FILE
percpu.c
File preview with syntax highlighting.
linux-6.19.14/mm/percpu.c
1// SPDX-License-Identifier: GPL-2.0-only 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 * Copyright (C) 2017 Facebook Inc. 9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org> 10 * 11 * The percpu allocator handles both static and dynamic areas. Percpu 12 * areas are allocated in chunks which are divided into units. There is 13 * a 1-to-1 mapping for units to possible cpus. These units are grouped 14 * based on NUMA properties of the machine. 15 * 16 * c0 c1 c2 17 * ------------------- ------------------- ------------ 18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 19 * ------------------- ...... ------------------- .... ------------ 20 * 21 * Allocation is done by offsets into a unit's address space. Ie., an 22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, 23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear 24 * and even sparse. Access is handled by configuring percpu base 25 * registers according to the cpu to unit mappings and offsetting the 26 * base address using pcpu_unit_size. 27 * 28 * There is special consideration for the first chunk which must handle 29 * the static percpu variables in the kernel image as allocation services 30 * are not online yet. In short, the first chunk is structured like so: 31 * 32 * <Static | [Reserved] | Dynamic> 33 * 34 * The static data is copied from the original section managed by the 35 * linker. The reserved section, if non-zero, primarily manages static 36 * percpu variables from kernel modules. Finally, the dynamic section 37 * takes care of normal allocations. 38 * 39 * The allocator organizes chunks into lists according to free size and 40 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT 41 * flag should be passed. All memcg-aware allocations are sharing one set 42 * of chunks and all unaccounted allocations and allocations performed 43 * by processes belonging to the root memory cgroup are using the second set. 44 * 45 * The allocator tries to allocate from the fullest chunk first. Each chunk 46 * is managed by a bitmap with metadata blocks. The allocation map is updated 47 * on every allocation and free to reflect the current state while the boundary 48 * map is only updated on allocation. Each metadata block contains 49 * information to help mitigate the need to iterate over large portions 50 * of the bitmap. The reverse mapping from page to chunk is stored in 51 * the page's index. Lastly, units are lazily backed and grow in unison. 52 * 53 * There is a unique conversion that goes on here between bytes and bits. 54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk 55 * tracks the number of pages it is responsible for in nr_pages. Helper 56 * functions are used to convert from between the bytes, bits, and blocks. 57 * All hints are managed in bits unless explicitly stated. 58 * 59 * To use this allocator, arch code should do the following: 60 * 61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 62 * regular address to percpu pointer and back if they need to be 63 * different from the default 64 * 65 * - use pcpu_setup_first_chunk() during percpu area initialization to 66 * setup the first chunk containing the kernel static percpu area 67 */ 68 69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 70 71#include <linux/bitmap.h> 72#include <linux/cpumask.h> 73#include <linux/memblock.h> 74#include <linux/err.h> 75#include <linux/list.h> 76#include <linux/log2.h> 77#include <linux/mm.h> 78#include <linux/module.h> 79#include <linux/mutex.h> 80#include <linux/percpu.h> 81#include <linux/pfn.h> 82#include <linux/slab.h> 83#include <linux/spinlock.h> 84#include <linux/vmalloc.h> 85#include <linux/workqueue.h> 86#include <linux/kmemleak.h> 87#include <linux/sched.h> 88#include <linux/sched/mm.h> 89#include <linux/memcontrol.h> 90 91#include <asm/cacheflush.h> 92#include <asm/sections.h> 93#include <asm/tlbflush.h> 94#include <asm/io.h> 95 96#define CREATE_TRACE_POINTS 97#include <trace/events/percpu.h> 98 99#include "percpu-internal.h" 100 101/* 102 * The slots are sorted by the size of the biggest continuous free area. 103 * 1-31 bytes share the same slot. 104 */ 105#define PCPU_SLOT_BASE_SHIFT 5 106/* chunks in slots below this are subject to being sidelined on failed alloc */ 107#define PCPU_SLOT_FAIL_THRESHOLD 3 108 109#define PCPU_EMPTY_POP_PAGES_LOW 2 110#define PCPU_EMPTY_POP_PAGES_HIGH 4 111 112#ifdef CONFIG_SMP 113/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 114#ifndef __addr_to_pcpu_ptr 115#define __addr_to_pcpu_ptr(addr) \ 116 (void __percpu *)((unsigned long)(addr) - \ 117 (unsigned long)pcpu_base_addr + \ 118 (unsigned long)__per_cpu_start) 119#endif 120#ifndef __pcpu_ptr_to_addr 121#define __pcpu_ptr_to_addr(ptr) \ 122 (void __force *)((unsigned long)(ptr) + \ 123 (unsigned long)pcpu_base_addr - \ 124 (unsigned long)__per_cpu_start) 125#endif 126#else /* CONFIG_SMP */ 127/* on UP, it's always identity mapped */ 128#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 129#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 130#endif /* CONFIG_SMP */ 131 132static int pcpu_unit_pages __ro_after_init; 133static int pcpu_unit_size __ro_after_init; 134static int pcpu_nr_units __ro_after_init; 135static int pcpu_atom_size __ro_after_init; 136int pcpu_nr_slots __ro_after_init; 137static int pcpu_free_slot __ro_after_init; 138int pcpu_sidelined_slot __ro_after_init; 139int pcpu_to_depopulate_slot __ro_after_init; 140static size_t pcpu_chunk_struct_size __ro_after_init; 141 142/* cpus with the lowest and highest unit addresses */ 143static unsigned int pcpu_low_unit_cpu __ro_after_init; 144static unsigned int pcpu_high_unit_cpu __ro_after_init; 145 146/* the address of the first chunk which starts with the kernel static area */ 147void *pcpu_base_addr __ro_after_init; 148 149static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ 150const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ 151 152/* group information, used for vm allocation */ 153static int pcpu_nr_groups __ro_after_init; 154static const unsigned long *pcpu_group_offsets __ro_after_init; 155static const size_t *pcpu_group_sizes __ro_after_init; 156 157/* 158 * The first chunk which always exists. Note that unlike other 159 * chunks, this one can be allocated and mapped in several different 160 * ways and thus often doesn't live in the vmalloc area. 161 */ 162struct pcpu_chunk *pcpu_first_chunk __ro_after_init; 163 164/* 165 * Optional reserved chunk. This chunk reserves part of the first 166 * chunk and serves it for reserved allocations. When the reserved 167 * region doesn't exist, the following variable is NULL. 168 */ 169struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; 170 171DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 172static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 173 174struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */ 175 176/* 177 * The number of empty populated pages, protected by pcpu_lock. 178 * The reserved chunk doesn't contribute to the count. 179 */ 180int pcpu_nr_empty_pop_pages; 181 182/* 183 * The number of populated pages in use by the allocator, protected by 184 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets 185 * allocated/deallocated, it is allocated/deallocated in all units of a chunk 186 * and increments/decrements this count by 1). 187 */ 188static unsigned long pcpu_nr_populated; 189 190/* 191 * Balance work is used to populate or destroy chunks asynchronously. We 192 * try to keep the number of populated free pages between 193 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 194 * empty chunk. 195 */ 196static void pcpu_balance_workfn(struct work_struct *work); 197static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 198static bool pcpu_async_enabled __read_mostly; 199static bool pcpu_atomic_alloc_failed; 200 201static void pcpu_schedule_balance_work(void) 202{ 203 if (pcpu_async_enabled) 204 schedule_work(&pcpu_balance_work); 205} 206 207/** 208 * pcpu_addr_in_chunk - check if the address is served from this chunk 209 * @chunk: chunk of interest 210 * @addr: percpu address 211 * 212 * RETURNS: 213 * True if the address is served from this chunk. 214 */ 215static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) 216{ 217 void *start_addr, *end_addr; 218 219 if (!chunk) 220 return false; 221 222 start_addr = chunk->base_addr + chunk->start_offset; 223 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - 224 chunk->end_offset; 225 226 return addr >= start_addr && addr < end_addr; 227} 228 229static int __pcpu_size_to_slot(int size) 230{ 231 int highbit = fls(size); /* size is in bytes */ 232 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 233} 234 235static int pcpu_size_to_slot(int size) 236{ 237 if (size == pcpu_unit_size) 238 return pcpu_free_slot; 239 return __pcpu_size_to_slot(size); 240} 241 242static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 243{ 244 const struct pcpu_block_md *chunk_md = &chunk->chunk_md; 245 246 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || 247 chunk_md->contig_hint == 0) 248 return 0; 249 250 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE); 251} 252 253/* set the pointer to a chunk in a page struct */ 254static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 255{ 256 page->private = (unsigned long)pcpu; 257} 258 259/* obtain pointer to a chunk from a page struct */ 260static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 261{ 262 return (struct pcpu_chunk *)page->private; 263} 264 265static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 266{ 267 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 268} 269 270static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 271{ 272 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 273} 274 275static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 276 unsigned int cpu, int page_idx) 277{ 278 return (unsigned long)chunk->base_addr + 279 pcpu_unit_page_offset(cpu, page_idx); 280} 281 282/* 283 * The following are helper functions to help access bitmaps and convert 284 * between bitmap offsets to address offsets. 285 */ 286static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 287{ 288 return chunk->alloc_map + 289 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 290} 291 292static unsigned long pcpu_off_to_block_index(int off) 293{ 294 return off / PCPU_BITMAP_BLOCK_BITS; 295} 296 297static unsigned long pcpu_off_to_block_off(int off) 298{ 299 return off & (PCPU_BITMAP_BLOCK_BITS - 1); 300} 301 302static unsigned long pcpu_block_off_to_off(int index, int off) 303{ 304 return index * PCPU_BITMAP_BLOCK_BITS + off; 305} 306 307/** 308 * pcpu_check_block_hint - check against the contig hint 309 * @block: block of interest 310 * @bits: size of allocation 311 * @align: alignment of area (max PAGE_SIZE) 312 * 313 * Check to see if the allocation can fit in the block's contig hint. 314 * Note, a chunk uses the same hints as a block so this can also check against 315 * the chunk's contig hint. 316 */ 317static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits, 318 size_t align) 319{ 320 int bit_off = ALIGN(block->contig_hint_start, align) - 321 block->contig_hint_start; 322 323 return bit_off + bits <= block->contig_hint; 324} 325 326/* 327 * pcpu_next_hint - determine which hint to use 328 * @block: block of interest 329 * @alloc_bits: size of allocation 330 * 331 * This determines if we should scan based on the scan_hint or first_free. 332 * In general, we want to scan from first_free to fulfill allocations by 333 * first fit. However, if we know a scan_hint at position scan_hint_start 334 * cannot fulfill an allocation, we can begin scanning from there knowing 335 * the contig_hint will be our fallback. 336 */ 337static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits) 338{ 339 /* 340 * The three conditions below determine if we can skip past the 341 * scan_hint. First, does the scan hint exist. Second, is the 342 * contig_hint after the scan_hint (possibly not true iff 343 * contig_hint == scan_hint). Third, is the allocation request 344 * larger than the scan_hint. 345 */ 346 if (block->scan_hint && 347 block->contig_hint_start > block->scan_hint_start && 348 alloc_bits > block->scan_hint) 349 return block->scan_hint_start + block->scan_hint; 350 351 return block->first_free; 352} 353 354/** 355 * pcpu_next_md_free_region - finds the next hint free area 356 * @chunk: chunk of interest 357 * @bit_off: chunk offset 358 * @bits: size of free area 359 * 360 * Helper function for pcpu_for_each_md_free_region. It checks 361 * block->contig_hint and performs aggregation across blocks to find the 362 * next hint. It modifies bit_off and bits in-place to be consumed in the 363 * loop. 364 */ 365static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, 366 int *bits) 367{ 368 int i = pcpu_off_to_block_index(*bit_off); 369 int block_off = pcpu_off_to_block_off(*bit_off); 370 struct pcpu_block_md *block; 371 372 *bits = 0; 373 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 374 block++, i++) { 375 /* handles contig area across blocks */ 376 if (*bits) { 377 *bits += block->left_free; 378 if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 379 continue; 380 return; 381 } 382 383 /* 384 * This checks three things. First is there a contig_hint to 385 * check. Second, have we checked this hint before by 386 * comparing the block_off. Third, is this the same as the 387 * right contig hint. In the last case, it spills over into 388 * the next block and should be handled by the contig area 389 * across blocks code. 390 */ 391 *bits = block->contig_hint; 392 if (*bits && block->contig_hint_start >= block_off && 393 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { 394 *bit_off = pcpu_block_off_to_off(i, 395 block->contig_hint_start); 396 return; 397 } 398 /* reset to satisfy the second predicate above */ 399 block_off = 0; 400 401 *bits = block->right_free; 402 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; 403 } 404} 405 406/** 407 * pcpu_next_fit_region - finds fit areas for a given allocation request 408 * @chunk: chunk of interest 409 * @alloc_bits: size of allocation 410 * @align: alignment of area (max PAGE_SIZE) 411 * @bit_off: chunk offset 412 * @bits: size of free area 413 * 414 * Finds the next free region that is viable for use with a given size and 415 * alignment. This only returns if there is a valid area to be used for this 416 * allocation. block->first_free is returned if the allocation request fits 417 * within the block to see if the request can be fulfilled prior to the contig 418 * hint. 419 */ 420static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, 421 int align, int *bit_off, int *bits) 422{ 423 int i = pcpu_off_to_block_index(*bit_off); 424 int block_off = pcpu_off_to_block_off(*bit_off); 425 struct pcpu_block_md *block; 426 427 *bits = 0; 428 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 429 block++, i++) { 430 /* handles contig area across blocks */ 431 if (*bits) { 432 *bits += block->left_free; 433 if (*bits >= alloc_bits) 434 return; 435 if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 436 continue; 437 } 438 439 /* check block->contig_hint */ 440 *bits = ALIGN(block->contig_hint_start, align) - 441 block->contig_hint_start; 442 /* 443 * This uses the block offset to determine if this has been 444 * checked in the prior iteration. 445 */ 446 if (block->contig_hint && 447 block->contig_hint_start >= block_off && 448 block->contig_hint >= *bits + alloc_bits) { 449 int start = pcpu_next_hint(block, alloc_bits); 450 451 *bits += alloc_bits + block->contig_hint_start - 452 start; 453 *bit_off = pcpu_block_off_to_off(i, start); 454 return; 455 } 456 /* reset to satisfy the second predicate above */ 457 block_off = 0; 458 459 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, 460 align); 461 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; 462 *bit_off = pcpu_block_off_to_off(i, *bit_off); 463 if (*bits >= alloc_bits) 464 return; 465 } 466 467 /* no valid offsets were found - fail condition */ 468 *bit_off = pcpu_chunk_map_bits(chunk); 469} 470 471/* 472 * Metadata free area iterators. These perform aggregation of free areas 473 * based on the metadata blocks and return the offset @bit_off and size in 474 * bits of the free area @bits. pcpu_for_each_fit_region only returns when 475 * a fit is found for the allocation request. 476 */ 477#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ 478 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ 479 (bit_off) < pcpu_chunk_map_bits((chunk)); \ 480 (bit_off) += (bits) + 1, \ 481 pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) 482 483#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ 484 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 485 &(bits)); \ 486 (bit_off) < pcpu_chunk_map_bits((chunk)); \ 487 (bit_off) += (bits), \ 488 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 489 &(bits))) 490 491/** 492 * pcpu_mem_zalloc - allocate memory 493 * @size: bytes to allocate 494 * @gfp: allocation flags 495 * 496 * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. 498 * This is to facilitate passing through whitelisted flags. The 499 * returned memory is always zeroed. 500 * 501 * RETURNS: 502 * Pointer to the allocated area on success, NULL on failure. 503 */ 504static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) 505{ 506 if (WARN_ON_ONCE(!slab_is_available())) 507 return NULL; 508 509 if (size <= PAGE_SIZE) 510 return kzalloc(size, gfp); 511 else 512 return __vmalloc(size, gfp | __GFP_ZERO); 513} 514 515/** 516 * pcpu_mem_free - free memory 517 * @ptr: memory to free 518 * 519 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 520 */ 521static void pcpu_mem_free(void *ptr) 522{ 523 kvfree(ptr); 524} 525 526static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot, 527 bool move_front) 528{ 529 if (chunk != pcpu_reserved_chunk) { 530 if (move_front) 531 list_move(&chunk->list, &pcpu_chunk_lists[slot]); 532 else 533 list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]); 534 } 535} 536 537static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot) 538{ 539 __pcpu_chunk_move(chunk, slot, true); 540} 541 542/** 543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 544 * @chunk: chunk of interest 545 * @oslot: the previous slot it was on 546 * 547 * This function is called after an allocation or free changed @chunk. 548 * New slot according to the changed state is determined and @chunk is 549 * moved to the slot. Note that the reserved chunk is never put on 550 * chunk slots. 551 * 552 * CONTEXT: 553 * pcpu_lock. 554 */ 555static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 556{ 557 int nslot = pcpu_chunk_slot(chunk); 558 559 /* leave isolated chunks in-place */ 560 if (chunk->isolated) 561 return; 562 563 if (oslot != nslot) 564 __pcpu_chunk_move(chunk, nslot, oslot < nslot); 565} 566 567static void pcpu_isolate_chunk(struct pcpu_chunk *chunk) 568{ 569 lockdep_assert_held(&pcpu_lock); 570 571 if (!chunk->isolated) { 572 chunk->isolated = true; 573 pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages; 574 } 575 list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]); 576} 577 578static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk) 579{ 580 lockdep_assert_held(&pcpu_lock); 581 582 if (chunk->isolated) { 583 chunk->isolated = false; 584 pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages; 585 pcpu_chunk_relocate(chunk, -1); 586 } 587} 588 589/* 590 * pcpu_update_empty_pages - update empty page counters 591 * @chunk: chunk of interest 592 * @nr: nr of empty pages 593 * 594 * This is used to keep track of the empty pages now based on the premise 595 * a md_block covers a page. The hint update functions recognize if a block 596 * is made full or broken to calculate deltas for keeping track of free pages. 597 */ 598static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr) 599{ 600 chunk->nr_empty_pop_pages += nr; 601 if (chunk != pcpu_reserved_chunk && !chunk->isolated) 602 pcpu_nr_empty_pop_pages += nr; 603} 604 605/* 606 * pcpu_region_overlap - determines if two regions overlap 607 * @a: start of first region, inclusive 608 * @b: end of first region, exclusive 609 * @x: start of second region, inclusive 610 * @y: end of second region, exclusive 611 * 612 * This is used to determine if the hint region [a, b) overlaps with the 613 * allocated region [x, y). 614 */ 615static inline bool pcpu_region_overlap(int a, int b, int x, int y) 616{ 617 return (a < y) && (x < b); 618} 619 620/** 621 * pcpu_block_update - updates a block given a free area 622 * @block: block of interest 623 * @start: start offset in block 624 * @end: end offset in block 625 * 626 * Updates a block given a known free area. The region [start, end) is 627 * expected to be the entirety of the free area within a block. Chooses 628 * the best starting offset if the contig hints are equal. 629 */ 630static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 631{ 632 int contig = end - start; 633 634 block->first_free = min(block->first_free, start); 635 if (start == 0) 636 block->left_free = contig; 637 638 if (end == block->nr_bits) 639 block->right_free = contig; 640 641 if (contig > block->contig_hint) { 642 /* promote the old contig_hint to be the new scan_hint */ 643 if (start > block->contig_hint_start) { 644 if (block->contig_hint > block->scan_hint) { 645 block->scan_hint_start = 646 block->contig_hint_start; 647 block->scan_hint = block->contig_hint; 648 } else if (start < block->scan_hint_start) { 649 /* 650 * The old contig_hint == scan_hint. But, the 651 * new contig is larger so hold the invariant 652 * scan_hint_start < contig_hint_start. 653 */ 654 block->scan_hint = 0; 655 } 656 } else { 657 block->scan_hint = 0; 658 } 659 block->contig_hint_start = start; 660 block->contig_hint = contig; 661 } else if (contig == block->contig_hint) { 662 if (block->contig_hint_start && 663 (!start || 664 __ffs(start) > __ffs(block->contig_hint_start))) { 665 /* start has a better alignment so use it */ 666 block->contig_hint_start = start; 667 if (start < block->scan_hint_start && 668 block->contig_hint > block->scan_hint) 669 block->scan_hint = 0; 670 } else if (start > block->scan_hint_start || 671 block->contig_hint > block->scan_hint) { 672 /* 673 * Knowing contig == contig_hint, update the scan_hint 674 * if it is farther than or larger than the current 675 * scan_hint. 676 */ 677 block->scan_hint_start = start; 678 block->scan_hint = contig; 679 } 680 } else { 681 /* 682 * The region is smaller than the contig_hint. So only update 683 * the scan_hint if it is larger than or equal and farther than 684 * the current scan_hint. 685 */ 686 if ((start < block->contig_hint_start && 687 (contig > block->scan_hint || 688 (contig == block->scan_hint && 689 start > block->scan_hint_start)))) { 690 block->scan_hint_start = start; 691 block->scan_hint = contig; 692 } 693 } 694} 695 696/* 697 * pcpu_block_update_scan - update a block given a free area from a scan 698 * @chunk: chunk of interest 699 * @bit_off: chunk offset 700 * @bits: size of free area 701 * 702 * Finding the final allocation spot first goes through pcpu_find_block_fit() 703 * to find a block that can hold the allocation and then pcpu_alloc_area() 704 * where a scan is used. When allocations require specific alignments, 705 * we can inadvertently create holes which will not be seen in the alloc 706 * or free paths. 707 * 708 * This takes a given free area hole and updates a block as it may change the 709 * scan_hint. We need to scan backwards to ensure we don't miss free bits 710 * from alignment. 711 */ 712static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off, 713 int bits) 714{ 715 int s_off = pcpu_off_to_block_off(bit_off); 716 int e_off = s_off + bits; 717 int s_index, l_bit; 718 struct pcpu_block_md *block; 719 720 if (e_off > PCPU_BITMAP_BLOCK_BITS) 721 return; 722 723 s_index = pcpu_off_to_block_index(bit_off); 724 block = chunk->md_blocks + s_index; 725 726 /* scan backwards in case of alignment skipping free bits */ 727 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off); 728 s_off = (s_off == l_bit) ? 0 : l_bit + 1; 729 730 pcpu_block_update(block, s_off, e_off); 731} 732 733/** 734 * pcpu_chunk_refresh_hint - updates metadata about a chunk 735 * @chunk: chunk of interest 736 * @full_scan: if we should scan from the beginning 737 * 738 * Iterates over the metadata blocks to find the largest contig area. 739 * A full scan can be avoided on the allocation path as this is triggered 740 * if we broke the contig_hint. In doing so, the scan_hint will be before 741 * the contig_hint or after if the scan_hint == contig_hint. This cannot 742 * be prevented on freeing as we want to find the largest area possibly 743 * spanning blocks. 744 */ 745static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan) 746{ 747 struct pcpu_block_md *chunk_md = &chunk->chunk_md; 748 int bit_off, bits; 749 750 /* promote scan_hint to contig_hint */ 751 if (!full_scan && chunk_md->scan_hint) { 752 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint; 753 chunk_md->contig_hint_start = chunk_md->scan_hint_start; 754 chunk_md->contig_hint = chunk_md->scan_hint; 755 chunk_md->scan_hint = 0; 756 } else { 757 bit_off = chunk_md->first_free; 758 chunk_md->contig_hint = 0; 759 } 760 761 bits = 0; 762 pcpu_for_each_md_free_region(chunk, bit_off, bits) 763 pcpu_block_update(chunk_md, bit_off, bit_off + bits); 764} 765 766/** 767 * pcpu_block_refresh_hint 768 * @chunk: chunk of interest 769 * @index: index of the metadata block 770 * 771 * Scans over the block beginning at first_free and updates the block 772 * metadata accordingly. 773 */ 774static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 775{ 776 struct pcpu_block_md *block = chunk->md_blocks + index; 777 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 778 unsigned int start, end; /* region start, region end */ 779 780 /* promote scan_hint to contig_hint */ 781 if (block->scan_hint) { 782 start = block->scan_hint_start + block->scan_hint; 783 block->contig_hint_start = block->scan_hint_start; 784 block->contig_hint = block->scan_hint; 785 block->scan_hint = 0; 786 } else { 787 start = block->first_free; 788 block->contig_hint = 0; 789 } 790 791 block->right_free = 0; 792 793 /* iterate over free areas and update the contig hints */ 794 for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS) 795 pcpu_block_update(block, start, end); 796} 797 798/** 799 * pcpu_block_update_hint_alloc - update hint on allocation path 800 * @chunk: chunk of interest 801 * @bit_off: chunk offset 802 * @bits: size of request 803 * 804 * Updates metadata for the allocation path. The metadata only has to be 805 * refreshed by a full scan iff the chunk's contig hint is broken. Block level 806 * scans are required if the block's contig hint is broken. 807 */ 808static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 809 int bits) 810{ 811 struct pcpu_block_md *chunk_md = &chunk->chunk_md; 812 int nr_empty_pages = 0; 813 struct pcpu_block_md *s_block, *e_block, *block; 814 int s_index, e_index; /* block indexes of the freed allocation */ 815 int s_off, e_off; /* block offsets of the freed allocation */ 816 817 /* 818 * Calculate per block offsets. 819 * The calculation uses an inclusive range, but the resulting offsets 820 * are [start, end). e_index always points to the last block in the 821 * range. 822 */ 823 s_index = pcpu_off_to_block_index(bit_off); 824 e_index = pcpu_off_to_block_index(bit_off + bits - 1); 825 s_off = pcpu_off_to_block_off(bit_off); 826 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 827 828 s_block = chunk->md_blocks + s_index; 829 e_block = chunk->md_blocks + e_index; 830 831 /* 832 * Update s_block. 833 */ 834 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 835 nr_empty_pages++; 836 837 /* 838 * block->first_free must be updated if the allocation takes its place. 839 * If the allocation breaks the contig_hint, a scan is required to 840 * restore this hint. 841 */ 842 if (s_off == s_block->first_free) 843 s_block->first_free = find_next_zero_bit( 844 pcpu_index_alloc_map(chunk, s_index), 845 PCPU_BITMAP_BLOCK_BITS, 846 s_off + bits); 847 848 if (pcpu_region_overlap(s_block->scan_hint_start, 849 s_block->scan_hint_start + s_block->scan_hint, 850 s_off, 851 s_off + bits)) 852 s_block->scan_hint = 0; 853 854 if (pcpu_region_overlap(s_block->contig_hint_start, 855 s_block->contig_hint_start + 856 s_block->contig_hint, 857 s_off, 858 s_off + bits)) { 859 /* block contig hint is broken - scan to fix it */ 860 if (!s_off) 861 s_block->left_free = 0; 862 pcpu_block_refresh_hint(chunk, s_index); 863 } else { 864 /* update left and right contig manually */ 865 s_block->left_free = min(s_block->left_free, s_off); 866 if (s_index == e_index) 867 s_block->right_free = min_t(int, s_block->right_free, 868 PCPU_BITMAP_BLOCK_BITS - e_off); 869 else 870 s_block->right_free = 0; 871 } 872 873 /* 874 * Update e_block. 875 */ 876 if (s_index != e_index) { 877 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 878 nr_empty_pages++; 879 880 /* 881 * When the allocation is across blocks, the end is along 882 * the left part of the e_block. 883 */ 884 e_block->first_free = find_next_zero_bit( 885 pcpu_index_alloc_map(chunk, e_index), 886 PCPU_BITMAP_BLOCK_BITS, e_off); 887 888 if (e_off == PCPU_BITMAP_BLOCK_BITS) { 889 /* reset the block */ 890 e_block++; 891 } else { 892 if (e_off > e_block->scan_hint_start) 893 e_block->scan_hint = 0; 894 895 e_block->left_free = 0; 896 if (e_off > e_block->contig_hint_start) { 897 /* contig hint is broken - scan to fix it */ 898 pcpu_block_refresh_hint(chunk, e_index); 899 } else { 900 e_block->right_free = 901 min_t(int, e_block->right_free, 902 PCPU_BITMAP_BLOCK_BITS - e_off); 903 } 904 } 905 906 /* update in-between md_blocks */ 907 nr_empty_pages += (e_index - s_index - 1); 908 for (block = s_block + 1; block < e_block; block++) { 909 block->scan_hint = 0; 910 block->contig_hint = 0; 911 block->left_free = 0; 912 block->right_free = 0; 913 } 914 } 915 916 /* 917 * If the allocation is not atomic, some blocks may not be 918 * populated with pages, while we account it here. The number 919 * of pages will be added back with pcpu_chunk_populated() 920 * when populating pages. 921 */ 922 if (nr_empty_pages) 923 pcpu_update_empty_pages(chunk, -nr_empty_pages); 924 925 if (pcpu_region_overlap(chunk_md->scan_hint_start, 926 chunk_md->scan_hint_start + 927 chunk_md->scan_hint, 928 bit_off, 929 bit_off + bits)) 930 chunk_md->scan_hint = 0; 931 932 /* 933 * The only time a full chunk scan is required is if the chunk 934 * contig hint is broken. Otherwise, it means a smaller space 935 * was used and therefore the chunk contig hint is still correct. 936 */ 937 if (pcpu_region_overlap(chunk_md->contig_hint_start, 938 chunk_md->contig_hint_start + 939 chunk_md->contig_hint, 940 bit_off, 941 bit_off + bits)) 942 pcpu_chunk_refresh_hint(chunk, false); 943} 944 945/** 946 * pcpu_block_update_hint_free - updates the block hints on the free path 947 * @chunk: chunk of interest 948 * @bit_off: chunk offset 949 * @bits: size of request 950 * 951 * Updates metadata for the allocation path. This avoids a blind block 952 * refresh by making use of the block contig hints. If this fails, it scans 953 * forward and backward to determine the extent of the free area. This is 954 * capped at the boundary of blocks. 955 * 956 * A chunk update is triggered if a page becomes free, a block becomes free, 957 * or the free spans across blocks. This tradeoff is to minimize iterating 958 * over the block metadata to update chunk_md->contig_hint. 959 * chunk_md->contig_hint may be off by up to a page, but it will never be more 960 * than the available space. If the contig hint is contained in one block, it 961 * will be accurate. 962 */ 963static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 964 int bits) 965{ 966 int nr_empty_pages = 0; 967 struct pcpu_block_md *s_block, *e_block, *block; 968 int s_index, e_index; /* block indexes of the freed allocation */ 969 int s_off, e_off; /* block offsets of the freed allocation */ 970 int start, end; /* start and end of the whole free area */ 971 972 /* 973 * Calculate per block offsets. 974 * The calculation uses an inclusive range, but the resulting offsets 975 * are [start, end). e_index always points to the last block in the 976 * range. 977 */ 978 s_index = pcpu_off_to_block_index(bit_off); 979 e_index = pcpu_off_to_block_index(bit_off + bits - 1); 980 s_off = pcpu_off_to_block_off(bit_off); 981 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 982 983 s_block = chunk->md_blocks + s_index; 984 e_block = chunk->md_blocks + e_index; 985 986 /* 987 * Check if the freed area aligns with the block->contig_hint. 988 * If it does, then the scan to find the beginning/end of the 989 * larger free area can be avoided. 990 * 991 * start and end refer to beginning and end of the free area 992 * within each their respective blocks. This is not necessarily 993 * the entire free area as it may span blocks past the beginning 994 * or end of the block. 995 */ 996 start = s_off; 997 if (s_off == s_block->contig_hint + s_block->contig_hint_start) { 998 start = s_block->contig_hint_start; 999 } else { 1000 /* 1001 * Scan backwards to find the extent of the free area. 1002 * find_last_bit returns the starting bit, so if the start bit 1003 * is returned, that means there was no last bit and the 1004 * remainder of the chunk is free. 1005 */ 1006 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), 1007 start); 1008 start = (start == l_bit) ? 0 : l_bit + 1; 1009 } 1010 1011 end = e_off; 1012 if (e_off == e_block->contig_hint_start) 1013 end = e_block->contig_hint_start + e_block->contig_hint; 1014 else 1015 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), 1016 PCPU_BITMAP_BLOCK_BITS, end); 1017 1018 /* update s_block */ 1019 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; 1020 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS) 1021 nr_empty_pages++; 1022 pcpu_block_update(s_block, start, e_off); 1023 1024 /* freeing in the same block */ 1025 if (s_index != e_index) { 1026 /* update e_block */ 1027 if (end == PCPU_BITMAP_BLOCK_BITS) 1028 nr_empty_pages++; 1029 pcpu_block_update(e_block, 0, end); 1030 1031 /* reset md_blocks in the middle */ 1032 nr_empty_pages += (e_index - s_index - 1); 1033 for (block = s_block + 1; block < e_block; block++) { 1034 block->first_free = 0; 1035 block->scan_hint = 0; 1036 block->contig_hint_start = 0; 1037 block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 1038 block->left_free = PCPU_BITMAP_BLOCK_BITS; 1039 block->right_free = PCPU_BITMAP_BLOCK_BITS; 1040 } 1041 } 1042 1043 if (nr_empty_pages) 1044 pcpu_update_empty_pages(chunk, nr_empty_pages); 1045 1046 /* 1047 * Refresh chunk metadata when the free makes a block free or spans 1048 * across blocks. The contig_hint may be off by up to a page, but if 1049 * the contig_hint is contained in a block, it will be accurate with 1050 * the else condition below. 1051 */ 1052 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index) 1053 pcpu_chunk_refresh_hint(chunk, true); 1054 else 1055 pcpu_block_update(&chunk->chunk_md, 1056 pcpu_block_off_to_off(s_index, start), 1057 end); 1058} 1059 1060/** 1061 * pcpu_is_populated - determines if the region is populated 1062 * @chunk: chunk of interest 1063 * @bit_off: chunk offset 1064 * @bits: size of area 1065 * @next_off: return value for the next offset to start searching 1066 * 1067 * For atomic allocations, check if the backing pages are populated. 1068 * 1069 * RETURNS: 1070 * Bool if the backing pages are populated. 1071 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 1072 */ 1073static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 1074 int *next_off) 1075{ 1076 unsigned int start, end; 1077 1078 start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 1079 end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 1080 1081 start = find_next_zero_bit(chunk->populated, end, start); 1082 if (start >= end) 1083 return true; 1084 1085 end = find_next_bit(chunk->populated, end, start + 1); 1086 1087 *next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 1088 return false; 1089} 1090 1091/** 1092 * pcpu_find_block_fit - finds the block index to start searching 1093 * @chunk: chunk of interest 1094 * @alloc_bits: size of request in allocation units 1095 * @align: alignment of area (max PAGE_SIZE bytes) 1096 * @pop_only: use populated regions only 1097 * 1098 * Given a chunk and an allocation spec, find the offset to begin searching 1099 * for a free region. This iterates over the bitmap metadata blocks to 1100 * find an offset that will be guaranteed to fit the requirements. It is 1101 * not quite first fit as if the allocation does not fit in the contig hint 1102 * of a block or chunk, it is skipped. This errs on the side of caution 1103 * to prevent excess iteration. Poor alignment can cause the allocator to 1104 * skip over blocks and chunks that have valid free areas. 1105 * 1106 * RETURNS: 1107 * The offset in the bitmap to begin searching. 1108 * -1 if no offset is found. 1109 */ 1110static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 1111 size_t align, bool pop_only) 1112{ 1113 struct pcpu_block_md *chunk_md = &chunk->chunk_md; 1114 int bit_off, bits, next_off; 1115 1116 /* 1117 * This is an optimization to prevent scanning by assuming if the 1118 * allocation cannot fit in the global hint, there is memory pressure 1119 * and creating a new chunk would happen soon. 1120 */ 1121 if (!pcpu_check_block_hint(chunk_md, alloc_bits, align)) 1122 return -1; 1123 1124 bit_off = pcpu_next_hint(chunk_md, alloc_bits); 1125 bits = 0; 1126 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { 1127 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 1128 &next_off)) 1129 break; 1130 1131 bit_off = next_off; 1132 bits = 0; 1133 } 1134 1135 if (bit_off == pcpu_chunk_map_bits(chunk)) 1136 return -1; 1137 1138 return bit_off; 1139} 1140 1141/* 1142 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off() 1143 * @map: the address to base the search on 1144 * @size: the bitmap size in bits 1145 * @start: the bitnumber to start searching at 1146 * @nr: the number of zeroed bits we're looking for 1147 * @align_mask: alignment mask for zero area 1148 * @largest_off: offset of the largest area skipped 1149 * @largest_bits: size of the largest area skipped 1150 * 1151 * The @align_mask should be one less than a power of 2. 1152 * 1153 * This is a modified version of bitmap_find_next_zero_area_off() to remember 1154 * the largest area that was skipped. This is imperfect, but in general is 1155 * good enough. The largest remembered region is the largest failed region 1156 * seen. This does not include anything we possibly skipped due to alignment. 1157 * pcpu_block_update_scan() does scan backwards to try and recover what was 1158 * lost to alignment. While this can cause scanning to miss earlier possible 1159 * free areas, smaller allocations will eventually fill those holes. 1160 */ 1161static unsigned long pcpu_find_zero_area(unsigned long *map, 1162 unsigned long size, 1163 unsigned long start, 1164 unsigned long nr, 1165 unsigned long align_mask, 1166 unsigned long *largest_off, 1167 unsigned long *largest_bits) 1168{ 1169 unsigned long index, end, i, area_off, area_bits; 1170again: 1171 index = find_next_zero_bit(map, size, start); 1172 1173 /* Align allocation */ 1174 index = __ALIGN_MASK(index, align_mask); 1175 area_off = index; 1176 1177 end = index + nr; 1178 if (end > size) 1179 return end; 1180 i = find_next_bit(map, end, index); 1181 if (i < end) { 1182 area_bits = i - area_off; 1183 /* remember largest unused area with best alignment */ 1184 if (area_bits > *largest_bits || 1185 (area_bits == *largest_bits && *largest_off && 1186 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) { 1187 *largest_off = area_off; 1188 *largest_bits = area_bits; 1189 } 1190 1191 start = i + 1; 1192 goto again; 1193 } 1194 return index; 1195} 1196 1197/** 1198 * pcpu_alloc_area - allocates an area from a pcpu_chunk 1199 * @chunk: chunk of interest 1200 * @alloc_bits: size of request in allocation units 1201 * @align: alignment of area (max PAGE_SIZE) 1202 * @start: bit_off to start searching 1203 * 1204 * This function takes in a @start offset to begin searching to fit an 1205 * allocation of @alloc_bits with alignment @align. It needs to scan 1206 * the allocation map because if it fits within the block's contig hint, 1207 * @start will be block->first_free. This is an attempt to fill the 1208 * allocation prior to breaking the contig hint. The allocation and 1209 * boundary maps are updated accordingly if it confirms a valid 1210 * free area. 1211 * 1212 * RETURNS: 1213 * Allocated addr offset in @chunk on success. 1214 * -1 if no matching area is found. 1215 */ 1216static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 1217 size_t align, int start) 1218{ 1219 struct pcpu_block_md *chunk_md = &chunk->chunk_md; 1220 size_t align_mask = (align) ? (align - 1) : 0; 1221 unsigned long area_off = 0, area_bits = 0; 1222 int bit_off, end, oslot; 1223 1224 lockdep_assert_held(&pcpu_lock); 1225 1226 oslot = pcpu_chunk_slot(chunk); 1227 1228 /* 1229 * Search to find a fit. 1230 */ 1231 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS, 1232 pcpu_chunk_map_bits(chunk)); 1233 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits, 1234 align_mask, &area_off, &area_bits); 1235 if (bit_off >= end) 1236 return -1; 1237 1238 if (area_bits) 1239 pcpu_block_update_scan(chunk, area_off, area_bits); 1240 1241 /* update alloc map */ 1242 bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 1243 1244 /* update boundary map */ 1245 set_bit(bit_off, chunk->bound_map); 1246 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 1247 set_bit(bit_off + alloc_bits, chunk->bound_map); 1248 1249 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 1250 1251 /* update first free bit */ 1252 if (bit_off == chunk_md->first_free) 1253 chunk_md->first_free = find_next_zero_bit( 1254 chunk->alloc_map, 1255 pcpu_chunk_map_bits(chunk), 1256 bit_off + alloc_bits); 1257 1258 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 1259 1260 pcpu_chunk_relocate(chunk, oslot); 1261 1262 return bit_off * PCPU_MIN_ALLOC_SIZE; 1263} 1264 1265/** 1266 * pcpu_free_area - frees the corresponding offset 1267 * @chunk: chunk of interest 1268 * @off: addr offset into chunk 1269 * 1270 * This function determines the size of an allocation to free using 1271 * the boundary bitmap and clears the allocation map. 1272 * 1273 * RETURNS: 1274 * Number of freed bytes. 1275 */ 1276static int pcpu_free_area(struct pcpu_chunk *chunk, int off) 1277{ 1278 struct pcpu_block_md *chunk_md = &chunk->chunk_md; 1279 int bit_off, bits, end, oslot, freed; 1280 1281 lockdep_assert_held(&pcpu_lock); 1282 pcpu_stats_area_dealloc(chunk); 1283 1284 oslot = pcpu_chunk_slot(chunk); 1285 1286 bit_off = off / PCPU_MIN_ALLOC_SIZE; 1287 1288 /* find end index */ 1289 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 1290 bit_off + 1); 1291 bits = end - bit_off; 1292 bitmap_clear(chunk->alloc_map, bit_off, bits); 1293 1294 freed = bits * PCPU_MIN_ALLOC_SIZE; 1295 1296 /* update metadata */ 1297 chunk->free_bytes += freed; 1298 1299 /* update first free bit */ 1300 chunk_md->first_free = min(chunk_md->first_free, bit_off); 1301 1302 pcpu_block_update_hint_free(chunk, bit_off, bits); 1303 1304 pcpu_chunk_relocate(chunk, oslot); 1305 1306 return freed; 1307} 1308 1309static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits) 1310{ 1311 block->scan_hint = 0; 1312 block->contig_hint = nr_bits; 1313 block->left_free = nr_bits; 1314 block->right_free = nr_bits; 1315 block->first_free = 0; 1316 block->nr_bits = nr_bits; 1317} 1318 1319static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 1320{ 1321 struct pcpu_block_md *md_block; 1322 1323 /* init the chunk's block */ 1324 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk)); 1325 1326 for (md_block = chunk->md_blocks; 1327 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 1328 md_block++) 1329 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS); 1330} 1331 1332/** 1333 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 1334 * @tmp_addr: the start of the region served 1335 * @map_size: size of the region served 1336 * 1337 * This is responsible for creating the chunks that serve the first chunk. The 1338 * base_addr is page aligned down of @tmp_addr while the region end is page 1339 * aligned up. Offsets are kept track of to determine the region served. All 1340 * this is done to appease the bitmap allocator in avoiding partial blocks. 1341 * 1342 * RETURNS: 1343 * Chunk serving the region at @tmp_addr of @map_size. 1344 */ 1345static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 1346 int map_size) 1347{ 1348 struct pcpu_chunk *chunk; 1349 unsigned long aligned_addr; 1350 int start_offset, offset_bits, region_size, region_bits; 1351 size_t alloc_size; 1352 1353 /* region calculations */ 1354 aligned_addr = tmp_addr & PAGE_MASK; 1355 1356 start_offset = tmp_addr - aligned_addr; 1357 region_size = ALIGN(start_offset + map_size, PAGE_SIZE); 1358 1359 /* allocate chunk */ 1360 alloc_size = struct_size(chunk, populated, 1361 BITS_TO_LONGS(region_size >> PAGE_SHIFT)); 1362 chunk = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 1363 1364 INIT_LIST_HEAD(&chunk->list); 1365 1366 chunk->base_addr = (void *)aligned_addr; 1367 chunk->start_offset = start_offset; 1368 chunk->end_offset = region_size - chunk->start_offset - map_size; 1369 1370 chunk->nr_pages = region_size >> PAGE_SHIFT; 1371 region_bits = pcpu_chunk_map_bits(chunk); 1372 1373 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]); 1374 chunk->alloc_map = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 1375 1376 alloc_size = 1377 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]); 1378 chunk->bound_map = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 1379 1380 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]); 1381 chunk->md_blocks = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 1382#ifdef NEED_PCPUOBJ_EXT 1383 /* first chunk is free to use */ 1384 chunk->obj_exts = NULL; 1385#endif 1386 pcpu_init_md_blocks(chunk); 1387 1388 /* manage populated page bitmap */ 1389 chunk->immutable = true; 1390 bitmap_fill(chunk->populated, chunk->nr_pages); 1391 chunk->nr_populated = chunk->nr_pages; 1392 chunk->nr_empty_pop_pages = chunk->nr_pages; 1393 1394 chunk->free_bytes = map_size; 1395 1396 if (chunk->start_offset) { 1397 /* hide the beginning of the bitmap */ 1398 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 1399 bitmap_set(chunk->alloc_map, 0, offset_bits); 1400 set_bit(0, chunk->bound_map); 1401 set_bit(offset_bits, chunk->bound_map); 1402 1403 chunk->chunk_md.first_free = offset_bits; 1404 1405 pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 1406 } 1407 1408 if (chunk->end_offset) { 1409 /* hide the end of the bitmap */ 1410 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 1411 bitmap_set(chunk->alloc_map, 1412 pcpu_chunk_map_bits(chunk) - offset_bits, 1413 offset_bits); 1414 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 1415 chunk->bound_map); 1416 set_bit(region_bits, chunk->bound_map); 1417 1418 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 1419 - offset_bits, offset_bits); 1420 } 1421 1422 return chunk; 1423} 1424 1425static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) 1426{ 1427 struct pcpu_chunk *chunk; 1428 int region_bits; 1429 1430 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); 1431 if (!chunk) 1432 return NULL; 1433 1434 INIT_LIST_HEAD(&chunk->list); 1435 chunk->nr_pages = pcpu_unit_pages; 1436 region_bits = pcpu_chunk_map_bits(chunk); 1437 1438 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 1439 sizeof(chunk->alloc_map[0]), gfp); 1440 if (!chunk->alloc_map) 1441 goto alloc_map_fail; 1442 1443 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 1444 sizeof(chunk->bound_map[0]), gfp); 1445 if (!chunk->bound_map) 1446 goto bound_map_fail; 1447 1448 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 1449 sizeof(chunk->md_blocks[0]), gfp); 1450 if (!chunk->md_blocks) 1451 goto md_blocks_fail; 1452 1453#ifdef NEED_PCPUOBJ_EXT 1454 if (need_pcpuobj_ext()) { 1455 chunk->obj_exts = 1456 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) * 1457 sizeof(struct pcpuobj_ext), gfp); 1458 if (!chunk->obj_exts) 1459 goto objcg_fail; 1460 } 1461#endif 1462 1463 pcpu_init_md_blocks(chunk); 1464 1465 /* init metadata */ 1466 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 1467 1468 return chunk; 1469 1470#ifdef NEED_PCPUOBJ_EXT 1471objcg_fail: 1472 pcpu_mem_free(chunk->md_blocks); 1473#endif 1474md_blocks_fail: 1475 pcpu_mem_free(chunk->bound_map); 1476bound_map_fail: 1477 pcpu_mem_free(chunk->alloc_map); 1478alloc_map_fail: 1479 pcpu_mem_free(chunk); 1480 1481 return NULL; 1482} 1483 1484static void pcpu_free_chunk(struct pcpu_chunk *chunk) 1485{ 1486 if (!chunk) 1487 return; 1488#ifdef NEED_PCPUOBJ_EXT 1489 pcpu_mem_free(chunk->obj_exts); 1490#endif 1491 pcpu_mem_free(chunk->md_blocks); 1492 pcpu_mem_free(chunk->bound_map); 1493 pcpu_mem_free(chunk->alloc_map); 1494 pcpu_mem_free(chunk); 1495} 1496 1497/** 1498 * pcpu_chunk_populated - post-population bookkeeping 1499 * @chunk: pcpu_chunk which got populated 1500 * @page_start: the start page 1501 * @page_end: the end page 1502 * 1503 * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1504 * the bookkeeping information accordingly. Must be called after each 1505 * successful population. 1506 */ 1507static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 1508 int page_end) 1509{ 1510 int nr = page_end - page_start; 1511 1512 lockdep_assert_held(&pcpu_lock); 1513 1514 bitmap_set(chunk->populated, page_start, nr); 1515 chunk->nr_populated += nr; 1516 pcpu_nr_populated += nr; 1517 1518 pcpu_update_empty_pages(chunk, nr); 1519} 1520 1521/** 1522 * pcpu_chunk_depopulated - post-depopulation bookkeeping 1523 * @chunk: pcpu_chunk which got depopulated 1524 * @page_start: the start page 1525 * @page_end: the end page 1526 * 1527 * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1528 * Update the bookkeeping information accordingly. Must be called after 1529 * each successful depopulation. 1530 */ 1531static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1532 int page_start, int page_end) 1533{ 1534 int nr = page_end - page_start; 1535 1536 lockdep_assert_held(&pcpu_lock); 1537 1538 bitmap_clear(chunk->populated, page_start, nr); 1539 chunk->nr_populated -= nr; 1540 pcpu_nr_populated -= nr; 1541 1542 pcpu_update_empty_pages(chunk, -nr); 1543} 1544 1545/* 1546 * Chunk management implementation. 1547 * 1548 * To allow different implementations, chunk alloc/free and 1549 * [de]population are implemented in a separate file which is pulled 1550 * into this file and compiled together. The following functions 1551 * should be implemented. 1552 * 1553 * pcpu_populate_chunk - populate the specified range of a chunk 1554 * pcpu_depopulate_chunk - depopulate the specified range of a chunk 1555 * pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk 1556 * pcpu_create_chunk - create a new chunk 1557 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 1558 * pcpu_addr_to_page - translate address to physical address 1559 * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1560 */ 1561static int pcpu_populate_chunk(struct pcpu_chunk *chunk, 1562 int page_start, int page_end, gfp_t gfp); 1563static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, 1564 int page_start, int page_end); 1565static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, 1566 int page_start, int page_end); 1567static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); 1568static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 1569static struct page *pcpu_addr_to_page(void *addr); 1570static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1571 1572#ifdef CONFIG_NEED_PER_CPU_KM 1573#include "percpu-km.c" 1574#else 1575#include "percpu-vm.c" 1576#endif 1577 1578/** 1579 * pcpu_chunk_addr_search - determine chunk containing specified address 1580 * @addr: address for which the chunk needs to be determined. 1581 * 1582 * This is an internal function that handles all but static allocations. 1583 * Static percpu address values should never be passed into the allocator. 1584 * 1585 * RETURNS: 1586 * The address of the found chunk. 1587 */ 1588static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 1589{ 1590 /* is it in the dynamic region (first chunk)? */ 1591 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1592 return pcpu_first_chunk; 1593 1594 /* is it in the reserved region? */ 1595 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 1596 return pcpu_reserved_chunk; 1597 1598 /* 1599 * The address is relative to unit0 which might be unused and 1600 * thus unmapped. Offset the address to the unit space of the 1601 * current processor before looking it up in the vmalloc 1602 * space. Note that any possible cpu id can be used here, so 1603 * there's no need to worry about preemption or cpu hotplug. 1604 */ 1605 addr += pcpu_unit_offsets[raw_smp_processor_id()]; 1606 return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 1607} 1608 1609#ifdef CONFIG_MEMCG 1610static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, 1611 struct obj_cgroup **objcgp) 1612{ 1613 struct obj_cgroup *objcg; 1614 1615 if (!memcg_kmem_online() || !(gfp & __GFP_ACCOUNT)) 1616 return true; 1617 1618 objcg = current_obj_cgroup(); 1619 if (!objcg) 1620 return true; 1621 1622 if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size))) 1623 return false; 1624 1625 *objcgp = objcg; 1626 return true; 1627} 1628 1629static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg, 1630 struct pcpu_chunk *chunk, int off, 1631 size_t size) 1632{ 1633 if (!objcg) 1634 return; 1635 1636 if (likely(chunk && chunk->obj_exts)) { 1637 obj_cgroup_get(objcg); 1638 chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup = objcg; 1639 1640 rcu_read_lock(); 1641 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B, 1642 pcpu_obj_full_size(size)); 1643 rcu_read_unlock(); 1644 } else { 1645 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size)); 1646 } 1647} 1648 1649static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size) 1650{ 1651 struct obj_cgroup *objcg; 1652 1653 if (unlikely(!chunk->obj_exts)) 1654 return; 1655 1656 objcg = chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup; 1657 if (!objcg) 1658 return; 1659 chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].cgroup = NULL; 1660 1661 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size)); 1662 1663 rcu_read_lock(); 1664 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B, 1665 -pcpu_obj_full_size(size)); 1666 rcu_read_unlock(); 1667 1668 obj_cgroup_put(objcg); 1669} 1670 1671#else /* CONFIG_MEMCG */ 1672static bool 1673pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp) 1674{ 1675 return true; 1676} 1677 1678static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg, 1679 struct pcpu_chunk *chunk, int off, 1680 size_t size) 1681{ 1682} 1683 1684static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size) 1685{ 1686} 1687#endif /* CONFIG_MEMCG */ 1688 1689#ifdef CONFIG_MEM_ALLOC_PROFILING 1690static void pcpu_alloc_tag_alloc_hook(struct pcpu_chunk *chunk, int off, 1691 size_t size) 1692{ 1693 if (mem_alloc_profiling_enabled() && likely(chunk->obj_exts)) { 1694 alloc_tag_add(&chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].tag, 1695 current->alloc_tag, size); 1696 } 1697} 1698 1699static void pcpu_alloc_tag_free_hook(struct pcpu_chunk *chunk, int off, size_t size) 1700{ 1701 if (mem_alloc_profiling_enabled() && likely(chunk->obj_exts)) 1702 alloc_tag_sub(&chunk->obj_exts[off >> PCPU_MIN_ALLOC_SHIFT].tag, size); 1703} 1704#else 1705static void pcpu_alloc_tag_alloc_hook(struct pcpu_chunk *chunk, int off, 1706 size_t size) 1707{ 1708} 1709 1710static void pcpu_alloc_tag_free_hook(struct pcpu_chunk *chunk, int off, size_t size) 1711{ 1712} 1713#endif 1714 1715/** 1716 * pcpu_alloc - the percpu allocator 1717 * @size: size of area to allocate in bytes 1718 * @align: alignment of area (max PAGE_SIZE) 1719 * @reserved: allocate from the reserved chunk if available 1720 * @gfp: allocation flags 1721 * 1722 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 1723 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN 1724 * then no warning will be triggered on invalid or failed allocation 1725 * requests. 1726 * 1727 * RETURNS: 1728 * Percpu pointer to the allocated area on success, NULL on failure. 1729 */ 1730void __percpu *pcpu_alloc_noprof(size_t size, size_t align, bool reserved, 1731 gfp_t gfp) 1732{ 1733 gfp_t pcpu_gfp; 1734 bool is_atomic; 1735 bool do_warn; 1736 struct obj_cgroup *objcg = NULL; 1737 static atomic_t warn_limit = ATOMIC_INIT(10); 1738 struct pcpu_chunk *chunk, *next; 1739 const char *err; 1740 int slot, off, cpu, ret; 1741 unsigned long flags; 1742 void __percpu *ptr; 1743 size_t bits, bit_align; 1744 1745 gfp = current_gfp_context(gfp); 1746 /* whitelisted flags that can be passed to the backing allocators */ 1747 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN); 1748 is_atomic = !gfpflags_allow_blocking(gfp); 1749 do_warn = !(gfp & __GFP_NOWARN); 1750 1751 /* 1752 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 1753 * therefore alignment must be a minimum of that many bytes. 1754 * An allocation may have internal fragmentation from rounding up 1755 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1756 */ 1757 if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1758 align = PCPU_MIN_ALLOC_SIZE; 1759 1760 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 1761 bits = size >> PCPU_MIN_ALLOC_SHIFT; 1762 bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 1763 1764 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 1765 !is_power_of_2(align))) { 1766 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1767 size, align); 1768 return NULL; 1769 } 1770 1771 if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg))) 1772 return NULL; 1773 1774 if (!is_atomic) { 1775 /* 1776 * pcpu_balance_workfn() allocates memory under this mutex, 1777 * and it may wait for memory reclaim. Allow current task 1778 * to become OOM victim, in case of memory pressure. 1779 */ 1780 if (gfp & __GFP_NOFAIL) { 1781 mutex_lock(&pcpu_alloc_mutex); 1782 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) { 1783 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size); 1784 return NULL; 1785 } 1786 } 1787 1788 spin_lock_irqsave(&pcpu_lock, flags); 1789 1790 /* serve reserved allocations from the reserved chunk if available */ 1791 if (reserved && pcpu_reserved_chunk) { 1792 chunk = pcpu_reserved_chunk; 1793 1794 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 1795 if (off < 0) { 1796 err = "alloc from reserved chunk failed"; 1797 goto fail_unlock; 1798 } 1799 1800 off = pcpu_alloc_area(chunk, bits, bit_align, off); 1801 if (off >= 0) 1802 goto area_found; 1803 1804 err = "alloc from reserved chunk failed"; 1805 goto fail_unlock; 1806 } 1807 1808restart: 1809 /* search through normal chunks */ 1810 for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) { 1811 list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot], 1812 list) { 1813 off = pcpu_find_block_fit(chunk, bits, bit_align, 1814 is_atomic); 1815 if (off < 0) { 1816 if (slot < PCPU_SLOT_FAIL_THRESHOLD) 1817 pcpu_chunk_move(chunk, 0); 1818 continue; 1819 } 1820 1821 off = pcpu_alloc_area(chunk, bits, bit_align, off); 1822 if (off >= 0) { 1823 pcpu_reintegrate_chunk(chunk); 1824 goto area_found; 1825 } 1826 } 1827 } 1828 1829 spin_unlock_irqrestore(&pcpu_lock, flags); 1830 1831 if (is_atomic) { 1832 err = "atomic alloc failed, no space left"; 1833 goto fail; 1834 } 1835 1836 /* No space left. Create a new chunk. */ 1837 if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) { 1838 chunk = pcpu_create_chunk(pcpu_gfp); 1839 if (!chunk) { 1840 err = "failed to allocate new chunk"; 1841 goto fail; 1842 } 1843 1844 spin_lock_irqsave(&pcpu_lock, flags); 1845 pcpu_chunk_relocate(chunk, -1); 1846 } else { 1847 spin_lock_irqsave(&pcpu_lock, flags); 1848 } 1849 1850 goto restart; 1851 1852area_found: 1853 pcpu_stats_area_alloc(chunk, size); 1854 1855 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 1856 pcpu_schedule_balance_work(); 1857 1858 spin_unlock_irqrestore(&pcpu_lock, flags); 1859 1860 /* populate if not all pages are already there */ 1861 if (!is_atomic) { 1862 unsigned int page_end, rs, re; 1863 1864 rs = PFN_DOWN(off); 1865 page_end = PFN_UP(off + size); 1866 1867 for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) { 1868 WARN_ON(chunk->immutable); 1869 1870 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp); 1871 1872 spin_lock_irqsave(&pcpu_lock, flags); 1873 if (ret) { 1874 pcpu_free_area(chunk, off); 1875 err = "failed to populate"; 1876 goto fail_unlock; 1877 } 1878 pcpu_chunk_populated(chunk, rs, re); 1879 spin_unlock_irqrestore(&pcpu_lock, flags); 1880 } 1881 1882 mutex_unlock(&pcpu_alloc_mutex); 1883 } 1884 1885 /* clear the areas and return address relative to base address */ 1886 for_each_possible_cpu(cpu) 1887 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1888 1889 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 1890 kmemleak_alloc_percpu(ptr, size, gfp); 1891 1892 trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align, 1893 chunk->base_addr, off, ptr, 1894 pcpu_obj_full_size(size), gfp); 1895 1896 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size); 1897 1898 pcpu_alloc_tag_alloc_hook(chunk, off, size); 1899 1900 return ptr; 1901 1902fail_unlock: 1903 spin_unlock_irqrestore(&pcpu_lock, flags); 1904fail: 1905 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1906 1907 if (do_warn) { 1908 int remaining = atomic_dec_if_positive(&warn_limit); 1909 1910 if (remaining >= 0) { 1911 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 1912 size, align, is_atomic, err); 1913 if (!is_atomic) 1914 dump_stack(); 1915 if (remaining == 0) 1916 pr_info("limit reached, disable warning\n"); 1917 } 1918 } 1919 1920 if (is_atomic) { 1921 /* see the flag handling in pcpu_balance_workfn() */ 1922 pcpu_atomic_alloc_failed = true; 1923 pcpu_schedule_balance_work(); 1924 } else { 1925 mutex_unlock(&pcpu_alloc_mutex); 1926 } 1927 1928 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size); 1929 1930 return NULL; 1931} 1932EXPORT_SYMBOL_GPL(pcpu_alloc_noprof); 1933 1934/** 1935 * pcpu_balance_free - manage the amount of free chunks 1936 * @empty_only: free chunks only if there are no populated pages 1937 * 1938 * If empty_only is %false, reclaim all fully free chunks regardless of the 1939 * number of populated pages. Otherwise, only reclaim chunks that have no 1940 * populated pages. 1941 * 1942 * CONTEXT: 1943 * pcpu_lock (can be dropped temporarily) 1944 */ 1945static void pcpu_balance_free(bool empty_only) 1946{ 1947 LIST_HEAD(to_free); 1948 struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot]; 1949 struct pcpu_chunk *chunk, *next; 1950 1951 lockdep_assert_held(&pcpu_lock); 1952 1953 /* 1954 * There's no reason to keep around multiple unused chunks and VM 1955 * areas can be scarce. Destroy all free chunks except for one. 1956 */ 1957 list_for_each_entry_safe(chunk, next, free_head, list) { 1958 WARN_ON(chunk->immutable); 1959 1960 /* spare the first one */ 1961 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1962 continue; 1963 1964 if (!empty_only || chunk->nr_empty_pop_pages == 0) 1965 list_move(&chunk->list, &to_free); 1966 } 1967 1968 if (list_empty(&to_free)) 1969 return; 1970 1971 spin_unlock_irq(&pcpu_lock); 1972 list_for_each_entry_safe(chunk, next, &to_free, list) { 1973 unsigned int rs, re; 1974 1975 for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) { 1976 pcpu_depopulate_chunk(chunk, rs, re); 1977 spin_lock_irq(&pcpu_lock); 1978 pcpu_chunk_depopulated(chunk, rs, re); 1979 spin_unlock_irq(&pcpu_lock); 1980 } 1981 pcpu_destroy_chunk(chunk); 1982 cond_resched(); 1983 } 1984 spin_lock_irq(&pcpu_lock); 1985} 1986 1987/** 1988 * pcpu_balance_populated - manage the amount of populated pages 1989 * 1990 * Maintain a certain amount of populated pages to satisfy atomic allocations. 1991 * It is possible that this is called when physical memory is scarce causing 1992 * OOM killer to be triggered. We should avoid doing so until an actual 1993 * allocation causes the failure as it is possible that requests can be 1994 * serviced from already backed regions. 1995 * 1996 * CONTEXT: 1997 * pcpu_lock (can be dropped temporarily) 1998 */ 1999static void pcpu_balance_populated(void) 2000{ 2001 /* gfp flags passed to underlying allocators */ 2002 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN; 2003 struct pcpu_chunk *chunk; 2004 int slot, nr_to_pop, ret; 2005 2006 lockdep_assert_held(&pcpu_lock); 2007 2008 /* 2009 * Ensure there are certain number of free populated pages for 2010 * atomic allocs. Fill up from the most packed so that atomic 2011 * allocs don't increase fragmentation. If atomic allocation 2012 * failed previously, always populate the maximum amount. This 2013 * should prevent atomic allocs larger than PAGE_SIZE from keeping 2014 * failing indefinitely; however, large atomic allocs are not 2015 * something we support properly and can be highly unreliable and 2016 * inefficient. 2017 */ 2018retry_pop: 2019 if (pcpu_atomic_alloc_failed) { 2020 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 2021 /* best effort anyway, don't worry about synchronization */ 2022 pcpu_atomic_alloc_failed = false; 2023 } else { 2024 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 2025 pcpu_nr_empty_pop_pages, 2026 0, PCPU_EMPTY_POP_PAGES_HIGH); 2027 } 2028 2029 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) { 2030 unsigned int nr_unpop = 0, rs, re; 2031 2032 if (!nr_to_pop) 2033 break; 2034 2035 list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) { 2036 nr_unpop = chunk->nr_pages - chunk->nr_populated; 2037 if (nr_unpop) 2038 break; 2039 } 2040 2041 if (!nr_unpop) 2042 continue; 2043 2044 /* @chunk can't go away while pcpu_alloc_mutex is held */ 2045 for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) { 2046 int nr = min_t(int, re - rs, nr_to_pop); 2047 2048 spin_unlock_irq(&pcpu_lock); 2049 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); 2050 cond_resched(); 2051 spin_lock_irq(&pcpu_lock); 2052 if (!ret) { 2053 nr_to_pop -= nr; 2054 pcpu_chunk_populated(chunk, rs, rs + nr); 2055 } else { 2056 nr_to_pop = 0; 2057 } 2058 2059 if (!nr_to_pop) 2060 break; 2061 } 2062 } 2063 2064 if (nr_to_pop) { 2065 /* ran out of chunks to populate, create a new one and retry */ 2066 spin_unlock_irq(&pcpu_lock); 2067 chunk = pcpu_create_chunk(gfp); 2068 cond_resched(); 2069 spin_lock_irq(&pcpu_lock); 2070 if (chunk) { 2071 pcpu_chunk_relocate(chunk, -1); 2072 goto retry_pop; 2073 } 2074 } 2075} 2076 2077/** 2078 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages 2079 * 2080 * Scan over chunks in the depopulate list and try to release unused populated 2081 * pages back to the system. Depopulated chunks are sidelined to prevent 2082 * repopulating these pages unless required. Fully free chunks are reintegrated 2083 * and freed accordingly (1 is kept around). If we drop below the empty 2084 * populated pages threshold, reintegrate the chunk if it has empty free pages. 2085 * Each chunk is scanned in the reverse order to keep populated pages close to 2086 * the beginning of the chunk. 2087 * 2088 * CONTEXT: 2089 * pcpu_lock (can be dropped temporarily) 2090 * 2091 */ 2092static void pcpu_reclaim_populated(void) 2093{ 2094 struct pcpu_chunk *chunk; 2095 struct pcpu_block_md *block; 2096 int freed_page_start, freed_page_end; 2097 int i, end; 2098 bool reintegrate; 2099 2100 lockdep_assert_held(&pcpu_lock); 2101 2102 /* 2103 * Once a chunk is isolated to the to_depopulate list, the chunk is no 2104 * longer discoverable to allocations whom may populate pages. The only 2105 * other accessor is the free path which only returns area back to the 2106 * allocator not touching the populated bitmap. 2107 */ 2108 while ((chunk = list_first_entry_or_null( 2109 &pcpu_chunk_lists[pcpu_to_depopulate_slot], 2110 struct pcpu_chunk, list))) { 2111 WARN_ON(chunk->immutable); 2112 2113 /* 2114 * Scan chunk's pages in the reverse order to keep populated 2115 * pages close to the beginning of the chunk. 2116 */ 2117 freed_page_start = chunk->nr_pages; 2118 freed_page_end = 0; 2119 reintegrate = false; 2120 for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) { 2121 /* no more work to do */ 2122 if (chunk->nr_empty_pop_pages == 0) 2123 break; 2124 2125 /* reintegrate chunk to prevent atomic alloc failures */ 2126 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) { 2127 reintegrate = true; 2128 break; 2129 } 2130 2131 /* 2132 * If the page is empty and populated, start or 2133 * extend the (i, end) range. If i == 0, decrease 2134 * i and perform the depopulation to cover the last 2135 * (first) page in the chunk. 2136 */ 2137 block = chunk->md_blocks + i; 2138 if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS && 2139 test_bit(i, chunk->populated)) { 2140 if (end == -1) 2141 end = i; 2142 if (i > 0) 2143 continue; 2144 i--; 2145 } 2146 2147 /* depopulate if there is an active range */ 2148 if (end == -1) 2149 continue; 2150 2151 spin_unlock_irq(&pcpu_lock); 2152 pcpu_depopulate_chunk(chunk, i + 1, end + 1); 2153 cond_resched(); 2154 spin_lock_irq(&pcpu_lock); 2155 2156 pcpu_chunk_depopulated(chunk, i + 1, end + 1); 2157 freed_page_start = min(freed_page_start, i + 1); 2158 freed_page_end = max(freed_page_end, end + 1); 2159 2160 /* reset the range and continue */ 2161 end = -1; 2162 } 2163 2164 /* batch tlb flush per chunk to amortize cost */ 2165 if (freed_page_start < freed_page_end) { 2166 spin_unlock_irq(&pcpu_lock); 2167 pcpu_post_unmap_tlb_flush(chunk, 2168 freed_page_start, 2169 freed_page_end); 2170 cond_resched(); 2171 spin_lock_irq(&pcpu_lock); 2172 } 2173 2174 if (reintegrate || chunk->free_bytes == pcpu_unit_size) 2175 pcpu_reintegrate_chunk(chunk); 2176 else 2177 list_move_tail(&chunk->list, 2178 &pcpu_chunk_lists[pcpu_sidelined_slot]); 2179 } 2180} 2181 2182/** 2183 * pcpu_balance_workfn - manage the amount of free chunks and populated pages 2184 * @work: unused 2185 * 2186 * For each chunk type, manage the number of fully free chunks and the number of 2187 * populated pages. An important thing to consider is when pages are freed and 2188 * how they contribute to the global counts. 2189 */ 2190static void pcpu_balance_workfn(struct work_struct *work) 2191{ 2192 /* 2193 * pcpu_balance_free() is called twice because the first time we may 2194 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us 2195 * to grow other chunks. This then gives pcpu_reclaim_populated() time 2196 * to move fully free chunks to the active list to be freed if 2197 * appropriate. 2198 * 2199 * Enforce GFP_NOIO allocations because we have pcpu_alloc users 2200 * constrained to GFP_NOIO/NOFS contexts and they could form lock 2201 * dependency through pcpu_alloc_mutex 2202 */ 2203 unsigned int flags = memalloc_noio_save(); 2204 mutex_lock(&pcpu_alloc_mutex); 2205 spin_lock_irq(&pcpu_lock); 2206 2207 pcpu_balance_free(false); 2208 pcpu_reclaim_populated(); 2209 pcpu_balance_populated(); 2210 pcpu_balance_free(true); 2211 2212 spin_unlock_irq(&pcpu_lock); 2213 mutex_unlock(&pcpu_alloc_mutex); 2214 memalloc_noio_restore(flags); 2215} 2216 2217/** 2218 * free_percpu - free percpu area 2219 * @ptr: pointer to area to free 2220 * 2221 * Free percpu area @ptr. 2222 * 2223 * CONTEXT: 2224 * Can be called from atomic context. 2225 */ 2226void free_percpu(void __percpu *ptr) 2227{ 2228 void *addr; 2229 struct pcpu_chunk *chunk; 2230 unsigned long flags; 2231 int size, off; 2232 bool need_balance = false; 2233 2234 if (!ptr) 2235 return; 2236 2237 kmemleak_free_percpu(ptr); 2238 2239 addr = __pcpu_ptr_to_addr(ptr); 2240 chunk = pcpu_chunk_addr_search(addr); 2241 off = addr - chunk->base_addr; 2242 2243 spin_lock_irqsave(&pcpu_lock, flags); 2244 size = pcpu_free_area(chunk, off); 2245 2246 pcpu_alloc_tag_free_hook(chunk, off, size); 2247 2248 pcpu_memcg_free_hook(chunk, off, size); 2249 2250 /* 2251 * If there are more than one fully free chunks, wake up grim reaper. 2252 * If the chunk is isolated, it may be in the process of being 2253 * reclaimed. Let reclaim manage cleaning up of that chunk. 2254 */ 2255 if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) { 2256 struct pcpu_chunk *pos; 2257 2258 list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list) 2259 if (pos != chunk) { 2260 need_balance = true; 2261 break; 2262 } 2263 } else if (pcpu_should_reclaim_chunk(chunk)) { 2264 pcpu_isolate_chunk(chunk); 2265 need_balance = true; 2266 } 2267 2268 trace_percpu_free_percpu(chunk->base_addr, off, ptr); 2269 2270 spin_unlock_irqrestore(&pcpu_lock, flags); 2271 2272 if (need_balance) 2273 pcpu_schedule_balance_work(); 2274} 2275EXPORT_SYMBOL_GPL(free_percpu); 2276 2277bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 2278{ 2279#ifdef CONFIG_SMP 2280 const size_t static_size = __per_cpu_end - __per_cpu_start; 2281 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 2282 unsigned int cpu; 2283 2284 for_each_possible_cpu(cpu) { 2285 void *start = per_cpu_ptr(base, cpu); 2286 void *va = (void *)addr; 2287 2288 if (va >= start && va < start + static_size) { 2289 if (can_addr) { 2290 *can_addr = (unsigned long) (va - start); 2291 *can_addr += (unsigned long) 2292 per_cpu_ptr(base, get_boot_cpu_id()); 2293 } 2294 return true; 2295 } 2296 } 2297#endif 2298 /* on UP, can't distinguish from other static vars, always false */ 2299 return false; 2300} 2301 2302/** 2303 * is_kernel_percpu_address - test whether address is from static percpu area 2304 * @addr: address to test 2305 * 2306 * Test whether @addr belongs to in-kernel static percpu area. Module 2307 * static percpu areas are not considered. For those, use 2308 * is_module_percpu_address(). 2309 * 2310 * RETURNS: 2311 * %true if @addr is from in-kernel static percpu area, %false otherwise. 2312 */ 2313bool is_kernel_percpu_address(unsigned long addr) 2314{ 2315 return __is_kernel_percpu_address(addr, NULL); 2316} 2317 2318/** 2319 * per_cpu_ptr_to_phys - convert translated percpu address to physical address 2320 * @addr: the address to be converted to physical address 2321 * 2322 * Given @addr which is dereferenceable address obtained via one of 2323 * percpu access macros, this function translates it into its physical 2324 * address. The caller is responsible for ensuring @addr stays valid 2325 * until this function finishes. 2326 * 2327 * percpu allocator has special setup for the first chunk, which currently 2328 * supports either embedding in linear address space or vmalloc mapping, 2329 * and, from the second one, the backing allocator (currently either vm or 2330 * km) provides translation. 2331 * 2332 * The addr can be translated simply without checking if it falls into the 2333 * first chunk. But the current code reflects better how percpu allocator 2334 * actually works, and the verification can discover both bugs in percpu 2335 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 2336 * code. 2337 * 2338 * RETURNS: 2339 * The physical address for @addr. 2340 */ 2341phys_addr_t per_cpu_ptr_to_phys(void *addr) 2342{ 2343 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 2344 bool in_first_chunk = false; 2345 unsigned long first_low, first_high; 2346 unsigned int cpu; 2347 2348 /* 2349 * The following test on unit_low/high isn't strictly 2350 * necessary but will speed up lookups of addresses which 2351 * aren't in the first chunk. 2352 * 2353 * The address check is against full chunk sizes. pcpu_base_addr 2354 * points to the beginning of the first chunk including the 2355 * static region. Assumes good intent as the first chunk may 2356 * not be full (ie. < pcpu_unit_pages in size). 2357 */ 2358 first_low = (unsigned long)pcpu_base_addr + 2359 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 2360 first_high = (unsigned long)pcpu_base_addr + 2361 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 2362 if ((unsigned long)addr >= first_low && 2363 (unsigned long)addr < first_high) { 2364 for_each_possible_cpu(cpu) { 2365 void *start = per_cpu_ptr(base, cpu); 2366 2367 if (addr >= start && addr < start + pcpu_unit_size) { 2368 in_first_chunk = true; 2369 break; 2370 } 2371 } 2372 } 2373 2374 if (in_first_chunk) { 2375 if (!is_vmalloc_addr(addr)) 2376 return __pa(addr); 2377 else 2378 return page_to_phys(vmalloc_to_page(addr)) + 2379 offset_in_page(addr); 2380 } else 2381 return page_to_phys(pcpu_addr_to_page(addr)) + 2382 offset_in_page(addr); 2383} 2384 2385/** 2386 * pcpu_alloc_alloc_info - allocate percpu allocation info 2387 * @nr_groups: the number of groups 2388 * @nr_units: the number of units 2389 * 2390 * Allocate ai which is large enough for @nr_groups groups containing 2391 * @nr_units units. The returned ai's groups[0].cpu_map points to the 2392 * cpu_map array which is long enough for @nr_units and filled with 2393 * NR_CPUS. It's the caller's responsibility to initialize cpu_map 2394 * pointer of other groups. 2395 * 2396 * RETURNS: 2397 * Pointer to the allocated pcpu_alloc_info on success, NULL on 2398 * failure. 2399 */ 2400struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 2401 int nr_units) 2402{ 2403 struct pcpu_alloc_info *ai; 2404 size_t base_size, ai_size; 2405 void *ptr; 2406 int unit; 2407 2408 base_size = ALIGN(struct_size(ai, groups, nr_groups), 2409 __alignof__(ai->groups[0].cpu_map[0])); 2410 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 2411 2412 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE); 2413 if (!ptr) 2414 return NULL; 2415 ai = ptr; 2416 ptr += base_size; 2417 2418 ai->groups[0].cpu_map = ptr; 2419 2420 for (unit = 0; unit < nr_units; unit++) 2421 ai->groups[0].cpu_map[unit] = NR_CPUS; 2422 2423 ai->nr_groups = nr_groups; 2424 ai->__ai_size = PFN_ALIGN(ai_size); 2425 2426 return ai; 2427} 2428 2429/** 2430 * pcpu_free_alloc_info - free percpu allocation info 2431 * @ai: pcpu_alloc_info to free 2432 * 2433 * Free @ai which was allocated by pcpu_alloc_alloc_info(). 2434 */ 2435void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 2436{ 2437 memblock_free(ai, ai->__ai_size); 2438} 2439 2440/** 2441 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 2442 * @lvl: loglevel 2443 * @ai: allocation info to dump 2444 * 2445 * Print out information about @ai using loglevel @lvl. 2446 */ 2447static void pcpu_dump_alloc_info(const char *lvl, 2448 const struct pcpu_alloc_info *ai) 2449{ 2450 int group_width = 1, cpu_width = 1, width; 2451 char empty_str[] = "--------"; 2452 int alloc = 0, alloc_end = 0; 2453 int group, v; 2454 int upa, apl; /* units per alloc, allocs per line */ 2455 2456 v = ai->nr_groups; 2457 while (v /= 10) 2458 group_width++; 2459 2460 v = num_possible_cpus(); 2461 while (v /= 10) 2462 cpu_width++; 2463 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 2464 2465 upa = ai->alloc_size / ai->unit_size; 2466 width = upa * (cpu_width + 1) + group_width + 3; 2467 apl = rounddown_pow_of_two(max(60 / width, 1)); 2468 2469 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 2470 lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 2471 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 2472 2473 for (group = 0; group < ai->nr_groups; group++) { 2474 const struct pcpu_group_info *gi = &ai->groups[group]; 2475 int unit = 0, unit_end = 0; 2476 2477 BUG_ON(gi->nr_units % upa); 2478 for (alloc_end += gi->nr_units / upa; 2479 alloc < alloc_end; alloc++) { 2480 if (!(alloc % apl)) { 2481 pr_cont("\n"); 2482 printk("%spcpu-alloc: ", lvl); 2483 } 2484 pr_cont("[%0*d] ", group_width, group); 2485 2486 for (unit_end += upa; unit < unit_end; unit++) 2487 if (gi->cpu_map[unit] != NR_CPUS) 2488 pr_cont("%0*d ", 2489 cpu_width, gi->cpu_map[unit]); 2490 else 2491 pr_cont("%s ", empty_str); 2492 } 2493 } 2494 pr_cont("\n"); 2495} 2496 2497/** 2498 * pcpu_setup_first_chunk - initialize the first percpu chunk 2499 * @ai: pcpu_alloc_info describing how to percpu area is shaped 2500 * @base_addr: mapped address 2501 * 2502 * Initialize the first percpu chunk which contains the kernel static 2503 * percpu area. This function is to be called from arch percpu area 2504 * setup path. 2505 * 2506 * @ai contains all information necessary to initialize the first 2507 * chunk and prime the dynamic percpu allocator. 2508 * 2509 * @ai->static_size is the size of static percpu area. 2510 * 2511 * @ai->reserved_size, if non-zero, specifies the amount of bytes to 2512 * reserve after the static area in the first chunk. This reserves 2513 * the first chunk such that it's available only through reserved 2514 * percpu allocation. This is primarily used to serve module percpu 2515 * static areas on architectures where the addressing model has 2516 * limited offset range for symbol relocations to guarantee module 2517 * percpu symbols fall inside the relocatable range. 2518 * 2519 * @ai->dyn_size determines the number of bytes available for dynamic 2520 * allocation in the first chunk. The area between @ai->static_size + 2521 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 2522 * 2523 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 2524 * and equal to or larger than @ai->static_size + @ai->reserved_size + 2525 * @ai->dyn_size. 2526 * 2527 * @ai->atom_size is the allocation atom size and used as alignment 2528 * for vm areas. 2529 * 2530 * @ai->alloc_size is the allocation size and always multiple of 2531 * @ai->atom_size. This is larger than @ai->atom_size if 2532 * @ai->unit_size is larger than @ai->atom_size. 2533 * 2534 * @ai->nr_groups and @ai->groups describe virtual memory layout of 2535 * percpu areas. Units which should be colocated are put into the 2536 * same group. Dynamic VM areas will be allocated according to these 2537 * groupings. If @ai->nr_groups is zero, a single group containing 2538 * all units is assumed. 2539 * 2540 * The caller should have mapped the first chunk at @base_addr and 2541 * copied static data to each unit. 2542 * 2543 * The first chunk will always contain a static and a dynamic region. 2544 * However, the static region is not managed by any chunk. If the first 2545 * chunk also contains a reserved region, it is served by two chunks - 2546 * one for the reserved region and one for the dynamic region. They 2547 * share the same vm, but use offset regions in the area allocation map. 2548 * The chunk serving the dynamic region is circulated in the chunk slots 2549 * and available for dynamic allocation like any other chunk. 2550 */ 2551void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 2552 void *base_addr) 2553{ 2554 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2555 size_t static_size, dyn_size; 2556 unsigned long *group_offsets; 2557 size_t *group_sizes; 2558 unsigned long *unit_off; 2559 unsigned int cpu; 2560 int *unit_map; 2561 int group, unit, i; 2562 unsigned long tmp_addr; 2563 size_t alloc_size; 2564 2565#define PCPU_SETUP_BUG_ON(cond) do { \ 2566 if (unlikely(cond)) { \ 2567 pr_emerg("failed to initialize, %s\n", #cond); \ 2568 pr_emerg("cpu_possible_mask=%*pb\n", \ 2569 cpumask_pr_args(cpu_possible_mask)); \ 2570 pcpu_dump_alloc_info(KERN_EMERG, ai); \ 2571 BUG(); \ 2572 } \ 2573} while (0) 2574 2575 /* sanity checks */ 2576 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 2577#ifdef CONFIG_SMP 2578 PCPU_SETUP_BUG_ON(!ai->static_size); 2579 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 2580#endif 2581 PCPU_SETUP_BUG_ON(!base_addr); 2582 PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 2583 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 2584 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 2585 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 2586 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 2587 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 2588 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 2589 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 2590 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 2591 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 2592 2593 /* process group information and build config tables accordingly */ 2594 alloc_size = ai->nr_groups * sizeof(group_offsets[0]); 2595 group_offsets = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 2596 2597 alloc_size = ai->nr_groups * sizeof(group_sizes[0]); 2598 group_sizes = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 2599 2600 alloc_size = nr_cpu_ids * sizeof(unit_map[0]); 2601 unit_map = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 2602 2603 alloc_size = nr_cpu_ids * sizeof(unit_off[0]); 2604 unit_off = memblock_alloc_or_panic(alloc_size, SMP_CACHE_BYTES); 2605 2606 for (cpu = 0; cpu < nr_cpu_ids; cpu++) 2607 unit_map[cpu] = UINT_MAX; 2608 2609 pcpu_low_unit_cpu = NR_CPUS; 2610 pcpu_high_unit_cpu = NR_CPUS; 2611 2612 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 2613 const struct pcpu_group_info *gi = &ai->groups[group]; 2614 2615 group_offsets[group] = gi->base_offset; 2616 group_sizes[group] = gi->nr_units * ai->unit_size; 2617 2618 for (i = 0; i < gi->nr_units; i++) { 2619 cpu = gi->cpu_map[i]; 2620 if (cpu == NR_CPUS) 2621 continue; 2622 2623 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 2624 PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 2625 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 2626 2627 unit_map[cpu] = unit + i; 2628 unit_off[cpu] = gi->base_offset + i * ai->unit_size; 2629 2630 /* determine low/high unit_cpu */ 2631 if (pcpu_low_unit_cpu == NR_CPUS || 2632 unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 2633 pcpu_low_unit_cpu = cpu; 2634 if (pcpu_high_unit_cpu == NR_CPUS || 2635 unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 2636 pcpu_high_unit_cpu = cpu; 2637 } 2638 } 2639 pcpu_nr_units = unit; 2640 2641 for_each_possible_cpu(cpu) 2642 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 2643 2644 /* we're done parsing the input, undefine BUG macro and dump config */ 2645#undef PCPU_SETUP_BUG_ON 2646 pcpu_dump_alloc_info(KERN_DEBUG, ai); 2647 2648 pcpu_nr_groups = ai->nr_groups; 2649 pcpu_group_offsets = group_offsets; 2650 pcpu_group_sizes = group_sizes; 2651 pcpu_unit_map = unit_map; 2652 pcpu_unit_offsets = unit_off; 2653 2654 /* determine basic parameters */ 2655 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 2656 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 2657 pcpu_atom_size = ai->atom_size; 2658 pcpu_chunk_struct_size = struct_size((struct pcpu_chunk *)0, populated, 2659 BITS_TO_LONGS(pcpu_unit_pages)); 2660 2661 pcpu_stats_save_ai(ai); 2662 2663 /* 2664 * Allocate chunk slots. The slots after the active slots are: 2665 * sidelined_slot - isolated, depopulated chunks 2666 * free_slot - fully free chunks 2667 * to_depopulate_slot - isolated, chunks to depopulate 2668 */ 2669 pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1; 2670 pcpu_free_slot = pcpu_sidelined_slot + 1; 2671 pcpu_to_depopulate_slot = pcpu_free_slot + 1; 2672 pcpu_nr_slots = pcpu_to_depopulate_slot + 1; 2673 pcpu_chunk_lists = memblock_alloc_or_panic(pcpu_nr_slots * 2674 sizeof(pcpu_chunk_lists[0]), 2675 SMP_CACHE_BYTES); 2676 2677 for (i = 0; i < pcpu_nr_slots; i++) 2678 INIT_LIST_HEAD(&pcpu_chunk_lists[i]); 2679 2680 /* 2681 * The end of the static region needs to be aligned with the 2682 * minimum allocation size as this offsets the reserved and 2683 * dynamic region. The first chunk ends page aligned by 2684 * expanding the dynamic region, therefore the dynamic region 2685 * can be shrunk to compensate while still staying above the 2686 * configured sizes. 2687 */ 2688 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 2689 dyn_size = ai->dyn_size - (static_size - ai->static_size); 2690 2691 /* 2692 * Initialize first chunk: 2693 * This chunk is broken up into 3 parts: 2694 * < static | [reserved] | dynamic > 2695 * - static - there is no backing chunk because these allocations can 2696 * never be freed. 2697 * - reserved (pcpu_reserved_chunk) - exists primarily to serve 2698 * allocations from module load. 2699 * - dynamic (pcpu_first_chunk) - serves the dynamic part of the first 2700 * chunk. 2701 */ 2702 tmp_addr = (unsigned long)base_addr + static_size; 2703 if (ai->reserved_size) 2704 pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr, 2705 ai->reserved_size); 2706 tmp_addr = (unsigned long)base_addr + static_size + ai->reserved_size; 2707 pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, dyn_size); 2708 2709 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 2710 pcpu_chunk_relocate(pcpu_first_chunk, -1); 2711 2712 /* include all regions of the first chunk */ 2713 pcpu_nr_populated += PFN_DOWN(size_sum); 2714 2715 pcpu_stats_chunk_alloc(); 2716 trace_percpu_create_chunk(base_addr); 2717 2718 /* we're done */ 2719 pcpu_base_addr = base_addr; 2720} 2721 2722#ifdef CONFIG_SMP 2723 2724const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 2725 [PCPU_FC_AUTO] = "auto", 2726 [PCPU_FC_EMBED] = "embed", 2727 [PCPU_FC_PAGE] = "page", 2728}; 2729 2730enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 2731 2732static int __init percpu_alloc_setup(char *str) 2733{ 2734 if (!str) 2735 return -EINVAL; 2736 2737 if (0) 2738 /* nada */; 2739#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 2740 else if (!strcmp(str, "embed")) 2741 pcpu_chosen_fc = PCPU_FC_EMBED; 2742#endif 2743#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2744 else if (!strcmp(str, "page")) 2745 pcpu_chosen_fc = PCPU_FC_PAGE; 2746#endif 2747 else 2748 pr_warn("unknown allocator %s specified\n", str); 2749 2750 return 0; 2751} 2752early_param("percpu_alloc", percpu_alloc_setup); 2753 2754/* 2755 * pcpu_embed_first_chunk() is used by the generic percpu setup. 2756 * Build it if needed by the arch config or the generic setup is going 2757 * to be used. 2758 */ 2759#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 2760 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 2761#define BUILD_EMBED_FIRST_CHUNK 2762#endif 2763 2764/* build pcpu_page_first_chunk() iff needed by the arch config */ 2765#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 2766#define BUILD_PAGE_FIRST_CHUNK 2767#endif 2768 2769/* pcpu_build_alloc_info() is used by both embed and page first chunk */ 2770#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 2771/** 2772 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2773 * @reserved_size: the size of reserved percpu area in bytes 2774 * @dyn_size: minimum free size for dynamic allocation in bytes 2775 * @atom_size: allocation atom size 2776 * @cpu_distance_fn: callback to determine distance between cpus, optional 2777 * 2778 * This function determines grouping of units, their mappings to cpus 2779 * and other parameters considering needed percpu size, allocation 2780 * atom size and distances between CPUs. 2781 * 2782 * Groups are always multiples of atom size and CPUs which are of 2783 * LOCAL_DISTANCE both ways are grouped together and share space for 2784 * units in the same group. The returned configuration is guaranteed 2785 * to have CPUs on different nodes on different groups and >=75% usage 2786 * of allocated virtual address space. 2787 * 2788 * RETURNS: 2789 * On success, pointer to the new allocation_info is returned. On 2790 * failure, ERR_PTR value is returned. 2791 */ 2792static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info( 2793 size_t reserved_size, size_t dyn_size, 2794 size_t atom_size, 2795 pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2796{ 2797 static int group_map[NR_CPUS] __initdata; 2798 static int group_cnt[NR_CPUS] __initdata; 2799 static struct cpumask mask __initdata; 2800 const size_t static_size = __per_cpu_end - __per_cpu_start; 2801 int nr_groups = 1, nr_units = 0; 2802 size_t size_sum, min_unit_size, alloc_size; 2803 int upa, max_upa, best_upa; /* units_per_alloc */ 2804 int last_allocs, group, unit; 2805 unsigned int cpu, tcpu; 2806 struct pcpu_alloc_info *ai; 2807 unsigned int *cpu_map; 2808 2809 /* this function may be called multiple times */ 2810 memset(group_map, 0, sizeof(group_map)); 2811 memset(group_cnt, 0, sizeof(group_cnt)); 2812 cpumask_clear(&mask); 2813 2814 /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2815 size_sum = PFN_ALIGN(static_size + reserved_size + 2816 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2817 dyn_size = size_sum - static_size - reserved_size; 2818 2819 /* 2820 * Determine min_unit_size, alloc_size and max_upa such that 2821 * alloc_size is multiple of atom_size and is the smallest 2822 * which can accommodate 4k aligned segments which are equal to 2823 * or larger than min_unit_size. 2824 */ 2825 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2826 2827 /* determine the maximum # of units that can fit in an allocation */ 2828 alloc_size = roundup(min_unit_size, atom_size); 2829 upa = alloc_size / min_unit_size; 2830 while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2831 upa--; 2832 max_upa = upa; 2833 2834 cpumask_copy(&mask, cpu_possible_mask); 2835 2836 /* group cpus according to their proximity */ 2837 for (group = 0; !cpumask_empty(&mask); group++) { 2838 /* pop the group's first cpu */ 2839 cpu = cpumask_first(&mask); 2840 group_map[cpu] = group; 2841 group_cnt[group]++; 2842 cpumask_clear_cpu(cpu, &mask); 2843 2844 for_each_cpu(tcpu, &mask) { 2845 if (!cpu_distance_fn || 2846 (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE && 2847 cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) { 2848 group_map[tcpu] = group; 2849 group_cnt[group]++; 2850 cpumask_clear_cpu(tcpu, &mask); 2851 } 2852 } 2853 } 2854 nr_groups = group; 2855 2856 /* 2857 * Wasted space is caused by a ratio imbalance of upa to group_cnt. 2858 * Expand the unit_size until we use >= 75% of the units allocated. 2859 * Related to atom_size, which could be much larger than the unit_size. 2860 */ 2861 last_allocs = INT_MAX; 2862 best_upa = 0; 2863 for (upa = max_upa; upa; upa--) { 2864 int allocs = 0, wasted = 0; 2865 2866 if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2867 continue; 2868 2869 for (group = 0; group < nr_groups; group++) { 2870 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2871 allocs += this_allocs; 2872 wasted += this_allocs * upa - group_cnt[group]; 2873 } 2874 2875 /* 2876 * Don't accept if wastage is over 1/3. The 2877 * greater-than comparison ensures upa==1 always 2878 * passes the following check. 2879 */ 2880 if (wasted > num_possible_cpus() / 3) 2881 continue; 2882 2883 /* and then don't consume more memory */ 2884 if (allocs > last_allocs) 2885 break; 2886 last_allocs = allocs; 2887 best_upa = upa; 2888 } 2889 BUG_ON(!best_upa); 2890 upa = best_upa; 2891 2892 /* allocate and fill alloc_info */ 2893 for (group = 0; group < nr_groups; group++) 2894 nr_units += roundup(group_cnt[group], upa); 2895 2896 ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2897 if (!ai) 2898 return ERR_PTR(-ENOMEM); 2899 cpu_map = ai->groups[0].cpu_map; 2900 2901 for (group = 0; group < nr_groups; group++) { 2902 ai->groups[group].cpu_map = cpu_map; 2903 cpu_map += roundup(group_cnt[group], upa); 2904 } 2905 2906 ai->static_size = static_size; 2907 ai->reserved_size = reserved_size; 2908 ai->dyn_size = dyn_size; 2909 ai->unit_size = alloc_size / upa; 2910 ai->atom_size = atom_size; 2911 ai->alloc_size = alloc_size; 2912 2913 for (group = 0, unit = 0; group < nr_groups; group++) { 2914 struct pcpu_group_info *gi = &ai->groups[group]; 2915 2916 /* 2917 * Initialize base_offset as if all groups are located 2918 * back-to-back. The caller should update this to 2919 * reflect actual allocation. 2920 */ 2921 gi->base_offset = unit * ai->unit_size; 2922 2923 for_each_possible_cpu(cpu) 2924 if (group_map[cpu] == group) 2925 gi->cpu_map[gi->nr_units++] = cpu; 2926 gi->nr_units = roundup(gi->nr_units, upa); 2927 unit += gi->nr_units; 2928 } 2929 BUG_ON(unit != nr_units); 2930 2931 return ai; 2932} 2933 2934static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align, 2935 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn) 2936{ 2937 const unsigned long goal = __pa(MAX_DMA_ADDRESS); 2938#ifdef CONFIG_NUMA 2939 int node = NUMA_NO_NODE; 2940 void *ptr; 2941 2942 if (cpu_to_nd_fn) 2943 node = cpu_to_nd_fn(cpu); 2944 2945 if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) { 2946 ptr = memblock_alloc_from(size, align, goal); 2947 pr_info("cpu %d has no node %d or node-local memory\n", 2948 cpu, node); 2949 pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n", 2950 cpu, size, (u64)__pa(ptr)); 2951 } else { 2952 ptr = memblock_alloc_try_nid(size, align, goal, 2953 MEMBLOCK_ALLOC_ACCESSIBLE, 2954 node); 2955 2956 pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n", 2957 cpu, size, node, (u64)__pa(ptr)); 2958 } 2959 return ptr; 2960#else 2961 return memblock_alloc_from(size, align, goal); 2962#endif 2963} 2964 2965static void __init pcpu_fc_free(void *ptr, size_t size) 2966{ 2967 memblock_free(ptr, size); 2968} 2969#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2970 2971#if defined(BUILD_EMBED_FIRST_CHUNK) 2972/** 2973 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 2974 * @reserved_size: the size of reserved percpu area in bytes 2975 * @dyn_size: minimum free size for dynamic allocation in bytes 2976 * @atom_size: allocation atom size 2977 * @cpu_distance_fn: callback to determine distance between cpus, optional 2978 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional 2979 * 2980 * This is a helper to ease setting up embedded first percpu chunk and 2981 * can be called where pcpu_setup_first_chunk() is expected. 2982 * 2983 * If this function is used to setup the first chunk, it is allocated 2984 * by calling pcpu_fc_alloc and used as-is without being mapped into 2985 * vmalloc area. Allocations are always whole multiples of @atom_size 2986 * aligned to @atom_size. 2987 * 2988 * This enables the first chunk to piggy back on the linear physical 2989 * mapping which often uses larger page size. Please note that this 2990 * can result in very sparse cpu->unit mapping on NUMA machines thus 2991 * requiring large vmalloc address space. Don't use this allocator if 2992 * vmalloc space is not orders of magnitude larger than distances 2993 * between node memory addresses (ie. 32bit NUMA machines). 2994 * 2995 * @dyn_size specifies the minimum dynamic area size. 2996 * 2997 * If the needed size is smaller than the minimum or specified unit 2998 * size, the leftover is returned using pcpu_fc_free. 2999 * 3000 * RETURNS: 3001 * 0 on success, -errno on failure. 3002 */ 3003int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 3004 size_t atom_size, 3005 pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 3006 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn) 3007{ 3008 void *base = (void *)ULONG_MAX; 3009 void **areas = NULL; 3010 struct pcpu_alloc_info *ai; 3011 size_t size_sum, areas_size; 3012 unsigned long max_distance; 3013 int group, i, highest_group, rc = 0; 3014 3015 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 3016 cpu_distance_fn); 3017 if (IS_ERR(ai)) 3018 return PTR_ERR(ai); 3019 3020 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 3021 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 3022 3023 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES); 3024 if (!areas) { 3025 rc = -ENOMEM; 3026 goto out_free; 3027 } 3028 3029 /* allocate, copy and determine base address & max_distance */ 3030 highest_group = 0; 3031 for (group = 0; group < ai->nr_groups; group++) { 3032 struct pcpu_group_info *gi = &ai->groups[group]; 3033 unsigned int cpu = NR_CPUS; 3034 void *ptr; 3035 3036 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 3037 cpu = gi->cpu_map[i]; 3038 BUG_ON(cpu == NR_CPUS); 3039 3040 /* allocate space for the whole group */ 3041 ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn); 3042 if (!ptr) { 3043 rc = -ENOMEM; 3044 goto out_free_areas; 3045 } 3046 /* kmemleak tracks the percpu allocations separately */ 3047 kmemleak_ignore_phys(__pa(ptr)); 3048 areas[group] = ptr; 3049 3050 base = min(ptr, base); 3051 if (ptr > areas[highest_group]) 3052 highest_group = group; 3053 } 3054 max_distance = areas[highest_group] - base; 3055 max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 3056 3057 /* warn if maximum distance is further than 75% of vmalloc space */ 3058 if (max_distance > VMALLOC_TOTAL * 3 / 4) { 3059 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 3060 max_distance, VMALLOC_TOTAL); 3061#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 3062 /* and fail if we have fallback */ 3063 rc = -EINVAL; 3064 goto out_free_areas; 3065#endif 3066 } 3067 3068 /* 3069 * Copy data and free unused parts. This should happen after all 3070 * allocations are complete; otherwise, we may end up with 3071 * overlapping groups. 3072 */ 3073 for (group = 0; group < ai->nr_groups; group++) { 3074 struct pcpu_group_info *gi = &ai->groups[group]; 3075 void *ptr = areas[group]; 3076 3077 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 3078 if (gi->cpu_map[i] == NR_CPUS) { 3079 /* unused unit, free whole */ 3080 pcpu_fc_free(ptr, ai->unit_size); 3081 continue; 3082 } 3083 /* copy and return the unused part */ 3084 memcpy(ptr, __per_cpu_start, ai->static_size); 3085 pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum); 3086 } 3087 } 3088 3089 /* base address is now known, determine group base offsets */ 3090 for (group = 0; group < ai->nr_groups; group++) { 3091 ai->groups[group].base_offset = areas[group] - base; 3092 } 3093 3094 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n", 3095 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size, 3096 ai->dyn_size, ai->unit_size); 3097 3098 pcpu_setup_first_chunk(ai, base); 3099 goto out_free; 3100 3101out_free_areas: 3102 for (group = 0; group < ai->nr_groups; group++) 3103 if (areas[group]) 3104 pcpu_fc_free(areas[group], 3105 ai->groups[group].nr_units * ai->unit_size); 3106out_free: 3107 pcpu_free_alloc_info(ai); 3108 if (areas) 3109 memblock_free(areas, areas_size); 3110 return rc; 3111} 3112#endif /* BUILD_EMBED_FIRST_CHUNK */ 3113 3114#ifdef BUILD_PAGE_FIRST_CHUNK 3115#include <linux/pgalloc.h> 3116 3117#ifndef P4D_TABLE_SIZE 3118#define P4D_TABLE_SIZE PAGE_SIZE 3119#endif 3120 3121#ifndef PUD_TABLE_SIZE 3122#define PUD_TABLE_SIZE PAGE_SIZE 3123#endif 3124 3125#ifndef PMD_TABLE_SIZE 3126#define PMD_TABLE_SIZE PAGE_SIZE 3127#endif 3128 3129#ifndef PTE_TABLE_SIZE 3130#define PTE_TABLE_SIZE PAGE_SIZE 3131#endif 3132void __init __weak pcpu_populate_pte(unsigned long addr) 3133{ 3134 pgd_t *pgd = pgd_offset_k(addr); 3135 p4d_t *p4d; 3136 pud_t *pud; 3137 pmd_t *pmd; 3138 3139 if (pgd_none(*pgd)) { 3140 p4d = memblock_alloc_or_panic(P4D_TABLE_SIZE, P4D_TABLE_SIZE); 3141 pgd_populate_kernel(addr, pgd, p4d); 3142 } 3143 3144 p4d = p4d_offset(pgd, addr); 3145 if (p4d_none(*p4d)) { 3146 pud = memblock_alloc_or_panic(PUD_TABLE_SIZE, PUD_TABLE_SIZE); 3147 p4d_populate_kernel(addr, p4d, pud); 3148 } 3149 3150 pud = pud_offset(p4d, addr); 3151 if (pud_none(*pud)) { 3152 pmd = memblock_alloc_or_panic(PMD_TABLE_SIZE, PMD_TABLE_SIZE); 3153 pud_populate(&init_mm, pud, pmd); 3154 } 3155 3156 pmd = pmd_offset(pud, addr); 3157 if (!pmd_present(*pmd)) { 3158 pte_t *new; 3159 3160 new = memblock_alloc_or_panic(PTE_TABLE_SIZE, PTE_TABLE_SIZE); 3161 pmd_populate_kernel(&init_mm, pmd, new); 3162 } 3163 3164 return; 3165} 3166 3167/** 3168 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 3169 * @reserved_size: the size of reserved percpu area in bytes 3170 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional 3171 * 3172 * This is a helper to ease setting up page-remapped first percpu 3173 * chunk and can be called where pcpu_setup_first_chunk() is expected. 3174 * 3175 * This is the basic allocator. Static percpu area is allocated 3176 * page-by-page into vmalloc area. 3177 * 3178 * RETURNS: 3179 * 0 on success, -errno on failure. 3180 */ 3181int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn) 3182{ 3183 static struct vm_struct vm; 3184 struct pcpu_alloc_info *ai; 3185 char psize_str[16]; 3186 int unit_pages; 3187 size_t pages_size; 3188 struct page **pages; 3189 int unit, i, j, rc = 0; 3190 int upa; 3191 int nr_g0_units; 3192 3193 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 3194 3195 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 3196 if (IS_ERR(ai)) 3197 return PTR_ERR(ai); 3198 BUG_ON(ai->nr_groups != 1); 3199 upa = ai->alloc_size/ai->unit_size; 3200 nr_g0_units = roundup(num_possible_cpus(), upa); 3201 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) { 3202 pcpu_free_alloc_info(ai); 3203 return -EINVAL; 3204 } 3205 3206 unit_pages = ai->unit_size >> PAGE_SHIFT; 3207 3208 /* unaligned allocations can't be freed, round up to page size */ 3209 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 3210 sizeof(pages[0])); 3211 pages = memblock_alloc_or_panic(pages_size, SMP_CACHE_BYTES); 3212 3213 /* allocate pages */ 3214 j = 0; 3215 for (unit = 0; unit < num_possible_cpus(); unit++) { 3216 unsigned int cpu = ai->groups[0].cpu_map[unit]; 3217 for (i = 0; i < unit_pages; i++) { 3218 void *ptr; 3219 3220 ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn); 3221 if (!ptr) { 3222 pr_warn("failed to allocate %s page for cpu%u\n", 3223 psize_str, cpu); 3224 goto enomem; 3225 } 3226 /* kmemleak tracks the percpu allocations separately */ 3227 kmemleak_ignore_phys(__pa(ptr)); 3228 pages[j++] = virt_to_page(ptr); 3229 } 3230 } 3231 3232 /* allocate vm area, map the pages and copy static data */ 3233 vm.flags = VM_ALLOC; 3234 vm.size = num_possible_cpus() * ai->unit_size; 3235 vm_area_register_early(&vm, PAGE_SIZE); 3236 3237 for (unit = 0; unit < num_possible_cpus(); unit++) { 3238 unsigned long unit_addr = 3239 (unsigned long)vm.addr + unit * ai->unit_size; 3240 3241 for (i = 0; i < unit_pages; i++) 3242 pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT)); 3243 3244 /* pte already populated, the following shouldn't fail */ 3245 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 3246 unit_pages); 3247 if (rc < 0) 3248 panic("failed to map percpu area, err=%d\n", rc); 3249 3250 flush_cache_vmap_early(unit_addr, unit_addr + ai->unit_size); 3251 3252 /* copy static data */ 3253 memcpy((void *)unit_addr, __per_cpu_start, ai->static_size); 3254 } 3255 3256 /* we're ready, commit */ 3257 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n", 3258 unit_pages, psize_str, ai->static_size, 3259 ai->reserved_size, ai->dyn_size); 3260 3261 pcpu_setup_first_chunk(ai, vm.addr); 3262 goto out_free_ar; 3263 3264enomem: 3265 while (--j >= 0) 3266 pcpu_fc_free(page_address(pages[j]), PAGE_SIZE); 3267 rc = -ENOMEM; 3268out_free_ar: 3269 memblock_free(pages, pages_size); 3270 pcpu_free_alloc_info(ai); 3271 return rc; 3272} 3273#endif /* BUILD_PAGE_FIRST_CHUNK */ 3274 3275#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 3276/* 3277 * Generic SMP percpu area setup. 3278 * 3279 * The embedding helper is used because its behavior closely resembles 3280 * the original non-dynamic generic percpu area setup. This is 3281 * important because many archs have addressing restrictions and might 3282 * fail if the percpu area is located far away from the previous 3283 * location. As an added bonus, in non-NUMA cases, embedding is 3284 * generally a good idea TLB-wise because percpu area can piggy back 3285 * on the physical linear memory mapping which uses large page 3286 * mappings on applicable archs. 3287 */ 3288unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 3289EXPORT_SYMBOL(__per_cpu_offset); 3290 3291void __init setup_per_cpu_areas(void) 3292{ 3293 unsigned long delta; 3294 unsigned int cpu; 3295 int rc; 3296 3297 /* 3298 * Always reserve area for module percpu variables. That's 3299 * what the legacy allocator did. 3300 */ 3301 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE, 3302 PAGE_SIZE, NULL, NULL); 3303 if (rc < 0) 3304 panic("Failed to initialize percpu areas."); 3305 3306 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 3307 for_each_possible_cpu(cpu) 3308 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 3309} 3310#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 3311 3312#else /* CONFIG_SMP */ 3313 3314/* 3315 * UP percpu area setup. 3316 * 3317 * UP always uses km-based percpu allocator with identity mapping. 3318 * Static percpu variables are indistinguishable from the usual static 3319 * variables and don't require any special preparation. 3320 */ 3321void __init setup_per_cpu_areas(void) 3322{ 3323 const size_t unit_size = 3324 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 3325 PERCPU_DYNAMIC_RESERVE)); 3326 struct pcpu_alloc_info *ai; 3327 void *fc; 3328 3329 ai = pcpu_alloc_alloc_info(1, 1); 3330 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); 3331 if (!ai || !fc) 3332 panic("Failed to allocate memory for percpu areas."); 3333 /* kmemleak tracks the percpu allocations separately */ 3334 kmemleak_ignore_phys(__pa(fc)); 3335 3336 ai->dyn_size = unit_size; 3337 ai->unit_size = unit_size; 3338 ai->atom_size = unit_size; 3339 ai->alloc_size = unit_size; 3340 ai->groups[0].nr_units = 1; 3341 ai->groups[0].cpu_map[0] = 0; 3342 3343 pcpu_setup_first_chunk(ai, fc); 3344 pcpu_free_alloc_info(ai); 3345} 3346 3347#endif /* CONFIG_SMP */ 3348 3349/* 3350 * pcpu_nr_pages - calculate total number of populated backing pages 3351 * 3352 * This reflects the number of pages populated to back chunks. Metadata is 3353 * excluded in the number exposed in meminfo as the number of backing pages 3354 * scales with the number of cpus and can quickly outweigh the memory used for 3355 * metadata. It also keeps this calculation nice and simple. 3356 * 3357 * RETURNS: 3358 * Total number of populated backing pages in use by the allocator. 3359 */ 3360unsigned long pcpu_nr_pages(void) 3361{ 3362 return data_race(READ_ONCE(pcpu_nr_populated)) * pcpu_nr_units; 3363} 3364 3365/* 3366 * Percpu allocator is initialized early during boot when neither slab or 3367 * workqueue is available. Plug async management until everything is up 3368 * and running. 3369 */ 3370static int __init percpu_enable_async(void) 3371{ 3372 pcpu_async_enabled = true; 3373 return 0; 3374} 3375subsys_initcall(percpu_enable_async);
~/
Read-only file preview.