| 1152 | * Ensure there are certain number of free populated pages for |
| 1153 | * atomic allocs. Fill up from the most packed so that atomic |
| 1154 | * allocs don't increase fragmentation. If atomic allocation |
| 1155 | * failed previously, always populate the maximum amount. This |
| 1156 | * should prevent atomic allocs larger than PAGE_SIZE from keeping |
| 1157 | * failing indefinitely; however, large atomic allocs are not |
| 1158 | * something we support properly and can be highly unreliable and |
| 1159 | * inefficient. |
| 1160 | */ |
| 1161 | retry_pop: |
| 1162 | if (pcpu_atomic_alloc_failed) { |
| 1163 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; |
| 1164 | /* best effort anyway, don't worry about synchronization */ |
| 1165 | pcpu_atomic_alloc_failed = false; |
| 1166 | } else { |
| 1167 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - |
| 1168 | pcpu_nr_empty_pop_pages, |
| 1169 | 0, PCPU_EMPTY_POP_PAGES_HIGH); |
| 1170 | } |
| 1171 | |
| 1172 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { |
| 1173 | int nr_unpop = 0, rs, re; |
| 1174 | |
| 1175 | if (!nr_to_pop) |
| 1176 | break; |
| 1177 | |
| 1178 | spin_lock_irq(&pcpu_lock); |
| 1179 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 1180 | nr_unpop = pcpu_unit_pages - chunk->nr_populated; |
| 1181 | if (nr_unpop) |
| 1182 | break; |
| 1183 | } |
| 1184 | spin_unlock_irq(&pcpu_lock); |
| 1185 | |
| 1186 | if (!nr_unpop) |
| 1187 | continue; |
| 1188 | |
| 1189 | /* @chunk can't go away while pcpu_alloc_mutex is held */ |
| 1190 | pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) { |
| 1191 | int nr = min(re - rs, nr_to_pop); |
| 1192 | |
| 1193 | ret = pcpu_populate_chunk(chunk, rs, rs + nr); |
| 1194 | if (!ret) { |
| 1195 | nr_to_pop -= nr; |
| 1196 | spin_lock_irq(&pcpu_lock); |
| 1197 | pcpu_chunk_populated(chunk, rs, rs + nr); |
| 1198 | spin_unlock_irq(&pcpu_lock); |
| 1199 | } else { |
| 1200 | nr_to_pop = 0; |
| 1201 | } |
| 1202 | |
| 1203 | if (!nr_to_pop) |
| 1204 | break; |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | if (nr_to_pop) { |
| 1209 | /* ran out of chunks to populate, create a new one and retry */ |
| 1210 | chunk = pcpu_create_chunk(); |
| 1211 | if (chunk) { |
| 1212 | spin_lock_irq(&pcpu_lock); |
| 1213 | pcpu_chunk_relocate(chunk, -1); |
| 1214 | spin_unlock_irq(&pcpu_lock); |
| 1215 | goto retry_pop; |
| 1216 | } |
| 1217 | } |
| 1218 | |
Tejun Heo | 971f391 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1219 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1220 | } |
| 1221 | |
| 1222 | /** |
| 1223 | * free_percpu - free percpu area |
| 1224 | * @ptr: pointer to area to free |
| 1225 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1226 | * Free percpu area @ptr. |
| 1227 | * |
| 1228 | * CONTEXT: |
| 1229 | * Can be called from atomic context. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1230 | */ |
Tejun Heo | 43cf38e | 2010-02-02 14:38:57 +0900 | [diff] [blame] | 1231 | void free_percpu(void __percpu *ptr) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1232 | { |
Andrew Morton | 129182e | 2010-01-08 14:42:39 -0800 | [diff] [blame] | 1233 | void *addr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1234 | struct pcpu_chunk *chunk; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1235 | unsigned long flags; |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1236 | int off, occ_pages; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1237 | |
| 1238 | if (!ptr) |
| 1239 | return; |
| 1240 | |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 1241 | kmemleak_free_percpu(ptr); |
| 1242 | |
Andrew Morton | 129182e | 2010-01-08 14:42:39 -0800 | [diff] [blame] | 1243 | addr = __pcpu_ptr_to_addr(ptr); |
| 1244 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1245 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1246 | |
| 1247 | chunk = pcpu_chunk_addr_search(addr); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1248 | off = addr - chunk->base_addr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1249 | |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1250 | pcpu_free_area(chunk, off, &occ_pages); |
| 1251 | |
| 1252 | if (chunk != pcpu_reserved_chunk) |
| 1253 | pcpu_nr_empty_pop_pages += occ_pages; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1254 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1255 | /* if there are more than one fully free chunks, wake up grim reaper */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1256 | if (chunk->free_size == pcpu_unit_size) { |
| 1257 | struct pcpu_chunk *pos; |
| 1258 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1259 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1260 | if (pos != chunk) { |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1261 | pcpu_schedule_balance_work(); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1262 | break; |
| 1263 | } |
| 1264 | } |
| 1265 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1266 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1267 | } |
| 1268 | EXPORT_SYMBOL_GPL(free_percpu); |
| 1269 | |
Vivek Goyal | 3b034b0 | 2009-11-24 15:50:03 +0900 | [diff] [blame] | 1270 | /** |
Tejun Heo | 10fad5e | 2010-03-10 18:57:54 +0900 | [diff] [blame] | 1271 | * is_kernel_percpu_address - test whether address is from static percpu area |
| 1272 | * @addr: address to test |
| 1273 | * |
| 1274 | * Test whether @addr belongs to in-kernel static percpu area. Module |
| 1275 | * static percpu areas are not considered. For those, use |
| 1276 | * is_module_percpu_address(). |
| 1277 | * |
| 1278 | * RETURNS: |
| 1279 | * %true if @addr is from in-kernel static percpu area, %false otherwise. |
| 1280 | */ |
| 1281 | bool is_kernel_percpu_address(unsigned long addr) |
| 1282 | { |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 1283 | #ifdef CONFIG_SMP |
Tejun Heo | 10fad5e | 2010-03-10 18:57:54 +0900 | [diff] [blame] | 1284 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
| 1285 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
| 1286 | unsigned int cpu; |
| 1287 | |
| 1288 | for_each_possible_cpu(cpu) { |
| 1289 | void *start = per_cpu_ptr(base, cpu); |
| 1290 | |
| 1291 | if ((void *)addr >= start && (void *)addr < start + static_size) |
| 1292 | return true; |
| 1293 | } |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 1294 | #endif |
| 1295 | /* on UP, can't distinguish from other static vars, always false */ |
Tejun Heo | 10fad5e | 2010-03-10 18:57:54 +0900 | [diff] [blame] | 1296 | return false; |
| 1297 | } |
| 1298 | |
| 1299 | /** |
Vivek Goyal | 3b034b0 | 2009-11-24 15:50:03 +0900 | [diff] [blame] | 1300 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address |
| 1301 | * @addr: the address to be converted to physical address |
| 1302 | * |
| 1303 | * Given @addr which is dereferenceable address obtained via one of |
| 1304 | * percpu access macros, this function translates it into its physical |
| 1305 | * address. The caller is responsible for ensuring @addr stays valid |
| 1306 | * until this function finishes. |
| 1307 | * |
Dave Young | 67589c7 | 2011-11-23 08:20:53 -0800 | [diff] [blame] | 1308 | * percpu allocator has special setup for the first chunk, which currently |
| 1309 | * supports either embedding in linear address space or vmalloc mapping, |
| 1310 | * and, from the second one, the backing allocator (currently either vm or |
| 1311 | * km) provides translation. |
| 1312 | * |
Yannick Guerrini | bffc437 | 2015-03-06 23:30:42 +0100 | [diff] [blame] | 1313 | * The addr can be translated simply without checking if it falls into the |
Dave Young | 67589c7 | 2011-11-23 08:20:53 -0800 | [diff] [blame] | 1314 | * first chunk. But the current code reflects better how percpu allocator |
| 1315 | * actually works, and the verification can discover both bugs in percpu |
