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 |
| 26 | * percpu base registers UNIT_SIZE apart. |
| 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. |
| 41 | * Chunks are also linked into a rb tree to ease address to chunk |
| 42 | * mapping during free. |
| 43 | * |
| 44 | * To use this allocator, arch code should do the followings. |
| 45 | * |
| 46 | * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA |
| 47 | * |
| 48 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
| 49 | * regular address to percpu pointer and back |
| 50 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 51 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
| 52 | * setup the first chunk containing the kernel static percpu area |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 53 | */ |
| 54 | |
| 55 | #include <linux/bitmap.h> |
| 56 | #include <linux/bootmem.h> |
| 57 | #include <linux/list.h> |
| 58 | #include <linux/mm.h> |
| 59 | #include <linux/module.h> |
| 60 | #include <linux/mutex.h> |
| 61 | #include <linux/percpu.h> |
| 62 | #include <linux/pfn.h> |
| 63 | #include <linux/rbtree.h> |
| 64 | #include <linux/slab.h> |
| 65 | #include <linux/vmalloc.h> |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 66 | #include <linux/workqueue.h> |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 67 | |
| 68 | #include <asm/cacheflush.h> |
| 69 | #include <asm/tlbflush.h> |
| 70 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 71 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
| 72 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ |
| 73 | |
| 74 | struct pcpu_chunk { |
| 75 | struct list_head list; /* linked to pcpu_slot lists */ |
| 76 | struct rb_node rb_node; /* key is chunk->vm->addr */ |
| 77 | int free_size; /* free bytes in the chunk */ |
| 78 | int contig_hint; /* max contiguous size hint */ |
| 79 | struct vm_struct *vm; /* mapped vmalloc region */ |
| 80 | int map_used; /* # of map entries used */ |
| 81 | int map_alloc; /* # of map entries allocated */ |
| 82 | int *map; /* allocation map */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 83 | bool immutable; /* no [de]population allowed */ |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 84 | struct page **page; /* points to page array */ |
| 85 | struct page *page_ar[]; /* #cpus * UNIT_PAGES */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 86 | }; |
| 87 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 88 | static int pcpu_unit_pages __read_mostly; |
| 89 | static int pcpu_unit_size __read_mostly; |
| 90 | static int pcpu_chunk_size __read_mostly; |
| 91 | static int pcpu_nr_slots __read_mostly; |
| 92 | static size_t pcpu_chunk_struct_size __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 93 | |
| 94 | /* 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] | 95 | void *pcpu_base_addr __read_mostly; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 96 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
| 97 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 98 | /* optional reserved chunk, only accessible for reserved allocations */ |
| 99 | static struct pcpu_chunk *pcpu_reserved_chunk; |
| 100 | /* offset limit of the reserved chunk */ |
| 101 | static int pcpu_reserved_chunk_limit; |
| 102 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 103 | /* |
| 104 | * One mutex to rule them all. |
| 105 | * |
| 106 | * The following mutex is grabbed in the outermost public alloc/free |
| 107 | * interface functions and released only when the operation is |
| 108 | * complete. As such, every function in this file other than the |
| 109 | * outermost functions are called under pcpu_mutex. |
| 110 | * |
| 111 | * It can easily be switched to use spinlock such that only the area |
| 112 | * allocation and page population commit are protected with it doing |
| 113 | * actual [de]allocation without holding any lock. However, given |
| 114 | * what this allocator does, I think it's better to let them run |
| 115 | * sequentially. |
| 116 | */ |
| 117 | static DEFINE_MUTEX(pcpu_mutex); |
| 118 | |
Tejun Heo | 40150d3 | 2009-02-24 12:32:28 +0900 | [diff] [blame] | 119 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 120 | static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */ |
| 121 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 122 | /* reclaim work to release fully free chunks, scheduled from free path */ |
| 123 | static void pcpu_reclaim(struct work_struct *work); |
| 124 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); |
| 125 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 126 | static int __pcpu_size_to_slot(int size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 127 | { |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 128 | int highbit = fls(size); /* size is in bytes */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 129 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
| 130 | } |
| 131 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 132 | static int pcpu_size_to_slot(int size) |
| 133 | { |
| 134 | if (size == pcpu_unit_size) |
| 135 | return pcpu_nr_slots - 1; |
| 136 | return __pcpu_size_to_slot(size); |
| 137 | } |
| 138 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 139 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
| 140 | { |
| 141 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) |
| 142 | return 0; |
| 143 | |
| 144 | return pcpu_size_to_slot(chunk->free_size); |
| 145 | } |
| 146 | |
| 147 | static int pcpu_page_idx(unsigned int cpu, int page_idx) |
| 148 | { |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 149 | return cpu * pcpu_unit_pages + page_idx; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 150 | } |
| 151 | |
| 152 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, |
| 153 | unsigned int cpu, int page_idx) |
| 154 | { |
| 155 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; |
| 156 | } |
| 157 | |
| 158 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
| 159 | unsigned int cpu, int page_idx) |
