Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1 | /* |
| 2 | * linux/mm/percpu.c - percpu memory allocator |
| 3 | * |
| 4 | * Copyright (C) 2009 SUSE Linux Products GmbH |
| 5 | * Copyright (C) 2009 Tejun Heo <tj@kernel.org> |
| 6 | * |
| 7 | * This file is released under the GPLv2. |
| 8 | * |
| 9 | * This is percpu allocator which can handle both static and dynamic |
| 10 | * areas. Percpu areas are allocated in chunks in vmalloc area. Each |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 11 | * chunk is consisted of boot-time determined number of units and the |
| 12 | * first chunk is used for static percpu variables in the kernel image |
| 13 | * (special boot time alloc/init handling necessary as these areas |
| 14 | * need to be brought up before allocation services are running). |
| 15 | * Unit grows as necessary and all units grow or shrink in unison. |
| 16 | * When a chunk is filled up, another chunk is allocated. ie. in |
| 17 | * vmalloc area |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 18 | * |
| 19 | * c0 c1 c2 |
| 20 | * ------------------- ------------------- ------------ |
| 21 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u |
| 22 | * ------------------- ...... ------------------- .... ------------ |
| 23 | * |
| 24 | * Allocation is done in offset-size areas of single unit space. Ie, |
| 25 | * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 26 | * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to |
| 27 | * cpus. On NUMA, the mapping can be non-linear and even sparse. |
| 28 | * Percpu access can be done by configuring percpu base registers |
| 29 | * according to cpu to unit mapping and pcpu_unit_size. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 30 | * |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 31 | * There are usually many small percpu allocations many of them being |
| 32 | * as small as 4 bytes. The allocator organizes chunks into lists |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 33 | * according to free size and tries to allocate from the fullest one. |
| 34 | * Each chunk keeps the maximum contiguous area size hint which is |
| 35 | * guaranteed to be eqaul to or larger than the maximum contiguous |
| 36 | * area in the chunk. This helps the allocator not to iterate the |
| 37 | * chunk maps unnecessarily. |
| 38 | * |
| 39 | * Allocation state in each chunk is kept using an array of integers |
| 40 | * on chunk->map. A positive value in the map represents a free |
| 41 | * region and negative allocated. Allocation inside a chunk is done |
| 42 | * by scanning this map sequentially and serving the first matching |
| 43 | * entry. This is mostly copied from the percpu_modalloc() allocator. |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 44 | * Chunks can be determined from the address using the index field |
| 45 | * in the page struct. The index field contains a pointer to the chunk. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 46 | * |
| 47 | * To use this allocator, arch code should do the followings. |
| 48 | * |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 49 | * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 50 | * |
| 51 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 52 | * regular address to percpu pointer and back if they need to be |
| 53 | * different from the default |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 54 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 55 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
| 56 | * setup the first chunk containing the kernel static percpu area |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 57 | */ |
| 58 | |
| 59 | #include <linux/bitmap.h> |
| 60 | #include <linux/bootmem.h> |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 61 | #include <linux/err.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 62 | #include <linux/list.h> |
Tejun Heo | a530b79 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 63 | #include <linux/log2.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 64 | #include <linux/mm.h> |
| 65 | #include <linux/module.h> |
| 66 | #include <linux/mutex.h> |
| 67 | #include <linux/percpu.h> |
| 68 | #include <linux/pfn.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 69 | #include <linux/slab.h> |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 70 | #include <linux/spinlock.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 71 | #include <linux/vmalloc.h> |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 72 | #include <linux/workqueue.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 73 | |
| 74 | #include <asm/cacheflush.h> |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 75 | #include <asm/sections.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 76 | #include <asm/tlbflush.h> |
| 77 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 78 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
| 79 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ |
| 80 | |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 81 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
| 82 | #ifndef __addr_to_pcpu_ptr |
| 83 | #define __addr_to_pcpu_ptr(addr) \ |
| 84 | (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ |
| 85 | + (unsigned long)__per_cpu_start) |
| 86 | #endif |
| 87 | #ifndef __pcpu_ptr_to_addr |
| 88 | #define __pcpu_ptr_to_addr(ptr) \ |
| 89 | (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ |
| 90 | - (unsigned long)__per_cpu_start) |
| 91 | #endif |
| 92 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 93 | struct pcpu_chunk { |
| 94 | struct list_head list; /* linked to pcpu_slot lists */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 95 | int free_size; /* free bytes in the chunk */ |
| 96 | int contig_hint; /* max contiguous size hint */ |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 97 | void *base_addr; /* base address of this chunk */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 98 | int map_used; /* # of map entries used */ |
| 99 | int map_alloc; /* # of map entries allocated */ |
| 100 | int *map; /* allocation map */ |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 101 | struct vm_struct **vms; /* mapped vmalloc regions */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 102 | bool immutable; /* no [de]population allowed */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 103 | unsigned long populated[]; /* populated bitmap */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 104 | }; |
| 105 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 106 | static int pcpu_unit_pages __read_mostly; |
| 107 | static int pcpu_unit_size __read_mostly; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 108 | static int pcpu_nr_units __read_mostly; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 109 | static int pcpu_atom_size __read_mostly; |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 110 | static int pcpu_nr_slots __read_mostly; |
| 111 | static size_t pcpu_chunk_struct_size __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 112 | |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 113 | /* cpus with the lowest and highest unit numbers */ |
| 114 | static unsigned int pcpu_first_unit_cpu __read_mostly; |
| 115 | static unsigned int pcpu_last_unit_cpu __read_mostly; |
| 116 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 117 | /* 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] | 118 | void *pcpu_base_addr __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 119 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
| 120 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 121 | static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ |
| 122 | const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 123 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 124 | /* group information, used for vm allocation */ |
| 125 | static int pcpu_nr_groups __read_mostly; |
| 126 | static const unsigned long *pcpu_group_offsets __read_mostly; |
| 127 | static const size_t *pcpu_group_sizes __read_mostly; |
| 128 | |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 129 | /* |
| 130 | * The first chunk which always exists. Note that unlike other |
| 131 | * chunks, this one can be allocated and mapped in several different |
| 132 | * ways and thus often doesn't live in the vmalloc area. |
| 133 | */ |
| 134 | static struct pcpu_chunk *pcpu_first_chunk; |
| 135 | |
| 136 | /* |
| 137 | * Optional reserved chunk. This chunk reserves part of the first |
| 138 | * chunk and serves it for reserved allocations. The amount of |
| 139 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved |
| 140 | * area doesn't exist, the following variables contain NULL and 0 |
| 141 | * respectively. |
| 142 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 143 | static struct pcpu_chunk *pcpu_reserved_chunk; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 144 | static int pcpu_reserved_chunk_limit; |
| 145 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 146 | /* |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 147 | * Synchronization rules. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 148 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 149 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 150 | * protects allocation/reclaim paths, chunks, populated bitmap and |
| 151 | * vmalloc mapping. The latter is a spinlock and protects the index |
| 152 | * data structures - chunk slots, chunks and area maps in chunks. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 153 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 154 | * During allocation, pcpu_alloc_mutex is kept locked all the time and |
| 155 | * pcpu_lock is grabbed and released as necessary. All actual memory |
| 156 | * allocations are done using GFP_KERNEL with pcpu_lock released. |
| 157 | * |
| 158 | * Free path accesses and alters only the index data structures, so it |
| 159 | * can be safely called from atomic context. When memory needs to be |
| 160 | * returned to the system, free path schedules reclaim_work which |
| 161 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be |
| 162 | * reclaimed, release both locks and frees the chunks. Note that it's |
| 163 | * necessary to grab both locks to remove a chunk from circulation as |
| 164 | * allocation path might be referencing the chunk with only |
| 165 | * pcpu_alloc_mutex locked. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 166 | */ |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 167 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
| 168 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 169 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 170 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 171 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 172 | /* reclaim work to release fully free chunks, scheduled from free path */ |
| 173 | static void pcpu_reclaim(struct work_struct *work); |
| 174 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); |
| 175 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 176 | static int __pcpu_size_to_slot(int size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 177 | { |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 178 | int highbit = fls(size); /* size is in bytes */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 179 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
| 180 | } |
| 181 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 182 | static int pcpu_size_to_slot(int size) |
| 183 | { |
| 184 | if (size == pcpu_unit_size) |
| 185 | return pcpu_nr_slots - 1; |
| 186 | return __pcpu_size_to_slot(size); |
| 187 | } |
| 188 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 189 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
| 190 | { |
| 191 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) |
| 192 | return 0; |
| 193 | |
| 194 | return pcpu_size_to_slot(chunk->free_size); |
| 195 | } |
| 196 | |
| 197 | static int pcpu_page_idx(unsigned int cpu, int page_idx) |
| 198 | { |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 199 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 200 | } |
| 201 | |
| 202 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
| 203 | unsigned int cpu, int page_idx) |
| 204 | { |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 205 | return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 206 | (page_idx << PAGE_SHIFT); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 207 | } |
| 208 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 209 | static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, |
| 210 | unsigned int cpu, int page_idx) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 211 | { |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 212 | /* must not be used on pre-mapped chunk */ |
| 213 | WARN_ON(chunk->immutable); |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 214 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 215 | return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 216 | } |
| 217 | |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 218 | /* set the pointer to a chunk in a page struct */ |
| 219 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) |
| 220 | { |
| 221 | page->index = (unsigned long)pcpu; |