| 1316 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current |
| 1317 | * code. |
| 1318 | * |
Vivek Goyal | 3b034b0 | 2009-11-24 15:50:03 +0900 | [diff] [blame] | 1319 | * RETURNS: |
| 1320 | * The physical address for @addr. |
| 1321 | */ |
| 1322 | phys_addr_t per_cpu_ptr_to_phys(void *addr) |
| 1323 | { |
Tejun Heo | 9983b6f0 | 2010-06-18 11:44:31 +0200 | [diff] [blame] | 1324 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
| 1325 | bool in_first_chunk = false; |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 1326 | unsigned long first_low, first_high; |
Tejun Heo | 9983b6f0 | 2010-06-18 11:44:31 +0200 | [diff] [blame] | 1327 | unsigned int cpu; |
| 1328 | |
| 1329 | /* |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 1330 | * The following test on unit_low/high isn't strictly |
Tejun Heo | 9983b6f0 | 2010-06-18 11:44:31 +0200 | [diff] [blame] | 1331 | * necessary but will speed up lookups of addresses which |
| 1332 | * aren't in the first chunk. |
| 1333 | */ |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 1334 | first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0); |
| 1335 | first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu, |
| 1336 | pcpu_unit_pages); |
| 1337 | if ((unsigned long)addr >= first_low && |
| 1338 | (unsigned long)addr < first_high) { |
Tejun Heo | 9983b6f0 | 2010-06-18 11:44:31 +0200 | [diff] [blame] | 1339 | for_each_possible_cpu(cpu) { |
| 1340 | void *start = per_cpu_ptr(base, cpu); |
| 1341 | |
| 1342 | if (addr >= start && addr < start + pcpu_unit_size) { |
| 1343 | in_first_chunk = true; |
| 1344 | break; |
| 1345 | } |
| 1346 | } |
| 1347 | } |
| 1348 | |
| 1349 | if (in_first_chunk) { |
David Howells | eac522e | 2011-03-28 12:53:29 +0100 | [diff] [blame] | 1350 | if (!is_vmalloc_addr(addr)) |
Tejun Heo | 020ec65 | 2010-04-09 18:57:00 +0900 | [diff] [blame] | 1351 | return __pa(addr); |
| 1352 | else |
Eugene Surovegin | 9f57bd4 | 2011-12-15 11:25:59 -0800 | [diff] [blame] | 1353 | return page_to_phys(vmalloc_to_page(addr)) + |
| 1354 | offset_in_page(addr); |
Tejun Heo | 020ec65 | 2010-04-09 18:57:00 +0900 | [diff] [blame] | 1355 | } else |
Eugene Surovegin | 9f57bd4 | 2011-12-15 11:25:59 -0800 | [diff] [blame] | 1356 | return page_to_phys(pcpu_addr_to_page(addr)) + |
| 1357 | offset_in_page(addr); |
Vivek Goyal | 3b034b0 | 2009-11-24 15:50:03 +0900 | [diff] [blame] | 1358 | } |
| 1359 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1360 | /** |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1361 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
| 1362 | * @nr_groups: the number of groups |
| 1363 | * @nr_units: the number of units |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1364 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1365 | * Allocate ai which is large enough for @nr_groups groups containing |
| 1366 | * @nr_units units. The returned ai's groups[0].cpu_map points to the |
| 1367 | * cpu_map array which is long enough for @nr_units and filled with |
| 1368 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map |
| 1369 | * pointer of other groups. |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1370 | * |
| 1371 | * RETURNS: |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1372 | * Pointer to the allocated pcpu_alloc_info on success, NULL on |
| 1373 | * failure. |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1374 | */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1375 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, |
| 1376 | int nr_units) |
| 1377 | { |
| 1378 | struct pcpu_alloc_info *ai; |
| 1379 | size_t base_size, ai_size; |
| 1380 | void *ptr; |
| 1381 | int unit; |
| 1382 | |
| 1383 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), |
| 1384 | __alignof__(ai->groups[0].cpu_map[0])); |
| 1385 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); |
| 1386 | |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1387 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1388 | if (!ptr) |
| 1389 | return NULL; |
| 1390 | ai = ptr; |
| 1391 | ptr += base_size; |
| 1392 | |
| 1393 | ai->groups[0].cpu_map = ptr; |
| 1394 | |
| 1395 | for (unit = 0; unit < nr_units; unit++) |
| 1396 | ai->groups[0].cpu_map[unit] = NR_CPUS; |
| 1397 | |
| 1398 | ai->nr_groups = nr_groups; |
| 1399 | ai->__ai_size = PFN_ALIGN(ai_size); |
| 1400 | |
| 1401 | return ai; |
| 1402 | } |
| 1403 | |
| 1404 | /** |
| 1405 | * pcpu_free_alloc_info - free percpu allocation info |
| 1406 | * @ai: pcpu_alloc_info to free |
| 1407 | * |
| 1408 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). |
| 1409 | */ |
| 1410 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) |
| 1411 | { |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1412 | memblock_free_early(__pa(ai), ai->__ai_size); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1413 | } |
| 1414 | |
| 1415 | /** |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1416 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info |
| 1417 | * @lvl: loglevel |
| 1418 | * @ai: allocation info to dump |
| 1419 | * |
| 1420 | * Print out information about @ai using loglevel @lvl. |
| 1421 | */ |
| 1422 | static void pcpu_dump_alloc_info(const char *lvl, |
| 1423 | const struct pcpu_alloc_info *ai) |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1424 | { |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1425 | int group_width = 1, cpu_width = 1, width; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1426 | char empty_str[] = "--------"; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1427 | int alloc = 0, alloc_end = 0; |
| 1428 | int group, v; |
| 1429 | int upa, apl; /* units per alloc, allocs per line */ |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1430 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1431 | v = ai->nr_groups; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1432 | while (v /= 10) |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1433 | group_width++; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1434 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1435 | v = num_possible_cpus(); |
| 1436 | while (v /= 10) |
| 1437 | cpu_width++; |
| 1438 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1439 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1440 | upa = ai->alloc_size / ai->unit_size; |
| 1441 | width = upa * (cpu_width + 1) + group_width + 3; |
| 1442 | apl = rounddown_pow_of_two(max(60 / width, 1)); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1443 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1444 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
| 1445 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, |
| 1446 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); |
| 1447 | |
| 1448 | for (group = 0; group < ai->nr_groups; group++) { |
| 1449 | const struct pcpu_group_info *gi = &ai->groups[group]; |
| 1450 | int unit = 0, unit_end = 0; |
| 1451 | |
| 1452 | BUG_ON(gi->nr_units % upa); |
| 1453 | for (alloc_end += gi->nr_units / upa; |
| 1454 | alloc < alloc_end; alloc++) { |
| 1455 | if (!(alloc % apl)) { |
Tejun Heo | cb12982 | 2012-03-29 09:45:58 -0700 | [diff] [blame] | 1456 | printk(KERN_CONT "\n"); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1457 | printk("%spcpu-alloc: ", lvl); |
| 1458 | } |
Tejun Heo | cb12982 | 2012-03-29 09:45:58 -0700 | [diff] [blame] | 1459 | printk(KERN_CONT "[%0*d] ", group_width, group); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1460 | |
| 1461 | for (unit_end += upa; unit < unit_end; unit++) |
| 1462 | if (gi->cpu_map[unit] != NR_CPUS) |
Tejun Heo | cb12982 | 2012-03-29 09:45:58 -0700 | [diff] [blame] | 1463 | printk(KERN_CONT "%0*d ", cpu_width, |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1464 | gi->cpu_map[unit]); |
| 1465 | else |
