1/* Modified by Broadcom Corp. Portions Copyright (c) Broadcom Corp, 2012. */ 2/* 3 * mm/percpu.c - percpu memory allocator 4 * 5 * Copyright (C) 2009 SUSE Linux Products GmbH 6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 7 * 8 * This file is released under the GPLv2. 9 * 10 * This is percpu allocator which can handle both static and dynamic 11 * areas. Percpu areas are allocated in chunks. Each chunk is 12 * consisted of boot-time determined number of units and the first 13 * chunk is used for static percpu variables in the kernel image 14 * (special boot time alloc/init handling necessary as these areas 15 * need to be brought up before allocation services are running). 16 * Unit grows as necessary and all units grow or shrink in unison. 17 * When a chunk is filled up, another chunk is allocated. 18 * 19 * c0 c1 c2 20 * ------------------- ------------------- ------------ 21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 22 * ------------------- ...... ------------------- .... ------------ 23 * 24 * Allocation is done in offset-size areas of single unit space. Ie, 25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, 26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to 27 * cpus. On NUMA, the mapping can be non-linear and even sparse. 28 * Percpu access can be done by configuring percpu base registers 29 * according to cpu to unit mapping and pcpu_unit_size. 30 * 31 * There are usually many small percpu allocations many of them being 32 * as small as 4 bytes. The allocator organizes chunks into lists 33 * according to free size and tries to allocate from the fullest one. 34 * Each chunk keeps the maximum contiguous area size hint which is 35 * guaranteed to be eqaul to or larger than the maximum contiguous 36 * area in the chunk. This helps the allocator not to iterate the 37 * chunk maps unnecessarily. 38 * 39 * Allocation state in each chunk is kept using an array of integers 40 * on chunk->map. A positive value in the map represents a free 41 * region and negative allocated. Allocation inside a chunk is done 42 * by scanning this map sequentially and serving the first matching 43 * entry. This is mostly copied from the percpu_modalloc() allocator. 44 * Chunks can be determined from the address using the index field 45 * in the page struct. The index field contains a pointer to the chunk. 46 * 47 * To use this allocator, arch code should do the followings. 48 * 49 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 50 * regular address to percpu pointer and back if they need to be 51 * different from the default 52 * 53 * - use pcpu_setup_first_chunk() during percpu area initialization to 54 * setup the first chunk containing the kernel static percpu area 55 */ 56 57#include <linux/bitmap.h> 58#include <linux/bootmem.h> 59#include <linux/err.h> 60#include <linux/list.h> 61#include <linux/log2.h> 62#include <linux/mm.h> 63#include <linux/module.h> 64#include <linux/mutex.h> 65#include <linux/percpu.h> 66#include <linux/pfn.h> 67#include <linux/slab.h> 68#include <linux/spinlock.h> 69#include <linux/vmalloc.h> 70#include <linux/workqueue.h> 71 72#include <asm/cacheflush.h> 73#include <asm/sections.h> 74#include <asm/tlbflush.h> 75#include <asm/io.h> 76 77#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 79 80/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 81#ifndef __addr_to_pcpu_ptr 82#define __addr_to_pcpu_ptr(addr) \ 83 (void __percpu *)((unsigned long)(addr) - \ 84 (unsigned long)pcpu_base_addr + \ 85 (unsigned long)__per_cpu_start) 86#endif 87#ifndef __pcpu_ptr_to_addr 88#define __pcpu_ptr_to_addr(ptr) \ 89 (void __force *)((unsigned long)(ptr) + \ 90 (unsigned long)pcpu_base_addr - \ 91 (unsigned long)__per_cpu_start) 92#endif 93 94struct pcpu_chunk { 95 struct list_head list; /* linked to pcpu_slot lists */ 96 int free_size; /* free bytes in the chunk */ 97 int contig_hint; /* max contiguous size hint */ 98 void *base_addr; /* base address of this chunk */ 99 int map_used; /* # of map entries used */ 100 int map_alloc; /* # of map entries allocated */ 101 int *map; /* allocation map */ 102 void *data; /* chunk data */ 103 bool immutable; /* no [de]population allowed */ 104 unsigned long populated[]; /* populated bitmap */ 105}; 106 107static int pcpu_unit_pages __read_mostly; 108static int pcpu_unit_size __read_mostly; 109static int pcpu_nr_units __read_mostly; 110static int pcpu_atom_size __read_mostly; 111static int pcpu_nr_slots __read_mostly; 112static size_t pcpu_chunk_struct_size __read_mostly; 113 114/* cpus with the lowest and highest unit numbers */ 115static unsigned int pcpu_first_unit_cpu __read_mostly; 116static unsigned int pcpu_last_unit_cpu __read_mostly; 117 118/* the address of the first chunk which starts with the kernel static area */ 119void *pcpu_base_addr __read_mostly; 120EXPORT_SYMBOL_GPL(pcpu_base_addr); 121 122static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 123const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 124 125/* group information, used for vm allocation */ 126static int pcpu_nr_groups __read_mostly; 127static const unsigned long *pcpu_group_offsets __read_mostly; 128static const size_t *pcpu_group_sizes __read_mostly; 129 130/* 131 * The first chunk which always exists. Note that unlike other 132 * chunks, this one can be allocated and mapped in several different 133 * ways and thus often doesn't live in the vmalloc area. 134 */ 135static struct pcpu_chunk *pcpu_first_chunk; 136 137/* 138 * Optional reserved chunk. This chunk reserves part of the first 139 * chunk and serves it for reserved allocations. The amount of 140 * reserved offset is in pcpu_reserved_chunk_limit. When reserved 141 * area doesn't exist, the following variables contain NULL and 0 142 * respectively. 143 */ 144static struct pcpu_chunk *pcpu_reserved_chunk; 145static int pcpu_reserved_chunk_limit; 146 147/* 148 * Synchronization rules. 149 * 150 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 151 * protects allocation/reclaim paths, chunks, populated bitmap and 152 * vmalloc mapping. The latter is a spinlock and protects the index 153 * data structures - chunk slots, chunks and area maps in chunks. 154 * 155 * During allocation, pcpu_alloc_mutex is kept locked all the time and 156 * pcpu_lock is grabbed and released as necessary. All actual memory 157 * allocations are done using GFP_KERNEL with pcpu_lock released. In 158 * general, percpu memory can't be allocated with irq off but 159 * irqsave/restore are still used in alloc path so that it can be used 160 * from early init path - sched_init() specifically. 161 * 162 * Free path accesses and alters only the index data structures, so it 163 * can be safely called from atomic context. When memory needs to be 164 * returned to the system, free path schedules reclaim_work which 165 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 166 * reclaimed, release both locks and frees the chunks. Note that it's 167 * necessary to grab both locks to remove a chunk from circulation as 168 * allocation path might be referencing the chunk with only 169 * pcpu_alloc_mutex locked. 