1fbf59bc9STejun Heo /* 288999a89STejun Heo * mm/percpu.c - percpu memory allocator 3fbf59bc9STejun Heo * 4fbf59bc9STejun Heo * Copyright (C) 2009 SUSE Linux Products GmbH 5fbf59bc9STejun Heo * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 6fbf59bc9STejun Heo * 75e81ee3eSDennis Zhou (Facebook) * Copyright (C) 2017 Facebook Inc. 85e81ee3eSDennis Zhou (Facebook) * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com> 95e81ee3eSDennis Zhou (Facebook) * 109c015162SDennis Zhou (Facebook) * This file is released under the GPLv2 license. 11fbf59bc9STejun Heo * 129c015162SDennis Zhou (Facebook) * The percpu allocator handles both static and dynamic areas. Percpu 139c015162SDennis Zhou (Facebook) * areas are allocated in chunks which are divided into units. There is 149c015162SDennis Zhou (Facebook) * a 1-to-1 mapping for units to possible cpus. These units are grouped 159c015162SDennis Zhou (Facebook) * based on NUMA properties of the machine. 16fbf59bc9STejun Heo * 17fbf59bc9STejun Heo * c0 c1 c2 18fbf59bc9STejun Heo * ------------------- ------------------- ------------ 19fbf59bc9STejun Heo * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 20fbf59bc9STejun Heo * ------------------- ...... ------------------- .... ------------ 21fbf59bc9STejun Heo * 229c015162SDennis Zhou (Facebook) * Allocation is done by offsets into a unit's address space. Ie., an 239c015162SDennis Zhou (Facebook) * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, 249c015162SDennis Zhou (Facebook) * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear 259c015162SDennis Zhou (Facebook) * and even sparse. Access is handled by configuring percpu base 269c015162SDennis Zhou (Facebook) * registers according to the cpu to unit mappings and offsetting the 279c015162SDennis Zhou (Facebook) * base address using pcpu_unit_size. 28fbf59bc9STejun Heo * 299c015162SDennis Zhou (Facebook) * There is special consideration for the first chunk which must handle 309c015162SDennis Zhou (Facebook) * the static percpu variables in the kernel image as allocation services 315e81ee3eSDennis Zhou (Facebook) * are not online yet. In short, the first chunk is structured like so: 329c015162SDennis Zhou (Facebook) * 339c015162SDennis Zhou (Facebook) * <Static | [Reserved] | Dynamic> 349c015162SDennis Zhou (Facebook) * 359c015162SDennis Zhou (Facebook) * The static data is copied from the original section managed by the 369c015162SDennis Zhou (Facebook) * linker. The reserved section, if non-zero, primarily manages static 379c015162SDennis Zhou (Facebook) * percpu variables from kernel modules. Finally, the dynamic section 389c015162SDennis Zhou (Facebook) * takes care of normal allocations. 39fbf59bc9STejun Heo * 405e81ee3eSDennis Zhou (Facebook) * The allocator organizes chunks into lists according to free size and 415e81ee3eSDennis Zhou (Facebook) * tries to allocate from the fullest chunk first. Each chunk is managed 425e81ee3eSDennis Zhou (Facebook) * by a bitmap with metadata blocks. The allocation map is updated on 435e81ee3eSDennis Zhou (Facebook) * every allocation and free to reflect the current state while the boundary 445e81ee3eSDennis Zhou (Facebook) * map is only updated on allocation. Each metadata block contains 455e81ee3eSDennis Zhou (Facebook) * information to help mitigate the need to iterate over large portions 465e81ee3eSDennis Zhou (Facebook) * of the bitmap. The reverse mapping from page to chunk is stored in 475e81ee3eSDennis Zhou (Facebook) * the page's index. Lastly, units are lazily backed and grow in unison. 48fbf59bc9STejun Heo * 495e81ee3eSDennis Zhou (Facebook) * There is a unique conversion that goes on here between bytes and bits. 505e81ee3eSDennis Zhou (Facebook) * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk 515e81ee3eSDennis Zhou (Facebook) * tracks the number of pages it is responsible for in nr_pages. Helper 525e81ee3eSDennis Zhou (Facebook) * functions are used to convert from between the bytes, bits, and blocks. 535e81ee3eSDennis Zhou (Facebook) * All hints are managed in bits unless explicitly stated. 549c015162SDennis Zhou (Facebook) * 554091fb95SMasahiro Yamada * To use this allocator, arch code should do the following: 56fbf59bc9STejun Heo * 57fbf59bc9STejun Heo * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 58e0100983STejun Heo * regular address to percpu pointer and back if they need to be 59e0100983STejun Heo * different from the default 60fbf59bc9STejun Heo * 618d408b4bSTejun Heo * - use pcpu_setup_first_chunk() during percpu area initialization to 628d408b4bSTejun Heo * setup the first chunk containing the kernel static percpu area 63fbf59bc9STejun Heo */ 64fbf59bc9STejun Heo 65870d4b12SJoe Perches #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 66870d4b12SJoe Perches 67fbf59bc9STejun Heo #include <linux/bitmap.h> 6857c8a661SMike Rapoport #include <linux/memblock.h> 69fd1e8a1fSTejun Heo #include <linux/err.h> 70ca460b3cSDennis Zhou (Facebook) #include <linux/lcm.h> 71fbf59bc9STejun Heo #include <linux/list.h> 72a530b795STejun Heo #include <linux/log2.h> 73fbf59bc9STejun Heo #include <linux/mm.h> 74fbf59bc9STejun Heo #include <linux/module.h> 75fbf59bc9STejun Heo #include <linux/mutex.h> 76fbf59bc9STejun Heo #include <linux/percpu.h> 77fbf59bc9STejun Heo #include <linux/pfn.h> 78fbf59bc9STejun Heo #include <linux/slab.h> 79ccea34b5STejun Heo #include <linux/spinlock.h> 80fbf59bc9STejun Heo #include <linux/vmalloc.h> 81a56dbddfSTejun Heo #include <linux/workqueue.h> 82f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 8371546d10STejun Heo #include <linux/sched.h> 84fbf59bc9STejun Heo 85fbf59bc9STejun Heo #include <asm/cacheflush.h> 86e0100983STejun Heo #include <asm/sections.h> 87fbf59bc9STejun Heo #include <asm/tlbflush.h> 883b034b0dSVivek Goyal #include <asm/io.h> 89fbf59bc9STejun Heo 90df95e795SDennis Zhou #define CREATE_TRACE_POINTS 91df95e795SDennis Zhou #include <trace/events/percpu.h> 92df95e795SDennis Zhou 938fa3ed80SDennis Zhou #include "percpu-internal.h" 948fa3ed80SDennis Zhou 9540064aecSDennis Zhou (Facebook) /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ 9640064aecSDennis Zhou (Facebook) #define PCPU_SLOT_BASE_SHIFT 5 978744d859SDennis Zhou /* chunks in slots below this are subject to being sidelined on failed alloc */ 988744d859SDennis Zhou #define PCPU_SLOT_FAIL_THRESHOLD 3 9940064aecSDennis Zhou (Facebook) 1001a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_LOW 2 1011a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_HIGH 4 102fbf59bc9STejun Heo 103bbddff05STejun Heo #ifdef CONFIG_SMP 104e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 105e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 106e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 10743cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 10843cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 10943cf38ebSTejun Heo (unsigned long)__per_cpu_start) 110e0100983STejun Heo #endif 111e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 112e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 11343cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 11443cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 11543cf38ebSTejun Heo (unsigned long)__per_cpu_start) 116e0100983STejun Heo #endif 117bbddff05STejun Heo #else /* CONFIG_SMP */ 118bbddff05STejun Heo /* on UP, it's always identity mapped */ 119bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 120bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 121bbddff05STejun Heo #endif /* CONFIG_SMP */ 122e0100983STejun Heo 1231328710bSDaniel Micay static int pcpu_unit_pages __ro_after_init; 1241328710bSDaniel Micay static int pcpu_unit_size __ro_after_init; 1251328710bSDaniel Micay static int pcpu_nr_units __ro_after_init; 1261328710bSDaniel Micay static int pcpu_atom_size __ro_after_init; 1278fa3ed80SDennis Zhou int pcpu_nr_slots __ro_after_init; 1281328710bSDaniel Micay static size_t pcpu_chunk_struct_size __ro_after_init; 129fbf59bc9STejun Heo 130a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 1311328710bSDaniel Micay static unsigned int pcpu_low_unit_cpu __ro_after_init; 1321328710bSDaniel Micay static unsigned int pcpu_high_unit_cpu __ro_after_init; 1332f39e637STejun Heo 134fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 1351328710bSDaniel Micay void *pcpu_base_addr __ro_after_init; 136fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 137fbf59bc9STejun Heo 1381328710bSDaniel Micay static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ 1391328710bSDaniel Micay const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ 1402f39e637STejun Heo 1416563297cSTejun Heo /* group information, used for vm allocation */ 1421328710bSDaniel Micay static int pcpu_nr_groups __ro_after_init; 1431328710bSDaniel Micay static const unsigned long *pcpu_group_offsets __ro_after_init; 1441328710bSDaniel Micay static const size_t *pcpu_group_sizes __ro_after_init; 1456563297cSTejun Heo 146ae9e6bc9STejun Heo /* 147ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 148ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 149ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 150ae9e6bc9STejun Heo */ 1518fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_first_chunk __ro_after_init; 152ae9e6bc9STejun Heo 153ae9e6bc9STejun Heo /* 154ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 155e2266705SDennis Zhou (Facebook) * chunk and serves it for reserved allocations. When the reserved 156e2266705SDennis Zhou (Facebook) * region doesn't exist, the following variable is NULL. 157ae9e6bc9STejun Heo */ 1588fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; 159edcb4639STejun Heo 1608fa3ed80SDennis Zhou DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 1616710e594STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 162fbf59bc9STejun Heo 1638fa3ed80SDennis Zhou struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ 164fbf59bc9STejun Heo 1654f996e23STejun Heo /* chunks which need their map areas extended, protected by pcpu_lock */ 1664f996e23STejun Heo static LIST_HEAD(pcpu_map_extend_chunks); 1674f996e23STejun Heo 168b539b87fSTejun Heo /* 169b539b87fSTejun Heo * The number of empty populated pages, protected by pcpu_lock. The 170b539b87fSTejun Heo * reserved chunk doesn't contribute to the count. 171b539b87fSTejun Heo */ 1726b9b6f39SDennis Zhou (Facebook) int pcpu_nr_empty_pop_pages; 173b539b87fSTejun Heo 1741a4d7607STejun Heo /* 1757e8a6304SDennis Zhou (Facebook) * The number of populated pages in use by the allocator, protected by 1767e8a6304SDennis Zhou (Facebook) * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets 1777e8a6304SDennis Zhou (Facebook) * allocated/deallocated, it is allocated/deallocated in all units of a chunk 1787e8a6304SDennis Zhou (Facebook) * and increments/decrements this count by 1). 1797e8a6304SDennis Zhou (Facebook) */ 1807e8a6304SDennis Zhou (Facebook) static unsigned long pcpu_nr_populated; 1817e8a6304SDennis Zhou (Facebook) 1827e8a6304SDennis Zhou (Facebook) /* 1831a4d7607STejun Heo * Balance work is used to populate or destroy chunks asynchronously. We 1841a4d7607STejun Heo * try to keep the number of populated free pages between 1851a4d7607STejun Heo * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 1861a4d7607STejun Heo * empty chunk. 1871a4d7607STejun Heo */ 188fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work); 189fe6bd8c3STejun Heo static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 1901a4d7607STejun Heo static bool pcpu_async_enabled __read_mostly; 1911a4d7607STejun Heo static bool pcpu_atomic_alloc_failed; 1921a4d7607STejun Heo 1931a4d7607STejun Heo static void pcpu_schedule_balance_work(void) 1941a4d7607STejun Heo { 1951a4d7607STejun Heo if (pcpu_async_enabled) 1961a4d7607STejun Heo schedule_work(&pcpu_balance_work); 1971a4d7607STejun Heo } 198a56dbddfSTejun Heo 199c0ebfdc3SDennis Zhou (Facebook) /** 200560f2c23SDennis Zhou (Facebook) * pcpu_addr_in_chunk - check if the address is served from this chunk 201560f2c23SDennis Zhou (Facebook) * @chunk: chunk of interest 202560f2c23SDennis Zhou (Facebook) * @addr: percpu address 203c0ebfdc3SDennis Zhou (Facebook) * 204c0ebfdc3SDennis Zhou (Facebook) * RETURNS: 205560f2c23SDennis Zhou (Facebook) * True if the address is served from this chunk. 206c0ebfdc3SDennis Zhou (Facebook) */ 207560f2c23SDennis Zhou (Facebook) static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) 208020ec653STejun Heo { 209c0ebfdc3SDennis Zhou (Facebook) void *start_addr, *end_addr; 210020ec653STejun Heo 211560f2c23SDennis Zhou (Facebook) if (!chunk) 212c0ebfdc3SDennis Zhou (Facebook) return false; 213c0ebfdc3SDennis Zhou (Facebook) 214560f2c23SDennis Zhou (Facebook) start_addr = chunk->base_addr + chunk->start_offset; 215560f2c23SDennis Zhou (Facebook) end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - 216560f2c23SDennis Zhou (Facebook) chunk->end_offset; 217c0ebfdc3SDennis Zhou (Facebook) 218c0ebfdc3SDennis Zhou (Facebook) return addr >= start_addr && addr < end_addr; 219020ec653STejun Heo } 220020ec653STejun Heo 221d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 222fbf59bc9STejun Heo { 223cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 224fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 225fbf59bc9STejun Heo } 226fbf59bc9STejun Heo 227d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 228d9b55eebSTejun Heo { 229d9b55eebSTejun Heo if (size == pcpu_unit_size) 230d9b55eebSTejun Heo return pcpu_nr_slots - 1; 231d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 232d9b55eebSTejun Heo } 233d9b55eebSTejun Heo 234fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 235fbf59bc9STejun Heo { 23640064aecSDennis Zhou (Facebook) if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) 237fbf59bc9STejun Heo return 0; 238fbf59bc9STejun Heo 2393e54097bSDennis Zhou return pcpu_size_to_slot(chunk->contig_bits * PCPU_MIN_ALLOC_SIZE); 240fbf59bc9STejun Heo } 241fbf59bc9STejun Heo 24288999a89STejun Heo /* set the pointer to a chunk in a page struct */ 24388999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 24488999a89STejun Heo { 24588999a89STejun Heo page->index = (unsigned long)pcpu; 24688999a89STejun Heo } 24788999a89STejun Heo 24888999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 24988999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 25088999a89STejun Heo { 25188999a89STejun Heo return (struct pcpu_chunk *)page->index; 25288999a89STejun Heo } 25388999a89STejun Heo 25488999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 255fbf59bc9STejun Heo { 2562f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 257fbf59bc9STejun Heo } 258fbf59bc9STejun Heo 259c0ebfdc3SDennis Zhou (Facebook) static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 260c0ebfdc3SDennis Zhou (Facebook) { 261c0ebfdc3SDennis Zhou (Facebook) return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 262c0ebfdc3SDennis Zhou (Facebook) } 263c0ebfdc3SDennis Zhou (Facebook) 2649983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 265fbf59bc9STejun Heo unsigned int cpu, int page_idx) 266fbf59bc9STejun Heo { 267c0ebfdc3SDennis Zhou (Facebook) return (unsigned long)chunk->base_addr + 268c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(cpu, page_idx); 269fbf59bc9STejun Heo } 270fbf59bc9STejun Heo 27191e914c5SDennis Zhou (Facebook) static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) 272ce3141a2STejun Heo { 27391e914c5SDennis Zhou (Facebook) *rs = find_next_zero_bit(bitmap, end, *rs); 27491e914c5SDennis Zhou (Facebook) *re = find_next_bit(bitmap, end, *rs + 1); 275ce3141a2STejun Heo } 276ce3141a2STejun Heo 27791e914c5SDennis Zhou (Facebook) static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) 278ce3141a2STejun Heo { 27991e914c5SDennis Zhou (Facebook) *rs = find_next_bit(bitmap, end, *rs); 28091e914c5SDennis Zhou (Facebook) *re = find_next_zero_bit(bitmap, end, *rs + 1); 281ce3141a2STejun Heo } 282ce3141a2STejun Heo 283ce3141a2STejun Heo /* 28491e914c5SDennis Zhou (Facebook) * Bitmap region iterators. Iterates over the bitmap between 28591e914c5SDennis Zhou (Facebook) * [@start, @end) in @chunk. @rs and @re should be integer variables 28691e914c5SDennis Zhou (Facebook) * and will be set to start and end index of the current free region. 287ce3141a2STejun Heo */ 28891e914c5SDennis Zhou (Facebook) #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ 28991e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ 290ce3141a2STejun Heo (rs) < (re); \ 29191e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) 292ce3141a2STejun Heo 29391e914c5SDennis Zhou (Facebook) #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ 29491e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ 295ce3141a2STejun Heo (rs) < (re); \ 29691e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) 297ce3141a2STejun Heo 298ca460b3cSDennis Zhou (Facebook) /* 299ca460b3cSDennis Zhou (Facebook) * The following are helper functions to help access bitmaps and convert 300ca460b3cSDennis Zhou (Facebook) * between bitmap offsets to address offsets. 301ca460b3cSDennis Zhou (Facebook) */ 302ca460b3cSDennis Zhou (Facebook) static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 303ca460b3cSDennis Zhou (Facebook) { 304ca460b3cSDennis Zhou (Facebook) return chunk->alloc_map + 305ca460b3cSDennis Zhou (Facebook) (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 306ca460b3cSDennis Zhou (Facebook) } 307ca460b3cSDennis Zhou (Facebook) 308ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_index(int off) 309ca460b3cSDennis Zhou (Facebook) { 310ca460b3cSDennis Zhou (Facebook) return off / PCPU_BITMAP_BLOCK_BITS; 311ca460b3cSDennis Zhou (Facebook) } 312ca460b3cSDennis Zhou (Facebook) 313ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_off(int off) 314ca460b3cSDennis Zhou (Facebook) { 315ca460b3cSDennis Zhou (Facebook) return off & (PCPU_BITMAP_BLOCK_BITS - 1); 316ca460b3cSDennis Zhou (Facebook) } 317ca460b3cSDennis Zhou (Facebook) 318b185cd0dSDennis Zhou (Facebook) static unsigned long pcpu_block_off_to_off(int index, int off) 319b185cd0dSDennis Zhou (Facebook) { 320b185cd0dSDennis Zhou (Facebook) return index * PCPU_BITMAP_BLOCK_BITS + off; 321b185cd0dSDennis Zhou (Facebook) } 322b185cd0dSDennis Zhou (Facebook) 323fbf59bc9STejun Heo /** 324525ca84dSDennis Zhou (Facebook) * pcpu_next_md_free_region - finds the next hint free area 325525ca84dSDennis Zhou (Facebook) * @chunk: chunk of interest 326525ca84dSDennis Zhou (Facebook) * @bit_off: chunk offset 327525ca84dSDennis Zhou (Facebook) * @bits: size of free area 328525ca84dSDennis Zhou (Facebook) * 329525ca84dSDennis Zhou (Facebook) * Helper function for pcpu_for_each_md_free_region. It checks 330525ca84dSDennis Zhou (Facebook) * block->contig_hint and performs aggregation across blocks to find the 331525ca84dSDennis Zhou (Facebook) * next hint. It modifies bit_off and bits in-place to be consumed in the 332525ca84dSDennis Zhou (Facebook) * loop. 333525ca84dSDennis Zhou (Facebook) */ 334525ca84dSDennis Zhou (Facebook) static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, 335525ca84dSDennis Zhou (Facebook) int *bits) 336525ca84dSDennis Zhou (Facebook) { 337525ca84dSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 338525ca84dSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 339525ca84dSDennis Zhou (Facebook) struct pcpu_block_md *block; 340525ca84dSDennis Zhou (Facebook) 341525ca84dSDennis Zhou (Facebook) *bits = 0; 342525ca84dSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 343525ca84dSDennis Zhou (Facebook) block++, i++) { 344525ca84dSDennis Zhou (Facebook) /* handles contig area across blocks */ 345525ca84dSDennis Zhou (Facebook) if (*bits) { 346525ca84dSDennis Zhou (Facebook) *bits += block->left_free; 347525ca84dSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 348525ca84dSDennis Zhou (Facebook) continue; 349525ca84dSDennis Zhou (Facebook) return; 350525ca84dSDennis Zhou (Facebook) } 351525ca84dSDennis Zhou (Facebook) 352525ca84dSDennis Zhou (Facebook) /* 353525ca84dSDennis Zhou (Facebook) * This checks three things. First is there a contig_hint to 354525ca84dSDennis Zhou (Facebook) * check. Second, have we checked this hint before by 355525ca84dSDennis Zhou (Facebook) * comparing the block_off. Third, is this the same as the 356525ca84dSDennis Zhou (Facebook) * right contig hint. In the last case, it spills over into 357525ca84dSDennis Zhou (Facebook) * the next block and should be handled by the contig area 358525ca84dSDennis Zhou (Facebook) * across blocks code. 359525ca84dSDennis Zhou (Facebook) */ 360525ca84dSDennis Zhou (Facebook) *bits = block->contig_hint; 361525ca84dSDennis Zhou (Facebook) if (*bits && block->contig_hint_start >= block_off && 362525ca84dSDennis Zhou (Facebook) *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { 363525ca84dSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, 364525ca84dSDennis Zhou (Facebook) block->contig_hint_start); 365525ca84dSDennis Zhou (Facebook) return; 366525ca84dSDennis Zhou (Facebook) } 3671fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 3681fa4df3eSDennis Zhou block_off = 0; 369525ca84dSDennis Zhou (Facebook) 370525ca84dSDennis Zhou (Facebook) *bits = block->right_free; 371525ca84dSDennis Zhou (Facebook) *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; 372525ca84dSDennis Zhou (Facebook) } 373525ca84dSDennis Zhou (Facebook) } 374525ca84dSDennis Zhou (Facebook) 375b4c2116cSDennis Zhou (Facebook) /** 376b4c2116cSDennis Zhou (Facebook) * pcpu_next_fit_region - finds fit areas for a given allocation request 377b4c2116cSDennis Zhou (Facebook) * @chunk: chunk of interest 378b4c2116cSDennis Zhou (Facebook) * @alloc_bits: size of allocation 379b4c2116cSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 380b4c2116cSDennis Zhou (Facebook) * @bit_off: chunk offset 381b4c2116cSDennis Zhou (Facebook) * @bits: size of free area 382b4c2116cSDennis Zhou (Facebook) * 383b4c2116cSDennis Zhou (Facebook) * Finds the next free region that is viable for use with a given size and 384b4c2116cSDennis Zhou (Facebook) * alignment. This only returns if there is a valid area to be used for this 385b4c2116cSDennis Zhou (Facebook) * allocation. block->first_free is returned if the allocation request fits 386b4c2116cSDennis Zhou (Facebook) * within the block to see if the request can be fulfilled prior to the contig 387b4c2116cSDennis Zhou (Facebook) * hint. 388b4c2116cSDennis Zhou (Facebook) */ 389b4c2116cSDennis Zhou (Facebook) static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, 390b4c2116cSDennis Zhou (Facebook) int align, int *bit_off, int *bits) 391b4c2116cSDennis Zhou (Facebook) { 392b4c2116cSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 393b4c2116cSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 394b4c2116cSDennis Zhou (Facebook) struct pcpu_block_md *block; 395b4c2116cSDennis Zhou (Facebook) 396b4c2116cSDennis Zhou (Facebook) *bits = 0; 397b4c2116cSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 398b4c2116cSDennis Zhou (Facebook) block++, i++) { 399b4c2116cSDennis Zhou (Facebook) /* handles contig area across blocks */ 400b4c2116cSDennis Zhou (Facebook) if (*bits) { 401b4c2116cSDennis Zhou (Facebook) *bits += block->left_free; 402b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 403b4c2116cSDennis Zhou (Facebook) return; 404b4c2116cSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 405b4c2116cSDennis Zhou (Facebook) continue; 406b4c2116cSDennis Zhou (Facebook) } 407b4c2116cSDennis Zhou (Facebook) 408b4c2116cSDennis Zhou (Facebook) /* check block->contig_hint */ 409b4c2116cSDennis Zhou (Facebook) *bits = ALIGN(block->contig_hint_start, align) - 410b4c2116cSDennis Zhou (Facebook) block->contig_hint_start; 411b4c2116cSDennis Zhou (Facebook) /* 412b4c2116cSDennis Zhou (Facebook) * This uses the block offset to determine if this has been 413b4c2116cSDennis Zhou (Facebook) * checked in the prior iteration. 414b4c2116cSDennis Zhou (Facebook) */ 415b4c2116cSDennis Zhou (Facebook) if (block->contig_hint && 416b4c2116cSDennis Zhou (Facebook) block->contig_hint_start >= block_off && 417b4c2116cSDennis Zhou (Facebook) block->contig_hint >= *bits + alloc_bits) { 418b4c2116cSDennis Zhou (Facebook) *bits += alloc_bits + block->contig_hint_start - 419b4c2116cSDennis Zhou (Facebook) block->first_free; 420b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, block->first_free); 421b4c2116cSDennis Zhou (Facebook) return; 422b4c2116cSDennis Zhou (Facebook) } 4231fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 4241fa4df3eSDennis Zhou block_off = 0; 425b4c2116cSDennis Zhou (Facebook) 426b4c2116cSDennis Zhou (Facebook) *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, 427b4c2116cSDennis Zhou (Facebook) align); 428b4c2116cSDennis Zhou (Facebook) *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; 429b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, *bit_off); 430b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 431b4c2116cSDennis Zhou (Facebook) return; 432b4c2116cSDennis Zhou (Facebook) } 433b4c2116cSDennis Zhou (Facebook) 434b4c2116cSDennis Zhou (Facebook) /* no valid offsets were found - fail condition */ 435b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_chunk_map_bits(chunk); 436b4c2116cSDennis Zhou (Facebook) } 437b4c2116cSDennis Zhou (Facebook) 438525ca84dSDennis Zhou (Facebook) /* 439525ca84dSDennis Zhou (Facebook) * Metadata free area iterators. These perform aggregation of free areas 440525ca84dSDennis Zhou (Facebook) * based on the metadata blocks and return the offset @bit_off and size in 441b4c2116cSDennis Zhou (Facebook) * bits of the free area @bits. pcpu_for_each_fit_region only returns when 442b4c2116cSDennis Zhou (Facebook) * a fit is found for the allocation request. 443525ca84dSDennis Zhou (Facebook) */ 444525ca84dSDennis Zhou (Facebook) #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ 445525ca84dSDennis Zhou (Facebook) for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ 446525ca84dSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 447525ca84dSDennis Zhou (Facebook) (bit_off) += (bits) + 1, \ 448525ca84dSDennis Zhou (Facebook) pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) 449525ca84dSDennis Zhou (Facebook) 450b4c2116cSDennis Zhou (Facebook) #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ 451b4c2116cSDennis Zhou (Facebook) for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 452b4c2116cSDennis Zhou (Facebook) &(bits)); \ 453b4c2116cSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 454b4c2116cSDennis Zhou (Facebook) (bit_off) += (bits), \ 455b4c2116cSDennis Zhou (Facebook) pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 456b4c2116cSDennis Zhou (Facebook) &(bits))) 457b4c2116cSDennis Zhou (Facebook) 458525ca84dSDennis Zhou (Facebook) /** 45990459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 4601880d93bSTejun Heo * @size: bytes to allocate 46147504ee0SDennis Zhou * @gfp: allocation flags 462fbf59bc9STejun Heo * 4631880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 46447504ee0SDennis Zhou * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. 46547504ee0SDennis Zhou * This is to facilitate passing through whitelisted flags. The 46647504ee0SDennis Zhou * returned memory is always zeroed. 467fbf59bc9STejun Heo * 468fbf59bc9STejun Heo * RETURNS: 4691880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 470fbf59bc9STejun Heo */ 47147504ee0SDennis Zhou static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) 472fbf59bc9STejun Heo { 473099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 474099a19d9STejun Heo return NULL; 475099a19d9STejun Heo 476fbf59bc9STejun Heo if (size <= PAGE_SIZE) 477554fef1cSDennis Zhou return kzalloc(size, gfp); 4787af4c093SJesper Juhl else 479554fef1cSDennis Zhou return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL); 4801880d93bSTejun Heo } 481fbf59bc9STejun Heo 4821880d93bSTejun Heo /** 4831880d93bSTejun Heo * pcpu_mem_free - free memory 4841880d93bSTejun Heo * @ptr: memory to free 4851880d93bSTejun Heo * 48690459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 4871880d93bSTejun Heo */ 4881d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 4891880d93bSTejun Heo { 4901d5cfdb0STetsuo Handa kvfree(ptr); 491fbf59bc9STejun Heo } 492fbf59bc9STejun Heo 4938744d859SDennis Zhou static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot, 4948744d859SDennis Zhou bool move_front) 4958744d859SDennis Zhou { 4968744d859SDennis Zhou if (chunk != pcpu_reserved_chunk) { 4978744d859SDennis Zhou if (move_front) 4988744d859SDennis Zhou list_move(&chunk->list, &pcpu_slot[slot]); 4998744d859SDennis Zhou else 5008744d859SDennis Zhou list_move_tail(&chunk->list, &pcpu_slot[slot]); 5018744d859SDennis Zhou } 5028744d859SDennis Zhou } 5038744d859SDennis Zhou 5048744d859SDennis Zhou static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot) 5058744d859SDennis Zhou { 5068744d859SDennis Zhou __pcpu_chunk_move(chunk, slot, true); 5078744d859SDennis Zhou } 5088744d859SDennis Zhou 509fbf59bc9STejun Heo /** 510fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 511fbf59bc9STejun Heo * @chunk: chunk of interest 512fbf59bc9STejun Heo * @oslot: the previous slot it was on 513fbf59bc9STejun Heo * 514fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 515fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 516edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 517edcb4639STejun Heo * chunk slots. 518ccea34b5STejun Heo * 519ccea34b5STejun Heo * CONTEXT: 520ccea34b5STejun Heo * pcpu_lock. 521fbf59bc9STejun Heo */ 522fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 523fbf59bc9STejun Heo { 524fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 525fbf59bc9STejun Heo 5268744d859SDennis Zhou if (oslot != nslot) 5278744d859SDennis Zhou __pcpu_chunk_move(chunk, nslot, oslot < nslot); 528fbf59bc9STejun Heo } 529fbf59bc9STejun Heo 53040064aecSDennis Zhou (Facebook) /* 531*b239f7daSDennis Zhou * pcpu_update_empty_pages - update empty page counters 532*b239f7daSDennis Zhou * @chunk: chunk of interest 533*b239f7daSDennis Zhou * @nr: nr of empty pages 534*b239f7daSDennis Zhou * 535*b239f7daSDennis Zhou * This is used to keep track of the empty pages now based on the premise 536*b239f7daSDennis Zhou * a md_block covers a page. The hint update functions recognize if a block 537*b239f7daSDennis Zhou * is made full or broken to calculate deltas for keeping track of free pages. 53840064aecSDennis Zhou (Facebook) */ 539*b239f7daSDennis Zhou static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr) 540*b239f7daSDennis Zhou { 541*b239f7daSDennis Zhou chunk->nr_empty_pop_pages += nr; 542*b239f7daSDennis Zhou if (chunk != pcpu_reserved_chunk) 543*b239f7daSDennis Zhou pcpu_nr_empty_pop_pages += nr; 54440064aecSDennis Zhou (Facebook) } 54540064aecSDennis Zhou (Facebook) 546d9f3a01eSDennis Zhou /* 547d9f3a01eSDennis Zhou * pcpu_region_overlap - determines if two regions overlap 548d9f3a01eSDennis Zhou * @a: start of first region, inclusive 549d9f3a01eSDennis Zhou * @b: end of first region, exclusive 550d9f3a01eSDennis Zhou * @x: start of second region, inclusive 551d9f3a01eSDennis Zhou * @y: end of second region, exclusive 552d9f3a01eSDennis Zhou * 553d9f3a01eSDennis Zhou * This is used to determine if the hint region [a, b) overlaps with the 554d9f3a01eSDennis Zhou * allocated region [x, y). 555d9f3a01eSDennis Zhou */ 556d9f3a01eSDennis Zhou static inline bool pcpu_region_overlap(int a, int b, int x, int y) 557d9f3a01eSDennis Zhou { 558d9f3a01eSDennis Zhou return (a < y) && (x < b); 559d9f3a01eSDennis Zhou } 560d9f3a01eSDennis Zhou 56140064aecSDennis Zhou (Facebook) /** 56240064aecSDennis Zhou (Facebook) * pcpu_chunk_update - updates the chunk metadata given a free area 56340064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 56440064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 56540064aecSDennis Zhou (Facebook) * @bits: size of free area 56640064aecSDennis Zhou (Facebook) * 56713f96637SDennis Zhou (Facebook) * This updates the chunk's contig hint and starting offset given a free area. 568268625a6SDennis Zhou (Facebook) * Choose the best starting offset if the contig hint is equal. 56940064aecSDennis Zhou (Facebook) */ 57040064aecSDennis Zhou (Facebook) static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) 57140064aecSDennis Zhou (Facebook) { 57213f96637SDennis Zhou (Facebook) if (bits > chunk->contig_bits) { 57313f96637SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 57440064aecSDennis Zhou (Facebook) chunk->contig_bits = bits; 575268625a6SDennis Zhou (Facebook) } else if (bits == chunk->contig_bits && chunk->contig_bits_start && 576268625a6SDennis Zhou (Facebook) (!bit_off || 577268625a6SDennis Zhou (Facebook) __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { 578268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 579268625a6SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 58040064aecSDennis Zhou (Facebook) } 58113f96637SDennis Zhou (Facebook) } 58240064aecSDennis Zhou (Facebook) 58340064aecSDennis Zhou (Facebook) /** 58440064aecSDennis Zhou (Facebook) * pcpu_chunk_refresh_hint - updates metadata about a chunk 58540064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 58640064aecSDennis Zhou (Facebook) * 587525ca84dSDennis Zhou (Facebook) * Iterates over the metadata blocks to find the largest contig area. 588525ca84dSDennis Zhou (Facebook) * It also counts the populated pages and uses the delta to update the 589525ca84dSDennis Zhou (Facebook) * global count. 59040064aecSDennis Zhou (Facebook) * 59140064aecSDennis Zhou (Facebook) * Updates: 59240064aecSDennis Zhou (Facebook) * chunk->contig_bits 59313f96637SDennis Zhou (Facebook) * chunk->contig_bits_start 59440064aecSDennis Zhou (Facebook) */ 59540064aecSDennis Zhou (Facebook) static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) 59640064aecSDennis Zhou (Facebook) { 597*b239f7daSDennis Zhou int bit_off, bits; 59840064aecSDennis Zhou (Facebook) 59940064aecSDennis Zhou (Facebook) /* clear metadata */ 60040064aecSDennis Zhou (Facebook) chunk->contig_bits = 0; 60140064aecSDennis Zhou (Facebook) 602525ca84dSDennis Zhou (Facebook) bit_off = chunk->first_bit; 603*b239f7daSDennis Zhou bits = 0; 604525ca84dSDennis Zhou (Facebook) pcpu_for_each_md_free_region(chunk, bit_off, bits) { 605525ca84dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, bit_off, bits); 60640064aecSDennis Zhou (Facebook) } 60740064aecSDennis Zhou (Facebook) } 60840064aecSDennis Zhou (Facebook) 60940064aecSDennis Zhou (Facebook) /** 610ca460b3cSDennis Zhou (Facebook) * pcpu_block_update - updates a block given a free area 611ca460b3cSDennis Zhou (Facebook) * @block: block of interest 612ca460b3cSDennis Zhou (Facebook) * @start: start offset in block 613ca460b3cSDennis Zhou (Facebook) * @end: end offset in block 614ca460b3cSDennis Zhou (Facebook) * 615ca460b3cSDennis Zhou (Facebook) * Updates a block given a known free area. The region [start, end) is 616268625a6SDennis Zhou (Facebook) * expected to be the entirety of the free area within a block. Chooses 617268625a6SDennis Zhou (Facebook) * the best starting offset if the contig hints are equal. 618ca460b3cSDennis Zhou (Facebook) */ 619ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 620ca460b3cSDennis Zhou (Facebook) { 621ca460b3cSDennis Zhou (Facebook) int contig = end - start; 622ca460b3cSDennis Zhou (Facebook) 623ca460b3cSDennis Zhou (Facebook) block->first_free = min(block->first_free, start); 624ca460b3cSDennis Zhou (Facebook) if (start == 0) 625ca460b3cSDennis Zhou (Facebook) block->left_free = contig; 626ca460b3cSDennis Zhou (Facebook) 627ca460b3cSDennis Zhou (Facebook) if (end == PCPU_BITMAP_BLOCK_BITS) 628ca460b3cSDennis Zhou (Facebook) block->right_free = contig; 629ca460b3cSDennis Zhou (Facebook) 630ca460b3cSDennis Zhou (Facebook) if (contig > block->contig_hint) { 631ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = start; 632ca460b3cSDennis Zhou (Facebook) block->contig_hint = contig; 633268625a6SDennis Zhou (Facebook) } else if (block->contig_hint_start && contig == block->contig_hint && 634268625a6SDennis Zhou (Facebook) (!start || __ffs(start) > __ffs(block->contig_hint_start))) { 635268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 636268625a6SDennis Zhou (Facebook) block->contig_hint_start = start; 637ca460b3cSDennis Zhou (Facebook) } 638ca460b3cSDennis Zhou (Facebook) } 639ca460b3cSDennis Zhou (Facebook) 640ca460b3cSDennis Zhou (Facebook) /** 641ca460b3cSDennis Zhou (Facebook) * pcpu_block_refresh_hint 642ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 643ca460b3cSDennis Zhou (Facebook) * @index: index of the metadata block 644ca460b3cSDennis Zhou (Facebook) * 645ca460b3cSDennis Zhou (Facebook) * Scans over the block beginning at first_free and updates the block 646ca460b3cSDennis Zhou (Facebook) * metadata accordingly. 647ca460b3cSDennis Zhou (Facebook) */ 648ca460b3cSDennis Zhou (Facebook) static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 649ca460b3cSDennis Zhou (Facebook) { 650ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *block = chunk->md_blocks + index; 651ca460b3cSDennis Zhou (Facebook) unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 652ca460b3cSDennis Zhou (Facebook) int rs, re; /* region start, region end */ 653ca460b3cSDennis Zhou (Facebook) 654ca460b3cSDennis Zhou (Facebook) /* clear hints */ 655ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 656ca460b3cSDennis Zhou (Facebook) block->left_free = block->right_free = 0; 657ca460b3cSDennis Zhou (Facebook) 658ca460b3cSDennis Zhou (Facebook) /* iterate over free areas and update the contig hints */ 659ca460b3cSDennis Zhou (Facebook) pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, 660ca460b3cSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS) { 661ca460b3cSDennis Zhou (Facebook) pcpu_block_update(block, rs, re); 662ca460b3cSDennis Zhou (Facebook) } 663ca460b3cSDennis Zhou (Facebook) } 664ca460b3cSDennis Zhou (Facebook) 665ca460b3cSDennis Zhou (Facebook) /** 666ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_alloc - update hint on allocation path 667ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 668ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 669ca460b3cSDennis Zhou (Facebook) * @bits: size of request 670fc304334SDennis Zhou (Facebook) * 671fc304334SDennis Zhou (Facebook) * Updates metadata for the allocation path. The metadata only has to be 672fc304334SDennis Zhou (Facebook) * refreshed by a full scan iff the chunk's contig hint is broken. Block level 673fc304334SDennis Zhou (Facebook) * scans are required if the block's contig hint is broken. 674ca460b3cSDennis Zhou (Facebook) */ 675ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 676ca460b3cSDennis Zhou (Facebook) int bits) 677ca460b3cSDennis Zhou (Facebook) { 678*b239f7daSDennis Zhou int nr_empty_pages = 0; 679ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 680ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 681ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 682ca460b3cSDennis Zhou (Facebook) 683ca460b3cSDennis Zhou (Facebook) /* 684ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 685ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 686ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 687ca460b3cSDennis Zhou (Facebook) * range. 688ca460b3cSDennis Zhou (Facebook) */ 689ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 690ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 691ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 692ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 693ca460b3cSDennis Zhou (Facebook) 694ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 695ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 696ca460b3cSDennis Zhou (Facebook) 697ca460b3cSDennis Zhou (Facebook) /* 698ca460b3cSDennis Zhou (Facebook) * Update s_block. 699fc304334SDennis Zhou (Facebook) * block->first_free must be updated if the allocation takes its place. 700fc304334SDennis Zhou (Facebook) * If the allocation breaks the contig_hint, a scan is required to 701fc304334SDennis Zhou (Facebook) * restore this hint. 702ca460b3cSDennis Zhou (Facebook) */ 703*b239f7daSDennis Zhou if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 704*b239f7daSDennis Zhou nr_empty_pages++; 705*b239f7daSDennis Zhou 706fc304334SDennis Zhou (Facebook) if (s_off == s_block->first_free) 707fc304334SDennis Zhou (Facebook) s_block->first_free = find_next_zero_bit( 708fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, s_index), 709fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, 710fc304334SDennis Zhou (Facebook) s_off + bits); 711fc304334SDennis Zhou (Facebook) 712d9f3a01eSDennis Zhou if (pcpu_region_overlap(s_block->contig_hint_start, 713d9f3a01eSDennis Zhou s_block->contig_hint_start + 714d9f3a01eSDennis Zhou s_block->contig_hint, 715d9f3a01eSDennis Zhou s_off, 716d9f3a01eSDennis Zhou s_off + bits)) { 717fc304334SDennis Zhou (Facebook) /* block contig hint is broken - scan to fix it */ 718ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 719fc304334SDennis Zhou (Facebook) } else { 720fc304334SDennis Zhou (Facebook) /* update left and right contig manually */ 721fc304334SDennis Zhou (Facebook) s_block->left_free = min(s_block->left_free, s_off); 722fc304334SDennis Zhou (Facebook) if (s_index == e_index) 723fc304334SDennis Zhou (Facebook) s_block->right_free = min_t(int, s_block->right_free, 724fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 725fc304334SDennis Zhou (Facebook) else 726fc304334SDennis Zhou (Facebook) s_block->right_free = 0; 727fc304334SDennis Zhou (Facebook) } 728ca460b3cSDennis Zhou (Facebook) 729ca460b3cSDennis Zhou (Facebook) /* 730ca460b3cSDennis Zhou (Facebook) * Update e_block. 731ca460b3cSDennis Zhou (Facebook) */ 732ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 733*b239f7daSDennis Zhou if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 734*b239f7daSDennis Zhou nr_empty_pages++; 735*b239f7daSDennis Zhou 736fc304334SDennis Zhou (Facebook) /* 737fc304334SDennis Zhou (Facebook) * When the allocation is across blocks, the end is along 738fc304334SDennis Zhou (Facebook) * the left part of the e_block. 739fc304334SDennis Zhou (Facebook) */ 740fc304334SDennis Zhou (Facebook) e_block->first_free = find_next_zero_bit( 741fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, e_index), 742fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, e_off); 743fc304334SDennis Zhou (Facebook) 744fc304334SDennis Zhou (Facebook) if (e_off == PCPU_BITMAP_BLOCK_BITS) { 745fc304334SDennis Zhou (Facebook) /* reset the block */ 746fc304334SDennis Zhou (Facebook) e_block++; 747fc304334SDennis Zhou (Facebook) } else { 748fc304334SDennis Zhou (Facebook) if (e_off > e_block->contig_hint_start) { 749fc304334SDennis Zhou (Facebook) /* contig hint is broken - scan to fix it */ 750ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 751fc304334SDennis Zhou (Facebook) } else { 752fc304334SDennis Zhou (Facebook) e_block->left_free = 0; 753fc304334SDennis Zhou (Facebook) e_block->right_free = 754fc304334SDennis Zhou (Facebook) min_t(int, e_block->right_free, 755fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 756fc304334SDennis Zhou (Facebook) } 757fc304334SDennis Zhou (Facebook) } 758ca460b3cSDennis Zhou (Facebook) 759ca460b3cSDennis Zhou (Facebook) /* update in-between md_blocks */ 760*b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 761ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 762ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 763ca460b3cSDennis Zhou (Facebook) block->left_free = 0; 764ca460b3cSDennis Zhou (Facebook) block->right_free = 0; 765ca460b3cSDennis Zhou (Facebook) } 766ca460b3cSDennis Zhou (Facebook) } 767ca460b3cSDennis Zhou (Facebook) 768*b239f7daSDennis Zhou if (nr_empty_pages) 769*b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr_empty_pages); 770*b239f7daSDennis Zhou 771fc304334SDennis Zhou (Facebook) /* 772fc304334SDennis Zhou (Facebook) * The only time a full chunk scan is required is if the chunk 773fc304334SDennis Zhou (Facebook) * contig hint is broken. Otherwise, it means a smaller space 774fc304334SDennis Zhou (Facebook) * was used and therefore the chunk contig hint is still correct. 775fc304334SDennis Zhou (Facebook) */ 776d9f3a01eSDennis Zhou if (pcpu_region_overlap(chunk->contig_bits_start, 777d9f3a01eSDennis Zhou chunk->contig_bits_start + chunk->contig_bits, 778d9f3a01eSDennis Zhou bit_off, 779d9f3a01eSDennis Zhou bit_off + bits)) 780ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 781ca460b3cSDennis Zhou (Facebook) } 782ca460b3cSDennis Zhou (Facebook) 783ca460b3cSDennis Zhou (Facebook) /** 784ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_free - updates the block hints on the free path 785ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 786ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 787ca460b3cSDennis Zhou (Facebook) * @bits: size of request 788b185cd0dSDennis Zhou (Facebook) * 789b185cd0dSDennis Zhou (Facebook) * Updates metadata for the allocation path. This avoids a blind block 790b185cd0dSDennis Zhou (Facebook) * refresh by making use of the block contig hints. If this fails, it scans 791b185cd0dSDennis Zhou (Facebook) * forward and backward to determine the extent of the free area. This is 792b185cd0dSDennis Zhou (Facebook) * capped at the boundary of blocks. 793b185cd0dSDennis Zhou (Facebook) * 794b185cd0dSDennis Zhou (Facebook) * A chunk update is triggered if a page becomes free, a block becomes free, 795b185cd0dSDennis Zhou (Facebook) * or the free spans across blocks. This tradeoff is to minimize iterating 796b185cd0dSDennis Zhou (Facebook) * over the block metadata to update chunk->contig_bits. chunk->contig_bits 797b185cd0dSDennis Zhou (Facebook) * may be off by up to a page, but it will never be more than the available 798b185cd0dSDennis Zhou (Facebook) * space. If the contig hint is contained in one block, it will be accurate. 799ca460b3cSDennis Zhou (Facebook) */ 800ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 801ca460b3cSDennis Zhou (Facebook) int bits) 802ca460b3cSDennis Zhou (Facebook) { 803*b239f7daSDennis Zhou int nr_empty_pages = 0; 804ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 805ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 806ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 807b185cd0dSDennis Zhou (Facebook) int start, end; /* start and end of the whole free area */ 808ca460b3cSDennis Zhou (Facebook) 809ca460b3cSDennis Zhou (Facebook) /* 810ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 811ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 812ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 813ca460b3cSDennis Zhou (Facebook) * range. 814ca460b3cSDennis Zhou (Facebook) */ 815ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 816ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 817ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 818ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 819ca460b3cSDennis Zhou (Facebook) 820ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 821ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 822ca460b3cSDennis Zhou (Facebook) 823b185cd0dSDennis Zhou (Facebook) /* 824b185cd0dSDennis Zhou (Facebook) * Check if the freed area aligns with the block->contig_hint. 825b185cd0dSDennis Zhou (Facebook) * If it does, then the scan to find the beginning/end of the 826b185cd0dSDennis Zhou (Facebook) * larger free area can be avoided. 827b185cd0dSDennis Zhou (Facebook) * 828b185cd0dSDennis Zhou (Facebook) * start and end refer to beginning and end of the free area 829b185cd0dSDennis Zhou (Facebook) * within each their respective blocks. This is not necessarily 830b185cd0dSDennis Zhou (Facebook) * the entire free area as it may span blocks past the beginning 831b185cd0dSDennis Zhou (Facebook) * or end of the block. 832b185cd0dSDennis Zhou (Facebook) */ 833b185cd0dSDennis Zhou (Facebook) start = s_off; 834b185cd0dSDennis Zhou (Facebook) if (s_off == s_block->contig_hint + s_block->contig_hint_start) { 835b185cd0dSDennis Zhou (Facebook) start = s_block->contig_hint_start; 836b185cd0dSDennis Zhou (Facebook) } else { 837b185cd0dSDennis Zhou (Facebook) /* 838b185cd0dSDennis Zhou (Facebook) * Scan backwards to find the extent of the free area. 839b185cd0dSDennis Zhou (Facebook) * find_last_bit returns the starting bit, so if the start bit 840b185cd0dSDennis Zhou (Facebook) * is returned, that means there was no last bit and the 841b185cd0dSDennis Zhou (Facebook) * remainder of the chunk is free. 842b185cd0dSDennis Zhou (Facebook) */ 843b185cd0dSDennis Zhou (Facebook) int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), 844b185cd0dSDennis Zhou (Facebook) start); 845b185cd0dSDennis Zhou (Facebook) start = (start == l_bit) ? 0 : l_bit + 1; 846b185cd0dSDennis Zhou (Facebook) } 847b185cd0dSDennis Zhou (Facebook) 848b185cd0dSDennis Zhou (Facebook) end = e_off; 849b185cd0dSDennis Zhou (Facebook) if (e_off == e_block->contig_hint_start) 850b185cd0dSDennis Zhou (Facebook) end = e_block->contig_hint_start + e_block->contig_hint; 851b185cd0dSDennis Zhou (Facebook) else 852b185cd0dSDennis Zhou (Facebook) end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), 853b185cd0dSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, end); 854b185cd0dSDennis Zhou (Facebook) 855ca460b3cSDennis Zhou (Facebook) /* update s_block */ 856b185cd0dSDennis Zhou (Facebook) e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; 857*b239f7daSDennis Zhou if (!start && e_off == PCPU_BITMAP_BLOCK_BITS) 858*b239f7daSDennis Zhou nr_empty_pages++; 859b185cd0dSDennis Zhou (Facebook) pcpu_block_update(s_block, start, e_off); 860ca460b3cSDennis Zhou (Facebook) 861ca460b3cSDennis Zhou (Facebook) /* freeing in the same block */ 862ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 863ca460b3cSDennis Zhou (Facebook) /* update e_block */ 864*b239f7daSDennis Zhou if (end == PCPU_BITMAP_BLOCK_BITS) 865*b239f7daSDennis Zhou nr_empty_pages++; 866b185cd0dSDennis Zhou (Facebook) pcpu_block_update(e_block, 0, end); 867ca460b3cSDennis Zhou (Facebook) 868ca460b3cSDennis Zhou (Facebook) /* reset md_blocks in the middle */ 869*b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 870ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 871ca460b3cSDennis Zhou (Facebook) block->first_free = 0; 872ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = 0; 873ca460b3cSDennis Zhou (Facebook) block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 874ca460b3cSDennis Zhou (Facebook) block->left_free = PCPU_BITMAP_BLOCK_BITS; 875ca460b3cSDennis Zhou (Facebook) block->right_free = PCPU_BITMAP_BLOCK_BITS; 876ca460b3cSDennis Zhou (Facebook) } 877ca460b3cSDennis Zhou (Facebook) } 878ca460b3cSDennis Zhou (Facebook) 879*b239f7daSDennis Zhou if (nr_empty_pages) 880*b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr_empty_pages); 881*b239f7daSDennis Zhou 882b185cd0dSDennis Zhou (Facebook) /* 883*b239f7daSDennis Zhou * Refresh chunk metadata when the free makes a block free or spans 884*b239f7daSDennis Zhou * across blocks. The contig_hint may be off by up to a page, but if 885*b239f7daSDennis Zhou * the contig_hint is contained in a block, it will be accurate with 886*b239f7daSDennis Zhou * the else condition below. 887b185cd0dSDennis Zhou (Facebook) */ 888*b239f7daSDennis Zhou if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index) 889ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 890b185cd0dSDennis Zhou (Facebook) else 891b185cd0dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start), 8928e5a2b98SDennis Zhou end - start); 893ca460b3cSDennis Zhou (Facebook) } 894ca460b3cSDennis Zhou (Facebook) 895ca460b3cSDennis Zhou (Facebook) /** 89640064aecSDennis Zhou (Facebook) * pcpu_is_populated - determines if the region is populated 89740064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 89840064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 89940064aecSDennis Zhou (Facebook) * @bits: size of area 90040064aecSDennis Zhou (Facebook) * @next_off: return value for the next offset to start searching 90140064aecSDennis Zhou (Facebook) * 90240064aecSDennis Zhou (Facebook) * For atomic allocations, check if the backing pages are populated. 90340064aecSDennis Zhou (Facebook) * 90440064aecSDennis Zhou (Facebook) * RETURNS: 90540064aecSDennis Zhou (Facebook) * Bool if the backing pages are populated. 90640064aecSDennis Zhou (Facebook) * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 90740064aecSDennis Zhou (Facebook) */ 90840064aecSDennis Zhou (Facebook) static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 90940064aecSDennis Zhou (Facebook) int *next_off) 91040064aecSDennis Zhou (Facebook) { 91140064aecSDennis Zhou (Facebook) int page_start, page_end, rs, re; 91240064aecSDennis Zhou (Facebook) 91340064aecSDennis Zhou (Facebook) page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 91440064aecSDennis Zhou (Facebook) page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 91540064aecSDennis Zhou (Facebook) 91640064aecSDennis Zhou (Facebook) rs = page_start; 91740064aecSDennis Zhou (Facebook) pcpu_next_unpop(chunk->populated, &rs, &re, page_end); 91840064aecSDennis Zhou (Facebook) if (rs >= page_end) 91940064aecSDennis Zhou (Facebook) return true; 92040064aecSDennis Zhou (Facebook) 92140064aecSDennis Zhou (Facebook) *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 92240064aecSDennis Zhou (Facebook) return false; 92340064aecSDennis Zhou (Facebook) } 92440064aecSDennis Zhou (Facebook) 92540064aecSDennis Zhou (Facebook) /** 92640064aecSDennis Zhou (Facebook) * pcpu_find_block_fit - finds the block index to start searching 92740064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 92840064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 92940064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE bytes) 93040064aecSDennis Zhou (Facebook) * @pop_only: use populated regions only 93140064aecSDennis Zhou (Facebook) * 932b4c2116cSDennis Zhou (Facebook) * Given a chunk and an allocation spec, find the offset to begin searching 933b4c2116cSDennis Zhou (Facebook) * for a free region. This iterates over the bitmap metadata blocks to 934b4c2116cSDennis Zhou (Facebook) * find an offset that will be guaranteed to fit the requirements. It is 935b4c2116cSDennis Zhou (Facebook) * not quite first fit as if the allocation does not fit in the contig hint 936b4c2116cSDennis Zhou (Facebook) * of a block or chunk, it is skipped. This errs on the side of caution 937b4c2116cSDennis Zhou (Facebook) * to prevent excess iteration. Poor alignment can cause the allocator to 938b4c2116cSDennis Zhou (Facebook) * skip over blocks and chunks that have valid free areas. 939b4c2116cSDennis Zhou (Facebook) * 94040064aecSDennis Zhou (Facebook) * RETURNS: 94140064aecSDennis Zhou (Facebook) * The offset in the bitmap to begin searching. 94240064aecSDennis Zhou (Facebook) * -1 if no offset is found. 94340064aecSDennis Zhou (Facebook) */ 94440064aecSDennis Zhou (Facebook) static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 94540064aecSDennis Zhou (Facebook) size_t align, bool pop_only) 94640064aecSDennis Zhou (Facebook) { 947b4c2116cSDennis Zhou (Facebook) int bit_off, bits, next_off; 94840064aecSDennis Zhou (Facebook) 94913f96637SDennis Zhou (Facebook) /* 95013f96637SDennis Zhou (Facebook) * Check to see if the allocation can fit in the chunk's contig hint. 95113f96637SDennis Zhou (Facebook) * This is an optimization to prevent scanning by assuming if it 95213f96637SDennis Zhou (Facebook) * cannot fit in the global hint, there is memory pressure and creating 95313f96637SDennis Zhou (Facebook) * a new chunk would happen soon. 95413f96637SDennis Zhou (Facebook) */ 95513f96637SDennis Zhou (Facebook) bit_off = ALIGN(chunk->contig_bits_start, align) - 95613f96637SDennis Zhou (Facebook) chunk->contig_bits_start; 95713f96637SDennis Zhou (Facebook) if (bit_off + alloc_bits > chunk->contig_bits) 95813f96637SDennis Zhou (Facebook) return -1; 95913f96637SDennis Zhou (Facebook) 960b4c2116cSDennis Zhou (Facebook) bit_off = chunk->first_bit; 961b4c2116cSDennis Zhou (Facebook) bits = 0; 962b4c2116cSDennis Zhou (Facebook) pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { 96340064aecSDennis Zhou (Facebook) if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 964b4c2116cSDennis Zhou (Facebook) &next_off)) 96540064aecSDennis Zhou (Facebook) break; 96640064aecSDennis Zhou (Facebook) 967b4c2116cSDennis Zhou (Facebook) bit_off = next_off; 96840064aecSDennis Zhou (Facebook) bits = 0; 96940064aecSDennis Zhou (Facebook) } 97040064aecSDennis Zhou (Facebook) 97140064aecSDennis Zhou (Facebook) if (bit_off == pcpu_chunk_map_bits(chunk)) 97240064aecSDennis Zhou (Facebook) return -1; 97340064aecSDennis Zhou (Facebook) 97440064aecSDennis Zhou (Facebook) return bit_off; 97540064aecSDennis Zhou (Facebook) } 97640064aecSDennis Zhou (Facebook) 97740064aecSDennis Zhou (Facebook) /** 97840064aecSDennis Zhou (Facebook) * pcpu_alloc_area - allocates an area from a pcpu_chunk 97940064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 98040064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 98140064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 98240064aecSDennis Zhou (Facebook) * @start: bit_off to start searching 98340064aecSDennis Zhou (Facebook) * 98440064aecSDennis Zhou (Facebook) * This function takes in a @start offset to begin searching to fit an 985b4c2116cSDennis Zhou (Facebook) * allocation of @alloc_bits with alignment @align. It needs to scan 986b4c2116cSDennis Zhou (Facebook) * the allocation map because if it fits within the block's contig hint, 987b4c2116cSDennis Zhou (Facebook) * @start will be block->first_free. This is an attempt to fill the 988b4c2116cSDennis Zhou (Facebook) * allocation prior to breaking the contig hint. The allocation and 989b4c2116cSDennis Zhou (Facebook) * boundary maps are updated accordingly if it confirms a valid 990b4c2116cSDennis Zhou (Facebook) * free area. 99140064aecSDennis Zhou (Facebook) * 99240064aecSDennis Zhou (Facebook) * RETURNS: 99340064aecSDennis Zhou (Facebook) * Allocated addr offset in @chunk on success. 99440064aecSDennis Zhou (Facebook) * -1 if no matching area is found. 99540064aecSDennis Zhou (Facebook) */ 99640064aecSDennis Zhou (Facebook) static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 99740064aecSDennis Zhou (Facebook) size_t align, int start) 99840064aecSDennis Zhou (Facebook) { 99940064aecSDennis Zhou (Facebook) size_t align_mask = (align) ? (align - 1) : 0; 100040064aecSDennis Zhou (Facebook) int bit_off, end, oslot; 10019f7dcf22STejun Heo 10024f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 10034f996e23STejun Heo 100440064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1005833af842STejun Heo 1006833af842STejun Heo /* 100740064aecSDennis Zhou (Facebook) * Search to find a fit. 1008833af842STejun Heo */ 10098c43004aSDennis Zhou end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS, 10108c43004aSDennis Zhou pcpu_chunk_map_bits(chunk)); 101140064aecSDennis Zhou (Facebook) bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, 101240064aecSDennis Zhou (Facebook) alloc_bits, align_mask); 101340064aecSDennis Zhou (Facebook) if (bit_off >= end) 1014a16037c8STejun Heo return -1; 1015a16037c8STejun Heo 101640064aecSDennis Zhou (Facebook) /* update alloc map */ 101740064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 1018a16037c8STejun Heo 101940064aecSDennis Zhou (Facebook) /* update boundary map */ 102040064aecSDennis Zhou (Facebook) set_bit(bit_off, chunk->bound_map); 102140064aecSDennis Zhou (Facebook) bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 102240064aecSDennis Zhou (Facebook) set_bit(bit_off + alloc_bits, chunk->bound_map); 1023a16037c8STejun Heo 102440064aecSDennis Zhou (Facebook) chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 102540064aecSDennis Zhou (Facebook) 102686b442fbSDennis Zhou (Facebook) /* update first free bit */ 102786b442fbSDennis Zhou (Facebook) if (bit_off == chunk->first_bit) 102886b442fbSDennis Zhou (Facebook) chunk->first_bit = find_next_zero_bit( 102986b442fbSDennis Zhou (Facebook) chunk->alloc_map, 103086b442fbSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk), 103186b442fbSDennis Zhou (Facebook) bit_off + alloc_bits); 103286b442fbSDennis Zhou (Facebook) 1033ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 103440064aecSDennis Zhou (Facebook) 103540064aecSDennis Zhou (Facebook) pcpu_chunk_relocate(chunk, oslot); 103640064aecSDennis Zhou (Facebook) 103740064aecSDennis Zhou (Facebook) return bit_off * PCPU_MIN_ALLOC_SIZE; 1038a16037c8STejun Heo } 1039a16037c8STejun Heo 1040a16037c8STejun Heo /** 104140064aecSDennis Zhou (Facebook) * pcpu_free_area - frees the corresponding offset 1042fbf59bc9STejun Heo * @chunk: chunk of interest 104340064aecSDennis Zhou (Facebook) * @off: addr offset into chunk 1044fbf59bc9STejun Heo * 104540064aecSDennis Zhou (Facebook) * This function determines the size of an allocation to free using 104640064aecSDennis Zhou (Facebook) * the boundary bitmap and clears the allocation map. 1047fbf59bc9STejun Heo */ 104840064aecSDennis Zhou (Facebook) static void pcpu_free_area(struct pcpu_chunk *chunk, int off) 1049fbf59bc9STejun Heo { 105040064aecSDennis Zhou (Facebook) int bit_off, bits, end, oslot; 1051fbf59bc9STejun Heo 10525ccd30e4SDennis Zhou lockdep_assert_held(&pcpu_lock); 105330a5b536SDennis Zhou pcpu_stats_area_dealloc(chunk); 10545ccd30e4SDennis Zhou 105540064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1056723ad1d9SAl Viro 105740064aecSDennis Zhou (Facebook) bit_off = off / PCPU_MIN_ALLOC_SIZE; 1058fbf59bc9STejun Heo 105940064aecSDennis Zhou (Facebook) /* find end index */ 106040064aecSDennis Zhou (Facebook) end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 106140064aecSDennis Zhou (Facebook) bit_off + 1); 106240064aecSDennis Zhou (Facebook) bits = end - bit_off; 106340064aecSDennis Zhou (Facebook) bitmap_clear(chunk->alloc_map, bit_off, bits); 10643d331ad7SAl Viro 106540064aecSDennis Zhou (Facebook) /* update metadata */ 106640064aecSDennis Zhou (Facebook) chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; 1067fbf59bc9STejun Heo 106886b442fbSDennis Zhou (Facebook) /* update first free bit */ 106986b442fbSDennis Zhou (Facebook) chunk->first_bit = min(chunk->first_bit, bit_off); 107086b442fbSDennis Zhou (Facebook) 1071ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_free(chunk, bit_off, bits); 1072b539b87fSTejun Heo 1073fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 1074fbf59bc9STejun Heo } 1075fbf59bc9STejun Heo 1076ca460b3cSDennis Zhou (Facebook) static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 1077ca460b3cSDennis Zhou (Facebook) { 1078ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *md_block; 1079ca460b3cSDennis Zhou (Facebook) 1080ca460b3cSDennis Zhou (Facebook) for (md_block = chunk->md_blocks; 1081ca460b3cSDennis Zhou (Facebook) md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 1082ca460b3cSDennis Zhou (Facebook) md_block++) { 1083ca460b3cSDennis Zhou (Facebook) md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 1084ca460b3cSDennis Zhou (Facebook) md_block->left_free = PCPU_BITMAP_BLOCK_BITS; 1085ca460b3cSDennis Zhou (Facebook) md_block->right_free = PCPU_BITMAP_BLOCK_BITS; 1086ca460b3cSDennis Zhou (Facebook) } 1087ca460b3cSDennis Zhou (Facebook) } 1088ca460b3cSDennis Zhou (Facebook) 108940064aecSDennis Zhou (Facebook) /** 109040064aecSDennis Zhou (Facebook) * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 109140064aecSDennis Zhou (Facebook) * @tmp_addr: the start of the region served 109240064aecSDennis Zhou (Facebook) * @map_size: size of the region served 109340064aecSDennis Zhou (Facebook) * 109440064aecSDennis Zhou (Facebook) * This is responsible for creating the chunks that serve the first chunk. The 109540064aecSDennis Zhou (Facebook) * base_addr is page aligned down of @tmp_addr while the region end is page 109640064aecSDennis Zhou (Facebook) * aligned up. Offsets are kept track of to determine the region served. All 109740064aecSDennis Zhou (Facebook) * this is done to appease the bitmap allocator in avoiding partial blocks. 109840064aecSDennis Zhou (Facebook) * 109940064aecSDennis Zhou (Facebook) * RETURNS: 110040064aecSDennis Zhou (Facebook) * Chunk serving the region at @tmp_addr of @map_size. 110140064aecSDennis Zhou (Facebook) */ 1102c0ebfdc3SDennis Zhou (Facebook) static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 110340064aecSDennis Zhou (Facebook) int map_size) 110410edf5b0SDennis Zhou (Facebook) { 110510edf5b0SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 1106ca460b3cSDennis Zhou (Facebook) unsigned long aligned_addr, lcm_align; 110740064aecSDennis Zhou (Facebook) int start_offset, offset_bits, region_size, region_bits; 1108f655f405SMike Rapoport size_t alloc_size; 1109c0ebfdc3SDennis Zhou (Facebook) 1110c0ebfdc3SDennis Zhou (Facebook) /* region calculations */ 1111c0ebfdc3SDennis Zhou (Facebook) aligned_addr = tmp_addr & PAGE_MASK; 1112c0ebfdc3SDennis Zhou (Facebook) 1113c0ebfdc3SDennis Zhou (Facebook) start_offset = tmp_addr - aligned_addr; 11146b9d7c8eSDennis Zhou (Facebook) 1115ca460b3cSDennis Zhou (Facebook) /* 1116ca460b3cSDennis Zhou (Facebook) * Align the end of the region with the LCM of PAGE_SIZE and 1117ca460b3cSDennis Zhou (Facebook) * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of 1118ca460b3cSDennis Zhou (Facebook) * the other. 1119ca460b3cSDennis Zhou (Facebook) */ 1120ca460b3cSDennis Zhou (Facebook) lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); 1121ca460b3cSDennis Zhou (Facebook) region_size = ALIGN(start_offset + map_size, lcm_align); 112210edf5b0SDennis Zhou (Facebook) 1123c0ebfdc3SDennis Zhou (Facebook) /* allocate chunk */ 1124f655f405SMike Rapoport alloc_size = sizeof(struct pcpu_chunk) + 1125f655f405SMike Rapoport BITS_TO_LONGS(region_size >> PAGE_SHIFT); 1126f655f405SMike Rapoport chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1127f655f405SMike Rapoport if (!chunk) 1128f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1129f655f405SMike Rapoport alloc_size); 1130c0ebfdc3SDennis Zhou (Facebook) 113110edf5b0SDennis Zhou (Facebook) INIT_LIST_HEAD(&chunk->list); 1132c0ebfdc3SDennis Zhou (Facebook) 1133c0ebfdc3SDennis Zhou (Facebook) chunk->base_addr = (void *)aligned_addr; 113410edf5b0SDennis Zhou (Facebook) chunk->start_offset = start_offset; 11356b9d7c8eSDennis Zhou (Facebook) chunk->end_offset = region_size - chunk->start_offset - map_size; 1136c0ebfdc3SDennis Zhou (Facebook) 11378ab16c43SDennis Zhou (Facebook) chunk->nr_pages = region_size >> PAGE_SHIFT; 113840064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 1139c0ebfdc3SDennis Zhou (Facebook) 1140f655f405SMike Rapoport alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]); 1141f655f405SMike Rapoport chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1142f655f405SMike Rapoport if (!chunk->alloc_map) 1143f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1144f655f405SMike Rapoport alloc_size); 1145f655f405SMike Rapoport 1146f655f405SMike Rapoport alloc_size = 1147f655f405SMike Rapoport BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]); 1148f655f405SMike Rapoport chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1149f655f405SMike Rapoport if (!chunk->bound_map) 1150f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1151f655f405SMike Rapoport alloc_size); 1152f655f405SMike Rapoport 1153f655f405SMike Rapoport alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]); 1154f655f405SMike Rapoport chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1155f655f405SMike Rapoport if (!chunk->md_blocks) 1156f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1157f655f405SMike Rapoport alloc_size); 1158f655f405SMike Rapoport 1159ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 116010edf5b0SDennis Zhou (Facebook) 116110edf5b0SDennis Zhou (Facebook) /* manage populated page bitmap */ 116210edf5b0SDennis Zhou (Facebook) chunk->immutable = true; 11638ab16c43SDennis Zhou (Facebook) bitmap_fill(chunk->populated, chunk->nr_pages); 11648ab16c43SDennis Zhou (Facebook) chunk->nr_populated = chunk->nr_pages; 1165*b239f7daSDennis Zhou chunk->nr_empty_pop_pages = chunk->nr_pages; 116610edf5b0SDennis Zhou (Facebook) 116740064aecSDennis Zhou (Facebook) chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; 116840064aecSDennis Zhou (Facebook) chunk->free_bytes = map_size; 1169c0ebfdc3SDennis Zhou (Facebook) 1170c0ebfdc3SDennis Zhou (Facebook) if (chunk->start_offset) { 1171c0ebfdc3SDennis Zhou (Facebook) /* hide the beginning of the bitmap */ 117240064aecSDennis Zhou (Facebook) offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 117340064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 0, offset_bits); 117440064aecSDennis Zhou (Facebook) set_bit(0, chunk->bound_map); 117540064aecSDennis Zhou (Facebook) set_bit(offset_bits, chunk->bound_map); 1176ca460b3cSDennis Zhou (Facebook) 117786b442fbSDennis Zhou (Facebook) chunk->first_bit = offset_bits; 117886b442fbSDennis Zhou (Facebook) 1179ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 1180c0ebfdc3SDennis Zhou (Facebook) } 1181c0ebfdc3SDennis Zhou (Facebook) 11826b9d7c8eSDennis Zhou (Facebook) if (chunk->end_offset) { 11836b9d7c8eSDennis Zhou (Facebook) /* hide the end of the bitmap */ 118440064aecSDennis Zhou (Facebook) offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 118540064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 118640064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk) - offset_bits, 118740064aecSDennis Zhou (Facebook) offset_bits); 118840064aecSDennis Zhou (Facebook) set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 118940064aecSDennis Zhou (Facebook) chunk->bound_map); 119040064aecSDennis Zhou (Facebook) set_bit(region_bits, chunk->bound_map); 11916b9d7c8eSDennis Zhou (Facebook) 1192ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 1193ca460b3cSDennis Zhou (Facebook) - offset_bits, offset_bits); 1194ca460b3cSDennis Zhou (Facebook) } 119540064aecSDennis Zhou (Facebook) 119610edf5b0SDennis Zhou (Facebook) return chunk; 119710edf5b0SDennis Zhou (Facebook) } 119810edf5b0SDennis Zhou (Facebook) 119947504ee0SDennis Zhou static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) 12006081089fSTejun Heo { 12016081089fSTejun Heo struct pcpu_chunk *chunk; 120240064aecSDennis Zhou (Facebook) int region_bits; 12036081089fSTejun Heo 120447504ee0SDennis Zhou chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); 12056081089fSTejun Heo if (!chunk) 12066081089fSTejun Heo return NULL; 12076081089fSTejun Heo 12086081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 1209c0ebfdc3SDennis Zhou (Facebook) chunk->nr_pages = pcpu_unit_pages; 121040064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 121140064aecSDennis Zhou (Facebook) 121240064aecSDennis Zhou (Facebook) chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 121347504ee0SDennis Zhou sizeof(chunk->alloc_map[0]), gfp); 121440064aecSDennis Zhou (Facebook) if (!chunk->alloc_map) 121540064aecSDennis Zhou (Facebook) goto alloc_map_fail; 121640064aecSDennis Zhou (Facebook) 121740064aecSDennis Zhou (Facebook) chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 121847504ee0SDennis Zhou sizeof(chunk->bound_map[0]), gfp); 121940064aecSDennis Zhou (Facebook) if (!chunk->bound_map) 122040064aecSDennis Zhou (Facebook) goto bound_map_fail; 122140064aecSDennis Zhou (Facebook) 1222ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 122347504ee0SDennis Zhou sizeof(chunk->md_blocks[0]), gfp); 1224ca460b3cSDennis Zhou (Facebook) if (!chunk->md_blocks) 1225ca460b3cSDennis Zhou (Facebook) goto md_blocks_fail; 1226ca460b3cSDennis Zhou (Facebook) 1227ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 1228ca460b3cSDennis Zhou (Facebook) 122940064aecSDennis Zhou (Facebook) /* init metadata */ 123040064aecSDennis Zhou (Facebook) chunk->contig_bits = region_bits; 123140064aecSDennis Zhou (Facebook) chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 1232c0ebfdc3SDennis Zhou (Facebook) 12336081089fSTejun Heo return chunk; 123440064aecSDennis Zhou (Facebook) 1235ca460b3cSDennis Zhou (Facebook) md_blocks_fail: 1236ca460b3cSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 123740064aecSDennis Zhou (Facebook) bound_map_fail: 123840064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 123940064aecSDennis Zhou (Facebook) alloc_map_fail: 124040064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk); 124140064aecSDennis Zhou (Facebook) 124240064aecSDennis Zhou (Facebook) return NULL; 12436081089fSTejun Heo } 12446081089fSTejun Heo 12456081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 12466081089fSTejun Heo { 12476081089fSTejun Heo if (!chunk) 12486081089fSTejun Heo return; 12496685b357SMike Rapoport pcpu_mem_free(chunk->md_blocks); 125040064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 125140064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 12521d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 12536081089fSTejun Heo } 12546081089fSTejun Heo 1255b539b87fSTejun Heo /** 1256b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 1257b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 1258b539b87fSTejun Heo * @page_start: the start page 1259b539b87fSTejun Heo * @page_end: the end page 1260b539b87fSTejun Heo * 1261b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1262b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 1263b539b87fSTejun Heo * successful population. 126440064aecSDennis Zhou (Facebook) * 126540064aecSDennis Zhou (Facebook) * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it 126640064aecSDennis Zhou (Facebook) * is to serve an allocation in that area. 1267b539b87fSTejun Heo */ 126840064aecSDennis Zhou (Facebook) static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 1269*b239f7daSDennis Zhou int page_end) 1270b539b87fSTejun Heo { 1271b539b87fSTejun Heo int nr = page_end - page_start; 1272b539b87fSTejun Heo 1273b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1274b539b87fSTejun Heo 1275b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 1276b539b87fSTejun Heo chunk->nr_populated += nr; 12777e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += nr; 127840064aecSDennis Zhou (Facebook) 1279*b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr); 128040064aecSDennis Zhou (Facebook) } 1281b539b87fSTejun Heo 1282b539b87fSTejun Heo /** 1283b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 1284b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 1285b539b87fSTejun Heo * @page_start: the start page 1286b539b87fSTejun Heo * @page_end: the end page 1287b539b87fSTejun Heo * 1288b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1289b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 1290b539b87fSTejun Heo * each successful depopulation. 1291b539b87fSTejun Heo */ 1292b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1293b539b87fSTejun Heo int page_start, int page_end) 1294b539b87fSTejun Heo { 1295b539b87fSTejun Heo int nr = page_end - page_start; 1296b539b87fSTejun Heo 1297b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1298b539b87fSTejun Heo 1299b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 1300b539b87fSTejun Heo chunk->nr_populated -= nr; 13017e8a6304SDennis Zhou (Facebook) pcpu_nr_populated -= nr; 1302*b239f7daSDennis Zhou 1303*b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr); 1304b539b87fSTejun Heo } 1305b539b87fSTejun Heo 1306fbf59bc9STejun Heo /* 13079f645532STejun Heo * Chunk management implementation. 1308fbf59bc9STejun Heo * 13099f645532STejun Heo * To allow different implementations, chunk alloc/free and 13109f645532STejun Heo * [de]population are implemented in a separate file which is pulled 13119f645532STejun Heo * into this file and compiled together. The following functions 13129f645532STejun Heo * should be implemented. 1313ccea34b5STejun Heo * 13149f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 13159f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 13169f645532STejun Heo * pcpu_create_chunk - create a new chunk 13179f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 13189f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 13199f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1320fbf59bc9STejun Heo */ 132115d9f3d1SDennis Zhou static int pcpu_populate_chunk(struct pcpu_chunk *chunk, 132247504ee0SDennis Zhou int page_start, int page_end, gfp_t gfp); 132315d9f3d1SDennis Zhou static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, 132415d9f3d1SDennis Zhou int page_start, int page_end); 132547504ee0SDennis Zhou static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); 13269f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 13279f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 13289f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1329fbf59bc9STejun Heo 1330b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 1331b0c9778bSTejun Heo #include "percpu-km.c" 1332b0c9778bSTejun Heo #else 13339f645532STejun Heo #include "percpu-vm.c" 1334b0c9778bSTejun Heo #endif 1335fbf59bc9STejun Heo 1336fbf59bc9STejun Heo /** 133788999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 133888999a89STejun Heo * @addr: address for which the chunk needs to be determined. 133988999a89STejun Heo * 1340c0ebfdc3SDennis Zhou (Facebook) * This is an internal function that handles all but static allocations. 1341c0ebfdc3SDennis Zhou (Facebook) * Static percpu address values should never be passed into the allocator. 1342c0ebfdc3SDennis Zhou (Facebook) * 134388999a89STejun Heo * RETURNS: 134488999a89STejun Heo * The address of the found chunk. 134588999a89STejun Heo */ 134688999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 134788999a89STejun Heo { 1348c0ebfdc3SDennis Zhou (Facebook) /* is it in the dynamic region (first chunk)? */ 1349560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1350c0ebfdc3SDennis Zhou (Facebook) return pcpu_first_chunk; 1351c0ebfdc3SDennis Zhou (Facebook) 1352c0ebfdc3SDennis Zhou (Facebook) /* is it in the reserved region? */ 1353560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 135488999a89STejun Heo return pcpu_reserved_chunk; 135588999a89STejun Heo 135688999a89STejun Heo /* 135788999a89STejun Heo * The address is relative to unit0 which might be unused and 135888999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 135988999a89STejun Heo * current processor before looking it up in the vmalloc 136088999a89STejun Heo * space. Note that any possible cpu id can be used here, so 136188999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 136288999a89STejun Heo */ 136388999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 13649f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 136588999a89STejun Heo } 136688999a89STejun Heo 136788999a89STejun Heo /** 1368edcb4639STejun Heo * pcpu_alloc - the percpu allocator 1369cae3aeb8STejun Heo * @size: size of area to allocate in bytes 1370fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 1371edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 13725835d96eSTejun Heo * @gfp: allocation flags 1373fbf59bc9STejun Heo * 13745835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 13750ea7eeecSDaniel Borkmann * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN 13760ea7eeecSDaniel Borkmann * then no warning will be triggered on invalid or failed allocation 13770ea7eeecSDaniel Borkmann * requests. 1378fbf59bc9STejun Heo * 1379fbf59bc9STejun Heo * RETURNS: 1380fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1381fbf59bc9STejun Heo */ 13825835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 13835835d96eSTejun Heo gfp_t gfp) 1384fbf59bc9STejun Heo { 1385554fef1cSDennis Zhou /* whitelisted flags that can be passed to the backing allocators */ 1386554fef1cSDennis Zhou gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN); 13870ea7eeecSDaniel Borkmann bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 13880ea7eeecSDaniel Borkmann bool do_warn = !(gfp & __GFP_NOWARN); 1389f2badb0cSTejun Heo static int warn_limit = 10; 13908744d859SDennis Zhou struct pcpu_chunk *chunk, *next; 1391f2badb0cSTejun Heo const char *err; 139240064aecSDennis Zhou (Facebook) int slot, off, cpu, ret; 1393403a91b1SJiri Kosina unsigned long flags; 1394f528f0b8SCatalin Marinas void __percpu *ptr; 139540064aecSDennis Zhou (Facebook) size_t bits, bit_align; 1396fbf59bc9STejun Heo 1397723ad1d9SAl Viro /* 139840064aecSDennis Zhou (Facebook) * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 139940064aecSDennis Zhou (Facebook) * therefore alignment must be a minimum of that many bytes. 140040064aecSDennis Zhou (Facebook) * An allocation may have internal fragmentation from rounding up 140140064aecSDennis Zhou (Facebook) * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1402723ad1d9SAl Viro */ 1403d2f3c384SDennis Zhou (Facebook) if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1404d2f3c384SDennis Zhou (Facebook) align = PCPU_MIN_ALLOC_SIZE; 1405723ad1d9SAl Viro 1406d2f3c384SDennis Zhou (Facebook) size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 140740064aecSDennis Zhou (Facebook) bits = size >> PCPU_MIN_ALLOC_SHIFT; 140840064aecSDennis Zhou (Facebook) bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 14092f69fa82SViro 14103ca45a46Szijun_hu if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 14113ca45a46Szijun_hu !is_power_of_2(align))) { 14120ea7eeecSDaniel Borkmann WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1413756a025fSJoe Perches size, align); 1414fbf59bc9STejun Heo return NULL; 1415fbf59bc9STejun Heo } 1416fbf59bc9STejun Heo 1417f52ba1feSKirill Tkhai if (!is_atomic) { 1418f52ba1feSKirill Tkhai /* 1419f52ba1feSKirill Tkhai * pcpu_balance_workfn() allocates memory under this mutex, 1420f52ba1feSKirill Tkhai * and it may wait for memory reclaim. Allow current task 1421f52ba1feSKirill Tkhai * to become OOM victim, in case of memory pressure. 1422f52ba1feSKirill Tkhai */ 1423f52ba1feSKirill Tkhai if (gfp & __GFP_NOFAIL) 14246710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 1425f52ba1feSKirill Tkhai else if (mutex_lock_killable(&pcpu_alloc_mutex)) 1426f52ba1feSKirill Tkhai return NULL; 1427f52ba1feSKirill Tkhai } 14286710e594STejun Heo 1429403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1430fbf59bc9STejun Heo 1431edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 1432edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 1433edcb4639STejun Heo chunk = pcpu_reserved_chunk; 1434833af842STejun Heo 143540064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 143640064aecSDennis Zhou (Facebook) if (off < 0) { 1437833af842STejun Heo err = "alloc from reserved chunk failed"; 1438ccea34b5STejun Heo goto fail_unlock; 1439f2badb0cSTejun Heo } 1440833af842STejun Heo 144140064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1442edcb4639STejun Heo if (off >= 0) 1443edcb4639STejun Heo goto area_found; 1444833af842STejun Heo 1445f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 1446ccea34b5STejun Heo goto fail_unlock; 1447edcb4639STejun Heo } 1448edcb4639STejun Heo 1449ccea34b5STejun Heo restart: 1450edcb4639STejun Heo /* search through normal chunks */ 1451fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 14528744d859SDennis Zhou list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) { 145340064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, 145440064aecSDennis Zhou (Facebook) is_atomic); 14558744d859SDennis Zhou if (off < 0) { 14568744d859SDennis Zhou if (slot < PCPU_SLOT_FAIL_THRESHOLD) 14578744d859SDennis Zhou pcpu_chunk_move(chunk, 0); 1458fbf59bc9STejun Heo continue; 14598744d859SDennis Zhou } 1460ccea34b5STejun Heo 146140064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1462fbf59bc9STejun Heo if (off >= 0) 1463fbf59bc9STejun Heo goto area_found; 146440064aecSDennis Zhou (Facebook) 1465fbf59bc9STejun Heo } 1466fbf59bc9STejun Heo } 1467fbf59bc9STejun Heo 1468403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1469ccea34b5STejun Heo 1470b38d08f3STejun Heo /* 1471b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 1472b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 1473b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 1474b38d08f3STejun Heo */ 147511df02bfSDennis Zhou if (is_atomic) { 147611df02bfSDennis Zhou err = "atomic alloc failed, no space left"; 14775835d96eSTejun Heo goto fail; 147811df02bfSDennis Zhou } 14795835d96eSTejun Heo 1480b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 1481554fef1cSDennis Zhou chunk = pcpu_create_chunk(pcpu_gfp); 1482f2badb0cSTejun Heo if (!chunk) { 1483f2badb0cSTejun Heo err = "failed to allocate new chunk"; 1484b38d08f3STejun Heo goto fail; 1485f2badb0cSTejun Heo } 1486ccea34b5STejun Heo 1487403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1488fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 1489b38d08f3STejun Heo } else { 1490b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1491b38d08f3STejun Heo } 1492b38d08f3STejun Heo 1493ccea34b5STejun Heo goto restart; 1494fbf59bc9STejun Heo 1495fbf59bc9STejun Heo area_found: 149630a5b536SDennis Zhou pcpu_stats_area_alloc(chunk, size); 1497403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1498ccea34b5STejun Heo 1499dca49645STejun Heo /* populate if not all pages are already there */ 15005835d96eSTejun Heo if (!is_atomic) { 1501e04d3208STejun Heo int page_start, page_end, rs, re; 1502e04d3208STejun Heo 1503dca49645STejun Heo page_start = PFN_DOWN(off); 1504dca49645STejun Heo page_end = PFN_UP(off + size); 1505dca49645STejun Heo 150691e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 150791e914c5SDennis Zhou (Facebook) page_start, page_end) { 1508dca49645STejun Heo WARN_ON(chunk->immutable); 1509dca49645STejun Heo 1510554fef1cSDennis Zhou ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp); 1511b38d08f3STejun Heo 1512403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1513b38d08f3STejun Heo if (ret) { 151440064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1515f2badb0cSTejun Heo err = "failed to populate"; 1516ccea34b5STejun Heo goto fail_unlock; 1517fbf59bc9STejun Heo } 1518*b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, re); 1519b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1520dca49645STejun Heo } 1521dca49645STejun Heo 1522ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1523e04d3208STejun Heo } 1524ccea34b5STejun Heo 15251a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 15261a4d7607STejun Heo pcpu_schedule_balance_work(); 15271a4d7607STejun Heo 1528dca49645STejun Heo /* clear the areas and return address relative to base address */ 1529dca49645STejun Heo for_each_possible_cpu(cpu) 1530dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1531dca49645STejun Heo 1532f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 15338a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1534df95e795SDennis Zhou 1535df95e795SDennis Zhou trace_percpu_alloc_percpu(reserved, is_atomic, size, align, 1536df95e795SDennis Zhou chunk->base_addr, off, ptr); 1537df95e795SDennis Zhou 1538f528f0b8SCatalin Marinas return ptr; 1539ccea34b5STejun Heo 1540ccea34b5STejun Heo fail_unlock: 1541403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1542b38d08f3STejun Heo fail: 1543df95e795SDennis Zhou trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1544df95e795SDennis Zhou 15450ea7eeecSDaniel Borkmann if (!is_atomic && do_warn && warn_limit) { 1546870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 15475835d96eSTejun Heo size, align, is_atomic, err); 1548f2badb0cSTejun Heo dump_stack(); 1549f2badb0cSTejun Heo if (!--warn_limit) 1550870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1551f2badb0cSTejun Heo } 15521a4d7607STejun Heo if (is_atomic) { 15531a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 15541a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 15551a4d7607STejun Heo pcpu_schedule_balance_work(); 15566710e594STejun Heo } else { 15576710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 15581a4d7607STejun Heo } 1559ccea34b5STejun Heo return NULL; 1560fbf59bc9STejun Heo } 1561edcb4639STejun Heo 1562edcb4639STejun Heo /** 15635835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1564edcb4639STejun Heo * @size: size of area to allocate in bytes 1565edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 15665835d96eSTejun Heo * @gfp: allocation flags 1567edcb4639STejun Heo * 15685835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 15695835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 15700ea7eeecSDaniel Borkmann * be called from any context but is a lot more likely to fail. If @gfp 15710ea7eeecSDaniel Borkmann * has __GFP_NOWARN then no warning will be triggered on invalid or failed 15720ea7eeecSDaniel Borkmann * allocation requests. 1573ccea34b5STejun Heo * 1574edcb4639STejun Heo * RETURNS: 1575edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1576edcb4639STejun Heo */ 15775835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 15785835d96eSTejun Heo { 15795835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 15805835d96eSTejun Heo } 15815835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 15825835d96eSTejun Heo 15835835d96eSTejun Heo /** 15845835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 15855835d96eSTejun Heo * @size: size of area to allocate in bytes 15865835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 15875835d96eSTejun Heo * 15885835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 15895835d96eSTejun Heo */ 159043cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1591edcb4639STejun Heo { 15925835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1593edcb4639STejun Heo } 1594fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1595fbf59bc9STejun Heo 1596edcb4639STejun Heo /** 1597edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1598edcb4639STejun Heo * @size: size of area to allocate in bytes 1599edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1600edcb4639STejun Heo * 16019329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 16029329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 16039329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 16049329ba97STejun Heo * Might trigger writeouts. 1605edcb4639STejun Heo * 1606ccea34b5STejun Heo * CONTEXT: 1607ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1608ccea34b5STejun Heo * 1609edcb4639STejun Heo * RETURNS: 1610edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1611edcb4639STejun Heo */ 161243cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1613edcb4639STejun Heo { 16145835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1615edcb4639STejun Heo } 1616edcb4639STejun Heo 1617a56dbddfSTejun Heo /** 16181a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1619a56dbddfSTejun Heo * @work: unused 1620a56dbddfSTejun Heo * 162147504ee0SDennis Zhou * Reclaim all fully free chunks except for the first one. This is also 162247504ee0SDennis Zhou * responsible for maintaining the pool of empty populated pages. However, 162347504ee0SDennis Zhou * it is possible that this is called when physical memory is scarce causing 162447504ee0SDennis Zhou * OOM killer to be triggered. We should avoid doing so until an actual 162547504ee0SDennis Zhou * allocation causes the failure as it is possible that requests can be 162647504ee0SDennis Zhou * serviced from already backed regions. 1627a56dbddfSTejun Heo */ 1628fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1629fbf59bc9STejun Heo { 163047504ee0SDennis Zhou /* gfp flags passed to underlying allocators */ 1631554fef1cSDennis Zhou const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN; 1632fe6bd8c3STejun Heo LIST_HEAD(to_free); 1633fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1634a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 16351a4d7607STejun Heo int slot, nr_to_pop, ret; 1636a56dbddfSTejun Heo 16371a4d7607STejun Heo /* 16381a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 16391a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 16401a4d7607STejun Heo */ 1641ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1642ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1643a56dbddfSTejun Heo 1644fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 16458d408b4bSTejun Heo WARN_ON(chunk->immutable); 1646a56dbddfSTejun Heo 1647a56dbddfSTejun Heo /* spare the first one */ 1648fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1649a56dbddfSTejun Heo continue; 1650a56dbddfSTejun Heo 1651fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1652a56dbddfSTejun Heo } 1653a56dbddfSTejun Heo 1654ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1655a56dbddfSTejun Heo 1656fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1657a93ace48STejun Heo int rs, re; 1658dca49645STejun Heo 165991e914c5SDennis Zhou (Facebook) pcpu_for_each_pop_region(chunk->populated, rs, re, 0, 166091e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1661a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1662b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1663b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1664b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1665a93ace48STejun Heo } 16666081089fSTejun Heo pcpu_destroy_chunk(chunk); 1667accd4f36SEric Dumazet cond_resched(); 1668fbf59bc9STejun Heo } 1669971f3918STejun Heo 16701a4d7607STejun Heo /* 16711a4d7607STejun Heo * Ensure there are certain number of free populated pages for 16721a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 16731a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 16741a4d7607STejun Heo * failed previously, always populate the maximum amount. This 16751a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 16761a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 16771a4d7607STejun Heo * something we support properly and can be highly unreliable and 16781a4d7607STejun Heo * inefficient. 16791a4d7607STejun Heo */ 16801a4d7607STejun Heo retry_pop: 16811a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 16821a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 16831a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 16841a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 16851a4d7607STejun Heo } else { 16861a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 16871a4d7607STejun Heo pcpu_nr_empty_pop_pages, 16881a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 16891a4d7607STejun Heo } 16901a4d7607STejun Heo 16911a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 16921a4d7607STejun Heo int nr_unpop = 0, rs, re; 16931a4d7607STejun Heo 16941a4d7607STejun Heo if (!nr_to_pop) 16951a4d7607STejun Heo break; 16961a4d7607STejun Heo 16971a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 16981a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 16998ab16c43SDennis Zhou (Facebook) nr_unpop = chunk->nr_pages - chunk->nr_populated; 17001a4d7607STejun Heo if (nr_unpop) 17011a4d7607STejun Heo break; 17021a4d7607STejun Heo } 17031a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 17041a4d7607STejun Heo 17051a4d7607STejun Heo if (!nr_unpop) 17061a4d7607STejun Heo continue; 17071a4d7607STejun Heo 17081a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 170991e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, 171091e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 17111a4d7607STejun Heo int nr = min(re - rs, nr_to_pop); 17121a4d7607STejun Heo 171347504ee0SDennis Zhou ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); 17141a4d7607STejun Heo if (!ret) { 17151a4d7607STejun Heo nr_to_pop -= nr; 17161a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 1717*b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, rs + nr); 17181a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 17191a4d7607STejun Heo } else { 17201a4d7607STejun Heo nr_to_pop = 0; 17211a4d7607STejun Heo } 17221a4d7607STejun Heo 17231a4d7607STejun Heo if (!nr_to_pop) 17241a4d7607STejun Heo break; 17251a4d7607STejun Heo } 17261a4d7607STejun Heo } 17271a4d7607STejun Heo 17281a4d7607STejun Heo if (nr_to_pop) { 17291a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 173047504ee0SDennis Zhou chunk = pcpu_create_chunk(gfp); 17311a4d7607STejun Heo if (chunk) { 17321a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 17331a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 17341a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 17351a4d7607STejun Heo goto retry_pop; 17361a4d7607STejun Heo } 17371a4d7607STejun Heo } 17381a4d7607STejun Heo 1739971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1740a56dbddfSTejun Heo } 1741fbf59bc9STejun Heo 1742fbf59bc9STejun Heo /** 1743fbf59bc9STejun Heo * free_percpu - free percpu area 1744fbf59bc9STejun Heo * @ptr: pointer to area to free 1745fbf59bc9STejun Heo * 1746ccea34b5STejun Heo * Free percpu area @ptr. 1747ccea34b5STejun Heo * 1748ccea34b5STejun Heo * CONTEXT: 1749ccea34b5STejun Heo * Can be called from atomic context. 1750fbf59bc9STejun Heo */ 175143cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1752fbf59bc9STejun Heo { 1753129182e5SAndrew Morton void *addr; 1754fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1755ccea34b5STejun Heo unsigned long flags; 175640064aecSDennis Zhou (Facebook) int off; 1757fbf59bc9STejun Heo 1758fbf59bc9STejun Heo if (!ptr) 1759fbf59bc9STejun Heo return; 1760fbf59bc9STejun Heo 1761f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1762f528f0b8SCatalin Marinas 1763129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1764129182e5SAndrew Morton 1765ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1766fbf59bc9STejun Heo 1767fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1768bba174f5STejun Heo off = addr - chunk->base_addr; 1769fbf59bc9STejun Heo 177040064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1771fbf59bc9STejun Heo 1772a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 177340064aecSDennis Zhou (Facebook) if (chunk->free_bytes == pcpu_unit_size) { 1774fbf59bc9STejun Heo struct pcpu_chunk *pos; 1775fbf59bc9STejun Heo 1776a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1777fbf59bc9STejun Heo if (pos != chunk) { 17781a4d7607STejun Heo pcpu_schedule_balance_work(); 1779fbf59bc9STejun Heo break; 1780fbf59bc9STejun Heo } 1781fbf59bc9STejun Heo } 1782fbf59bc9STejun Heo 1783df95e795SDennis Zhou trace_percpu_free_percpu(chunk->base_addr, off, ptr); 1784df95e795SDennis Zhou 1785ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1786fbf59bc9STejun Heo } 1787fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1788fbf59bc9STejun Heo 1789383776faSThomas Gleixner bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 1790383776faSThomas Gleixner { 1791383776faSThomas Gleixner #ifdef CONFIG_SMP 1792383776faSThomas Gleixner const size_t static_size = __per_cpu_end - __per_cpu_start; 1793383776faSThomas Gleixner void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 1794383776faSThomas Gleixner unsigned int cpu; 1795383776faSThomas Gleixner 1796383776faSThomas Gleixner for_each_possible_cpu(cpu) { 1797383776faSThomas Gleixner void *start = per_cpu_ptr(base, cpu); 1798383776faSThomas Gleixner void *va = (void *)addr; 1799383776faSThomas Gleixner 1800383776faSThomas Gleixner if (va >= start && va < start + static_size) { 18018ce371f9SPeter Zijlstra if (can_addr) { 1802383776faSThomas Gleixner *can_addr = (unsigned long) (va - start); 18038ce371f9SPeter Zijlstra *can_addr += (unsigned long) 18048ce371f9SPeter Zijlstra per_cpu_ptr(base, get_boot_cpu_id()); 18058ce371f9SPeter Zijlstra } 1806383776faSThomas Gleixner return true; 1807383776faSThomas Gleixner } 1808383776faSThomas Gleixner } 1809383776faSThomas Gleixner #endif 1810383776faSThomas Gleixner /* on UP, can't distinguish from other static vars, always false */ 1811383776faSThomas Gleixner return false; 1812383776faSThomas Gleixner } 1813383776faSThomas Gleixner 18143b034b0dSVivek Goyal /** 181510fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 181610fad5e4STejun Heo * @addr: address to test 181710fad5e4STejun Heo * 181810fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 181910fad5e4STejun Heo * static percpu areas are not considered. For those, use 182010fad5e4STejun Heo * is_module_percpu_address(). 182110fad5e4STejun Heo * 182210fad5e4STejun Heo * RETURNS: 182310fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 182410fad5e4STejun Heo */ 182510fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 182610fad5e4STejun Heo { 1827383776faSThomas Gleixner return __is_kernel_percpu_address(addr, NULL); 182810fad5e4STejun Heo } 182910fad5e4STejun Heo 183010fad5e4STejun Heo /** 18313b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 18323b034b0dSVivek Goyal * @addr: the address to be converted to physical address 18333b034b0dSVivek Goyal * 18343b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 18353b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 18363b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 18373b034b0dSVivek Goyal * until this function finishes. 18383b034b0dSVivek Goyal * 183967589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 184067589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 184167589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 184267589c71SDave Young * km) provides translation. 184367589c71SDave Young * 1844bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 184567589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 184667589c71SDave Young * actually works, and the verification can discover both bugs in percpu 184767589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 184867589c71SDave Young * code. 184967589c71SDave Young * 18503b034b0dSVivek Goyal * RETURNS: 18513b034b0dSVivek Goyal * The physical address for @addr. 18523b034b0dSVivek Goyal */ 18533b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 18543b034b0dSVivek Goyal { 18559983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 18569983b6f0STejun Heo bool in_first_chunk = false; 1857a855b84cSTejun Heo unsigned long first_low, first_high; 18589983b6f0STejun Heo unsigned int cpu; 18599983b6f0STejun Heo 18609983b6f0STejun Heo /* 1861a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 18629983b6f0STejun Heo * necessary but will speed up lookups of addresses which 18639983b6f0STejun Heo * aren't in the first chunk. 1864c0ebfdc3SDennis Zhou (Facebook) * 1865c0ebfdc3SDennis Zhou (Facebook) * The address check is against full chunk sizes. pcpu_base_addr 1866c0ebfdc3SDennis Zhou (Facebook) * points to the beginning of the first chunk including the 1867c0ebfdc3SDennis Zhou (Facebook) * static region. Assumes good intent as the first chunk may 1868c0ebfdc3SDennis Zhou (Facebook) * not be full (ie. < pcpu_unit_pages in size). 18699983b6f0STejun Heo */ 1870c0ebfdc3SDennis Zhou (Facebook) first_low = (unsigned long)pcpu_base_addr + 1871c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 1872c0ebfdc3SDennis Zhou (Facebook) first_high = (unsigned long)pcpu_base_addr + 1873c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 1874a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1875a855b84cSTejun Heo (unsigned long)addr < first_high) { 18769983b6f0STejun Heo for_each_possible_cpu(cpu) { 18779983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 18789983b6f0STejun Heo 18799983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 18809983b6f0STejun Heo in_first_chunk = true; 18819983b6f0STejun Heo break; 18829983b6f0STejun Heo } 18839983b6f0STejun Heo } 18849983b6f0STejun Heo } 18859983b6f0STejun Heo 18869983b6f0STejun Heo if (in_first_chunk) { 1887eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 18883b034b0dSVivek Goyal return __pa(addr); 18893b034b0dSVivek Goyal else 18909f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 18919f57bd4dSEugene Surovegin offset_in_page(addr); 1892020ec653STejun Heo } else 18939f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 18949f57bd4dSEugene Surovegin offset_in_page(addr); 18953b034b0dSVivek Goyal } 18963b034b0dSVivek Goyal 1897fbf59bc9STejun Heo /** 1898fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1899fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1900fd1e8a1fSTejun Heo * @nr_units: the number of units 1901033e48fbSTejun Heo * 1902fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1903fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1904fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1905fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1906fd1e8a1fSTejun Heo * pointer of other groups. 1907033e48fbSTejun Heo * 1908033e48fbSTejun Heo * RETURNS: 1909fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1910fd1e8a1fSTejun Heo * failure. 1911033e48fbSTejun Heo */ 1912fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1913fd1e8a1fSTejun Heo int nr_units) 1914fd1e8a1fSTejun Heo { 1915fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 1916fd1e8a1fSTejun Heo size_t base_size, ai_size; 1917fd1e8a1fSTejun Heo void *ptr; 1918fd1e8a1fSTejun Heo int unit; 1919fd1e8a1fSTejun Heo 1920fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1921fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 1922fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1923fd1e8a1fSTejun Heo 192426fb3daeSMike Rapoport ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE); 1925fd1e8a1fSTejun Heo if (!ptr) 1926fd1e8a1fSTejun Heo return NULL; 1927fd1e8a1fSTejun Heo ai = ptr; 1928fd1e8a1fSTejun Heo ptr += base_size; 1929fd1e8a1fSTejun Heo 1930fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 1931fd1e8a1fSTejun Heo 1932fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 1933fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 1934fd1e8a1fSTejun Heo 1935fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 1936fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 1937fd1e8a1fSTejun Heo 1938fd1e8a1fSTejun Heo return ai; 1939fd1e8a1fSTejun Heo } 1940fd1e8a1fSTejun Heo 1941fd1e8a1fSTejun Heo /** 1942fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 1943fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 1944fd1e8a1fSTejun Heo * 1945fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1946fd1e8a1fSTejun Heo */ 1947fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1948fd1e8a1fSTejun Heo { 1949999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 1950fd1e8a1fSTejun Heo } 1951fd1e8a1fSTejun Heo 1952fd1e8a1fSTejun Heo /** 1953fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1954fd1e8a1fSTejun Heo * @lvl: loglevel 1955fd1e8a1fSTejun Heo * @ai: allocation info to dump 1956fd1e8a1fSTejun Heo * 1957fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 1958fd1e8a1fSTejun Heo */ 1959fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 1960fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 1961033e48fbSTejun Heo { 1962fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 1963033e48fbSTejun Heo char empty_str[] = "--------"; 1964fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 1965fd1e8a1fSTejun Heo int group, v; 1966fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 1967033e48fbSTejun Heo 1968fd1e8a1fSTejun Heo v = ai->nr_groups; 1969033e48fbSTejun Heo while (v /= 10) 1970fd1e8a1fSTejun Heo group_width++; 1971033e48fbSTejun Heo 1972fd1e8a1fSTejun Heo v = num_possible_cpus(); 1973fd1e8a1fSTejun Heo while (v /= 10) 1974fd1e8a1fSTejun Heo cpu_width++; 1975fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1976033e48fbSTejun Heo 1977fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 1978fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 1979fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 1980033e48fbSTejun Heo 1981fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1982fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1983fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1984fd1e8a1fSTejun Heo 1985fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 1986fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 1987fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 1988fd1e8a1fSTejun Heo 1989fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 1990fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 1991fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 1992fd1e8a1fSTejun Heo if (!(alloc % apl)) { 19931170532bSJoe Perches pr_cont("\n"); 1994fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 1995033e48fbSTejun Heo } 19961170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 1997fd1e8a1fSTejun Heo 1998fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 1999fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 20001170532bSJoe Perches pr_cont("%0*d ", 20011170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 2002033e48fbSTejun Heo else 20031170532bSJoe Perches pr_cont("%s ", empty_str); 2004033e48fbSTejun Heo } 2005fd1e8a1fSTejun Heo } 20061170532bSJoe Perches pr_cont("\n"); 2007033e48fbSTejun Heo } 2008033e48fbSTejun Heo 2009fbf59bc9STejun Heo /** 20108d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 2011fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 201238a6be52STejun Heo * @base_addr: mapped address 2013fbf59bc9STejun Heo * 20148d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 20158d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 201638a6be52STejun Heo * setup path. 20178d408b4bSTejun Heo * 2018fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 2019fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 20208d408b4bSTejun Heo * 2021fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 2022fd1e8a1fSTejun Heo * 2023fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 2024edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 2025edcb4639STejun Heo * the first chunk such that it's available only through reserved 2026edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 2027edcb4639STejun Heo * static areas on architectures where the addressing model has 2028edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 2029edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 2030edcb4639STejun Heo * 2031fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 2032fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 2033fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 20346074d5b0STejun Heo * 2035fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 2036fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 2037fd1e8a1fSTejun Heo * @ai->dyn_size. 20388d408b4bSTejun Heo * 2039fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 2040fd1e8a1fSTejun Heo * for vm areas. 20418d408b4bSTejun Heo * 2042fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 2043fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 2044fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 2045fd1e8a1fSTejun Heo * 2046fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 2047fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 2048fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 2049fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 2050fd1e8a1fSTejun Heo * all units is assumed. 20518d408b4bSTejun Heo * 205238a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 205338a6be52STejun Heo * copied static data to each unit. 2054fbf59bc9STejun Heo * 2055c0ebfdc3SDennis Zhou (Facebook) * The first chunk will always contain a static and a dynamic region. 2056c0ebfdc3SDennis Zhou (Facebook) * However, the static region is not managed by any chunk. If the first 2057c0ebfdc3SDennis Zhou (Facebook) * chunk also contains a reserved region, it is served by two chunks - 2058c0ebfdc3SDennis Zhou (Facebook) * one for the reserved region and one for the dynamic region. They 2059c0ebfdc3SDennis Zhou (Facebook) * share the same vm, but use offset regions in the area allocation map. 2060c0ebfdc3SDennis Zhou (Facebook) * The chunk serving the dynamic region is circulated in the chunk slots 2061c0ebfdc3SDennis Zhou (Facebook) * and available for dynamic allocation like any other chunk. 2062edcb4639STejun Heo * 2063fbf59bc9STejun Heo * RETURNS: 2064fb435d52STejun Heo * 0 on success, -errno on failure. 2065fbf59bc9STejun Heo */ 2066fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 2067fd1e8a1fSTejun Heo void *base_addr) 2068fbf59bc9STejun Heo { 2069b9c39442SDennis Zhou (Facebook) size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2070d2f3c384SDennis Zhou (Facebook) size_t static_size, dyn_size; 20710c4169c3SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 20726563297cSTejun Heo unsigned long *group_offsets; 20736563297cSTejun Heo size_t *group_sizes; 2074fb435d52STejun Heo unsigned long *unit_off; 2075fbf59bc9STejun Heo unsigned int cpu; 2076fd1e8a1fSTejun Heo int *unit_map; 2077fd1e8a1fSTejun Heo int group, unit, i; 2078c0ebfdc3SDennis Zhou (Facebook) int map_size; 2079c0ebfdc3SDennis Zhou (Facebook) unsigned long tmp_addr; 2080f655f405SMike Rapoport size_t alloc_size; 2081fbf59bc9STejun Heo 2082635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 2083635b75fcSTejun Heo if (unlikely(cond)) { \ 2084870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 2085870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 2086807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 2087635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 2088635b75fcSTejun Heo BUG(); \ 2089635b75fcSTejun Heo } \ 2090635b75fcSTejun Heo } while (0) 2091635b75fcSTejun Heo 20922f39e637STejun Heo /* sanity checks */ 2093635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 2094bbddff05STejun Heo #ifdef CONFIG_SMP 2095635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 2096f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 2097bbddff05STejun Heo #endif 2098635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 2099f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 2100635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 2101f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 2102635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 2103ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 2104099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 2105fb29a2ccSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!ai->dyn_size); 2106d2f3c384SDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 2107ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 2108ca460b3cSDennis Zhou (Facebook) IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 21099f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 21108d408b4bSTejun Heo 21116563297cSTejun Heo /* process group information and build config tables accordingly */ 2112f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_offsets[0]); 2113f655f405SMike Rapoport group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2114f655f405SMike Rapoport if (!group_offsets) 2115f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2116f655f405SMike Rapoport alloc_size); 2117f655f405SMike Rapoport 2118f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_sizes[0]); 2119f655f405SMike Rapoport group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2120f655f405SMike Rapoport if (!group_sizes) 2121f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2122f655f405SMike Rapoport alloc_size); 2123f655f405SMike Rapoport 2124f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_map[0]); 2125f655f405SMike Rapoport unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2126f655f405SMike Rapoport if (!unit_map) 2127f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2128f655f405SMike Rapoport alloc_size); 2129f655f405SMike Rapoport 2130f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_off[0]); 2131f655f405SMike Rapoport unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2132f655f405SMike Rapoport if (!unit_off) 2133f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2134f655f405SMike Rapoport alloc_size); 21352f39e637STejun Heo 2136fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 2137ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 2138a855b84cSTejun Heo 2139a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 2140a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 21412f39e637STejun Heo 2142fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 2143fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 21442f39e637STejun Heo 21456563297cSTejun Heo group_offsets[group] = gi->base_offset; 21466563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 21476563297cSTejun Heo 2148fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 2149fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 2150fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 2151fd1e8a1fSTejun Heo continue; 2152fd1e8a1fSTejun Heo 21539f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 2154635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 2155635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 2156fd1e8a1fSTejun Heo 2157fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 2158fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 2159fb435d52STejun Heo 2160a855b84cSTejun Heo /* determine low/high unit_cpu */ 2161a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 2162a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 2163a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 2164a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 2165a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 2166a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 21670fc0531eSLinus Torvalds } 21680fc0531eSLinus Torvalds } 2169fd1e8a1fSTejun Heo pcpu_nr_units = unit; 21702f39e637STejun Heo 21712f39e637STejun Heo for_each_possible_cpu(cpu) 2172635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 2173635b75fcSTejun Heo 2174635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 2175635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 2176bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 21772f39e637STejun Heo 21786563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 21796563297cSTejun Heo pcpu_group_offsets = group_offsets; 21806563297cSTejun Heo pcpu_group_sizes = group_sizes; 2181fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 2182fb435d52STejun Heo pcpu_unit_offsets = unit_off; 21832f39e637STejun Heo 21842f39e637STejun Heo /* determine basic parameters */ 2185fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 2186d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 21876563297cSTejun Heo pcpu_atom_size = ai->atom_size; 2188ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 2189ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 2190cafe8816STejun Heo 219130a5b536SDennis Zhou pcpu_stats_save_ai(ai); 219230a5b536SDennis Zhou 2193d9b55eebSTejun Heo /* 2194d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 2195d9b55eebSTejun Heo * empty chunks. 2196d9b55eebSTejun Heo */ 2197d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 21987e1c4e27SMike Rapoport pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]), 21997e1c4e27SMike Rapoport SMP_CACHE_BYTES); 2200f655f405SMike Rapoport if (!pcpu_slot) 2201f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2202f655f405SMike Rapoport pcpu_nr_slots * sizeof(pcpu_slot[0])); 2203fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 2204fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 2205fbf59bc9STejun Heo 2206edcb4639STejun Heo /* 2207d2f3c384SDennis Zhou (Facebook) * The end of the static region needs to be aligned with the 2208d2f3c384SDennis Zhou (Facebook) * minimum allocation size as this offsets the reserved and 2209d2f3c384SDennis Zhou (Facebook) * dynamic region. The first chunk ends page aligned by 2210d2f3c384SDennis Zhou (Facebook) * expanding the dynamic region, therefore the dynamic region 2211d2f3c384SDennis Zhou (Facebook) * can be shrunk to compensate while still staying above the 2212d2f3c384SDennis Zhou (Facebook) * configured sizes. 2213d2f3c384SDennis Zhou (Facebook) */ 2214d2f3c384SDennis Zhou (Facebook) static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 2215d2f3c384SDennis Zhou (Facebook) dyn_size = ai->dyn_size - (static_size - ai->static_size); 2216d2f3c384SDennis Zhou (Facebook) 2217d2f3c384SDennis Zhou (Facebook) /* 2218c0ebfdc3SDennis Zhou (Facebook) * Initialize first chunk. 2219c0ebfdc3SDennis Zhou (Facebook) * If the reserved_size is non-zero, this initializes the reserved 2220c0ebfdc3SDennis Zhou (Facebook) * chunk. If the reserved_size is zero, the reserved chunk is NULL 2221c0ebfdc3SDennis Zhou (Facebook) * and the dynamic region is initialized here. The first chunk, 2222c0ebfdc3SDennis Zhou (Facebook) * pcpu_first_chunk, will always point to the chunk that serves 2223c0ebfdc3SDennis Zhou (Facebook) * the dynamic region. 2224edcb4639STejun Heo */ 2225d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size; 2226d2f3c384SDennis Zhou (Facebook) map_size = ai->reserved_size ?: dyn_size; 222740064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 222861ace7faSTejun Heo 2229edcb4639STejun Heo /* init dynamic chunk if necessary */ 2230b9c39442SDennis Zhou (Facebook) if (ai->reserved_size) { 22310c4169c3SDennis Zhou (Facebook) pcpu_reserved_chunk = chunk; 2232b9c39442SDennis Zhou (Facebook) 2233d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size + 2234c0ebfdc3SDennis Zhou (Facebook) ai->reserved_size; 2235d2f3c384SDennis Zhou (Facebook) map_size = dyn_size; 223640064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 2237edcb4639STejun Heo } 2238edcb4639STejun Heo 22392441d15cSTejun Heo /* link the first chunk in */ 22400c4169c3SDennis Zhou (Facebook) pcpu_first_chunk = chunk; 22410cecf50cSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 2242ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 2243fbf59bc9STejun Heo 22447e8a6304SDennis Zhou (Facebook) /* include all regions of the first chunk */ 22457e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += PFN_DOWN(size_sum); 22467e8a6304SDennis Zhou (Facebook) 224730a5b536SDennis Zhou pcpu_stats_chunk_alloc(); 2248df95e795SDennis Zhou trace_percpu_create_chunk(base_addr); 224930a5b536SDennis Zhou 2250fbf59bc9STejun Heo /* we're done */ 2251bba174f5STejun Heo pcpu_base_addr = base_addr; 2252fb435d52STejun Heo return 0; 2253fbf59bc9STejun Heo } 225466c3a757STejun Heo 2255bbddff05STejun Heo #ifdef CONFIG_SMP 2256bbddff05STejun Heo 225717f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 2258f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 2259f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 2260f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 2261f58dc01bSTejun Heo }; 226266c3a757STejun Heo 2263f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 2264f58dc01bSTejun Heo 2265f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 226666c3a757STejun Heo { 22675479c78aSCyrill Gorcunov if (!str) 22685479c78aSCyrill Gorcunov return -EINVAL; 22695479c78aSCyrill Gorcunov 2270f58dc01bSTejun Heo if (0) 2271f58dc01bSTejun Heo /* nada */; 2272f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 2273f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 2274f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 2275f58dc01bSTejun Heo #endif 2276f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2277f58dc01bSTejun Heo else if (!strcmp(str, "page")) 2278f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 2279f58dc01bSTejun Heo #endif 2280f58dc01bSTejun Heo else 2281870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 228266c3a757STejun Heo 2283f58dc01bSTejun Heo return 0; 228466c3a757STejun Heo } 2285f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 228666c3a757STejun Heo 22873c9a024fSTejun Heo /* 22883c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 22893c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 22903c9a024fSTejun Heo * to be used. 22913c9a024fSTejun Heo */ 229208fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 229308fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 22943c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 22953c9a024fSTejun Heo #endif 22963c9a024fSTejun Heo 22973c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 22983c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 22993c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 23003c9a024fSTejun Heo #endif 23013c9a024fSTejun Heo 23023c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 23033c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 23043c9a024fSTejun Heo /** 2305fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2306fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2307fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2308fbf59bc9STejun Heo * @atom_size: allocation atom size 2309fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2310fbf59bc9STejun Heo * 2311fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 2312fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 2313fbf59bc9STejun Heo * atom size and distances between CPUs. 2314fbf59bc9STejun Heo * 2315bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 2316fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 2317fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 2318fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 2319fbf59bc9STejun Heo * of allocated virtual address space. 2320fbf59bc9STejun Heo * 2321fbf59bc9STejun Heo * RETURNS: 2322fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 2323fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 2324fbf59bc9STejun Heo */ 2325fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 2326fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 2327fbf59bc9STejun Heo size_t atom_size, 2328fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2329fbf59bc9STejun Heo { 2330fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 2331fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 2332fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 2333fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 2334fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 2335fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 2336fbf59bc9STejun Heo int last_allocs, group, unit; 2337fbf59bc9STejun Heo unsigned int cpu, tcpu; 2338fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 2339fbf59bc9STejun Heo unsigned int *cpu_map; 2340fbf59bc9STejun Heo 2341fbf59bc9STejun Heo /* this function may be called multiple times */ 2342fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 2343fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 2344fbf59bc9STejun Heo 2345fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2346fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 2347fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2348fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 2349fbf59bc9STejun Heo 2350fbf59bc9STejun Heo /* 2351fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 2352fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 235325985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 2354fbf59bc9STejun Heo * or larger than min_unit_size. 2355fbf59bc9STejun Heo */ 2356fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2357fbf59bc9STejun Heo 23589c015162SDennis Zhou (Facebook) /* determine the maximum # of units that can fit in an allocation */ 2359fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 2360fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 2361f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2362fbf59bc9STejun Heo upa--; 2363fbf59bc9STejun Heo max_upa = upa; 2364fbf59bc9STejun Heo 2365fbf59bc9STejun Heo /* group cpus according to their proximity */ 2366fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 2367fbf59bc9STejun Heo group = 0; 2368fbf59bc9STejun Heo next_group: 2369fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 2370fbf59bc9STejun Heo if (cpu == tcpu) 2371fbf59bc9STejun Heo break; 2372fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 2373fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 2374fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 2375fbf59bc9STejun Heo group++; 2376fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 2377fbf59bc9STejun Heo goto next_group; 2378fbf59bc9STejun Heo } 2379fbf59bc9STejun Heo } 2380fbf59bc9STejun Heo group_map[cpu] = group; 2381fbf59bc9STejun Heo group_cnt[group]++; 2382fbf59bc9STejun Heo } 2383fbf59bc9STejun Heo 2384fbf59bc9STejun Heo /* 23859c015162SDennis Zhou (Facebook) * Wasted space is caused by a ratio imbalance of upa to group_cnt. 23869c015162SDennis Zhou (Facebook) * Expand the unit_size until we use >= 75% of the units allocated. 23879c015162SDennis Zhou (Facebook) * Related to atom_size, which could be much larger than the unit_size. 2388fbf59bc9STejun Heo */ 2389fbf59bc9STejun Heo last_allocs = INT_MAX; 2390fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 2391fbf59bc9STejun Heo int allocs = 0, wasted = 0; 2392fbf59bc9STejun Heo 2393f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2394fbf59bc9STejun Heo continue; 2395fbf59bc9STejun Heo 2396fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2397fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2398fbf59bc9STejun Heo allocs += this_allocs; 2399fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 2400fbf59bc9STejun Heo } 2401fbf59bc9STejun Heo 2402fbf59bc9STejun Heo /* 2403fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 2404fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 2405fbf59bc9STejun Heo * passes the following check. 2406fbf59bc9STejun Heo */ 2407fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 2408fbf59bc9STejun Heo continue; 2409fbf59bc9STejun Heo 2410fbf59bc9STejun Heo /* and then don't consume more memory */ 2411fbf59bc9STejun Heo if (allocs > last_allocs) 2412fbf59bc9STejun Heo break; 2413fbf59bc9STejun Heo last_allocs = allocs; 2414fbf59bc9STejun Heo best_upa = upa; 2415fbf59bc9STejun Heo } 2416fbf59bc9STejun Heo upa = best_upa; 2417fbf59bc9STejun Heo 2418fbf59bc9STejun Heo /* allocate and fill alloc_info */ 2419fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 2420fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 2421fbf59bc9STejun Heo 2422fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2423fbf59bc9STejun Heo if (!ai) 2424fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 2425fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 2426fbf59bc9STejun Heo 2427fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2428fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 2429fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 2430fbf59bc9STejun Heo } 2431fbf59bc9STejun Heo 2432fbf59bc9STejun Heo ai->static_size = static_size; 2433fbf59bc9STejun Heo ai->reserved_size = reserved_size; 2434fbf59bc9STejun Heo ai->dyn_size = dyn_size; 2435fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 2436fbf59bc9STejun Heo ai->atom_size = atom_size; 2437fbf59bc9STejun Heo ai->alloc_size = alloc_size; 2438fbf59bc9STejun Heo 24392de7852fSPeng Fan for (group = 0, unit = 0; group < nr_groups; group++) { 2440fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2441fbf59bc9STejun Heo 2442fbf59bc9STejun Heo /* 2443fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 2444fbf59bc9STejun Heo * back-to-back. The caller should update this to 2445fbf59bc9STejun Heo * reflect actual allocation. 2446fbf59bc9STejun Heo */ 2447fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 2448fbf59bc9STejun Heo 2449fbf59bc9STejun Heo for_each_possible_cpu(cpu) 2450fbf59bc9STejun Heo if (group_map[cpu] == group) 2451fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 2452fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 2453fbf59bc9STejun Heo unit += gi->nr_units; 2454fbf59bc9STejun Heo } 2455fbf59bc9STejun Heo BUG_ON(unit != nr_units); 2456fbf59bc9STejun Heo 2457fbf59bc9STejun Heo return ai; 2458fbf59bc9STejun Heo } 24593c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2460fbf59bc9STejun Heo 24613c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 246266c3a757STejun Heo /** 246366c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 246466c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 24654ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2466c8826dd5STejun Heo * @atom_size: allocation atom size 2467c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2468c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 246925985edcSLucas De Marchi * @free_fn: function to free percpu page 247066c3a757STejun Heo * 247166c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 247266c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 247366c3a757STejun Heo * 247466c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 2475c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 2476c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 2477c8826dd5STejun Heo * aligned to @atom_size. 2478c8826dd5STejun Heo * 2479c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 2480c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 2481c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 2482c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 2483c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 2484c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 248566c3a757STejun Heo * 24864ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 248766c3a757STejun Heo * 248866c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 2489c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 249066c3a757STejun Heo * 249166c3a757STejun Heo * RETURNS: 2492fb435d52STejun Heo * 0 on success, -errno on failure. 249366c3a757STejun Heo */ 24944ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 2495c8826dd5STejun Heo size_t atom_size, 2496c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 2497c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2498c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 249966c3a757STejun Heo { 2500c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 2501c8826dd5STejun Heo void **areas = NULL; 2502fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 250393c76b6bSzijun_hu size_t size_sum, areas_size; 250493c76b6bSzijun_hu unsigned long max_distance; 25059b739662Szijun_hu int group, i, highest_group, rc; 250666c3a757STejun Heo 2507c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 2508c8826dd5STejun Heo cpu_distance_fn); 2509fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2510fd1e8a1fSTejun Heo return PTR_ERR(ai); 251166c3a757STejun Heo 2512fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2513c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 251466c3a757STejun Heo 251526fb3daeSMike Rapoport areas = memblock_alloc(areas_size, SMP_CACHE_BYTES); 2516c8826dd5STejun Heo if (!areas) { 2517fb435d52STejun Heo rc = -ENOMEM; 2518c8826dd5STejun Heo goto out_free; 2519fa8a7094STejun Heo } 252066c3a757STejun Heo 25219b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 25229b739662Szijun_hu highest_group = 0; 2523c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2524c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2525c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 2526c8826dd5STejun Heo void *ptr; 252766c3a757STejun Heo 2528c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 2529c8826dd5STejun Heo cpu = gi->cpu_map[i]; 2530c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 2531c8826dd5STejun Heo 2532c8826dd5STejun Heo /* allocate space for the whole group */ 2533c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 2534c8826dd5STejun Heo if (!ptr) { 2535c8826dd5STejun Heo rc = -ENOMEM; 2536c8826dd5STejun Heo goto out_free_areas; 2537c8826dd5STejun Heo } 2538f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2539f528f0b8SCatalin Marinas kmemleak_free(ptr); 2540c8826dd5STejun Heo areas[group] = ptr; 2541c8826dd5STejun Heo 2542c8826dd5STejun Heo base = min(ptr, base); 25439b739662Szijun_hu if (ptr > areas[highest_group]) 25449b739662Szijun_hu highest_group = group; 25459b739662Szijun_hu } 25469b739662Szijun_hu max_distance = areas[highest_group] - base; 25479b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 25489b739662Szijun_hu 25499b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 25509b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 25519b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 25529b739662Szijun_hu max_distance, VMALLOC_TOTAL); 25539b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 25549b739662Szijun_hu /* and fail if we have fallback */ 25559b739662Szijun_hu rc = -EINVAL; 25569b739662Szijun_hu goto out_free_areas; 25579b739662Szijun_hu #endif 255842b64281STejun Heo } 255942b64281STejun Heo 256042b64281STejun Heo /* 256142b64281STejun Heo * Copy data and free unused parts. This should happen after all 256242b64281STejun Heo * allocations are complete; otherwise, we may end up with 256342b64281STejun Heo * overlapping groups. 256442b64281STejun Heo */ 256542b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 256642b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 256742b64281STejun Heo void *ptr = areas[group]; 2568c8826dd5STejun Heo 2569c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2570c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2571c8826dd5STejun Heo /* unused unit, free whole */ 2572c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2573c8826dd5STejun Heo continue; 2574c8826dd5STejun Heo } 2575c8826dd5STejun Heo /* copy and return the unused part */ 2576fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2577c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2578c8826dd5STejun Heo } 257966c3a757STejun Heo } 258066c3a757STejun Heo 2581c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 25826ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2583c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 25846ea529a2STejun Heo } 2585c8826dd5STejun Heo 2586870d4b12SJoe Perches pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 2587fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 2588fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 258966c3a757STejun Heo 2590fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 2591c8826dd5STejun Heo goto out_free; 2592c8826dd5STejun Heo 2593c8826dd5STejun Heo out_free_areas: 2594c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2595f851c8d8SMichael Holzheu if (areas[group]) 2596c8826dd5STejun Heo free_fn(areas[group], 2597c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2598c8826dd5STejun Heo out_free: 2599fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2600c8826dd5STejun Heo if (areas) 2601999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2602fb435d52STejun Heo return rc; 2603d4b95f80STejun Heo } 26043c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2605d4b95f80STejun Heo 26063c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2607d4b95f80STejun Heo /** 260800ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2609d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2610d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 261125985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2612d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2613d4b95f80STejun Heo * 261400ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 261500ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2616d4b95f80STejun Heo * 2617d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2618d4b95f80STejun Heo * page-by-page into vmalloc area. 2619d4b95f80STejun Heo * 2620d4b95f80STejun Heo * RETURNS: 2621fb435d52STejun Heo * 0 on success, -errno on failure. 2622d4b95f80STejun Heo */ 2623fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2624d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2625d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2626d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2627d4b95f80STejun Heo { 26288f05a6a6STejun Heo static struct vm_struct vm; 2629fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 263000ae4064STejun Heo char psize_str[16]; 2631ce3141a2STejun Heo int unit_pages; 2632d4b95f80STejun Heo size_t pages_size; 2633ce3141a2STejun Heo struct page **pages; 2634fb435d52STejun Heo int unit, i, j, rc; 26358f606604Szijun_hu int upa; 26368f606604Szijun_hu int nr_g0_units; 2637d4b95f80STejun Heo 263800ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 263900ae4064STejun Heo 26404ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2641fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2642fd1e8a1fSTejun Heo return PTR_ERR(ai); 2643fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 26448f606604Szijun_hu upa = ai->alloc_size/ai->unit_size; 26458f606604Szijun_hu nr_g0_units = roundup(num_possible_cpus(), upa); 26460b59c25fSIgor Stoppa if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) { 26478f606604Szijun_hu pcpu_free_alloc_info(ai); 26488f606604Szijun_hu return -EINVAL; 26498f606604Szijun_hu } 2650fd1e8a1fSTejun Heo 2651fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2652d4b95f80STejun Heo 2653d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2654fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2655fd1e8a1fSTejun Heo sizeof(pages[0])); 26567e1c4e27SMike Rapoport pages = memblock_alloc(pages_size, SMP_CACHE_BYTES); 2657f655f405SMike Rapoport if (!pages) 2658f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2659f655f405SMike Rapoport pages_size); 2660d4b95f80STejun Heo 26618f05a6a6STejun Heo /* allocate pages */ 2662d4b95f80STejun Heo j = 0; 26638f606604Szijun_hu for (unit = 0; unit < num_possible_cpus(); unit++) { 2664fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 26658f606604Szijun_hu for (i = 0; i < unit_pages; i++) { 2666d4b95f80STejun Heo void *ptr; 2667d4b95f80STejun Heo 26683cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2669d4b95f80STejun Heo if (!ptr) { 2670870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2671598d8091SJoe Perches psize_str, cpu); 2672d4b95f80STejun Heo goto enomem; 2673d4b95f80STejun Heo } 2674f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2675f528f0b8SCatalin Marinas kmemleak_free(ptr); 2676ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2677d4b95f80STejun Heo } 26788f606604Szijun_hu } 2679d4b95f80STejun Heo 26808f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 26818f05a6a6STejun Heo vm.flags = VM_ALLOC; 2682fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 26838f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 26848f05a6a6STejun Heo 2685fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 26861d9d3257STejun Heo unsigned long unit_addr = 2687fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 26888f05a6a6STejun Heo 2689ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 26908f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 26918f05a6a6STejun Heo 26928f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2693fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2694ce3141a2STejun Heo unit_pages); 2695fb435d52STejun Heo if (rc < 0) 2696fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 26978f05a6a6STejun Heo 26988f05a6a6STejun Heo /* 26998f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 27008f05a6a6STejun Heo * cache for the linear mapping here - something 27018f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 27028f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 27038f05a6a6STejun Heo * data structures are not set up yet. 27048f05a6a6STejun Heo */ 27058f05a6a6STejun Heo 27068f05a6a6STejun Heo /* copy static data */ 2707fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 270866c3a757STejun Heo } 270966c3a757STejun Heo 271066c3a757STejun Heo /* we're ready, commit */ 2711870d4b12SJoe Perches pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", 2712fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 2713fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 271466c3a757STejun Heo 2715fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 2716d4b95f80STejun Heo goto out_free_ar; 2717d4b95f80STejun Heo 2718d4b95f80STejun Heo enomem: 2719d4b95f80STejun Heo while (--j >= 0) 2720ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2721fb435d52STejun Heo rc = -ENOMEM; 2722d4b95f80STejun Heo out_free_ar: 2723999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2724fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2725fb435d52STejun Heo return rc; 272666c3a757STejun Heo } 27273c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2728d4b95f80STejun Heo 2729bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 27308c4bfc6eSTejun Heo /* 2731bbddff05STejun Heo * Generic SMP percpu area setup. 2732e74e3962STejun Heo * 2733e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2734e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2735e74e3962STejun Heo * important because many archs have addressing restrictions and might 2736e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2737e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2738e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2739e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2740e74e3962STejun Heo * mappings on applicable archs. 2741e74e3962STejun Heo */ 2742e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2743e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2744e74e3962STejun Heo 2745c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2746c8826dd5STejun Heo size_t align) 2747c8826dd5STejun Heo { 274826fb3daeSMike Rapoport return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS)); 2749c8826dd5STejun Heo } 2750c8826dd5STejun Heo 2751c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2752c8826dd5STejun Heo { 2753999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2754c8826dd5STejun Heo } 2755c8826dd5STejun Heo 2756e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2757e74e3962STejun Heo { 2758e74e3962STejun Heo unsigned long delta; 2759e74e3962STejun Heo unsigned int cpu; 2760fb435d52STejun Heo int rc; 2761e74e3962STejun Heo 2762e74e3962STejun Heo /* 2763e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2764e74e3962STejun Heo * what the legacy allocator did. 2765e74e3962STejun Heo */ 2766fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2767c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2768c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2769fb435d52STejun Heo if (rc < 0) 2770bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2771e74e3962STejun Heo 2772e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2773e74e3962STejun Heo for_each_possible_cpu(cpu) 2774fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2775e74e3962STejun Heo } 2776e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2777099a19d9STejun Heo 2778bbddff05STejun Heo #else /* CONFIG_SMP */ 2779bbddff05STejun Heo 2780bbddff05STejun Heo /* 2781bbddff05STejun Heo * UP percpu area setup. 2782bbddff05STejun Heo * 2783bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2784bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2785bbddff05STejun Heo * variables and don't require any special preparation. 2786bbddff05STejun Heo */ 2787bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2788bbddff05STejun Heo { 2789bbddff05STejun Heo const size_t unit_size = 2790bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2791bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2792bbddff05STejun Heo struct pcpu_alloc_info *ai; 2793bbddff05STejun Heo void *fc; 2794bbddff05STejun Heo 2795bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 279626fb3daeSMike Rapoport fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); 2797bbddff05STejun Heo if (!ai || !fc) 2798bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2799100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2800100d13c3SCatalin Marinas kmemleak_free(fc); 2801bbddff05STejun Heo 2802bbddff05STejun Heo ai->dyn_size = unit_size; 2803bbddff05STejun Heo ai->unit_size = unit_size; 2804bbddff05STejun Heo ai->atom_size = unit_size; 2805bbddff05STejun Heo ai->alloc_size = unit_size; 2806bbddff05STejun Heo ai->groups[0].nr_units = 1; 2807bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2808bbddff05STejun Heo 2809bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 2810bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2811438a5061SNicolas Pitre pcpu_free_alloc_info(ai); 2812bbddff05STejun Heo } 2813bbddff05STejun Heo 2814bbddff05STejun Heo #endif /* CONFIG_SMP */ 2815bbddff05STejun Heo 2816099a19d9STejun Heo /* 28177e8a6304SDennis Zhou (Facebook) * pcpu_nr_pages - calculate total number of populated backing pages 28187e8a6304SDennis Zhou (Facebook) * 28197e8a6304SDennis Zhou (Facebook) * This reflects the number of pages populated to back chunks. Metadata is 28207e8a6304SDennis Zhou (Facebook) * excluded in the number exposed in meminfo as the number of backing pages 28217e8a6304SDennis Zhou (Facebook) * scales with the number of cpus and can quickly outweigh the memory used for 28227e8a6304SDennis Zhou (Facebook) * metadata. It also keeps this calculation nice and simple. 28237e8a6304SDennis Zhou (Facebook) * 28247e8a6304SDennis Zhou (Facebook) * RETURNS: 28257e8a6304SDennis Zhou (Facebook) * Total number of populated backing pages in use by the allocator. 28267e8a6304SDennis Zhou (Facebook) */ 28277e8a6304SDennis Zhou (Facebook) unsigned long pcpu_nr_pages(void) 28287e8a6304SDennis Zhou (Facebook) { 28297e8a6304SDennis Zhou (Facebook) return pcpu_nr_populated * pcpu_nr_units; 28307e8a6304SDennis Zhou (Facebook) } 28317e8a6304SDennis Zhou (Facebook) 28327e8a6304SDennis Zhou (Facebook) /* 28331a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 28341a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 28351a4d7607STejun Heo * and running. 28361a4d7607STejun Heo */ 28371a4d7607STejun Heo static int __init percpu_enable_async(void) 28381a4d7607STejun Heo { 28391a4d7607STejun Heo pcpu_async_enabled = true; 28401a4d7607STejun Heo return 0; 28411a4d7607STejun Heo } 28421a4d7607STejun Heo subsys_initcall(percpu_enable_async); 2843