Tejun Heo | cb12982 | 2012-03-29 09:45:58 -0700 | [diff] [blame] | 1466 | printk(KERN_CONT "%s ", empty_str); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1467 | } |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1468 | } |
Tejun Heo | cb12982 | 2012-03-29 09:45:58 -0700 | [diff] [blame] | 1469 | printk(KERN_CONT "\n"); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1470 | } |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1471 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1472 | /** |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1473 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1474 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1475 | * @base_addr: mapped address |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1476 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1477 | * Initialize the first percpu chunk which contains the kernel static |
| 1478 | * perpcu area. This function is to be called from arch percpu area |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1479 | * setup path. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1480 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1481 | * @ai contains all information necessary to initialize the first |
| 1482 | * chunk and prime the dynamic percpu allocator. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1483 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1484 | * @ai->static_size is the size of static percpu area. |
| 1485 | * |
| 1486 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1487 | * reserve after the static area in the first chunk. This reserves |
| 1488 | * the first chunk such that it's available only through reserved |
| 1489 | * percpu allocation. This is primarily used to serve module percpu |
| 1490 | * static areas on architectures where the addressing model has |
| 1491 | * limited offset range for symbol relocations to guarantee module |
| 1492 | * percpu symbols fall inside the relocatable range. |
| 1493 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1494 | * @ai->dyn_size determines the number of bytes available for dynamic |
| 1495 | * allocation in the first chunk. The area between @ai->static_size + |
| 1496 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. |
Tejun Heo | 6074d5b | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1497 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1498 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
| 1499 | * and equal to or larger than @ai->static_size + @ai->reserved_size + |
| 1500 | * @ai->dyn_size. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1501 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1502 | * @ai->atom_size is the allocation atom size and used as alignment |
| 1503 | * for vm areas. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1504 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1505 | * @ai->alloc_size is the allocation size and always multiple of |
| 1506 | * @ai->atom_size. This is larger than @ai->atom_size if |
| 1507 | * @ai->unit_size is larger than @ai->atom_size. |
| 1508 | * |
| 1509 | * @ai->nr_groups and @ai->groups describe virtual memory layout of |
| 1510 | * percpu areas. Units which should be colocated are put into the |
| 1511 | * same group. Dynamic VM areas will be allocated according to these |
| 1512 | * groupings. If @ai->nr_groups is zero, a single group containing |
| 1513 | * all units is assumed. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1514 | * |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1515 | * The caller should have mapped the first chunk at @base_addr and |
| 1516 | * copied static data to each unit. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1517 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1518 | * If the first chunk ends up with both reserved and dynamic areas, it |
| 1519 | * is served by two chunks - one to serve the core static and reserved |
| 1520 | * areas and the other for the dynamic area. They share the same vm |
| 1521 | * and page map but uses different area allocation map to stay away |
| 1522 | * from each other. The latter chunk is circulated in the chunk slots |
| 1523 | * and available for dynamic allocation like any other chunks. |
| 1524 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1525 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1526 | * 0 on success, -errno on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1527 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1528 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
| 1529 | void *base_addr) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1530 | { |
Tejun Heo | 099a19d | 2010-06-27 18:50:00 +0200 | [diff] [blame] | 1531 | static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; |
| 1532 | static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1533 | size_t dyn_size = ai->dyn_size; |
| 1534 | size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1535 | struct pcpu_chunk *schunk, *dchunk = NULL; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1536 | unsigned long *group_offsets; |
| 1537 | size_t *group_sizes; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1538 | unsigned long *unit_off; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1539 | unsigned int cpu; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1540 | int *unit_map; |
| 1541 | int group, unit, i; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1542 | |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1543 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
| 1544 | if (unlikely(cond)) { \ |
| 1545 | pr_emerg("PERCPU: failed to initialize, %s", #cond); \ |
Tejun Heo | 807de07 | 2015-02-13 14:37:34 -0800 | [diff] [blame] | 1546 | pr_emerg("PERCPU: cpu_possible_mask=%*pb\n", \ |
| 1547 | cpumask_pr_args(cpu_possible_mask)); \ |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1548 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
| 1549 | BUG(); \ |
| 1550 | } \ |
| 1551 | } while (0) |
| 1552 | |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1553 | /* sanity checks */ |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1554 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 1555 | #ifdef CONFIG_SMP |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1556 | PCPU_SETUP_BUG_ON(!ai->static_size); |
Tejun Heo | 0415b00d1 | 2011-03-24 18:50:09 +0100 | [diff] [blame] | 1557 | PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK); |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 1558 | #endif |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1559 | PCPU_SETUP_BUG_ON(!base_addr); |
Tejun Heo | 0415b00d1 | 2011-03-24 18:50:09 +0100 | [diff] [blame] | 1560 | PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK); |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1561 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
| 1562 | PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); |
| 1563 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
Tejun Heo | 099a19d | 2010-06-27 18:50:00 +0200 | [diff] [blame] | 1564 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
Tejun Heo | 9f64553 | 2010-04-09 18:57:01 +0900 | [diff] [blame] | 1565 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1566 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1567 | /* process group information and build config tables accordingly */ |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1568 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
| 1569 | sizeof(group_offsets[0]), 0); |
| 1570 | group_sizes = memblock_virt_alloc(ai->nr_groups * |
| 1571 | sizeof(group_sizes[0]), 0); |
| 1572 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); |
| 1573 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1574 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1575 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