| 222 | } |
| 223 | |
| 224 | /* obtain pointer to a chunk from a page struct */ |
| 225 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) |
| 226 | { |
| 227 | return (struct pcpu_chunk *)page->index; |
| 228 | } |
| 229 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 230 | static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) |
| 231 | { |
| 232 | *rs = find_next_zero_bit(chunk->populated, end, *rs); |
| 233 | *re = find_next_bit(chunk->populated, end, *rs + 1); |
| 234 | } |
| 235 | |
| 236 | static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) |
| 237 | { |
| 238 | *rs = find_next_bit(chunk->populated, end, *rs); |
| 239 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); |
| 240 | } |
| 241 | |
| 242 | /* |
| 243 | * (Un)populated page region iterators. Iterate over (un)populated |
| 244 | * page regions betwen @start and @end in @chunk. @rs and @re should |
| 245 | * be integer variables and will be set to start and end page index of |
| 246 | * the current region. |
| 247 | */ |
| 248 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ |
| 249 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ |
| 250 | (rs) < (re); \ |
| 251 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) |
| 252 | |
| 253 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ |
| 254 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ |
| 255 | (rs) < (re); \ |
| 256 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) |
| 257 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 258 | /** |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 259 | * pcpu_mem_alloc - allocate memory |
| 260 | * @size: bytes to allocate |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 261 | * |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 262 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
| 263 | * kzalloc() is used; otherwise, vmalloc() is used. The returned |
| 264 | * memory is always zeroed. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 265 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 266 | * CONTEXT: |
| 267 | * Does GFP_KERNEL allocation. |
| 268 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 269 | * RETURNS: |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 270 | * Pointer to the allocated area on success, NULL on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 271 | */ |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 272 | static void *pcpu_mem_alloc(size_t size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 273 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 274 | if (size <= PAGE_SIZE) |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 275 | return kzalloc(size, GFP_KERNEL); |
| 276 | else { |
| 277 | void *ptr = vmalloc(size); |
| 278 | if (ptr) |
| 279 | memset(ptr, 0, size); |
| 280 | return ptr; |
| 281 | } |
| 282 | } |
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 | /** |
| 285 | * pcpu_mem_free - free memory |
| 286 | * @ptr: memory to free |
| 287 | * @size: size of the area |
| 288 | * |
| 289 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). |
| 290 | */ |
| 291 | static void pcpu_mem_free(void *ptr, size_t size) |
| 292 | { |
| 293 | if (size <= PAGE_SIZE) |
| 294 | kfree(ptr); |
| 295 | else |
| 296 | vfree(ptr); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 297 | } |
| 298 | |
| 299 | /** |
| 300 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot |
| 301 | * @chunk: chunk of interest |
| 302 | * @oslot: the previous slot it was on |
| 303 | * |
| 304 | * This function is called after an allocation or free changed @chunk. |
| 305 | * New slot according to the changed state is determined and @chunk is |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 306 | * moved to the slot. Note that the reserved chunk is never put on |
| 307 | * chunk slots. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 308 | * |
| 309 | * CONTEXT: |
| 310 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 311 | */ |
| 312 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) |
| 313 | { |
| 314 | int nslot = pcpu_chunk_slot(chunk); |
| 315 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 316 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 317 | if (oslot < nslot) |
| 318 | list_move(&chunk->list, &pcpu_slot[nslot]); |
| 319 | else |
| 320 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); |
| 321 | } |
| 322 | } |
| 323 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 324 | /** |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 325 | * pcpu_chunk_addr_search - determine chunk containing specified address |
| 326 | * @addr: address for which the chunk needs to be determined. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 327 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 328 | * RETURNS: |
| 329 | * The address of the found chunk. |
| 330 | */ |
| 331 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) |
| 332 | { |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 333 | void *first_start = pcpu_first_chunk->base_addr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 334 | |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 335 | /* is it in the first chunk? */ |
Tejun Heo | 79ba6ac | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 336 | if (addr >= first_start && addr < first_start + pcpu_unit_size) { |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 337 | /* is it in the reserved area? */ |
| 338 | if (addr < first_start + pcpu_reserved_chunk_limit) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 339 | return pcpu_reserved_chunk; |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 340 | return pcpu_first_chunk; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 341 | } |
| 342 | |
Tejun Heo | 04a13c7 | 2009-09-01 21:12:28 +0900 | [diff] [blame] | 343 | /* |
| 344 | * The address is relative to unit0 which might be unused and |
| 345 | * thus unmapped. Offset the address to the unit space of the |
| 346 | * current processor before looking it up in the vmalloc |
| 347 | * space. Note that any possible cpu id can be used here, so |
| 348 | * there's no need to worry about preemption or cpu hotplug. |
| 349 | */ |
Tejun Heo | 5579fd7 | 2009-09-15 09:57:19 +0900 | [diff] [blame] | 350 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 351 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 352 | } |
| 353 | |
| 354 | /** |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 355 | * pcpu_extend_area_map - extend area map for allocation |
| 356 | * @chunk: target chunk |
| 357 | * |
| 358 | * Extend area map of @chunk so that it can accomodate an allocation. |
| 359 | * A single allocation can split an area into three areas, so this |
| 360 | * function makes sure that @chunk->map has at least two extra slots. |
| 361 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 362 | * CONTEXT: |
| 363 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired |
| 364 | * if area map is extended. |
| 365 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 366 | * RETURNS: |
| 367 | * 0 if noop, 1 if successfully extended, -errno on failure. |
| 368 | */ |
| 369 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) |
| 370 | { |
| 371 | int new_alloc; |
| 372 | int *new; |
| 373 | size_t size; |
| 374 | |
| 375 | /* has enough? */ |
| 376 | if (chunk->map_alloc >= chunk->map_used + 2) |
| 377 | return 0; |
| 378 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 379 | spin_unlock_irq(&pcpu_lock); |
| 380 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 381 | new_alloc = PCPU_DFL_MAP_ALLOC; |
| 382 | while (new_alloc < chunk->map_used + 2) |
| 383 | new_alloc *= 2; |
| 384 | |
| 385 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 386 | if (!new) { |
| 387 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 388 | return -ENOMEM; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 389 | } |
| 390 | |
| 391 | /* |
| 392 | * Acquire pcpu_lock and switch to new area map. Only free |
| 393 | * could have happened inbetween, so map_used couldn't have |
| 394 | * grown. |
| 395 | */ |
| 396 | spin_lock_irq(&pcpu_lock); |
| 397 | BUG_ON(new_alloc < chunk->map_used + 2); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 398 | |
| 399 | size = chunk->map_alloc * sizeof(chunk->map[0]); |
| 400 | memcpy(new, chunk->map, size); |
| 401 | |
| 402 | /* |
| 403 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is |
| 404 | * one of the first chunks and still using static map. |
| 405 | */ |
| 406 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) |
| 407 | pcpu_mem_free(chunk->map, size); |
| 408 | |
| 409 | chunk->map_alloc = new_alloc; |
| 410 | chunk->map = new; |
| 411 | return 0; |
| 412 | } |
| 413 | |
| 414 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 415 | * pcpu_split_block - split a map block |
| 416 | * @chunk: chunk of interest |
| 417 | * @i: index of map block to split |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 418 | * @head: head size in bytes (can be 0) |
| 419 | * @tail: tail size in bytes (can be 0) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 420 | * |
| 421 | * Split the @i'th map block into two or three blocks. If @head is |
| 422 | * non-zero, @head bytes block is inserted before block @i moving it |
| 423 | * to @i+1 and reducing its size by @head bytes. |
| 424 | * |
| 425 | * If @tail is non-zero, the target block, which can be @i or @i+1 |
| 426 | * depending on @head, is reduced by @tail bytes and @tail byte block |
| 427 | * is inserted after the target block. |
| 428 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 429 | * @chunk->map must have enough free slots to accomodate the split. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 430 | * |
| 431 | * CONTEXT: |
| 432 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 433 | */ |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 434 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
| 435 | int head, int tail) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 436 | { |
| 437 | int nr_extra = !!head + !!tail; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 438 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 439 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 440 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 441 | /* insert new subblocks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 442 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
| 443 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
| 444 | chunk->map_used += nr_extra; |
| 445 | |
| 446 | if (head) { |
| 447 | chunk->map[i + 1] = chunk->map[i] - head; |
| 448 | chunk->map[i++] = head; |
| 449 | } |
| 450 | if (tail) { |
| 451 | chunk->map[i++] -= tail; |
| 452 | chunk->map[i] = tail; |
| 453 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 454 | } |
| 455 | |
| 456 | /** |
| 457 | * pcpu_alloc_area - allocate area from a pcpu_chunk |
| 458 | * @chunk: chunk of interest |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 459 | * @size: wanted size in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 460 | * @align: wanted align |
| 461 | * |
| 462 | * Try to allocate @size bytes area aligned at @align from @chunk. |
| 463 | * Note that this function only allocates the offset. It doesn't |
| 464 | * populate or map the area. |
| 465 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 466 | * @chunk->map must have at least two free slots. |
| 467 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 468 | * CONTEXT: |
| 469 | * pcpu_lock. |
| 470 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 471 | * RETURNS: |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 472 | * Allocated offset in @chunk on success, -1 if no matching area is |
| 473 | * found. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 474 | */ |
| 475 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) |
| 476 | { |
| 477 | int oslot = pcpu_chunk_slot(chunk); |
| 478 | int max_contig = 0; |
| 479 | int i, off; |
| 480 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 481 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
| 482 | bool is_last = i + 1 == chunk->map_used; |
| 483 | int head, tail; |
| 484 | |
| 485 | /* extra for alignment requirement */ |
| 486 | head = ALIGN(off, align) - off; |
| 487 | BUG_ON(i == 0 && head != 0); |
| 488 | |
| 489 | if (chunk->map[i] < 0) |
| 490 | continue; |
| 491 | if (chunk->map[i] < head + size) { |
| 492 | max_contig = max(chunk->map[i], max_contig); |
| 493 | continue; |
| 494 | } |
| 495 | |
| 496 | /* |
| 497 | * If head is small or the previous block is free, |
| 498 | * merge'em. Note that 'small' is defined as smaller |
| 499 | * than sizeof(int), which is very small but isn't too |
| 500 | * uncommon for percpu allocations. |
| 501 | */ |
| 502 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { |
| 503 | if (chunk->map[i - 1] > 0) |
| 504 | chunk->map[i - 1] += head; |
| 505 | else { |
| 506 | chunk->map[i - 1] -= head; |
| 507 | chunk->free_size -= head; |
| 508 | } |
| 509 | chunk->map[i] -= head; |
| 510 | off += head; |
| 511 | head = 0; |
| 512 | } |
| 513 | |
| 514 | /* if tail is small, just keep it around */ |
| 515 | tail = chunk->map[i] - head - size; |
| 516 | if (tail < sizeof(int)) |
| 517 | tail = 0; |
| 518 | |
| 519 | /* split if warranted */ |
