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 |
| 11 | * chunk is consisted of num_possible_cpus() units and the first chunk |
| 12 | * is used for static percpu variables in the kernel image (special |
| 13 | * boot time alloc/init handling necessary as these areas need to be |
| 14 | * brought up before allocation services are running). Unit grows as |
| 15 | * necessary and all units grow or shrink in unison. When a chunk is |
| 16 | * filled up, another chunk is allocated. ie. in vmalloc area |
| 17 | * |
| 18 | * c0 c1 c2 |
| 19 | * ------------------- ------------------- ------------ |
| 20 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u |
| 21 | * ------------------- ...... ------------------- .... ------------ |
| 22 | * |
| 23 | * Allocation is done in offset-size areas of single unit space. Ie, |
| 24 | * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, |
| 25 | * c1:u1, c1:u2 and c1:u3. Percpu access can be done by configuring |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 26 | * percpu base registers pcpu_unit_size apart. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 27 | * |
| 28 | * There are usually many small percpu allocations many of them as |
| 29 | * small as 4 bytes. The allocator organizes chunks into lists |
| 30 | * according to free size and tries to allocate from the fullest one. |
| 31 | * Each chunk keeps the maximum contiguous area size hint which is |
| 32 | * guaranteed to be eqaul to or larger than the maximum contiguous |
| 33 | * area in the chunk. This helps the allocator not to iterate the |
| 34 | * chunk maps unnecessarily. |
| 35 | * |
| 36 | * Allocation state in each chunk is kept using an array of integers |
| 37 | * on chunk->map. A positive value in the map represents a free |
| 38 | * region and negative allocated. Allocation inside a chunk is done |
| 39 | * by scanning this map sequentially and serving the first matching |
| 40 | * entry. This is mostly copied from the percpu_modalloc() allocator. |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 41 | * Chunks can be determined from the address using the index field |
| 42 | * in the page struct. The index field contains a pointer to the chunk. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 43 | * |
| 44 | * To use this allocator, arch code should do the followings. |
| 45 | * |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 46 | * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 47 | * |
| 48 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 49 | * regular address to percpu pointer and back if they need to be |
| 50 | * different from the default |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 51 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 52 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
| 53 | * setup the first chunk containing the kernel static percpu area |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 54 | */ |
| 55 | |
| 56 | #include <linux/bitmap.h> |
| 57 | #include <linux/bootmem.h> |
| 58 | #include <linux/list.h> |
| 59 | #include <linux/mm.h> |
| 60 | #include <linux/module.h> |
| 61 | #include <linux/mutex.h> |
| 62 | #include <linux/percpu.h> |
| 63 | #include <linux/pfn.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 64 | #include <linux/slab.h> |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 65 | #include <linux/spinlock.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 66 | #include <linux/vmalloc.h> |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 67 | #include <linux/workqueue.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 68 | |
| 69 | #include <asm/cacheflush.h> |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 70 | #include <asm/sections.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 71 | #include <asm/tlbflush.h> |
| 72 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 73 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
| 74 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ |
| 75 | |
Tejun Heo | e010098 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 76 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
| 77 | #ifndef __addr_to_pcpu_ptr |
| 78 | #define __addr_to_pcpu_ptr(addr) \ |
| 79 | (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ |
| 80 | + (unsigned long)__per_cpu_start) |
| 81 | #endif |
| 82 | #ifndef __pcpu_ptr_to_addr |
| 83 | #define __pcpu_ptr_to_addr(ptr) \ |
| 84 | (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ |
| 85 | - (unsigned long)__per_cpu_start) |
| 86 | #endif |
| 87 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 88 | struct pcpu_chunk { |
| 89 | struct list_head list; /* linked to pcpu_slot lists */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 90 | int free_size; /* free bytes in the chunk */ |
| 91 | int contig_hint; /* max contiguous size hint */ |
| 92 | struct vm_struct *vm; /* mapped vmalloc region */ |
| 93 | int map_used; /* # of map entries used */ |
| 94 | int map_alloc; /* # of map entries allocated */ |
| 95 | int *map; /* allocation map */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 96 | bool immutable; /* no [de]population allowed */ |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 97 | struct page **page; /* points to page array */ |
| 98 | struct page *page_ar[]; /* #cpus * UNIT_PAGES */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 99 | }; |
| 100 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 101 | static int pcpu_unit_pages __read_mostly; |
| 102 | static int pcpu_unit_size __read_mostly; |
| 103 | static int pcpu_chunk_size __read_mostly; |
| 104 | static int pcpu_nr_slots __read_mostly; |
| 105 | static size_t pcpu_chunk_struct_size __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 106 | |
| 107 | /* 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] | 108 | void *pcpu_base_addr __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 109 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
| 110 | |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 111 | /* |
| 112 | * The first chunk which always exists. Note that unlike other |
| 113 | * chunks, this one can be allocated and mapped in several different |
| 114 | * ways and thus often doesn't live in the vmalloc area. |
| 115 | */ |
| 116 | static struct pcpu_chunk *pcpu_first_chunk; |
| 117 | |
| 118 | /* |
| 119 | * Optional reserved chunk. This chunk reserves part of the first |
| 120 | * chunk and serves it for reserved allocations. The amount of |
| 121 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved |
| 122 | * area doesn't exist, the following variables contain NULL and 0 |
| 123 | * respectively. |
| 124 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 125 | static struct pcpu_chunk *pcpu_reserved_chunk; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 126 | static int pcpu_reserved_chunk_limit; |
| 127 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 128 | /* |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 129 | * Synchronization rules. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 130 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 131 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former |
| 132 | * protects allocation/reclaim paths, chunks and chunk->page arrays. |
| 133 | * The latter is a spinlock and protects the index data structures - |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 134 | * chunk slots, chunks and area maps in chunks. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 135 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 136 | * During allocation, pcpu_alloc_mutex is kept locked all the time and |
| 137 | * pcpu_lock is grabbed and released as necessary. All actual memory |
| 138 | * allocations are done using GFP_KERNEL with pcpu_lock released. |
| 139 | * |
| 140 | * Free path accesses and alters only the index data structures, so it |
| 141 | * can be safely called from atomic context. When memory needs to be |
| 142 | * returned to the system, free path schedules reclaim_work which |
| 143 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be |
| 144 | * reclaimed, release both locks and frees the chunks. Note that it's |
| 145 | * necessary to grab both locks to remove a chunk from circulation as |
| 146 | * allocation path might be referencing the chunk with only |
| 147 | * pcpu_alloc_mutex locked. |
Tejun 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 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
| 150 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 151 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 152 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 153 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 154 | /* reclaim work to release fully free chunks, scheduled from free path */ |
| 155 | static void pcpu_reclaim(struct work_struct *work); |
| 156 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); |
| 157 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 158 | static int __pcpu_size_to_slot(int size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 159 | { |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 160 | int highbit = fls(size); /* size is in bytes */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 161 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
| 162 | } |