| 160 | { |
| 161 | return (unsigned long)chunk->vm->addr + |
| 162 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); |
| 163 | } |
| 164 | |
| 165 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, |
| 166 | int page_idx) |
| 167 | { |
| 168 | return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL; |
| 169 | } |
| 170 | |
| 171 | /** |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 172 | * pcpu_mem_alloc - allocate memory |
| 173 | * @size: bytes to allocate |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 174 | * |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 175 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
| 176 | * kzalloc() is used; otherwise, vmalloc() is used. The returned |
| 177 | * memory is always zeroed. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 178 | * |
| 179 | * RETURNS: |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 180 | * Pointer to the allocated area on success, NULL on failure. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 181 | */ |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 182 | static void *pcpu_mem_alloc(size_t size) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 183 | { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 184 | if (size <= PAGE_SIZE) |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 185 | return kzalloc(size, GFP_KERNEL); |
| 186 | else { |
| 187 | void *ptr = vmalloc(size); |
| 188 | if (ptr) |
| 189 | memset(ptr, 0, size); |
| 190 | return ptr; |
| 191 | } |
| 192 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 193 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 194 | /** |
| 195 | * pcpu_mem_free - free memory |
| 196 | * @ptr: memory to free |
| 197 | * @size: size of the area |
| 198 | * |
| 199 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). |
| 200 | */ |
| 201 | static void pcpu_mem_free(void *ptr, size_t size) |
| 202 | { |
| 203 | if (size <= PAGE_SIZE) |
| 204 | kfree(ptr); |
| 205 | else |
| 206 | vfree(ptr); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 207 | } |
| 208 | |
| 209 | /** |
| 210 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot |
| 211 | * @chunk: chunk of interest |
| 212 | * @oslot: the previous slot it was on |
| 213 | * |
| 214 | * This function is called after an allocation or free changed @chunk. |
| 215 | * New slot according to the changed state is determined and @chunk is |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 216 | * moved to the slot. Note that the reserved chunk is never put on |
| 217 | * chunk slots. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 218 | */ |
| 219 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) |
| 220 | { |
| 221 | int nslot = pcpu_chunk_slot(chunk); |
| 222 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 223 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 224 | if (oslot < nslot) |
| 225 | list_move(&chunk->list, &pcpu_slot[nslot]); |
| 226 | else |
| 227 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | static struct rb_node **pcpu_chunk_rb_search(void *addr, |
| 232 | struct rb_node **parentp) |
| 233 | { |
| 234 | struct rb_node **p = &pcpu_addr_root.rb_node; |
| 235 | struct rb_node *parent = NULL; |
| 236 | struct pcpu_chunk *chunk; |
| 237 | |
| 238 | while (*p) { |
| 239 | parent = *p; |
| 240 | chunk = rb_entry(parent, struct pcpu_chunk, rb_node); |
| 241 | |
| 242 | if (addr < chunk->vm->addr) |
| 243 | p = &(*p)->rb_left; |
| 244 | else if (addr > chunk->vm->addr) |
| 245 | p = &(*p)->rb_right; |
| 246 | else |
| 247 | break; |
| 248 | } |
| 249 | |
| 250 | if (parentp) |
| 251 | *parentp = parent; |
| 252 | return p; |
| 253 | } |
| 254 | |
| 255 | /** |
| 256 | * pcpu_chunk_addr_search - search for chunk containing specified address |
| 257 | * @addr: address to search for |
| 258 | * |
| 259 | * Look for chunk which might contain @addr. More specifically, it |
| 260 | * searchs for the chunk with the highest start address which isn't |
| 261 | * beyond @addr. |
| 262 | * |
| 263 | * RETURNS: |
| 264 | * The address of the found chunk. |
| 265 | */ |
| 266 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) |
| 267 | { |
| 268 | struct rb_node *n, *parent; |
| 269 | struct pcpu_chunk *chunk; |
| 270 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 271 | /* is it in the reserved chunk? */ |
| 272 | if (pcpu_reserved_chunk) { |
| 273 | void *start = pcpu_reserved_chunk->vm->addr; |
| 274 | |
| 275 | if (addr >= start && addr < start + pcpu_reserved_chunk_limit) |
| 276 | return pcpu_reserved_chunk; |
| 277 | } |
| 278 | |
| 279 | /* nah... search the regular ones */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 280 | n = *pcpu_chunk_rb_search(addr, &parent); |
| 281 | if (!n) { |
| 282 | /* no exactly matching chunk, the parent is the closest */ |
| 283 | n = parent; |
| 284 | BUG_ON(!n); |
| 285 | } |
| 286 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); |
| 287 | |
| 288 | if (addr < chunk->vm->addr) { |
| 289 | /* the parent was the next one, look for the previous one */ |
| 290 | n = rb_prev(n); |
| 291 | BUG_ON(!n); |
| 292 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); |
| 293 | } |
| 294 | |
| 295 | return chunk; |
| 296 | } |
| 297 | |
| 298 | /** |
| 299 | * pcpu_chunk_addr_insert - insert chunk into address rb tree |
| 300 | * @new: chunk to insert |
| 301 | * |
| 302 | * Insert @new into address rb tree. |
| 303 | */ |
| 304 | static void pcpu_chunk_addr_insert(struct pcpu_chunk *new) |
| 305 | { |
| 306 | struct rb_node **p, *parent; |
| 307 | |
| 308 | p = pcpu_chunk_rb_search(new->vm->addr, &parent); |
| 309 | BUG_ON(*p); |
| 310 | rb_link_node(&new->rb_node, parent, p); |
| 311 | rb_insert_color(&new->rb_node, &pcpu_addr_root); |
| 312 | } |
| 313 | |
| 314 | /** |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 315 | * pcpu_extend_area_map - extend area map for allocation |
| 316 | * @chunk: target chunk |
| 317 | * |
| 318 | * Extend area map of @chunk so that it can accomodate an allocation. |
