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 | * |
| 51 | * - use pcpu_setup_static() during percpu area initialization to |
| 52 | * setup kernel static percpu area |
| 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> |
| 66 | |
| 67 | #include <asm/cacheflush.h> |
| 68 | #include <asm/tlbflush.h> |
| 69 | |
| 70 | #define PCPU_MIN_UNIT_PAGES_SHIFT 4 /* also max alloc size */ |
| 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 */ |
| 83 | struct page *page[]; /* #cpus * UNIT_PAGES */ |
| 84 | }; |
| 85 | |
| 86 | static int pcpu_unit_pages_shift; |
| 87 | static int pcpu_unit_pages; |
| 88 | static int pcpu_unit_shift; |
| 89 | static int pcpu_unit_size; |
| 90 | static int pcpu_chunk_size; |
| 91 | static int pcpu_nr_slots; |
| 92 | static size_t pcpu_chunk_struct_size; |
| 93 | |
| 94 | /* the address of the first chunk which starts with the kernel static area */ |
| 95 | void *pcpu_base_addr; |
| 96 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
| 97 | |
| 98 | /* the size of kernel static area */ |
| 99 | static int pcpu_static_size; |
| 100 | |
| 101 | /* |
| 102 | * One mutex to rule them all. |
| 103 | * |
| 104 | * The following mutex is grabbed in the outermost public alloc/free |
| 105 | * interface functions and released only when the operation is |
| 106 | * complete. As such, every function in this file other than the |
| 107 | * outermost functions are called under pcpu_mutex. |
| 108 | * |
| 109 | * It can easily be switched to use spinlock such that only the area |
| 110 | * allocation and page population commit are protected with it doing |
| 111 | * actual [de]allocation without holding any lock. However, given |
| 112 | * what this allocator does, I think it's better to let them run |
| 113 | * sequentially. |
| 114 | */ |
| 115 | static DEFINE_MUTEX(pcpu_mutex); |
| 116 | |
| 117 | static struct list_head *pcpu_slot; /* chunk list slots */ |
| 118 | static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */ |
| 119 | |
| 120 | static int pcpu_size_to_slot(int size) |
| 121 | { |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 122 | int highbit = fls(size); /* size is in bytes */ |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 123 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
| 124 | } |
| 125 | |
| 126 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
| 127 | { |
| 128 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) |
| 129 | return 0; |
| 130 | |
| 131 | return pcpu_size_to_slot(chunk->free_size); |
| 132 | } |
| 133 | |
| 134 | static int pcpu_page_idx(unsigned int cpu, int page_idx) |
| 135 | { |
| 136 | return (cpu << pcpu_unit_pages_shift) + page_idx; |
| 137 | } |
| 138 | |
| 139 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, |
| 140 | unsigned int cpu, int page_idx) |
| 141 | { |
| 142 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; |
| 143 | } |
| 144 | |
| 145 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
| 146 | unsigned int cpu, int page_idx) |
| 147 | { |
| 148 | return (unsigned long)chunk->vm->addr + |
| 149 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); |
| 150 | } |
| 151 | |
| 152 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, |
| 153 | int page_idx) |
| 154 | { |
| 155 | return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL; |
| 156 | } |
| 157 | |
| 158 | /** |
| 159 | * pcpu_realloc - versatile realloc |
| 160 | * @p: the current pointer (can be NULL for new allocations) |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 161 | * @size: the current size in bytes (can be 0 for new allocations) |
| 162 | * @new_size: the wanted new size in bytes (can be 0 for free) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 163 | * |
