155716d26SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 2fbf59bc9STejun Heo /* 388999a89STejun Heo * mm/percpu.c - percpu memory allocator 4fbf59bc9STejun Heo * 5fbf59bc9STejun Heo * Copyright (C) 2009 SUSE Linux Products GmbH 6fbf59bc9STejun Heo * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 7fbf59bc9STejun Heo * 85e81ee3eSDennis Zhou (Facebook) * Copyright (C) 2017 Facebook Inc. 9bfacd38fSDennis Zhou * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org> 105e81ee3eSDennis Zhou (Facebook) * 119c015162SDennis Zhou (Facebook) * The percpu allocator handles both static and dynamic areas. Percpu 129c015162SDennis Zhou (Facebook) * areas are allocated in chunks which are divided into units. There is 139c015162SDennis Zhou (Facebook) * a 1-to-1 mapping for units to possible cpus. These units are grouped 149c015162SDennis Zhou (Facebook) * based on NUMA properties of the machine. 15fbf59bc9STejun Heo * 16fbf59bc9STejun Heo * c0 c1 c2 17fbf59bc9STejun Heo * ------------------- ------------------- ------------ 18fbf59bc9STejun Heo * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 19fbf59bc9STejun Heo * ------------------- ...... ------------------- .... ------------ 20fbf59bc9STejun Heo * 219c015162SDennis Zhou (Facebook) * Allocation is done by offsets into a unit's address space. Ie., an 229c015162SDennis Zhou (Facebook) * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, 239c015162SDennis Zhou (Facebook) * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear 249c015162SDennis Zhou (Facebook) * and even sparse. Access is handled by configuring percpu base 259c015162SDennis Zhou (Facebook) * registers according to the cpu to unit mappings and offsetting the 269c015162SDennis Zhou (Facebook) * base address using pcpu_unit_size. 27fbf59bc9STejun Heo * 289c015162SDennis Zhou (Facebook) * There is special consideration for the first chunk which must handle 299c015162SDennis Zhou (Facebook) * the static percpu variables in the kernel image as allocation services 305e81ee3eSDennis Zhou (Facebook) * are not online yet. In short, the first chunk is structured like so: 319c015162SDennis Zhou (Facebook) * 329c015162SDennis Zhou (Facebook) * <Static | [Reserved] | Dynamic> 339c015162SDennis Zhou (Facebook) * 349c015162SDennis Zhou (Facebook) * The static data is copied from the original section managed by the 359c015162SDennis Zhou (Facebook) * linker. The reserved section, if non-zero, primarily manages static 369c015162SDennis Zhou (Facebook) * percpu variables from kernel modules. Finally, the dynamic section 379c015162SDennis Zhou (Facebook) * takes care of normal allocations. 38fbf59bc9STejun Heo * 395e81ee3eSDennis Zhou (Facebook) * The allocator organizes chunks into lists according to free size and 405e81ee3eSDennis Zhou (Facebook) * tries to allocate from the fullest chunk first. Each chunk is managed 415e81ee3eSDennis Zhou (Facebook) * by a bitmap with metadata blocks. The allocation map is updated on 425e81ee3eSDennis Zhou (Facebook) * every allocation and free to reflect the current state while the boundary 435e81ee3eSDennis Zhou (Facebook) * map is only updated on allocation. Each metadata block contains 445e81ee3eSDennis Zhou (Facebook) * information to help mitigate the need to iterate over large portions 455e81ee3eSDennis Zhou (Facebook) * of the bitmap. The reverse mapping from page to chunk is stored in 465e81ee3eSDennis Zhou (Facebook) * the page's index. Lastly, units are lazily backed and grow in unison. 47fbf59bc9STejun Heo * 485e81ee3eSDennis Zhou (Facebook) * There is a unique conversion that goes on here between bytes and bits. 495e81ee3eSDennis Zhou (Facebook) * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk 505e81ee3eSDennis Zhou (Facebook) * tracks the number of pages it is responsible for in nr_pages. Helper 515e81ee3eSDennis Zhou (Facebook) * functions are used to convert from between the bytes, bits, and blocks. 525e81ee3eSDennis Zhou (Facebook) * All hints are managed in bits unless explicitly stated. 539c015162SDennis Zhou (Facebook) * 544091fb95SMasahiro Yamada * To use this allocator, arch code should do the following: 55fbf59bc9STejun Heo * 56fbf59bc9STejun Heo * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 57e0100983STejun Heo * regular address to percpu pointer and back if they need to be 58e0100983STejun Heo * different from the default 59fbf59bc9STejun Heo * 608d408b4bSTejun Heo * - use pcpu_setup_first_chunk() during percpu area initialization to 618d408b4bSTejun Heo * setup the first chunk containing the kernel static percpu area 62fbf59bc9STejun Heo */ 63fbf59bc9STejun Heo 64870d4b12SJoe Perches #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 65870d4b12SJoe Perches 66fbf59bc9STejun Heo #include <linux/bitmap.h> 6757c8a661SMike Rapoport #include <linux/memblock.h> 68fd1e8a1fSTejun Heo #include <linux/err.h> 69ca460b3cSDennis Zhou (Facebook) #include <linux/lcm.h> 70fbf59bc9STejun Heo #include <linux/list.h> 71a530b795STejun Heo #include <linux/log2.h> 72fbf59bc9STejun Heo #include <linux/mm.h> 73fbf59bc9STejun Heo #include <linux/module.h> 74fbf59bc9STejun Heo #include <linux/mutex.h> 75fbf59bc9STejun Heo #include <linux/percpu.h> 76fbf59bc9STejun Heo #include <linux/pfn.h> 77fbf59bc9STejun Heo #include <linux/slab.h> 78ccea34b5STejun Heo #include <linux/spinlock.h> 79fbf59bc9STejun Heo #include <linux/vmalloc.h> 80a56dbddfSTejun Heo #include <linux/workqueue.h> 81f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 8271546d10STejun Heo #include <linux/sched.h> 8328307d93SFilipe Manana #include <linux/sched/mm.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 { 23692c14cabSDennis Zhou const struct pcpu_block_md *chunk_md = &chunk->chunk_md; 23792c14cabSDennis Zhou 23892c14cabSDennis Zhou if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || 23992c14cabSDennis Zhou chunk_md->contig_hint == 0) 240fbf59bc9STejun Heo return 0; 241fbf59bc9STejun Heo 24292c14cabSDennis Zhou return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE); 243fbf59bc9STejun Heo } 244fbf59bc9STejun Heo 24588999a89STejun Heo /* set the pointer to a chunk in a page struct */ 24688999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 24788999a89STejun Heo { 24888999a89STejun Heo page->index = (unsigned long)pcpu; 24988999a89STejun Heo } 25088999a89STejun Heo 25188999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 25288999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 25388999a89STejun Heo { 25488999a89STejun Heo return (struct pcpu_chunk *)page->index; 25588999a89STejun Heo } 25688999a89STejun Heo 25788999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 258fbf59bc9STejun Heo { 2592f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 260fbf59bc9STejun Heo } 261fbf59bc9STejun Heo 262c0ebfdc3SDennis Zhou (Facebook) static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 263c0ebfdc3SDennis Zhou (Facebook) { 264c0ebfdc3SDennis Zhou (Facebook) return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 265c0ebfdc3SDennis Zhou (Facebook) } 266c0ebfdc3SDennis Zhou (Facebook) 2679983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 268fbf59bc9STejun Heo unsigned int cpu, int page_idx) 269fbf59bc9STejun Heo { 270c0ebfdc3SDennis Zhou (Facebook) return (unsigned long)chunk->base_addr + 271c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(cpu, page_idx); 272fbf59bc9STejun Heo } 273fbf59bc9STejun Heo 274ca460b3cSDennis Zhou (Facebook) /* 275ca460b3cSDennis Zhou (Facebook) * The following are helper functions to help access bitmaps and convert 276ca460b3cSDennis Zhou (Facebook) * between bitmap offsets to address offsets. 277ca460b3cSDennis Zhou (Facebook) */ 278ca460b3cSDennis Zhou (Facebook) static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 279ca460b3cSDennis Zhou (Facebook) { 280ca460b3cSDennis Zhou (Facebook) return chunk->alloc_map + 281ca460b3cSDennis Zhou (Facebook) (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 282ca460b3cSDennis Zhou (Facebook) } 283ca460b3cSDennis Zhou (Facebook) 284ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_index(int off) 285ca460b3cSDennis Zhou (Facebook) { 286ca460b3cSDennis Zhou (Facebook) return off / PCPU_BITMAP_BLOCK_BITS; 287ca460b3cSDennis Zhou (Facebook) } 288ca460b3cSDennis Zhou (Facebook) 289ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_off(int off) 290ca460b3cSDennis Zhou (Facebook) { 291ca460b3cSDennis Zhou (Facebook) return off & (PCPU_BITMAP_BLOCK_BITS - 1); 292ca460b3cSDennis Zhou (Facebook) } 293ca460b3cSDennis Zhou (Facebook) 294b185cd0dSDennis Zhou (Facebook) static unsigned long pcpu_block_off_to_off(int index, int off) 295b185cd0dSDennis Zhou (Facebook) { 296b185cd0dSDennis Zhou (Facebook) return index * PCPU_BITMAP_BLOCK_BITS + off; 297b185cd0dSDennis Zhou (Facebook) } 298b185cd0dSDennis Zhou (Facebook) 299382b88e9SDennis Zhou /* 300382b88e9SDennis Zhou * pcpu_next_hint - determine which hint to use 301382b88e9SDennis Zhou * @block: block of interest 302382b88e9SDennis Zhou * @alloc_bits: size of allocation 303382b88e9SDennis Zhou * 304382b88e9SDennis Zhou * This determines if we should scan based on the scan_hint or first_free. 305382b88e9SDennis Zhou * In general, we want to scan from first_free to fulfill allocations by 306382b88e9SDennis Zhou * first fit. However, if we know a scan_hint at position scan_hint_start 307382b88e9SDennis Zhou * cannot fulfill an allocation, we can begin scanning from there knowing 308382b88e9SDennis Zhou * the contig_hint will be our fallback. 309382b88e9SDennis Zhou */ 310382b88e9SDennis Zhou static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits) 311382b88e9SDennis Zhou { 312382b88e9SDennis Zhou /* 313382b88e9SDennis Zhou * The three conditions below determine if we can skip past the 314382b88e9SDennis Zhou * scan_hint. First, does the scan hint exist. Second, is the 315382b88e9SDennis Zhou * contig_hint after the scan_hint (possibly not true iff 316382b88e9SDennis Zhou * contig_hint == scan_hint). Third, is the allocation request 317382b88e9SDennis Zhou * larger than the scan_hint. 318382b88e9SDennis Zhou */ 319382b88e9SDennis Zhou if (block->scan_hint && 320382b88e9SDennis Zhou block->contig_hint_start > block->scan_hint_start && 321382b88e9SDennis Zhou alloc_bits > block->scan_hint) 322382b88e9SDennis Zhou return block->scan_hint_start + block->scan_hint; 323382b88e9SDennis Zhou 324382b88e9SDennis Zhou return block->first_free; 325382b88e9SDennis Zhou } 326382b88e9SDennis Zhou 327fbf59bc9STejun Heo /** 328525ca84dSDennis Zhou (Facebook) * pcpu_next_md_free_region - finds the next hint free area 329525ca84dSDennis Zhou (Facebook) * @chunk: chunk of interest 330525ca84dSDennis Zhou (Facebook) * @bit_off: chunk offset 331525ca84dSDennis Zhou (Facebook) * @bits: size of free area 332525ca84dSDennis Zhou (Facebook) * 333525ca84dSDennis Zhou (Facebook) * Helper function for pcpu_for_each_md_free_region. It checks 334525ca84dSDennis Zhou (Facebook) * block->contig_hint and performs aggregation across blocks to find the 335525ca84dSDennis Zhou (Facebook) * next hint. It modifies bit_off and bits in-place to be consumed in the 336525ca84dSDennis Zhou (Facebook) * loop. 337525ca84dSDennis Zhou (Facebook) */ 338525ca84dSDennis Zhou (Facebook) static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, 339525ca84dSDennis Zhou (Facebook) int *bits) 340525ca84dSDennis Zhou (Facebook) { 341525ca84dSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 342525ca84dSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 343525ca84dSDennis Zhou (Facebook) struct pcpu_block_md *block; 344525ca84dSDennis Zhou (Facebook) 345525ca84dSDennis Zhou (Facebook) *bits = 0; 346525ca84dSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 347525ca84dSDennis Zhou (Facebook) block++, i++) { 348525ca84dSDennis Zhou (Facebook) /* handles contig area across blocks */ 349525ca84dSDennis Zhou (Facebook) if (*bits) { 350525ca84dSDennis Zhou (Facebook) *bits += block->left_free; 351525ca84dSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 352525ca84dSDennis Zhou (Facebook) continue; 353525ca84dSDennis Zhou (Facebook) return; 354525ca84dSDennis Zhou (Facebook) } 355525ca84dSDennis Zhou (Facebook) 356525ca84dSDennis Zhou (Facebook) /* 357525ca84dSDennis Zhou (Facebook) * This checks three things. First is there a contig_hint to 358525ca84dSDennis Zhou (Facebook) * check. Second, have we checked this hint before by 359525ca84dSDennis Zhou (Facebook) * comparing the block_off. Third, is this the same as the 360525ca84dSDennis Zhou (Facebook) * right contig hint. In the last case, it spills over into 361525ca84dSDennis Zhou (Facebook) * the next block and should be handled by the contig area 362525ca84dSDennis Zhou (Facebook) * across blocks code. 363525ca84dSDennis Zhou (Facebook) */ 364525ca84dSDennis Zhou (Facebook) *bits = block->contig_hint; 365525ca84dSDennis Zhou (Facebook) if (*bits && block->contig_hint_start >= block_off && 366525ca84dSDennis Zhou (Facebook) *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { 367525ca84dSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, 368525ca84dSDennis Zhou (Facebook) block->contig_hint_start); 369525ca84dSDennis Zhou (Facebook) return; 370525ca84dSDennis Zhou (Facebook) } 3711fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 3721fa4df3eSDennis Zhou block_off = 0; 373525ca84dSDennis Zhou (Facebook) 374525ca84dSDennis Zhou (Facebook) *bits = block->right_free; 375525ca84dSDennis Zhou (Facebook) *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; 376525ca84dSDennis Zhou (Facebook) } 377525ca84dSDennis Zhou (Facebook) } 378525ca84dSDennis Zhou (Facebook) 379b4c2116cSDennis Zhou (Facebook) /** 380b4c2116cSDennis Zhou (Facebook) * pcpu_next_fit_region - finds fit areas for a given allocation request 381b4c2116cSDennis Zhou (Facebook) * @chunk: chunk of interest 382b4c2116cSDennis Zhou (Facebook) * @alloc_bits: size of allocation 383b4c2116cSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 384b4c2116cSDennis Zhou (Facebook) * @bit_off: chunk offset 385b4c2116cSDennis Zhou (Facebook) * @bits: size of free area 386b4c2116cSDennis Zhou (Facebook) * 387b4c2116cSDennis Zhou (Facebook) * Finds the next free region that is viable for use with a given size and 388b4c2116cSDennis Zhou (Facebook) * alignment. This only returns if there is a valid area to be used for this 389b4c2116cSDennis Zhou (Facebook) * allocation. block->first_free is returned if the allocation request fits 390b4c2116cSDennis Zhou (Facebook) * within the block to see if the request can be fulfilled prior to the contig 391b4c2116cSDennis Zhou (Facebook) * hint. 392b4c2116cSDennis Zhou (Facebook) */ 393b4c2116cSDennis Zhou (Facebook) static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, 394b4c2116cSDennis Zhou (Facebook) int align, int *bit_off, int *bits) 395b4c2116cSDennis Zhou (Facebook) { 396b4c2116cSDennis Zhou (Facebook) int i = pcpu_off_to_block_index(*bit_off); 397b4c2116cSDennis Zhou (Facebook) int block_off = pcpu_off_to_block_off(*bit_off); 398b4c2116cSDennis Zhou (Facebook) struct pcpu_block_md *block; 399b4c2116cSDennis Zhou (Facebook) 400b4c2116cSDennis Zhou (Facebook) *bits = 0; 401b4c2116cSDennis Zhou (Facebook) for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); 402b4c2116cSDennis Zhou (Facebook) block++, i++) { 403b4c2116cSDennis Zhou (Facebook) /* handles contig area across blocks */ 404b4c2116cSDennis Zhou (Facebook) if (*bits) { 405b4c2116cSDennis Zhou (Facebook) *bits += block->left_free; 406b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 407b4c2116cSDennis Zhou (Facebook) return; 408b4c2116cSDennis Zhou (Facebook) if (block->left_free == PCPU_BITMAP_BLOCK_BITS) 409b4c2116cSDennis Zhou (Facebook) continue; 410b4c2116cSDennis Zhou (Facebook) } 411b4c2116cSDennis Zhou (Facebook) 412b4c2116cSDennis Zhou (Facebook) /* check block->contig_hint */ 413b4c2116cSDennis Zhou (Facebook) *bits = ALIGN(block->contig_hint_start, align) - 414b4c2116cSDennis Zhou (Facebook) block->contig_hint_start; 415b4c2116cSDennis Zhou (Facebook) /* 416b4c2116cSDennis Zhou (Facebook) * This uses the block offset to determine if this has been 417b4c2116cSDennis Zhou (Facebook) * checked in the prior iteration. 418b4c2116cSDennis Zhou (Facebook) */ 419b4c2116cSDennis Zhou (Facebook) if (block->contig_hint && 420b4c2116cSDennis Zhou (Facebook) block->contig_hint_start >= block_off && 421b4c2116cSDennis Zhou (Facebook) block->contig_hint >= *bits + alloc_bits) { 422382b88e9SDennis Zhou int start = pcpu_next_hint(block, alloc_bits); 423382b88e9SDennis Zhou 424b4c2116cSDennis Zhou (Facebook) *bits += alloc_bits + block->contig_hint_start - 425382b88e9SDennis Zhou start; 426382b88e9SDennis Zhou *bit_off = pcpu_block_off_to_off(i, start); 427b4c2116cSDennis Zhou (Facebook) return; 428b4c2116cSDennis Zhou (Facebook) } 4291fa4df3eSDennis Zhou /* reset to satisfy the second predicate above */ 4301fa4df3eSDennis Zhou block_off = 0; 431b4c2116cSDennis Zhou (Facebook) 432b4c2116cSDennis Zhou (Facebook) *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, 433b4c2116cSDennis Zhou (Facebook) align); 434b4c2116cSDennis Zhou (Facebook) *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; 435b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_block_off_to_off(i, *bit_off); 436b4c2116cSDennis Zhou (Facebook) if (*bits >= alloc_bits) 437b4c2116cSDennis Zhou (Facebook) return; 438b4c2116cSDennis Zhou (Facebook) } 439b4c2116cSDennis Zhou (Facebook) 440b4c2116cSDennis Zhou (Facebook) /* no valid offsets were found - fail condition */ 441b4c2116cSDennis Zhou (Facebook) *bit_off = pcpu_chunk_map_bits(chunk); 442b4c2116cSDennis Zhou (Facebook) } 443b4c2116cSDennis Zhou (Facebook) 444525ca84dSDennis Zhou (Facebook) /* 445525ca84dSDennis Zhou (Facebook) * Metadata free area iterators. These perform aggregation of free areas 446525ca84dSDennis Zhou (Facebook) * based on the metadata blocks and return the offset @bit_off and size in 447b4c2116cSDennis Zhou (Facebook) * bits of the free area @bits. pcpu_for_each_fit_region only returns when 448b4c2116cSDennis Zhou (Facebook) * a fit is found for the allocation request. 449525ca84dSDennis Zhou (Facebook) */ 450525ca84dSDennis Zhou (Facebook) #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ 451525ca84dSDennis Zhou (Facebook) for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ 452525ca84dSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 453525ca84dSDennis Zhou (Facebook) (bit_off) += (bits) + 1, \ 454525ca84dSDennis Zhou (Facebook) pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) 455525ca84dSDennis Zhou (Facebook) 456b4c2116cSDennis Zhou (Facebook) #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ 457b4c2116cSDennis Zhou (Facebook) for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 458b4c2116cSDennis Zhou (Facebook) &(bits)); \ 459b4c2116cSDennis Zhou (Facebook) (bit_off) < pcpu_chunk_map_bits((chunk)); \ 460b4c2116cSDennis Zhou (Facebook) (bit_off) += (bits), \ 461b4c2116cSDennis Zhou (Facebook) pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ 462b4c2116cSDennis Zhou (Facebook) &(bits))) 463b4c2116cSDennis Zhou (Facebook) 464525ca84dSDennis Zhou (Facebook) /** 46590459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 4661880d93bSTejun Heo * @size: bytes to allocate 46747504ee0SDennis Zhou * @gfp: allocation flags 468fbf59bc9STejun Heo * 4691880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 47047504ee0SDennis Zhou * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. 