| 520 | if (head || tail) { |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 521 | pcpu_split_block(chunk, i, head, tail); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 522 | if (head) { |
| 523 | i++; |
| 524 | off += head; |
| 525 | max_contig = max(chunk->map[i - 1], max_contig); |
| 526 | } |
| 527 | if (tail) |
| 528 | max_contig = max(chunk->map[i + 1], max_contig); |
| 529 | } |
| 530 | |
| 531 | /* update hint and mark allocated */ |
| 532 | if (is_last) |
| 533 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 534 | else |
| 535 | chunk->contig_hint = max(chunk->contig_hint, |
| 536 | max_contig); |
| 537 | |
| 538 | chunk->free_size -= chunk->map[i]; |
| 539 | chunk->map[i] = -chunk->map[i]; |
| 540 | |
| 541 | pcpu_chunk_relocate(chunk, oslot); |
| 542 | return off; |
| 543 | } |
| 544 | |
| 545 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 546 | pcpu_chunk_relocate(chunk, oslot); |
| 547 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 548 | /* tell the upper layer that this chunk has no matching area */ |
| 549 | return -1; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 550 | } |
| 551 | |
| 552 | /** |
| 553 | * pcpu_free_area - free area to a pcpu_chunk |
| 554 | * @chunk: chunk of interest |
| 555 | * @freeme: offset of area to free |
| 556 | * |
| 557 | * Free area starting from @freeme to @chunk. Note that this function |
| 558 | * only modifies the allocation map. It doesn't depopulate or unmap |
| 559 | * the area. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 560 | * |
| 561 | * CONTEXT: |
| 562 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 563 | */ |
| 564 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) |
| 565 | { |
| 566 | int oslot = pcpu_chunk_slot(chunk); |
| 567 | int i, off; |
| 568 | |
| 569 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) |
| 570 | if (off == freeme) |
| 571 | break; |
| 572 | BUG_ON(off != freeme); |
| 573 | BUG_ON(chunk->map[i] > 0); |
| 574 | |
| 575 | chunk->map[i] = -chunk->map[i]; |
| 576 | chunk->free_size += chunk->map[i]; |
| 577 | |
| 578 | /* merge with previous? */ |
| 579 | if (i > 0 && chunk->map[i - 1] >= 0) { |
| 580 | chunk->map[i - 1] += chunk->map[i]; |
| 581 | chunk->map_used--; |
| 582 | memmove(&chunk->map[i], &chunk->map[i + 1], |
| 583 | (chunk->map_used - i) * sizeof(chunk->map[0])); |
| 584 | i--; |
| 585 | } |
| 586 | /* merge with next? */ |
| 587 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { |
| 588 | chunk->map[i] += chunk->map[i + 1]; |
| 589 | chunk->map_used--; |
| 590 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], |
| 591 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); |
| 592 | } |
| 593 | |
| 594 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); |
| 595 | pcpu_chunk_relocate(chunk, oslot); |
| 596 | } |
| 597 | |
| 598 | /** |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 599 | * pcpu_get_pages_and_bitmap - get temp pages array and bitmap |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 600 | * @chunk: chunk of interest |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 601 | * @bitmapp: output parameter for bitmap |
| 602 | * @may_alloc: may allocate the array |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 603 | * |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 604 | * Returns pointer to array of pointers to struct page and bitmap, |
| 605 | * both of which can be indexed with pcpu_page_idx(). The returned |
| 606 | * array is cleared to zero and *@bitmapp is copied from |
| 607 | * @chunk->populated. Note that there is only one array and bitmap |
| 608 | * and access exclusion is the caller's responsibility. |
| 609 | * |
| 610 | * CONTEXT: |
| 611 | * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. |
| 612 | * Otherwise, don't care. |
| 613 | * |
| 614 | * RETURNS: |
| 615 | * Pointer to temp pages array on success, NULL on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 616 | */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 617 | static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, |
| 618 | unsigned long **bitmapp, |
| 619 | bool may_alloc) |
| 620 | { |
| 621 | static struct page **pages; |
| 622 | static unsigned long *bitmap; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 623 | size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 624 | size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * |
| 625 | sizeof(unsigned long); |
| 626 | |
| 627 | if (!pages || !bitmap) { |
| 628 | if (may_alloc && !pages) |
| 629 | pages = pcpu_mem_alloc(pages_size); |
| 630 | if (may_alloc && !bitmap) |
| 631 | bitmap = pcpu_mem_alloc(bitmap_size); |
| 632 | if (!pages || !bitmap) |
| 633 | return NULL; |
| 634 | } |
| 635 | |
| 636 | memset(pages, 0, pages_size); |
| 637 | bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); |
| 638 | |
| 639 | *bitmapp = bitmap; |
| 640 | return pages; |
| 641 | } |
| 642 | |
| 643 | /** |
| 644 | * pcpu_free_pages - free pages which were allocated for @chunk |
| 645 | * @chunk: chunk pages were allocated for |
| 646 | * @pages: array of pages to be freed, indexed by pcpu_page_idx() |
| 647 | * @populated: populated bitmap |
| 648 | * @page_start: page index of the first page to be freed |
| 649 | * @page_end: page index of the last page to be freed + 1 |
| 650 | * |
| 651 | * Free pages [@page_start and @page_end) in @pages for all units. |
| 652 | * The pages were allocated for @chunk. |
| 653 | */ |
| 654 | static void pcpu_free_pages(struct pcpu_chunk *chunk, |
| 655 | struct page **pages, unsigned long *populated, |
| 656 | int page_start, int page_end) |
| 657 | { |
| 658 | unsigned int cpu; |
| 659 | int i; |
| 660 | |
| 661 | for_each_possible_cpu(cpu) { |
| 662 | for (i = page_start; i < page_end; i++) { |
| 663 | struct page *page = pages[pcpu_page_idx(cpu, i)]; |
| 664 | |
| 665 | if (page) |
| 666 | __free_page(page); |
| 667 | } |
| 668 | } |
| 669 | } |
| 670 | |
| 671 | /** |
| 672 | * pcpu_alloc_pages - allocates pages for @chunk |
| 673 | * @chunk: target chunk |
| 674 | * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() |
| 675 | * @populated: populated bitmap |
| 676 | * @page_start: page index of the first page to be allocated |
| 677 | * @page_end: page index of the last page to be allocated + 1 |
| 678 | * |
| 679 | * Allocate pages [@page_start,@page_end) into @pages for all units. |
| 680 | * The allocation is for @chunk. Percpu core doesn't care about the |
| 681 | * content of @pages and will pass it verbatim to pcpu_map_pages(). |
| 682 | */ |
| 683 | static int pcpu_alloc_pages(struct pcpu_chunk *chunk, |
| 684 | struct page **pages, unsigned long *populated, |
| 685 | int page_start, int page_end) |
| 686 | { |
| 687 | const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; |
| 688 | unsigned int cpu; |
| 689 | int i; |
| 690 | |
| 691 | for_each_possible_cpu(cpu) { |
| 692 | for (i = page_start; i < page_end; i++) { |
| 693 | struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; |
| 694 | |
| 695 | *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); |
| 696 | if (!*pagep) { |
| 697 | pcpu_free_pages(chunk, pages, populated, |
| 698 | page_start, page_end); |
| 699 | return -ENOMEM; |
| 700 | } |
| 701 | } |
| 702 | } |
| 703 | return 0; |
| 704 | } |
| 705 | |
| 706 | /** |
| 707 | * pcpu_pre_unmap_flush - flush cache prior to unmapping |
| 708 | * @chunk: chunk the regions to be flushed belongs to |
| 709 | * @page_start: page index of the first page to be flushed |
| 710 | * @page_end: page index of the last page to be flushed + 1 |
| 711 | * |
| 712 | * Pages in [@page_start,@page_end) of @chunk are about to be |
| 713 | * unmapped. Flush cache. As each flushing trial can be very |
| 714 | * expensive, issue flush on the whole region at once rather than |
| 715 | * doing it for each cpu. This could be an overkill but is more |
| 716 | * scalable. |
| 717 | */ |
| 718 | static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, |
| 719 | int page_start, int page_end) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 720 | { |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 721 | flush_cache_vunmap( |
| 722 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), |
| 723 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 724 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 725 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 726 | static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) |
| 727 | { |
| 728 | unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); |
| 729 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 730 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 731 | /** |
| 732 | * pcpu_unmap_pages - unmap pages out of a pcpu_chunk |
| 733 | * @chunk: chunk of interest |
| 734 | * @pages: pages array which can be used to pass information to free |
| 735 | * @populated: populated bitmap |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 736 | * @page_start: page index of the first page to unmap |
| 737 | * @page_end: page index of the last page to unmap + 1 |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 738 | * |
| 739 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 740 | * Corresponding elements in @pages were cleared by the caller and can |
| 741 | * be used to carry information to pcpu_free_pages() which will be |
| 742 | * called after all unmaps are finished. The caller should call |
| 743 | * proper pre/post flush functions. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 744 | */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 745 | static void pcpu_unmap_pages(struct pcpu_chunk *chunk, |
| 746 | struct page **pages, unsigned long *populated, |
| 747 | int page_start, int page_end) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 748 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 749 | unsigned int cpu; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 750 | int i; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 751 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 752 | for_each_possible_cpu(cpu) { |
| 753 | for (i = page_start; i < page_end; i++) { |
| 754 | struct page *page; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 755 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 756 | page = pcpu_chunk_page(chunk, cpu, i); |
| 757 | WARN_ON(!page); |
| 758 | pages[pcpu_page_idx(cpu, i)] = page; |
| 759 | } |
| 760 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), |
| 761 | page_end - page_start); |
| 762 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 763 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 764 | for (i = page_start; i < page_end; i++) |
| 765 | __clear_bit(i, populated); |
| 766 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 767 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 768 | /** |
| 769 | * pcpu_post_unmap_tlb_flush - flush TLB after unmapping |
| 770 | * @chunk: pcpu_chunk the regions to be flushed belong to |
| 771 | * @page_start: page index of the first page to be flushed |
| 772 | * @page_end: page index of the last page to be flushed + 1 |
| 773 | * |
| 774 | * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush |
| 775 | * TLB for the regions. This can be skipped if the area is to be |
| 776 | * returned to vmalloc as vmalloc will handle TLB flushing lazily. |
| 777 | * |
| 778 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once |
| 779 | * for the whole region. |
| 780 | */ |
| 781 | static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, |
| 782 | int page_start, int page_end) |
| 783 | { |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 784 | flush_tlb_kernel_range( |
| 785 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), |
| 786 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 787 | } |
| 788 | |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 789 | static int __pcpu_map_pages(unsigned long addr, struct page **pages, |
| 790 | int nr_pages) |
| 791 | { |
| 792 | return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, |
| 793 | PAGE_KERNEL, pages); |
| 794 | } |
| 795 | |
| 796 | /** |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 797 | * pcpu_map_pages - map pages into a pcpu_chunk |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 798 | * @chunk: chunk of interest |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 799 | * @pages: pages array containing pages to be mapped |
| 800 | * @populated: populated bitmap |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 801 | * @page_start: page index of the first page to map |
| 802 | * @page_end: page index of the last page to map + 1 |
| 803 | * |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 804 | * For each cpu, map pages [@page_start,@page_end) into @chunk. The |
| 805 | * caller is responsible for calling pcpu_post_map_flush() after all |
| 806 | * mappings are complete. |
| 807 | * |
| 808 | * This function is responsible for setting corresponding bits in |
| 809 | * @chunk->populated bitmap and whatever is necessary for reverse |
| 810 | * lookup (addr -> chunk). |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 811 | */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 812 | static int pcpu_map_pages(struct pcpu_chunk *chunk, |
| 813 | struct page **pages, unsigned long *populated, |
