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 56fbf59bc9STejun Heo #include <linux/bitmap.h> 57fbf59bc9STejun Heo #include <linux/bootmem.h> 58fd1e8a1fSTejun Heo #include <linux/err.h> 59fbf59bc9STejun Heo #include <linux/list.h> 60a530b795STejun Heo #include <linux/log2.h> 61fbf59bc9STejun Heo #include <linux/mm.h> 62fbf59bc9STejun Heo #include <linux/module.h> 63fbf59bc9STejun Heo #include <linux/mutex.h> 64fbf59bc9STejun Heo #include <linux/percpu.h> 65fbf59bc9STejun Heo #include <linux/pfn.h> 66fbf59bc9STejun Heo #include <linux/slab.h> 67ccea34b5STejun Heo #include <linux/spinlock.h> 68fbf59bc9STejun Heo #include <linux/vmalloc.h> 69a56dbddfSTejun Heo #include <linux/workqueue.h> 70f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 71fbf59bc9STejun Heo 72fbf59bc9STejun Heo #include <asm/cacheflush.h> 73e0100983STejun Heo #include <asm/sections.h> 74fbf59bc9STejun Heo #include <asm/tlbflush.h> 753b034b0dSVivek Goyal #include <asm/io.h> 76fbf59bc9STejun Heo 77fbf59bc9STejun Heo #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ 78fbf59bc9STejun Heo #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ 79fbf59bc9STejun Heo 80bbddff05STejun Heo #ifdef CONFIG_SMP 81e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 82e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 83e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 8443cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 8543cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 8643cf38ebSTejun Heo (unsigned long)__per_cpu_start) 87e0100983STejun Heo #endif 88e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 89e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 9043cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 9143cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 9243cf38ebSTejun Heo (unsigned long)__per_cpu_start) 93e0100983STejun Heo #endif 94bbddff05STejun Heo #else /* CONFIG_SMP */ 95bbddff05STejun Heo /* on UP, it's always identity mapped */ 96bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 97bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 98bbddff05STejun Heo #endif /* CONFIG_SMP */ 99e0100983STejun Heo 100fbf59bc9STejun Heo struct pcpu_chunk { 101fbf59bc9STejun Heo struct list_head list; /* linked to pcpu_slot lists */ 102fbf59bc9STejun Heo int free_size; /* free bytes in the chunk */ 103fbf59bc9STejun Heo int contig_hint; /* max contiguous size hint */ 104bba174f5STejun Heo void *base_addr; /* base address of this chunk */ 105fbf59bc9STejun Heo int map_used; /* # of map entries used */ 106fbf59bc9STejun Heo int map_alloc; /* # of map entries allocated */ 107fbf59bc9STejun Heo int *map; /* allocation map */ 10888999a89STejun Heo void *data; /* chunk data */ 1098d408b4bSTejun Heo bool immutable; /* no [de]population allowed */ 110ce3141a2STejun Heo unsigned long populated[]; /* populated bitmap */ 111fbf59bc9STejun Heo }; 112fbf59bc9STejun Heo 11340150d37STejun Heo static int pcpu_unit_pages __read_mostly; 11440150d37STejun Heo static int pcpu_unit_size __read_mostly; 1152f39e637STejun Heo static int pcpu_nr_units __read_mostly; 1166563297cSTejun Heo static int pcpu_atom_size __read_mostly; 11740150d37STejun Heo static int pcpu_nr_slots __read_mostly; 11840150d37STejun Heo static size_t pcpu_chunk_struct_size __read_mostly; 119fbf59bc9STejun Heo 120a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 121a855b84cSTejun Heo static unsigned int pcpu_low_unit_cpu __read_mostly; 122a855b84cSTejun Heo static unsigned int pcpu_high_unit_cpu __read_mostly; 1232f39e637STejun Heo 124fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 12540150d37STejun Heo void *pcpu_base_addr __read_mostly; 126fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 127fbf59bc9STejun Heo 128fb435d52STejun Heo static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ 129fb435d52STejun Heo const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ 1302f39e637STejun Heo 1316563297cSTejun Heo /* group information, used for vm allocation */ 1326563297cSTejun Heo static int pcpu_nr_groups __read_mostly; 1336563297cSTejun Heo static const unsigned long *pcpu_group_offsets __read_mostly; 1346563297cSTejun Heo static const size_t *pcpu_group_sizes __read_mostly; 1356563297cSTejun Heo 136ae9e6bc9STejun Heo /* 137ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 138ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 139ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 140ae9e6bc9STejun Heo */ 141ae9e6bc9STejun Heo static struct pcpu_chunk *pcpu_first_chunk; 142ae9e6bc9STejun Heo 143ae9e6bc9STejun Heo /* 144ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 145ae9e6bc9STejun Heo * chunk and serves it for reserved allocations. The amount of 146ae9e6bc9STejun Heo * reserved offset is in pcpu_reserved_chunk_limit. When reserved 147ae9e6bc9STejun Heo * area doesn't exist, the following variables contain NULL and 0 148ae9e6bc9STejun Heo * respectively. 149ae9e6bc9STejun Heo */ 150edcb4639STejun Heo static struct pcpu_chunk *pcpu_reserved_chunk; 151edcb4639STejun Heo static int pcpu_reserved_chunk_limit; 152edcb4639STejun Heo 153fbf59bc9STejun Heo /* 154ccea34b5STejun Heo * Synchronization rules. 155fbf59bc9STejun Heo * 156ccea34b5STejun Heo * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 157ce3141a2STejun Heo * protects allocation/reclaim paths, chunks, populated bitmap and 158ce3141a2STejun Heo * vmalloc mapping. The latter is a spinlock and protects the index 159ce3141a2STejun Heo * data structures - chunk slots, chunks and area maps in chunks. 160fbf59bc9STejun Heo * 161ccea34b5STejun Heo * During allocation, pcpu_alloc_mutex is kept locked all the time and 162ccea34b5STejun Heo * pcpu_lock is grabbed and released as necessary. All actual memory 163403a91b1SJiri Kosina * allocations are done using GFP_KERNEL with pcpu_lock released. In 164403a91b1SJiri Kosina * general, percpu memory can't be allocated with irq off but 165403a91b1SJiri Kosina * irqsave/restore are still used in alloc path so that it can be used 166403a91b1SJiri Kosina * from early init path - sched_init() specifically. 167ccea34b5STejun Heo * 168ccea34b5STejun Heo * Free path accesses and alters only the index data structures, so it 169ccea34b5STejun Heo * can be safely called from atomic context. When memory needs to be 170ccea34b5STejun Heo * returned to the system, free path schedules reclaim_work which 171ccea34b5STejun Heo * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be 172ccea34b5STejun Heo * reclaimed, release both locks and frees the chunks. Note that it's 173ccea34b5STejun Heo * necessary to grab both locks to remove a chunk from circulation as 174ccea34b5STejun Heo * allocation path might be referencing the chunk with only 175ccea34b5STejun Heo * pcpu_alloc_mutex locked. 176fbf59bc9STejun Heo */ 177ccea34b5STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ 178ccea34b5STejun Heo static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ 179fbf59bc9STejun Heo 18040150d37STejun Heo static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ 181fbf59bc9STejun Heo 182a56dbddfSTejun Heo /* reclaim work to release fully free chunks, scheduled from free path */ 183a56dbddfSTejun Heo static void pcpu_reclaim(struct work_struct *work); 184a56dbddfSTejun Heo static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); 185a56dbddfSTejun Heo 186020ec653STejun Heo static bool pcpu_addr_in_first_chunk(void *addr) 187020ec653STejun Heo { 188020ec653STejun Heo void *first_start = pcpu_first_chunk->base_addr; 189020ec653STejun Heo 190020ec653STejun Heo return addr >= first_start && addr < first_start + pcpu_unit_size; 191020ec653STejun Heo } 192020ec653STejun Heo 193020ec653STejun Heo static bool pcpu_addr_in_reserved_chunk(void *addr) 194020ec653STejun Heo { 195020ec653STejun Heo void *first_start = pcpu_first_chunk->base_addr; 196020ec653STejun Heo 197020ec653STejun Heo return addr >= first_start && 198020ec653STejun Heo addr < first_start + pcpu_reserved_chunk_limit; 199020ec653STejun Heo } 200020ec653STejun Heo 201d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 202fbf59bc9STejun Heo { 203cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 204fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 205fbf59bc9STejun Heo } 206fbf59bc9STejun Heo 207d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 208d9b55eebSTejun Heo { 209d9b55eebSTejun Heo if (size == pcpu_unit_size) 210d9b55eebSTejun Heo return pcpu_nr_slots - 1; 211d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 212d9b55eebSTejun Heo } 213d9b55eebSTejun Heo 214fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 215fbf59bc9STejun Heo { 216fbf59bc9STejun Heo if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) 217fbf59bc9STejun Heo return 0; 218fbf59bc9STejun Heo 219fbf59bc9STejun Heo return pcpu_size_to_slot(chunk->free_size); 220fbf59bc9STejun Heo } 221fbf59bc9STejun Heo 22288999a89STejun Heo /* set the pointer to a chunk in a page struct */ 22388999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 22488999a89STejun Heo { 22588999a89STejun Heo page->index = (unsigned long)pcpu; 22688999a89STejun Heo } 22788999a89STejun Heo 22888999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 22988999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 23088999a89STejun Heo { 23188999a89STejun Heo return (struct pcpu_chunk *)page->index; 23288999a89STejun Heo } 23388999a89STejun Heo 23488999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 235fbf59bc9STejun Heo { 2362f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 237fbf59bc9STejun Heo } 238fbf59bc9STejun Heo 2399983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 240fbf59bc9STejun Heo unsigned int cpu, int page_idx) 241fbf59bc9STejun Heo { 242bba174f5STejun Heo return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + 243fb435d52STejun Heo (page_idx << PAGE_SHIFT); 244fbf59bc9STejun Heo } 245fbf59bc9STejun Heo 24688999a89STejun Heo static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, 24788999a89STejun Heo int *rs, int *re, int end) 248ce3141a2STejun Heo { 249ce3141a2STejun Heo *rs = find_next_zero_bit(chunk->populated, end, *rs); 250ce3141a2STejun Heo *re = find_next_bit(chunk->populated, end, *rs + 1); 251ce3141a2STejun Heo } 252ce3141a2STejun Heo 25388999a89STejun Heo static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, 25488999a89STejun Heo int *rs, int *re, int end) 255ce3141a2STejun Heo { 256ce3141a2STejun Heo *rs = find_next_bit(chunk->populated, end, *rs); 257ce3141a2STejun Heo *re = find_next_zero_bit(chunk->populated, end, *rs + 1); 258ce3141a2STejun Heo } 259ce3141a2STejun Heo 260ce3141a2STejun Heo /* 261ce3141a2STejun Heo * (Un)populated page region iterators. Iterate over (un)populated 262b595076aSUwe Kleine-König * page regions between @start and @end in @chunk. @rs and @re should 263ce3141a2STejun Heo * be integer variables and will be set to start and end page index of 264ce3141a2STejun Heo * the current region. 