47147504ee0SDennis Zhou * This is to facilitate passing through whitelisted flags. The 47247504ee0SDennis Zhou * returned memory is always zeroed. 473fbf59bc9STejun Heo * 474fbf59bc9STejun Heo * RETURNS: 4751880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 476fbf59bc9STejun Heo */ 47747504ee0SDennis Zhou static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) 478fbf59bc9STejun Heo { 479099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 480099a19d9STejun Heo return NULL; 481099a19d9STejun Heo 482fbf59bc9STejun Heo if (size <= PAGE_SIZE) 483554fef1cSDennis Zhou return kzalloc(size, gfp); 4847af4c093SJesper Juhl else 48588dca4caSChristoph Hellwig return __vmalloc(size, gfp | __GFP_ZERO); 4861880d93bSTejun Heo } 487fbf59bc9STejun Heo 4881880d93bSTejun Heo /** 4891880d93bSTejun Heo * pcpu_mem_free - free memory 4901880d93bSTejun Heo * @ptr: memory to free 4911880d93bSTejun Heo * 49290459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 4931880d93bSTejun Heo */ 4941d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 4951880d93bSTejun Heo { 4961d5cfdb0STetsuo Handa kvfree(ptr); 497fbf59bc9STejun Heo } 498fbf59bc9STejun Heo 4998744d859SDennis Zhou static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot, 5008744d859SDennis Zhou bool move_front) 5018744d859SDennis Zhou { 5028744d859SDennis Zhou if (chunk != pcpu_reserved_chunk) { 5038744d859SDennis Zhou if (move_front) 5048744d859SDennis Zhou list_move(&chunk->list, &pcpu_slot[slot]); 5058744d859SDennis Zhou else 5068744d859SDennis Zhou list_move_tail(&chunk->list, &pcpu_slot[slot]); 5078744d859SDennis Zhou } 5088744d859SDennis Zhou } 5098744d859SDennis Zhou 5108744d859SDennis Zhou static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot) 5118744d859SDennis Zhou { 5128744d859SDennis Zhou __pcpu_chunk_move(chunk, slot, true); 5138744d859SDennis Zhou } 5148744d859SDennis Zhou 515fbf59bc9STejun Heo /** 516fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 517fbf59bc9STejun Heo * @chunk: chunk of interest 518fbf59bc9STejun Heo * @oslot: the previous slot it was on 519fbf59bc9STejun Heo * 520fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 521fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 522edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 523edcb4639STejun Heo * chunk slots. 524ccea34b5STejun Heo * 525ccea34b5STejun Heo * CONTEXT: 526ccea34b5STejun Heo * pcpu_lock. 527fbf59bc9STejun Heo */ 528fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 529fbf59bc9STejun Heo { 530fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 531fbf59bc9STejun Heo 5328744d859SDennis Zhou if (oslot != nslot) 5338744d859SDennis Zhou __pcpu_chunk_move(chunk, nslot, oslot < nslot); 53440064aecSDennis Zhou (Facebook) } 53540064aecSDennis Zhou (Facebook) 53640064aecSDennis Zhou (Facebook) /* 537b239f7daSDennis Zhou * pcpu_update_empty_pages - update empty page counters 538b239f7daSDennis Zhou * @chunk: chunk of interest 539b239f7daSDennis Zhou * @nr: nr of empty pages 54040064aecSDennis Zhou (Facebook) * 541b239f7daSDennis Zhou * This is used to keep track of the empty pages now based on the premise 542b239f7daSDennis Zhou * a md_block covers a page. The hint update functions recognize if a block 543b239f7daSDennis Zhou * is made full or broken to calculate deltas for keeping track of free pages. 54440064aecSDennis Zhou (Facebook) */ 545b239f7daSDennis Zhou static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr) 546b239f7daSDennis Zhou { 547b239f7daSDennis Zhou chunk->nr_empty_pop_pages += nr; 54840064aecSDennis Zhou (Facebook) if (chunk != pcpu_reserved_chunk) 549b239f7daSDennis Zhou pcpu_nr_empty_pop_pages += nr; 55040064aecSDennis Zhou (Facebook) } 55140064aecSDennis Zhou (Facebook) 552d9f3a01eSDennis Zhou /* 553d9f3a01eSDennis Zhou * pcpu_region_overlap - determines if two regions overlap 554d9f3a01eSDennis Zhou * @a: start of first region, inclusive 555d9f3a01eSDennis Zhou * @b: end of first region, exclusive 556d9f3a01eSDennis Zhou * @x: start of second region, inclusive 557d9f3a01eSDennis Zhou * @y: end of second region, exclusive 558d9f3a01eSDennis Zhou * 559d9f3a01eSDennis Zhou * This is used to determine if the hint region [a, b) overlaps with the 560d9f3a01eSDennis Zhou * allocated region [x, y). 561d9f3a01eSDennis Zhou */ 562d9f3a01eSDennis Zhou static inline bool pcpu_region_overlap(int a, int b, int x, int y) 563d9f3a01eSDennis Zhou { 564d9f3a01eSDennis Zhou return (a < y) && (x < b); 56540064aecSDennis Zhou (Facebook) } 56640064aecSDennis Zhou (Facebook) 56740064aecSDennis Zhou (Facebook) /** 568ca460b3cSDennis Zhou (Facebook) * pcpu_block_update - updates a block given a free area 569ca460b3cSDennis Zhou (Facebook) * @block: block of interest 570ca460b3cSDennis Zhou (Facebook) * @start: start offset in block 571ca460b3cSDennis Zhou (Facebook) * @end: end offset in block 572ca460b3cSDennis Zhou (Facebook) * 573ca460b3cSDennis Zhou (Facebook) * Updates a block given a known free area. The region [start, end) is 574268625a6SDennis Zhou (Facebook) * expected to be the entirety of the free area within a block. Chooses 575268625a6SDennis Zhou (Facebook) * the best starting offset if the contig hints are equal. 576ca460b3cSDennis Zhou (Facebook) */ 577ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 578ca460b3cSDennis Zhou (Facebook) { 579ca460b3cSDennis Zhou (Facebook) int contig = end - start; 580ca460b3cSDennis Zhou (Facebook) 581ca460b3cSDennis Zhou (Facebook) block->first_free = min(block->first_free, start); 582ca460b3cSDennis Zhou (Facebook) if (start == 0) 583ca460b3cSDennis Zhou (Facebook) block->left_free = contig; 584ca460b3cSDennis Zhou (Facebook) 585047924c9SDennis Zhou if (end == block->nr_bits) 586ca460b3cSDennis Zhou (Facebook) block->right_free = contig; 587ca460b3cSDennis Zhou (Facebook) 588ca460b3cSDennis Zhou (Facebook) if (contig > block->contig_hint) { 589382b88e9SDennis Zhou /* promote the old contig_hint to be the new scan_hint */ 590382b88e9SDennis Zhou if (start > block->contig_hint_start) { 591382b88e9SDennis Zhou if (block->contig_hint > block->scan_hint) { 592382b88e9SDennis Zhou block->scan_hint_start = 593382b88e9SDennis Zhou block->contig_hint_start; 594382b88e9SDennis Zhou block->scan_hint = block->contig_hint; 595382b88e9SDennis Zhou } else if (start < block->scan_hint_start) { 596382b88e9SDennis Zhou /* 597382b88e9SDennis Zhou * The old contig_hint == scan_hint. But, the 598382b88e9SDennis Zhou * new contig is larger so hold the invariant 599382b88e9SDennis Zhou * scan_hint_start < contig_hint_start. 600382b88e9SDennis Zhou */ 601382b88e9SDennis Zhou block->scan_hint = 0; 602382b88e9SDennis Zhou } 603382b88e9SDennis Zhou } else { 604382b88e9SDennis Zhou block->scan_hint = 0; 605382b88e9SDennis Zhou } 606ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = start; 607ca460b3cSDennis Zhou (Facebook) block->contig_hint = contig; 608382b88e9SDennis Zhou } else if (contig == block->contig_hint) { 609382b88e9SDennis Zhou if (block->contig_hint_start && 610382b88e9SDennis Zhou (!start || 611382b88e9SDennis Zhou __ffs(start) > __ffs(block->contig_hint_start))) { 612382b88e9SDennis Zhou /* start has a better alignment so use it */ 613268625a6SDennis Zhou (Facebook) block->contig_hint_start = start; 614382b88e9SDennis Zhou if (start < block->scan_hint_start && 615382b88e9SDennis Zhou block->contig_hint > block->scan_hint) 616382b88e9SDennis Zhou block->scan_hint = 0; 617382b88e9SDennis Zhou } else if (start > block->scan_hint_start || 618382b88e9SDennis Zhou block->contig_hint > block->scan_hint) { 619382b88e9SDennis Zhou /* 620382b88e9SDennis Zhou * Knowing contig == contig_hint, update the scan_hint 621382b88e9SDennis Zhou * if it is farther than or larger than the current 622382b88e9SDennis Zhou * scan_hint. 623382b88e9SDennis Zhou */ 624382b88e9SDennis Zhou block->scan_hint_start = start; 625382b88e9SDennis Zhou block->scan_hint = contig; 626382b88e9SDennis Zhou } 627382b88e9SDennis Zhou } else { 628382b88e9SDennis Zhou /* 629382b88e9SDennis Zhou * The region is smaller than the contig_hint. So only update 630382b88e9SDennis Zhou * the scan_hint if it is larger than or equal and farther than 631382b88e9SDennis Zhou * the current scan_hint. 632382b88e9SDennis Zhou */ 633382b88e9SDennis Zhou if ((start < block->contig_hint_start && 634382b88e9SDennis Zhou (contig > block->scan_hint || 635382b88e9SDennis Zhou (contig == block->scan_hint && 636382b88e9SDennis Zhou start > block->scan_hint_start)))) { 637382b88e9SDennis Zhou block->scan_hint_start = start; 638382b88e9SDennis Zhou block->scan_hint = contig; 639382b88e9SDennis Zhou } 640ca460b3cSDennis Zhou (Facebook) } 641ca460b3cSDennis Zhou (Facebook) } 642ca460b3cSDennis Zhou (Facebook) 643b89462a9SDennis Zhou /* 644b89462a9SDennis Zhou * pcpu_block_update_scan - update a block given a free area from a scan 645b89462a9SDennis Zhou * @chunk: chunk of interest 646b89462a9SDennis Zhou * @bit_off: chunk offset 647b89462a9SDennis Zhou * @bits: size of free area 648b89462a9SDennis Zhou * 649b89462a9SDennis Zhou * Finding the final allocation spot first goes through pcpu_find_block_fit() 650b89462a9SDennis Zhou * to find a block that can hold the allocation and then pcpu_alloc_area() 651b89462a9SDennis Zhou * where a scan is used. When allocations require specific alignments, 652b89462a9SDennis Zhou * we can inadvertently create holes which will not be seen in the alloc 653b89462a9SDennis Zhou * or free paths. 654b89462a9SDennis Zhou * 655b89462a9SDennis Zhou * This takes a given free area hole and updates a block as it may change the 656b89462a9SDennis Zhou * scan_hint. We need to scan backwards to ensure we don't miss free bits 657b89462a9SDennis Zhou * from alignment. 658b89462a9SDennis Zhou */ 659b89462a9SDennis Zhou static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off, 660b89462a9SDennis Zhou int bits) 661b89462a9SDennis Zhou { 662b89462a9SDennis Zhou int s_off = pcpu_off_to_block_off(bit_off); 663b89462a9SDennis Zhou int e_off = s_off + bits; 664b89462a9SDennis Zhou int s_index, l_bit; 665b89462a9SDennis Zhou struct pcpu_block_md *block; 666b89462a9SDennis Zhou 667b89462a9SDennis Zhou if (e_off > PCPU_BITMAP_BLOCK_BITS) 668b89462a9SDennis Zhou return; 669b89462a9SDennis Zhou 670b89462a9SDennis Zhou s_index = pcpu_off_to_block_index(bit_off); 671b89462a9SDennis Zhou block = chunk->md_blocks + s_index; 672b89462a9SDennis Zhou 673b89462a9SDennis Zhou /* scan backwards in case of alignment skipping free bits */ 674b89462a9SDennis Zhou l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off); 675b89462a9SDennis Zhou s_off = (s_off == l_bit) ? 0 : l_bit + 1; 676b89462a9SDennis Zhou 677b89462a9SDennis Zhou pcpu_block_update(block, s_off, e_off); 678b89462a9SDennis Zhou } 679b89462a9SDennis Zhou 680ca460b3cSDennis Zhou (Facebook) /** 68192c14cabSDennis Zhou * pcpu_chunk_refresh_hint - updates metadata about a chunk 68292c14cabSDennis Zhou * @chunk: chunk of interest 683d33d9f3dSDennis Zhou * @full_scan: if we should scan from the beginning 68492c14cabSDennis Zhou * 68592c14cabSDennis Zhou * Iterates over the metadata blocks to find the largest contig area. 686d33d9f3dSDennis Zhou * A full scan can be avoided on the allocation path as this is triggered 687d33d9f3dSDennis Zhou * if we broke the contig_hint. In doing so, the scan_hint will be before 688d33d9f3dSDennis Zhou * the contig_hint or after if the scan_hint == contig_hint. This cannot 689d33d9f3dSDennis Zhou * be prevented on freeing as we want to find the largest area possibly 690d33d9f3dSDennis Zhou * spanning blocks. 69192c14cabSDennis Zhou */ 692d33d9f3dSDennis Zhou static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan) 69392c14cabSDennis Zhou { 69492c14cabSDennis Zhou struct pcpu_block_md *chunk_md = &chunk->chunk_md; 69592c14cabSDennis Zhou int bit_off, bits; 69692c14cabSDennis Zhou 697d33d9f3dSDennis Zhou /* promote scan_hint to contig_hint */ 698d33d9f3dSDennis Zhou if (!full_scan && chunk_md->scan_hint) { 699d33d9f3dSDennis Zhou bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint; 700d33d9f3dSDennis Zhou chunk_md->contig_hint_start = chunk_md->scan_hint_start; 701d33d9f3dSDennis Zhou chunk_md->contig_hint = chunk_md->scan_hint; 702d33d9f3dSDennis Zhou chunk_md->scan_hint = 0; 703d33d9f3dSDennis Zhou } else { 70492c14cabSDennis Zhou bit_off = chunk_md->first_free; 705d33d9f3dSDennis Zhou chunk_md->contig_hint = 0; 706d33d9f3dSDennis Zhou } 707d33d9f3dSDennis Zhou 70892c14cabSDennis Zhou bits = 0; 709e837dfdeSDennis Zhou pcpu_for_each_md_free_region(chunk, bit_off, bits) 71092c14cabSDennis Zhou pcpu_block_update(chunk_md, bit_off, bit_off + bits); 711ca460b3cSDennis Zhou (Facebook) } 712ca460b3cSDennis Zhou (Facebook) 713ca460b3cSDennis Zhou (Facebook) /** 714ca460b3cSDennis Zhou (Facebook) * pcpu_block_refresh_hint 715ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 716ca460b3cSDennis Zhou (Facebook) * @index: index of the metadata block 717ca460b3cSDennis Zhou (Facebook) * 718ca460b3cSDennis Zhou (Facebook) * Scans over the block beginning at first_free and updates the block 719ca460b3cSDennis Zhou (Facebook) * metadata accordingly. 720ca460b3cSDennis Zhou (Facebook) */ 721ca460b3cSDennis Zhou (Facebook) static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 722ca460b3cSDennis Zhou (Facebook) { 723ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *block = chunk->md_blocks + index; 724ca460b3cSDennis Zhou (Facebook) unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 725e837dfdeSDennis Zhou unsigned int rs, re, start; /* region start, region end */ 726ca460b3cSDennis Zhou (Facebook) 727da3afdd5SDennis Zhou /* promote scan_hint to contig_hint */ 728da3afdd5SDennis Zhou if (block->scan_hint) { 729da3afdd5SDennis Zhou start = block->scan_hint_start + block->scan_hint; 730da3afdd5SDennis Zhou block->contig_hint_start = block->scan_hint_start; 731da3afdd5SDennis Zhou block->contig_hint = block->scan_hint; 732da3afdd5SDennis Zhou block->scan_hint = 0; 733da3afdd5SDennis Zhou } else { 734da3afdd5SDennis Zhou start = block->first_free; 735ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 736da3afdd5SDennis Zhou } 737da3afdd5SDennis Zhou 738da3afdd5SDennis Zhou block->right_free = 0; 739ca460b3cSDennis Zhou (Facebook) 740ca460b3cSDennis Zhou (Facebook) /* iterate over free areas and update the contig hints */ 741e837dfdeSDennis Zhou bitmap_for_each_clear_region(alloc_map, rs, re, start, 742e837dfdeSDennis Zhou PCPU_BITMAP_BLOCK_BITS) 743ca460b3cSDennis Zhou (Facebook) pcpu_block_update(block, rs, re); 744ca460b3cSDennis Zhou (Facebook) } 745ca460b3cSDennis Zhou (Facebook) 746ca460b3cSDennis Zhou (Facebook) /** 747ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_alloc - update hint on allocation path 748ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 749ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 750ca460b3cSDennis Zhou (Facebook) * @bits: size of request 751fc304334SDennis Zhou (Facebook) * 752fc304334SDennis Zhou (Facebook) * Updates metadata for the allocation path. The metadata only has to be 753fc304334SDennis Zhou (Facebook) * refreshed by a full scan iff the chunk's contig hint is broken. Block level 754fc304334SDennis Zhou (Facebook) * scans are required if the block's contig hint is broken. 755ca460b3cSDennis Zhou (Facebook) */ 756ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 757ca460b3cSDennis Zhou (Facebook) int bits) 758ca460b3cSDennis Zhou (Facebook) { 75992c14cabSDennis Zhou struct pcpu_block_md *chunk_md = &chunk->chunk_md; 760b239f7daSDennis Zhou int nr_empty_pages = 0; 761ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 762ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 763ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 764ca460b3cSDennis Zhou (Facebook) 765ca460b3cSDennis Zhou (Facebook) /* 766ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 767ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 768ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 769ca460b3cSDennis Zhou (Facebook) * range. 770ca460b3cSDennis Zhou (Facebook) */ 771ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 772ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 773ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 774ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 775ca460b3cSDennis Zhou (Facebook) 776ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 777ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 778ca460b3cSDennis Zhou (Facebook) 779ca460b3cSDennis Zhou (Facebook) /* 780ca460b3cSDennis Zhou (Facebook) * Update s_block. 781fc304334SDennis Zhou (Facebook) * block->first_free must be updated if the allocation takes its place. 782fc304334SDennis Zhou (Facebook) * If the allocation breaks the contig_hint, a scan is required to 783fc304334SDennis Zhou (Facebook) * restore this hint. 784ca460b3cSDennis Zhou (Facebook) */ 785b239f7daSDennis Zhou if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 786b239f7daSDennis Zhou nr_empty_pages++; 787b239f7daSDennis Zhou 788fc304334SDennis Zhou (Facebook) if (s_off == s_block->first_free) 789fc304334SDennis Zhou (Facebook) s_block->first_free = find_next_zero_bit( 790fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, s_index), 791fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, 792fc304334SDennis Zhou (Facebook) s_off + bits); 793fc304334SDennis Zhou (Facebook) 794382b88e9SDennis Zhou if (pcpu_region_overlap(s_block->scan_hint_start, 795382b88e9SDennis Zhou s_block->scan_hint_start + s_block->scan_hint, 796382b88e9SDennis Zhou s_off, 797382b88e9SDennis Zhou s_off + bits)) 798382b88e9SDennis Zhou s_block->scan_hint = 0; 799382b88e9SDennis Zhou 800d9f3a01eSDennis Zhou if (pcpu_region_overlap(s_block->contig_hint_start, 801d9f3a01eSDennis Zhou s_block->contig_hint_start + 802d9f3a01eSDennis Zhou s_block->contig_hint, 803d9f3a01eSDennis Zhou s_off, 804d9f3a01eSDennis Zhou s_off + bits)) { 805fc304334SDennis Zhou (Facebook) /* block contig hint is broken - scan to fix it */ 806da3afdd5SDennis Zhou if (!s_off) 807da3afdd5SDennis Zhou s_block->left_free = 0; 808ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 809fc304334SDennis Zhou (Facebook) } else { 810fc304334SDennis Zhou (Facebook) /* update left and right contig manually */ 811fc304334SDennis Zhou (Facebook) s_block->left_free = min(s_block->left_free, s_off); 812fc304334SDennis Zhou (Facebook) if (s_index == e_index) 813fc304334SDennis Zhou (Facebook) s_block->right_free = min_t(int, s_block->right_free, 814fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 815fc304334SDennis Zhou (Facebook) else 816fc304334SDennis Zhou (Facebook) s_block->right_free = 0; 817fc304334SDennis Zhou (Facebook) } 818ca460b3cSDennis Zhou (Facebook) 819ca460b3cSDennis Zhou (Facebook) /* 820ca460b3cSDennis Zhou (Facebook) * Update e_block. 821ca460b3cSDennis Zhou (Facebook) */ 822ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 823b239f7daSDennis Zhou if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS) 824b239f7daSDennis Zhou nr_empty_pages++; 825b239f7daSDennis Zhou 826fc304334SDennis Zhou (Facebook) /* 827fc304334SDennis Zhou (Facebook) * When the allocation is across blocks, the end is along 828fc304334SDennis Zhou (Facebook) * the left part of the e_block. 829fc304334SDennis Zhou (Facebook) */ 830fc304334SDennis Zhou (Facebook) e_block->first_free = find_next_zero_bit( 831fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, e_index), 832fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, e_off); 833fc304334SDennis Zhou (Facebook) 834fc304334SDennis Zhou (Facebook) if (e_off == PCPU_BITMAP_BLOCK_BITS) { 835fc304334SDennis Zhou (Facebook) /* reset the block */ 836fc304334SDennis Zhou (Facebook) e_block++; 837fc304334SDennis Zhou (Facebook) } else { 838382b88e9SDennis Zhou if (e_off > e_block->scan_hint_start) 839382b88e9SDennis Zhou e_block->scan_hint = 0; 840382b88e9SDennis Zhou 841da3afdd5SDennis Zhou e_block->left_free = 0; 842fc304334SDennis Zhou (Facebook) if (e_off > e_block->contig_hint_start) { 843fc304334SDennis Zhou (Facebook) /* contig hint is broken - scan to fix it */ 844ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 845fc304334SDennis Zhou (Facebook) } else { 846fc304334SDennis Zhou (Facebook) e_block->right_free = 847fc304334SDennis Zhou (Facebook) min_t(int, e_block->right_free, 848fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 849fc304334SDennis Zhou (Facebook) } 850fc304334SDennis Zhou (Facebook) } 851ca460b3cSDennis Zhou (Facebook) 852ca460b3cSDennis Zhou (Facebook) /* update in-between md_blocks */ 853b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 854ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 855382b88e9SDennis Zhou block->scan_hint = 0; 856ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 857ca460b3cSDennis Zhou (Facebook) block->left_free = 0; 858ca460b3cSDennis Zhou (Facebook) block->right_free = 0; 859ca460b3cSDennis Zhou (Facebook) } 860ca460b3cSDennis Zhou (Facebook) } 861ca460b3cSDennis Zhou (Facebook) 862b239f7daSDennis Zhou if (nr_empty_pages) 863b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr_empty_pages); 864b239f7daSDennis Zhou 865d33d9f3dSDennis Zhou if (pcpu_region_overlap(chunk_md->scan_hint_start, 866d33d9f3dSDennis Zhou chunk_md->scan_hint_start + 867d33d9f3dSDennis Zhou chunk_md->scan_hint, 868d33d9f3dSDennis Zhou bit_off, 869d33d9f3dSDennis Zhou bit_off + bits)) 870d33d9f3dSDennis Zhou chunk_md->scan_hint = 0; 871d33d9f3dSDennis Zhou 872fc304334SDennis Zhou (Facebook) /* 873fc304334SDennis Zhou (Facebook) * The only time a full chunk scan is required is if the chunk 874fc304334SDennis Zhou (Facebook) * contig hint is broken. Otherwise, it means a smaller space 875fc304334SDennis Zhou (Facebook) * was used and therefore the chunk contig hint is still correct. 876fc304334SDennis Zhou (Facebook) */ 87792c14cabSDennis Zhou if (pcpu_region_overlap(chunk_md->contig_hint_start, 87892c14cabSDennis Zhou chunk_md->contig_hint_start + 87992c14cabSDennis Zhou chunk_md->contig_hint, 880d9f3a01eSDennis Zhou bit_off, 881d9f3a01eSDennis Zhou bit_off + bits)) 882d33d9f3dSDennis Zhou pcpu_chunk_refresh_hint(chunk, false); 883ca460b3cSDennis Zhou (Facebook) } 884ca460b3cSDennis Zhou (Facebook) 885ca460b3cSDennis Zhou (Facebook) /** 886ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_free - updates the block hints on the free path 887ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 888ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 889ca460b3cSDennis Zhou (Facebook) * @bits: size of request 890b185cd0dSDennis Zhou (Facebook) * 891b185cd0dSDennis Zhou (Facebook) * Updates metadata for the allocation path. This avoids a blind block 892b185cd0dSDennis Zhou (Facebook) * refresh by making use of the block contig hints. If this fails, it scans 893b185cd0dSDennis Zhou (Facebook) * forward and backward to determine the extent of the free area. This is 894b185cd0dSDennis Zhou (Facebook) * capped at the boundary of blocks. 895b185cd0dSDennis Zhou (Facebook) * 896b185cd0dSDennis Zhou (Facebook) * A chunk update is triggered if a page becomes free, a block becomes free, 897b185cd0dSDennis Zhou (Facebook) * or the free spans across blocks. This tradeoff is to minimize iterating 89892c14cabSDennis Zhou * over the block metadata to update chunk_md->contig_hint. 89992c14cabSDennis Zhou * chunk_md->contig_hint may be off by up to a page, but it will never be more 90092c14cabSDennis Zhou * than the available space. If the contig hint is contained in one block, it 90192c14cabSDennis Zhou * will be accurate. 902ca460b3cSDennis Zhou (Facebook) */ 903ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 904ca460b3cSDennis Zhou (Facebook) int bits) 905ca460b3cSDennis Zhou (Facebook) { 906b239f7daSDennis Zhou int nr_empty_pages = 0; 907ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 908ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 909ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 910b185cd0dSDennis Zhou (Facebook) int start, end; /* start and end of the whole free area */ 911ca460b3cSDennis Zhou (Facebook) 912ca460b3cSDennis Zhou (Facebook) /* 913ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 914ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 915ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 916ca460b3cSDennis Zhou (Facebook) * range. 917ca460b3cSDennis Zhou (Facebook) */ 918ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 919ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 920ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 921ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 922ca460b3cSDennis Zhou (Facebook) 923ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 924ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 925ca460b3cSDennis Zhou (Facebook) 926b185cd0dSDennis Zhou (Facebook) /* 927b185cd0dSDennis Zhou (Facebook) * Check if the freed area aligns with the block->contig_hint. 928b185cd0dSDennis Zhou (Facebook) * If it does, then the scan to find the beginning/end of the 929b185cd0dSDennis Zhou (Facebook) * larger free area can be avoided. 930b185cd0dSDennis Zhou (Facebook) * 931b185cd0dSDennis Zhou (Facebook) * start and end refer to beginning and end of the free area 932b185cd0dSDennis Zhou (Facebook) * within each their respective blocks. This is not necessarily 933b185cd0dSDennis Zhou (Facebook) * the entire free area as it may span blocks past the beginning 934b185cd0dSDennis Zhou (Facebook) * or end of the block. 935b185cd0dSDennis Zhou (Facebook) */ 936b185cd0dSDennis Zhou (Facebook) start = s_off; 937b185cd0dSDennis Zhou (Facebook) if (s_off == s_block->contig_hint + s_block->contig_hint_start) { 938b185cd0dSDennis Zhou (Facebook) start = s_block->contig_hint_start; 939b185cd0dSDennis Zhou (Facebook) } else { 940b185cd0dSDennis Zhou (Facebook) /* 941b185cd0dSDennis Zhou (Facebook) * Scan backwards to find the extent of the free area. 942b185cd0dSDennis Zhou (Facebook) * find_last_bit returns the starting bit, so if the start bit 943b185cd0dSDennis Zhou (Facebook) * is returned, that means there was no last bit and the 944b185cd0dSDennis Zhou (Facebook) * remainder of the chunk is free. 945b185cd0dSDennis Zhou (Facebook) */ 946b185cd0dSDennis Zhou (Facebook) int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), 947b185cd0dSDennis Zhou (Facebook) start); 948b185cd0dSDennis Zhou (Facebook) start = (start == l_bit) ? 0 : l_bit + 1; 949b185cd0dSDennis Zhou (Facebook) } 950b185cd0dSDennis Zhou (Facebook) 951b185cd0dSDennis Zhou (Facebook) end = e_off; 952b185cd0dSDennis Zhou (Facebook) if (e_off == e_block->contig_hint_start) 953b185cd0dSDennis Zhou (Facebook) end = e_block->contig_hint_start + e_block->contig_hint; 954b185cd0dSDennis Zhou (Facebook) else 955b185cd0dSDennis Zhou (Facebook) end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), 956b185cd0dSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, end); 957b185cd0dSDennis Zhou (Facebook) 958ca460b3cSDennis Zhou (Facebook) /* update s_block */ 959b185cd0dSDennis Zhou (Facebook) e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; 960b239f7daSDennis Zhou if (!start && e_off == PCPU_BITMAP_BLOCK_BITS) 961b239f7daSDennis Zhou nr_empty_pages++; 962b185cd0dSDennis Zhou (Facebook) pcpu_block_update(s_block, start, e_off); 963ca460b3cSDennis Zhou (Facebook) 964ca460b3cSDennis Zhou (Facebook) /* freeing in the same block */ 965ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 966ca460b3cSDennis Zhou (Facebook) /* update e_block */ 967b239f7daSDennis Zhou if (end == PCPU_BITMAP_BLOCK_BITS) 968b239f7daSDennis Zhou nr_empty_pages++; 969b185cd0dSDennis Zhou (Facebook) pcpu_block_update(e_block, 0, end); 970ca460b3cSDennis Zhou (Facebook) 971ca460b3cSDennis Zhou (Facebook) /* reset md_blocks in the middle */ 972b239f7daSDennis Zhou nr_empty_pages += (e_index - s_index - 1); 973ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 974ca460b3cSDennis Zhou (Facebook) block->first_free = 0; 975382b88e9SDennis Zhou block->scan_hint = 0; 976ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = 0; 977ca460b3cSDennis Zhou (Facebook) block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 978ca460b3cSDennis Zhou (Facebook) block->left_free = PCPU_BITMAP_BLOCK_BITS; 979ca460b3cSDennis Zhou (Facebook) block->right_free = PCPU_BITMAP_BLOCK_BITS; 980ca460b3cSDennis Zhou (Facebook) } 981ca460b3cSDennis Zhou (Facebook) } 982ca460b3cSDennis Zhou (Facebook) 983b239f7daSDennis Zhou if (nr_empty_pages) 984b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr_empty_pages); 985b239f7daSDennis Zhou 986b185cd0dSDennis Zhou (Facebook) /* 987b239f7daSDennis Zhou * Refresh chunk metadata when the free makes a block free or spans 988b239f7daSDennis Zhou * across blocks. The contig_hint may be off by up to a page, but if 989b239f7daSDennis Zhou * the contig_hint is contained in a block, it will be accurate with 990b239f7daSDennis Zhou * the else condition below. 991b185cd0dSDennis Zhou (Facebook) */ 992b239f7daSDennis Zhou if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index) 993d33d9f3dSDennis Zhou pcpu_chunk_refresh_hint(chunk, true); 994b185cd0dSDennis Zhou (Facebook) else 99592c14cabSDennis Zhou pcpu_block_update(&chunk->chunk_md, 99692c14cabSDennis Zhou pcpu_block_off_to_off(s_index, start), 99792c14cabSDennis Zhou end); 998ca460b3cSDennis Zhou (Facebook) } 999ca460b3cSDennis Zhou (Facebook) 1000ca460b3cSDennis Zhou (Facebook) /** 100140064aecSDennis Zhou (Facebook) * pcpu_is_populated - determines if the region is populated 100240064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 100340064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 100440064aecSDennis Zhou (Facebook) * @bits: size of area 100540064aecSDennis Zhou (Facebook) * @next_off: return value for the next offset to start searching 100640064aecSDennis Zhou (Facebook) * 100740064aecSDennis Zhou (Facebook) * For atomic allocations, check if the backing pages are populated. 100840064aecSDennis Zhou (Facebook) * 100940064aecSDennis Zhou (Facebook) * RETURNS: 101040064aecSDennis Zhou (Facebook) * Bool if the backing pages are populated. 101140064aecSDennis Zhou (Facebook) * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 101240064aecSDennis Zhou (Facebook) */ 101340064aecSDennis Zhou (Facebook) static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 101440064aecSDennis Zhou (Facebook) int *next_off) 101540064aecSDennis Zhou (Facebook) { 1016e837dfdeSDennis Zhou unsigned int page_start, page_end, rs, re; 101740064aecSDennis Zhou (Facebook) 101840064aecSDennis Zhou (Facebook) page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 101940064aecSDennis Zhou (Facebook) page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 102040064aecSDennis Zhou (Facebook) 102140064aecSDennis Zhou (Facebook) rs = page_start; 1022e837dfdeSDennis Zhou bitmap_next_clear_region(chunk->populated, &rs, &re, page_end); 102340064aecSDennis Zhou (Facebook) if (rs >= page_end) 102440064aecSDennis Zhou (Facebook) return true; 102540064aecSDennis Zhou (Facebook) 102640064aecSDennis Zhou (Facebook) *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 102740064aecSDennis Zhou (Facebook) return false; 102840064aecSDennis Zhou (Facebook) } 102940064aecSDennis Zhou (Facebook) 103040064aecSDennis Zhou (Facebook) /** 103140064aecSDennis Zhou (Facebook) * pcpu_find_block_fit - finds the block index to start searching 103240064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 103340064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 103440064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE bytes) 103540064aecSDennis Zhou (Facebook) * @pop_only: use populated regions only 103640064aecSDennis Zhou (Facebook) * 1037b4c2116cSDennis Zhou (Facebook) * Given a chunk and an allocation spec, find the offset to begin searching 1038b4c2116cSDennis Zhou (Facebook) * for a free region. This iterates over the bitmap metadata blocks to 1039b4c2116cSDennis Zhou (Facebook) * find an offset that will be guaranteed to fit the requirements. It is 1040b4c2116cSDennis Zhou (Facebook) * not quite first fit as if the allocation does not fit in the contig hint 1041b4c2116cSDennis Zhou (Facebook) * of a block or chunk, it is skipped. This errs on the side of caution 1042b4c2116cSDennis Zhou (Facebook) * to prevent excess iteration. Poor alignment can cause the allocator to 1043b4c2116cSDennis Zhou (Facebook) * skip over blocks and chunks that have valid free areas. 1044b4c2116cSDennis Zhou (Facebook) * 104540064aecSDennis Zhou (Facebook) * RETURNS: 104640064aecSDennis Zhou (Facebook) * The offset in the bitmap to begin searching. 104740064aecSDennis Zhou (Facebook) * -1 if no offset is found. 104840064aecSDennis Zhou (Facebook) */ 104940064aecSDennis Zhou (Facebook) static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 105040064aecSDennis Zhou (Facebook) size_t align, bool pop_only) 105140064aecSDennis Zhou (Facebook) { 105292c14cabSDennis Zhou struct pcpu_block_md *chunk_md = &chunk->chunk_md; 1053b4c2116cSDennis Zhou (Facebook) int bit_off, bits, next_off; 105440064aecSDennis Zhou (Facebook) 105513f96637SDennis Zhou (Facebook) /* 105613f96637SDennis Zhou (Facebook) * Check to see if the allocation can fit in the chunk's contig hint. 105713f96637SDennis Zhou (Facebook) * This is an optimization to prevent scanning by assuming if it 105813f96637SDennis Zhou (Facebook) * cannot fit in the global hint, there is memory pressure and creating 105913f96637SDennis Zhou (Facebook) * a new chunk would happen soon. 106013f96637SDennis Zhou (Facebook) */ 106192c14cabSDennis Zhou bit_off = ALIGN(chunk_md->contig_hint_start, align) - 106292c14cabSDennis Zhou chunk_md->contig_hint_start; 106392c14cabSDennis Zhou if (bit_off + alloc_bits > chunk_md->contig_hint) 106413f96637SDennis Zhou (Facebook) return -1; 106513f96637SDennis Zhou (Facebook) 1066d33d9f3dSDennis Zhou bit_off = pcpu_next_hint(chunk_md, alloc_bits); 1067b4c2116cSDennis Zhou (Facebook) bits = 0; 1068b4c2116cSDennis Zhou (Facebook) pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { 106940064aecSDennis Zhou (Facebook) if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 1070b4c2116cSDennis Zhou (Facebook) &next_off)) 107140064aecSDennis Zhou (Facebook) break; 107240064aecSDennis Zhou (Facebook) 1073b4c2116cSDennis Zhou (Facebook) bit_off = next_off; 107440064aecSDennis Zhou (Facebook) bits = 0; 107540064aecSDennis Zhou (Facebook) } 107640064aecSDennis Zhou (Facebook) 107740064aecSDennis Zhou (Facebook) if (bit_off == pcpu_chunk_map_bits(chunk)) 107840064aecSDennis Zhou (Facebook) return -1; 107940064aecSDennis Zhou (Facebook) 108040064aecSDennis Zhou (Facebook) return bit_off; 108140064aecSDennis Zhou (Facebook) } 108240064aecSDennis Zhou (Facebook) 1083b89462a9SDennis Zhou /* 1084b89462a9SDennis Zhou * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off() 1085b89462a9SDennis Zhou * @map: the address to base the search on 1086b89462a9SDennis Zhou * @size: the bitmap size in bits 1087b89462a9SDennis Zhou * @start: the bitnumber to start searching at 1088b89462a9SDennis Zhou * @nr: the number of zeroed bits we're looking for 1089b89462a9SDennis Zhou * @align_mask: alignment mask for zero area 1090b89462a9SDennis Zhou * @largest_off: offset of the largest area skipped 1091b89462a9SDennis Zhou * @largest_bits: size of the largest area skipped 1092b89462a9SDennis Zhou * 1093b89462a9SDennis Zhou * The @align_mask should be one less than a power of 2. 1094b89462a9SDennis Zhou * 1095b89462a9SDennis Zhou * This is a modified version of bitmap_find_next_zero_area_off() to remember 1096b89462a9SDennis Zhou * the largest area that was skipped. This is imperfect, but in general is 1097b89462a9SDennis Zhou * good enough. The largest remembered region is the largest failed region 1098b89462a9SDennis Zhou * seen. This does not include anything we possibly skipped due to alignment. 1099b89462a9SDennis Zhou * pcpu_block_update_scan() does scan backwards to try and recover what was 1100b89462a9SDennis Zhou * lost to alignment. While this can cause scanning to miss earlier possible 1101b89462a9SDennis Zhou * free areas, smaller allocations will eventually fill those holes. 1102b89462a9SDennis Zhou */ 1103b89462a9SDennis Zhou static unsigned long pcpu_find_zero_area(unsigned long *map, 1104b89462a9SDennis Zhou unsigned long size, 1105b89462a9SDennis Zhou unsigned long start, 1106b89462a9SDennis Zhou unsigned long nr, 1107b89462a9SDennis Zhou unsigned long align_mask, 1108b89462a9SDennis Zhou unsigned long *largest_off, 1109b89462a9SDennis Zhou unsigned long *largest_bits) 1110b89462a9SDennis Zhou { 1111b89462a9SDennis Zhou unsigned long index, end, i, area_off, area_bits; 1112b89462a9SDennis Zhou again: 1113b89462a9SDennis Zhou index = find_next_zero_bit(map, size, start); 1114b89462a9SDennis Zhou 1115b89462a9SDennis Zhou /* Align allocation */ 1116b89462a9SDennis Zhou index = __ALIGN_MASK(index, align_mask); 1117b89462a9SDennis Zhou area_off = index; 1118b89462a9SDennis Zhou 1119b89462a9SDennis Zhou end = index + nr; 1120b89462a9SDennis Zhou if (end > size) 1121b89462a9SDennis Zhou return end; 1122b89462a9SDennis Zhou i = find_next_bit(map, end, index); 1123b89462a9SDennis Zhou if (i < end) { 1124b89462a9SDennis Zhou area_bits = i - area_off; 1125b89462a9SDennis Zhou /* remember largest unused area with best alignment */ 1126b89462a9SDennis Zhou if (area_bits > *largest_bits || 1127b89462a9SDennis Zhou (area_bits == *largest_bits && *largest_off && 1128b89462a9SDennis Zhou (!area_off || __ffs(area_off) > __ffs(*largest_off)))) { 1129b89462a9SDennis Zhou *largest_off = area_off; 1130b89462a9SDennis Zhou *largest_bits = area_bits; 1131b89462a9SDennis Zhou } 1132b89462a9SDennis Zhou 1133b89462a9SDennis Zhou start = i + 1; 1134b89462a9SDennis Zhou goto again; 1135b89462a9SDennis Zhou } 1136b89462a9SDennis Zhou return index; 1137b89462a9SDennis Zhou } 1138b89462a9SDennis Zhou 113940064aecSDennis Zhou (Facebook) /** 114040064aecSDennis Zhou (Facebook) * pcpu_alloc_area - allocates an area from a pcpu_chunk 114140064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 114240064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 114340064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 114440064aecSDennis Zhou (Facebook) * @start: bit_off to start searching 114540064aecSDennis Zhou (Facebook) * 114640064aecSDennis Zhou (Facebook) * This function takes in a @start offset to begin searching to fit an 1147b4c2116cSDennis Zhou (Facebook) * allocation of @alloc_bits with alignment @align. It needs to scan 1148b4c2116cSDennis Zhou (Facebook) * the allocation map because if it fits within the block's contig hint, 1149b4c2116cSDennis Zhou (Facebook) * @start will be block->first_free. This is an attempt to fill the 1150b4c2116cSDennis Zhou (Facebook) * allocation prior to breaking the contig hint. The allocation and 1151b4c2116cSDennis Zhou (Facebook) * boundary maps are updated accordingly if it confirms a valid 1152b4c2116cSDennis Zhou (Facebook) * free area. 115340064aecSDennis Zhou (Facebook) * 115440064aecSDennis Zhou (Facebook) * RETURNS: 115540064aecSDennis Zhou (Facebook) * Allocated addr offset in @chunk on success. 115640064aecSDennis Zhou (Facebook) * -1 if no matching area is found. 115740064aecSDennis Zhou (Facebook) */ 115840064aecSDennis Zhou (Facebook) static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 115940064aecSDennis Zhou (Facebook) size_t align, int start) 116040064aecSDennis Zhou (Facebook) { 116192c14cabSDennis Zhou struct pcpu_block_md *chunk_md = &chunk->chunk_md; 116240064aecSDennis Zhou (Facebook) size_t align_mask = (align) ? (align - 1) : 0; 1163b89462a9SDennis Zhou unsigned long area_off = 0, area_bits = 0; 116440064aecSDennis Zhou (Facebook) int bit_off, end, oslot; 11659f7dcf22STejun Heo 11664f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 11674f996e23STejun Heo 116840064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1169833af842STejun Heo 1170833af842STejun Heo /* 117140064aecSDennis Zhou (Facebook) * Search to find a fit. 1172833af842STejun Heo */ 11738c43004aSDennis Zhou end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS, 11748c43004aSDennis Zhou pcpu_chunk_map_bits(chunk)); 1175b89462a9SDennis Zhou bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits, 1176b89462a9SDennis Zhou align_mask, &area_off, &area_bits); 117740064aecSDennis Zhou (Facebook) if (bit_off >= end) 1178a16037c8STejun Heo return -1; 1179a16037c8STejun Heo 1180b89462a9SDennis Zhou if (area_bits) 1181b89462a9SDennis Zhou pcpu_block_update_scan(chunk, area_off, area_bits); 1182b89462a9SDennis Zhou 118340064aecSDennis Zhou (Facebook) /* update alloc map */ 118440064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 1185a16037c8STejun Heo 118640064aecSDennis Zhou (Facebook) /* update boundary map */ 118740064aecSDennis Zhou (Facebook) set_bit(bit_off, chunk->bound_map); 118840064aecSDennis Zhou (Facebook) bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 118940064aecSDennis Zhou (Facebook) set_bit(bit_off + alloc_bits, chunk->bound_map); 1190a16037c8STejun Heo 119140064aecSDennis Zhou (Facebook) chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 119240064aecSDennis Zhou (Facebook) 119386b442fbSDennis Zhou (Facebook) /* update first free bit */ 119492c14cabSDennis Zhou if (bit_off == chunk_md->first_free) 119592c14cabSDennis Zhou chunk_md->first_free = find_next_zero_bit( 119686b442fbSDennis Zhou (Facebook) chunk->alloc_map, 119786b442fbSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk), 119886b442fbSDennis Zhou (Facebook) bit_off + alloc_bits); 119986b442fbSDennis Zhou (Facebook) 1200ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 120140064aecSDennis Zhou (Facebook) 120240064aecSDennis Zhou (Facebook) pcpu_chunk_relocate(chunk, oslot); 120340064aecSDennis Zhou (Facebook) 120440064aecSDennis Zhou (Facebook) return bit_off * PCPU_MIN_ALLOC_SIZE; 1205a16037c8STejun Heo } 1206a16037c8STejun Heo 1207a16037c8STejun Heo /** 120840064aecSDennis Zhou (Facebook) * pcpu_free_area - frees the corresponding offset 1209fbf59bc9STejun Heo * @chunk: chunk of interest 121040064aecSDennis Zhou (Facebook) * @off: addr offset into chunk 1211fbf59bc9STejun Heo * 121240064aecSDennis Zhou (Facebook) * This function determines the size of an allocation to free using 121340064aecSDennis Zhou (Facebook) * the boundary bitmap and clears the allocation map. 1214*5b32af91SRoman Gushchin * 1215*5b32af91SRoman Gushchin * RETURNS: 1216*5b32af91SRoman Gushchin * Number of freed bytes. 1217fbf59bc9STejun Heo */ 1218*5b32af91SRoman Gushchin static int pcpu_free_area(struct pcpu_chunk *chunk, int off) 1219fbf59bc9STejun Heo { 122092c14cabSDennis Zhou struct pcpu_block_md *chunk_md = &chunk->chunk_md; 1221*5b32af91SRoman Gushchin int bit_off, bits, end, oslot, freed; 1222fbf59bc9STejun Heo 12235ccd30e4SDennis Zhou lockdep_assert_held(&pcpu_lock); 122430a5b536SDennis Zhou pcpu_stats_area_dealloc(chunk); 12255ccd30e4SDennis Zhou 122640064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 1227723ad1d9SAl Viro 122840064aecSDennis Zhou (Facebook) bit_off = off / PCPU_MIN_ALLOC_SIZE; 1229fbf59bc9STejun Heo 123040064aecSDennis Zhou (Facebook) /* find end index */ 123140064aecSDennis Zhou (Facebook) end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 123240064aecSDennis Zhou (Facebook) bit_off + 1); 123340064aecSDennis Zhou (Facebook) bits = end - bit_off; 123440064aecSDennis Zhou (Facebook) bitmap_clear(chunk->alloc_map, bit_off, bits); 12353d331ad7SAl Viro 1236*5b32af91SRoman Gushchin freed = bits * PCPU_MIN_ALLOC_SIZE; 1237*5b32af91SRoman Gushchin 123840064aecSDennis Zhou (Facebook) /* update metadata */ 1239*5b32af91SRoman Gushchin chunk->free_bytes += freed; 1240fbf59bc9STejun Heo 124186b442fbSDennis Zhou (Facebook) /* update first free bit */ 124292c14cabSDennis Zhou chunk_md->first_free = min(chunk_md->first_free, bit_off); 124386b442fbSDennis Zhou (Facebook) 1244ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_free(chunk, bit_off, bits); 1245b539b87fSTejun Heo 1246fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 1247*5b32af91SRoman Gushchin 1248*5b32af91SRoman Gushchin return freed; 1249fbf59bc9STejun Heo } 1250fbf59bc9STejun Heo 1251047924c9SDennis Zhou static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits) 1252047924c9SDennis Zhou { 1253047924c9SDennis Zhou block->scan_hint = 0; 1254047924c9SDennis Zhou block->contig_hint = nr_bits; 1255047924c9SDennis Zhou block->left_free = nr_bits; 1256047924c9SDennis Zhou block->right_free = nr_bits; 1257047924c9SDennis Zhou block->first_free = 0; 1258047924c9SDennis Zhou block->nr_bits = nr_bits; 1259047924c9SDennis Zhou } 1260047924c9SDennis Zhou 1261ca460b3cSDennis Zhou (Facebook) static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 1262ca460b3cSDennis Zhou (Facebook) { 1263ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *md_block; 1264ca460b3cSDennis Zhou (Facebook) 126592c14cabSDennis Zhou /* init the chunk's block */ 126692c14cabSDennis Zhou pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk)); 126792c14cabSDennis Zhou 1268ca460b3cSDennis Zhou (Facebook) for (md_block = chunk->md_blocks; 1269ca460b3cSDennis Zhou (Facebook) md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 1270047924c9SDennis Zhou md_block++) 1271047924c9SDennis Zhou pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS); 1272ca460b3cSDennis Zhou (Facebook) } 1273ca460b3cSDennis Zhou (Facebook) 127440064aecSDennis Zhou (Facebook) /** 127540064aecSDennis Zhou (Facebook) * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 127640064aecSDennis Zhou (Facebook) * @tmp_addr: the start of the region served 127740064aecSDennis Zhou (Facebook) * @map_size: size of the region served 127840064aecSDennis Zhou (Facebook) * 127940064aecSDennis Zhou (Facebook) * This is responsible for creating the chunks that serve the first chunk. The 128040064aecSDennis Zhou (Facebook) * base_addr is page aligned down of @tmp_addr while the region end is page 128140064aecSDennis Zhou (Facebook) * aligned up. Offsets are kept track of to determine the region served. All 128240064aecSDennis Zhou (Facebook) * this is done to appease the bitmap allocator in avoiding partial blocks. 128340064aecSDennis Zhou (Facebook) * 128440064aecSDennis Zhou (Facebook) * RETURNS: 128540064aecSDennis Zhou (Facebook) * Chunk serving the region at @tmp_addr of @map_size. 128640064aecSDennis Zhou (Facebook) */ 1287c0ebfdc3SDennis Zhou (Facebook) static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 128840064aecSDennis Zhou (Facebook) int map_size) 128910edf5b0SDennis Zhou (Facebook) { 129010edf5b0SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 1291ca460b3cSDennis Zhou (Facebook) unsigned long aligned_addr, lcm_align; 129240064aecSDennis Zhou (Facebook) int start_offset, offset_bits, region_size, region_bits; 1293f655f405SMike Rapoport size_t alloc_size; 1294c0ebfdc3SDennis Zhou (Facebook) 1295c0ebfdc3SDennis Zhou (Facebook) /* region calculations */ 1296c0ebfdc3SDennis Zhou (Facebook) aligned_addr = tmp_addr & PAGE_MASK; 1297c0ebfdc3SDennis Zhou (Facebook) 1298c0ebfdc3SDennis Zhou (Facebook) start_offset = tmp_addr - aligned_addr; 12996b9d7c8eSDennis Zhou (Facebook) 1300ca460b3cSDennis Zhou (Facebook) /* 1301ca460b3cSDennis Zhou (Facebook) * Align the end of the region with the LCM of PAGE_SIZE and 1302ca460b3cSDennis Zhou (Facebook) * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of 1303ca460b3cSDennis Zhou (Facebook) * the other. 1304ca460b3cSDennis Zhou (Facebook) */ 1305ca460b3cSDennis Zhou (Facebook) lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); 1306ca460b3cSDennis Zhou (Facebook) region_size = ALIGN(start_offset + map_size, lcm_align); 130710edf5b0SDennis Zhou (Facebook) 1308c0ebfdc3SDennis Zhou (Facebook) /* allocate chunk */ 1309f655f405SMike Rapoport alloc_size = sizeof(struct pcpu_chunk) + 1310f655f405SMike Rapoport BITS_TO_LONGS(region_size >> PAGE_SHIFT); 1311f655f405SMike Rapoport chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1312f655f405SMike Rapoport if (!chunk) 1313f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1314f655f405SMike Rapoport alloc_size); 1315c0ebfdc3SDennis Zhou (Facebook) 131610edf5b0SDennis Zhou (Facebook) INIT_LIST_HEAD(&chunk->list); 1317c0ebfdc3SDennis Zhou (Facebook) 1318c0ebfdc3SDennis Zhou (Facebook) chunk->base_addr = (void *)aligned_addr; 131910edf5b0SDennis Zhou (Facebook) chunk->start_offset = start_offset; 13206b9d7c8eSDennis Zhou (Facebook) chunk->end_offset = region_size - chunk->start_offset - map_size; 1321c0ebfdc3SDennis Zhou (Facebook) 13228ab16c43SDennis Zhou (Facebook) chunk->nr_pages = region_size >> PAGE_SHIFT; 132340064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 1324c0ebfdc3SDennis Zhou (Facebook) 1325f655f405SMike Rapoport alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]); 1326f655f405SMike Rapoport chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1327f655f405SMike Rapoport if (!chunk->alloc_map) 1328f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1329f655f405SMike Rapoport alloc_size); 1330f655f405SMike Rapoport 1331f655f405SMike Rapoport alloc_size = 1332f655f405SMike Rapoport BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]); 1333f655f405SMike Rapoport chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1334f655f405SMike Rapoport if (!chunk->bound_map) 1335f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1336f655f405SMike Rapoport alloc_size); 1337f655f405SMike Rapoport 1338f655f405SMike Rapoport alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]); 1339f655f405SMike Rapoport chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 1340f655f405SMike Rapoport if (!chunk->md_blocks) 1341f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 1342f655f405SMike Rapoport alloc_size); 1343f655f405SMike Rapoport 1344ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 134510edf5b0SDennis Zhou (Facebook) 134610edf5b0SDennis Zhou (Facebook) /* manage populated page bitmap */ 134710edf5b0SDennis Zhou (Facebook) chunk->immutable = true; 13488ab16c43SDennis Zhou (Facebook) bitmap_fill(chunk->populated, chunk->nr_pages); 13498ab16c43SDennis Zhou (Facebook) chunk->nr_populated = chunk->nr_pages; 1350b239f7daSDennis Zhou chunk->nr_empty_pop_pages = chunk->nr_pages; 135110edf5b0SDennis Zhou (Facebook) 135240064aecSDennis Zhou (Facebook) chunk->free_bytes = map_size; 1353c0ebfdc3SDennis Zhou (Facebook) 1354c0ebfdc3SDennis Zhou (Facebook) if (chunk->start_offset) { 1355c0ebfdc3SDennis Zhou (Facebook) /* hide the beginning of the bitmap */ 135640064aecSDennis Zhou (Facebook) offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 135740064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 0, offset_bits); 135840064aecSDennis Zhou (Facebook) set_bit(0, chunk->bound_map); 135940064aecSDennis Zhou (Facebook) set_bit(offset_bits, chunk->bound_map); 1360ca460b3cSDennis Zhou (Facebook) 136192c14cabSDennis Zhou chunk->chunk_md.first_free = offset_bits; 136286b442fbSDennis Zhou (Facebook) 1363ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 1364c0ebfdc3SDennis Zhou (Facebook) } 1365c0ebfdc3SDennis Zhou (Facebook) 13666b9d7c8eSDennis Zhou (Facebook) if (chunk->end_offset) { 13676b9d7c8eSDennis Zhou (Facebook) /* hide the end of the bitmap */ 136840064aecSDennis Zhou (Facebook) offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 136940064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 137040064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk) - offset_bits, 137140064aecSDennis Zhou (Facebook) offset_bits); 137240064aecSDennis Zhou (Facebook) set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 137340064aecSDennis Zhou (Facebook) chunk->bound_map); 137440064aecSDennis Zhou (Facebook) set_bit(region_bits, chunk->bound_map); 13756b9d7c8eSDennis Zhou (Facebook) 1376ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 1377ca460b3cSDennis Zhou (Facebook) - offset_bits, offset_bits); 1378ca460b3cSDennis Zhou (Facebook) } 137940064aecSDennis Zhou (Facebook) 138010edf5b0SDennis Zhou (Facebook) return chunk; 138110edf5b0SDennis Zhou (Facebook) } 138210edf5b0SDennis Zhou (Facebook) 138347504ee0SDennis Zhou static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) 13846081089fSTejun Heo { 13856081089fSTejun Heo struct pcpu_chunk *chunk; 138640064aecSDennis Zhou (Facebook) int region_bits; 13876081089fSTejun Heo 138847504ee0SDennis Zhou chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); 13896081089fSTejun Heo if (!chunk) 13906081089fSTejun Heo return NULL; 13916081089fSTejun Heo 13926081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 1393c0ebfdc3SDennis Zhou (Facebook) chunk->nr_pages = pcpu_unit_pages; 139440064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 139540064aecSDennis Zhou (Facebook) 139640064aecSDennis Zhou (Facebook) chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 139747504ee0SDennis Zhou sizeof(chunk->alloc_map[0]), gfp); 139840064aecSDennis Zhou (Facebook) if (!chunk->alloc_map) 139940064aecSDennis Zhou (Facebook) goto alloc_map_fail; 140040064aecSDennis Zhou (Facebook) 140140064aecSDennis Zhou (Facebook) chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 140247504ee0SDennis Zhou sizeof(chunk->bound_map[0]), gfp); 140340064aecSDennis Zhou (Facebook) if (!chunk->bound_map) 140440064aecSDennis Zhou (Facebook) goto bound_map_fail; 140540064aecSDennis Zhou (Facebook) 1406ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 140747504ee0SDennis Zhou sizeof(chunk->md_blocks[0]), gfp); 1408ca460b3cSDennis Zhou (Facebook) if (!chunk->md_blocks) 1409ca460b3cSDennis Zhou (Facebook) goto md_blocks_fail; 1410ca460b3cSDennis Zhou (Facebook) 1411ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 1412ca460b3cSDennis Zhou (Facebook) 141340064aecSDennis Zhou (Facebook) /* init metadata */ 141440064aecSDennis Zhou (Facebook) chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 1415c0ebfdc3SDennis Zhou (Facebook) 14166081089fSTejun Heo return chunk; 141740064aecSDennis Zhou (Facebook) 1418ca460b3cSDennis Zhou (Facebook) md_blocks_fail: 1419ca460b3cSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 142040064aecSDennis Zhou (Facebook) bound_map_fail: 142140064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 142240064aecSDennis Zhou (Facebook) alloc_map_fail: 142340064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk); 142440064aecSDennis Zhou (Facebook) 142540064aecSDennis Zhou (Facebook) return NULL; 14266081089fSTejun Heo } 14276081089fSTejun Heo 14286081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 14296081089fSTejun Heo { 14306081089fSTejun Heo if (!chunk) 14316081089fSTejun Heo return; 14326685b357SMike Rapoport pcpu_mem_free(chunk->md_blocks); 143340064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 143440064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 14351d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 14366081089fSTejun Heo } 14376081089fSTejun Heo 1438b539b87fSTejun Heo /** 1439b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 1440b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 1441b539b87fSTejun Heo * @page_start: the start page 1442b539b87fSTejun Heo * @page_end: the end page 1443b539b87fSTejun Heo * 1444b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1445b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 1446b539b87fSTejun Heo * successful population. 144740064aecSDennis Zhou (Facebook) * 144840064aecSDennis Zhou (Facebook) * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it 144940064aecSDennis Zhou (Facebook) * is to serve an allocation in that area. 1450b539b87fSTejun Heo */ 145140064aecSDennis Zhou (Facebook) static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 1452b239f7daSDennis Zhou int page_end) 1453b539b87fSTejun Heo { 1454b539b87fSTejun Heo int nr = page_end - page_start; 1455b539b87fSTejun Heo 1456b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1457b539b87fSTejun Heo 1458b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 1459b539b87fSTejun Heo chunk->nr_populated += nr; 14607e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += nr; 146140064aecSDennis Zhou (Facebook) 1462b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, nr); 146340064aecSDennis Zhou (Facebook) } 1464b539b87fSTejun Heo 1465b539b87fSTejun Heo /** 1466b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 1467b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 1468b539b87fSTejun Heo * @page_start: the start page 1469b539b87fSTejun Heo * @page_end: the end page 1470b539b87fSTejun Heo * 1471b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1472b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 1473b539b87fSTejun Heo * each successful depopulation. 1474b539b87fSTejun Heo */ 1475b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1476b539b87fSTejun Heo int page_start, int page_end) 1477b539b87fSTejun Heo { 1478b539b87fSTejun Heo int nr = page_end - page_start; 1479b539b87fSTejun Heo 1480b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1481b539b87fSTejun Heo 1482b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 1483b539b87fSTejun Heo chunk->nr_populated -= nr; 14847e8a6304SDennis Zhou (Facebook) pcpu_nr_populated -= nr; 1485b239f7daSDennis Zhou 1486b239f7daSDennis Zhou pcpu_update_empty_pages(chunk, -nr); 1487b539b87fSTejun Heo } 1488b539b87fSTejun Heo 1489fbf59bc9STejun Heo /* 14909f645532STejun Heo * Chunk management implementation. 1491fbf59bc9STejun Heo * 14929f645532STejun Heo * To allow different implementations, chunk alloc/free and 14939f645532STejun Heo * [de]population are implemented in a separate file which is pulled 14949f645532STejun Heo * into this file and compiled together. The following functions 14959f645532STejun Heo * should be implemented. 1496ccea34b5STejun Heo * 14979f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 14989f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 14999f645532STejun Heo * pcpu_create_chunk - create a new chunk 15009f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 15019f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 15029f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1503fbf59bc9STejun Heo */ 150415d9f3d1SDennis Zhou static int pcpu_populate_chunk(struct pcpu_chunk *chunk, 150547504ee0SDennis Zhou int page_start, int page_end, gfp_t gfp); 150615d9f3d1SDennis Zhou static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, 150715d9f3d1SDennis Zhou int page_start, int page_end); 150847504ee0SDennis Zhou static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); 15099f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 15109f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 15119f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1512fbf59bc9STejun Heo 1513b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 1514b0c9778bSTejun Heo #include "percpu-km.c" 1515b0c9778bSTejun Heo #else 15169f645532STejun Heo #include "percpu-vm.c" 1517b0c9778bSTejun Heo #endif 1518fbf59bc9STejun Heo 1519fbf59bc9STejun Heo /** 152088999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 152188999a89STejun Heo * @addr: address for which the chunk needs to be determined. 152288999a89STejun Heo * 1523c0ebfdc3SDennis Zhou (Facebook) * This is an internal function that handles all but static allocations. 1524c0ebfdc3SDennis Zhou (Facebook) * Static percpu address values should never be passed into the allocator. 1525c0ebfdc3SDennis Zhou (Facebook) * 152688999a89STejun Heo * RETURNS: 152788999a89STejun Heo * The address of the found chunk. 152888999a89STejun Heo */ 152988999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 153088999a89STejun Heo { 1531c0ebfdc3SDennis Zhou (Facebook) /* is it in the dynamic region (first chunk)? */ 1532560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1533c0ebfdc3SDennis Zhou (Facebook) return pcpu_first_chunk; 1534c0ebfdc3SDennis Zhou (Facebook) 1535c0ebfdc3SDennis Zhou (Facebook) /* is it in the reserved region? */ 1536560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 153788999a89STejun Heo return pcpu_reserved_chunk; 153888999a89STejun Heo 153988999a89STejun Heo /* 154088999a89STejun Heo * The address is relative to unit0 which might be unused and 154188999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 154288999a89STejun Heo * current processor before looking it up in the vmalloc 154388999a89STejun Heo * space. Note that any possible cpu id can be used here, so 154488999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 154588999a89STejun Heo */ 154688999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 15479f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 154888999a89STejun Heo } 154988999a89STejun Heo 155088999a89STejun Heo /** 1551edcb4639STejun Heo * pcpu_alloc - the percpu allocator 1552cae3aeb8STejun Heo * @size: size of area to allocate in bytes 1553fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 1554edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 15555835d96eSTejun Heo * @gfp: allocation flags 1556fbf59bc9STejun Heo * 15575835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 15580ea7eeecSDaniel Borkmann * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN 15590ea7eeecSDaniel Borkmann * then no warning will be triggered on invalid or failed allocation 15600ea7eeecSDaniel Borkmann * requests. 1561fbf59bc9STejun Heo * 1562fbf59bc9STejun Heo * RETURNS: 1563fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1564fbf59bc9STejun Heo */ 15655835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 15665835d96eSTejun Heo gfp_t gfp) 1567fbf59bc9STejun Heo { 156828307d93SFilipe Manana gfp_t pcpu_gfp; 156928307d93SFilipe Manana bool is_atomic; 157028307d93SFilipe Manana bool do_warn; 1571f2badb0cSTejun Heo static int warn_limit = 10; 15728744d859SDennis Zhou struct pcpu_chunk *chunk, *next; 1573f2badb0cSTejun Heo const char *err; 157440064aecSDennis Zhou (Facebook) int slot, off, cpu, ret; 1575403a91b1SJiri Kosina unsigned long flags; 1576f528f0b8SCatalin Marinas void __percpu *ptr; 157740064aecSDennis Zhou (Facebook) size_t bits, bit_align; 1578fbf59bc9STejun Heo 157928307d93SFilipe Manana gfp = current_gfp_context(gfp); 158028307d93SFilipe Manana /* whitelisted flags that can be passed to the backing allocators */ 158128307d93SFilipe Manana pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN); 158228307d93SFilipe Manana is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 158328307d93SFilipe Manana do_warn = !(gfp & __GFP_NOWARN); 158428307d93SFilipe Manana 1585723ad1d9SAl Viro /* 158640064aecSDennis Zhou (Facebook) * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 158740064aecSDennis Zhou (Facebook) * therefore alignment must be a minimum of that many bytes. 158840064aecSDennis Zhou (Facebook) * An allocation may have internal fragmentation from rounding up 158940064aecSDennis Zhou (Facebook) * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1590723ad1d9SAl Viro */ 1591d2f3c384SDennis Zhou (Facebook) if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1592d2f3c384SDennis Zhou (Facebook) align = PCPU_MIN_ALLOC_SIZE; 1593723ad1d9SAl Viro 1594d2f3c384SDennis Zhou (Facebook) size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 159540064aecSDennis Zhou (Facebook) bits = size >> PCPU_MIN_ALLOC_SHIFT; 159640064aecSDennis Zhou (Facebook) bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 15972f69fa82SViro 15983ca45a46Szijun_hu if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 15993ca45a46Szijun_hu !is_power_of_2(align))) { 16000ea7eeecSDaniel Borkmann WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1601756a025fSJoe Perches size, align); 1602fbf59bc9STejun Heo return NULL; 1603fbf59bc9STejun Heo } 1604fbf59bc9STejun Heo 1605f52ba1feSKirill Tkhai if (!is_atomic) { 1606f52ba1feSKirill Tkhai /* 1607f52ba1feSKirill Tkhai * pcpu_balance_workfn() allocates memory under this mutex, 1608f52ba1feSKirill Tkhai * and it may wait for memory reclaim. Allow current task 1609f52ba1feSKirill Tkhai * to become OOM victim, in case of memory pressure. 1610f52ba1feSKirill Tkhai */ 1611f52ba1feSKirill Tkhai if (gfp & __GFP_NOFAIL) 16126710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 1613f52ba1feSKirill Tkhai else if (mutex_lock_killable(&pcpu_alloc_mutex)) 1614f52ba1feSKirill Tkhai return NULL; 1615f52ba1feSKirill Tkhai } 16166710e594STejun Heo 1617403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1618fbf59bc9STejun Heo 1619edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 1620edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 1621edcb4639STejun Heo chunk = pcpu_reserved_chunk; 1622833af842STejun Heo 162340064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 162440064aecSDennis Zhou (Facebook) if (off < 0) { 1625833af842STejun Heo err = "alloc from reserved chunk failed"; 1626ccea34b5STejun Heo goto fail_unlock; 1627f2badb0cSTejun Heo } 1628833af842STejun Heo 162940064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1630edcb4639STejun Heo if (off >= 0) 1631edcb4639STejun Heo goto area_found; 1632833af842STejun Heo 1633f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 1634ccea34b5STejun Heo goto fail_unlock; 1635edcb4639STejun Heo } 1636edcb4639STejun Heo 1637ccea34b5STejun Heo restart: 1638edcb4639STejun Heo /* search through normal chunks */ 1639fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 16408744d859SDennis Zhou list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) { 164140064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, 164240064aecSDennis Zhou (Facebook) is_atomic); 16438744d859SDennis Zhou if (off < 0) { 16448744d859SDennis Zhou if (slot < PCPU_SLOT_FAIL_THRESHOLD) 16458744d859SDennis Zhou pcpu_chunk_move(chunk, 0); 1646fbf59bc9STejun Heo continue; 16478744d859SDennis Zhou } 1648ccea34b5STejun Heo 164940064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1650fbf59bc9STejun Heo if (off >= 0) 1651fbf59bc9STejun Heo goto area_found; 165240064aecSDennis Zhou (Facebook) 1653fbf59bc9STejun Heo } 1654fbf59bc9STejun Heo } 1655fbf59bc9STejun Heo 1656403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1657ccea34b5STejun Heo 1658b38d08f3STejun Heo /* 1659b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 1660b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 1661b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 1662b38d08f3STejun Heo */ 166311df02bfSDennis Zhou if (is_atomic) { 166411df02bfSDennis Zhou err = "atomic alloc failed, no space left"; 16655835d96eSTejun Heo goto fail; 166611df02bfSDennis Zhou } 16675835d96eSTejun Heo 1668b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 1669554fef1cSDennis Zhou chunk = pcpu_create_chunk(pcpu_gfp); 1670f2badb0cSTejun Heo if (!chunk) { 1671f2badb0cSTejun Heo err = "failed to allocate new chunk"; 1672b38d08f3STejun Heo goto fail; 1673f2badb0cSTejun Heo } 1674ccea34b5STejun Heo 1675403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1676fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 1677b38d08f3STejun Heo } else { 1678b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1679b38d08f3STejun Heo } 1680b38d08f3STejun Heo 1681ccea34b5STejun Heo goto restart; 1682fbf59bc9STejun Heo 1683fbf59bc9STejun Heo area_found: 168430a5b536SDennis Zhou pcpu_stats_area_alloc(chunk, size); 1685403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1686ccea34b5STejun Heo 1687dca49645STejun Heo /* populate if not all pages are already there */ 16885835d96eSTejun Heo if (!is_atomic) { 1689e837dfdeSDennis Zhou unsigned int page_start, page_end, rs, re; 1690e04d3208STejun Heo 1691dca49645STejun Heo page_start = PFN_DOWN(off); 1692dca49645STejun Heo page_end = PFN_UP(off + size); 1693dca49645STejun Heo 1694e837dfdeSDennis Zhou bitmap_for_each_clear_region(chunk->populated, rs, re, 169591e914c5SDennis Zhou (Facebook) page_start, page_end) { 1696dca49645STejun Heo WARN_ON(chunk->immutable); 1697dca49645STejun Heo 1698554fef1cSDennis Zhou ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp); 1699b38d08f3STejun Heo 1700403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1701b38d08f3STejun Heo if (ret) { 170240064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1703f2badb0cSTejun Heo err = "failed to populate"; 1704ccea34b5STejun Heo goto fail_unlock; 1705fbf59bc9STejun Heo } 1706b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, re); 1707b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1708dca49645STejun Heo } 1709dca49645STejun Heo 1710ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1711e04d3208STejun Heo } 1712ccea34b5STejun Heo 17131a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 17141a4d7607STejun Heo pcpu_schedule_balance_work(); 17151a4d7607STejun Heo 1716dca49645STejun Heo /* clear the areas and return address relative to base address */ 1717dca49645STejun Heo for_each_possible_cpu(cpu) 1718dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1719dca49645STejun Heo 1720f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 17218a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1722df95e795SDennis Zhou 1723df95e795SDennis Zhou trace_percpu_alloc_percpu(reserved, is_atomic, size, align, 1724df95e795SDennis Zhou chunk->base_addr, off, ptr); 1725df95e795SDennis Zhou 1726f528f0b8SCatalin Marinas return ptr; 1727ccea34b5STejun Heo 1728ccea34b5STejun Heo fail_unlock: 1729403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1730b38d08f3STejun Heo fail: 1731df95e795SDennis Zhou trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1732df95e795SDennis Zhou 17330ea7eeecSDaniel Borkmann if (!is_atomic && do_warn && warn_limit) { 1734870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 17355835d96eSTejun Heo size, align, is_atomic, err); 1736f2badb0cSTejun Heo dump_stack(); 1737f2badb0cSTejun Heo if (!--warn_limit) 1738870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1739f2badb0cSTejun Heo } 17401a4d7607STejun Heo if (is_atomic) { 17411a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 17421a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 17431a4d7607STejun Heo pcpu_schedule_balance_work(); 17446710e594STejun Heo } else { 17456710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 17461a4d7607STejun Heo } 1747ccea34b5STejun Heo return NULL; 1748fbf59bc9STejun Heo } 1749edcb4639STejun Heo 1750edcb4639STejun Heo /** 17515835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1752edcb4639STejun Heo * @size: size of area to allocate in bytes 1753edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 17545835d96eSTejun Heo * @gfp: allocation flags 1755edcb4639STejun Heo * 17565835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 17575835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 17580ea7eeecSDaniel Borkmann * be called from any context but is a lot more likely to fail. If @gfp 17590ea7eeecSDaniel Borkmann * has __GFP_NOWARN then no warning will be triggered on invalid or failed 17600ea7eeecSDaniel Borkmann * allocation requests. 1761ccea34b5STejun Heo * 1762edcb4639STejun Heo * RETURNS: 1763edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1764edcb4639STejun Heo */ 17655835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 17665835d96eSTejun Heo { 17675835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 17685835d96eSTejun Heo } 17695835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 17705835d96eSTejun Heo 17715835d96eSTejun Heo /** 17725835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 17735835d96eSTejun Heo * @size: size of area to allocate in bytes 17745835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 17755835d96eSTejun Heo * 17765835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 17775835d96eSTejun Heo */ 177843cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1779edcb4639STejun Heo { 17805835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1781edcb4639STejun Heo } 1782fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1783fbf59bc9STejun Heo 1784edcb4639STejun Heo /** 1785edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1786edcb4639STejun Heo * @size: size of area to allocate in bytes 1787edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1788edcb4639STejun Heo * 17899329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 17909329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 17919329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 17929329ba97STejun Heo * Might trigger writeouts. 1793edcb4639STejun Heo * 1794ccea34b5STejun Heo * CONTEXT: 1795ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1796ccea34b5STejun Heo * 1797edcb4639STejun Heo * RETURNS: 1798edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1799edcb4639STejun Heo */ 180043cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1801edcb4639STejun Heo { 18025835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1803edcb4639STejun Heo } 1804edcb4639STejun Heo 1805a56dbddfSTejun Heo /** 18061a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1807a56dbddfSTejun Heo * @work: unused 1808a56dbddfSTejun Heo * 180947504ee0SDennis Zhou * Reclaim all fully free chunks except for the first one. This is also 181047504ee0SDennis Zhou * responsible for maintaining the pool of empty populated pages. However, 181147504ee0SDennis Zhou * it is possible that this is called when physical memory is scarce causing 181247504ee0SDennis Zhou * OOM killer to be triggered. We should avoid doing so until an actual 181347504ee0SDennis Zhou * allocation causes the failure as it is possible that requests can be 181447504ee0SDennis Zhou * serviced from already backed regions. 1815a56dbddfSTejun Heo */ 1816fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1817fbf59bc9STejun Heo { 181847504ee0SDennis Zhou /* gfp flags passed to underlying allocators */ 1819554fef1cSDennis Zhou const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN; 1820fe6bd8c3STejun Heo LIST_HEAD(to_free); 1821fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1822a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 18231a4d7607STejun Heo int slot, nr_to_pop, ret; 1824a56dbddfSTejun Heo 18251a4d7607STejun Heo /* 18261a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 18271a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 18281a4d7607STejun Heo */ 1829ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1830ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1831a56dbddfSTejun Heo 1832fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 18338d408b4bSTejun Heo WARN_ON(chunk->immutable); 1834a56dbddfSTejun Heo 1835a56dbddfSTejun Heo /* spare the first one */ 1836fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1837a56dbddfSTejun Heo continue; 1838a56dbddfSTejun Heo 1839fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1840a56dbddfSTejun Heo } 1841a56dbddfSTejun Heo 1842ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1843a56dbddfSTejun Heo 1844fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1845e837dfdeSDennis Zhou unsigned int rs, re; 1846dca49645STejun Heo 1847e837dfdeSDennis Zhou bitmap_for_each_set_region(chunk->populated, rs, re, 0, 184891e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1849a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1850b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1851b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1852b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1853a93ace48STejun Heo } 18546081089fSTejun Heo pcpu_destroy_chunk(chunk); 1855accd4f36SEric Dumazet cond_resched(); 1856fbf59bc9STejun Heo } 1857971f3918STejun Heo 18581a4d7607STejun Heo /* 18591a4d7607STejun Heo * Ensure there are certain number of free populated pages for 18601a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 18611a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 18621a4d7607STejun Heo * failed previously, always populate the maximum amount. This 18631a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 18641a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 18651a4d7607STejun Heo * something we support properly and can be highly unreliable and 18661a4d7607STejun Heo * inefficient. 18671a4d7607STejun Heo */ 18681a4d7607STejun Heo retry_pop: 18691a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 18701a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 18711a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 18721a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 18731a4d7607STejun Heo } else { 18741a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 18751a4d7607STejun Heo pcpu_nr_empty_pop_pages, 18761a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 18771a4d7607STejun Heo } 18781a4d7607STejun Heo 18791a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 1880e837dfdeSDennis Zhou unsigned int nr_unpop = 0, rs, re; 18811a4d7607STejun Heo 18821a4d7607STejun Heo if (!nr_to_pop) 18831a4d7607STejun Heo break; 18841a4d7607STejun Heo 18851a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 18861a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 18878ab16c43SDennis Zhou (Facebook) nr_unpop = chunk->nr_pages - chunk->nr_populated; 18881a4d7607STejun Heo if (nr_unpop) 18891a4d7607STejun Heo break; 18901a4d7607STejun Heo } 18911a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 18921a4d7607STejun Heo 18931a4d7607STejun Heo if (!nr_unpop) 18941a4d7607STejun Heo continue; 18951a4d7607STejun Heo 18961a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 1897e837dfdeSDennis Zhou bitmap_for_each_clear_region(chunk->populated, rs, re, 0, 189891e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1899e837dfdeSDennis Zhou int nr = min_t(int, re - rs, nr_to_pop); 19001a4d7607STejun Heo 190147504ee0SDennis Zhou ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); 19021a4d7607STejun Heo if (!ret) { 19031a4d7607STejun Heo nr_to_pop -= nr; 19041a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 1905b239f7daSDennis Zhou pcpu_chunk_populated(chunk, rs, rs + nr); 19061a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 19071a4d7607STejun Heo } else { 19081a4d7607STejun Heo nr_to_pop = 0; 19091a4d7607STejun Heo } 19101a4d7607STejun Heo 19111a4d7607STejun Heo if (!nr_to_pop) 19121a4d7607STejun Heo break; 19131a4d7607STejun Heo } 19141a4d7607STejun Heo } 19151a4d7607STejun Heo 19161a4d7607STejun Heo if (nr_to_pop) { 19171a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 191847504ee0SDennis Zhou chunk = pcpu_create_chunk(gfp); 19191a4d7607STejun Heo if (chunk) { 19201a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 19211a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 19221a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 19231a4d7607STejun Heo goto retry_pop; 19241a4d7607STejun Heo } 19251a4d7607STejun Heo } 19261a4d7607STejun Heo 1927971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1928a56dbddfSTejun Heo } 1929fbf59bc9STejun Heo 1930fbf59bc9STejun Heo /** 1931fbf59bc9STejun Heo * free_percpu - free percpu area 1932fbf59bc9STejun Heo * @ptr: pointer to area to free 1933fbf59bc9STejun Heo * 1934ccea34b5STejun Heo * Free percpu area @ptr. 1935ccea34b5STejun Heo * 1936ccea34b5STejun Heo * CONTEXT: 1937ccea34b5STejun Heo * Can be called from atomic context. 1938fbf59bc9STejun Heo */ 193943cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1940fbf59bc9STejun Heo { 1941129182e5SAndrew Morton void *addr; 1942fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1943ccea34b5STejun Heo unsigned long flags; 194440064aecSDennis Zhou (Facebook) int off; 1945198790d9SJohn Sperbeck bool need_balance = false; 1946fbf59bc9STejun Heo 1947fbf59bc9STejun Heo if (!ptr) 1948fbf59bc9STejun Heo return; 1949fbf59bc9STejun Heo 1950f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1951f528f0b8SCatalin Marinas 1952129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1953129182e5SAndrew Morton 1954ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1955fbf59bc9STejun Heo 1956fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1957bba174f5STejun Heo off = addr - chunk->base_addr; 1958fbf59bc9STejun Heo 195940064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1960fbf59bc9STejun Heo 1961a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 196240064aecSDennis Zhou (Facebook) if (chunk->free_bytes == pcpu_unit_size) { 1963fbf59bc9STejun Heo struct pcpu_chunk *pos; 1964fbf59bc9STejun Heo 1965a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1966fbf59bc9STejun Heo if (pos != chunk) { 1967198790d9SJohn Sperbeck need_balance = true; 1968fbf59bc9STejun Heo break; 1969fbf59bc9STejun Heo } 1970fbf59bc9STejun Heo } 1971fbf59bc9STejun Heo 1972df95e795SDennis Zhou trace_percpu_free_percpu(chunk->base_addr, off, ptr); 1973df95e795SDennis Zhou 1974ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1975198790d9SJohn Sperbeck 1976198790d9SJohn Sperbeck if (need_balance) 1977198790d9SJohn Sperbeck pcpu_schedule_balance_work(); 1978fbf59bc9STejun Heo } 1979fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1980fbf59bc9STejun Heo 1981383776faSThomas Gleixner bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 1982383776faSThomas Gleixner { 1983383776faSThomas Gleixner #ifdef CONFIG_SMP 1984383776faSThomas Gleixner const size_t static_size = __per_cpu_end - __per_cpu_start; 1985383776faSThomas Gleixner void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 1986383776faSThomas Gleixner unsigned int cpu; 1987383776faSThomas Gleixner 1988383776faSThomas Gleixner for_each_possible_cpu(cpu) { 1989383776faSThomas Gleixner void *start = per_cpu_ptr(base, cpu); 1990383776faSThomas Gleixner void *va = (void *)addr; 1991383776faSThomas Gleixner 1992383776faSThomas Gleixner if (va >= start && va < start + static_size) { 19938ce371f9SPeter Zijlstra if (can_addr) { 1994383776faSThomas Gleixner *can_addr = (unsigned long) (va - start); 19958ce371f9SPeter Zijlstra *can_addr += (unsigned long) 19968ce371f9SPeter Zijlstra per_cpu_ptr(base, get_boot_cpu_id()); 19978ce371f9SPeter Zijlstra } 1998383776faSThomas Gleixner return true; 1999383776faSThomas Gleixner } 2000383776faSThomas Gleixner } 2001383776faSThomas Gleixner #endif 2002383776faSThomas Gleixner /* on UP, can't distinguish from other static vars, always false */ 2003383776faSThomas Gleixner return false; 2004383776faSThomas Gleixner } 2005383776faSThomas Gleixner 20063b034b0dSVivek Goyal /** 200710fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 200810fad5e4STejun Heo * @addr: address to test 200910fad5e4STejun Heo * 201010fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 201110fad5e4STejun Heo * static percpu areas are not considered. For those, use 201210fad5e4STejun Heo * is_module_percpu_address(). 201310fad5e4STejun Heo * 201410fad5e4STejun Heo * RETURNS: 201510fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 201610fad5e4STejun Heo */ 201710fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 201810fad5e4STejun Heo { 2019383776faSThomas Gleixner return __is_kernel_percpu_address(addr, NULL); 202010fad5e4STejun Heo } 202110fad5e4STejun Heo 202210fad5e4STejun Heo /** 20233b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 20243b034b0dSVivek Goyal * @addr: the address to be converted to physical address 20253b034b0dSVivek Goyal * 20263b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 20273b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 20283b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 20293b034b0dSVivek Goyal * until this function finishes. 20303b034b0dSVivek Goyal * 203167589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 203267589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 203367589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 203467589c71SDave Young * km) provides translation. 203567589c71SDave Young * 2036bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 203767589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 203867589c71SDave Young * actually works, and the verification can discover both bugs in percpu 203967589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 204067589c71SDave Young * code. 204167589c71SDave Young * 20423b034b0dSVivek Goyal * RETURNS: 20433b034b0dSVivek Goyal * The physical address for @addr. 20443b034b0dSVivek Goyal */ 20453b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 20463b034b0dSVivek Goyal { 20479983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 20489983b6f0STejun Heo bool in_first_chunk = false; 2049a855b84cSTejun Heo unsigned long first_low, first_high; 20509983b6f0STejun Heo unsigned int cpu; 20519983b6f0STejun Heo 20529983b6f0STejun Heo /* 2053a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 20549983b6f0STejun Heo * necessary but will speed up lookups of addresses which 20559983b6f0STejun Heo * aren't in the first chunk. 2056c0ebfdc3SDennis Zhou (Facebook) * 2057c0ebfdc3SDennis Zhou (Facebook) * The address check is against full chunk sizes. pcpu_base_addr 2058c0ebfdc3SDennis Zhou (Facebook) * points to the beginning of the first chunk including the 2059c0ebfdc3SDennis Zhou (Facebook) * static region. Assumes good intent as the first chunk may 2060c0ebfdc3SDennis Zhou (Facebook) * not be full (ie. < pcpu_unit_pages in size). 20619983b6f0STejun Heo */ 2062c0ebfdc3SDennis Zhou (Facebook) first_low = (unsigned long)pcpu_base_addr + 2063c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 2064c0ebfdc3SDennis Zhou (Facebook) first_high = (unsigned long)pcpu_base_addr + 2065c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 2066a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 2067a855b84cSTejun Heo (unsigned long)addr < first_high) { 20689983b6f0STejun Heo for_each_possible_cpu(cpu) { 20699983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 20709983b6f0STejun Heo 20719983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 20729983b6f0STejun Heo in_first_chunk = true; 20739983b6f0STejun Heo break; 20749983b6f0STejun Heo } 20759983b6f0STejun Heo } 20769983b6f0STejun Heo } 20779983b6f0STejun Heo 20789983b6f0STejun Heo if (in_first_chunk) { 2079eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 20803b034b0dSVivek Goyal return __pa(addr); 20813b034b0dSVivek Goyal else 20829f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 20839f57bd4dSEugene Surovegin offset_in_page(addr); 2084020ec653STejun Heo } else 20859f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 20869f57bd4dSEugene Surovegin offset_in_page(addr); 20873b034b0dSVivek Goyal } 20883b034b0dSVivek Goyal 2089fbf59bc9STejun Heo /** 2090fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 2091fd1e8a1fSTejun Heo * @nr_groups: the number of groups 2092fd1e8a1fSTejun Heo * @nr_units: the number of units 2093033e48fbSTejun Heo * 2094fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 2095fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 2096fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 2097fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 2098fd1e8a1fSTejun Heo * pointer of other groups. 2099033e48fbSTejun Heo * 2100033e48fbSTejun Heo * RETURNS: 2101fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 2102fd1e8a1fSTejun Heo * failure. 2103033e48fbSTejun Heo */ 2104fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 2105fd1e8a1fSTejun Heo int nr_units) 2106fd1e8a1fSTejun Heo { 2107fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 2108fd1e8a1fSTejun Heo size_t base_size, ai_size; 2109fd1e8a1fSTejun Heo void *ptr; 2110fd1e8a1fSTejun Heo int unit; 2111fd1e8a1fSTejun Heo 211214d37612SGustavo A. R. Silva base_size = ALIGN(struct_size(ai, groups, nr_groups), 2113fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 2114fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 2115fd1e8a1fSTejun Heo 211626fb3daeSMike Rapoport ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE); 2117fd1e8a1fSTejun Heo if (!ptr) 2118fd1e8a1fSTejun Heo return NULL; 2119fd1e8a1fSTejun Heo ai = ptr; 2120fd1e8a1fSTejun Heo ptr += base_size; 2121fd1e8a1fSTejun Heo 2122fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 2123fd1e8a1fSTejun Heo 2124fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 2125fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 2126fd1e8a1fSTejun Heo 2127fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 2128fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 2129fd1e8a1fSTejun Heo 2130fd1e8a1fSTejun Heo return ai; 2131fd1e8a1fSTejun Heo } 2132fd1e8a1fSTejun Heo 2133fd1e8a1fSTejun Heo /** 2134fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 2135fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 2136fd1e8a1fSTejun Heo * 2137fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 2138fd1e8a1fSTejun Heo */ 2139fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 2140fd1e8a1fSTejun Heo { 2141999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 2142fd1e8a1fSTejun Heo } 2143fd1e8a1fSTejun Heo 2144fd1e8a1fSTejun Heo /** 2145fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 2146fd1e8a1fSTejun Heo * @lvl: loglevel 2147fd1e8a1fSTejun Heo * @ai: allocation info to dump 2148fd1e8a1fSTejun Heo * 2149fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 2150fd1e8a1fSTejun Heo */ 2151fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 2152fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 2153033e48fbSTejun Heo { 2154fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 2155033e48fbSTejun Heo char empty_str[] = "--------"; 2156fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 2157fd1e8a1fSTejun Heo int group, v; 2158fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 2159033e48fbSTejun Heo 2160fd1e8a1fSTejun Heo v = ai->nr_groups; 2161033e48fbSTejun Heo while (v /= 10) 2162fd1e8a1fSTejun Heo group_width++; 2163033e48fbSTejun Heo 2164fd1e8a1fSTejun Heo v = num_possible_cpus(); 2165fd1e8a1fSTejun Heo while (v /= 10) 2166fd1e8a1fSTejun Heo cpu_width++; 2167fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 2168033e48fbSTejun Heo 2169fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 2170fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 2171fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 2172033e48fbSTejun Heo 2173fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 2174fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 2175fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 2176fd1e8a1fSTejun Heo 2177fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 2178fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 2179fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 2180fd1e8a1fSTejun Heo 2181fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 2182fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 2183fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 2184fd1e8a1fSTejun Heo if (!(alloc % apl)) { 21851170532bSJoe Perches pr_cont("\n"); 2186fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 2187033e48fbSTejun Heo } 21881170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 2189fd1e8a1fSTejun Heo 2190fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 2191fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 21921170532bSJoe Perches pr_cont("%0*d ", 21931170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 2194033e48fbSTejun Heo else 21951170532bSJoe Perches pr_cont("%s ", empty_str); 2196033e48fbSTejun Heo } 2197fd1e8a1fSTejun Heo } 21981170532bSJoe Perches pr_cont("\n"); 2199033e48fbSTejun Heo } 2200033e48fbSTejun Heo 2201fbf59bc9STejun Heo /** 22028d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 2203fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 220438a6be52STejun Heo * @base_addr: mapped address 2205fbf59bc9STejun Heo * 22068d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 220769ab285bSChristophe JAILLET * percpu area. This function is to be called from arch percpu area 220838a6be52STejun Heo * setup path. 22098d408b4bSTejun Heo * 2210fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 2211fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 22128d408b4bSTejun Heo * 2213fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 2214fd1e8a1fSTejun Heo * 2215fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 2216edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 2217edcb4639STejun Heo * the first chunk such that it's available only through reserved 2218edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 2219edcb4639STejun Heo * static areas on architectures where the addressing model has 2220edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 2221edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 2222edcb4639STejun Heo * 2223fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 2224fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 2225fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 22266074d5b0STejun Heo * 2227fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 2228fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 2229fd1e8a1fSTejun Heo * @ai->dyn_size. 22308d408b4bSTejun Heo * 2231fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 2232fd1e8a1fSTejun Heo * for vm areas. 22338d408b4bSTejun Heo * 2234fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 2235fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 2236fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 2237fd1e8a1fSTejun Heo * 2238fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 2239fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 2240fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 2241fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 2242fd1e8a1fSTejun Heo * all units is assumed. 22438d408b4bSTejun Heo * 224438a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 224538a6be52STejun Heo * copied static data to each unit. 2246fbf59bc9STejun Heo * 2247c0ebfdc3SDennis Zhou (Facebook) * The first chunk will always contain a static and a dynamic region. 2248c0ebfdc3SDennis Zhou (Facebook) * However, the static region is not managed by any chunk. If the first 2249c0ebfdc3SDennis Zhou (Facebook) * chunk also contains a reserved region, it is served by two chunks - 2250c0ebfdc3SDennis Zhou (Facebook) * one for the reserved region and one for the dynamic region. They 2251c0ebfdc3SDennis Zhou (Facebook) * share the same vm, but use offset regions in the area allocation map. 2252c0ebfdc3SDennis Zhou (Facebook) * The chunk serving the dynamic region is circulated in the chunk slots 2253c0ebfdc3SDennis Zhou (Facebook) * and available for dynamic allocation like any other chunk. 