Tejun Heo | ffe0d5a | 2009-09-29 09:17:56 +0900 | [diff] [blame] | 1576 | unit_map[cpu] = UINT_MAX; |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 1577 | |
| 1578 | pcpu_low_unit_cpu = NR_CPUS; |
| 1579 | pcpu_high_unit_cpu = NR_CPUS; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1580 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1581 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
| 1582 | const struct pcpu_group_info *gi = &ai->groups[group]; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1583 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1584 | group_offsets[group] = gi->base_offset; |
| 1585 | group_sizes[group] = gi->nr_units * ai->unit_size; |
| 1586 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1587 | for (i = 0; i < gi->nr_units; i++) { |
| 1588 | cpu = gi->cpu_map[i]; |
| 1589 | if (cpu == NR_CPUS) |
| 1590 | continue; |
| 1591 | |
Dan Carpenter | 9f29566 | 2014-10-29 11:45:04 +0300 | [diff] [blame] | 1592 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1593 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
| 1594 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1595 | |
| 1596 | unit_map[cpu] = unit + i; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1597 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
| 1598 | |
Tejun Heo | a855b84 | 2011-11-18 10:55:35 -0800 | [diff] [blame] | 1599 | /* determine low/high unit_cpu */ |
| 1600 | if (pcpu_low_unit_cpu == NR_CPUS || |
| 1601 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) |
| 1602 | pcpu_low_unit_cpu = cpu; |
| 1603 | if (pcpu_high_unit_cpu == NR_CPUS || |
| 1604 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) |
| 1605 | pcpu_high_unit_cpu = cpu; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1606 | } |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1607 | } |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1608 | pcpu_nr_units = unit; |
| 1609 | |
| 1610 | for_each_possible_cpu(cpu) |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1611 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
| 1612 | |
| 1613 | /* we're done parsing the input, undefine BUG macro and dump config */ |
| 1614 | #undef PCPU_SETUP_BUG_ON |
Tejun Heo | bcbea79 | 2010-12-22 14:19:14 +0100 | [diff] [blame] | 1615 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1616 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1617 | pcpu_nr_groups = ai->nr_groups; |
| 1618 | pcpu_group_offsets = group_offsets; |
| 1619 | pcpu_group_sizes = group_sizes; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1620 | pcpu_unit_map = unit_map; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1621 | pcpu_unit_offsets = unit_off; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1622 | |
| 1623 | /* determine basic parameters */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1624 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1625 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1626 | pcpu_atom_size = ai->atom_size; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1627 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
| 1628 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1629 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1630 | /* |
| 1631 | * Allocate chunk slots. The additional last slot is for |
| 1632 | * empty chunks. |
| 1633 | */ |
| 1634 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1635 | pcpu_slot = memblock_virt_alloc( |
| 1636 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1637 | for (i = 0; i < pcpu_nr_slots; i++) |
| 1638 | INIT_LIST_HEAD(&pcpu_slot[i]); |
| 1639 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1640 | /* |
| 1641 | * Initialize static chunk. If reserved_size is zero, the |
| 1642 | * static chunk covers static area + dynamic allocation area |
| 1643 | * in the first chunk. If reserved_size is not zero, it |
| 1644 | * covers static area + reserved area (mostly used for module |
| 1645 | * static percpu allocation). |
| 1646 | */ |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1647 | schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1648 | INIT_LIST_HEAD(&schunk->list); |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1649 | INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1650 | schunk->base_addr = base_addr; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1651 | schunk->map = smap; |
| 1652 | schunk->map_alloc = ARRAY_SIZE(smap); |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1653 | schunk->immutable = true; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1654 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1655 | schunk->nr_populated = pcpu_unit_pages; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1656 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1657 | if (ai->reserved_size) { |
| 1658 | schunk->free_size = ai->reserved_size; |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1659 | pcpu_reserved_chunk = schunk; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1660 | pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1661 | } else { |
| 1662 | schunk->free_size = dyn_size; |
| 1663 | dyn_size = 0; /* dynamic area covered */ |
| 1664 | } |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1665 | schunk->contig_hint = schunk->free_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1666 | |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 1667 | schunk->map[0] = 1; |
| 1668 | schunk->map[1] = ai->static_size; |
| 1669 | schunk->map_used = 1; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1670 | if (schunk->free_size) |
Baoquan He | 292c24a | 2015-07-20 22:55:28 +0800 | [diff] [blame] | 1671 | schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size; |
| 1672 | schunk->map[schunk->map_used] |= 1; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1673 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1674 | /* init dynamic chunk if necessary */ |
| 1675 | if (dyn_size) { |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1676 | dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1677 | INIT_LIST_HEAD(&dchunk->list); |
Tejun Heo | 9c824b6 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1678 | INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1679 | dchunk->base_addr = base_addr; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1680 | dchunk->map = dmap; |
| 1681 | dchunk->map_alloc = ARRAY_SIZE(dmap); |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1682 | dchunk->immutable = true; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1683 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1684 | dchunk->nr_populated = pcpu_unit_pages; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1685 | |
| 1686 | dchunk->contig_hint = dchunk->free_size = dyn_size; |
Al Viro | 723ad1d | 2014-03-06 21:13:18 -0500 | [diff] [blame] | 1687 | dchunk->map[0] = 1; |
| 1688 | dchunk->map[1] = pcpu_reserved_chunk_limit; |
| 1689 | dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1; |
| 1690 | dchunk->map_used = 2; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1691 | } |
| 1692 | |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1693 | /* link the first chunk in */ |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1694 | pcpu_first_chunk = dchunk ?: schunk; |
Tejun Heo | b539b87 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 1695 | pcpu_nr_empty_pop_pages += |
| 1696 | pcpu_count_occupied_pages(pcpu_first_chunk, 1); |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1697 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1698 | |
| 1699 | /* we're done */ |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1700 | pcpu_base_addr = base_addr; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1701 | return 0; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1702 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1703 | |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 1704 | #ifdef CONFIG_SMP |
