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> 68fbf59bc9STejun Heo #include <linux/bootmem.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> 83fbf59bc9STejun Heo 84fbf59bc9STejun Heo #include <asm/cacheflush.h> 85e0100983STejun Heo #include <asm/sections.h> 86fbf59bc9STejun Heo #include <asm/tlbflush.h> 873b034b0dSVivek Goyal #include <asm/io.h> 88fbf59bc9STejun Heo 89df95e795SDennis Zhou #define CREATE_TRACE_POINTS 90df95e795SDennis Zhou #include <trace/events/percpu.h> 91df95e795SDennis Zhou 928fa3ed80SDennis Zhou #include "percpu-internal.h" 938fa3ed80SDennis Zhou 9440064aecSDennis Zhou (Facebook) /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ 9540064aecSDennis Zhou (Facebook) #define PCPU_SLOT_BASE_SHIFT 5 9640064aecSDennis Zhou (Facebook) 971a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_LOW 2 981a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_HIGH 4 99fbf59bc9STejun Heo 100bbddff05STejun Heo #ifdef CONFIG_SMP 101e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 102e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 103e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 10443cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 10543cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 10643cf38ebSTejun Heo (unsigned long)__per_cpu_start) 107e0100983STejun Heo #endif 108e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 109e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 11043cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 11143cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 11243cf38ebSTejun Heo (unsigned long)__per_cpu_start) 113e0100983STejun Heo #endif 114bbddff05STejun Heo #else /* CONFIG_SMP */ 115bbddff05STejun Heo /* on UP, it's always identity mapped */ 116bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 117bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 118bbddff05STejun Heo #endif /* CONFIG_SMP */ 119e0100983STejun Heo 1201328710bSDaniel Micay static int pcpu_unit_pages __ro_after_init; 1211328710bSDaniel Micay static int pcpu_unit_size __ro_after_init; 1221328710bSDaniel Micay static int pcpu_nr_units __ro_after_init; 1231328710bSDaniel Micay static int pcpu_atom_size __ro_after_init; 1248fa3ed80SDennis Zhou int pcpu_nr_slots __ro_after_init; 1251328710bSDaniel Micay static size_t pcpu_chunk_struct_size __ro_after_init; 126fbf59bc9STejun Heo 127a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 1281328710bSDaniel Micay static unsigned int pcpu_low_unit_cpu __ro_after_init; 1291328710bSDaniel Micay static unsigned int pcpu_high_unit_cpu __ro_after_init; 1302f39e637STejun Heo 131fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 1321328710bSDaniel Micay void *pcpu_base_addr __ro_after_init; 133fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 134fbf59bc9STejun Heo 1351328710bSDaniel Micay static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ 1361328710bSDaniel Micay const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ 1372f39e637STejun Heo 1386563297cSTejun Heo /* group information, used for vm allocation */ 1391328710bSDaniel Micay static int pcpu_nr_groups __ro_after_init; 1401328710bSDaniel Micay static const unsigned long *pcpu_group_offsets __ro_after_init; 1411328710bSDaniel Micay static const size_t *pcpu_group_sizes __ro_after_init; 1426563297cSTejun Heo 143ae9e6bc9STejun Heo /* 144ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 145ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 146ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 147ae9e6bc9STejun Heo */ 1488fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_first_chunk __ro_after_init; 149ae9e6bc9STejun Heo 150ae9e6bc9STejun Heo /* 151ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 152e2266705SDennis Zhou (Facebook) * chunk and serves it for reserved allocations. When the reserved 153e2266705SDennis Zhou (Facebook) * region doesn't exist, the following variable is NULL. 154ae9e6bc9STejun Heo */ 1558fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; 156edcb4639STejun Heo 1578fa3ed80SDennis Zhou DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 1586710e594STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 159fbf59bc9STejun Heo 1608fa3ed80SDennis Zhou struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ 161fbf59bc9STejun Heo 1624f996e23STejun Heo /* chunks which need their map areas extended, protected by pcpu_lock */ 1634f996e23STejun Heo static LIST_HEAD(pcpu_map_extend_chunks); 1644f996e23STejun Heo 165b539b87fSTejun Heo /* 166b539b87fSTejun Heo * The number of empty populated pages, protected by pcpu_lock. The 167b539b87fSTejun Heo * reserved chunk doesn't contribute to the count. 168b539b87fSTejun Heo */ 1696b9b6f39SDennis Zhou (Facebook) int pcpu_nr_empty_pop_pages; 170b539b87fSTejun Heo 1711a4d7607STejun Heo /* 1721a4d7607STejun Heo * Balance work is used to populate or destroy chunks asynchronously. We 1731a4d7607STejun Heo * try to keep the number of populated free pages between 1741a4d7607STejun Heo * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 1751a4d7607STejun Heo * empty chunk. 1761a4d7607STejun Heo */ 177fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work); 178fe6bd8c3STejun Heo static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 1791a4d7607STejun Heo static bool pcpu_async_enabled __read_mostly; 1801a4d7607STejun Heo static bool pcpu_atomic_alloc_failed; 1811a4d7607STejun Heo 1821a4d7607STejun Heo static void pcpu_schedule_balance_work(void) 1831a4d7607STejun Heo { 1841a4d7607STejun Heo if (pcpu_async_enabled) 1851a4d7607STejun Heo schedule_work(&pcpu_balance_work); 1861a4d7607STejun Heo } 187a56dbddfSTejun Heo 188c0ebfdc3SDennis Zhou (Facebook) /** 189560f2c23SDennis Zhou (Facebook) * pcpu_addr_in_chunk - check if the address is served from this chunk 190560f2c23SDennis Zhou (Facebook) * @chunk: chunk of interest 191560f2c23SDennis Zhou (Facebook) * @addr: percpu address 192c0ebfdc3SDennis Zhou (Facebook) * 193c0ebfdc3SDennis Zhou (Facebook) * RETURNS: 194560f2c23SDennis Zhou (Facebook) * True if the address is served from this chunk. 195c0ebfdc3SDennis Zhou (Facebook) */ 196560f2c23SDennis Zhou (Facebook) static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) 197020ec653STejun Heo { 198c0ebfdc3SDennis Zhou (Facebook) void *start_addr, *end_addr; 199020ec653STejun Heo 200560f2c23SDennis Zhou (Facebook) if (!chunk) 201c0ebfdc3SDennis Zhou (Facebook) return false; 202c0ebfdc3SDennis Zhou (Facebook) 203560f2c23SDennis Zhou (Facebook) start_addr = chunk->base_addr + chunk->start_offset; 204560f2c23SDennis Zhou (Facebook) end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - 205560f2c23SDennis Zhou (Facebook) chunk->end_offset; 206c0ebfdc3SDennis Zhou (Facebook) 207c0ebfdc3SDennis Zhou (Facebook) return addr >= start_addr && addr < end_addr; 208020ec653STejun Heo } 209020ec653STejun Heo 210d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 211fbf59bc9STejun Heo { 212cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 213fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 214fbf59bc9STejun Heo } 215fbf59bc9STejun Heo 216d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 217d9b55eebSTejun Heo { 218d9b55eebSTejun Heo if (size == pcpu_unit_size) 219d9b55eebSTejun Heo return pcpu_nr_slots - 1; 220d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 221d9b55eebSTejun Heo } 222d9b55eebSTejun Heo 223fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 224fbf59bc9STejun Heo { 22540064aecSDennis Zhou (Facebook) if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) 226fbf59bc9STejun Heo return 0; 227fbf59bc9STejun Heo 22840064aecSDennis Zhou (Facebook) return pcpu_size_to_slot(chunk->free_bytes); 229fbf59bc9STejun Heo } 230fbf59bc9STejun Heo 23188999a89STejun Heo /* set the pointer to a chunk in a page struct */ 23288999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 23388999a89STejun Heo { 23488999a89STejun Heo page->index = (unsigned long)pcpu; 23588999a89STejun Heo } 23688999a89STejun Heo 23788999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 23888999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 23988999a89STejun Heo { 24088999a89STejun Heo return (struct pcpu_chunk *)page->index; 24188999a89STejun Heo } 24288999a89STejun Heo 24388999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 244fbf59bc9STejun Heo { 2452f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 246fbf59bc9STejun Heo } 247fbf59bc9STejun Heo 248c0ebfdc3SDennis Zhou (Facebook) static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 249c0ebfdc3SDennis Zhou (Facebook) { 250c0ebfdc3SDennis Zhou (Facebook) return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 251c0ebfdc3SDennis Zhou (Facebook) } 252c0ebfdc3SDennis Zhou (Facebook) 2539983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 254fbf59bc9STejun Heo unsigned int cpu, int page_idx) 255fbf59bc9STejun Heo { 256c0ebfdc3SDennis Zhou (Facebook) return (unsigned long)chunk->base_addr + 257c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(cpu, page_idx); 258fbf59bc9STejun Heo } 259fbf59bc9STejun Heo 26091e914c5SDennis Zhou (Facebook) static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) 261ce3141a2STejun Heo { 26291e914c5SDennis Zhou (Facebook) *rs = find_next_zero_bit(bitmap, end, *rs); 26391e914c5SDennis Zhou (Facebook) *re = find_next_bit(bitmap, end, *rs + 1); 264ce3141a2STejun Heo } 265ce3141a2STejun Heo 26691e914c5SDennis Zhou (Facebook) static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) 267ce3141a2STejun Heo { 26891e914c5SDennis Zhou (Facebook) *rs = find_next_bit(bitmap, end, *rs); 26991e914c5SDennis Zhou (Facebook) *re = find_next_zero_bit(bitmap, end, *rs + 1); 270ce3141a2STejun Heo } 271ce3141a2STejun Heo 272ce3141a2STejun Heo /* 27391e914c5SDennis Zhou (Facebook) * Bitmap region iterators. Iterates over the bitmap between 27491e914c5SDennis Zhou (Facebook) * [@start, @end) in @chunk. @rs and @re should be integer variables 27591e914c5SDennis Zhou (Facebook) * and will be set to start and end index of the current free region. 276ce3141a2STejun Heo */ 27791e914c5SDennis Zhou (Facebook) #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ 27891e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ 279ce3141a2STejun Heo (rs) < (re); \ 28091e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) 281ce3141a2STejun Heo 28291e914c5SDennis Zhou (Facebook) #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ 28391e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ 284ce3141a2STejun Heo (rs) < (re); \ 28591e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) 286ce3141a2STejun Heo 287ca460b3cSDennis Zhou (Facebook) /* 288ca460b3cSDennis Zhou (Facebook) * The following are helper functions to help access bitmaps and convert 289ca460b3cSDennis Zhou (Facebook) * between bitmap offsets to address offsets. 290ca460b3cSDennis Zhou (Facebook) */ 291ca460b3cSDennis Zhou (Facebook) static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 292ca460b3cSDennis Zhou (Facebook) { 293ca460b3cSDennis Zhou (Facebook) return chunk->alloc_map + 294ca460b3cSDennis Zhou (Facebook) (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 295ca460b3cSDennis Zhou (Facebook) } 296ca460b3cSDennis Zhou (Facebook) 297ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_index(int off) 298ca460b3cSDennis Zhou (Facebook) { 299ca460b3cSDennis Zhou (Facebook) return off / PCPU_BITMAP_BLOCK_BITS; 300ca460b3cSDennis Zhou (Facebook) } 301ca460b3cSDennis Zhou (Facebook) 302ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_off(int off) 303ca460b3cSDennis Zhou (Facebook) { 304ca460b3cSDennis Zhou (Facebook) return off & (PCPU_BITMAP_BLOCK_BITS - 1); 305ca460b3cSDennis Zhou (Facebook) } 306ca460b3cSDennis Zhou (Facebook) 307b185cd0dSDennis Zhou (Facebook) static unsigned long pcpu_block_off_to_off(int index, int off) 308b185cd0dSDennis Zhou (Facebook) { 309b185cd0dSDennis Zhou (Facebook) return index * PCPU_BITMAP_BLOCK_BITS + off; 310b185cd0dSDennis Zhou (Facebook) } 311b185cd0dSDennis Zhou (Facebook) 312fbf59bc9STejun Heo /** 313525ca84dSDennis Zhou (Facebook) * pcpu_next_md_free_region - finds the next hint free area 314525ca84dSDennis Zhou (Facebook) * @chunk: chunk of interest 315525ca84dSDennis Zhou (Facebook) * @bit_off: chunk offset 316525ca84dSDennis Zhou (Facebook) * @bits: size of free area 317525ca84dSDennis Zhou (Facebook) * 318525ca84dSDennis Zhou (Facebook) * Helper function for pcpu_for_each_md_free_region. It checks 319525ca84dSDennis Zhou (Facebook) * block->contig_hint and performs aggregation across blocks to find the 320525ca84dSDennis Zhou (Facebook) * next hint. It modifies bit_off and bits in-place to be consumed in the 321525ca84dSDennis Zhou (Facebook) * loop. 322525ca84dSDennis Zhou (Facebook) */ 323525ca84dSDennis Zhou (Facebook) static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, 324525ca84dSDennis Zhou (Facebook) int *bits) 325525ca84dSDennis Zhou (Facebook) { 326525ca84dSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 327525ca84dSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 328525ca84dSDennis Zhou (Facebook) struct pcpu_block_md *block; 329525ca84dSDennis Zhou (Facebook) 330525ca84dSDennis Zhou (Facebook) *bits = 0; 331525ca84dSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 332525ca84dSDennis Zhou (Facebook) block++, i++) { 333525ca84dSDennis Zhou (Facebook) /* handles contig area across blocks */ 334525ca84dSDennis Zhou (Facebook) if (*bits) { 335525ca84dSDennis Zhou (Facebook) *bits += block->left_free; 336525ca84dSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 337525ca84dSDennis Zhou (Facebook) continue; 338525ca84dSDennis Zhou (Facebook) return; 339525ca84dSDennis Zhou (Facebook) } 340525ca84dSDennis Zhou (Facebook) 341525ca84dSDennis Zhou (Facebook) /* 342525ca84dSDennis Zhou (Facebook) * This checks three things. First is there a contig_hint to 343525ca84dSDennis Zhou (Facebook) * check. Second, have we checked this hint before by 344525ca84dSDennis Zhou (Facebook) * comparing the block_off. Third, is this the same as the 345525ca84dSDennis Zhou (Facebook) * right contig hint. In the last case, it spills over into 346525ca84dSDennis Zhou (Facebook) * the next block and should be handled by the contig area 347525ca84dSDennis Zhou (Facebook) * across blocks code. 348525ca84dSDennis Zhou (Facebook) */ 349525ca84dSDennis Zhou (Facebook) *bits = block->contig_hint; 350525ca84dSDennis Zhou (Facebook) if (*bits && block->contig_hint_start >= block_off && 351525ca84dSDennis Zhou (Facebook) *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { 352525ca84dSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, 353525ca84dSDennis Zhou (Facebook) block->contig_hint_start); 354525ca84dSDennis Zhou (Facebook) return; 355525ca84dSDennis Zhou (Facebook) } 3561fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 3571fa4df3eSDennis Zhou block_off = 0; 358525ca84dSDennis Zhou (Facebook) 359525ca84dSDennis Zhou (Facebook) *bits = block->right_free; 360525ca84dSDennis Zhou (Facebook) *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; 361525ca84dSDennis Zhou (Facebook) } 362525ca84dSDennis Zhou (Facebook) } 363525ca84dSDennis Zhou (Facebook) 364b4c2116cSDennis Zhou (Facebook) /** 365b4c2116cSDennis Zhou (Facebook) * pcpu_next_fit_region - finds fit areas for a given allocation request 366b4c2116cSDennis Zhou (Facebook) * @chunk: chunk of interest 367b4c2116cSDennis Zhou (Facebook) * @alloc_bits: size of allocation 368b4c2116cSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 369b4c2116cSDennis Zhou (Facebook) * @bit_off: chunk offset 370b4c2116cSDennis Zhou (Facebook) * @bits: size of free area 371b4c2116cSDennis Zhou (Facebook) * 372b4c2116cSDennis Zhou (Facebook) * Finds the next free region that is viable for use with a given size and 373b4c2116cSDennis Zhou (Facebook) * alignment. This only returns if there is a valid area to be used for this 374b4c2116cSDennis Zhou (Facebook) * allocation. block->first_free is returned if the allocation request fits 375b4c2116cSDennis Zhou (Facebook) * within the block to see if the request can be fulfilled prior to the contig 376b4c2116cSDennis Zhou (Facebook) * hint. 377b4c2116cSDennis Zhou (Facebook) */ 378b4c2116cSDennis Zhou (Facebook) static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, 379b4c2116cSDennis Zhou (Facebook) int align, int *bit_off, int *bits) 380b4c2116cSDennis Zhou (Facebook) { 381b4c2116cSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 382b4c2116cSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 383b4c2116cSDennis Zhou (Facebook) struct pcpu_block_md *block; 384b4c2116cSDennis Zhou (Facebook) 385b4c2116cSDennis Zhou (Facebook) *bits = 0; 386b4c2116cSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 387b4c2116cSDennis Zhou (Facebook) block++, i++) { 388b4c2116cSDennis Zhou (Facebook) /* handles contig area across blocks */ 389b4c2116cSDennis Zhou (Facebook) if (*bits) { 390b4c2116cSDennis Zhou (Facebook) *bits += block->left_free; 391b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 392b4c2116cSDennis Zhou (Facebook) return; 393b4c2116cSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 394b4c2116cSDennis Zhou (Facebook) continue; 395b4c2116cSDennis Zhou (Facebook) } 396b4c2116cSDennis Zhou (Facebook) 397b4c2116cSDennis Zhou (Facebook) /* check block->contig_hint */ 398b4c2116cSDennis Zhou (Facebook) *bits = ALIGN(block->contig_hint_start, align) - 399b4c2116cSDennis Zhou (Facebook) block->contig_hint_start; 400b4c2116cSDennis Zhou (Facebook) /* 401b4c2116cSDennis Zhou (Facebook) * This uses the block offset to determine if this has been 402b4c2116cSDennis Zhou (Facebook) * checked in the prior iteration. 