2254fbf59bc9STejun Heo */ 2255163fa234SKefeng Wang void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 2256fd1e8a1fSTejun Heo void *base_addr) 2257fbf59bc9STejun Heo { 2258b9c39442SDennis Zhou (Facebook) size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2259d2f3c384SDennis Zhou (Facebook) size_t static_size, dyn_size; 22600c4169c3SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 22616563297cSTejun Heo unsigned long *group_offsets; 22626563297cSTejun Heo size_t *group_sizes; 2263fb435d52STejun Heo unsigned long *unit_off; 2264fbf59bc9STejun Heo unsigned int cpu; 2265fd1e8a1fSTejun Heo int *unit_map; 2266fd1e8a1fSTejun Heo int group, unit, i; 2267c0ebfdc3SDennis Zhou (Facebook) int map_size; 2268c0ebfdc3SDennis Zhou (Facebook) unsigned long tmp_addr; 2269f655f405SMike Rapoport size_t alloc_size; 2270fbf59bc9STejun Heo 2271635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 2272635b75fcSTejun Heo if (unlikely(cond)) { \ 2273870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 2274870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 2275807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 2276635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 2277635b75fcSTejun Heo BUG(); \ 2278635b75fcSTejun Heo } \ 2279635b75fcSTejun Heo } while (0) 2280635b75fcSTejun Heo 22812f39e637STejun Heo /* sanity checks */ 2282635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 2283bbddff05STejun Heo #ifdef CONFIG_SMP 2284635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 2285f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 2286bbddff05STejun Heo #endif 2287635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 2288f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 2289635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 2290f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 2291635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 2292ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 2293099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 2294fb29a2ccSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!ai->dyn_size); 2295d2f3c384SDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 2296ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 2297ca460b3cSDennis Zhou (Facebook) IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 22989f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 22998d408b4bSTejun Heo 23006563297cSTejun Heo /* process group information and build config tables accordingly */ 2301f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_offsets[0]); 2302f655f405SMike Rapoport group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2303f655f405SMike Rapoport if (!group_offsets) 2304f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2305f655f405SMike Rapoport alloc_size); 2306f655f405SMike Rapoport 2307f655f405SMike Rapoport alloc_size = ai->nr_groups * sizeof(group_sizes[0]); 2308f655f405SMike Rapoport group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2309f655f405SMike Rapoport if (!group_sizes) 2310f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2311f655f405SMike Rapoport alloc_size); 2312f655f405SMike Rapoport 2313f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_map[0]); 2314f655f405SMike Rapoport unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2315f655f405SMike Rapoport if (!unit_map) 2316f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2317f655f405SMike Rapoport alloc_size); 2318f655f405SMike Rapoport 2319f655f405SMike Rapoport alloc_size = nr_cpu_ids * sizeof(unit_off[0]); 2320f655f405SMike Rapoport unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES); 2321f655f405SMike Rapoport if (!unit_off) 2322f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2323f655f405SMike Rapoport alloc_size); 23242f39e637STejun Heo 2325fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 2326ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 2327a855b84cSTejun Heo 2328a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 2329a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 23302f39e637STejun Heo 2331fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 2332fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 23332f39e637STejun Heo 23346563297cSTejun Heo group_offsets[group] = gi->base_offset; 23356563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 23366563297cSTejun Heo 2337fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 2338fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 2339fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 2340fd1e8a1fSTejun Heo continue; 2341fd1e8a1fSTejun Heo 23429f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 2343635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 2344635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 2345fd1e8a1fSTejun Heo 2346fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 2347fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 2348fb435d52STejun Heo 2349a855b84cSTejun Heo /* determine low/high unit_cpu */ 2350a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 2351a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 2352a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 2353a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 2354a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 2355a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 23560fc0531eSLinus Torvalds } 23570fc0531eSLinus Torvalds } 2358fd1e8a1fSTejun Heo pcpu_nr_units = unit; 23592f39e637STejun Heo 23602f39e637STejun Heo for_each_possible_cpu(cpu) 2361635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 2362635b75fcSTejun Heo 2363635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 2364635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 2365bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 23662f39e637STejun Heo 23676563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 23686563297cSTejun Heo pcpu_group_offsets = group_offsets; 23696563297cSTejun Heo pcpu_group_sizes = group_sizes; 2370fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 2371fb435d52STejun Heo pcpu_unit_offsets = unit_off; 23722f39e637STejun Heo 23732f39e637STejun Heo /* determine basic parameters */ 2374fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 2375d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 23766563297cSTejun Heo pcpu_atom_size = ai->atom_size; 2377ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 2378ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 2379cafe8816STejun Heo 238030a5b536SDennis Zhou pcpu_stats_save_ai(ai); 238130a5b536SDennis Zhou 2382d9b55eebSTejun Heo /* 2383d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 2384d9b55eebSTejun Heo * empty chunks. 2385d9b55eebSTejun Heo */ 2386d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 23877e1c4e27SMike Rapoport pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]), 23887e1c4e27SMike Rapoport SMP_CACHE_BYTES); 2389f655f405SMike Rapoport if (!pcpu_slot) 2390f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2391f655f405SMike Rapoport pcpu_nr_slots * sizeof(pcpu_slot[0])); 2392fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 2393fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 2394fbf59bc9STejun Heo 2395edcb4639STejun Heo /* 2396d2f3c384SDennis Zhou (Facebook) * The end of the static region needs to be aligned with the 2397d2f3c384SDennis Zhou (Facebook) * minimum allocation size as this offsets the reserved and 2398d2f3c384SDennis Zhou (Facebook) * dynamic region. The first chunk ends page aligned by 2399d2f3c384SDennis Zhou (Facebook) * expanding the dynamic region, therefore the dynamic region 2400d2f3c384SDennis Zhou (Facebook) * can be shrunk to compensate while still staying above the 2401d2f3c384SDennis Zhou (Facebook) * configured sizes. 2402d2f3c384SDennis Zhou (Facebook) */ 2403d2f3c384SDennis Zhou (Facebook) static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 2404d2f3c384SDennis Zhou (Facebook) dyn_size = ai->dyn_size - (static_size - ai->static_size); 2405d2f3c384SDennis Zhou (Facebook) 2406d2f3c384SDennis Zhou (Facebook) /* 2407c0ebfdc3SDennis Zhou (Facebook) * Initialize first chunk. 2408c0ebfdc3SDennis Zhou (Facebook) * If the reserved_size is non-zero, this initializes the reserved 2409c0ebfdc3SDennis Zhou (Facebook) * chunk. If the reserved_size is zero, the reserved chunk is NULL 2410c0ebfdc3SDennis Zhou (Facebook) * and the dynamic region is initialized here. The first chunk, 2411c0ebfdc3SDennis Zhou (Facebook) * pcpu_first_chunk, will always point to the chunk that serves 2412c0ebfdc3SDennis Zhou (Facebook) * the dynamic region. 2413edcb4639STejun Heo */ 2414d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size; 2415d2f3c384SDennis Zhou (Facebook) map_size = ai->reserved_size ?: dyn_size; 241640064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 241761ace7faSTejun Heo 2418edcb4639STejun Heo /* init dynamic chunk if necessary */ 2419b9c39442SDennis Zhou (Facebook) if (ai->reserved_size) { 24200c4169c3SDennis Zhou (Facebook) pcpu_reserved_chunk = chunk; 2421b9c39442SDennis Zhou (Facebook) 2422d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size + 2423c0ebfdc3SDennis Zhou (Facebook) ai->reserved_size; 2424d2f3c384SDennis Zhou (Facebook) map_size = dyn_size; 242540064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 2426edcb4639STejun Heo } 2427edcb4639STejun Heo 24282441d15cSTejun Heo /* link the first chunk in */ 24290c4169c3SDennis Zhou (Facebook) pcpu_first_chunk = chunk; 24300cecf50cSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 2431ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 2432fbf59bc9STejun Heo 24337e8a6304SDennis Zhou (Facebook) /* include all regions of the first chunk */ 24347e8a6304SDennis Zhou (Facebook) pcpu_nr_populated += PFN_DOWN(size_sum); 24357e8a6304SDennis Zhou (Facebook) 243630a5b536SDennis Zhou pcpu_stats_chunk_alloc(); 2437df95e795SDennis Zhou trace_percpu_create_chunk(base_addr); 243830a5b536SDennis Zhou 2439fbf59bc9STejun Heo /* we're done */ 2440bba174f5STejun Heo pcpu_base_addr = base_addr; 2441fbf59bc9STejun Heo } 244266c3a757STejun Heo 2443bbddff05STejun Heo #ifdef CONFIG_SMP 2444bbddff05STejun Heo 244517f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 2446f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 2447f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 2448f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 2449f58dc01bSTejun Heo }; 245066c3a757STejun Heo 2451f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 2452f58dc01bSTejun Heo 2453f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 245466c3a757STejun Heo { 24555479c78aSCyrill Gorcunov if (!str) 24565479c78aSCyrill Gorcunov return -EINVAL; 24575479c78aSCyrill Gorcunov 2458f58dc01bSTejun Heo if (0) 2459f58dc01bSTejun Heo /* nada */; 2460f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 2461f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 2462f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 2463f58dc01bSTejun Heo #endif 2464f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2465f58dc01bSTejun Heo else if (!strcmp(str, "page")) 2466f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 2467f58dc01bSTejun Heo #endif 2468f58dc01bSTejun Heo else 2469870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 247066c3a757STejun Heo 2471f58dc01bSTejun Heo return 0; 247266c3a757STejun Heo } 2473f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 247466c3a757STejun Heo 24753c9a024fSTejun Heo /* 24763c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 24773c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 24783c9a024fSTejun Heo * to be used. 24793c9a024fSTejun Heo */ 248008fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 248108fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 24823c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 24833c9a024fSTejun Heo #endif 24843c9a024fSTejun Heo 24853c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 24863c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 24873c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 24883c9a024fSTejun Heo #endif 24893c9a024fSTejun Heo 24903c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 24913c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 24923c9a024fSTejun Heo /** 2493fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2494fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2495fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2496fbf59bc9STejun Heo * @atom_size: allocation atom size 2497fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2498fbf59bc9STejun Heo * 2499fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 2500fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 2501fbf59bc9STejun Heo * atom size and distances between CPUs. 2502fbf59bc9STejun Heo * 2503bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 2504fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 2505fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 2506fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 2507fbf59bc9STejun Heo * of allocated virtual address space. 2508fbf59bc9STejun Heo * 2509fbf59bc9STejun Heo * RETURNS: 2510fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 2511fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 2512fbf59bc9STejun Heo */ 2513fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 2514fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 2515fbf59bc9STejun Heo size_t atom_size, 2516fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2517fbf59bc9STejun Heo { 2518fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 2519fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 2520fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 2521fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 2522fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 25233f649ab7SKees Cook int upa, max_upa, best_upa; /* units_per_alloc */ 2524fbf59bc9STejun Heo int last_allocs, group, unit; 2525fbf59bc9STejun Heo unsigned int cpu, tcpu; 2526fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 2527fbf59bc9STejun Heo unsigned int *cpu_map; 2528fbf59bc9STejun Heo 2529fbf59bc9STejun Heo /* this function may be called multiple times */ 2530fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 2531fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 2532fbf59bc9STejun Heo 2533fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2534fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 2535fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2536fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 2537fbf59bc9STejun Heo 2538fbf59bc9STejun Heo /* 2539fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 2540fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 254125985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 2542fbf59bc9STejun Heo * or larger than min_unit_size. 2543fbf59bc9STejun Heo */ 2544fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2545fbf59bc9STejun Heo 25469c015162SDennis Zhou (Facebook) /* determine the maximum # of units that can fit in an allocation */ 2547fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 2548fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 2549f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2550fbf59bc9STejun Heo upa--; 2551fbf59bc9STejun Heo max_upa = upa; 2552fbf59bc9STejun Heo 2553fbf59bc9STejun Heo /* group cpus according to their proximity */ 2554fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 2555fbf59bc9STejun Heo group = 0; 2556fbf59bc9STejun Heo next_group: 2557fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 2558fbf59bc9STejun Heo if (cpu == tcpu) 2559fbf59bc9STejun Heo break; 2560fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 2561fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 2562fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 2563fbf59bc9STejun Heo group++; 2564fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 2565fbf59bc9STejun Heo goto next_group; 2566fbf59bc9STejun Heo } 2567fbf59bc9STejun Heo } 2568fbf59bc9STejun Heo group_map[cpu] = group; 2569fbf59bc9STejun Heo group_cnt[group]++; 2570fbf59bc9STejun Heo } 2571fbf59bc9STejun Heo 2572fbf59bc9STejun Heo /* 25739c015162SDennis Zhou (Facebook) * Wasted space is caused by a ratio imbalance of upa to group_cnt. 25749c015162SDennis Zhou (Facebook) * Expand the unit_size until we use >= 75% of the units allocated. 25759c015162SDennis Zhou (Facebook) * Related to atom_size, which could be much larger than the unit_size. 2576fbf59bc9STejun Heo */ 2577fbf59bc9STejun Heo last_allocs = INT_MAX; 2578fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 2579fbf59bc9STejun Heo int allocs = 0, wasted = 0; 2580fbf59bc9STejun Heo 2581f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2582fbf59bc9STejun Heo continue; 2583fbf59bc9STejun Heo 2584fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2585fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2586fbf59bc9STejun Heo allocs += this_allocs; 2587fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 2588fbf59bc9STejun Heo } 2589fbf59bc9STejun Heo 2590fbf59bc9STejun Heo /* 2591fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 2592fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 2593fbf59bc9STejun Heo * passes the following check. 2594fbf59bc9STejun Heo */ 2595fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 2596fbf59bc9STejun Heo continue; 2597fbf59bc9STejun Heo 2598fbf59bc9STejun Heo /* and then don't consume more memory */ 2599fbf59bc9STejun Heo if (allocs > last_allocs) 2600fbf59bc9STejun Heo break; 2601fbf59bc9STejun Heo last_allocs = allocs; 2602fbf59bc9STejun Heo best_upa = upa; 2603fbf59bc9STejun Heo } 2604fbf59bc9STejun Heo upa = best_upa; 2605fbf59bc9STejun Heo 2606fbf59bc9STejun Heo /* allocate and fill alloc_info */ 2607fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 2608fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 2609fbf59bc9STejun Heo 2610fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2611fbf59bc9STejun Heo if (!ai) 2612fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 2613fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 2614fbf59bc9STejun Heo 2615fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2616fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 2617fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 2618fbf59bc9STejun Heo } 2619fbf59bc9STejun Heo 2620fbf59bc9STejun Heo ai->static_size = static_size; 2621fbf59bc9STejun Heo ai->reserved_size = reserved_size; 2622fbf59bc9STejun Heo ai->dyn_size = dyn_size; 2623fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 2624fbf59bc9STejun Heo ai->atom_size = atom_size; 2625fbf59bc9STejun Heo ai->alloc_size = alloc_size; 2626fbf59bc9STejun Heo 26272de7852fSPeng Fan for (group = 0, unit = 0; group < nr_groups; group++) { 2628fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2629fbf59bc9STejun Heo 2630fbf59bc9STejun Heo /* 2631fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 2632fbf59bc9STejun Heo * back-to-back. The caller should update this to 2633fbf59bc9STejun Heo * reflect actual allocation. 2634fbf59bc9STejun Heo */ 2635fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 2636fbf59bc9STejun Heo 2637fbf59bc9STejun Heo for_each_possible_cpu(cpu) 2638fbf59bc9STejun Heo if (group_map[cpu] == group) 2639fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 2640fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 2641fbf59bc9STejun Heo unit += gi->nr_units; 2642fbf59bc9STejun Heo } 2643fbf59bc9STejun Heo BUG_ON(unit != nr_units); 2644fbf59bc9STejun Heo 2645fbf59bc9STejun Heo return ai; 2646fbf59bc9STejun Heo } 26473c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2648fbf59bc9STejun Heo 26493c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 265066c3a757STejun Heo /** 265166c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 265266c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 26534ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2654c8826dd5STejun Heo * @atom_size: allocation atom size 2655c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2656c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 265725985edcSLucas De Marchi * @free_fn: function to free percpu page 265866c3a757STejun Heo * 265966c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 266066c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 266166c3a757STejun Heo * 266266c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 2663c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 2664c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 2665c8826dd5STejun Heo * aligned to @atom_size. 2666c8826dd5STejun Heo * 2667c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 2668c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 2669c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 2670c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 2671c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 2672c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 267366c3a757STejun Heo * 26744ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 267566c3a757STejun Heo * 267666c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 2677c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 267866c3a757STejun Heo * 267966c3a757STejun Heo * RETURNS: 2680fb435d52STejun Heo * 0 on success, -errno on failure. 268166c3a757STejun Heo */ 26824ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 2683c8826dd5STejun Heo size_t atom_size, 2684c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 2685c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2686c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 268766c3a757STejun Heo { 2688c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 2689c8826dd5STejun Heo void **areas = NULL; 2690fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 269193c76b6bSzijun_hu size_t size_sum, areas_size; 269293c76b6bSzijun_hu unsigned long max_distance; 2693163fa234SKefeng Wang int group, i, highest_group, rc = 0; 269466c3a757STejun Heo 2695c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 2696c8826dd5STejun Heo cpu_distance_fn); 2697fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2698fd1e8a1fSTejun Heo return PTR_ERR(ai); 269966c3a757STejun Heo 2700fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2701c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 270266c3a757STejun Heo 270326fb3daeSMike Rapoport areas = memblock_alloc(areas_size, SMP_CACHE_BYTES); 2704c8826dd5STejun Heo if (!areas) { 2705fb435d52STejun Heo rc = -ENOMEM; 2706c8826dd5STejun Heo goto out_free; 2707fa8a7094STejun Heo } 270866c3a757STejun Heo 27099b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 27109b739662Szijun_hu highest_group = 0; 2711c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2712c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2713c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 2714c8826dd5STejun Heo void *ptr; 271566c3a757STejun Heo 2716c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 2717c8826dd5STejun Heo cpu = gi->cpu_map[i]; 2718c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 2719c8826dd5STejun Heo 2720c8826dd5STejun Heo /* allocate space for the whole group */ 2721c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 2722c8826dd5STejun Heo if (!ptr) { 2723c8826dd5STejun Heo rc = -ENOMEM; 2724c8826dd5STejun Heo goto out_free_areas; 2725c8826dd5STejun Heo } 2726f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2727f528f0b8SCatalin Marinas kmemleak_free(ptr); 2728c8826dd5STejun Heo areas[group] = ptr; 2729c8826dd5STejun Heo 2730c8826dd5STejun Heo base = min(ptr, base); 27319b739662Szijun_hu if (ptr > areas[highest_group]) 27329b739662Szijun_hu highest_group = group; 27339b739662Szijun_hu } 27349b739662Szijun_hu max_distance = areas[highest_group] - base; 27359b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 27369b739662Szijun_hu 27379b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 27389b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 27399b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 27409b739662Szijun_hu max_distance, VMALLOC_TOTAL); 27419b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 27429b739662Szijun_hu /* and fail if we have fallback */ 27439b739662Szijun_hu rc = -EINVAL; 27449b739662Szijun_hu goto out_free_areas; 27459b739662Szijun_hu #endif 274642b64281STejun Heo } 274742b64281STejun Heo 274842b64281STejun Heo /* 274942b64281STejun Heo * Copy data and free unused parts. This should happen after all 275042b64281STejun Heo * allocations are complete; otherwise, we may end up with 275142b64281STejun Heo * overlapping groups. 275242b64281STejun Heo */ 275342b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 275442b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 275542b64281STejun Heo void *ptr = areas[group]; 2756c8826dd5STejun Heo 2757c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2758c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2759c8826dd5STejun Heo /* unused unit, free whole */ 2760c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2761c8826dd5STejun Heo continue; 2762c8826dd5STejun Heo } 2763c8826dd5STejun Heo /* copy and return the unused part */ 2764fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2765c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2766c8826dd5STejun Heo } 276766c3a757STejun Heo } 276866c3a757STejun Heo 2769c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 27706ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2771c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 27726ea529a2STejun Heo } 2773c8826dd5STejun Heo 277400206a69SMatteo Croce pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n", 277500206a69SMatteo Croce PFN_DOWN(size_sum), ai->static_size, ai->reserved_size, 2776fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 277766c3a757STejun Heo 2778163fa234SKefeng Wang pcpu_setup_first_chunk(ai, base); 2779c8826dd5STejun Heo goto out_free; 2780c8826dd5STejun Heo 2781c8826dd5STejun Heo out_free_areas: 2782c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2783f851c8d8SMichael Holzheu if (areas[group]) 2784c8826dd5STejun Heo free_fn(areas[group], 2785c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2786c8826dd5STejun Heo out_free: 2787fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2788c8826dd5STejun Heo if (areas) 2789999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2790fb435d52STejun Heo return rc; 2791d4b95f80STejun Heo } 27923c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2793d4b95f80STejun Heo 27943c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2795d4b95f80STejun Heo /** 279600ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2797d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2798d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 279925985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2800d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2801d4b95f80STejun Heo * 280200ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 280300ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2804d4b95f80STejun Heo * 2805d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2806d4b95f80STejun Heo * page-by-page into vmalloc area. 2807d4b95f80STejun Heo * 2808d4b95f80STejun Heo * RETURNS: 2809fb435d52STejun Heo * 0 on success, -errno on failure. 2810d4b95f80STejun Heo */ 2811fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2812d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2813d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2814d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2815d4b95f80STejun Heo { 28168f05a6a6STejun Heo static struct vm_struct vm; 2817fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 281800ae4064STejun Heo char psize_str[16]; 2819ce3141a2STejun Heo int unit_pages; 2820d4b95f80STejun Heo size_t pages_size; 2821ce3141a2STejun Heo struct page **pages; 2822163fa234SKefeng Wang int unit, i, j, rc = 0; 28238f606604Szijun_hu int upa; 28248f606604Szijun_hu int nr_g0_units; 2825d4b95f80STejun Heo 282600ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 282700ae4064STejun Heo 28284ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2829fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2830fd1e8a1fSTejun Heo return PTR_ERR(ai); 2831fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 28328f606604Szijun_hu upa = ai->alloc_size/ai->unit_size; 28338f606604Szijun_hu nr_g0_units = roundup(num_possible_cpus(), upa); 28340b59c25fSIgor Stoppa if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) { 28358f606604Szijun_hu pcpu_free_alloc_info(ai); 28368f606604Szijun_hu return -EINVAL; 28378f606604Szijun_hu } 2838fd1e8a1fSTejun Heo 2839fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2840d4b95f80STejun Heo 2841d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2842fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2843fd1e8a1fSTejun Heo sizeof(pages[0])); 28447e1c4e27SMike Rapoport pages = memblock_alloc(pages_size, SMP_CACHE_BYTES); 2845f655f405SMike Rapoport if (!pages) 2846f655f405SMike Rapoport panic("%s: Failed to allocate %zu bytes\n", __func__, 2847f655f405SMike Rapoport pages_size); 2848d4b95f80STejun Heo 28498f05a6a6STejun Heo /* allocate pages */ 2850d4b95f80STejun Heo j = 0; 28518f606604Szijun_hu for (unit = 0; unit < num_possible_cpus(); unit++) { 2852fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 28538f606604Szijun_hu for (i = 0; i < unit_pages; i++) { 2854d4b95f80STejun Heo void *ptr; 2855d4b95f80STejun Heo 28563cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2857d4b95f80STejun Heo if (!ptr) { 2858870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2859598d8091SJoe Perches psize_str, cpu); 2860d4b95f80STejun Heo goto enomem; 2861d4b95f80STejun Heo } 2862f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2863f528f0b8SCatalin Marinas kmemleak_free(ptr); 2864ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2865d4b95f80STejun Heo } 28668f606604Szijun_hu } 2867d4b95f80STejun Heo 28688f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 28698f05a6a6STejun Heo vm.flags = VM_ALLOC; 2870fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 28718f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 28728f05a6a6STejun Heo 2873fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 28741d9d3257STejun Heo unsigned long unit_addr = 2875fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 28768f05a6a6STejun Heo 2877ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 28788f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 28798f05a6a6STejun Heo 28808f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2881fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2882ce3141a2STejun Heo unit_pages); 2883fb435d52STejun Heo if (rc < 0) 2884fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 28858f05a6a6STejun Heo 28868f05a6a6STejun Heo /* 28878f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 28888f05a6a6STejun Heo * cache for the linear mapping here - something 28898f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 28908f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 28918f05a6a6STejun Heo * data structures are not set up yet. 28928f05a6a6STejun Heo */ 28938f05a6a6STejun Heo 28948f05a6a6STejun Heo /* copy static data */ 2895fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 289666c3a757STejun Heo } 289766c3a757STejun Heo 289866c3a757STejun Heo /* we're ready, commit */ 289900206a69SMatteo Croce pr_info("%d %s pages/cpu s%zu r%zu d%zu\n", 290000206a69SMatteo Croce unit_pages, psize_str, ai->static_size, 2901fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 290266c3a757STejun Heo 2903163fa234SKefeng Wang pcpu_setup_first_chunk(ai, vm.addr); 2904d4b95f80STejun Heo goto out_free_ar; 2905d4b95f80STejun Heo 2906d4b95f80STejun Heo enomem: 2907d4b95f80STejun Heo while (--j >= 0) 2908ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2909fb435d52STejun Heo rc = -ENOMEM; 2910d4b95f80STejun Heo out_free_ar: 2911999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2912fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2913fb435d52STejun Heo return rc; 291466c3a757STejun Heo } 29153c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2916d4b95f80STejun Heo 2917bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 29188c4bfc6eSTejun Heo /* 2919bbddff05STejun Heo * Generic SMP percpu area setup. 2920e74e3962STejun Heo * 2921e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2922e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2923e74e3962STejun Heo * important because many archs have addressing restrictions and might 2924e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2925e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2926e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2927e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2928e74e3962STejun Heo * mappings on applicable archs. 2929e74e3962STejun Heo */ 2930e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2931e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2932e74e3962STejun Heo 2933c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2934c8826dd5STejun Heo size_t align) 2935c8826dd5STejun Heo { 293626fb3daeSMike Rapoport return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS)); 2937c8826dd5STejun Heo } 2938c8826dd5STejun Heo 2939c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2940c8826dd5STejun Heo { 2941999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2942c8826dd5STejun Heo } 2943c8826dd5STejun Heo 2944e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2945e74e3962STejun Heo { 2946e74e3962STejun Heo unsigned long delta; 2947e74e3962STejun Heo unsigned int cpu; 2948fb435d52STejun Heo int rc; 2949e74e3962STejun Heo 2950e74e3962STejun Heo /* 2951e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2952e74e3962STejun Heo * what the legacy allocator did. 2953e74e3962STejun Heo */ 2954fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2955c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2956c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2957fb435d52STejun Heo if (rc < 0) 2958bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2959e74e3962STejun Heo 2960e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2961e74e3962STejun Heo for_each_possible_cpu(cpu) 2962fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2963e74e3962STejun Heo } 2964e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2965099a19d9STejun Heo 2966bbddff05STejun Heo #else /* CONFIG_SMP */ 2967bbddff05STejun Heo 2968bbddff05STejun Heo /* 2969bbddff05STejun Heo * UP percpu area setup. 2970bbddff05STejun Heo * 2971bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2972bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2973bbddff05STejun Heo * variables and don't require any special preparation. 2974bbddff05STejun Heo */ 2975bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2976bbddff05STejun Heo { 2977bbddff05STejun Heo const size_t unit_size = 2978bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2979bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2980bbddff05STejun Heo struct pcpu_alloc_info *ai; 2981bbddff05STejun Heo void *fc; 2982bbddff05STejun Heo 2983bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 298426fb3daeSMike Rapoport fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); 2985bbddff05STejun Heo if (!ai || !fc) 2986bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2987100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2988100d13c3SCatalin Marinas kmemleak_free(fc); 2989bbddff05STejun Heo 2990bbddff05STejun Heo ai->dyn_size = unit_size; 2991bbddff05STejun Heo ai->unit_size = unit_size; 2992bbddff05STejun Heo ai->atom_size = unit_size; 2993bbddff05STejun Heo ai->alloc_size = unit_size; 2994bbddff05STejun Heo ai->groups[0].nr_units = 1; 2995bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2996bbddff05STejun Heo 2997163fa234SKefeng Wang pcpu_setup_first_chunk(ai, fc); 2998438a5061SNicolas Pitre pcpu_free_alloc_info(ai); 2999bbddff05STejun Heo } 3000bbddff05STejun Heo 3001bbddff05STejun Heo #endif /* CONFIG_SMP */ 3002bbddff05STejun Heo 3003099a19d9STejun Heo /* 30047e8a6304SDennis Zhou (Facebook) * pcpu_nr_pages - calculate total number of populated backing pages 30057e8a6304SDennis Zhou (Facebook) * 30067e8a6304SDennis Zhou (Facebook) * This reflects the number of pages populated to back chunks. Metadata is 30077e8a6304SDennis Zhou (Facebook) * excluded in the number exposed in meminfo as the number of backing pages 30087e8a6304SDennis Zhou (Facebook) * scales with the number of cpus and can quickly outweigh the memory used for 30097e8a6304SDennis Zhou (Facebook) * metadata. It also keeps this calculation nice and simple. 30107e8a6304SDennis Zhou (Facebook) * 30117e8a6304SDennis Zhou (Facebook) * RETURNS: 30127e8a6304SDennis Zhou (Facebook) * Total number of populated backing pages in use by the allocator. 30137e8a6304SDennis Zhou (Facebook) */ 30147e8a6304SDennis Zhou (Facebook) unsigned long pcpu_nr_pages(void) 30157e8a6304SDennis Zhou (Facebook) { 30167e8a6304SDennis Zhou (Facebook) return pcpu_nr_populated * pcpu_nr_units; 30177e8a6304SDennis Zhou (Facebook) } 30187e8a6304SDennis Zhou (Facebook) 30197e8a6304SDennis Zhou (Facebook) /* 30201a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 30211a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 30221a4d7607STejun Heo * and running. 30231a4d7607STejun Heo */ 30241a4d7607STejun Heo static int __init percpu_enable_async(void) 30251a4d7607STejun Heo { 30261a4d7607STejun Heo pcpu_async_enabled = true; 30271a4d7607STejun Heo return 0; 30281a4d7607STejun Heo } 30291a4d7607STejun Heo subsys_initcall(percpu_enable_async); 3030