| 814 | int page_start, int page_end) |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 815 | { |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 816 | unsigned int cpu, tcpu; |
| 817 | int i, err; |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 818 | |
| 819 | for_each_possible_cpu(cpu) { |
| 820 | err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 821 | &pages[pcpu_page_idx(cpu, page_start)], |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 822 | page_end - page_start); |
| 823 | if (err < 0) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 824 | goto err; |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 825 | } |
| 826 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 827 | /* mapping successful, link chunk and mark populated */ |
| 828 | for (i = page_start; i < page_end; i++) { |
| 829 | for_each_possible_cpu(cpu) |
| 830 | pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], |
| 831 | chunk); |
| 832 | __set_bit(i, populated); |
| 833 | } |
| 834 | |
| 835 | return 0; |
| 836 | |
| 837 | err: |
| 838 | for_each_possible_cpu(tcpu) { |
| 839 | if (tcpu == cpu) |
| 840 | break; |
| 841 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), |
| 842 | page_end - page_start); |
| 843 | } |
| 844 | return err; |
| 845 | } |
| 846 | |
| 847 | /** |
| 848 | * pcpu_post_map_flush - flush cache after mapping |
| 849 | * @chunk: pcpu_chunk the regions to be flushed belong to |
| 850 | * @page_start: page index of the first page to be flushed |
| 851 | * @page_end: page index of the last page to be flushed + 1 |
| 852 | * |
| 853 | * Pages [@page_start,@page_end) of @chunk have been mapped. Flush |
| 854 | * cache. |
| 855 | * |
| 856 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once |
| 857 | * for the whole region. |
| 858 | */ |
| 859 | static void pcpu_post_map_flush(struct pcpu_chunk *chunk, |
| 860 | int page_start, int page_end) |
| 861 | { |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 862 | flush_cache_vmap( |
| 863 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), |
| 864 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 865 | } |
| 866 | |
| 867 | /** |
| 868 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk |
| 869 | * @chunk: chunk to depopulate |
| 870 | * @off: offset to the area to depopulate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 871 | * @size: size of the area to depopulate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 872 | * @flush: whether to flush cache and tlb or not |
| 873 | * |
| 874 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) |
| 875 | * from @chunk. If @flush is true, vcache is flushed before unmapping |
| 876 | * and tlb after. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 877 | * |
| 878 | * CONTEXT: |
| 879 | * pcpu_alloc_mutex. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 880 | */ |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 881 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 882 | { |
| 883 | int page_start = PFN_DOWN(off); |
| 884 | int page_end = PFN_UP(off + size); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 885 | struct page **pages; |
| 886 | unsigned long *populated; |
| 887 | int rs, re; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 888 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 889 | /* quick path, check whether it's empty already */ |
| 890 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
| 891 | if (rs == page_start && re == page_end) |
| 892 | return; |
| 893 | break; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 894 | } |
| 895 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 896 | /* immutable chunks can't be depopulated */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 897 | WARN_ON(chunk->immutable); |
| 898 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 899 | /* |
| 900 | * If control reaches here, there must have been at least one |
| 901 | * successful population attempt so the temp pages array must |
| 902 | * be available now. |
| 903 | */ |
| 904 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); |
| 905 | BUG_ON(!pages); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 906 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 907 | /* unmap and free */ |
| 908 | pcpu_pre_unmap_flush(chunk, page_start, page_end); |
| 909 | |
| 910 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) |
| 911 | pcpu_unmap_pages(chunk, pages, populated, rs, re); |
| 912 | |
| 913 | /* no need to flush tlb, vmalloc will handle it lazily */ |
| 914 | |
| 915 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) |
| 916 | pcpu_free_pages(chunk, pages, populated, rs, re); |
| 917 | |
| 918 | /* commit new bitmap */ |
| 919 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 920 | } |
| 921 | |
| 922 | /** |
| 923 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk |
| 924 | * @chunk: chunk of interest |
| 925 | * @off: offset to the area to populate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 926 | * @size: size of the area to populate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 927 | * |
| 928 | * For each cpu, populate and map pages [@page_start,@page_end) into |
| 929 | * @chunk. The area is cleared on return. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 930 | * |
| 931 | * CONTEXT: |
| 932 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 933 | */ |
| 934 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) |
| 935 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 936 | int page_start = PFN_DOWN(off); |
| 937 | int page_end = PFN_UP(off + size); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 938 | int free_end = page_start, unmap_end = page_start; |
| 939 | struct page **pages; |
| 940 | unsigned long *populated; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 941 | unsigned int cpu; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 942 | int rs, re, rc; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 943 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 944 | /* quick path, check whether all pages are already there */ |
| 945 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { |
| 946 | if (rs == page_start && re == page_end) |
| 947 | goto clear; |
| 948 | break; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 949 | } |
| 950 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 951 | /* need to allocate and map pages, this chunk can't be immutable */ |
| 952 | WARN_ON(chunk->immutable); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 953 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 954 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); |
| 955 | if (!pages) |
| 956 | return -ENOMEM; |
| 957 | |
| 958 | /* alloc and map */ |
| 959 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
| 960 | rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); |
| 961 | if (rc) |
| 962 | goto err_free; |
| 963 | free_end = re; |
| 964 | } |
| 965 | |
| 966 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
| 967 | rc = pcpu_map_pages(chunk, pages, populated, rs, re); |
| 968 | if (rc) |
| 969 | goto err_unmap; |
| 970 | unmap_end = re; |
| 971 | } |
| 972 | pcpu_post_map_flush(chunk, page_start, page_end); |
| 973 | |
| 974 | /* commit new bitmap */ |
| 975 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); |
| 976 | clear: |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 977 | for_each_possible_cpu(cpu) |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 978 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 979 | return 0; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 980 | |
| 981 | err_unmap: |
| 982 | pcpu_pre_unmap_flush(chunk, page_start, unmap_end); |
| 983 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) |
| 984 | pcpu_unmap_pages(chunk, pages, populated, rs, re); |
| 985 | pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); |
| 986 | err_free: |
| 987 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) |
| 988 | pcpu_free_pages(chunk, pages, populated, rs, re); |
| 989 | return rc; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 990 | } |
| 991 | |
| 992 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) |
| 993 | { |
| 994 | if (!chunk) |
| 995 | return; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 996 | if (chunk->vms) |
| 997 | pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups); |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 998 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 999 | kfree(chunk); |
| 1000 | } |
| 1001 | |
| 1002 | static struct pcpu_chunk *alloc_pcpu_chunk(void) |
| 1003 | { |
| 1004 | struct pcpu_chunk *chunk; |
| 1005 | |
| 1006 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); |
| 1007 | if (!chunk) |
| 1008 | return NULL; |
| 1009 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 1010 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1011 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
| 1012 | chunk->map[chunk->map_used++] = pcpu_unit_size; |
| 1013 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1014 | chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes, |
| 1015 | pcpu_nr_groups, pcpu_atom_size, |
| 1016 | GFP_KERNEL); |
| 1017 | if (!chunk->vms) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1018 | free_pcpu_chunk(chunk); |
| 1019 | return NULL; |
| 1020 | } |
| 1021 | |
| 1022 | INIT_LIST_HEAD(&chunk->list); |
| 1023 | chunk->free_size = pcpu_unit_size; |
| 1024 | chunk->contig_hint = pcpu_unit_size; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1025 | chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0]; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1026 | |
| 1027 | return chunk; |
| 1028 | } |
| 1029 | |
| 1030 | /** |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1031 | * pcpu_alloc - the percpu allocator |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 1032 | * @size: size of area to allocate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1033 | * @align: alignment of area (max PAGE_SIZE) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1034 | * @reserved: allocate from the reserved chunk if available |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1035 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1036 | * Allocate percpu area of @size bytes aligned at @align. |
| 1037 | * |
| 1038 | * CONTEXT: |
| 1039 | * Does GFP_KERNEL allocation. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1040 | * |
| 1041 | * RETURNS: |
| 1042 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 1043 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1044 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1045 | { |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1046 | static int warn_limit = 10; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1047 | struct pcpu_chunk *chunk; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1048 | const char *err; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1049 | int slot, off; |
| 1050 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1051 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1052 | WARN(true, "illegal size (%zu) or align (%zu) for " |
| 1053 | "percpu allocation\n", size, align); |
| 1054 | return NULL; |
| 1055 | } |
| 1056 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1057 | mutex_lock(&pcpu_alloc_mutex); |
| 1058 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1059 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1060 | /* serve reserved allocations from the reserved chunk if available */ |
| 1061 | if (reserved && pcpu_reserved_chunk) { |
| 1062 | chunk = pcpu_reserved_chunk; |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 1063 | if (size > chunk->contig_hint || |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1064 | pcpu_extend_area_map(chunk) < 0) { |
| 1065 | err = "failed to extend area map of reserved chunk"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1066 | goto fail_unlock; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1067 | } |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1068 | off = pcpu_alloc_area(chunk, size, align); |
| 1069 | if (off >= 0) |
| 1070 | goto area_found; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1071 | err = "alloc from reserved chunk failed"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1072 | goto fail_unlock; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1073 | } |
| 1074 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1075 | restart: |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1076 | /* search through normal chunks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1077 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
| 1078 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 1079 | if (size > chunk->contig_hint) |
| 1080 | continue; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1081 | |
| 1082 | switch (pcpu_extend_area_map(chunk)) { |
| 1083 | case 0: |
| 1084 | break; |
| 1085 | case 1: |
| 1086 | goto restart; /* pcpu_lock dropped, restart */ |
| 1087 | default: |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1088 | err = "failed to extend area map"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1089 | goto fail_unlock; |
| 1090 | } |
| 1091 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1092 | off = pcpu_alloc_area(chunk, size, align); |
| 1093 | if (off >= 0) |