170 */ 171static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 172static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 173 174static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 175 176/* reclaim work to release fully free chunks, scheduled from free path */ 177static void pcpu_reclaim(struct work_struct *work); 178static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 179 180static bool pcpu_addr_in_first_chunk(void *addr) 181{ 182 void *first_start = pcpu_first_chunk->base_addr; 183 184 return addr >= first_start && addr < first_start + pcpu_unit_size; 185} 186 187static bool pcpu_addr_in_reserved_chunk(void *addr) 188{ 189 void *first_start = pcpu_first_chunk->base_addr; 190 191 return addr >= first_start && 192 addr < first_start + pcpu_reserved_chunk_limit; 193} 194 195static int __pcpu_size_to_slot(int size) 196{ 197 int highbit = fls(size); /* size is in bytes */ 198 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 199} 200 201static int pcpu_size_to_slot(int size) 202{ 203 if (size == pcpu_unit_size) 204 return pcpu_nr_slots - 1; 205 return __pcpu_size_to_slot(size); 206} 207 208static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 209{ 210 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 211 return 0; 212 213 return pcpu_size_to_slot(chunk->free_size); 214} 215 216/* set the pointer to a chunk in a page struct */ 217static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 218{ 219 page->index = (unsigned long)pcpu; 220} 221 222/* obtain pointer to a chunk from a page struct */ 223static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 224{ 225 return (struct pcpu_chunk *)page->index; 226} 227 228static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 229{ 230 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 231} 232 233static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 234 unsigned int cpu, int page_idx) 235{ 236 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 237 (page_idx << PAGE_SHIFT); 238} 239 240static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, 241 int *rs, int *re, int end) 242{ 243 *rs = find_next_zero_bit(chunk->populated, end, *rs); 244 *re = find_next_bit(chunk->populated, end, *rs + 1); 245} 246 247static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, 248 int *rs, int *re, int end) 249{ 250 *rs = find_next_bit(chunk->populated, end, *rs); 251 *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 252} 253 254/* 255 * (Un)populated page region iterators. Iterate over (un)populated 256 * page regions betwen @start and @end in @chunk. @rs and @re should 257 * be integer variables and will be set to start and end page index of 258 * the current region. 259 */ 260#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 261 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 262 (rs) < (re); \ 263 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 264 265#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 266 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 267 (rs) < (re); \ 268 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 269 270/** 271 * pcpu_mem_alloc - allocate memory 272 * @size: bytes to allocate 273 * 274 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 275 * kzalloc() is used; otherwise, vmalloc() is used. The returned 276 * memory is always zeroed. 277 * 278 * CONTEXT: 279 * Does GFP_KERNEL allocation. 280 * 281 * RETURNS: 282 * Pointer to the allocated area on success, NULL on failure. 283 */ 284static void *pcpu_mem_alloc(size_t size) 285{ 286 if (WARN_ON_ONCE(!slab_is_available())) 287 return NULL; 288 289 if (size <= PAGE_SIZE) 290 return kzalloc(size, GFP_KERNEL); 291 else { 292 void *ptr = vmalloc(size); 293 if (ptr) 294 memset(ptr, 0, size); 295 return ptr; 296 } 297} 298 299/** 300 * pcpu_mem_free - free memory 301 * @ptr: memory to free 302 * @size: size of the area 303 * 304 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). 305 */ 306static void pcpu_mem_free(void *ptr, size_t size) 307{ 308 if (size <= PAGE_SIZE) 309 kfree(ptr); 310 else 311 vfree(ptr); 312} 313 314/** 315 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 316 * @chunk: chunk of interest 317 * @oslot: the previous slot it was on 318 * 319 * This function is called after an allocation or free changed @chunk. 320 * New slot according to the changed state is determined and @chunk is 321 * moved to the slot. Note that the reserved chunk is never put on 322 * chunk slots. 323 * 324 * CONTEXT: 325 * pcpu_lock. 326 */ 327static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 328{ 329 int nslot = pcpu_chunk_slot(chunk); 330 331 if (chunk != pcpu_reserved_chunk && oslot != nslot) { 332 if (oslot < nslot) 333 list_move(&chunk->list, &pcpu_slot[nslot]); 334 else 335 list_move_tail(&chunk->list, &pcpu_slot[nslot]); 336 } 337} 338 339/** 340 * pcpu_need_to_extend - determine whether chunk area map needs to be extended 341 * @chunk: chunk of interest 342 * 343 * Determine whether area map of @chunk needs to be extended to 344 * accomodate a new allocation. 345 * 346 * CONTEXT: 347 * pcpu_lock. 348 * 349 * RETURNS: 350 * New target map allocation length if extension is necessary, 0 351 * otherwise. 352 */ 353static int pcpu_need_to_extend(struct pcpu_chunk *chunk) 354{ 355 int new_alloc; 356 357 if (chunk->map_alloc >= chunk->map_used + 2) 358 return 0; 359 360 new_alloc = PCPU_DFL_MAP_ALLOC; 361 while (new_alloc < chunk->map_used + 2) 362 new_alloc *= 2; 363 364 return new_alloc; 365} 366 367/** 368 * pcpu_extend_area_map - extend area map of a chunk 369 * @chunk: chunk of interest 370 * @new_alloc: new target allocation length of the area map 371 * 372 * Extend area map of @chunk to have @new_alloc entries. 373 * 374 * CONTEXT: 375 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. 376 * 377 * RETURNS: 378 * 0 on success, -errno on failure. 379 */ 380static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) 381{ 382 int *old = NULL, *new = NULL; 383 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); 384 unsigned long flags; 385 386 new = pcpu_mem_alloc(new_size); 387 if (!new) 388 return -ENOMEM; 389 390 /* acquire pcpu_lock and switch to new area map */ 391 spin_lock_irqsave(&pcpu_lock, flags); 392 393 if (new_alloc <= chunk->map_alloc) 394 goto out_unlock; 395 396 old_size = chunk->map_alloc * sizeof(chunk->map[0]); 397 old = chunk->map; 398 399 memcpy(new, old, old_size); 400 401 chunk->map_alloc = new_alloc; 402 chunk->map = new; 403 new = NULL; 404 405out_unlock: 406 spin_unlock_irqrestore(&pcpu_lock, flags); 407 408 /* 409 * pcpu_mem_free() might end up calling vfree() which uses 410 * IRQ-unsafe lock and thus can't be called under pcpu_lock. 411 */ 412 pcpu_mem_free(old, old_size); 413 pcpu_mem_free(new, new_size); 414 415 return 0; 416} 417 418/** 419 * pcpu_split_block - split a map block 420 * @chunk: chunk of interest 421 * @i: index of map block to split 422 * @head: head size in bytes (can be 0) 423 * @tail: tail size in bytes (can be 0) 424 * 425 * Split the @i'th map block into two or three blocks. If @head is 426 * non-zero, @head bytes block is inserted before block @i moving it 427 * to @i+1 and reducing its size by @head bytes. 428 * 429 * If @tail is non-zero, the target block, which can be @i or @i+1 430 * depending on @head, is reduced by @tail bytes and @tail byte block 431 * is inserted after the target block. 432 * 433 * @chunk->map must have enough free slots to accomodate the split. 434 * 435 * CONTEXT: 436 * pcpu_lock. 437 */ 438static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 439 int head, int tail) 440{ 441 int nr_extra = !!head + !!tail; 442 443 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 444 445 /* insert new subblocks */ 446 memmove(&chunk->map[i + nr_extra], &chunk->map[i], 447 sizeof(chunk->map[0]) * (chunk->map_used - i)); 448 chunk->map_used += nr_extra; 449 450 if (head) { 451 chunk->map[i + 1] = chunk->map[i] - head; 452 chunk->map[i++] = head; 453 } 454 if (tail) { 455 chunk->map[i++] -= tail; 456 chunk->map[i] = tail; 457 } 458} 459 460/** 461 * pcpu_alloc_area - allocate area from a pcpu_chunk 462 * @chunk: chunk of interest 463 * @size: wanted size in bytes 464 * @align: wanted align 465 * 466 * Try to allocate @size bytes area aligned at @align from @chunk. 467 * Note that this function only allocates the offset. It doesn't 468 * populate or map the area. 469 * 470 * @chunk->map must have at least two free slots. 471 * 472 * CONTEXT: 473 * pcpu_lock. 474 * 475 * RETURNS: 476 * Allocated offset in @chunk on success, -1 if no matching area is 477 * found. 478 */ 479static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 480{ 481 int oslot = pcpu_chunk_slot(chunk); 482 int max_contig = 0; 483 int i, off; 484 485 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 486 bool is_last = i + 1 == chunk->map_used; 487 int head, tail; 488 489 /* extra for alignment requirement */ 490 head = ALIGN(off, align) - off; 491 BUG_ON(i == 0 && head != 0); 492 493 if (chunk->map[i] < 0) 494 continue; 495 if (chunk->map[i] < head + size) { 496 max_contig = max(chunk->map[i], max_contig); 497 continue; 498 } 499 500 /* 501 * If head is small or the previous block is free, 502 * merge'em. Note that 'small' is defined as smaller 503 * than sizeof(int), which is very small but isn't too 504 * uncommon for percpu allocations. 505 */ 506 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 507 if (chunk->map[i - 1] > 0) 508 chunk->map[i - 1] += head; 509 else { 510 chunk->map[i - 1] -= head; 511 chunk->free_size -= head; 512 } 513 chunk->map[i] -= head; 514 off += head; 515 head = 0; 516 } 517 518 /* if tail is small, just keep it around */ 519 tail = chunk->map[i] - head - size; 520 if (tail < sizeof(int)) 521 tail = 0; 522 523 /* split if warranted */ 524 if (head || tail) { 525 pcpu_split_block(chunk, i, head, tail); 526 if (head) { 527 i++; 528 off += head; 529 max_contig = max(chunk->map[i - 1], max_contig); 530 } 531 if (tail) 532 max_contig = max(chunk->map[i + 1], max_contig); 533 } 534 535 /* update hint and mark allocated */ 536 if (is_last) 537 chunk->contig_hint = max_contig; /* fully scanned */ 538 else 539 chunk->contig_hint = max(chunk->contig_hint, 540 max_contig); 541 542 chunk->free_size -= chunk->map[i]; 543 chunk->map[i] = -chunk->map[i]; 544 545 pcpu_chunk_relocate(chunk, oslot); 546 return off; 547 } 548 549 chunk->contig_hint = max_contig; /* fully scanned */ 550 pcpu_chunk_relocate(chunk, oslot); 551 552 /* tell the upper layer that this chunk has no matching area */ 553 return -1; 554} 555 556/** 557 * pcpu_free_area - free area to a pcpu_chunk 558 * @chunk: chunk of interest 559 * @freeme: offset of area to free 560 * 561 * Free area starting from @freeme to @chunk. Note that this function 562 * only modifies the allocation map. It doesn't depopulate or unmap 563 * the area. 564 * 565 * CONTEXT: 566 * pcpu_lock. 567 */ 568static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 569{ 570 int oslot = pcpu_chunk_slot(chunk); 571 int i, off; 572 573 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 574 if (off == freeme) 575 break; 576 BUG_ON(off != freeme); 577 BUG_ON(chunk->map[i] > 0); 578 579 chunk->map[i] = -chunk->map[i]; 580 chunk->free_size += chunk->map[i]; 581 582 /* merge with previous? */ 583 if (i > 0 && chunk->map[i - 1] >= 0) { 584 chunk->map[i - 1] += chunk->map[i]; 585 chunk->map_used--; 586 memmove(&chunk->map[i], &chunk->map[i + 1], 587 (chunk->map_used - i) * sizeof(chunk->map[0])); 588 i--; 589 } 590 /* merge with next? */ 591 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 592 chunk->map[i] += chunk->map[i + 1]; 593 chunk->map_used--; 594 memmove(&chunk->map[i + 1], &chunk->map[i + 2], 595 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 596 } 597 598 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 599 pcpu_chunk_relocate(chunk, oslot); 600} 601 602static struct pcpu_chunk *pcpu_alloc_chunk(void) 603{ 604 struct pcpu_chunk *chunk; 605 606 chunk = pcpu_mem_alloc(pcpu_chunk_struct_size); 607 if (!chunk) 608 return NULL; 609 610 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 611 if (!chunk->map) { 612 kfree(chunk); 613 return NULL; 614 } 615 616 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 617 chunk->map[chunk->map_used++] = pcpu_unit_size; 618 619 INIT_LIST_HEAD(&chunk->list); 620 chunk->free_size = pcpu_unit_size; 621 chunk->contig_hint = pcpu_unit_size; 622 623 return chunk; 624} 625 626static void pcpu_free_chunk(struct pcpu_chunk *chunk) 627{ 628 if (!chunk) 629 return; 630 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 631 kfree(chunk); 632} 633 634/* 635 * Chunk management implementation. 636 * 637 * To allow different implementations, chunk alloc/free and 638 * [de]population are implemented in a separate file which is pulled 639 * into this file and compiled together. The following functions 640 * should be implemented. 641 * 642 * pcpu_populate_chunk - populate the specified range of a chunk 643 * pcpu_depopulate_chunk - depopulate the specified range of a chunk 644 * pcpu_create_chunk - create a new chunk 645 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 646 * pcpu_addr_to_page - translate address to physical address 647 * pcpu_verify_alloc_info - check alloc_info is acceptable during init 648 */ 649static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 650static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 651static struct pcpu_chunk *pcpu_create_chunk(void); 652static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 653static struct page *pcpu_addr_to_page(void *addr); 654static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 655 656#ifdef CONFIG_NEED_PER_CPU_KM 657#include "percpu-km.c" 658#else 659#include "percpu-vm.c" 660#endif 661 662/** 663 * pcpu_chunk_addr_search - determine chunk containing specified address 664 * @addr: address for which the chunk needs to be determined. 665 * 666 * RETURNS: 667 * The address of the found chunk. 668 */ 669static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 670{ 671 /* is it in the first chunk? */ 672 if (pcpu_addr_in_first_chunk(addr)) { 673 /* is it in the reserved area? */ 674 if (pcpu_addr_in_reserved_chunk(addr)) 675 return pcpu_reserved_chunk; 676 return pcpu_first_chunk; 677 } 678 679 /* 680 * The address is relative to unit0 which might be unused and 681 * thus unmapped. Offset the address to the unit space of the 682 * current processor before looking it up in the vmalloc 683 * space. Note that any possible cpu id can be used here, so 684 * there's no need to worry about preemption or cpu hotplug. 685 */ 686 addr += pcpu_unit_offsets[raw_smp_processor_id()]; 687 return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 688} 689 690/** 691 * pcpu_alloc - the percpu allocator 692 * @size: size of area to allocate in bytes 693 * @align: alignment of area (max PAGE_SIZE) 694 * @reserved: allocate from the reserved chunk if available 695 * 696 * Allocate percpu area of @size bytes aligned at @align. 