265ce3141a2STejun Heo */ 266ce3141a2STejun Heo #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ 267ce3141a2STejun Heo for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ 268ce3141a2STejun Heo (rs) < (re); \ 269ce3141a2STejun Heo (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) 270ce3141a2STejun Heo 271ce3141a2STejun Heo #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ 272ce3141a2STejun Heo for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ 273ce3141a2STejun Heo (rs) < (re); \ 274ce3141a2STejun Heo (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) 275ce3141a2STejun Heo 276fbf59bc9STejun Heo /** 27790459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 2781880d93bSTejun Heo * @size: bytes to allocate 279fbf59bc9STejun Heo * 2801880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 28190459ce0SBob Liu * kzalloc() is used; otherwise, vzalloc() is used. The returned 2821880d93bSTejun Heo * memory is always zeroed. 283fbf59bc9STejun Heo * 284ccea34b5STejun Heo * CONTEXT: 285ccea34b5STejun Heo * Does GFP_KERNEL allocation. 286ccea34b5STejun Heo * 287fbf59bc9STejun Heo * RETURNS: 2881880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 289fbf59bc9STejun Heo */ 29090459ce0SBob Liu static void *pcpu_mem_zalloc(size_t size) 291fbf59bc9STejun Heo { 292099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 293099a19d9STejun Heo return NULL; 294099a19d9STejun Heo 295fbf59bc9STejun Heo if (size <= PAGE_SIZE) 2961880d93bSTejun Heo return kzalloc(size, GFP_KERNEL); 2977af4c093SJesper Juhl else 2987af4c093SJesper Juhl return vzalloc(size); 2991880d93bSTejun Heo } 300fbf59bc9STejun Heo 3011880d93bSTejun Heo /** 3021880d93bSTejun Heo * pcpu_mem_free - free memory 3031880d93bSTejun Heo * @ptr: memory to free 3041880d93bSTejun Heo * @size: size of the area 3051880d93bSTejun Heo * 30690459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 3071880d93bSTejun Heo */ 3081880d93bSTejun Heo static void pcpu_mem_free(void *ptr, size_t size) 3091880d93bSTejun Heo { 3101880d93bSTejun Heo if (size <= PAGE_SIZE) 3111880d93bSTejun Heo kfree(ptr); 3121880d93bSTejun Heo else 3131880d93bSTejun Heo vfree(ptr); 314fbf59bc9STejun Heo } 315fbf59bc9STejun Heo 316fbf59bc9STejun Heo /** 317fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 318fbf59bc9STejun Heo * @chunk: chunk of interest 319fbf59bc9STejun Heo * @oslot: the previous slot it was on 320fbf59bc9STejun Heo * 321fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 322fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 323edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 324edcb4639STejun Heo * chunk slots. 325ccea34b5STejun Heo * 326ccea34b5STejun Heo * CONTEXT: 327ccea34b5STejun Heo * pcpu_lock. 328fbf59bc9STejun Heo */ 329fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 330fbf59bc9STejun Heo { 331fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 332fbf59bc9STejun Heo 333edcb4639STejun Heo if (chunk != pcpu_reserved_chunk && oslot != nslot) { 334fbf59bc9STejun Heo if (oslot < nslot) 335fbf59bc9STejun Heo list_move(&chunk->list, &pcpu_slot[nslot]); 336fbf59bc9STejun Heo else 337fbf59bc9STejun Heo list_move_tail(&chunk->list, &pcpu_slot[nslot]); 338fbf59bc9STejun Heo } 339fbf59bc9STejun Heo } 340fbf59bc9STejun Heo 341fbf59bc9STejun Heo /** 342833af842STejun Heo * pcpu_need_to_extend - determine whether chunk area map needs to be extended 343833af842STejun Heo * @chunk: chunk of interest 3449f7dcf22STejun Heo * 345833af842STejun Heo * Determine whether area map of @chunk needs to be extended to 34625985edcSLucas De Marchi * accommodate a new allocation. 3479f7dcf22STejun Heo * 348ccea34b5STejun Heo * CONTEXT: 349833af842STejun Heo * pcpu_lock. 350ccea34b5STejun Heo * 3519f7dcf22STejun Heo * RETURNS: 352833af842STejun Heo * New target map allocation length if extension is necessary, 0 353833af842STejun Heo * otherwise. 3549f7dcf22STejun Heo */ 355833af842STejun Heo static int pcpu_need_to_extend(struct pcpu_chunk *chunk) 3569f7dcf22STejun Heo { 3579f7dcf22STejun Heo int new_alloc; 3589f7dcf22STejun Heo 3599f7dcf22STejun Heo if (chunk->map_alloc >= chunk->map_used + 2) 3609f7dcf22STejun Heo return 0; 3619f7dcf22STejun Heo 3629f7dcf22STejun Heo new_alloc = PCPU_DFL_MAP_ALLOC; 3639f7dcf22STejun Heo while (new_alloc < chunk->map_used + 2) 3649f7dcf22STejun Heo new_alloc *= 2; 3659f7dcf22STejun Heo 366833af842STejun Heo return new_alloc; 367ccea34b5STejun Heo } 368ccea34b5STejun Heo 369833af842STejun Heo /** 370833af842STejun Heo * pcpu_extend_area_map - extend area map of a chunk 371833af842STejun Heo * @chunk: chunk of interest 372833af842STejun Heo * @new_alloc: new target allocation length of the area map 373833af842STejun Heo * 374833af842STejun Heo * Extend area map of @chunk to have @new_alloc entries. 375833af842STejun Heo * 376833af842STejun Heo * CONTEXT: 377833af842STejun Heo * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. 378833af842STejun Heo * 379833af842STejun Heo * RETURNS: 380833af842STejun Heo * 0 on success, -errno on failure. 381ccea34b5STejun Heo */ 382833af842STejun Heo static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) 383833af842STejun Heo { 384833af842STejun Heo int *old = NULL, *new = NULL; 385833af842STejun Heo size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); 386833af842STejun Heo unsigned long flags; 3879f7dcf22STejun Heo 38890459ce0SBob Liu new = pcpu_mem_zalloc(new_size); 389833af842STejun Heo if (!new) 390833af842STejun Heo return -ENOMEM; 391833af842STejun Heo 392833af842STejun Heo /* acquire pcpu_lock and switch to new area map */ 393833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 394833af842STejun Heo 395833af842STejun Heo if (new_alloc <= chunk->map_alloc) 396833af842STejun Heo goto out_unlock; 397833af842STejun Heo 398833af842STejun Heo old_size = chunk->map_alloc * sizeof(chunk->map[0]); 399a002d148SHuang Shijie old = chunk->map; 400a002d148SHuang Shijie 401a002d148SHuang Shijie memcpy(new, old, old_size); 4029f7dcf22STejun Heo 4039f7dcf22STejun Heo chunk->map_alloc = new_alloc; 4049f7dcf22STejun Heo chunk->map = new; 405833af842STejun Heo new = NULL; 406833af842STejun Heo 407833af842STejun Heo out_unlock: 408833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 409833af842STejun Heo 410833af842STejun Heo /* 411833af842STejun Heo * pcpu_mem_free() might end up calling vfree() which uses 412833af842STejun Heo * IRQ-unsafe lock and thus can't be called under pcpu_lock. 413833af842STejun Heo */ 414833af842STejun Heo pcpu_mem_free(old, old_size); 415833af842STejun Heo pcpu_mem_free(new, new_size); 416833af842STejun Heo 4179f7dcf22STejun Heo return 0; 4189f7dcf22STejun Heo } 4199f7dcf22STejun Heo 4209f7dcf22STejun Heo /** 421fbf59bc9STejun Heo * pcpu_split_block - split a map block 422fbf59bc9STejun Heo * @chunk: chunk of interest 423fbf59bc9STejun Heo * @i: index of map block to split 424cae3aeb8STejun Heo * @head: head size in bytes (can be 0) 425cae3aeb8STejun Heo * @tail: tail size in bytes (can be 0) 426fbf59bc9STejun Heo * 427fbf59bc9STejun Heo * Split the @i'th map block into two or three blocks. If @head is 428fbf59bc9STejun Heo * non-zero, @head bytes block is inserted before block @i moving it 429fbf59bc9STejun Heo * to @i+1 and reducing its size by @head bytes. 430fbf59bc9STejun Heo * 431fbf59bc9STejun Heo * If @tail is non-zero, the target block, which can be @i or @i+1 432fbf59bc9STejun Heo * depending on @head, is reduced by @tail bytes and @tail byte block 433fbf59bc9STejun Heo * is inserted after the target block. 434fbf59bc9STejun Heo * 43525985edcSLucas De Marchi * @chunk->map must have enough free slots to accommodate the split. 436ccea34b5STejun Heo * 437ccea34b5STejun Heo * CONTEXT: 438ccea34b5STejun Heo * pcpu_lock. 439fbf59bc9STejun Heo */ 4409f7dcf22STejun Heo static void pcpu_split_block(struct pcpu_chunk *chunk, int i, 4419f7dcf22STejun Heo int head, int tail) 442fbf59bc9STejun Heo { 443fbf59bc9STejun Heo int nr_extra = !!head + !!tail; 444fbf59bc9STejun Heo 4459f7dcf22STejun Heo BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); 446fbf59bc9STejun Heo 4479f7dcf22STejun Heo /* insert new subblocks */ 448fbf59bc9STejun Heo memmove(&chunk->map[i + nr_extra], &chunk->map[i], 449fbf59bc9STejun Heo sizeof(chunk->map[0]) * (chunk->map_used - i)); 450fbf59bc9STejun Heo chunk->map_used += nr_extra; 451fbf59bc9STejun Heo 452fbf59bc9STejun Heo if (head) { 453fbf59bc9STejun Heo chunk->map[i + 1] = chunk->map[i] - head; 454fbf59bc9STejun Heo chunk->map[i++] = head; 455fbf59bc9STejun Heo } 456fbf59bc9STejun Heo if (tail) { 457fbf59bc9STejun Heo chunk->map[i++] -= tail; 458fbf59bc9STejun Heo chunk->map[i] = tail; 459fbf59bc9STejun Heo } 460fbf59bc9STejun Heo } 461fbf59bc9STejun Heo 462fbf59bc9STejun Heo /** 463fbf59bc9STejun Heo * pcpu_alloc_area - allocate area from a pcpu_chunk 464fbf59bc9STejun Heo * @chunk: chunk of interest 465cae3aeb8STejun Heo * @size: wanted size in bytes 466fbf59bc9STejun Heo * @align: wanted align 467fbf59bc9STejun Heo * 468fbf59bc9STejun Heo * Try to allocate @size bytes area aligned at @align from @chunk. 469fbf59bc9STejun Heo * Note that this function only allocates the offset. It doesn't 470fbf59bc9STejun Heo * populate or map the area. 471fbf59bc9STejun Heo * 4729f7dcf22STejun Heo * @chunk->map must have at least two free slots. 4739f7dcf22STejun Heo * 474ccea34b5STejun Heo * CONTEXT: 475ccea34b5STejun Heo * pcpu_lock. 476ccea34b5STejun Heo * 477fbf59bc9STejun Heo * RETURNS: 4789f7dcf22STejun Heo * Allocated offset in @chunk on success, -1 if no matching area is 4799f7dcf22STejun Heo * found. 480fbf59bc9STejun Heo */ 481fbf59bc9STejun Heo static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) 482fbf59bc9STejun Heo { 483fbf59bc9STejun Heo int oslot = pcpu_chunk_slot(chunk); 484fbf59bc9STejun Heo int max_contig = 0; 485fbf59bc9STejun Heo int i, off; 486fbf59bc9STejun Heo 487fbf59bc9STejun Heo for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { 488fbf59bc9STejun Heo bool is_last = i + 1 == chunk->map_used; 489fbf59bc9STejun Heo int head, tail; 490fbf59bc9STejun Heo 491fbf59bc9STejun Heo /* extra for alignment requirement */ 492fbf59bc9STejun Heo head = ALIGN(off, align) - off; 493fbf59bc9STejun Heo BUG_ON(i == 0 && head != 0); 494fbf59bc9STejun Heo 495fbf59bc9STejun Heo if (chunk->map[i] < 0) 496fbf59bc9STejun Heo continue; 497fbf59bc9STejun Heo if (chunk->map[i] < head + size) { 498fbf59bc9STejun Heo max_contig = max(chunk->map[i], max_contig); 499fbf59bc9STejun Heo continue; 500fbf59bc9STejun Heo } 501fbf59bc9STejun Heo 502fbf59bc9STejun Heo /* 503fbf59bc9STejun Heo * If head is small or the previous block is free, 504fbf59bc9STejun Heo * merge'em. Note that 'small' is defined as smaller 505fbf59bc9STejun Heo * than sizeof(int), which is very small but isn't too 506fbf59bc9STejun Heo * uncommon for percpu allocations. 