| 1705 | |
Andi Kleen | 17f3609 | 2012-10-04 17:12:07 -0700 | [diff] [blame] | 1706 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1707 | [PCPU_FC_AUTO] = "auto", |
| 1708 | [PCPU_FC_EMBED] = "embed", |
| 1709 | [PCPU_FC_PAGE] = "page", |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1710 | }; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1711 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1712 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
| 1713 | |
| 1714 | static int __init percpu_alloc_setup(char *str) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1715 | { |
Cyrill Gorcunov | 5479c78 | 2012-11-25 01:17:13 +0400 | [diff] [blame] | 1716 | if (!str) |
| 1717 | return -EINVAL; |
| 1718 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1719 | if (0) |
| 1720 | /* nada */; |
| 1721 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK |
| 1722 | else if (!strcmp(str, "embed")) |
| 1723 | pcpu_chosen_fc = PCPU_FC_EMBED; |
| 1724 | #endif |
| 1725 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
| 1726 | else if (!strcmp(str, "page")) |
| 1727 | pcpu_chosen_fc = PCPU_FC_PAGE; |
| 1728 | #endif |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1729 | else |
| 1730 | pr_warning("PERCPU: unknown allocator %s specified\n", str); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1731 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1732 | return 0; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1733 | } |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1734 | early_param("percpu_alloc", percpu_alloc_setup); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1735 | |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 1736 | /* |
| 1737 | * pcpu_embed_first_chunk() is used by the generic percpu setup. |
| 1738 | * Build it if needed by the arch config or the generic setup is going |
| 1739 | * to be used. |
| 1740 | */ |
Tejun Heo | 08fc458 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1741 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
| 1742 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 1743 | #define BUILD_EMBED_FIRST_CHUNK |
| 1744 | #endif |
| 1745 | |
| 1746 | /* build pcpu_page_first_chunk() iff needed by the arch config */ |
| 1747 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) |
| 1748 | #define BUILD_PAGE_FIRST_CHUNK |
| 1749 | #endif |
| 1750 | |
| 1751 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ |
| 1752 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) |
| 1753 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1754 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs |
| 1755 | * @reserved_size: the size of reserved percpu area in bytes |
| 1756 | * @dyn_size: minimum free size for dynamic allocation in bytes |
| 1757 | * @atom_size: allocation atom size |
| 1758 | * @cpu_distance_fn: callback to determine distance between cpus, optional |
| 1759 | * |
| 1760 | * This function determines grouping of units, their mappings to cpus |
| 1761 | * and other parameters considering needed percpu size, allocation |
| 1762 | * atom size and distances between CPUs. |
| 1763 | * |
Yannick Guerrini | bffc437 | 2015-03-06 23:30:42 +0100 | [diff] [blame] | 1764 | * Groups are always multiples of atom size and CPUs which are of |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1765 | * LOCAL_DISTANCE both ways are grouped together and share space for |
| 1766 | * units in the same group. The returned configuration is guaranteed |
| 1767 | * to have CPUs on different nodes on different groups and >=75% usage |
| 1768 | * of allocated virtual address space. |
| 1769 | * |
| 1770 | * RETURNS: |
| 1771 | * On success, pointer to the new allocation_info is returned. On |
| 1772 | * failure, ERR_PTR value is returned. |
| 1773 | */ |
| 1774 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( |
| 1775 | size_t reserved_size, size_t dyn_size, |
| 1776 | size_t atom_size, |
| 1777 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) |
| 1778 | { |
| 1779 | static int group_map[NR_CPUS] __initdata; |
| 1780 | static int group_cnt[NR_CPUS] __initdata; |
| 1781 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
| 1782 | int nr_groups = 1, nr_units = 0; |
| 1783 | size_t size_sum, min_unit_size, alloc_size; |
| 1784 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ |
| 1785 | int last_allocs, group, unit; |
| 1786 | unsigned int cpu, tcpu; |
| 1787 | struct pcpu_alloc_info *ai; |
| 1788 | unsigned int *cpu_map; |
| 1789 | |
| 1790 | /* this function may be called multiple times */ |
| 1791 | memset(group_map, 0, sizeof(group_map)); |
| 1792 | memset(group_cnt, 0, sizeof(group_cnt)); |
| 1793 | |
| 1794 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ |
| 1795 | size_sum = PFN_ALIGN(static_size + reserved_size + |
| 1796 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); |
| 1797 | dyn_size = size_sum - static_size - reserved_size; |
| 1798 | |
| 1799 | /* |
| 1800 | * Determine min_unit_size, alloc_size and max_upa such that |
| 1801 | * alloc_size is multiple of atom_size and is the smallest |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1802 | * which can accommodate 4k aligned segments which are equal to |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1803 | * or larger than min_unit_size. |
| 1804 | */ |
| 1805 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
| 1806 | |
| 1807 | alloc_size = roundup(min_unit_size, atom_size); |
| 1808 | upa = alloc_size / min_unit_size; |
| 1809 | while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) |
| 1810 | upa--; |
| 1811 | max_upa = upa; |
| 1812 | |
| 1813 | /* group cpus according to their proximity */ |
| 1814 | for_each_possible_cpu(cpu) { |
| 1815 | group = 0; |
| 1816 | next_group: |
| 1817 | for_each_possible_cpu(tcpu) { |
| 1818 | if (cpu == tcpu) |
| 1819 | break; |
| 1820 | if (group_map[tcpu] == group && cpu_distance_fn && |
| 1821 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || |
| 1822 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { |
| 1823 | group++; |
| 1824 | nr_groups = max(nr_groups, group + 1); |
| 1825 | goto next_group; |
| 1826 | } |
| 1827 | } |
| 1828 | group_map[cpu] = group; |
| 1829 | group_cnt[group]++; |
| 1830 | } |
| 1831 | |
| 1832 | /* |
| 1833 | * Expand unit size until address space usage goes over 75% |
| 1834 | * and then as much as possible without using more address |
| 1835 | * space. |
| 1836 | */ |
| 1837 | last_allocs = INT_MAX; |
| 1838 | for (upa = max_upa; upa; upa--) { |
| 1839 | int allocs = 0, wasted = 0; |
| 1840 | |
| 1841 | if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) |
| 1842 | continue; |
| 1843 | |
| 1844 | for (group = 0; group < nr_groups; group++) { |
| 1845 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); |
| 1846 | allocs += this_allocs; |
| 1847 | wasted += this_allocs * upa - group_cnt[group]; |
| 1848 | } |
| 1849 | |
| 1850 | /* |
| 1851 | * Don't accept if wastage is over 1/3. The |
| 1852 | * greater-than comparison ensures upa==1 always |
| 1853 | * passes the following check. |
| 1854 | */ |
| 1855 | if (wasted > num_possible_cpus() / 3) |
| 1856 | continue; |
| 1857 | |
| 1858 | /* and then don't consume more memory */ |
| 1859 | if (allocs > last_allocs) |
| 1860 | break; |
| 1861 | last_allocs = allocs; |
| 1862 | best_upa = upa; |
| 1863 | } |
| 1864 | upa = best_upa; |
| 1865 | |
| 1866 | /* allocate and fill alloc_info */ |
| 1867 | for (group = 0; group < nr_groups; group++) |
| 1868 | nr_units += roundup(group_cnt[group], upa); |
| 1869 | |
| 1870 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); |
| 1871 | if (!ai) |
| 1872 | return ERR_PTR(-ENOMEM); |
| 1873 | cpu_map = ai->groups[0].cpu_map; |
| 1874 | |
| 1875 | for (group = 0; group < nr_groups; group++) { |
| 1876 | ai->groups[group].cpu_map = cpu_map; |