403b4c2116cSDennis Zhou (Facebook) */ 404b4c2116cSDennis Zhou (Facebook) if (block->contig_hint && 405b4c2116cSDennis Zhou (Facebook) block->contig_hint_start >= block_off && 406b4c2116cSDennis Zhou (Facebook) block->contig_hint >= *bits + alloc_bits) { 407b4c2116cSDennis Zhou (Facebook) *bits += alloc_bits + block->contig_hint_start - 408b4c2116cSDennis Zhou (Facebook) block->first_free; 409b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, block->first_free); 410b4c2116cSDennis Zhou (Facebook) return; 411b4c2116cSDennis Zhou (Facebook) } 4121fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 4131fa4df3eSDennis Zhou block_off = 0; 414b4c2116cSDennis Zhou (Facebook) 415b4c2116cSDennis Zhou (Facebook) *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, 416b4c2116cSDennis Zhou (Facebook) align); 417b4c2116cSDennis Zhou (Facebook) *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; 418b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, *bit_off); 419b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 420b4c2116cSDennis Zhou (Facebook) return; 421b4c2116cSDennis Zhou (Facebook) } 422b4c2116cSDennis Zhou (Facebook) 423b4c2116cSDennis Zhou (Facebook) /* no valid offsets were found - fail condition */ 424b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_chunk_map_bits(chunk); 425b4c2116cSDennis Zhou (Facebook) } 426b4c2116cSDennis Zhou (Facebook) 427525ca84dSDennis Zhou (Facebook) /* 428525ca84dSDennis Zhou (Facebook) * Metadata free area iterators. These perform aggregation of free areas 429525ca84dSDennis Zhou (Facebook) * based on the metadata blocks and return the offset @bit_off and size in 430b4c2116cSDennis Zhou (Facebook) * bits of the free area @bits. pcpu_for_each_fit_region only returns when 431b4c2116cSDennis Zhou (Facebook) * a fit is found for the allocation request. 432525ca84dSDennis Zhou (Facebook) */ 433525ca84dSDennis Zhou (Facebook) #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ 434525ca84dSDennis Zhou (Facebook) for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ 435525ca84dSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 436525ca84dSDennis Zhou (Facebook) (bit_off) += (bits) + 1, \ 437525ca84dSDennis Zhou (Facebook) pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) 438525ca84dSDennis Zhou (Facebook) 439b4c2116cSDennis Zhou (Facebook) #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ 440b4c2116cSDennis Zhou (Facebook) for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 441b4c2116cSDennis Zhou (Facebook) &(bits)); \ 442b4c2116cSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 443b4c2116cSDennis Zhou (Facebook) (bit_off) += (bits), \ 444b4c2116cSDennis Zhou (Facebook) pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 445b4c2116cSDennis Zhou (Facebook) &(bits))) 446b4c2116cSDennis Zhou (Facebook) 447525ca84dSDennis Zhou (Facebook) /** 44890459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 4491880d93bSTejun Heo * @size: bytes to allocate 45047504ee0SDennis Zhou * @gfp: allocation flags 451fbf59bc9STejun Heo * 4521880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 45347504ee0SDennis Zhou * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. 45447504ee0SDennis Zhou * This is to facilitate passing through whitelisted flags. The 45547504ee0SDennis Zhou * returned memory is always zeroed. 456fbf59bc9STejun Heo * 457fbf59bc9STejun Heo * RETURNS: 4581880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 459fbf59bc9STejun Heo */ 46047504ee0SDennis Zhou static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) 461fbf59bc9STejun Heo { 462099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 463099a19d9STejun Heo return NULL; 464099a19d9STejun Heo 465fbf59bc9STejun Heo if (size <= PAGE_SIZE) 466*554fef1cSDennis Zhou return kzalloc(size, gfp); 4677af4c093SJesper Juhl else 468*554fef1cSDennis Zhou return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL); 4691880d93bSTejun Heo } 470fbf59bc9STejun Heo 4711880d93bSTejun Heo /** 4721880d93bSTejun Heo * pcpu_mem_free - free memory 4731880d93bSTejun Heo * @ptr: memory to free 4741880d93bSTejun Heo * 47590459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 4761880d93bSTejun Heo */ 4771d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 4781880d93bSTejun Heo { 4791d5cfdb0STetsuo Handa kvfree(ptr); 480fbf59bc9STejun Heo } 481fbf59bc9STejun Heo 482fbf59bc9STejun Heo /** 483fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 484fbf59bc9STejun Heo * @chunk: chunk of interest 485fbf59bc9STejun Heo * @oslot: the previous slot it was on 486fbf59bc9STejun Heo * 487fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 488fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 489edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 490edcb4639STejun Heo * chunk slots. 491ccea34b5STejun Heo * 492ccea34b5STejun Heo * CONTEXT: 493ccea34b5STejun Heo * pcpu_lock. 494fbf59bc9STejun Heo */ 495fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 496fbf59bc9STejun Heo { 497fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 498fbf59bc9STejun Heo 499edcb4639STejun Heo if (chunk != pcpu_reserved_chunk && oslot != nslot) { 500fbf59bc9STejun Heo if (oslot < nslot) 501fbf59bc9STejun Heo list_move(&chunk->list, &pcpu_slot[nslot]); 502fbf59bc9STejun Heo else 503fbf59bc9STejun Heo list_move_tail(&chunk->list, &pcpu_slot[nslot]); 504fbf59bc9STejun Heo } 505fbf59bc9STejun Heo } 506fbf59bc9STejun Heo 507fbf59bc9STejun Heo /** 50840064aecSDennis Zhou (Facebook) * pcpu_cnt_pop_pages- counts populated backing pages in range 509833af842STejun Heo * @chunk: chunk of interest 51040064aecSDennis Zhou (Facebook) * @bit_off: start offset 51140064aecSDennis Zhou (Facebook) * @bits: size of area to check 5129f7dcf22STejun Heo * 51340064aecSDennis Zhou (Facebook) * Calculates the number of populated pages in the region 51440064aecSDennis Zhou (Facebook) * [page_start, page_end). This keeps track of how many empty populated 51540064aecSDennis Zhou (Facebook) * pages are available and decide if async work should be scheduled. 516ccea34b5STejun Heo * 5179f7dcf22STejun Heo * RETURNS: 51840064aecSDennis Zhou (Facebook) * The nr of populated pages. 5199f7dcf22STejun Heo */ 52040064aecSDennis Zhou (Facebook) static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off, 52140064aecSDennis Zhou (Facebook) int bits) 5229f7dcf22STejun Heo { 52340064aecSDennis Zhou (Facebook) int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE); 52440064aecSDennis Zhou (Facebook) int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 52540064aecSDennis Zhou (Facebook) 52640064aecSDennis Zhou (Facebook) if (page_start >= page_end) 52740064aecSDennis Zhou (Facebook) return 0; 52840064aecSDennis Zhou (Facebook) 52940064aecSDennis Zhou (Facebook) /* 53040064aecSDennis Zhou (Facebook) * bitmap_weight counts the number of bits set in a bitmap up to 53140064aecSDennis Zhou (Facebook) * the specified number of bits. This is counting the populated 53240064aecSDennis Zhou (Facebook) * pages up to page_end and then subtracting the populated pages 53340064aecSDennis Zhou (Facebook) * up to page_start to count the populated pages in 53440064aecSDennis Zhou (Facebook) * [page_start, page_end). 53540064aecSDennis Zhou (Facebook) */ 53640064aecSDennis Zhou (Facebook) return bitmap_weight(chunk->populated, page_end) - 53740064aecSDennis Zhou (Facebook) bitmap_weight(chunk->populated, page_start); 53840064aecSDennis Zhou (Facebook) } 53940064aecSDennis Zhou (Facebook) 54040064aecSDennis Zhou (Facebook) /** 54140064aecSDennis Zhou (Facebook) * pcpu_chunk_update - updates the chunk metadata given a free area 54240064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 54340064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 54440064aecSDennis Zhou (Facebook) * @bits: size of free area 54540064aecSDennis Zhou (Facebook) * 54613f96637SDennis Zhou (Facebook) * This updates the chunk's contig hint and starting offset given a free area. 547268625a6SDennis Zhou (Facebook) * Choose the best starting offset if the contig hint is equal. 54840064aecSDennis Zhou (Facebook) */ 54940064aecSDennis Zhou (Facebook) static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) 55040064aecSDennis Zhou (Facebook) { 55113f96637SDennis Zhou (Facebook) if (bits > chunk->contig_bits) { 55213f96637SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 55340064aecSDennis Zhou (Facebook) chunk->contig_bits = bits; 554268625a6SDennis Zhou (Facebook) } else if (bits == chunk->contig_bits && chunk->contig_bits_start && 555268625a6SDennis Zhou (Facebook) (!bit_off || 556268625a6SDennis Zhou (Facebook) __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { 557268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 558268625a6SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 55940064aecSDennis Zhou (Facebook) } 56013f96637SDennis Zhou (Facebook) } 56140064aecSDennis Zhou (Facebook) 56240064aecSDennis Zhou (Facebook) /** 56340064aecSDennis Zhou (Facebook) * pcpu_chunk_refresh_hint - updates metadata about a chunk 56440064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 56540064aecSDennis Zhou (Facebook) * 566525ca84dSDennis Zhou (Facebook) * Iterates over the metadata blocks to find the largest contig area. 567525ca84dSDennis Zhou (Facebook) * It also counts the populated pages and uses the delta to update the 568525ca84dSDennis Zhou (Facebook) * global count. 56940064aecSDennis Zhou (Facebook) * 57040064aecSDennis Zhou (Facebook) * Updates: 57140064aecSDennis Zhou (Facebook) * chunk->contig_bits 57213f96637SDennis Zhou (Facebook) * chunk->contig_bits_start 573525ca84dSDennis Zhou (Facebook) * nr_empty_pop_pages (chunk and global) 57440064aecSDennis Zhou (Facebook) */ 57540064aecSDennis Zhou (Facebook) static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) 57640064aecSDennis Zhou (Facebook) { 577525ca84dSDennis Zhou (Facebook) int bit_off, bits, nr_empty_pop_pages; 57840064aecSDennis Zhou (Facebook) 57940064aecSDennis Zhou (Facebook) /* clear metadata */ 58040064aecSDennis Zhou (Facebook) chunk->contig_bits = 0; 58140064aecSDennis Zhou (Facebook) 582525ca84dSDennis Zhou (Facebook) bit_off = chunk->first_bit; 58340064aecSDennis Zhou (Facebook) bits = nr_empty_pop_pages = 0; 584525ca84dSDennis Zhou (Facebook) pcpu_for_each_md_free_region(chunk, bit_off, bits) { 585525ca84dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, bit_off, bits); 58640064aecSDennis Zhou (Facebook) 587525ca84dSDennis Zhou (Facebook) nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits); 58840064aecSDennis Zhou (Facebook) } 58940064aecSDennis Zhou (Facebook) 59040064aecSDennis Zhou (Facebook) /* 59140064aecSDennis Zhou (Facebook) * Keep track of nr_empty_pop_pages. 59240064aecSDennis Zhou (Facebook) * 59340064aecSDennis Zhou (Facebook) * The chunk maintains the previous number of free pages it held, 59440064aecSDennis Zhou (Facebook) * so the delta is used to update the global counter. The reserved 59540064aecSDennis Zhou (Facebook) * chunk is not part of the free page count as they are populated 59640064aecSDennis Zhou (Facebook) * at init and are special to serving reserved allocations. 59740064aecSDennis Zhou (Facebook) */ 59840064aecSDennis Zhou (Facebook) if (chunk != pcpu_reserved_chunk) 59940064aecSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages += 60040064aecSDennis Zhou (Facebook) (nr_empty_pop_pages - chunk->nr_empty_pop_pages); 60140064aecSDennis Zhou (Facebook) 60240064aecSDennis Zhou (Facebook) chunk->nr_empty_pop_pages = nr_empty_pop_pages; 60340064aecSDennis Zhou (Facebook) } 60440064aecSDennis Zhou (Facebook) 60540064aecSDennis Zhou (Facebook) /** 606ca460b3cSDennis Zhou (Facebook) * pcpu_block_update - updates a block given a free area 607ca460b3cSDennis Zhou (Facebook) * @block: block of interest 608ca460b3cSDennis Zhou (Facebook) * @start: start offset in block 609ca460b3cSDennis Zhou (Facebook) * @end: end offset in block 610ca460b3cSDennis Zhou (Facebook) * 611ca460b3cSDennis Zhou (Facebook) * Updates a block given a known free area. The region [start, end) is 612268625a6SDennis Zhou (Facebook) * expected to be the entirety of the free area within a block. Chooses 613268625a6SDennis Zhou (Facebook) * the best starting offset if the contig hints are equal. 614ca460b3cSDennis Zhou (Facebook) */ 615ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 616ca460b3cSDennis Zhou (Facebook) { 617ca460b3cSDennis Zhou (Facebook) int contig = end - start; 618ca460b3cSDennis Zhou (Facebook) 619ca460b3cSDennis Zhou (Facebook) block->first_free = min(block->first_free, start); 620ca460b3cSDennis Zhou (Facebook) if (start == 0) 621ca460b3cSDennis Zhou (Facebook) block->left_free = contig; 622ca460b3cSDennis Zhou (Facebook) 623ca460b3cSDennis Zhou (Facebook) if (end == PCPU_BITMAP_BLOCK_BITS) 624ca460b3cSDennis Zhou (Facebook) block->right_free = contig; 625ca460b3cSDennis Zhou (Facebook) 626ca460b3cSDennis Zhou (Facebook) if (contig > block->contig_hint) { 627ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = start; 628ca460b3cSDennis Zhou (Facebook) block->contig_hint = contig; 629268625a6SDennis Zhou (Facebook) } else if (block->contig_hint_start && contig == block->contig_hint && 630268625a6SDennis Zhou (Facebook) (!start || __ffs(start) > __ffs(block->contig_hint_start))) { 631268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 632268625a6SDennis Zhou (Facebook) block->contig_hint_start = start; 633ca460b3cSDennis Zhou (Facebook) } 634ca460b3cSDennis Zhou (Facebook) } 635ca460b3cSDennis Zhou (Facebook) 636ca460b3cSDennis Zhou (Facebook) /** 637ca460b3cSDennis Zhou (Facebook) * pcpu_block_refresh_hint 638ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 639ca460b3cSDennis Zhou (Facebook) * @index: index of the metadata block 640ca460b3cSDennis Zhou (Facebook) * 641ca460b3cSDennis Zhou (Facebook) * Scans over the block beginning at first_free and updates the block 642ca460b3cSDennis Zhou (Facebook) * metadata accordingly. 643ca460b3cSDennis Zhou (Facebook) */ 644ca460b3cSDennis Zhou (Facebook) static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 645ca460b3cSDennis Zhou (Facebook) { 646ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *block = chunk->md_blocks + index; 647ca460b3cSDennis Zhou (Facebook) unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 648ca460b3cSDennis Zhou (Facebook) int rs, re; /* region start, region end */ 649ca460b3cSDennis Zhou (Facebook) 650ca460b3cSDennis Zhou (Facebook) /* clear hints */ 651ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 652ca460b3cSDennis Zhou (Facebook) block->left_free = block->right_free = 0; 653ca460b3cSDennis Zhou (Facebook) 654ca460b3cSDennis Zhou (Facebook) /* iterate over free areas and update the contig hints */ 655ca460b3cSDennis Zhou (Facebook) pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, 656ca460b3cSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS) { 657ca460b3cSDennis Zhou (Facebook) pcpu_block_update(block, rs, re); 658ca460b3cSDennis Zhou (Facebook) } 659ca460b3cSDennis Zhou (Facebook) } 660ca460b3cSDennis Zhou (Facebook) 661ca460b3cSDennis Zhou (Facebook) /** 662ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_alloc - update hint on allocation path 663ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 664ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 665ca460b3cSDennis Zhou (Facebook) * @bits: size of request 666fc304334SDennis Zhou (Facebook) * 667fc304334SDennis Zhou (Facebook) * Updates metadata for the allocation path. The metadata only has to be 668fc304334SDennis Zhou (Facebook) * refreshed by a full scan iff the chunk's contig hint is broken. Block level 669fc304334SDennis Zhou (Facebook) * scans are required if the block's contig hint is broken. 670ca460b3cSDennis Zhou (Facebook) */ 671ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 672ca460b3cSDennis Zhou (Facebook) int bits) 673ca460b3cSDennis Zhou (Facebook) { 674ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 675ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 676ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 677ca460b3cSDennis Zhou (Facebook) 678ca460b3cSDennis Zhou (Facebook) /* 679ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 680ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 681ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 682ca460b3cSDennis Zhou (Facebook) * range. 683ca460b3cSDennis Zhou (Facebook) */ 684ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 685ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 686ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 687ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 688ca460b3cSDennis Zhou (Facebook) 689ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 690ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 691ca460b3cSDennis Zhou (Facebook) 692ca460b3cSDennis Zhou (Facebook) /* 693ca460b3cSDennis Zhou (Facebook) * Update s_block. 694fc304334SDennis Zhou (Facebook) * block->first_free must be updated if the allocation takes its place. 