| 163 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 164 | static int pcpu_size_to_slot(int size) |
| 165 | { |
| 166 | if (size == pcpu_unit_size) |
| 167 | return pcpu_nr_slots - 1; |
| 168 | return __pcpu_size_to_slot(size); |
| 169 | } |
| 170 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 171 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
| 172 | { |
| 173 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) |
| 174 | return 0; |
| 175 | |
| 176 | return pcpu_size_to_slot(chunk->free_size); |
| 177 | } |
| 178 | |
| 179 | static int pcpu_page_idx(unsigned int cpu, int page_idx) |
| 180 | { |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 181 | return cpu * pcpu_unit_pages + page_idx; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 182 | } |
| 183 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 184 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
| 185 | unsigned int cpu, int page_idx) |
| 186 | { |
| 187 | return (unsigned long)chunk->vm->addr + |
| 188 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); |
| 189 | } |
| 190 | |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame^] | 191 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, |
| 192 | unsigned int cpu, int page_idx) |
| 193 | { |
| 194 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; |
| 195 | } |
| 196 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 197 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, |
| 198 | int page_idx) |
| 199 | { |
| 200 | return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL; |
| 201 | } |
| 202 | |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 203 | /* set the pointer to a chunk in a page struct */ |
| 204 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) |
| 205 | { |
| 206 | page->index = (unsigned long)pcpu; |
| 207 | } |
| 208 | |
| 209 | /* obtain pointer to a chunk from a page struct */ |
| 210 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) |
| 211 | { |
| 212 | return (struct pcpu_chunk *)page->index; |
| 213 | } |
| 214 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 215 | /** |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 216 | * pcpu_mem_alloc - allocate memory |
| 217 | * @size: bytes to allocate |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 218 | * |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 219 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
| 220 | * kzalloc() is used; otherwise, vmalloc() is used. The returned |
| 221 | * memory is always zeroed. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 222 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 223 | * CONTEXT: |
| 224 | * Does GFP_KERNEL allocation. |
| 225 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 226 | * RETURNS: |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 227 | * Pointer to the allocated area on success, NULL on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 228 | */ |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 229 | static void *pcpu_mem_alloc(size_t size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 230 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 231 | if (size <= PAGE_SIZE) |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 232 | return kzalloc(size, GFP_KERNEL); |
| 233 | else { |
| 234 | void *ptr = vmalloc(size); |
| 235 | if (ptr) |
| 236 | memset(ptr, 0, size); |
| 237 | return ptr; |
| 238 | } |
| 239 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 240 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 241 | /** |
| 242 | * pcpu_mem_free - free memory |
| 243 | * @ptr: memory to free |
| 244 | * @size: size of the area |
| 245 | * |
| 246 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). |
| 247 | */ |
| 248 | static void pcpu_mem_free(void *ptr, size_t size) |
| 249 | { |
| 250 | if (size <= PAGE_SIZE) |
| 251 | kfree(ptr); |
| 252 | else |
| 253 | vfree(ptr); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 254 | } |
| 255 | |
| 256 | /** |
| 257 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot |
| 258 | * @chunk: chunk of interest |
| 259 | * @oslot: the previous slot it was on |
| 260 | * |
| 261 | * This function is called after an allocation or free changed @chunk. |
| 262 | * New slot according to the changed state is determined and @chunk is |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 263 | * moved to the slot. Note that the reserved chunk is never put on |
| 264 | * chunk slots. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 265 | * |
| 266 | * CONTEXT: |
| 267 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 268 | */ |
| 269 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) |
| 270 | { |
| 271 | int nslot = pcpu_chunk_slot(chunk); |
| 272 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 273 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 274 | if (oslot < nslot) |
| 275 | list_move(&chunk->list, &pcpu_slot[nslot]); |
| 276 | else |
| 277 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); |
| 278 | } |
| 279 | } |
| 280 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 281 | /** |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 282 | * pcpu_chunk_addr_search - determine chunk containing specified address |
| 283 | * @addr: address for which the chunk needs to be determined. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 284 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 285 | * RETURNS: |
| 286 | * The address of the found chunk. |
| 287 | */ |
| 288 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) |
| 289 | { |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 290 | void *first_start = pcpu_first_chunk->vm->addr; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 291 | |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 292 | /* is it in the first chunk? */ |
Tejun Heo | 79ba6ac | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 293 | if (addr >= first_start && addr < first_start + pcpu_unit_size) { |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 294 | /* is it in the reserved area? */ |
| 295 | if (addr < first_start + pcpu_reserved_chunk_limit) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 296 | return pcpu_reserved_chunk; |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 297 | return pcpu_first_chunk; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 298 | } |
| 299 | |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 300 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 301 | } |
| 302 | |
| 303 | /** |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 304 | * pcpu_extend_area_map - extend area map for allocation |
| 305 | * @chunk: target chunk |
| 306 | * |
| 307 | * Extend area map of @chunk so that it can accomodate an allocation. |
| 308 | * A single allocation can split an area into three areas, so this |
| 309 | * function makes sure that @chunk->map has at least two extra slots. |
| 310 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 311 | * CONTEXT: |
| 312 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired |
| 313 | * if area map is extended. |
| 314 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 315 | * RETURNS: |
| 316 | * 0 if noop, 1 if successfully extended, -errno on failure. |
| 317 | */ |
| 318 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) |
| 319 | { |
| 320 | int new_alloc; |
| 321 | int *new; |
| 322 | size_t size; |
| 323 | |
| 324 | /* has enough? */ |
| 325 | if (chunk->map_alloc >= chunk->map_used + 2) |
| 326 | return 0; |
| 327 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 328 | spin_unlock_irq(&pcpu_lock); |
| 329 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 330 | new_alloc = PCPU_DFL_MAP_ALLOC; |
| 331 | while (new_alloc < chunk->map_used + 2) |
| 332 | new_alloc *= 2; |
| 333 | |
| 334 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 335 | if (!new) { |
| 336 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 337 | return -ENOMEM; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 338 | } |
| 339 | |
| 340 | /* |
| 341 | * Acquire pcpu_lock and switch to new area map. Only free |
| 342 | * could have happened inbetween, so map_used couldn't have |
| 343 | * grown. |
| 344 | */ |
| 345 | spin_lock_irq(&pcpu_lock); |
| 346 | BUG_ON(new_alloc < chunk->map_used + 2); |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 347 | |
| 348 | size = chunk->map_alloc * sizeof(chunk->map[0]); |
| 349 | memcpy(new, chunk->map, size); |
| 350 | |
| 351 | /* |
| 352 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is |
| 353 | * one of the first chunks and still using static map. |
| 354 | */ |
| 355 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) |
| 356 | pcpu_mem_free(chunk->map, size); |
| 357 | |
| 358 | chunk->map_alloc = new_alloc; |
| 359 | chunk->map = new; |
| 360 | return 0; |
| 361 | } |
| 362 | |
| 363 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 364 | * pcpu_split_block - split a map block |
| 365 | * @chunk: chunk of interest |
| 366 | * @i: index of map block to split |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 367 | * @head: head size in bytes (can be 0) |
| 368 | * @tail: tail size in bytes (can be 0) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 369 | * |
| 370 | * Split the @i'th map block into two or three blocks. If @head is |