| 1094 | goto area_found; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1095 | } |
| 1096 | } |
| 1097 | |
| 1098 | /* hmmm... no space left, create a new chunk */ |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1099 | spin_unlock_irq(&pcpu_lock); |
| 1100 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1101 | chunk = alloc_pcpu_chunk(); |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1102 | if (!chunk) { |
| 1103 | err = "failed to allocate new chunk"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1104 | goto fail_unlock_mutex; |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1105 | } |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1106 | |
| 1107 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1108 | pcpu_chunk_relocate(chunk, -1); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1109 | goto restart; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1110 | |
| 1111 | area_found: |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1112 | spin_unlock_irq(&pcpu_lock); |
| 1113 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1114 | /* populate, map and clear the area */ |
| 1115 | if (pcpu_populate_chunk(chunk, off, size)) { |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1116 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1117 | pcpu_free_area(chunk, off); |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1118 | err = "failed to populate"; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1119 | goto fail_unlock; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1120 | } |
| 1121 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1122 | mutex_unlock(&pcpu_alloc_mutex); |
| 1123 | |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1124 | /* return address relative to base address */ |
| 1125 | return __addr_to_pcpu_ptr(chunk->base_addr + off); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1126 | |
| 1127 | fail_unlock: |
| 1128 | spin_unlock_irq(&pcpu_lock); |
| 1129 | fail_unlock_mutex: |
| 1130 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | f2badb0 | 2009-09-29 09:17:58 +0900 | [diff] [blame] | 1131 | if (warn_limit) { |
| 1132 | pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " |
| 1133 | "%s\n", size, align, err); |
| 1134 | dump_stack(); |
| 1135 | if (!--warn_limit) |
| 1136 | pr_info("PERCPU: limit reached, disable warning\n"); |
| 1137 | } |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1138 | return NULL; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1139 | } |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1140 | |
| 1141 | /** |
| 1142 | * __alloc_percpu - allocate dynamic percpu area |
| 1143 | * @size: size of area to allocate in bytes |
| 1144 | * @align: alignment of area (max PAGE_SIZE) |
| 1145 | * |
| 1146 | * Allocate percpu area of @size bytes aligned at @align. Might |
| 1147 | * sleep. Might trigger writeouts. |
| 1148 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1149 | * CONTEXT: |
| 1150 | * Does GFP_KERNEL allocation. |
| 1151 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1152 | * RETURNS: |
| 1153 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 1154 | */ |
| 1155 | void *__alloc_percpu(size_t size, size_t align) |
| 1156 | { |
| 1157 | return pcpu_alloc(size, align, false); |
| 1158 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1159 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
| 1160 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1161 | /** |
| 1162 | * __alloc_reserved_percpu - allocate reserved percpu area |
| 1163 | * @size: size of area to allocate in bytes |
| 1164 | * @align: alignment of area (max PAGE_SIZE) |
| 1165 | * |
| 1166 | * Allocate percpu area of @size bytes aligned at @align from reserved |
| 1167 | * percpu area if arch has set it up; otherwise, allocation is served |
| 1168 | * from the same dynamic area. Might sleep. Might trigger writeouts. |
| 1169 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1170 | * CONTEXT: |
| 1171 | * Does GFP_KERNEL allocation. |
| 1172 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1173 | * RETURNS: |
| 1174 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 1175 | */ |
| 1176 | void *__alloc_reserved_percpu(size_t size, size_t align) |
| 1177 | { |
| 1178 | return pcpu_alloc(size, align, true); |
| 1179 | } |
| 1180 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1181 | /** |
| 1182 | * pcpu_reclaim - reclaim fully free chunks, workqueue function |
| 1183 | * @work: unused |
| 1184 | * |
| 1185 | * Reclaim all fully free chunks except for the first one. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1186 | * |
| 1187 | * CONTEXT: |
| 1188 | * workqueue context. |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1189 | */ |
| 1190 | static void pcpu_reclaim(struct work_struct *work) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1191 | { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1192 | LIST_HEAD(todo); |
| 1193 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; |
| 1194 | struct pcpu_chunk *chunk, *next; |
| 1195 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1196 | mutex_lock(&pcpu_alloc_mutex); |
| 1197 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1198 | |
| 1199 | list_for_each_entry_safe(chunk, next, head, list) { |
| 1200 | WARN_ON(chunk->immutable); |
| 1201 | |
| 1202 | /* spare the first one */ |
| 1203 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) |
| 1204 | continue; |
| 1205 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1206 | list_move(&chunk->list, &todo); |
| 1207 | } |
| 1208 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1209 | spin_unlock_irq(&pcpu_lock); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1210 | |
| 1211 | list_for_each_entry_safe(chunk, next, &todo, list) { |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1212 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1213 | free_pcpu_chunk(chunk); |
| 1214 | } |
Tejun Heo | 971f391 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1215 | |
| 1216 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1217 | } |
| 1218 | |
| 1219 | /** |
| 1220 | * free_percpu - free percpu area |
| 1221 | * @ptr: pointer to area to free |
| 1222 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1223 | * Free percpu area @ptr. |
| 1224 | * |
| 1225 | * CONTEXT: |
| 1226 | * Can be called from atomic context. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1227 | */ |
| 1228 | void free_percpu(void *ptr) |
| 1229 | { |
| 1230 | void *addr = __pcpu_ptr_to_addr(ptr); |
| 1231 | struct pcpu_chunk *chunk; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1232 | unsigned long flags; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1233 | int off; |
| 1234 | |
| 1235 | if (!ptr) |
| 1236 | return; |
| 1237 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1238 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1239 | |
| 1240 | chunk = pcpu_chunk_addr_search(addr); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1241 | off = addr - chunk->base_addr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1242 | |
| 1243 | pcpu_free_area(chunk, off); |
| 1244 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1245 | /* 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] | 1246 | if (chunk->free_size == pcpu_unit_size) { |
| 1247 | struct pcpu_chunk *pos; |
| 1248 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1249 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1250 | if (pos != chunk) { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 1251 | schedule_work(&pcpu_reclaim_work); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1252 | break; |
| 1253 | } |
| 1254 | } |
| 1255 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 1256 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1257 | } |
| 1258 | EXPORT_SYMBOL_GPL(free_percpu); |
| 1259 | |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1260 | static inline size_t pcpu_calc_fc_sizes(size_t static_size, |
| 1261 | size_t reserved_size, |
| 1262 | ssize_t *dyn_sizep) |
| 1263 | { |
| 1264 | size_t size_sum; |
| 1265 | |
| 1266 | size_sum = PFN_ALIGN(static_size + reserved_size + |
| 1267 | (*dyn_sizep >= 0 ? *dyn_sizep : 0)); |
| 1268 | if (*dyn_sizep != 0) |
| 1269 | *dyn_sizep = size_sum - static_size - reserved_size; |
| 1270 | |
| 1271 | return size_sum; |
| 1272 | } |
| 1273 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1274 | /** |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1275 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
| 1276 | * @nr_groups: the number of groups |
| 1277 | * @nr_units: the number of units |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1278 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1279 | * Allocate ai which is large enough for @nr_groups groups containing |
| 1280 | * @nr_units units. The returned ai's groups[0].cpu_map points to the |
| 1281 | * cpu_map array which is long enough for @nr_units and filled with |
| 1282 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map |
| 1283 | * pointer of other groups. |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1284 | * |
| 1285 | * RETURNS: |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1286 | * Pointer to the allocated pcpu_alloc_info on success, NULL on |
| 1287 | * failure. |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1288 | */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1289 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, |
| 1290 | int nr_units) |
| 1291 | { |
| 1292 | struct pcpu_alloc_info *ai; |
| 1293 | size_t base_size, ai_size; |
| 1294 | void *ptr; |
| 1295 | int unit; |
| 1296 | |
| 1297 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), |
| 1298 | __alignof__(ai->groups[0].cpu_map[0])); |
| 1299 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); |
| 1300 | |
| 1301 | ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); |
| 1302 | if (!ptr) |
| 1303 | return NULL; |
| 1304 | ai = ptr; |
| 1305 | ptr += base_size; |
| 1306 | |
| 1307 | ai->groups[0].cpu_map = ptr; |
| 1308 | |
| 1309 | for (unit = 0; unit < nr_units; unit++) |
| 1310 | ai->groups[0].cpu_map[unit] = NR_CPUS; |
| 1311 | |
| 1312 | ai->nr_groups = nr_groups; |
| 1313 | ai->__ai_size = PFN_ALIGN(ai_size); |
| 1314 | |
| 1315 | return ai; |
| 1316 | } |
| 1317 | |
| 1318 | /** |
| 1319 | * pcpu_free_alloc_info - free percpu allocation info |
| 1320 | * @ai: pcpu_alloc_info to free |
| 1321 | * |
| 1322 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). |
| 1323 | */ |
| 1324 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) |
| 1325 | { |
| 1326 | free_bootmem(__pa(ai), ai->__ai_size); |
| 1327 | } |
| 1328 | |
| 1329 | /** |
| 1330 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1331 | * @reserved_size: the size of reserved percpu area in bytes |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1332 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1333 | * @atom_size: allocation atom size |
| 1334 | * @cpu_distance_fn: callback to determine distance between cpus, optional |
| 1335 | * |
| 1336 | * This function determines grouping of units, their mappings to cpus |
| 1337 | * and other parameters considering needed percpu size, allocation |
| 1338 | * atom size and distances between CPUs. |
| 1339 | * |
| 1340 | * Groups are always mutliples of atom size and CPUs which are of |
| 1341 | * LOCAL_DISTANCE both ways are grouped together and share space for |
| 1342 | * units in the same group. The returned configuration is guaranteed |
| 1343 | * to have CPUs on different nodes on different groups and >=75% usage |
| 1344 | * of allocated virtual address space. |
| 1345 | * |
| 1346 | * RETURNS: |
| 1347 | * On success, pointer to the new allocation_info is returned. On |
| 1348 | * failure, ERR_PTR value is returned. |
| 1349 | */ |
| 1350 | struct pcpu_alloc_info * __init pcpu_build_alloc_info( |
| 1351 | size_t reserved_size, ssize_t dyn_size, |
| 1352 | size_t atom_size, |
| 1353 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1354 | { |
| 1355 | static int group_map[NR_CPUS] __initdata; |
| 1356 | static int group_cnt[NR_CPUS] __initdata; |
| 1357 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1358 | int group_cnt_max = 0, nr_groups = 1, nr_units = 0; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1359 | size_t size_sum, min_unit_size, alloc_size; |
| 1360 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1361 | int last_allocs, group, unit; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1362 | unsigned int cpu, tcpu; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1363 | struct pcpu_alloc_info *ai; |
| 1364 | unsigned int *cpu_map; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1365 | |
Tejun Heo | fb59e72 | 2009-09-24 18:50:34 +0900 | [diff] [blame] | 1366 | /* this function may be called multiple times */ |
| 1367 | memset(group_map, 0, sizeof(group_map)); |
| 1368 | memset(group_cnt, 0, sizeof(group_map)); |
| 1369 | |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1370 | /* |
| 1371 | * Determine min_unit_size, alloc_size and max_upa such that |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1372 | * alloc_size is multiple of atom_size and is the smallest |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1373 | * which can accomodate 4k aligned segments which are equal to |