| 1877 | cpu_map += roundup(group_cnt[group], upa); |
| 1878 | } |
| 1879 | |
| 1880 | ai->static_size = static_size; |
| 1881 | ai->reserved_size = reserved_size; |
| 1882 | ai->dyn_size = dyn_size; |
| 1883 | ai->unit_size = alloc_size / upa; |
| 1884 | ai->atom_size = atom_size; |
| 1885 | ai->alloc_size = alloc_size; |
| 1886 | |
| 1887 | for (group = 0, unit = 0; group_cnt[group]; group++) { |
| 1888 | struct pcpu_group_info *gi = &ai->groups[group]; |
| 1889 | |
| 1890 | /* |
| 1891 | * Initialize base_offset as if all groups are located |
| 1892 | * back-to-back. The caller should update this to |
| 1893 | * reflect actual allocation. |
| 1894 | */ |
| 1895 | gi->base_offset = unit * ai->unit_size; |
| 1896 | |
| 1897 | for_each_possible_cpu(cpu) |
| 1898 | if (group_map[cpu] == group) |
| 1899 | gi->cpu_map[gi->nr_units++] = cpu; |
| 1900 | gi->nr_units = roundup(gi->nr_units, upa); |
| 1901 | unit += gi->nr_units; |
| 1902 | } |
| 1903 | BUG_ON(unit != nr_units); |
| 1904 | |
| 1905 | return ai; |
| 1906 | } |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 1907 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1908 | |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 1909 | #if defined(BUILD_EMBED_FIRST_CHUNK) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1910 | /** |
| 1911 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1912 | * @reserved_size: the size of reserved percpu area in bytes |
Tejun Heo | 4ba6ce2 | 2010-06-27 18:49:59 +0200 | [diff] [blame] | 1913 | * @dyn_size: minimum free size for dynamic allocation in bytes |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1914 | * @atom_size: allocation atom size |
| 1915 | * @cpu_distance_fn: callback to determine distance between cpus, optional |
| 1916 | * @alloc_fn: function to allocate percpu page |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 1917 | * @free_fn: function to free percpu page |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1918 | * |
| 1919 | * This is a helper to ease setting up embedded first percpu chunk and |
| 1920 | * can be called where pcpu_setup_first_chunk() is expected. |
| 1921 | * |
| 1922 | * If this function is used to setup the first chunk, it is allocated |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1923 | * by calling @alloc_fn and used as-is without being mapped into |
| 1924 | * vmalloc area. Allocations are always whole multiples of @atom_size |
| 1925 | * aligned to @atom_size. |
| 1926 | * |
| 1927 | * This enables the first chunk to piggy back on the linear physical |
| 1928 | * mapping which often uses larger page size. Please note that this |
| 1929 | * can result in very sparse cpu->unit mapping on NUMA machines thus |
| 1930 | * requiring large vmalloc address space. Don't use this allocator if |
| 1931 | * vmalloc space is not orders of magnitude larger than distances |
| 1932 | * between node memory addresses (ie. 32bit NUMA machines). |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1933 | * |
Tejun Heo | 4ba6ce2 | 2010-06-27 18:49:59 +0200 | [diff] [blame] | 1934 | * @dyn_size specifies the minimum dynamic area size. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1935 | * |
| 1936 | * If the needed size is smaller than the minimum or specified unit |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1937 | * size, the leftover is returned using @free_fn. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1938 | * |
| 1939 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1940 | * 0 on success, -errno on failure. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1941 | */ |
Tejun Heo | 4ba6ce2 | 2010-06-27 18:49:59 +0200 | [diff] [blame] | 1942 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1943 | size_t atom_size, |
| 1944 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, |
| 1945 | pcpu_fc_alloc_fn_t alloc_fn, |
| 1946 | pcpu_fc_free_fn_t free_fn) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1947 | { |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1948 | void *base = (void *)ULONG_MAX; |
| 1949 | void **areas = NULL; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1950 | struct pcpu_alloc_info *ai; |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 1951 | size_t size_sum, areas_size, max_distance; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1952 | int group, i, rc; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1953 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1954 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
| 1955 | cpu_distance_fn); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1956 | if (IS_ERR(ai)) |
| 1957 | return PTR_ERR(ai); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1958 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1959 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1960 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1961 | |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 1962 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1963 | if (!areas) { |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1964 | rc = -ENOMEM; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1965 | goto out_free; |
Tejun Heo | fa8a709 | 2009-06-22 11:56:24 +0900 | [diff] [blame] | 1966 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1967 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1968 | /* allocate, copy and determine base address */ |
| 1969 | for (group = 0; group < ai->nr_groups; group++) { |
| 1970 | struct pcpu_group_info *gi = &ai->groups[group]; |
| 1971 | unsigned int cpu = NR_CPUS; |
| 1972 | void *ptr; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1973 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1974 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) |
| 1975 | cpu = gi->cpu_map[i]; |
| 1976 | BUG_ON(cpu == NR_CPUS); |
| 1977 | |
| 1978 | /* allocate space for the whole group */ |
| 1979 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); |
| 1980 | if (!ptr) { |
| 1981 | rc = -ENOMEM; |
| 1982 | goto out_free_areas; |
| 1983 | } |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 1984 | /* kmemleak tracks the percpu allocations separately */ |
| 1985 | kmemleak_free(ptr); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1986 | areas[group] = ptr; |
| 1987 | |
| 1988 | base = min(ptr, base); |
Tejun Heo | 42b6428 | 2012-04-27 08:42:53 -0700 | [diff] [blame] | 1989 | } |
| 1990 | |
| 1991 | /* |
| 1992 | * Copy data and free unused parts. This should happen after all |
| 1993 | * allocations are complete; otherwise, we may end up with |
| 1994 | * overlapping groups. |
| 1995 | */ |
| 1996 | for (group = 0; group < ai->nr_groups; group++) { |
| 1997 | struct pcpu_group_info *gi = &ai->groups[group]; |
| 1998 | void *ptr = areas[group]; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1999 | |
| 2000 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { |
| 2001 | if (gi->cpu_map[i] == NR_CPUS) { |
| 2002 | /* unused unit, free whole */ |
| 2003 | free_fn(ptr, ai->unit_size); |
| 2004 | continue; |
| 2005 | } |
| 2006 | /* copy and return the unused part */ |
| 2007 | memcpy(ptr, __per_cpu_load, ai->static_size); |
| 2008 | free_fn(ptr + size_sum, ai->unit_size - size_sum); |
| 2009 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2010 | } |
| 2011 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2012 | /* base address is now known, determine group base offsets */ |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 2013 | max_distance = 0; |
| 2014 | for (group = 0; group < ai->nr_groups; group++) { |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2015 | ai->groups[group].base_offset = areas[group] - base; |
Tejun Heo | 1a0c329 | 2009-10-04 09:31:05 +0900 | [diff] [blame] | 2016 | max_distance = max_t(size_t, max_distance, |
| 2017 | ai->groups[group].base_offset); |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 2018 | } |
