1fbf59bc9STejun Heo /* 288999a89STejun Heo * mm/percpu.c - percpu memory allocator 3fbf59bc9STejun Heo * 4fbf59bc9STejun Heo * Copyright (C) 2009 SUSE Linux Products GmbH 5fbf59bc9STejun Heo * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 6fbf59bc9STejun Heo * 75e81ee3eSDennis Zhou (Facebook) * Copyright (C) 2017 Facebook Inc. 85e81ee3eSDennis Zhou (Facebook) * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com> 95e81ee3eSDennis Zhou (Facebook) * 109c015162SDennis Zhou (Facebook) * This file is released under the GPLv2 license. 11fbf59bc9STejun Heo * 129c015162SDennis Zhou (Facebook) * The percpu allocator handles both static and dynamic areas. Percpu 139c015162SDennis Zhou (Facebook) * areas are allocated in chunks which are divided into units. There is 149c015162SDennis Zhou (Facebook) * a 1-to-1 mapping for units to possible cpus. These units are grouped 159c015162SDennis Zhou (Facebook) * based on NUMA properties of the machine. 16fbf59bc9STejun Heo * 17fbf59bc9STejun Heo * c0 c1 c2 18fbf59bc9STejun Heo * ------------------- ------------------- ------------ 19fbf59bc9STejun Heo * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 20fbf59bc9STejun Heo * ------------------- ...... ------------------- .... ------------ 21fbf59bc9STejun Heo * 229c015162SDennis Zhou (Facebook) * Allocation is done by offsets into a unit's address space. Ie., an 239c015162SDennis Zhou (Facebook) * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, 249c015162SDennis Zhou (Facebook) * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear 259c015162SDennis Zhou (Facebook) * and even sparse. Access is handled by configuring percpu base 269c015162SDennis Zhou (Facebook) * registers according to the cpu to unit mappings and offsetting the 279c015162SDennis Zhou (Facebook) * base address using pcpu_unit_size. 28fbf59bc9STejun Heo * 299c015162SDennis Zhou (Facebook) * There is special consideration for the first chunk which must handle 309c015162SDennis Zhou (Facebook) * the static percpu variables in the kernel image as allocation services 315e81ee3eSDennis Zhou (Facebook) * are not online yet. In short, the first chunk is structured like so: 329c015162SDennis Zhou (Facebook) * 339c015162SDennis Zhou (Facebook) * <Static | [Reserved] | Dynamic> 349c015162SDennis Zhou (Facebook) * 359c015162SDennis Zhou (Facebook) * The static data is copied from the original section managed by the 369c015162SDennis Zhou (Facebook) * linker. The reserved section, if non-zero, primarily manages static 379c015162SDennis Zhou (Facebook) * percpu variables from kernel modules. Finally, the dynamic section 389c015162SDennis Zhou (Facebook) * takes care of normal allocations. 39fbf59bc9STejun Heo * 405e81ee3eSDennis Zhou (Facebook) * The allocator organizes chunks into lists according to free size and 415e81ee3eSDennis Zhou (Facebook) * tries to allocate from the fullest chunk first. Each chunk is managed 425e81ee3eSDennis Zhou (Facebook) * by a bitmap with metadata blocks. The allocation map is updated on 435e81ee3eSDennis Zhou (Facebook) * every allocation and free to reflect the current state while the boundary 445e81ee3eSDennis Zhou (Facebook) * map is only updated on allocation. Each metadata block contains 455e81ee3eSDennis Zhou (Facebook) * information to help mitigate the need to iterate over large portions 465e81ee3eSDennis Zhou (Facebook) * of the bitmap. The reverse mapping from page to chunk is stored in 475e81ee3eSDennis Zhou (Facebook) * the page's index. Lastly, units are lazily backed and grow in unison. 48fbf59bc9STejun Heo * 495e81ee3eSDennis Zhou (Facebook) * There is a unique conversion that goes on here between bytes and bits. 505e81ee3eSDennis Zhou (Facebook) * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk 515e81ee3eSDennis Zhou (Facebook) * tracks the number of pages it is responsible for in nr_pages. Helper 525e81ee3eSDennis Zhou (Facebook) * functions are used to convert from between the bytes, bits, and blocks. 535e81ee3eSDennis Zhou (Facebook) * All hints are managed in bits unless explicitly stated. 549c015162SDennis Zhou (Facebook) * 554091fb95SMasahiro Yamada * To use this allocator, arch code should do the following: 56fbf59bc9STejun Heo * 57fbf59bc9STejun Heo * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 58e0100983STejun Heo * regular address to percpu pointer and back if they need to be 59e0100983STejun Heo * different from the default 60fbf59bc9STejun Heo * 618d408b4bSTejun Heo * - use pcpu_setup_first_chunk() during percpu area initialization to 628d408b4bSTejun Heo * setup the first chunk containing the kernel static percpu area 63fbf59bc9STejun Heo */ 64fbf59bc9STejun Heo 65870d4b12SJoe Perches #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 66870d4b12SJoe Perches 67fbf59bc9STejun Heo #include <linux/bitmap.h> 6857c8a661SMike Rapoport #include <linux/memblock.h> 69fd1e8a1fSTejun Heo #include <linux/err.h> 70ca460b3cSDennis Zhou (Facebook) #include <linux/lcm.h> 71fbf59bc9STejun Heo #include <linux/list.h> 72a530b795STejun Heo #include <linux/log2.h> 73fbf59bc9STejun Heo #include <linux/mm.h> 74fbf59bc9STejun Heo #include <linux/module.h> 75fbf59bc9STejun Heo #include <linux/mutex.h> 76fbf59bc9STejun Heo #include <linux/percpu.h> 77fbf59bc9STejun Heo #include <linux/pfn.h> 78fbf59bc9STejun Heo #include <linux/slab.h> 79ccea34b5STejun Heo #include <linux/spinlock.h> 80fbf59bc9STejun Heo #include <linux/vmalloc.h> 81a56dbddfSTejun Heo #include <linux/workqueue.h> 82f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 8371546d10STejun Heo #include <linux/sched.h> 84fbf59bc9STejun Heo 85fbf59bc9STejun Heo #include <asm/cacheflush.h> 86e0100983STejun Heo #include <asm/sections.h> 87fbf59bc9STejun Heo #include <asm/tlbflush.h> 883b034b0dSVivek Goyal #include <asm/io.h> 89fbf59bc9STejun Heo 90df95e795SDennis Zhou #define CREATE_TRACE_POINTS 91df95e795SDennis Zhou #include <trace/events/percpu.h> 92df95e795SDennis Zhou 938fa3ed80SDennis Zhou #include "percpu-internal.h" 948fa3ed80SDennis Zhou 9540064aecSDennis Zhou (Facebook) /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ 9640064aecSDennis Zhou (Facebook) #define PCPU_SLOT_BASE_SHIFT 5 978744d859SDennis Zhou /* chunks in slots below this are subject to being sidelined on failed alloc */ 988744d859SDennis Zhou #define PCPU_SLOT_FAIL_THRESHOLD 3 9940064aecSDennis Zhou (Facebook) 1001a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_LOW 2 1011a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_HIGH 4 102fbf59bc9STejun Heo 103bbddff05STejun Heo #ifdef CONFIG_SMP 104e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 105e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 106e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 10743cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 10843cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 10943cf38ebSTejun Heo (unsigned long)__per_cpu_start) 110e0100983STejun Heo #endif 111e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 112e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 11343cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 11443cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 11543cf38ebSTejun Heo (unsigned long)__per_cpu_start) 116e0100983STejun Heo #endif 117bbddff05STejun Heo #else /* CONFIG_SMP */ 118bbddff05STejun Heo /* on UP, it's always identity mapped */ 119bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 120bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 121bbddff05STejun Heo #endif /* CONFIG_SMP */ 122e0100983STejun Heo 1231328710bSDaniel Micay static int pcpu_unit_pages __ro_after_init; 1241328710bSDaniel Micay static int pcpu_unit_size __ro_after_init; 1251328710bSDaniel Micay static int pcpu_nr_units __ro_after_init; 1261328710bSDaniel Micay static int pcpu_atom_size __ro_after_init; 1278fa3ed80SDennis Zhou int pcpu_nr_slots __ro_after_init; 1281328710bSDaniel Micay static size_t pcpu_chunk_struct_size __ro_after_init; 129fbf59bc9STejun Heo 130a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 1311328710bSDaniel Micay static unsigned int pcpu_low_unit_cpu __ro_after_init; 1321328710bSDaniel Micay static unsigned int pcpu_high_unit_cpu __ro_after_init; 1332f39e637STejun Heo 134fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 1351328710bSDaniel Micay void *pcpu_base_addr __ro_after_init; 136fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 137fbf59bc9STejun Heo 1381328710bSDaniel Micay static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ 1391328710bSDaniel Micay const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ 1402f39e637STejun Heo 1416563297cSTejun Heo /* group information, used for vm allocation */ 1421328710bSDaniel Micay static int pcpu_nr_groups __ro_after_init; 1431328710bSDaniel Micay static const unsigned long *pcpu_group_offsets __ro_after_init; 1441328710bSDaniel Micay static const size_t *pcpu_group_sizes __ro_after_init; 1456563297cSTejun Heo 146ae9e6bc9STejun Heo /* 147ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 148ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 149ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 150ae9e6bc9STejun Heo */ 1518fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_first_chunk __ro_after_init; 152ae9e6bc9STejun Heo 153ae9e6bc9STejun Heo /* 154ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 155e2266705SDennis Zhou (Facebook) * chunk and serves it for reserved allocations. When the reserved 156e2266705SDennis Zhou (Facebook) * region doesn't exist, the following variable is NULL. 157ae9e6bc9STejun Heo */ 1588fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; 159edcb4639STejun Heo 1608fa3ed80SDennis Zhou DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 1616710e594STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 162fbf59bc9STejun Heo 1638fa3ed80SDennis Zhou struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ 164fbf59bc9STejun Heo 1654f996e23STejun Heo /* chunks which need their map areas extended, protected by pcpu_lock */ 1664f996e23STejun Heo static LIST_HEAD(pcpu_map_extend_chunks); 1674f996e23STejun Heo 168b539b87fSTejun Heo /* 169b539b87fSTejun Heo * The number of empty populated pages, protected by pcpu_lock. The 170b539b87fSTejun Heo * reserved chunk doesn't contribute to the count. 171b539b87fSTejun Heo */ 1726b9b6f39SDennis Zhou (Facebook) int pcpu_nr_empty_pop_pages; 173b539b87fSTejun Heo 1741a4d7607STejun Heo /* 1757e8a6304SDennis Zhou (Facebook) * The number of populated pages in use by the allocator, protected by 1767e8a6304SDennis Zhou (Facebook) * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets 1777e8a6304SDennis Zhou (Facebook) * allocated/deallocated, it is allocated/deallocated in all units of a chunk 1787e8a6304SDennis Zhou (Facebook) * and increments/decrements this count by 1). 1797e8a6304SDennis Zhou (Facebook) */ 1807e8a6304SDennis Zhou (Facebook) static unsigned long pcpu_nr_populated; 1817e8a6304SDennis Zhou (Facebook) 1827e8a6304SDennis Zhou (Facebook) /* 1831a4d7607STejun Heo * Balance work is used to populate or destroy chunks asynchronously. We 1841a4d7607STejun Heo * try to keep the number of populated free pages between 1851a4d7607STejun Heo * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 1861a4d7607STejun Heo * empty chunk. 1871a4d7607STejun Heo */ 188fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work); 189fe6bd8c3STejun Heo static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 1901a4d7607STejun Heo static bool pcpu_async_enabled __read_mostly; 1911a4d7607STejun Heo static bool pcpu_atomic_alloc_failed; 1921a4d7607STejun Heo 1931a4d7607STejun Heo static void pcpu_schedule_balance_work(void) 1941a4d7607STejun Heo { 1951a4d7607STejun Heo if (pcpu_async_enabled) 1961a4d7607STejun Heo schedule_work(&pcpu_balance_work); 1971a4d7607STejun Heo } 198a56dbddfSTejun Heo 199c0ebfdc3SDennis Zhou (Facebook) /** 200560f2c23SDennis Zhou (Facebook) * pcpu_addr_in_chunk - check if the address is served from this chunk 201560f2c23SDennis Zhou (Facebook) * @chunk: chunk of interest 202560f2c23SDennis Zhou (Facebook) * @addr: percpu address 203c0ebfdc3SDennis Zhou (Facebook) * 204c0ebfdc3SDennis Zhou (Facebook) * RETURNS: 205560f2c23SDennis Zhou (Facebook) * True if the address is served from this chunk. 206c0ebfdc3SDennis Zhou (Facebook) */ 207560f2c23SDennis Zhou (Facebook) static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) 208020ec653STejun Heo { 209c0ebfdc3SDennis Zhou (Facebook) void *start_addr, *end_addr; 210020ec653STejun Heo 211560f2c23SDennis Zhou (Facebook) if (!chunk) 212c0ebfdc3SDennis Zhou (Facebook) return false; 213c0ebfdc3SDennis Zhou (Facebook) 214560f2c23SDennis Zhou (Facebook) start_addr = chunk->base_addr + chunk->start_offset; 215560f2c23SDennis Zhou (Facebook) end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - 216560f2c23SDennis Zhou (Facebook) chunk->end_offset; 217c0ebfdc3SDennis Zhou (Facebook) 218c0ebfdc3SDennis Zhou (Facebook) return addr >= start_addr && addr < end_addr; 219020ec653STejun Heo } 220020ec653STejun Heo 221d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 222fbf59bc9STejun Heo { 223cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 224fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 225fbf59bc9STejun Heo } 226fbf59bc9STejun Heo 227d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 228d9b55eebSTejun Heo { 229d9b55eebSTejun Heo if (size == pcpu_unit_size) 230d9b55eebSTejun Heo return pcpu_nr_slots - 1; 231d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 232d9b55eebSTejun Heo } 233d9b55eebSTejun Heo 234fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 235fbf59bc9STejun Heo { 23640064aecSDennis Zhou (Facebook) if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) 237fbf59bc9STejun Heo return 0; 238fbf59bc9STejun Heo 2393e54097bSDennis Zhou return pcpu_size_to_slot(chunk->contig_bits * PCPU_MIN_ALLOC_SIZE); 240fbf59bc9STejun Heo } 241fbf59bc9STejun Heo 24288999a89STejun Heo /* set the pointer to a chunk in a page struct */ 24388999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 24488999a89STejun Heo { 24588999a89STejun Heo page->index = (unsigned long)pcpu; 24688999a89STejun Heo } 24788999a89STejun Heo 24888999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 24988999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 25088999a89STejun Heo { 25188999a89STejun Heo return (struct pcpu_chunk *)page->index; 25288999a89STejun Heo } 25388999a89STejun Heo 25488999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 255fbf59bc9STejun Heo { 2562f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 257fbf59bc9STejun Heo } 258fbf59bc9STejun Heo 259c0ebfdc3SDennis Zhou (Facebook) static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 260c0ebfdc3SDennis Zhou (Facebook) { 261c0ebfdc3SDennis Zhou (Facebook) return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 262c0ebfdc3SDennis Zhou (Facebook) } 263c0ebfdc3SDennis Zhou (Facebook) 2649983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 265fbf59bc9STejun Heo unsigned int cpu, int page_idx) 266fbf59bc9STejun Heo { 267c0ebfdc3SDennis Zhou (Facebook) return (unsigned long)chunk->base_addr + 268c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(cpu, page_idx); 269fbf59bc9STejun Heo } 270fbf59bc9STejun Heo 27191e914c5SDennis Zhou (Facebook) static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) 272ce3141a2STejun Heo { 27391e914c5SDennis Zhou (Facebook) *rs = find_next_zero_bit(bitmap, end, *rs); 27491e914c5SDennis Zhou (Facebook) *re = find_next_bit(bitmap, end, *rs + 1); 275ce3141a2STejun Heo } 276ce3141a2STejun Heo 27791e914c5SDennis Zhou (Facebook) static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) 278ce3141a2STejun Heo { 27991e914c5SDennis Zhou (Facebook) *rs = find_next_bit(bitmap, end, *rs); 28091e914c5SDennis Zhou (Facebook) *re = find_next_zero_bit(bitmap, end, *rs + 1); 281ce3141a2STejun Heo } 282ce3141a2STejun Heo 283ce3141a2STejun Heo /* 28491e914c5SDennis Zhou (Facebook) * Bitmap region iterators. Iterates over the bitmap between 28591e914c5SDennis Zhou (Facebook) * [@start, @end) in @chunk. @rs and @re should be integer variables 28691e914c5SDennis Zhou (Facebook) * and will be set to start and end index of the current free region. 287ce3141a2STejun Heo */ 28891e914c5SDennis Zhou (Facebook) #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ 28991e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ 290ce3141a2STejun Heo (rs) < (re); \ 29191e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) 292ce3141a2STejun Heo 29391e914c5SDennis Zhou (Facebook) #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ 29491e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ 295ce3141a2STejun Heo (rs) < (re); \ 29691e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) 297ce3141a2STejun Heo 298ca460b3cSDennis Zhou (Facebook) /* 299ca460b3cSDennis Zhou (Facebook) * The following are helper functions to help access bitmaps and convert 300ca460b3cSDennis Zhou (Facebook) * between bitmap offsets to address offsets. 301ca460b3cSDennis Zhou (Facebook) */ 302ca460b3cSDennis Zhou (Facebook) static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 303ca460b3cSDennis Zhou (Facebook) { 304ca460b3cSDennis Zhou (Facebook) return chunk->alloc_map + 305ca460b3cSDennis Zhou (Facebook) (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 306ca460b3cSDennis Zhou (Facebook) } 307ca460b3cSDennis Zhou (Facebook) 308ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_index(int off) 309ca460b3cSDennis Zhou (Facebook) { 310ca460b3cSDennis Zhou (Facebook) return off / PCPU_BITMAP_BLOCK_BITS; 311ca460b3cSDennis Zhou (Facebook) } 312ca460b3cSDennis Zhou (Facebook) 313ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_off(int off) 314ca460b3cSDennis Zhou (Facebook) { 315ca460b3cSDennis Zhou (Facebook) return off & (PCPU_BITMAP_BLOCK_BITS - 1); 316ca460b3cSDennis Zhou (Facebook) } 317ca460b3cSDennis Zhou (Facebook) 318b185cd0dSDennis Zhou (Facebook) static unsigned long pcpu_block_off_to_off(int index, int off) 319b185cd0dSDennis Zhou (Facebook) { 320b185cd0dSDennis Zhou (Facebook) return index * PCPU_BITMAP_BLOCK_BITS + off; 321b185cd0dSDennis Zhou (Facebook) } 322b185cd0dSDennis Zhou (Facebook) 323*382b88e9SDennis Zhou /* 324*382b88e9SDennis Zhou * pcpu_next_hint - determine which hint to use 325*382b88e9SDennis Zhou * @block: block of interest 326*382b88e9SDennis Zhou * @alloc_bits: size of allocation 327*382b88e9SDennis Zhou * 328*382b88e9SDennis Zhou * This determines if we should scan based on the scan_hint or first_free. 329*382b88e9SDennis Zhou * In general, we want to scan from first_free to fulfill allocations by 330*382b88e9SDennis Zhou * first fit. However, if we know a scan_hint at position scan_hint_start 331*382b88e9SDennis Zhou * cannot fulfill an allocation, we can begin scanning from there knowing 332*382b88e9SDennis Zhou * the contig_hint will be our fallback. 333*382b88e9SDennis Zhou */ 334*382b88e9SDennis Zhou static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits) 335*382b88e9SDennis Zhou { 336*382b88e9SDennis Zhou /* 337*382b88e9SDennis Zhou * The three conditions below determine if we can skip past the 338*382b88e9SDennis Zhou * scan_hint. First, does the scan hint exist. Second, is the 339*382b88e9SDennis Zhou * contig_hint after the scan_hint (possibly not true iff 340*382b88e9SDennis Zhou * contig_hint == scan_hint). Third, is the allocation request 341*382b88e9SDennis Zhou * larger than the scan_hint. 342*382b88e9SDennis Zhou */ 343*382b88e9SDennis Zhou if (block->scan_hint && 344*382b88e9SDennis Zhou block->contig_hint_start > block->scan_hint_start && 345*382b88e9SDennis Zhou alloc_bits > block->scan_hint) 346*382b88e9SDennis Zhou return block->scan_hint_start + block->scan_hint; 347*382b88e9SDennis Zhou 348*382b88e9SDennis Zhou return block->first_free; 349*382b88e9SDennis Zhou } 350*382b88e9SDennis Zhou 351fbf59bc9STejun Heo /** 352525ca84dSDennis Zhou (Facebook) * pcpu_next_md_free_region - finds the next hint free area 353525ca84dSDennis Zhou (Facebook) * @chunk: chunk of interest 354525ca84dSDennis Zhou (Facebook) * @bit_off: chunk offset 355525ca84dSDennis Zhou (Facebook) * @bits: size of free area 356525ca84dSDennis Zhou (Facebook) * 357525ca84dSDennis Zhou (Facebook) * Helper function for pcpu_for_each_md_free_region. It checks 358525ca84dSDennis Zhou (Facebook) * block->contig_hint and performs aggregation across blocks to find the 359525ca84dSDennis Zhou (Facebook) * next hint. It modifies bit_off and bits in-place to be consumed in the 360525ca84dSDennis Zhou (Facebook) * loop. 361525ca84dSDennis Zhou (Facebook) */ 362525ca84dSDennis Zhou (Facebook) static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, 363525ca84dSDennis Zhou (Facebook) int *bits) 364525ca84dSDennis Zhou (Facebook) { 365525ca84dSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 366525ca84dSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 367525ca84dSDennis Zhou (Facebook) struct pcpu_block_md *block; 368525ca84dSDennis Zhou (Facebook) 369525ca84dSDennis Zhou (Facebook) *bits = 0; 370525ca84dSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 371525ca84dSDennis Zhou (Facebook) block++, i++) { 372525ca84dSDennis Zhou (Facebook) /* handles contig area across blocks */ 373525ca84dSDennis Zhou (Facebook) if (*bits) { 374525ca84dSDennis Zhou (Facebook) *bits += block->left_free; 375525ca84dSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 376525ca84dSDennis Zhou (Facebook) continue; 377525ca84dSDennis Zhou (Facebook) return; 378525ca84dSDennis Zhou (Facebook) } 379525ca84dSDennis Zhou (Facebook) 380525ca84dSDennis Zhou (Facebook) /* 381525ca84dSDennis Zhou (Facebook) * This checks three things. First is there a contig_hint to 382525ca84dSDennis Zhou (Facebook) * check. Second, have we checked this hint before by 383525ca84dSDennis Zhou (Facebook) * comparing the block_off. Third, is this the same as the 384525ca84dSDennis Zhou (Facebook) * right contig hint. In the last case, it spills over into 385525ca84dSDennis Zhou (Facebook) * the next block and should be handled by the contig area 386525ca84dSDennis Zhou (Facebook) * across blocks code. 387525ca84dSDennis Zhou (Facebook) */ 388525ca84dSDennis Zhou (Facebook) *bits = block->contig_hint; 389525ca84dSDennis Zhou (Facebook) if (*bits && block->contig_hint_start >= block_off && 390525ca84dSDennis Zhou (Facebook) *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { 391525ca84dSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, 392525ca84dSDennis Zhou (Facebook) block->contig_hint_start); 393525ca84dSDennis Zhou (Facebook) return; 394525ca84dSDennis Zhou (Facebook) } 3951fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 3961fa4df3eSDennis Zhou block_off = 0; 397525ca84dSDennis Zhou (Facebook) 398525ca84dSDennis Zhou (Facebook) *bits = block->right_free; 399525ca84dSDennis Zhou (Facebook) *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; 400525ca84dSDennis Zhou (Facebook) } 401525ca84dSDennis Zhou (Facebook) } 402525ca84dSDennis Zhou (Facebook) 403b4c2116cSDennis Zhou (Facebook) /** 404b4c2116cSDennis Zhou (Facebook) * pcpu_next_fit_region - finds fit areas for a given allocation request 405b4c2116cSDennis Zhou (Facebook) * @chunk: chunk of interest 406b4c2116cSDennis Zhou (Facebook) * @alloc_bits: size of allocation 407b4c2116cSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 408b4c2116cSDennis Zhou (Facebook) * @bit_off: chunk offset 409b4c2116cSDennis Zhou (Facebook) * @bits: size of free area 410b4c2116cSDennis Zhou (Facebook) * 411b4c2116cSDennis Zhou (Facebook) * Finds the next free region that is viable for use with a given size and 412b4c2116cSDennis Zhou (Facebook) * alignment. This only returns if there is a valid area to be used for this 413b4c2116cSDennis Zhou (Facebook) * allocation. block->first_free is returned if the allocation request fits 414b4c2116cSDennis Zhou (Facebook) * within the block to see if the request can be fulfilled prior to the contig 415b4c2116cSDennis Zhou (Facebook) * hint. 416b4c2116cSDennis Zhou (Facebook) */ 417b4c2116cSDennis Zhou (Facebook) static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, 418b4c2116cSDennis Zhou (Facebook) int align, int *bit_off, int *bits) 419b4c2116cSDennis Zhou (Facebook) { 420b4c2116cSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 421b4c2116cSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 422b4c2116cSDennis Zhou (Facebook) struct pcpu_block_md *block; 423b4c2116cSDennis Zhou (Facebook) 424b4c2116cSDennis Zhou (Facebook) *bits = 0; 425b4c2116cSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 426b4c2116cSDennis Zhou (Facebook) block++, i++) { 427b4c2116cSDennis Zhou (Facebook) /* handles contig area across blocks */ 428b4c2116cSDennis Zhou (Facebook) if (*bits) { 429b4c2116cSDennis Zhou (Facebook) *bits += block->left_free; 430b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 431b4c2116cSDennis Zhou (Facebook) return; 432b4c2116cSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 433b4c2116cSDennis Zhou (Facebook) continue; 434b4c2116cSDennis Zhou (Facebook) } 435b4c2116cSDennis Zhou (Facebook) 436b4c2116cSDennis Zhou (Facebook) /* check block->contig_hint */ 437b4c2116cSDennis Zhou (Facebook) *bits = ALIGN(block->contig_hint_start, align) - 438b4c2116cSDennis Zhou (Facebook) block->contig_hint_start; 439b4c2116cSDennis Zhou (Facebook) /* 440b4c2116cSDennis Zhou (Facebook) * This uses the block offset to determine if this has been 441b4c2116cSDennis Zhou (Facebook) * checked in the prior iteration. 442b4c2116cSDennis Zhou (Facebook) */ 443b4c2116cSDennis Zhou (Facebook) if (block->contig_hint && 444b4c2116cSDennis Zhou (Facebook) block->contig_hint_start >= block_off && 445b4c2116cSDennis Zhou (Facebook) block->contig_hint >= *bits + alloc_bits) { 446*382b88e9SDennis Zhou int start = pcpu_next_hint(block, alloc_bits); 447*382b88e9SDennis Zhou 448b4c2116cSDennis Zhou (Facebook) *bits += alloc_bits + block->contig_hint_start - 449*382b88e9SDennis Zhou start; 450*382b88e9SDennis Zhou *bit_off = pcpu_block_off_to_off(i, start); 451b4c2116cSDennis Zhou (Facebook) return; 452b4c2116cSDennis Zhou (Facebook) } 4531fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 4541fa4df3eSDennis Zhou block_off = 0; 455b4c2116cSDennis Zhou (Facebook) 456b4c2116cSDennis Zhou (Facebook) *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, 457b4c2116cSDennis Zhou (Facebook) align); 458b4c2116cSDennis Zhou (Facebook) *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; 459b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, *bit_off); 460b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 461b4c2116cSDennis Zhou (Facebook) return; 462b4c2116cSDennis Zhou (Facebook) } 463b4c2116cSDennis Zhou (Facebook) 464b4c2116cSDennis Zhou (Facebook) /* no valid offsets were found - fail condition */ 465b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_chunk_map_bits(chunk); 466b4c2116cSDennis Zhou (Facebook) } 467b4c2116cSDennis Zhou (Facebook) 468525ca84dSDennis Zhou (Facebook) /* 469525ca84dSDennis Zhou (Facebook) * Metadata free area iterators. These perform aggregation of free areas 470525ca84dSDennis Zhou (Facebook) * based on the metadata blocks and return the offset @bit_off and size in 471b4c2116cSDennis Zhou (Facebook) * bits of the free area @bits. pcpu_for_each_fit_region only returns when 472b4c2116cSDennis Zhou (Facebook) * a fit is found for the allocation request. 473525ca84dSDennis Zhou (Facebook) */ 474525ca84dSDennis Zhou (Facebook) #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ 475525ca84dSDennis Zhou (Facebook) for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ 476525ca84dSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 477525ca84dSDennis Zhou (Facebook) (bit_off) += (bits) + 1, \ 478525ca84dSDennis Zhou (Facebook) pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) 479525ca84dSDennis Zhou (Facebook) 480b4c2116cSDennis Zhou (Facebook) #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ 481b4c2116cSDennis Zhou (Facebook) for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 482b4c2116cSDennis Zhou (Facebook) &(bits)); \ 483b4c2116cSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 484b4c2116cSDennis Zhou (Facebook) (bit_off) += (bits), \ 485b4c2116cSDennis Zhou (Facebook) pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 486b4c2116cSDennis Zhou (Facebook) &(bits))) 487b4c2116cSDennis Zhou (Facebook) 488525ca84dSDennis Zhou (Facebook) /** 48990459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 4901880d93bSTejun Heo * @size: bytes to allocate 49147504ee0SDennis Zhou * @gfp: allocation flags 492fbf59bc9STejun Heo * 4931880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 49447504ee0SDennis Zhou * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. 49547504ee0SDennis Zhou * This is to facilitate passing through whitelisted flags. The 49647504ee0SDennis Zhou * returned memory is always zeroed. 497fbf59bc9STejun Heo * 498fbf59bc9STejun Heo * RETURNS: 4991880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 500fbf59bc9STejun Heo */ 50147504ee0SDennis Zhou static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) 502fbf59bc9STejun Heo { 503099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 504099a19d9STejun Heo return NULL; 505099a19d9STejun Heo 506fbf59bc9STejun Heo if (size <= PAGE_SIZE) 507554fef1cSDennis Zhou return kzalloc(size, gfp); 5087af4c093SJesper Juhl else 509554fef1cSDennis Zhou return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL); 5101880d93bSTejun Heo } 511fbf59bc9STejun Heo 5121880d93bSTejun Heo /** 5131880d93bSTejun Heo * pcpu_mem_free - free memory 5141880d93bSTejun Heo * @ptr: memory to free 5151880d93bSTejun Heo * 51690459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 5171880d93bSTejun Heo */ 5181d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 5191880d93bSTejun Heo { 5201d5cfdb0STetsuo Handa kvfree(ptr); 521fbf59bc9STejun Heo } 522fbf59bc9STejun Heo 5238744d859SDennis Zhou static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot, 5248744d859SDennis Zhou bool move_front) 5258744d859SDennis Zhou { 5268744d859SDennis Zhou if (chunk != pcpu_reserved_chunk) { 5278744d859SDennis Zhou if (move_front) 5288744d859SDennis Zhou list_move(&chunk->list, &pcpu_slot[slot]); 5298744d859SDennis Zhou else 5308744d859SDennis Zhou list_move_tail(&chunk->list, &pcpu_slot[slot]); 5318744d859SDennis Zhou } 5328744d859SDennis Zhou } 5338744d859SDennis Zhou 5348744d859SDennis Zhou static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot) 5358744d859SDennis Zhou { 5368744d859SDennis Zhou __pcpu_chunk_move(chunk, slot, true); 5378744d859SDennis Zhou } 5388744d859SDennis Zhou 539fbf59bc9STejun Heo /** 540fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 541fbf59bc9STejun Heo * @chunk: chunk of interest 542fbf59bc9STejun Heo * @oslot: the previous slot it was on 543fbf59bc9STejun Heo * 544fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 545fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 546edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 547edcb4639STejun Heo * chunk slots. 548ccea34b5STejun Heo * 549ccea34b5STejun Heo * CONTEXT: 550ccea34b5STejun Heo * pcpu_lock. 551fbf59bc9STejun Heo */ 552fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 553fbf59bc9STejun Heo { 554fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 555fbf59bc9STejun Heo 5568744d859SDennis Zhou if (oslot != nslot) 5578744d859SDennis Zhou __pcpu_chunk_move(chunk, nslot, oslot < nslot); 558fbf59bc9STejun Heo } 559fbf59bc9STejun Heo 56040064aecSDennis Zhou (Facebook) /* 561b239f7daSDennis Zhou * pcpu_update_empty_pages - update empty page counters 562b239f7daSDennis Zhou * @chunk: chunk of interest 563b239f7daSDennis Zhou * @nr: nr of empty pages 564b239f7daSDennis Zhou * 565b239f7daSDennis Zhou * This is used to keep track of the empty pages now based on the premise 566b239f7daSDennis Zhou * a md_block covers a page. The hint update functions recognize if a block 567b239f7daSDennis Zhou * is made full or broken to calculate deltas for keeping track of free pages. 56840064aecSDennis Zhou (Facebook) */ 569b239f7daSDennis Zhou static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr) 570b239f7daSDennis Zhou { 571b239f7daSDennis Zhou chunk->nr_empty_pop_pages += nr; 572b239f7daSDennis Zhou if (chunk != pcpu_reserved_chunk) 573b239f7daSDennis Zhou pcpu_nr_empty_pop_pages += nr; 57440064aecSDennis Zhou (Facebook) } 57540064aecSDennis Zhou (Facebook) 576d9f3a01eSDennis Zhou /* 577d9f3a01eSDennis Zhou * pcpu_region_overlap - determines if two regions overlap 578d9f3a01eSDennis Zhou * @a: start of first region, inclusive 579d9f3a01eSDennis Zhou * @b: end of first region, exclusive 580d9f3a01eSDennis Zhou * @x: start of second region, inclusive 581d9f3a01eSDennis Zhou * @y: end of second region, exclusive 582d9f3a01eSDennis Zhou * 583d9f3a01eSDennis Zhou * This is used to determine if the hint region [a, b) overlaps with the 584d9f3a01eSDennis Zhou * allocated region [x, y). 585d9f3a01eSDennis Zhou */ 586d9f3a01eSDennis Zhou static inline bool pcpu_region_overlap(int a, int b, int x, int y) 587d9f3a01eSDennis Zhou { 588d9f3a01eSDennis Zhou return (a < y) && (x < b); 589d9f3a01eSDennis Zhou } 590d9f3a01eSDennis Zhou 59140064aecSDennis Zhou (Facebook) /** 59240064aecSDennis Zhou (Facebook) * pcpu_chunk_update - updates the chunk metadata given a free area 59340064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 59440064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 59540064aecSDennis Zhou (Facebook) * @bits: size of free area 59640064aecSDennis Zhou (Facebook) * 59713f96637SDennis Zhou (Facebook) * This updates the chunk's contig hint and starting offset given a free area. 598268625a6SDennis Zhou (Facebook) * Choose the best starting offset if the contig hint is equal. 59940064aecSDennis Zhou (Facebook) */ 60040064aecSDennis Zhou (Facebook) static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) 60140064aecSDennis Zhou (Facebook) { 60213f96637SDennis Zhou (Facebook) if (bits > chunk->contig_bits) { 60313f96637SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 60440064aecSDennis Zhou (Facebook) chunk->contig_bits = bits; 605268625a6SDennis Zhou (Facebook) } else if (bits == chunk->contig_bits && chunk->contig_bits_start && 606268625a6SDennis Zhou (Facebook) (!bit_off || 607268625a6SDennis Zhou (Facebook) __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { 608268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 609268625a6SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 61040064aecSDennis Zhou (Facebook) } 61113f96637SDennis Zhou (Facebook) } 61240064aecSDennis Zhou (Facebook) 61340064aecSDennis Zhou (Facebook) /** 61440064aecSDennis Zhou (Facebook) * pcpu_chunk_refresh_hint - updates metadata about a chunk 61540064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 61640064aecSDennis Zhou (Facebook) * 617525ca84dSDennis Zhou (Facebook) * Iterates over the metadata blocks to find the largest contig area. 618525ca84dSDennis Zhou (Facebook) * It also counts the populated pages and uses the delta to update the 619525ca84dSDennis Zhou (Facebook) * global count. 62040064aecSDennis Zhou (Facebook) * 62140064aecSDennis Zhou (Facebook) * Updates: 62240064aecSDennis Zhou (Facebook) * chunk->contig_bits 62313f96637SDennis Zhou (Facebook) * chunk->contig_bits_start 62440064aecSDennis Zhou (Facebook) */ 62540064aecSDennis Zhou (Facebook) static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) 62640064aecSDennis Zhou (Facebook) { 627b239f7daSDennis Zhou int bit_off, bits; 62840064aecSDennis Zhou (Facebook) 62940064aecSDennis Zhou (Facebook) /* clear metadata */ 63040064aecSDennis Zhou (Facebook) chunk->contig_bits = 0; 63140064aecSDennis Zhou (Facebook) 632525ca84dSDennis Zhou (Facebook) bit_off = chunk->first_bit; 633b239f7daSDennis Zhou bits = 0; 634525ca84dSDennis Zhou (Facebook) pcpu_for_each_md_free_region(chunk, bit_off, bits) { 635525ca84dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, bit_off, bits); 63640064aecSDennis Zhou (Facebook) } 63740064aecSDennis Zhou (Facebook) } 63840064aecSDennis Zhou (Facebook) 63940064aecSDennis Zhou (Facebook) /** 640ca460b3cSDennis Zhou (Facebook) * pcpu_block_update - updates a block given a free area 641ca460b3cSDennis Zhou (Facebook) * @block: block of interest 642ca460b3cSDennis Zhou (Facebook) * @start: start offset in block 643ca460b3cSDennis Zhou (Facebook) * @end: end offset in block 644ca460b3cSDennis Zhou (Facebook) * 645ca460b3cSDennis Zhou (Facebook) * Updates a block given a known free area. The region [start, end) is 646268625a6SDennis Zhou (Facebook) * expected to be the entirety of the free area within a block. Chooses 647268625a6SDennis Zhou (Facebook) * the best starting offset if the contig hints are equal. 648ca460b3cSDennis Zhou (Facebook) */ 649ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 650ca460b3cSDennis Zhou (Facebook) { 651ca460b3cSDennis Zhou (Facebook) int contig = end - start; 652ca460b3cSDennis Zhou (Facebook) 653ca460b3cSDennis Zhou (Facebook) block->first_free = min(block->first_free, start); 654ca460b3cSDennis Zhou (Facebook) if (start == 0) 655ca460b3cSDennis Zhou (Facebook) block->left_free = contig; 656ca460b3cSDennis Zhou (Facebook) 657ca460b3cSDennis Zhou (Facebook) if (end == PCPU_BITMAP_BLOCK_BITS) 658ca460b3cSDennis Zhou (Facebook) block->right_free = contig; 659ca460b3cSDennis Zhou (Facebook) 660ca460b3cSDennis Zhou (Facebook) if (contig > block->contig_hint) { 661*382b88e9SDennis Zhou /* promote the old contig_hint to be the new scan_hint */ 662*382b88e9SDennis Zhou if (start > block->contig_hint_start) { 663*382b88e9SDennis Zhou if (block->contig_hint > block->scan_hint) { 664*382b88e9SDennis Zhou block->scan_hint_start = 665*382b88e9SDennis Zhou block->contig_hint_start; 666*382b88e9SDennis Zhou block->scan_hint = block->contig_hint; 667*382b88e9SDennis Zhou } else if (start < block->scan_hint_start) { 668*382b88e9SDennis Zhou /* 669*382b88e9SDennis Zhou * The old contig_hint == scan_hint. But, the 670*382b88e9SDennis Zhou * new contig is larger so hold the invariant 671*382b88e9SDennis Zhou * scan_hint_start < contig_hint_start. 672*382b88e9SDennis Zhou */ 673*382b88e9SDennis Zhou block->scan_hint = 0; 674*382b88e9SDennis Zhou } 675*382b88e9SDennis Zhou } else { 676*382b88e9SDennis Zhou block->scan_hint = 0; 677*382b88e9SDennis Zhou } 678ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = start; 679ca460b3cSDennis Zhou (Facebook) block->contig_hint = contig; 680*382b88e9SDennis Zhou } else if (contig == block->contig_hint) { 681*382b88e9SDennis Zhou if (block->contig_hint_start && 682*382b88e9SDennis Zhou (!start || 683*382b88e9SDennis Zhou __ffs(start) > __ffs(block->contig_hint_start))) { 684*382b88e9SDennis Zhou /* start has a better alignment so use it */ 685268625a6SDennis Zhou (Facebook) block->contig_hint_start = start; 686*382b88e9SDennis Zhou if (start < block->scan_hint_start && 687*382b88e9SDennis Zhou block->contig_hint > block->scan_hint) 688*382b88e9SDennis Zhou block->scan_hint = 0; 689*382b88e9SDennis Zhou } else if (start > block->scan_hint_start || 690*382b88e9SDennis Zhou block->contig_hint > block->scan_hint) { 691*382b88e9SDennis Zhou /* 692*382b88e9SDennis Zhou * Knowing contig == contig_hint, update the scan_hint 693*382b88e9SDennis Zhou * if it is farther than or larger than the current 694*382b88e9SDennis Zhou * scan_hint. 695*382b88e9SDennis Zhou */ 696*382b88e9SDennis Zhou block->scan_hint_start = start; 697*382b88e9SDennis Zhou block->scan_hint = contig; 698*382b88e9SDennis Zhou } 699*382b88e9SDennis Zhou } else { 700*382b88e9SDennis Zhou /* 701*382b88e9SDennis Zhou * The region is smaller than the contig_hint. So only update 702*382b88e9SDennis Zhou * the scan_hint if it is larger than or equal and farther than 703*382b88e9SDennis Zhou * the current scan_hint. 704*382b88e9SDennis Zhou */ 705*382b88e9SDennis Zhou if ((start < block->contig_hint_start && 706*382b88e9SDennis Zhou (contig > block->scan_hint || 707*382b88e9SDennis Zhou (contig == block->scan_hint && 708*382b88e9SDennis Zhou start > block->scan_hint_start)))) { 709*382b88e9SDennis Zhou block->scan_hint_start = start; 710*382b88e9SDennis Zhou block->scan_hint = contig; 711*382b88e9SDennis Zhou } 712ca460b3cSDennis Zhou (Facebook) } 713ca460b3cSDennis Zhou (Facebook) } 714ca460b3cSDennis Zhou (Facebook) 715ca460b3cSDennis Zhou (Facebook) /** 716ca460b3cSDennis Zhou (Facebook) * pcpu_block_refresh_hint 717ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 718ca460b3cSDennis Zhou (Facebook) * @index: index of the metadata block 719ca460b3cSDennis Zhou (Facebook) * 720ca460b3cSDennis Zhou (Facebook) * Scans over the block beginning at first_free and updates the block 721ca460b3cSDennis Zhou (Facebook) * metadata accordingly. 722ca460b3cSDennis Zhou (Facebook) */ 723ca460b3cSDennis Zhou (Facebook) static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 724ca460b3cSDennis Zhou (Facebook) { 725ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *block = chunk->md_blocks + index; 726ca460b3cSDennis Zhou (Facebook) unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 727ca460b3cSDennis Zhou (Facebook) int rs, re; /* region start, region end */ 728ca460b3cSDennis Zhou (Facebook) 729ca460b3cSDennis Zhou (Facebook) /* clear hints */ 730*382b88e9SDennis Zhou block->contig_hint = block->scan_hint = 0; 731ca460b3cSDennis Zhou (Facebook) block->left_free = block->right_free = 0; 732ca460b3cSDennis Zhou (Facebook) 733ca460b3cSDennis Zhou (Facebook) /* iterate over free areas and update the contig hints */ 734ca460b3cSDennis Zhou (Facebook) pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, 735ca460b3cSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS) { 736ca460b3cSDennis Zhou (Facebook) pcpu_block_update(block, rs, re); 737ca460b3cSDennis Zhou (Facebook) } 738ca460b3cSDennis Zhou (Facebook) } 739ca460b3cSDennis Zhou (Facebook) 740ca460b3cSDennis Zhou (Facebook) /** 741ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_alloc - update hint on allocation path 742ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 743ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 744ca460b3cSDennis Zhou (Facebook) * @bits: size of request 745fc304334SDennis Zhou (Facebook) * 746fc304334SDennis Zhou (Facebook) * Updates metadata for the allocation path. The metadata only has to be 747fc304334SDennis Zhou (Facebook) * refreshed by a full scan iff the chunk's contig hint is broken. Block level 748fc304334SDennis Zhou (Facebook) * scans are required if the block's contig hint is broken. 749ca460b3cSDennis Zhou (Facebook) */ 750ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 751ca460b3cSDennis Zhou (Facebook) int bits) 752ca460b3cSDennis Zhou (Facebook) { 753b239f7daSDennis Zhou int nr_empty_pages = 0; 754ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 755ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 756ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 757ca460b3cSDennis Zhou (Facebook) 758ca460b3cSDennis Zhou (Facebook) /* 759ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 760ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 761ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 762ca460b3cSDennis Zhou (Facebook) * range. 763ca460b3cSDennis Zhou (Facebook) */ 764ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 765ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 766ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 767ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 768ca460b3cSDennis Zhou (Facebook) 769ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 770ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 771ca460b3cSDennis Zhou (Facebook) 772ca460b3cSDennis Zhou (Facebook) /* 773ca460b3cSDennis Zhou (Facebook) * Update s_block. 774fc304334SDennis Zhou (Facebook) * block->first_free must be updated if the allocation takes its place. 775fc304334SDennis Zhou (Facebook) * If the allocation breaks the contig_hint, a scan is required to 776fc304334SDennis Zhou (Facebook) * restore this hint. 777ca460b3cSDennis Zhou (Facebook) */ 778b239f7daSDennis Zhou if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 779b239f7daSDennis Zhou nr_empty_pages++; 780b239f7daSDennis Zhou 781fc304334SDennis Zhou (Facebook) if (s_off == s_block->first_free) 782fc304334SDennis Zhou (Facebook) s_block->first_free = find_next_zero_bit( 783fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, s_index), 784fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, 785fc304334SDennis Zhou (Facebook) s_off + bits); 786fc304334SDennis Zhou (Facebook) 787*382b88e9SDennis Zhou if (pcpu_region_overlap(s_block->scan_hint_start, 788*382b88e9SDennis Zhou s_block->scan_hint_start + s_block->scan_hint, 789*382b88e9SDennis Zhou s_off, 790*382b88e9SDennis Zhou s_off + bits)) 791*382b88e9SDennis Zhou s_block->scan_hint = 0; 792*382b88e9SDennis Zhou 793d9f3a01eSDennis Zhou if (pcpu_region_overlap(s_block->contig_hint_start, 794d9f3a01eSDennis Zhou s_block->contig_hint_start + 795d9f3a01eSDennis Zhou s_block->contig_hint, 796d9f3a01eSDennis Zhou s_off, 797d9f3a01eSDennis Zhou s_off + bits)) { 798fc304334SDennis Zhou (Facebook) /* block contig hint is broken - scan to fix it */ 799ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 800fc304334SDennis Zhou (Facebook) } else { 801fc304334SDennis Zhou (Facebook) /* update left and right contig manually */ 802fc304334SDennis Zhou (Facebook) s_block->left_free = min(s_block->left_free, s_off); 803fc304334SDennis Zhou (Facebook) if (s_index == e_index) 804fc304334SDennis Zhou (Facebook) s_block->right_free = min_t(int, s_block->right_free, 805fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 806fc304334SDennis Zhou (Facebook) else 807fc304334SDennis Zhou (Facebook) s_block->right_free = 0; 808fc304334SDennis Zhou (Facebook) } 809ca460b3cSDennis Zhou (Facebook) 810ca460b3cSDennis Zhou (Facebook) /* 811ca460b3cSDennis Zhou (Facebook) * Update e_block. 812ca460b3cSDennis Zhou (Facebook) */ 813ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 814b239f7daSDennis Zhou if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 815b239f7daSDennis Zhou nr_empty_pages++; 816b239f7daSDennis Zhou 817fc304334SDennis Zhou (Facebook) /* 818fc304334SDennis Zhou (Facebook) * When the allocation is across blocks, the end is along 819fc304334SDennis Zhou (Facebook) * the left part of the e_block. 820fc304334SDennis Zhou (Facebook) */ 821fc304334SDennis Zhou (Facebook) e_block->first_free = find_next_zero_bit( 822fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, e_index), 823fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, e_off); 824fc304334SDennis Zhou (Facebook) 825fc304334SDennis Zhou (Facebook) if (e_off == PCPU_BITMAP_BLOCK_BITS) { 826fc304334SDennis Zhou (Facebook) /* reset the block */ 827fc304334SDennis Zhou (Facebook) e_block++; 828fc304334SDennis Zhou (Facebook) } else { 829*382b88e9SDennis Zhou if (e_off > e_block->scan_hint_start) 830*382b88e9SDennis Zhou e_block->scan_hint = 0; 831*382b88e9SDennis Zhou 832fc304334SDennis Zhou (Facebook) if (e_off > e_block->contig_hint_start) { 833fc304334SDennis Zhou (Facebook) /* contig hint is broken - scan to fix it */ 834ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 835fc304334SDennis Zhou (Facebook) } else { 836fc304334SDennis Zhou (Facebook) e_block->left_free = 0; 837fc304334SDennis Zhou (Facebook) e_block->right_free = 838fc304334SDennis Zhou (Facebook) min_t(int, e_block->right_free, 839fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 840fc304334SDennis Zhou (Facebook) } 841fc304334SDennis Zhou (Facebook) } 842ca460b3cSDennis Zhou (Facebook) 843ca460b3cSDennis Zhou (Facebook) /* update in-between md_blocks */ 844b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 845ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 846*382b88e9SDennis Zhou block->scan_hint = 0; 847ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 848ca460b3cSDennis Zhou (Facebook) block->left_free = 0; 849ca460b3cSDennis Zhou (Facebook) block->right_free = 0; 850ca460b3cSDennis Zhou (Facebook) } 851ca460b3cSDennis Zhou (Facebook) } 852ca460b3cSDennis Zhou (Facebook) 853b239f7daSDennis Zhou if (nr_empty_pages) 854b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr_empty_pages); 855b239f7daSDennis Zhou 856fc304334SDennis Zhou (Facebook) /* 857fc304334SDennis Zhou (Facebook) * The only time a full chunk scan is required is if the chunk 858fc304334SDennis Zhou (Facebook) * contig hint is broken. Otherwise, it means a smaller space 859fc304334SDennis Zhou (Facebook) * was used and therefore the chunk contig hint is still correct. 860fc304334SDennis Zhou (Facebook) */ 861d9f3a01eSDennis Zhou if (pcpu_region_overlap(chunk->contig_bits_start, 862d9f3a01eSDennis Zhou chunk->contig_bits_start + chunk->contig_bits, 863d9f3a01eSDennis Zhou bit_off, 864d9f3a01eSDennis Zhou bit_off + bits)) 865ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 866ca460b3cSDennis Zhou (Facebook) } 867ca460b3cSDennis Zhou (Facebook) 868ca460b3cSDennis Zhou (Facebook) /** 869ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_free - updates the block hints on the free path 870ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 871ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 872ca460b3cSDennis Zhou (Facebook) * @bits: size of request 873b185cd0dSDennis Zhou (Facebook) * 874b185cd0dSDennis Zhou (Facebook) * Updates metadata for the allocation path. This avoids a blind block 875b185cd0dSDennis Zhou (Facebook) * refresh by making use of the block contig hints. If this fails, it scans 876b185cd0dSDennis Zhou (Facebook) * forward and backward to determine the extent of the free area. This is 877b185cd0dSDennis Zhou (Facebook) * capped at the boundary of blocks. 878b185cd0dSDennis Zhou (Facebook) * 879b185cd0dSDennis Zhou (Facebook) * A chunk update is triggered if a page becomes free, a block becomes free, 880b185cd0dSDennis Zhou (Facebook) * or the free spans across blocks. This tradeoff is to minimize iterating 881b185cd0dSDennis Zhou (Facebook) * over the block metadata to update chunk->contig_bits. chunk->contig_bits 882b185cd0dSDennis Zhou (Facebook) * may be off by up to a page, but it will never be more than the available 883b185cd0dSDennis Zhou (Facebook) * space. If the contig hint is contained in one block, it will be accurate. 884ca460b3cSDennis Zhou (Facebook) */ 885ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 886ca460b3cSDennis Zhou (Facebook) int bits) 887ca460b3cSDennis Zhou (Facebook) { 888b239f7daSDennis Zhou int nr_empty_pages = 0; 889ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 890ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 891ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 892b185cd0dSDennis Zhou (Facebook) int start, end; /* start and end of the whole free area */ 893ca460b3cSDennis Zhou (Facebook) 894ca460b3cSDennis Zhou (Facebook) /* 895ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 896ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 897ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 898ca460b3cSDennis Zhou (Facebook) * range. 899ca460b3cSDennis Zhou (Facebook) */ 900ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 901ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 902ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 903ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 904ca460b3cSDennis Zhou (Facebook) 905ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 906ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 907ca460b3cSDennis Zhou (Facebook) 908b185cd0dSDennis Zhou (Facebook) /* 909b185cd0dSDennis Zhou (Facebook) * Check if the freed area aligns with the block->contig_hint. 910b185cd0dSDennis Zhou (Facebook) * If it does, then the scan to find the beginning/end of the 911b185cd0dSDennis Zhou (Facebook) * larger free area can be avoided. 912b185cd0dSDennis Zhou (Facebook) * 913b185cd0dSDennis Zhou (Facebook) * start and end refer to beginning and end of the free area 914b185cd0dSDennis Zhou (Facebook) * within each their respective blocks. This is not necessarily 915b185cd0dSDennis Zhou (Facebook) * the entire free area as it may span blocks past the beginning 916b185cd0dSDennis Zhou (Facebook) * or end of the block. 917b185cd0dSDennis Zhou (Facebook) */ 918b185cd0dSDennis Zhou (Facebook) start = s_off; 919b185cd0dSDennis Zhou (Facebook) if (s_off == s_block->contig_hint + s_block->contig_hint_start) { 920b185cd0dSDennis Zhou (Facebook) start = s_block->contig_hint_start; 921b185cd0dSDennis Zhou (Facebook) } else { 922b185cd0dSDennis Zhou (Facebook) /* 923b185cd0dSDennis Zhou (Facebook) * Scan backwards to find the extent of the free area. 924b185cd0dSDennis Zhou (Facebook) * find_last_bit returns the starting bit, so if the start bit 925b185cd0dSDennis Zhou (Facebook) * is returned, that means there was no last bit and the 926b185cd0dSDennis Zhou (Facebook) * remainder of the chunk is free. 927b185cd0dSDennis Zhou (Facebook) */ 928b185cd0dSDennis Zhou (Facebook) int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), 929b185cd0dSDennis Zhou (Facebook) start); 930b185cd0dSDennis Zhou (Facebook) start = (start == l_bit) ? 0 : l_bit + 1; 931b185cd0dSDennis Zhou (Facebook) } 932b185cd0dSDennis Zhou (Facebook) 933b185cd0dSDennis Zhou (Facebook) end = e_off; 934b185cd0dSDennis Zhou (Facebook) if (e_off == e_block->contig_hint_start) 935b185cd0dSDennis Zhou (Facebook) end = e_block->contig_hint_start + e_block->contig_hint; 936b185cd0dSDennis Zhou (Facebook) else 937b185cd0dSDennis Zhou (Facebook) end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), 938b185cd0dSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, end); 939b185cd0dSDennis Zhou (Facebook) 940ca460b3cSDennis Zhou (Facebook) /* update s_block */ 941b185cd0dSDennis Zhou (Facebook) e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; 942b239f7daSDennis Zhou if (!start && e_off == PCPU_BITMAP_BLOCK_BITS) 943b239f7daSDennis Zhou nr_empty_pages++; 944b185cd0dSDennis Zhou (Facebook) pcpu_block_update(s_block, start, e_off); 945ca460b3cSDennis Zhou (Facebook) 946ca460b3cSDennis Zhou (Facebook) /* freeing in the same block */ 947ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 948ca460b3cSDennis Zhou (Facebook) /* update e_block */ 949b239f7daSDennis Zhou if (end == PCPU_BITMAP_BLOCK_BITS) 950b239f7daSDennis Zhou nr_empty_pages++; 951b185cd0dSDennis Zhou (Facebook) pcpu_block_update(e_block, 0, end); 952ca460b3cSDennis Zhou (Facebook) 953ca460b3cSDennis Zhou (Facebook) /* reset md_blocks in the middle */ 954b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 955ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 956ca460b3cSDennis Zhou (Facebook) block->first_free = 0; 957*382b88e9SDennis Zhou block->scan_hint = 0; 958ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = 0; 959ca460b3cSDennis Zhou (Facebook) block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 960ca460b3cSDennis Zhou (Facebook) block->left_free = PCPU_BITMAP_BLOCK_BITS; 961ca460b3cSDennis Zhou (Facebook) block->right_free = PCPU_BITMAP_BLOCK_BITS; 962ca460b3cSDennis Zhou (Facebook) } 963ca460b3cSDennis Zhou (Facebook) } 964ca460b3cSDennis Zhou (Facebook) 965b239f7daSDennis Zhou if (nr_empty_pages) 966b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr_empty_pages); 967b239f7daSDennis Zhou 968b185cd0dSDennis Zhou (Facebook) /* 969b239f7daSDennis Zhou * Refresh chunk metadata when the free makes a block free or spans 970b239f7daSDennis Zhou * across blocks. The contig_hint may be off by up to a page, but if 971b239f7daSDennis Zhou * the contig_hint is contained in a block, it will be accurate with 972b239f7daSDennis Zhou * the else condition below. 973b185cd0dSDennis Zhou (Facebook) */ 974b239f7daSDennis Zhou if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index) 975ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 976b185cd0dSDennis Zhou (Facebook) else 977b185cd0dSDennis Zhou (Facebook) pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start), 9788e5a2b98SDennis Zhou end - start); 979ca460b3cSDennis Zhou (Facebook) } 980ca460b3cSDennis Zhou (Facebook) 981ca460b3cSDennis Zhou (Facebook) /** 98240064aecSDennis Zhou (Facebook) * pcpu_is_populated - determines if the region is populated 98340064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 98440064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 98540064aecSDennis Zhou (Facebook) * @bits: size of area 98640064aecSDennis Zhou (Facebook) * @next_off: return value for the next offset to start searching 98740064aecSDennis Zhou (Facebook) * 98840064aecSDennis Zhou (Facebook) * For atomic allocations, check if the backing pages are populated. 98940064aecSDennis Zhou (Facebook) * 99040064aecSDennis Zhou (Facebook) * RETURNS: 99140064aecSDennis Zhou (Facebook) * Bool if the backing pages are populated. 99240064aecSDennis Zhou (Facebook) * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 99340064aecSDennis Zhou (Facebook) */ 99440064aecSDennis Zhou (Facebook) static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 99540064aecSDennis Zhou (Facebook) int *next_off) 99640064aecSDennis Zhou (Facebook) { 99740064aecSDennis Zhou (Facebook) int page_start, page_end, rs, re; 99840064aecSDennis Zhou (Facebook) 99940064aecSDennis Zhou (Facebook) page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 100040064aecSDennis Zhou (Facebook) page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 100140064aecSDennis Zhou (Facebook) 100240064aecSDennis Zhou (Facebook) rs = page_start; 100340064aecSDennis Zhou (Facebook) pcpu_next_unpop(chunk->populated, &rs, &re, page_end); 100440064aecSDennis Zhou (Facebook) if (rs >= page_end) 100540064aecSDennis Zhou (Facebook) return true; 100640064aecSDennis Zhou (Facebook) 100740064aecSDennis Zhou (Facebook) *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 100840064aecSDennis Zhou (Facebook) return false; 100940064aecSDennis Zhou (Facebook) } 101040064aecSDennis Zhou (Facebook) 101140064aecSDennis Zhou (Facebook) /** 101240064aecSDennis Zhou (Facebook) * pcpu_find_block_fit - finds the block index to start searching 101340064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 101440064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 101540064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE bytes) 101640064aecSDennis Zhou (Facebook) * @pop_only: use populated regions only 101740064aecSDennis Zhou (Facebook) * 1018b4c2116cSDennis Zhou (Facebook) * Given a chunk and an allocation spec, find the offset to begin searching 1019b4c2116cSDennis Zhou (Facebook) * for a free region. This iterates over the bitmap metadata blocks to 1020b4c2116cSDennis Zhou (Facebook) * find an offset that will be guaranteed to fit the requirements. It is 1021b4c2116cSDennis Zhou (Facebook) * not quite first fit as if the allocation does not fit in the contig hint 1022b4c2116cSDennis Zhou (Facebook) * of a block or chunk, it is skipped. This errs on the side of caution 1023b4c2116cSDennis Zhou (Facebook) * to prevent excess iteration. Poor alignment can cause the allocator to 1024b4c2116cSDennis Zhou (Facebook) * skip over blocks and chunks that have valid free areas. 1025b4c2116cSDennis Zhou (Facebook) * 102640064aecSDennis Zhou (Facebook) * RETURNS: 102740064aecSDennis Zhou (Facebook) * The offset in the bitmap to begin searching. 102840064aecSDennis Zhou (Facebook) * -1 if no offset is found. 102940064aecSDennis Zhou (Facebook) */ 103040064aecSDennis Zhou (Facebook) static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 103140064aecSDennis Zhou (Facebook) size_t align, bool pop_only) 103240064aecSDennis Zhou (Facebook) { 1033b4c2116cSDennis Zhou (Facebook) int bit_off, bits, next_off; 103440064aecSDennis Zhou (Facebook) 103513f96637SDennis Zhou (Facebook) /* 103613f96637SDennis Zhou (Facebook) * Check to see if the allocation can fit in the chunk's contig hint. 103713f96637SDennis Zhou (Facebook) * This is an optimization to prevent scanning by assuming if it 103813f96637SDennis Zhou (Facebook) * cannot fit in the global hint, there is memory pressure and creating 103913f96637SDennis Zhou (Facebook) * a new chunk would happen soon. 104013f96637SDennis Zhou (Facebook) */ 104113f96637SDennis Zhou (Facebook) bit_off = ALIGN(chunk->contig_bits_start, align) - 104213f96637SDennis Zhou (Facebook) chunk->contig_bits_start; 104313f96637SDennis Zhou (Facebook) if (bit_off + alloc_bits > chunk->contig_bits) 104413f96637SDennis Zhou (Facebook) return -1; 104513f96637SDennis Zhou (Facebook) 1046b4c2116cSDennis Zhou (Facebook) bit_off = chunk->first_bit; 1047b4c2116cSDennis Zhou (Facebook) bits = 0; 1048b4c2116cSDennis Zhou (Facebook) pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { 104940064aecSDennis Zhou (Facebook) if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 1050b4c2116cSDennis Zhou (Facebook) &next_off)) 105140064aecSDennis Zhou (Facebook) break; 105240064aecSDennis Zhou (Facebook) 1053b4c2116cSDennis Zhou (Facebook) bit_off = next_off; 105440064aecSDennis Zhou (Facebook) bits = 0; 105540064aecSDennis Zhou (Facebook) } 105640064aecSDennis Zhou (Facebook) 105740064aecSDennis Zhou (Facebook) if (bit_off == pcpu_chunk_map_bits(chunk)) 105840064aecSDennis Zhou (Facebook) return -1; 105940064aecSDennis Zhou (Facebook) 106040064aecSDennis Zhou (Facebook) return bit_off; 106140064aecSDennis Zhou (Facebook) } 106240064aecSDennis Zhou (Facebook) 106340064aecSDennis Zhou (Facebook) /** 106440064aecSDennis Zhou (Facebook) * pcpu_alloc_area - allocates an area from a pcpu_chunk 106540064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 106640064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 106740064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 106840064aecSDennis Zhou (Facebook) * @start: bit_off to start searching 106940064aecSDennis Zhou (Facebook) * 107040064aecSDennis Zhou (Facebook) * This function takes in a @start offset to begin searching to fit an 1071b4c2116cSDennis Zhou (Facebook) * allocation of @alloc_bits with alignment @align. It needs to scan 1072b4c2116cSDennis Zhou (Facebook) * the allocation map because if it fits within the block's contig hint, 1073b4c2116cSDennis Zhou (Facebook) * @start will be block->first_free. This is an attempt to fill the 1074b4c2116cSDennis Zhou (Facebook) * allocation prior to breaking the contig hint. The allocation and 1075b4c2116cSDennis Zhou (Facebook) * boundary maps are updated accordingly if it confirms a valid 1076b4c2116cSDennis Zhou (Facebook) * free area. 107740064aecSDennis Zhou (Facebook) * 107840064aecSDennis Zhou (Facebook) * RETURNS: 107940064aecSDennis Zhou (Facebook) * Allocated addr offset in @chunk on success. 108040064aecSDennis Zhou (Facebook) * -1 if no matching area is found. 108140064aecSDennis Zhou (Facebook) */ 108240064aecSDennis Zhou (Facebook) static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 108340064aecSDennis Zhou (Facebook) size_t align, int start) 108440064aecSDennis Zhou (Facebook) { 108540064aecSDennis Zhou (Facebook) size_t align_mask = (align) ? (align - 1) : 0; 108640064aecSDennis Zhou (Facebook) int bit_off, end, oslot; 10879f7dcf22STejun Heo 10884f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 10894f996e23STejun Heo 109040064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1091833af842STejun Heo 1092833af842STejun Heo /* 109340064aecSDennis Zhou (Facebook) * Search to find a fit. 1094833af842STejun Heo */ 10958c43004aSDennis Zhou end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS, 10968c43004aSDennis Zhou pcpu_chunk_map_bits(chunk)); 109740064aecSDennis Zhou (Facebook) bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, 109840064aecSDennis Zhou (Facebook) alloc_bits, align_mask); 109940064aecSDennis Zhou (Facebook) if (bit_off >= end) 1100a16037c8STejun Heo return -1; 1101a16037c8STejun Heo 110240064aecSDennis Zhou (Facebook) /* update alloc map */ 110340064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 1104a16037c8STejun Heo 110540064aecSDennis Zhou (Facebook) /* update boundary map */ 110640064aecSDennis Zhou (Facebook) set_bit(bit_off, chunk->bound_map); 110740064aecSDennis Zhou (Facebook) bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 110840064aecSDennis Zhou (Facebook) set_bit(bit_off + alloc_bits, chunk->bound_map); 1109a16037c8STejun Heo 111040064aecSDennis Zhou (Facebook) chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 111140064aecSDennis Zhou (Facebook) 111286b442fbSDennis Zhou (Facebook) /* update first free bit */ 111386b442fbSDennis Zhou (Facebook) if (bit_off == chunk->first_bit) 111486b442fbSDennis Zhou (Facebook) chunk->first_bit = find_next_zero_bit( 111586b442fbSDennis Zhou (Facebook) chunk->alloc_map, 111686b442fbSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk), 111786b442fbSDennis Zhou (Facebook) bit_off + alloc_bits); 111886b442fbSDennis Zhou (Facebook) 1119ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 112040064aecSDennis Zhou (Facebook) 112140064aecSDennis Zhou (Facebook) pcpu_chunk_relocate(chunk, oslot); 112240064aecSDennis Zhou (Facebook) 112340064aecSDennis Zhou (Facebook) return bit_off * PCPU_MIN_ALLOC_SIZE; 1124a16037c8STejun Heo } 1125a16037c8STejun Heo 1126a16037c8STejun Heo /** 112740064aecSDennis Zhou (Facebook) * pcpu_free_area - frees the corresponding offset 1128fbf59bc9STejun Heo * @chunk: chunk of interest 112940064aecSDennis Zhou (Facebook) * @off: addr offset into chunk 1130fbf59bc9STejun Heo * 113140064aecSDennis Zhou (Facebook) * This function determines the size of an allocation to free using 113240064aecSDennis Zhou (Facebook) * the boundary bitmap and clears the allocation map. 1133fbf59bc9STejun Heo */ 113440064aecSDennis Zhou (Facebook) static void pcpu_free_area(struct pcpu_chunk *chunk, int off) 1135fbf59bc9STejun Heo { 113640064aecSDennis Zhou (Facebook) int bit_off, bits, end, oslot; 1137fbf59bc9STejun Heo 11385ccd30e4SDennis Zhou lockdep_assert_held(&pcpu_lock); 113930a5b536SDennis Zhou pcpu_stats_area_dealloc(chunk); 11405ccd30e4SDennis Zhou 114140064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1142723ad1d9SAl Viro 114340064aecSDennis Zhou (Facebook) bit_off = off / PCPU_MIN_ALLOC_SIZE; 1144fbf59bc9STejun Heo 114540064aecSDennis Zhou (Facebook) /* find end index */ 114640064aecSDennis Zhou (Facebook) end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 114740064aecSDennis Zhou (Facebook) bit_off + 1); 114840064aecSDennis Zhou (Facebook) bits = end - bit_off; 114940064aecSDennis Zhou (Facebook) bitmap_clear(chunk->alloc_map, bit_off, bits); 11503d331ad7SAl Viro 115140064aecSDennis Zhou (Facebook) /* update metadata */ 115240064aecSDennis Zhou (Facebook) chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; 1153fbf59bc9STejun Heo 115486b442fbSDennis Zhou (Facebook) /* update first free bit */ 115586b442fbSDennis Zhou (Facebook) chunk->first_bit = min(chunk->first_bit, bit_off); 115686b442fbSDennis Zhou (Facebook) 1157ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_free(chunk, bit_off, bits); 1158b539b87fSTejun Heo 1159fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 1160fbf59bc9STejun Heo } 1161fbf59bc9STejun Heo 1162ca460b3cSDennis Zhou (Facebook) static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 1163ca460b3cSDennis Zhou (Facebook) { 1164ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *md_block; 1165ca460b3cSDennis Zhou (Facebook) 1166ca460b3cSDennis Zhou (Facebook) for (md_block = chunk->md_blocks; 1167ca460b3cSDennis Zhou (Facebook) md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 1168ca460b3cSDennis Zhou (Facebook) md_block++) { 1169*382b88e9SDennis Zhou md_block->scan_hint = 0; 1170ca460b3cSDennis Zhou (Facebook) md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 1171ca460b3cSDennis Zhou (Facebook) md_block->left_free = PCPU_BITMAP_BLOCK_BITS; 1172ca460b3cSDennis Zhou (Facebook) md_block->right_free = PCPU_BITMAP_BLOCK_BITS; 1173ca460b3cSDennis Zhou (Facebook) } 1174ca460b3cSDennis Zhou (Facebook) } 1175ca460b3cSDennis Zhou (Facebook) 117640064aecSDennis Zhou (Facebook) /** 117740064aecSDennis Zhou (Facebook) * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 117840064aecSDennis Zhou (Facebook) * @tmp_addr: the start of the region served 117940064aecSDennis Zhou (Facebook) * @map_size: size of the region served 118040064aecSDennis Zhou (Facebook) * 118140064aecSDennis Zhou (Facebook) * This is responsible for creating the chunks that serve the first chunk. The 118240064aecSDennis Zhou (Facebook) * base_addr is page aligned down of @tmp_addr while the region end is page 118340064aecSDennis Zhou (Facebook) * aligned up. Offsets are kept track of to determine the region served. All 118440064aecSDennis Zhou (Facebook) * this is done to appease the bitmap allocator in avoiding partial blocks. 118540064aecSDennis Zhou (Facebook) * 118640064aecSDennis Zhou (Facebook) * RETURNS: 118740064aecSDennis Zhou (Facebook) * Chunk serving the region at @tmp_addr of @map_size. 118840064aecSDennis Zhou (Facebook) */ 1189c0ebfdc3SDennis Zhou (Facebook) static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 119040064aecSDennis Zhou (Facebook) int map_size) 119110edf5b0SDennis Zhou (Facebook) { 119210edf5b0SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 1193ca460b3cSDennis Zhou (Facebook) unsigned long aligned_addr, lcm_align; 119440064aecSDennis Zhou (Facebook) int start_offset, offset_bits, region_size, region_bits; 1195f655f405SMike Rapoport size_t alloc_size; 1196c0ebfdc3SDennis Zhou (Facebook) 1197c0ebfdc3SDennis Zhou (Facebook) /* region calculations */ 1198c0ebfdc3SDennis Zhou (Facebook) aligned_addr = tmp_addr & PAGE_MASK; 1199c0ebfdc3SDennis Zhou (Facebook) 1200c0ebfdc3SDennis Zhou (Facebook) start_offset = tmp_addr - aligned_addr; 12016b9d7c8eSDennis Zhou (Facebook) 1202ca460b3cSDennis Zhou (Facebook) /* 1203ca460b3cSDennis Zhou (Facebook) * Align the end of the region with the LCM of PAGE_SIZE and 1204ca460b3cSDennis Zhou (Facebook) * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of 1205ca460b3cSDennis Zhou (Facebook) * the other. 1206ca460b3cSDennis Zhou (Facebook) */ 1207ca460b3cSDennis Zhou (Facebook) lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); 1208ca460b3cSDennis Zhou (Facebook) region_size = ALIGN(start_offset + map_size, lcm_align); 120910edf5b0SDennis Zhou (Facebook) 1210c0ebfdc3SDennis Zhou (Facebook) /* allocate chunk */ 1211f655f405SMike Rapoport alloc_size = sizeof(struct pcpu_chunk) + 1212f655f405SMike Rapoport BITS_TO_LONGS(region_size >> PAGE_SHIFT); 1213f655f405SMike Rapoport chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1214f655f405SMike Rapoport if (!chunk) 1215f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1216f655f405SMike Rapoport alloc_size); 1217c0ebfdc3SDennis Zhou (Facebook) 121810edf5b0SDennis Zhou (Facebook) INIT_LIST_HEAD(&chunk->list); 1219c0ebfdc3SDennis Zhou (Facebook) 1220c0ebfdc3SDennis Zhou (Facebook) chunk->base_addr = (void *)aligned_addr; 122110edf5b0SDennis Zhou (Facebook) chunk->start_offset = start_offset; 12226b9d7c8eSDennis Zhou (Facebook) chunk->end_offset = region_size - chunk->start_offset - map_size; 1223c0ebfdc3SDennis Zhou (Facebook) 12248ab16c43SDennis Zhou (Facebook) chunk->nr_pages = region_size >> PAGE_SHIFT; 122540064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 1226c0ebfdc3SDennis Zhou (Facebook) 1227f655f405SMike Rapoport alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]); 1228f655f405SMike Rapoport chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1229f655f405SMike Rapoport if (!chunk->alloc_map) 1230f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1231f655f405SMike Rapoport alloc_size); 1232f655f405SMike Rapoport 1233f655f405SMike Rapoport alloc_size = 1234f655f405SMike Rapoport BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]); 1235f655f405SMike Rapoport chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1236f655f405SMike Rapoport if (!chunk->bound_map) 1237f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1238f655f405SMike Rapoport alloc_size); 1239f655f405SMike Rapoport 1240f655f405SMike Rapoport alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]); 1241f655f405SMike Rapoport chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1242f655f405SMike Rapoport if (!chunk->md_blocks) 1243f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1244f655f405SMike Rapoport alloc_size); 1245f655f405SMike Rapoport 1246ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 124710edf5b0SDennis Zhou (Facebook) 124810edf5b0SDennis Zhou (Facebook) /* manage populated page bitmap */ 124910edf5b0SDennis Zhou (Facebook) chunk->immutable = true; 12508ab16c43SDennis Zhou (Facebook) bitmap_fill(chunk->populated, chunk->nr_pages); 12518ab16c43SDennis Zhou (Facebook) chunk->nr_populated = chunk->nr_pages; 1252b239f7daSDennis Zhou chunk->nr_empty_pop_pages = chunk->nr_pages; 125310edf5b0SDennis Zhou (Facebook) 125440064aecSDennis Zhou (Facebook) chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; 125540064aecSDennis Zhou (Facebook) chunk->free_bytes = map_size; 1256c0ebfdc3SDennis Zhou (Facebook) 1257c0ebfdc3SDennis Zhou (Facebook) if (chunk->start_offset) { 1258c0ebfdc3SDennis Zhou (Facebook) /* hide the beginning of the bitmap */ 125940064aecSDennis Zhou (Facebook) offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 126040064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 0, offset_bits); 126140064aecSDennis Zhou (Facebook) set_bit(0, chunk->bound_map); 126240064aecSDennis Zhou (Facebook) set_bit(offset_bits, chunk->bound_map); 1263ca460b3cSDennis Zhou (Facebook) 126486b442fbSDennis Zhou (Facebook) chunk->first_bit = offset_bits; 126586b442fbSDennis Zhou (Facebook) 1266ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 1267c0ebfdc3SDennis Zhou (Facebook) } 1268c0ebfdc3SDennis Zhou (Facebook) 12696b9d7c8eSDennis Zhou (Facebook) if (chunk->end_offset) { 12706b9d7c8eSDennis Zhou (Facebook) /* hide the end of the bitmap */ 127140064aecSDennis Zhou (Facebook) offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 127240064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 127340064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk) - offset_bits, 127440064aecSDennis Zhou (Facebook) offset_bits); 127540064aecSDennis Zhou (Facebook) set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 127640064aecSDennis Zhou (Facebook) chunk->bound_map); 127740064aecSDennis Zhou (Facebook) set_bit(region_bits, chunk->bound_map); 12786b9d7c8eSDennis Zhou (Facebook) 1279ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 1280ca460b3cSDennis Zhou (Facebook) - offset_bits, offset_bits); 1281ca460b3cSDennis Zhou (Facebook) } 128240064aecSDennis Zhou (Facebook) 128310edf5b0SDennis Zhou (Facebook) return chunk; 128410edf5b0SDennis Zhou (Facebook) } 128510edf5b0SDennis Zhou (Facebook) 128647504ee0SDennis Zhou static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) 12876081089fSTejun Heo { 12886081089fSTejun Heo struct pcpu_chunk *chunk; 128940064aecSDennis Zhou (Facebook) int region_bits; 12906081089fSTejun Heo 129147504ee0SDennis Zhou chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); 12926081089fSTejun Heo if (!chunk) 12936081089fSTejun Heo return NULL; 12946081089fSTejun Heo 12956081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 1296c0ebfdc3SDennis Zhou (Facebook) chunk->nr_pages = pcpu_unit_pages; 129740064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 129840064aecSDennis Zhou (Facebook) 129940064aecSDennis Zhou (Facebook) chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 130047504ee0SDennis Zhou sizeof(chunk->alloc_map[0]), gfp); 130140064aecSDennis Zhou (Facebook) if (!chunk->alloc_map) 130240064aecSDennis Zhou (Facebook) goto alloc_map_fail; 130340064aecSDennis Zhou (Facebook) 130440064aecSDennis Zhou (Facebook) chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 130547504ee0SDennis Zhou sizeof(chunk->bound_map[0]), gfp); 130640064aecSDennis Zhou (Facebook) if (!chunk->bound_map) 130740064aecSDennis Zhou (Facebook) goto bound_map_fail; 130840064aecSDennis Zhou (Facebook) 1309ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 131047504ee0SDennis Zhou sizeof(chunk->md_blocks[0]), gfp); 1311ca460b3cSDennis Zhou (Facebook) if (!chunk->md_blocks) 1312ca460b3cSDennis Zhou (Facebook) goto md_blocks_fail; 1313ca460b3cSDennis Zhou (Facebook) 1314ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 1315ca460b3cSDennis Zhou (Facebook) 131640064aecSDennis Zhou (Facebook) /* init metadata */ 131740064aecSDennis Zhou (Facebook) chunk->contig_bits = region_bits; 131840064aecSDennis Zhou (Facebook) chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 1319c0ebfdc3SDennis Zhou (Facebook) 13206081089fSTejun Heo return chunk; 132140064aecSDennis Zhou (Facebook) 1322ca460b3cSDennis Zhou (Facebook) md_blocks_fail: 1323ca460b3cSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 132440064aecSDennis Zhou (Facebook) bound_map_fail: 132540064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 132640064aecSDennis Zhou (Facebook) alloc_map_fail: 132740064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk); 132840064aecSDennis Zhou (Facebook) 132940064aecSDennis Zhou (Facebook) return NULL; 13306081089fSTejun Heo } 13316081089fSTejun Heo 13326081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 13336081089fSTejun Heo { 13346081089fSTejun Heo if (!chunk) 13356081089fSTejun Heo return; 13366685b357SMike Rapoport pcpu_mem_free(chunk->md_blocks); 133740064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 133840064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 13391d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 13406081089fSTejun Heo } 13416081089fSTejun Heo 1342b539b87fSTejun Heo /** 1343b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 1344b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 1345b539b87fSTejun Heo * @page_start: the start page 1346b539b87fSTejun Heo * @page_end: the end page 1347b539b87fSTejun Heo * 1348b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1349b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 1350b539b87fSTejun Heo * successful population. 135140064aecSDennis Zhou (Facebook) * 135240064aecSDennis Zhou (Facebook) * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it 135340064aecSDennis Zhou (Facebook) * is to serve an allocation in that area. 1354b539b87fSTejun Heo */ 135540064aecSDennis Zhou (Facebook) static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 1356b239f7daSDennis Zhou int page_end) 1357b539b87fSTejun Heo { 1358b539b87fSTejun Heo int nr = page_end - page_start; 1359b539b87fSTejun Heo 1360b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1361b539b87fSTejun Heo 1362b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 1363b539b87fSTejun Heo chunk->nr_populated += nr; 13647e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += nr; 136540064aecSDennis Zhou (Facebook) 1366b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr); 136740064aecSDennis Zhou (Facebook) } 1368b539b87fSTejun Heo 1369b539b87fSTejun Heo /** 1370b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 1371b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 1372b539b87fSTejun Heo * @page_start: the start page 1373b539b87fSTejun Heo * @page_end: the end page 1374b539b87fSTejun Heo * 1375b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1376b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 1377b539b87fSTejun Heo * each successful depopulation. 1378b539b87fSTejun Heo */ 1379b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1380b539b87fSTejun Heo int page_start, int page_end) 1381b539b87fSTejun Heo { 1382b539b87fSTejun Heo int nr = page_end - page_start; 1383b539b87fSTejun Heo 1384b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1385b539b87fSTejun Heo 1386b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 1387b539b87fSTejun Heo chunk->nr_populated -= nr; 13887e8a6304SDennis Zhou (Facebook) pcpu_nr_populated -= nr; 1389b239f7daSDennis Zhou 1390b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr); 1391b539b87fSTejun Heo } 1392b539b87fSTejun Heo 1393fbf59bc9STejun Heo /* 13949f645532STejun Heo * Chunk management implementation. 1395fbf59bc9STejun Heo * 13969f645532STejun Heo * To allow different implementations, chunk alloc/free and 13979f645532STejun Heo * [de]population are implemented in a separate file which is pulled 13989f645532STejun Heo * into this file and compiled together. The following functions 13999f645532STejun Heo * should be implemented. 1400ccea34b5STejun Heo * 14019f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 14029f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 14039f645532STejun Heo * pcpu_create_chunk - create a new chunk 14049f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 14059f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 14069f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1407fbf59bc9STejun Heo */ 140815d9f3d1SDennis Zhou static int pcpu_populate_chunk(struct pcpu_chunk *chunk, 140947504ee0SDennis Zhou int page_start, int page_end, gfp_t gfp); 141015d9f3d1SDennis Zhou static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, 141115d9f3d1SDennis Zhou int page_start, int page_end); 141247504ee0SDennis Zhou static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); 14139f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 14149f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 14159f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1416fbf59bc9STejun Heo 1417b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 1418b0c9778bSTejun Heo #include "percpu-km.c" 1419b0c9778bSTejun Heo #else 14209f645532STejun Heo #include "percpu-vm.c" 1421b0c9778bSTejun Heo #endif 1422fbf59bc9STejun Heo 1423fbf59bc9STejun Heo /** 142488999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 142588999a89STejun Heo * @addr: address for which the chunk needs to be determined. 142688999a89STejun Heo * 1427c0ebfdc3SDennis Zhou (Facebook) * This is an internal function that handles all but static allocations. 1428c0ebfdc3SDennis Zhou (Facebook) * Static percpu address values should never be passed into the allocator. 1429c0ebfdc3SDennis Zhou (Facebook) * 143088999a89STejun Heo * RETURNS: 143188999a89STejun Heo * The address of the found chunk. 143288999a89STejun Heo */ 143388999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 143488999a89STejun Heo { 1435c0ebfdc3SDennis Zhou (Facebook) /* is it in the dynamic region (first chunk)? */ 1436560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1437c0ebfdc3SDennis Zhou (Facebook) return pcpu_first_chunk; 1438c0ebfdc3SDennis Zhou (Facebook) 1439c0ebfdc3SDennis Zhou (Facebook) /* is it in the reserved region? */ 1440560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 144188999a89STejun Heo return pcpu_reserved_chunk; 144288999a89STejun Heo 144388999a89STejun Heo /* 144488999a89STejun Heo * The address is relative to unit0 which might be unused and 144588999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 144688999a89STejun Heo * current processor before looking it up in the vmalloc 144788999a89STejun Heo * space. Note that any possible cpu id can be used here, so 144888999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 144988999a89STejun Heo */ 145088999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 14519f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 145288999a89STejun Heo } 145388999a89STejun Heo 145488999a89STejun Heo /** 1455edcb4639STejun Heo * pcpu_alloc - the percpu allocator 1456cae3aeb8STejun Heo * @size: size of area to allocate in bytes 1457fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 1458edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 14595835d96eSTejun Heo * @gfp: allocation flags 1460fbf59bc9STejun Heo * 14615835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 14620ea7eeecSDaniel Borkmann * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN 14630ea7eeecSDaniel Borkmann * then no warning will be triggered on invalid or failed allocation 14640ea7eeecSDaniel Borkmann * requests. 1465fbf59bc9STejun Heo * 1466fbf59bc9STejun Heo * RETURNS: 1467fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1468fbf59bc9STejun Heo */ 14695835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 14705835d96eSTejun Heo gfp_t gfp) 1471fbf59bc9STejun Heo { 1472554fef1cSDennis Zhou /* whitelisted flags that can be passed to the backing allocators */ 1473554fef1cSDennis Zhou gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN); 14740ea7eeecSDaniel Borkmann bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 14750ea7eeecSDaniel Borkmann bool do_warn = !(gfp & __GFP_NOWARN); 1476f2badb0cSTejun Heo static int warn_limit = 10; 14778744d859SDennis Zhou struct pcpu_chunk *chunk, *next; 1478f2badb0cSTejun Heo const char *err; 147940064aecSDennis Zhou (Facebook) int slot, off, cpu, ret; 1480403a91b1SJiri Kosina unsigned long flags; 1481f528f0b8SCatalin Marinas void __percpu *ptr; 148240064aecSDennis Zhou (Facebook) size_t bits, bit_align; 1483fbf59bc9STejun Heo 1484723ad1d9SAl Viro /* 148540064aecSDennis Zhou (Facebook) * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 148640064aecSDennis Zhou (Facebook) * therefore alignment must be a minimum of that many bytes. 148740064aecSDennis Zhou (Facebook) * An allocation may have internal fragmentation from rounding up 148840064aecSDennis Zhou (Facebook) * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1489723ad1d9SAl Viro */ 1490d2f3c384SDennis Zhou (Facebook) if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1491d2f3c384SDennis Zhou (Facebook) align = PCPU_MIN_ALLOC_SIZE; 1492723ad1d9SAl Viro 1493d2f3c384SDennis Zhou (Facebook) size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 149440064aecSDennis Zhou (Facebook) bits = size >> PCPU_MIN_ALLOC_SHIFT; 149540064aecSDennis Zhou (Facebook) bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 14962f69fa82SViro 14973ca45a46Szijun_hu if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 14983ca45a46Szijun_hu !is_power_of_2(align))) { 14990ea7eeecSDaniel Borkmann WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1500756a025fSJoe Perches size, align); 1501fbf59bc9STejun Heo return NULL; 1502fbf59bc9STejun Heo } 1503fbf59bc9STejun Heo 1504f52ba1feSKirill Tkhai if (!is_atomic) { 1505f52ba1feSKirill Tkhai /* 1506f52ba1feSKirill Tkhai * pcpu_balance_workfn() allocates memory under this mutex, 1507f52ba1feSKirill Tkhai * and it may wait for memory reclaim. Allow current task 1508f52ba1feSKirill Tkhai * to become OOM victim, in case of memory pressure. 1509f52ba1feSKirill Tkhai */ 1510f52ba1feSKirill Tkhai if (gfp & __GFP_NOFAIL) 15116710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 1512f52ba1feSKirill Tkhai else if (mutex_lock_killable(&pcpu_alloc_mutex)) 1513f52ba1feSKirill Tkhai return NULL; 1514f52ba1feSKirill Tkhai } 15156710e594STejun Heo 1516403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1517fbf59bc9STejun Heo 1518edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 1519edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 1520edcb4639STejun Heo chunk = pcpu_reserved_chunk; 1521833af842STejun Heo 152240064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 152340064aecSDennis Zhou (Facebook) if (off < 0) { 1524833af842STejun Heo err = "alloc from reserved chunk failed"; 1525ccea34b5STejun Heo goto fail_unlock; 1526f2badb0cSTejun Heo } 1527833af842STejun Heo 152840064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1529edcb4639STejun Heo if (off >= 0) 1530edcb4639STejun Heo goto area_found; 1531833af842STejun Heo 1532f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 1533ccea34b5STejun Heo goto fail_unlock; 1534edcb4639STejun Heo } 1535edcb4639STejun Heo 1536ccea34b5STejun Heo restart: 1537edcb4639STejun Heo /* search through normal chunks */ 1538fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 15398744d859SDennis Zhou list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) { 154040064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, 154140064aecSDennis Zhou (Facebook) is_atomic); 15428744d859SDennis Zhou if (off < 0) { 15438744d859SDennis Zhou if (slot < PCPU_SLOT_FAIL_THRESHOLD) 15448744d859SDennis Zhou pcpu_chunk_move(chunk, 0); 1545fbf59bc9STejun Heo continue; 15468744d859SDennis Zhou } 1547ccea34b5STejun Heo 154840064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1549fbf59bc9STejun Heo if (off >= 0) 1550fbf59bc9STejun Heo goto area_found; 155140064aecSDennis Zhou (Facebook) 1552fbf59bc9STejun Heo } 1553fbf59bc9STejun Heo } 1554fbf59bc9STejun Heo 1555403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1556ccea34b5STejun Heo 1557b38d08f3STejun Heo /* 1558b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 1559b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 1560b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 1561b38d08f3STejun Heo */ 156211df02bfSDennis Zhou if (is_atomic) { 156311df02bfSDennis Zhou err = "atomic alloc failed, no space left"; 15645835d96eSTejun Heo goto fail; 156511df02bfSDennis Zhou } 15665835d96eSTejun Heo 1567b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 1568554fef1cSDennis Zhou chunk = pcpu_create_chunk(pcpu_gfp); 1569f2badb0cSTejun Heo if (!chunk) { 1570f2badb0cSTejun Heo err = "failed to allocate new chunk"; 1571b38d08f3STejun Heo goto fail; 1572f2badb0cSTejun Heo } 1573ccea34b5STejun Heo 1574403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1575fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 1576b38d08f3STejun Heo } else { 1577b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1578b38d08f3STejun Heo } 1579b38d08f3STejun Heo 1580ccea34b5STejun Heo goto restart; 1581fbf59bc9STejun Heo 1582fbf59bc9STejun Heo area_found: 158330a5b536SDennis Zhou pcpu_stats_area_alloc(chunk, size); 1584403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1585ccea34b5STejun Heo 1586dca49645STejun Heo /* populate if not all pages are already there */ 15875835d96eSTejun Heo if (!is_atomic) { 1588e04d3208STejun Heo int page_start, page_end, rs, re; 1589e04d3208STejun Heo 1590dca49645STejun Heo page_start = PFN_DOWN(off); 1591dca49645STejun Heo page_end = PFN_UP(off + size); 1592dca49645STejun Heo 159391e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 159491e914c5SDennis Zhou (Facebook) page_start, page_end) { 1595dca49645STejun Heo WARN_ON(chunk->immutable); 1596dca49645STejun Heo 1597554fef1cSDennis Zhou ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp); 1598b38d08f3STejun Heo 1599403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1600b38d08f3STejun Heo if (ret) { 160140064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1602f2badb0cSTejun Heo err = "failed to populate"; 1603ccea34b5STejun Heo goto fail_unlock; 1604fbf59bc9STejun Heo } 1605b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, re); 1606b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1607dca49645STejun Heo } 1608dca49645STejun Heo 1609ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1610e04d3208STejun Heo } 1611ccea34b5STejun Heo 16121a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 16131a4d7607STejun Heo pcpu_schedule_balance_work(); 16141a4d7607STejun Heo 1615dca49645STejun Heo /* clear the areas and return address relative to base address */ 1616dca49645STejun Heo for_each_possible_cpu(cpu) 1617dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1618dca49645STejun Heo 1619f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 16208a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1621df95e795SDennis Zhou 1622df95e795SDennis Zhou trace_percpu_alloc_percpu(reserved, is_atomic, size, align, 1623df95e795SDennis Zhou chunk->base_addr, off, ptr); 1624df95e795SDennis Zhou 1625f528f0b8SCatalin Marinas return ptr; 1626ccea34b5STejun Heo 1627ccea34b5STejun Heo fail_unlock: 1628403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1629b38d08f3STejun Heo fail: 1630df95e795SDennis Zhou trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1631df95e795SDennis Zhou 16320ea7eeecSDaniel Borkmann if (!is_atomic && do_warn && warn_limit) { 1633870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 16345835d96eSTejun Heo size, align, is_atomic, err); 1635f2badb0cSTejun Heo dump_stack(); 1636f2badb0cSTejun Heo if (!--warn_limit) 1637870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1638f2badb0cSTejun Heo } 16391a4d7607STejun Heo if (is_atomic) { 16401a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 16411a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 16421a4d7607STejun Heo pcpu_schedule_balance_work(); 16436710e594STejun Heo } else { 16446710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 16451a4d7607STejun Heo } 1646ccea34b5STejun Heo return NULL; 1647fbf59bc9STejun Heo } 1648edcb4639STejun Heo 1649edcb4639STejun Heo /** 16505835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1651edcb4639STejun Heo * @size: size of area to allocate in bytes 1652edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 16535835d96eSTejun Heo * @gfp: allocation flags 1654edcb4639STejun Heo * 16555835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 16565835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 16570ea7eeecSDaniel Borkmann * be called from any context but is a lot more likely to fail. If @gfp 16580ea7eeecSDaniel Borkmann * has __GFP_NOWARN then no warning will be triggered on invalid or failed 16590ea7eeecSDaniel Borkmann * allocation requests. 1660ccea34b5STejun Heo * 1661edcb4639STejun Heo * RETURNS: 1662edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1663edcb4639STejun Heo */ 16645835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 16655835d96eSTejun Heo { 16665835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 16675835d96eSTejun Heo } 16685835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 16695835d96eSTejun Heo 16705835d96eSTejun Heo /** 16715835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 16725835d96eSTejun Heo * @size: size of area to allocate in bytes 16735835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 16745835d96eSTejun Heo * 16755835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 16765835d96eSTejun Heo */ 167743cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1678edcb4639STejun Heo { 16795835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1680edcb4639STejun Heo } 1681fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1682fbf59bc9STejun Heo 1683edcb4639STejun Heo /** 1684edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1685edcb4639STejun Heo * @size: size of area to allocate in bytes 1686edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1687edcb4639STejun Heo * 16889329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 16899329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 16909329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 16919329ba97STejun Heo * Might trigger writeouts. 1692edcb4639STejun Heo * 1693ccea34b5STejun Heo * CONTEXT: 1694ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1695ccea34b5STejun Heo * 1696edcb4639STejun Heo * RETURNS: 1697edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1698edcb4639STejun Heo */ 169943cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1700edcb4639STejun Heo { 17015835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1702edcb4639STejun Heo } 1703edcb4639STejun Heo 1704a56dbddfSTejun Heo /** 17051a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1706a56dbddfSTejun Heo * @work: unused 1707a56dbddfSTejun Heo * 170847504ee0SDennis Zhou * Reclaim all fully free chunks except for the first one. This is also 170947504ee0SDennis Zhou * responsible for maintaining the pool of empty populated pages. However, 171047504ee0SDennis Zhou * it is possible that this is called when physical memory is scarce causing 171147504ee0SDennis Zhou * OOM killer to be triggered. We should avoid doing so until an actual 171247504ee0SDennis Zhou * allocation causes the failure as it is possible that requests can be 171347504ee0SDennis Zhou * serviced from already backed regions. 1714a56dbddfSTejun Heo */ 1715fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1716fbf59bc9STejun Heo { 171747504ee0SDennis Zhou /* gfp flags passed to underlying allocators */ 1718554fef1cSDennis Zhou const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN; 1719fe6bd8c3STejun Heo LIST_HEAD(to_free); 1720fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1721a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 17221a4d7607STejun Heo int slot, nr_to_pop, ret; 1723a56dbddfSTejun Heo 17241a4d7607STejun Heo /* 17251a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 17261a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 17271a4d7607STejun Heo */ 1728ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1729ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1730a56dbddfSTejun Heo 1731fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 17328d408b4bSTejun Heo WARN_ON(chunk->immutable); 1733a56dbddfSTejun Heo 1734a56dbddfSTejun Heo /* spare the first one */ 1735fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1736a56dbddfSTejun Heo continue; 1737a56dbddfSTejun Heo 1738fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1739a56dbddfSTejun Heo } 1740a56dbddfSTejun Heo 1741ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1742a56dbddfSTejun Heo 1743fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1744a93ace48STejun Heo int rs, re; 1745dca49645STejun Heo 174691e914c5SDennis Zhou (Facebook) pcpu_for_each_pop_region(chunk->populated, rs, re, 0, 174791e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1748a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1749b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1750b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1751b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1752a93ace48STejun Heo } 17536081089fSTejun Heo pcpu_destroy_chunk(chunk); 1754accd4f36SEric Dumazet cond_resched(); 1755fbf59bc9STejun Heo } 1756971f3918STejun Heo 17571a4d7607STejun Heo /* 17581a4d7607STejun Heo * Ensure there are certain number of free populated pages for 17591a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 17601a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 17611a4d7607STejun Heo * failed previously, always populate the maximum amount. This 17621a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 17631a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 17641a4d7607STejun Heo * something we support properly and can be highly unreliable and 17651a4d7607STejun Heo * inefficient. 17661a4d7607STejun Heo */ 17671a4d7607STejun Heo retry_pop: 17681a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 17691a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 17701a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 17711a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 17721a4d7607STejun Heo } else { 17731a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 17741a4d7607STejun Heo pcpu_nr_empty_pop_pages, 17751a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 17761a4d7607STejun Heo } 17771a4d7607STejun Heo 17781a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 17791a4d7607STejun Heo int nr_unpop = 0, rs, re; 17801a4d7607STejun Heo 17811a4d7607STejun Heo if (!nr_to_pop) 17821a4d7607STejun Heo break; 17831a4d7607STejun Heo 17841a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 17851a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 17868ab16c43SDennis Zhou (Facebook) nr_unpop = chunk->nr_pages - chunk->nr_populated; 17871a4d7607STejun Heo if (nr_unpop) 17881a4d7607STejun Heo break; 17891a4d7607STejun Heo } 17901a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 17911a4d7607STejun Heo 17921a4d7607STejun Heo if (!nr_unpop) 17931a4d7607STejun Heo continue; 17941a4d7607STejun Heo 17951a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 179691e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, 179791e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 17981a4d7607STejun Heo int nr = min(re - rs, nr_to_pop); 17991a4d7607STejun Heo 180047504ee0SDennis Zhou ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); 18011a4d7607STejun Heo if (!ret) { 18021a4d7607STejun Heo nr_to_pop -= nr; 18031a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 1804b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, rs + nr); 18051a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 18061a4d7607STejun Heo } else { 18071a4d7607STejun Heo nr_to_pop = 0; 18081a4d7607STejun Heo } 18091a4d7607STejun Heo 18101a4d7607STejun Heo if (!nr_to_pop) 18111a4d7607STejun Heo break; 18121a4d7607STejun Heo } 18131a4d7607STejun Heo } 18141a4d7607STejun Heo 18151a4d7607STejun Heo if (nr_to_pop) { 18161a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 181747504ee0SDennis Zhou chunk = pcpu_create_chunk(gfp); 18181a4d7607STejun Heo if (chunk) { 18191a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 18201a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 18211a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 18221a4d7607STejun Heo goto retry_pop; 18231a4d7607STejun Heo } 18241a4d7607STejun Heo } 18251a4d7607STejun Heo 1826971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1827a56dbddfSTejun Heo } 1828fbf59bc9STejun Heo 1829fbf59bc9STejun Heo /** 1830fbf59bc9STejun Heo * free_percpu - free percpu area 1831fbf59bc9STejun Heo * @ptr: pointer to area to free 1832fbf59bc9STejun Heo * 1833ccea34b5STejun Heo * Free percpu area @ptr. 1834ccea34b5STejun Heo * 1835ccea34b5STejun Heo * CONTEXT: 1836ccea34b5STejun Heo * Can be called from atomic context. 1837fbf59bc9STejun Heo */ 183843cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1839fbf59bc9STejun Heo { 1840129182e5SAndrew Morton void *addr; 1841fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1842ccea34b5STejun Heo unsigned long flags; 184340064aecSDennis Zhou (Facebook) int off; 1844fbf59bc9STejun Heo 1845fbf59bc9STejun Heo if (!ptr) 1846fbf59bc9STejun Heo return; 1847fbf59bc9STejun Heo 1848f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1849f528f0b8SCatalin Marinas 1850129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1851129182e5SAndrew Morton 1852ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1853fbf59bc9STejun Heo 1854fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1855bba174f5STejun Heo off = addr - chunk->base_addr; 1856fbf59bc9STejun Heo 185740064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1858fbf59bc9STejun Heo 1859a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 186040064aecSDennis Zhou (Facebook) if (chunk->free_bytes == pcpu_unit_size) { 1861fbf59bc9STejun Heo struct pcpu_chunk *pos; 1862fbf59bc9STejun Heo 1863a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1864fbf59bc9STejun Heo if (pos != chunk) { 18651a4d7607STejun Heo pcpu_schedule_balance_work(); 1866fbf59bc9STejun Heo break; 1867fbf59bc9STejun Heo } 1868fbf59bc9STejun Heo } 1869fbf59bc9STejun Heo 1870df95e795SDennis Zhou trace_percpu_free_percpu(chunk->base_addr, off, ptr); 1871df95e795SDennis Zhou 1872ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1873fbf59bc9STejun Heo } 1874fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1875fbf59bc9STejun Heo 1876383776faSThomas Gleixner bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 1877383776faSThomas Gleixner { 1878383776faSThomas Gleixner #ifdef CONFIG_SMP 1879383776faSThomas Gleixner const size_t static_size = __per_cpu_end - __per_cpu_start; 1880383776faSThomas Gleixner void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 1881383776faSThomas Gleixner unsigned int cpu; 1882383776faSThomas Gleixner 1883383776faSThomas Gleixner for_each_possible_cpu(cpu) { 1884383776faSThomas Gleixner void *start = per_cpu_ptr(base, cpu); 1885383776faSThomas Gleixner void *va = (void *)addr; 1886383776faSThomas Gleixner 1887383776faSThomas Gleixner if (va >= start && va < start + static_size) { 18888ce371f9SPeter Zijlstra if (can_addr) { 1889383776faSThomas Gleixner *can_addr = (unsigned long) (va - start); 18908ce371f9SPeter Zijlstra *can_addr += (unsigned long) 18918ce371f9SPeter Zijlstra per_cpu_ptr(base, get_boot_cpu_id()); 18928ce371f9SPeter Zijlstra } 1893383776faSThomas Gleixner return true; 1894383776faSThomas Gleixner } 1895383776faSThomas Gleixner } 1896383776faSThomas Gleixner #endif 1897383776faSThomas Gleixner /* on UP, can't distinguish from other static vars, always false */ 1898383776faSThomas Gleixner return false; 1899383776faSThomas Gleixner } 1900383776faSThomas Gleixner 19013b034b0dSVivek Goyal /** 190210fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 190310fad5e4STejun Heo * @addr: address to test 190410fad5e4STejun Heo * 190510fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 190610fad5e4STejun Heo * static percpu areas are not considered. For those, use 190710fad5e4STejun Heo * is_module_percpu_address(). 190810fad5e4STejun Heo * 190910fad5e4STejun Heo * RETURNS: 191010fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 191110fad5e4STejun Heo */ 191210fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 191310fad5e4STejun Heo { 1914383776faSThomas Gleixner return __is_kernel_percpu_address(addr, NULL); 191510fad5e4STejun Heo } 191610fad5e4STejun Heo 191710fad5e4STejun Heo /** 19183b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 19193b034b0dSVivek Goyal * @addr: the address to be converted to physical address 19203b034b0dSVivek Goyal * 19213b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 19223b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 19233b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 19243b034b0dSVivek Goyal * until this function finishes. 19253b034b0dSVivek Goyal * 192667589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 192767589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 192867589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 192967589c71SDave Young * km) provides translation. 193067589c71SDave Young * 1931bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 193267589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 193367589c71SDave Young * actually works, and the verification can discover both bugs in percpu 193467589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 193567589c71SDave Young * code. 193667589c71SDave Young * 19373b034b0dSVivek Goyal * RETURNS: 19383b034b0dSVivek Goyal * The physical address for @addr. 19393b034b0dSVivek Goyal */ 19403b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 19413b034b0dSVivek Goyal { 19429983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 19439983b6f0STejun Heo bool in_first_chunk = false; 1944a855b84cSTejun Heo unsigned long first_low, first_high; 19459983b6f0STejun Heo unsigned int cpu; 19469983b6f0STejun Heo 19479983b6f0STejun Heo /* 1948a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 19499983b6f0STejun Heo * necessary but will speed up lookups of addresses which 19509983b6f0STejun Heo * aren't in the first chunk. 1951c0ebfdc3SDennis Zhou (Facebook) * 1952c0ebfdc3SDennis Zhou (Facebook) * The address check is against full chunk sizes. pcpu_base_addr 1953c0ebfdc3SDennis Zhou (Facebook) * points to the beginning of the first chunk including the 1954c0ebfdc3SDennis Zhou (Facebook) * static region. Assumes good intent as the first chunk may 1955c0ebfdc3SDennis Zhou (Facebook) * not be full (ie. < pcpu_unit_pages in size). 19569983b6f0STejun Heo */ 1957c0ebfdc3SDennis Zhou (Facebook) first_low = (unsigned long)pcpu_base_addr + 1958c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 1959c0ebfdc3SDennis Zhou (Facebook) first_high = (unsigned long)pcpu_base_addr + 1960c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 1961a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1962a855b84cSTejun Heo (unsigned long)addr < first_high) { 19639983b6f0STejun Heo for_each_possible_cpu(cpu) { 19649983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 19659983b6f0STejun Heo 19669983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 19679983b6f0STejun Heo in_first_chunk = true; 19689983b6f0STejun Heo break; 19699983b6f0STejun Heo } 19709983b6f0STejun Heo } 19719983b6f0STejun Heo } 19729983b6f0STejun Heo 19739983b6f0STejun Heo if (in_first_chunk) { 1974eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 19753b034b0dSVivek Goyal return __pa(addr); 19763b034b0dSVivek Goyal else 19779f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 19789f57bd4dSEugene Surovegin offset_in_page(addr); 1979020ec653STejun Heo } else 19809f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 19819f57bd4dSEugene Surovegin offset_in_page(addr); 19823b034b0dSVivek Goyal } 19833b034b0dSVivek Goyal 1984fbf59bc9STejun Heo /** 1985fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1986fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1987fd1e8a1fSTejun Heo * @nr_units: the number of units 1988033e48fbSTejun Heo * 1989fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1990fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1991fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1992fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1993fd1e8a1fSTejun Heo * pointer of other groups. 1994033e48fbSTejun Heo * 1995033e48fbSTejun Heo * RETURNS: 1996fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1997fd1e8a1fSTejun Heo * failure. 1998033e48fbSTejun Heo */ 1999fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 2000fd1e8a1fSTejun Heo int nr_units) 2001fd1e8a1fSTejun Heo { 2002fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 2003fd1e8a1fSTejun Heo size_t base_size, ai_size; 2004fd1e8a1fSTejun Heo void *ptr; 2005fd1e8a1fSTejun Heo int unit; 2006fd1e8a1fSTejun Heo 2007fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 2008fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 2009fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 2010fd1e8a1fSTejun Heo 201126fb3daeSMike Rapoport ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE); 2012fd1e8a1fSTejun Heo if (!ptr) 2013fd1e8a1fSTejun Heo return NULL; 2014fd1e8a1fSTejun Heo ai = ptr; 2015fd1e8a1fSTejun Heo ptr += base_size; 2016fd1e8a1fSTejun Heo 2017fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 2018fd1e8a1fSTejun Heo 2019fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 2020fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 2021fd1e8a1fSTejun Heo 2022fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 2023fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 2024fd1e8a1fSTejun Heo 2025fd1e8a1fSTejun Heo return ai; 2026fd1e8a1fSTejun Heo } 2027fd1e8a1fSTejun Heo 2028fd1e8a1fSTejun Heo /** 2029fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 2030fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 2031fd1e8a1fSTejun Heo * 2032fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 2033fd1e8a1fSTejun Heo */ 2034fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 2035fd1e8a1fSTejun Heo { 2036999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 2037fd1e8a1fSTejun Heo } 2038fd1e8a1fSTejun Heo 2039fd1e8a1fSTejun Heo /** 2040fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 2041fd1e8a1fSTejun Heo * @lvl: loglevel 2042fd1e8a1fSTejun Heo * @ai: allocation info to dump 2043fd1e8a1fSTejun Heo * 2044fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 2045fd1e8a1fSTejun Heo */ 2046fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 2047fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 2048033e48fbSTejun Heo { 2049fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 2050033e48fbSTejun Heo char empty_str[] = "--------"; 2051fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 2052fd1e8a1fSTejun Heo int group, v; 2053fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 2054033e48fbSTejun Heo 2055fd1e8a1fSTejun Heo v = ai->nr_groups; 2056033e48fbSTejun Heo while (v /= 10) 2057fd1e8a1fSTejun Heo group_width++; 2058033e48fbSTejun Heo 2059fd1e8a1fSTejun Heo v = num_possible_cpus(); 2060fd1e8a1fSTejun Heo while (v /= 10) 2061fd1e8a1fSTejun Heo cpu_width++; 2062fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 2063033e48fbSTejun Heo 2064fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 2065fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 2066fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 2067033e48fbSTejun Heo 2068fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 2069fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 2070fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 2071fd1e8a1fSTejun Heo 2072fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 2073fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 2074fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 2075fd1e8a1fSTejun Heo 2076fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 2077fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 2078fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 2079fd1e8a1fSTejun Heo if (!(alloc % apl)) { 20801170532bSJoe Perches pr_cont("\n"); 2081fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 2082033e48fbSTejun Heo } 20831170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 2084fd1e8a1fSTejun Heo 2085fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 2086fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 20871170532bSJoe Perches pr_cont("%0*d ", 20881170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 2089033e48fbSTejun Heo else 20901170532bSJoe Perches pr_cont("%s ", empty_str); 2091033e48fbSTejun Heo } 2092fd1e8a1fSTejun Heo } 20931170532bSJoe Perches pr_cont("\n"); 2094033e48fbSTejun Heo } 2095033e48fbSTejun Heo 2096fbf59bc9STejun Heo /** 20978d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 2098fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 209938a6be52STejun Heo * @base_addr: mapped address 2100fbf59bc9STejun Heo * 21018d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 21028d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 210338a6be52STejun Heo * setup path. 21048d408b4bSTejun Heo * 2105fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 2106fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 21078d408b4bSTejun Heo * 2108fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 2109fd1e8a1fSTejun Heo * 2110fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 2111edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 2112edcb4639STejun Heo * the first chunk such that it's available only through reserved 2113edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 2114edcb4639STejun Heo * static areas on architectures where the addressing model has 2115edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 2116edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 2117edcb4639STejun Heo * 2118fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 2119fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 2120fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 21216074d5b0STejun Heo * 2122fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 2123fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 2124fd1e8a1fSTejun Heo * @ai->dyn_size. 21258d408b4bSTejun Heo * 2126fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 2127fd1e8a1fSTejun Heo * for vm areas. 21288d408b4bSTejun Heo * 2129fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 2130fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 2131fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 2132fd1e8a1fSTejun Heo * 2133fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 2134fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 2135fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 2136fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 2137fd1e8a1fSTejun Heo * all units is assumed. 21388d408b4bSTejun Heo * 213938a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 214038a6be52STejun Heo * copied static data to each unit. 2141fbf59bc9STejun Heo * 2142c0ebfdc3SDennis Zhou (Facebook) * The first chunk will always contain a static and a dynamic region. 2143c0ebfdc3SDennis Zhou (Facebook) * However, the static region is not managed by any chunk. If the first 2144c0ebfdc3SDennis Zhou (Facebook) * chunk also contains a reserved region, it is served by two chunks - 2145c0ebfdc3SDennis Zhou (Facebook) * one for the reserved region and one for the dynamic region. They 2146c0ebfdc3SDennis Zhou (Facebook) * share the same vm, but use offset regions in the area allocation map. 2147c0ebfdc3SDennis Zhou (Facebook) * The chunk serving the dynamic region is circulated in the chunk slots 2148c0ebfdc3SDennis Zhou (Facebook) * and available for dynamic allocation like any other chunk. 2149edcb4639STejun Heo * 2150fbf59bc9STejun Heo * RETURNS: 2151fb435d52STejun Heo * 0 on success, -errno on failure. 2152fbf59bc9STejun Heo */ 2153fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 2154fd1e8a1fSTejun Heo void *base_addr) 2155fbf59bc9STejun Heo { 2156b9c39442SDennis Zhou (Facebook) size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2157d2f3c384SDennis Zhou (Facebook) size_t static_size, dyn_size; 21580c4169c3SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 21596563297cSTejun Heo unsigned long *group_offsets; 21606563297cSTejun Heo size_t *group_sizes; 2161fb435d52STejun Heo unsigned long *unit_off; 2162fbf59bc9STejun Heo unsigned int cpu; 2163fd1e8a1fSTejun Heo int *unit_map; 2164fd1e8a1fSTejun Heo int group, unit, i; 2165c0ebfdc3SDennis Zhou (Facebook) int map_size; 2166c0ebfdc3SDennis Zhou (Facebook) unsigned long tmp_addr; 2167f655f405SMike Rapoport size_t alloc_size; 2168fbf59bc9STejun Heo 2169635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 2170635b75fcSTejun Heo if (unlikely(cond)) { \ 2171870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 2172870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 2173807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 2174635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 2175635b75fcSTejun Heo BUG(); \ 2176635b75fcSTejun Heo } \ 2177635b75fcSTejun Heo } while (0) 2178635b75fcSTejun Heo 21792f39e637STejun Heo /* sanity checks */ 2180635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 2181bbddff05STejun Heo #ifdef CONFIG_SMP 2182635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 2183f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 2184bbddff05STejun Heo #endif 2185635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 2186f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 2187635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 2188f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 2189635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 2190ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 2191099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 2192fb29a2ccSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!ai->dyn_size); 2193d2f3c384SDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 2194ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 2195ca460b3cSDennis Zhou (Facebook) IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 21969f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 21978d408b4bSTejun Heo 21986563297cSTejun Heo /* process group information and build config tables accordingly */ 2199f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_offsets[0]); 2200f655f405SMike Rapoport group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2201f655f405SMike Rapoport if (!group_offsets) 2202f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2203f655f405SMike Rapoport alloc_size); 2204f655f405SMike Rapoport 2205f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_sizes[0]); 2206f655f405SMike Rapoport group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2207f655f405SMike Rapoport if (!group_sizes) 2208f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2209f655f405SMike Rapoport alloc_size); 2210f655f405SMike Rapoport 2211f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_map[0]); 2212f655f405SMike Rapoport unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2213f655f405SMike Rapoport if (!unit_map) 2214f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2215f655f405SMike Rapoport alloc_size); 2216f655f405SMike Rapoport 2217f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_off[0]); 2218f655f405SMike Rapoport unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2219f655f405SMike Rapoport if (!unit_off) 2220f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2221f655f405SMike Rapoport alloc_size); 22222f39e637STejun Heo 2223fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 2224ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 2225a855b84cSTejun Heo 2226a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 2227a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 22282f39e637STejun Heo 2229fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 2230fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 22312f39e637STejun Heo 22326563297cSTejun Heo group_offsets[group] = gi->base_offset; 22336563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 22346563297cSTejun Heo 2235fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 2236fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 2237fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 2238fd1e8a1fSTejun Heo continue; 2239fd1e8a1fSTejun Heo 22409f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 2241635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 2242635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 2243fd1e8a1fSTejun Heo 2244fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 2245fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 2246fb435d52STejun Heo 2247a855b84cSTejun Heo /* determine low/high unit_cpu */ 2248a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 2249a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 2250a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 2251a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 2252a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 2253a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 22540fc0531eSLinus Torvalds } 22550fc0531eSLinus Torvalds } 2256fd1e8a1fSTejun Heo pcpu_nr_units = unit; 22572f39e637STejun Heo 22582f39e637STejun Heo for_each_possible_cpu(cpu) 2259635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 2260635b75fcSTejun Heo 2261635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 2262635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 2263bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 22642f39e637STejun Heo 22656563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 22666563297cSTejun Heo pcpu_group_offsets = group_offsets; 22676563297cSTejun Heo pcpu_group_sizes = group_sizes; 2268fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 2269fb435d52STejun Heo pcpu_unit_offsets = unit_off; 22702f39e637STejun Heo 22712f39e637STejun Heo /* determine basic parameters */ 2272fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 2273d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 22746563297cSTejun Heo pcpu_atom_size = ai->atom_size; 2275ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 2276ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 2277cafe8816STejun Heo 227830a5b536SDennis Zhou pcpu_stats_save_ai(ai); 227930a5b536SDennis Zhou 2280d9b55eebSTejun Heo /* 2281d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 2282d9b55eebSTejun Heo * empty chunks. 2283d9b55eebSTejun Heo */ 2284d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 22857e1c4e27SMike Rapoport pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]), 22867e1c4e27SMike Rapoport SMP_CACHE_BYTES); 2287f655f405SMike Rapoport if (!pcpu_slot) 2288f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2289f655f405SMike Rapoport pcpu_nr_slots * sizeof(pcpu_slot[0])); 2290fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 2291fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 2292fbf59bc9STejun Heo 2293edcb4639STejun Heo /* 2294d2f3c384SDennis Zhou (Facebook) * The end of the static region needs to be aligned with the 2295d2f3c384SDennis Zhou (Facebook) * minimum allocation size as this offsets the reserved and 2296d2f3c384SDennis Zhou (Facebook) * dynamic region. The first chunk ends page aligned by 2297d2f3c384SDennis Zhou (Facebook) * expanding the dynamic region, therefore the dynamic region 2298d2f3c384SDennis Zhou (Facebook) * can be shrunk to compensate while still staying above the 2299d2f3c384SDennis Zhou (Facebook) * configured sizes. 2300d2f3c384SDennis Zhou (Facebook) */ 2301d2f3c384SDennis Zhou (Facebook) static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 2302d2f3c384SDennis Zhou (Facebook) dyn_size = ai->dyn_size - (static_size - ai->static_size); 2303d2f3c384SDennis Zhou (Facebook) 2304d2f3c384SDennis Zhou (Facebook) /* 2305c0ebfdc3SDennis Zhou (Facebook) * Initialize first chunk. 2306c0ebfdc3SDennis Zhou (Facebook) * If the reserved_size is non-zero, this initializes the reserved 2307c0ebfdc3SDennis Zhou (Facebook) * chunk. If the reserved_size is zero, the reserved chunk is NULL 2308c0ebfdc3SDennis Zhou (Facebook) * and the dynamic region is initialized here. The first chunk, 2309c0ebfdc3SDennis Zhou (Facebook) * pcpu_first_chunk, will always point to the chunk that serves 2310c0ebfdc3SDennis Zhou (Facebook) * the dynamic region. 2311edcb4639STejun Heo */ 2312d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size; 2313d2f3c384SDennis Zhou (Facebook) map_size = ai->reserved_size ?: dyn_size; 231440064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 231561ace7faSTejun Heo 2316edcb4639STejun Heo /* init dynamic chunk if necessary */ 2317b9c39442SDennis Zhou (Facebook) if (ai->reserved_size) { 23180c4169c3SDennis Zhou (Facebook) pcpu_reserved_chunk = chunk; 2319b9c39442SDennis Zhou (Facebook) 2320d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size + 2321c0ebfdc3SDennis Zhou (Facebook) ai->reserved_size; 2322d2f3c384SDennis Zhou (Facebook) map_size = dyn_size; 232340064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 2324edcb4639STejun Heo } 2325edcb4639STejun Heo 23262441d15cSTejun Heo /* link the first chunk in */ 23270c4169c3SDennis Zhou (Facebook) pcpu_first_chunk = chunk; 23280cecf50cSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 2329ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 2330fbf59bc9STejun Heo 23317e8a6304SDennis Zhou (Facebook) /* include all regions of the first chunk */ 23327e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += PFN_DOWN(size_sum); 23337e8a6304SDennis Zhou (Facebook) 233430a5b536SDennis Zhou pcpu_stats_chunk_alloc(); 2335df95e795SDennis Zhou trace_percpu_create_chunk(base_addr); 233630a5b536SDennis Zhou 2337fbf59bc9STejun Heo /* we're done */ 2338bba174f5STejun Heo pcpu_base_addr = base_addr; 2339fb435d52STejun Heo return 0; 2340fbf59bc9STejun Heo } 234166c3a757STejun Heo 2342bbddff05STejun Heo #ifdef CONFIG_SMP 2343bbddff05STejun Heo 234417f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 2345f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 2346f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 2347f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 2348f58dc01bSTejun Heo }; 234966c3a757STejun Heo 2350f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 2351f58dc01bSTejun Heo 2352f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 235366c3a757STejun Heo { 23545479c78aSCyrill Gorcunov if (!str) 23555479c78aSCyrill Gorcunov return -EINVAL; 23565479c78aSCyrill Gorcunov 2357f58dc01bSTejun Heo if (0) 2358f58dc01bSTejun Heo /* nada */; 2359f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 2360f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 2361f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 2362f58dc01bSTejun Heo #endif 2363f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2364f58dc01bSTejun Heo else if (!strcmp(str, "page")) 2365f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 2366f58dc01bSTejun Heo #endif 2367f58dc01bSTejun Heo else 2368870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 236966c3a757STejun Heo 2370f58dc01bSTejun Heo return 0; 237166c3a757STejun Heo } 2372f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 237366c3a757STejun Heo 23743c9a024fSTejun Heo /* 23753c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 23763c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 23773c9a024fSTejun Heo * to be used. 23783c9a024fSTejun Heo */ 237908fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 238008fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 23813c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 23823c9a024fSTejun Heo #endif 23833c9a024fSTejun Heo 23843c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 23853c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 23863c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 23873c9a024fSTejun Heo #endif 23883c9a024fSTejun Heo 23893c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 23903c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 23913c9a024fSTejun Heo /** 2392fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2393fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2394fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2395fbf59bc9STejun Heo * @atom_size: allocation atom size 2396fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2397fbf59bc9STejun Heo * 2398fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 2399fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 2400fbf59bc9STejun Heo * atom size and distances between CPUs. 2401fbf59bc9STejun Heo * 2402bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 2403fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 2404fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 2405fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 2406fbf59bc9STejun Heo * of allocated virtual address space. 2407fbf59bc9STejun Heo * 2408fbf59bc9STejun Heo * RETURNS: 2409fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 2410fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 2411fbf59bc9STejun Heo */ 2412fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 2413fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 2414fbf59bc9STejun Heo size_t atom_size, 2415fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2416fbf59bc9STejun Heo { 2417fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 2418fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 2419fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 2420fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 2421fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 2422fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 2423fbf59bc9STejun Heo int last_allocs, group, unit; 2424fbf59bc9STejun Heo unsigned int cpu, tcpu; 2425fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 2426fbf59bc9STejun Heo unsigned int *cpu_map; 2427fbf59bc9STejun Heo 2428fbf59bc9STejun Heo /* this function may be called multiple times */ 2429fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 2430fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 2431fbf59bc9STejun Heo 2432fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2433fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 2434fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2435fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 2436fbf59bc9STejun Heo 2437fbf59bc9STejun Heo /* 2438fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 2439fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 244025985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 2441fbf59bc9STejun Heo * or larger than min_unit_size. 2442fbf59bc9STejun Heo */ 2443fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2444fbf59bc9STejun Heo 24459c015162SDennis Zhou (Facebook) /* determine the maximum # of units that can fit in an allocation */ 2446fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 2447fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 2448f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2449fbf59bc9STejun Heo upa--; 2450fbf59bc9STejun Heo max_upa = upa; 2451fbf59bc9STejun Heo 2452fbf59bc9STejun Heo /* group cpus according to their proximity */ 2453fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 2454fbf59bc9STejun Heo group = 0; 2455fbf59bc9STejun Heo next_group: 2456fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 2457fbf59bc9STejun Heo if (cpu == tcpu) 2458fbf59bc9STejun Heo break; 2459fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 2460fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 2461fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 2462fbf59bc9STejun Heo group++; 2463fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 2464fbf59bc9STejun Heo goto next_group; 2465fbf59bc9STejun Heo } 2466fbf59bc9STejun Heo } 2467fbf59bc9STejun Heo group_map[cpu] = group; 2468fbf59bc9STejun Heo group_cnt[group]++; 2469fbf59bc9STejun Heo } 2470fbf59bc9STejun Heo 2471fbf59bc9STejun Heo /* 24729c015162SDennis Zhou (Facebook) * Wasted space is caused by a ratio imbalance of upa to group_cnt. 24739c015162SDennis Zhou (Facebook) * Expand the unit_size until we use >= 75% of the units allocated. 24749c015162SDennis Zhou (Facebook) * Related to atom_size, which could be much larger than the unit_size. 2475fbf59bc9STejun Heo */ 2476fbf59bc9STejun Heo last_allocs = INT_MAX; 2477fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 2478fbf59bc9STejun Heo int allocs = 0, wasted = 0; 2479fbf59bc9STejun Heo 2480f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2481fbf59bc9STejun Heo continue; 2482fbf59bc9STejun Heo 2483fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2484fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2485fbf59bc9STejun Heo allocs += this_allocs; 2486fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 2487fbf59bc9STejun Heo } 2488fbf59bc9STejun Heo 2489fbf59bc9STejun Heo /* 2490fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 2491fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 2492fbf59bc9STejun Heo * passes the following check. 2493fbf59bc9STejun Heo */ 2494fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 2495fbf59bc9STejun Heo continue; 2496fbf59bc9STejun Heo 2497fbf59bc9STejun Heo /* and then don't consume more memory */ 2498fbf59bc9STejun Heo if (allocs > last_allocs) 2499fbf59bc9STejun Heo break; 2500fbf59bc9STejun Heo last_allocs = allocs; 2501fbf59bc9STejun Heo best_upa = upa; 2502fbf59bc9STejun Heo } 2503fbf59bc9STejun Heo upa = best_upa; 2504fbf59bc9STejun Heo 2505fbf59bc9STejun Heo /* allocate and fill alloc_info */ 2506fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 2507fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 2508fbf59bc9STejun Heo 2509fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2510fbf59bc9STejun Heo if (!ai) 2511fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 2512fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 2513fbf59bc9STejun Heo 2514fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2515fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 2516fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 2517fbf59bc9STejun Heo } 2518fbf59bc9STejun Heo 2519fbf59bc9STejun Heo ai->static_size = static_size; 2520fbf59bc9STejun Heo ai->reserved_size = reserved_size; 2521fbf59bc9STejun Heo ai->dyn_size = dyn_size; 2522fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 2523fbf59bc9STejun Heo ai->atom_size = atom_size; 2524fbf59bc9STejun Heo ai->alloc_size = alloc_size; 2525fbf59bc9STejun Heo 25262de7852fSPeng Fan for (group = 0, unit = 0; group < nr_groups; group++) { 2527fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2528fbf59bc9STejun Heo 2529fbf59bc9STejun Heo /* 2530fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 2531fbf59bc9STejun Heo * back-to-back. The caller should update this to 2532fbf59bc9STejun Heo * reflect actual allocation. 2533fbf59bc9STejun Heo */ 2534fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 2535fbf59bc9STejun Heo 2536fbf59bc9STejun Heo for_each_possible_cpu(cpu) 2537fbf59bc9STejun Heo if (group_map[cpu] == group) 2538fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 2539fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 2540fbf59bc9STejun Heo unit += gi->nr_units; 2541fbf59bc9STejun Heo } 2542fbf59bc9STejun Heo BUG_ON(unit != nr_units); 2543fbf59bc9STejun Heo 2544fbf59bc9STejun Heo return ai; 2545fbf59bc9STejun Heo } 25463c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2547fbf59bc9STejun Heo 25483c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 254966c3a757STejun Heo /** 255066c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 255166c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 25524ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2553c8826dd5STejun Heo * @atom_size: allocation atom size 2554c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2555c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 255625985edcSLucas De Marchi * @free_fn: function to free percpu page 255766c3a757STejun Heo * 255866c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 255966c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 256066c3a757STejun Heo * 256166c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 2562c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 2563c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 2564c8826dd5STejun Heo * aligned to @atom_size. 2565c8826dd5STejun Heo * 2566c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 2567c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 2568c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 2569c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 2570c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 2571c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 257266c3a757STejun Heo * 25734ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 257466c3a757STejun Heo * 257566c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 2576c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 257766c3a757STejun Heo * 257866c3a757STejun Heo * RETURNS: 2579fb435d52STejun Heo * 0 on success, -errno on failure. 258066c3a757STejun Heo */ 25814ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 2582c8826dd5STejun Heo size_t atom_size, 2583c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 2584c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2585c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 258666c3a757STejun Heo { 2587c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 2588c8826dd5STejun Heo void **areas = NULL; 2589fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 259093c76b6bSzijun_hu size_t size_sum, areas_size; 259193c76b6bSzijun_hu unsigned long max_distance; 25929b739662Szijun_hu int group, i, highest_group, rc; 259366c3a757STejun Heo 2594c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 2595c8826dd5STejun Heo cpu_distance_fn); 2596fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2597fd1e8a1fSTejun Heo return PTR_ERR(ai); 259866c3a757STejun Heo 2599fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2600c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 260166c3a757STejun Heo 260226fb3daeSMike Rapoport areas = memblock_alloc(areas_size, SMP_CACHE_BYTES); 2603c8826dd5STejun Heo if (!areas) { 2604fb435d52STejun Heo rc = -ENOMEM; 2605c8826dd5STejun Heo goto out_free; 2606fa8a7094STejun Heo } 260766c3a757STejun Heo 26089b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 26099b739662Szijun_hu highest_group = 0; 2610c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2611c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2612c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 2613c8826dd5STejun Heo void *ptr; 261466c3a757STejun Heo 2615c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 2616c8826dd5STejun Heo cpu = gi->cpu_map[i]; 2617c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 2618c8826dd5STejun Heo 2619c8826dd5STejun Heo /* allocate space for the whole group */ 2620c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 2621c8826dd5STejun Heo if (!ptr) { 2622c8826dd5STejun Heo rc = -ENOMEM; 2623c8826dd5STejun Heo goto out_free_areas; 2624c8826dd5STejun Heo } 2625f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2626f528f0b8SCatalin Marinas kmemleak_free(ptr); 2627c8826dd5STejun Heo areas[group] = ptr; 2628c8826dd5STejun Heo 2629c8826dd5STejun Heo base = min(ptr, base); 26309b739662Szijun_hu if (ptr > areas[highest_group]) 26319b739662Szijun_hu highest_group = group; 26329b739662Szijun_hu } 26339b739662Szijun_hu max_distance = areas[highest_group] - base; 26349b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 26359b739662Szijun_hu 26369b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 26379b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 26389b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 26399b739662Szijun_hu max_distance, VMALLOC_TOTAL); 26409b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 26419b739662Szijun_hu /* and fail if we have fallback */ 26429b739662Szijun_hu rc = -EINVAL; 26439b739662Szijun_hu goto out_free_areas; 26449b739662Szijun_hu #endif 264542b64281STejun Heo } 264642b64281STejun Heo 264742b64281STejun Heo /* 264842b64281STejun Heo * Copy data and free unused parts. This should happen after all 264942b64281STejun Heo * allocations are complete; otherwise, we may end up with 265042b64281STejun Heo * overlapping groups. 265142b64281STejun Heo */ 265242b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 265342b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 265442b64281STejun Heo void *ptr = areas[group]; 2655c8826dd5STejun Heo 2656c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2657c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2658c8826dd5STejun Heo /* unused unit, free whole */ 2659c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2660c8826dd5STejun Heo continue; 2661c8826dd5STejun Heo } 2662c8826dd5STejun Heo /* copy and return the unused part */ 2663fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2664c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2665c8826dd5STejun Heo } 266666c3a757STejun Heo } 266766c3a757STejun Heo 2668c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 26696ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2670c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 26716ea529a2STejun Heo } 2672c8826dd5STejun Heo 2673870d4b12SJoe Perches pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 2674fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 2675fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 267666c3a757STejun Heo 2677fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 2678c8826dd5STejun Heo goto out_free; 2679c8826dd5STejun Heo 2680c8826dd5STejun Heo out_free_areas: 2681c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2682f851c8d8SMichael Holzheu if (areas[group]) 2683c8826dd5STejun Heo free_fn(areas[group], 2684c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2685c8826dd5STejun Heo out_free: 2686fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2687c8826dd5STejun Heo if (areas) 2688999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2689fb435d52STejun Heo return rc; 2690d4b95f80STejun Heo } 26913c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2692d4b95f80STejun Heo 26933c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2694d4b95f80STejun Heo /** 269500ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2696d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2697d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 269825985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2699d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2700d4b95f80STejun Heo * 270100ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 270200ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2703d4b95f80STejun Heo * 2704d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2705d4b95f80STejun Heo * page-by-page into vmalloc area. 2706d4b95f80STejun Heo * 2707d4b95f80STejun Heo * RETURNS: 2708fb435d52STejun Heo * 0 on success, -errno on failure. 2709d4b95f80STejun Heo */ 2710fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2711d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2712d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2713d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2714d4b95f80STejun Heo { 27158f05a6a6STejun Heo static struct vm_struct vm; 2716fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 271700ae4064STejun Heo char psize_str[16]; 2718ce3141a2STejun Heo int unit_pages; 2719d4b95f80STejun Heo size_t pages_size; 2720ce3141a2STejun Heo struct page **pages; 2721fb435d52STejun Heo int unit, i, j, rc; 27228f606604Szijun_hu int upa; 27238f606604Szijun_hu int nr_g0_units; 2724d4b95f80STejun Heo 272500ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 272600ae4064STejun Heo 27274ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2728fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2729fd1e8a1fSTejun Heo return PTR_ERR(ai); 2730fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 27318f606604Szijun_hu upa = ai->alloc_size/ai->unit_size; 27328f606604Szijun_hu nr_g0_units = roundup(num_possible_cpus(), upa); 27330b59c25fSIgor Stoppa if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) { 27348f606604Szijun_hu pcpu_free_alloc_info(ai); 27358f606604Szijun_hu return -EINVAL; 27368f606604Szijun_hu } 2737fd1e8a1fSTejun Heo 2738fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2739d4b95f80STejun Heo 2740d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2741fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2742fd1e8a1fSTejun Heo sizeof(pages[0])); 27437e1c4e27SMike Rapoport pages = memblock_alloc(pages_size, SMP_CACHE_BYTES); 2744f655f405SMike Rapoport if (!pages) 2745f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2746f655f405SMike Rapoport pages_size); 2747d4b95f80STejun Heo 27488f05a6a6STejun Heo /* allocate pages */ 2749d4b95f80STejun Heo j = 0; 27508f606604Szijun_hu for (unit = 0; unit < num_possible_cpus(); unit++) { 2751fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 27528f606604Szijun_hu for (i = 0; i < unit_pages; i++) { 2753d4b95f80STejun Heo void *ptr; 2754d4b95f80STejun Heo 27553cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2756d4b95f80STejun Heo if (!ptr) { 2757870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2758598d8091SJoe Perches psize_str, cpu); 2759d4b95f80STejun Heo goto enomem; 2760d4b95f80STejun Heo } 2761f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2762f528f0b8SCatalin Marinas kmemleak_free(ptr); 2763ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2764d4b95f80STejun Heo } 27658f606604Szijun_hu } 2766d4b95f80STejun Heo 27678f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 27688f05a6a6STejun Heo vm.flags = VM_ALLOC; 2769fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 27708f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 27718f05a6a6STejun Heo 2772fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 27731d9d3257STejun Heo unsigned long unit_addr = 2774fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 27758f05a6a6STejun Heo 2776ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 27778f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 27788f05a6a6STejun Heo 27798f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2780fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2781ce3141a2STejun Heo unit_pages); 2782fb435d52STejun Heo if (rc < 0) 2783fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 27848f05a6a6STejun Heo 27858f05a6a6STejun Heo /* 27868f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 27878f05a6a6STejun Heo * cache for the linear mapping here - something 27888f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 27898f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 27908f05a6a6STejun Heo * data structures are not set up yet. 27918f05a6a6STejun Heo */ 27928f05a6a6STejun Heo 27938f05a6a6STejun Heo /* copy static data */ 2794fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 279566c3a757STejun Heo } 279666c3a757STejun Heo 279766c3a757STejun Heo /* we're ready, commit */ 2798870d4b12SJoe Perches pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", 2799fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 2800fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 280166c3a757STejun Heo 2802fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 2803d4b95f80STejun Heo goto out_free_ar; 2804d4b95f80STejun Heo 2805d4b95f80STejun Heo enomem: 2806d4b95f80STejun Heo while (--j >= 0) 2807ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2808fb435d52STejun Heo rc = -ENOMEM; 2809d4b95f80STejun Heo out_free_ar: 2810999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2811fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2812fb435d52STejun Heo return rc; 281366c3a757STejun Heo } 28143c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2815d4b95f80STejun Heo 2816bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 28178c4bfc6eSTejun Heo /* 2818bbddff05STejun Heo * Generic SMP percpu area setup. 2819e74e3962STejun Heo * 2820e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2821e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2822e74e3962STejun Heo * important because many archs have addressing restrictions and might 2823e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2824e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2825e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2826e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2827e74e3962STejun Heo * mappings on applicable archs. 2828e74e3962STejun Heo */ 2829e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2830e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2831e74e3962STejun Heo 2832c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2833c8826dd5STejun Heo size_t align) 2834c8826dd5STejun Heo { 283526fb3daeSMike Rapoport return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS)); 2836c8826dd5STejun Heo } 2837c8826dd5STejun Heo 2838c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2839c8826dd5STejun Heo { 2840999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2841c8826dd5STejun Heo } 2842c8826dd5STejun Heo 2843e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2844e74e3962STejun Heo { 2845e74e3962STejun Heo unsigned long delta; 2846e74e3962STejun Heo unsigned int cpu; 2847fb435d52STejun Heo int rc; 2848e74e3962STejun Heo 2849e74e3962STejun Heo /* 2850e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2851e74e3962STejun Heo * what the legacy allocator did. 2852e74e3962STejun Heo */ 2853fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2854c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2855c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2856fb435d52STejun Heo if (rc < 0) 2857bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2858e74e3962STejun Heo 2859e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2860e74e3962STejun Heo for_each_possible_cpu(cpu) 2861fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2862e74e3962STejun Heo } 2863e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2864099a19d9STejun Heo 2865bbddff05STejun Heo #else /* CONFIG_SMP */ 2866bbddff05STejun Heo 2867bbddff05STejun Heo /* 2868bbddff05STejun Heo * UP percpu area setup. 2869bbddff05STejun Heo * 2870bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2871bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2872bbddff05STejun Heo * variables and don't require any special preparation. 2873bbddff05STejun Heo */ 2874bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2875bbddff05STejun Heo { 2876bbddff05STejun Heo const size_t unit_size = 2877bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2878bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2879bbddff05STejun Heo struct pcpu_alloc_info *ai; 2880bbddff05STejun Heo void *fc; 2881bbddff05STejun Heo 2882bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 288326fb3daeSMike Rapoport fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); 2884bbddff05STejun Heo if (!ai || !fc) 2885bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2886100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2887100d13c3SCatalin Marinas kmemleak_free(fc); 2888bbddff05STejun Heo 2889bbddff05STejun Heo ai->dyn_size = unit_size; 2890bbddff05STejun Heo ai->unit_size = unit_size; 2891bbddff05STejun Heo ai->atom_size = unit_size; 2892bbddff05STejun Heo ai->alloc_size = unit_size; 2893bbddff05STejun Heo ai->groups[0].nr_units = 1; 2894bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2895bbddff05STejun Heo 2896bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 2897bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2898438a5061SNicolas Pitre pcpu_free_alloc_info(ai); 2899bbddff05STejun Heo } 2900bbddff05STejun Heo 2901bbddff05STejun Heo #endif /* CONFIG_SMP */ 2902bbddff05STejun Heo 2903099a19d9STejun Heo /* 29047e8a6304SDennis Zhou (Facebook) * pcpu_nr_pages - calculate total number of populated backing pages 29057e8a6304SDennis Zhou (Facebook) * 29067e8a6304SDennis Zhou (Facebook) * This reflects the number of pages populated to back chunks. Metadata is 29077e8a6304SDennis Zhou (Facebook) * excluded in the number exposed in meminfo as the number of backing pages 29087e8a6304SDennis Zhou (Facebook) * scales with the number of cpus and can quickly outweigh the memory used for 29097e8a6304SDennis Zhou (Facebook) * metadata. It also keeps this calculation nice and simple. 29107e8a6304SDennis Zhou (Facebook) * 29117e8a6304SDennis Zhou (Facebook) * RETURNS: 29127e8a6304SDennis Zhou (Facebook) * Total number of populated backing pages in use by the allocator. 29137e8a6304SDennis Zhou (Facebook) */ 29147e8a6304SDennis Zhou (Facebook) unsigned long pcpu_nr_pages(void) 29157e8a6304SDennis Zhou (Facebook) { 29167e8a6304SDennis Zhou (Facebook) return pcpu_nr_populated * pcpu_nr_units; 29177e8a6304SDennis Zhou (Facebook) } 29187e8a6304SDennis Zhou (Facebook) 29197e8a6304SDennis Zhou (Facebook) /* 29201a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 29211a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 29221a4d7607STejun Heo * and running. 29231a4d7607STejun Heo */ 29241a4d7607STejun Heo static int __init percpu_enable_async(void) 29251a4d7607STejun Heo { 29261a4d7607STejun Heo pcpu_async_enabled = true; 29271a4d7607STejun Heo return 0; 29281a4d7607STejun Heo } 29291a4d7607STejun Heo subsys_initcall(percpu_enable_async); 2930