| 1374 | * or larger than min_unit_size. |
| 1375 | */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1376 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1377 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
| 1378 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1379 | alloc_size = roundup(min_unit_size, atom_size); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1380 | upa = alloc_size / min_unit_size; |
| 1381 | while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) |
| 1382 | upa--; |
| 1383 | max_upa = upa; |
| 1384 | |
| 1385 | /* group cpus according to their proximity */ |
| 1386 | for_each_possible_cpu(cpu) { |
| 1387 | group = 0; |
| 1388 | next_group: |
| 1389 | for_each_possible_cpu(tcpu) { |
| 1390 | if (cpu == tcpu) |
| 1391 | break; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1392 | if (group_map[tcpu] == group && cpu_distance_fn && |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1393 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || |
| 1394 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { |
| 1395 | group++; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1396 | nr_groups = max(nr_groups, group + 1); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1397 | goto next_group; |
| 1398 | } |
| 1399 | } |
| 1400 | group_map[cpu] = group; |
| 1401 | group_cnt[group]++; |
| 1402 | group_cnt_max = max(group_cnt_max, group_cnt[group]); |
| 1403 | } |
| 1404 | |
| 1405 | /* |
| 1406 | * Expand unit size until address space usage goes over 75% |
| 1407 | * and then as much as possible without using more address |
| 1408 | * space. |
| 1409 | */ |
| 1410 | last_allocs = INT_MAX; |
| 1411 | for (upa = max_upa; upa; upa--) { |
| 1412 | int allocs = 0, wasted = 0; |
| 1413 | |
| 1414 | if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) |
| 1415 | continue; |
| 1416 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1417 | for (group = 0; group < nr_groups; group++) { |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1418 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); |
| 1419 | allocs += this_allocs; |
| 1420 | wasted += this_allocs * upa - group_cnt[group]; |
| 1421 | } |
| 1422 | |
| 1423 | /* |
| 1424 | * Don't accept if wastage is over 25%. The |
| 1425 | * greater-than comparison ensures upa==1 always |
| 1426 | * passes the following check. |
| 1427 | */ |
| 1428 | if (wasted > num_possible_cpus() / 3) |
| 1429 | continue; |
| 1430 | |
| 1431 | /* and then don't consume more memory */ |
| 1432 | if (allocs > last_allocs) |
| 1433 | break; |
| 1434 | last_allocs = allocs; |
| 1435 | best_upa = upa; |
| 1436 | } |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1437 | upa = best_upa; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1438 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1439 | /* allocate and fill alloc_info */ |
| 1440 | for (group = 0; group < nr_groups; group++) |
| 1441 | nr_units += roundup(group_cnt[group], upa); |
| 1442 | |
| 1443 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); |
| 1444 | if (!ai) |
| 1445 | return ERR_PTR(-ENOMEM); |
| 1446 | cpu_map = ai->groups[0].cpu_map; |
| 1447 | |
| 1448 | for (group = 0; group < nr_groups; group++) { |
| 1449 | ai->groups[group].cpu_map = cpu_map; |
| 1450 | cpu_map += roundup(group_cnt[group], upa); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1451 | } |
| 1452 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1453 | ai->static_size = static_size; |
| 1454 | ai->reserved_size = reserved_size; |
| 1455 | ai->dyn_size = dyn_size; |
| 1456 | ai->unit_size = alloc_size / upa; |
| 1457 | ai->atom_size = atom_size; |
| 1458 | ai->alloc_size = alloc_size; |
| 1459 | |
| 1460 | for (group = 0, unit = 0; group_cnt[group]; group++) { |
| 1461 | struct pcpu_group_info *gi = &ai->groups[group]; |
| 1462 | |
| 1463 | /* |
| 1464 | * Initialize base_offset as if all groups are located |
| 1465 | * back-to-back. The caller should update this to |
| 1466 | * reflect actual allocation. |
| 1467 | */ |
| 1468 | gi->base_offset = unit * ai->unit_size; |
| 1469 | |
| 1470 | for_each_possible_cpu(cpu) |
| 1471 | if (group_map[cpu] == group) |
| 1472 | gi->cpu_map[gi->nr_units++] = cpu; |
| 1473 | gi->nr_units = roundup(gi->nr_units, upa); |
| 1474 | unit += gi->nr_units; |
| 1475 | } |
| 1476 | BUG_ON(unit != nr_units); |
| 1477 | |
| 1478 | return ai; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1479 | } |
| 1480 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1481 | /** |
| 1482 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info |
| 1483 | * @lvl: loglevel |
| 1484 | * @ai: allocation info to dump |
| 1485 | * |
| 1486 | * Print out information about @ai using loglevel @lvl. |
| 1487 | */ |
| 1488 | static void pcpu_dump_alloc_info(const char *lvl, |
| 1489 | const struct pcpu_alloc_info *ai) |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1490 | { |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1491 | int group_width = 1, cpu_width = 1, width; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1492 | char empty_str[] = "--------"; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1493 | int alloc = 0, alloc_end = 0; |
| 1494 | int group, v; |
| 1495 | int upa, apl; /* units per alloc, allocs per line */ |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1496 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1497 | v = ai->nr_groups; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1498 | while (v /= 10) |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1499 | group_width++; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1500 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1501 | v = num_possible_cpus(); |
| 1502 | while (v /= 10) |
| 1503 | cpu_width++; |
| 1504 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1505 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1506 | upa = ai->alloc_size / ai->unit_size; |
| 1507 | width = upa * (cpu_width + 1) + group_width + 3; |
| 1508 | apl = rounddown_pow_of_two(max(60 / width, 1)); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1509 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1510 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
| 1511 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, |
| 1512 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); |
| 1513 | |
| 1514 | for (group = 0; group < ai->nr_groups; group++) { |
| 1515 | const struct pcpu_group_info *gi = &ai->groups[group]; |
| 1516 | int unit = 0, unit_end = 0; |
| 1517 | |
| 1518 | BUG_ON(gi->nr_units % upa); |
| 1519 | for (alloc_end += gi->nr_units / upa; |
| 1520 | alloc < alloc_end; alloc++) { |
| 1521 | if (!(alloc % apl)) { |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1522 | printk("\n"); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1523 | printk("%spcpu-alloc: ", lvl); |
| 1524 | } |
| 1525 | printk("[%0*d] ", group_width, group); |
| 1526 | |
| 1527 | for (unit_end += upa; unit < unit_end; unit++) |
| 1528 | if (gi->cpu_map[unit] != NR_CPUS) |
| 1529 | printk("%0*d ", cpu_width, |
| 1530 | gi->cpu_map[unit]); |
| 1531 | else |
| 1532 | printk("%s ", empty_str); |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1533 | } |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1534 | } |
| 1535 | printk("\n"); |
| 1536 | } |
Tejun Heo | 033e48f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1537 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1538 | /** |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1539 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1540 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1541 | * @base_addr: mapped address |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1542 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1543 | * Initialize the first percpu chunk which contains the kernel static |
| 1544 | * perpcu area. This function is to be called from arch percpu area |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1545 | * setup path. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1546 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1547 | * @ai contains all information necessary to initialize the first |
| 1548 | * chunk and prime the dynamic percpu allocator. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1549 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1550 | * @ai->static_size is the size of static percpu area. |
| 1551 | * |
| 1552 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1553 | * reserve after the static area in the first chunk. This reserves |
| 1554 | * the first chunk such that it's available only through reserved |
| 1555 | * percpu allocation. This is primarily used to serve module percpu |
| 1556 | * static areas on architectures where the addressing model has |
| 1557 | * limited offset range for symbol relocations to guarantee module |
| 1558 | * percpu symbols fall inside the relocatable range. |
| 1559 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1560 | * @ai->dyn_size determines the number of bytes available for dynamic |
| 1561 | * allocation in the first chunk. The area between @ai->static_size + |
| 1562 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. |
Tejun Heo | 6074d5b | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1563 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1564 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
| 1565 | * and equal to or larger than @ai->static_size + @ai->reserved_size + |
| 1566 | * @ai->dyn_size. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1567 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1568 | * @ai->atom_size is the allocation atom size and used as alignment |
| 1569 | * for vm areas. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1570 | * |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1571 | * @ai->alloc_size is the allocation size and always multiple of |
| 1572 | * @ai->atom_size. This is larger than @ai->atom_size if |
| 1573 | * @ai->unit_size is larger than @ai->atom_size. |
| 1574 | * |
| 1575 | * @ai->nr_groups and @ai->groups describe virtual memory layout of |
| 1576 | * percpu areas. Units which should be colocated are put into the |
| 1577 | * same group. Dynamic VM areas will be allocated according to these |
| 1578 | * groupings. If @ai->nr_groups is zero, a single group containing |
| 1579 | * all units is assumed. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1580 | * |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1581 | * The caller should have mapped the first chunk at @base_addr and |
| 1582 | * copied static data to each unit. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1583 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1584 | * If the first chunk ends up with both reserved and dynamic areas, it |
| 1585 | * is served by two chunks - one to serve the core static and reserved |
| 1586 | * areas and the other for the dynamic area. They share the same vm |
| 1587 | * and page map but uses different area allocation map to stay away |
| 1588 | * from each other. The latter chunk is circulated in the chunk slots |
| 1589 | * and available for dynamic allocation like any other chunks. |
| 1590 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1591 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1592 | * 0 on success, -errno on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1593 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1594 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
| 1595 | void *base_addr) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1596 | { |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1597 | static char cpus_buf[4096] __initdata; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1598 | static int smap[2], dmap[2]; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1599 | size_t dyn_size = ai->dyn_size; |
| 1600 | size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1601 | struct pcpu_chunk *schunk, *dchunk = NULL; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1602 | unsigned long *group_offsets; |
| 1603 | size_t *group_sizes; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1604 | unsigned long *unit_off; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1605 | unsigned int cpu; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1606 | int *unit_map; |
| 1607 | int group, unit, i; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1608 | |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1609 | cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); |
| 1610 | |
| 1611 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
| 1612 | if (unlikely(cond)) { \ |
| 1613 | pr_emerg("PERCPU: failed to initialize, %s", #cond); \ |
| 1614 | pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ |
| 1615 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
| 1616 | BUG(); \ |
| 1617 | } \ |
| 1618 | } while (0) |
| 1619 | |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1620 | /* sanity checks */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1621 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
| 1622 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1623 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
| 1624 | PCPU_SETUP_BUG_ON(!ai->static_size); |
| 1625 | PCPU_SETUP_BUG_ON(!base_addr); |
