1fbf59bc9STejun Heo /* 288999a89STejun Heo * mm/percpu.c - percpu memory allocator 3fbf59bc9STejun Heo * 4fbf59bc9STejun Heo * Copyright (C) 2009 SUSE Linux Products GmbH 5fbf59bc9STejun Heo * Copyright (C) 2009 Tejun Heo <tj@kernel.org> 6fbf59bc9STejun Heo * 7fbf59bc9STejun Heo * This file is released under the GPLv2. 8fbf59bc9STejun Heo * 9fbf59bc9STejun Heo * This is percpu allocator which can handle both static and dynamic 1088999a89STejun Heo * areas. Percpu areas are allocated in chunks. Each chunk is 1188999a89STejun Heo * consisted of boot-time determined number of units and the first 1288999a89STejun Heo * chunk is used for static percpu variables in the kernel image 132f39e637STejun Heo * (special boot time alloc/init handling necessary as these areas 142f39e637STejun Heo * need to be brought up before allocation services are running). 152f39e637STejun Heo * Unit grows as necessary and all units grow or shrink in unison. 1688999a89STejun Heo * When a chunk is filled up, another chunk is allocated. 17fbf59bc9STejun Heo * 18fbf59bc9STejun Heo * c0 c1 c2 19fbf59bc9STejun Heo * ------------------- ------------------- ------------ 20fbf59bc9STejun Heo * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 21fbf59bc9STejun Heo * ------------------- ...... ------------------- .... ------------ 22fbf59bc9STejun Heo * 23fbf59bc9STejun Heo * Allocation is done in offset-size areas of single unit space. Ie, 24fbf59bc9STejun Heo * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, 252f39e637STejun Heo * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to 262f39e637STejun Heo * cpus. On NUMA, the mapping can be non-linear and even sparse. 272f39e637STejun Heo * Percpu access can be done by configuring percpu base registers 282f39e637STejun Heo * according to cpu to unit mapping and pcpu_unit_size. 29fbf59bc9STejun Heo * 302f39e637STejun Heo * There are usually many small percpu allocations many of them being 312f39e637STejun Heo * as small as 4 bytes. The allocator organizes chunks into lists 32fbf59bc9STejun Heo * according to free size and tries to allocate from the fullest one. 33fbf59bc9STejun Heo * Each chunk keeps the maximum contiguous area size hint which is 344785879eSNamhyung Kim * guaranteed to be equal to or larger than the maximum contiguous 35fbf59bc9STejun Heo * area in the chunk. This helps the allocator not to iterate the 36fbf59bc9STejun Heo * chunk maps unnecessarily. 37fbf59bc9STejun Heo * 38fbf59bc9STejun Heo * Allocation state in each chunk is kept using an array of integers 39fbf59bc9STejun Heo * on chunk->map. A positive value in the map represents a free 40fbf59bc9STejun Heo * region and negative allocated. Allocation inside a chunk is done 41fbf59bc9STejun Heo * by scanning this map sequentially and serving the first matching 42fbf59bc9STejun Heo * entry. This is mostly copied from the percpu_modalloc() allocator. 43e1b9aa3fSChristoph Lameter * Chunks can be determined from the address using the index field 44e1b9aa3fSChristoph Lameter * in the page struct. The index field contains a pointer to the chunk. 45fbf59bc9STejun Heo * 46fbf59bc9STejun Heo * To use this allocator, arch code should do the followings. 47fbf59bc9STejun Heo * 48fbf59bc9STejun Heo * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 49e0100983STejun Heo * regular address to percpu pointer and back if they need to be 50e0100983STejun Heo * different from the default 51fbf59bc9STejun Heo * 528d408b4bSTejun Heo * - use pcpu_setup_first_chunk() during percpu area initialization to 538d408b4bSTejun Heo * setup the first chunk containing the kernel static percpu area 54fbf59bc9STejun Heo */ 55fbf59bc9STejun Heo 56870d4b12SJoe Perches #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 57870d4b12SJoe Perches 58fbf59bc9STejun Heo #include <linux/bitmap.h> 59fbf59bc9STejun Heo #include <linux/bootmem.h> 60fd1e8a1fSTejun Heo #include <linux/err.h> 61fbf59bc9STejun Heo #include <linux/list.h> 62a530b795STejun Heo #include <linux/log2.h> 63fbf59bc9STejun Heo #include <linux/mm.h> 64fbf59bc9STejun Heo #include <linux/module.h> 65fbf59bc9STejun Heo #include <linux/mutex.h> 66fbf59bc9STejun Heo #include <linux/percpu.h> 67fbf59bc9STejun Heo #include <linux/pfn.h> 68fbf59bc9STejun Heo #include <linux/slab.h> 69ccea34b5STejun Heo #include <linux/spinlock.h> 70fbf59bc9STejun Heo #include <linux/vmalloc.h> 71a56dbddfSTejun Heo #include <linux/workqueue.h> 72f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 73fbf59bc9STejun Heo 74fbf59bc9STejun Heo #include <asm/cacheflush.h> 75e0100983STejun Heo #include <asm/sections.h> 76fbf59bc9STejun Heo #include <asm/tlbflush.h> 773b034b0dSVivek Goyal #include <asm/io.h> 78fbf59bc9STejun Heo 79fbf59bc9STejun Heo #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 80fbf59bc9STejun Heo #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 819c824b6aSTejun Heo #define PCPU_ATOMIC_MAP_MARGIN_LOW 32 829c824b6aSTejun Heo #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64 831a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_LOW 2 841a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_HIGH 4 85fbf59bc9STejun Heo 86bbddff05STejun Heo #ifdef CONFIG_SMP 87e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 88e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 89e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 9043cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 9143cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 9243cf38ebSTejun Heo (unsigned long)__per_cpu_start) 93e0100983STejun Heo #endif 94e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 95e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 9643cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 9743cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 9843cf38ebSTejun Heo (unsigned long)__per_cpu_start) 99e0100983STejun Heo #endif 100bbddff05STejun Heo #else /* CONFIG_SMP */ 101bbddff05STejun Heo /* on UP, it's always identity mapped */ 102bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 103bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 104bbddff05STejun Heo #endif /* CONFIG_SMP */ 105e0100983STejun Heo 106fbf59bc9STejun Heo struct pcpu_chunk { 107fbf59bc9STejun Heo struct list_head list; /* linked to pcpu_slot lists */ 108fbf59bc9STejun Heo int free_size; /* free bytes in the chunk */ 109fbf59bc9STejun Heo int contig_hint; /* max contiguous size hint */ 110bba174f5STejun Heo void *base_addr; /* base address of this chunk */ 1119c824b6aSTejun Heo 112723ad1d9SAl Viro int map_used; /* # of map entries used before the sentry */ 113fbf59bc9STejun Heo int map_alloc; /* # of map entries allocated */ 114fbf59bc9STejun Heo int *map; /* allocation map */ 1154f996e23STejun Heo struct list_head map_extend_list;/* on pcpu_map_extend_chunks */ 1169c824b6aSTejun Heo 11788999a89STejun Heo void *data; /* chunk data */ 1183d331ad7SAl Viro int first_free; /* no free below this */ 1198d408b4bSTejun Heo bool immutable; /* no [de]population allowed */ 120b539b87fSTejun Heo int nr_populated; /* # of populated pages */ 121ce3141a2STejun Heo unsigned long populated[]; /* populated bitmap */ 122fbf59bc9STejun Heo }; 123fbf59bc9STejun Heo 12440150d37STejun Heo static int pcpu_unit_pages __read_mostly; 12540150d37STejun Heo static int pcpu_unit_size __read_mostly; 1262f39e637STejun Heo static int pcpu_nr_units __read_mostly; 1276563297cSTejun Heo static int pcpu_atom_size __read_mostly; 12840150d37STejun Heo static int pcpu_nr_slots __read_mostly; 12940150d37STejun Heo static size_t pcpu_chunk_struct_size __read_mostly; 130fbf59bc9STejun Heo 131a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 132a855b84cSTejun Heo static unsigned int pcpu_low_unit_cpu __read_mostly; 133a855b84cSTejun Heo static unsigned int pcpu_high_unit_cpu __read_mostly; 1342f39e637STejun Heo 135fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 13640150d37STejun Heo void *pcpu_base_addr __read_mostly; 137fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 138fbf59bc9STejun Heo 139fb435d52STejun Heo static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 140fb435d52STejun Heo const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 1412f39e637STejun Heo 1426563297cSTejun Heo /* group information, used for vm allocation */ 1436563297cSTejun Heo static int pcpu_nr_groups __read_mostly; 1446563297cSTejun Heo static const unsigned long *pcpu_group_offsets __read_mostly; 1456563297cSTejun Heo static const size_t *pcpu_group_sizes __read_mostly; 1466563297cSTejun Heo 147ae9e6bc9STejun Heo /* 148ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 149ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 150ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 151ae9e6bc9STejun Heo */ 152ae9e6bc9STejun Heo static struct pcpu_chunk *pcpu_first_chunk; 153ae9e6bc9STejun Heo 154ae9e6bc9STejun Heo /* 155ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 156ae9e6bc9STejun Heo * chunk and serves it for reserved allocations. The amount of 157ae9e6bc9STejun Heo * reserved offset is in pcpu_reserved_chunk_limit. When reserved 158ae9e6bc9STejun Heo * area doesn't exist, the following variables contain NULL and 0 159ae9e6bc9STejun Heo * respectively. 160ae9e6bc9STejun Heo */ 161edcb4639STejun Heo static struct pcpu_chunk *pcpu_reserved_chunk; 162edcb4639STejun Heo static int pcpu_reserved_chunk_limit; 163edcb4639STejun Heo 164b38d08f3STejun Heo static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 1656710e594STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 166fbf59bc9STejun Heo 16740150d37STejun Heo static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 168fbf59bc9STejun Heo 1694f996e23STejun Heo /* chunks which need their map areas extended, protected by pcpu_lock */ 1704f996e23STejun Heo static LIST_HEAD(pcpu_map_extend_chunks); 1714f996e23STejun Heo 172b539b87fSTejun Heo /* 173b539b87fSTejun Heo * The number of empty populated pages, protected by pcpu_lock. The 174b539b87fSTejun Heo * reserved chunk doesn't contribute to the count. 175b539b87fSTejun Heo */ 176b539b87fSTejun Heo static int pcpu_nr_empty_pop_pages; 177b539b87fSTejun Heo 1781a4d7607STejun Heo /* 1791a4d7607STejun Heo * Balance work is used to populate or destroy chunks asynchronously. We 1801a4d7607STejun Heo * try to keep the number of populated free pages between 1811a4d7607STejun Heo * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 1821a4d7607STejun Heo * empty chunk. 1831a4d7607STejun Heo */ 184fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work); 185fe6bd8c3STejun Heo static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 1861a4d7607STejun Heo static bool pcpu_async_enabled __read_mostly; 1871a4d7607STejun Heo static bool pcpu_atomic_alloc_failed; 1881a4d7607STejun Heo 1891a4d7607STejun Heo static void pcpu_schedule_balance_work(void) 1901a4d7607STejun Heo { 1911a4d7607STejun Heo if (pcpu_async_enabled) 1921a4d7607STejun Heo schedule_work(&pcpu_balance_work); 1931a4d7607STejun Heo } 194a56dbddfSTejun Heo 195020ec653STejun Heo static bool pcpu_addr_in_first_chunk(void *addr) 196020ec653STejun Heo { 197020ec653STejun Heo void *first_start = pcpu_first_chunk->base_addr; 198020ec653STejun Heo 199020ec653STejun Heo return addr >= first_start && addr < first_start + pcpu_unit_size; 200020ec653STejun Heo } 201020ec653STejun Heo 202020ec653STejun Heo static bool pcpu_addr_in_reserved_chunk(void *addr) 203020ec653STejun Heo { 204020ec653STejun Heo void *first_start = pcpu_first_chunk->base_addr; 205020ec653STejun Heo 206020ec653STejun Heo return addr >= first_start && 207020ec653STejun Heo addr < first_start + pcpu_reserved_chunk_limit; 208020ec653STejun Heo } 209020ec653STejun Heo 210d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 211fbf59bc9STejun Heo { 212cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 213fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 214fbf59bc9STejun Heo } 215fbf59bc9STejun Heo 216d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 217d9b55eebSTejun Heo { 218d9b55eebSTejun Heo if (size == pcpu_unit_size) 219d9b55eebSTejun Heo return pcpu_nr_slots - 1; 220d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 221d9b55eebSTejun Heo } 222d9b55eebSTejun Heo 223fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 224fbf59bc9STejun Heo { 225fbf59bc9STejun Heo if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 226fbf59bc9STejun Heo return 0; 227fbf59bc9STejun Heo 228fbf59bc9STejun Heo return pcpu_size_to_slot(chunk->free_size); 229fbf59bc9STejun Heo } 230fbf59bc9STejun Heo 23188999a89STejun Heo /* set the pointer to a chunk in a page struct */ 23288999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 23388999a89STejun Heo { 23488999a89STejun Heo page->index = (unsigned long)pcpu; 23588999a89STejun Heo } 23688999a89STejun Heo 23788999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 23888999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 23988999a89STejun Heo { 24088999a89STejun Heo return (struct pcpu_chunk *)page->index; 24188999a89STejun Heo } 24288999a89STejun Heo 24388999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 244fbf59bc9STejun Heo { 2452f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 246fbf59bc9STejun Heo } 247fbf59bc9STejun Heo 2489983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 249fbf59bc9STejun Heo unsigned int cpu, int page_idx) 250fbf59bc9STejun Heo { 251bba174f5STejun Heo return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 252fb435d52STejun Heo (page_idx << PAGE_SHIFT); 253fbf59bc9STejun Heo } 254fbf59bc9STejun Heo 25588999a89STejun Heo static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, 25688999a89STejun Heo int *rs, int *re, int end) 257ce3141a2STejun Heo { 258ce3141a2STejun Heo *rs = find_next_zero_bit(chunk->populated, end, *rs); 259ce3141a2STejun Heo *re = find_next_bit(chunk->populated, end, *rs + 1); 260ce3141a2STejun Heo } 261ce3141a2STejun Heo 26288999a89STejun Heo static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, 26388999a89STejun Heo int *rs, int *re, int end) 264ce3141a2STejun Heo { 265ce3141a2STejun Heo *rs = find_next_bit(chunk->populated, end, *rs); 266ce3141a2STejun Heo *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 267ce3141a2STejun Heo } 268ce3141a2STejun Heo 269ce3141a2STejun Heo /* 270ce3141a2STejun Heo * (Un)populated page region iterators. Iterate over (un)populated 271b595076aSUwe Kleine-König * page regions between @start and @end in @chunk. @rs and @re should 272ce3141a2STejun Heo * be integer variables and will be set to start and end page index of 273ce3141a2STejun Heo * the current region. 274ce3141a2STejun Heo */ 275ce3141a2STejun Heo #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 276ce3141a2STejun Heo for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 277ce3141a2STejun Heo (rs) < (re); \ 278ce3141a2STejun Heo (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 279ce3141a2STejun Heo 280ce3141a2STejun Heo #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 281ce3141a2STejun Heo for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 282ce3141a2STejun Heo (rs) < (re); \ 283ce3141a2STejun Heo (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 284ce3141a2STejun Heo 285fbf59bc9STejun Heo /** 28690459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 2871880d93bSTejun Heo * @size: bytes to allocate 288fbf59bc9STejun Heo * 2891880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 29090459ce0SBob Liu * kzalloc() is used; otherwise, vzalloc() is used. The returned 2911880d93bSTejun Heo * memory is always zeroed. 292fbf59bc9STejun Heo * 293ccea34b5STejun Heo * CONTEXT: 294ccea34b5STejun Heo * Does GFP_KERNEL allocation. 295ccea34b5STejun Heo * 296fbf59bc9STejun Heo * RETURNS: 2971880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 298fbf59bc9STejun Heo */ 29990459ce0SBob Liu static void *pcpu_mem_zalloc(size_t size) 300fbf59bc9STejun Heo { 301099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 302099a19d9STejun Heo return NULL; 303099a19d9STejun Heo 304fbf59bc9STejun Heo if (size <= PAGE_SIZE) 3051880d93bSTejun Heo return kzalloc(size, GFP_KERNEL); 3067af4c093SJesper Juhl else 3077af4c093SJesper Juhl return vzalloc(size); 3081880d93bSTejun Heo } 309fbf59bc9STejun Heo 3101880d93bSTejun Heo /** 3111880d93bSTejun Heo * pcpu_mem_free - free memory 3121880d93bSTejun Heo * @ptr: memory to free 3131880d93bSTejun Heo * 31490459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 3151880d93bSTejun Heo */ 3161d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 3171880d93bSTejun Heo { 3181d5cfdb0STetsuo Handa kvfree(ptr); 319fbf59bc9STejun Heo } 320fbf59bc9STejun Heo 321fbf59bc9STejun Heo /** 322b539b87fSTejun Heo * pcpu_count_occupied_pages - count the number of pages an area occupies 323b539b87fSTejun Heo * @chunk: chunk of interest 324b539b87fSTejun Heo * @i: index of the area in question 325b539b87fSTejun Heo * 326b539b87fSTejun Heo * Count the number of pages chunk's @i'th area occupies. When the area's 327b539b87fSTejun Heo * start and/or end address isn't aligned to page boundary, the straddled 328b539b87fSTejun Heo * page is included in the count iff the rest of the page is free. 329b539b87fSTejun Heo */ 330b539b87fSTejun Heo static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i) 331b539b87fSTejun Heo { 332b539b87fSTejun Heo int off = chunk->map[i] & ~1; 333b539b87fSTejun Heo int end = chunk->map[i + 1] & ~1; 334b539b87fSTejun Heo 335b539b87fSTejun Heo if (!PAGE_ALIGNED(off) && i > 0) { 336b539b87fSTejun Heo int prev = chunk->map[i - 1]; 337b539b87fSTejun Heo 338b539b87fSTejun Heo if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE)) 339b539b87fSTejun Heo off = round_down(off, PAGE_SIZE); 340b539b87fSTejun Heo } 341b539b87fSTejun Heo 342b539b87fSTejun Heo if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) { 343b539b87fSTejun Heo int next = chunk->map[i + 1]; 344b539b87fSTejun Heo int nend = chunk->map[i + 2] & ~1; 345b539b87fSTejun Heo 346b539b87fSTejun Heo if (!(next & 1) && nend >= round_up(end, PAGE_SIZE)) 347b539b87fSTejun Heo end = round_up(end, PAGE_SIZE); 348b539b87fSTejun Heo } 349b539b87fSTejun Heo 350b539b87fSTejun Heo return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0); 351b539b87fSTejun Heo } 352b539b87fSTejun Heo 353b539b87fSTejun Heo /** 354fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 355fbf59bc9STejun Heo * @chunk: chunk of interest 356fbf59bc9STejun Heo * @oslot: the previous slot it was on 357fbf59bc9STejun Heo * 358fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 359fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 360edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 361edcb4639STejun Heo * chunk slots. 362ccea34b5STejun Heo * 363ccea34b5STejun Heo * CONTEXT: 364ccea34b5STejun Heo * pcpu_lock. 365fbf59bc9STejun Heo */ 366fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 367fbf59bc9STejun Heo { 368fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 369fbf59bc9STejun Heo 370edcb4639STejun Heo if (chunk != pcpu_reserved_chunk && oslot != nslot) { 371fbf59bc9STejun Heo if (oslot < nslot) 372fbf59bc9STejun Heo list_move(&chunk->list, &pcpu_slot[nslot]); 373fbf59bc9STejun Heo else 374fbf59bc9STejun Heo list_move_tail(&chunk->list, &pcpu_slot[nslot]); 375fbf59bc9STejun Heo } 376fbf59bc9STejun Heo } 377fbf59bc9STejun Heo 378fbf59bc9STejun Heo /** 379833af842STejun Heo * pcpu_need_to_extend - determine whether chunk area map needs to be extended 380833af842STejun Heo * @chunk: chunk of interest 3819c824b6aSTejun Heo * @is_atomic: the allocation context 3829f7dcf22STejun Heo * 3839c824b6aSTejun Heo * Determine whether area map of @chunk needs to be extended. If 3849c824b6aSTejun Heo * @is_atomic, only the amount necessary for a new allocation is 3859c824b6aSTejun Heo * considered; however, async extension is scheduled if the left amount is 3869c824b6aSTejun Heo * low. If !@is_atomic, it aims for more empty space. Combined, this 3879c824b6aSTejun Heo * ensures that the map is likely to have enough available space to 3889c824b6aSTejun Heo * accomodate atomic allocations which can't extend maps directly. 3899f7dcf22STejun Heo * 390ccea34b5STejun Heo * CONTEXT: 391833af842STejun Heo * pcpu_lock. 392ccea34b5STejun Heo * 3939f7dcf22STejun Heo * RETURNS: 394833af842STejun Heo * New target map allocation length if extension is necessary, 0 395833af842STejun Heo * otherwise. 3969f7dcf22STejun Heo */ 3979c824b6aSTejun Heo static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic) 3989f7dcf22STejun Heo { 3999c824b6aSTejun Heo int margin, new_alloc; 4009f7dcf22STejun Heo 4014f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 4024f996e23STejun Heo 4039c824b6aSTejun Heo if (is_atomic) { 4049c824b6aSTejun Heo margin = 3; 4059c824b6aSTejun Heo 4069c824b6aSTejun Heo if (chunk->map_alloc < 4074f996e23STejun Heo chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) { 4084f996e23STejun Heo if (list_empty(&chunk->map_extend_list)) { 4094f996e23STejun Heo list_add_tail(&chunk->map_extend_list, 4104f996e23STejun Heo &pcpu_map_extend_chunks); 4114f996e23STejun Heo pcpu_schedule_balance_work(); 4124f996e23STejun Heo } 4134f996e23STejun Heo } 4149c824b6aSTejun Heo } else { 4159c824b6aSTejun Heo margin = PCPU_ATOMIC_MAP_MARGIN_HIGH; 4169c824b6aSTejun Heo } 4179c824b6aSTejun Heo 4189c824b6aSTejun Heo if (chunk->map_alloc >= chunk->map_used + margin) 4199f7dcf22STejun Heo return 0; 4209f7dcf22STejun Heo 4219f7dcf22STejun Heo new_alloc = PCPU_DFL_MAP_ALLOC; 4229c824b6aSTejun Heo while (new_alloc < chunk->map_used + margin) 4239f7dcf22STejun Heo new_alloc *= 2; 4249f7dcf22STejun Heo 425833af842STejun Heo return new_alloc; 426ccea34b5STejun Heo } 427ccea34b5STejun Heo 428833af842STejun Heo /** 429833af842STejun Heo * pcpu_extend_area_map - extend area map of a chunk 430833af842STejun Heo * @chunk: chunk of interest 431833af842STejun Heo * @new_alloc: new target allocation length of the area map 432833af842STejun Heo * 433833af842STejun Heo * Extend area map of @chunk to have @new_alloc entries. 434833af842STejun Heo * 435833af842STejun Heo * CONTEXT: 436833af842STejun Heo * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. 437833af842STejun Heo * 438833af842STejun Heo * RETURNS: 439833af842STejun Heo * 0 on success, -errno on failure. 440ccea34b5STejun Heo */ 441833af842STejun Heo static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) 442833af842STejun Heo { 443833af842STejun Heo int *old = NULL, *new = NULL; 444833af842STejun Heo size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); 445833af842STejun Heo unsigned long flags; 4469f7dcf22STejun Heo 4476710e594STejun Heo lockdep_assert_held(&pcpu_alloc_mutex); 4486710e594STejun Heo 44990459ce0SBob Liu new = pcpu_mem_zalloc(new_size); 450833af842STejun Heo if (!new) 451833af842STejun Heo return -ENOMEM; 452833af842STejun Heo 453833af842STejun Heo /* acquire pcpu_lock and switch to new area map */ 454833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 455833af842STejun Heo 456833af842STejun Heo if (new_alloc <= chunk->map_alloc) 457833af842STejun Heo goto out_unlock; 458833af842STejun Heo 459833af842STejun Heo old_size = chunk->map_alloc * sizeof(chunk->map[0]); 460a002d148SHuang Shijie old = chunk->map; 461a002d148SHuang Shijie 462a002d148SHuang Shijie memcpy(new, old, old_size); 4639f7dcf22STejun Heo 4649f7dcf22STejun Heo chunk->map_alloc = new_alloc; 4659f7dcf22STejun Heo chunk->map = new; 466833af842STejun Heo new = NULL; 467833af842STejun Heo 468833af842STejun Heo out_unlock: 469833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 470833af842STejun Heo 471833af842STejun Heo /* 472833af842STejun Heo * pcpu_mem_free() might end up calling vfree() which uses 473833af842STejun Heo * IRQ-unsafe lock and thus can't be called under pcpu_lock. 474833af842STejun Heo */ 4751d5cfdb0STetsuo Handa pcpu_mem_free(old); 4761d5cfdb0STetsuo Handa pcpu_mem_free(new); 477833af842STejun Heo 4789f7dcf22STejun Heo return 0; 4799f7dcf22STejun Heo } 4809f7dcf22STejun Heo 4819f7dcf22STejun Heo /** 482a16037c8STejun Heo * pcpu_fit_in_area - try to fit the requested allocation in a candidate area 483a16037c8STejun Heo * @chunk: chunk the candidate area belongs to 484a16037c8STejun Heo * @off: the offset to the start of the candidate area 485a16037c8STejun Heo * @this_size: the size of the candidate area 486a16037c8STejun Heo * @size: the size of the target allocation 487a16037c8STejun Heo * @align: the alignment of the target allocation 488a16037c8STejun Heo * @pop_only: only allocate from already populated region 489a16037c8STejun Heo * 490a16037c8STejun Heo * We're trying to allocate @size bytes aligned at @align. @chunk's area 491a16037c8STejun Heo * at @off sized @this_size is a candidate. This function determines 492a16037c8STejun Heo * whether the target allocation fits in the candidate area and returns the 493a16037c8STejun Heo * number of bytes to pad after @off. If the target area doesn't fit, -1 494a16037c8STejun Heo * is returned. 495a16037c8STejun Heo * 496a16037c8STejun Heo * If @pop_only is %true, this function only considers the already 497a16037c8STejun Heo * populated part of the candidate area. 498a16037c8STejun Heo */ 499a16037c8STejun Heo static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size, 500a16037c8STejun Heo int size, int align, bool pop_only) 501a16037c8STejun Heo { 502a16037c8STejun Heo int cand_off = off; 503a16037c8STejun Heo 504a16037c8STejun Heo while (true) { 505a16037c8STejun Heo int head = ALIGN(cand_off, align) - off; 506a16037c8STejun Heo int page_start, page_end, rs, re; 507a16037c8STejun Heo 508a16037c8STejun Heo if (this_size < head + size) 509a16037c8STejun Heo return -1; 510a16037c8STejun Heo 511a16037c8STejun Heo if (!pop_only) 512a16037c8STejun Heo return head; 513a16037c8STejun Heo 514a16037c8STejun Heo /* 515a16037c8STejun Heo * If the first unpopulated page is beyond the end of the 516a16037c8STejun Heo * allocation, the whole allocation is populated; 517a16037c8STejun Heo * otherwise, retry from the end of the unpopulated area. 518a16037c8STejun Heo */ 519a16037c8STejun Heo page_start = PFN_DOWN(head + off); 520a16037c8STejun Heo page_end = PFN_UP(head + off + size); 521a16037c8STejun Heo 522a16037c8STejun Heo rs = page_start; 523a16037c8STejun Heo pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size)); 524a16037c8STejun Heo if (rs >= page_end) 525a16037c8STejun Heo return head; 526a16037c8STejun Heo cand_off = re * PAGE_SIZE; 527a16037c8STejun Heo } 528a16037c8STejun Heo } 529a16037c8STejun Heo 530a16037c8STejun Heo /** 531fbf59bc9STejun Heo * pcpu_alloc_area - allocate area from a pcpu_chunk 532fbf59bc9STejun Heo * @chunk: chunk of interest 533cae3aeb8STejun Heo * @size: wanted size in bytes 534fbf59bc9STejun Heo * @align: wanted align 535a16037c8STejun Heo * @pop_only: allocate only from the populated area 536b539b87fSTejun Heo * @occ_pages_p: out param for the number of pages the area occupies 537fbf59bc9STejun Heo * 538fbf59bc9STejun Heo * Try to allocate @size bytes area aligned at @align from @chunk. 539fbf59bc9STejun Heo * Note that this function only allocates the offset. It doesn't 540fbf59bc9STejun Heo * populate or map the area. 541fbf59bc9STejun Heo * 5429f7dcf22STejun Heo * @chunk->map must have at least two free slots. 5439f7dcf22STejun Heo * 544ccea34b5STejun Heo * CONTEXT: 545ccea34b5STejun Heo * pcpu_lock. 546ccea34b5STejun Heo * 547fbf59bc9STejun Heo * RETURNS: 5489f7dcf22STejun Heo * Allocated offset in @chunk on success, -1 if no matching area is 5499f7dcf22STejun Heo * found. 550fbf59bc9STejun Heo */ 551a16037c8STejun Heo static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align, 552b539b87fSTejun Heo bool pop_only, int *occ_pages_p) 553fbf59bc9STejun Heo { 554fbf59bc9STejun Heo int oslot = pcpu_chunk_slot(chunk); 555fbf59bc9STejun Heo int max_contig = 0; 556fbf59bc9STejun Heo int i, off; 5573d331ad7SAl Viro bool seen_free = false; 558723ad1d9SAl Viro int *p; 559fbf59bc9STejun Heo 5603d331ad7SAl Viro for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) { 561fbf59bc9STejun Heo int head, tail; 562723ad1d9SAl Viro int this_size; 563723ad1d9SAl Viro 564723ad1d9SAl Viro off = *p; 565723ad1d9SAl Viro if (off & 1) 566723ad1d9SAl Viro continue; 567fbf59bc9STejun Heo 568723ad1d9SAl Viro this_size = (p[1] & ~1) - off; 569a16037c8STejun Heo 570a16037c8STejun Heo head = pcpu_fit_in_area(chunk, off, this_size, size, align, 571a16037c8STejun Heo pop_only); 572a16037c8STejun Heo if (head < 0) { 5733d331ad7SAl Viro if (!seen_free) { 5743d331ad7SAl Viro chunk->first_free = i; 5753d331ad7SAl Viro seen_free = true; 5763d331ad7SAl Viro } 577723ad1d9SAl Viro max_contig = max(this_size, max_contig); 578fbf59bc9STejun Heo continue; 579fbf59bc9STejun Heo } 580fbf59bc9STejun Heo 581fbf59bc9STejun Heo /* 582fbf59bc9STejun Heo * If head is small or the previous block is free, 583fbf59bc9STejun Heo * merge'em. Note that 'small' is defined as smaller 584fbf59bc9STejun Heo * than sizeof(int), which is very small but isn't too 585fbf59bc9STejun Heo * uncommon for percpu allocations. 586fbf59bc9STejun Heo */ 587723ad1d9SAl Viro if (head && (head < sizeof(int) || !(p[-1] & 1))) { 58821ddfd38SJianyu Zhan *p = off += head; 589723ad1d9SAl Viro if (p[-1] & 1) 590fbf59bc9STejun Heo chunk->free_size -= head; 59121ddfd38SJianyu Zhan else 59221ddfd38SJianyu Zhan max_contig = max(*p - p[-1], max_contig); 593723ad1d9SAl Viro this_size -= head; 594fbf59bc9STejun Heo head = 0; 595fbf59bc9STejun Heo } 596fbf59bc9STejun Heo 597fbf59bc9STejun Heo /* if tail is small, just keep it around */ 598723ad1d9SAl Viro tail = this_size - head - size; 599723ad1d9SAl Viro if (tail < sizeof(int)) { 600fbf59bc9STejun Heo tail = 0; 601723ad1d9SAl Viro size = this_size - head; 602723ad1d9SAl Viro } 603fbf59bc9STejun Heo 604fbf59bc9STejun Heo /* split if warranted */ 605fbf59bc9STejun Heo if (head || tail) { 606706c16f2SAl Viro int nr_extra = !!head + !!tail; 607706c16f2SAl Viro 608706c16f2SAl Viro /* insert new subblocks */ 609723ad1d9SAl Viro memmove(p + nr_extra + 1, p + 1, 610706c16f2SAl Viro sizeof(chunk->map[0]) * (chunk->map_used - i)); 611706c16f2SAl Viro chunk->map_used += nr_extra; 612706c16f2SAl Viro 613fbf59bc9STejun Heo if (head) { 6143d331ad7SAl Viro if (!seen_free) { 6153d331ad7SAl Viro chunk->first_free = i; 6163d331ad7SAl Viro seen_free = true; 6173d331ad7SAl Viro } 618723ad1d9SAl Viro *++p = off += head; 619723ad1d9SAl Viro ++i; 620706c16f2SAl Viro max_contig = max(head, max_contig); 621fbf59bc9STejun Heo } 622706c16f2SAl Viro if (tail) { 623723ad1d9SAl Viro p[1] = off + size; 624706c16f2SAl Viro max_contig = max(tail, max_contig); 625706c16f2SAl Viro } 626fbf59bc9STejun Heo } 627fbf59bc9STejun Heo 6283d331ad7SAl Viro if (!seen_free) 6293d331ad7SAl Viro chunk->first_free = i + 1; 6303d331ad7SAl Viro 631fbf59bc9STejun Heo /* update hint and mark allocated */ 632723ad1d9SAl Viro if (i + 1 == chunk->map_used) 633fbf59bc9STejun Heo chunk->contig_hint = max_contig; /* fully scanned */ 634fbf59bc9STejun Heo else 635fbf59bc9STejun Heo chunk->contig_hint = max(chunk->contig_hint, 636fbf59bc9STejun Heo max_contig); 637fbf59bc9STejun Heo 638723ad1d9SAl Viro chunk->free_size -= size; 639723ad1d9SAl Viro *p |= 1; 640fbf59bc9STejun Heo 641b539b87fSTejun Heo *occ_pages_p = pcpu_count_occupied_pages(chunk, i); 642fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 643fbf59bc9STejun Heo return off; 644fbf59bc9STejun Heo } 645fbf59bc9STejun Heo 646fbf59bc9STejun Heo chunk->contig_hint = max_contig; /* fully scanned */ 647fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 648fbf59bc9STejun Heo 6499f7dcf22STejun Heo /* tell the upper layer that this chunk has no matching area */ 6509f7dcf22STejun Heo return -1; 651fbf59bc9STejun Heo } 652fbf59bc9STejun Heo 653fbf59bc9STejun Heo /** 654fbf59bc9STejun Heo * pcpu_free_area - free area to a pcpu_chunk 655fbf59bc9STejun Heo * @chunk: chunk of interest 656fbf59bc9STejun Heo * @freeme: offset of area to free 657b539b87fSTejun Heo * @occ_pages_p: out param for the number of pages the area occupies 658fbf59bc9STejun Heo * 659fbf59bc9STejun Heo * Free area starting from @freeme to @chunk. Note that this function 660fbf59bc9STejun Heo * only modifies the allocation map. It doesn't depopulate or unmap 661fbf59bc9STejun Heo * the area. 662ccea34b5STejun Heo * 663ccea34b5STejun Heo * CONTEXT: 664ccea34b5STejun Heo * pcpu_lock. 665fbf59bc9STejun Heo */ 666b539b87fSTejun Heo static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme, 667b539b87fSTejun Heo int *occ_pages_p) 668fbf59bc9STejun Heo { 669fbf59bc9STejun Heo int oslot = pcpu_chunk_slot(chunk); 670723ad1d9SAl Viro int off = 0; 671723ad1d9SAl Viro unsigned i, j; 672723ad1d9SAl Viro int to_free = 0; 673723ad1d9SAl Viro int *p; 674fbf59bc9STejun Heo 675723ad1d9SAl Viro freeme |= 1; /* we are searching for <given offset, in use> pair */ 676723ad1d9SAl Viro 677723ad1d9SAl Viro i = 0; 678723ad1d9SAl Viro j = chunk->map_used; 679723ad1d9SAl Viro while (i != j) { 680723ad1d9SAl Viro unsigned k = (i + j) / 2; 681723ad1d9SAl Viro off = chunk->map[k]; 682723ad1d9SAl Viro if (off < freeme) 683723ad1d9SAl Viro i = k + 1; 684723ad1d9SAl Viro else if (off > freeme) 685723ad1d9SAl Viro j = k; 686723ad1d9SAl Viro else 687723ad1d9SAl Viro i = j = k; 688723ad1d9SAl Viro } 689fbf59bc9STejun Heo BUG_ON(off != freeme); 690fbf59bc9STejun Heo 6913d331ad7SAl Viro if (i < chunk->first_free) 6923d331ad7SAl Viro chunk->first_free = i; 6933d331ad7SAl Viro 694723ad1d9SAl Viro p = chunk->map + i; 695723ad1d9SAl Viro *p = off &= ~1; 696723ad1d9SAl Viro chunk->free_size += (p[1] & ~1) - off; 697fbf59bc9STejun Heo 698b539b87fSTejun Heo *occ_pages_p = pcpu_count_occupied_pages(chunk, i); 699b539b87fSTejun Heo 700fbf59bc9STejun Heo /* merge with next? */ 701723ad1d9SAl Viro if (!(p[1] & 1)) 702723ad1d9SAl Viro to_free++; 703723ad1d9SAl Viro /* merge with previous? */ 704723ad1d9SAl Viro if (i > 0 && !(p[-1] & 1)) { 705723ad1d9SAl Viro to_free++; 706723ad1d9SAl Viro i--; 707723ad1d9SAl Viro p--; 708723ad1d9SAl Viro } 709723ad1d9SAl Viro if (to_free) { 710723ad1d9SAl Viro chunk->map_used -= to_free; 711723ad1d9SAl Viro memmove(p + 1, p + 1 + to_free, 712723ad1d9SAl Viro (chunk->map_used - i) * sizeof(chunk->map[0])); 713fbf59bc9STejun Heo } 714fbf59bc9STejun Heo 715723ad1d9SAl Viro chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint); 716fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 717fbf59bc9STejun Heo } 718fbf59bc9STejun Heo 7196081089fSTejun Heo static struct pcpu_chunk *pcpu_alloc_chunk(void) 7206081089fSTejun Heo { 7216081089fSTejun Heo struct pcpu_chunk *chunk; 7226081089fSTejun Heo 72390459ce0SBob Liu chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); 7246081089fSTejun Heo if (!chunk) 7256081089fSTejun Heo return NULL; 7266081089fSTejun Heo 72790459ce0SBob Liu chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * 72890459ce0SBob Liu sizeof(chunk->map[0])); 7296081089fSTejun Heo if (!chunk->map) { 7301d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 7316081089fSTejun Heo return NULL; 7326081089fSTejun Heo } 7336081089fSTejun Heo 7346081089fSTejun Heo chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 735723ad1d9SAl Viro chunk->map[0] = 0; 736723ad1d9SAl Viro chunk->map[1] = pcpu_unit_size | 1; 737723ad1d9SAl Viro chunk->map_used = 1; 7386081089fSTejun Heo 7396081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 7404f996e23STejun Heo INIT_LIST_HEAD(&chunk->map_extend_list); 7416081089fSTejun Heo chunk->free_size = pcpu_unit_size; 7426081089fSTejun Heo chunk->contig_hint = pcpu_unit_size; 7436081089fSTejun Heo 7446081089fSTejun Heo return chunk; 7456081089fSTejun Heo } 7466081089fSTejun Heo 7476081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 7486081089fSTejun Heo { 7496081089fSTejun Heo if (!chunk) 7506081089fSTejun Heo return; 7511d5cfdb0STetsuo Handa pcpu_mem_free(chunk->map); 7521d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 7536081089fSTejun Heo } 7546081089fSTejun Heo 755b539b87fSTejun Heo /** 756b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 757b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 758b539b87fSTejun Heo * @page_start: the start page 759b539b87fSTejun Heo * @page_end: the end page 760b539b87fSTejun Heo * 761b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 762b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 763b539b87fSTejun Heo * successful population. 764b539b87fSTejun Heo */ 765b539b87fSTejun Heo static void pcpu_chunk_populated(struct pcpu_chunk *chunk, 766b539b87fSTejun Heo int page_start, int page_end) 767b539b87fSTejun Heo { 768b539b87fSTejun Heo int nr = page_end - page_start; 769b539b87fSTejun Heo 770b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 771b539b87fSTejun Heo 772b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 773b539b87fSTejun Heo chunk->nr_populated += nr; 774b539b87fSTejun Heo pcpu_nr_empty_pop_pages += nr; 775b539b87fSTejun Heo } 776b539b87fSTejun Heo 777b539b87fSTejun Heo /** 778b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 779b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 780b539b87fSTejun Heo * @page_start: the start page 781b539b87fSTejun Heo * @page_end: the end page 782b539b87fSTejun Heo * 783b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 784b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 785b539b87fSTejun Heo * each successful depopulation. 786b539b87fSTejun Heo */ 787b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 788b539b87fSTejun Heo int page_start, int page_end) 789b539b87fSTejun Heo { 790b539b87fSTejun Heo int nr = page_end - page_start; 791b539b87fSTejun Heo 792b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 793b539b87fSTejun Heo 794b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 795b539b87fSTejun Heo chunk->nr_populated -= nr; 796b539b87fSTejun Heo pcpu_nr_empty_pop_pages -= nr; 797b539b87fSTejun Heo } 798b539b87fSTejun Heo 799fbf59bc9STejun Heo /* 8009f645532STejun Heo * Chunk management implementation. 801fbf59bc9STejun Heo * 8029f645532STejun Heo * To allow different implementations, chunk alloc/free and 8039f645532STejun Heo * [de]population are implemented in a separate file which is pulled 8049f645532STejun Heo * into this file and compiled together. The following functions 8059f645532STejun Heo * should be implemented. 806ccea34b5STejun Heo * 8079f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 8089f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 8099f645532STejun Heo * pcpu_create_chunk - create a new chunk 8109f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 8119f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 8129f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 813fbf59bc9STejun Heo */ 8149f645532STejun Heo static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 8159f645532STejun Heo static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 8169f645532STejun Heo static struct pcpu_chunk *pcpu_create_chunk(void); 8179f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 8189f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 8199f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 820fbf59bc9STejun Heo 821b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 822b0c9778bSTejun Heo #include "percpu-km.c" 823b0c9778bSTejun Heo #else 8249f645532STejun Heo #include "percpu-vm.c" 825b0c9778bSTejun Heo #endif 826fbf59bc9STejun Heo 827fbf59bc9STejun Heo /** 82888999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 82988999a89STejun Heo * @addr: address for which the chunk needs to be determined. 83088999a89STejun Heo * 83188999a89STejun Heo * RETURNS: 83288999a89STejun Heo * The address of the found chunk. 83388999a89STejun Heo */ 83488999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 83588999a89STejun Heo { 83688999a89STejun Heo /* is it in the first chunk? */ 83788999a89STejun Heo if (pcpu_addr_in_first_chunk(addr)) { 83888999a89STejun Heo /* is it in the reserved area? */ 83988999a89STejun Heo if (pcpu_addr_in_reserved_chunk(addr)) 84088999a89STejun Heo return pcpu_reserved_chunk; 84188999a89STejun Heo return pcpu_first_chunk; 84288999a89STejun Heo } 84388999a89STejun Heo 84488999a89STejun Heo /* 84588999a89STejun Heo * The address is relative to unit0 which might be unused and 84688999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 84788999a89STejun Heo * current processor before looking it up in the vmalloc 84888999a89STejun Heo * space. Note that any possible cpu id can be used here, so 84988999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 85088999a89STejun Heo */ 85188999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 8529f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 85388999a89STejun Heo } 85488999a89STejun Heo 85588999a89STejun Heo /** 856edcb4639STejun Heo * pcpu_alloc - the percpu allocator 857cae3aeb8STejun Heo * @size: size of area to allocate in bytes 858fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 859edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 8605835d96eSTejun Heo * @gfp: allocation flags 861fbf59bc9STejun Heo * 8625835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 8635835d96eSTejun Heo * contain %GFP_KERNEL, the allocation is atomic. 864fbf59bc9STejun Heo * 865fbf59bc9STejun Heo * RETURNS: 866fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 867fbf59bc9STejun Heo */ 8685835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 8695835d96eSTejun Heo gfp_t gfp) 870fbf59bc9STejun Heo { 871f2badb0cSTejun Heo static int warn_limit = 10; 872fbf59bc9STejun Heo struct pcpu_chunk *chunk; 873f2badb0cSTejun Heo const char *err; 8746ae833c7STejun Heo bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 875b539b87fSTejun Heo int occ_pages = 0; 876b38d08f3STejun Heo int slot, off, new_alloc, cpu, ret; 877403a91b1SJiri Kosina unsigned long flags; 878f528f0b8SCatalin Marinas void __percpu *ptr; 879fbf59bc9STejun Heo 880723ad1d9SAl Viro /* 881723ad1d9SAl Viro * We want the lowest bit of offset available for in-use/free 8822f69fa82SViro * indicator, so force >= 16bit alignment and make size even. 883723ad1d9SAl Viro */ 884723ad1d9SAl Viro if (unlikely(align < 2)) 885723ad1d9SAl Viro align = 2; 886723ad1d9SAl Viro 887fb009e3aSChristoph Lameter size = ALIGN(size, 2); 8882f69fa82SViro 8898d408b4bSTejun Heo if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 890756a025fSJoe Perches WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", 891756a025fSJoe Perches size, align); 892fbf59bc9STejun Heo return NULL; 893fbf59bc9STejun Heo } 894fbf59bc9STejun Heo 8956710e594STejun Heo if (!is_atomic) 8966710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 8976710e594STejun Heo 898403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 899fbf59bc9STejun Heo 900edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 901edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 902edcb4639STejun Heo chunk = pcpu_reserved_chunk; 903833af842STejun Heo 904833af842STejun Heo if (size > chunk->contig_hint) { 905833af842STejun Heo err = "alloc from reserved chunk failed"; 906ccea34b5STejun Heo goto fail_unlock; 907f2badb0cSTejun Heo } 908833af842STejun Heo 9099c824b6aSTejun Heo while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) { 910833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 9115835d96eSTejun Heo if (is_atomic || 9125835d96eSTejun Heo pcpu_extend_area_map(chunk, new_alloc) < 0) { 913833af842STejun Heo err = "failed to extend area map of reserved chunk"; 914b38d08f3STejun Heo goto fail; 915833af842STejun Heo } 916833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 917833af842STejun Heo } 918833af842STejun Heo 919b539b87fSTejun Heo off = pcpu_alloc_area(chunk, size, align, is_atomic, 920b539b87fSTejun Heo &occ_pages); 921edcb4639STejun Heo if (off >= 0) 922edcb4639STejun Heo goto area_found; 923833af842STejun Heo 924f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 925ccea34b5STejun Heo goto fail_unlock; 926edcb4639STejun Heo } 927edcb4639STejun Heo 928ccea34b5STejun Heo restart: 929edcb4639STejun Heo /* search through normal chunks */ 930fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 931fbf59bc9STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 932fbf59bc9STejun Heo if (size > chunk->contig_hint) 933fbf59bc9STejun Heo continue; 934ccea34b5STejun Heo 9359c824b6aSTejun Heo new_alloc = pcpu_need_to_extend(chunk, is_atomic); 936833af842STejun Heo if (new_alloc) { 9375835d96eSTejun Heo if (is_atomic) 9385835d96eSTejun Heo continue; 939833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 940833af842STejun Heo if (pcpu_extend_area_map(chunk, 941833af842STejun Heo new_alloc) < 0) { 942f2badb0cSTejun Heo err = "failed to extend area map"; 943b38d08f3STejun Heo goto fail; 944833af842STejun Heo } 945833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 946833af842STejun Heo /* 947833af842STejun Heo * pcpu_lock has been dropped, need to 948833af842STejun Heo * restart cpu_slot list walking. 949833af842STejun Heo */ 950833af842STejun Heo goto restart; 951ccea34b5STejun Heo } 952ccea34b5STejun Heo 953b539b87fSTejun Heo off = pcpu_alloc_area(chunk, size, align, is_atomic, 954b539b87fSTejun Heo &occ_pages); 955fbf59bc9STejun Heo if (off >= 0) 956fbf59bc9STejun Heo goto area_found; 957fbf59bc9STejun Heo } 958fbf59bc9STejun Heo } 959fbf59bc9STejun Heo 960403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 961ccea34b5STejun Heo 962b38d08f3STejun Heo /* 963b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 964b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 965b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 966b38d08f3STejun Heo */ 9675835d96eSTejun Heo if (is_atomic) 9685835d96eSTejun Heo goto fail; 9695835d96eSTejun Heo 970b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 9716081089fSTejun Heo chunk = pcpu_create_chunk(); 972f2badb0cSTejun Heo if (!chunk) { 973f2badb0cSTejun Heo err = "failed to allocate new chunk"; 974b38d08f3STejun Heo goto fail; 975f2badb0cSTejun Heo } 976ccea34b5STejun Heo 977403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 978fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 979b38d08f3STejun Heo } else { 980b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 981b38d08f3STejun Heo } 982b38d08f3STejun Heo 983ccea34b5STejun Heo goto restart; 984fbf59bc9STejun Heo 985fbf59bc9STejun Heo area_found: 986403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 987ccea34b5STejun Heo 988dca49645STejun Heo /* populate if not all pages are already there */ 9895835d96eSTejun Heo if (!is_atomic) { 990e04d3208STejun Heo int page_start, page_end, rs, re; 991e04d3208STejun Heo 992dca49645STejun Heo page_start = PFN_DOWN(off); 993dca49645STejun Heo page_end = PFN_UP(off + size); 994dca49645STejun Heo 995a93ace48STejun Heo pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { 996dca49645STejun Heo WARN_ON(chunk->immutable); 997dca49645STejun Heo 998b38d08f3STejun Heo ret = pcpu_populate_chunk(chunk, rs, re); 999b38d08f3STejun Heo 1000403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1001b38d08f3STejun Heo if (ret) { 1002b539b87fSTejun Heo pcpu_free_area(chunk, off, &occ_pages); 1003f2badb0cSTejun Heo err = "failed to populate"; 1004ccea34b5STejun Heo goto fail_unlock; 1005fbf59bc9STejun Heo } 1006b539b87fSTejun Heo pcpu_chunk_populated(chunk, rs, re); 1007b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1008dca49645STejun Heo } 1009dca49645STejun Heo 1010ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1011e04d3208STejun Heo } 1012ccea34b5STejun Heo 1013b539b87fSTejun Heo if (chunk != pcpu_reserved_chunk) 1014b539b87fSTejun Heo pcpu_nr_empty_pop_pages -= occ_pages; 1015b539b87fSTejun Heo 10161a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 10171a4d7607STejun Heo pcpu_schedule_balance_work(); 10181a4d7607STejun Heo 1019dca49645STejun Heo /* clear the areas and return address relative to base address */ 1020dca49645STejun Heo for_each_possible_cpu(cpu) 1021dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1022dca49645STejun Heo 1023f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 10248a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1025f528f0b8SCatalin Marinas return ptr; 1026ccea34b5STejun Heo 1027ccea34b5STejun Heo fail_unlock: 1028403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1029b38d08f3STejun Heo fail: 10305835d96eSTejun Heo if (!is_atomic && warn_limit) { 1031870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 10325835d96eSTejun Heo size, align, is_atomic, err); 1033f2badb0cSTejun Heo dump_stack(); 1034f2badb0cSTejun Heo if (!--warn_limit) 1035870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1036f2badb0cSTejun Heo } 10371a4d7607STejun Heo if (is_atomic) { 10381a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 10391a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 10401a4d7607STejun Heo pcpu_schedule_balance_work(); 10416710e594STejun Heo } else { 10426710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 10431a4d7607STejun Heo } 1044ccea34b5STejun Heo return NULL; 1045fbf59bc9STejun Heo } 1046edcb4639STejun Heo 1047edcb4639STejun Heo /** 10485835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1049edcb4639STejun Heo * @size: size of area to allocate in bytes 1050edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 10515835d96eSTejun Heo * @gfp: allocation flags 1052edcb4639STejun Heo * 10535835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 10545835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 10555835d96eSTejun Heo * be called from any context but is a lot more likely to fail. 1056ccea34b5STejun Heo * 1057edcb4639STejun Heo * RETURNS: 1058edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1059edcb4639STejun Heo */ 10605835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 10615835d96eSTejun Heo { 10625835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 10635835d96eSTejun Heo } 10645835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 10655835d96eSTejun Heo 10665835d96eSTejun Heo /** 10675835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 10685835d96eSTejun Heo * @size: size of area to allocate in bytes 10695835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 10705835d96eSTejun Heo * 10715835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 10725835d96eSTejun Heo */ 107343cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1074edcb4639STejun Heo { 10755835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1076edcb4639STejun Heo } 1077fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1078fbf59bc9STejun Heo 1079edcb4639STejun Heo /** 1080edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1081edcb4639STejun Heo * @size: size of area to allocate in bytes 1082edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1083edcb4639STejun Heo * 10849329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 10859329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 10869329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 10879329ba97STejun Heo * Might trigger writeouts. 1088edcb4639STejun Heo * 1089ccea34b5STejun Heo * CONTEXT: 1090ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1091ccea34b5STejun Heo * 1092edcb4639STejun Heo * RETURNS: 1093edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1094edcb4639STejun Heo */ 109543cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1096edcb4639STejun Heo { 10975835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1098edcb4639STejun Heo } 1099edcb4639STejun Heo 1100a56dbddfSTejun Heo /** 11011a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1102a56dbddfSTejun Heo * @work: unused 1103a56dbddfSTejun Heo * 1104a56dbddfSTejun Heo * Reclaim all fully free chunks except for the first one. 1105a56dbddfSTejun Heo */ 1106fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1107fbf59bc9STejun Heo { 1108fe6bd8c3STejun Heo LIST_HEAD(to_free); 1109fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1110a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 11111a4d7607STejun Heo int slot, nr_to_pop, ret; 1112a56dbddfSTejun Heo 11131a4d7607STejun Heo /* 11141a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 11151a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 11161a4d7607STejun Heo */ 1117ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1118ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1119a56dbddfSTejun Heo 1120fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 11218d408b4bSTejun Heo WARN_ON(chunk->immutable); 1122a56dbddfSTejun Heo 1123a56dbddfSTejun Heo /* spare the first one */ 1124fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1125a56dbddfSTejun Heo continue; 1126a56dbddfSTejun Heo 11274f996e23STejun Heo list_del_init(&chunk->map_extend_list); 1128fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1129a56dbddfSTejun Heo } 1130a56dbddfSTejun Heo 1131ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1132a56dbddfSTejun Heo 1133fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1134a93ace48STejun Heo int rs, re; 1135dca49645STejun Heo 1136a93ace48STejun Heo pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) { 1137a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1138b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1139b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1140b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1141a93ace48STejun Heo } 11426081089fSTejun Heo pcpu_destroy_chunk(chunk); 1143fbf59bc9STejun Heo } 1144971f3918STejun Heo 11454f996e23STejun Heo /* service chunks which requested async area map extension */ 11464f996e23STejun Heo do { 11474f996e23STejun Heo int new_alloc = 0; 11484f996e23STejun Heo 11494f996e23STejun Heo spin_lock_irq(&pcpu_lock); 11504f996e23STejun Heo 11514f996e23STejun Heo chunk = list_first_entry_or_null(&pcpu_map_extend_chunks, 11524f996e23STejun Heo struct pcpu_chunk, map_extend_list); 11534f996e23STejun Heo if (chunk) { 11544f996e23STejun Heo list_del_init(&chunk->map_extend_list); 11554f996e23STejun Heo new_alloc = pcpu_need_to_extend(chunk, false); 11564f996e23STejun Heo } 11574f996e23STejun Heo 11584f996e23STejun Heo spin_unlock_irq(&pcpu_lock); 11594f996e23STejun Heo 11604f996e23STejun Heo if (new_alloc) 11614f996e23STejun Heo pcpu_extend_area_map(chunk, new_alloc); 11624f996e23STejun Heo } while (chunk); 11634f996e23STejun Heo 11641a4d7607STejun Heo /* 11651a4d7607STejun Heo * Ensure there are certain number of free populated pages for 11661a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 11671a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 11681a4d7607STejun Heo * failed previously, always populate the maximum amount. This 11691a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 11701a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 11711a4d7607STejun Heo * something we support properly and can be highly unreliable and 11721a4d7607STejun Heo * inefficient. 11731a4d7607STejun Heo */ 11741a4d7607STejun Heo retry_pop: 11751a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 11761a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 11771a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 11781a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 11791a4d7607STejun Heo } else { 11801a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 11811a4d7607STejun Heo pcpu_nr_empty_pop_pages, 11821a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 11831a4d7607STejun Heo } 11841a4d7607STejun Heo 11851a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 11861a4d7607STejun Heo int nr_unpop = 0, rs, re; 11871a4d7607STejun Heo 11881a4d7607STejun Heo if (!nr_to_pop) 11891a4d7607STejun Heo break; 11901a4d7607STejun Heo 11911a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 11921a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 11931a4d7607STejun Heo nr_unpop = pcpu_unit_pages - chunk->nr_populated; 11941a4d7607STejun Heo if (nr_unpop) 11951a4d7607STejun Heo break; 11961a4d7607STejun Heo } 11971a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 11981a4d7607STejun Heo 11991a4d7607STejun Heo if (!nr_unpop) 12001a4d7607STejun Heo continue; 12011a4d7607STejun Heo 12021a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 12031a4d7607STejun Heo pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) { 12041a4d7607STejun Heo int nr = min(re - rs, nr_to_pop); 12051a4d7607STejun Heo 12061a4d7607STejun Heo ret = pcpu_populate_chunk(chunk, rs, rs + nr); 12071a4d7607STejun Heo if (!ret) { 12081a4d7607STejun Heo nr_to_pop -= nr; 12091a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 12101a4d7607STejun Heo pcpu_chunk_populated(chunk, rs, rs + nr); 12111a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 12121a4d7607STejun Heo } else { 12131a4d7607STejun Heo nr_to_pop = 0; 12141a4d7607STejun Heo } 12151a4d7607STejun Heo 12161a4d7607STejun Heo if (!nr_to_pop) 12171a4d7607STejun Heo break; 12181a4d7607STejun Heo } 12191a4d7607STejun Heo } 12201a4d7607STejun Heo 12211a4d7607STejun Heo if (nr_to_pop) { 12221a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 12231a4d7607STejun Heo chunk = pcpu_create_chunk(); 12241a4d7607STejun Heo if (chunk) { 12251a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 12261a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 12271a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 12281a4d7607STejun Heo goto retry_pop; 12291a4d7607STejun Heo } 12301a4d7607STejun Heo } 12311a4d7607STejun Heo 1232971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1233a56dbddfSTejun Heo } 1234fbf59bc9STejun Heo 1235fbf59bc9STejun Heo /** 1236fbf59bc9STejun Heo * free_percpu - free percpu area 1237fbf59bc9STejun Heo * @ptr: pointer to area to free 1238fbf59bc9STejun Heo * 1239ccea34b5STejun Heo * Free percpu area @ptr. 1240ccea34b5STejun Heo * 1241ccea34b5STejun Heo * CONTEXT: 1242ccea34b5STejun Heo * Can be called from atomic context. 1243fbf59bc9STejun Heo */ 124443cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1245fbf59bc9STejun Heo { 1246129182e5SAndrew Morton void *addr; 1247fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1248ccea34b5STejun Heo unsigned long flags; 1249b539b87fSTejun Heo int off, occ_pages; 1250fbf59bc9STejun Heo 1251fbf59bc9STejun Heo if (!ptr) 1252fbf59bc9STejun Heo return; 1253fbf59bc9STejun Heo 1254f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1255f528f0b8SCatalin Marinas 1256129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1257129182e5SAndrew Morton 1258ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1259fbf59bc9STejun Heo 1260fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1261bba174f5STejun Heo off = addr - chunk->base_addr; 1262fbf59bc9STejun Heo 1263b539b87fSTejun Heo pcpu_free_area(chunk, off, &occ_pages); 1264b539b87fSTejun Heo 1265b539b87fSTejun Heo if (chunk != pcpu_reserved_chunk) 1266b539b87fSTejun Heo pcpu_nr_empty_pop_pages += occ_pages; 1267fbf59bc9STejun Heo 1268a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 1269fbf59bc9STejun Heo if (chunk->free_size == pcpu_unit_size) { 1270fbf59bc9STejun Heo struct pcpu_chunk *pos; 1271fbf59bc9STejun Heo 1272a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1273fbf59bc9STejun Heo if (pos != chunk) { 12741a4d7607STejun Heo pcpu_schedule_balance_work(); 1275fbf59bc9STejun Heo break; 1276fbf59bc9STejun Heo } 1277fbf59bc9STejun Heo } 1278fbf59bc9STejun Heo 1279ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1280fbf59bc9STejun Heo } 1281fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1282fbf59bc9STejun Heo 12833b034b0dSVivek Goyal /** 128410fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 128510fad5e4STejun Heo * @addr: address to test 128610fad5e4STejun Heo * 128710fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 128810fad5e4STejun Heo * static percpu areas are not considered. For those, use 128910fad5e4STejun Heo * is_module_percpu_address(). 129010fad5e4STejun Heo * 129110fad5e4STejun Heo * RETURNS: 129210fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 129310fad5e4STejun Heo */ 129410fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 129510fad5e4STejun Heo { 1296bbddff05STejun Heo #ifdef CONFIG_SMP 129710fad5e4STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 129810fad5e4STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 129910fad5e4STejun Heo unsigned int cpu; 130010fad5e4STejun Heo 130110fad5e4STejun Heo for_each_possible_cpu(cpu) { 130210fad5e4STejun Heo void *start = per_cpu_ptr(base, cpu); 130310fad5e4STejun Heo 130410fad5e4STejun Heo if ((void *)addr >= start && (void *)addr < start + static_size) 130510fad5e4STejun Heo return true; 130610fad5e4STejun Heo } 1307bbddff05STejun Heo #endif 1308bbddff05STejun Heo /* on UP, can't distinguish from other static vars, always false */ 130910fad5e4STejun Heo return false; 131010fad5e4STejun Heo } 131110fad5e4STejun Heo 131210fad5e4STejun Heo /** 13133b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 13143b034b0dSVivek Goyal * @addr: the address to be converted to physical address 13153b034b0dSVivek Goyal * 13163b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 13173b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 13183b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 13193b034b0dSVivek Goyal * until this function finishes. 13203b034b0dSVivek Goyal * 132167589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 132267589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 132367589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 132467589c71SDave Young * km) provides translation. 132567589c71SDave Young * 1326bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 132767589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 132867589c71SDave Young * actually works, and the verification can discover both bugs in percpu 132967589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 133067589c71SDave Young * code. 133167589c71SDave Young * 13323b034b0dSVivek Goyal * RETURNS: 13333b034b0dSVivek Goyal * The physical address for @addr. 13343b034b0dSVivek Goyal */ 13353b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 13363b034b0dSVivek Goyal { 13379983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 13389983b6f0STejun Heo bool in_first_chunk = false; 1339a855b84cSTejun Heo unsigned long first_low, first_high; 13409983b6f0STejun Heo unsigned int cpu; 13419983b6f0STejun Heo 13429983b6f0STejun Heo /* 1343a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 13449983b6f0STejun Heo * necessary but will speed up lookups of addresses which 13459983b6f0STejun Heo * aren't in the first chunk. 13469983b6f0STejun Heo */ 1347a855b84cSTejun Heo first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0); 1348a855b84cSTejun Heo first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu, 13499983b6f0STejun Heo pcpu_unit_pages); 1350a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1351a855b84cSTejun Heo (unsigned long)addr < first_high) { 13529983b6f0STejun Heo for_each_possible_cpu(cpu) { 13539983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 13549983b6f0STejun Heo 13559983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 13569983b6f0STejun Heo in_first_chunk = true; 13579983b6f0STejun Heo break; 13589983b6f0STejun Heo } 13599983b6f0STejun Heo } 13609983b6f0STejun Heo } 13619983b6f0STejun Heo 13629983b6f0STejun Heo if (in_first_chunk) { 1363eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 13643b034b0dSVivek Goyal return __pa(addr); 13653b034b0dSVivek Goyal else 13669f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 13679f57bd4dSEugene Surovegin offset_in_page(addr); 1368020ec653STejun Heo } else 13699f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 13709f57bd4dSEugene Surovegin offset_in_page(addr); 13713b034b0dSVivek Goyal } 13723b034b0dSVivek Goyal 1373fbf59bc9STejun Heo /** 1374fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1375fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1376fd1e8a1fSTejun Heo * @nr_units: the number of units 1377033e48fbSTejun Heo * 1378fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1379fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1380fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1381fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1382fd1e8a1fSTejun Heo * pointer of other groups. 1383033e48fbSTejun Heo * 1384033e48fbSTejun Heo * RETURNS: 1385fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1386fd1e8a1fSTejun Heo * failure. 1387033e48fbSTejun Heo */ 1388fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1389fd1e8a1fSTejun Heo int nr_units) 1390fd1e8a1fSTejun Heo { 1391fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 1392fd1e8a1fSTejun Heo size_t base_size, ai_size; 1393fd1e8a1fSTejun Heo void *ptr; 1394fd1e8a1fSTejun Heo int unit; 1395fd1e8a1fSTejun Heo 1396fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1397fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 1398fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1399fd1e8a1fSTejun Heo 1400999c17e3SSantosh Shilimkar ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); 1401fd1e8a1fSTejun Heo if (!ptr) 1402fd1e8a1fSTejun Heo return NULL; 1403fd1e8a1fSTejun Heo ai = ptr; 1404fd1e8a1fSTejun Heo ptr += base_size; 1405fd1e8a1fSTejun Heo 1406fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 1407fd1e8a1fSTejun Heo 1408fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 1409fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 1410fd1e8a1fSTejun Heo 1411fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 1412fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 1413fd1e8a1fSTejun Heo 1414fd1e8a1fSTejun Heo return ai; 1415fd1e8a1fSTejun Heo } 1416fd1e8a1fSTejun Heo 1417fd1e8a1fSTejun Heo /** 1418fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 1419fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 1420fd1e8a1fSTejun Heo * 1421fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1422fd1e8a1fSTejun Heo */ 1423fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1424fd1e8a1fSTejun Heo { 1425999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 1426fd1e8a1fSTejun Heo } 1427fd1e8a1fSTejun Heo 1428fd1e8a1fSTejun Heo /** 1429fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1430fd1e8a1fSTejun Heo * @lvl: loglevel 1431fd1e8a1fSTejun Heo * @ai: allocation info to dump 1432fd1e8a1fSTejun Heo * 1433fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 1434fd1e8a1fSTejun Heo */ 1435fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 1436fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 1437033e48fbSTejun Heo { 1438fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 1439033e48fbSTejun Heo char empty_str[] = "--------"; 1440fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 1441fd1e8a1fSTejun Heo int group, v; 1442fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 1443033e48fbSTejun Heo 1444fd1e8a1fSTejun Heo v = ai->nr_groups; 1445033e48fbSTejun Heo while (v /= 10) 1446fd1e8a1fSTejun Heo group_width++; 1447033e48fbSTejun Heo 1448fd1e8a1fSTejun Heo v = num_possible_cpus(); 1449fd1e8a1fSTejun Heo while (v /= 10) 1450fd1e8a1fSTejun Heo cpu_width++; 1451fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1452033e48fbSTejun Heo 1453fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 1454fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 1455fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 1456033e48fbSTejun Heo 1457fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1458fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1459fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1460fd1e8a1fSTejun Heo 1461fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 1462fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 1463fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 1464fd1e8a1fSTejun Heo 1465fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 1466fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 1467fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 1468fd1e8a1fSTejun Heo if (!(alloc % apl)) { 14691170532bSJoe Perches pr_cont("\n"); 1470fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 1471033e48fbSTejun Heo } 14721170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 1473fd1e8a1fSTejun Heo 1474fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 1475fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 14761170532bSJoe Perches pr_cont("%0*d ", 14771170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 1478033e48fbSTejun Heo else 14791170532bSJoe Perches pr_cont("%s ", empty_str); 1480033e48fbSTejun Heo } 1481fd1e8a1fSTejun Heo } 14821170532bSJoe Perches pr_cont("\n"); 1483033e48fbSTejun Heo } 1484033e48fbSTejun Heo 1485fbf59bc9STejun Heo /** 14868d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 1487fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 148838a6be52STejun Heo * @base_addr: mapped address 1489fbf59bc9STejun Heo * 14908d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 14918d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 149238a6be52STejun Heo * setup path. 14938d408b4bSTejun Heo * 1494fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 1495fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 14968d408b4bSTejun Heo * 1497fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 1498fd1e8a1fSTejun Heo * 1499fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1500edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 1501edcb4639STejun Heo * the first chunk such that it's available only through reserved 1502edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 1503edcb4639STejun Heo * static areas on architectures where the addressing model has 1504edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 1505edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 1506edcb4639STejun Heo * 1507fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 1508fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 1509fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 15106074d5b0STejun Heo * 1511fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1512fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 1513fd1e8a1fSTejun Heo * @ai->dyn_size. 15148d408b4bSTejun Heo * 1515fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 1516fd1e8a1fSTejun Heo * for vm areas. 15178d408b4bSTejun Heo * 1518fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 1519fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 1520fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 1521fd1e8a1fSTejun Heo * 1522fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 1523fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 1524fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 1525fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 1526fd1e8a1fSTejun Heo * all units is assumed. 15278d408b4bSTejun Heo * 152838a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 152938a6be52STejun Heo * copied static data to each unit. 1530fbf59bc9STejun Heo * 1531edcb4639STejun Heo * If the first chunk ends up with both reserved and dynamic areas, it 1532edcb4639STejun Heo * is served by two chunks - one to serve the core static and reserved 1533edcb4639STejun Heo * areas and the other for the dynamic area. They share the same vm 1534edcb4639STejun Heo * and page map but uses different area allocation map to stay away 1535edcb4639STejun Heo * from each other. The latter chunk is circulated in the chunk slots 1536edcb4639STejun Heo * and available for dynamic allocation like any other chunks. 1537edcb4639STejun Heo * 1538fbf59bc9STejun Heo * RETURNS: 1539fb435d52STejun Heo * 0 on success, -errno on failure. 1540fbf59bc9STejun Heo */ 1541fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1542fd1e8a1fSTejun Heo void *base_addr) 1543fbf59bc9STejun Heo { 1544099a19d9STejun Heo static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1545099a19d9STejun Heo static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1546fd1e8a1fSTejun Heo size_t dyn_size = ai->dyn_size; 1547fd1e8a1fSTejun Heo size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1548edcb4639STejun Heo struct pcpu_chunk *schunk, *dchunk = NULL; 15496563297cSTejun Heo unsigned long *group_offsets; 15506563297cSTejun Heo size_t *group_sizes; 1551fb435d52STejun Heo unsigned long *unit_off; 1552fbf59bc9STejun Heo unsigned int cpu; 1553fd1e8a1fSTejun Heo int *unit_map; 1554fd1e8a1fSTejun Heo int group, unit, i; 1555fbf59bc9STejun Heo 1556635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 1557635b75fcSTejun Heo if (unlikely(cond)) { \ 1558870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 1559870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 1560807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 1561635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1562635b75fcSTejun Heo BUG(); \ 1563635b75fcSTejun Heo } \ 1564635b75fcSTejun Heo } while (0) 1565635b75fcSTejun Heo 15662f39e637STejun Heo /* sanity checks */ 1567635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1568bbddff05STejun Heo #ifdef CONFIG_SMP 1569635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 1570f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 1571bbddff05STejun Heo #endif 1572635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 1573f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 1574635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1575f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 1576635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1577099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 15789f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 15798d408b4bSTejun Heo 15806563297cSTejun Heo /* process group information and build config tables accordingly */ 1581999c17e3SSantosh Shilimkar group_offsets = memblock_virt_alloc(ai->nr_groups * 1582999c17e3SSantosh Shilimkar sizeof(group_offsets[0]), 0); 1583999c17e3SSantosh Shilimkar group_sizes = memblock_virt_alloc(ai->nr_groups * 1584999c17e3SSantosh Shilimkar sizeof(group_sizes[0]), 0); 1585999c17e3SSantosh Shilimkar unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); 1586999c17e3SSantosh Shilimkar unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); 15872f39e637STejun Heo 1588fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1589ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 1590a855b84cSTejun Heo 1591a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 1592a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 15932f39e637STejun Heo 1594fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1595fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 15962f39e637STejun Heo 15976563297cSTejun Heo group_offsets[group] = gi->base_offset; 15986563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 15996563297cSTejun Heo 1600fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 1601fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 1602fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 1603fd1e8a1fSTejun Heo continue; 1604fd1e8a1fSTejun Heo 16059f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 1606635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1607635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1608fd1e8a1fSTejun Heo 1609fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 1610fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1611fb435d52STejun Heo 1612a855b84cSTejun Heo /* determine low/high unit_cpu */ 1613a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 1614a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 1615a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 1616a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 1617a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 1618a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 16190fc0531eSLinus Torvalds } 16200fc0531eSLinus Torvalds } 1621fd1e8a1fSTejun Heo pcpu_nr_units = unit; 16222f39e637STejun Heo 16232f39e637STejun Heo for_each_possible_cpu(cpu) 1624635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1625635b75fcSTejun Heo 1626635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 1627635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 1628bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 16292f39e637STejun Heo 16306563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 16316563297cSTejun Heo pcpu_group_offsets = group_offsets; 16326563297cSTejun Heo pcpu_group_sizes = group_sizes; 1633fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 1634fb435d52STejun Heo pcpu_unit_offsets = unit_off; 16352f39e637STejun Heo 16362f39e637STejun Heo /* determine basic parameters */ 1637fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1638d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 16396563297cSTejun Heo pcpu_atom_size = ai->atom_size; 1640ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1641ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1642cafe8816STejun Heo 1643d9b55eebSTejun Heo /* 1644d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 1645d9b55eebSTejun Heo * empty chunks. 1646d9b55eebSTejun Heo */ 1647d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1648999c17e3SSantosh Shilimkar pcpu_slot = memblock_virt_alloc( 1649999c17e3SSantosh Shilimkar pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); 1650fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 1651fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 1652fbf59bc9STejun Heo 1653edcb4639STejun Heo /* 1654edcb4639STejun Heo * Initialize static chunk. If reserved_size is zero, the 1655edcb4639STejun Heo * static chunk covers static area + dynamic allocation area 1656edcb4639STejun Heo * in the first chunk. If reserved_size is not zero, it 1657edcb4639STejun Heo * covers static area + reserved area (mostly used for module 1658edcb4639STejun Heo * static percpu allocation). 1659edcb4639STejun Heo */ 1660999c17e3SSantosh Shilimkar schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); 16612441d15cSTejun Heo INIT_LIST_HEAD(&schunk->list); 16624f996e23STejun Heo INIT_LIST_HEAD(&schunk->map_extend_list); 1663bba174f5STejun Heo schunk->base_addr = base_addr; 166461ace7faSTejun Heo schunk->map = smap; 166561ace7faSTejun Heo schunk->map_alloc = ARRAY_SIZE(smap); 166638a6be52STejun Heo schunk->immutable = true; 1667ce3141a2STejun Heo bitmap_fill(schunk->populated, pcpu_unit_pages); 1668b539b87fSTejun Heo schunk->nr_populated = pcpu_unit_pages; 1669edcb4639STejun Heo 1670fd1e8a1fSTejun Heo if (ai->reserved_size) { 1671fd1e8a1fSTejun Heo schunk->free_size = ai->reserved_size; 1672ae9e6bc9STejun Heo pcpu_reserved_chunk = schunk; 1673fd1e8a1fSTejun Heo pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1674edcb4639STejun Heo } else { 16752441d15cSTejun Heo schunk->free_size = dyn_size; 1676edcb4639STejun Heo dyn_size = 0; /* dynamic area covered */ 1677edcb4639STejun Heo } 16782441d15cSTejun Heo schunk->contig_hint = schunk->free_size; 1679fbf59bc9STejun Heo 1680723ad1d9SAl Viro schunk->map[0] = 1; 1681723ad1d9SAl Viro schunk->map[1] = ai->static_size; 1682723ad1d9SAl Viro schunk->map_used = 1; 168361ace7faSTejun Heo if (schunk->free_size) 1684292c24a0SBaoquan He schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size; 1685292c24a0SBaoquan He schunk->map[schunk->map_used] |= 1; 168661ace7faSTejun Heo 1687edcb4639STejun Heo /* init dynamic chunk if necessary */ 1688edcb4639STejun Heo if (dyn_size) { 1689999c17e3SSantosh Shilimkar dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); 1690edcb4639STejun Heo INIT_LIST_HEAD(&dchunk->list); 16914f996e23STejun Heo INIT_LIST_HEAD(&dchunk->map_extend_list); 1692bba174f5STejun Heo dchunk->base_addr = base_addr; 1693edcb4639STejun Heo dchunk->map = dmap; 1694edcb4639STejun Heo dchunk->map_alloc = ARRAY_SIZE(dmap); 169538a6be52STejun Heo dchunk->immutable = true; 1696ce3141a2STejun Heo bitmap_fill(dchunk->populated, pcpu_unit_pages); 1697b539b87fSTejun Heo dchunk->nr_populated = pcpu_unit_pages; 1698edcb4639STejun Heo 1699edcb4639STejun Heo dchunk->contig_hint = dchunk->free_size = dyn_size; 1700723ad1d9SAl Viro dchunk->map[0] = 1; 1701723ad1d9SAl Viro dchunk->map[1] = pcpu_reserved_chunk_limit; 1702723ad1d9SAl Viro dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1; 1703723ad1d9SAl Viro dchunk->map_used = 2; 1704edcb4639STejun Heo } 1705edcb4639STejun Heo 17062441d15cSTejun Heo /* link the first chunk in */ 1707ae9e6bc9STejun Heo pcpu_first_chunk = dchunk ?: schunk; 1708b539b87fSTejun Heo pcpu_nr_empty_pop_pages += 1709b539b87fSTejun Heo pcpu_count_occupied_pages(pcpu_first_chunk, 1); 1710ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 1711fbf59bc9STejun Heo 1712fbf59bc9STejun Heo /* we're done */ 1713bba174f5STejun Heo pcpu_base_addr = base_addr; 1714fb435d52STejun Heo return 0; 1715fbf59bc9STejun Heo } 171666c3a757STejun Heo 1717bbddff05STejun Heo #ifdef CONFIG_SMP 1718bbddff05STejun Heo 171917f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 1720f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 1721f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 1722f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 1723f58dc01bSTejun Heo }; 172466c3a757STejun Heo 1725f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1726f58dc01bSTejun Heo 1727f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 172866c3a757STejun Heo { 17295479c78aSCyrill Gorcunov if (!str) 17305479c78aSCyrill Gorcunov return -EINVAL; 17315479c78aSCyrill Gorcunov 1732f58dc01bSTejun Heo if (0) 1733f58dc01bSTejun Heo /* nada */; 1734f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1735f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 1736f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 1737f58dc01bSTejun Heo #endif 1738f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1739f58dc01bSTejun Heo else if (!strcmp(str, "page")) 1740f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 1741f58dc01bSTejun Heo #endif 1742f58dc01bSTejun Heo else 1743870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 174466c3a757STejun Heo 1745f58dc01bSTejun Heo return 0; 174666c3a757STejun Heo } 1747f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 174866c3a757STejun Heo 17493c9a024fSTejun Heo /* 17503c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 17513c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 17523c9a024fSTejun Heo * to be used. 17533c9a024fSTejun Heo */ 175408fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 175508fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 17563c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 17573c9a024fSTejun Heo #endif 17583c9a024fSTejun Heo 17593c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 17603c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 17613c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 17623c9a024fSTejun Heo #endif 17633c9a024fSTejun Heo 17643c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 17653c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 17663c9a024fSTejun Heo /** 1767fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1768fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 1769fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 1770fbf59bc9STejun Heo * @atom_size: allocation atom size 1771fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 1772fbf59bc9STejun Heo * 1773fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 1774fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 1775fbf59bc9STejun Heo * atom size and distances between CPUs. 1776fbf59bc9STejun Heo * 1777bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 1778fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 1779fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 1780fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 1781fbf59bc9STejun Heo * of allocated virtual address space. 1782fbf59bc9STejun Heo * 1783fbf59bc9STejun Heo * RETURNS: 1784fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 1785fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 1786fbf59bc9STejun Heo */ 1787fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1788fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 1789fbf59bc9STejun Heo size_t atom_size, 1790fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1791fbf59bc9STejun Heo { 1792fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 1793fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 1794fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 1795fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 1796fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 1797fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1798fbf59bc9STejun Heo int last_allocs, group, unit; 1799fbf59bc9STejun Heo unsigned int cpu, tcpu; 1800fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 1801fbf59bc9STejun Heo unsigned int *cpu_map; 1802fbf59bc9STejun Heo 1803fbf59bc9STejun Heo /* this function may be called multiple times */ 1804fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 1805fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 1806fbf59bc9STejun Heo 1807fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 1808fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 1809fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 1810fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 1811fbf59bc9STejun Heo 1812fbf59bc9STejun Heo /* 1813fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 1814fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 181525985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 1816fbf59bc9STejun Heo * or larger than min_unit_size. 1817fbf59bc9STejun Heo */ 1818fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1819fbf59bc9STejun Heo 1820fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 1821fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 1822f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 1823fbf59bc9STejun Heo upa--; 1824fbf59bc9STejun Heo max_upa = upa; 1825fbf59bc9STejun Heo 1826fbf59bc9STejun Heo /* group cpus according to their proximity */ 1827fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 1828fbf59bc9STejun Heo group = 0; 1829fbf59bc9STejun Heo next_group: 1830fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 1831fbf59bc9STejun Heo if (cpu == tcpu) 1832fbf59bc9STejun Heo break; 1833fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 1834fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1835fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1836fbf59bc9STejun Heo group++; 1837fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 1838fbf59bc9STejun Heo goto next_group; 1839fbf59bc9STejun Heo } 1840fbf59bc9STejun Heo } 1841fbf59bc9STejun Heo group_map[cpu] = group; 1842fbf59bc9STejun Heo group_cnt[group]++; 1843fbf59bc9STejun Heo } 1844fbf59bc9STejun Heo 1845fbf59bc9STejun Heo /* 1846fbf59bc9STejun Heo * Expand unit size until address space usage goes over 75% 1847fbf59bc9STejun Heo * and then as much as possible without using more address 1848fbf59bc9STejun Heo * space. 1849fbf59bc9STejun Heo */ 1850fbf59bc9STejun Heo last_allocs = INT_MAX; 1851fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 1852fbf59bc9STejun Heo int allocs = 0, wasted = 0; 1853fbf59bc9STejun Heo 1854f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 1855fbf59bc9STejun Heo continue; 1856fbf59bc9STejun Heo 1857fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 1858fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1859fbf59bc9STejun Heo allocs += this_allocs; 1860fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 1861fbf59bc9STejun Heo } 1862fbf59bc9STejun Heo 1863fbf59bc9STejun Heo /* 1864fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 1865fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 1866fbf59bc9STejun Heo * passes the following check. 1867fbf59bc9STejun Heo */ 1868fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 1869fbf59bc9STejun Heo continue; 1870fbf59bc9STejun Heo 1871fbf59bc9STejun Heo /* and then don't consume more memory */ 1872fbf59bc9STejun Heo if (allocs > last_allocs) 1873fbf59bc9STejun Heo break; 1874fbf59bc9STejun Heo last_allocs = allocs; 1875fbf59bc9STejun Heo best_upa = upa; 1876fbf59bc9STejun Heo } 1877fbf59bc9STejun Heo upa = best_upa; 1878fbf59bc9STejun Heo 1879fbf59bc9STejun Heo /* allocate and fill alloc_info */ 1880fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 1881fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 1882fbf59bc9STejun Heo 1883fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1884fbf59bc9STejun Heo if (!ai) 1885fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 1886fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 1887fbf59bc9STejun Heo 1888fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 1889fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 1890fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 1891fbf59bc9STejun Heo } 1892fbf59bc9STejun Heo 1893fbf59bc9STejun Heo ai->static_size = static_size; 1894fbf59bc9STejun Heo ai->reserved_size = reserved_size; 1895fbf59bc9STejun Heo ai->dyn_size = dyn_size; 1896fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 1897fbf59bc9STejun Heo ai->atom_size = atom_size; 1898fbf59bc9STejun Heo ai->alloc_size = alloc_size; 1899fbf59bc9STejun Heo 1900fbf59bc9STejun Heo for (group = 0, unit = 0; group_cnt[group]; group++) { 1901fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 1902fbf59bc9STejun Heo 1903fbf59bc9STejun Heo /* 1904fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 1905fbf59bc9STejun Heo * back-to-back. The caller should update this to 1906fbf59bc9STejun Heo * reflect actual allocation. 1907fbf59bc9STejun Heo */ 1908fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 1909fbf59bc9STejun Heo 1910fbf59bc9STejun Heo for_each_possible_cpu(cpu) 1911fbf59bc9STejun Heo if (group_map[cpu] == group) 1912fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 1913fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 1914fbf59bc9STejun Heo unit += gi->nr_units; 1915fbf59bc9STejun Heo } 1916fbf59bc9STejun Heo BUG_ON(unit != nr_units); 1917fbf59bc9STejun Heo 1918fbf59bc9STejun Heo return ai; 1919fbf59bc9STejun Heo } 19203c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 1921fbf59bc9STejun Heo 19223c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 192366c3a757STejun Heo /** 192466c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 192566c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 19264ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 1927c8826dd5STejun Heo * @atom_size: allocation atom size 1928c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 1929c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 193025985edcSLucas De Marchi * @free_fn: function to free percpu page 193166c3a757STejun Heo * 193266c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 193366c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 193466c3a757STejun Heo * 193566c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 1936c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 1937c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 1938c8826dd5STejun Heo * aligned to @atom_size. 1939c8826dd5STejun Heo * 1940c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 1941c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 1942c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 1943c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 1944c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 1945c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 194666c3a757STejun Heo * 19474ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 194866c3a757STejun Heo * 194966c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 1950c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 195166c3a757STejun Heo * 195266c3a757STejun Heo * RETURNS: 1953fb435d52STejun Heo * 0 on success, -errno on failure. 195466c3a757STejun Heo */ 19554ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 1956c8826dd5STejun Heo size_t atom_size, 1957c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1958c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 1959c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 196066c3a757STejun Heo { 1961c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 1962c8826dd5STejun Heo void **areas = NULL; 1963fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 196493c76b6bSzijun_hu size_t size_sum, areas_size; 196593c76b6bSzijun_hu unsigned long max_distance; 1966*9b739662Szijun_hu int group, i, highest_group, rc; 196766c3a757STejun Heo 1968c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1969c8826dd5STejun Heo cpu_distance_fn); 1970fd1e8a1fSTejun Heo if (IS_ERR(ai)) 1971fd1e8a1fSTejun Heo return PTR_ERR(ai); 197266c3a757STejun Heo 1973fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1974c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 197566c3a757STejun Heo 1976999c17e3SSantosh Shilimkar areas = memblock_virt_alloc_nopanic(areas_size, 0); 1977c8826dd5STejun Heo if (!areas) { 1978fb435d52STejun Heo rc = -ENOMEM; 1979c8826dd5STejun Heo goto out_free; 1980fa8a7094STejun Heo } 198166c3a757STejun Heo 1982*9b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 1983*9b739662Szijun_hu highest_group = 0; 1984c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 1985c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 1986c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 1987c8826dd5STejun Heo void *ptr; 198866c3a757STejun Heo 1989c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1990c8826dd5STejun Heo cpu = gi->cpu_map[i]; 1991c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 1992c8826dd5STejun Heo 1993c8826dd5STejun Heo /* allocate space for the whole group */ 1994c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1995c8826dd5STejun Heo if (!ptr) { 1996c8826dd5STejun Heo rc = -ENOMEM; 1997c8826dd5STejun Heo goto out_free_areas; 1998c8826dd5STejun Heo } 1999f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2000f528f0b8SCatalin Marinas kmemleak_free(ptr); 2001c8826dd5STejun Heo areas[group] = ptr; 2002c8826dd5STejun Heo 2003c8826dd5STejun Heo base = min(ptr, base); 2004*9b739662Szijun_hu if (ptr > areas[highest_group]) 2005*9b739662Szijun_hu highest_group = group; 2006*9b739662Szijun_hu } 2007*9b739662Szijun_hu max_distance = areas[highest_group] - base; 2008*9b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 2009*9b739662Szijun_hu 2010*9b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 2011*9b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 2012*9b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 2013*9b739662Szijun_hu max_distance, VMALLOC_TOTAL); 2014*9b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 2015*9b739662Szijun_hu /* and fail if we have fallback */ 2016*9b739662Szijun_hu rc = -EINVAL; 2017*9b739662Szijun_hu goto out_free_areas; 2018*9b739662Szijun_hu #endif 201942b64281STejun Heo } 202042b64281STejun Heo 202142b64281STejun Heo /* 202242b64281STejun Heo * Copy data and free unused parts. This should happen after all 202342b64281STejun Heo * allocations are complete; otherwise, we may end up with 202442b64281STejun Heo * overlapping groups. 202542b64281STejun Heo */ 202642b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 202742b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 202842b64281STejun Heo void *ptr = areas[group]; 2029c8826dd5STejun Heo 2030c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2031c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2032c8826dd5STejun Heo /* unused unit, free whole */ 2033c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2034c8826dd5STejun Heo continue; 2035c8826dd5STejun Heo } 2036c8826dd5STejun Heo /* copy and return the unused part */ 2037fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2038c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2039c8826dd5STejun Heo } 204066c3a757STejun Heo } 204166c3a757STejun Heo 2042c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 20436ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2044c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 20456ea529a2STejun Heo } 2046c8826dd5STejun Heo 2047870d4b12SJoe Perches pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 2048fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 2049fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 205066c3a757STejun Heo 2051fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 2052c8826dd5STejun Heo goto out_free; 2053c8826dd5STejun Heo 2054c8826dd5STejun Heo out_free_areas: 2055c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2056f851c8d8SMichael Holzheu if (areas[group]) 2057c8826dd5STejun Heo free_fn(areas[group], 2058c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2059c8826dd5STejun Heo out_free: 2060fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2061c8826dd5STejun Heo if (areas) 2062999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2063fb435d52STejun Heo return rc; 2064d4b95f80STejun Heo } 20653c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2066d4b95f80STejun Heo 20673c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2068d4b95f80STejun Heo /** 206900ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2070d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2071d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 207225985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2073d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2074d4b95f80STejun Heo * 207500ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 207600ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2077d4b95f80STejun Heo * 2078d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2079d4b95f80STejun Heo * page-by-page into vmalloc area. 2080d4b95f80STejun Heo * 2081d4b95f80STejun Heo * RETURNS: 2082fb435d52STejun Heo * 0 on success, -errno on failure. 2083d4b95f80STejun Heo */ 2084fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2085d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2086d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2087d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2088d4b95f80STejun Heo { 20898f05a6a6STejun Heo static struct vm_struct vm; 2090fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 209100ae4064STejun Heo char psize_str[16]; 2092ce3141a2STejun Heo int unit_pages; 2093d4b95f80STejun Heo size_t pages_size; 2094ce3141a2STejun Heo struct page **pages; 2095fb435d52STejun Heo int unit, i, j, rc; 2096d4b95f80STejun Heo 209700ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 209800ae4064STejun Heo 20994ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2100fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2101fd1e8a1fSTejun Heo return PTR_ERR(ai); 2102fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 2103fd1e8a1fSTejun Heo BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 2104fd1e8a1fSTejun Heo 2105fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2106d4b95f80STejun Heo 2107d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2108fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2109fd1e8a1fSTejun Heo sizeof(pages[0])); 2110999c17e3SSantosh Shilimkar pages = memblock_virt_alloc(pages_size, 0); 2111d4b95f80STejun Heo 21128f05a6a6STejun Heo /* allocate pages */ 2113d4b95f80STejun Heo j = 0; 2114fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) 2115ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) { 2116fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 2117d4b95f80STejun Heo void *ptr; 2118d4b95f80STejun Heo 21193cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2120d4b95f80STejun Heo if (!ptr) { 2121870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2122598d8091SJoe Perches psize_str, cpu); 2123d4b95f80STejun Heo goto enomem; 2124d4b95f80STejun Heo } 2125f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2126f528f0b8SCatalin Marinas kmemleak_free(ptr); 2127ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2128d4b95f80STejun Heo } 2129d4b95f80STejun Heo 21308f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 21318f05a6a6STejun Heo vm.flags = VM_ALLOC; 2132fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 21338f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 21348f05a6a6STejun Heo 2135fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 21361d9d3257STejun Heo unsigned long unit_addr = 2137fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 21388f05a6a6STejun Heo 2139ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 21408f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 21418f05a6a6STejun Heo 21428f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2143fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2144ce3141a2STejun Heo unit_pages); 2145fb435d52STejun Heo if (rc < 0) 2146fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 21478f05a6a6STejun Heo 21488f05a6a6STejun Heo /* 21498f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 21508f05a6a6STejun Heo * cache for the linear mapping here - something 21518f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 21528f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 21538f05a6a6STejun Heo * data structures are not set up yet. 21548f05a6a6STejun Heo */ 21558f05a6a6STejun Heo 21568f05a6a6STejun Heo /* copy static data */ 2157fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 215866c3a757STejun Heo } 215966c3a757STejun Heo 216066c3a757STejun Heo /* we're ready, commit */ 2161870d4b12SJoe Perches pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", 2162fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 2163fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 216466c3a757STejun Heo 2165fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 2166d4b95f80STejun Heo goto out_free_ar; 2167d4b95f80STejun Heo 2168d4b95f80STejun Heo enomem: 2169d4b95f80STejun Heo while (--j >= 0) 2170ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2171fb435d52STejun Heo rc = -ENOMEM; 2172d4b95f80STejun Heo out_free_ar: 2173999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2174fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2175fb435d52STejun Heo return rc; 217666c3a757STejun Heo } 21773c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2178d4b95f80STejun Heo 2179bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 21808c4bfc6eSTejun Heo /* 2181bbddff05STejun Heo * Generic SMP percpu area setup. 2182e74e3962STejun Heo * 2183e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2184e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2185e74e3962STejun Heo * important because many archs have addressing restrictions and might 2186e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2187e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2188e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2189e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2190e74e3962STejun Heo * mappings on applicable archs. 2191e74e3962STejun Heo */ 2192e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2193e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2194e74e3962STejun Heo 2195c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2196c8826dd5STejun Heo size_t align) 2197c8826dd5STejun Heo { 2198999c17e3SSantosh Shilimkar return memblock_virt_alloc_from_nopanic( 2199999c17e3SSantosh Shilimkar size, align, __pa(MAX_DMA_ADDRESS)); 2200c8826dd5STejun Heo } 2201c8826dd5STejun Heo 2202c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2203c8826dd5STejun Heo { 2204999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2205c8826dd5STejun Heo } 2206c8826dd5STejun Heo 2207e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2208e74e3962STejun Heo { 2209e74e3962STejun Heo unsigned long delta; 2210e74e3962STejun Heo unsigned int cpu; 2211fb435d52STejun Heo int rc; 2212e74e3962STejun Heo 2213e74e3962STejun Heo /* 2214e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2215e74e3962STejun Heo * what the legacy allocator did. 2216e74e3962STejun Heo */ 2217fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2218c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2219c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2220fb435d52STejun Heo if (rc < 0) 2221bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2222e74e3962STejun Heo 2223e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2224e74e3962STejun Heo for_each_possible_cpu(cpu) 2225fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2226e74e3962STejun Heo } 2227e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2228099a19d9STejun Heo 2229bbddff05STejun Heo #else /* CONFIG_SMP */ 2230bbddff05STejun Heo 2231bbddff05STejun Heo /* 2232bbddff05STejun Heo * UP percpu area setup. 2233bbddff05STejun Heo * 2234bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2235bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2236bbddff05STejun Heo * variables and don't require any special preparation. 2237bbddff05STejun Heo */ 2238bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2239bbddff05STejun Heo { 2240bbddff05STejun Heo const size_t unit_size = 2241bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2242bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2243bbddff05STejun Heo struct pcpu_alloc_info *ai; 2244bbddff05STejun Heo void *fc; 2245bbddff05STejun Heo 2246bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 2247999c17e3SSantosh Shilimkar fc = memblock_virt_alloc_from_nopanic(unit_size, 2248999c17e3SSantosh Shilimkar PAGE_SIZE, 2249999c17e3SSantosh Shilimkar __pa(MAX_DMA_ADDRESS)); 2250bbddff05STejun Heo if (!ai || !fc) 2251bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2252100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2253100d13c3SCatalin Marinas kmemleak_free(fc); 2254bbddff05STejun Heo 2255bbddff05STejun Heo ai->dyn_size = unit_size; 2256bbddff05STejun Heo ai->unit_size = unit_size; 2257bbddff05STejun Heo ai->atom_size = unit_size; 2258bbddff05STejun Heo ai->alloc_size = unit_size; 2259bbddff05STejun Heo ai->groups[0].nr_units = 1; 2260bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2261bbddff05STejun Heo 2262bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 2263bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2264bbddff05STejun Heo } 2265bbddff05STejun Heo 2266bbddff05STejun Heo #endif /* CONFIG_SMP */ 2267bbddff05STejun Heo 2268099a19d9STejun Heo /* 2269099a19d9STejun Heo * First and reserved chunks are initialized with temporary allocation 2270099a19d9STejun Heo * map in initdata so that they can be used before slab is online. 2271099a19d9STejun Heo * This function is called after slab is brought up and replaces those 2272099a19d9STejun Heo * with properly allocated maps. 2273099a19d9STejun Heo */ 2274099a19d9STejun Heo void __init percpu_init_late(void) 2275099a19d9STejun Heo { 2276099a19d9STejun Heo struct pcpu_chunk *target_chunks[] = 2277099a19d9STejun Heo { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; 2278099a19d9STejun Heo struct pcpu_chunk *chunk; 2279099a19d9STejun Heo unsigned long flags; 2280099a19d9STejun Heo int i; 2281099a19d9STejun Heo 2282099a19d9STejun Heo for (i = 0; (chunk = target_chunks[i]); i++) { 2283099a19d9STejun Heo int *map; 2284099a19d9STejun Heo const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); 2285099a19d9STejun Heo 2286099a19d9STejun Heo BUILD_BUG_ON(size > PAGE_SIZE); 2287099a19d9STejun Heo 228890459ce0SBob Liu map = pcpu_mem_zalloc(size); 2289099a19d9STejun Heo BUG_ON(!map); 2290099a19d9STejun Heo 2291099a19d9STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 2292099a19d9STejun Heo memcpy(map, chunk->map, size); 2293099a19d9STejun Heo chunk->map = map; 2294099a19d9STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 2295099a19d9STejun Heo } 2296099a19d9STejun Heo } 22971a4d7607STejun Heo 22981a4d7607STejun Heo /* 22991a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 23001a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 23011a4d7607STejun Heo * and running. 23021a4d7607STejun Heo */ 23031a4d7607STejun Heo static int __init percpu_enable_async(void) 23041a4d7607STejun Heo { 23051a4d7607STejun Heo pcpu_async_enabled = true; 23061a4d7607STejun Heo return 0; 23071a4d7607STejun Heo } 23081a4d7607STejun Heo subsys_initcall(percpu_enable_async); 2309