| 371 | * non-zero, @head bytes block is inserted before block @i moving it |
| 372 | * to @i+1 and reducing its size by @head bytes. |
| 373 | * |
| 374 | * If @tail is non-zero, the target block, which can be @i or @i+1 |
| 375 | * depending on @head, is reduced by @tail bytes and @tail byte block |
| 376 | * is inserted after the target block. |
| 377 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 378 | * @chunk->map must have enough free slots to accomodate the split. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 379 | * |
| 380 | * CONTEXT: |
| 381 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 382 | */ |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 383 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
| 384 | int head, int tail) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 385 | { |
| 386 | int nr_extra = !!head + !!tail; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 387 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 388 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 389 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 390 | /* insert new subblocks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 391 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
| 392 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
| 393 | chunk->map_used += nr_extra; |
| 394 | |
| 395 | if (head) { |
| 396 | chunk->map[i + 1] = chunk->map[i] - head; |
| 397 | chunk->map[i++] = head; |
| 398 | } |
| 399 | if (tail) { |
| 400 | chunk->map[i++] -= tail; |
| 401 | chunk->map[i] = tail; |
| 402 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 403 | } |
| 404 | |
| 405 | /** |
| 406 | * pcpu_alloc_area - allocate area from a pcpu_chunk |
| 407 | * @chunk: chunk of interest |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 408 | * @size: wanted size in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 409 | * @align: wanted align |
| 410 | * |
| 411 | * Try to allocate @size bytes area aligned at @align from @chunk. |
| 412 | * Note that this function only allocates the offset. It doesn't |
| 413 | * populate or map the area. |
| 414 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 415 | * @chunk->map must have at least two free slots. |
| 416 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 417 | * CONTEXT: |
| 418 | * pcpu_lock. |
| 419 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 420 | * RETURNS: |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 421 | * Allocated offset in @chunk on success, -1 if no matching area is |
| 422 | * found. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 423 | */ |
| 424 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) |
| 425 | { |
| 426 | int oslot = pcpu_chunk_slot(chunk); |
| 427 | int max_contig = 0; |
| 428 | int i, off; |
| 429 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 430 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
| 431 | bool is_last = i + 1 == chunk->map_used; |
| 432 | int head, tail; |
| 433 | |
| 434 | /* extra for alignment requirement */ |
| 435 | head = ALIGN(off, align) - off; |
| 436 | BUG_ON(i == 0 && head != 0); |
| 437 | |
| 438 | if (chunk->map[i] < 0) |
| 439 | continue; |
| 440 | if (chunk->map[i] < head + size) { |
| 441 | max_contig = max(chunk->map[i], max_contig); |
| 442 | continue; |
| 443 | } |
| 444 | |
| 445 | /* |
| 446 | * If head is small or the previous block is free, |
| 447 | * merge'em. Note that 'small' is defined as smaller |
| 448 | * than sizeof(int), which is very small but isn't too |
| 449 | * uncommon for percpu allocations. |
| 450 | */ |
| 451 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { |
| 452 | if (chunk->map[i - 1] > 0) |
| 453 | chunk->map[i - 1] += head; |
| 454 | else { |
| 455 | chunk->map[i - 1] -= head; |
| 456 | chunk->free_size -= head; |
| 457 | } |
| 458 | chunk->map[i] -= head; |
| 459 | off += head; |
| 460 | head = 0; |
| 461 | } |
| 462 | |
| 463 | /* if tail is small, just keep it around */ |
| 464 | tail = chunk->map[i] - head - size; |
| 465 | if (tail < sizeof(int)) |
| 466 | tail = 0; |
| 467 | |
| 468 | /* split if warranted */ |
| 469 | if (head || tail) { |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 470 | pcpu_split_block(chunk, i, head, tail); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 471 | if (head) { |
| 472 | i++; |
| 473 | off += head; |
| 474 | max_contig = max(chunk->map[i - 1], max_contig); |
| 475 | } |
| 476 | if (tail) |
| 477 | max_contig = max(chunk->map[i + 1], max_contig); |
| 478 | } |
| 479 | |
| 480 | /* update hint and mark allocated */ |
| 481 | if (is_last) |
| 482 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 483 | else |
| 484 | chunk->contig_hint = max(chunk->contig_hint, |
| 485 | max_contig); |
| 486 | |
| 487 | chunk->free_size -= chunk->map[i]; |
| 488 | chunk->map[i] = -chunk->map[i]; |
| 489 | |
| 490 | pcpu_chunk_relocate(chunk, oslot); |
| 491 | return off; |
| 492 | } |
| 493 | |
| 494 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 495 | pcpu_chunk_relocate(chunk, oslot); |
| 496 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 497 | /* tell the upper layer that this chunk has no matching area */ |
| 498 | return -1; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 499 | } |
| 500 | |
| 501 | /** |
| 502 | * pcpu_free_area - free area to a pcpu_chunk |
| 503 | * @chunk: chunk of interest |
| 504 | * @freeme: offset of area to free |
| 505 | * |
| 506 | * Free area starting from @freeme to @chunk. Note that this function |
| 507 | * only modifies the allocation map. It doesn't depopulate or unmap |
| 508 | * the area. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 509 | * |
| 510 | * CONTEXT: |
| 511 | * pcpu_lock. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 512 | */ |
| 513 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) |
| 514 | { |
| 515 | int oslot = pcpu_chunk_slot(chunk); |
| 516 | int i, off; |
| 517 | |
| 518 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) |
| 519 | if (off == freeme) |
| 520 | break; |
| 521 | BUG_ON(off != freeme); |
| 522 | BUG_ON(chunk->map[i] > 0); |
| 523 | |
| 524 | chunk->map[i] = -chunk->map[i]; |
| 525 | chunk->free_size += chunk->map[i]; |
| 526 | |
| 527 | /* merge with previous? */ |
| 528 | if (i > 0 && chunk->map[i - 1] >= 0) { |
| 529 | chunk->map[i - 1] += chunk->map[i]; |
| 530 | chunk->map_used--; |
| 531 | memmove(&chunk->map[i], &chunk->map[i + 1], |
| 532 | (chunk->map_used - i) * sizeof(chunk->map[0])); |
| 533 | i--; |
| 534 | } |
| 535 | /* merge with next? */ |
| 536 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { |
| 537 | chunk->map[i] += chunk->map[i + 1]; |
| 538 | chunk->map_used--; |
| 539 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], |
| 540 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); |
| 541 | } |
| 542 | |
| 543 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); |
| 544 | pcpu_chunk_relocate(chunk, oslot); |
| 545 | } |
| 546 | |
| 547 | /** |
| 548 | * pcpu_unmap - unmap pages out of a pcpu_chunk |
| 549 | * @chunk: chunk of interest |
| 550 | * @page_start: page index of the first page to unmap |
| 551 | * @page_end: page index of the last page to unmap + 1 |
Tejun Heo | 85ae87c | 2009-06-22 11:56:23 +0900 | [diff] [blame] | 552 | * @flush_tlb: whether to flush tlb or not |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 553 | * |
| 554 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. |
| 555 | * If @flush is true, vcache is flushed before unmapping and tlb |
| 556 | * after. |
| 557 | */ |
| 558 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, |
Tejun Heo | 85ae87c | 2009-06-22 11:56:23 +0900 | [diff] [blame] | 559 | bool flush_tlb) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 560 | { |
| 561 | unsigned int last = num_possible_cpus() - 1; |
| 562 | unsigned int cpu; |
| 563 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 564 | /* unmap must not be done on immutable chunk */ |
| 565 | WARN_ON(chunk->immutable); |
| 566 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 567 | /* |
| 568 | * Each flushing trial can be very expensive, issue flush on |
| 569 | * the whole region at once rather than doing it for each cpu. |
| 570 | * This could be an overkill but is more scalable. |
| 571 | */ |
Tejun Heo | 85ae87c | 2009-06-22 11:56:23 +0900 | [diff] [blame] | 572 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 573 | pcpu_chunk_addr(chunk, last, page_end)); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 574 | |
| 575 | for_each_possible_cpu(cpu) |
| 576 | unmap_kernel_range_noflush( |
| 577 | pcpu_chunk_addr(chunk, cpu, page_start), |
| 578 | (page_end - page_start) << PAGE_SHIFT); |
| 579 | |
| 580 | /* ditto as flush_cache_vunmap() */ |
Tejun Heo | 85ae87c | 2009-06-22 11:56:23 +0900 | [diff] [blame] | 581 | if (flush_tlb) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 582 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), |
| 583 | pcpu_chunk_addr(chunk, last, page_end)); |
| 584 | } |
| 585 | |
Tejun Heo | c8a51be | 2009-07-04 08:10:59 +0900 | [diff] [blame^] | 586 | static int __pcpu_map_pages(unsigned long addr, struct page **pages, |
| 587 | int nr_pages) |
| 588 | { |
| 589 | return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, |
| 590 | PAGE_KERNEL, pages); |
| 591 | } |
| 592 | |