695fc304334SDennis Zhou (Facebook) * If the allocation breaks the contig_hint, a scan is required to 696fc304334SDennis Zhou (Facebook) * restore this hint. 697ca460b3cSDennis Zhou (Facebook) */ 698fc304334SDennis Zhou (Facebook) if (s_off == s_block->first_free) 699fc304334SDennis Zhou (Facebook) s_block->first_free = find_next_zero_bit( 700fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, s_index), 701fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, 702fc304334SDennis Zhou (Facebook) s_off + bits); 703fc304334SDennis Zhou (Facebook) 704fc304334SDennis Zhou (Facebook) if (s_off >= s_block->contig_hint_start && 705fc304334SDennis Zhou (Facebook) s_off < s_block->contig_hint_start + s_block->contig_hint) { 706fc304334SDennis Zhou (Facebook) /* block contig hint is broken - scan to fix it */ 707ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 708fc304334SDennis Zhou (Facebook) } else { 709fc304334SDennis Zhou (Facebook) /* update left and right contig manually */ 710fc304334SDennis Zhou (Facebook) s_block->left_free = min(s_block->left_free, s_off); 711fc304334SDennis Zhou (Facebook) if (s_index == e_index) 712fc304334SDennis Zhou (Facebook) s_block->right_free = min_t(int, s_block->right_free, 713fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 714fc304334SDennis Zhou (Facebook) else 715fc304334SDennis Zhou (Facebook) s_block->right_free = 0; 716fc304334SDennis Zhou (Facebook) } 717ca460b3cSDennis Zhou (Facebook) 718ca460b3cSDennis Zhou (Facebook) /* 719ca460b3cSDennis Zhou (Facebook) * Update e_block. 720ca460b3cSDennis Zhou (Facebook) */ 721ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 722fc304334SDennis Zhou (Facebook) /* 723fc304334SDennis Zhou (Facebook) * When the allocation is across blocks, the end is along 724fc304334SDennis Zhou (Facebook) * the left part of the e_block. 725fc304334SDennis Zhou (Facebook) */ 726fc304334SDennis Zhou (Facebook) e_block->first_free = find_next_zero_bit( 727fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, e_index), 728fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, e_off); 729fc304334SDennis Zhou (Facebook) 730fc304334SDennis Zhou (Facebook) if (e_off == PCPU_BITMAP_BLOCK_BITS) { 731fc304334SDennis Zhou (Facebook) /* reset the block */ 732fc304334SDennis Zhou (Facebook) e_block++; 733fc304334SDennis Zhou (Facebook) } else { 734fc304334SDennis Zhou (Facebook) if (e_off > e_block->contig_hint_start) { 735fc304334SDennis Zhou (Facebook) /* contig hint is broken - scan to fix it */ 736ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 737fc304334SDennis Zhou (Facebook) } else { 738fc304334SDennis Zhou (Facebook) e_block->left_free = 0; 739fc304334SDennis Zhou (Facebook) e_block->right_free = 740fc304334SDennis Zhou (Facebook) min_t(int, e_block->right_free, 741fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 742fc304334SDennis Zhou (Facebook) } 743fc304334SDennis Zhou (Facebook) } 744ca460b3cSDennis Zhou (Facebook) 745ca460b3cSDennis Zhou (Facebook) /* update in-between md_blocks */ 746ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 747ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 748ca460b3cSDennis Zhou (Facebook) block->left_free = 0; 749ca460b3cSDennis Zhou (Facebook) block->right_free = 0; 750ca460b3cSDennis Zhou (Facebook) } 751ca460b3cSDennis Zhou (Facebook) } 752ca460b3cSDennis Zhou (Facebook) 753fc304334SDennis Zhou (Facebook) /* 754fc304334SDennis Zhou (Facebook) * The only time a full chunk scan is required is if the chunk 755fc304334SDennis Zhou (Facebook) * contig hint is broken. Otherwise, it means a smaller space 756fc304334SDennis Zhou (Facebook) * was used and therefore the chunk contig hint is still correct. 757fc304334SDennis Zhou (Facebook) */ 758fc304334SDennis Zhou (Facebook) if (bit_off >= chunk->contig_bits_start && 759fc304334SDennis Zhou (Facebook) bit_off < chunk->contig_bits_start + chunk->contig_bits) 760ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 761ca460b3cSDennis Zhou (Facebook) } 762ca460b3cSDennis Zhou (Facebook) 763ca460b3cSDennis Zhou (Facebook) /** 764ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_free - updates the block hints on the free path 765ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 766ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 767ca460b3cSDennis Zhou (Facebook) * @bits: size of request 768b185cd0dSDennis Zhou (Facebook) * 769b185cd0dSDennis Zhou (Facebook) * Updates metadata for the allocation path. This avoids a blind block 770b185cd0dSDennis Zhou (Facebook) * refresh by making use of the block contig hints. If this fails, it scans 771b185cd0dSDennis Zhou (Facebook) * forward and backward to determine the extent of the free area. This is 772b185cd0dSDennis Zhou (Facebook) * capped at the boundary of blocks. 773b185cd0dSDennis Zhou (Facebook) * 774b185cd0dSDennis Zhou (Facebook) * A chunk update is triggered if a page becomes free, a block becomes free, 775b185cd0dSDennis Zhou (Facebook) * or the free spans across blocks. This tradeoff is to minimize iterating 776b185cd0dSDennis Zhou (Facebook) * over the block metadata to update chunk->contig_bits. chunk->contig_bits 777b185cd0dSDennis Zhou (Facebook) * may be off by up to a page, but it will never be more than the available 778b185cd0dSDennis Zhou (Facebook) * space. If the contig hint is contained in one block, it will be accurate. 779ca460b3cSDennis Zhou (Facebook) */ 780ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 781ca460b3cSDennis Zhou (Facebook) int bits) 782ca460b3cSDennis Zhou (Facebook) { 783ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 784ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 785ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 786b185cd0dSDennis Zhou (Facebook) int start, end; /* start and end of the whole free area */ 787ca460b3cSDennis Zhou (Facebook) 788ca460b3cSDennis Zhou (Facebook) /* 789ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 790ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 791ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 792ca460b3cSDennis Zhou (Facebook) * range. 793ca460b3cSDennis Zhou (Facebook) */ 794ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 795ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 796ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 797ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 798ca460b3cSDennis Zhou (Facebook) 799ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 800ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 801ca460b3cSDennis Zhou (Facebook) 802b185cd0dSDennis Zhou (Facebook) /* 803b185cd0dSDennis Zhou (Facebook) * Check if the freed area aligns with the block->contig_hint. 804b185cd0dSDennis Zhou (Facebook) * If it does, then the scan to find the beginning/end of the 805b185cd0dSDennis Zhou (Facebook) * larger free area can be avoided. 806b185cd0dSDennis Zhou (Facebook) * 807b185cd0dSDennis Zhou (Facebook) * start and end refer to beginning and end of the free area 808b185cd0dSDennis Zhou (Facebook) * within each their respective blocks. This is not necessarily 809b185cd0dSDennis Zhou (Facebook) * the entire free area as it may span blocks past the beginning 810b185cd0dSDennis Zhou (Facebook) * or end of the block. 811b185cd0dSDennis Zhou (Facebook) */ 812b185cd0dSDennis Zhou (Facebook) start = s_off; 813b185cd0dSDennis Zhou (Facebook) if (s_off == s_block->contig_hint + s_block->contig_hint_start) { 814b185cd0dSDennis Zhou (Facebook) start = s_block->contig_hint_start; 815b185cd0dSDennis Zhou (Facebook) } else { 816b185cd0dSDennis Zhou (Facebook) /* 817b185cd0dSDennis Zhou (Facebook) * Scan backwards to find the extent of the free area. 818b185cd0dSDennis Zhou (Facebook) * find_last_bit returns the starting bit, so if the start bit 819b185cd0dSDennis Zhou (Facebook) * is returned, that means there was no last bit and the 820b185cd0dSDennis Zhou (Facebook) * remainder of the chunk is free. 821b185cd0dSDennis Zhou (Facebook) */ 822b185cd0dSDennis Zhou (Facebook) int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), 823b185cd0dSDennis Zhou (Facebook) start); 824b185cd0dSDennis Zhou (Facebook) start = (start == l_bit) ? 0 : l_bit + 1; 825b185cd0dSDennis Zhou (Facebook) } 826b185cd0dSDennis Zhou (Facebook) 827b185cd0dSDennis Zhou (Facebook) end = e_off; 828b185cd0dSDennis Zhou (Facebook) if (e_off == e_block->contig_hint_start) 829b185cd0dSDennis Zhou (Facebook) end = e_block->contig_hint_start + e_block->contig_hint; 830b185cd0dSDennis Zhou (Facebook) else 831b185cd0dSDennis Zhou (Facebook) end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), 832b185cd0dSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, end); 833b185cd0dSDennis Zhou (Facebook) 834ca460b3cSDennis Zhou (Facebook) /* update s_block */ 835b185cd0dSDennis Zhou (Facebook) e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; 836b185cd0dSDennis Zhou (Facebook) pcpu_block_update(s_block, start, e_off); 837ca460b3cSDennis Zhou (Facebook) 838ca460b3cSDennis Zhou (Facebook) /* freeing in the same block */ 839ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 840ca460b3cSDennis Zhou (Facebook) /* update e_block */ 841b185cd0dSDennis Zhou (Facebook) pcpu_block_update(e_block, 0, end); 842ca460b3cSDennis Zhou (Facebook) 843ca460b3cSDennis Zhou (Facebook) /* reset md_blocks in the middle */ 844ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 845ca460b3cSDennis Zhou (Facebook) block->first_free = 0; 846ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = 0; 847ca460b3cSDennis Zhou (Facebook) block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 848ca460b3cSDennis Zhou (Facebook) block->left_free = PCPU_BITMAP_BLOCK_BITS; 849ca460b3cSDennis Zhou (Facebook) block->right_free = PCPU_BITMAP_BLOCK_BITS; 850ca460b3cSDennis Zhou (Facebook) } 851ca460b3cSDennis Zhou (Facebook) } 852ca460b3cSDennis Zhou (Facebook) 853b185cd0dSDennis Zhou (Facebook) /* 854b185cd0dSDennis Zhou (Facebook) * Refresh chunk metadata when the free makes a page free, a block 855b185cd0dSDennis Zhou (Facebook) * free, or spans across blocks. The contig hint may be off by up to 856b185cd0dSDennis Zhou (Facebook) * a page, but if the hint is contained in a block, it will be accurate 857b185cd0dSDennis Zhou (Facebook) * with the else condition below. 858b185cd0dSDennis Zhou (Facebook) */ 859b185cd0dSDennis Zhou (Facebook) if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) > 860b185cd0dSDennis Zhou (Facebook) ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) || 861b185cd0dSDennis Zhou (Facebook) s_index != e_index) 862ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 863b185cd0dSDennis Zhou (Facebook) else 864b185cd0dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start), 865b185cd0dSDennis Zhou (Facebook) s_block->contig_hint); 866ca460b3cSDennis Zhou (Facebook) } 867ca460b3cSDennis Zhou (Facebook) 868ca460b3cSDennis Zhou (Facebook) /** 86940064aecSDennis Zhou (Facebook) * pcpu_is_populated - determines if the region is populated 87040064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 87140064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 87240064aecSDennis Zhou (Facebook) * @bits: size of area 87340064aecSDennis Zhou (Facebook) * @next_off: return value for the next offset to start searching 87440064aecSDennis Zhou (Facebook) * 87540064aecSDennis Zhou (Facebook) * For atomic allocations, check if the backing pages are populated. 87640064aecSDennis Zhou (Facebook) * 87740064aecSDennis Zhou (Facebook) * RETURNS: 87840064aecSDennis Zhou (Facebook) * Bool if the backing pages are populated. 87940064aecSDennis Zhou (Facebook) * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 88040064aecSDennis Zhou (Facebook) */ 88140064aecSDennis Zhou (Facebook) static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 88240064aecSDennis Zhou (Facebook) int *next_off) 88340064aecSDennis Zhou (Facebook) { 88440064aecSDennis Zhou (Facebook) int page_start, page_end, rs, re; 88540064aecSDennis Zhou (Facebook) 88640064aecSDennis Zhou (Facebook) page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 88740064aecSDennis Zhou (Facebook) page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 88840064aecSDennis Zhou (Facebook) 88940064aecSDennis Zhou (Facebook) rs = page_start; 89040064aecSDennis Zhou (Facebook) pcpu_next_unpop(chunk->populated, &rs, &re, page_end); 89140064aecSDennis Zhou (Facebook) if (rs >= page_end) 89240064aecSDennis Zhou (Facebook) return true; 89340064aecSDennis Zhou (Facebook) 89440064aecSDennis Zhou (Facebook) *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 89540064aecSDennis Zhou (Facebook) return false; 89640064aecSDennis Zhou (Facebook) } 89740064aecSDennis Zhou (Facebook) 89840064aecSDennis Zhou (Facebook) /** 89940064aecSDennis Zhou (Facebook) * pcpu_find_block_fit - finds the block index to start searching 90040064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 90140064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 90240064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE bytes) 90340064aecSDennis Zhou (Facebook) * @pop_only: use populated regions only 90440064aecSDennis Zhou (Facebook) * 905b4c2116cSDennis Zhou (Facebook) * Given a chunk and an allocation spec, find the offset to begin searching 906b4c2116cSDennis Zhou (Facebook) * for a free region. This iterates over the bitmap metadata blocks to 907b4c2116cSDennis Zhou (Facebook) * find an offset that will be guaranteed to fit the requirements. It is 908b4c2116cSDennis Zhou (Facebook) * not quite first fit as if the allocation does not fit in the contig hint 909b4c2116cSDennis Zhou (Facebook) * of a block or chunk, it is skipped. This errs on the side of caution 910b4c2116cSDennis Zhou (Facebook) * to prevent excess iteration. Poor alignment can cause the allocator to 911b4c2116cSDennis Zhou (Facebook) * skip over blocks and chunks that have valid free areas. 912b4c2116cSDennis Zhou (Facebook) * 91340064aecSDennis Zhou (Facebook) * RETURNS: 91440064aecSDennis Zhou (Facebook) * The offset in the bitmap to begin searching. 91540064aecSDennis Zhou (Facebook) * -1 if no offset is found. 91640064aecSDennis Zhou (Facebook) */ 91740064aecSDennis Zhou (Facebook) static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 91840064aecSDennis Zhou (Facebook) size_t align, bool pop_only) 91940064aecSDennis Zhou (Facebook) { 920b4c2116cSDennis Zhou (Facebook) int bit_off, bits, next_off; 92140064aecSDennis Zhou (Facebook) 92213f96637SDennis Zhou (Facebook) /* 92313f96637SDennis Zhou (Facebook) * Check to see if the allocation can fit in the chunk's contig hint. 92413f96637SDennis Zhou (Facebook) * This is an optimization to prevent scanning by assuming if it 92513f96637SDennis Zhou (Facebook) * cannot fit in the global hint, there is memory pressure and creating 92613f96637SDennis Zhou (Facebook) * a new chunk would happen soon. 92713f96637SDennis Zhou (Facebook) */ 92813f96637SDennis Zhou (Facebook) bit_off = ALIGN(chunk->contig_bits_start, align) - 92913f96637SDennis Zhou (Facebook) chunk->contig_bits_start; 93013f96637SDennis Zhou (Facebook) if (bit_off + alloc_bits > chunk->contig_bits) 93113f96637SDennis Zhou (Facebook) return -1; 93213f96637SDennis Zhou (Facebook) 933b4c2116cSDennis Zhou (Facebook) bit_off = chunk->first_bit; 934b4c2116cSDennis Zhou (Facebook) bits = 0; 935b4c2116cSDennis Zhou (Facebook) pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { 93640064aecSDennis Zhou (Facebook) if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 937b4c2116cSDennis Zhou (Facebook) &next_off)) 93840064aecSDennis Zhou (Facebook) break; 93940064aecSDennis Zhou (Facebook) 940b4c2116cSDennis Zhou (Facebook) bit_off = next_off; 94140064aecSDennis Zhou (Facebook) bits = 0; 94240064aecSDennis Zhou (Facebook) } 94340064aecSDennis Zhou (Facebook) 94440064aecSDennis Zhou (Facebook) if (bit_off == pcpu_chunk_map_bits(chunk)) 94540064aecSDennis Zhou (Facebook) return -1; 94640064aecSDennis Zhou (Facebook) 94740064aecSDennis Zhou (Facebook) return bit_off; 94840064aecSDennis Zhou (Facebook) } 94940064aecSDennis Zhou (Facebook) 95040064aecSDennis Zhou (Facebook) /** 95140064aecSDennis Zhou (Facebook) * pcpu_alloc_area - allocates an area from a pcpu_chunk 95240064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 95340064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 95440064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 95540064aecSDennis Zhou (Facebook) * @start: bit_off to start searching 95640064aecSDennis Zhou (Facebook) * 95740064aecSDennis Zhou (Facebook) * This function takes in a @start offset to begin searching to fit an 958b4c2116cSDennis Zhou (Facebook) * allocation of @alloc_bits with alignment @align. It needs to scan 959b4c2116cSDennis Zhou (Facebook) * the allocation map because if it fits within the block's contig hint, 960b4c2116cSDennis Zhou (Facebook) * @start will be block->first_free. This is an attempt to fill the 961b4c2116cSDennis Zhou (Facebook) * allocation prior to breaking the contig hint. The allocation and 962b4c2116cSDennis Zhou (Facebook) * boundary maps are updated accordingly if it confirms a valid 963b4c2116cSDennis Zhou (Facebook) * free area. 96440064aecSDennis Zhou (Facebook) * 96540064aecSDennis Zhou (Facebook) * RETURNS: 96640064aecSDennis Zhou (Facebook) * Allocated addr offset in @chunk on success. 96740064aecSDennis Zhou (Facebook) * -1 if no matching area is found. 96840064aecSDennis Zhou (Facebook) */ 96940064aecSDennis Zhou (Facebook) static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 97040064aecSDennis Zhou (Facebook) size_t align, int start) 97140064aecSDennis Zhou (Facebook) { 97240064aecSDennis Zhou (Facebook) size_t align_mask = (align) ? (align - 1) : 0; 97340064aecSDennis Zhou (Facebook) int bit_off, end, oslot; 9749f7dcf22STejun Heo 9754f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 9764f996e23STejun Heo 97740064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 978833af842STejun Heo 979833af842STejun Heo /* 98040064aecSDennis Zhou (Facebook) * Search to find a fit. 981833af842STejun Heo */ 982b4c2116cSDennis Zhou (Facebook) end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS; 98340064aecSDennis Zhou (Facebook) bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, 98440064aecSDennis Zhou (Facebook) alloc_bits, align_mask); 98540064aecSDennis Zhou (Facebook) if (bit_off >= end) 986a16037c8STejun Heo return -1; 987a16037c8STejun Heo 98840064aecSDennis Zhou (Facebook) /* update alloc map */ 98940064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 990a16037c8STejun Heo 99140064aecSDennis Zhou (Facebook) /* update boundary map */ 99240064aecSDennis Zhou (Facebook) set_bit(bit_off, chunk->bound_map); 99340064aecSDennis Zhou (Facebook) bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 99440064aecSDennis Zhou (Facebook) set_bit(bit_off + alloc_bits, chunk->bound_map); 995a16037c8STejun Heo 99640064aecSDennis Zhou (Facebook) chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 99740064aecSDennis Zhou (Facebook) 99886b442fbSDennis Zhou (Facebook) /* update first free bit */ 99986b442fbSDennis Zhou (Facebook) if (bit_off == chunk->first_bit) 100086b442fbSDennis Zhou (Facebook) chunk->first_bit = find_next_zero_bit( 100186b442fbSDennis Zhou (Facebook) chunk->alloc_map, 100286b442fbSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk), 100386b442fbSDennis Zhou (Facebook) bit_off + alloc_bits); 100486b442fbSDennis Zhou (Facebook) 1005ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 100640064aecSDennis Zhou (Facebook) 100740064aecSDennis Zhou (Facebook) pcpu_chunk_relocate(chunk, oslot); 100840064aecSDennis Zhou (Facebook) 100940064aecSDennis Zhou (Facebook) return bit_off * PCPU_MIN_ALLOC_SIZE; 1010a16037c8STejun Heo } 1011a16037c8STejun Heo 1012a16037c8STejun Heo /** 101340064aecSDennis Zhou (Facebook) * pcpu_free_area - frees the corresponding offset 1014fbf59bc9STejun Heo * @chunk: chunk of interest 101540064aecSDennis Zhou (Facebook) * @off: addr offset into chunk 1016fbf59bc9STejun Heo * 101740064aecSDennis Zhou (Facebook) * This function determines the size of an allocation to free using 101840064aecSDennis Zhou (Facebook) * the boundary bitmap and clears the allocation map. 1019fbf59bc9STejun Heo */ 102040064aecSDennis Zhou (Facebook) static void pcpu_free_area(struct pcpu_chunk *chunk, int off) 1021fbf59bc9STejun Heo { 102240064aecSDennis Zhou (Facebook) int bit_off, bits, end, oslot; 1023fbf59bc9STejun Heo 10245ccd30e4SDennis Zhou lockdep_assert_held(&pcpu_lock); 102530a5b536SDennis Zhou pcpu_stats_area_dealloc(chunk); 10265ccd30e4SDennis Zhou 102740064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1028723ad1d9SAl Viro 102940064aecSDennis Zhou (Facebook) bit_off = off / PCPU_MIN_ALLOC_SIZE; 1030fbf59bc9STejun Heo 103140064aecSDennis Zhou (Facebook) /* find end index */ 103240064aecSDennis Zhou (Facebook) end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 103340064aecSDennis Zhou (Facebook) bit_off + 1); 103440064aecSDennis Zhou (Facebook) bits = end - bit_off; 103540064aecSDennis Zhou (Facebook) bitmap_clear(chunk->alloc_map, bit_off, bits); 10363d331ad7SAl Viro 103740064aecSDennis Zhou (Facebook) /* update metadata */ 103840064aecSDennis Zhou (Facebook) chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; 1039fbf59bc9STejun Heo 104086b442fbSDennis Zhou (Facebook) /* update first free bit */ 104186b442fbSDennis Zhou (Facebook) chunk->first_bit = min(chunk->first_bit, bit_off); 104286b442fbSDennis Zhou (Facebook) 1043ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_free(chunk, bit_off, bits); 1044b539b87fSTejun Heo 1045fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 1046fbf59bc9STejun Heo } 1047fbf59bc9STejun Heo 1048ca460b3cSDennis Zhou (Facebook) static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 1049ca460b3cSDennis Zhou (Facebook) { 1050ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *md_block; 1051ca460b3cSDennis Zhou (Facebook) 1052ca460b3cSDennis Zhou (Facebook) for (md_block = chunk->md_blocks; 1053ca460b3cSDennis Zhou (Facebook) md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 1054ca460b3cSDennis Zhou (Facebook) md_block++) { 1055ca460b3cSDennis Zhou (Facebook) md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 1056ca460b3cSDennis Zhou (Facebook) md_block->left_free = PCPU_BITMAP_BLOCK_BITS; 1057ca460b3cSDennis Zhou (Facebook) md_block->right_free = PCPU_BITMAP_BLOCK_BITS; 1058ca460b3cSDennis Zhou (Facebook) } 1059ca460b3cSDennis Zhou (Facebook) } 1060ca460b3cSDennis Zhou (Facebook) 106140064aecSDennis Zhou (Facebook) /** 106240064aecSDennis Zhou (Facebook) * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 106340064aecSDennis Zhou (Facebook) * @tmp_addr: the start of the region served 106440064aecSDennis Zhou (Facebook) * @map_size: size of the region served 106540064aecSDennis Zhou (Facebook) * 106640064aecSDennis Zhou (Facebook) * This is responsible for creating the chunks that serve the first chunk. The 106740064aecSDennis Zhou (Facebook) * base_addr is page aligned down of @tmp_addr while the region end is page 106840064aecSDennis Zhou (Facebook) * aligned up. Offsets are kept track of to determine the region served. All 106940064aecSDennis Zhou (Facebook) * this is done to appease the bitmap allocator in avoiding partial blocks. 107040064aecSDennis Zhou (Facebook) * 107140064aecSDennis Zhou (Facebook) * RETURNS: 107240064aecSDennis Zhou (Facebook) * Chunk serving the region at @tmp_addr of @map_size. 107340064aecSDennis Zhou (Facebook) */ 1074c0ebfdc3SDennis Zhou (Facebook) static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 107540064aecSDennis Zhou (Facebook) int map_size) 107610edf5b0SDennis Zhou (Facebook) { 107710edf5b0SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 1078ca460b3cSDennis Zhou (Facebook) unsigned long aligned_addr, lcm_align; 107940064aecSDennis Zhou (Facebook) int start_offset, offset_bits, region_size, region_bits; 1080c0ebfdc3SDennis Zhou (Facebook) 1081c0ebfdc3SDennis Zhou (Facebook) /* region calculations */ 1082c0ebfdc3SDennis Zhou (Facebook) aligned_addr = tmp_addr & PAGE_MASK; 1083c0ebfdc3SDennis Zhou (Facebook) 1084c0ebfdc3SDennis Zhou (Facebook) start_offset = tmp_addr - aligned_addr; 10856b9d7c8eSDennis Zhou (Facebook) 1086ca460b3cSDennis Zhou (Facebook) /* 1087ca460b3cSDennis Zhou (Facebook) * Align the end of the region with the LCM of PAGE_SIZE and 1088ca460b3cSDennis Zhou (Facebook) * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of 1089ca460b3cSDennis Zhou (Facebook) * the other. 1090ca460b3cSDennis Zhou (Facebook) */ 1091ca460b3cSDennis Zhou (Facebook) lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); 1092ca460b3cSDennis Zhou (Facebook) region_size = ALIGN(start_offset + map_size, lcm_align); 109310edf5b0SDennis Zhou (Facebook) 1094c0ebfdc3SDennis Zhou (Facebook) /* allocate chunk */ 10958ab16c43SDennis Zhou (Facebook) chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + 10968ab16c43SDennis Zhou (Facebook) BITS_TO_LONGS(region_size >> PAGE_SHIFT), 10978ab16c43SDennis Zhou (Facebook) 0); 1098c0ebfdc3SDennis Zhou (Facebook) 109910edf5b0SDennis Zhou (Facebook) INIT_LIST_HEAD(&chunk->list); 1100c0ebfdc3SDennis Zhou (Facebook) 1101c0ebfdc3SDennis Zhou (Facebook) chunk->base_addr = (void *)aligned_addr; 110210edf5b0SDennis Zhou (Facebook) chunk->start_offset = start_offset; 11036b9d7c8eSDennis Zhou (Facebook) chunk->end_offset = region_size - chunk->start_offset - map_size; 1104c0ebfdc3SDennis Zhou (Facebook) 11058ab16c43SDennis Zhou (Facebook) chunk->nr_pages = region_size >> PAGE_SHIFT; 110640064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 1107c0ebfdc3SDennis Zhou (Facebook) 1108ca460b3cSDennis Zhou (Facebook) chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) * 110940064aecSDennis Zhou (Facebook) sizeof(chunk->alloc_map[0]), 0); 1110ca460b3cSDennis Zhou (Facebook) chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) * 111140064aecSDennis Zhou (Facebook) sizeof(chunk->bound_map[0]), 0); 1112ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) * 1113ca460b3cSDennis Zhou (Facebook) sizeof(chunk->md_blocks[0]), 0); 1114ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 111510edf5b0SDennis Zhou (Facebook) 111610edf5b0SDennis Zhou (Facebook) /* manage populated page bitmap */ 111710edf5b0SDennis Zhou (Facebook) chunk->immutable = true; 11188ab16c43SDennis Zhou (Facebook) bitmap_fill(chunk->populated, chunk->nr_pages); 11198ab16c43SDennis Zhou (Facebook) chunk->nr_populated = chunk->nr_pages; 112040064aecSDennis Zhou (Facebook) chunk->nr_empty_pop_pages = 112140064aecSDennis Zhou (Facebook) pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE, 112240064aecSDennis Zhou (Facebook) map_size / PCPU_MIN_ALLOC_SIZE); 112310edf5b0SDennis Zhou (Facebook) 112440064aecSDennis Zhou (Facebook) chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; 112540064aecSDennis Zhou (Facebook) chunk->free_bytes = map_size; 1126c0ebfdc3SDennis Zhou (Facebook) 1127c0ebfdc3SDennis Zhou (Facebook) if (chunk->start_offset) { 1128c0ebfdc3SDennis Zhou (Facebook) /* hide the beginning of the bitmap */ 112940064aecSDennis Zhou (Facebook) offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 113040064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 0, offset_bits); 113140064aecSDennis Zhou (Facebook) set_bit(0, chunk->bound_map); 113240064aecSDennis Zhou (Facebook) set_bit(offset_bits, chunk->bound_map); 1133ca460b3cSDennis Zhou (Facebook) 113486b442fbSDennis Zhou (Facebook) chunk->first_bit = offset_bits; 113586b442fbSDennis Zhou (Facebook) 1136ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 1137c0ebfdc3SDennis Zhou (Facebook) } 1138c0ebfdc3SDennis Zhou (Facebook) 11396b9d7c8eSDennis Zhou (Facebook) if (chunk->end_offset) { 11406b9d7c8eSDennis Zhou (Facebook) /* hide the end of the bitmap */ 114140064aecSDennis Zhou (Facebook) offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 114240064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 114340064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk) - offset_bits, 114440064aecSDennis Zhou (Facebook) offset_bits); 114540064aecSDennis Zhou (Facebook) set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 114640064aecSDennis Zhou (Facebook) chunk->bound_map); 114740064aecSDennis Zhou (Facebook) set_bit(region_bits, chunk->bound_map); 11486b9d7c8eSDennis Zhou (Facebook) 1149ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 1150ca460b3cSDennis Zhou (Facebook) - offset_bits, offset_bits); 1151ca460b3cSDennis Zhou (Facebook) } 115240064aecSDennis Zhou (Facebook) 115310edf5b0SDennis Zhou (Facebook) return chunk; 115410edf5b0SDennis Zhou (Facebook) } 115510edf5b0SDennis Zhou (Facebook) 115647504ee0SDennis Zhou static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) 11576081089fSTejun Heo { 11586081089fSTejun Heo struct pcpu_chunk *chunk; 115940064aecSDennis Zhou (Facebook) int region_bits; 11606081089fSTejun Heo 116147504ee0SDennis Zhou chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); 11626081089fSTejun Heo if (!chunk) 11636081089fSTejun Heo return NULL; 11646081089fSTejun Heo 11656081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 1166c0ebfdc3SDennis Zhou (Facebook) chunk->nr_pages = pcpu_unit_pages; 116740064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 116840064aecSDennis Zhou (Facebook) 116940064aecSDennis Zhou (Facebook) chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 117047504ee0SDennis Zhou sizeof(chunk->alloc_map[0]), gfp); 117140064aecSDennis Zhou (Facebook) if (!chunk->alloc_map) 117240064aecSDennis Zhou (Facebook) goto alloc_map_fail; 117340064aecSDennis Zhou (Facebook) 117440064aecSDennis Zhou (Facebook) chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 117547504ee0SDennis Zhou sizeof(chunk->bound_map[0]), gfp); 117640064aecSDennis Zhou (Facebook) if (!chunk->bound_map) 117740064aecSDennis Zhou (Facebook) goto bound_map_fail; 117840064aecSDennis Zhou (Facebook) 1179ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 118047504ee0SDennis Zhou sizeof(chunk->md_blocks[0]), gfp); 1181ca460b3cSDennis Zhou (Facebook) if (!chunk->md_blocks) 1182ca460b3cSDennis Zhou (Facebook) goto md_blocks_fail; 1183ca460b3cSDennis Zhou (Facebook) 1184ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 1185ca460b3cSDennis Zhou (Facebook) 118640064aecSDennis Zhou (Facebook) /* init metadata */ 118740064aecSDennis Zhou (Facebook) chunk->contig_bits = region_bits; 118840064aecSDennis Zhou (Facebook) chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 1189c0ebfdc3SDennis Zhou (Facebook) 11906081089fSTejun Heo return chunk; 119140064aecSDennis Zhou (Facebook) 1192ca460b3cSDennis Zhou (Facebook) md_blocks_fail: 1193ca460b3cSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 119440064aecSDennis Zhou (Facebook) bound_map_fail: 119540064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 119640064aecSDennis Zhou (Facebook) alloc_map_fail: 119740064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk); 119840064aecSDennis Zhou (Facebook) 119940064aecSDennis Zhou (Facebook) return NULL; 12006081089fSTejun Heo } 12016081089fSTejun Heo 12026081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 12036081089fSTejun Heo { 12046081089fSTejun Heo if (!chunk) 12056081089fSTejun Heo return; 120640064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 120740064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 12081d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 12096081089fSTejun Heo } 12106081089fSTejun Heo 1211b539b87fSTejun Heo /** 1212b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 1213b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 1214b539b87fSTejun Heo * @page_start: the start page 1215b539b87fSTejun Heo * @page_end: the end page 121640064aecSDennis Zhou (Facebook) * @for_alloc: if this is to populate for allocation 1217b539b87fSTejun Heo * 1218b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1219b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 1220b539b87fSTejun Heo * successful population. 122140064aecSDennis Zhou (Facebook) * 122240064aecSDennis Zhou (Facebook) * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it 122340064aecSDennis Zhou (Facebook) * is to serve an allocation in that area. 1224b539b87fSTejun Heo */ 122540064aecSDennis Zhou (Facebook) static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 122640064aecSDennis Zhou (Facebook) int page_end, bool for_alloc) 1227b539b87fSTejun Heo { 1228b539b87fSTejun Heo int nr = page_end - page_start; 1229b539b87fSTejun Heo 1230b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1231b539b87fSTejun Heo 1232b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 1233b539b87fSTejun Heo chunk->nr_populated += nr; 123440064aecSDennis Zhou (Facebook) 123540064aecSDennis Zhou (Facebook) if (!for_alloc) { 12360cecf50cSDennis Zhou (Facebook) chunk->nr_empty_pop_pages += nr; 1237b539b87fSTejun Heo pcpu_nr_empty_pop_pages += nr; 1238b539b87fSTejun Heo } 123940064aecSDennis Zhou (Facebook) } 1240b539b87fSTejun Heo 1241b539b87fSTejun Heo /** 1242b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 1243b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 1244b539b87fSTejun Heo * @page_start: the start page 1245b539b87fSTejun Heo * @page_end: the end page 1246b539b87fSTejun Heo * 1247b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1248b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 1249b539b87fSTejun Heo * each successful depopulation. 1250b539b87fSTejun Heo */ 1251b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1252b539b87fSTejun Heo int page_start, int page_end) 1253b539b87fSTejun Heo { 1254b539b87fSTejun Heo int nr = page_end - page_start; 1255b539b87fSTejun Heo 1256b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1257b539b87fSTejun Heo 1258b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 1259b539b87fSTejun Heo chunk->nr_populated -= nr; 12600cecf50cSDennis Zhou (Facebook) chunk->nr_empty_pop_pages -= nr; 1261b539b87fSTejun Heo pcpu_nr_empty_pop_pages -= nr; 1262b539b87fSTejun Heo } 1263b539b87fSTejun Heo 1264fbf59bc9STejun Heo /* 12659f645532STejun Heo * Chunk management implementation. 1266fbf59bc9STejun Heo * 12679f645532STejun Heo * To allow different implementations, chunk alloc/free and 12689f645532STejun Heo * [de]population are implemented in a separate file which is pulled 12699f645532STejun Heo * into this file and compiled together. The following functions 12709f645532STejun Heo * should be implemented. 1271ccea34b5STejun Heo * 12729f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 12739f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 12749f645532STejun Heo * pcpu_create_chunk - create a new chunk 12759f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 12769f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 12779f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1278fbf59bc9STejun Heo */ 127915d9f3d1SDennis Zhou static int pcpu_populate_chunk(struct pcpu_chunk *chunk, 128047504ee0SDennis Zhou int page_start, int page_end, gfp_t gfp); 128115d9f3d1SDennis Zhou static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, 128215d9f3d1SDennis Zhou int page_start, int page_end); 128347504ee0SDennis Zhou static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); 12849f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 12859f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 12869f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1287fbf59bc9STejun Heo 1288b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 1289b0c9778bSTejun Heo #include "percpu-km.c" 1290b0c9778bSTejun Heo #else 12919f645532STejun Heo #include "percpu-vm.c" 1292b0c9778bSTejun Heo #endif 1293fbf59bc9STejun Heo 1294fbf59bc9STejun Heo /** 129588999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 129688999a89STejun Heo * @addr: address for which the chunk needs to be determined. 129788999a89STejun Heo * 1298c0ebfdc3SDennis Zhou (Facebook) * This is an internal function that handles all but static allocations. 1299c0ebfdc3SDennis Zhou (Facebook) * Static percpu address values should never be passed into the allocator. 1300c0ebfdc3SDennis Zhou (Facebook) * 130188999a89STejun Heo * RETURNS: 130288999a89STejun Heo * The address of the found chunk. 130388999a89STejun Heo */ 130488999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 130588999a89STejun Heo { 1306c0ebfdc3SDennis Zhou (Facebook) /* is it in the dynamic region (first chunk)? */ 1307560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1308c0ebfdc3SDennis Zhou (Facebook) return pcpu_first_chunk; 1309c0ebfdc3SDennis Zhou (Facebook) 1310c0ebfdc3SDennis Zhou (Facebook) /* is it in the reserved region? */ 1311560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 131288999a89STejun Heo return pcpu_reserved_chunk; 131388999a89STejun Heo 131488999a89STejun Heo /* 131588999a89STejun Heo * The address is relative to unit0 which might be unused and 131688999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 131788999a89STejun Heo * current processor before looking it up in the vmalloc 131888999a89STejun Heo * space. Note that any possible cpu id can be used here, so 131988999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 132088999a89STejun Heo */ 132188999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 13229f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 132388999a89STejun Heo } 132488999a89STejun Heo 132588999a89STejun Heo /** 1326edcb4639STejun Heo * pcpu_alloc - the percpu allocator 1327cae3aeb8STejun Heo * @size: size of area to allocate in bytes 1328fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 1329edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 13305835d96eSTejun Heo * @gfp: allocation flags 1331fbf59bc9STejun Heo * 13325835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 13330ea7eeecSDaniel Borkmann * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN 13340ea7eeecSDaniel Borkmann * then no warning will be triggered on invalid or failed allocation 13350ea7eeecSDaniel Borkmann * requests. 1336fbf59bc9STejun Heo * 1337fbf59bc9STejun Heo * RETURNS: 1338fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1339fbf59bc9STejun Heo */ 13405835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 13415835d96eSTejun Heo gfp_t gfp) 1342fbf59bc9STejun Heo { 1343*554fef1cSDennis Zhou /* whitelisted flags that can be passed to the backing allocators */ 1344*554fef1cSDennis Zhou gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN); 13450ea7eeecSDaniel Borkmann bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 13460ea7eeecSDaniel Borkmann bool do_warn = !(gfp & __GFP_NOWARN); 1347f2badb0cSTejun Heo static int warn_limit = 10; 1348fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1349f2badb0cSTejun Heo const char *err; 135040064aecSDennis Zhou (Facebook) int slot, off, cpu, ret; 1351403a91b1SJiri Kosina unsigned long flags; 1352f528f0b8SCatalin Marinas void __percpu *ptr; 135340064aecSDennis Zhou (Facebook) size_t bits, bit_align; 1354fbf59bc9STejun Heo 1355723ad1d9SAl Viro /* 135640064aecSDennis Zhou (Facebook) * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 135740064aecSDennis Zhou (Facebook) * therefore alignment must be a minimum of that many bytes. 135840064aecSDennis Zhou (Facebook) * An allocation may have internal fragmentation from rounding up 135940064aecSDennis Zhou (Facebook) * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1360723ad1d9SAl Viro */ 1361d2f3c384SDennis Zhou (Facebook) if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1362d2f3c384SDennis Zhou (Facebook) align = PCPU_MIN_ALLOC_SIZE; 1363723ad1d9SAl Viro 1364d2f3c384SDennis Zhou (Facebook) size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 136540064aecSDennis Zhou (Facebook) bits = size >> PCPU_MIN_ALLOC_SHIFT; 136640064aecSDennis Zhou (Facebook) bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 13672f69fa82SViro 13683ca45a46Szijun_hu if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 13693ca45a46Szijun_hu !is_power_of_2(align))) { 13700ea7eeecSDaniel Borkmann WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1371756a025fSJoe Perches size, align); 1372fbf59bc9STejun Heo return NULL; 1373fbf59bc9STejun Heo } 1374fbf59bc9STejun Heo 13756710e594STejun Heo if (!is_atomic) 13766710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 13776710e594STejun Heo 1378403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1379fbf59bc9STejun Heo 1380edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 1381edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 1382edcb4639STejun Heo chunk = pcpu_reserved_chunk; 1383833af842STejun Heo 138440064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 138540064aecSDennis Zhou (Facebook) if (off < 0) { 1386833af842STejun Heo err = "alloc from reserved chunk failed"; 1387ccea34b5STejun Heo goto fail_unlock; 1388f2badb0cSTejun Heo } 1389833af842STejun Heo 139040064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1391edcb4639STejun Heo if (off >= 0) 1392edcb4639STejun Heo goto area_found; 1393833af842STejun Heo 1394f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 1395ccea34b5STejun Heo goto fail_unlock; 1396edcb4639STejun Heo } 1397edcb4639STejun Heo 1398ccea34b5STejun Heo restart: 1399edcb4639STejun Heo /* search through normal chunks */ 1400fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 1401fbf59bc9STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 140240064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, 140340064aecSDennis Zhou (Facebook) is_atomic); 140440064aecSDennis Zhou (Facebook) if (off < 0) 1405fbf59bc9STejun Heo continue; 1406ccea34b5STejun Heo 140740064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1408fbf59bc9STejun Heo if (off >= 0) 1409fbf59bc9STejun Heo goto area_found; 141040064aecSDennis Zhou (Facebook) 1411fbf59bc9STejun Heo } 1412fbf59bc9STejun Heo } 1413fbf59bc9STejun Heo 1414403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1415ccea34b5STejun Heo 1416b38d08f3STejun Heo /* 1417b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 1418b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 1419b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 1420b38d08f3STejun Heo */ 142111df02bfSDennis Zhou if (is_atomic) { 142211df02bfSDennis Zhou err = "atomic alloc failed, no space left"; 14235835d96eSTejun Heo goto fail; 142411df02bfSDennis Zhou } 14255835d96eSTejun Heo 1426b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 1427*554fef1cSDennis Zhou chunk = pcpu_create_chunk(pcpu_gfp); 1428f2badb0cSTejun Heo if (!chunk) { 1429f2badb0cSTejun Heo err = "failed to allocate new chunk"; 1430b38d08f3STejun Heo goto fail; 1431f2badb0cSTejun Heo } 1432ccea34b5STejun Heo 1433403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1434fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 1435b38d08f3STejun Heo } else { 1436b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1437b38d08f3STejun Heo } 1438b38d08f3STejun Heo 1439ccea34b5STejun Heo goto restart; 1440fbf59bc9STejun Heo 1441fbf59bc9STejun Heo area_found: 144230a5b536SDennis Zhou pcpu_stats_area_alloc(chunk, size); 1443403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1444ccea34b5STejun Heo 1445dca49645STejun Heo /* populate if not all pages are already there */ 14465835d96eSTejun Heo if (!is_atomic) { 1447e04d3208STejun Heo int page_start, page_end, rs, re; 1448e04d3208STejun Heo 1449dca49645STejun Heo page_start = PFN_DOWN(off); 1450dca49645STejun Heo page_end = PFN_UP(off + size); 1451dca49645STejun Heo 145291e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 145391e914c5SDennis Zhou (Facebook) page_start, page_end) { 1454dca49645STejun Heo WARN_ON(chunk->immutable); 1455dca49645STejun Heo 1456*554fef1cSDennis Zhou ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp); 1457b38d08f3STejun Heo 1458403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1459b38d08f3STejun Heo if (ret) { 146040064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1461f2badb0cSTejun Heo err = "failed to populate"; 1462ccea34b5STejun Heo goto fail_unlock; 1463fbf59bc9STejun Heo } 146440064aecSDennis Zhou (Facebook) pcpu_chunk_populated(chunk, rs, re, true); 1465b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1466dca49645STejun Heo } 1467dca49645STejun Heo 1468ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1469e04d3208STejun Heo } 1470ccea34b5STejun Heo 14711a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 14721a4d7607STejun Heo pcpu_schedule_balance_work(); 14731a4d7607STejun Heo 1474dca49645STejun Heo /* clear the areas and return address relative to base address */ 1475dca49645STejun Heo for_each_possible_cpu(cpu) 1476dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1477dca49645STejun Heo 1478f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 14798a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1480df95e795SDennis Zhou 1481df95e795SDennis Zhou trace_percpu_alloc_percpu(reserved, is_atomic, size, align, 1482df95e795SDennis Zhou chunk->base_addr, off, ptr); 1483df95e795SDennis Zhou 1484f528f0b8SCatalin Marinas return ptr; 1485ccea34b5STejun Heo 1486ccea34b5STejun Heo fail_unlock: 1487403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1488b38d08f3STejun Heo fail: 1489df95e795SDennis Zhou trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1490df95e795SDennis Zhou 14910ea7eeecSDaniel Borkmann if (!is_atomic && do_warn && warn_limit) { 1492870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 14935835d96eSTejun Heo size, align, is_atomic, err); 1494f2badb0cSTejun Heo dump_stack(); 1495f2badb0cSTejun Heo if (!--warn_limit) 1496870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1497f2badb0cSTejun Heo } 14981a4d7607STejun Heo if (is_atomic) { 14991a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 15001a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 15011a4d7607STejun Heo pcpu_schedule_balance_work(); 15026710e594STejun Heo } else { 15036710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 15041a4d7607STejun Heo } 1505ccea34b5STejun Heo return NULL; 1506fbf59bc9STejun Heo } 1507edcb4639STejun Heo 1508edcb4639STejun Heo /** 15095835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1510edcb4639STejun Heo * @size: size of area to allocate in bytes 1511edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 15125835d96eSTejun Heo * @gfp: allocation flags 1513edcb4639STejun Heo * 15145835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 15155835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 15160ea7eeecSDaniel Borkmann * be called from any context but is a lot more likely to fail. If @gfp 15170ea7eeecSDaniel Borkmann * has __GFP_NOWARN then no warning will be triggered on invalid or failed 15180ea7eeecSDaniel Borkmann * allocation requests. 1519ccea34b5STejun Heo * 1520edcb4639STejun Heo * RETURNS: 1521edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1522edcb4639STejun Heo */ 15235835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 15245835d96eSTejun Heo { 15255835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 15265835d96eSTejun Heo } 15275835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 15285835d96eSTejun Heo 15295835d96eSTejun Heo /** 15305835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 15315835d96eSTejun Heo * @size: size of area to allocate in bytes 15325835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 15335835d96eSTejun Heo * 15345835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 15355835d96eSTejun Heo */ 153643cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1537edcb4639STejun Heo { 15385835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1539edcb4639STejun Heo } 1540fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1541fbf59bc9STejun Heo 1542edcb4639STejun Heo /** 1543edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1544edcb4639STejun Heo * @size: size of area to allocate in bytes 1545edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1546edcb4639STejun Heo * 15479329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 15489329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 15499329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 15509329ba97STejun Heo * Might trigger writeouts. 1551edcb4639STejun Heo * 1552ccea34b5STejun Heo * CONTEXT: 1553ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1554ccea34b5STejun Heo * 1555edcb4639STejun Heo * RETURNS: 1556edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1557edcb4639STejun Heo */ 155843cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1559edcb4639STejun Heo { 15605835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1561edcb4639STejun Heo } 1562edcb4639STejun Heo 1563a56dbddfSTejun Heo /** 15641a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1565a56dbddfSTejun Heo * @work: unused 1566a56dbddfSTejun Heo * 156747504ee0SDennis Zhou * Reclaim all fully free chunks except for the first one. This is also 156847504ee0SDennis Zhou * responsible for maintaining the pool of empty populated pages. However, 156947504ee0SDennis Zhou * it is possible that this is called when physical memory is scarce causing 157047504ee0SDennis Zhou * OOM killer to be triggered. We should avoid doing so until an actual 157147504ee0SDennis Zhou * allocation causes the failure as it is possible that requests can be 157247504ee0SDennis Zhou * serviced from already backed regions. 1573a56dbddfSTejun Heo */ 1574fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1575fbf59bc9STejun Heo { 157647504ee0SDennis Zhou /* gfp flags passed to underlying allocators */ 1577*554fef1cSDennis Zhou const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN; 1578fe6bd8c3STejun Heo LIST_HEAD(to_free); 1579fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1580a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 15811a4d7607STejun Heo int slot, nr_to_pop, ret; 1582a56dbddfSTejun Heo 15831a4d7607STejun Heo /* 15841a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 15851a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 15861a4d7607STejun Heo */ 1587ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1588ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1589a56dbddfSTejun Heo 1590fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 15918d408b4bSTejun Heo WARN_ON(chunk->immutable); 1592a56dbddfSTejun Heo 1593a56dbddfSTejun Heo /* spare the first one */ 1594fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1595a56dbddfSTejun Heo continue; 1596a56dbddfSTejun Heo 1597fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1598a56dbddfSTejun Heo } 1599a56dbddfSTejun Heo 1600ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1601a56dbddfSTejun Heo 1602fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1603a93ace48STejun Heo int rs, re; 1604dca49645STejun Heo 160591e914c5SDennis Zhou (Facebook) pcpu_for_each_pop_region(chunk->populated, rs, re, 0, 160691e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1607a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1608b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1609b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1610b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1611a93ace48STejun Heo } 16126081089fSTejun Heo pcpu_destroy_chunk(chunk); 1613fbf59bc9STejun Heo } 1614971f3918STejun Heo 16151a4d7607STejun Heo /* 16161a4d7607STejun Heo * Ensure there are certain number of free populated pages for 16171a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 16181a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 16191a4d7607STejun Heo * failed previously, always populate the maximum amount. This 16201a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 16211a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 16221a4d7607STejun Heo * something we support properly and can be highly unreliable and 16231a4d7607STejun Heo * inefficient. 16241a4d7607STejun Heo */ 16251a4d7607STejun Heo retry_pop: 16261a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 16271a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 16281a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 16291a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 16301a4d7607STejun Heo } else { 16311a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 16321a4d7607STejun Heo pcpu_nr_empty_pop_pages, 16331a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 16341a4d7607STejun Heo } 16351a4d7607STejun Heo 16361a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 16371a4d7607STejun Heo int nr_unpop = 0, rs, re; 16381a4d7607STejun Heo 16391a4d7607STejun Heo if (!nr_to_pop) 16401a4d7607STejun Heo break; 16411a4d7607STejun Heo 16421a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 16431a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 16448ab16c43SDennis Zhou (Facebook) nr_unpop = chunk->nr_pages - chunk->nr_populated; 16451a4d7607STejun Heo if (nr_unpop) 16461a4d7607STejun Heo break; 16471a4d7607STejun Heo } 16481a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 16491a4d7607STejun Heo 16501a4d7607STejun Heo if (!nr_unpop) 16511a4d7607STejun Heo continue; 16521a4d7607STejun Heo 16531a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 165491e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, 165591e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 16561a4d7607STejun Heo int nr = min(re - rs, nr_to_pop); 16571a4d7607STejun Heo 165847504ee0SDennis Zhou ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); 16591a4d7607STejun Heo if (!ret) { 16601a4d7607STejun Heo nr_to_pop -= nr; 16611a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 166240064aecSDennis Zhou (Facebook) pcpu_chunk_populated(chunk, rs, rs + nr, false); 16631a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 16641a4d7607STejun Heo } else { 16651a4d7607STejun Heo nr_to_pop = 0; 16661a4d7607STejun Heo } 16671a4d7607STejun Heo 16681a4d7607STejun Heo if (!nr_to_pop) 16691a4d7607STejun Heo break; 16701a4d7607STejun Heo } 16711a4d7607STejun Heo } 16721a4d7607STejun Heo 16731a4d7607STejun Heo if (nr_to_pop) { 16741a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 167547504ee0SDennis Zhou chunk = pcpu_create_chunk(gfp); 16761a4d7607STejun Heo if (chunk) { 16771a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 16781a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 16791a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 16801a4d7607STejun Heo goto retry_pop; 16811a4d7607STejun Heo } 16821a4d7607STejun Heo } 16831a4d7607STejun Heo 1684971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1685a56dbddfSTejun Heo } 1686fbf59bc9STejun Heo 1687fbf59bc9STejun Heo /** 1688fbf59bc9STejun Heo * free_percpu - free percpu area 1689fbf59bc9STejun Heo * @ptr: pointer to area to free 1690fbf59bc9STejun Heo * 1691ccea34b5STejun Heo * Free percpu area @ptr. 1692ccea34b5STejun Heo * 1693ccea34b5STejun Heo * CONTEXT: 1694ccea34b5STejun Heo * Can be called from atomic context. 1695fbf59bc9STejun Heo */ 169643cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1697fbf59bc9STejun Heo { 1698129182e5SAndrew Morton void *addr; 1699fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1700ccea34b5STejun Heo unsigned long flags; 170140064aecSDennis Zhou (Facebook) int off; 1702fbf59bc9STejun Heo 1703fbf59bc9STejun Heo if (!ptr) 1704fbf59bc9STejun Heo return; 1705fbf59bc9STejun Heo 1706f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1707f528f0b8SCatalin Marinas 1708129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1709129182e5SAndrew Morton 1710ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1711fbf59bc9STejun Heo 1712fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1713bba174f5STejun Heo off = addr - chunk->base_addr; 1714fbf59bc9STejun Heo 171540064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1716fbf59bc9STejun Heo 1717a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 171840064aecSDennis Zhou (Facebook) if (chunk->free_bytes == pcpu_unit_size) { 1719fbf59bc9STejun Heo struct pcpu_chunk *pos; 1720fbf59bc9STejun Heo 1721a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1722fbf59bc9STejun Heo if (pos != chunk) { 17231a4d7607STejun Heo pcpu_schedule_balance_work(); 1724fbf59bc9STejun Heo break; 1725fbf59bc9STejun Heo } 1726fbf59bc9STejun Heo } 1727fbf59bc9STejun Heo 1728df95e795SDennis Zhou trace_percpu_free_percpu(chunk->base_addr, off, ptr); 1729df95e795SDennis Zhou 1730ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1731fbf59bc9STejun Heo } 1732fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1733fbf59bc9STejun Heo 1734383776faSThomas Gleixner bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 1735383776faSThomas Gleixner { 1736383776faSThomas Gleixner #ifdef CONFIG_SMP 1737383776faSThomas Gleixner const size_t static_size = __per_cpu_end - __per_cpu_start; 1738383776faSThomas Gleixner void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 1739383776faSThomas Gleixner unsigned int cpu; 1740383776faSThomas Gleixner 1741383776faSThomas Gleixner for_each_possible_cpu(cpu) { 1742383776faSThomas Gleixner void *start = per_cpu_ptr(base, cpu); 1743383776faSThomas Gleixner void *va = (void *)addr; 1744383776faSThomas Gleixner 1745383776faSThomas Gleixner if (va >= start && va < start + static_size) { 17468ce371f9SPeter Zijlstra if (can_addr) { 1747383776faSThomas Gleixner *can_addr = (unsigned long) (va - start); 17488ce371f9SPeter Zijlstra *can_addr += (unsigned long) 17498ce371f9SPeter Zijlstra per_cpu_ptr(base, get_boot_cpu_id()); 17508ce371f9SPeter Zijlstra } 1751383776faSThomas Gleixner return true; 1752383776faSThomas Gleixner } 1753383776faSThomas Gleixner } 1754383776faSThomas Gleixner #endif 1755383776faSThomas Gleixner /* on UP, can't distinguish from other static vars, always false */ 1756383776faSThomas Gleixner return false; 1757383776faSThomas Gleixner } 1758383776faSThomas Gleixner 17593b034b0dSVivek Goyal /** 176010fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 176110fad5e4STejun Heo * @addr: address to test 176210fad5e4STejun Heo * 176310fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 176410fad5e4STejun Heo * static percpu areas are not considered. For those, use 176510fad5e4STejun Heo * is_module_percpu_address(). 176610fad5e4STejun Heo * 176710fad5e4STejun Heo * RETURNS: 176810fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 176910fad5e4STejun Heo */ 177010fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 177110fad5e4STejun Heo { 1772383776faSThomas Gleixner return __is_kernel_percpu_address(addr, NULL); 177310fad5e4STejun Heo } 177410fad5e4STejun Heo 177510fad5e4STejun Heo /** 17763b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 17773b034b0dSVivek Goyal * @addr: the address to be converted to physical address 17783b034b0dSVivek Goyal * 17793b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 17803b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 17813b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 17823b034b0dSVivek Goyal * until this function finishes. 17833b034b0dSVivek Goyal * 178467589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 178567589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 178667589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 178767589c71SDave Young * km) provides translation. 178867589c71SDave Young * 1789bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 179067589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 179167589c71SDave Young * actually works, and the verification can discover both bugs in percpu 179267589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 179367589c71SDave Young * code. 179467589c71SDave Young * 17953b034b0dSVivek Goyal * RETURNS: 17963b034b0dSVivek Goyal * The physical address for @addr. 17973b034b0dSVivek Goyal */ 17983b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 17993b034b0dSVivek Goyal { 18009983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 18019983b6f0STejun Heo bool in_first_chunk = false; 1802a855b84cSTejun Heo unsigned long first_low, first_high; 18039983b6f0STejun Heo unsigned int cpu; 18049983b6f0STejun Heo 18059983b6f0STejun Heo /* 1806a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 18079983b6f0STejun Heo * necessary but will speed up lookups of addresses which 18089983b6f0STejun Heo * aren't in the first chunk. 1809c0ebfdc3SDennis Zhou (Facebook) * 1810c0ebfdc3SDennis Zhou (Facebook) * The address check is against full chunk sizes. pcpu_base_addr 1811c0ebfdc3SDennis Zhou (Facebook) * points to the beginning of the first chunk including the 1812c0ebfdc3SDennis Zhou (Facebook) * static region. Assumes good intent as the first chunk may 1813c0ebfdc3SDennis Zhou (Facebook) * not be full (ie. < pcpu_unit_pages in size). 18149983b6f0STejun Heo */ 1815c0ebfdc3SDennis Zhou (Facebook) first_low = (unsigned long)pcpu_base_addr + 1816c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 1817c0ebfdc3SDennis Zhou (Facebook) first_high = (unsigned long)pcpu_base_addr + 1818c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 1819a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1820a855b84cSTejun Heo (unsigned long)addr < first_high) { 18219983b6f0STejun Heo for_each_possible_cpu(cpu) { 18229983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 18239983b6f0STejun Heo 18249983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 18259983b6f0STejun Heo in_first_chunk = true; 18269983b6f0STejun Heo break; 18279983b6f0STejun Heo } 18289983b6f0STejun Heo } 18299983b6f0STejun Heo } 18309983b6f0STejun Heo 18319983b6f0STejun Heo if (in_first_chunk) { 1832eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 18333b034b0dSVivek Goyal return __pa(addr); 18343b034b0dSVivek Goyal else 18359f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 18369f57bd4dSEugene Surovegin offset_in_page(addr); 1837020ec653STejun Heo } else 18389f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 18399f57bd4dSEugene Surovegin offset_in_page(addr); 18403b034b0dSVivek Goyal } 18413b034b0dSVivek Goyal 1842fbf59bc9STejun Heo /** 1843fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1844fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1845fd1e8a1fSTejun Heo * @nr_units: the number of units 1846033e48fbSTejun Heo * 1847fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1848fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1849fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1850fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1851fd1e8a1fSTejun Heo * pointer of other groups. 1852033e48fbSTejun Heo * 1853033e48fbSTejun Heo * RETURNS: 1854fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1855fd1e8a1fSTejun Heo * failure. 1856033e48fbSTejun Heo */ 1857fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1858fd1e8a1fSTejun Heo int nr_units) 1859fd1e8a1fSTejun Heo { 1860fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 1861fd1e8a1fSTejun Heo size_t base_size, ai_size; 1862fd1e8a1fSTejun Heo void *ptr; 1863fd1e8a1fSTejun Heo int unit; 1864fd1e8a1fSTejun Heo 1865fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1866fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 1867fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1868fd1e8a1fSTejun Heo 1869438a5061SNicolas Pitre ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), PAGE_SIZE); 1870fd1e8a1fSTejun Heo if (!ptr) 1871fd1e8a1fSTejun Heo return NULL; 1872fd1e8a1fSTejun Heo ai = ptr; 1873fd1e8a1fSTejun Heo ptr += base_size; 1874fd1e8a1fSTejun Heo 1875fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 1876fd1e8a1fSTejun Heo 1877fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 1878fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 1879fd1e8a1fSTejun Heo 1880fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 1881fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 1882fd1e8a1fSTejun Heo 1883fd1e8a1fSTejun Heo return ai; 1884fd1e8a1fSTejun Heo } 1885fd1e8a1fSTejun Heo 1886fd1e8a1fSTejun Heo /** 1887fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 1888fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 1889fd1e8a1fSTejun Heo * 1890fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1891fd1e8a1fSTejun Heo */ 1892fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1893fd1e8a1fSTejun Heo { 1894999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 1895fd1e8a1fSTejun Heo } 1896fd1e8a1fSTejun Heo 1897fd1e8a1fSTejun Heo /** 1898fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1899fd1e8a1fSTejun Heo * @lvl: loglevel 1900fd1e8a1fSTejun Heo * @ai: allocation info to dump 1901fd1e8a1fSTejun Heo * 1902fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 1903fd1e8a1fSTejun Heo */ 1904fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 1905fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 1906033e48fbSTejun Heo { 1907fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 1908033e48fbSTejun Heo char empty_str[] = "--------"; 1909fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 1910fd1e8a1fSTejun Heo int group, v; 1911fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 1912033e48fbSTejun Heo 1913fd1e8a1fSTejun Heo v = ai->nr_groups; 1914033e48fbSTejun Heo while (v /= 10) 1915fd1e8a1fSTejun Heo group_width++; 1916033e48fbSTejun Heo 1917fd1e8a1fSTejun Heo v = num_possible_cpus(); 1918fd1e8a1fSTejun Heo while (v /= 10) 1919fd1e8a1fSTejun Heo cpu_width++; 1920fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1921033e48fbSTejun Heo 1922fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 1923fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 1924fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 1925033e48fbSTejun Heo 1926fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1927fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1928fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1929fd1e8a1fSTejun Heo 1930fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 1931fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 1932fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 1933fd1e8a1fSTejun Heo 1934fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 1935fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 1936fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 1937fd1e8a1fSTejun Heo if (!(alloc % apl)) { 19381170532bSJoe Perches pr_cont("\n"); 1939fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 1940033e48fbSTejun Heo } 19411170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 1942fd1e8a1fSTejun Heo 1943fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 1944fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 19451170532bSJoe Perches pr_cont("%0*d ", 19461170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 1947033e48fbSTejun Heo else 19481170532bSJoe Perches pr_cont("%s ", empty_str); 1949033e48fbSTejun Heo } 1950fd1e8a1fSTejun Heo } 19511170532bSJoe Perches pr_cont("\n"); 1952033e48fbSTejun Heo } 1953033e48fbSTejun Heo 1954fbf59bc9STejun Heo /** 19558d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 1956fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 195738a6be52STejun Heo * @base_addr: mapped address 1958fbf59bc9STejun Heo * 19598d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 19608d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 196138a6be52STejun Heo * setup path. 19628d408b4bSTejun Heo * 1963fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 1964fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 19658d408b4bSTejun Heo * 1966fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 1967fd1e8a1fSTejun Heo * 1968fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1969edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 1970edcb4639STejun Heo * the first chunk such that it's available only through reserved 1971edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 1972edcb4639STejun Heo * static areas on architectures where the addressing model has 1973edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 1974edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 1975edcb4639STejun Heo * 1976fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 1977fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 1978fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 19796074d5b0STejun Heo * 1980fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1981fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 1982fd1e8a1fSTejun Heo * @ai->dyn_size. 19838d408b4bSTejun Heo * 1984fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 1985fd1e8a1fSTejun Heo * for vm areas. 19868d408b4bSTejun Heo * 1987fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 1988fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 1989fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 1990fd1e8a1fSTejun Heo * 1991fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 1992fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 1993fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 1994fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 1995fd1e8a1fSTejun Heo * all units is assumed. 19968d408b4bSTejun Heo * 199738a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 199838a6be52STejun Heo * copied static data to each unit. 1999fbf59bc9STejun Heo * 2000c0ebfdc3SDennis Zhou (Facebook) * The first chunk will always contain a static and a dynamic region. 2001c0ebfdc3SDennis Zhou (Facebook) * However, the static region is not managed by any chunk. If the first 2002c0ebfdc3SDennis Zhou (Facebook) * chunk also contains a reserved region, it is served by two chunks - 2003c0ebfdc3SDennis Zhou (Facebook) * one for the reserved region and one for the dynamic region. They 2004c0ebfdc3SDennis Zhou (Facebook) * share the same vm, but use offset regions in the area allocation map. 2005c0ebfdc3SDennis Zhou (Facebook) * The chunk serving the dynamic region is circulated in the chunk slots 2006c0ebfdc3SDennis Zhou (Facebook) * and available for dynamic allocation like any other chunk. 2007edcb4639STejun Heo * 2008fbf59bc9STejun Heo * RETURNS: 2009fb435d52STejun Heo * 0 on success, -errno on failure. 