| 164 | * More robust realloc which can be used to allocate, resize or free a |
| 165 | * memory area of arbitrary size. If the needed size goes over |
| 166 | * PAGE_SIZE, kernel VM is used. |
| 167 | * |
| 168 | * RETURNS: |
| 169 | * The new pointer on success, NULL on failure. |
| 170 | */ |
| 171 | static void *pcpu_realloc(void *p, size_t size, size_t new_size) |
| 172 | { |
| 173 | void *new; |
| 174 | |
| 175 | if (new_size <= PAGE_SIZE) |
| 176 | new = kmalloc(new_size, GFP_KERNEL); |
| 177 | else |
| 178 | new = vmalloc(new_size); |
| 179 | if (new_size && !new) |
| 180 | return NULL; |
| 181 | |
| 182 | memcpy(new, p, min(size, new_size)); |
| 183 | if (new_size > size) |
| 184 | memset(new + size, 0, new_size - size); |
| 185 | |
| 186 | if (size <= PAGE_SIZE) |
| 187 | kfree(p); |
| 188 | else |
| 189 | vfree(p); |
| 190 | |
| 191 | return new; |
| 192 | } |
| 193 | |
| 194 | /** |
| 195 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot |
| 196 | * @chunk: chunk of interest |
| 197 | * @oslot: the previous slot it was on |
| 198 | * |
| 199 | * This function is called after an allocation or free changed @chunk. |
| 200 | * New slot according to the changed state is determined and @chunk is |
| 201 | * moved to the slot. |
| 202 | */ |
| 203 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) |
| 204 | { |
| 205 | int nslot = pcpu_chunk_slot(chunk); |
| 206 | |
| 207 | if (oslot != nslot) { |
| 208 | if (oslot < nslot) |
| 209 | list_move(&chunk->list, &pcpu_slot[nslot]); |
| 210 | else |
| 211 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | static struct rb_node **pcpu_chunk_rb_search(void *addr, |
| 216 | struct rb_node **parentp) |
| 217 | { |
| 218 | struct rb_node **p = &pcpu_addr_root.rb_node; |
| 219 | struct rb_node *parent = NULL; |
| 220 | struct pcpu_chunk *chunk; |
| 221 | |
| 222 | while (*p) { |
| 223 | parent = *p; |
| 224 | chunk = rb_entry(parent, struct pcpu_chunk, rb_node); |
| 225 | |
| 226 | if (addr < chunk->vm->addr) |
| 227 | p = &(*p)->rb_left; |
| 228 | else if (addr > chunk->vm->addr) |
| 229 | p = &(*p)->rb_right; |
| 230 | else |
| 231 | break; |
| 232 | } |
| 233 | |
| 234 | if (parentp) |
| 235 | *parentp = parent; |
| 236 | return p; |
| 237 | } |
| 238 | |
| 239 | /** |
| 240 | * pcpu_chunk_addr_search - search for chunk containing specified address |
| 241 | * @addr: address to search for |
| 242 | * |
| 243 | * Look for chunk which might contain @addr. More specifically, it |
| 244 | * searchs for the chunk with the highest start address which isn't |
| 245 | * beyond @addr. |
| 246 | * |
| 247 | * RETURNS: |
| 248 | * The address of the found chunk. |
| 249 | */ |
| 250 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) |
| 251 | { |
| 252 | struct rb_node *n, *parent; |
| 253 | struct pcpu_chunk *chunk; |
| 254 | |
| 255 | n = *pcpu_chunk_rb_search(addr, &parent); |
| 256 | if (!n) { |
| 257 | /* no exactly matching chunk, the parent is the closest */ |
| 258 | n = parent; |
| 259 | BUG_ON(!n); |
| 260 | } |
| 261 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); |
| 262 | |
| 263 | if (addr < chunk->vm->addr) { |
| 264 | /* the parent was the next one, look for the previous one */ |
| 265 | n = rb_prev(n); |
| 266 | BUG_ON(!n); |
| 267 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); |
| 268 | } |
| 269 | |
| 270 | return chunk; |
| 271 | } |
| 272 | |
| 273 | /** |
| 274 | * pcpu_chunk_addr_insert - insert chunk into address rb tree |
| 275 | * @new: chunk to insert |
| 276 | * |
| 277 | * Insert @new into address rb tree. |
| 278 | */ |
| 279 | static void pcpu_chunk_addr_insert(struct pcpu_chunk *new) |
| 280 | { |
| 281 | struct rb_node **p, *parent; |
| 282 | |
| 283 | p = pcpu_chunk_rb_search(new->vm->addr, &parent); |
| 284 | BUG_ON(*p); |
| 285 | rb_link_node(&new->rb_node, parent, p); |
| 286 | rb_insert_color(&new->rb_node, &pcpu_addr_root); |
| 287 | } |
| 288 | |
| 289 | /** |
| 290 | * pcpu_split_block - split a map block |
| 291 | * @chunk: chunk of interest |
| 292 | * @i: index of map block to split |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 293 | * @head: head size in bytes (can be 0) |
| 294 | * @tail: tail size in bytes (can be 0) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 295 | * |
| 296 | * Split the @i'th map block into two or three blocks. If @head is |
| 297 | * non-zero, @head bytes block is inserted before block @i moving it |
| 298 | * to @i+1 and reducing its size by @head bytes. |
| 299 | * |
| 300 | * If @tail is non-zero, the target block, which can be @i or @i+1 |
| 301 | * depending on @head, is reduced by @tail bytes and @tail byte block |
| 302 | * is inserted after the target block. |
| 303 | * |
| 304 | * RETURNS: |
| 305 | * 0 on success, -errno on failure. |
| 306 | */ |
| 307 | static int pcpu_split_block(struct pcpu_chunk *chunk, int i, int head, int tail) |
| 308 | { |
| 309 | int nr_extra = !!head + !!tail; |
| 310 | int target = chunk->map_used + nr_extra; |
| 311 | |
| 312 | /* reallocation required? */ |
| 313 | if (chunk->map_alloc < target) { |
| 314 | int new_alloc = chunk->map_alloc; |
| 315 | int *new; |
| 316 | |
| 317 | while (new_alloc < target) |
| 318 | new_alloc *= 2; |
| 319 | |
| 320 | new = pcpu_realloc(chunk->map, |
| 321 | chunk->map_alloc * sizeof(new[0]), |
| 322 | new_alloc * sizeof(new[0])); |
| 323 | if (!new) |
| 324 | return -ENOMEM; |
| 325 | |
| 326 | chunk->map_alloc = new_alloc; |
| 327 | chunk->map = new; |
| 328 | } |
| 329 | |
| 330 | /* insert a new subblock */ |
| 331 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
| 332 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
| 333 | chunk->map_used += nr_extra; |
| 334 | |
| 335 | if (head) { |
| 336 | chunk->map[i + 1] = chunk->map[i] - head; |
| 337 | chunk->map[i++] = head; |
| 338 | } |
| 339 | if (tail) { |
| 340 | chunk->map[i++] -= tail; |
| 341 | chunk->map[i] = tail; |
| 342 | } |
| 343 | return 0; |
| 344 | } |
| 345 | |
| 346 | /** |
| 347 | * pcpu_alloc_area - allocate area from a pcpu_chunk |
| 348 | * @chunk: chunk of interest |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 349 | * @size: wanted size in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 350 | * @align: wanted align |
| 351 | * |
| 352 | * Try to allocate @size bytes area aligned at @align from @chunk. |
| 353 | * Note that this function only allocates the offset. It doesn't |
| 354 | * populate or map the area. |
| 355 | * |
| 356 | * RETURNS: |
| 357 | * Allocated offset in @chunk on success, -errno on failure. |
| 358 | */ |
| 359 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) |
| 360 | { |
| 361 | int oslot = pcpu_chunk_slot(chunk); |
| 362 | int max_contig = 0; |
| 363 | int i, off; |
| 364 | |
| 365 | /* |
| 366 | * The static chunk initially doesn't have map attached |
| 367 | * because kmalloc wasn't available during init. Give it one. |
| 368 | */ |
| 369 | if (unlikely(!chunk->map)) { |
| 370 | chunk->map = pcpu_realloc(NULL, 0, |
| 371 | PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
| 372 | if (!chunk->map) |
| 373 | return -ENOMEM; |
| 374 | |
| 375 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
| 376 | chunk->map[chunk->map_used++] = -pcpu_static_size; |
| 377 | if (chunk->free_size) |
| 378 | chunk->map[chunk->map_used++] = chunk->free_size; |
| 379 | } |
| 380 | |
| 381 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
| 382 | bool is_last = i + 1 == chunk->map_used; |
| 383 | int head, tail; |
| 384 | |
| 385 | /* extra for alignment requirement */ |
| 386 | head = ALIGN(off, align) - off; |
| 387 | BUG_ON(i == 0 && head != 0); |
| 388 | |
| 389 | if (chunk->map[i] < 0) |
| 390 | continue; |
| 391 | if (chunk->map[i] < head + size) { |
| 392 | max_contig = max(chunk->map[i], max_contig); |
| 393 | continue; |
| 394 | } |
| 395 | |
| 396 | /* |
| 397 | * If head is small or the previous block is free, |
| 398 | * merge'em. Note that 'small' is defined as smaller |
| 399 | * than sizeof(int), which is very small but isn't too |
| 400 | * uncommon for percpu allocations. |
| 401 | */ |
| 402 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { |
| 403 | if (chunk->map[i - 1] > 0) |
| 404 | chunk->map[i - 1] += head; |
| 405 | else { |
| 406 | chunk->map[i - 1] -= head; |
| 407 | chunk->free_size -= head; |
| 408 | } |
| 409 | chunk->map[i] -= head; |
| 410 | off += head; |
| 411 | head = 0; |
| 412 | } |
| 413 | |
| 414 | /* if tail is small, just keep it around */ |
| 415 | tail = chunk->map[i] - head - size; |
| 416 | if (tail < sizeof(int)) |
| 417 | tail = 0; |
| 418 | |
| 419 | /* split if warranted */ |
| 420 | if (head || tail) { |
| 421 | if (pcpu_split_block(chunk, i, head, tail)) |
| 422 | return -ENOMEM; |
| 423 | if (head) { |
| 424 | i++; |
| 425 | off += head; |
| 426 | max_contig = max(chunk->map[i - 1], max_contig); |
| 427 | } |
| 428 | if (tail) |
| 429 | max_contig = max(chunk->map[i + 1], max_contig); |
| 430 | } |
| 431 | |
| 432 | /* update hint and mark allocated */ |
| 433 | if (is_last) |
| 434 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 435 | else |
| 436 | chunk->contig_hint = max(chunk->contig_hint, |
| 437 | max_contig); |
| 438 | |
| 439 | chunk->free_size -= chunk->map[i]; |
| 440 | chunk->map[i] = -chunk->map[i]; |
| 441 | |
| 442 | pcpu_chunk_relocate(chunk, oslot); |
| 443 | return off; |
| 444 | } |
| 445 | |
| 446 | chunk->contig_hint = max_contig; /* fully scanned */ |
| 447 | pcpu_chunk_relocate(chunk, oslot); |
| 448 | |
| 449 | /* |
| 450 | * Tell the upper layer that this chunk has no area left. |
| 451 | * Note that this is not an error condition but a notification |
| 452 | * to upper layer that it needs to look at other chunks. |
| 453 | * -ENOSPC is chosen as it isn't used in memory subsystem and |
| 454 | * matches the meaning in a way. |
| 455 | */ |
| 456 | return -ENOSPC; |
| 457 | } |
| 458 | |
| 459 | /** |
| 460 | * pcpu_free_area - free area to a pcpu_chunk |
| 461 | * @chunk: chunk of interest |
| 462 | * @freeme: offset of area to free |
| 463 | * |
| 464 | * Free area starting from @freeme to @chunk. Note that this function |
| 465 | * only modifies the allocation map. It doesn't depopulate or unmap |
| 466 | * the area. |
| 467 | */ |
| 468 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) |
| 469 | { |
| 470 | int oslot = pcpu_chunk_slot(chunk); |
| 471 | int i, off; |
| 472 | |
| 473 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) |
| 474 | if (off == freeme) |
| 475 | break; |
| 476 | BUG_ON(off != freeme); |
| 477 | BUG_ON(chunk->map[i] > 0); |
| 478 | |
| 479 | chunk->map[i] = -chunk->map[i]; |
| 480 | chunk->free_size += chunk->map[i]; |
| 481 | |
| 482 | /* merge with previous? */ |
| 483 | if (i > 0 && chunk->map[i - 1] >= 0) { |
| 484 | chunk->map[i - 1] += chunk->map[i]; |
| 485 | chunk->map_used--; |
| 486 | memmove(&chunk->map[i], &chunk->map[i + 1], |
| 487 | (chunk->map_used - i) * sizeof(chunk->map[0])); |
| 488 | i--; |
| 489 | } |
| 490 | /* merge with next? */ |
| 491 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { |
| 492 | chunk->map[i] += chunk->map[i + 1]; |
| 493 | chunk->map_used--; |
| 494 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], |
| 495 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); |
| 496 | } |
| 497 | |
| 498 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); |
| 499 | pcpu_chunk_relocate(chunk, oslot); |
| 500 | } |
| 501 | |
| 502 | /** |
| 503 | * pcpu_unmap - unmap pages out of a pcpu_chunk |
| 504 | * @chunk: chunk of interest |
| 505 | * @page_start: page index of the first page to unmap |
| 506 | * @page_end: page index of the last page to unmap + 1 |
| 507 | * @flush: whether to flush cache and tlb or not |
| 508 | * |
| 509 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. |
| 510 | * If @flush is true, vcache is flushed before unmapping and tlb |
| 511 | * after. |
| 512 | */ |
| 513 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, |
| 514 | bool flush) |
| 515 | { |
| 516 | unsigned int last = num_possible_cpus() - 1; |
| 517 | unsigned int cpu; |
| 518 | |
| 519 | /* |
| 520 | * Each flushing trial can be very expensive, issue flush on |
| 521 | * the whole region at once rather than doing it for each cpu. |
| 522 | * This could be an overkill but is more scalable. |
| 523 | */ |
| 524 | if (flush) |
| 525 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 526 | pcpu_chunk_addr(chunk, last, page_end)); |
| 527 | |
| 528 | for_each_possible_cpu(cpu) |
| 529 | unmap_kernel_range_noflush( |
| 530 | pcpu_chunk_addr(chunk, cpu, page_start), |
| 531 | (page_end - page_start) << PAGE_SHIFT); |
| 532 | |
| 533 | /* ditto as flush_cache_vunmap() */ |
| 534 | if (flush) |
| 535 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), |
| 536 | pcpu_chunk_addr(chunk, last, page_end)); |
| 537 | } |
| 538 | |
| 539 | /** |
| 540 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk |
| 541 | * @chunk: chunk to depopulate |
| 542 | * @off: offset to the area to depopulate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 543 | * @size: size of the area to depopulate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 544 | * @flush: whether to flush cache and tlb or not |
| 545 | * |
| 546 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) |
| 547 | * from @chunk. If @flush is true, vcache is flushed before unmapping |
| 548 | * and tlb after. |
| 549 | */ |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 550 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
| 551 | bool flush) |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 552 | { |
| 553 | int page_start = PFN_DOWN(off); |
| 554 | int page_end = PFN_UP(off + size); |
| 555 | int unmap_start = -1; |
| 556 | int uninitialized_var(unmap_end); |
| 557 | unsigned int cpu; |
| 558 | int i; |
| 559 | |
| 560 | for (i = page_start; i < page_end; i++) { |
| 561 | for_each_possible_cpu(cpu) { |
| 562 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 563 | |
| 564 | if (!*pagep) |
| 565 | continue; |
| 566 | |
| 567 | __free_page(*pagep); |
| 568 | |
| 569 | /* |
| 570 | * If it's partial depopulation, it might get |
| 571 | * populated or depopulated again. Mark the |
| 572 | * page gone. |
| 573 | */ |
| 574 | *pagep = NULL; |
| 575 | |
| 576 | unmap_start = unmap_start < 0 ? i : unmap_start; |
| 577 | unmap_end = i + 1; |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | if (unmap_start >= 0) |
| 582 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); |
| 583 | } |
| 584 | |
| 585 | /** |
| 586 | * pcpu_map - map pages into a pcpu_chunk |
| 587 | * @chunk: chunk of interest |