| 593 | /** |
| 594 | * pcpu_map - map pages into a pcpu_chunk |
| 595 | * @chunk: chunk of interest |
| 596 | * @page_start: page index of the first page to map |
| 597 | * @page_end: page index of the last page to map + 1 |
| 598 | * |
| 599 | * For each cpu, map pages [@page_start,@page_end) into @chunk. |
| 600 | * vcache is flushed afterwards. |
| 601 | */ |
| 602 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) |
| 603 | { |
| 604 | unsigned int last = num_possible_cpus() - 1; |
| 605 | unsigned int cpu; |
| 606 | int err; |
| 607 | |
| 608 | /* map must not be done on immutable chunk */ |
| 609 | WARN_ON(chunk->immutable); |
| 610 | |
| 611 | for_each_possible_cpu(cpu) { |
| 612 | err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), |
| 613 | pcpu_chunk_pagep(chunk, cpu, page_start), |
| 614 | page_end - page_start); |
| 615 | if (err < 0) |
| 616 | return err; |
| 617 | } |
| 618 | |
| 619 | /* flush at once, please read comments in pcpu_unmap() */ |
| 620 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 621 | pcpu_chunk_addr(chunk, last, page_end)); |
| 622 | return 0; |
| 623 | } |
| 624 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 625 | /** |
| 626 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk |
| 627 | * @chunk: chunk to depopulate |
| 628 | * @off: offset to the area to depopulate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 629 | * @size: size of the area to depopulate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 630 | * @flush: whether to flush cache and tlb or not |
| 631 | * |
| 632 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) |
| 633 | * from @chunk. If @flush is true, vcache is flushed before unmapping |
| 634 | * and tlb after. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 635 | * |
| 636 | * CONTEXT: |
| 637 | * pcpu_alloc_mutex. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 638 | */ |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 639 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
| 640 | bool flush) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 641 | { |
| 642 | int page_start = PFN_DOWN(off); |
| 643 | int page_end = PFN_UP(off + size); |
| 644 | int unmap_start = -1; |
| 645 | int uninitialized_var(unmap_end); |
| 646 | unsigned int cpu; |
| 647 | int i; |
| 648 | |
| 649 | for (i = page_start; i < page_end; i++) { |
| 650 | for_each_possible_cpu(cpu) { |
| 651 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 652 | |
| 653 | if (!*pagep) |
| 654 | continue; |
| 655 | |
| 656 | __free_page(*pagep); |
| 657 | |
| 658 | /* |
| 659 | * If it's partial depopulation, it might get |
| 660 | * populated or depopulated again. Mark the |
| 661 | * page gone. |
| 662 | */ |
| 663 | *pagep = NULL; |
| 664 | |
| 665 | unmap_start = unmap_start < 0 ? i : unmap_start; |
| 666 | unmap_end = i + 1; |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | if (unmap_start >= 0) |
| 671 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); |
| 672 | } |
| 673 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 674 | /** |
| 675 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk |
| 676 | * @chunk: chunk of interest |
| 677 | * @off: offset to the area to populate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 678 | * @size: size of the area to populate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 679 | * |
| 680 | * For each cpu, populate and map pages [@page_start,@page_end) into |
| 681 | * @chunk. The area is cleared on return. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 682 | * |
| 683 | * CONTEXT: |
| 684 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 685 | */ |
| 686 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) |
| 687 | { |
| 688 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; |
| 689 | int page_start = PFN_DOWN(off); |
| 690 | int page_end = PFN_UP(off + size); |
| 691 | int map_start = -1; |
Tejun Heo | 02d51fdf | 2009-03-01 15:42:36 +0900 | [diff] [blame] | 692 | int uninitialized_var(map_end); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 693 | unsigned int cpu; |
| 694 | int i; |
| 695 | |
| 696 | for (i = page_start; i < page_end; i++) { |
| 697 | if (pcpu_chunk_page_occupied(chunk, i)) { |
| 698 | if (map_start >= 0) { |
| 699 | if (pcpu_map(chunk, map_start, map_end)) |
| 700 | goto err; |
| 701 | map_start = -1; |
| 702 | } |
| 703 | continue; |
| 704 | } |
| 705 | |
| 706 | map_start = map_start < 0 ? i : map_start; |
| 707 | map_end = i + 1; |
| 708 | |
| 709 | for_each_possible_cpu(cpu) { |
| 710 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 711 | |
| 712 | *pagep = alloc_pages_node(cpu_to_node(cpu), |
| 713 | alloc_mask, 0); |
| 714 | if (!*pagep) |
| 715 | goto err; |
Christoph Lameter | e1b9aa3 | 2009-04-02 13:21:44 +0900 | [diff] [blame] | 716 | pcpu_set_page_chunk(*pagep, chunk); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 717 | } |
| 718 | } |
| 719 | |
| 720 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) |
| 721 | goto err; |
| 722 | |
| 723 | for_each_possible_cpu(cpu) |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 724 | memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 725 | size); |
| 726 | |
| 727 | return 0; |
| 728 | err: |
| 729 | /* likely under heavy memory pressure, give memory back */ |
| 730 | pcpu_depopulate_chunk(chunk, off, size, true); |
| 731 | return -ENOMEM; |
| 732 | } |
| 733 | |
| 734 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) |
| 735 | { |
| 736 | if (!chunk) |
| 737 | return; |
| 738 | if (chunk->vm) |
| 739 | free_vm_area(chunk->vm); |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 740 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 741 | kfree(chunk); |
| 742 | } |
| 743 | |
| 744 | static struct pcpu_chunk *alloc_pcpu_chunk(void) |
| 745 | { |
| 746 | struct pcpu_chunk *chunk; |
| 747 | |
| 748 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); |
| 749 | if (!chunk) |
| 750 | return NULL; |
| 751 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 752 | 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] | 753 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
| 754 | chunk->map[chunk->map_used++] = pcpu_unit_size; |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 755 | chunk->page = chunk->page_ar; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 756 | |
| 757 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); |
| 758 | if (!chunk->vm) { |
| 759 | free_pcpu_chunk(chunk); |
| 760 | return NULL; |
| 761 | } |
| 762 | |
| 763 | INIT_LIST_HEAD(&chunk->list); |
| 764 | chunk->free_size = pcpu_unit_size; |
| 765 | chunk->contig_hint = pcpu_unit_size; |
| 766 | |
| 767 | return chunk; |
| 768 | } |
| 769 | |
| 770 | /** |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 771 | * pcpu_alloc - the percpu allocator |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 772 | * @size: size of area to allocate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 773 | * @align: alignment of area (max PAGE_SIZE) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 774 | * @reserved: allocate from the reserved chunk if available |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 775 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 776 | * Allocate percpu area of @size bytes aligned at @align. |
| 777 | * |
| 778 | * CONTEXT: |
| 779 | * Does GFP_KERNEL allocation. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 780 | * |
| 781 | * RETURNS: |
| 782 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 783 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 784 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 785 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 786 | struct pcpu_chunk *chunk; |
| 787 | int slot, off; |
| 788 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 789 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 790 | WARN(true, "illegal size (%zu) or align (%zu) for " |
| 791 | "percpu allocation\n", size, align); |
| 792 | return NULL; |
| 793 | } |
| 794 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 795 | mutex_lock(&pcpu_alloc_mutex); |
| 796 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 797 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 798 | /* serve reserved allocations from the reserved chunk if available */ |
| 799 | if (reserved && pcpu_reserved_chunk) { |
| 800 | chunk = pcpu_reserved_chunk; |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 801 | if (size > chunk->contig_hint || |
| 802 | pcpu_extend_area_map(chunk) < 0) |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 803 | goto fail_unlock; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 804 | off = pcpu_alloc_area(chunk, size, align); |
| 805 | if (off >= 0) |
| 806 | goto area_found; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 807 | goto fail_unlock; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 808 | } |
| 809 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 810 | restart: |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 811 | /* search through normal chunks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 812 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
| 813 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 814 | if (size > chunk->contig_hint) |
| 815 | continue; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 816 | |