507fbf59bc9STejun Heo */ 508fbf59bc9STejun Heo if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { 509fbf59bc9STejun Heo if (chunk->map[i - 1] > 0) 510fbf59bc9STejun Heo chunk->map[i - 1] += head; 511fbf59bc9STejun Heo else { 512fbf59bc9STejun Heo chunk->map[i - 1] -= head; 513fbf59bc9STejun Heo chunk->free_size -= head; 514fbf59bc9STejun Heo } 515fbf59bc9STejun Heo chunk->map[i] -= head; 516fbf59bc9STejun Heo off += head; 517fbf59bc9STejun Heo head = 0; 518fbf59bc9STejun Heo } 519fbf59bc9STejun Heo 520fbf59bc9STejun Heo /* if tail is small, just keep it around */ 521fbf59bc9STejun Heo tail = chunk->map[i] - head - size; 522fbf59bc9STejun Heo if (tail < sizeof(int)) 523fbf59bc9STejun Heo tail = 0; 524fbf59bc9STejun Heo 525fbf59bc9STejun Heo /* split if warranted */ 526fbf59bc9STejun Heo if (head || tail) { 5279f7dcf22STejun Heo pcpu_split_block(chunk, i, head, tail); 528fbf59bc9STejun Heo if (head) { 529fbf59bc9STejun Heo i++; 530fbf59bc9STejun Heo off += head; 531fbf59bc9STejun Heo max_contig = max(chunk->map[i - 1], max_contig); 532fbf59bc9STejun Heo } 533fbf59bc9STejun Heo if (tail) 534fbf59bc9STejun Heo max_contig = max(chunk->map[i + 1], max_contig); 535fbf59bc9STejun Heo } 536fbf59bc9STejun Heo 537fbf59bc9STejun Heo /* update hint and mark allocated */ 538fbf59bc9STejun Heo if (is_last) 539fbf59bc9STejun Heo chunk->contig_hint = max_contig; /* fully scanned */ 540fbf59bc9STejun Heo else 541fbf59bc9STejun Heo chunk->contig_hint = max(chunk->contig_hint, 542fbf59bc9STejun Heo max_contig); 543fbf59bc9STejun Heo 544fbf59bc9STejun Heo chunk->free_size -= chunk->map[i]; 545fbf59bc9STejun Heo chunk->map[i] = -chunk->map[i]; 546fbf59bc9STejun Heo 547fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 548fbf59bc9STejun Heo return off; 549fbf59bc9STejun Heo } 550fbf59bc9STejun Heo 551fbf59bc9STejun Heo chunk->contig_hint = max_contig; /* fully scanned */ 552fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 553fbf59bc9STejun Heo 5549f7dcf22STejun Heo /* tell the upper layer that this chunk has no matching area */ 5559f7dcf22STejun Heo return -1; 556fbf59bc9STejun Heo } 557fbf59bc9STejun Heo 558fbf59bc9STejun Heo /** 559fbf59bc9STejun Heo * pcpu_free_area - free area to a pcpu_chunk 560fbf59bc9STejun Heo * @chunk: chunk of interest 561fbf59bc9STejun Heo * @freeme: offset of area to free 562fbf59bc9STejun Heo * 563fbf59bc9STejun Heo * Free area starting from @freeme to @chunk. Note that this function 564fbf59bc9STejun Heo * only modifies the allocation map. It doesn't depopulate or unmap 565fbf59bc9STejun Heo * the area. 566ccea34b5STejun Heo * 567ccea34b5STejun Heo * CONTEXT: 568ccea34b5STejun Heo * pcpu_lock. 569fbf59bc9STejun Heo */ 570fbf59bc9STejun Heo static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) 571fbf59bc9STejun Heo { 572fbf59bc9STejun Heo int oslot = pcpu_chunk_slot(chunk); 573fbf59bc9STejun Heo int i, off; 574fbf59bc9STejun Heo 575fbf59bc9STejun Heo for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) 576fbf59bc9STejun Heo if (off == freeme) 577fbf59bc9STejun Heo break; 578fbf59bc9STejun Heo BUG_ON(off != freeme); 579fbf59bc9STejun Heo BUG_ON(chunk->map[i] > 0); 580fbf59bc9STejun Heo 581fbf59bc9STejun Heo chunk->map[i] = -chunk->map[i]; 582fbf59bc9STejun Heo chunk->free_size += chunk->map[i]; 583fbf59bc9STejun Heo 584fbf59bc9STejun Heo /* merge with previous? */ 585fbf59bc9STejun Heo if (i > 0 && chunk->map[i - 1] >= 0) { 586fbf59bc9STejun Heo chunk->map[i - 1] += chunk->map[i]; 587fbf59bc9STejun Heo chunk->map_used--; 588fbf59bc9STejun Heo memmove(&chunk->map[i], &chunk->map[i + 1], 589fbf59bc9STejun Heo (chunk->map_used - i) * sizeof(chunk->map[0])); 590fbf59bc9STejun Heo i--; 591fbf59bc9STejun Heo } 592fbf59bc9STejun Heo /* merge with next? */ 593fbf59bc9STejun Heo if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { 594fbf59bc9STejun Heo chunk->map[i] += chunk->map[i + 1]; 595fbf59bc9STejun Heo chunk->map_used--; 596fbf59bc9STejun Heo memmove(&chunk->map[i + 1], &chunk->map[i + 2], 597fbf59bc9STejun Heo (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); 598fbf59bc9STejun Heo } 599fbf59bc9STejun Heo 600fbf59bc9STejun Heo chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); 601fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 602fbf59bc9STejun Heo } 603fbf59bc9STejun Heo 6046081089fSTejun Heo static struct pcpu_chunk *pcpu_alloc_chunk(void) 6056081089fSTejun Heo { 6066081089fSTejun Heo struct pcpu_chunk *chunk; 6076081089fSTejun Heo 60890459ce0SBob Liu chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); 6096081089fSTejun Heo if (!chunk) 6106081089fSTejun Heo return NULL; 6116081089fSTejun Heo 61290459ce0SBob Liu chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * 61390459ce0SBob Liu sizeof(chunk->map[0])); 6146081089fSTejun Heo if (!chunk->map) { 6156081089fSTejun Heo kfree(chunk); 6166081089fSTejun Heo return NULL; 6176081089fSTejun Heo } 6186081089fSTejun Heo 6196081089fSTejun Heo chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 6206081089fSTejun Heo chunk->map[chunk->map_used++] = pcpu_unit_size; 6216081089fSTejun Heo 6226081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 6236081089fSTejun Heo chunk->free_size = pcpu_unit_size; 6246081089fSTejun Heo chunk->contig_hint = pcpu_unit_size; 6256081089fSTejun Heo 6266081089fSTejun Heo return chunk; 6276081089fSTejun Heo } 6286081089fSTejun Heo 6296081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 6306081089fSTejun Heo { 6316081089fSTejun Heo if (!chunk) 6326081089fSTejun Heo return; 6336081089fSTejun Heo pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 634b4916cb1SJoonsoo Kim pcpu_mem_free(chunk, pcpu_chunk_struct_size); 6356081089fSTejun Heo } 6366081089fSTejun Heo 637fbf59bc9STejun Heo /* 6389f645532STejun Heo * Chunk management implementation. 639fbf59bc9STejun Heo * 6409f645532STejun Heo * To allow different implementations, chunk alloc/free and 6419f645532STejun Heo * [de]population are implemented in a separate file which is pulled 6429f645532STejun Heo * into this file and compiled together. The following functions 6439f645532STejun Heo * should be implemented. 644ccea34b5STejun Heo * 6459f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 6469f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 6479f645532STejun Heo * pcpu_create_chunk - create a new chunk 6489f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 6499f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 6509f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 651fbf59bc9STejun Heo */ 6529f645532STejun Heo static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 6539f645532STejun Heo static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 6549f645532STejun Heo static struct pcpu_chunk *pcpu_create_chunk(void); 6559f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 6569f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 6579f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 658fbf59bc9STejun Heo 659b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 660b0c9778bSTejun Heo #include "percpu-km.c" 661b0c9778bSTejun Heo #else 6629f645532STejun Heo #include "percpu-vm.c" 663b0c9778bSTejun Heo #endif 664fbf59bc9STejun Heo 665fbf59bc9STejun Heo /** 66688999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 66788999a89STejun Heo * @addr: address for which the chunk needs to be determined. 66888999a89STejun Heo * 66988999a89STejun Heo * RETURNS: 67088999a89STejun Heo * The address of the found chunk. 67188999a89STejun Heo */ 67288999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 67388999a89STejun Heo { 67488999a89STejun Heo /* is it in the first chunk? */ 67588999a89STejun Heo if (pcpu_addr_in_first_chunk(addr)) { 67688999a89STejun Heo /* is it in the reserved area? */ 67788999a89STejun Heo if (pcpu_addr_in_reserved_chunk(addr)) 67888999a89STejun Heo return pcpu_reserved_chunk; 67988999a89STejun Heo return pcpu_first_chunk; 68088999a89STejun Heo } 68188999a89STejun Heo 68288999a89STejun Heo /* 68388999a89STejun Heo * The address is relative to unit0 which might be unused and 68488999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 68588999a89STejun Heo * current processor before looking it up in the vmalloc 68688999a89STejun Heo * space. Note that any possible cpu id can be used here, so 68788999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 68888999a89STejun Heo */ 68988999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 6909f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 69188999a89STejun Heo } 69288999a89STejun Heo 69388999a89STejun Heo /** 694edcb4639STejun Heo * pcpu_alloc - the percpu allocator 695cae3aeb8STejun Heo * @size: size of area to allocate in bytes 696fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 697edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 698fbf59bc9STejun Heo * 699ccea34b5STejun Heo * Allocate percpu area of @size bytes aligned at @align. 700ccea34b5STejun Heo * 701ccea34b5STejun Heo * CONTEXT: 702ccea34b5STejun Heo * Does GFP_KERNEL allocation. 703fbf59bc9STejun Heo * 704fbf59bc9STejun Heo * RETURNS: 705fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 706fbf59bc9STejun Heo */ 70743cf38ebSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved) 708fbf59bc9STejun Heo { 709f2badb0cSTejun Heo static int warn_limit = 10; 710fbf59bc9STejun Heo struct pcpu_chunk *chunk; 711f2badb0cSTejun Heo const char *err; 712833af842STejun Heo int slot, off, new_alloc; 713403a91b1SJiri Kosina unsigned long flags; 714f528f0b8SCatalin Marinas void __percpu *ptr; 715fbf59bc9STejun Heo 7168d408b4bSTejun Heo if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { 717fbf59bc9STejun Heo WARN(true, "illegal size (%zu) or align (%zu) for " 718fbf59bc9STejun Heo "percpu allocation\n", size, align); 719fbf59bc9STejun Heo return NULL; 720fbf59bc9STejun Heo } 721fbf59bc9STejun Heo 722ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 723403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 724fbf59bc9STejun Heo 725edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 726edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 727edcb4639STejun Heo chunk = pcpu_reserved_chunk; 728833af842STejun Heo 729833af842STejun Heo if (size > chunk->contig_hint) { 730833af842STejun Heo err = "alloc from reserved chunk failed"; 731ccea34b5STejun Heo goto fail_unlock; 732f2badb0cSTejun Heo } 733833af842STejun Heo 734833af842STejun Heo while ((new_alloc = pcpu_need_to_extend(chunk))) { 735833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 736833af842STejun Heo if (pcpu_extend_area_map(chunk, new_alloc) < 0) { 737833af842STejun Heo err = "failed to extend area map of reserved chunk"; 738833af842STejun Heo goto fail_unlock_mutex; 739833af842STejun Heo } 740833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 741833af842STejun Heo } 742833af842STejun Heo 743edcb4639STejun Heo off = pcpu_alloc_area(chunk, size, align); 744edcb4639STejun Heo if (off >= 0) 745edcb4639STejun Heo goto area_found; 746833af842STejun Heo 747f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 748ccea34b5STejun Heo goto fail_unlock; 749edcb4639STejun Heo } 750edcb4639STejun Heo 751ccea34b5STejun Heo restart: 752edcb4639STejun Heo /* search through normal chunks */ 753fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 754fbf59bc9STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 755fbf59bc9STejun Heo if (size > chunk->contig_hint) 756fbf59bc9STejun Heo continue; 757ccea34b5STejun Heo 758833af842STejun Heo new_alloc = pcpu_need_to_extend(chunk); 759833af842STejun Heo if (new_alloc) { 760833af842STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 761833af842STejun Heo if (pcpu_extend_area_map(chunk, 762833af842STejun Heo new_alloc) < 0) { 763f2badb0cSTejun Heo err = "failed to extend area map"; 764833af842STejun Heo goto fail_unlock_mutex; 765833af842STejun Heo } 766833af842STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 767833af842STejun Heo /* 768833af842STejun Heo * pcpu_lock has been dropped, need to 769833af842STejun Heo * restart cpu_slot list walking. 