| 2019 | max_distance += ai->unit_size; |
| 2020 | |
| 2021 | /* warn if maximum distance is further than 75% of vmalloc space */ |
Laura Abbott | 8a09217 | 2014-01-02 13:53:21 -0800 | [diff] [blame] | 2022 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { |
Tejun Heo | 1a0c329 | 2009-10-04 09:31:05 +0900 | [diff] [blame] | 2023 | pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " |
Mike Frysinger | 787e5b0 | 2011-03-23 08:23:52 +0100 | [diff] [blame] | 2024 | "space 0x%lx\n", max_distance, |
Laura Abbott | 8a09217 | 2014-01-02 13:53:21 -0800 | [diff] [blame] | 2025 | VMALLOC_TOTAL); |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 2026 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
| 2027 | /* and fail if we have fallback */ |
| 2028 | rc = -EINVAL; |
| 2029 | goto out_free; |
| 2030 | #endif |
| 2031 | } |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2032 | |
Tejun Heo | 004018e | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2033 | pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2034 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
| 2035 | ai->dyn_size, ai->unit_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2036 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2037 | rc = pcpu_setup_first_chunk(ai, base); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2038 | goto out_free; |
| 2039 | |
| 2040 | out_free_areas: |
| 2041 | for (group = 0; group < ai->nr_groups; group++) |
Michael Holzheu | f851c8d | 2013-09-17 16:57:34 +0200 | [diff] [blame] | 2042 | if (areas[group]) |
| 2043 | free_fn(areas[group], |
| 2044 | ai->groups[group].nr_units * ai->unit_size); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2045 | out_free: |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2046 | pcpu_free_alloc_info(ai); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2047 | if (areas) |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2048 | memblock_free_early(__pa(areas), areas_size); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2049 | return rc; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2050 | } |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 2051 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2052 | |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 2053 | #ifdef BUILD_PAGE_FIRST_CHUNK |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2054 | /** |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2055 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2056 | * @reserved_size: the size of reserved percpu area in bytes |
| 2057 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 2058 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2059 | * @populate_pte_fn: function to populate pte |
| 2060 | * |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2061 | * This is a helper to ease setting up page-remapped first percpu |
| 2062 | * chunk and can be called where pcpu_setup_first_chunk() is expected. |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2063 | * |
| 2064 | * This is the basic allocator. Static percpu area is allocated |
| 2065 | * page-by-page into vmalloc area. |
| 2066 | * |
| 2067 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2068 | * 0 on success, -errno on failure. |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2069 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2070 | int __init pcpu_page_first_chunk(size_t reserved_size, |
| 2071 | pcpu_fc_alloc_fn_t alloc_fn, |
| 2072 | pcpu_fc_free_fn_t free_fn, |
| 2073 | pcpu_fc_populate_pte_fn_t populate_pte_fn) |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2074 | { |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2075 | static struct vm_struct vm; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2076 | struct pcpu_alloc_info *ai; |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2077 | char psize_str[16]; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2078 | int unit_pages; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2079 | size_t pages_size; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2080 | struct page **pages; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2081 | int unit, i, j, rc; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2082 | |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2083 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
| 2084 | |
Tejun Heo | 4ba6ce2 | 2010-06-27 18:49:59 +0200 | [diff] [blame] | 2085 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2086 | if (IS_ERR(ai)) |
| 2087 | return PTR_ERR(ai); |
| 2088 | BUG_ON(ai->nr_groups != 1); |
| 2089 | BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); |
| 2090 | |
| 2091 | unit_pages = ai->unit_size >> PAGE_SHIFT; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2092 | |
| 2093 | /* unaligned allocations can't be freed, round up to page size */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2094 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
| 2095 | sizeof(pages[0])); |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2096 | pages = memblock_virt_alloc(pages_size, 0); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2097 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2098 | /* allocate pages */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2099 | j = 0; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2100 | for (unit = 0; unit < num_possible_cpus(); unit++) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2101 | for (i = 0; i < unit_pages; i++) { |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2102 | unsigned int cpu = ai->groups[0].cpu_map[unit]; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2103 | void *ptr; |
| 2104 | |
Tejun Heo | 3cbc856 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 2105 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2106 | if (!ptr) { |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2107 | pr_warning("PERCPU: failed to allocate %s page " |
| 2108 | "for cpu%u\n", psize_str, cpu); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2109 | goto enomem; |
| 2110 | } |
Catalin Marinas | f528f0b | 2011-09-26 17:12:53 +0100 | [diff] [blame] | 2111 | /* kmemleak tracks the percpu allocations separately */ |
| 2112 | kmemleak_free(ptr); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2113 | pages[j++] = virt_to_page(ptr); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2114 | } |
| 2115 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2116 | /* allocate vm area, map the pages and copy static data */ |
| 2117 | vm.flags = VM_ALLOC; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2118 | vm.size = num_possible_cpus() * ai->unit_size; |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2119 | vm_area_register_early(&vm, PAGE_SIZE); |
| 2120 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2121 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
Tejun Heo | 1d9d325 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 2122 | unsigned long unit_addr = |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2123 | (unsigned long)vm.addr + unit * ai->unit_size; |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2124 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2125 | for (i = 0; i < unit_pages; i++) |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2126 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
| 2127 | |
| 2128 | /* pte already populated, the following shouldn't fail */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2129 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
| 2130 | unit_pages); |
| 2131 | if (rc < 0) |
| 2132 | panic("failed to map percpu area, err=%d\n", rc); |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2133 | |
| 2134 | /* |
| 2135 | * FIXME: Archs with virtual cache should flush local |
| 2136 | * cache for the linear mapping here - something |