| 817 | switch (pcpu_extend_area_map(chunk)) { |
| 818 | case 0: |
| 819 | break; |
| 820 | case 1: |
| 821 | goto restart; /* pcpu_lock dropped, restart */ |
| 822 | default: |
| 823 | goto fail_unlock; |
| 824 | } |
| 825 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 826 | off = pcpu_alloc_area(chunk, size, align); |
| 827 | if (off >= 0) |
| 828 | goto area_found; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 829 | } |
| 830 | } |
| 831 | |
| 832 | /* hmmm... no space left, create a new chunk */ |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 833 | spin_unlock_irq(&pcpu_lock); |
| 834 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 835 | chunk = alloc_pcpu_chunk(); |
| 836 | if (!chunk) |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 837 | goto fail_unlock_mutex; |
| 838 | |
| 839 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 840 | pcpu_chunk_relocate(chunk, -1); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 841 | goto restart; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 842 | |
| 843 | area_found: |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 844 | spin_unlock_irq(&pcpu_lock); |
| 845 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 846 | /* populate, map and clear the area */ |
| 847 | if (pcpu_populate_chunk(chunk, off, size)) { |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 848 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 849 | pcpu_free_area(chunk, off); |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 850 | goto fail_unlock; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 851 | } |
| 852 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 853 | mutex_unlock(&pcpu_alloc_mutex); |
| 854 | |
| 855 | return __addr_to_pcpu_ptr(chunk->vm->addr + off); |
| 856 | |
| 857 | fail_unlock: |
| 858 | spin_unlock_irq(&pcpu_lock); |
| 859 | fail_unlock_mutex: |
| 860 | mutex_unlock(&pcpu_alloc_mutex); |
| 861 | return NULL; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 862 | } |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 863 | |
| 864 | /** |
| 865 | * __alloc_percpu - allocate dynamic percpu area |
| 866 | * @size: size of area to allocate in bytes |
| 867 | * @align: alignment of area (max PAGE_SIZE) |
| 868 | * |
| 869 | * Allocate percpu area of @size bytes aligned at @align. Might |
| 870 | * sleep. Might trigger writeouts. |
| 871 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 872 | * CONTEXT: |
| 873 | * Does GFP_KERNEL allocation. |
| 874 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 875 | * RETURNS: |
| 876 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 877 | */ |
| 878 | void *__alloc_percpu(size_t size, size_t align) |
| 879 | { |
| 880 | return pcpu_alloc(size, align, false); |
| 881 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 882 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
| 883 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 884 | /** |
| 885 | * __alloc_reserved_percpu - allocate reserved percpu area |
| 886 | * @size: size of area to allocate in bytes |
| 887 | * @align: alignment of area (max PAGE_SIZE) |
| 888 | * |
| 889 | * Allocate percpu area of @size bytes aligned at @align from reserved |
| 890 | * percpu area if arch has set it up; otherwise, allocation is served |
| 891 | * from the same dynamic area. Might sleep. Might trigger writeouts. |
| 892 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 893 | * CONTEXT: |
| 894 | * Does GFP_KERNEL allocation. |
| 895 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 896 | * RETURNS: |
| 897 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 898 | */ |
| 899 | void *__alloc_reserved_percpu(size_t size, size_t align) |
| 900 | { |
| 901 | return pcpu_alloc(size, align, true); |
| 902 | } |
| 903 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 904 | /** |
| 905 | * pcpu_reclaim - reclaim fully free chunks, workqueue function |
| 906 | * @work: unused |
| 907 | * |
| 908 | * Reclaim all fully free chunks except for the first one. |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 909 | * |
| 910 | * CONTEXT: |
| 911 | * workqueue context. |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 912 | */ |
| 913 | static void pcpu_reclaim(struct work_struct *work) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 914 | { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 915 | LIST_HEAD(todo); |
| 916 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; |
| 917 | struct pcpu_chunk *chunk, *next; |
| 918 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 919 | mutex_lock(&pcpu_alloc_mutex); |
| 920 | spin_lock_irq(&pcpu_lock); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 921 | |
| 922 | list_for_each_entry_safe(chunk, next, head, list) { |
| 923 | WARN_ON(chunk->immutable); |
| 924 | |
| 925 | /* spare the first one */ |
| 926 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) |
| 927 | continue; |
| 928 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 929 | list_move(&chunk->list, &todo); |
| 930 | } |
| 931 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 932 | spin_unlock_irq(&pcpu_lock); |
| 933 | mutex_unlock(&pcpu_alloc_mutex); |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 934 | |
| 935 | list_for_each_entry_safe(chunk, next, &todo, list) { |
| 936 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); |
| 937 | free_pcpu_chunk(chunk); |
| 938 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 939 | } |
| 940 | |
| 941 | /** |
| 942 | * free_percpu - free percpu area |
| 943 | * @ptr: pointer to area to free |
| 944 | * |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 945 | * Free percpu area @ptr. |
| 946 | * |
| 947 | * CONTEXT: |
| 948 | * Can be called from atomic context. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 949 | */ |
| 950 | void free_percpu(void *ptr) |
| 951 | { |
| 952 | void *addr = __pcpu_ptr_to_addr(ptr); |
| 953 | struct pcpu_chunk *chunk; |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 954 | unsigned long flags; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 955 | int off; |
| 956 | |
| 957 | if (!ptr) |
| 958 | return; |
| 959 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 960 | spin_lock_irqsave(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 961 | |
| 962 | chunk = pcpu_chunk_addr_search(addr); |
| 963 | off = addr - chunk->vm->addr; |
| 964 | |
| 965 | pcpu_free_area(chunk, off); |
| 966 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 967 | /* 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] | 968 | if (chunk->free_size == pcpu_unit_size) { |
| 969 | struct pcpu_chunk *pos; |
| 970 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 971 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 972 | if (pos != chunk) { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame] | 973 | schedule_work(&pcpu_reclaim_work); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 974 | break; |
| 975 | } |
| 976 | } |
| 977 | |
Tejun Heo | ccea34b | 2009-03-07 00:44:13 +0900 | [diff] [blame] | 978 | spin_unlock_irqrestore(&pcpu_lock, flags); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 979 | } |
| 980 | EXPORT_SYMBOL_GPL(free_percpu); |
| 981 | |
| 982 | /** |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 983 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
| 984 | * @get_page_fn: callback to fetch page pointer |
| 985 | * @static_size: the size of static percpu area in bytes |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 986 | * @reserved_size: the size of reserved percpu area in bytes, 0 for none |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 987 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 988 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE |
| 989 | * @base_addr: mapped address |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 990 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 991 | * Initialize the first percpu chunk which contains the kernel static |
| 992 | * perpcu area. This function is to be called from arch percpu area |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 993 | * setup path. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 994 | * |
| 995 | * @get_page_fn() should return pointer to percpu page given cpu |
| 996 | * number and page number. It should at least return enough pages to |
| 997 | * cover the static area. The returned pages for static area should |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 998 | * have been initialized with valid data. It can also return pages |
| 999 | * after the static area. NULL return indicates end of pages for the |
| 1000 | * cpu. Note that @get_page_fn() must return the same number of pages |
| 1001 | * for all cpus. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1002 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1003 | * @reserved_size, if non-zero, specifies the amount of bytes to |
| 1004 | * reserve after the static area in the first chunk. This reserves |
| 1005 | * the first chunk such that it's available only through reserved |
| 1006 | * percpu allocation. This is primarily used to serve module percpu |
| 1007 | * static areas on architectures where the addressing model has |
| 1008 | * limited offset range for symbol relocations to guarantee module |