770833af842STejun Heo */ 771833af842STejun Heo goto restart; 772ccea34b5STejun Heo } 773ccea34b5STejun Heo 774fbf59bc9STejun Heo off = pcpu_alloc_area(chunk, size, align); 775fbf59bc9STejun Heo if (off >= 0) 776fbf59bc9STejun Heo goto area_found; 777fbf59bc9STejun Heo } 778fbf59bc9STejun Heo } 779fbf59bc9STejun Heo 780fbf59bc9STejun Heo /* hmmm... no space left, create a new chunk */ 781403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 782ccea34b5STejun Heo 7836081089fSTejun Heo chunk = pcpu_create_chunk(); 784f2badb0cSTejun Heo if (!chunk) { 785f2badb0cSTejun Heo err = "failed to allocate new chunk"; 786ccea34b5STejun Heo goto fail_unlock_mutex; 787f2badb0cSTejun Heo } 788ccea34b5STejun Heo 789403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 790fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 791ccea34b5STejun Heo goto restart; 792fbf59bc9STejun Heo 793fbf59bc9STejun Heo area_found: 794403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 795ccea34b5STejun Heo 796fbf59bc9STejun Heo /* populate, map and clear the area */ 797fbf59bc9STejun Heo if (pcpu_populate_chunk(chunk, off, size)) { 798403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 799fbf59bc9STejun Heo pcpu_free_area(chunk, off); 800f2badb0cSTejun Heo err = "failed to populate"; 801ccea34b5STejun Heo goto fail_unlock; 802fbf59bc9STejun Heo } 803fbf59bc9STejun Heo 804ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 805ccea34b5STejun Heo 806bba174f5STejun Heo /* return address relative to base address */ 807f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 808f528f0b8SCatalin Marinas kmemleak_alloc_percpu(ptr, size); 809f528f0b8SCatalin Marinas return ptr; 810ccea34b5STejun Heo 811ccea34b5STejun Heo fail_unlock: 812403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 813ccea34b5STejun Heo fail_unlock_mutex: 814ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 815f2badb0cSTejun Heo if (warn_limit) { 816f2badb0cSTejun Heo pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " 817f2badb0cSTejun Heo "%s\n", size, align, err); 818f2badb0cSTejun Heo dump_stack(); 819f2badb0cSTejun Heo if (!--warn_limit) 820f2badb0cSTejun Heo pr_info("PERCPU: limit reached, disable warning\n"); 821f2badb0cSTejun Heo } 822ccea34b5STejun Heo return NULL; 823fbf59bc9STejun Heo } 824edcb4639STejun Heo 825edcb4639STejun Heo /** 826edcb4639STejun Heo * __alloc_percpu - allocate dynamic percpu area 827edcb4639STejun Heo * @size: size of area to allocate in bytes 828edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 829edcb4639STejun Heo * 8309329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. 8319329ba97STejun Heo * Might sleep. Might trigger writeouts. 832edcb4639STejun Heo * 833ccea34b5STejun Heo * CONTEXT: 834ccea34b5STejun Heo * Does GFP_KERNEL allocation. 835ccea34b5STejun Heo * 836edcb4639STejun Heo * RETURNS: 837edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 838edcb4639STejun Heo */ 83943cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 840edcb4639STejun Heo { 841edcb4639STejun Heo return pcpu_alloc(size, align, false); 842edcb4639STejun Heo } 843fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 844fbf59bc9STejun Heo 845edcb4639STejun Heo /** 846edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 847edcb4639STejun Heo * @size: size of area to allocate in bytes 848edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 849edcb4639STejun Heo * 8509329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 8519329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 8529329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 8539329ba97STejun Heo * Might trigger writeouts. 854edcb4639STejun Heo * 855ccea34b5STejun Heo * CONTEXT: 856ccea34b5STejun Heo * Does GFP_KERNEL allocation. 857ccea34b5STejun Heo * 858edcb4639STejun Heo * RETURNS: 859edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 860edcb4639STejun Heo */ 86143cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 862edcb4639STejun Heo { 863edcb4639STejun Heo return pcpu_alloc(size, align, true); 864edcb4639STejun Heo } 865edcb4639STejun Heo 866a56dbddfSTejun Heo /** 867a56dbddfSTejun Heo * pcpu_reclaim - reclaim fully free chunks, workqueue function 868a56dbddfSTejun Heo * @work: unused 869a56dbddfSTejun Heo * 870a56dbddfSTejun Heo * Reclaim all fully free chunks except for the first one. 871ccea34b5STejun Heo * 872ccea34b5STejun Heo * CONTEXT: 873ccea34b5STejun Heo * workqueue context. 874a56dbddfSTejun Heo */ 875a56dbddfSTejun Heo static void pcpu_reclaim(struct work_struct *work) 876fbf59bc9STejun Heo { 877a56dbddfSTejun Heo LIST_HEAD(todo); 878a56dbddfSTejun Heo struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; 879a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 880a56dbddfSTejun Heo 881ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 882ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 883a56dbddfSTejun Heo 884a56dbddfSTejun Heo list_for_each_entry_safe(chunk, next, head, list) { 8858d408b4bSTejun Heo WARN_ON(chunk->immutable); 886a56dbddfSTejun Heo 887a56dbddfSTejun Heo /* spare the first one */ 888a56dbddfSTejun Heo if (chunk == list_first_entry(head, struct pcpu_chunk, list)) 889a56dbddfSTejun Heo continue; 890a56dbddfSTejun Heo 891a56dbddfSTejun Heo list_move(&chunk->list, &todo); 892a56dbddfSTejun Heo } 893a56dbddfSTejun Heo 894ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 895a56dbddfSTejun Heo 896a56dbddfSTejun Heo list_for_each_entry_safe(chunk, next, &todo, list) { 897ce3141a2STejun Heo pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); 8986081089fSTejun Heo pcpu_destroy_chunk(chunk); 899fbf59bc9STejun Heo } 900971f3918STejun Heo 901971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 902a56dbddfSTejun Heo } 903fbf59bc9STejun Heo 904fbf59bc9STejun Heo /** 905fbf59bc9STejun Heo * free_percpu - free percpu area 906fbf59bc9STejun Heo * @ptr: pointer to area to free 907fbf59bc9STejun Heo * 908ccea34b5STejun Heo * Free percpu area @ptr. 909ccea34b5STejun Heo * 910ccea34b5STejun Heo * CONTEXT: 911ccea34b5STejun Heo * Can be called from atomic context. 912fbf59bc9STejun Heo */ 91343cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 914fbf59bc9STejun Heo { 915129182e5SAndrew Morton void *addr; 916fbf59bc9STejun Heo struct pcpu_chunk *chunk; 917ccea34b5STejun Heo unsigned long flags; 918fbf59bc9STejun Heo int off; 919fbf59bc9STejun Heo 920fbf59bc9STejun Heo if (!ptr) 921fbf59bc9STejun Heo return; 922fbf59bc9STejun Heo 923f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 924f528f0b8SCatalin Marinas 925129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 926129182e5SAndrew Morton 927ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 928fbf59bc9STejun Heo 929fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 930bba174f5STejun Heo off = addr - chunk->base_addr; 931fbf59bc9STejun Heo 932fbf59bc9STejun Heo pcpu_free_area(chunk, off); 933fbf59bc9STejun Heo 934a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 935fbf59bc9STejun Heo if (chunk->free_size == pcpu_unit_size) { 936fbf59bc9STejun Heo struct pcpu_chunk *pos; 937fbf59bc9STejun Heo 938a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 939fbf59bc9STejun Heo if (pos != chunk) { 940a56dbddfSTejun Heo schedule_work(&pcpu_reclaim_work); 941fbf59bc9STejun Heo break; 942fbf59bc9STejun Heo } 943fbf59bc9STejun Heo } 944fbf59bc9STejun Heo 945ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 946fbf59bc9STejun Heo } 947fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 948fbf59bc9STejun Heo 9493b034b0dSVivek Goyal /** 95010fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 95110fad5e4STejun Heo * @addr: address to test 95210fad5e4STejun Heo * 95310fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 95410fad5e4STejun Heo * static percpu areas are not considered. For those, use 95510fad5e4STejun Heo * is_module_percpu_address(). 95610fad5e4STejun Heo * 95710fad5e4STejun Heo * RETURNS: 95810fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 95910fad5e4STejun Heo */ 96010fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 96110fad5e4STejun Heo { 962bbddff05STejun Heo #ifdef CONFIG_SMP 96310fad5e4STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 96410fad5e4STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 96510fad5e4STejun Heo unsigned int cpu; 96610fad5e4STejun Heo 96710fad5e4STejun Heo for_each_possible_cpu(cpu) { 96810fad5e4STejun Heo void *start = per_cpu_ptr(base, cpu); 96910fad5e4STejun Heo 97010fad5e4STejun Heo if ((void *)addr >= start && (void *)addr < start + static_size) 97110fad5e4STejun Heo return true; 97210fad5e4STejun Heo } 973bbddff05STejun Heo #endif 974bbddff05STejun Heo /* on UP, can't distinguish from other static vars, always false */ 97510fad5e4STejun Heo return false; 97610fad5e4STejun Heo } 97710fad5e4STejun Heo 97810fad5e4STejun Heo /** 9793b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 9803b034b0dSVivek Goyal * @addr: the address to be converted to physical address 9813b034b0dSVivek Goyal * 9823b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 9833b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 9843b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 9853b034b0dSVivek Goyal * until this function finishes. 9863b034b0dSVivek Goyal * 98767589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 98867589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 98967589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 99067589c71SDave Young * km) provides translation. 99167589c71SDave Young * 99267589c71SDave Young * The addr can be tranlated simply without checking if it falls into the 99367589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 99467589c71SDave Young * actually works, and the verification can discover both bugs in percpu 99567589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 99667589c71SDave Young * code. 99767589c71SDave Young * 9983b034b0dSVivek Goyal * RETURNS: 9993b034b0dSVivek Goyal * The physical address for @addr. 10003b034b0dSVivek Goyal */ 10013b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 10023b034b0dSVivek Goyal { 10039983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 10049983b6f0STejun Heo bool in_first_chunk = false; 1005a855b84cSTejun Heo unsigned long first_low, first_high; 10069983b6f0STejun Heo unsigned int cpu; 10079983b6f0STejun Heo 10089983b6f0STejun Heo /* 1009a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 10109983b6f0STejun Heo * necessary but will speed up lookups of addresses which 10119983b6f0STejun Heo * aren't in the first chunk. 10129983b6f0STejun Heo */ 1013a855b84cSTejun Heo first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0); 1014a855b84cSTejun Heo first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu, 10159983b6f0STejun Heo pcpu_unit_pages); 1016a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1017a855b84cSTejun Heo (unsigned long)addr < first_high) { 10189983b6f0STejun Heo for_each_possible_cpu(cpu) { 10199983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 10209983b6f0STejun Heo 10219983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 10229983b6f0STejun Heo in_first_chunk = true; 10239983b6f0STejun Heo break; 10249983b6f0STejun Heo } 10259983b6f0STejun Heo } 10269983b6f0STejun Heo } 10279983b6f0STejun Heo 10289983b6f0STejun Heo if (in_first_chunk) { 1029eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 10303b034b0dSVivek Goyal return __pa(addr); 10313b034b0dSVivek Goyal else 10329f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 10339f57bd4dSEugene Surovegin offset_in_page(addr); 1034020ec653STejun Heo } else 10359f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 10369f57bd4dSEugene Surovegin offset_in_page(addr); 10373b034b0dSVivek Goyal } 10383b034b0dSVivek Goyal 1039fbf59bc9STejun Heo /** 1040fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1041fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1042fd1e8a1fSTejun Heo * @nr_units: the number of units 1043033e48fbSTejun Heo * 1044fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1045fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1046fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1047fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1048fd1e8a1fSTejun Heo * pointer of other groups. 1049033e48fbSTejun Heo * 1050033e48fbSTejun Heo * RETURNS: 1051fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1052fd1e8a1fSTejun Heo * failure. 1053033e48fbSTejun Heo */ 1054fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1055fd1e8a1fSTejun Heo int nr_units) 1056fd1e8a1fSTejun Heo { 1057fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 1058fd1e8a1fSTejun Heo size_t base_size, ai_size; 1059fd1e8a1fSTejun Heo void *ptr; 1060fd1e8a1fSTejun Heo int unit; 1061fd1e8a1fSTejun Heo 1062fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1063fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 1064fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1065fd1e8a1fSTejun Heo 1066fd1e8a1fSTejun Heo ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); 1067fd1e8a1fSTejun Heo if (!ptr) 1068fd1e8a1fSTejun Heo return NULL; 1069fd1e8a1fSTejun Heo ai = ptr; 1070fd1e8a1fSTejun Heo ptr += base_size; 1071fd1e8a1fSTejun Heo 1072fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 1073fd1e8a1fSTejun Heo 1074fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 1075fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 1076fd1e8a1fSTejun Heo 1077fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 1078fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 1079fd1e8a1fSTejun Heo 1080fd1e8a1fSTejun Heo return ai; 1081fd1e8a1fSTejun Heo } 1082fd1e8a1fSTejun Heo 1083fd1e8a1fSTejun Heo /** 1084fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 1085fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 1086fd1e8a1fSTejun Heo * 1087fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1088fd1e8a1fSTejun Heo */ 1089fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1090fd1e8a1fSTejun Heo { 1091fd1e8a1fSTejun Heo free_bootmem(__pa(ai), ai->__ai_size); 1092fd1e8a1fSTejun Heo } 1093fd1e8a1fSTejun Heo 1094fd1e8a1fSTejun Heo /** 1095fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1096fd1e8a1fSTejun Heo * @lvl: loglevel 1097fd1e8a1fSTejun Heo * @ai: allocation info to dump 1098fd1e8a1fSTejun Heo * 1099fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 1100fd1e8a1fSTejun Heo */ 1101fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 1102fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 1103033e48fbSTejun Heo { 1104fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 1105033e48fbSTejun Heo char empty_str[] = "--------"; 1106fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 1107fd1e8a1fSTejun Heo int group, v; 1108fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 1109033e48fbSTejun Heo 1110fd1e8a1fSTejun Heo v = ai->nr_groups; 1111033e48fbSTejun Heo while (v /= 10) 1112fd1e8a1fSTejun Heo group_width++; 1113033e48fbSTejun Heo 1114fd1e8a1fSTejun Heo v = num_possible_cpus(); 1115fd1e8a1fSTejun Heo while (v /= 10) 1116fd1e8a1fSTejun Heo cpu_width++; 1117fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1118033e48fbSTejun Heo 1119fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 1120fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 1121fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 1122033e48fbSTejun Heo 1123fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1124fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1125fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1126fd1e8a1fSTejun Heo 1127fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 1128fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 1129fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 1130fd1e8a1fSTejun Heo 1131fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 1132fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 1133fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 1134fd1e8a1fSTejun Heo if (!(alloc % apl)) { 1135cb129820STejun Heo printk(KERN_CONT "\n"); 1136fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 1137033e48fbSTejun Heo } 1138cb129820STejun Heo printk(KERN_CONT "[%0*d] ", group_width, group); 1139fd1e8a1fSTejun Heo 1140fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 1141fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 1142cb129820STejun Heo printk(KERN_CONT "%0*d ", cpu_width, 1143fd1e8a1fSTejun Heo gi->cpu_map[unit]); 1144033e48fbSTejun Heo else 1145cb129820STejun Heo printk(KERN_CONT "%s ", empty_str); 1146033e48fbSTejun Heo } 1147fd1e8a1fSTejun Heo } 1148cb129820STejun Heo printk(KERN_CONT "\n"); 1149033e48fbSTejun Heo } 1150033e48fbSTejun Heo 1151fbf59bc9STejun Heo /** 11528d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 1153fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 115438a6be52STejun Heo * @base_addr: mapped address 1155fbf59bc9STejun Heo * 11568d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 11578d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 115838a6be52STejun Heo * setup path. 11598d408b4bSTejun Heo * 1160fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 1161fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 11628d408b4bSTejun Heo * 1163fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 1164fd1e8a1fSTejun Heo * 1165fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1166edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 1167edcb4639STejun Heo * the first chunk such that it's available only through reserved 1168edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 1169edcb4639STejun Heo * static areas on architectures where the addressing model has 1170edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 1171edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 1172edcb4639STejun Heo * 1173fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 1174fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 1175fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 11766074d5b0STejun Heo * 1177fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1178fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 1179fd1e8a1fSTejun Heo * @ai->dyn_size. 11808d408b4bSTejun Heo * 1181fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 1182fd1e8a1fSTejun Heo * for vm areas. 11838d408b4bSTejun Heo * 1184fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 1185fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 1186fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 1187fd1e8a1fSTejun Heo * 1188fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 1189fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 1190fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 1191fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 1192fd1e8a1fSTejun Heo * all units is assumed. 11938d408b4bSTejun Heo * 119438a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 119538a6be52STejun Heo * copied static data to each unit. 1196fbf59bc9STejun Heo * 1197edcb4639STejun Heo * If the first chunk ends up with both reserved and dynamic areas, it 1198edcb4639STejun Heo * is served by two chunks - one to serve the core static and reserved 1199edcb4639STejun Heo * areas and the other for the dynamic area. They share the same vm 1200edcb4639STejun Heo * and page map but uses different area allocation map to stay away 1201edcb4639STejun Heo * from each other. The latter chunk is circulated in the chunk slots 1202edcb4639STejun Heo * and available for dynamic allocation like any other chunks. 1203edcb4639STejun Heo * 1204fbf59bc9STejun Heo * RETURNS: 1205fb435d52STejun Heo * 0 on success, -errno on failure. 1206fbf59bc9STejun Heo */ 1207fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1208fd1e8a1fSTejun Heo void *base_addr) 1209fbf59bc9STejun Heo { 1210635b75fcSTejun Heo static char cpus_buf[4096] __initdata; 1211099a19d9STejun Heo static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1212099a19d9STejun Heo static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; 1213fd1e8a1fSTejun Heo size_t dyn_size = ai->dyn_size; 1214fd1e8a1fSTejun Heo size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; 1215edcb4639STejun Heo struct pcpu_chunk *schunk, *dchunk = NULL; 12166563297cSTejun Heo unsigned long *group_offsets; 12176563297cSTejun Heo size_t *group_sizes; 1218fb435d52STejun Heo unsigned long *unit_off; 1219fbf59bc9STejun Heo unsigned int cpu; 1220fd1e8a1fSTejun Heo int *unit_map; 1221fd1e8a1fSTejun Heo int group, unit, i; 1222fbf59bc9STejun Heo 1223635b75fcSTejun Heo cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); 1224635b75fcSTejun Heo 1225635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 1226635b75fcSTejun Heo if (unlikely(cond)) { \ 1227635b75fcSTejun Heo pr_emerg("PERCPU: failed to initialize, %s", #cond); \ 1228635b75fcSTejun Heo pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ 1229635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1230635b75fcSTejun Heo BUG(); \ 1231635b75fcSTejun Heo } \ 1232635b75fcSTejun Heo } while (0) 1233635b75fcSTejun Heo 12342f39e637STejun Heo /* sanity checks */ 1235635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1236bbddff05STejun Heo #ifdef CONFIG_SMP 1237635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 12380415b00dSTejun Heo PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK); 1239bbddff05STejun Heo #endif 1240635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 12410415b00dSTejun Heo PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK); 1242635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1243635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); 1244635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1245099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 