697 * 698 * CONTEXT: 699 * Does GFP_KERNEL allocation. 700 * 701 * RETURNS: 702 * Percpu pointer to the allocated area on success, NULL on failure. 703 */ 704static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved) 705{ 706 static int warn_limit = 10; 707 struct pcpu_chunk *chunk; 708 const char *err; 709 int slot, off, new_alloc; 710 unsigned long flags; 711 712 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 713 WARN(true, "illegal size (%zu) or align (%zu) for " 714 "percpu allocation\n", size, align); 715 return NULL; 716 } 717 718 mutex_lock(&pcpu_alloc_mutex); 719 spin_lock_irqsave(&pcpu_lock, flags); 720 721 /* serve reserved allocations from the reserved chunk if available */ 722 if (reserved && pcpu_reserved_chunk) { 723 chunk = pcpu_reserved_chunk; 724 725 if (size > chunk->contig_hint) { 726 err = "alloc from reserved chunk failed"; 727 goto fail_unlock; 728 } 729 730 while ((new_alloc = pcpu_need_to_extend(chunk))) { 731 spin_unlock_irqrestore(&pcpu_lock, flags); 732 if (pcpu_extend_area_map(chunk, new_alloc) < 0) { 733 err = "failed to extend area map of reserved chunk"; 734 goto fail_unlock_mutex; 735 } 736 spin_lock_irqsave(&pcpu_lock, flags); 737 } 738 739 off = pcpu_alloc_area(chunk, size, align); 740 if (off >= 0) 741 goto area_found; 742 743 err = "alloc from reserved chunk failed"; 744 goto fail_unlock; 745 } 746 747restart: 748 /* search through normal chunks */ 749 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 750 list_for_each_entry(chunk, &pcpu_slot[slot], list) { 751 if (size > chunk->contig_hint) 752 continue; 753 754 new_alloc = pcpu_need_to_extend(chunk); 755 if (new_alloc) { 756 spin_unlock_irqrestore(&pcpu_lock, flags); 757 if (pcpu_extend_area_map(chunk, 758 new_alloc) < 0) { 759 err = "failed to extend area map"; 760 goto fail_unlock_mutex; 761 } 762 spin_lock_irqsave(&pcpu_lock, flags); 763 /* 764 * pcpu_lock has been dropped, need to 765 * restart cpu_slot list walking. 766 */ 767 goto restart; 768 } 769 770 off = pcpu_alloc_area(chunk, size, align); 771 if (off >= 0) 772 goto area_found; 773 } 774 } 775 776 /* hmmm... no space left, create a new chunk */ 777 spin_unlock_irqrestore(&pcpu_lock, flags); 778 779 chunk = pcpu_create_chunk(); 780 if (!chunk) { 781 err = "failed to allocate new chunk"; 782 goto fail_unlock_mutex; 783 } 784 785 spin_lock_irqsave(&pcpu_lock, flags); 786 pcpu_chunk_relocate(chunk, -1); 787 goto restart; 788 789area_found: 790 spin_unlock_irqrestore(&pcpu_lock, flags); 791 792 /* populate, map and clear the area */ 793 if (pcpu_populate_chunk(chunk, off, size)) { 794 spin_lock_irqsave(&pcpu_lock, flags); 795 pcpu_free_area(chunk, off); 796 err = "failed to populate"; 797 goto fail_unlock; 798 } 799 800 mutex_unlock(&pcpu_alloc_mutex); 801 802 /* return address relative to base address */ 803 return __addr_to_pcpu_ptr(chunk->base_addr + off); 804 805fail_unlock: 806 spin_unlock_irqrestore(&pcpu_lock, flags); 807fail_unlock_mutex: 808 mutex_unlock(&pcpu_alloc_mutex); 809 if (warn_limit) { 810 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " 811 "%s\n", size, align, err); 812 dump_stack(); 813 if (!--warn_limit) 814 pr_info("PERCPU: limit reached, disable warning\n"); 815 } 816 return NULL; 817} 818 819/** 820 * __alloc_percpu - allocate dynamic percpu area 821 * @size: size of area to allocate in bytes 822 * @align: alignment of area (max PAGE_SIZE) 823 * 824 * Allocate percpu area of @size bytes aligned at @align. Might 825 * sleep. Might trigger writeouts. 826 * 827 * CONTEXT: 828 * Does GFP_KERNEL allocation. 829 * 830 * RETURNS: 831 * Percpu pointer to the allocated area on success, NULL on failure. 832 */ 833void __percpu *__alloc_percpu(size_t size, size_t align) 834{ 835 return pcpu_alloc(size, align, false); 836} 837EXPORT_SYMBOL_GPL(__alloc_percpu); 838 839/** 840 * __alloc_reserved_percpu - allocate reserved percpu area 841 * @size: size of area to allocate in bytes 842 * @align: alignment of area (max PAGE_SIZE) 843 * 844 * Allocate percpu area of @size bytes aligned at @align from reserved 845 * percpu area if arch has set it up; otherwise, allocation is served 846 * from the same dynamic area. Might sleep. Might trigger writeouts. 847 * 848 * CONTEXT: 849 * Does GFP_KERNEL allocation. 850 * 851 * RETURNS: 852 * Percpu pointer to the allocated area on success, NULL on failure. 853 */ 854void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 855{ 856 return pcpu_alloc(size, align, true); 857} 858 859/** 860 * pcpu_reclaim - reclaim fully free chunks, workqueue function 861 * @work: unused 862 * 863 * Reclaim all fully free chunks except for the first one. 864 * 865 * CONTEXT: 866 * workqueue context. 867 */ 868static void pcpu_reclaim(struct work_struct *work) 869{ 870 LIST_HEAD(todo); 871 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 872 struct pcpu_chunk *chunk, *next; 873 874 mutex_lock(&pcpu_alloc_mutex); 875 spin_lock_irq(&pcpu_lock); 876 877 list_for_each_entry_safe(chunk, next, head, list) { 878 WARN_ON(chunk->immutable); 879 880 /* spare the first one */ 881 if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 882 continue; 883 884 list_move(&chunk->list, &todo); 885 } 886 887 spin_unlock_irq(&pcpu_lock); 888 889 list_for_each_entry_safe(chunk, next, &todo, list) { 890 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 891 pcpu_destroy_chunk(chunk); 892 } 893 894 mutex_unlock(&pcpu_alloc_mutex); 895} 896 897/** 898 * free_percpu - free percpu area 899 * @ptr: pointer to area to free 900 * 901 * Free percpu area @ptr. 902 * 903 * CONTEXT: 904 * Can be called from atomic context. 905 */ 906void free_percpu(void __percpu *ptr) 907{ 908 void *addr; 909 struct pcpu_chunk *chunk; 910 unsigned long flags; 911 int off; 912 913 if (!ptr) 914 return; 915 916 addr = __pcpu_ptr_to_addr(ptr); 917 918 spin_lock_irqsave(&pcpu_lock, flags); 919 920 chunk = pcpu_chunk_addr_search(addr); 921 off = addr - chunk->base_addr; 922 923 pcpu_free_area(chunk, off); 924 925 /* if there are more than one fully free chunks, wake up grim reaper */ 926 if (chunk->free_size == pcpu_unit_size) { 927 struct pcpu_chunk *pos; 928 929 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 930 if (pos != chunk) { 931 schedule_work(&pcpu_reclaim_work); 932 break; 933 } 934 } 935 936 spin_unlock_irqrestore(&pcpu_lock, flags); 937} 938EXPORT_SYMBOL_GPL(free_percpu); 939 940/** 941 * is_kernel_percpu_address - test whether address is from static percpu area 942 * @addr: address to test 943 * 944 * Test whether @addr belongs to in-kernel static percpu area. Module 945 * static percpu areas are not considered. For those, use 946 * is_module_percpu_address(). 947 * 948 * RETURNS: 949 * %true if @addr is from in-kernel static percpu area, %false otherwise. 950 */ 951bool is_kernel_percpu_address(unsigned long addr) 952{ 953 const size_t static_size = __per_cpu_end - __per_cpu_start; 954 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 955 unsigned int cpu; 956 957 for_each_possible_cpu(cpu) { 958 void *start = per_cpu_ptr(base, cpu); 959 960 if ((void *)addr >= start && (void *)addr < start + static_size) 961 return true; 962 } 963 return false; 964} 965 966/** 967 * per_cpu_ptr_to_phys - convert translated percpu address to physical address 968 * @addr: the address to be converted to physical address 969 * 970 * Given @addr which is dereferenceable address obtained via one of 971 * percpu access macros, this function translates it into its physical 972 * address. The caller is responsible for ensuring @addr stays valid 973 * until this function finishes. 974 * 975 * RETURNS: 976 * The physical address for @addr. 977 */ 978phys_addr_t per_cpu_ptr_to_phys(void *addr) 979{ 980 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 981 bool in_first_chunk = false; 982 unsigned long first_start, first_end; 983 unsigned int cpu; 984 985 /* 986 * The following test on first_start/end isn't strictly 987 * necessary but will speed up lookups of addresses which 988 * aren't in the first chunk. 989 */ 990 first_start = pcpu_chunk_addr(pcpu_first_chunk, pcpu_first_unit_cpu, 0); 991 first_end = pcpu_chunk_addr(pcpu_first_chunk, pcpu_last_unit_cpu, 992 pcpu_unit_pages); 993 if ((unsigned long)addr >= first_start && 994 (unsigned long)addr < first_end) { 995 for_each_possible_cpu(cpu) { 996 void *start = per_cpu_ptr(base, cpu); 997 998 if (addr >= start && addr < start + pcpu_unit_size) { 999 in_first_chunk = true; 1000 break; 1001 } 1002 } 1003 } 1004 1005 if (in_first_chunk) { 1006 if ((unsigned long)addr < VMALLOC_START || 1007 (unsigned long)addr >= VMALLOC_END) 1008 return __pa(addr); 1009 else 1010 return page_to_phys(vmalloc_to_page(addr)); 1011 } else 1012 return page_to_phys(pcpu_addr_to_page(addr)); 1013} 1014 1015/** 1016 * pcpu_alloc_alloc_info - allocate percpu allocation info 1017 * @nr_groups: the number of groups 1018 * @nr_units: the number of units 1019 * 1020 * Allocate ai which is large enough for @nr_groups groups containing 1021 * @nr_units units. The returned ai's groups[0].cpu_map points to the 1022 * cpu_map array which is long enough for @nr_units and filled with 1023 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1024 * pointer of other groups. 1025 * 1026 * RETURNS: 1027 * Pointer to the allocated pcpu_alloc_info on success, NULL on 1028 * failure. 1029 */ 1030struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1031 int nr_units) 1032{ 1033 struct pcpu_alloc_info *ai; 1034 size_t base_size, ai_size; 1035 void *ptr; 1036 int unit; 1037 1038 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1039 __alignof__(ai->groups[0].cpu_map[0])); 1040 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1041 1042 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1043 if (!ptr) 1044 return NULL; 1045 ai = ptr; 1046 ptr += base_size; 1047 1048 ai->groups[0].cpu_map = ptr; 1049 1050 for (unit = 0; unit < nr_units; unit++) 1051 ai->groups[0].cpu_map[unit] = NR_CPUS; 1052 1053 ai->nr_groups = nr_groups; 1054 ai->__ai_size = PFN_ALIGN(ai_size); 1055 1056 return ai; 1057} 1058 1059/** 1060 * pcpu_free_alloc_info - free percpu allocation info 1061 * @ai: pcpu_alloc_info to free 1062 * 1063 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1064 */ 1065void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1066{ 1067 free_bootmem(__pa(ai), ai->__ai_size); 1068} 1069 1070/** 1071 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1072 * @reserved_size: the size of reserved percpu area in bytes 1073 * @dyn_size: minimum free size for dynamic allocation in bytes 1074 * @atom_size: allocation atom size 1075 * @cpu_distance_fn: callback to determine distance between cpus, optional 1076 * 1077 * This function determines grouping of units, their mappings to cpus 1078 * and other parameters considering needed percpu size, allocation 1079 * atom size and distances between CPUs. 1080 * 1081 * Groups are always mutliples of atom size and CPUs which are of 1082 * LOCAL_DISTANCE both ways are grouped together and share space for 1083 * units in the same group. The returned configuration is guaranteed 1084 * to have CPUs on different nodes on different groups and >=75% usage 1085 * of allocated virtual address space. 1086 * 1087 * RETURNS: 1088 * On success, pointer to the new allocation_info is returned. On 1089 * failure, ERR_PTR value is returned. 1090 */ 1091static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1092 size_t reserved_size, size_t dyn_size, 1093 size_t atom_size, 1094 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1095{ 1096 static int group_map[NR_CPUS] __initdata; 1097 static int group_cnt[NR_CPUS] __initdata; 1098 const size_t static_size = __per_cpu_end - __per_cpu_start; 1099 int nr_groups = 1, nr_units = 0; 1100 size_t size_sum, min_unit_size, alloc_size; 1101 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1102 int last_allocs, group, unit; 1103 unsigned int cpu, tcpu; 1104 struct pcpu_alloc_info *ai; 1105 unsigned int *cpu_map; 1106 1107 /* this function may be called multiple times */ 1108 memset(group_map, 0, sizeof(group_map)); 1109 memset(group_cnt, 0, sizeof(group_cnt)); 1110 1111 /* calculate size_sum and ensure dyn_size is enough for early alloc */ 1112 size_sum = PFN_ALIGN(static_size + reserved_size + 1113 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 1114 dyn_size = size_sum - static_size - reserved_size; 1115 1116 /* 1117 * Determine min_unit_size, alloc_size and max_upa such that 1118 * alloc_size is multiple of atom_size and is the smallest 1119 * which can accomodate 4k aligned segments which are equal to 1120 * or larger than min_unit_size. 1121 */ 1122 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1123 1124 alloc_size = roundup(min_unit_size, atom_size); 1125 upa = alloc_size / min_unit_size; 1126 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1127 upa--; 1128 max_upa = upa; 1129 1130 /* group cpus according to their proximity */ 1131 for_each_possible_cpu(cpu) { 1132 group = 0; 1133 next_group: 1134 for_each_possible_cpu(tcpu) { 1135 if (cpu == tcpu) 1136 break; 1137 if (group_map[tcpu] == group && cpu_distance_fn && 1138 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1139 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1140 group++; 1141 nr_groups = max(nr_groups, group + 1); 1142 goto next_group; 1143 } 1144 } 1145 group_map[cpu] = group; 1146 group_cnt[group]++; 1147 } 1148 1149 /* 1150 * Expand unit size until address space usage goes over 75% 1151 * and then as much as possible without using more address 1152 * space. 1153 */ 1154 last_allocs = INT_MAX; 1155 for (upa = max_upa; upa; upa--) { 1156 int allocs = 0, wasted = 0; 1157 1158 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1159 continue; 1160 1161 for (group = 0; group < nr_groups; group++) { 1162 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1163 allocs += this_allocs; 1164 wasted += this_allocs * upa - group_cnt[group]; 1165 } 1166 1167 /* 1168 * Don't accept if wastage is over 1/3. The 1169 * greater-than comparison ensures upa==1 always 1170 * passes the following check. 1171 */ 1172 if (wasted > num_possible_cpus() / 3) 1173 continue; 1174 1175 /* and then don't consume more memory */ 1176 if (allocs > last_allocs) 1177 break; 1178 last_allocs = allocs; 1179 best_upa = upa; 1180 } 1181 upa = best_upa; 1182 1183 /* allocate and fill alloc_info */ 1184 for (group = 0; group < nr_groups; group++) 1185 nr_units += roundup(group_cnt[group], upa); 1186 1187 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1188 if (!ai) 1189 return ERR_PTR(-ENOMEM); 1190 cpu_map = ai->groups[0].cpu_map; 1191 1192 for (group = 0; group < nr_groups; group++) { 1193 ai->groups[group].cpu_map = cpu_map; 1194 cpu_map += roundup(group_cnt[group], upa); 1195 } 1196 1197 ai->static_size = static_size; 1198 ai->reserved_size = reserved_size; 1199 ai->dyn_size = dyn_size; 1200 ai->unit_size = alloc_size / upa; 1201 ai->atom_size = atom_size; 1202 ai->alloc_size = alloc_size; 1203 1204 for (group = 0, unit = 0; group_cnt[group]; group++) { 1205 struct pcpu_group_info *gi = &ai->groups[group]; 1206 1207 /* 1208 * Initialize base_offset as if all groups are located 1209 * back-to-back. The caller should update this to 1210 * reflect actual allocation. 1211 */ 1212 gi->base_offset = unit * ai->unit_size; 1213 1214 for_each_possible_cpu(cpu) 1215 if (group_map[cpu] == group) 1216 gi->cpu_map[gi->nr_units++] = cpu; 1217 gi->nr_units = roundup(gi->nr_units, upa); 1218 unit += gi->nr_units; 1219 } 1220 BUG_ON(unit != nr_units); 1221 1222 return ai; 1223} 1224 1225/** 1226 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1227 * @lvl: loglevel 1228 * @ai: allocation info to dump 1229 * 1230 * Print out information about @ai using loglevel @lvl. 1231 */ 1232static void pcpu_dump_alloc_info(const char *lvl, 1233 const struct pcpu_alloc_info *ai) 1234{ 1235 int group_width = 1, cpu_width = 1, width; 1236 char * empty_str = "--------"; 1237 int alloc = 0, alloc_end = 0; 1238 int group, v; 1239 int upa, apl; /* units per alloc, allocs per line */ 1240 1241 v = ai->nr_groups; 1242 while (v /= 10) 1243 group_width++; 1244 1245 v = num_possible_cpus(); 1246 while (v /= 10) 1247 cpu_width++; 1248 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1249 1250 upa = ai->alloc_size / ai->unit_size; 1251 width = upa * (cpu_width + 1) + group_width + 3; 1252 apl = rounddown_pow_of_two(max(60 / width, 1)); 1253 1254 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1255 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1256 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1257 1258 for (group = 0; group < ai->nr_groups; group++) { 1259 const struct pcpu_group_info *gi = &ai->groups[group]; 1260 int unit = 0, unit_end = 0; 1261 1262 BUG_ON(gi->nr_units % upa); 1263 for (alloc_end += gi->nr_units / upa; 1264 alloc < alloc_end; alloc++) { 1265 if (!(alloc % apl)) { 1266 printk("\n"); 1267 printk("%spcpu-alloc: ", lvl); 1268 } 1269 printk("[%0*d] ", group_width, group); 1270 1271 for (unit_end += upa; unit < unit_end; unit++) 1272 if (gi->cpu_map[unit] != NR_CPUS) 1273 printk("%0*d ", cpu_width, 1274 gi->cpu_map[unit]); 1275 else 1276 printk("%s ", empty_str); 1277 } 1278 } 1279 printk("\n"); 1280} 1281 1282/** 1283 * pcpu_setup_first_chunk - initialize the first percpu chunk 1284 * @ai: pcpu_alloc_info describing how to percpu area is shaped 1285 * @base_addr: mapped address 1286 * 1287 * Initialize the first percpu chunk which contains the kernel static 1288 * perpcu area. This function is to be called from arch percpu area 1289 * setup path. 1290 * 1291 * @ai contains all information necessary to initialize the first 1292 * chunk and prime the dynamic percpu allocator. 1293 * 1294 * @ai->static_size is the size of static percpu area. 1295 * 1296 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1297 * reserve after the static area in the first chunk. This reserves 1298 * the first chunk such that it's available only through reserved 1299 * percpu allocation. This is primarily used to serve module percpu 1300 * static areas on architectures where the addressing model has 1301 * limited offset range for symbol relocations to guarantee module 1302 * percpu symbols fall inside the relocatable range. 1303 * 1304 * @ai->dyn_size determines the number of bytes available for dynamic 1305 * allocation in the first chunk. The area between @ai->static_size + 1306 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 1307 * 1308 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1309 * and equal to or larger than @ai->static_size + @ai->reserved_size + 1310 * @ai->dyn_size. 1311 * 1312 * @ai->atom_size is the allocation atom size and used as alignment 1313 * for vm areas. 1314 * 1315 * @ai->alloc_size is the allocation size and always multiple of 1316 * @ai->atom_size. This is larger than @ai->atom_size if 1317 * @ai->unit_size is larger than @ai->atom_size. 1318 * 1319 * @ai->nr_groups and @ai->groups describe virtual memory layout of 1320 * percpu areas. Units which should be colocated are put into the 1321 * same group. Dynamic VM areas will be allocated according to these 1322 * groupings. If @ai->nr_groups is zero, a single group containing 1323 * all units is assumed. 1324 * 1325 * The caller should have mapped the first chunk at @base_addr and 1326 * copied static data to each unit. 1327 * 1328 * If the first chunk ends up with both reserved and dynamic areas, it 1329 * is served by two chunks - one to serve the core static and reserved 1330 * areas and the other for the dynamic area. They share the same vm 1331 * and page map but uses different area allocation map to stay away 1332 * from each other. The latter chunk is circulated in the chunk slots 1333 * and available for dynamic allocation like any other chunks. 1334 * 1335 * RETURNS: 1336 * 0 on success, -errno on failure. 1337 */ 1338int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1339 void *base_addr) 1340{ 1341 static char cpus_buf[4096] __initdata; 1342 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1343 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1344 size_t dyn_size = ai->dyn_size; 1345 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1346 struct pcpu_chunk *schunk, *dchunk = NULL; 1347 unsigned long *group_offsets; 1348 size_t *group_sizes; 1349 unsigned long *unit_off; 1350 unsigned int cpu; 1351 int *unit_map; 1352 int group, unit, i; 1353 1354 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); 1355 1356#define PCPU_SETUP_BUG_ON(cond) do { \ 1357 if (unlikely(cond)) { \ 1358 pr_emerg("PERCPU: failed to initialize, %s", #cond); \ 1359 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ 1360 pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1361 BUG(); \ 1362 } \ 1363} while (0) 1364 1365 /* sanity checks */ 1366 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1367 PCPU_SETUP_BUG_ON(!ai->static_size); 1368 PCPU_SETUP_BUG_ON(!base_addr); 1369 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1370 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); 1371 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1372 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 1373 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 1374 1375 /* process group information and build config tables accordingly */ 1376 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 1377 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1378 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1379 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 1380 1381 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1382 unit_map[cpu] = UINT_MAX; 1383 pcpu_first_unit_cpu = NR_CPUS; 1384 1385 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1386 const struct pcpu_group_info *gi = &ai->groups[group]; 1387 1388 group_offsets[group] = gi->base_offset; 1389 group_sizes[group] = gi->nr_units * ai->unit_size; 1390 1391 for (i = 0; i < gi->nr_units; i++) { 1392 cpu = gi->cpu_map[i]; 1393 if (cpu == NR_CPUS) 1394 continue; 1395 1396 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); 1397 PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1398 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1399 1400 unit_map[cpu] = unit + i; 1401 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1402 1403 if (pcpu_first_unit_cpu == NR_CPUS) 1404 pcpu_first_unit_cpu = cpu; 1405 pcpu_last_unit_cpu = cpu; 1406 } 1407 } 1408 pcpu_nr_units = unit; 1409 1410 for_each_possible_cpu(cpu) 1411 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1412 1413 /* we're done parsing the input, undefine BUG macro and dump config */ 1414#undef PCPU_SETUP_BUG_ON 1415 pcpu_dump_alloc_info(KERN_INFO, ai); 1416 1417 pcpu_nr_groups = ai->nr_groups; 1418 pcpu_group_offsets = group_offsets; 1419 pcpu_group_sizes = group_sizes; 1420 pcpu_unit_map = unit_map; 1421 pcpu_unit_offsets = unit_off; 1422 1423 /* determine basic parameters */ 1424 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1425 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1426 pcpu_atom_size = ai->atom_size; 1427 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1428 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1429 1430 /* 1431 * Allocate chunk slots. The additional last slot is for 1432 * empty chunks. 1433 */ 1434 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1435 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1436 for (i = 0; i < pcpu_nr_slots; i++) 1437 INIT_LIST_HEAD(&pcpu_slot[i]); 1438 1439 /* 1440 * Initialize static chunk. If reserved_size is zero, the 1441 * static chunk covers static area + dynamic allocation area 1442 * in the first chunk. If reserved_size is not zero, it 1443 * covers static area + reserved area (mostly used for module 1444 * static percpu allocation). 1445 */ 1446 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1447 INIT_LIST_HEAD(&schunk->list); 1448 schunk->base_addr = base_addr; 1449 schunk->map = smap; 1450 schunk->map_alloc = ARRAY_SIZE(smap); 1451 schunk->immutable = true; 1452 bitmap_fill(schunk->populated, pcpu_unit_pages); 1453 1454 if (ai->reserved_size) { 1455 schunk->free_size = ai->reserved_size; 1456 pcpu_reserved_chunk = schunk; 1457 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1458 } else { 1459 schunk->free_size = dyn_size; 1460 dyn_size = 0; /* dynamic area covered */ 1461 } 1462 schunk->contig_hint = schunk->free_size; 1463 1464 schunk->map[schunk->map_used++] = -ai->static_size; 1465 if (schunk->free_size) 1466 schunk->map[schunk->map_used++] = schunk->free_size; 1467 1468 /* init dynamic chunk if necessary */ 1469 if (dyn_size) { 1470 dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1471 INIT_LIST_HEAD(&dchunk->list); 1472 dchunk->base_addr = base_addr; 1473 dchunk->map = dmap; 1474 dchunk->map_alloc = ARRAY_SIZE(dmap); 1475 dchunk->immutable = true; 1476 bitmap_fill(dchunk->populated, pcpu_unit_pages); 1477 1478 dchunk->contig_hint = dchunk->free_size = dyn_size; 1479 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1480 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1481 } 1482 1483 /* link the first chunk in */ 1484 pcpu_first_chunk = dchunk ?: schunk; 1485 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1486 1487 /* we're done */ 1488 pcpu_base_addr = base_addr; 1489 return 0; 1490} 1491 1492const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { 1493 [PCPU_FC_AUTO] = "auto", 1494 [PCPU_FC_EMBED] = "embed", 1495 [PCPU_FC_PAGE] = "page", 1496}; 1497 1498enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1499 1500static int __init percpu_alloc_setup(char *str) 1501{ 1502 if (0) 1503 /* nada */; 1504#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1505 else if (!strcmp(str, "embed")) 1506 pcpu_chosen_fc = PCPU_FC_EMBED; 1507#endif 1508#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1509 else if (!strcmp(str, "page")) 1510 pcpu_chosen_fc = PCPU_FC_PAGE; 1511#endif 1512 else 1513 pr_warning("PERCPU: unknown allocator %s specified\n", str); 1514 1515 return 0; 1516} 1517early_param("percpu_alloc", percpu_alloc_setup); 1518 1519#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 1520 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 1521/** 1522 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1523 * @reserved_size: the size of reserved percpu area in bytes 1524 * @dyn_size: minimum free size for dynamic allocation in bytes 1525 * @atom_size: allocation atom size 1526 * @cpu_distance_fn: callback to determine distance between cpus, optional 1527 * @alloc_fn: function to allocate percpu page 1528 * @free_fn: funtion to free percpu page 1529 * 1530 * This is a helper to ease setting up embedded first percpu chunk and 1531 * can be called where pcpu_setup_first_chunk() is expected. 1532 * 1533 * If this function is used to setup the first chunk, it is allocated 1534 * by calling @alloc_fn and used as-is without being mapped into 1535 * vmalloc area. Allocations are always whole multiples of @atom_size 1536 * aligned to @atom_size. 1537 * 1538 * This enables the first chunk to piggy back on the linear physical 1539 * mapping which often uses larger page size. Please note that this 1540 * can result in very sparse cpu->unit mapping on NUMA machines thus 1541 * requiring large vmalloc address space. Don't use this allocator if 1542 * vmalloc space is not orders of magnitude larger than distances 1543 * between node memory addresses (ie. 32bit NUMA machines). 1544 * 1545 * @dyn_size specifies the minimum dynamic area size. 1546 * 1547 * If the needed size is smaller than the minimum or specified unit 1548 * size, the leftover is returned using @free_fn. 1549 * 1550 * RETURNS: 1551 * 0 on success, -errno on failure. 1552 */ 1553int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 1554 size_t atom_size, 1555 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1556 pcpu_fc_alloc_fn_t alloc_fn, 1557 pcpu_fc_free_fn_t free_fn) 1558{ 1559 void *base = (void *)ULONG_MAX; 1560 void **areas = NULL; 1561 struct pcpu_alloc_info *ai; 1562 size_t size_sum, areas_size, max_distance; 1563 int group, i, rc; 1564 1565 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1566 cpu_distance_fn); 1567 if (IS_ERR(ai)) 1568 return PTR_ERR(ai); 1569 1570 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1571 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 1572 1573 areas = alloc_bootmem_nopanic(areas_size); 1574 if (!areas) { 1575 rc = -ENOMEM; 1576 goto out_free; 1577 } 1578 1579 /* allocate, copy and determine base address */ 1580 for (group = 0; group < ai->nr_groups; group++) { 1581 struct pcpu_group_info *gi = &ai->groups[group]; 1582 unsigned int cpu = NR_CPUS; 1583 void *ptr; 1584 1585 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1586 cpu = gi->cpu_map[i]; 1587 BUG_ON(cpu == NR_CPUS); 1588 1589 /* allocate space for the whole group */ 1590 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1591 if (!ptr) { 1592 rc = -ENOMEM; 1593 goto out_free_areas; 1594 } 1595 areas[group] = ptr; 1596 1597 base = min(ptr, base); 1598 1599 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1600 if (gi->cpu_map[i] == NR_CPUS) { 1601 /* unused unit, free whole */ 1602 free_fn(ptr, ai->unit_size); 1603 continue; 1604 } 1605 /* copy and return the unused part */ 1606 memcpy(ptr, __per_cpu_load, ai->static_size); 1607 free_fn(ptr + size_sum, ai->unit_size - size_sum); 