| 1626 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
| 1627 | PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); |
| 1628 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1629 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1630 | /* process group information and build config tables accordingly */ |
| 1631 | group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); |
| 1632 | group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1633 | unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1634 | unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1635 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1636 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
Tejun Heo | ffe0d5a | 2009-09-29 09:17:56 +0900 | [diff] [blame] | 1637 | unit_map[cpu] = UINT_MAX; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1638 | pcpu_first_unit_cpu = NR_CPUS; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1639 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1640 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
| 1641 | const struct pcpu_group_info *gi = &ai->groups[group]; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1642 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1643 | group_offsets[group] = gi->base_offset; |
| 1644 | group_sizes[group] = gi->nr_units * ai->unit_size; |
| 1645 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1646 | for (i = 0; i < gi->nr_units; i++) { |
| 1647 | cpu = gi->cpu_map[i]; |
| 1648 | if (cpu == NR_CPUS) |
| 1649 | continue; |
| 1650 | |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1651 | PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); |
| 1652 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
| 1653 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1654 | |
| 1655 | unit_map[cpu] = unit + i; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1656 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
| 1657 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1658 | if (pcpu_first_unit_cpu == NR_CPUS) |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1659 | pcpu_first_unit_cpu = cpu; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1660 | } |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1661 | } |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1662 | pcpu_last_unit_cpu = cpu; |
| 1663 | pcpu_nr_units = unit; |
| 1664 | |
| 1665 | for_each_possible_cpu(cpu) |
Tejun Heo | 635b75f | 2009-09-24 09:43:11 +0900 | [diff] [blame] | 1666 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
| 1667 | |
| 1668 | /* we're done parsing the input, undefine BUG macro and dump config */ |
| 1669 | #undef PCPU_SETUP_BUG_ON |
| 1670 | pcpu_dump_alloc_info(KERN_INFO, ai); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1671 | |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1672 | pcpu_nr_groups = ai->nr_groups; |
| 1673 | pcpu_group_offsets = group_offsets; |
| 1674 | pcpu_group_sizes = group_sizes; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1675 | pcpu_unit_map = unit_map; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1676 | pcpu_unit_offsets = unit_off; |
Tejun Heo | 2f39e63 | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1677 | |
| 1678 | /* determine basic parameters */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1679 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1680 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
Tejun Heo | 6563297 | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1681 | pcpu_atom_size = ai->atom_size; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1682 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
| 1683 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1684 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1685 | /* |
| 1686 | * Allocate chunk slots. The additional last slot is for |
| 1687 | * empty chunks. |
| 1688 | */ |
| 1689 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1690 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
| 1691 | for (i = 0; i < pcpu_nr_slots; i++) |
| 1692 | INIT_LIST_HEAD(&pcpu_slot[i]); |
| 1693 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1694 | /* |
| 1695 | * Initialize static chunk. If reserved_size is zero, the |
| 1696 | * static chunk covers static area + dynamic allocation area |
| 1697 | * in the first chunk. If reserved_size is not zero, it |
| 1698 | * covers static area + reserved area (mostly used for module |
| 1699 | * static percpu allocation). |
| 1700 | */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1701 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
| 1702 | INIT_LIST_HEAD(&schunk->list); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1703 | schunk->base_addr = base_addr; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1704 | schunk->map = smap; |
| 1705 | schunk->map_alloc = ARRAY_SIZE(smap); |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1706 | schunk->immutable = true; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1707 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1708 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1709 | if (ai->reserved_size) { |
| 1710 | schunk->free_size = ai->reserved_size; |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1711 | pcpu_reserved_chunk = schunk; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1712 | pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1713 | } else { |
| 1714 | schunk->free_size = dyn_size; |
| 1715 | dyn_size = 0; /* dynamic area covered */ |
| 1716 | } |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1717 | schunk->contig_hint = schunk->free_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1718 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1719 | schunk->map[schunk->map_used++] = -ai->static_size; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1720 | if (schunk->free_size) |
| 1721 | schunk->map[schunk->map_used++] = schunk->free_size; |
| 1722 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1723 | /* init dynamic chunk if necessary */ |
| 1724 | if (dyn_size) { |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1725 | dchunk = alloc_bootmem(pcpu_chunk_struct_size); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1726 | INIT_LIST_HEAD(&dchunk->list); |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1727 | dchunk->base_addr = base_addr; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1728 | dchunk->map = dmap; |
| 1729 | dchunk->map_alloc = ARRAY_SIZE(dmap); |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1730 | dchunk->immutable = true; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1731 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1732 | |
| 1733 | dchunk->contig_hint = dchunk->free_size = dyn_size; |
| 1734 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; |
| 1735 | dchunk->map[dchunk->map_used++] = dchunk->free_size; |
| 1736 | } |
| 1737 | |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1738 | /* link the first chunk in */ |
Tejun Heo | ae9e6bc9 | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1739 | pcpu_first_chunk = dchunk ?: schunk; |
| 1740 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1741 | |
| 1742 | /* we're done */ |
Tejun Heo | bba174f | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1743 | pcpu_base_addr = base_addr; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1744 | return 0; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1745 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1746 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1747 | const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { |
| 1748 | [PCPU_FC_AUTO] = "auto", |
| 1749 | [PCPU_FC_EMBED] = "embed", |
| 1750 | [PCPU_FC_PAGE] = "page", |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1751 | }; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1752 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1753 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
| 1754 | |
| 1755 | static int __init percpu_alloc_setup(char *str) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1756 | { |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1757 | if (0) |
| 1758 | /* nada */; |
| 1759 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK |
| 1760 | else if (!strcmp(str, "embed")) |
| 1761 | pcpu_chosen_fc = PCPU_FC_EMBED; |
| 1762 | #endif |
| 1763 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
| 1764 | else if (!strcmp(str, "page")) |
| 1765 | pcpu_chosen_fc = PCPU_FC_PAGE; |
| 1766 | #endif |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1767 | else |
| 1768 | pr_warning("PERCPU: unknown allocator %s specified\n", str); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1769 | |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1770 | return 0; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1771 | } |
Tejun Heo | f58dc01 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1772 | early_param("percpu_alloc", percpu_alloc_setup); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1773 | |
Tejun Heo | 08fc458 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1774 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
| 1775 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1776 | /** |
| 1777 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1778 | * @reserved_size: the size of reserved percpu area in bytes |
| 1779 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1780 | * @atom_size: allocation atom size |
| 1781 | * @cpu_distance_fn: callback to determine distance between cpus, optional |
| 1782 | * @alloc_fn: function to allocate percpu page |
| 1783 | * @free_fn: funtion to free percpu page |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1784 | * |
| 1785 | * This is a helper to ease setting up embedded first percpu chunk and |
| 1786 | * can be called where pcpu_setup_first_chunk() is expected. |
| 1787 | * |
| 1788 | * 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] | 1789 | * by calling @alloc_fn and used as-is without being mapped into |
| 1790 | * vmalloc area. Allocations are always whole multiples of @atom_size |
| 1791 | * aligned to @atom_size. |
| 1792 | * |
| 1793 | * This enables the first chunk to piggy back on the linear physical |
| 1794 | * mapping which often uses larger page size. Please note that this |
| 1795 | * can result in very sparse cpu->unit mapping on NUMA machines thus |
| 1796 | * requiring large vmalloc address space. Don't use this allocator if |
| 1797 | * vmalloc space is not orders of magnitude larger than distances |
| 1798 | * between node memory addresses (ie. 32bit NUMA machines). |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1799 | * |
| 1800 | * When @dyn_size is positive, dynamic area might be larger than |
Tejun Heo | 788e5ab | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 1801 | * specified to fill page alignment. When @dyn_size is auto, |
| 1802 | * @dyn_size is just big enough to fill page alignment after static |
| 1803 | * and reserved areas. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1804 | * |
| 1805 | * If the needed size is smaller than the minimum or specified unit |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1806 | * size, the leftover is returned using @free_fn. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1807 | * |
| 1808 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1809 | * 0 on success, -errno on failure. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1810 | */ |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1811 | int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size, |
| 1812 | size_t atom_size, |
| 1813 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, |
| 1814 | pcpu_fc_alloc_fn_t alloc_fn, |
| 1815 | pcpu_fc_free_fn_t free_fn) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1816 | { |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1817 | void *base = (void *)ULONG_MAX; |
| 1818 | void **areas = NULL; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1819 | struct pcpu_alloc_info *ai; |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 1820 | size_t size_sum, areas_size, max_distance; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1821 | int group, i, rc; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1822 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1823 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
| 1824 | cpu_distance_fn); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1825 | if (IS_ERR(ai)) |
| 1826 | return PTR_ERR(ai); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1827 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1828 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1829 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1830 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1831 | areas = alloc_bootmem_nopanic(areas_size); |
| 1832 | if (!areas) { |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1833 | rc = -ENOMEM; |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1834 | goto out_free; |
Tejun Heo | fa8a709 | 2009-06-22 11:56:24 +0900 | [diff] [blame] | 1835 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1836 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1837 | /* allocate, copy and determine base address */ |
| 1838 | for (group = 0; group < ai->nr_groups; group++) { |