| 319 | * A single allocation can split an area into three areas, so this |
| 320 | * function makes sure that @chunk->map has at least two extra slots. |
| 321 | * |
| 322 | * RETURNS: |
| 323 | * 0 if noop, 1 if successfully extended, -errno on failure. |
| 324 | */ |
| 325 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) |
| 326 | { |
| 327 | int new_alloc; |
| 328 | int *new; |
| 329 | size_t size; |
| 330 | |
| 331 | /* has enough? */ |
| 332 | if (chunk->map_alloc >= chunk->map_used + 2) |
| 333 | return 0; |
| 334 | |
| 335 | new_alloc = PCPU_DFL_MAP_ALLOC; |
| 336 | while (new_alloc < chunk->map_used + 2) |
| 337 | new_alloc *= 2; |
| 338 | |
| 339 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); |
| 340 | if (!new) |
| 341 | return -ENOMEM; |
| 342 | |
| 343 | size = chunk->map_alloc * sizeof(chunk->map[0]); |
| 344 | memcpy(new, chunk->map, size); |
| 345 | |
| 346 | /* |
| 347 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is |
| 348 | * one of the first chunks and still using static map. |
| 349 | */ |
| 350 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) |
| 351 | pcpu_mem_free(chunk->map, size); |
| 352 | |
| 353 | chunk->map_alloc = new_alloc; |
| 354 | chunk->map = new; |
| 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | /** |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 359 | * pcpu_split_block - split a map block |
| 360 | * @chunk: chunk of interest |
| 361 | * @i: index of map block to split |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 362 | * @head: head size in bytes (can be 0) |
| 363 | * @tail: tail size in bytes (can be 0) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 364 | * |
| 365 | * Split the @i'th map block into two or three blocks. If @head is |
| 366 | * non-zero, @head bytes block is inserted before block @i moving it |
| 367 | * to @i+1 and reducing its size by @head bytes. |
| 368 | * |
| 369 | * If @tail is non-zero, the target block, which can be @i or @i+1 |
| 370 | * depending on @head, is reduced by @tail bytes and @tail byte block |
| 371 | * is inserted after the target block. |
| 372 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 373 | * @chunk->map must have enough free slots to accomodate the split. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 374 | */ |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 375 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
| 376 | int head, int tail) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 377 | { |
| 378 | int nr_extra = !!head + !!tail; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 379 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 380 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 381 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 382 | /* insert new subblocks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 383 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
| 384 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
| 385 | chunk->map_used += nr_extra; |
| 386 | |
| 387 | if (head) { |
| 388 | chunk->map[i + 1] = chunk->map[i] - head; |
| 389 | chunk->map[i++] = head; |
| 390 | } |
| 391 | if (tail) { |
| 392 | chunk->map[i++] -= tail; |
| 393 | chunk->map[i] = tail; |
| 394 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 395 | } |
| 396 | |
| 397 | /** |
| 398 | * pcpu_alloc_area - allocate area from a pcpu_chunk |
| 399 | * @chunk: chunk of interest |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 400 | * @size: wanted size in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 401 | * @align: wanted align |
| 402 | * |
| 403 | * Try to allocate @size bytes area aligned at @align from @chunk. |
| 404 | * Note that this function only allocates the offset. It doesn't |
| 405 | * populate or map the area. |
| 406 | * |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 407 | * @chunk->map must have at least two free slots. |
| 408 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 409 | * RETURNS: |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 410 | * Allocated offset in @chunk on success, -1 if no matching area is |
| 411 | * found. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 412 | */ |
| 413 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) |
| 414 | { |
| 415 | int oslot = pcpu_chunk_slot(chunk); |
| 416 | int max_contig = 0; |
| 417 | int i, off; |
| 418 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 419 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
| 420 | bool is_last = i + 1 == chunk->map_used; |
| 421 | int head, tail; |
| 422 | |
| 423 | /* extra for alignment requirement */ |
| 424 | head = ALIGN(off, align) - off; |
| 425 | BUG_ON(i == 0 && head != 0); |
| 426 | |
| 427 | if (chunk->map[i] < 0) |
| 428 | continue; |
| 429 | if (chunk->map[i] < head + size) { |
| 430 | max_contig = max(chunk->map[i], max_contig); |
| 431 | continue; |
| 432 | } |
| 433 | |
| 434 | /* |
| 435 | * If head is small or the previous block is free, |
| 436 | * merge'em. Note that 'small' is defined as smaller |
| 437 | * than sizeof(int), which is very small but isn't too |
| 438 | * uncommon for percpu allocations. |
| 439 | */ |
| 440 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { |
| 441 | if (chunk->map[i - 1] > 0) |
| 442 | chunk->map[i - 1] += head; |
| 443 | else { |
| 444 | chunk->map[i - 1] -= head; |
| 445 | chunk->free_size -= head; |
| 446 | } |
| 447 | chunk->map[i] -= head; |
| 448 | off += head; |
| 449 | head = 0; |
| 450 | } |
| 451 | |
| 452 | /* if tail is small, just keep it around */ |
| 453 | tail = chunk->map[i] - head - size; |
| 454 | if (tail < sizeof(int)) |
| 455 | tail = 0; |
| 456 | |
| 457 | /* split if warranted */ |
| 458 | if (head || tail) { |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 459 | pcpu_split_block(chunk, i, head, tail); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 460 | if (head) { |
| 461 | i++; |
| 462 | off += head; |
| 463 | max_contig = max(chunk->map[i - 1], max_contig); |
| 464 | } |
| 465 | if (tail) |