| 1009 | * percpu symbols fall inside the relocatable range. |
| 1010 | * |
Tejun Heo | 6074d5b | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1011 | * @dyn_size, if non-negative, determines the number of bytes |
| 1012 | * available for dynamic allocation in the first chunk. Specifying |
| 1013 | * non-negative value makes percpu leave alone the area beyond |
| 1014 | * @static_size + @reserved_size + @dyn_size. |
| 1015 | * |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1016 | * @unit_size specifies unit size and must be aligned to PAGE_SIZE and |
| 1017 | * equal to or larger than @static_size + @reserved_size + if |
| 1018 | * non-negative, @dyn_size. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1019 | * |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1020 | * The caller should have mapped the first chunk at @base_addr and |
| 1021 | * copied static data to each unit. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1022 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1023 | * If the first chunk ends up with both reserved and dynamic areas, it |
| 1024 | * is served by two chunks - one to serve the core static and reserved |
| 1025 | * areas and the other for the dynamic area. They share the same vm |
| 1026 | * and page map but uses different area allocation map to stay away |
| 1027 | * from each other. The latter chunk is circulated in the chunk slots |
| 1028 | * and available for dynamic allocation like any other chunks. |
| 1029 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1030 | * RETURNS: |
| 1031 | * The determined pcpu_unit_size which can be used to initialize |
| 1032 | * percpu access. |
| 1033 | */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1034 | size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn, |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1035 | size_t static_size, size_t reserved_size, |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1036 | ssize_t dyn_size, size_t unit_size, |
| 1037 | void *base_addr) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1038 | { |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1039 | static struct vm_struct first_vm; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1040 | static int smap[2], dmap[2]; |
Tejun Heo | 6074d5b | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1041 | size_t size_sum = static_size + reserved_size + |
| 1042 | (dyn_size >= 0 ? dyn_size : 0); |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1043 | struct pcpu_chunk *schunk, *dchunk = NULL; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1044 | unsigned int cpu; |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1045 | int i, nr_pages; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1046 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1047 | /* santiy checks */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1048 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
| 1049 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1050 | BUG_ON(!static_size); |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1051 | BUG_ON(!base_addr); |
| 1052 | BUG_ON(unit_size < size_sum); |
| 1053 | BUG_ON(unit_size & ~PAGE_MASK); |
| 1054 | BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1055 | |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1056 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1057 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1058 | pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1059 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
Tejun Heo | cb83b42 | 2009-02-24 11:57:20 +0900 | [diff] [blame] | 1060 | + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1061 | |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1062 | if (dyn_size < 0) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1063 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1064 | |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1065 | first_vm.flags = VM_ALLOC; |
| 1066 | first_vm.size = pcpu_chunk_size; |
| 1067 | first_vm.addr = base_addr; |
| 1068 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1069 | /* |
| 1070 | * Allocate chunk slots. The additional last slot is for |
| 1071 | * empty chunks. |
| 1072 | */ |
| 1073 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1074 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
| 1075 | for (i = 0; i < pcpu_nr_slots; i++) |
| 1076 | INIT_LIST_HEAD(&pcpu_slot[i]); |
| 1077 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1078 | /* |
| 1079 | * Initialize static chunk. If reserved_size is zero, the |
| 1080 | * static chunk covers static area + dynamic allocation area |
| 1081 | * in the first chunk. If reserved_size is not zero, it |
| 1082 | * covers static area + reserved area (mostly used for module |
| 1083 | * static percpu allocation). |
| 1084 | */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1085 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
| 1086 | INIT_LIST_HEAD(&schunk->list); |
| 1087 | schunk->vm = &first_vm; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1088 | schunk->map = smap; |
| 1089 | schunk->map_alloc = ARRAY_SIZE(smap); |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1090 | schunk->page = schunk->page_ar; |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1091 | schunk->immutable = true; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1092 | |
| 1093 | if (reserved_size) { |
| 1094 | schunk->free_size = reserved_size; |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1095 | pcpu_reserved_chunk = schunk; |
| 1096 | pcpu_reserved_chunk_limit = static_size + reserved_size; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1097 | } else { |
| 1098 | schunk->free_size = dyn_size; |
| 1099 | dyn_size = 0; /* dynamic area covered */ |
| 1100 | } |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1101 | schunk->contig_hint = schunk->free_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1102 | |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1103 | schunk->map[schunk->map_used++] = -static_size; |
| 1104 | if (schunk->free_size) |
| 1105 | schunk->map[schunk->map_used++] = schunk->free_size; |
| 1106 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1107 | /* init dynamic chunk if necessary */ |
| 1108 | if (dyn_size) { |
| 1109 | dchunk = alloc_bootmem(sizeof(struct pcpu_chunk)); |
| 1110 | INIT_LIST_HEAD(&dchunk->list); |
| 1111 | dchunk->vm = &first_vm; |
| 1112 | dchunk->map = dmap; |
| 1113 | dchunk->map_alloc = ARRAY_SIZE(dmap); |
| 1114 | dchunk->page = schunk->page_ar; /* share page map with schunk */ |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1115 | dchunk->immutable = true; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1116 | |
| 1117 | dchunk->contig_hint = dchunk->free_size = dyn_size; |
| 1118 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; |
| 1119 | dchunk->map[dchunk->map_used++] = dchunk->free_size; |
| 1120 | } |
| 1121 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1122 | /* assign pages */ |
| 1123 | nr_pages = -1; |
| 1124 | for_each_possible_cpu(cpu) { |
| 1125 | for (i = 0; i < pcpu_unit_pages; i++) { |
| 1126 | struct page *page = get_page_fn(cpu, i); |
| 1127 | |
| 1128 | if (!page) |
| 1129 | break; |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1130 | *pcpu_chunk_pagep(schunk, cpu, i) = page; |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1131 | } |
| 1132 | |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1133 | BUG_ON(i < PFN_UP(static_size)); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1134 | |
| 1135 | if (nr_pages < 0) |
| 1136 | nr_pages = i; |
| 1137 | else |
| 1138 | BUG_ON(nr_pages != i); |
| 1139 | } |
| 1140 | |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1141 | /* link the first chunk in */ |
Tejun Heo | ae9e6bc | 2009-04-02 13:19:54 +0900 | [diff] [blame] | 1142 | pcpu_first_chunk = dchunk ?: schunk; |
| 1143 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1144 | |
| 1145 | /* we're done */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1146 | pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1147 | return pcpu_unit_size; |
| 1148 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1149 | |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1150 | static size_t pcpu_calc_fc_sizes(size_t static_size, size_t reserved_size, |
| 1151 | ssize_t *dyn_sizep) |
| 1152 | { |
| 1153 | size_t size_sum; |
| 1154 | |
| 1155 | size_sum = PFN_ALIGN(static_size + reserved_size + |
| 1156 | (*dyn_sizep >= 0 ? *dyn_sizep : 0)); |
| 1157 | if (*dyn_sizep != 0) |
| 1158 | *dyn_sizep = size_sum - static_size - reserved_size; |
| 1159 | |
| 1160 | return size_sum; |
| 1161 | } |
| 1162 | |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1163 | /* |
| 1164 | * Embedding first chunk setup helper. |
| 1165 | */ |
| 1166 | static void *pcpue_ptr __initdata; |
| 1167 | static size_t pcpue_size __initdata; |
| 1168 | static size_t pcpue_unit_size __initdata; |
| 1169 | |
| 1170 | static struct page * __init pcpue_get_page(unsigned int cpu, int pageno) |
| 1171 | { |
| 1172 | size_t off = (size_t)pageno << PAGE_SHIFT; |
| 1173 | |
| 1174 | if (off >= pcpue_size) |
| 1175 | return NULL; |
| 1176 | |
| 1177 | return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off); |
| 1178 | } |
| 1179 | |