| 1839 | struct pcpu_group_info *gi = &ai->groups[group]; |
| 1840 | unsigned int cpu = NR_CPUS; |
| 1841 | void *ptr; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1842 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1843 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) |
| 1844 | cpu = gi->cpu_map[i]; |
| 1845 | BUG_ON(cpu == NR_CPUS); |
| 1846 | |
| 1847 | /* allocate space for the whole group */ |
| 1848 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); |
| 1849 | if (!ptr) { |
| 1850 | rc = -ENOMEM; |
| 1851 | goto out_free_areas; |
| 1852 | } |
| 1853 | areas[group] = ptr; |
| 1854 | |
| 1855 | base = min(ptr, base); |
| 1856 | |
| 1857 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { |
| 1858 | if (gi->cpu_map[i] == NR_CPUS) { |
| 1859 | /* unused unit, free whole */ |
| 1860 | free_fn(ptr, ai->unit_size); |
| 1861 | continue; |
| 1862 | } |
| 1863 | /* copy and return the unused part */ |
| 1864 | memcpy(ptr, __per_cpu_load, ai->static_size); |
| 1865 | free_fn(ptr + size_sum, ai->unit_size - size_sum); |
| 1866 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1867 | } |
| 1868 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1869 | /* base address is now known, determine group base offsets */ |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 1870 | max_distance = 0; |
| 1871 | for (group = 0; group < ai->nr_groups; group++) { |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1872 | ai->groups[group].base_offset = areas[group] - base; |
Tejun Heo | 1a0c329 | 2009-10-04 09:31:05 +0900 | [diff] [blame] | 1873 | max_distance = max_t(size_t, max_distance, |
| 1874 | ai->groups[group].base_offset); |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 1875 | } |
| 1876 | max_distance += ai->unit_size; |
| 1877 | |
| 1878 | /* warn if maximum distance is further than 75% of vmalloc space */ |
| 1879 | if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { |
Tejun Heo | 1a0c329 | 2009-10-04 09:31:05 +0900 | [diff] [blame] | 1880 | pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " |
Tejun Heo | 6ea529a | 2009-09-24 18:46:01 +0900 | [diff] [blame] | 1881 | "space 0x%lx\n", |
| 1882 | max_distance, VMALLOC_END - VMALLOC_START); |
| 1883 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
| 1884 | /* and fail if we have fallback */ |
| 1885 | rc = -EINVAL; |
| 1886 | goto out_free; |
| 1887 | #endif |
| 1888 | } |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1889 | |
Tejun Heo | 004018e | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1890 | 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] | 1891 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
| 1892 | ai->dyn_size, ai->unit_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1893 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1894 | rc = pcpu_setup_first_chunk(ai, base); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1895 | goto out_free; |
| 1896 | |
| 1897 | out_free_areas: |
| 1898 | for (group = 0; group < ai->nr_groups; group++) |
| 1899 | free_fn(areas[group], |
| 1900 | ai->groups[group].nr_units * ai->unit_size); |
| 1901 | out_free: |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1902 | pcpu_free_alloc_info(ai); |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 1903 | if (areas) |
| 1904 | free_bootmem(__pa(areas), areas_size); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1905 | return rc; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1906 | } |
Tejun Heo | 08fc458 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1907 | #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || |
| 1908 | !CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1909 | |
Tejun Heo | 08fc458 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1910 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1911 | /** |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1912 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1913 | * @reserved_size: the size of reserved percpu area in bytes |
| 1914 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE |
| 1915 | * @free_fn: funtion to free percpu page, always called with PAGE_SIZE |
| 1916 | * @populate_pte_fn: function to populate pte |
| 1917 | * |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1918 | * This is a helper to ease setting up page-remapped first percpu |
| 1919 | * chunk and can be called where pcpu_setup_first_chunk() is expected. |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1920 | * |
| 1921 | * This is the basic allocator. Static percpu area is allocated |
| 1922 | * page-by-page into vmalloc area. |
| 1923 | * |
| 1924 | * RETURNS: |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1925 | * 0 on success, -errno on failure. |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1926 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1927 | int __init pcpu_page_first_chunk(size_t reserved_size, |
| 1928 | pcpu_fc_alloc_fn_t alloc_fn, |
| 1929 | pcpu_fc_free_fn_t free_fn, |
| 1930 | pcpu_fc_populate_pte_fn_t populate_pte_fn) |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1931 | { |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1932 | static struct vm_struct vm; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1933 | struct pcpu_alloc_info *ai; |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1934 | char psize_str[16]; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1935 | int unit_pages; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1936 | size_t pages_size; |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1937 | struct page **pages; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1938 | int unit, i, j, rc; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1939 | |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1940 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
| 1941 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1942 | ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); |
| 1943 | if (IS_ERR(ai)) |
| 1944 | return PTR_ERR(ai); |
| 1945 | BUG_ON(ai->nr_groups != 1); |
| 1946 | BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); |
| 1947 | |
| 1948 | unit_pages = ai->unit_size >> PAGE_SHIFT; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1949 | |
| 1950 | /* unaligned allocations can't be freed, round up to page size */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1951 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
| 1952 | sizeof(pages[0])); |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1953 | pages = alloc_bootmem(pages_size); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1954 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1955 | /* allocate pages */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1956 | j = 0; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1957 | for (unit = 0; unit < num_possible_cpus(); unit++) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1958 | for (i = 0; i < unit_pages; i++) { |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1959 | unsigned int cpu = ai->groups[0].cpu_map[unit]; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1960 | void *ptr; |
| 1961 | |
Tejun Heo | 3cbc856 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1962 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1963 | if (!ptr) { |
Tejun Heo | 00ae406 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 1964 | pr_warning("PERCPU: failed to allocate %s page " |
| 1965 | "for cpu%u\n", psize_str, cpu); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1966 | goto enomem; |
| 1967 | } |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1968 | pages[j++] = virt_to_page(ptr); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1969 | } |
| 1970 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1971 | /* allocate vm area, map the pages and copy static data */ |
| 1972 | vm.flags = VM_ALLOC; |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1973 | vm.size = num_possible_cpus() * ai->unit_size; |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1974 | vm_area_register_early(&vm, PAGE_SIZE); |
| 1975 | |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1976 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
Tejun Heo | 1d9d325 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 1977 | unsigned long unit_addr = |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1978 | (unsigned long)vm.addr + unit * ai->unit_size; |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1979 | |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 1980 | for (i = 0; i < unit_pages; i++) |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1981 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
| 1982 | |
| 1983 | /* pte already populated, the following shouldn't fail */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1984 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
| 1985 | unit_pages); |
| 1986 | if (rc < 0) |
| 1987 | panic("failed to map percpu area, err=%d\n", rc); |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1988 | |
| 1989 | /* |
| 1990 | * FIXME: Archs with virtual cache should flush local |
| 1991 | * cache for the linear mapping here - something |
| 1992 | * equivalent to flush_cache_vmap() on the local cpu. |
| 1993 | * flush_cache_vmap() can't be used as most supporting |
| 1994 | * data structures are not set up yet. |
| 1995 | */ |
| 1996 | |
| 1997 | /* copy static data */ |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 1998 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1999 | } |
| 2000 | |
| 2001 | /* we're ready, commit */ |
Tejun Heo | 1d9d325 | 2009-08-14 15:00:50 +0900 | [diff] [blame] | 2002 | 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] | 2003 | unit_pages, psize_str, vm.addr, ai->static_size, |
| 2004 | ai->reserved_size, ai->dyn_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2005 | |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2006 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2007 | goto out_free_ar; |
| 2008 | |
| 2009 | enomem: |
| 2010 | while (--j >= 0) |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2011 | free_fn(page_address(pages[j]), PAGE_SIZE); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2012 | rc = -ENOMEM; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2013 | out_free_ar: |
Tejun Heo | ce3141a | 2009-07-04 08:11:00 +0900 | [diff] [blame] | 2014 | free_bootmem(__pa(pages), pages_size); |
Tejun Heo | fd1e8a1 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2015 | pcpu_free_alloc_info(ai); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2016 | return rc; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 2017 | } |
Tejun Heo | 08fc458 | 2009-08-14 15:00:49 +0900 | [diff] [blame] | 2018 | #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2019 | |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 2020 | /* |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2021 | * Generic percpu area setup. |
| 2022 | * |
| 2023 | * The embedding helper is used because its behavior closely resembles |
| 2024 | * the original non-dynamic generic percpu area setup. This is |
| 2025 | * important because many archs have addressing restrictions and might |
| 2026 | * fail if the percpu area is located far away from the previous |
| 2027 | * location. As an added bonus, in non-NUMA cases, embedding is |
| 2028 | * generally a good idea TLB-wise because percpu area can piggy back |
| 2029 | * on the physical linear memory mapping which uses large page |
| 2030 | * mappings on applicable archs. |
| 2031 | */ |
| 2032 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
| 2033 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
| 2034 | EXPORT_SYMBOL(__per_cpu_offset); |
| 2035 | |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2036 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
| 2037 | size_t align) |
| 2038 | { |
| 2039 | return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); |
| 2040 | } |
| 2041 | |
| 2042 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
| 2043 | { |
| 2044 | free_bootmem(__pa(ptr), size); |
| 2045 | } |
| 2046 | |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2047 | void __init setup_per_cpu_areas(void) |
| 2048 | { |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2049 | unsigned long delta; |
| 2050 | unsigned int cpu; |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2051 | int rc; |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2052 | |
| 2053 | /* |
| 2054 | * Always reserve area for module percpu variables. That's |
| 2055 | * what the legacy allocator did. |
| 2056 | */ |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2057 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
Tejun Heo | c8826dd | 2009-08-14 15:00:52 +0900 | [diff] [blame] | 2058 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
| 2059 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2060 | if (rc < 0) |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2061 | panic("Failed to initialized percpu areas."); |
| 2062 | |
| 2063 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
| 2064 | for_each_possible_cpu(cpu) |
Tejun Heo | fb435d5 | 2009-08-14 15:00:51 +0900 | [diff] [blame] | 2065 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 2066 | } |
| 2067 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |