percpu.c revision 9f6455325618821dcf6775d7972881fde32e77c5
1/* 2 * 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. Each chunk is 11 * consisted of boot-time determined number of units and the first 12 * chunk is used for static percpu variables in the kernel image 13 * (special boot time alloc/init handling necessary as these areas 14 * need to be brought up before allocation services are running). 15 * Unit grows as necessary and all units grow or shrink in unison. 16 * When a chunk is filled up, another chunk is allocated. 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. On UMA, units corresponds directly to 26 * cpus. On NUMA, the mapping can be non-linear and even sparse. 27 * Percpu access can be done by configuring percpu base registers 28 * according to cpu to unit mapping and pcpu_unit_size. 29 * 30 * There are usually many small percpu allocations many of them being 31 * as small as 4 bytes. The allocator organizes chunks into lists 32 * according to free size and tries to allocate from the fullest one. 33 * Each chunk keeps the maximum contiguous area size hint which is 34 * guaranteed to be eqaul to or larger than the maximum contiguous 35 * area in the chunk. This helps the allocator not to iterate the 36 * chunk maps unnecessarily. 37 * 38 * Allocation state in each chunk is kept using an array of integers 39 * on chunk->map. A positive value in the map represents a free 40 * region and negative allocated. Allocation inside a chunk is done 41 * by scanning this map sequentially and serving the first matching 42 * entry. This is mostly copied from the percpu_modalloc() allocator. 43 * Chunks can be determined from the address using the index field 44 * in the page struct. The index field contains a pointer to the chunk. 45 * 46 * To use this allocator, arch code should do the followings. 47 * 48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 49 * regular address to percpu pointer and back if they need to be 50 * different from the default 51 * 52 * - use pcpu_setup_first_chunk() during percpu area initialization to 53 * setup the first chunk containing the kernel static percpu area 54 */ 55 56#include <linux/bitmap.h> 57#include <linux/bootmem.h> 58#include <linux/err.h> 59#include <linux/list.h> 60#include <linux/log2.h> 61#include <linux/mm.h> 62#include <linux/module.h> 63#include <linux/mutex.h> 64#include <linux/percpu.h> 65#include <linux/pfn.h> 66#include <linux/slab.h> 67#include <linux/spinlock.h> 68#include <linux/vmalloc.h> 69#include <linux/workqueue.h> 70 71#include <asm/cacheflush.h> 72#include <asm/sections.h> 73#include <asm/tlbflush.h> 74#include <asm/io.h> 75 76#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 77#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 78 79/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 80#ifndef __addr_to_pcpu_ptr 81#define __addr_to_pcpu_ptr(addr) \ 82 (void __percpu *)((unsigned long)(addr) - \ 83 (unsigned long)pcpu_base_addr + \ 84 (unsigned long)__per_cpu_start) 85#endif 86#ifndef __pcpu_ptr_to_addr 87#define __pcpu_ptr_to_addr(ptr) \ 88 (void __force *)((unsigned long)(ptr) + \ 89 (unsigned long)pcpu_base_addr - \ 90 (unsigned long)__per_cpu_start) 91#endif 92 93struct pcpu_chunk { 94 struct list_head list; /* linked to pcpu_slot lists */ 95 int free_size; /* free bytes in the chunk */ 96 int contig_hint; /* max contiguous size hint */ 97 void *base_addr; /* base address of this chunk */ 98 int map_used; /* # of map entries used */ 99 int map_alloc; /* # of map entries allocated */ 100 int *map; /* allocation map */ 101 void *data; /* chunk data */ 102 bool immutable; /* no [de]population allowed */ 103 unsigned long populated[]; /* populated bitmap */ 104}; 105 106static int pcpu_unit_pages __read_mostly; 107static int pcpu_unit_size __read_mostly; 108static int pcpu_nr_units __read_mostly; 109static int pcpu_atom_size __read_mostly; 110static int pcpu_nr_slots __read_mostly; 111static size_t pcpu_chunk_struct_size __read_mostly; 112 113/* cpus with the lowest and highest unit numbers */ 114static unsigned int pcpu_first_unit_cpu __read_mostly; 115static unsigned int pcpu_last_unit_cpu __read_mostly; 116 117/* the address of the first chunk which starts with the kernel static area */ 118void *pcpu_base_addr __read_mostly; 119EXPORT_SYMBOL_GPL(pcpu_base_addr); 120 121static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 122const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 123 124/* group information, used for vm allocation */ 125static int pcpu_nr_groups __read_mostly; 126static const unsigned long *pcpu_group_offsets __read_mostly; 127static const size_t *pcpu_group_sizes __read_mostly; 128 129/* 130 * The first chunk which always exists. Note that unlike other 131 * chunks, this one can be allocated and mapped in several different 132 * ways and thus often doesn't live in the vmalloc area. 133 */ 134static struct pcpu_chunk *pcpu_first_chunk; 135 136/* 137 * Optional reserved chunk. This chunk reserves part of the first 138 * chunk and serves it for reserved allocations. The amount of 139 * reserved offset is in pcpu_reserved_chunk_limit. When reserved 140 * area doesn't exist, the following variables contain NULL and 0 141 * respectively. 142 */ 143static struct pcpu_chunk *pcpu_reserved_chunk; 144static int pcpu_reserved_chunk_limit; 145 146/* 147 * Synchronization rules. 148 * 149 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 150 * protects allocation/reclaim paths, chunks, populated bitmap and 151 * vmalloc mapping. The latter is a spinlock and protects the index 152 * data structures - chunk slots, chunks and area maps in chunks. 153 * 154 * During allocation, pcpu_alloc_mutex is kept locked all the time and 155 * pcpu_lock is grabbed and released as necessary. All actual memory 156 * allocations are done using GFP_KERNEL with pcpu_lock released. In 157 * general, percpu memory can't be allocated with irq off but 158 * irqsave/restore are still used in alloc path so that it can be used 159 * from early init path - sched_init() specifically. 160 * 161 * Free path accesses and alters only the index data structures, so it 162 * can be safely called from atomic context. When memory needs to be 163 * returned to the system, free path schedules reclaim_work which 164 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 165 * reclaimed, release both locks and frees the chunks. Note that it's 166 * necessary to grab both locks to remove a chunk from circulation as 167 * allocation path might be referencing the chunk with only 168 * pcpu_alloc_mutex locked. 169 */ 170static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 171static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 172 173static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 174 175/* reclaim work to release fully free chunks, scheduled from free path */ 176static void pcpu_reclaim(struct work_struct *work); 177static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 178 179static bool pcpu_addr_in_first_chunk(void *addr) 180{ 181 void *first_start = pcpu_first_chunk->base_addr; 182 183 return addr >= first_start && addr < first_start + pcpu_unit_size; 184} 185 186static bool pcpu_addr_in_reserved_chunk(void *addr) 187{ 188 void *first_start = pcpu_first_chunk->base_addr; 189 190 return addr >= first_start && 191 addr < first_start + pcpu_reserved_chunk_limit; 192} 193 194static int __pcpu_size_to_slot(int size) 195{ 196 int highbit = fls(size); /* size is in bytes */ 197 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 198} 199 200static int pcpu_size_to_slot(int size) 201{ 202 if (size == pcpu_unit_size) 203 return pcpu_nr_slots - 1; 204 return __pcpu_size_to_slot(size); 205} 206 207static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 208{ 209 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 210 return 0; 211 212 return pcpu_size_to_slot(chunk->free_size); 213} 214 215/* set the pointer to a chunk in a page struct */ 216static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 217{ 218 page->index = (unsigned long)pcpu; 219} 220 221/* obtain pointer to a chunk from a page struct */ 222static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 223{ 224 return (struct pcpu_chunk *)page->index; 225} 226 227static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 228{ 229 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 230} 231 232static unsigned long __maybe_unused pcpu_chunk_addr(struct pcpu_chunk *chunk, 233 unsigned int cpu, int page_idx) 234{ 235 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 236 (page_idx << PAGE_SHIFT); 237} 238 239static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, 240 int *rs, int *re, int end) 241{ 242 *rs = find_next_zero_bit(chunk->populated, end, *rs); 243 *re = find_next_bit(chunk->populated, end, *rs + 1); 244} 245 246static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, 247 int *rs, int *re, int end) 248{ 249 *rs = find_next_bit(chunk->populated, end, *rs); 250 *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 251} 252 253/* 254 * (Un)populated page region iterators. Iterate over (un)populated 255 * page regions betwen @start and @end in @chunk. @rs and @re should 256 * be integer variables and will be set to start and end page index of 257 * the current region. 258 */ 259#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 260 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 261 (rs) < (re); \ 262 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 263 264#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 265 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 266 (rs) < (re); \ 267 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 268 269/** 270 * pcpu_mem_alloc - allocate memory 271 * @size: bytes to allocate 272 * 273 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 274 * kzalloc() is used; otherwise, vmalloc() is used. The returned 275 * memory is always zeroed. 276 * 277 * CONTEXT: 278 * Does GFP_KERNEL allocation. 279 * 280 * RETURNS: 281 * Pointer to the allocated area on success, NULL on failure. 282 */ 283static void *pcpu_mem_alloc(size_t size) 284{ 285 if (size <= PAGE_SIZE) 286 return kzalloc(size, GFP_KERNEL); 287 else { 288 void *ptr = vmalloc(size); 289 if (ptr) 290 memset(ptr, 0, size); 291 return ptr; 292 } 293} 294 295/** 296 * pcpu_mem_free - free memory 297 * @ptr: memory to free 298 * @size: size of the area 299 * 300 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). 301 */ 302static void pcpu_mem_free(void *ptr, size_t size) 303{ 304 if (size <= PAGE_SIZE) 305 kfree(ptr); 306 else 307 vfree(ptr); 308} 309 310/** 311 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 312 * @chunk: chunk of interest 313 * @oslot: the previous slot it was on 314 * 315 * This function is called after an allocation or free changed @chunk. 316 * New slot according to the changed state is determined and @chunk is 317 * moved to the slot. Note that the reserved chunk is never put on 318 * chunk slots. 319 * 320 * CONTEXT: 321 * pcpu_lock. 322 */ 323static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 324{ 325 int nslot = pcpu_chunk_slot(chunk); 326 327 if (chunk != pcpu_reserved_chunk && oslot != nslot) { 328 if (oslot < nslot) 329 list_move(&chunk->list, &pcpu_slot[nslot]); 330 else 331 list_move_tail(&chunk->list, &pcpu_slot[nslot]); 332 } 333} 334 335/** 336 * pcpu_need_to_extend - determine whether chunk area map needs to be extended 337 * @chunk: chunk of interest 338 * 339 * Determine whether area map of @chunk needs to be extended to 340 * accomodate a new allocation. 341 * 342 * CONTEXT: 343 * pcpu_lock. 344 * 345 * RETURNS: 346 * New target map allocation length if extension is necessary, 0 347 * otherwise. 348 */ 349static int pcpu_need_to_extend(struct pcpu_chunk *chunk) 350{ 351 int new_alloc; 352 353 if (chunk->map_alloc >= chunk->map_used + 2) 354 return 0; 355 356 new_alloc = PCPU_DFL_MAP_ALLOC; 357 while (new_alloc < chunk->map_used + 2) 358 new_alloc *= 2; 359 360 return new_alloc; 361} 362 363/** 364 * pcpu_extend_area_map - extend area map of a chunk 365 * @chunk: chunk of interest 366 * @new_alloc: new target allocation length of the area map 367 * 368 * Extend area map of @chunk to have @new_alloc entries. 369 * 370 * CONTEXT: 371 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. 372 * 373 * RETURNS: 374 * 0 on success, -errno on failure. 375 */ 376static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) 377{ 378 int *old = NULL, *new = NULL; 379 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); 380 unsigned long flags; 381 382 new = pcpu_mem_alloc(new_size); 383 if (!new) 384 return -ENOMEM; 385 386 /* acquire pcpu_lock and switch to new area map */ 387 spin_lock_irqsave(&pcpu_lock, flags); 388 389 if (new_alloc <= chunk->map_alloc) 390 goto out_unlock; 391 392 old_size = chunk->map_alloc * sizeof(chunk->map[0]); 393 memcpy(new, chunk->map, old_size); 394 395 /* 396 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is 397 * one of the first chunks and still using static map. 398 */ 399 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) 400 old = chunk->map; 401 402 chunk->map_alloc = new_alloc; 403 chunk->map = new; 404 new = NULL; 405 406out_unlock: 407 spin_unlock_irqrestore(&pcpu_lock, flags); 408 409 /* 410 * pcpu_mem_free() might end up calling vfree() which uses 411 * IRQ-unsafe lock and thus can't be called under pcpu_lock. 412 */ 413 pcpu_mem_free(old, old_size); 414 pcpu_mem_free(new, new_size); 415 416 return 0; 417} 418 419/** 420 * pcpu_split_block - split a map block 421 * @chunk: chunk of interest 422 * @i: index of map block to split 423 * @head: head size in bytes (can be 0) 424 * @tail: tail size in bytes (can be 0) 425 * 426 * Split the @i'th map block into two or three blocks. If @head is 427 * non-zero, @head bytes block is inserted before block @i moving it 428 * to @i+1 and reducing its size by @head bytes. 429 * 430 * If @tail is non-zero, the target block, which can be @i or @i+1 431 * depending on @head, is reduced by @tail bytes and @tail byte block 432 * is inserted after the target block. 433 * 434 * @chunk->map must have enough free slots to accomodate the split. 435 * 436 * CONTEXT: 437 * pcpu_lock. 438 */ 439static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 440 int head, int tail) 441{ 442 int nr_extra = !!head + !!tail; 443 444 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 445 446 /* insert new subblocks */ 447 memmove(&chunk->map[i + nr_extra], &chunk->map[i], 448 sizeof(chunk->map[0]) * (chunk->map_used - i)); 449 chunk->map_used += nr_extra; 450 451 if (head) { 452 chunk->map[i + 1] = chunk->map[i] - head; 453 chunk->map[i++] = head; 454 } 455 if (tail) { 456 chunk->map[i++] -= tail; 457 chunk->map[i] = tail; 458 } 459} 460 461/** 462 * pcpu_alloc_area - allocate area from a pcpu_chunk 463 * @chunk: chunk of interest 464 * @size: wanted size in bytes 465 * @align: wanted align 466 * 467 * Try to allocate @size bytes area aligned at @align from @chunk. 468 * Note that this function only allocates the offset. It doesn't 469 * populate or map the area. 470 * 471 * @chunk->map must have at least two free slots. 472 * 473 * CONTEXT: 474 * pcpu_lock. 475 * 476 * RETURNS: 477 * Allocated offset in @chunk on success, -1 if no matching area is 478 * found. 479 */ 480static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 481{ 482 int oslot = pcpu_chunk_slot(chunk); 483 int max_contig = 0; 484 int i, off; 485 486 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 487 bool is_last = i + 1 == chunk->map_used; 488 int head, tail; 489 490 /* extra for alignment requirement */ 491 head = ALIGN(off, align) - off; 492 BUG_ON(i == 0 && head != 0); 493 494 if (chunk->map[i] < 0) 495 continue; 496 if (chunk->map[i] < head + size) { 497 max_contig = max(chunk->map[i], max_contig); 498 continue; 499 } 500 501 /* 502 * If head is small or the previous block is free, 503 * merge'em. Note that 'small' is defined as smaller 504 * than sizeof(int), which is very small but isn't too 505 * uncommon for percpu allocations. 506 */ 507 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 508 if (chunk->map[i - 1] > 0) 509 chunk->map[i - 1] += head; 510 else { 511 chunk->map[i - 1] -= head; 512 chunk->free_size -= head; 513 } 514 chunk->map[i] -= head; 515 off += head; 516 head = 0; 517 } 518 519 /* if tail is small, just keep it around */ 520 tail = chunk->map[i] - head - size; 521 if (tail < sizeof(int)) 522 tail = 0; 523 524 /* split if warranted */ 525 if (head || tail) { 526 pcpu_split_block(chunk, i, head, tail); 527 if (head) { 528 i++; 529 off += head; 530 max_contig = max(chunk->map[i - 1], max_contig); 531 } 532 if (tail) 533 max_contig = max(chunk->map[i + 1], max_contig); 534 } 535 536 /* update hint and mark allocated */ 537 if (is_last) 538 chunk->contig_hint = max_contig; /* fully scanned */ 539 else 540 chunk->contig_hint = max(chunk->contig_hint, 541 max_contig); 542 543 chunk->free_size -= chunk->map[i]; 544 chunk->map[i] = -chunk->map[i]; 545 546 pcpu_chunk_relocate(chunk, oslot); 547 return off; 548 } 549 550 chunk->contig_hint = max_contig; /* fully scanned */ 551 pcpu_chunk_relocate(chunk, oslot); 552 553 /* tell the upper layer that this chunk has no matching area */ 554 return -1; 555} 556 557/** 558 * pcpu_free_area - free area to a pcpu_chunk 559 * @chunk: chunk of interest 560 * @freeme: offset of area to free 561 * 562 * Free area starting from @freeme to @chunk. Note that this function 563 * only modifies the allocation map. It doesn't depopulate or unmap 564 * the area. 565 * 566 * CONTEXT: 567 * pcpu_lock. 568 */ 569static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 570{ 571 int oslot = pcpu_chunk_slot(chunk); 572 int i, off; 573 574 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 575 if (off == freeme) 576 break; 577 BUG_ON(off != freeme); 578 BUG_ON(chunk->map[i] > 0); 579 580 chunk->map[i] = -chunk->map[i]; 581 chunk->free_size += chunk->map[i]; 582 583 /* merge with previous? */ 584 if (i > 0 && chunk->map[i - 1] >= 0) { 585 chunk->map[i - 1] += chunk->map[i]; 586 chunk->map_used--; 587 memmove(&chunk->map[i], &chunk->map[i + 1], 588 (chunk->map_used - i) * sizeof(chunk->map[0])); 589 i--; 590 } 591 /* merge with next? */ 592 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 593 chunk->map[i] += chunk->map[i + 1]; 594 chunk->map_used--; 595 memmove(&chunk->map[i + 1], &chunk->map[i + 2], 596 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 597 } 598 599 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 600 pcpu_chunk_relocate(chunk, oslot); 601} 602 603static struct pcpu_chunk *pcpu_alloc_chunk(void) 604{ 605 struct pcpu_chunk *chunk; 606 607 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); 608 if (!chunk) 609 return NULL; 610 611 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 612 if (!chunk->map) { 613 kfree(chunk); 614 return NULL; 615 } 616 617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 618 chunk->map[chunk->map_used++] = pcpu_unit_size; 619 620 INIT_LIST_HEAD(&chunk->list); 621 chunk->free_size = pcpu_unit_size; 622 chunk->contig_hint = pcpu_unit_size; 623 624 return chunk; 625} 626 627static void pcpu_free_chunk(struct pcpu_chunk *chunk) 628{ 629 if (!chunk) 630 return; 631 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 632 kfree(chunk); 633} 634 635/* 636 * Chunk management implementation. 637 * 638 * To allow different implementations, chunk alloc/free and 639 * [de]population are implemented in a separate file which is pulled 640 * into this file and compiled together. The following functions 641 * should be implemented. 642 * 643 * pcpu_populate_chunk - populate the specified range of a chunk 644 * pcpu_depopulate_chunk - depopulate the specified range of a chunk 645 * pcpu_create_chunk - create a new chunk 646 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 647 * pcpu_addr_to_page - translate address to physical address 648 * pcpu_verify_alloc_info - check alloc_info is acceptable during init 649 */ 650static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 651static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 652static struct pcpu_chunk *pcpu_create_chunk(void); 653static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 654static struct page *pcpu_addr_to_page(void *addr); 655static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 656 657#include "percpu-vm.c" 658 659/** 660 * pcpu_chunk_addr_search - determine chunk containing specified address 661 * @addr: address for which the chunk needs to be determined. 662 * 663 * RETURNS: 664 * The address of the found chunk. 665 */ 666static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 667{ 668 /* is it in the first chunk? */ 669 if (pcpu_addr_in_first_chunk(addr)) { 670 /* is it in the reserved area? */ 671 if (pcpu_addr_in_reserved_chunk(addr)) 672 return pcpu_reserved_chunk; 673 return pcpu_first_chunk; 674 } 675 676 /* 677 * The address is relative to unit0 which might be unused and 678 * thus unmapped. Offset the address to the unit space of the 679 * current processor before looking it up in the vmalloc 680 * space. Note that any possible cpu id can be used here, so 681 * there's no need to worry about preemption or cpu hotplug. 682 */ 683 addr += pcpu_unit_offsets[raw_smp_processor_id()]; 684 return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 685} 686 687/** 688 * pcpu_alloc - the percpu allocator 689 * @size: size of area to allocate in bytes 690 * @align: alignment of area (max PAGE_SIZE) 691 * @reserved: allocate from the reserved chunk if available 692 * 693 * Allocate percpu area of @size bytes aligned at @align. 694 * 695 * CONTEXT: 696 * Does GFP_KERNEL allocation. 697 * 698 * RETURNS: 699 * Percpu pointer to the allocated area on success, NULL on failure. 700 */ 701static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved) 702{ 703 static int warn_limit = 10; 704 struct pcpu_chunk *chunk; 705 const char *err; 706 int slot, off, new_alloc; 707 unsigned long flags; 708 709 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 710 WARN(true, "illegal size (%zu) or align (%zu) for " 711 "percpu allocation\n", size, align); 712 return NULL; 713 } 714 715 mutex_lock(&pcpu_alloc_mutex); 716 spin_lock_irqsave(&pcpu_lock, flags); 717 718 /* serve reserved allocations from the reserved chunk if available */ 719 if (reserved && pcpu_reserved_chunk) { 720 chunk = pcpu_reserved_chunk; 721 722 if (size > chunk->contig_hint) { 723 err = "alloc from reserved chunk failed"; 724 goto fail_unlock; 725 } 726 727 while ((new_alloc = pcpu_need_to_extend(chunk))) { 728 spin_unlock_irqrestore(&pcpu_lock, flags); 729 if (pcpu_extend_area_map(chunk, new_alloc) < 0) { 730 err = "failed to extend area map of reserved chunk"; 731 goto fail_unlock_mutex; 732 } 733 spin_lock_irqsave(&pcpu_lock, flags); 734 } 735 736 off = pcpu_alloc_area(chunk, size, align); 737 if (off >= 0) 738 goto area_found; 739 740 err = "alloc from reserved chunk failed"; 741 goto fail_unlock; 742 } 743 744restart: 745 /* search through normal chunks */ 746 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 747 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 748 if (size > chunk->contig_hint) 749 continue; 750 751 new_alloc = pcpu_need_to_extend(chunk); 752 if (new_alloc) { 753 spin_unlock_irqrestore(&pcpu_lock, flags); 754 if (pcpu_extend_area_map(chunk, 755 new_alloc) < 0) { 756 err = "failed to extend area map"; 757 goto fail_unlock_mutex; 758 } 759 spin_lock_irqsave(&pcpu_lock, flags); 760 /* 761 * pcpu_lock has been dropped, need to 762 * restart cpu_slot list walking. 763 */ 764 goto restart; 765 } 766 767 off = pcpu_alloc_area(chunk, size, align); 768 if (off >= 0) 769 goto area_found; 770 } 771 } 772 773 /* hmmm... no space left, create a new chunk */ 774 spin_unlock_irqrestore(&pcpu_lock, flags); 775 776 chunk = pcpu_create_chunk(); 777 if (!chunk) { 778 err = "failed to allocate new chunk"; 779 goto fail_unlock_mutex; 780 } 781 782 spin_lock_irqsave(&pcpu_lock, flags); 783 pcpu_chunk_relocate(chunk, -1); 784 goto restart; 785 786area_found: 787 spin_unlock_irqrestore(&pcpu_lock, flags); 788 789 /* populate, map and clear the area */ 790 if (pcpu_populate_chunk(chunk, off, size)) { 791 spin_lock_irqsave(&pcpu_lock, flags); 792 pcpu_free_area(chunk, off); 793 err = "failed to populate"; 794 goto fail_unlock; 795 } 796 797 mutex_unlock(&pcpu_alloc_mutex); 798 799 /* return address relative to base address */ 800 return __addr_to_pcpu_ptr(chunk->base_addr + off); 801 802fail_unlock: 803 spin_unlock_irqrestore(&pcpu_lock, flags); 804fail_unlock_mutex: 805 mutex_unlock(&pcpu_alloc_mutex); 806 if (warn_limit) { 807 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " 808 "%s\n", size, align, err); 809 dump_stack(); 810 if (!--warn_limit) 811 pr_info("PERCPU: limit reached, disable warning\n"); 812 } 813 return NULL; 814} 815 816/** 817 * __alloc_percpu - allocate dynamic percpu area 818 * @size: size of area to allocate in bytes 819 * @align: alignment of area (max PAGE_SIZE) 820 * 821 * Allocate percpu area of @size bytes aligned at @align. Might 822 * sleep. Might trigger writeouts. 823 * 824 * CONTEXT: 825 * Does GFP_KERNEL allocation. 826 * 827 * RETURNS: 828 * Percpu pointer to the allocated area on success, NULL on failure. 829 */ 830void __percpu *__alloc_percpu(size_t size, size_t align) 831{ 832 return pcpu_alloc(size, align, false); 833} 834EXPORT_SYMBOL_GPL(__alloc_percpu); 835 836/** 837 * __alloc_reserved_percpu - allocate reserved percpu area 838 * @size: size of area to allocate in bytes 839 * @align: alignment of area (max PAGE_SIZE) 840 * 841 * Allocate percpu area of @size bytes aligned at @align from reserved 842 * percpu area if arch has set it up; otherwise, allocation is served 843 * from the same dynamic area. Might sleep. Might trigger writeouts. 844 * 845 * CONTEXT: 846 * Does GFP_KERNEL allocation. 847 * 848 * RETURNS: 849 * Percpu pointer to the allocated area on success, NULL on failure. 850 */ 851void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 852{ 853 return pcpu_alloc(size, align, true); 854} 855 856/** 857 * pcpu_reclaim - reclaim fully free chunks, workqueue function 858 * @work: unused 859 * 860 * Reclaim all fully free chunks except for the first one. 861 * 862 * CONTEXT: 863 * workqueue context. 864 */ 865static void pcpu_reclaim(struct work_struct *work) 866{ 867 LIST_HEAD(todo); 868 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 869 struct pcpu_chunk *chunk, *next; 870 871 mutex_lock(&pcpu_alloc_mutex); 872 spin_lock_irq(&pcpu_lock); 873 874 list_for_each_entry_safe(chunk, next, head, list) { 875 WARN_ON(chunk->immutable); 876 877 /* spare the first one */ 878 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 879 continue; 880 881 list_move(&chunk->list, &todo); 882 } 883 884 spin_unlock_irq(&pcpu_lock); 885 886 list_for_each_entry_safe(chunk, next, &todo, list) { 887 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 888 pcpu_destroy_chunk(chunk); 889 } 890 891 mutex_unlock(&pcpu_alloc_mutex); 892} 893 894/** 895 * free_percpu - free percpu area 896 * @ptr: pointer to area to free 897 * 898 * Free percpu area @ptr. 899 * 900 * CONTEXT: 901 * Can be called from atomic context. 902 */ 903void free_percpu(void __percpu *ptr) 904{ 905 void *addr; 906 struct pcpu_chunk *chunk; 907 unsigned long flags; 908 int off; 909 910 if (!ptr) 911 return; 912 913 addr = __pcpu_ptr_to_addr(ptr); 914 915 spin_lock_irqsave(&pcpu_lock, flags); 916 917 chunk = pcpu_chunk_addr_search(addr); 918 off = addr - chunk->base_addr; 919 920 pcpu_free_area(chunk, off); 921 922 /* if there are more than one fully free chunks, wake up grim reaper */ 923 if (chunk->free_size == pcpu_unit_size) { 924 struct pcpu_chunk *pos; 925 926 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 927 if (pos != chunk) { 928 schedule_work(&pcpu_reclaim_work); 929 break; 930 } 931 } 932 933 spin_unlock_irqrestore(&pcpu_lock, flags); 934} 935EXPORT_SYMBOL_GPL(free_percpu); 936 937/** 938 * is_kernel_percpu_address - test whether address is from static percpu area 939 * @addr: address to test 940 * 941 * Test whether @addr belongs to in-kernel static percpu area. Module 942 * static percpu areas are not considered. For those, use 943 * is_module_percpu_address(). 944 * 945 * RETURNS: 946 * %true if @addr is from in-kernel static percpu area, %false otherwise. 947 */ 948bool is_kernel_percpu_address(unsigned long addr) 949{ 950 const size_t static_size = __per_cpu_end - __per_cpu_start; 951 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 952 unsigned int cpu; 953 954 for_each_possible_cpu(cpu) { 955 void *start = per_cpu_ptr(base, cpu); 956 957 if ((void *)addr >= start && (void *)addr < start + static_size) 958 return true; 959 } 960 return false; 961} 962 963/** 964 * per_cpu_ptr_to_phys - convert translated percpu address to physical address 965 * @addr: the address to be converted to physical address 966 * 967 * Given @addr which is dereferenceable address obtained via one of 968 * percpu access macros, this function translates it into its physical 969 * address. The caller is responsible for ensuring @addr stays valid 970 * until this function finishes. 971 * 972 * RETURNS: 973 * The physical address for @addr. 974 */ 975phys_addr_t per_cpu_ptr_to_phys(void *addr) 976{ 977 if (pcpu_addr_in_first_chunk(addr)) { 978 if ((unsigned long)addr < VMALLOC_START || 979 (unsigned long)addr >= VMALLOC_END) 980 return __pa(addr); 981 else 982 return page_to_phys(vmalloc_to_page(addr)); 983 } else 984 return page_to_phys(pcpu_addr_to_page(addr)); 985} 986 987static inline size_t pcpu_calc_fc_sizes(size_t static_size, 988 size_t reserved_size, 989 ssize_t *dyn_sizep) 990{ 991 size_t size_sum; 992 993 size_sum = PFN_ALIGN(static_size + reserved_size + 994 (*dyn_sizep >= 0 ? *dyn_sizep : 0)); 995 if (*dyn_sizep != 0) 996 *dyn_sizep = size_sum - static_size - reserved_size; 997 998 return size_sum; 999} 1000 1001/** 1002 * pcpu_alloc_alloc_info - allocate percpu allocation info 1003 * @nr_groups: the number of groups 1004 * @nr_units: the number of units 1005 * 1006 * Allocate ai which is large enough for @nr_groups groups containing 1007 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1008 * cpu_map array which is long enough for @nr_units and filled with 1009 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1010 * pointer of other groups. 1011 * 1012 * RETURNS: 1013 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1014 * failure. 1015 */ 1016struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1017 int nr_units) 1018{ 1019 struct pcpu_alloc_info *ai; 1020 size_t base_size, ai_size; 1021 void *ptr; 1022 int unit; 1023 1024 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1025 __alignof__(ai->groups[0].cpu_map[0])); 1026 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1027 1028 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1029 if (!ptr) 1030 return NULL; 1031 ai = ptr; 1032 ptr += base_size; 1033 1034 ai->groups[0].cpu_map = ptr; 1035 1036 for (unit = 0; unit < nr_units; unit++) 1037 ai->groups[0].cpu_map[unit] = NR_CPUS; 1038 1039 ai->nr_groups = nr_groups; 1040 ai->__ai_size = PFN_ALIGN(ai_size); 1041 1042 return ai; 1043} 1044 1045/** 1046 * pcpu_free_alloc_info - free percpu allocation info 1047 * @ai: pcpu_alloc_info to free 1048 * 1049 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1050 */ 1051void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1052{ 1053 free_bootmem(__pa(ai), ai->__ai_size); 1054} 1055 1056/** 1057 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1058 * @reserved_size: the size of reserved percpu area in bytes 1059 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1060 * @atom_size: allocation atom size 1061 * @cpu_distance_fn: callback to determine distance between cpus, optional 1062 * 1063 * This function determines grouping of units, their mappings to cpus 1064 * and other parameters considering needed percpu size, allocation 1065 * atom size and distances between CPUs. 1066 * 1067 * Groups are always mutliples of atom size and CPUs which are of 1068 * LOCAL_DISTANCE both ways are grouped together and share space for 1069 * units in the same group. The returned configuration is guaranteed 1070 * to have CPUs on different nodes on different groups and >=75% usage 1071 * of allocated virtual address space. 1072 * 1073 * RETURNS: 1074 * On success, pointer to the new allocation_info is returned. On 1075 * failure, ERR_PTR value is returned. 1076 */ 1077struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1078 size_t reserved_size, ssize_t dyn_size, 1079 size_t atom_size, 1080 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1081{ 1082 static int group_map[NR_CPUS] __initdata; 1083 static int group_cnt[NR_CPUS] __initdata; 1084 const size_t static_size = __per_cpu_end - __per_cpu_start; 1085 int group_cnt_max = 0, nr_groups = 1, nr_units = 0; 1086 size_t size_sum, min_unit_size, alloc_size; 1087 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1088 int last_allocs, group, unit; 1089 unsigned int cpu, tcpu; 1090 struct pcpu_alloc_info *ai; 1091 unsigned int *cpu_map; 1092 1093 /* this function may be called multiple times */ 1094 memset(group_map, 0, sizeof(group_map)); 1095 memset(group_cnt, 0, sizeof(group_map)); 1096 1097 /* 1098 * Determine min_unit_size, alloc_size and max_upa such that 1099 * alloc_size is multiple of atom_size and is the smallest 1100 * which can accomodate 4k aligned segments which are equal to 1101 * or larger than min_unit_size. 1102 */ 1103 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); 1104 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1105 1106 alloc_size = roundup(min_unit_size, atom_size); 1107 upa = alloc_size / min_unit_size; 1108 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1109 upa--; 1110 max_upa = upa; 1111 1112 /* group cpus according to their proximity */ 1113 for_each_possible_cpu(cpu) { 1114 group = 0; 1115 next_group: 1116 for_each_possible_cpu(tcpu) { 1117 if (cpu == tcpu) 1118 break; 1119 if (group_map[tcpu] == group && cpu_distance_fn && 1120 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1121 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1122 group++; 1123 nr_groups = max(nr_groups, group + 1); 1124 goto next_group; 1125 } 1126 } 1127 group_map[cpu] = group; 1128 group_cnt[group]++; 1129 group_cnt_max = max(group_cnt_max, group_cnt[group]); 1130 } 1131 1132 /* 1133 * Expand unit size until address space usage goes over 75% 1134 * and then as much as possible without using more address 1135 * space. 1136 */ 1137 last_allocs = INT_MAX; 1138 for (upa = max_upa; upa; upa--) { 1139 int allocs = 0, wasted = 0; 1140 1141 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1142 continue; 1143 1144 for (group = 0; group < nr_groups; group++) { 1145 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1146 allocs += this_allocs; 1147 wasted += this_allocs * upa - group_cnt[group]; 1148 } 1149 1150 /* 1151 * Don't accept if wastage is over 25%. The 1152 * greater-than comparison ensures upa==1 always 1153 * passes the following check. 1154 */ 1155 if (wasted > num_possible_cpus() / 3) 1156 continue; 1157 1158 /* and then don't consume more memory */ 1159 if (allocs > last_allocs) 1160 break; 1161 last_allocs = allocs; 1162 best_upa = upa; 1163 } 1164 upa = best_upa; 1165 1166 /* allocate and fill alloc_info */ 1167 for (group = 0; group < nr_groups; group++) 1168 nr_units += roundup(group_cnt[group], upa); 1169 1170 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1171 if (!ai) 1172 return ERR_PTR(-ENOMEM); 1173 cpu_map = ai->groups[0].cpu_map; 1174 1175 for (group = 0; group < nr_groups; group++) { 1176 ai->groups[group].cpu_map = cpu_map; 1177 cpu_map += roundup(group_cnt[group], upa); 1178 } 1179 1180 ai->static_size = static_size; 1181 ai->reserved_size = reserved_size; 1182 ai->dyn_size = dyn_size; 1183 ai->unit_size = alloc_size / upa; 1184 ai->atom_size = atom_size; 1185 ai->alloc_size = alloc_size; 1186 1187 for (group = 0, unit = 0; group_cnt[group]; group++) { 1188 struct pcpu_group_info *gi = &ai->groups[group]; 1189 1190 /* 1191 * Initialize base_offset as if all groups are located 1192 * back-to-back. The caller should update this to 1193 * reflect actual allocation. 1194 */ 1195 gi->base_offset = unit * ai->unit_size; 1196 1197 for_each_possible_cpu(cpu) 1198 if (group_map[cpu] == group) 1199 gi->cpu_map[gi->nr_units++] = cpu; 1200 gi->nr_units = roundup(gi->nr_units, upa); 1201 unit += gi->nr_units; 1202 } 1203 BUG_ON(unit != nr_units); 1204 1205 return ai; 1206} 1207 1208/** 1209 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1210 * @lvl: loglevel 1211 * @ai: allocation info to dump 1212 * 1213 * Print out information about @ai using loglevel @lvl. 1214 */ 1215static void pcpu_dump_alloc_info(const char *lvl, 1216 const struct pcpu_alloc_info *ai) 1217{ 1218 int group_width = 1, cpu_width = 1, width; 1219 char empty_str[] = "--------"; 1220 int alloc = 0, alloc_end = 0; 1221 int group, v; 1222 int upa, apl; /* units per alloc, allocs per line */ 1223 1224 v = ai->nr_groups; 1225 while (v /= 10) 1226 group_width++; 1227 1228 v = num_possible_cpus(); 1229 while (v /= 10) 1230 cpu_width++; 1231 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1232 1233 upa = ai->alloc_size / ai->unit_size; 1234 width = upa * (cpu_width + 1) + group_width + 3; 1235 apl = rounddown_pow_of_two(max(60 / width, 1)); 1236 1237 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1238 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1239 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1240 1241 for (group = 0; group < ai->nr_groups; group++) { 1242 const struct pcpu_group_info *gi = &ai->groups[group]; 1243 int unit = 0, unit_end = 0; 1244 1245 BUG_ON(gi->nr_units % upa); 1246 for (alloc_end += gi->nr_units / upa; 1247 alloc < alloc_end; alloc++) { 1248 if (!(alloc % apl)) { 1249 printk("\n"); 1250 printk("%spcpu-alloc: ", lvl); 1251 } 1252 printk("[%0*d] ", group_width, group); 1253 1254 for (unit_end += upa; unit < unit_end; unit++) 1255 if (gi->cpu_map[unit] != NR_CPUS) 1256 printk("%0*d ", cpu_width, 1257 gi->cpu_map[unit]); 1258 else 1259 printk("%s ", empty_str); 1260 } 1261 } 1262 printk("\n"); 1263} 1264 1265/** 1266 * pcpu_setup_first_chunk - initialize the first percpu chunk 1267 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1268 * @base_addr: mapped address 1269 * 1270 * Initialize the first percpu chunk which contains the kernel static 1271 * perpcu area. This function is to be called from arch percpu area 1272 * setup path. 1273 * 1274 * @ai contains all information necessary to initialize the first 1275 * chunk and prime the dynamic percpu allocator. 1276 * 1277 * @ai->static_size is the size of static percpu area. 1278 * 1279 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1280 * reserve after the static area in the first chunk. This reserves 1281 * the first chunk such that it's available only through reserved 1282 * percpu allocation. This is primarily used to serve module percpu 1283 * static areas on architectures where the addressing model has 1284 * limited offset range for symbol relocations to guarantee module 1285 * percpu symbols fall inside the relocatable range. 1286 * 1287 * @ai->dyn_size determines the number of bytes available for dynamic 1288 * allocation in the first chunk. The area between @ai->static_size + 1289 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1290 * 1291 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1292 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1293 * @ai->dyn_size. 1294 * 1295 * @ai->atom_size is the allocation atom size and used as alignment 1296 * for vm areas. 1297 * 1298 * @ai->alloc_size is the allocation size and always multiple of 1299 * @ai->atom_size. This is larger than @ai->atom_size if 1300 * @ai->unit_size is larger than @ai->atom_size. 1301 * 1302 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1303 * percpu areas. Units which should be colocated are put into the 1304 * same group. Dynamic VM areas will be allocated according to these 1305 * groupings. If @ai->nr_groups is zero, a single group containing 1306 * all units is assumed. 1307 * 1308 * The caller should have mapped the first chunk at @base_addr and 1309 * copied static data to each unit. 1310 * 1311 * If the first chunk ends up with both reserved and dynamic areas, it 1312 * is served by two chunks - one to serve the core static and reserved 1313 * areas and the other for the dynamic area. They share the same vm 1314 * and page map but uses different area allocation map to stay away 1315 * from each other. The latter chunk is circulated in the chunk slots 1316 * and available for dynamic allocation like any other chunks. 1317 * 1318 * RETURNS: 1319 * 0 on success, -errno on failure. 1320 */ 1321int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1322 void *base_addr) 1323{ 1324 static char cpus_buf[4096] __initdata; 1325 static int smap[2], dmap[2]; 1326 size_t dyn_size = ai->dyn_size; 1327 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1328 struct pcpu_chunk *schunk, *dchunk = NULL; 1329 unsigned long *group_offsets; 1330 size_t *group_sizes; 1331 unsigned long *unit_off; 1332 unsigned int cpu; 1333 int *unit_map; 1334 int group, unit, i; 1335 1336 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); 1337 1338#define PCPU_SETUP_BUG_ON(cond) do { \ 1339 if (unlikely(cond)) { \ 1340 pr_emerg("PERCPU: failed to initialize, %s", #cond); \ 1341 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ 1342 pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1343 BUG(); \ 1344 } \ 1345} while (0) 1346 1347 /* sanity checks */ 1348 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || 1349 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); 1350 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1351 PCPU_SETUP_BUG_ON(!ai->static_size); 1352 PCPU_SETUP_BUG_ON(!base_addr); 1353 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1354 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); 1355 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1356 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 1357 1358 /* process group information and build config tables accordingly */ 1359 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 1360 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1361 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1362 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1363 1364 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1365 unit_map[cpu] = UINT_MAX; 1366 pcpu_first_unit_cpu = NR_CPUS; 1367 1368 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1369 const struct pcpu_group_info *gi = &ai->groups[group]; 1370 1371 group_offsets[group] = gi->base_offset; 1372 group_sizes[group] = gi->nr_units * ai->unit_size; 1373 1374 for (i = 0; i < gi->nr_units; i++) { 1375 cpu = gi->cpu_map[i]; 1376 if (cpu == NR_CPUS) 1377 continue; 1378 1379 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); 1380 PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1381 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1382 1383 unit_map[cpu] = unit + i; 1384 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1385 1386 if (pcpu_first_unit_cpu == NR_CPUS) 1387 pcpu_first_unit_cpu = cpu; 1388 } 1389 } 1390 pcpu_last_unit_cpu = cpu; 1391 pcpu_nr_units = unit; 1392 1393 for_each_possible_cpu(cpu) 1394 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1395 1396 /* we're done parsing the input, undefine BUG macro and dump config */ 1397#undef PCPU_SETUP_BUG_ON 1398 pcpu_dump_alloc_info(KERN_INFO, ai); 1399 1400 pcpu_nr_groups = ai->nr_groups; 1401 pcpu_group_offsets = group_offsets; 1402 pcpu_group_sizes = group_sizes; 1403 pcpu_unit_map = unit_map; 1404 pcpu_unit_offsets = unit_off; 1405 1406 /* determine basic parameters */ 1407 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1408 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1409 pcpu_atom_size = ai->atom_size; 1410 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1411 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1412 1413 /* 1414 * Allocate chunk slots. The additional last slot is for 1415 * empty chunks. 1416 */ 1417 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1418 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1419 for (i = 0; i < pcpu_nr_slots; i++) 1420 INIT_LIST_HEAD(&pcpu_slot[i]); 1421 1422 /* 1423 * Initialize static chunk. If reserved_size is zero, the 1424 * static chunk covers static area + dynamic allocation area 1425 * in the first chunk. If reserved_size is not zero, it 1426 * covers static area + reserved area (mostly used for module 1427 * static percpu allocation). 1428 */ 1429 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1430 INIT_LIST_HEAD(&schunk->list); 1431 schunk->base_addr = base_addr; 1432 schunk->map = smap; 1433 schunk->map_alloc = ARRAY_SIZE(smap); 1434 schunk->immutable = true; 1435 bitmap_fill(schunk->populated, pcpu_unit_pages); 1436 1437 if (ai->reserved_size) { 1438 schunk->free_size = ai->reserved_size; 1439 pcpu_reserved_chunk = schunk; 1440 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1441 } else { 1442 schunk->free_size = dyn_size; 1443 dyn_size = 0; /* dynamic area covered */ 1444 } 1445 schunk->contig_hint = schunk->free_size; 1446 1447 schunk->map[schunk->map_used++] = -ai->static_size; 1448 if (schunk->free_size) 1449 schunk->map[schunk->map_used++] = schunk->free_size; 1450 1451 /* init dynamic chunk if necessary */ 1452 if (dyn_size) { 1453 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1454 INIT_LIST_HEAD(&dchunk->list); 1455 dchunk->base_addr = base_addr; 1456 dchunk->map = dmap; 1457 dchunk->map_alloc = ARRAY_SIZE(dmap); 1458 dchunk->immutable = true; 1459 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1460 1461 dchunk->contig_hint = dchunk->free_size = dyn_size; 1462 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1463 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1464 } 1465 1466 /* link the first chunk in */ 1467 pcpu_first_chunk = dchunk ?: schunk; 1468 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1469 1470 /* we're done */ 1471 pcpu_base_addr = base_addr; 1472 return 0; 1473} 1474 1475const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1476 [PCPU_FC_AUTO] = "auto", 1477 [PCPU_FC_EMBED] = "embed", 1478 [PCPU_FC_PAGE] = "page", 1479}; 1480 1481enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1482 1483static int __init percpu_alloc_setup(char *str) 1484{ 1485 if (0) 1486 /* nada */; 1487#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1488 else if (!strcmp(str, "embed")) 1489 pcpu_chosen_fc = PCPU_FC_EMBED; 1490#endif 1491#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1492 else if (!strcmp(str, "page")) 1493 pcpu_chosen_fc = PCPU_FC_PAGE; 1494#endif 1495 else 1496 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1497 1498 return 0; 1499} 1500early_param("percpu_alloc", percpu_alloc_setup); 1501 1502#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1503 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1504/** 1505 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1506 * @reserved_size: the size of reserved percpu area in bytes 1507 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1508 * @atom_size: allocation atom size 1509 * @cpu_distance_fn: callback to determine distance between cpus, optional 1510 * @alloc_fn: function to allocate percpu page 1511 * @free_fn: funtion to free percpu page 1512 * 1513 * This is a helper to ease setting up embedded first percpu chunk and 1514 * can be called where pcpu_setup_first_chunk() is expected. 1515 * 1516 * If this function is used to setup the first chunk, it is allocated 1517 * by calling @alloc_fn and used as-is without being mapped into 1518 * vmalloc area. Allocations are always whole multiples of @atom_size 1519 * aligned to @atom_size. 1520 * 1521 * This enables the first chunk to piggy back on the linear physical 1522 * mapping which often uses larger page size. Please note that this 1523 * can result in very sparse cpu->unit mapping on NUMA machines thus 1524 * requiring large vmalloc address space. Don't use this allocator if 1525 * vmalloc space is not orders of magnitude larger than distances 1526 * between node memory addresses (ie. 32bit NUMA machines). 1527 * 1528 * When @dyn_size is positive, dynamic area might be larger than 1529 * specified to fill page alignment. When @dyn_size is auto, 1530 * @dyn_size is just big enough to fill page alignment after static 1531 * and reserved areas. 1532 * 1533 * If the needed size is smaller than the minimum or specified unit 1534 * size, the leftover is returned using @free_fn. 1535 * 1536 * RETURNS: 1537 * 0 on success, -errno on failure. 1538 */ 1539int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size, 1540 size_t atom_size, 1541 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1542 pcpu_fc_alloc_fn_t alloc_fn, 1543 pcpu_fc_free_fn_t free_fn) 1544{ 1545 void *base = (void *)ULONG_MAX; 1546 void **areas = NULL; 1547 struct pcpu_alloc_info *ai; 1548 size_t size_sum, areas_size, max_distance; 1549 int group, i, rc; 1550 1551 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1552 cpu_distance_fn); 1553 if (IS_ERR(ai)) 1554 return PTR_ERR(ai); 1555 1556 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1557 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 1558 1559 areas = alloc_bootmem_nopanic(areas_size); 1560 if (!areas) { 1561 rc = -ENOMEM; 1562 goto out_free; 1563 } 1564 1565 /* allocate, copy and determine base address */ 1566 for (group = 0; group < ai->nr_groups; group++) { 1567 struct pcpu_group_info *gi = &ai->groups[group]; 1568 unsigned int cpu = NR_CPUS; 1569 void *ptr; 1570 1571 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1572 cpu = gi->cpu_map[i]; 1573 BUG_ON(cpu == NR_CPUS); 1574 1575 /* allocate space for the whole group */ 1576 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1577 if (!ptr) { 1578 rc = -ENOMEM; 1579 goto out_free_areas; 1580 } 1581 areas[group] = ptr; 1582 1583 base = min(ptr, base); 1584 1585 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1586 if (gi->cpu_map[i] == NR_CPUS) { 1587 /* unused unit, free whole */ 1588 free_fn(ptr, ai->unit_size); 1589 continue; 1590 } 1591 /* copy and return the unused part */ 1592 memcpy(ptr, __per_cpu_load, ai->static_size); 1593 free_fn(ptr + size_sum, ai->unit_size - size_sum); 1594 } 1595 } 1596 1597 /* base address is now known, determine group base offsets */ 1598 max_distance = 0; 1599 for (group = 0; group < ai->nr_groups; group++) { 1600 ai->groups[group].base_offset = areas[group] - base; 1601 max_distance = max_t(size_t, max_distance, 1602 ai->groups[group].base_offset); 1603 } 1604 max_distance += ai->unit_size; 1605 1606 /* warn if maximum distance is further than 75% of vmalloc space */ 1607 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { 1608 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " 1609 "space 0x%lx\n", 1610 max_distance, VMALLOC_END - VMALLOC_START); 1611#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1612 /* and fail if we have fallback */ 1613 rc = -EINVAL; 1614 goto out_free; 1615#endif 1616 } 1617 1618 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1619 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1620 ai->dyn_size, ai->unit_size); 1621 1622 rc = pcpu_setup_first_chunk(ai, base); 1623 goto out_free; 1624 1625out_free_areas: 1626 for (group = 0; group < ai->nr_groups; group++) 1627 free_fn(areas[group], 1628 ai->groups[group].nr_units * ai->unit_size); 1629out_free: 1630 pcpu_free_alloc_info(ai); 1631 if (areas) 1632 free_bootmem(__pa(areas), areas_size); 1633 return rc; 1634} 1635#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1636 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1637 1638#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1639/** 1640 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1641 * @reserved_size: the size of reserved percpu area in bytes 1642 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1643 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1644 * @populate_pte_fn: function to populate pte 1645 * 1646 * This is a helper to ease setting up page-remapped first percpu 1647 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1648 * 1649 * This is the basic allocator. Static percpu area is allocated 1650 * page-by-page into vmalloc area. 1651 * 1652 * RETURNS: 1653 * 0 on success, -errno on failure. 1654 */ 1655int __init pcpu_page_first_chunk(size_t reserved_size, 1656 pcpu_fc_alloc_fn_t alloc_fn, 1657 pcpu_fc_free_fn_t free_fn, 1658 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1659{ 1660 static struct vm_struct vm; 1661 struct pcpu_alloc_info *ai; 1662 char psize_str[16]; 1663 int unit_pages; 1664 size_t pages_size; 1665 struct page **pages; 1666 int unit, i, j, rc; 1667 1668 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1669 1670 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); 1671 if (IS_ERR(ai)) 1672 return PTR_ERR(ai); 1673 BUG_ON(ai->nr_groups != 1); 1674 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1675 1676 unit_pages = ai->unit_size >> PAGE_SHIFT; 1677 1678 /* unaligned allocations can't be freed, round up to page size */ 1679 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1680 sizeof(pages[0])); 1681 pages = alloc_bootmem(pages_size); 1682 1683 /* allocate pages */ 1684 j = 0; 1685 for (unit = 0; unit < num_possible_cpus(); unit++) 1686 for (i = 0; i < unit_pages; i++) { 1687 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1688 void *ptr; 1689 1690 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1691 if (!ptr) { 1692 pr_warning("PERCPU: failed to allocate %s page " 1693 "for cpu%u\n", psize_str, cpu); 1694 goto enomem; 1695 } 1696 pages[j++] = virt_to_page(ptr); 1697 } 1698 1699 /* allocate vm area, map the pages and copy static data */ 1700 vm.flags = VM_ALLOC; 1701 vm.size = num_possible_cpus() * ai->unit_size; 1702 vm_area_register_early(&vm, PAGE_SIZE); 1703 1704 for (unit = 0; unit < num_possible_cpus(); unit++) { 1705 unsigned long unit_addr = 1706 (unsigned long)vm.addr + unit * ai->unit_size; 1707 1708 for (i = 0; i < unit_pages; i++) 1709 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1710 1711 /* pte already populated, the following shouldn't fail */ 1712 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1713 unit_pages); 1714 if (rc < 0) 1715 panic("failed to map percpu area, err=%d\n", rc); 1716 1717 /* 1718 * FIXME: Archs with virtual cache should flush local 1719 * cache for the linear mapping here - something 1720 * equivalent to flush_cache_vmap() on the local cpu. 1721 * flush_cache_vmap() can't be used as most supporting 1722 * data structures are not set up yet. 1723 */ 1724 1725 /* copy static data */ 1726 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1727 } 1728 1729 /* we're ready, commit */ 1730 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1731 unit_pages, psize_str, vm.addr, ai->static_size, 1732 ai->reserved_size, ai->dyn_size); 1733 1734 rc = pcpu_setup_first_chunk(ai, vm.addr); 1735 goto out_free_ar; 1736 1737enomem: 1738 while (--j >= 0) 1739 free_fn(page_address(pages[j]), PAGE_SIZE); 1740 rc = -ENOMEM; 1741out_free_ar: 1742 free_bootmem(__pa(pages), pages_size); 1743 pcpu_free_alloc_info(ai); 1744 return rc; 1745} 1746#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 1747 1748/* 1749 * Generic percpu area setup. 1750 * 1751 * The embedding helper is used because its behavior closely resembles 1752 * the original non-dynamic generic percpu area setup. This is 1753 * important because many archs have addressing restrictions and might 1754 * fail if the percpu area is located far away from the previous 1755 * location. As an added bonus, in non-NUMA cases, embedding is 1756 * generally a good idea TLB-wise because percpu area can piggy back 1757 * on the physical linear memory mapping which uses large page 1758 * mappings on applicable archs. 1759 */ 1760#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 1761unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 1762EXPORT_SYMBOL(__per_cpu_offset); 1763 1764static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 1765 size_t align) 1766{ 1767 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 1768} 1769 1770static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 1771{ 1772 free_bootmem(__pa(ptr), size); 1773} 1774 1775void __init setup_per_cpu_areas(void) 1776{ 1777 unsigned long delta; 1778 unsigned int cpu; 1779 int rc; 1780 1781 /* 1782 * Always reserve area for module percpu variables. That's 1783 * what the legacy allocator did. 1784 */ 1785 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 1786 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 1787 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 1788 if (rc < 0) 1789 panic("Failed to initialized percpu areas."); 1790 1791 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 1792 for_each_possible_cpu(cpu) 1793 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 1794} 1795#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1796