12469f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 12478d408b4bSTejun Heo 12486563297cSTejun Heo /* process group information and build config tables accordingly */ 12496563297cSTejun Heo group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); 12506563297cSTejun Heo group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); 1251fd1e8a1fSTejun Heo unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); 1252fb435d52STejun Heo unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); 12532f39e637STejun Heo 1254fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1255ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 1256a855b84cSTejun Heo 1257a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 1258a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 12592f39e637STejun Heo 1260fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1261fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 12622f39e637STejun Heo 12636563297cSTejun Heo group_offsets[group] = gi->base_offset; 12646563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 12656563297cSTejun Heo 1266fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 1267fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 1268fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 1269fd1e8a1fSTejun Heo continue; 1270fd1e8a1fSTejun Heo 1271635b75fcSTejun Heo PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); 1272635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1273635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1274fd1e8a1fSTejun Heo 1275fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 1276fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1277fb435d52STejun Heo 1278a855b84cSTejun Heo /* determine low/high unit_cpu */ 1279a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 1280a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 1281a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 1282a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 1283a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 1284a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 12850fc0531eSLinus Torvalds } 12860fc0531eSLinus Torvalds } 1287fd1e8a1fSTejun Heo pcpu_nr_units = unit; 12882f39e637STejun Heo 12892f39e637STejun Heo for_each_possible_cpu(cpu) 1290635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1291635b75fcSTejun Heo 1292635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 1293635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 1294bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 12952f39e637STejun Heo 12966563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 12976563297cSTejun Heo pcpu_group_offsets = group_offsets; 12986563297cSTejun Heo pcpu_group_sizes = group_sizes; 1299fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 1300fb435d52STejun Heo pcpu_unit_offsets = unit_off; 13012f39e637STejun Heo 13022f39e637STejun Heo /* determine basic parameters */ 1303fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1304d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 13056563297cSTejun Heo pcpu_atom_size = ai->atom_size; 1306ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1307ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1308cafe8816STejun Heo 1309d9b55eebSTejun Heo /* 1310d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 1311d9b55eebSTejun Heo * empty chunks. 1312d9b55eebSTejun Heo */ 1313d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1314fbf59bc9STejun Heo pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); 1315fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 1316fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 1317fbf59bc9STejun Heo 1318edcb4639STejun Heo /* 1319edcb4639STejun Heo * Initialize static chunk. If reserved_size is zero, the 1320edcb4639STejun Heo * static chunk covers static area + dynamic allocation area 1321edcb4639STejun Heo * in the first chunk. If reserved_size is not zero, it 1322edcb4639STejun Heo * covers static area + reserved area (mostly used for module 1323edcb4639STejun Heo * static percpu allocation). 1324edcb4639STejun Heo */ 13252441d15cSTejun Heo schunk = alloc_bootmem(pcpu_chunk_struct_size); 13262441d15cSTejun Heo INIT_LIST_HEAD(&schunk->list); 1327bba174f5STejun Heo schunk->base_addr = base_addr; 132861ace7faSTejun Heo schunk->map = smap; 132961ace7faSTejun Heo schunk->map_alloc = ARRAY_SIZE(smap); 133038a6be52STejun Heo schunk->immutable = true; 1331ce3141a2STejun Heo bitmap_fill(schunk->populated, pcpu_unit_pages); 1332edcb4639STejun Heo 1333fd1e8a1fSTejun Heo if (ai->reserved_size) { 1334fd1e8a1fSTejun Heo schunk->free_size = ai->reserved_size; 1335ae9e6bc9STejun Heo pcpu_reserved_chunk = schunk; 1336fd1e8a1fSTejun Heo pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; 1337edcb4639STejun Heo } else { 13382441d15cSTejun Heo schunk->free_size = dyn_size; 1339edcb4639STejun Heo dyn_size = 0; /* dynamic area covered */ 1340edcb4639STejun Heo } 13412441d15cSTejun Heo schunk->contig_hint = schunk->free_size; 1342fbf59bc9STejun Heo 1343fd1e8a1fSTejun Heo schunk->map[schunk->map_used++] = -ai->static_size; 134461ace7faSTejun Heo if (schunk->free_size) 134561ace7faSTejun Heo schunk->map[schunk->map_used++] = schunk->free_size; 134661ace7faSTejun Heo 1347edcb4639STejun Heo /* init dynamic chunk if necessary */ 1348edcb4639STejun Heo if (dyn_size) { 1349ce3141a2STejun Heo dchunk = alloc_bootmem(pcpu_chunk_struct_size); 1350edcb4639STejun Heo INIT_LIST_HEAD(&dchunk->list); 1351bba174f5STejun Heo dchunk->base_addr = base_addr; 1352edcb4639STejun Heo dchunk->map = dmap; 1353edcb4639STejun Heo dchunk->map_alloc = ARRAY_SIZE(dmap); 135438a6be52STejun Heo dchunk->immutable = true; 1355ce3141a2STejun Heo bitmap_fill(dchunk->populated, pcpu_unit_pages); 1356edcb4639STejun Heo 1357edcb4639STejun Heo dchunk->contig_hint = dchunk->free_size = dyn_size; 1358edcb4639STejun Heo dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1359edcb4639STejun Heo dchunk->map[dchunk->map_used++] = dchunk->free_size; 1360edcb4639STejun Heo } 1361edcb4639STejun Heo 13622441d15cSTejun Heo /* link the first chunk in */ 1363ae9e6bc9STejun Heo pcpu_first_chunk = dchunk ?: schunk; 1364ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 1365fbf59bc9STejun Heo 1366fbf59bc9STejun Heo /* we're done */ 1367bba174f5STejun Heo pcpu_base_addr = base_addr; 1368fb435d52STejun Heo return 0; 1369fbf59bc9STejun Heo } 137066c3a757STejun Heo 1371bbddff05STejun Heo #ifdef CONFIG_SMP 1372bbddff05STejun Heo 137317f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 1374f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 1375f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 1376f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 1377f58dc01bSTejun Heo }; 137866c3a757STejun Heo 1379f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1380f58dc01bSTejun Heo 1381f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 138266c3a757STejun Heo { 1383*5479c78aSCyrill Gorcunov if (!str) 1384*5479c78aSCyrill Gorcunov return -EINVAL; 1385*5479c78aSCyrill Gorcunov 1386f58dc01bSTejun Heo if (0) 1387f58dc01bSTejun Heo /* nada */; 1388f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1389f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 1390f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 1391f58dc01bSTejun Heo #endif 1392f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1393f58dc01bSTejun Heo else if (!strcmp(str, "page")) 1394f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 1395f58dc01bSTejun Heo #endif 1396f58dc01bSTejun Heo else 1397f58dc01bSTejun Heo pr_warning("PERCPU: unknown allocator %s specified\n", str); 139866c3a757STejun Heo 1399f58dc01bSTejun Heo return 0; 140066c3a757STejun Heo } 1401f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 140266c3a757STejun Heo 14033c9a024fSTejun Heo /* 14043c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 14053c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 14063c9a024fSTejun Heo * to be used. 14073c9a024fSTejun Heo */ 140808fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 140908fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 14103c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 14113c9a024fSTejun Heo #endif 14123c9a024fSTejun Heo 14133c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 14143c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 14153c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 14163c9a024fSTejun Heo #endif 14173c9a024fSTejun Heo 14183c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 14193c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 14203c9a024fSTejun Heo /** 1421fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 1422fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 1423fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 1424fbf59bc9STejun Heo * @atom_size: allocation atom size 1425fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 1426fbf59bc9STejun Heo * 1427fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 1428fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 1429fbf59bc9STejun Heo * atom size and distances between CPUs. 1430fbf59bc9STejun Heo * 1431fbf59bc9STejun Heo * Groups are always mutliples of atom size and CPUs which are of 1432fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 1433fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 1434fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 1435fbf59bc9STejun Heo * of allocated virtual address space. 1436fbf59bc9STejun Heo * 1437fbf59bc9STejun Heo * RETURNS: 1438fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 1439fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 1440fbf59bc9STejun Heo */ 1441fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 1442fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 1443fbf59bc9STejun Heo size_t atom_size, 1444fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 1445fbf59bc9STejun Heo { 1446fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 1447fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 1448fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 1449fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 1450fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 1451fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 1452fbf59bc9STejun Heo int last_allocs, group, unit; 1453fbf59bc9STejun Heo unsigned int cpu, tcpu; 1454fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 1455fbf59bc9STejun Heo unsigned int *cpu_map; 1456fbf59bc9STejun Heo 1457fbf59bc9STejun Heo /* this function may be called multiple times */ 1458fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 1459fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 1460fbf59bc9STejun Heo 1461fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 1462fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 1463fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 1464fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 1465fbf59bc9STejun Heo 1466fbf59bc9STejun Heo /* 1467fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 1468fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 146925985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 1470fbf59bc9STejun Heo * or larger than min_unit_size. 1471fbf59bc9STejun Heo */ 1472fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 1473fbf59bc9STejun Heo 1474fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 1475fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 1476fbf59bc9STejun Heo while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1477fbf59bc9STejun Heo upa--; 1478fbf59bc9STejun Heo max_upa = upa; 1479fbf59bc9STejun Heo 1480fbf59bc9STejun Heo /* group cpus according to their proximity */ 1481fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 1482fbf59bc9STejun Heo group = 0; 1483fbf59bc9STejun Heo next_group: 1484fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 1485fbf59bc9STejun Heo if (cpu == tcpu) 1486fbf59bc9STejun Heo break; 1487fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 1488fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 1489fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 1490fbf59bc9STejun Heo group++; 1491fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 1492fbf59bc9STejun Heo goto next_group; 1493fbf59bc9STejun Heo } 1494fbf59bc9STejun Heo } 1495fbf59bc9STejun Heo group_map[cpu] = group; 1496fbf59bc9STejun Heo group_cnt[group]++; 1497fbf59bc9STejun Heo } 1498fbf59bc9STejun Heo 1499fbf59bc9STejun Heo /* 1500fbf59bc9STejun Heo * Expand unit size until address space usage goes over 75% 1501fbf59bc9STejun Heo * and then as much as possible without using more address 1502fbf59bc9STejun Heo * space. 1503fbf59bc9STejun Heo */ 1504fbf59bc9STejun Heo last_allocs = INT_MAX; 1505fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 1506fbf59bc9STejun Heo int allocs = 0, wasted = 0; 1507fbf59bc9STejun Heo 1508fbf59bc9STejun Heo if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) 1509fbf59bc9STejun Heo continue; 1510fbf59bc9STejun Heo 1511fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 1512fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 1513fbf59bc9STejun Heo allocs += this_allocs; 1514fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 1515fbf59bc9STejun Heo } 1516fbf59bc9STejun Heo 1517fbf59bc9STejun Heo /* 1518fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 1519fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 1520fbf59bc9STejun Heo * passes the following check. 1521fbf59bc9STejun Heo */ 1522fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 1523fbf59bc9STejun Heo continue; 1524fbf59bc9STejun Heo 1525fbf59bc9STejun Heo /* and then don't consume more memory */ 1526fbf59bc9STejun Heo if (allocs > last_allocs) 1527fbf59bc9STejun Heo break; 1528fbf59bc9STejun Heo last_allocs = allocs; 1529fbf59bc9STejun Heo best_upa = upa; 1530fbf59bc9STejun Heo } 1531fbf59bc9STejun Heo upa = best_upa; 1532fbf59bc9STejun Heo 1533fbf59bc9STejun Heo /* allocate and fill alloc_info */ 1534fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 1535fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 1536fbf59bc9STejun Heo 1537fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 1538fbf59bc9STejun Heo if (!ai) 1539fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 1540fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 1541fbf59bc9STejun Heo 1542fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 1543fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 1544fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 1545fbf59bc9STejun Heo } 1546fbf59bc9STejun Heo 1547fbf59bc9STejun Heo ai->static_size = static_size; 1548fbf59bc9STejun Heo ai->reserved_size = reserved_size; 1549fbf59bc9STejun Heo ai->dyn_size = dyn_size; 1550fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 1551fbf59bc9STejun Heo ai->atom_size = atom_size; 1552fbf59bc9STejun Heo ai->alloc_size = alloc_size; 1553fbf59bc9STejun Heo 1554fbf59bc9STejun Heo for (group = 0, unit = 0; group_cnt[group]; group++) { 1555fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 1556fbf59bc9STejun Heo 1557fbf59bc9STejun Heo /* 1558fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 1559fbf59bc9STejun Heo * back-to-back. The caller should update this to 1560fbf59bc9STejun Heo * reflect actual allocation. 1561fbf59bc9STejun Heo */ 1562fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 1563fbf59bc9STejun Heo 1564fbf59bc9STejun Heo for_each_possible_cpu(cpu) 1565fbf59bc9STejun Heo if (group_map[cpu] == group) 1566fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 1567fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 1568fbf59bc9STejun Heo unit += gi->nr_units; 1569fbf59bc9STejun Heo } 1570fbf59bc9STejun Heo BUG_ON(unit != nr_units); 1571fbf59bc9STejun Heo 1572fbf59bc9STejun Heo return ai; 1573fbf59bc9STejun Heo } 15743c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 1575fbf59bc9STejun Heo 15763c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 157766c3a757STejun Heo /** 157866c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 157966c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 15804ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 1581c8826dd5STejun Heo * @atom_size: allocation atom size 1582c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 1583c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 158425985edcSLucas De Marchi * @free_fn: function to free percpu page 158566c3a757STejun Heo * 158666c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 158766c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 158866c3a757STejun Heo * 158966c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 1590c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 1591c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 1592c8826dd5STejun Heo * aligned to @atom_size. 1593c8826dd5STejun Heo * 1594c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 1595c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 1596c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 1597c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 1598c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 1599c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 160066c3a757STejun Heo * 16014ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 160266c3a757STejun Heo * 160366c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 1604c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 160566c3a757STejun Heo * 160666c3a757STejun Heo * RETURNS: 1607fb435d52STejun Heo * 0 on success, -errno on failure. 160866c3a757STejun Heo */ 16094ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 1610c8826dd5STejun Heo size_t atom_size, 1611c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 1612c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 1613c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 161466c3a757STejun Heo { 1615c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 1616c8826dd5STejun Heo void **areas = NULL; 1617fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 16186ea529a2STejun Heo size_t size_sum, areas_size, max_distance; 1619c8826dd5STejun Heo int group, i, rc; 162066c3a757STejun Heo 1621c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 1622c8826dd5STejun Heo cpu_distance_fn); 1623fd1e8a1fSTejun Heo if (IS_ERR(ai)) 1624fd1e8a1fSTejun Heo return PTR_ERR(ai); 162566c3a757STejun Heo 1626fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1627c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 162866c3a757STejun Heo 1629c8826dd5STejun Heo areas = alloc_bootmem_nopanic(areas_size); 1630c8826dd5STejun Heo if (!areas) { 1631fb435d52STejun Heo rc = -ENOMEM; 1632c8826dd5STejun Heo goto out_free; 1633fa8a7094STejun Heo } 163466c3a757STejun Heo 1635c8826dd5STejun Heo /* allocate, copy and determine base address */ 1636c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 1637c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 1638c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 1639c8826dd5STejun Heo void *ptr; 164066c3a757STejun Heo 1641c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 1642c8826dd5STejun Heo cpu = gi->cpu_map[i]; 1643c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 1644c8826dd5STejun Heo 1645c8826dd5STejun Heo /* allocate space for the whole group */ 1646c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 1647c8826dd5STejun Heo if (!ptr) { 1648c8826dd5STejun Heo rc = -ENOMEM; 1649c8826dd5STejun Heo goto out_free_areas; 1650c8826dd5STejun Heo } 1651f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 1652f528f0b8SCatalin Marinas kmemleak_free(ptr); 1653c8826dd5STejun Heo areas[group] = ptr; 1654c8826dd5STejun Heo 1655c8826dd5STejun Heo base = min(ptr, base); 165642b64281STejun Heo } 165742b64281STejun Heo 165842b64281STejun Heo /* 165942b64281STejun Heo * Copy data and free unused parts. This should happen after all 166042b64281STejun Heo * allocations are complete; otherwise, we may end up with 166142b64281STejun Heo * overlapping groups. 166242b64281STejun Heo */ 166342b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 166442b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 166542b64281STejun Heo void *ptr = areas[group]; 1666c8826dd5STejun Heo 1667c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 1668c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 1669c8826dd5STejun Heo /* unused unit, free whole */ 1670c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 1671c8826dd5STejun Heo continue; 1672c8826dd5STejun Heo } 1673c8826dd5STejun Heo /* copy and return the unused part */ 1674fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 1675c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 1676c8826dd5STejun Heo } 167766c3a757STejun Heo } 167866c3a757STejun Heo 1679c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 16806ea529a2STejun Heo max_distance = 0; 16816ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 1682c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 16831a0c3298STejun Heo max_distance = max_t(size_t, max_distance, 16841a0c3298STejun Heo ai->groups[group].base_offset); 16856ea529a2STejun Heo } 16866ea529a2STejun Heo max_distance += ai->unit_size; 16876ea529a2STejun Heo 16886ea529a2STejun Heo /* warn if maximum distance is further than 75% of vmalloc space */ 16896ea529a2STejun Heo if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { 16901a0c3298STejun Heo pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " 1691787e5b06SMike Frysinger "space 0x%lx\n", max_distance, 1692787e5b06SMike Frysinger (unsigned long)(VMALLOC_END - VMALLOC_START)); 16936ea529a2STejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 16946ea529a2STejun Heo /* and fail if we have fallback */ 16956ea529a2STejun Heo rc = -EINVAL; 16966ea529a2STejun Heo goto out_free; 16976ea529a2STejun Heo #endif 16986ea529a2STejun Heo } 1699c8826dd5STejun Heo 1700004018e2STejun Heo pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 1701fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 1702fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 170366c3a757STejun Heo 1704fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 1705c8826dd5STejun Heo goto out_free; 1706c8826dd5STejun Heo 1707c8826dd5STejun Heo out_free_areas: 1708c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 1709c8826dd5STejun Heo free_fn(areas[group], 1710c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 1711c8826dd5STejun Heo out_free: 1712fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 1713c8826dd5STejun Heo if (areas) 1714c8826dd5STejun Heo free_bootmem(__pa(areas), areas_size); 1715fb435d52STejun Heo return rc; 1716d4b95f80STejun Heo } 17173c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 1718d4b95f80STejun Heo 17193c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 1720d4b95f80STejun Heo /** 172100ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 1722d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 1723d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 172425985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 1725d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 1726d4b95f80STejun Heo * 172700ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 172800ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 1729d4b95f80STejun Heo * 1730d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 1731d4b95f80STejun Heo * page-by-page into vmalloc area. 1732d4b95f80STejun Heo * 1733d4b95f80STejun Heo * RETURNS: 1734fb435d52STejun Heo * 0 on success, -errno on failure. 