| 1180 | /** |
| 1181 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem |
| 1182 | * @static_size: the size of static percpu area in bytes |
| 1183 | * @reserved_size: the size of reserved percpu area in bytes |
| 1184 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1185 | * |
| 1186 | * This is a helper to ease setting up embedded first percpu chunk and |
| 1187 | * can be called where pcpu_setup_first_chunk() is expected. |
| 1188 | * |
| 1189 | * If this function is used to setup the first chunk, it is allocated |
| 1190 | * as a contiguous area using bootmem allocator and used as-is without |
| 1191 | * being mapped into vmalloc area. This enables the first chunk to |
| 1192 | * piggy back on the linear physical mapping which often uses larger |
| 1193 | * page size. |
| 1194 | * |
| 1195 | * When @dyn_size is positive, dynamic area might be larger than |
Tejun Heo | 788e5ab | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 1196 | * specified to fill page alignment. When @dyn_size is auto, |
| 1197 | * @dyn_size is just big enough to fill page alignment after static |
| 1198 | * and reserved areas. |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1199 | * |
| 1200 | * If the needed size is smaller than the minimum or specified unit |
| 1201 | * size, the leftover is returned to the bootmem allocator. |
| 1202 | * |
| 1203 | * RETURNS: |
| 1204 | * The determined pcpu_unit_size which can be used to initialize |
| 1205 | * percpu access on success, -errno on failure. |
| 1206 | */ |
| 1207 | ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size, |
Tejun Heo | 788e5ab | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 1208 | ssize_t dyn_size) |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1209 | { |
Tejun Heo | fa8a709 | 2009-06-22 11:56:24 +0900 | [diff] [blame] | 1210 | size_t chunk_size; |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1211 | unsigned int cpu; |
| 1212 | |
| 1213 | /* determine parameters and allocate */ |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1214 | pcpue_size = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1215 | |
Tejun Heo | 788e5ab | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 1216 | pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE); |
Tejun Heo | fa8a709 | 2009-06-22 11:56:24 +0900 | [diff] [blame] | 1217 | chunk_size = pcpue_unit_size * num_possible_cpus(); |
| 1218 | |
| 1219 | pcpue_ptr = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, |
| 1220 | __pa(MAX_DMA_ADDRESS)); |
| 1221 | if (!pcpue_ptr) { |
| 1222 | pr_warning("PERCPU: failed to allocate %zu bytes for " |
| 1223 | "embedding\n", chunk_size); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1224 | return -ENOMEM; |
Tejun Heo | fa8a709 | 2009-06-22 11:56:24 +0900 | [diff] [blame] | 1225 | } |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1226 | |
| 1227 | /* return the leftover and copy */ |
| 1228 | for_each_possible_cpu(cpu) { |
| 1229 | void *ptr = pcpue_ptr + cpu * pcpue_unit_size; |
| 1230 | |
| 1231 | free_bootmem(__pa(ptr + pcpue_size), |
| 1232 | pcpue_unit_size - pcpue_size); |
| 1233 | memcpy(ptr, __per_cpu_load, static_size); |
| 1234 | } |
| 1235 | |
| 1236 | /* we're ready, commit */ |
| 1237 | pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n", |
| 1238 | pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size); |
| 1239 | |
| 1240 | return pcpu_setup_first_chunk(pcpue_get_page, static_size, |
| 1241 | reserved_size, dyn_size, |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1242 | pcpue_unit_size, pcpue_ptr); |
Tejun Heo | 66c3a75 | 2009-03-10 16:27:48 +0900 | [diff] [blame] | 1243 | } |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 1244 | |
| 1245 | /* |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1246 | * 4k page first chunk setup helper. |
| 1247 | */ |
| 1248 | static struct page **pcpu4k_pages __initdata; |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1249 | static int pcpu4k_unit_pages __initdata; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1250 | |
| 1251 | static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno) |
| 1252 | { |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1253 | if (pageno < pcpu4k_unit_pages) |
| 1254 | return pcpu4k_pages[cpu * pcpu4k_unit_pages + pageno]; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1255 | return NULL; |
| 1256 | } |
| 1257 | |
| 1258 | /** |
| 1259 | * pcpu_4k_first_chunk - map the first chunk using PAGE_SIZE pages |
| 1260 | * @static_size: the size of static percpu area in bytes |
| 1261 | * @reserved_size: the size of reserved percpu area in bytes |
| 1262 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE |
| 1263 | * @free_fn: funtion to free percpu page, always called with PAGE_SIZE |
| 1264 | * @populate_pte_fn: function to populate pte |
| 1265 | * |
| 1266 | * This is a helper to ease setting up embedded first percpu chunk and |
| 1267 | * can be called where pcpu_setup_first_chunk() is expected. |
| 1268 | * |
| 1269 | * This is the basic allocator. Static percpu area is allocated |
| 1270 | * page-by-page into vmalloc area. |
| 1271 | * |
| 1272 | * RETURNS: |
| 1273 | * The determined pcpu_unit_size which can be used to initialize |
| 1274 | * percpu access on success, -errno on failure. |
| 1275 | */ |
| 1276 | ssize_t __init pcpu_4k_first_chunk(size_t static_size, size_t reserved_size, |
| 1277 | pcpu_fc_alloc_fn_t alloc_fn, |
| 1278 | pcpu_fc_free_fn_t free_fn, |
| 1279 | pcpu_fc_populate_pte_fn_t populate_pte_fn) |
| 1280 | { |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1281 | static struct vm_struct vm; |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1282 | size_t pages_size; |
| 1283 | unsigned int cpu; |
| 1284 | int i, j; |
| 1285 | ssize_t ret; |
| 1286 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1287 | pcpu4k_unit_pages = PFN_UP(max_t(size_t, static_size + reserved_size, |
| 1288 | PCPU_MIN_UNIT_SIZE)); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1289 | |
| 1290 | /* unaligned allocations can't be freed, round up to page size */ |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1291 | pages_size = PFN_ALIGN(pcpu4k_unit_pages * num_possible_cpus() * |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1292 | sizeof(pcpu4k_pages[0])); |
| 1293 | pcpu4k_pages = alloc_bootmem(pages_size); |
| 1294 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1295 | /* allocate pages */ |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1296 | j = 0; |
| 1297 | for_each_possible_cpu(cpu) |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1298 | for (i = 0; i < pcpu4k_unit_pages; i++) { |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1299 | void *ptr; |
| 1300 | |
| 1301 | ptr = alloc_fn(cpu, PAGE_SIZE); |
| 1302 | if (!ptr) { |
| 1303 | pr_warning("PERCPU: failed to allocate " |
| 1304 | "4k page for cpu%u\n", cpu); |
| 1305 | goto enomem; |
| 1306 | } |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1307 | pcpu4k_pages[j++] = virt_to_page(ptr); |
| 1308 | } |
| 1309 | |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1310 | /* allocate vm area, map the pages and copy static data */ |
| 1311 | vm.flags = VM_ALLOC; |
| 1312 | vm.size = num_possible_cpus() * pcpu4k_unit_pages << PAGE_SHIFT; |
| 1313 | vm_area_register_early(&vm, PAGE_SIZE); |
| 1314 | |
| 1315 | for_each_possible_cpu(cpu) { |
| 1316 | unsigned long unit_addr = (unsigned long)vm.addr + |
| 1317 | (cpu * pcpu4k_unit_pages << PAGE_SHIFT); |
| 1318 | |
| 1319 | for (i = 0; i < pcpu4k_unit_pages; i++) |
| 1320 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
| 1321 | |
| 1322 | /* pte already populated, the following shouldn't fail */ |
| 1323 | ret = __pcpu_map_pages(unit_addr, |
| 1324 | &pcpu4k_pages[cpu * pcpu4k_unit_pages], |
| 1325 | pcpu4k_unit_pages); |
| 1326 | if (ret < 0) |
| 1327 | panic("failed to map percpu area, err=%zd\n", ret); |
| 1328 | |
| 1329 | /* |
| 1330 | * FIXME: Archs with virtual cache should flush local |
| 1331 | * cache for the linear mapping here - something |
| 1332 | * equivalent to flush_cache_vmap() on the local cpu. |
| 1333 | * flush_cache_vmap() can't be used as most supporting |
| 1334 | * data structures are not set up yet. |
| 1335 | */ |
| 1336 | |
| 1337 | /* copy static data */ |
| 1338 | memcpy((void *)unit_addr, __per_cpu_load, static_size); |
| 1339 | } |
| 1340 | |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1341 | /* we're ready, commit */ |
Tejun Heo | 8f05a6a | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1342 | pr_info("PERCPU: %d 4k pages per cpu, static data %zu bytes\n", |
| 1343 | pcpu4k_unit_pages, static_size); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1344 | |
| 1345 | ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size, |
| 1346 | reserved_size, -1, |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1347 | pcpu4k_unit_pages << PAGE_SHIFT, vm.addr); |
Tejun Heo | d4b95f8 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1348 | goto out_free_ar; |
| 1349 | |
| 1350 | enomem: |
| 1351 | while (--j >= 0) |
| 1352 | free_fn(page_address(pcpu4k_pages[j]), PAGE_SIZE); |
| 1353 | ret = -ENOMEM; |
| 1354 | out_free_ar: |
| 1355 | free_bootmem(__pa(pcpu4k_pages), pages_size); |
| 1356 | return ret; |
| 1357 | } |
| 1358 | |