| 466 | max_contig = max(chunk->map[i + 1], max_contig); |
| 467 | } |
| 468 | |
| 469 | /* update hint and mark allocated */ |
| 470 | if (is_last) |
| 471 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 472 | else |
| 473 | chunk->contig_hint = max(chunk->contig_hint, |
| 474 | max_contig); |
| 475 | |
| 476 | chunk->free_size -= chunk->map[i]; |
| 477 | chunk->map[i] = -chunk->map[i]; |
| 478 | |
| 479 | pcpu_chunk_relocate(chunk, oslot); |
| 480 | return off; |
| 481 | } |
| 482 | |
| 483 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 484 | pcpu_chunk_relocate(chunk, oslot); |
| 485 | |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 486 | /* tell the upper layer that this chunk has no matching area */ |
| 487 | return -1; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 488 | } |
| 489 | |
| 490 | /** |
| 491 | * pcpu_free_area - free area to a pcpu_chunk |
| 492 | * @chunk: chunk of interest |
| 493 | * @freeme: offset of area to free |
| 494 | * |
| 495 | * Free area starting from @freeme to @chunk. Note that this function |
| 496 | * only modifies the allocation map. It doesn't depopulate or unmap |
| 497 | * the area. |
| 498 | */ |
| 499 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) |
| 500 | { |
| 501 | int oslot = pcpu_chunk_slot(chunk); |
| 502 | int i, off; |
| 503 | |
| 504 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) |
| 505 | if (off == freeme) |
| 506 | break; |
| 507 | BUG_ON(off != freeme); |
| 508 | BUG_ON(chunk->map[i] > 0); |
| 509 | |
| 510 | chunk->map[i] = -chunk->map[i]; |
| 511 | chunk->free_size += chunk->map[i]; |
| 512 | |
| 513 | /* merge with previous? */ |
| 514 | if (i > 0 && chunk->map[i - 1] >= 0) { |
| 515 | chunk->map[i - 1] += chunk->map[i]; |
| 516 | chunk->map_used--; |
| 517 | memmove(&chunk->map[i], &chunk->map[i + 1], |
| 518 | (chunk->map_used - i) * sizeof(chunk->map[0])); |
| 519 | i--; |
| 520 | } |
| 521 | /* merge with next? */ |
| 522 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { |
| 523 | chunk->map[i] += chunk->map[i + 1]; |
| 524 | chunk->map_used--; |
| 525 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], |
| 526 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); |
| 527 | } |
| 528 | |
| 529 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); |
| 530 | pcpu_chunk_relocate(chunk, oslot); |
| 531 | } |
| 532 | |
| 533 | /** |
| 534 | * pcpu_unmap - unmap pages out of a pcpu_chunk |
| 535 | * @chunk: chunk of interest |
| 536 | * @page_start: page index of the first page to unmap |
| 537 | * @page_end: page index of the last page to unmap + 1 |
| 538 | * @flush: whether to flush cache and tlb or not |
| 539 | * |
| 540 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. |
| 541 | * If @flush is true, vcache is flushed before unmapping and tlb |
| 542 | * after. |
| 543 | */ |
| 544 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, |
| 545 | bool flush) |
| 546 | { |
| 547 | unsigned int last = num_possible_cpus() - 1; |
| 548 | unsigned int cpu; |
| 549 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 550 | /* unmap must not be done on immutable chunk */ |
| 551 | WARN_ON(chunk->immutable); |
| 552 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 553 | /* |
| 554 | * Each flushing trial can be very expensive, issue flush on |
| 555 | * the whole region at once rather than doing it for each cpu. |
| 556 | * This could be an overkill but is more scalable. |
| 557 | */ |
| 558 | if (flush) |
| 559 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 560 | pcpu_chunk_addr(chunk, last, page_end)); |
| 561 | |
| 562 | for_each_possible_cpu(cpu) |
| 563 | unmap_kernel_range_noflush( |
| 564 | pcpu_chunk_addr(chunk, cpu, page_start), |
| 565 | (page_end - page_start) << PAGE_SHIFT); |
| 566 | |
| 567 | /* ditto as flush_cache_vunmap() */ |
| 568 | if (flush) |
| 569 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), |
| 570 | pcpu_chunk_addr(chunk, last, page_end)); |
| 571 | } |
| 572 | |
| 573 | /** |
| 574 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk |
| 575 | * @chunk: chunk to depopulate |
| 576 | * @off: offset to the area to depopulate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 577 | * @size: size of the area to depopulate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 578 | * @flush: whether to flush cache and tlb or not |
| 579 | * |
| 580 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) |
| 581 | * from @chunk. If @flush is true, vcache is flushed before unmapping |
| 582 | * and tlb after. |
| 583 | */ |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 584 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
| 585 | bool flush) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 586 | { |
| 587 | int page_start = PFN_DOWN(off); |
| 588 | int page_end = PFN_UP(off + size); |
| 589 | int unmap_start = -1; |
| 590 | int uninitialized_var(unmap_end); |
| 591 | unsigned int cpu; |
| 592 | int i; |
| 593 | |
| 594 | for (i = page_start; i < page_end; i++) { |
| 595 | for_each_possible_cpu(cpu) { |
| 596 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 597 | |
| 598 | if (!*pagep) |
| 599 | continue; |
| 600 | |
| 601 | __free_page(*pagep); |
| 602 | |
| 603 | /* |
| 604 | * If it's partial depopulation, it might get |
| 605 | * populated or depopulated again. Mark the |
| 606 | * page gone. |
| 607 | */ |
| 608 | *pagep = NULL; |
| 609 | |
| 610 | unmap_start = unmap_start < 0 ? i : unmap_start; |
| 611 | unmap_end = i + 1; |
| 612 | } |
| 613 | } |
| 614 | |
| 615 | if (unmap_start >= 0) |
| 616 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); |
| 617 | } |
| 618 | |
| 619 | /** |
| 620 | * pcpu_map - map pages into a pcpu_chunk |
| 621 | * @chunk: chunk of interest |
| 622 | * @page_start: page index of the first page to map |
| 623 | * @page_end: page index of the last page to map + 1 |
| 624 | * |
| 625 | * For each cpu, map pages [@page_start,@page_end) into @chunk. |
| 626 | * vcache is flushed afterwards. |
| 627 | */ |
| 628 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) |
| 629 | { |
| 630 | unsigned int last = num_possible_cpus() - 1; |
| 631 | unsigned int cpu; |
| 632 | int err; |
| 633 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 634 | /* map must not be done on immutable chunk */ |
| 635 | WARN_ON(chunk->immutable); |
| 636 | |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 637 | for_each_possible_cpu(cpu) { |
| 638 | err = map_kernel_range_noflush( |
| 639 | pcpu_chunk_addr(chunk, cpu, page_start), |
| 640 | (page_end - page_start) << PAGE_SHIFT, |
| 641 | PAGE_KERNEL, |
| 642 | pcpu_chunk_pagep(chunk, cpu, page_start)); |
| 643 | if (err < 0) |
| 644 | return err; |
| 645 | } |
| 646 | |
| 647 | /* flush at once, please read comments in pcpu_unmap() */ |
| 648 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 649 | pcpu_chunk_addr(chunk, last, page_end)); |
| 650 | return 0; |
| 651 | } |
| 652 | |
| 653 | /** |
| 654 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk |
| 655 | * @chunk: chunk of interest |
| 656 | * @off: offset to the area to populate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 657 | * @size: size of the area to populate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 658 | * |
| 659 | * For each cpu, populate and map pages [@page_start,@page_end) into |
| 660 | * @chunk. The area is cleared on return. |
| 661 | */ |
| 662 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) |
| 663 | { |
| 664 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; |
| 665 | int page_start = PFN_DOWN(off); |
| 666 | int page_end = PFN_UP(off + size); |
| 667 | int map_start = -1; |
Tejun Heo | 02d51fdf | 2009-03-01 15:42:36 +0900 | [diff] [blame] | 668 | int uninitialized_var(map_end); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 669 | unsigned int cpu; |
| 670 | int i; |
| 671 | |
| 672 | for (i = page_start; i < page_end; i++) { |
| 673 | if (pcpu_chunk_page_occupied(chunk, i)) { |
| 674 | if (map_start >= 0) { |
| 675 | if (pcpu_map(chunk, map_start, map_end)) |
| 676 | goto err; |
| 677 | map_start = -1; |
| 678 | } |
| 679 | continue; |
| 680 | } |
| 681 | |
| 682 | map_start = map_start < 0 ? i : map_start; |
| 683 | map_end = i + 1; |
| 684 | |
| 685 | for_each_possible_cpu(cpu) { |
| 686 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 687 | |
| 688 | *pagep = alloc_pages_node(cpu_to_node(cpu), |
| 689 | alloc_mask, 0); |
| 690 | if (!*pagep) |
| 691 | goto err; |
| 692 | } |
| 693 | } |
| 694 | |
| 695 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) |
| 696 | goto err; |
| 697 | |
| 698 | for_each_possible_cpu(cpu) |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 699 | memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 700 | size); |
| 701 | |
| 702 | return 0; |
| 703 | err: |
| 704 | /* likely under heavy memory pressure, give memory back */ |
| 705 | pcpu_depopulate_chunk(chunk, off, size, true); |
| 706 | return -ENOMEM; |
| 707 | } |
| 708 | |
| 709 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) |
| 710 | { |
| 711 | if (!chunk) |
| 712 | return; |
| 713 | if (chunk->vm) |
| 714 | free_vm_area(chunk->vm); |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 715 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 716 | kfree(chunk); |
| 717 | } |
| 718 | |
| 719 | static struct pcpu_chunk *alloc_pcpu_chunk(void) |
| 720 | { |
| 721 | struct pcpu_chunk *chunk; |
| 722 | |
| 723 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); |
| 724 | if (!chunk) |
| 725 | return NULL; |
| 726 | |
Tejun Heo | 1880d93 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 727 | 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] | 728 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
| 729 | chunk->map[chunk->map_used++] = pcpu_unit_size; |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 730 | chunk->page = chunk->page_ar; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 731 | |
| 732 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); |
| 733 | if (!chunk->vm) { |
| 734 | free_pcpu_chunk(chunk); |
| 735 | return NULL; |
| 736 | } |
| 737 | |
| 738 | INIT_LIST_HEAD(&chunk->list); |
| 739 | chunk->free_size = pcpu_unit_size; |
| 740 | chunk->contig_hint = pcpu_unit_size; |
| 741 | |
| 742 | return chunk; |
| 743 | } |
| 744 | |
| 745 | /** |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 746 | * pcpu_alloc - the percpu allocator |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 747 | * @size: size of area to allocate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 748 | * @align: alignment of area (max PAGE_SIZE) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 749 | * @reserved: allocate from the reserved chunk if available |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 750 | * |
| 751 | * Allocate percpu area of @size bytes aligned at @align. Might |
| 752 | * sleep. Might trigger writeouts. |
| 753 | * |
| 754 | * RETURNS: |
| 755 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 756 | */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 757 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 758 | { |
| 759 | void *ptr = NULL; |
| 760 | struct pcpu_chunk *chunk; |
| 761 | int slot, off; |
| 762 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 763 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 764 | WARN(true, "illegal size (%zu) or align (%zu) for " |
| 765 | "percpu allocation\n", size, align); |
| 766 | return NULL; |
| 767 | } |
| 768 | |
| 769 | mutex_lock(&pcpu_mutex); |
| 770 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 771 | /* serve reserved allocations from the reserved chunk if available */ |
| 772 | if (reserved && pcpu_reserved_chunk) { |
| 773 | chunk = pcpu_reserved_chunk; |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 774 | if (size > chunk->contig_hint || |
| 775 | pcpu_extend_area_map(chunk) < 0) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 776 | goto out_unlock; |
| 777 | off = pcpu_alloc_area(chunk, size, align); |
| 778 | if (off >= 0) |
| 779 | goto area_found; |
| 780 | goto out_unlock; |
| 781 | } |
| 782 | |
| 783 | /* search through normal chunks */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 784 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
| 785 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 786 | if (size > chunk->contig_hint) |
| 787 | continue; |
Tejun Heo | 9f7dcf2 | 2009-03-07 00:44:09 +0900 | [diff] [blame] | 788 | if (pcpu_extend_area_map(chunk) < 0) |
| 789 | goto out_unlock; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 790 | off = pcpu_alloc_area(chunk, size, align); |
| 791 | if (off >= 0) |
| 792 | goto area_found; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 793 | } |
| 794 | } |
| 795 | |
| 796 | /* hmmm... no space left, create a new chunk */ |
| 797 | chunk = alloc_pcpu_chunk(); |
| 798 | if (!chunk) |
| 799 | goto out_unlock; |
| 800 | pcpu_chunk_relocate(chunk, -1); |
| 801 | pcpu_chunk_addr_insert(chunk); |
| 802 | |
| 803 | off = pcpu_alloc_area(chunk, size, align); |
| 804 | if (off < 0) |
| 805 | goto out_unlock; |
| 806 | |
| 807 | area_found: |
| 808 | /* populate, map and clear the area */ |
| 809 | if (pcpu_populate_chunk(chunk, off, size)) { |
| 810 | pcpu_free_area(chunk, off); |
| 811 | goto out_unlock; |
| 812 | } |
| 813 | |
| 814 | ptr = __addr_to_pcpu_ptr(chunk->vm->addr + off); |
| 815 | out_unlock: |
| 816 | mutex_unlock(&pcpu_mutex); |
| 817 | return ptr; |
| 818 | } |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 819 | |
| 820 | /** |
| 821 | * __alloc_percpu - allocate dynamic percpu area |
| 822 | * @size: size of area to allocate in bytes |
| 823 | * @align: alignment of area (max PAGE_SIZE) |
| 824 | * |
| 825 | * Allocate percpu area of @size bytes aligned at @align. Might |
| 826 | * sleep. Might trigger writeouts. |
| 827 | * |
| 828 | * RETURNS: |
| 829 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 830 | */ |
| 831 | void *__alloc_percpu(size_t size, size_t align) |
| 832 | { |
| 833 | return pcpu_alloc(size, align, false); |
| 834 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 835 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
| 836 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 837 | /** |
| 838 | * __alloc_reserved_percpu - allocate reserved percpu area |
| 839 | * @size: size of area to allocate in bytes |
| 840 | * @align: alignment of area (max PAGE_SIZE) |
| 841 | * |
| 842 | * Allocate percpu area of @size bytes aligned at @align from reserved |
| 843 | * percpu area if arch has set it up; otherwise, allocation is served |
| 844 | * from the same dynamic area. Might sleep. Might trigger writeouts. |
| 845 | * |
| 846 | * RETURNS: |
| 847 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 848 | */ |
| 849 | void *__alloc_reserved_percpu(size_t size, size_t align) |
| 850 | { |
| 851 | return pcpu_alloc(size, align, true); |
| 852 | } |
| 853 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 854 | /** |
| 855 | * pcpu_reclaim - reclaim fully free chunks, workqueue function |
| 856 | * @work: unused |
| 857 | * |
| 858 | * Reclaim all fully free chunks except for the first one. |
| 859 | */ |
| 860 | static void pcpu_reclaim(struct work_struct *work) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 861 | { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 862 | LIST_HEAD(todo); |
| 863 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; |
| 864 | struct pcpu_chunk *chunk, *next; |
| 865 | |
| 866 | mutex_lock(&pcpu_mutex); |
| 867 | |
| 868 | list_for_each_entry_safe(chunk, next, head, list) { |
| 869 | WARN_ON(chunk->immutable); |
| 870 | |
| 871 | /* spare the first one */ |
| 872 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) |
| 873 | continue; |
| 874 | |
| 875 | rb_erase(&chunk->rb_node, &pcpu_addr_root); |
| 876 | list_move(&chunk->list, &todo); |
| 877 | } |
| 878 | |
| 879 | mutex_unlock(&pcpu_mutex); |
| 880 | |
| 881 | list_for_each_entry_safe(chunk, next, &todo, list) { |
| 882 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); |
| 883 | free_pcpu_chunk(chunk); |
| 884 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 885 | } |
| 886 | |
| 887 | /** |
| 888 | * free_percpu - free percpu area |
| 889 | * @ptr: pointer to area to free |
| 890 | * |
| 891 | * Free percpu area @ptr. Might sleep. |
| 892 | */ |
| 893 | void free_percpu(void *ptr) |
| 894 | { |
| 895 | void *addr = __pcpu_ptr_to_addr(ptr); |
| 896 | struct pcpu_chunk *chunk; |
| 897 | int off; |
| 898 | |
| 899 | if (!ptr) |
| 900 | return; |
| 901 | |
| 902 | mutex_lock(&pcpu_mutex); |
| 903 | |
| 904 | chunk = pcpu_chunk_addr_search(addr); |
| 905 | off = addr - chunk->vm->addr; |
| 906 | |
| 907 | pcpu_free_area(chunk, off); |
| 908 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 909 | /* 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] | 910 | if (chunk->free_size == pcpu_unit_size) { |
| 911 | struct pcpu_chunk *pos; |
| 912 | |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 913 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 914 | if (pos != chunk) { |
Tejun Heo | a56dbdd | 2009-03-07 00:44:11 +0900 | [diff] [blame^] | 915 | schedule_work(&pcpu_reclaim_work); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 916 | break; |
| 917 | } |
| 918 | } |
| 919 | |
| 920 | mutex_unlock(&pcpu_mutex); |
| 921 | } |
| 922 | EXPORT_SYMBOL_GPL(free_percpu); |
| 923 | |
| 924 | /** |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 925 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
| 926 | * @get_page_fn: callback to fetch page pointer |
| 927 | * @static_size: the size of static percpu area in bytes |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 928 | * @reserved_size: the size of reserved percpu area in bytes |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 929 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto |
| 930 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 931 | * @base_addr: mapped address, NULL for auto |
| 932 | * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 933 | * |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 934 | * Initialize the first percpu chunk which contains the kernel static |