1735d4b95f80STejun Heo */ 1736fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 1737d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 1738d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 1739d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 1740d4b95f80STejun Heo { 17418f05a6a6STejun Heo static struct vm_struct vm; 1742fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 174300ae4064STejun Heo char psize_str[16]; 1744ce3141a2STejun Heo int unit_pages; 1745d4b95f80STejun Heo size_t pages_size; 1746ce3141a2STejun Heo struct page **pages; 1747fb435d52STejun Heo int unit, i, j, rc; 1748d4b95f80STejun Heo 174900ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 175000ae4064STejun Heo 17514ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 1752fd1e8a1fSTejun Heo if (IS_ERR(ai)) 1753fd1e8a1fSTejun Heo return PTR_ERR(ai); 1754fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 1755fd1e8a1fSTejun Heo BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); 1756fd1e8a1fSTejun Heo 1757fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 1758d4b95f80STejun Heo 1759d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 1760fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 1761fd1e8a1fSTejun Heo sizeof(pages[0])); 1762ce3141a2STejun Heo pages = alloc_bootmem(pages_size); 1763d4b95f80STejun Heo 17648f05a6a6STejun Heo /* allocate pages */ 1765d4b95f80STejun Heo j = 0; 1766fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) 1767ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) { 1768fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 1769d4b95f80STejun Heo void *ptr; 1770d4b95f80STejun Heo 17713cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 1772d4b95f80STejun Heo if (!ptr) { 177300ae4064STejun Heo pr_warning("PERCPU: failed to allocate %s page " 177400ae4064STejun Heo "for cpu%u\n", psize_str, cpu); 1775d4b95f80STejun Heo goto enomem; 1776d4b95f80STejun Heo } 1777f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 1778f528f0b8SCatalin Marinas kmemleak_free(ptr); 1779ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 1780d4b95f80STejun Heo } 1781d4b95f80STejun Heo 17828f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 17838f05a6a6STejun Heo vm.flags = VM_ALLOC; 1784fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 17858f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 17868f05a6a6STejun Heo 1787fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 17881d9d3257STejun Heo unsigned long unit_addr = 1789fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 17908f05a6a6STejun Heo 1791ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 17928f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 17938f05a6a6STejun Heo 17948f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 1795fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 1796ce3141a2STejun Heo unit_pages); 1797fb435d52STejun Heo if (rc < 0) 1798fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 17998f05a6a6STejun Heo 18008f05a6a6STejun Heo /* 18018f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 18028f05a6a6STejun Heo * cache for the linear mapping here - something 18038f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 18048f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 18058f05a6a6STejun Heo * data structures are not set up yet. 18068f05a6a6STejun Heo */ 18078f05a6a6STejun Heo 18088f05a6a6STejun Heo /* copy static data */ 1809fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 181066c3a757STejun Heo } 181166c3a757STejun Heo 181266c3a757STejun Heo /* we're ready, commit */ 18131d9d3257STejun Heo pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", 1814fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 1815fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 181666c3a757STejun Heo 1817fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 1818d4b95f80STejun Heo goto out_free_ar; 1819d4b95f80STejun Heo 1820d4b95f80STejun Heo enomem: 1821d4b95f80STejun Heo while (--j >= 0) 1822ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 1823fb435d52STejun Heo rc = -ENOMEM; 1824d4b95f80STejun Heo out_free_ar: 1825ce3141a2STejun Heo free_bootmem(__pa(pages), pages_size); 1826fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 1827fb435d52STejun Heo return rc; 182866c3a757STejun Heo } 18293c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 1830d4b95f80STejun Heo 1831bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 18328c4bfc6eSTejun Heo /* 1833bbddff05STejun Heo * Generic SMP percpu area setup. 1834e74e3962STejun Heo * 1835e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 1836e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 1837e74e3962STejun Heo * important because many archs have addressing restrictions and might 1838e74e3962STejun Heo * fail if the percpu area is located far away from the previous 1839e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 1840e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 1841e74e3962STejun Heo * on the physical linear memory mapping which uses large page 1842e74e3962STejun Heo * mappings on applicable archs. 1843e74e3962STejun Heo */ 1844e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 1845e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 1846e74e3962STejun Heo 1847c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 1848c8826dd5STejun Heo size_t align) 1849c8826dd5STejun Heo { 1850c8826dd5STejun Heo return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); 1851c8826dd5STejun Heo } 1852c8826dd5STejun Heo 1853c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 1854c8826dd5STejun Heo { 1855c8826dd5STejun Heo free_bootmem(__pa(ptr), size); 1856c8826dd5STejun Heo } 1857c8826dd5STejun Heo 1858e74e3962STejun Heo void __init setup_per_cpu_areas(void) 1859e74e3962STejun Heo { 1860e74e3962STejun Heo unsigned long delta; 1861e74e3962STejun Heo unsigned int cpu; 1862fb435d52STejun Heo int rc; 1863e74e3962STejun Heo 1864e74e3962STejun Heo /* 1865e74e3962STejun Heo * Always reserve area for module percpu variables. That's 1866e74e3962STejun Heo * what the legacy allocator did. 1867e74e3962STejun Heo */ 1868fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 1869c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 1870c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 1871fb435d52STejun Heo if (rc < 0) 1872bbddff05STejun Heo panic("Failed to initialize percpu areas."); 1873e74e3962STejun Heo 1874e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 1875e74e3962STejun Heo for_each_possible_cpu(cpu) 1876fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 1877e74e3962STejun Heo } 1878e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 1879099a19d9STejun Heo 1880bbddff05STejun Heo #else /* CONFIG_SMP */ 1881bbddff05STejun Heo 1882bbddff05STejun Heo /* 1883bbddff05STejun Heo * UP percpu area setup. 1884bbddff05STejun Heo * 1885bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 1886bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 1887bbddff05STejun Heo * variables and don't require any special preparation. 1888bbddff05STejun Heo */ 1889bbddff05STejun Heo void __init setup_per_cpu_areas(void) 1890bbddff05STejun Heo { 1891bbddff05STejun Heo const size_t unit_size = 1892bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 1893bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 1894bbddff05STejun Heo struct pcpu_alloc_info *ai; 1895bbddff05STejun Heo void *fc; 1896bbddff05STejun Heo 1897bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 1898bbddff05STejun Heo fc = __alloc_bootmem(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS)); 1899bbddff05STejun Heo if (!ai || !fc) 1900bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 1901100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 1902100d13c3SCatalin Marinas kmemleak_free(fc); 1903bbddff05STejun Heo 1904bbddff05STejun Heo ai->dyn_size = unit_size; 1905bbddff05STejun Heo ai->unit_size = unit_size; 1906bbddff05STejun Heo ai->atom_size = unit_size; 1907bbddff05STejun Heo ai->alloc_size = unit_size; 1908bbddff05STejun Heo ai->groups[0].nr_units = 1; 1909bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 1910bbddff05STejun Heo 1911bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 1912bbddff05STejun Heo panic("Failed to initialize percpu areas."); 1913bbddff05STejun Heo } 1914bbddff05STejun Heo 1915bbddff05STejun Heo #endif /* CONFIG_SMP */ 1916bbddff05STejun Heo 1917099a19d9STejun Heo /* 1918099a19d9STejun Heo * First and reserved chunks are initialized with temporary allocation 1919099a19d9STejun Heo * map in initdata so that they can be used before slab is online. 1920099a19d9STejun Heo * This function is called after slab is brought up and replaces those 1921099a19d9STejun Heo * with properly allocated maps. 1922099a19d9STejun Heo */ 1923099a19d9STejun Heo void __init percpu_init_late(void) 1924099a19d9STejun Heo { 1925099a19d9STejun Heo struct pcpu_chunk *target_chunks[] = 1926099a19d9STejun Heo { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; 1927099a19d9STejun Heo struct pcpu_chunk *chunk; 1928099a19d9STejun Heo unsigned long flags; 1929099a19d9STejun Heo int i; 1930099a19d9STejun Heo 1931099a19d9STejun Heo for (i = 0; (chunk = target_chunks[i]); i++) { 1932099a19d9STejun Heo int *map; 1933099a19d9STejun Heo const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); 1934099a19d9STejun Heo 1935099a19d9STejun Heo BUILD_BUG_ON(size > PAGE_SIZE); 1936099a19d9STejun Heo 193790459ce0SBob Liu map = pcpu_mem_zalloc(size); 1938099a19d9STejun Heo BUG_ON(!map); 1939099a19d9STejun Heo 1940099a19d9STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1941099a19d9STejun Heo memcpy(map, chunk->map, size); 1942099a19d9STejun Heo chunk->map = map; 1943099a19d9STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1944099a19d9STejun Heo } 1945099a19d9STejun Heo } 1946