2010fbf59bc9STejun Heo */ 2011fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 2012fd1e8a1fSTejun Heo void *base_addr) 2013fbf59bc9STejun Heo { 2014b9c39442SDennis Zhou (Facebook) size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2015d2f3c384SDennis Zhou (Facebook) size_t static_size, dyn_size; 20160c4169c3SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 20176563297cSTejun Heo unsigned long *group_offsets; 20186563297cSTejun Heo size_t *group_sizes; 2019fb435d52STejun Heo unsigned long *unit_off; 2020fbf59bc9STejun Heo unsigned int cpu; 2021fd1e8a1fSTejun Heo int *unit_map; 2022fd1e8a1fSTejun Heo int group, unit, i; 2023c0ebfdc3SDennis Zhou (Facebook) int map_size; 2024c0ebfdc3SDennis Zhou (Facebook) unsigned long tmp_addr; 2025fbf59bc9STejun Heo 2026635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 2027635b75fcSTejun Heo if (unlikely(cond)) { \ 2028870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 2029870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 2030807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 2031635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 2032635b75fcSTejun Heo BUG(); \ 2033635b75fcSTejun Heo } \ 2034635b75fcSTejun Heo } while (0) 2035635b75fcSTejun Heo 20362f39e637STejun Heo /* sanity checks */ 2037635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 2038bbddff05STejun Heo #ifdef CONFIG_SMP 2039635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 2040f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 2041bbddff05STejun Heo #endif 2042635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 2043f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 2044635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 2045f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 2046635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 2047ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 2048099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 2049fb29a2ccSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!ai->dyn_size); 2050d2f3c384SDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 2051ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 2052ca460b3cSDennis Zhou (Facebook) IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 20539f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 20548d408b4bSTejun Heo 20556563297cSTejun Heo /* process group information and build config tables accordingly */ 2056999c17e3SSantosh Shilimkar group_offsets = memblock_virt_alloc(ai->nr_groups * 2057999c17e3SSantosh Shilimkar sizeof(group_offsets[0]), 0); 2058999c17e3SSantosh Shilimkar group_sizes = memblock_virt_alloc(ai->nr_groups * 2059999c17e3SSantosh Shilimkar sizeof(group_sizes[0]), 0); 2060999c17e3SSantosh Shilimkar unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); 2061999c17e3SSantosh Shilimkar unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); 20622f39e637STejun Heo 2063fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 2064ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 2065a855b84cSTejun Heo 2066a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 2067a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 20682f39e637STejun Heo 2069fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 2070fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 20712f39e637STejun Heo 20726563297cSTejun Heo group_offsets[group] = gi->base_offset; 20736563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 20746563297cSTejun Heo 2075fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 2076fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 2077fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 2078fd1e8a1fSTejun Heo continue; 2079fd1e8a1fSTejun Heo 20809f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 2081635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 2082635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 2083fd1e8a1fSTejun Heo 2084fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 2085fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 2086fb435d52STejun Heo 2087a855b84cSTejun Heo /* determine low/high unit_cpu */ 2088a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 2089a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 2090a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 2091a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 2092a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 2093a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 20940fc0531eSLinus Torvalds } 20950fc0531eSLinus Torvalds } 2096fd1e8a1fSTejun Heo pcpu_nr_units = unit; 20972f39e637STejun Heo 20982f39e637STejun Heo for_each_possible_cpu(cpu) 2099635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 2100635b75fcSTejun Heo 2101635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 2102635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 2103bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 21042f39e637STejun Heo 21056563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 21066563297cSTejun Heo pcpu_group_offsets = group_offsets; 21076563297cSTejun Heo pcpu_group_sizes = group_sizes; 2108fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 2109fb435d52STejun Heo pcpu_unit_offsets = unit_off; 21102f39e637STejun Heo 21112f39e637STejun Heo /* determine basic parameters */ 2112fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 2113d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 21146563297cSTejun Heo pcpu_atom_size = ai->atom_size; 2115ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 2116ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 2117cafe8816STejun Heo 211830a5b536SDennis Zhou pcpu_stats_save_ai(ai); 211930a5b536SDennis Zhou 2120d9b55eebSTejun Heo /* 2121d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 2122d9b55eebSTejun Heo * empty chunks. 2123d9b55eebSTejun Heo */ 2124d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 2125999c17e3SSantosh Shilimkar pcpu_slot = memblock_virt_alloc( 2126999c17e3SSantosh Shilimkar pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); 2127fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 2128fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 2129fbf59bc9STejun Heo 2130edcb4639STejun Heo /* 2131d2f3c384SDennis Zhou (Facebook) * The end of the static region needs to be aligned with the 2132d2f3c384SDennis Zhou (Facebook) * minimum allocation size as this offsets the reserved and 2133d2f3c384SDennis Zhou (Facebook) * dynamic region. The first chunk ends page aligned by 2134d2f3c384SDennis Zhou (Facebook) * expanding the dynamic region, therefore the dynamic region 2135d2f3c384SDennis Zhou (Facebook) * can be shrunk to compensate while still staying above the 2136d2f3c384SDennis Zhou (Facebook) * configured sizes. 2137d2f3c384SDennis Zhou (Facebook) */ 2138d2f3c384SDennis Zhou (Facebook) static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 2139d2f3c384SDennis Zhou (Facebook) dyn_size = ai->dyn_size - (static_size - ai->static_size); 2140d2f3c384SDennis Zhou (Facebook) 2141d2f3c384SDennis Zhou (Facebook) /* 2142c0ebfdc3SDennis Zhou (Facebook) * Initialize first chunk. 2143c0ebfdc3SDennis Zhou (Facebook) * If the reserved_size is non-zero, this initializes the reserved 2144c0ebfdc3SDennis Zhou (Facebook) * chunk. If the reserved_size is zero, the reserved chunk is NULL 2145c0ebfdc3SDennis Zhou (Facebook) * and the dynamic region is initialized here. The first chunk, 2146c0ebfdc3SDennis Zhou (Facebook) * pcpu_first_chunk, will always point to the chunk that serves 2147c0ebfdc3SDennis Zhou (Facebook) * the dynamic region. 2148edcb4639STejun Heo */ 2149d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size; 2150d2f3c384SDennis Zhou (Facebook) map_size = ai->reserved_size ?: dyn_size; 215140064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 215261ace7faSTejun Heo 2153edcb4639STejun Heo /* init dynamic chunk if necessary */ 2154b9c39442SDennis Zhou (Facebook) if (ai->reserved_size) { 21550c4169c3SDennis Zhou (Facebook) pcpu_reserved_chunk = chunk; 2156b9c39442SDennis Zhou (Facebook) 2157d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size + 2158c0ebfdc3SDennis Zhou (Facebook) ai->reserved_size; 2159d2f3c384SDennis Zhou (Facebook) map_size = dyn_size; 216040064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 2161edcb4639STejun Heo } 2162edcb4639STejun Heo 21632441d15cSTejun Heo /* link the first chunk in */ 21640c4169c3SDennis Zhou (Facebook) pcpu_first_chunk = chunk; 21650cecf50cSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 2166ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 2167fbf59bc9STejun Heo 216830a5b536SDennis Zhou pcpu_stats_chunk_alloc(); 2169df95e795SDennis Zhou trace_percpu_create_chunk(base_addr); 217030a5b536SDennis Zhou 2171fbf59bc9STejun Heo /* we're done */ 2172bba174f5STejun Heo pcpu_base_addr = base_addr; 2173fb435d52STejun Heo return 0; 2174fbf59bc9STejun Heo } 217566c3a757STejun Heo 2176bbddff05STejun Heo #ifdef CONFIG_SMP 2177bbddff05STejun Heo 217817f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 2179f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 2180f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 2181f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 2182f58dc01bSTejun Heo }; 218366c3a757STejun Heo 2184f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 2185f58dc01bSTejun Heo 2186f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 218766c3a757STejun Heo { 21885479c78aSCyrill Gorcunov if (!str) 21895479c78aSCyrill Gorcunov return -EINVAL; 21905479c78aSCyrill Gorcunov 2191f58dc01bSTejun Heo if (0) 2192f58dc01bSTejun Heo /* nada */; 2193f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 2194f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 2195f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 2196f58dc01bSTejun Heo #endif 2197f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2198f58dc01bSTejun Heo else if (!strcmp(str, "page")) 2199f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 2200f58dc01bSTejun Heo #endif 2201f58dc01bSTejun Heo else 2202870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 220366c3a757STejun Heo 2204f58dc01bSTejun Heo return 0; 220566c3a757STejun Heo } 2206f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 220766c3a757STejun Heo 22083c9a024fSTejun Heo /* 22093c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 22103c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 22113c9a024fSTejun Heo * to be used. 22123c9a024fSTejun Heo */ 221308fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 221408fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 22153c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 22163c9a024fSTejun Heo #endif 22173c9a024fSTejun Heo 22183c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 22193c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 22203c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 22213c9a024fSTejun Heo #endif 22223c9a024fSTejun Heo 22233c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 22243c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 22253c9a024fSTejun Heo /** 2226fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2227fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2228fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2229fbf59bc9STejun Heo * @atom_size: allocation atom size 2230fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2231fbf59bc9STejun Heo * 2232fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 2233fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 2234fbf59bc9STejun Heo * atom size and distances between CPUs. 2235fbf59bc9STejun Heo * 2236bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 2237fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 2238fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 2239fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 2240fbf59bc9STejun Heo * of allocated virtual address space. 2241fbf59bc9STejun Heo * 2242fbf59bc9STejun Heo * RETURNS: 2243fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 2244fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 2245fbf59bc9STejun Heo */ 2246fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 2247fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 2248fbf59bc9STejun Heo size_t atom_size, 2249fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2250fbf59bc9STejun Heo { 2251fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 2252fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 2253fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 2254fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 2255fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 2256fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 2257fbf59bc9STejun Heo int last_allocs, group, unit; 2258fbf59bc9STejun Heo unsigned int cpu, tcpu; 2259fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 2260fbf59bc9STejun Heo unsigned int *cpu_map; 2261fbf59bc9STejun Heo 2262fbf59bc9STejun Heo /* this function may be called multiple times */ 2263fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 2264fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 2265fbf59bc9STejun Heo 2266fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2267fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 2268fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2269fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 2270fbf59bc9STejun Heo 2271fbf59bc9STejun Heo /* 2272fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 2273fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 227425985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 2275fbf59bc9STejun Heo * or larger than min_unit_size. 2276fbf59bc9STejun Heo */ 2277fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2278fbf59bc9STejun Heo 22799c015162SDennis Zhou (Facebook) /* determine the maximum # of units that can fit in an allocation */ 2280fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 2281fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 2282f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2283fbf59bc9STejun Heo upa--; 2284fbf59bc9STejun Heo max_upa = upa; 2285fbf59bc9STejun Heo 2286fbf59bc9STejun Heo /* group cpus according to their proximity */ 2287fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 2288fbf59bc9STejun Heo group = 0; 2289fbf59bc9STejun Heo next_group: 2290fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 2291fbf59bc9STejun Heo if (cpu == tcpu) 2292fbf59bc9STejun Heo break; 2293fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 2294fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 2295fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 2296fbf59bc9STejun Heo group++; 2297fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 2298fbf59bc9STejun Heo goto next_group; 2299fbf59bc9STejun Heo } 2300fbf59bc9STejun Heo } 2301fbf59bc9STejun Heo group_map[cpu] = group; 2302fbf59bc9STejun Heo group_cnt[group]++; 2303fbf59bc9STejun Heo } 2304fbf59bc9STejun Heo 2305fbf59bc9STejun Heo /* 23069c015162SDennis Zhou (Facebook) * Wasted space is caused by a ratio imbalance of upa to group_cnt. 23079c015162SDennis Zhou (Facebook) * Expand the unit_size until we use >= 75% of the units allocated. 23089c015162SDennis Zhou (Facebook) * Related to atom_size, which could be much larger than the unit_size. 2309fbf59bc9STejun Heo */ 2310fbf59bc9STejun Heo last_allocs = INT_MAX; 2311fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 2312fbf59bc9STejun Heo int allocs = 0, wasted = 0; 2313fbf59bc9STejun Heo 2314f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2315fbf59bc9STejun Heo continue; 2316fbf59bc9STejun Heo 2317fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2318fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2319fbf59bc9STejun Heo allocs += this_allocs; 2320fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 2321fbf59bc9STejun Heo } 2322fbf59bc9STejun Heo 2323fbf59bc9STejun Heo /* 2324fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 2325fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 2326fbf59bc9STejun Heo * passes the following check. 2327fbf59bc9STejun Heo */ 2328fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 2329fbf59bc9STejun Heo continue; 2330fbf59bc9STejun Heo 2331fbf59bc9STejun Heo /* and then don't consume more memory */ 2332fbf59bc9STejun Heo if (allocs > last_allocs) 2333fbf59bc9STejun Heo break; 2334fbf59bc9STejun Heo last_allocs = allocs; 2335fbf59bc9STejun Heo best_upa = upa; 2336fbf59bc9STejun Heo } 2337fbf59bc9STejun Heo upa = best_upa; 2338fbf59bc9STejun Heo 2339fbf59bc9STejun Heo /* allocate and fill alloc_info */ 2340fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 2341fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 2342fbf59bc9STejun Heo 2343fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2344fbf59bc9STejun Heo if (!ai) 2345fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 2346fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 2347fbf59bc9STejun Heo 2348fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2349fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 2350fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 2351fbf59bc9STejun Heo } 2352fbf59bc9STejun Heo 2353fbf59bc9STejun Heo ai->static_size = static_size; 2354fbf59bc9STejun Heo ai->reserved_size = reserved_size; 2355fbf59bc9STejun Heo ai->dyn_size = dyn_size; 2356fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 2357fbf59bc9STejun Heo ai->atom_size = atom_size; 2358fbf59bc9STejun Heo ai->alloc_size = alloc_size; 2359fbf59bc9STejun Heo 2360fbf59bc9STejun Heo for (group = 0, unit = 0; group_cnt[group]; group++) { 2361fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2362fbf59bc9STejun Heo 2363fbf59bc9STejun Heo /* 2364fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 2365fbf59bc9STejun Heo * back-to-back. The caller should update this to 2366fbf59bc9STejun Heo * reflect actual allocation. 2367fbf59bc9STejun Heo */ 2368fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 2369fbf59bc9STejun Heo 2370fbf59bc9STejun Heo for_each_possible_cpu(cpu) 2371fbf59bc9STejun Heo if (group_map[cpu] == group) 2372fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 2373fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 2374fbf59bc9STejun Heo unit += gi->nr_units; 2375fbf59bc9STejun Heo } 2376fbf59bc9STejun Heo BUG_ON(unit != nr_units); 2377fbf59bc9STejun Heo 2378fbf59bc9STejun Heo return ai; 2379fbf59bc9STejun Heo } 23803c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2381fbf59bc9STejun Heo 23823c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 238366c3a757STejun Heo /** 238466c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 238566c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 23864ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2387c8826dd5STejun Heo * @atom_size: allocation atom size 2388c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2389c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 239025985edcSLucas De Marchi * @free_fn: function to free percpu page 239166c3a757STejun Heo * 239266c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 239366c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 239466c3a757STejun Heo * 239566c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 2396c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 2397c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 2398c8826dd5STejun Heo * aligned to @atom_size. 2399c8826dd5STejun Heo * 2400c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 2401c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 2402c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 2403c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 2404c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 2405c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 240666c3a757STejun Heo * 24074ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 240866c3a757STejun Heo * 240966c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 2410c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 241166c3a757STejun Heo * 241266c3a757STejun Heo * RETURNS: 2413fb435d52STejun Heo * 0 on success, -errno on failure. 241466c3a757STejun Heo */ 24154ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 2416c8826dd5STejun Heo size_t atom_size, 2417c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 2418c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2419c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 242066c3a757STejun Heo { 2421c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 2422c8826dd5STejun Heo void **areas = NULL; 2423fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 242493c76b6bSzijun_hu size_t size_sum, areas_size; 242593c76b6bSzijun_hu unsigned long max_distance; 24269b739662Szijun_hu int group, i, highest_group, rc; 242766c3a757STejun Heo 2428c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 2429c8826dd5STejun Heo cpu_distance_fn); 2430fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2431fd1e8a1fSTejun Heo return PTR_ERR(ai); 243266c3a757STejun Heo 2433fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2434c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 243566c3a757STejun Heo 2436999c17e3SSantosh Shilimkar areas = memblock_virt_alloc_nopanic(areas_size, 0); 2437c8826dd5STejun Heo if (!areas) { 2438fb435d52STejun Heo rc = -ENOMEM; 2439c8826dd5STejun Heo goto out_free; 2440fa8a7094STejun Heo } 244166c3a757STejun Heo 24429b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 24439b739662Szijun_hu highest_group = 0; 2444c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2445c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2446c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 2447c8826dd5STejun Heo void *ptr; 244866c3a757STejun Heo 2449c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 2450c8826dd5STejun Heo cpu = gi->cpu_map[i]; 2451c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 2452c8826dd5STejun Heo 2453c8826dd5STejun Heo /* allocate space for the whole group */ 2454c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 2455c8826dd5STejun Heo if (!ptr) { 2456c8826dd5STejun Heo rc = -ENOMEM; 2457c8826dd5STejun Heo goto out_free_areas; 2458c8826dd5STejun Heo } 2459f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2460f528f0b8SCatalin Marinas kmemleak_free(ptr); 2461c8826dd5STejun Heo areas[group] = ptr; 2462c8826dd5STejun Heo 2463c8826dd5STejun Heo base = min(ptr, base); 24649b739662Szijun_hu if (ptr > areas[highest_group]) 24659b739662Szijun_hu highest_group = group; 24669b739662Szijun_hu } 24679b739662Szijun_hu max_distance = areas[highest_group] - base; 24689b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 24699b739662Szijun_hu 24709b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 24719b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 24729b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 24739b739662Szijun_hu max_distance, VMALLOC_TOTAL); 24749b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 24759b739662Szijun_hu /* and fail if we have fallback */ 24769b739662Szijun_hu rc = -EINVAL; 24779b739662Szijun_hu goto out_free_areas; 24789b739662Szijun_hu #endif 247942b64281STejun Heo } 248042b64281STejun Heo 248142b64281STejun Heo /* 248242b64281STejun Heo * Copy data and free unused parts. This should happen after all 248342b64281STejun Heo * allocations are complete; otherwise, we may end up with 248442b64281STejun Heo * overlapping groups. 248542b64281STejun Heo */ 248642b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 248742b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 248842b64281STejun Heo void *ptr = areas[group]; 2489c8826dd5STejun Heo 2490c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2491c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2492c8826dd5STejun Heo /* unused unit, free whole */ 2493c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2494c8826dd5STejun Heo continue; 2495c8826dd5STejun Heo } 2496c8826dd5STejun Heo /* copy and return the unused part */ 2497fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2498c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2499c8826dd5STejun Heo } 250066c3a757STejun Heo } 250166c3a757STejun Heo 2502c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 25036ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2504c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 25056ea529a2STejun Heo } 2506c8826dd5STejun Heo 2507870d4b12SJoe Perches pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 2508fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 2509fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 251066c3a757STejun Heo 2511fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 2512c8826dd5STejun Heo goto out_free; 2513c8826dd5STejun Heo 2514c8826dd5STejun Heo out_free_areas: 2515c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2516f851c8d8SMichael Holzheu if (areas[group]) 2517c8826dd5STejun Heo free_fn(areas[group], 2518c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2519c8826dd5STejun Heo out_free: 2520fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2521c8826dd5STejun Heo if (areas) 2522999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2523fb435d52STejun Heo return rc; 2524d4b95f80STejun Heo } 25253c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2526d4b95f80STejun Heo 25273c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2528d4b95f80STejun Heo /** 252900ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2530d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2531d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 253225985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2533d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2534d4b95f80STejun Heo * 253500ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 253600ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2537d4b95f80STejun Heo * 2538d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2539d4b95f80STejun Heo * page-by-page into vmalloc area. 2540d4b95f80STejun Heo * 2541d4b95f80STejun Heo * RETURNS: 2542fb435d52STejun Heo * 0 on success, -errno on failure. 2543d4b95f80STejun Heo */ 2544fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2545d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2546d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2547d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2548d4b95f80STejun Heo { 25498f05a6a6STejun Heo static struct vm_struct vm; 2550fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 255100ae4064STejun Heo char psize_str[16]; 2552ce3141a2STejun Heo int unit_pages; 2553d4b95f80STejun Heo size_t pages_size; 2554ce3141a2STejun Heo struct page **pages; 2555fb435d52STejun Heo int unit, i, j, rc; 25568f606604Szijun_hu int upa; 25578f606604Szijun_hu int nr_g0_units; 2558d4b95f80STejun Heo 255900ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 256000ae4064STejun Heo 25614ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2562fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2563fd1e8a1fSTejun Heo return PTR_ERR(ai); 2564fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 25658f606604Szijun_hu upa = ai->alloc_size/ai->unit_size; 25668f606604Szijun_hu nr_g0_units = roundup(num_possible_cpus(), upa); 25678f606604Szijun_hu if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { 25688f606604Szijun_hu pcpu_free_alloc_info(ai); 25698f606604Szijun_hu return -EINVAL; 25708f606604Szijun_hu } 2571fd1e8a1fSTejun Heo 2572fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2573d4b95f80STejun Heo 2574d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2575fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2576fd1e8a1fSTejun Heo sizeof(pages[0])); 2577999c17e3SSantosh Shilimkar pages = memblock_virt_alloc(pages_size, 0); 2578d4b95f80STejun Heo 25798f05a6a6STejun Heo /* allocate pages */ 2580d4b95f80STejun Heo j = 0; 25818f606604Szijun_hu for (unit = 0; unit < num_possible_cpus(); unit++) { 2582fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 25838f606604Szijun_hu for (i = 0; i < unit_pages; i++) { 2584d4b95f80STejun Heo void *ptr; 2585d4b95f80STejun Heo 25863cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2587d4b95f80STejun Heo if (!ptr) { 2588870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2589598d8091SJoe Perches psize_str, cpu); 2590d4b95f80STejun Heo goto enomem; 2591d4b95f80STejun Heo } 2592f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2593f528f0b8SCatalin Marinas kmemleak_free(ptr); 2594ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2595d4b95f80STejun Heo } 25968f606604Szijun_hu } 2597d4b95f80STejun Heo 25988f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 25998f05a6a6STejun Heo vm.flags = VM_ALLOC; 2600fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 26018f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 26028f05a6a6STejun Heo 2603fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 26041d9d3257STejun Heo unsigned long unit_addr = 2605fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 26068f05a6a6STejun Heo 2607ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 26088f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 26098f05a6a6STejun Heo 26108f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2611fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2612ce3141a2STejun Heo unit_pages); 2613fb435d52STejun Heo if (rc < 0) 2614fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 26158f05a6a6STejun Heo 26168f05a6a6STejun Heo /* 26178f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 26188f05a6a6STejun Heo * cache for the linear mapping here - something 26198f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 26208f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 26218f05a6a6STejun Heo * data structures are not set up yet. 26228f05a6a6STejun Heo */ 26238f05a6a6STejun Heo 26248f05a6a6STejun Heo /* copy static data */ 2625fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 262666c3a757STejun Heo } 262766c3a757STejun Heo 262866c3a757STejun Heo /* we're ready, commit */ 2629870d4b12SJoe Perches pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", 2630fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 2631fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 263266c3a757STejun Heo 2633fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 2634d4b95f80STejun Heo goto out_free_ar; 2635d4b95f80STejun Heo 2636d4b95f80STejun Heo enomem: 2637d4b95f80STejun Heo while (--j >= 0) 2638ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2639fb435d52STejun Heo rc = -ENOMEM; 2640d4b95f80STejun Heo out_free_ar: 2641999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2642fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2643fb435d52STejun Heo return rc; 264466c3a757STejun Heo } 26453c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2646d4b95f80STejun Heo 2647bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 26488c4bfc6eSTejun Heo /* 2649bbddff05STejun Heo * Generic SMP percpu area setup. 2650e74e3962STejun Heo * 2651e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2652e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2653e74e3962STejun Heo * important because many archs have addressing restrictions and might 2654e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2655e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2656e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2657e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2658e74e3962STejun Heo * mappings on applicable archs. 2659e74e3962STejun Heo */ 2660e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2661e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2662e74e3962STejun Heo 2663c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2664c8826dd5STejun Heo size_t align) 2665c8826dd5STejun Heo { 2666999c17e3SSantosh Shilimkar return memblock_virt_alloc_from_nopanic( 2667999c17e3SSantosh Shilimkar size, align, __pa(MAX_DMA_ADDRESS)); 2668c8826dd5STejun Heo } 2669c8826dd5STejun Heo 2670c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2671c8826dd5STejun Heo { 2672999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2673c8826dd5STejun Heo } 2674c8826dd5STejun Heo 2675e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2676e74e3962STejun Heo { 2677e74e3962STejun Heo unsigned long delta; 2678e74e3962STejun Heo unsigned int cpu; 2679fb435d52STejun Heo int rc; 2680e74e3962STejun Heo 2681e74e3962STejun Heo /* 2682e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2683e74e3962STejun Heo * what the legacy allocator did. 2684e74e3962STejun Heo */ 2685fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2686c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2687c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2688fb435d52STejun Heo if (rc < 0) 2689bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2690e74e3962STejun Heo 2691e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2692e74e3962STejun Heo for_each_possible_cpu(cpu) 2693fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2694e74e3962STejun Heo } 2695e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2696099a19d9STejun Heo 2697bbddff05STejun Heo #else /* CONFIG_SMP */ 2698bbddff05STejun Heo 2699bbddff05STejun Heo /* 2700bbddff05STejun Heo * UP percpu area setup. 2701bbddff05STejun Heo * 2702bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2703bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2704bbddff05STejun Heo * variables and don't require any special preparation. 2705bbddff05STejun Heo */ 2706bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2707bbddff05STejun Heo { 2708bbddff05STejun Heo const size_t unit_size = 2709bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2710bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2711bbddff05STejun Heo struct pcpu_alloc_info *ai; 2712bbddff05STejun Heo void *fc; 2713bbddff05STejun Heo 2714bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 2715999c17e3SSantosh Shilimkar fc = memblock_virt_alloc_from_nopanic(unit_size, 2716999c17e3SSantosh Shilimkar PAGE_SIZE, 2717999c17e3SSantosh Shilimkar __pa(MAX_DMA_ADDRESS)); 2718bbddff05STejun Heo if (!ai || !fc) 2719bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2720100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2721100d13c3SCatalin Marinas kmemleak_free(fc); 2722bbddff05STejun Heo 2723bbddff05STejun Heo ai->dyn_size = unit_size; 2724bbddff05STejun Heo ai->unit_size = unit_size; 2725bbddff05STejun Heo ai->atom_size = unit_size; 2726bbddff05STejun Heo ai->alloc_size = unit_size; 2727bbddff05STejun Heo ai->groups[0].nr_units = 1; 2728bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2729bbddff05STejun Heo 2730bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 2731bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2732abee2105SNicolas Pitre #ifdef CONFIG_CRIS 2733abee2105SNicolas Pitre #warning "the CRIS architecture has physical and virtual addresses confused" 2734abee2105SNicolas Pitre #else 2735438a5061SNicolas Pitre pcpu_free_alloc_info(ai); 2736abee2105SNicolas Pitre #endif 2737bbddff05STejun Heo } 2738bbddff05STejun Heo 2739bbddff05STejun Heo #endif /* CONFIG_SMP */ 2740bbddff05STejun Heo 2741099a19d9STejun Heo /* 27421a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 27431a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 27441a4d7607STejun Heo * and running. 27451a4d7607STejun Heo */ 27461a4d7607STejun Heo static int __init percpu_enable_async(void) 27471a4d7607STejun Heo { 27481a4d7607STejun Heo pcpu_async_enabled = true; 27491a4d7607STejun Heo return 0; 27501a4d7607STejun Heo } 27511a4d7607STejun Heo subsys_initcall(percpu_enable_async); 2752