| 935 | * perpcu area. This function is to be called from arch percpu area |
| 936 | * setup path. The first two parameters are mandatory. The rest are |
| 937 | * optional. |
| 938 | * |
| 939 | * @get_page_fn() should return pointer to percpu page given cpu |
| 940 | * number and page number. It should at least return enough pages to |
| 941 | * cover the static area. The returned pages for static area should |
| 942 | * have been initialized with valid data. If @unit_size is specified, |
| 943 | * it can also return pages after the static area. NULL return |
| 944 | * indicates end of pages for the cpu. Note that @get_page_fn() must |
| 945 | * return the same number of pages for all cpus. |
| 946 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 947 | * @reserved_size, if non-zero, specifies the amount of bytes to |
| 948 | * reserve after the static area in the first chunk. This reserves |
| 949 | * the first chunk such that it's available only through reserved |
| 950 | * percpu allocation. This is primarily used to serve module percpu |
| 951 | * static areas on architectures where the addressing model has |
| 952 | * limited offset range for symbol relocations to guarantee module |
| 953 | * percpu symbols fall inside the relocatable range. |
| 954 | * |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 955 | * @unit_size, if non-negative, specifies unit size and must be |
| 956 | * aligned to PAGE_SIZE and equal to or larger than @static_size + |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 957 | * @reserved_size + @dyn_size. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 958 | * |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 959 | * @dyn_size, if non-negative, limits the number of bytes available |
| 960 | * for dynamic allocation in the first chunk. Specifying non-negative |
| 961 | * value make percpu leave alone the area beyond @static_size + |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 962 | * @reserved_size + @dyn_size. |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 963 | * |
| 964 | * Non-null @base_addr means that the caller already allocated virtual |
| 965 | * region for the first chunk and mapped it. percpu must not mess |
| 966 | * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL |
| 967 | * @populate_pte_fn doesn't make any sense. |
| 968 | * |
| 969 | * @populate_pte_fn is used to populate the pagetable. NULL means the |
| 970 | * caller already populated the pagetable. |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 971 | * |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 972 | * If the first chunk ends up with both reserved and dynamic areas, it |
| 973 | * is served by two chunks - one to serve the core static and reserved |
| 974 | * areas and the other for the dynamic area. They share the same vm |
| 975 | * and page map but uses different area allocation map to stay away |
| 976 | * from each other. The latter chunk is circulated in the chunk slots |
| 977 | * and available for dynamic allocation like any other chunks. |
| 978 | * |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 979 | * RETURNS: |
| 980 | * The determined pcpu_unit_size which can be used to initialize |
| 981 | * percpu access. |
| 982 | */ |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 983 | 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] | 984 | size_t static_size, size_t reserved_size, |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 985 | ssize_t unit_size, ssize_t dyn_size, |
| 986 | void *base_addr, |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 987 | pcpu_populate_pte_fn_t populate_pte_fn) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 988 | { |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 989 | static struct vm_struct first_vm; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 990 | static int smap[2], dmap[2]; |
| 991 | struct pcpu_chunk *schunk, *dchunk = NULL; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 992 | unsigned int cpu; |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 993 | int nr_pages; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 994 | int err, i; |
| 995 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 996 | /* santiy checks */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 997 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
| 998 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 999 | BUG_ON(!static_size); |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1000 | if (unit_size >= 0) { |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1001 | BUG_ON(unit_size < static_size + reserved_size + |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1002 | (dyn_size >= 0 ? dyn_size : 0)); |
| 1003 | BUG_ON(unit_size & ~PAGE_MASK); |
| 1004 | } else { |
| 1005 | BUG_ON(dyn_size >= 0); |
| 1006 | BUG_ON(base_addr); |
| 1007 | } |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1008 | BUG_ON(base_addr && populate_pte_fn); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1009 | |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1010 | if (unit_size >= 0) |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1011 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
| 1012 | else |
| 1013 | pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT, |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1014 | PFN_UP(static_size + reserved_size)); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1015 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1016 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1017 | pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1018 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
Tejun Heo | cb83b42 | 2009-02-24 11:57:20 +0900 | [diff] [blame] | 1019 | + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1020 | |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1021 | if (dyn_size < 0) |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1022 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
Tejun Heo | cafe881 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1023 | |
Tejun Heo | d9b55ee | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1024 | /* |
| 1025 | * Allocate chunk slots. The additional last slot is for |
| 1026 | * empty chunks. |
| 1027 | */ |
| 1028 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1029 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
| 1030 | for (i = 0; i < pcpu_nr_slots; i++) |
| 1031 | INIT_LIST_HEAD(&pcpu_slot[i]); |
| 1032 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1033 | /* |
| 1034 | * Initialize static chunk. If reserved_size is zero, the |
| 1035 | * static chunk covers static area + dynamic allocation area |
| 1036 | * in the first chunk. If reserved_size is not zero, it |
| 1037 | * covers static area + reserved area (mostly used for module |
| 1038 | * static percpu allocation). |
| 1039 | */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1040 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
| 1041 | INIT_LIST_HEAD(&schunk->list); |
| 1042 | schunk->vm = &first_vm; |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1043 | schunk->map = smap; |
| 1044 | schunk->map_alloc = ARRAY_SIZE(smap); |
Tejun Heo | 3e24aa5 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1045 | schunk->page = schunk->page_ar; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1046 | |
| 1047 | if (reserved_size) { |
| 1048 | schunk->free_size = reserved_size; |
| 1049 | pcpu_reserved_chunk = schunk; /* not for dynamic alloc */ |
| 1050 | } else { |
| 1051 | schunk->free_size = dyn_size; |
| 1052 | dyn_size = 0; /* dynamic area covered */ |
| 1053 | } |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1054 | schunk->contig_hint = schunk->free_size; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1055 | |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1056 | schunk->map[schunk->map_used++] = -static_size; |
| 1057 | if (schunk->free_size) |
| 1058 | schunk->map[schunk->map_used++] = schunk->free_size; |
| 1059 | |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1060 | pcpu_reserved_chunk_limit = static_size + schunk->free_size; |
| 1061 | |
| 1062 | /* init dynamic chunk if necessary */ |
| 1063 | if (dyn_size) { |
| 1064 | dchunk = alloc_bootmem(sizeof(struct pcpu_chunk)); |
| 1065 | INIT_LIST_HEAD(&dchunk->list); |
| 1066 | dchunk->vm = &first_vm; |
| 1067 | dchunk->map = dmap; |
| 1068 | dchunk->map_alloc = ARRAY_SIZE(dmap); |
| 1069 | dchunk->page = schunk->page_ar; /* share page map with schunk */ |
| 1070 | |
| 1071 | dchunk->contig_hint = dchunk->free_size = dyn_size; |
| 1072 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; |
| 1073 | dchunk->map[dchunk->map_used++] = dchunk->free_size; |
| 1074 | } |
| 1075 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1076 | /* allocate vm address */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1077 | first_vm.flags = VM_ALLOC; |
| 1078 | first_vm.size = pcpu_chunk_size; |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1079 | |
| 1080 | if (!base_addr) |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1081 | vm_area_register_early(&first_vm, PAGE_SIZE); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1082 | else { |
| 1083 | /* |
| 1084 | * Pages already mapped. No need to remap into |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1085 | * vmalloc area. In this case the first chunks can't |
| 1086 | * be mapped or unmapped by percpu and are marked |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1087 | * immutable. |
| 1088 | */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1089 | first_vm.addr = base_addr; |
| 1090 | schunk->immutable = true; |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1091 | if (dchunk) |
| 1092 | dchunk->immutable = true; |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1093 | } |
| 1094 | |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1095 | /* assign pages */ |
| 1096 | nr_pages = -1; |
| 1097 | for_each_possible_cpu(cpu) { |
| 1098 | for (i = 0; i < pcpu_unit_pages; i++) { |
| 1099 | struct page *page = get_page_fn(cpu, i); |
| 1100 | |
| 1101 | if (!page) |
| 1102 | break; |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1103 | *pcpu_chunk_pagep(schunk, cpu, i) = page; |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1104 | } |
| 1105 | |
Tejun Heo | 61ace7f | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1106 | BUG_ON(i < PFN_UP(static_size)); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1107 | |
| 1108 | if (nr_pages < 0) |
| 1109 | nr_pages = i; |
| 1110 | else |
| 1111 | BUG_ON(nr_pages != i); |
| 1112 | } |
| 1113 | |
| 1114 | /* map them */ |
| 1115 | if (populate_pte_fn) { |
| 1116 | for_each_possible_cpu(cpu) |
| 1117 | for (i = 0; i < nr_pages; i++) |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1118 | populate_pte_fn(pcpu_chunk_addr(schunk, |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1119 | cpu, i)); |
| 1120 | |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1121 | err = pcpu_map(schunk, 0, nr_pages); |
Tejun Heo | 8d408b4 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 1122 | if (err) |
| 1123 | panic("failed to setup static percpu area, err=%d\n", |
| 1124 | err); |
| 1125 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1126 | |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1127 | /* link the first chunk in */ |
Tejun Heo | edcb463 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1128 | if (!dchunk) { |
| 1129 | pcpu_chunk_relocate(schunk, -1); |
| 1130 | pcpu_chunk_addr_insert(schunk); |
| 1131 | } else { |
| 1132 | pcpu_chunk_relocate(dchunk, -1); |
| 1133 | pcpu_chunk_addr_insert(dchunk); |
| 1134 | } |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1135 | |
| 1136 | /* we're done */ |
Tejun Heo | 2441d15 | 2009-03-06 14:33:59 +0900 | [diff] [blame] | 1137 | pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 1138 | return pcpu_unit_size; |
| 1139 | } |