1608 } 1609 } 1610 1611 /* base address is now known, determine group base offsets */ 1612 max_distance = 0; 1613 for (group = 0; group < ai->nr_groups; group++) { 1614 ai->groups[group].base_offset = areas[group] - base; 1615 max_distance = max_t(size_t, max_distance, 1616 ai->groups[group].base_offset); 1617 } 1618 max_distance += ai->unit_size; 1619 1620 /* warn if maximum distance is further than 75% of vmalloc space */ 1621 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { 1622 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " 1623 "space 0x%lx\n", 1624 max_distance, VMALLOC_END - VMALLOC_START); 1625#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1626 /* and fail if we have fallback */ 1627 rc = -EINVAL; 1628 goto out_free; 1629#endif 1630 } 1631 1632 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1633 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1634 ai->dyn_size, ai->unit_size); 1635 1636 rc = pcpu_setup_first_chunk(ai, base); 1637 goto out_free; 1638 1639out_free_areas: 1640 for (group = 0; group < ai->nr_groups; group++) 1641 free_fn(areas[group], 1642 ai->groups[group].nr_units * ai->unit_size); 1643out_free: 1644 pcpu_free_alloc_info(ai); 1645 if (areas) 1646 free_bootmem(__pa(areas), areas_size); 1647 return rc; 1648} 1649#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || 1650 !CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1651 1652#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1653/** 1654 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1655 * @reserved_size: the size of reserved percpu area in bytes 1656 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 1657 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE 1658 * @populate_pte_fn: function to populate pte 1659 * 1660 * This is a helper to ease setting up page-remapped first percpu 1661 * chunk and can be called where pcpu_setup_first_chunk() is expected. 1662 * 1663 * This is the basic allocator. Static percpu area is allocated 1664 * page-by-page into vmalloc area. 1665 * 1666 * RETURNS: 1667 * 0 on success, -errno on failure. 1668 */ 1669int __init pcpu_page_first_chunk(size_t reserved_size, 1670 pcpu_fc_alloc_fn_t alloc_fn, 1671 pcpu_fc_free_fn_t free_fn, 1672 pcpu_fc_populate_pte_fn_t populate_pte_fn) 1673{ 1674 static struct vm_struct vm; 1675 struct pcpu_alloc_info *ai; 1676 char psize_str[16]; 1677 int unit_pages; 1678 size_t pages_size; 1679 struct page **pages; 1680 int unit, i, j, rc; 1681 1682 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 1683 1684 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 1685 if (IS_ERR(ai)) 1686 return PTR_ERR(ai); 1687 BUG_ON(ai->nr_groups != 1); 1688 BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1689 1690 unit_pages = ai->unit_size >> PAGE_SHIFT; 1691 1692 /* unaligned allocations can't be freed, round up to page size */ 1693 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1694 sizeof(pages[0])); 1695 pages = alloc_bootmem(pages_size); 1696 1697 /* allocate pages */ 1698 j = 0; 1699 for (unit = 0; unit < num_possible_cpus(); unit++) 1700 for (i = 0; i < unit_pages; i++) { 1701 unsigned int cpu = ai->groups[0].cpu_map[unit]; 1702 void *ptr; 1703 1704 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1705 if (!ptr) { 1706 pr_warning("PERCPU: failed to allocate %s page " 1707 "for cpu%u\n", psize_str, cpu); 1708 goto enomem; 1709 } 1710 pages[j++] = virt_to_page(ptr); 1711 } 1712 1713 /* allocate vm area, map the pages and copy static data */ 1714 vm.flags = VM_ALLOC; 1715 vm.size = num_possible_cpus() * ai->unit_size; 1716 vm_area_register_early(&vm, PAGE_SIZE); 1717 1718 for (unit = 0; unit < num_possible_cpus(); unit++) { 1719 unsigned long unit_addr = 1720 (unsigned long)vm.addr + unit * ai->unit_size; 1721 1722 for (i = 0; i < unit_pages; i++) 1723 populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 1724 1725 /* pte already populated, the following shouldn't fail */ 1726 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1727 unit_pages); 1728 if (rc < 0) 1729 panic("failed to map percpu area, err=%d\n", rc); 1730 1731 1732 /* copy static data */ 1733 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 1734 } 1735 1736 /* we're ready, commit */ 1737 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1738 unit_pages, psize_str, vm.addr, ai->static_size, 1739 ai->reserved_size, ai->dyn_size); 1740 1741 rc = pcpu_setup_first_chunk(ai, vm.addr); 1742 goto out_free_ar; 1743 1744enomem: 1745 while (--j >= 0) 1746 free_fn(page_address(pages[j]), PAGE_SIZE); 1747 rc = -ENOMEM; 1748out_free_ar: 1749 free_bootmem(__pa(pages), pages_size); 1750 pcpu_free_alloc_info(ai); 1751 return rc; 1752} 1753#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ 1754 1755/* 1756 * Generic percpu area setup. 1757 * 1758 * The embedding helper is used because its behavior closely resembles 1759 * the original non-dynamic generic percpu area setup. This is 1760 * important because many archs have addressing restrictions and might 1761 * fail if the percpu area is located far away from the previous 1762 * location. As an added bonus, in non-NUMA cases, embedding is 1763 * generally a good idea TLB-wise because percpu area can piggy back 1764 * on the physical linear memory mapping which uses large page 1765 * mappings on applicable archs. 1766 */ 1767#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 1768unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 1769EXPORT_SYMBOL(__per_cpu_offset); 1770 1771static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 1772 size_t align) 1773{ 1774 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 1775} 1776 1777static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 1778{ 1779 free_bootmem(__pa(ptr), size); 1780} 1781 1782void __init setup_per_cpu_areas(void) 1783{ 1784 unsigned long delta; 1785 unsigned int cpu; 1786 int rc; 1787 1788 /* 1789 * Always reserve area for module percpu variables. That's 1790 * what the legacy allocator did. 1791 */ 1792 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 1793 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 1794 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 1795 if (rc < 0) 1796 panic("Failed to initialized percpu areas."); 1797 1798 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 1799 for_each_possible_cpu(cpu) 1800 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 1801} 1802#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1803 1804/* 1805 * First and reserved chunks are initialized with temporary allocation 1806 * map in initdata so that they can be used before slab is online. 1807 * This function is called after slab is brought up and replaces those 1808 * with properly allocated maps. 1809 */ 1810void __init percpu_init_late(void) 1811{ 1812 struct pcpu_chunk *target_chunks[] = 1813 { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; 1814 struct pcpu_chunk *chunk; 1815 unsigned long flags; 1816 int i; 1817 1818 for (i = 0; (chunk = target_chunks[i]); i++) { 1819 int *map; 1820 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); 1821 1822 BUILD_BUG_ON(size > PAGE_SIZE); 1823 1824 map = pcpu_mem_alloc(size); 1825 BUG_ON(!map); 1826 1827 spin_lock_irqsave(&pcpu_lock, flags); 1828 memcpy(map, chunk->map, size); 1829 chunk->map = map; 1830 spin_unlock_irqrestore(&pcpu_lock, flags); 1831 } 1832} 1833