| 588 | * @page_start: page index of the first page to map |
| 589 | * @page_end: page index of the last page to map + 1 |
| 590 | * |
| 591 | * For each cpu, map pages [@page_start,@page_end) into @chunk. |
| 592 | * vcache is flushed afterwards. |
| 593 | */ |
| 594 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) |
| 595 | { |
| 596 | unsigned int last = num_possible_cpus() - 1; |
| 597 | unsigned int cpu; |
| 598 | int err; |
| 599 | |
| 600 | for_each_possible_cpu(cpu) { |
| 601 | err = map_kernel_range_noflush( |
| 602 | pcpu_chunk_addr(chunk, cpu, page_start), |
| 603 | (page_end - page_start) << PAGE_SHIFT, |
| 604 | PAGE_KERNEL, |
| 605 | pcpu_chunk_pagep(chunk, cpu, page_start)); |
| 606 | if (err < 0) |
| 607 | return err; |
| 608 | } |
| 609 | |
| 610 | /* flush at once, please read comments in pcpu_unmap() */ |
| 611 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), |
| 612 | pcpu_chunk_addr(chunk, last, page_end)); |
| 613 | return 0; |
| 614 | } |
| 615 | |
| 616 | /** |
| 617 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk |
| 618 | * @chunk: chunk of interest |
| 619 | * @off: offset to the area to populate |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 620 | * @size: size of the area to populate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 621 | * |
| 622 | * For each cpu, populate and map pages [@page_start,@page_end) into |
| 623 | * @chunk. The area is cleared on return. |
| 624 | */ |
| 625 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) |
| 626 | { |
| 627 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; |
| 628 | int page_start = PFN_DOWN(off); |
| 629 | int page_end = PFN_UP(off + size); |
| 630 | int map_start = -1; |
| 631 | int map_end; |
| 632 | unsigned int cpu; |
| 633 | int i; |
| 634 | |
| 635 | for (i = page_start; i < page_end; i++) { |
| 636 | if (pcpu_chunk_page_occupied(chunk, i)) { |
| 637 | if (map_start >= 0) { |
| 638 | if (pcpu_map(chunk, map_start, map_end)) |
| 639 | goto err; |
| 640 | map_start = -1; |
| 641 | } |
| 642 | continue; |
| 643 | } |
| 644 | |
| 645 | map_start = map_start < 0 ? i : map_start; |
| 646 | map_end = i + 1; |
| 647 | |
| 648 | for_each_possible_cpu(cpu) { |
| 649 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); |
| 650 | |
| 651 | *pagep = alloc_pages_node(cpu_to_node(cpu), |
| 652 | alloc_mask, 0); |
| 653 | if (!*pagep) |
| 654 | goto err; |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) |
| 659 | goto err; |
| 660 | |
| 661 | for_each_possible_cpu(cpu) |
| 662 | memset(chunk->vm->addr + (cpu << pcpu_unit_shift) + off, 0, |
| 663 | size); |
| 664 | |
| 665 | return 0; |
| 666 | err: |
| 667 | /* likely under heavy memory pressure, give memory back */ |
| 668 | pcpu_depopulate_chunk(chunk, off, size, true); |
| 669 | return -ENOMEM; |
| 670 | } |
| 671 | |
| 672 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) |
| 673 | { |
| 674 | if (!chunk) |
| 675 | return; |
| 676 | if (chunk->vm) |
| 677 | free_vm_area(chunk->vm); |
| 678 | pcpu_realloc(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]), 0); |
| 679 | kfree(chunk); |
| 680 | } |
| 681 | |
| 682 | static struct pcpu_chunk *alloc_pcpu_chunk(void) |
| 683 | { |
| 684 | struct pcpu_chunk *chunk; |
| 685 | |
| 686 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); |
| 687 | if (!chunk) |
| 688 | return NULL; |
| 689 | |
| 690 | chunk->map = pcpu_realloc(NULL, 0, |
| 691 | PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
| 692 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
| 693 | chunk->map[chunk->map_used++] = pcpu_unit_size; |
| 694 | |
| 695 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); |
| 696 | if (!chunk->vm) { |
| 697 | free_pcpu_chunk(chunk); |
| 698 | return NULL; |
| 699 | } |
| 700 | |
| 701 | INIT_LIST_HEAD(&chunk->list); |
| 702 | chunk->free_size = pcpu_unit_size; |
| 703 | chunk->contig_hint = pcpu_unit_size; |
| 704 | |
| 705 | return chunk; |
| 706 | } |
| 707 | |
| 708 | /** |
| 709 | * __alloc_percpu - allocate percpu area |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 710 | * @size: size of area to allocate in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 711 | * @align: alignment of area (max PAGE_SIZE) |
| 712 | * |
| 713 | * Allocate percpu area of @size bytes aligned at @align. Might |
| 714 | * sleep. Might trigger writeouts. |
| 715 | * |
| 716 | * RETURNS: |
| 717 | * Percpu pointer to the allocated area on success, NULL on failure. |
| 718 | */ |
| 719 | void *__alloc_percpu(size_t size, size_t align) |
| 720 | { |
| 721 | void *ptr = NULL; |
| 722 | struct pcpu_chunk *chunk; |
| 723 | int slot, off; |
| 724 | |
| 725 | if (unlikely(!size || size > PAGE_SIZE << PCPU_MIN_UNIT_PAGES_SHIFT || |
| 726 | align > PAGE_SIZE)) { |
| 727 | WARN(true, "illegal size (%zu) or align (%zu) for " |
| 728 | "percpu allocation\n", size, align); |
| 729 | return NULL; |
| 730 | } |
| 731 | |
| 732 | mutex_lock(&pcpu_mutex); |
| 733 | |
| 734 | /* allocate area */ |
| 735 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
| 736 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { |
| 737 | if (size > chunk->contig_hint) |
| 738 | continue; |
| 739 | off = pcpu_alloc_area(chunk, size, align); |
| 740 | if (off >= 0) |
| 741 | goto area_found; |
| 742 | if (off != -ENOSPC) |
| 743 | goto out_unlock; |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | /* hmmm... no space left, create a new chunk */ |
| 748 | chunk = alloc_pcpu_chunk(); |
| 749 | if (!chunk) |
| 750 | goto out_unlock; |
| 751 | pcpu_chunk_relocate(chunk, -1); |
| 752 | pcpu_chunk_addr_insert(chunk); |
| 753 | |
| 754 | off = pcpu_alloc_area(chunk, size, align); |
| 755 | if (off < 0) |
| 756 | goto out_unlock; |
| 757 | |
| 758 | area_found: |
| 759 | /* populate, map and clear the area */ |
| 760 | if (pcpu_populate_chunk(chunk, off, size)) { |
| 761 | pcpu_free_area(chunk, off); |
| 762 | goto out_unlock; |
| 763 | } |
| 764 | |
| 765 | ptr = __addr_to_pcpu_ptr(chunk->vm->addr + off); |
| 766 | out_unlock: |
| 767 | mutex_unlock(&pcpu_mutex); |
| 768 | return ptr; |
| 769 | } |
| 770 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
| 771 | |
| 772 | static void pcpu_kill_chunk(struct pcpu_chunk *chunk) |
| 773 | { |
| 774 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); |
| 775 | list_del(&chunk->list); |
| 776 | rb_erase(&chunk->rb_node, &pcpu_addr_root); |
| 777 | free_pcpu_chunk(chunk); |
| 778 | } |
| 779 | |
| 780 | /** |
| 781 | * free_percpu - free percpu area |
| 782 | * @ptr: pointer to area to free |
| 783 | * |
| 784 | * Free percpu area @ptr. Might sleep. |
| 785 | */ |
| 786 | void free_percpu(void *ptr) |
| 787 | { |
| 788 | void *addr = __pcpu_ptr_to_addr(ptr); |
| 789 | struct pcpu_chunk *chunk; |
| 790 | int off; |
| 791 | |
| 792 | if (!ptr) |
| 793 | return; |
| 794 | |
| 795 | mutex_lock(&pcpu_mutex); |
| 796 | |
| 797 | chunk = pcpu_chunk_addr_search(addr); |
| 798 | off = addr - chunk->vm->addr; |
| 799 | |
| 800 | pcpu_free_area(chunk, off); |
| 801 | |
| 802 | /* the chunk became fully free, kill one if there are other free ones */ |
| 803 | if (chunk->free_size == pcpu_unit_size) { |
| 804 | struct pcpu_chunk *pos; |
| 805 | |
| 806 | list_for_each_entry(pos, |
| 807 | &pcpu_slot[pcpu_chunk_slot(chunk)], list) |
| 808 | if (pos != chunk) { |
| 809 | pcpu_kill_chunk(pos); |
| 810 | break; |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | mutex_unlock(&pcpu_mutex); |
| 815 | } |
| 816 | EXPORT_SYMBOL_GPL(free_percpu); |
| 817 | |
| 818 | /** |
| 819 | * pcpu_setup_static - initialize kernel static percpu area |
| 820 | * @populate_pte_fn: callback to allocate pagetable |
| 821 | * @pages: num_possible_cpus() * PFN_UP(cpu_size) pages |
Tejun Heo | cae3aeb | 2009-02-21 16:56:23 +0900 | [diff] [blame] | 822 | * @cpu_size: the size of static percpu area in bytes |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 823 | * |
| 824 | * Initialize kernel static percpu area. The caller should allocate |
| 825 | * all the necessary pages and pass them in @pages. |
| 826 | * @populate_pte_fn() is called on each page to be used for percpu |
| 827 | * mapping and is responsible for making sure all the necessary page |
| 828 | * tables for the page is allocated. |
| 829 | * |
| 830 | * RETURNS: |
| 831 | * The determined pcpu_unit_size which can be used to initialize |
| 832 | * percpu access. |
| 833 | */ |
| 834 | size_t __init pcpu_setup_static(pcpu_populate_pte_fn_t populate_pte_fn, |
| 835 | struct page **pages, size_t cpu_size) |
| 836 | { |
| 837 | static struct vm_struct static_vm; |
| 838 | struct pcpu_chunk *static_chunk; |
| 839 | int nr_cpu_pages = DIV_ROUND_UP(cpu_size, PAGE_SIZE); |
| 840 | unsigned int cpu; |
| 841 | int err, i; |
| 842 | |
| 843 | pcpu_unit_pages_shift = max_t(int, PCPU_MIN_UNIT_PAGES_SHIFT, |
| 844 | order_base_2(cpu_size) - PAGE_SHIFT); |
| 845 | |
| 846 | pcpu_static_size = cpu_size; |
| 847 | pcpu_unit_pages = 1 << pcpu_unit_pages_shift; |
| 848 | pcpu_unit_shift = PAGE_SHIFT + pcpu_unit_pages_shift; |
| 849 | pcpu_unit_size = 1 << pcpu_unit_shift; |
| 850 | pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size; |
| 851 | pcpu_nr_slots = pcpu_size_to_slot(pcpu_unit_size) + 1; |
| 852 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
Tejun Heo | cb83b42 | 2009-02-24 11:57:20 +0900 | [diff] [blame] | 853 | + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 854 | |
| 855 | /* allocate chunk slots */ |
| 856 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
| 857 | for (i = 0; i < pcpu_nr_slots; i++) |
| 858 | INIT_LIST_HEAD(&pcpu_slot[i]); |
| 859 | |
| 860 | /* init and register vm area */ |
| 861 | static_vm.flags = VM_ALLOC; |
| 862 | static_vm.size = pcpu_chunk_size; |
Tejun Heo | c0c0a29 | 2009-02-24 11:57:21 +0900 | [diff] [blame] | 863 | vm_area_register_early(&static_vm, PAGE_SIZE); |
Tejun Heo | fbf59bc | 2009-02-20 16:29:08 +0900 | [diff] [blame] | 864 | |
| 865 | /* init static_chunk */ |
| 866 | static_chunk = alloc_bootmem(pcpu_chunk_struct_size); |
| 867 | INIT_LIST_HEAD(&static_chunk->list); |
| 868 | static_chunk->vm = &static_vm; |
| 869 | static_chunk->free_size = pcpu_unit_size - pcpu_static_size; |
| 870 | static_chunk->contig_hint = static_chunk->free_size; |
| 871 | |
| 872 | /* assign pages and map them */ |
| 873 | for_each_possible_cpu(cpu) { |
| 874 | for (i = 0; i < nr_cpu_pages; i++) { |
| 875 | *pcpu_chunk_pagep(static_chunk, cpu, i) = *pages++; |
| 876 | populate_pte_fn(pcpu_chunk_addr(static_chunk, cpu, i)); |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | err = pcpu_map(static_chunk, 0, nr_cpu_pages); |
| 881 | if (err) |
| 882 | panic("failed to setup static percpu area, err=%d\n", err); |
| 883 | |
| 884 | /* link static_chunk in */ |
| 885 | pcpu_chunk_relocate(static_chunk, -1); |
| 886 | pcpu_chunk_addr_insert(static_chunk); |
| 887 | |
| 888 | /* we're done */ |
| 889 | pcpu_base_addr = (void *)pcpu_chunk_addr(static_chunk, 0, 0); |
| 890 | return pcpu_unit_size; |
| 891 | } |