| 2137 | * equivalent to flush_cache_vmap() on the local cpu. |
| 2138 | * flush_cache_vmap() can't be used as most supporting |
| 2139 | * data structures are not set up yet. |
| 2140 | */ |
| 2141 | |
| 2142 | /* copy static data */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2143 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2144 | } |
| 2145 | |
| 2146 | /* we're ready, commit */ |
Tejun Heo | 1d9d325 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 2147 | pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2148 | unit_pages, psize_str, vm.addr, ai->static_size, |
| 2149 | ai->reserved_size, ai->dyn_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2150 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2151 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2152 | goto out_free_ar; |
| 2153 | |
| 2154 | enomem: |
| 2155 | while (--j >= 0) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2156 | free_fn(page_address(pages[j]), PAGE_SIZE); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2157 | rc = -ENOMEM; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2158 | out_free_ar: |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2159 | memblock_free_early(__pa(pages), pages_size); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2160 | pcpu_free_alloc_info(ai); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2161 | return rc; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2162 | } |
Tejun Heo | 3c9a024 | 2010-09-09 18:00:15 +0200 | [diff] [blame] | 2163 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2164 | |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2165 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2166 | /* |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2167 | * Generic SMP percpu area setup. |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2168 | * |
| 2169 | * The embedding helper is used because its behavior closely resembles |
| 2170 | * the original non-dynamic generic percpu area setup. This is |
| 2171 | * important because many archs have addressing restrictions and might |
| 2172 | * fail if the percpu area is located far away from the previous |
| 2173 | * location. As an added bonus, in non-NUMA cases, embedding is |
| 2174 | * generally a good idea TLB-wise because percpu area can piggy back |
| 2175 | * on the physical linear memory mapping which uses large page |
| 2176 | * mappings on applicable archs. |
| 2177 | */ |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2178 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
| 2179 | EXPORT_SYMBOL(__per_cpu_offset); |
| 2180 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2181 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
| 2182 | size_t align) |
| 2183 | { |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2184 | return memblock_virt_alloc_from_nopanic( |
| 2185 | size, align, __pa(MAX_DMA_ADDRESS)); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2186 | } |
| 2187 | |
| 2188 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
| 2189 | { |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2190 | memblock_free_early(__pa(ptr), size); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2191 | } |
| 2192 | |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2193 | void __init setup_per_cpu_areas(void) |
| 2194 | { |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2195 | unsigned long delta; |
| 2196 | unsigned int cpu; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2197 | int rc; |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2198 | |
| 2199 | /* |
| 2200 | * Always reserve area for module percpu variables. That's |
| 2201 | * what the legacy allocator did. |
| 2202 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2203 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2204 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
| 2205 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2206 | if (rc < 0) |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2207 | panic("Failed to initialize percpu areas."); |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2208 | |
| 2209 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
| 2210 | for_each_possible_cpu(cpu) |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2211 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2212 | } |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2213 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
| 2214 | |
| 2215 | #else /* CONFIG_SMP */ |
| 2216 | |
| 2217 | /* |
| 2218 | * UP percpu area setup. |
| 2219 | * |
| 2220 | * UP always uses km-based percpu allocator with identity mapping. |
| 2221 | * Static percpu variables are indistinguishable from the usual static |
| 2222 | * variables and don't require any special preparation. |
| 2223 | */ |
| 2224 | void __init setup_per_cpu_areas(void) |
| 2225 | { |
| 2226 | const size_t unit_size = |
| 2227 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, |
| 2228 | PERCPU_DYNAMIC_RESERVE)); |
| 2229 | struct pcpu_alloc_info *ai; |
| 2230 | void *fc; |
| 2231 | |
| 2232 | ai = pcpu_alloc_alloc_info(1, 1); |
Santosh Shilimkar | 999c17e | 2014-01-21 15:50:40 -0800 | [diff] [blame] | 2233 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
| 2234 | PAGE_SIZE, |
| 2235 | __pa(MAX_DMA_ADDRESS)); |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2236 | if (!ai || !fc) |
| 2237 | panic("Failed to allocate memory for percpu areas."); |
Catalin Marinas | 100d13c | 2012-05-09 16:55:19 +0100 | [diff] [blame] | 2238 | /* kmemleak tracks the percpu allocations separately */ |
| 2239 | kmemleak_free(fc); |
Tejun Heo | bbddff0 | 2010-09-03 18:22:48 +0200 | [diff] [blame] | 2240 | |
| 2241 | ai->dyn_size = unit_size; |
| 2242 | ai->unit_size = unit_size; |
| 2243 | ai->atom_size = unit_size; |
| 2244 | ai->alloc_size = unit_size; |
| 2245 | ai->groups[0].nr_units = 1; |
| 2246 | ai->groups[0].cpu_map[0] = 0; |
| 2247 | |
| 2248 | if (pcpu_setup_first_chunk(ai, fc) < 0) |
| 2249 | panic("Failed to initialize percpu areas."); |
| 2250 | } |
| 2251 | |
| 2252 | #endif /* CONFIG_SMP */ |
Tejun Heo | 099a19d | 2010-06-27 18:50:00 +0200 | [diff] [blame] | 2253 | |
| 2254 | /* |
| 2255 | * First and reserved chunks are initialized with temporary allocation |
| 2256 | * map in initdata so that they can be used before slab is online. |
| 2257 | * This function is called after slab is brought up and replaces those |
| 2258 | * with properly allocated maps. |
| 2259 | */ |
| 2260 | void __init percpu_init_late(void) |
| 2261 | { |
| 2262 | struct pcpu_chunk *target_chunks[] = |
| 2263 | { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; |
| 2264 | struct pcpu_chunk *chunk; |
| 2265 | unsigned long flags; |
| 2266 | int i; |
| 2267 | |
| 2268 | for (i = 0; (chunk = target_chunks[i]); i++) { |
| 2269 | int *map; |
| 2270 | const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); |
| 2271 | |
| 2272 | BUILD_BUG_ON(size > PAGE_SIZE); |
| 2273 | |
Bob Liu | 90459ce0 | 2011-08-04 11:02:33 +0200 | [diff] [blame] | 2274 | map = pcpu_mem_zalloc(size); |
Tejun Heo | 099a19d | 2010-06-27 18:50:00 +0200 | [diff] [blame] | 2275 | BUG_ON(!map); |
| 2276 | |
| 2277 | spin_lock_irqsave(&pcpu_lock, flags); |
| 2278 | memcpy(map, chunk->map, size); |
| 2279 | chunk->map = map; |
| 2280 | spin_unlock_irqrestore(&pcpu_lock, flags); |
| 2281 | } |
| 2282 | } |
Tejun Heo | 1a4d760 | 2014-09-02 14:46:05 -0400 | [diff] [blame] | 2283 | |
| 2284 | /* |
| 2285 | * Percpu allocator is initialized early during boot when neither slab or |
| 2286 | * workqueue is available. Plug async management until everything is up |
| 2287 | * and running. |
| 2288 | */ |
| 2289 | static int __init percpu_enable_async(void) |
| 2290 | { |
| 2291 | pcpu_async_enabled = true; |
| 2292 | return 0; |
| 2293 | } |
| 2294 | subsys_initcall(percpu_enable_async); |