| 1359 | /* |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1360 | * Large page remapping first chunk setup helper |
| 1361 | */ |
| 1362 | #ifdef CONFIG_NEED_MULTIPLE_NODES |
| 1363 | struct pcpul_ent { |
| 1364 | unsigned int cpu; |
| 1365 | void *ptr; |
| 1366 | }; |
| 1367 | |
| 1368 | static size_t pcpul_size; |
| 1369 | static size_t pcpul_unit_size; |
| 1370 | static struct pcpul_ent *pcpul_map; |
| 1371 | static struct vm_struct pcpul_vm; |
| 1372 | |
| 1373 | static struct page * __init pcpul_get_page(unsigned int cpu, int pageno) |
| 1374 | { |
| 1375 | size_t off = (size_t)pageno << PAGE_SHIFT; |
| 1376 | |
| 1377 | if (off >= pcpul_size) |
| 1378 | return NULL; |
| 1379 | |
| 1380 | return virt_to_page(pcpul_map[cpu].ptr + off); |
| 1381 | } |
| 1382 | |
| 1383 | /** |
| 1384 | * pcpu_lpage_first_chunk - remap the first percpu chunk using large page |
| 1385 | * @static_size: the size of static percpu area in bytes |
| 1386 | * @reserved_size: the size of reserved percpu area in bytes |
| 1387 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
| 1388 | * @lpage_size: the size of a large page |
| 1389 | * @alloc_fn: function to allocate percpu lpage, always called with lpage_size |
| 1390 | * @free_fn: function to free percpu memory, @size <= lpage_size |
| 1391 | * @map_fn: function to map percpu lpage, always called with lpage_size |
| 1392 | * |
| 1393 | * This allocator uses large page as unit. A large page is allocated |
| 1394 | * for each cpu and each is remapped into vmalloc area using large |
| 1395 | * page mapping. As large page can be quite large, only part of it is |
| 1396 | * used for the first chunk. Unused part is returned to the bootmem |
| 1397 | * allocator. |
| 1398 | * |
| 1399 | * So, the large pages are mapped twice - once to the physical mapping |
| 1400 | * and to the vmalloc area for the first percpu chunk. The double |
| 1401 | * mapping does add one more large TLB entry pressure but still is |
| 1402 | * much better than only using 4k mappings while still being NUMA |
| 1403 | * friendly. |
| 1404 | * |
| 1405 | * RETURNS: |
| 1406 | * The determined pcpu_unit_size which can be used to initialize |
| 1407 | * percpu access on success, -errno on failure. |
| 1408 | */ |
| 1409 | ssize_t __init pcpu_lpage_first_chunk(size_t static_size, size_t reserved_size, |
| 1410 | ssize_t dyn_size, size_t lpage_size, |
| 1411 | pcpu_fc_alloc_fn_t alloc_fn, |
| 1412 | pcpu_fc_free_fn_t free_fn, |
| 1413 | pcpu_fc_map_fn_t map_fn) |
| 1414 | { |
| 1415 | size_t size_sum; |
| 1416 | size_t map_size; |
| 1417 | unsigned int cpu; |
| 1418 | int i, j; |
| 1419 | ssize_t ret; |
| 1420 | |
| 1421 | /* |
| 1422 | * Currently supports only single page. Supporting multiple |
| 1423 | * pages won't be too difficult if it ever becomes necessary. |
| 1424 | */ |
| 1425 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
| 1426 | |
| 1427 | pcpul_unit_size = lpage_size; |
| 1428 | pcpul_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
| 1429 | if (pcpul_size > pcpul_unit_size) { |
| 1430 | pr_warning("PERCPU: static data is larger than large page, " |
| 1431 | "can't use large page\n"); |
| 1432 | return -EINVAL; |
| 1433 | } |
| 1434 | |
| 1435 | /* allocate pointer array and alloc large pages */ |
| 1436 | map_size = PFN_ALIGN(num_possible_cpus() * sizeof(pcpul_map[0])); |
| 1437 | pcpul_map = alloc_bootmem(map_size); |
| 1438 | |
| 1439 | for_each_possible_cpu(cpu) { |
| 1440 | void *ptr; |
| 1441 | |
| 1442 | ptr = alloc_fn(cpu, lpage_size); |
| 1443 | if (!ptr) { |
| 1444 | pr_warning("PERCPU: failed to allocate large page " |
| 1445 | "for cpu%u\n", cpu); |
| 1446 | goto enomem; |
| 1447 | } |
| 1448 | |
| 1449 | /* |
| 1450 | * Only use pcpul_size bytes and give back the rest. |
| 1451 | * |
| 1452 | * Ingo: The lpage_size up-rounding bootmem is needed |
| 1453 | * to make sure the partial lpage is still fully RAM - |
| 1454 | * it's not well-specified to have a incompatible area |
| 1455 | * (unmapped RAM, device memory, etc.) in that hole. |
| 1456 | */ |
| 1457 | free_fn(ptr + pcpul_size, lpage_size - pcpul_size); |
| 1458 | |
| 1459 | pcpul_map[cpu].cpu = cpu; |
| 1460 | pcpul_map[cpu].ptr = ptr; |
| 1461 | |
| 1462 | memcpy(ptr, __per_cpu_load, static_size); |
| 1463 | } |
| 1464 | |
| 1465 | /* allocate address and map */ |
| 1466 | pcpul_vm.flags = VM_ALLOC; |
| 1467 | pcpul_vm.size = num_possible_cpus() * pcpul_unit_size; |
| 1468 | vm_area_register_early(&pcpul_vm, pcpul_unit_size); |
| 1469 | |
| 1470 | for_each_possible_cpu(cpu) |
| 1471 | map_fn(pcpul_map[cpu].ptr, pcpul_unit_size, |
| 1472 | pcpul_vm.addr + cpu * pcpul_unit_size); |
| 1473 | |
| 1474 | /* we're ready, commit */ |
| 1475 | pr_info("PERCPU: Remapped at %p with large pages, static data " |
| 1476 | "%zu bytes\n", pcpul_vm.addr, static_size); |
| 1477 | |
| 1478 | ret = pcpu_setup_first_chunk(pcpul_get_page, static_size, |
| 1479 | reserved_size, dyn_size, pcpul_unit_size, |
Tejun Heo | 38a6be5 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1480 | pcpul_vm.addr); |
Tejun Heo | 8c4bfc6 | 2009-07-04 08:10:59 +0900 | [diff] [blame] | 1481 | |
| 1482 | /* sort pcpul_map array for pcpu_lpage_remapped() */ |
| 1483 | for (i = 0; i < num_possible_cpus() - 1; i++) |
| 1484 | for (j = i + 1; j < num_possible_cpus(); j++) |
| 1485 | if (pcpul_map[i].ptr > pcpul_map[j].ptr) { |
| 1486 | struct pcpul_ent tmp = pcpul_map[i]; |
| 1487 | pcpul_map[i] = pcpul_map[j]; |
| 1488 | pcpul_map[j] = tmp; |
| 1489 | } |
| 1490 | |
| 1491 | return ret; |
| 1492 | |
| 1493 | enomem: |
| 1494 | for_each_possible_cpu(cpu) |
| 1495 | if (pcpul_map[cpu].ptr) |
| 1496 | free_fn(pcpul_map[cpu].ptr, pcpul_size); |
| 1497 | free_bootmem(__pa(pcpul_map), map_size); |
| 1498 | return -ENOMEM; |
| 1499 | } |
| 1500 | |
| 1501 | /** |
| 1502 | * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area |
| 1503 | * @kaddr: the kernel address in question |
| 1504 | * |
| 1505 | * Determine whether @kaddr falls in the pcpul recycled area. This is |
| 1506 | * used by pageattr to detect VM aliases and break up the pcpu large |
| 1507 | * page mapping such that the same physical page is not mapped under |
| 1508 | * different attributes. |
| 1509 | * |
| 1510 | * The recycled area is always at the tail of a partially used large |
| 1511 | * page. |
| 1512 | * |
| 1513 | * RETURNS: |
| 1514 | * Address of corresponding remapped pcpu address if match is found; |
| 1515 | * otherwise, NULL. |
| 1516 | */ |
| 1517 | void *pcpu_lpage_remapped(void *kaddr) |
| 1518 | { |
| 1519 | unsigned long unit_mask = pcpul_unit_size - 1; |
| 1520 | void *lpage_addr = (void *)((unsigned long)kaddr & ~unit_mask); |
| 1521 | unsigned long offset = (unsigned long)kaddr & unit_mask; |
| 1522 | int left = 0, right = num_possible_cpus() - 1; |
| 1523 | int pos; |
| 1524 | |
| 1525 | /* pcpul in use at all? */ |
| 1526 | if (!pcpul_map) |
| 1527 | return NULL; |
| 1528 | |
| 1529 | /* okay, perform binary search */ |
| 1530 | while (left <= right) { |
| 1531 | pos = (left + right) / 2; |
| 1532 | |
| 1533 | if (pcpul_map[pos].ptr < lpage_addr) |
| 1534 | left = pos + 1; |
| 1535 | else if (pcpul_map[pos].ptr > lpage_addr) |
| 1536 | right = pos - 1; |
| 1537 | else { |
| 1538 | /* it shouldn't be in the area for the first chunk */ |
| 1539 | WARN_ON(offset < pcpul_size); |
| 1540 | |
| 1541 | return pcpul_vm.addr + |
| 1542 | pcpul_map[pos].cpu * pcpul_unit_size + offset; |
| 1543 | } |
| 1544 | } |
| 1545 | |
| 1546 | return NULL; |
| 1547 | } |
| 1548 | #endif |
| 1549 | |
| 1550 | /* |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 1551 | * Generic percpu area setup. |
| 1552 | * |
| 1553 | * The embedding helper is used because its behavior closely resembles |
| 1554 | * the original non-dynamic generic percpu area setup. This is |
| 1555 | * important because many archs have addressing restrictions and might |
| 1556 | * fail if the percpu area is located far away from the previous |
| 1557 | * location. As an added bonus, in non-NUMA cases, embedding is |
| 1558 | * generally a good idea TLB-wise because percpu area can piggy back |
| 1559 | * on the physical linear memory mapping which uses large page |
| 1560 | * mappings on applicable archs. |
| 1561 | */ |
| 1562 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
| 1563 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
| 1564 | EXPORT_SYMBOL(__per_cpu_offset); |
| 1565 | |
| 1566 | void __init setup_per_cpu_areas(void) |
| 1567 | { |
| 1568 | size_t static_size = __per_cpu_end - __per_cpu_start; |
| 1569 | ssize_t unit_size; |
| 1570 | unsigned long delta; |
| 1571 | unsigned int cpu; |
| 1572 | |
| 1573 | /* |
| 1574 | * Always reserve area for module percpu variables. That's |
| 1575 | * what the legacy allocator did. |
| 1576 | */ |
| 1577 | unit_size = pcpu_embed_first_chunk(static_size, PERCPU_MODULE_RESERVE, |
Tejun Heo | 788e5ab | 2009-07-04 08:10:58 +0900 | [diff] [blame] | 1578 | PERCPU_DYNAMIC_RESERVE); |
Tejun Heo | e74e396 | 2009-03-30 19:07:44 +0900 | [diff] [blame] | 1579 | if (unit_size < 0) |
| 1580 | panic("Failed to initialized percpu areas."); |
| 1581 | |
| 1582 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; |
| 1583 | for_each_possible_cpu(cpu) |
| 1584 | __per_cpu_offset[cpu] = delta + cpu * unit_size; |
| 1585 | } |
| 1586 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |