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 * 79c015162SDennis Zhou (Facebook) * This file is released under the GPLv2 license. 8fbf59bc9STejun Heo * 99c015162SDennis Zhou (Facebook) * The percpu allocator handles both static and dynamic areas. Percpu 109c015162SDennis Zhou (Facebook) * areas are allocated in chunks which are divided into units. There is 119c015162SDennis Zhou (Facebook) * a 1-to-1 mapping for units to possible cpus. These units are grouped 129c015162SDennis Zhou (Facebook) * based on NUMA properties of the machine. 13fbf59bc9STejun Heo * 14fbf59bc9STejun Heo * c0 c1 c2 15fbf59bc9STejun Heo * ------------------- ------------------- ------------ 16fbf59bc9STejun Heo * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u 17fbf59bc9STejun Heo * ------------------- ...... ------------------- .... ------------ 18fbf59bc9STejun Heo * 199c015162SDennis Zhou (Facebook) * Allocation is done by offsets into a unit's address space. Ie., an 209c015162SDennis Zhou (Facebook) * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, 219c015162SDennis Zhou (Facebook) * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear 229c015162SDennis Zhou (Facebook) * and even sparse. Access is handled by configuring percpu base 239c015162SDennis Zhou (Facebook) * registers according to the cpu to unit mappings and offsetting the 249c015162SDennis Zhou (Facebook) * base address using pcpu_unit_size. 25fbf59bc9STejun Heo * 269c015162SDennis Zhou (Facebook) * There is special consideration for the first chunk which must handle 279c015162SDennis Zhou (Facebook) * the static percpu variables in the kernel image as allocation services 289c015162SDennis Zhou (Facebook) * are not online yet. In short, the first chunk is structure like so: 299c015162SDennis Zhou (Facebook) * 309c015162SDennis Zhou (Facebook) * <Static | [Reserved] | Dynamic> 319c015162SDennis Zhou (Facebook) * 329c015162SDennis Zhou (Facebook) * The static data is copied from the original section managed by the 339c015162SDennis Zhou (Facebook) * linker. The reserved section, if non-zero, primarily manages static 349c015162SDennis Zhou (Facebook) * percpu variables from kernel modules. Finally, the dynamic section 359c015162SDennis Zhou (Facebook) * takes care of normal allocations. 36fbf59bc9STejun Heo * 37fbf59bc9STejun Heo * Allocation state in each chunk is kept using an array of integers 38fbf59bc9STejun Heo * on chunk->map. A positive value in the map represents a free 39fbf59bc9STejun Heo * region and negative allocated. Allocation inside a chunk is done 40fbf59bc9STejun Heo * by scanning this map sequentially and serving the first matching 41fbf59bc9STejun Heo * entry. This is mostly copied from the percpu_modalloc() allocator. 42e1b9aa3fSChristoph Lameter * Chunks can be determined from the address using the index field 43e1b9aa3fSChristoph Lameter * in the page struct. The index field contains a pointer to the chunk. 44fbf59bc9STejun Heo * 459c015162SDennis Zhou (Facebook) * These chunks are organized into lists according to free_size and 469c015162SDennis Zhou (Facebook) * tries to allocate from the fullest chunk first. Each chunk maintains 479c015162SDennis Zhou (Facebook) * a maximum contiguous area size hint which is guaranteed to be equal 489c015162SDennis Zhou (Facebook) * to or larger than the maximum contiguous area in the chunk. This 499c015162SDennis Zhou (Facebook) * helps prevent the allocator from iterating over chunks unnecessarily. 509c015162SDennis Zhou (Facebook) * 514091fb95SMasahiro Yamada * To use this allocator, arch code should do the following: 52fbf59bc9STejun Heo * 53fbf59bc9STejun Heo * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 54e0100983STejun Heo * regular address to percpu pointer and back if they need to be 55e0100983STejun Heo * different from the default 56fbf59bc9STejun Heo * 578d408b4bSTejun Heo * - use pcpu_setup_first_chunk() during percpu area initialization to 588d408b4bSTejun Heo * setup the first chunk containing the kernel static percpu area 59fbf59bc9STejun Heo */ 60fbf59bc9STejun Heo 61870d4b12SJoe Perches #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 62870d4b12SJoe Perches 63fbf59bc9STejun Heo #include <linux/bitmap.h> 64fbf59bc9STejun Heo #include <linux/bootmem.h> 65fd1e8a1fSTejun Heo #include <linux/err.h> 66ca460b3cSDennis Zhou (Facebook) #include <linux/lcm.h> 67fbf59bc9STejun Heo #include <linux/list.h> 68a530b795STejun Heo #include <linux/log2.h> 69fbf59bc9STejun Heo #include <linux/mm.h> 70fbf59bc9STejun Heo #include <linux/module.h> 71fbf59bc9STejun Heo #include <linux/mutex.h> 72fbf59bc9STejun Heo #include <linux/percpu.h> 73fbf59bc9STejun Heo #include <linux/pfn.h> 74fbf59bc9STejun Heo #include <linux/slab.h> 75ccea34b5STejun Heo #include <linux/spinlock.h> 76fbf59bc9STejun Heo #include <linux/vmalloc.h> 77a56dbddfSTejun Heo #include <linux/workqueue.h> 78f528f0b8SCatalin Marinas #include <linux/kmemleak.h> 79fbf59bc9STejun Heo 80fbf59bc9STejun Heo #include <asm/cacheflush.h> 81e0100983STejun Heo #include <asm/sections.h> 82fbf59bc9STejun Heo #include <asm/tlbflush.h> 833b034b0dSVivek Goyal #include <asm/io.h> 84fbf59bc9STejun Heo 85df95e795SDennis Zhou #define CREATE_TRACE_POINTS 86df95e795SDennis Zhou #include <trace/events/percpu.h> 87df95e795SDennis Zhou 888fa3ed80SDennis Zhou #include "percpu-internal.h" 898fa3ed80SDennis Zhou 9040064aecSDennis Zhou (Facebook) /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ 9140064aecSDennis Zhou (Facebook) #define PCPU_SLOT_BASE_SHIFT 5 9240064aecSDennis Zhou (Facebook) 931a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_LOW 2 941a4d7607STejun Heo #define PCPU_EMPTY_POP_PAGES_HIGH 4 95fbf59bc9STejun Heo 96bbddff05STejun Heo #ifdef CONFIG_SMP 97e0100983STejun Heo /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ 98e0100983STejun Heo #ifndef __addr_to_pcpu_ptr 99e0100983STejun Heo #define __addr_to_pcpu_ptr(addr) \ 10043cf38ebSTejun Heo (void __percpu *)((unsigned long)(addr) - \ 10143cf38ebSTejun Heo (unsigned long)pcpu_base_addr + \ 10243cf38ebSTejun Heo (unsigned long)__per_cpu_start) 103e0100983STejun Heo #endif 104e0100983STejun Heo #ifndef __pcpu_ptr_to_addr 105e0100983STejun Heo #define __pcpu_ptr_to_addr(ptr) \ 10643cf38ebSTejun Heo (void __force *)((unsigned long)(ptr) + \ 10743cf38ebSTejun Heo (unsigned long)pcpu_base_addr - \ 10843cf38ebSTejun Heo (unsigned long)__per_cpu_start) 109e0100983STejun Heo #endif 110bbddff05STejun Heo #else /* CONFIG_SMP */ 111bbddff05STejun Heo /* on UP, it's always identity mapped */ 112bbddff05STejun Heo #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) 113bbddff05STejun Heo #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) 114bbddff05STejun Heo #endif /* CONFIG_SMP */ 115e0100983STejun Heo 1161328710bSDaniel Micay static int pcpu_unit_pages __ro_after_init; 1171328710bSDaniel Micay static int pcpu_unit_size __ro_after_init; 1181328710bSDaniel Micay static int pcpu_nr_units __ro_after_init; 1191328710bSDaniel Micay static int pcpu_atom_size __ro_after_init; 1208fa3ed80SDennis Zhou int pcpu_nr_slots __ro_after_init; 1211328710bSDaniel Micay static size_t pcpu_chunk_struct_size __ro_after_init; 122fbf59bc9STejun Heo 123a855b84cSTejun Heo /* cpus with the lowest and highest unit addresses */ 1241328710bSDaniel Micay static unsigned int pcpu_low_unit_cpu __ro_after_init; 1251328710bSDaniel Micay static unsigned int pcpu_high_unit_cpu __ro_after_init; 1262f39e637STejun Heo 127fbf59bc9STejun Heo /* the address of the first chunk which starts with the kernel static area */ 1281328710bSDaniel Micay void *pcpu_base_addr __ro_after_init; 129fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(pcpu_base_addr); 130fbf59bc9STejun Heo 1311328710bSDaniel Micay static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ 1321328710bSDaniel Micay const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ 1332f39e637STejun Heo 1346563297cSTejun Heo /* group information, used for vm allocation */ 1351328710bSDaniel Micay static int pcpu_nr_groups __ro_after_init; 1361328710bSDaniel Micay static const unsigned long *pcpu_group_offsets __ro_after_init; 1371328710bSDaniel Micay static const size_t *pcpu_group_sizes __ro_after_init; 1386563297cSTejun Heo 139ae9e6bc9STejun Heo /* 140ae9e6bc9STejun Heo * The first chunk which always exists. Note that unlike other 141ae9e6bc9STejun Heo * chunks, this one can be allocated and mapped in several different 142ae9e6bc9STejun Heo * ways and thus often doesn't live in the vmalloc area. 143ae9e6bc9STejun Heo */ 1448fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_first_chunk __ro_after_init; 145ae9e6bc9STejun Heo 146ae9e6bc9STejun Heo /* 147ae9e6bc9STejun Heo * Optional reserved chunk. This chunk reserves part of the first 148e2266705SDennis Zhou (Facebook) * chunk and serves it for reserved allocations. When the reserved 149e2266705SDennis Zhou (Facebook) * region doesn't exist, the following variable is NULL. 150ae9e6bc9STejun Heo */ 1518fa3ed80SDennis Zhou struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; 152edcb4639STejun Heo 1538fa3ed80SDennis Zhou DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ 1546710e594STejun Heo static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ 155fbf59bc9STejun Heo 1568fa3ed80SDennis Zhou struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ 157fbf59bc9STejun Heo 1584f996e23STejun Heo /* chunks which need their map areas extended, protected by pcpu_lock */ 1594f996e23STejun Heo static LIST_HEAD(pcpu_map_extend_chunks); 1604f996e23STejun Heo 161b539b87fSTejun Heo /* 162b539b87fSTejun Heo * The number of empty populated pages, protected by pcpu_lock. The 163b539b87fSTejun Heo * reserved chunk doesn't contribute to the count. 164b539b87fSTejun Heo */ 1656b9b6f39SDennis Zhou (Facebook) int pcpu_nr_empty_pop_pages; 166b539b87fSTejun Heo 1671a4d7607STejun Heo /* 1681a4d7607STejun Heo * Balance work is used to populate or destroy chunks asynchronously. We 1691a4d7607STejun Heo * try to keep the number of populated free pages between 1701a4d7607STejun Heo * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one 1711a4d7607STejun Heo * empty chunk. 1721a4d7607STejun Heo */ 173fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work); 174fe6bd8c3STejun Heo static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); 1751a4d7607STejun Heo static bool pcpu_async_enabled __read_mostly; 1761a4d7607STejun Heo static bool pcpu_atomic_alloc_failed; 1771a4d7607STejun Heo 1781a4d7607STejun Heo static void pcpu_schedule_balance_work(void) 1791a4d7607STejun Heo { 1801a4d7607STejun Heo if (pcpu_async_enabled) 1811a4d7607STejun Heo schedule_work(&pcpu_balance_work); 1821a4d7607STejun Heo } 183a56dbddfSTejun Heo 184c0ebfdc3SDennis Zhou (Facebook) /** 185560f2c23SDennis Zhou (Facebook) * pcpu_addr_in_chunk - check if the address is served from this chunk 186560f2c23SDennis Zhou (Facebook) * @chunk: chunk of interest 187560f2c23SDennis Zhou (Facebook) * @addr: percpu address 188c0ebfdc3SDennis Zhou (Facebook) * 189c0ebfdc3SDennis Zhou (Facebook) * RETURNS: 190560f2c23SDennis Zhou (Facebook) * True if the address is served from this chunk. 191c0ebfdc3SDennis Zhou (Facebook) */ 192560f2c23SDennis Zhou (Facebook) static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) 193020ec653STejun Heo { 194c0ebfdc3SDennis Zhou (Facebook) void *start_addr, *end_addr; 195020ec653STejun Heo 196560f2c23SDennis Zhou (Facebook) if (!chunk) 197c0ebfdc3SDennis Zhou (Facebook) return false; 198c0ebfdc3SDennis Zhou (Facebook) 199560f2c23SDennis Zhou (Facebook) start_addr = chunk->base_addr + chunk->start_offset; 200560f2c23SDennis Zhou (Facebook) end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - 201560f2c23SDennis Zhou (Facebook) chunk->end_offset; 202c0ebfdc3SDennis Zhou (Facebook) 203c0ebfdc3SDennis Zhou (Facebook) return addr >= start_addr && addr < end_addr; 204020ec653STejun Heo } 205020ec653STejun Heo 206d9b55eebSTejun Heo static int __pcpu_size_to_slot(int size) 207fbf59bc9STejun Heo { 208cae3aeb8STejun Heo int highbit = fls(size); /* size is in bytes */ 209fbf59bc9STejun Heo return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); 210fbf59bc9STejun Heo } 211fbf59bc9STejun Heo 212d9b55eebSTejun Heo static int pcpu_size_to_slot(int size) 213d9b55eebSTejun Heo { 214d9b55eebSTejun Heo if (size == pcpu_unit_size) 215d9b55eebSTejun Heo return pcpu_nr_slots - 1; 216d9b55eebSTejun Heo return __pcpu_size_to_slot(size); 217d9b55eebSTejun Heo } 218d9b55eebSTejun Heo 219fbf59bc9STejun Heo static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) 220fbf59bc9STejun Heo { 22140064aecSDennis Zhou (Facebook) if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) 222fbf59bc9STejun Heo return 0; 223fbf59bc9STejun Heo 22440064aecSDennis Zhou (Facebook) return pcpu_size_to_slot(chunk->free_bytes); 225fbf59bc9STejun Heo } 226fbf59bc9STejun Heo 22788999a89STejun Heo /* set the pointer to a chunk in a page struct */ 22888999a89STejun Heo static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) 22988999a89STejun Heo { 23088999a89STejun Heo page->index = (unsigned long)pcpu; 23188999a89STejun Heo } 23288999a89STejun Heo 23388999a89STejun Heo /* obtain pointer to a chunk from a page struct */ 23488999a89STejun Heo static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) 23588999a89STejun Heo { 23688999a89STejun Heo return (struct pcpu_chunk *)page->index; 23788999a89STejun Heo } 23888999a89STejun Heo 23988999a89STejun Heo static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) 240fbf59bc9STejun Heo { 2412f39e637STejun Heo return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; 242fbf59bc9STejun Heo } 243fbf59bc9STejun Heo 244c0ebfdc3SDennis Zhou (Facebook) static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) 245c0ebfdc3SDennis Zhou (Facebook) { 246c0ebfdc3SDennis Zhou (Facebook) return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); 247c0ebfdc3SDennis Zhou (Facebook) } 248c0ebfdc3SDennis Zhou (Facebook) 2499983b6f0STejun Heo static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 250fbf59bc9STejun Heo unsigned int cpu, int page_idx) 251fbf59bc9STejun Heo { 252c0ebfdc3SDennis Zhou (Facebook) return (unsigned long)chunk->base_addr + 253c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(cpu, page_idx); 254fbf59bc9STejun Heo } 255fbf59bc9STejun Heo 25691e914c5SDennis Zhou (Facebook) static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) 257ce3141a2STejun Heo { 25891e914c5SDennis Zhou (Facebook) *rs = find_next_zero_bit(bitmap, end, *rs); 25991e914c5SDennis Zhou (Facebook) *re = find_next_bit(bitmap, end, *rs + 1); 260ce3141a2STejun Heo } 261ce3141a2STejun Heo 26291e914c5SDennis Zhou (Facebook) static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) 263ce3141a2STejun Heo { 26491e914c5SDennis Zhou (Facebook) *rs = find_next_bit(bitmap, end, *rs); 26591e914c5SDennis Zhou (Facebook) *re = find_next_zero_bit(bitmap, end, *rs + 1); 266ce3141a2STejun Heo } 267ce3141a2STejun Heo 268ce3141a2STejun Heo /* 26991e914c5SDennis Zhou (Facebook) * Bitmap region iterators. Iterates over the bitmap between 27091e914c5SDennis Zhou (Facebook) * [@start, @end) in @chunk. @rs and @re should be integer variables 27191e914c5SDennis Zhou (Facebook) * and will be set to start and end index of the current free region. 272ce3141a2STejun Heo */ 27391e914c5SDennis Zhou (Facebook) #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ 27491e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ 275ce3141a2STejun Heo (rs) < (re); \ 27691e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) 277ce3141a2STejun Heo 27891e914c5SDennis Zhou (Facebook) #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ 27991e914c5SDennis Zhou (Facebook) for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ 280ce3141a2STejun Heo (rs) < (re); \ 28191e914c5SDennis Zhou (Facebook) (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) 282ce3141a2STejun Heo 283ca460b3cSDennis Zhou (Facebook) /* 284ca460b3cSDennis Zhou (Facebook) * The following are helper functions to help access bitmaps and convert 285ca460b3cSDennis Zhou (Facebook) * between bitmap offsets to address offsets. 286ca460b3cSDennis Zhou (Facebook) */ 287ca460b3cSDennis Zhou (Facebook) static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) 288ca460b3cSDennis Zhou (Facebook) { 289ca460b3cSDennis Zhou (Facebook) return chunk->alloc_map + 290ca460b3cSDennis Zhou (Facebook) (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); 291ca460b3cSDennis Zhou (Facebook) } 292ca460b3cSDennis Zhou (Facebook) 293ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_index(int off) 294ca460b3cSDennis Zhou (Facebook) { 295ca460b3cSDennis Zhou (Facebook) return off / PCPU_BITMAP_BLOCK_BITS; 296ca460b3cSDennis Zhou (Facebook) } 297ca460b3cSDennis Zhou (Facebook) 298ca460b3cSDennis Zhou (Facebook) static unsigned long pcpu_off_to_block_off(int off) 299ca460b3cSDennis Zhou (Facebook) { 300ca460b3cSDennis Zhou (Facebook) return off & (PCPU_BITMAP_BLOCK_BITS - 1); 301ca460b3cSDennis Zhou (Facebook) } 302ca460b3cSDennis Zhou (Facebook) 303fbf59bc9STejun Heo /** 30490459ce0SBob Liu * pcpu_mem_zalloc - allocate memory 3051880d93bSTejun Heo * @size: bytes to allocate 306fbf59bc9STejun Heo * 3071880d93bSTejun Heo * Allocate @size bytes. If @size is smaller than PAGE_SIZE, 30890459ce0SBob Liu * kzalloc() is used; otherwise, vzalloc() is used. The returned 3091880d93bSTejun Heo * memory is always zeroed. 310fbf59bc9STejun Heo * 311ccea34b5STejun Heo * CONTEXT: 312ccea34b5STejun Heo * Does GFP_KERNEL allocation. 313ccea34b5STejun Heo * 314fbf59bc9STejun Heo * RETURNS: 3151880d93bSTejun Heo * Pointer to the allocated area on success, NULL on failure. 316fbf59bc9STejun Heo */ 31790459ce0SBob Liu static void *pcpu_mem_zalloc(size_t size) 318fbf59bc9STejun Heo { 319099a19d9STejun Heo if (WARN_ON_ONCE(!slab_is_available())) 320099a19d9STejun Heo return NULL; 321099a19d9STejun Heo 322fbf59bc9STejun Heo if (size <= PAGE_SIZE) 3231880d93bSTejun Heo return kzalloc(size, GFP_KERNEL); 3247af4c093SJesper Juhl else 3257af4c093SJesper Juhl return vzalloc(size); 3261880d93bSTejun Heo } 327fbf59bc9STejun Heo 3281880d93bSTejun Heo /** 3291880d93bSTejun Heo * pcpu_mem_free - free memory 3301880d93bSTejun Heo * @ptr: memory to free 3311880d93bSTejun Heo * 33290459ce0SBob Liu * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). 3331880d93bSTejun Heo */ 3341d5cfdb0STetsuo Handa static void pcpu_mem_free(void *ptr) 3351880d93bSTejun Heo { 3361d5cfdb0STetsuo Handa kvfree(ptr); 337fbf59bc9STejun Heo } 338fbf59bc9STejun Heo 339fbf59bc9STejun Heo /** 340fbf59bc9STejun Heo * pcpu_chunk_relocate - put chunk in the appropriate chunk slot 341fbf59bc9STejun Heo * @chunk: chunk of interest 342fbf59bc9STejun Heo * @oslot: the previous slot it was on 343fbf59bc9STejun Heo * 344fbf59bc9STejun Heo * This function is called after an allocation or free changed @chunk. 345fbf59bc9STejun Heo * New slot according to the changed state is determined and @chunk is 346edcb4639STejun Heo * moved to the slot. Note that the reserved chunk is never put on 347edcb4639STejun Heo * chunk slots. 348ccea34b5STejun Heo * 349ccea34b5STejun Heo * CONTEXT: 350ccea34b5STejun Heo * pcpu_lock. 351fbf59bc9STejun Heo */ 352fbf59bc9STejun Heo static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) 353fbf59bc9STejun Heo { 354fbf59bc9STejun Heo int nslot = pcpu_chunk_slot(chunk); 355fbf59bc9STejun Heo 356edcb4639STejun Heo if (chunk != pcpu_reserved_chunk && oslot != nslot) { 357fbf59bc9STejun Heo if (oslot < nslot) 358fbf59bc9STejun Heo list_move(&chunk->list, &pcpu_slot[nslot]); 359fbf59bc9STejun Heo else 360fbf59bc9STejun Heo list_move_tail(&chunk->list, &pcpu_slot[nslot]); 361fbf59bc9STejun Heo } 362fbf59bc9STejun Heo } 363fbf59bc9STejun Heo 364fbf59bc9STejun Heo /** 36540064aecSDennis Zhou (Facebook) * pcpu_cnt_pop_pages- counts populated backing pages in range 366833af842STejun Heo * @chunk: chunk of interest 36740064aecSDennis Zhou (Facebook) * @bit_off: start offset 36840064aecSDennis Zhou (Facebook) * @bits: size of area to check 3699f7dcf22STejun Heo * 37040064aecSDennis Zhou (Facebook) * Calculates the number of populated pages in the region 37140064aecSDennis Zhou (Facebook) * [page_start, page_end). This keeps track of how many empty populated 37240064aecSDennis Zhou (Facebook) * pages are available and decide if async work should be scheduled. 373ccea34b5STejun Heo * 3749f7dcf22STejun Heo * RETURNS: 37540064aecSDennis Zhou (Facebook) * The nr of populated pages. 3769f7dcf22STejun Heo */ 37740064aecSDennis Zhou (Facebook) static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off, 37840064aecSDennis Zhou (Facebook) int bits) 3799f7dcf22STejun Heo { 38040064aecSDennis Zhou (Facebook) int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE); 38140064aecSDennis Zhou (Facebook) int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 38240064aecSDennis Zhou (Facebook) 38340064aecSDennis Zhou (Facebook) if (page_start >= page_end) 38440064aecSDennis Zhou (Facebook) return 0; 38540064aecSDennis Zhou (Facebook) 38640064aecSDennis Zhou (Facebook) /* 38740064aecSDennis Zhou (Facebook) * bitmap_weight counts the number of bits set in a bitmap up to 38840064aecSDennis Zhou (Facebook) * the specified number of bits. This is counting the populated 38940064aecSDennis Zhou (Facebook) * pages up to page_end and then subtracting the populated pages 39040064aecSDennis Zhou (Facebook) * up to page_start to count the populated pages in 39140064aecSDennis Zhou (Facebook) * [page_start, page_end). 39240064aecSDennis Zhou (Facebook) */ 39340064aecSDennis Zhou (Facebook) return bitmap_weight(chunk->populated, page_end) - 39440064aecSDennis Zhou (Facebook) bitmap_weight(chunk->populated, page_start); 39540064aecSDennis Zhou (Facebook) } 39640064aecSDennis Zhou (Facebook) 39740064aecSDennis Zhou (Facebook) /** 39840064aecSDennis Zhou (Facebook) * pcpu_chunk_update - updates the chunk metadata given a free area 39940064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 40040064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 40140064aecSDennis Zhou (Facebook) * @bits: size of free area 40240064aecSDennis Zhou (Facebook) * 40313f96637SDennis Zhou (Facebook) * This updates the chunk's contig hint and starting offset given a free area. 404268625a6SDennis Zhou (Facebook) * Choose the best starting offset if the contig hint is equal. 40540064aecSDennis Zhou (Facebook) */ 40640064aecSDennis Zhou (Facebook) static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) 40740064aecSDennis Zhou (Facebook) { 40813f96637SDennis Zhou (Facebook) if (bits > chunk->contig_bits) { 40913f96637SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 41040064aecSDennis Zhou (Facebook) chunk->contig_bits = bits; 411268625a6SDennis Zhou (Facebook) } else if (bits == chunk->contig_bits && chunk->contig_bits_start && 412268625a6SDennis Zhou (Facebook) (!bit_off || 413268625a6SDennis Zhou (Facebook) __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { 414268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 415268625a6SDennis Zhou (Facebook) chunk->contig_bits_start = bit_off; 41640064aecSDennis Zhou (Facebook) } 41713f96637SDennis Zhou (Facebook) } 41840064aecSDennis Zhou (Facebook) 41940064aecSDennis Zhou (Facebook) /** 42040064aecSDennis Zhou (Facebook) * pcpu_chunk_refresh_hint - updates metadata about a chunk 42140064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 42240064aecSDennis Zhou (Facebook) * 42340064aecSDennis Zhou (Facebook) * Iterates over the chunk to find the largest free area. 42440064aecSDennis Zhou (Facebook) * 42540064aecSDennis Zhou (Facebook) * Updates: 42640064aecSDennis Zhou (Facebook) * chunk->contig_bits 42713f96637SDennis Zhou (Facebook) * chunk->contig_bits_start 42840064aecSDennis Zhou (Facebook) * nr_empty_pop_pages 42940064aecSDennis Zhou (Facebook) */ 43040064aecSDennis Zhou (Facebook) static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) 43140064aecSDennis Zhou (Facebook) { 43240064aecSDennis Zhou (Facebook) int bits, nr_empty_pop_pages; 43340064aecSDennis Zhou (Facebook) int rs, re; /* region start, region end */ 43440064aecSDennis Zhou (Facebook) 43540064aecSDennis Zhou (Facebook) /* clear metadata */ 43640064aecSDennis Zhou (Facebook) chunk->contig_bits = 0; 43740064aecSDennis Zhou (Facebook) 43840064aecSDennis Zhou (Facebook) bits = nr_empty_pop_pages = 0; 43986b442fbSDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->alloc_map, rs, re, chunk->first_bit, 44040064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk)) { 44140064aecSDennis Zhou (Facebook) bits = re - rs; 44240064aecSDennis Zhou (Facebook) 44340064aecSDennis Zhou (Facebook) pcpu_chunk_update(chunk, rs, bits); 44440064aecSDennis Zhou (Facebook) 44540064aecSDennis Zhou (Facebook) nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, rs, bits); 44640064aecSDennis Zhou (Facebook) } 44740064aecSDennis Zhou (Facebook) 44840064aecSDennis Zhou (Facebook) /* 44940064aecSDennis Zhou (Facebook) * Keep track of nr_empty_pop_pages. 45040064aecSDennis Zhou (Facebook) * 45140064aecSDennis Zhou (Facebook) * The chunk maintains the previous number of free pages it held, 45240064aecSDennis Zhou (Facebook) * so the delta is used to update the global counter. The reserved 45340064aecSDennis Zhou (Facebook) * chunk is not part of the free page count as they are populated 45440064aecSDennis Zhou (Facebook) * at init and are special to serving reserved allocations. 45540064aecSDennis Zhou (Facebook) */ 45640064aecSDennis Zhou (Facebook) if (chunk != pcpu_reserved_chunk) 45740064aecSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages += 45840064aecSDennis Zhou (Facebook) (nr_empty_pop_pages - chunk->nr_empty_pop_pages); 45940064aecSDennis Zhou (Facebook) 46040064aecSDennis Zhou (Facebook) chunk->nr_empty_pop_pages = nr_empty_pop_pages; 46140064aecSDennis Zhou (Facebook) } 46240064aecSDennis Zhou (Facebook) 46340064aecSDennis Zhou (Facebook) /** 464ca460b3cSDennis Zhou (Facebook) * pcpu_block_update - updates a block given a free area 465ca460b3cSDennis Zhou (Facebook) * @block: block of interest 466ca460b3cSDennis Zhou (Facebook) * @start: start offset in block 467ca460b3cSDennis Zhou (Facebook) * @end: end offset in block 468ca460b3cSDennis Zhou (Facebook) * 469ca460b3cSDennis Zhou (Facebook) * Updates a block given a known free area. The region [start, end) is 470268625a6SDennis Zhou (Facebook) * expected to be the entirety of the free area within a block. Chooses 471268625a6SDennis Zhou (Facebook) * the best starting offset if the contig hints are equal. 472ca460b3cSDennis Zhou (Facebook) */ 473ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) 474ca460b3cSDennis Zhou (Facebook) { 475ca460b3cSDennis Zhou (Facebook) int contig = end - start; 476ca460b3cSDennis Zhou (Facebook) 477ca460b3cSDennis Zhou (Facebook) block->first_free = min(block->first_free, start); 478ca460b3cSDennis Zhou (Facebook) if (start == 0) 479ca460b3cSDennis Zhou (Facebook) block->left_free = contig; 480ca460b3cSDennis Zhou (Facebook) 481ca460b3cSDennis Zhou (Facebook) if (end == PCPU_BITMAP_BLOCK_BITS) 482ca460b3cSDennis Zhou (Facebook) block->right_free = contig; 483ca460b3cSDennis Zhou (Facebook) 484ca460b3cSDennis Zhou (Facebook) if (contig > block->contig_hint) { 485ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = start; 486ca460b3cSDennis Zhou (Facebook) block->contig_hint = contig; 487268625a6SDennis Zhou (Facebook) } else if (block->contig_hint_start && contig == block->contig_hint && 488268625a6SDennis Zhou (Facebook) (!start || __ffs(start) > __ffs(block->contig_hint_start))) { 489268625a6SDennis Zhou (Facebook) /* use the start with the best alignment */ 490268625a6SDennis Zhou (Facebook) block->contig_hint_start = start; 491ca460b3cSDennis Zhou (Facebook) } 492ca460b3cSDennis Zhou (Facebook) } 493ca460b3cSDennis Zhou (Facebook) 494ca460b3cSDennis Zhou (Facebook) /** 495ca460b3cSDennis Zhou (Facebook) * pcpu_block_refresh_hint 496ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 497ca460b3cSDennis Zhou (Facebook) * @index: index of the metadata block 498ca460b3cSDennis Zhou (Facebook) * 499ca460b3cSDennis Zhou (Facebook) * Scans over the block beginning at first_free and updates the block 500ca460b3cSDennis Zhou (Facebook) * metadata accordingly. 501ca460b3cSDennis Zhou (Facebook) */ 502ca460b3cSDennis Zhou (Facebook) static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) 503ca460b3cSDennis Zhou (Facebook) { 504ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *block = chunk->md_blocks + index; 505ca460b3cSDennis Zhou (Facebook) unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); 506ca460b3cSDennis Zhou (Facebook) int rs, re; /* region start, region end */ 507ca460b3cSDennis Zhou (Facebook) 508ca460b3cSDennis Zhou (Facebook) /* clear hints */ 509ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 510ca460b3cSDennis Zhou (Facebook) block->left_free = block->right_free = 0; 511ca460b3cSDennis Zhou (Facebook) 512ca460b3cSDennis Zhou (Facebook) /* iterate over free areas and update the contig hints */ 513ca460b3cSDennis Zhou (Facebook) pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, 514ca460b3cSDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS) { 515ca460b3cSDennis Zhou (Facebook) pcpu_block_update(block, rs, re); 516ca460b3cSDennis Zhou (Facebook) } 517ca460b3cSDennis Zhou (Facebook) } 518ca460b3cSDennis Zhou (Facebook) 519ca460b3cSDennis Zhou (Facebook) /** 520ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_alloc - update hint on allocation path 521ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 522ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 523ca460b3cSDennis Zhou (Facebook) * @bits: size of request 524*fc304334SDennis Zhou (Facebook) * 525*fc304334SDennis Zhou (Facebook) * Updates metadata for the allocation path. The metadata only has to be 526*fc304334SDennis Zhou (Facebook) * refreshed by a full scan iff the chunk's contig hint is broken. Block level 527*fc304334SDennis Zhou (Facebook) * scans are required if the block's contig hint is broken. 528ca460b3cSDennis Zhou (Facebook) */ 529ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, 530ca460b3cSDennis Zhou (Facebook) int bits) 531ca460b3cSDennis Zhou (Facebook) { 532ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 533ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 534ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 535ca460b3cSDennis Zhou (Facebook) 536ca460b3cSDennis Zhou (Facebook) /* 537ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 538ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 539ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 540ca460b3cSDennis Zhou (Facebook) * range. 541ca460b3cSDennis Zhou (Facebook) */ 542ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 543ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 544ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 545ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 546ca460b3cSDennis Zhou (Facebook) 547ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 548ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 549ca460b3cSDennis Zhou (Facebook) 550ca460b3cSDennis Zhou (Facebook) /* 551ca460b3cSDennis Zhou (Facebook) * Update s_block. 552*fc304334SDennis Zhou (Facebook) * block->first_free must be updated if the allocation takes its place. 553*fc304334SDennis Zhou (Facebook) * If the allocation breaks the contig_hint, a scan is required to 554*fc304334SDennis Zhou (Facebook) * restore this hint. 555ca460b3cSDennis Zhou (Facebook) */ 556*fc304334SDennis Zhou (Facebook) if (s_off == s_block->first_free) 557*fc304334SDennis Zhou (Facebook) s_block->first_free = find_next_zero_bit( 558*fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, s_index), 559*fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, 560*fc304334SDennis Zhou (Facebook) s_off + bits); 561*fc304334SDennis Zhou (Facebook) 562*fc304334SDennis Zhou (Facebook) if (s_off >= s_block->contig_hint_start && 563*fc304334SDennis Zhou (Facebook) s_off < s_block->contig_hint_start + s_block->contig_hint) { 564*fc304334SDennis Zhou (Facebook) /* block contig hint is broken - scan to fix it */ 565ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 566*fc304334SDennis Zhou (Facebook) } else { 567*fc304334SDennis Zhou (Facebook) /* update left and right contig manually */ 568*fc304334SDennis Zhou (Facebook) s_block->left_free = min(s_block->left_free, s_off); 569*fc304334SDennis Zhou (Facebook) if (s_index == e_index) 570*fc304334SDennis Zhou (Facebook) s_block->right_free = min_t(int, s_block->right_free, 571*fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 572*fc304334SDennis Zhou (Facebook) else 573*fc304334SDennis Zhou (Facebook) s_block->right_free = 0; 574*fc304334SDennis Zhou (Facebook) } 575ca460b3cSDennis Zhou (Facebook) 576ca460b3cSDennis Zhou (Facebook) /* 577ca460b3cSDennis Zhou (Facebook) * Update e_block. 578ca460b3cSDennis Zhou (Facebook) */ 579ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 580*fc304334SDennis Zhou (Facebook) /* 581*fc304334SDennis Zhou (Facebook) * When the allocation is across blocks, the end is along 582*fc304334SDennis Zhou (Facebook) * the left part of the e_block. 583*fc304334SDennis Zhou (Facebook) */ 584*fc304334SDennis Zhou (Facebook) e_block->first_free = find_next_zero_bit( 585*fc304334SDennis Zhou (Facebook) pcpu_index_alloc_map(chunk, e_index), 586*fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS, e_off); 587*fc304334SDennis Zhou (Facebook) 588*fc304334SDennis Zhou (Facebook) if (e_off == PCPU_BITMAP_BLOCK_BITS) { 589*fc304334SDennis Zhou (Facebook) /* reset the block */ 590*fc304334SDennis Zhou (Facebook) e_block++; 591*fc304334SDennis Zhou (Facebook) } else { 592*fc304334SDennis Zhou (Facebook) if (e_off > e_block->contig_hint_start) { 593*fc304334SDennis Zhou (Facebook) /* contig hint is broken - scan to fix it */ 594ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 595*fc304334SDennis Zhou (Facebook) } else { 596*fc304334SDennis Zhou (Facebook) e_block->left_free = 0; 597*fc304334SDennis Zhou (Facebook) e_block->right_free = 598*fc304334SDennis Zhou (Facebook) min_t(int, e_block->right_free, 599*fc304334SDennis Zhou (Facebook) PCPU_BITMAP_BLOCK_BITS - e_off); 600*fc304334SDennis Zhou (Facebook) } 601*fc304334SDennis Zhou (Facebook) } 602ca460b3cSDennis Zhou (Facebook) 603ca460b3cSDennis Zhou (Facebook) /* update in-between md_blocks */ 604ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 605ca460b3cSDennis Zhou (Facebook) block->contig_hint = 0; 606ca460b3cSDennis Zhou (Facebook) block->left_free = 0; 607ca460b3cSDennis Zhou (Facebook) block->right_free = 0; 608ca460b3cSDennis Zhou (Facebook) } 609ca460b3cSDennis Zhou (Facebook) } 610ca460b3cSDennis Zhou (Facebook) 611*fc304334SDennis Zhou (Facebook) /* 612*fc304334SDennis Zhou (Facebook) * The only time a full chunk scan is required is if the chunk 613*fc304334SDennis Zhou (Facebook) * contig hint is broken. Otherwise, it means a smaller space 614*fc304334SDennis Zhou (Facebook) * was used and therefore the chunk contig hint is still correct. 615*fc304334SDennis Zhou (Facebook) */ 616*fc304334SDennis Zhou (Facebook) if (bit_off >= chunk->contig_bits_start && 617*fc304334SDennis Zhou (Facebook) bit_off < chunk->contig_bits_start + chunk->contig_bits) 618ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 619ca460b3cSDennis Zhou (Facebook) } 620ca460b3cSDennis Zhou (Facebook) 621ca460b3cSDennis Zhou (Facebook) /** 622ca460b3cSDennis Zhou (Facebook) * pcpu_block_update_hint_free - updates the block hints on the free path 623ca460b3cSDennis Zhou (Facebook) * @chunk: chunk of interest 624ca460b3cSDennis Zhou (Facebook) * @bit_off: chunk offset 625ca460b3cSDennis Zhou (Facebook) * @bits: size of request 626ca460b3cSDennis Zhou (Facebook) */ 627ca460b3cSDennis Zhou (Facebook) static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, 628ca460b3cSDennis Zhou (Facebook) int bits) 629ca460b3cSDennis Zhou (Facebook) { 630ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *s_block, *e_block, *block; 631ca460b3cSDennis Zhou (Facebook) int s_index, e_index; /* block indexes of the freed allocation */ 632ca460b3cSDennis Zhou (Facebook) int s_off, e_off; /* block offsets of the freed allocation */ 633ca460b3cSDennis Zhou (Facebook) 634ca460b3cSDennis Zhou (Facebook) /* 635ca460b3cSDennis Zhou (Facebook) * Calculate per block offsets. 636ca460b3cSDennis Zhou (Facebook) * The calculation uses an inclusive range, but the resulting offsets 637ca460b3cSDennis Zhou (Facebook) * are [start, end). e_index always points to the last block in the 638ca460b3cSDennis Zhou (Facebook) * range. 639ca460b3cSDennis Zhou (Facebook) */ 640ca460b3cSDennis Zhou (Facebook) s_index = pcpu_off_to_block_index(bit_off); 641ca460b3cSDennis Zhou (Facebook) e_index = pcpu_off_to_block_index(bit_off + bits - 1); 642ca460b3cSDennis Zhou (Facebook) s_off = pcpu_off_to_block_off(bit_off); 643ca460b3cSDennis Zhou (Facebook) e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; 644ca460b3cSDennis Zhou (Facebook) 645ca460b3cSDennis Zhou (Facebook) s_block = chunk->md_blocks + s_index; 646ca460b3cSDennis Zhou (Facebook) e_block = chunk->md_blocks + e_index; 647ca460b3cSDennis Zhou (Facebook) 648ca460b3cSDennis Zhou (Facebook) /* update s_block */ 649ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, s_index); 650ca460b3cSDennis Zhou (Facebook) 651ca460b3cSDennis Zhou (Facebook) /* freeing in the same block */ 652ca460b3cSDennis Zhou (Facebook) if (s_index != e_index) { 653ca460b3cSDennis Zhou (Facebook) /* update e_block */ 654ca460b3cSDennis Zhou (Facebook) pcpu_block_refresh_hint(chunk, e_index); 655ca460b3cSDennis Zhou (Facebook) 656ca460b3cSDennis Zhou (Facebook) /* reset md_blocks in the middle */ 657ca460b3cSDennis Zhou (Facebook) for (block = s_block + 1; block < e_block; block++) { 658ca460b3cSDennis Zhou (Facebook) block->first_free = 0; 659ca460b3cSDennis Zhou (Facebook) block->contig_hint_start = 0; 660ca460b3cSDennis Zhou (Facebook) block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 661ca460b3cSDennis Zhou (Facebook) block->left_free = PCPU_BITMAP_BLOCK_BITS; 662ca460b3cSDennis Zhou (Facebook) block->right_free = PCPU_BITMAP_BLOCK_BITS; 663ca460b3cSDennis Zhou (Facebook) } 664ca460b3cSDennis Zhou (Facebook) } 665ca460b3cSDennis Zhou (Facebook) 666ca460b3cSDennis Zhou (Facebook) pcpu_chunk_refresh_hint(chunk); 667ca460b3cSDennis Zhou (Facebook) } 668ca460b3cSDennis Zhou (Facebook) 669ca460b3cSDennis Zhou (Facebook) /** 67040064aecSDennis Zhou (Facebook) * pcpu_is_populated - determines if the region is populated 67140064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 67240064aecSDennis Zhou (Facebook) * @bit_off: chunk offset 67340064aecSDennis Zhou (Facebook) * @bits: size of area 67440064aecSDennis Zhou (Facebook) * @next_off: return value for the next offset to start searching 67540064aecSDennis Zhou (Facebook) * 67640064aecSDennis Zhou (Facebook) * For atomic allocations, check if the backing pages are populated. 67740064aecSDennis Zhou (Facebook) * 67840064aecSDennis Zhou (Facebook) * RETURNS: 67940064aecSDennis Zhou (Facebook) * Bool if the backing pages are populated. 68040064aecSDennis Zhou (Facebook) * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. 68140064aecSDennis Zhou (Facebook) */ 68240064aecSDennis Zhou (Facebook) static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, 68340064aecSDennis Zhou (Facebook) int *next_off) 68440064aecSDennis Zhou (Facebook) { 68540064aecSDennis Zhou (Facebook) int page_start, page_end, rs, re; 68640064aecSDennis Zhou (Facebook) 68740064aecSDennis Zhou (Facebook) page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); 68840064aecSDennis Zhou (Facebook) page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); 68940064aecSDennis Zhou (Facebook) 69040064aecSDennis Zhou (Facebook) rs = page_start; 69140064aecSDennis Zhou (Facebook) pcpu_next_unpop(chunk->populated, &rs, &re, page_end); 69240064aecSDennis Zhou (Facebook) if (rs >= page_end) 69340064aecSDennis Zhou (Facebook) return true; 69440064aecSDennis Zhou (Facebook) 69540064aecSDennis Zhou (Facebook) *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; 69640064aecSDennis Zhou (Facebook) return false; 69740064aecSDennis Zhou (Facebook) } 69840064aecSDennis Zhou (Facebook) 69940064aecSDennis Zhou (Facebook) /** 70040064aecSDennis Zhou (Facebook) * pcpu_find_block_fit - finds the block index to start searching 70140064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 70240064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 70340064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE bytes) 70440064aecSDennis Zhou (Facebook) * @pop_only: use populated regions only 70540064aecSDennis Zhou (Facebook) * 70640064aecSDennis Zhou (Facebook) * RETURNS: 70740064aecSDennis Zhou (Facebook) * The offset in the bitmap to begin searching. 70840064aecSDennis Zhou (Facebook) * -1 if no offset is found. 70940064aecSDennis Zhou (Facebook) */ 71040064aecSDennis Zhou (Facebook) static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, 71140064aecSDennis Zhou (Facebook) size_t align, bool pop_only) 71240064aecSDennis Zhou (Facebook) { 71340064aecSDennis Zhou (Facebook) int bit_off, bits; 71440064aecSDennis Zhou (Facebook) int re; /* region end */ 71540064aecSDennis Zhou (Facebook) 71613f96637SDennis Zhou (Facebook) /* 71713f96637SDennis Zhou (Facebook) * Check to see if the allocation can fit in the chunk's contig hint. 71813f96637SDennis Zhou (Facebook) * This is an optimization to prevent scanning by assuming if it 71913f96637SDennis Zhou (Facebook) * cannot fit in the global hint, there is memory pressure and creating 72013f96637SDennis Zhou (Facebook) * a new chunk would happen soon. 72113f96637SDennis Zhou (Facebook) */ 72213f96637SDennis Zhou (Facebook) bit_off = ALIGN(chunk->contig_bits_start, align) - 72313f96637SDennis Zhou (Facebook) chunk->contig_bits_start; 72413f96637SDennis Zhou (Facebook) if (bit_off + alloc_bits > chunk->contig_bits) 72513f96637SDennis Zhou (Facebook) return -1; 72613f96637SDennis Zhou (Facebook) 72786b442fbSDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->alloc_map, bit_off, re, 72886b442fbSDennis Zhou (Facebook) chunk->first_bit, 72940064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk)) { 73040064aecSDennis Zhou (Facebook) bits = re - bit_off; 73140064aecSDennis Zhou (Facebook) 73240064aecSDennis Zhou (Facebook) /* check alignment */ 73340064aecSDennis Zhou (Facebook) bits -= ALIGN(bit_off, align) - bit_off; 73440064aecSDennis Zhou (Facebook) bit_off = ALIGN(bit_off, align); 73540064aecSDennis Zhou (Facebook) if (bits < alloc_bits) 73640064aecSDennis Zhou (Facebook) continue; 73740064aecSDennis Zhou (Facebook) 73840064aecSDennis Zhou (Facebook) bits = alloc_bits; 73940064aecSDennis Zhou (Facebook) if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, 74040064aecSDennis Zhou (Facebook) &bit_off)) 74140064aecSDennis Zhou (Facebook) break; 74240064aecSDennis Zhou (Facebook) 74340064aecSDennis Zhou (Facebook) bits = 0; 74440064aecSDennis Zhou (Facebook) } 74540064aecSDennis Zhou (Facebook) 74640064aecSDennis Zhou (Facebook) if (bit_off == pcpu_chunk_map_bits(chunk)) 74740064aecSDennis Zhou (Facebook) return -1; 74840064aecSDennis Zhou (Facebook) 74940064aecSDennis Zhou (Facebook) return bit_off; 75040064aecSDennis Zhou (Facebook) } 75140064aecSDennis Zhou (Facebook) 75240064aecSDennis Zhou (Facebook) /** 75340064aecSDennis Zhou (Facebook) * pcpu_alloc_area - allocates an area from a pcpu_chunk 75440064aecSDennis Zhou (Facebook) * @chunk: chunk of interest 75540064aecSDennis Zhou (Facebook) * @alloc_bits: size of request in allocation units 75640064aecSDennis Zhou (Facebook) * @align: alignment of area (max PAGE_SIZE) 75740064aecSDennis Zhou (Facebook) * @start: bit_off to start searching 75840064aecSDennis Zhou (Facebook) * 75940064aecSDennis Zhou (Facebook) * This function takes in a @start offset to begin searching to fit an 76040064aecSDennis Zhou (Facebook) * allocation of @alloc_bits with alignment @align. If it confirms a 76140064aecSDennis Zhou (Facebook) * valid free area, it then updates the allocation and boundary maps 76240064aecSDennis Zhou (Facebook) * accordingly. 76340064aecSDennis Zhou (Facebook) * 76440064aecSDennis Zhou (Facebook) * RETURNS: 76540064aecSDennis Zhou (Facebook) * Allocated addr offset in @chunk on success. 76640064aecSDennis Zhou (Facebook) * -1 if no matching area is found. 76740064aecSDennis Zhou (Facebook) */ 76840064aecSDennis Zhou (Facebook) static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, 76940064aecSDennis Zhou (Facebook) size_t align, int start) 77040064aecSDennis Zhou (Facebook) { 77140064aecSDennis Zhou (Facebook) size_t align_mask = (align) ? (align - 1) : 0; 77240064aecSDennis Zhou (Facebook) int bit_off, end, oslot; 7739f7dcf22STejun Heo 7744f996e23STejun Heo lockdep_assert_held(&pcpu_lock); 7754f996e23STejun Heo 77640064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 777833af842STejun Heo 778833af842STejun Heo /* 77940064aecSDennis Zhou (Facebook) * Search to find a fit. 780833af842STejun Heo */ 78140064aecSDennis Zhou (Facebook) end = start + alloc_bits; 78240064aecSDennis Zhou (Facebook) bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, 78340064aecSDennis Zhou (Facebook) alloc_bits, align_mask); 78440064aecSDennis Zhou (Facebook) if (bit_off >= end) 785a16037c8STejun Heo return -1; 786a16037c8STejun Heo 78740064aecSDennis Zhou (Facebook) /* update alloc map */ 78840064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, bit_off, alloc_bits); 789a16037c8STejun Heo 79040064aecSDennis Zhou (Facebook) /* update boundary map */ 79140064aecSDennis Zhou (Facebook) set_bit(bit_off, chunk->bound_map); 79240064aecSDennis Zhou (Facebook) bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); 79340064aecSDennis Zhou (Facebook) set_bit(bit_off + alloc_bits, chunk->bound_map); 794a16037c8STejun Heo 79540064aecSDennis Zhou (Facebook) chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; 79640064aecSDennis Zhou (Facebook) 79786b442fbSDennis Zhou (Facebook) /* update first free bit */ 79886b442fbSDennis Zhou (Facebook) if (bit_off == chunk->first_bit) 79986b442fbSDennis Zhou (Facebook) chunk->first_bit = find_next_zero_bit( 80086b442fbSDennis Zhou (Facebook) chunk->alloc_map, 80186b442fbSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk), 80286b442fbSDennis Zhou (Facebook) bit_off + alloc_bits); 80386b442fbSDennis Zhou (Facebook) 804ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); 80540064aecSDennis Zhou (Facebook) 80640064aecSDennis Zhou (Facebook) pcpu_chunk_relocate(chunk, oslot); 80740064aecSDennis Zhou (Facebook) 80840064aecSDennis Zhou (Facebook) return bit_off * PCPU_MIN_ALLOC_SIZE; 809a16037c8STejun Heo } 810a16037c8STejun Heo 811a16037c8STejun Heo /** 81240064aecSDennis Zhou (Facebook) * pcpu_free_area - frees the corresponding offset 813fbf59bc9STejun Heo * @chunk: chunk of interest 81440064aecSDennis Zhou (Facebook) * @off: addr offset into chunk 815fbf59bc9STejun Heo * 81640064aecSDennis Zhou (Facebook) * This function determines the size of an allocation to free using 81740064aecSDennis Zhou (Facebook) * the boundary bitmap and clears the allocation map. 818fbf59bc9STejun Heo */ 81940064aecSDennis Zhou (Facebook) static void pcpu_free_area(struct pcpu_chunk *chunk, int off) 820fbf59bc9STejun Heo { 82140064aecSDennis Zhou (Facebook) int bit_off, bits, end, oslot; 822fbf59bc9STejun Heo 8235ccd30e4SDennis Zhou lockdep_assert_held(&pcpu_lock); 82430a5b536SDennis Zhou pcpu_stats_area_dealloc(chunk); 8255ccd30e4SDennis Zhou 82640064aecSDennis Zhou (Facebook) oslot = pcpu_chunk_slot(chunk); 827723ad1d9SAl Viro 82840064aecSDennis Zhou (Facebook) bit_off = off / PCPU_MIN_ALLOC_SIZE; 829fbf59bc9STejun Heo 83040064aecSDennis Zhou (Facebook) /* find end index */ 83140064aecSDennis Zhou (Facebook) end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), 83240064aecSDennis Zhou (Facebook) bit_off + 1); 83340064aecSDennis Zhou (Facebook) bits = end - bit_off; 83440064aecSDennis Zhou (Facebook) bitmap_clear(chunk->alloc_map, bit_off, bits); 8353d331ad7SAl Viro 83640064aecSDennis Zhou (Facebook) /* update metadata */ 83740064aecSDennis Zhou (Facebook) chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; 838fbf59bc9STejun Heo 83986b442fbSDennis Zhou (Facebook) /* update first free bit */ 84086b442fbSDennis Zhou (Facebook) chunk->first_bit = min(chunk->first_bit, bit_off); 84186b442fbSDennis Zhou (Facebook) 842ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_free(chunk, bit_off, bits); 843b539b87fSTejun Heo 844fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, oslot); 845fbf59bc9STejun Heo } 846fbf59bc9STejun Heo 847ca460b3cSDennis Zhou (Facebook) static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) 848ca460b3cSDennis Zhou (Facebook) { 849ca460b3cSDennis Zhou (Facebook) struct pcpu_block_md *md_block; 850ca460b3cSDennis Zhou (Facebook) 851ca460b3cSDennis Zhou (Facebook) for (md_block = chunk->md_blocks; 852ca460b3cSDennis Zhou (Facebook) md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); 853ca460b3cSDennis Zhou (Facebook) md_block++) { 854ca460b3cSDennis Zhou (Facebook) md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; 855ca460b3cSDennis Zhou (Facebook) md_block->left_free = PCPU_BITMAP_BLOCK_BITS; 856ca460b3cSDennis Zhou (Facebook) md_block->right_free = PCPU_BITMAP_BLOCK_BITS; 857ca460b3cSDennis Zhou (Facebook) } 858ca460b3cSDennis Zhou (Facebook) } 859ca460b3cSDennis Zhou (Facebook) 86040064aecSDennis Zhou (Facebook) /** 86140064aecSDennis Zhou (Facebook) * pcpu_alloc_first_chunk - creates chunks that serve the first chunk 86240064aecSDennis Zhou (Facebook) * @tmp_addr: the start of the region served 86340064aecSDennis Zhou (Facebook) * @map_size: size of the region served 86440064aecSDennis Zhou (Facebook) * 86540064aecSDennis Zhou (Facebook) * This is responsible for creating the chunks that serve the first chunk. The 86640064aecSDennis Zhou (Facebook) * base_addr is page aligned down of @tmp_addr while the region end is page 86740064aecSDennis Zhou (Facebook) * aligned up. Offsets are kept track of to determine the region served. All 86840064aecSDennis Zhou (Facebook) * this is done to appease the bitmap allocator in avoiding partial blocks. 86940064aecSDennis Zhou (Facebook) * 87040064aecSDennis Zhou (Facebook) * RETURNS: 87140064aecSDennis Zhou (Facebook) * Chunk serving the region at @tmp_addr of @map_size. 87240064aecSDennis Zhou (Facebook) */ 873c0ebfdc3SDennis Zhou (Facebook) static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, 87440064aecSDennis Zhou (Facebook) int map_size) 87510edf5b0SDennis Zhou (Facebook) { 87610edf5b0SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 877ca460b3cSDennis Zhou (Facebook) unsigned long aligned_addr, lcm_align; 87840064aecSDennis Zhou (Facebook) int start_offset, offset_bits, region_size, region_bits; 879c0ebfdc3SDennis Zhou (Facebook) 880c0ebfdc3SDennis Zhou (Facebook) /* region calculations */ 881c0ebfdc3SDennis Zhou (Facebook) aligned_addr = tmp_addr & PAGE_MASK; 882c0ebfdc3SDennis Zhou (Facebook) 883c0ebfdc3SDennis Zhou (Facebook) start_offset = tmp_addr - aligned_addr; 8846b9d7c8eSDennis Zhou (Facebook) 885ca460b3cSDennis Zhou (Facebook) /* 886ca460b3cSDennis Zhou (Facebook) * Align the end of the region with the LCM of PAGE_SIZE and 887ca460b3cSDennis Zhou (Facebook) * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of 888ca460b3cSDennis Zhou (Facebook) * the other. 889ca460b3cSDennis Zhou (Facebook) */ 890ca460b3cSDennis Zhou (Facebook) lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); 891ca460b3cSDennis Zhou (Facebook) region_size = ALIGN(start_offset + map_size, lcm_align); 89210edf5b0SDennis Zhou (Facebook) 893c0ebfdc3SDennis Zhou (Facebook) /* allocate chunk */ 8948ab16c43SDennis Zhou (Facebook) chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + 8958ab16c43SDennis Zhou (Facebook) BITS_TO_LONGS(region_size >> PAGE_SHIFT), 8968ab16c43SDennis Zhou (Facebook) 0); 897c0ebfdc3SDennis Zhou (Facebook) 89810edf5b0SDennis Zhou (Facebook) INIT_LIST_HEAD(&chunk->list); 899c0ebfdc3SDennis Zhou (Facebook) 900c0ebfdc3SDennis Zhou (Facebook) chunk->base_addr = (void *)aligned_addr; 90110edf5b0SDennis Zhou (Facebook) chunk->start_offset = start_offset; 9026b9d7c8eSDennis Zhou (Facebook) chunk->end_offset = region_size - chunk->start_offset - map_size; 903c0ebfdc3SDennis Zhou (Facebook) 9048ab16c43SDennis Zhou (Facebook) chunk->nr_pages = region_size >> PAGE_SHIFT; 90540064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 906c0ebfdc3SDennis Zhou (Facebook) 907ca460b3cSDennis Zhou (Facebook) chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) * 90840064aecSDennis Zhou (Facebook) sizeof(chunk->alloc_map[0]), 0); 909ca460b3cSDennis Zhou (Facebook) chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) * 91040064aecSDennis Zhou (Facebook) sizeof(chunk->bound_map[0]), 0); 911ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) * 912ca460b3cSDennis Zhou (Facebook) sizeof(chunk->md_blocks[0]), 0); 913ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 91410edf5b0SDennis Zhou (Facebook) 91510edf5b0SDennis Zhou (Facebook) /* manage populated page bitmap */ 91610edf5b0SDennis Zhou (Facebook) chunk->immutable = true; 9178ab16c43SDennis Zhou (Facebook) bitmap_fill(chunk->populated, chunk->nr_pages); 9188ab16c43SDennis Zhou (Facebook) chunk->nr_populated = chunk->nr_pages; 91940064aecSDennis Zhou (Facebook) chunk->nr_empty_pop_pages = 92040064aecSDennis Zhou (Facebook) pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE, 92140064aecSDennis Zhou (Facebook) map_size / PCPU_MIN_ALLOC_SIZE); 92210edf5b0SDennis Zhou (Facebook) 92340064aecSDennis Zhou (Facebook) chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; 92440064aecSDennis Zhou (Facebook) chunk->free_bytes = map_size; 925c0ebfdc3SDennis Zhou (Facebook) 926c0ebfdc3SDennis Zhou (Facebook) if (chunk->start_offset) { 927c0ebfdc3SDennis Zhou (Facebook) /* hide the beginning of the bitmap */ 92840064aecSDennis Zhou (Facebook) offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; 92940064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 0, offset_bits); 93040064aecSDennis Zhou (Facebook) set_bit(0, chunk->bound_map); 93140064aecSDennis Zhou (Facebook) set_bit(offset_bits, chunk->bound_map); 932ca460b3cSDennis Zhou (Facebook) 93386b442fbSDennis Zhou (Facebook) chunk->first_bit = offset_bits; 93486b442fbSDennis Zhou (Facebook) 935ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, 0, offset_bits); 936c0ebfdc3SDennis Zhou (Facebook) } 937c0ebfdc3SDennis Zhou (Facebook) 9386b9d7c8eSDennis Zhou (Facebook) if (chunk->end_offset) { 9396b9d7c8eSDennis Zhou (Facebook) /* hide the end of the bitmap */ 94040064aecSDennis Zhou (Facebook) offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; 94140064aecSDennis Zhou (Facebook) bitmap_set(chunk->alloc_map, 94240064aecSDennis Zhou (Facebook) pcpu_chunk_map_bits(chunk) - offset_bits, 94340064aecSDennis Zhou (Facebook) offset_bits); 94440064aecSDennis Zhou (Facebook) set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, 94540064aecSDennis Zhou (Facebook) chunk->bound_map); 94640064aecSDennis Zhou (Facebook) set_bit(region_bits, chunk->bound_map); 9476b9d7c8eSDennis Zhou (Facebook) 948ca460b3cSDennis Zhou (Facebook) pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) 949ca460b3cSDennis Zhou (Facebook) - offset_bits, offset_bits); 950ca460b3cSDennis Zhou (Facebook) } 95140064aecSDennis Zhou (Facebook) 95210edf5b0SDennis Zhou (Facebook) return chunk; 95310edf5b0SDennis Zhou (Facebook) } 95410edf5b0SDennis Zhou (Facebook) 9556081089fSTejun Heo static struct pcpu_chunk *pcpu_alloc_chunk(void) 9566081089fSTejun Heo { 9576081089fSTejun Heo struct pcpu_chunk *chunk; 95840064aecSDennis Zhou (Facebook) int region_bits; 9596081089fSTejun Heo 96090459ce0SBob Liu chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); 9616081089fSTejun Heo if (!chunk) 9626081089fSTejun Heo return NULL; 9636081089fSTejun Heo 9646081089fSTejun Heo INIT_LIST_HEAD(&chunk->list); 965c0ebfdc3SDennis Zhou (Facebook) chunk->nr_pages = pcpu_unit_pages; 96640064aecSDennis Zhou (Facebook) region_bits = pcpu_chunk_map_bits(chunk); 96740064aecSDennis Zhou (Facebook) 96840064aecSDennis Zhou (Facebook) chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * 96940064aecSDennis Zhou (Facebook) sizeof(chunk->alloc_map[0])); 97040064aecSDennis Zhou (Facebook) if (!chunk->alloc_map) 97140064aecSDennis Zhou (Facebook) goto alloc_map_fail; 97240064aecSDennis Zhou (Facebook) 97340064aecSDennis Zhou (Facebook) chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * 97440064aecSDennis Zhou (Facebook) sizeof(chunk->bound_map[0])); 97540064aecSDennis Zhou (Facebook) if (!chunk->bound_map) 97640064aecSDennis Zhou (Facebook) goto bound_map_fail; 97740064aecSDennis Zhou (Facebook) 978ca460b3cSDennis Zhou (Facebook) chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * 979ca460b3cSDennis Zhou (Facebook) sizeof(chunk->md_blocks[0])); 980ca460b3cSDennis Zhou (Facebook) if (!chunk->md_blocks) 981ca460b3cSDennis Zhou (Facebook) goto md_blocks_fail; 982ca460b3cSDennis Zhou (Facebook) 983ca460b3cSDennis Zhou (Facebook) pcpu_init_md_blocks(chunk); 984ca460b3cSDennis Zhou (Facebook) 98540064aecSDennis Zhou (Facebook) /* init metadata */ 98640064aecSDennis Zhou (Facebook) chunk->contig_bits = region_bits; 98740064aecSDennis Zhou (Facebook) chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; 988c0ebfdc3SDennis Zhou (Facebook) 9896081089fSTejun Heo return chunk; 99040064aecSDennis Zhou (Facebook) 991ca460b3cSDennis Zhou (Facebook) md_blocks_fail: 992ca460b3cSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 99340064aecSDennis Zhou (Facebook) bound_map_fail: 99440064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 99540064aecSDennis Zhou (Facebook) alloc_map_fail: 99640064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk); 99740064aecSDennis Zhou (Facebook) 99840064aecSDennis Zhou (Facebook) return NULL; 9996081089fSTejun Heo } 10006081089fSTejun Heo 10016081089fSTejun Heo static void pcpu_free_chunk(struct pcpu_chunk *chunk) 10026081089fSTejun Heo { 10036081089fSTejun Heo if (!chunk) 10046081089fSTejun Heo return; 100540064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->bound_map); 100640064aecSDennis Zhou (Facebook) pcpu_mem_free(chunk->alloc_map); 10071d5cfdb0STetsuo Handa pcpu_mem_free(chunk); 10086081089fSTejun Heo } 10096081089fSTejun Heo 1010b539b87fSTejun Heo /** 1011b539b87fSTejun Heo * pcpu_chunk_populated - post-population bookkeeping 1012b539b87fSTejun Heo * @chunk: pcpu_chunk which got populated 1013b539b87fSTejun Heo * @page_start: the start page 1014b539b87fSTejun Heo * @page_end: the end page 101540064aecSDennis Zhou (Facebook) * @for_alloc: if this is to populate for allocation 1016b539b87fSTejun Heo * 1017b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been populated to @chunk. Update 1018b539b87fSTejun Heo * the bookkeeping information accordingly. Must be called after each 1019b539b87fSTejun Heo * successful population. 102040064aecSDennis Zhou (Facebook) * 102140064aecSDennis Zhou (Facebook) * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it 102240064aecSDennis Zhou (Facebook) * is to serve an allocation in that area. 1023b539b87fSTejun Heo */ 102440064aecSDennis Zhou (Facebook) static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, 102540064aecSDennis Zhou (Facebook) int page_end, bool for_alloc) 1026b539b87fSTejun Heo { 1027b539b87fSTejun Heo int nr = page_end - page_start; 1028b539b87fSTejun Heo 1029b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1030b539b87fSTejun Heo 1031b539b87fSTejun Heo bitmap_set(chunk->populated, page_start, nr); 1032b539b87fSTejun Heo chunk->nr_populated += nr; 103340064aecSDennis Zhou (Facebook) 103440064aecSDennis Zhou (Facebook) if (!for_alloc) { 10350cecf50cSDennis Zhou (Facebook) chunk->nr_empty_pop_pages += nr; 1036b539b87fSTejun Heo pcpu_nr_empty_pop_pages += nr; 1037b539b87fSTejun Heo } 103840064aecSDennis Zhou (Facebook) } 1039b539b87fSTejun Heo 1040b539b87fSTejun Heo /** 1041b539b87fSTejun Heo * pcpu_chunk_depopulated - post-depopulation bookkeeping 1042b539b87fSTejun Heo * @chunk: pcpu_chunk which got depopulated 1043b539b87fSTejun Heo * @page_start: the start page 1044b539b87fSTejun Heo * @page_end: the end page 1045b539b87fSTejun Heo * 1046b539b87fSTejun Heo * Pages in [@page_start,@page_end) have been depopulated from @chunk. 1047b539b87fSTejun Heo * Update the bookkeeping information accordingly. Must be called after 1048b539b87fSTejun Heo * each successful depopulation. 1049b539b87fSTejun Heo */ 1050b539b87fSTejun Heo static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, 1051b539b87fSTejun Heo int page_start, int page_end) 1052b539b87fSTejun Heo { 1053b539b87fSTejun Heo int nr = page_end - page_start; 1054b539b87fSTejun Heo 1055b539b87fSTejun Heo lockdep_assert_held(&pcpu_lock); 1056b539b87fSTejun Heo 1057b539b87fSTejun Heo bitmap_clear(chunk->populated, page_start, nr); 1058b539b87fSTejun Heo chunk->nr_populated -= nr; 10590cecf50cSDennis Zhou (Facebook) chunk->nr_empty_pop_pages -= nr; 1060b539b87fSTejun Heo pcpu_nr_empty_pop_pages -= nr; 1061b539b87fSTejun Heo } 1062b539b87fSTejun Heo 1063fbf59bc9STejun Heo /* 10649f645532STejun Heo * Chunk management implementation. 1065fbf59bc9STejun Heo * 10669f645532STejun Heo * To allow different implementations, chunk alloc/free and 10679f645532STejun Heo * [de]population are implemented in a separate file which is pulled 10689f645532STejun Heo * into this file and compiled together. The following functions 10699f645532STejun Heo * should be implemented. 1070ccea34b5STejun Heo * 10719f645532STejun Heo * pcpu_populate_chunk - populate the specified range of a chunk 10729f645532STejun Heo * pcpu_depopulate_chunk - depopulate the specified range of a chunk 10739f645532STejun Heo * pcpu_create_chunk - create a new chunk 10749f645532STejun Heo * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop 10759f645532STejun Heo * pcpu_addr_to_page - translate address to physical address 10769f645532STejun Heo * pcpu_verify_alloc_info - check alloc_info is acceptable during init 1077fbf59bc9STejun Heo */ 10789f645532STejun Heo static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); 10799f645532STejun Heo static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); 10809f645532STejun Heo static struct pcpu_chunk *pcpu_create_chunk(void); 10819f645532STejun Heo static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); 10829f645532STejun Heo static struct page *pcpu_addr_to_page(void *addr); 10839f645532STejun Heo static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); 1084fbf59bc9STejun Heo 1085b0c9778bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_KM 1086b0c9778bSTejun Heo #include "percpu-km.c" 1087b0c9778bSTejun Heo #else 10889f645532STejun Heo #include "percpu-vm.c" 1089b0c9778bSTejun Heo #endif 1090fbf59bc9STejun Heo 1091fbf59bc9STejun Heo /** 109288999a89STejun Heo * pcpu_chunk_addr_search - determine chunk containing specified address 109388999a89STejun Heo * @addr: address for which the chunk needs to be determined. 109488999a89STejun Heo * 1095c0ebfdc3SDennis Zhou (Facebook) * This is an internal function that handles all but static allocations. 1096c0ebfdc3SDennis Zhou (Facebook) * Static percpu address values should never be passed into the allocator. 1097c0ebfdc3SDennis Zhou (Facebook) * 109888999a89STejun Heo * RETURNS: 109988999a89STejun Heo * The address of the found chunk. 110088999a89STejun Heo */ 110188999a89STejun Heo static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 110288999a89STejun Heo { 1103c0ebfdc3SDennis Zhou (Facebook) /* is it in the dynamic region (first chunk)? */ 1104560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) 1105c0ebfdc3SDennis Zhou (Facebook) return pcpu_first_chunk; 1106c0ebfdc3SDennis Zhou (Facebook) 1107c0ebfdc3SDennis Zhou (Facebook) /* is it in the reserved region? */ 1108560f2c23SDennis Zhou (Facebook) if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) 110988999a89STejun Heo return pcpu_reserved_chunk; 111088999a89STejun Heo 111188999a89STejun Heo /* 111288999a89STejun Heo * The address is relative to unit0 which might be unused and 111388999a89STejun Heo * thus unmapped. Offset the address to the unit space of the 111488999a89STejun Heo * current processor before looking it up in the vmalloc 111588999a89STejun Heo * space. Note that any possible cpu id can be used here, so 111688999a89STejun Heo * there's no need to worry about preemption or cpu hotplug. 111788999a89STejun Heo */ 111888999a89STejun Heo addr += pcpu_unit_offsets[raw_smp_processor_id()]; 11199f645532STejun Heo return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); 112088999a89STejun Heo } 112188999a89STejun Heo 112288999a89STejun Heo /** 1123edcb4639STejun Heo * pcpu_alloc - the percpu allocator 1124cae3aeb8STejun Heo * @size: size of area to allocate in bytes 1125fbf59bc9STejun Heo * @align: alignment of area (max PAGE_SIZE) 1126edcb4639STejun Heo * @reserved: allocate from the reserved chunk if available 11275835d96eSTejun Heo * @gfp: allocation flags 1128fbf59bc9STejun Heo * 11295835d96eSTejun Heo * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't 11305835d96eSTejun Heo * contain %GFP_KERNEL, the allocation is atomic. 1131fbf59bc9STejun Heo * 1132fbf59bc9STejun Heo * RETURNS: 1133fbf59bc9STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1134fbf59bc9STejun Heo */ 11355835d96eSTejun Heo static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, 11365835d96eSTejun Heo gfp_t gfp) 1137fbf59bc9STejun Heo { 1138f2badb0cSTejun Heo static int warn_limit = 10; 1139fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1140f2badb0cSTejun Heo const char *err; 11416ae833c7STejun Heo bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; 114240064aecSDennis Zhou (Facebook) int slot, off, cpu, ret; 1143403a91b1SJiri Kosina unsigned long flags; 1144f528f0b8SCatalin Marinas void __percpu *ptr; 114540064aecSDennis Zhou (Facebook) size_t bits, bit_align; 1146fbf59bc9STejun Heo 1147723ad1d9SAl Viro /* 114840064aecSDennis Zhou (Facebook) * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, 114940064aecSDennis Zhou (Facebook) * therefore alignment must be a minimum of that many bytes. 115040064aecSDennis Zhou (Facebook) * An allocation may have internal fragmentation from rounding up 115140064aecSDennis Zhou (Facebook) * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. 1152723ad1d9SAl Viro */ 1153d2f3c384SDennis Zhou (Facebook) if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) 1154d2f3c384SDennis Zhou (Facebook) align = PCPU_MIN_ALLOC_SIZE; 1155723ad1d9SAl Viro 1156d2f3c384SDennis Zhou (Facebook) size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); 115740064aecSDennis Zhou (Facebook) bits = size >> PCPU_MIN_ALLOC_SHIFT; 115840064aecSDennis Zhou (Facebook) bit_align = align >> PCPU_MIN_ALLOC_SHIFT; 11592f69fa82SViro 11603ca45a46Szijun_hu if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || 11613ca45a46Szijun_hu !is_power_of_2(align))) { 1162756a025fSJoe Perches WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", 1163756a025fSJoe Perches size, align); 1164fbf59bc9STejun Heo return NULL; 1165fbf59bc9STejun Heo } 1166fbf59bc9STejun Heo 11676710e594STejun Heo if (!is_atomic) 11686710e594STejun Heo mutex_lock(&pcpu_alloc_mutex); 11696710e594STejun Heo 1170403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1171fbf59bc9STejun Heo 1172edcb4639STejun Heo /* serve reserved allocations from the reserved chunk if available */ 1173edcb4639STejun Heo if (reserved && pcpu_reserved_chunk) { 1174edcb4639STejun Heo chunk = pcpu_reserved_chunk; 1175833af842STejun Heo 117640064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); 117740064aecSDennis Zhou (Facebook) if (off < 0) { 1178833af842STejun Heo err = "alloc from reserved chunk failed"; 1179ccea34b5STejun Heo goto fail_unlock; 1180f2badb0cSTejun Heo } 1181833af842STejun Heo 118240064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1183edcb4639STejun Heo if (off >= 0) 1184edcb4639STejun Heo goto area_found; 1185833af842STejun Heo 1186f2badb0cSTejun Heo err = "alloc from reserved chunk failed"; 1187ccea34b5STejun Heo goto fail_unlock; 1188edcb4639STejun Heo } 1189edcb4639STejun Heo 1190ccea34b5STejun Heo restart: 1191edcb4639STejun Heo /* search through normal chunks */ 1192fbf59bc9STejun Heo for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { 1193fbf59bc9STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 119440064aecSDennis Zhou (Facebook) off = pcpu_find_block_fit(chunk, bits, bit_align, 119540064aecSDennis Zhou (Facebook) is_atomic); 119640064aecSDennis Zhou (Facebook) if (off < 0) 1197fbf59bc9STejun Heo continue; 1198ccea34b5STejun Heo 119940064aecSDennis Zhou (Facebook) off = pcpu_alloc_area(chunk, bits, bit_align, off); 1200fbf59bc9STejun Heo if (off >= 0) 1201fbf59bc9STejun Heo goto area_found; 120240064aecSDennis Zhou (Facebook) 1203fbf59bc9STejun Heo } 1204fbf59bc9STejun Heo } 1205fbf59bc9STejun Heo 1206403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1207ccea34b5STejun Heo 1208b38d08f3STejun Heo /* 1209b38d08f3STejun Heo * No space left. Create a new chunk. We don't want multiple 1210b38d08f3STejun Heo * tasks to create chunks simultaneously. Serialize and create iff 1211b38d08f3STejun Heo * there's still no empty chunk after grabbing the mutex. 1212b38d08f3STejun Heo */ 121311df02bfSDennis Zhou if (is_atomic) { 121411df02bfSDennis Zhou err = "atomic alloc failed, no space left"; 12155835d96eSTejun Heo goto fail; 121611df02bfSDennis Zhou } 12175835d96eSTejun Heo 1218b38d08f3STejun Heo if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { 12196081089fSTejun Heo chunk = pcpu_create_chunk(); 1220f2badb0cSTejun Heo if (!chunk) { 1221f2badb0cSTejun Heo err = "failed to allocate new chunk"; 1222b38d08f3STejun Heo goto fail; 1223f2badb0cSTejun Heo } 1224ccea34b5STejun Heo 1225403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1226fbf59bc9STejun Heo pcpu_chunk_relocate(chunk, -1); 1227b38d08f3STejun Heo } else { 1228b38d08f3STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1229b38d08f3STejun Heo } 1230b38d08f3STejun Heo 1231ccea34b5STejun Heo goto restart; 1232fbf59bc9STejun Heo 1233fbf59bc9STejun Heo area_found: 123430a5b536SDennis Zhou pcpu_stats_area_alloc(chunk, size); 1235403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1236ccea34b5STejun Heo 1237dca49645STejun Heo /* populate if not all pages are already there */ 12385835d96eSTejun Heo if (!is_atomic) { 1239e04d3208STejun Heo int page_start, page_end, rs, re; 1240e04d3208STejun Heo 1241dca49645STejun Heo page_start = PFN_DOWN(off); 1242dca49645STejun Heo page_end = PFN_UP(off + size); 1243dca49645STejun Heo 124491e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 124591e914c5SDennis Zhou (Facebook) page_start, page_end) { 1246dca49645STejun Heo WARN_ON(chunk->immutable); 1247dca49645STejun Heo 1248b38d08f3STejun Heo ret = pcpu_populate_chunk(chunk, rs, re); 1249b38d08f3STejun Heo 1250403a91b1SJiri Kosina spin_lock_irqsave(&pcpu_lock, flags); 1251b38d08f3STejun Heo if (ret) { 125240064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1253f2badb0cSTejun Heo err = "failed to populate"; 1254ccea34b5STejun Heo goto fail_unlock; 1255fbf59bc9STejun Heo } 125640064aecSDennis Zhou (Facebook) pcpu_chunk_populated(chunk, rs, re, true); 1257b38d08f3STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1258dca49645STejun Heo } 1259dca49645STejun Heo 1260ccea34b5STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1261e04d3208STejun Heo } 1262ccea34b5STejun Heo 12631a4d7607STejun Heo if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) 12641a4d7607STejun Heo pcpu_schedule_balance_work(); 12651a4d7607STejun Heo 1266dca49645STejun Heo /* clear the areas and return address relative to base address */ 1267dca49645STejun Heo for_each_possible_cpu(cpu) 1268dca49645STejun Heo memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); 1269dca49645STejun Heo 1270f528f0b8SCatalin Marinas ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); 12718a8c35faSLarry Finger kmemleak_alloc_percpu(ptr, size, gfp); 1272df95e795SDennis Zhou 1273df95e795SDennis Zhou trace_percpu_alloc_percpu(reserved, is_atomic, size, align, 1274df95e795SDennis Zhou chunk->base_addr, off, ptr); 1275df95e795SDennis Zhou 1276f528f0b8SCatalin Marinas return ptr; 1277ccea34b5STejun Heo 1278ccea34b5STejun Heo fail_unlock: 1279403a91b1SJiri Kosina spin_unlock_irqrestore(&pcpu_lock, flags); 1280b38d08f3STejun Heo fail: 1281df95e795SDennis Zhou trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); 1282df95e795SDennis Zhou 12835835d96eSTejun Heo if (!is_atomic && warn_limit) { 1284870d4b12SJoe Perches pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", 12855835d96eSTejun Heo size, align, is_atomic, err); 1286f2badb0cSTejun Heo dump_stack(); 1287f2badb0cSTejun Heo if (!--warn_limit) 1288870d4b12SJoe Perches pr_info("limit reached, disable warning\n"); 1289f2badb0cSTejun Heo } 12901a4d7607STejun Heo if (is_atomic) { 12911a4d7607STejun Heo /* see the flag handling in pcpu_blance_workfn() */ 12921a4d7607STejun Heo pcpu_atomic_alloc_failed = true; 12931a4d7607STejun Heo pcpu_schedule_balance_work(); 12946710e594STejun Heo } else { 12956710e594STejun Heo mutex_unlock(&pcpu_alloc_mutex); 12961a4d7607STejun Heo } 1297ccea34b5STejun Heo return NULL; 1298fbf59bc9STejun Heo } 1299edcb4639STejun Heo 1300edcb4639STejun Heo /** 13015835d96eSTejun Heo * __alloc_percpu_gfp - allocate dynamic percpu area 1302edcb4639STejun Heo * @size: size of area to allocate in bytes 1303edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 13045835d96eSTejun Heo * @gfp: allocation flags 1305edcb4639STejun Heo * 13065835d96eSTejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align. If 13075835d96eSTejun Heo * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can 13085835d96eSTejun Heo * be called from any context but is a lot more likely to fail. 1309ccea34b5STejun Heo * 1310edcb4639STejun Heo * RETURNS: 1311edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1312edcb4639STejun Heo */ 13135835d96eSTejun Heo void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) 13145835d96eSTejun Heo { 13155835d96eSTejun Heo return pcpu_alloc(size, align, false, gfp); 13165835d96eSTejun Heo } 13175835d96eSTejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); 13185835d96eSTejun Heo 13195835d96eSTejun Heo /** 13205835d96eSTejun Heo * __alloc_percpu - allocate dynamic percpu area 13215835d96eSTejun Heo * @size: size of area to allocate in bytes 13225835d96eSTejun Heo * @align: alignment of area (max PAGE_SIZE) 13235835d96eSTejun Heo * 13245835d96eSTejun Heo * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). 13255835d96eSTejun Heo */ 132643cf38ebSTejun Heo void __percpu *__alloc_percpu(size_t size, size_t align) 1327edcb4639STejun Heo { 13285835d96eSTejun Heo return pcpu_alloc(size, align, false, GFP_KERNEL); 1329edcb4639STejun Heo } 1330fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(__alloc_percpu); 1331fbf59bc9STejun Heo 1332edcb4639STejun Heo /** 1333edcb4639STejun Heo * __alloc_reserved_percpu - allocate reserved percpu area 1334edcb4639STejun Heo * @size: size of area to allocate in bytes 1335edcb4639STejun Heo * @align: alignment of area (max PAGE_SIZE) 1336edcb4639STejun Heo * 13379329ba97STejun Heo * Allocate zero-filled percpu area of @size bytes aligned at @align 13389329ba97STejun Heo * from reserved percpu area if arch has set it up; otherwise, 13399329ba97STejun Heo * allocation is served from the same dynamic area. Might sleep. 13409329ba97STejun Heo * Might trigger writeouts. 1341edcb4639STejun Heo * 1342ccea34b5STejun Heo * CONTEXT: 1343ccea34b5STejun Heo * Does GFP_KERNEL allocation. 1344ccea34b5STejun Heo * 1345edcb4639STejun Heo * RETURNS: 1346edcb4639STejun Heo * Percpu pointer to the allocated area on success, NULL on failure. 1347edcb4639STejun Heo */ 134843cf38ebSTejun Heo void __percpu *__alloc_reserved_percpu(size_t size, size_t align) 1349edcb4639STejun Heo { 13505835d96eSTejun Heo return pcpu_alloc(size, align, true, GFP_KERNEL); 1351edcb4639STejun Heo } 1352edcb4639STejun Heo 1353a56dbddfSTejun Heo /** 13541a4d7607STejun Heo * pcpu_balance_workfn - manage the amount of free chunks and populated pages 1355a56dbddfSTejun Heo * @work: unused 1356a56dbddfSTejun Heo * 1357a56dbddfSTejun Heo * Reclaim all fully free chunks except for the first one. 1358a56dbddfSTejun Heo */ 1359fe6bd8c3STejun Heo static void pcpu_balance_workfn(struct work_struct *work) 1360fbf59bc9STejun Heo { 1361fe6bd8c3STejun Heo LIST_HEAD(to_free); 1362fe6bd8c3STejun Heo struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; 1363a56dbddfSTejun Heo struct pcpu_chunk *chunk, *next; 13641a4d7607STejun Heo int slot, nr_to_pop, ret; 1365a56dbddfSTejun Heo 13661a4d7607STejun Heo /* 13671a4d7607STejun Heo * There's no reason to keep around multiple unused chunks and VM 13681a4d7607STejun Heo * areas can be scarce. Destroy all free chunks except for one. 13691a4d7607STejun Heo */ 1370ccea34b5STejun Heo mutex_lock(&pcpu_alloc_mutex); 1371ccea34b5STejun Heo spin_lock_irq(&pcpu_lock); 1372a56dbddfSTejun Heo 1373fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, free_head, list) { 13748d408b4bSTejun Heo WARN_ON(chunk->immutable); 1375a56dbddfSTejun Heo 1376a56dbddfSTejun Heo /* spare the first one */ 1377fe6bd8c3STejun Heo if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) 1378a56dbddfSTejun Heo continue; 1379a56dbddfSTejun Heo 1380fe6bd8c3STejun Heo list_move(&chunk->list, &to_free); 1381a56dbddfSTejun Heo } 1382a56dbddfSTejun Heo 1383ccea34b5STejun Heo spin_unlock_irq(&pcpu_lock); 1384a56dbddfSTejun Heo 1385fe6bd8c3STejun Heo list_for_each_entry_safe(chunk, next, &to_free, list) { 1386a93ace48STejun Heo int rs, re; 1387dca49645STejun Heo 138891e914c5SDennis Zhou (Facebook) pcpu_for_each_pop_region(chunk->populated, rs, re, 0, 138991e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 1390a93ace48STejun Heo pcpu_depopulate_chunk(chunk, rs, re); 1391b539b87fSTejun Heo spin_lock_irq(&pcpu_lock); 1392b539b87fSTejun Heo pcpu_chunk_depopulated(chunk, rs, re); 1393b539b87fSTejun Heo spin_unlock_irq(&pcpu_lock); 1394a93ace48STejun Heo } 13956081089fSTejun Heo pcpu_destroy_chunk(chunk); 1396fbf59bc9STejun Heo } 1397971f3918STejun Heo 13981a4d7607STejun Heo /* 13991a4d7607STejun Heo * Ensure there are certain number of free populated pages for 14001a4d7607STejun Heo * atomic allocs. Fill up from the most packed so that atomic 14011a4d7607STejun Heo * allocs don't increase fragmentation. If atomic allocation 14021a4d7607STejun Heo * failed previously, always populate the maximum amount. This 14031a4d7607STejun Heo * should prevent atomic allocs larger than PAGE_SIZE from keeping 14041a4d7607STejun Heo * failing indefinitely; however, large atomic allocs are not 14051a4d7607STejun Heo * something we support properly and can be highly unreliable and 14061a4d7607STejun Heo * inefficient. 14071a4d7607STejun Heo */ 14081a4d7607STejun Heo retry_pop: 14091a4d7607STejun Heo if (pcpu_atomic_alloc_failed) { 14101a4d7607STejun Heo nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; 14111a4d7607STejun Heo /* best effort anyway, don't worry about synchronization */ 14121a4d7607STejun Heo pcpu_atomic_alloc_failed = false; 14131a4d7607STejun Heo } else { 14141a4d7607STejun Heo nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - 14151a4d7607STejun Heo pcpu_nr_empty_pop_pages, 14161a4d7607STejun Heo 0, PCPU_EMPTY_POP_PAGES_HIGH); 14171a4d7607STejun Heo } 14181a4d7607STejun Heo 14191a4d7607STejun Heo for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { 14201a4d7607STejun Heo int nr_unpop = 0, rs, re; 14211a4d7607STejun Heo 14221a4d7607STejun Heo if (!nr_to_pop) 14231a4d7607STejun Heo break; 14241a4d7607STejun Heo 14251a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 14261a4d7607STejun Heo list_for_each_entry(chunk, &pcpu_slot[slot], list) { 14278ab16c43SDennis Zhou (Facebook) nr_unpop = chunk->nr_pages - chunk->nr_populated; 14281a4d7607STejun Heo if (nr_unpop) 14291a4d7607STejun Heo break; 14301a4d7607STejun Heo } 14311a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 14321a4d7607STejun Heo 14331a4d7607STejun Heo if (!nr_unpop) 14341a4d7607STejun Heo continue; 14351a4d7607STejun Heo 14361a4d7607STejun Heo /* @chunk can't go away while pcpu_alloc_mutex is held */ 143791e914c5SDennis Zhou (Facebook) pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, 143891e914c5SDennis Zhou (Facebook) chunk->nr_pages) { 14391a4d7607STejun Heo int nr = min(re - rs, nr_to_pop); 14401a4d7607STejun Heo 14411a4d7607STejun Heo ret = pcpu_populate_chunk(chunk, rs, rs + nr); 14421a4d7607STejun Heo if (!ret) { 14431a4d7607STejun Heo nr_to_pop -= nr; 14441a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 144540064aecSDennis Zhou (Facebook) pcpu_chunk_populated(chunk, rs, rs + nr, false); 14461a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 14471a4d7607STejun Heo } else { 14481a4d7607STejun Heo nr_to_pop = 0; 14491a4d7607STejun Heo } 14501a4d7607STejun Heo 14511a4d7607STejun Heo if (!nr_to_pop) 14521a4d7607STejun Heo break; 14531a4d7607STejun Heo } 14541a4d7607STejun Heo } 14551a4d7607STejun Heo 14561a4d7607STejun Heo if (nr_to_pop) { 14571a4d7607STejun Heo /* ran out of chunks to populate, create a new one and retry */ 14581a4d7607STejun Heo chunk = pcpu_create_chunk(); 14591a4d7607STejun Heo if (chunk) { 14601a4d7607STejun Heo spin_lock_irq(&pcpu_lock); 14611a4d7607STejun Heo pcpu_chunk_relocate(chunk, -1); 14621a4d7607STejun Heo spin_unlock_irq(&pcpu_lock); 14631a4d7607STejun Heo goto retry_pop; 14641a4d7607STejun Heo } 14651a4d7607STejun Heo } 14661a4d7607STejun Heo 1467971f3918STejun Heo mutex_unlock(&pcpu_alloc_mutex); 1468a56dbddfSTejun Heo } 1469fbf59bc9STejun Heo 1470fbf59bc9STejun Heo /** 1471fbf59bc9STejun Heo * free_percpu - free percpu area 1472fbf59bc9STejun Heo * @ptr: pointer to area to free 1473fbf59bc9STejun Heo * 1474ccea34b5STejun Heo * Free percpu area @ptr. 1475ccea34b5STejun Heo * 1476ccea34b5STejun Heo * CONTEXT: 1477ccea34b5STejun Heo * Can be called from atomic context. 1478fbf59bc9STejun Heo */ 147943cf38ebSTejun Heo void free_percpu(void __percpu *ptr) 1480fbf59bc9STejun Heo { 1481129182e5SAndrew Morton void *addr; 1482fbf59bc9STejun Heo struct pcpu_chunk *chunk; 1483ccea34b5STejun Heo unsigned long flags; 148440064aecSDennis Zhou (Facebook) int off; 1485fbf59bc9STejun Heo 1486fbf59bc9STejun Heo if (!ptr) 1487fbf59bc9STejun Heo return; 1488fbf59bc9STejun Heo 1489f528f0b8SCatalin Marinas kmemleak_free_percpu(ptr); 1490f528f0b8SCatalin Marinas 1491129182e5SAndrew Morton addr = __pcpu_ptr_to_addr(ptr); 1492129182e5SAndrew Morton 1493ccea34b5STejun Heo spin_lock_irqsave(&pcpu_lock, flags); 1494fbf59bc9STejun Heo 1495fbf59bc9STejun Heo chunk = pcpu_chunk_addr_search(addr); 1496bba174f5STejun Heo off = addr - chunk->base_addr; 1497fbf59bc9STejun Heo 149840064aecSDennis Zhou (Facebook) pcpu_free_area(chunk, off); 1499fbf59bc9STejun Heo 1500a56dbddfSTejun Heo /* if there are more than one fully free chunks, wake up grim reaper */ 150140064aecSDennis Zhou (Facebook) if (chunk->free_bytes == pcpu_unit_size) { 1502fbf59bc9STejun Heo struct pcpu_chunk *pos; 1503fbf59bc9STejun Heo 1504a56dbddfSTejun Heo list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) 1505fbf59bc9STejun Heo if (pos != chunk) { 15061a4d7607STejun Heo pcpu_schedule_balance_work(); 1507fbf59bc9STejun Heo break; 1508fbf59bc9STejun Heo } 1509fbf59bc9STejun Heo } 1510fbf59bc9STejun Heo 1511df95e795SDennis Zhou trace_percpu_free_percpu(chunk->base_addr, off, ptr); 1512df95e795SDennis Zhou 1513ccea34b5STejun Heo spin_unlock_irqrestore(&pcpu_lock, flags); 1514fbf59bc9STejun Heo } 1515fbf59bc9STejun Heo EXPORT_SYMBOL_GPL(free_percpu); 1516fbf59bc9STejun Heo 1517383776faSThomas Gleixner bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) 1518383776faSThomas Gleixner { 1519383776faSThomas Gleixner #ifdef CONFIG_SMP 1520383776faSThomas Gleixner const size_t static_size = __per_cpu_end - __per_cpu_start; 1521383776faSThomas Gleixner void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 1522383776faSThomas Gleixner unsigned int cpu; 1523383776faSThomas Gleixner 1524383776faSThomas Gleixner for_each_possible_cpu(cpu) { 1525383776faSThomas Gleixner void *start = per_cpu_ptr(base, cpu); 1526383776faSThomas Gleixner void *va = (void *)addr; 1527383776faSThomas Gleixner 1528383776faSThomas Gleixner if (va >= start && va < start + static_size) { 15298ce371f9SPeter Zijlstra if (can_addr) { 1530383776faSThomas Gleixner *can_addr = (unsigned long) (va - start); 15318ce371f9SPeter Zijlstra *can_addr += (unsigned long) 15328ce371f9SPeter Zijlstra per_cpu_ptr(base, get_boot_cpu_id()); 15338ce371f9SPeter Zijlstra } 1534383776faSThomas Gleixner return true; 1535383776faSThomas Gleixner } 1536383776faSThomas Gleixner } 1537383776faSThomas Gleixner #endif 1538383776faSThomas Gleixner /* on UP, can't distinguish from other static vars, always false */ 1539383776faSThomas Gleixner return false; 1540383776faSThomas Gleixner } 1541383776faSThomas Gleixner 15423b034b0dSVivek Goyal /** 154310fad5e4STejun Heo * is_kernel_percpu_address - test whether address is from static percpu area 154410fad5e4STejun Heo * @addr: address to test 154510fad5e4STejun Heo * 154610fad5e4STejun Heo * Test whether @addr belongs to in-kernel static percpu area. Module 154710fad5e4STejun Heo * static percpu areas are not considered. For those, use 154810fad5e4STejun Heo * is_module_percpu_address(). 154910fad5e4STejun Heo * 155010fad5e4STejun Heo * RETURNS: 155110fad5e4STejun Heo * %true if @addr is from in-kernel static percpu area, %false otherwise. 155210fad5e4STejun Heo */ 155310fad5e4STejun Heo bool is_kernel_percpu_address(unsigned long addr) 155410fad5e4STejun Heo { 1555383776faSThomas Gleixner return __is_kernel_percpu_address(addr, NULL); 155610fad5e4STejun Heo } 155710fad5e4STejun Heo 155810fad5e4STejun Heo /** 15593b034b0dSVivek Goyal * per_cpu_ptr_to_phys - convert translated percpu address to physical address 15603b034b0dSVivek Goyal * @addr: the address to be converted to physical address 15613b034b0dSVivek Goyal * 15623b034b0dSVivek Goyal * Given @addr which is dereferenceable address obtained via one of 15633b034b0dSVivek Goyal * percpu access macros, this function translates it into its physical 15643b034b0dSVivek Goyal * address. The caller is responsible for ensuring @addr stays valid 15653b034b0dSVivek Goyal * until this function finishes. 15663b034b0dSVivek Goyal * 156767589c71SDave Young * percpu allocator has special setup for the first chunk, which currently 156867589c71SDave Young * supports either embedding in linear address space or vmalloc mapping, 156967589c71SDave Young * and, from the second one, the backing allocator (currently either vm or 157067589c71SDave Young * km) provides translation. 157167589c71SDave Young * 1572bffc4375SYannick Guerrini * The addr can be translated simply without checking if it falls into the 157367589c71SDave Young * first chunk. But the current code reflects better how percpu allocator 157467589c71SDave Young * actually works, and the verification can discover both bugs in percpu 157567589c71SDave Young * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current 157667589c71SDave Young * code. 157767589c71SDave Young * 15783b034b0dSVivek Goyal * RETURNS: 15793b034b0dSVivek Goyal * The physical address for @addr. 15803b034b0dSVivek Goyal */ 15813b034b0dSVivek Goyal phys_addr_t per_cpu_ptr_to_phys(void *addr) 15823b034b0dSVivek Goyal { 15839983b6f0STejun Heo void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); 15849983b6f0STejun Heo bool in_first_chunk = false; 1585a855b84cSTejun Heo unsigned long first_low, first_high; 15869983b6f0STejun Heo unsigned int cpu; 15879983b6f0STejun Heo 15889983b6f0STejun Heo /* 1589a855b84cSTejun Heo * The following test on unit_low/high isn't strictly 15909983b6f0STejun Heo * necessary but will speed up lookups of addresses which 15919983b6f0STejun Heo * aren't in the first chunk. 1592c0ebfdc3SDennis Zhou (Facebook) * 1593c0ebfdc3SDennis Zhou (Facebook) * The address check is against full chunk sizes. pcpu_base_addr 1594c0ebfdc3SDennis Zhou (Facebook) * points to the beginning of the first chunk including the 1595c0ebfdc3SDennis Zhou (Facebook) * static region. Assumes good intent as the first chunk may 1596c0ebfdc3SDennis Zhou (Facebook) * not be full (ie. < pcpu_unit_pages in size). 15979983b6f0STejun Heo */ 1598c0ebfdc3SDennis Zhou (Facebook) first_low = (unsigned long)pcpu_base_addr + 1599c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); 1600c0ebfdc3SDennis Zhou (Facebook) first_high = (unsigned long)pcpu_base_addr + 1601c0ebfdc3SDennis Zhou (Facebook) pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); 1602a855b84cSTejun Heo if ((unsigned long)addr >= first_low && 1603a855b84cSTejun Heo (unsigned long)addr < first_high) { 16049983b6f0STejun Heo for_each_possible_cpu(cpu) { 16059983b6f0STejun Heo void *start = per_cpu_ptr(base, cpu); 16069983b6f0STejun Heo 16079983b6f0STejun Heo if (addr >= start && addr < start + pcpu_unit_size) { 16089983b6f0STejun Heo in_first_chunk = true; 16099983b6f0STejun Heo break; 16109983b6f0STejun Heo } 16119983b6f0STejun Heo } 16129983b6f0STejun Heo } 16139983b6f0STejun Heo 16149983b6f0STejun Heo if (in_first_chunk) { 1615eac522efSDavid Howells if (!is_vmalloc_addr(addr)) 16163b034b0dSVivek Goyal return __pa(addr); 16173b034b0dSVivek Goyal else 16189f57bd4dSEugene Surovegin return page_to_phys(vmalloc_to_page(addr)) + 16199f57bd4dSEugene Surovegin offset_in_page(addr); 1620020ec653STejun Heo } else 16219f57bd4dSEugene Surovegin return page_to_phys(pcpu_addr_to_page(addr)) + 16229f57bd4dSEugene Surovegin offset_in_page(addr); 16233b034b0dSVivek Goyal } 16243b034b0dSVivek Goyal 1625fbf59bc9STejun Heo /** 1626fd1e8a1fSTejun Heo * pcpu_alloc_alloc_info - allocate percpu allocation info 1627fd1e8a1fSTejun Heo * @nr_groups: the number of groups 1628fd1e8a1fSTejun Heo * @nr_units: the number of units 1629033e48fbSTejun Heo * 1630fd1e8a1fSTejun Heo * Allocate ai which is large enough for @nr_groups groups containing 1631fd1e8a1fSTejun Heo * @nr_units units. The returned ai's groups[0].cpu_map points to the 1632fd1e8a1fSTejun Heo * cpu_map array which is long enough for @nr_units and filled with 1633fd1e8a1fSTejun Heo * NR_CPUS. It's the caller's responsibility to initialize cpu_map 1634fd1e8a1fSTejun Heo * pointer of other groups. 1635033e48fbSTejun Heo * 1636033e48fbSTejun Heo * RETURNS: 1637fd1e8a1fSTejun Heo * Pointer to the allocated pcpu_alloc_info on success, NULL on 1638fd1e8a1fSTejun Heo * failure. 1639033e48fbSTejun Heo */ 1640fd1e8a1fSTejun Heo struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, 1641fd1e8a1fSTejun Heo int nr_units) 1642fd1e8a1fSTejun Heo { 1643fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 1644fd1e8a1fSTejun Heo size_t base_size, ai_size; 1645fd1e8a1fSTejun Heo void *ptr; 1646fd1e8a1fSTejun Heo int unit; 1647fd1e8a1fSTejun Heo 1648fd1e8a1fSTejun Heo base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), 1649fd1e8a1fSTejun Heo __alignof__(ai->groups[0].cpu_map[0])); 1650fd1e8a1fSTejun Heo ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); 1651fd1e8a1fSTejun Heo 1652999c17e3SSantosh Shilimkar ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); 1653fd1e8a1fSTejun Heo if (!ptr) 1654fd1e8a1fSTejun Heo return NULL; 1655fd1e8a1fSTejun Heo ai = ptr; 1656fd1e8a1fSTejun Heo ptr += base_size; 1657fd1e8a1fSTejun Heo 1658fd1e8a1fSTejun Heo ai->groups[0].cpu_map = ptr; 1659fd1e8a1fSTejun Heo 1660fd1e8a1fSTejun Heo for (unit = 0; unit < nr_units; unit++) 1661fd1e8a1fSTejun Heo ai->groups[0].cpu_map[unit] = NR_CPUS; 1662fd1e8a1fSTejun Heo 1663fd1e8a1fSTejun Heo ai->nr_groups = nr_groups; 1664fd1e8a1fSTejun Heo ai->__ai_size = PFN_ALIGN(ai_size); 1665fd1e8a1fSTejun Heo 1666fd1e8a1fSTejun Heo return ai; 1667fd1e8a1fSTejun Heo } 1668fd1e8a1fSTejun Heo 1669fd1e8a1fSTejun Heo /** 1670fd1e8a1fSTejun Heo * pcpu_free_alloc_info - free percpu allocation info 1671fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info to free 1672fd1e8a1fSTejun Heo * 1673fd1e8a1fSTejun Heo * Free @ai which was allocated by pcpu_alloc_alloc_info(). 1674fd1e8a1fSTejun Heo */ 1675fd1e8a1fSTejun Heo void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) 1676fd1e8a1fSTejun Heo { 1677999c17e3SSantosh Shilimkar memblock_free_early(__pa(ai), ai->__ai_size); 1678fd1e8a1fSTejun Heo } 1679fd1e8a1fSTejun Heo 1680fd1e8a1fSTejun Heo /** 1681fd1e8a1fSTejun Heo * pcpu_dump_alloc_info - print out information about pcpu_alloc_info 1682fd1e8a1fSTejun Heo * @lvl: loglevel 1683fd1e8a1fSTejun Heo * @ai: allocation info to dump 1684fd1e8a1fSTejun Heo * 1685fd1e8a1fSTejun Heo * Print out information about @ai using loglevel @lvl. 1686fd1e8a1fSTejun Heo */ 1687fd1e8a1fSTejun Heo static void pcpu_dump_alloc_info(const char *lvl, 1688fd1e8a1fSTejun Heo const struct pcpu_alloc_info *ai) 1689033e48fbSTejun Heo { 1690fd1e8a1fSTejun Heo int group_width = 1, cpu_width = 1, width; 1691033e48fbSTejun Heo char empty_str[] = "--------"; 1692fd1e8a1fSTejun Heo int alloc = 0, alloc_end = 0; 1693fd1e8a1fSTejun Heo int group, v; 1694fd1e8a1fSTejun Heo int upa, apl; /* units per alloc, allocs per line */ 1695033e48fbSTejun Heo 1696fd1e8a1fSTejun Heo v = ai->nr_groups; 1697033e48fbSTejun Heo while (v /= 10) 1698fd1e8a1fSTejun Heo group_width++; 1699033e48fbSTejun Heo 1700fd1e8a1fSTejun Heo v = num_possible_cpus(); 1701fd1e8a1fSTejun Heo while (v /= 10) 1702fd1e8a1fSTejun Heo cpu_width++; 1703fd1e8a1fSTejun Heo empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; 1704033e48fbSTejun Heo 1705fd1e8a1fSTejun Heo upa = ai->alloc_size / ai->unit_size; 1706fd1e8a1fSTejun Heo width = upa * (cpu_width + 1) + group_width + 3; 1707fd1e8a1fSTejun Heo apl = rounddown_pow_of_two(max(60 / width, 1)); 1708033e48fbSTejun Heo 1709fd1e8a1fSTejun Heo printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", 1710fd1e8a1fSTejun Heo lvl, ai->static_size, ai->reserved_size, ai->dyn_size, 1711fd1e8a1fSTejun Heo ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); 1712fd1e8a1fSTejun Heo 1713fd1e8a1fSTejun Heo for (group = 0; group < ai->nr_groups; group++) { 1714fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 1715fd1e8a1fSTejun Heo int unit = 0, unit_end = 0; 1716fd1e8a1fSTejun Heo 1717fd1e8a1fSTejun Heo BUG_ON(gi->nr_units % upa); 1718fd1e8a1fSTejun Heo for (alloc_end += gi->nr_units / upa; 1719fd1e8a1fSTejun Heo alloc < alloc_end; alloc++) { 1720fd1e8a1fSTejun Heo if (!(alloc % apl)) { 17211170532bSJoe Perches pr_cont("\n"); 1722fd1e8a1fSTejun Heo printk("%spcpu-alloc: ", lvl); 1723033e48fbSTejun Heo } 17241170532bSJoe Perches pr_cont("[%0*d] ", group_width, group); 1725fd1e8a1fSTejun Heo 1726fd1e8a1fSTejun Heo for (unit_end += upa; unit < unit_end; unit++) 1727fd1e8a1fSTejun Heo if (gi->cpu_map[unit] != NR_CPUS) 17281170532bSJoe Perches pr_cont("%0*d ", 17291170532bSJoe Perches cpu_width, gi->cpu_map[unit]); 1730033e48fbSTejun Heo else 17311170532bSJoe Perches pr_cont("%s ", empty_str); 1732033e48fbSTejun Heo } 1733fd1e8a1fSTejun Heo } 17341170532bSJoe Perches pr_cont("\n"); 1735033e48fbSTejun Heo } 1736033e48fbSTejun Heo 1737fbf59bc9STejun Heo /** 17388d408b4bSTejun Heo * pcpu_setup_first_chunk - initialize the first percpu chunk 1739fd1e8a1fSTejun Heo * @ai: pcpu_alloc_info describing how to percpu area is shaped 174038a6be52STejun Heo * @base_addr: mapped address 1741fbf59bc9STejun Heo * 17428d408b4bSTejun Heo * Initialize the first percpu chunk which contains the kernel static 17438d408b4bSTejun Heo * perpcu area. This function is to be called from arch percpu area 174438a6be52STejun Heo * setup path. 17458d408b4bSTejun Heo * 1746fd1e8a1fSTejun Heo * @ai contains all information necessary to initialize the first 1747fd1e8a1fSTejun Heo * chunk and prime the dynamic percpu allocator. 17488d408b4bSTejun Heo * 1749fd1e8a1fSTejun Heo * @ai->static_size is the size of static percpu area. 1750fd1e8a1fSTejun Heo * 1751fd1e8a1fSTejun Heo * @ai->reserved_size, if non-zero, specifies the amount of bytes to 1752edcb4639STejun Heo * reserve after the static area in the first chunk. This reserves 1753edcb4639STejun Heo * the first chunk such that it's available only through reserved 1754edcb4639STejun Heo * percpu allocation. This is primarily used to serve module percpu 1755edcb4639STejun Heo * static areas on architectures where the addressing model has 1756edcb4639STejun Heo * limited offset range for symbol relocations to guarantee module 1757edcb4639STejun Heo * percpu symbols fall inside the relocatable range. 1758edcb4639STejun Heo * 1759fd1e8a1fSTejun Heo * @ai->dyn_size determines the number of bytes available for dynamic 1760fd1e8a1fSTejun Heo * allocation in the first chunk. The area between @ai->static_size + 1761fd1e8a1fSTejun Heo * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. 17626074d5b0STejun Heo * 1763fd1e8a1fSTejun Heo * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE 1764fd1e8a1fSTejun Heo * and equal to or larger than @ai->static_size + @ai->reserved_size + 1765fd1e8a1fSTejun Heo * @ai->dyn_size. 17668d408b4bSTejun Heo * 1767fd1e8a1fSTejun Heo * @ai->atom_size is the allocation atom size and used as alignment 1768fd1e8a1fSTejun Heo * for vm areas. 17698d408b4bSTejun Heo * 1770fd1e8a1fSTejun Heo * @ai->alloc_size is the allocation size and always multiple of 1771fd1e8a1fSTejun Heo * @ai->atom_size. This is larger than @ai->atom_size if 1772fd1e8a1fSTejun Heo * @ai->unit_size is larger than @ai->atom_size. 1773fd1e8a1fSTejun Heo * 1774fd1e8a1fSTejun Heo * @ai->nr_groups and @ai->groups describe virtual memory layout of 1775fd1e8a1fSTejun Heo * percpu areas. Units which should be colocated are put into the 1776fd1e8a1fSTejun Heo * same group. Dynamic VM areas will be allocated according to these 1777fd1e8a1fSTejun Heo * groupings. If @ai->nr_groups is zero, a single group containing 1778fd1e8a1fSTejun Heo * all units is assumed. 17798d408b4bSTejun Heo * 178038a6be52STejun Heo * The caller should have mapped the first chunk at @base_addr and 178138a6be52STejun Heo * copied static data to each unit. 1782fbf59bc9STejun Heo * 1783c0ebfdc3SDennis Zhou (Facebook) * The first chunk will always contain a static and a dynamic region. 1784c0ebfdc3SDennis Zhou (Facebook) * However, the static region is not managed by any chunk. If the first 1785c0ebfdc3SDennis Zhou (Facebook) * chunk also contains a reserved region, it is served by two chunks - 1786c0ebfdc3SDennis Zhou (Facebook) * one for the reserved region and one for the dynamic region. They 1787c0ebfdc3SDennis Zhou (Facebook) * share the same vm, but use offset regions in the area allocation map. 1788c0ebfdc3SDennis Zhou (Facebook) * The chunk serving the dynamic region is circulated in the chunk slots 1789c0ebfdc3SDennis Zhou (Facebook) * and available for dynamic allocation like any other chunk. 1790edcb4639STejun Heo * 1791fbf59bc9STejun Heo * RETURNS: 1792fb435d52STejun Heo * 0 on success, -errno on failure. 1793fbf59bc9STejun Heo */ 1794fb435d52STejun Heo int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, 1795fd1e8a1fSTejun Heo void *base_addr) 1796fbf59bc9STejun Heo { 1797b9c39442SDennis Zhou (Facebook) size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 1798d2f3c384SDennis Zhou (Facebook) size_t static_size, dyn_size; 17990c4169c3SDennis Zhou (Facebook) struct pcpu_chunk *chunk; 18006563297cSTejun Heo unsigned long *group_offsets; 18016563297cSTejun Heo size_t *group_sizes; 1802fb435d52STejun Heo unsigned long *unit_off; 1803fbf59bc9STejun Heo unsigned int cpu; 1804fd1e8a1fSTejun Heo int *unit_map; 1805fd1e8a1fSTejun Heo int group, unit, i; 1806c0ebfdc3SDennis Zhou (Facebook) int map_size; 1807c0ebfdc3SDennis Zhou (Facebook) unsigned long tmp_addr; 1808fbf59bc9STejun Heo 1809635b75fcSTejun Heo #define PCPU_SETUP_BUG_ON(cond) do { \ 1810635b75fcSTejun Heo if (unlikely(cond)) { \ 1811870d4b12SJoe Perches pr_emerg("failed to initialize, %s\n", #cond); \ 1812870d4b12SJoe Perches pr_emerg("cpu_possible_mask=%*pb\n", \ 1813807de073STejun Heo cpumask_pr_args(cpu_possible_mask)); \ 1814635b75fcSTejun Heo pcpu_dump_alloc_info(KERN_EMERG, ai); \ 1815635b75fcSTejun Heo BUG(); \ 1816635b75fcSTejun Heo } \ 1817635b75fcSTejun Heo } while (0) 1818635b75fcSTejun Heo 18192f39e637STejun Heo /* sanity checks */ 1820635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); 1821bbddff05STejun Heo #ifdef CONFIG_SMP 1822635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!ai->static_size); 1823f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); 1824bbddff05STejun Heo #endif 1825635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!base_addr); 1826f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); 1827635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); 1828f09f1243SAlexander Kuleshov PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); 1829635b75fcSTejun Heo PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); 1830ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); 1831099a19d9STejun Heo PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); 1832fb29a2ccSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!ai->dyn_size); 1833d2f3c384SDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); 1834ca460b3cSDennis Zhou (Facebook) PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || 1835ca460b3cSDennis Zhou (Facebook) IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); 18369f645532STejun Heo PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); 18378d408b4bSTejun Heo 18386563297cSTejun Heo /* process group information and build config tables accordingly */ 1839999c17e3SSantosh Shilimkar group_offsets = memblock_virt_alloc(ai->nr_groups * 1840999c17e3SSantosh Shilimkar sizeof(group_offsets[0]), 0); 1841999c17e3SSantosh Shilimkar group_sizes = memblock_virt_alloc(ai->nr_groups * 1842999c17e3SSantosh Shilimkar sizeof(group_sizes[0]), 0); 1843999c17e3SSantosh Shilimkar unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); 1844999c17e3SSantosh Shilimkar unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); 18452f39e637STejun Heo 1846fd1e8a1fSTejun Heo for (cpu = 0; cpu < nr_cpu_ids; cpu++) 1847ffe0d5a5STejun Heo unit_map[cpu] = UINT_MAX; 1848a855b84cSTejun Heo 1849a855b84cSTejun Heo pcpu_low_unit_cpu = NR_CPUS; 1850a855b84cSTejun Heo pcpu_high_unit_cpu = NR_CPUS; 18512f39e637STejun Heo 1852fd1e8a1fSTejun Heo for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { 1853fd1e8a1fSTejun Heo const struct pcpu_group_info *gi = &ai->groups[group]; 18542f39e637STejun Heo 18556563297cSTejun Heo group_offsets[group] = gi->base_offset; 18566563297cSTejun Heo group_sizes[group] = gi->nr_units * ai->unit_size; 18576563297cSTejun Heo 1858fd1e8a1fSTejun Heo for (i = 0; i < gi->nr_units; i++) { 1859fd1e8a1fSTejun Heo cpu = gi->cpu_map[i]; 1860fd1e8a1fSTejun Heo if (cpu == NR_CPUS) 1861fd1e8a1fSTejun Heo continue; 1862fd1e8a1fSTejun Heo 18639f295664SDan Carpenter PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); 1864635b75fcSTejun Heo PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); 1865635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); 1866fd1e8a1fSTejun Heo 1867fd1e8a1fSTejun Heo unit_map[cpu] = unit + i; 1868fb435d52STejun Heo unit_off[cpu] = gi->base_offset + i * ai->unit_size; 1869fb435d52STejun Heo 1870a855b84cSTejun Heo /* determine low/high unit_cpu */ 1871a855b84cSTejun Heo if (pcpu_low_unit_cpu == NR_CPUS || 1872a855b84cSTejun Heo unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) 1873a855b84cSTejun Heo pcpu_low_unit_cpu = cpu; 1874a855b84cSTejun Heo if (pcpu_high_unit_cpu == NR_CPUS || 1875a855b84cSTejun Heo unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) 1876a855b84cSTejun Heo pcpu_high_unit_cpu = cpu; 18770fc0531eSLinus Torvalds } 18780fc0531eSLinus Torvalds } 1879fd1e8a1fSTejun Heo pcpu_nr_units = unit; 18802f39e637STejun Heo 18812f39e637STejun Heo for_each_possible_cpu(cpu) 1882635b75fcSTejun Heo PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); 1883635b75fcSTejun Heo 1884635b75fcSTejun Heo /* we're done parsing the input, undefine BUG macro and dump config */ 1885635b75fcSTejun Heo #undef PCPU_SETUP_BUG_ON 1886bcbea798STejun Heo pcpu_dump_alloc_info(KERN_DEBUG, ai); 18872f39e637STejun Heo 18886563297cSTejun Heo pcpu_nr_groups = ai->nr_groups; 18896563297cSTejun Heo pcpu_group_offsets = group_offsets; 18906563297cSTejun Heo pcpu_group_sizes = group_sizes; 1891fd1e8a1fSTejun Heo pcpu_unit_map = unit_map; 1892fb435d52STejun Heo pcpu_unit_offsets = unit_off; 18932f39e637STejun Heo 18942f39e637STejun Heo /* determine basic parameters */ 1895fd1e8a1fSTejun Heo pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; 1896d9b55eebSTejun Heo pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 18976563297cSTejun Heo pcpu_atom_size = ai->atom_size; 1898ce3141a2STejun Heo pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + 1899ce3141a2STejun Heo BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); 1900cafe8816STejun Heo 190130a5b536SDennis Zhou pcpu_stats_save_ai(ai); 190230a5b536SDennis Zhou 1903d9b55eebSTejun Heo /* 1904d9b55eebSTejun Heo * Allocate chunk slots. The additional last slot is for 1905d9b55eebSTejun Heo * empty chunks. 1906d9b55eebSTejun Heo */ 1907d9b55eebSTejun Heo pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; 1908999c17e3SSantosh Shilimkar pcpu_slot = memblock_virt_alloc( 1909999c17e3SSantosh Shilimkar pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); 1910fbf59bc9STejun Heo for (i = 0; i < pcpu_nr_slots; i++) 1911fbf59bc9STejun Heo INIT_LIST_HEAD(&pcpu_slot[i]); 1912fbf59bc9STejun Heo 1913edcb4639STejun Heo /* 1914d2f3c384SDennis Zhou (Facebook) * The end of the static region needs to be aligned with the 1915d2f3c384SDennis Zhou (Facebook) * minimum allocation size as this offsets the reserved and 1916d2f3c384SDennis Zhou (Facebook) * dynamic region. The first chunk ends page aligned by 1917d2f3c384SDennis Zhou (Facebook) * expanding the dynamic region, therefore the dynamic region 1918d2f3c384SDennis Zhou (Facebook) * can be shrunk to compensate while still staying above the 1919d2f3c384SDennis Zhou (Facebook) * configured sizes. 1920d2f3c384SDennis Zhou (Facebook) */ 1921d2f3c384SDennis Zhou (Facebook) static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); 1922d2f3c384SDennis Zhou (Facebook) dyn_size = ai->dyn_size - (static_size - ai->static_size); 1923d2f3c384SDennis Zhou (Facebook) 1924d2f3c384SDennis Zhou (Facebook) /* 1925c0ebfdc3SDennis Zhou (Facebook) * Initialize first chunk. 1926c0ebfdc3SDennis Zhou (Facebook) * If the reserved_size is non-zero, this initializes the reserved 1927c0ebfdc3SDennis Zhou (Facebook) * chunk. If the reserved_size is zero, the reserved chunk is NULL 1928c0ebfdc3SDennis Zhou (Facebook) * and the dynamic region is initialized here. The first chunk, 1929c0ebfdc3SDennis Zhou (Facebook) * pcpu_first_chunk, will always point to the chunk that serves 1930c0ebfdc3SDennis Zhou (Facebook) * the dynamic region. 1931edcb4639STejun Heo */ 1932d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size; 1933d2f3c384SDennis Zhou (Facebook) map_size = ai->reserved_size ?: dyn_size; 193440064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 193561ace7faSTejun Heo 1936edcb4639STejun Heo /* init dynamic chunk if necessary */ 1937b9c39442SDennis Zhou (Facebook) if (ai->reserved_size) { 19380c4169c3SDennis Zhou (Facebook) pcpu_reserved_chunk = chunk; 1939b9c39442SDennis Zhou (Facebook) 1940d2f3c384SDennis Zhou (Facebook) tmp_addr = (unsigned long)base_addr + static_size + 1941c0ebfdc3SDennis Zhou (Facebook) ai->reserved_size; 1942d2f3c384SDennis Zhou (Facebook) map_size = dyn_size; 194340064aecSDennis Zhou (Facebook) chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); 1944edcb4639STejun Heo } 1945edcb4639STejun Heo 19462441d15cSTejun Heo /* link the first chunk in */ 19470c4169c3SDennis Zhou (Facebook) pcpu_first_chunk = chunk; 19480cecf50cSDennis Zhou (Facebook) pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; 1949ae9e6bc9STejun Heo pcpu_chunk_relocate(pcpu_first_chunk, -1); 1950fbf59bc9STejun Heo 195130a5b536SDennis Zhou pcpu_stats_chunk_alloc(); 1952df95e795SDennis Zhou trace_percpu_create_chunk(base_addr); 195330a5b536SDennis Zhou 1954fbf59bc9STejun Heo /* we're done */ 1955bba174f5STejun Heo pcpu_base_addr = base_addr; 1956fb435d52STejun Heo return 0; 1957fbf59bc9STejun Heo } 195866c3a757STejun Heo 1959bbddff05STejun Heo #ifdef CONFIG_SMP 1960bbddff05STejun Heo 196117f3609cSAndi Kleen const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { 1962f58dc01bSTejun Heo [PCPU_FC_AUTO] = "auto", 1963f58dc01bSTejun Heo [PCPU_FC_EMBED] = "embed", 1964f58dc01bSTejun Heo [PCPU_FC_PAGE] = "page", 1965f58dc01bSTejun Heo }; 196666c3a757STejun Heo 1967f58dc01bSTejun Heo enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; 1968f58dc01bSTejun Heo 1969f58dc01bSTejun Heo static int __init percpu_alloc_setup(char *str) 197066c3a757STejun Heo { 19715479c78aSCyrill Gorcunov if (!str) 19725479c78aSCyrill Gorcunov return -EINVAL; 19735479c78aSCyrill Gorcunov 1974f58dc01bSTejun Heo if (0) 1975f58dc01bSTejun Heo /* nada */; 1976f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK 1977f58dc01bSTejun Heo else if (!strcmp(str, "embed")) 1978f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_EMBED; 1979f58dc01bSTejun Heo #endif 1980f58dc01bSTejun Heo #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 1981f58dc01bSTejun Heo else if (!strcmp(str, "page")) 1982f58dc01bSTejun Heo pcpu_chosen_fc = PCPU_FC_PAGE; 1983f58dc01bSTejun Heo #endif 1984f58dc01bSTejun Heo else 1985870d4b12SJoe Perches pr_warn("unknown allocator %s specified\n", str); 198666c3a757STejun Heo 1987f58dc01bSTejun Heo return 0; 198866c3a757STejun Heo } 1989f58dc01bSTejun Heo early_param("percpu_alloc", percpu_alloc_setup); 199066c3a757STejun Heo 19913c9a024fSTejun Heo /* 19923c9a024fSTejun Heo * pcpu_embed_first_chunk() is used by the generic percpu setup. 19933c9a024fSTejun Heo * Build it if needed by the arch config or the generic setup is going 19943c9a024fSTejun Heo * to be used. 19953c9a024fSTejun Heo */ 199608fc4580STejun Heo #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ 199708fc4580STejun Heo !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) 19983c9a024fSTejun Heo #define BUILD_EMBED_FIRST_CHUNK 19993c9a024fSTejun Heo #endif 20003c9a024fSTejun Heo 20013c9a024fSTejun Heo /* build pcpu_page_first_chunk() iff needed by the arch config */ 20023c9a024fSTejun Heo #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) 20033c9a024fSTejun Heo #define BUILD_PAGE_FIRST_CHUNK 20043c9a024fSTejun Heo #endif 20053c9a024fSTejun Heo 20063c9a024fSTejun Heo /* pcpu_build_alloc_info() is used by both embed and page first chunk */ 20073c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) 20083c9a024fSTejun Heo /** 2009fbf59bc9STejun Heo * pcpu_build_alloc_info - build alloc_info considering distances between CPUs 2010fbf59bc9STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2011fbf59bc9STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2012fbf59bc9STejun Heo * @atom_size: allocation atom size 2013fbf59bc9STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2014fbf59bc9STejun Heo * 2015fbf59bc9STejun Heo * This function determines grouping of units, their mappings to cpus 2016fbf59bc9STejun Heo * and other parameters considering needed percpu size, allocation 2017fbf59bc9STejun Heo * atom size and distances between CPUs. 2018fbf59bc9STejun Heo * 2019bffc4375SYannick Guerrini * Groups are always multiples of atom size and CPUs which are of 2020fbf59bc9STejun Heo * LOCAL_DISTANCE both ways are grouped together and share space for 2021fbf59bc9STejun Heo * units in the same group. The returned configuration is guaranteed 2022fbf59bc9STejun Heo * to have CPUs on different nodes on different groups and >=75% usage 2023fbf59bc9STejun Heo * of allocated virtual address space. 2024fbf59bc9STejun Heo * 2025fbf59bc9STejun Heo * RETURNS: 2026fbf59bc9STejun Heo * On success, pointer to the new allocation_info is returned. On 2027fbf59bc9STejun Heo * failure, ERR_PTR value is returned. 2028fbf59bc9STejun Heo */ 2029fbf59bc9STejun Heo static struct pcpu_alloc_info * __init pcpu_build_alloc_info( 2030fbf59bc9STejun Heo size_t reserved_size, size_t dyn_size, 2031fbf59bc9STejun Heo size_t atom_size, 2032fbf59bc9STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn) 2033fbf59bc9STejun Heo { 2034fbf59bc9STejun Heo static int group_map[NR_CPUS] __initdata; 2035fbf59bc9STejun Heo static int group_cnt[NR_CPUS] __initdata; 2036fbf59bc9STejun Heo const size_t static_size = __per_cpu_end - __per_cpu_start; 2037fbf59bc9STejun Heo int nr_groups = 1, nr_units = 0; 2038fbf59bc9STejun Heo size_t size_sum, min_unit_size, alloc_size; 2039fbf59bc9STejun Heo int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ 2040fbf59bc9STejun Heo int last_allocs, group, unit; 2041fbf59bc9STejun Heo unsigned int cpu, tcpu; 2042fbf59bc9STejun Heo struct pcpu_alloc_info *ai; 2043fbf59bc9STejun Heo unsigned int *cpu_map; 2044fbf59bc9STejun Heo 2045fbf59bc9STejun Heo /* this function may be called multiple times */ 2046fbf59bc9STejun Heo memset(group_map, 0, sizeof(group_map)); 2047fbf59bc9STejun Heo memset(group_cnt, 0, sizeof(group_cnt)); 2048fbf59bc9STejun Heo 2049fbf59bc9STejun Heo /* calculate size_sum and ensure dyn_size is enough for early alloc */ 2050fbf59bc9STejun Heo size_sum = PFN_ALIGN(static_size + reserved_size + 2051fbf59bc9STejun Heo max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); 2052fbf59bc9STejun Heo dyn_size = size_sum - static_size - reserved_size; 2053fbf59bc9STejun Heo 2054fbf59bc9STejun Heo /* 2055fbf59bc9STejun Heo * Determine min_unit_size, alloc_size and max_upa such that 2056fbf59bc9STejun Heo * alloc_size is multiple of atom_size and is the smallest 205725985edcSLucas De Marchi * which can accommodate 4k aligned segments which are equal to 2058fbf59bc9STejun Heo * or larger than min_unit_size. 2059fbf59bc9STejun Heo */ 2060fbf59bc9STejun Heo min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); 2061fbf59bc9STejun Heo 20629c015162SDennis Zhou (Facebook) /* determine the maximum # of units that can fit in an allocation */ 2063fbf59bc9STejun Heo alloc_size = roundup(min_unit_size, atom_size); 2064fbf59bc9STejun Heo upa = alloc_size / min_unit_size; 2065f09f1243SAlexander Kuleshov while (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2066fbf59bc9STejun Heo upa--; 2067fbf59bc9STejun Heo max_upa = upa; 2068fbf59bc9STejun Heo 2069fbf59bc9STejun Heo /* group cpus according to their proximity */ 2070fbf59bc9STejun Heo for_each_possible_cpu(cpu) { 2071fbf59bc9STejun Heo group = 0; 2072fbf59bc9STejun Heo next_group: 2073fbf59bc9STejun Heo for_each_possible_cpu(tcpu) { 2074fbf59bc9STejun Heo if (cpu == tcpu) 2075fbf59bc9STejun Heo break; 2076fbf59bc9STejun Heo if (group_map[tcpu] == group && cpu_distance_fn && 2077fbf59bc9STejun Heo (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || 2078fbf59bc9STejun Heo cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { 2079fbf59bc9STejun Heo group++; 2080fbf59bc9STejun Heo nr_groups = max(nr_groups, group + 1); 2081fbf59bc9STejun Heo goto next_group; 2082fbf59bc9STejun Heo } 2083fbf59bc9STejun Heo } 2084fbf59bc9STejun Heo group_map[cpu] = group; 2085fbf59bc9STejun Heo group_cnt[group]++; 2086fbf59bc9STejun Heo } 2087fbf59bc9STejun Heo 2088fbf59bc9STejun Heo /* 20899c015162SDennis Zhou (Facebook) * Wasted space is caused by a ratio imbalance of upa to group_cnt. 20909c015162SDennis Zhou (Facebook) * Expand the unit_size until we use >= 75% of the units allocated. 20919c015162SDennis Zhou (Facebook) * Related to atom_size, which could be much larger than the unit_size. 2092fbf59bc9STejun Heo */ 2093fbf59bc9STejun Heo last_allocs = INT_MAX; 2094fbf59bc9STejun Heo for (upa = max_upa; upa; upa--) { 2095fbf59bc9STejun Heo int allocs = 0, wasted = 0; 2096fbf59bc9STejun Heo 2097f09f1243SAlexander Kuleshov if (alloc_size % upa || (offset_in_page(alloc_size / upa))) 2098fbf59bc9STejun Heo continue; 2099fbf59bc9STejun Heo 2100fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2101fbf59bc9STejun Heo int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); 2102fbf59bc9STejun Heo allocs += this_allocs; 2103fbf59bc9STejun Heo wasted += this_allocs * upa - group_cnt[group]; 2104fbf59bc9STejun Heo } 2105fbf59bc9STejun Heo 2106fbf59bc9STejun Heo /* 2107fbf59bc9STejun Heo * Don't accept if wastage is over 1/3. The 2108fbf59bc9STejun Heo * greater-than comparison ensures upa==1 always 2109fbf59bc9STejun Heo * passes the following check. 2110fbf59bc9STejun Heo */ 2111fbf59bc9STejun Heo if (wasted > num_possible_cpus() / 3) 2112fbf59bc9STejun Heo continue; 2113fbf59bc9STejun Heo 2114fbf59bc9STejun Heo /* and then don't consume more memory */ 2115fbf59bc9STejun Heo if (allocs > last_allocs) 2116fbf59bc9STejun Heo break; 2117fbf59bc9STejun Heo last_allocs = allocs; 2118fbf59bc9STejun Heo best_upa = upa; 2119fbf59bc9STejun Heo } 2120fbf59bc9STejun Heo upa = best_upa; 2121fbf59bc9STejun Heo 2122fbf59bc9STejun Heo /* allocate and fill alloc_info */ 2123fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) 2124fbf59bc9STejun Heo nr_units += roundup(group_cnt[group], upa); 2125fbf59bc9STejun Heo 2126fbf59bc9STejun Heo ai = pcpu_alloc_alloc_info(nr_groups, nr_units); 2127fbf59bc9STejun Heo if (!ai) 2128fbf59bc9STejun Heo return ERR_PTR(-ENOMEM); 2129fbf59bc9STejun Heo cpu_map = ai->groups[0].cpu_map; 2130fbf59bc9STejun Heo 2131fbf59bc9STejun Heo for (group = 0; group < nr_groups; group++) { 2132fbf59bc9STejun Heo ai->groups[group].cpu_map = cpu_map; 2133fbf59bc9STejun Heo cpu_map += roundup(group_cnt[group], upa); 2134fbf59bc9STejun Heo } 2135fbf59bc9STejun Heo 2136fbf59bc9STejun Heo ai->static_size = static_size; 2137fbf59bc9STejun Heo ai->reserved_size = reserved_size; 2138fbf59bc9STejun Heo ai->dyn_size = dyn_size; 2139fbf59bc9STejun Heo ai->unit_size = alloc_size / upa; 2140fbf59bc9STejun Heo ai->atom_size = atom_size; 2141fbf59bc9STejun Heo ai->alloc_size = alloc_size; 2142fbf59bc9STejun Heo 2143fbf59bc9STejun Heo for (group = 0, unit = 0; group_cnt[group]; group++) { 2144fbf59bc9STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2145fbf59bc9STejun Heo 2146fbf59bc9STejun Heo /* 2147fbf59bc9STejun Heo * Initialize base_offset as if all groups are located 2148fbf59bc9STejun Heo * back-to-back. The caller should update this to 2149fbf59bc9STejun Heo * reflect actual allocation. 2150fbf59bc9STejun Heo */ 2151fbf59bc9STejun Heo gi->base_offset = unit * ai->unit_size; 2152fbf59bc9STejun Heo 2153fbf59bc9STejun Heo for_each_possible_cpu(cpu) 2154fbf59bc9STejun Heo if (group_map[cpu] == group) 2155fbf59bc9STejun Heo gi->cpu_map[gi->nr_units++] = cpu; 2156fbf59bc9STejun Heo gi->nr_units = roundup(gi->nr_units, upa); 2157fbf59bc9STejun Heo unit += gi->nr_units; 2158fbf59bc9STejun Heo } 2159fbf59bc9STejun Heo BUG_ON(unit != nr_units); 2160fbf59bc9STejun Heo 2161fbf59bc9STejun Heo return ai; 2162fbf59bc9STejun Heo } 21633c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ 2164fbf59bc9STejun Heo 21653c9a024fSTejun Heo #if defined(BUILD_EMBED_FIRST_CHUNK) 216666c3a757STejun Heo /** 216766c3a757STejun Heo * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 216866c3a757STejun Heo * @reserved_size: the size of reserved percpu area in bytes 21694ba6ce25STejun Heo * @dyn_size: minimum free size for dynamic allocation in bytes 2170c8826dd5STejun Heo * @atom_size: allocation atom size 2171c8826dd5STejun Heo * @cpu_distance_fn: callback to determine distance between cpus, optional 2172c8826dd5STejun Heo * @alloc_fn: function to allocate percpu page 217325985edcSLucas De Marchi * @free_fn: function to free percpu page 217466c3a757STejun Heo * 217566c3a757STejun Heo * This is a helper to ease setting up embedded first percpu chunk and 217666c3a757STejun Heo * can be called where pcpu_setup_first_chunk() is expected. 217766c3a757STejun Heo * 217866c3a757STejun Heo * If this function is used to setup the first chunk, it is allocated 2179c8826dd5STejun Heo * by calling @alloc_fn and used as-is without being mapped into 2180c8826dd5STejun Heo * vmalloc area. Allocations are always whole multiples of @atom_size 2181c8826dd5STejun Heo * aligned to @atom_size. 2182c8826dd5STejun Heo * 2183c8826dd5STejun Heo * This enables the first chunk to piggy back on the linear physical 2184c8826dd5STejun Heo * mapping which often uses larger page size. Please note that this 2185c8826dd5STejun Heo * can result in very sparse cpu->unit mapping on NUMA machines thus 2186c8826dd5STejun Heo * requiring large vmalloc address space. Don't use this allocator if 2187c8826dd5STejun Heo * vmalloc space is not orders of magnitude larger than distances 2188c8826dd5STejun Heo * between node memory addresses (ie. 32bit NUMA machines). 218966c3a757STejun Heo * 21904ba6ce25STejun Heo * @dyn_size specifies the minimum dynamic area size. 219166c3a757STejun Heo * 219266c3a757STejun Heo * If the needed size is smaller than the minimum or specified unit 2193c8826dd5STejun Heo * size, the leftover is returned using @free_fn. 219466c3a757STejun Heo * 219566c3a757STejun Heo * RETURNS: 2196fb435d52STejun Heo * 0 on success, -errno on failure. 219766c3a757STejun Heo */ 21984ba6ce25STejun Heo int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, 2199c8826dd5STejun Heo size_t atom_size, 2200c8826dd5STejun Heo pcpu_fc_cpu_distance_fn_t cpu_distance_fn, 2201c8826dd5STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2202c8826dd5STejun Heo pcpu_fc_free_fn_t free_fn) 220366c3a757STejun Heo { 2204c8826dd5STejun Heo void *base = (void *)ULONG_MAX; 2205c8826dd5STejun Heo void **areas = NULL; 2206fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 220793c76b6bSzijun_hu size_t size_sum, areas_size; 220893c76b6bSzijun_hu unsigned long max_distance; 22099b739662Szijun_hu int group, i, highest_group, rc; 221066c3a757STejun Heo 2211c8826dd5STejun Heo ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, 2212c8826dd5STejun Heo cpu_distance_fn); 2213fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2214fd1e8a1fSTejun Heo return PTR_ERR(ai); 221566c3a757STejun Heo 2216fd1e8a1fSTejun Heo size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; 2217c8826dd5STejun Heo areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); 221866c3a757STejun Heo 2219999c17e3SSantosh Shilimkar areas = memblock_virt_alloc_nopanic(areas_size, 0); 2220c8826dd5STejun Heo if (!areas) { 2221fb435d52STejun Heo rc = -ENOMEM; 2222c8826dd5STejun Heo goto out_free; 2223fa8a7094STejun Heo } 222466c3a757STejun Heo 22259b739662Szijun_hu /* allocate, copy and determine base address & max_distance */ 22269b739662Szijun_hu highest_group = 0; 2227c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2228c8826dd5STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 2229c8826dd5STejun Heo unsigned int cpu = NR_CPUS; 2230c8826dd5STejun Heo void *ptr; 223166c3a757STejun Heo 2232c8826dd5STejun Heo for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) 2233c8826dd5STejun Heo cpu = gi->cpu_map[i]; 2234c8826dd5STejun Heo BUG_ON(cpu == NR_CPUS); 2235c8826dd5STejun Heo 2236c8826dd5STejun Heo /* allocate space for the whole group */ 2237c8826dd5STejun Heo ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); 2238c8826dd5STejun Heo if (!ptr) { 2239c8826dd5STejun Heo rc = -ENOMEM; 2240c8826dd5STejun Heo goto out_free_areas; 2241c8826dd5STejun Heo } 2242f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2243f528f0b8SCatalin Marinas kmemleak_free(ptr); 2244c8826dd5STejun Heo areas[group] = ptr; 2245c8826dd5STejun Heo 2246c8826dd5STejun Heo base = min(ptr, base); 22479b739662Szijun_hu if (ptr > areas[highest_group]) 22489b739662Szijun_hu highest_group = group; 22499b739662Szijun_hu } 22509b739662Szijun_hu max_distance = areas[highest_group] - base; 22519b739662Szijun_hu max_distance += ai->unit_size * ai->groups[highest_group].nr_units; 22529b739662Szijun_hu 22539b739662Szijun_hu /* warn if maximum distance is further than 75% of vmalloc space */ 22549b739662Szijun_hu if (max_distance > VMALLOC_TOTAL * 3 / 4) { 22559b739662Szijun_hu pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", 22569b739662Szijun_hu max_distance, VMALLOC_TOTAL); 22579b739662Szijun_hu #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK 22589b739662Szijun_hu /* and fail if we have fallback */ 22599b739662Szijun_hu rc = -EINVAL; 22609b739662Szijun_hu goto out_free_areas; 22619b739662Szijun_hu #endif 226242b64281STejun Heo } 226342b64281STejun Heo 226442b64281STejun Heo /* 226542b64281STejun Heo * Copy data and free unused parts. This should happen after all 226642b64281STejun Heo * allocations are complete; otherwise, we may end up with 226742b64281STejun Heo * overlapping groups. 226842b64281STejun Heo */ 226942b64281STejun Heo for (group = 0; group < ai->nr_groups; group++) { 227042b64281STejun Heo struct pcpu_group_info *gi = &ai->groups[group]; 227142b64281STejun Heo void *ptr = areas[group]; 2272c8826dd5STejun Heo 2273c8826dd5STejun Heo for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { 2274c8826dd5STejun Heo if (gi->cpu_map[i] == NR_CPUS) { 2275c8826dd5STejun Heo /* unused unit, free whole */ 2276c8826dd5STejun Heo free_fn(ptr, ai->unit_size); 2277c8826dd5STejun Heo continue; 2278c8826dd5STejun Heo } 2279c8826dd5STejun Heo /* copy and return the unused part */ 2280fd1e8a1fSTejun Heo memcpy(ptr, __per_cpu_load, ai->static_size); 2281c8826dd5STejun Heo free_fn(ptr + size_sum, ai->unit_size - size_sum); 2282c8826dd5STejun Heo } 228366c3a757STejun Heo } 228466c3a757STejun Heo 2285c8826dd5STejun Heo /* base address is now known, determine group base offsets */ 22866ea529a2STejun Heo for (group = 0; group < ai->nr_groups; group++) { 2287c8826dd5STejun Heo ai->groups[group].base_offset = areas[group] - base; 22886ea529a2STejun Heo } 2289c8826dd5STejun Heo 2290870d4b12SJoe Perches pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", 2291fd1e8a1fSTejun Heo PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, 2292fd1e8a1fSTejun Heo ai->dyn_size, ai->unit_size); 229366c3a757STejun Heo 2294fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, base); 2295c8826dd5STejun Heo goto out_free; 2296c8826dd5STejun Heo 2297c8826dd5STejun Heo out_free_areas: 2298c8826dd5STejun Heo for (group = 0; group < ai->nr_groups; group++) 2299f851c8d8SMichael Holzheu if (areas[group]) 2300c8826dd5STejun Heo free_fn(areas[group], 2301c8826dd5STejun Heo ai->groups[group].nr_units * ai->unit_size); 2302c8826dd5STejun Heo out_free: 2303fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2304c8826dd5STejun Heo if (areas) 2305999c17e3SSantosh Shilimkar memblock_free_early(__pa(areas), areas_size); 2306fb435d52STejun Heo return rc; 2307d4b95f80STejun Heo } 23083c9a024fSTejun Heo #endif /* BUILD_EMBED_FIRST_CHUNK */ 2309d4b95f80STejun Heo 23103c9a024fSTejun Heo #ifdef BUILD_PAGE_FIRST_CHUNK 2311d4b95f80STejun Heo /** 231200ae4064STejun Heo * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages 2313d4b95f80STejun Heo * @reserved_size: the size of reserved percpu area in bytes 2314d4b95f80STejun Heo * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE 231525985edcSLucas De Marchi * @free_fn: function to free percpu page, always called with PAGE_SIZE 2316d4b95f80STejun Heo * @populate_pte_fn: function to populate pte 2317d4b95f80STejun Heo * 231800ae4064STejun Heo * This is a helper to ease setting up page-remapped first percpu 231900ae4064STejun Heo * chunk and can be called where pcpu_setup_first_chunk() is expected. 2320d4b95f80STejun Heo * 2321d4b95f80STejun Heo * This is the basic allocator. Static percpu area is allocated 2322d4b95f80STejun Heo * page-by-page into vmalloc area. 2323d4b95f80STejun Heo * 2324d4b95f80STejun Heo * RETURNS: 2325fb435d52STejun Heo * 0 on success, -errno on failure. 2326d4b95f80STejun Heo */ 2327fb435d52STejun Heo int __init pcpu_page_first_chunk(size_t reserved_size, 2328d4b95f80STejun Heo pcpu_fc_alloc_fn_t alloc_fn, 2329d4b95f80STejun Heo pcpu_fc_free_fn_t free_fn, 2330d4b95f80STejun Heo pcpu_fc_populate_pte_fn_t populate_pte_fn) 2331d4b95f80STejun Heo { 23328f05a6a6STejun Heo static struct vm_struct vm; 2333fd1e8a1fSTejun Heo struct pcpu_alloc_info *ai; 233400ae4064STejun Heo char psize_str[16]; 2335ce3141a2STejun Heo int unit_pages; 2336d4b95f80STejun Heo size_t pages_size; 2337ce3141a2STejun Heo struct page **pages; 2338fb435d52STejun Heo int unit, i, j, rc; 23398f606604Szijun_hu int upa; 23408f606604Szijun_hu int nr_g0_units; 2341d4b95f80STejun Heo 234200ae4064STejun Heo snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); 234300ae4064STejun Heo 23444ba6ce25STejun Heo ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); 2345fd1e8a1fSTejun Heo if (IS_ERR(ai)) 2346fd1e8a1fSTejun Heo return PTR_ERR(ai); 2347fd1e8a1fSTejun Heo BUG_ON(ai->nr_groups != 1); 23488f606604Szijun_hu upa = ai->alloc_size/ai->unit_size; 23498f606604Szijun_hu nr_g0_units = roundup(num_possible_cpus(), upa); 23508f606604Szijun_hu if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { 23518f606604Szijun_hu pcpu_free_alloc_info(ai); 23528f606604Szijun_hu return -EINVAL; 23538f606604Szijun_hu } 2354fd1e8a1fSTejun Heo 2355fd1e8a1fSTejun Heo unit_pages = ai->unit_size >> PAGE_SHIFT; 2356d4b95f80STejun Heo 2357d4b95f80STejun Heo /* unaligned allocations can't be freed, round up to page size */ 2358fd1e8a1fSTejun Heo pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * 2359fd1e8a1fSTejun Heo sizeof(pages[0])); 2360999c17e3SSantosh Shilimkar pages = memblock_virt_alloc(pages_size, 0); 2361d4b95f80STejun Heo 23628f05a6a6STejun Heo /* allocate pages */ 2363d4b95f80STejun Heo j = 0; 23648f606604Szijun_hu for (unit = 0; unit < num_possible_cpus(); unit++) { 2365fd1e8a1fSTejun Heo unsigned int cpu = ai->groups[0].cpu_map[unit]; 23668f606604Szijun_hu for (i = 0; i < unit_pages; i++) { 2367d4b95f80STejun Heo void *ptr; 2368d4b95f80STejun Heo 23693cbc8565STejun Heo ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); 2370d4b95f80STejun Heo if (!ptr) { 2371870d4b12SJoe Perches pr_warn("failed to allocate %s page for cpu%u\n", 2372598d8091SJoe Perches psize_str, cpu); 2373d4b95f80STejun Heo goto enomem; 2374d4b95f80STejun Heo } 2375f528f0b8SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2376f528f0b8SCatalin Marinas kmemleak_free(ptr); 2377ce3141a2STejun Heo pages[j++] = virt_to_page(ptr); 2378d4b95f80STejun Heo } 23798f606604Szijun_hu } 2380d4b95f80STejun Heo 23818f05a6a6STejun Heo /* allocate vm area, map the pages and copy static data */ 23828f05a6a6STejun Heo vm.flags = VM_ALLOC; 2383fd1e8a1fSTejun Heo vm.size = num_possible_cpus() * ai->unit_size; 23848f05a6a6STejun Heo vm_area_register_early(&vm, PAGE_SIZE); 23858f05a6a6STejun Heo 2386fd1e8a1fSTejun Heo for (unit = 0; unit < num_possible_cpus(); unit++) { 23871d9d3257STejun Heo unsigned long unit_addr = 2388fd1e8a1fSTejun Heo (unsigned long)vm.addr + unit * ai->unit_size; 23898f05a6a6STejun Heo 2390ce3141a2STejun Heo for (i = 0; i < unit_pages; i++) 23918f05a6a6STejun Heo populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); 23928f05a6a6STejun Heo 23938f05a6a6STejun Heo /* pte already populated, the following shouldn't fail */ 2394fb435d52STejun Heo rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], 2395ce3141a2STejun Heo unit_pages); 2396fb435d52STejun Heo if (rc < 0) 2397fb435d52STejun Heo panic("failed to map percpu area, err=%d\n", rc); 23988f05a6a6STejun Heo 23998f05a6a6STejun Heo /* 24008f05a6a6STejun Heo * FIXME: Archs with virtual cache should flush local 24018f05a6a6STejun Heo * cache for the linear mapping here - something 24028f05a6a6STejun Heo * equivalent to flush_cache_vmap() on the local cpu. 24038f05a6a6STejun Heo * flush_cache_vmap() can't be used as most supporting 24048f05a6a6STejun Heo * data structures are not set up yet. 24058f05a6a6STejun Heo */ 24068f05a6a6STejun Heo 24078f05a6a6STejun Heo /* copy static data */ 2408fd1e8a1fSTejun Heo memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); 240966c3a757STejun Heo } 241066c3a757STejun Heo 241166c3a757STejun Heo /* we're ready, commit */ 2412870d4b12SJoe Perches pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", 2413fd1e8a1fSTejun Heo unit_pages, psize_str, vm.addr, ai->static_size, 2414fd1e8a1fSTejun Heo ai->reserved_size, ai->dyn_size); 241566c3a757STejun Heo 2416fb435d52STejun Heo rc = pcpu_setup_first_chunk(ai, vm.addr); 2417d4b95f80STejun Heo goto out_free_ar; 2418d4b95f80STejun Heo 2419d4b95f80STejun Heo enomem: 2420d4b95f80STejun Heo while (--j >= 0) 2421ce3141a2STejun Heo free_fn(page_address(pages[j]), PAGE_SIZE); 2422fb435d52STejun Heo rc = -ENOMEM; 2423d4b95f80STejun Heo out_free_ar: 2424999c17e3SSantosh Shilimkar memblock_free_early(__pa(pages), pages_size); 2425fd1e8a1fSTejun Heo pcpu_free_alloc_info(ai); 2426fb435d52STejun Heo return rc; 242766c3a757STejun Heo } 24283c9a024fSTejun Heo #endif /* BUILD_PAGE_FIRST_CHUNK */ 2429d4b95f80STejun Heo 2430bbddff05STejun Heo #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA 24318c4bfc6eSTejun Heo /* 2432bbddff05STejun Heo * Generic SMP percpu area setup. 2433e74e3962STejun Heo * 2434e74e3962STejun Heo * The embedding helper is used because its behavior closely resembles 2435e74e3962STejun Heo * the original non-dynamic generic percpu area setup. This is 2436e74e3962STejun Heo * important because many archs have addressing restrictions and might 2437e74e3962STejun Heo * fail if the percpu area is located far away from the previous 2438e74e3962STejun Heo * location. As an added bonus, in non-NUMA cases, embedding is 2439e74e3962STejun Heo * generally a good idea TLB-wise because percpu area can piggy back 2440e74e3962STejun Heo * on the physical linear memory mapping which uses large page 2441e74e3962STejun Heo * mappings on applicable archs. 2442e74e3962STejun Heo */ 2443e74e3962STejun Heo unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; 2444e74e3962STejun Heo EXPORT_SYMBOL(__per_cpu_offset); 2445e74e3962STejun Heo 2446c8826dd5STejun Heo static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, 2447c8826dd5STejun Heo size_t align) 2448c8826dd5STejun Heo { 2449999c17e3SSantosh Shilimkar return memblock_virt_alloc_from_nopanic( 2450999c17e3SSantosh Shilimkar size, align, __pa(MAX_DMA_ADDRESS)); 2451c8826dd5STejun Heo } 2452c8826dd5STejun Heo 2453c8826dd5STejun Heo static void __init pcpu_dfl_fc_free(void *ptr, size_t size) 2454c8826dd5STejun Heo { 2455999c17e3SSantosh Shilimkar memblock_free_early(__pa(ptr), size); 2456c8826dd5STejun Heo } 2457c8826dd5STejun Heo 2458e74e3962STejun Heo void __init setup_per_cpu_areas(void) 2459e74e3962STejun Heo { 2460e74e3962STejun Heo unsigned long delta; 2461e74e3962STejun Heo unsigned int cpu; 2462fb435d52STejun Heo int rc; 2463e74e3962STejun Heo 2464e74e3962STejun Heo /* 2465e74e3962STejun Heo * Always reserve area for module percpu variables. That's 2466e74e3962STejun Heo * what the legacy allocator did. 2467e74e3962STejun Heo */ 2468fb435d52STejun Heo rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, 2469c8826dd5STejun Heo PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, 2470c8826dd5STejun Heo pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); 2471fb435d52STejun Heo if (rc < 0) 2472bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2473e74e3962STejun Heo 2474e74e3962STejun Heo delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; 2475e74e3962STejun Heo for_each_possible_cpu(cpu) 2476fb435d52STejun Heo __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; 2477e74e3962STejun Heo } 2478e74e3962STejun Heo #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ 2479099a19d9STejun Heo 2480bbddff05STejun Heo #else /* CONFIG_SMP */ 2481bbddff05STejun Heo 2482bbddff05STejun Heo /* 2483bbddff05STejun Heo * UP percpu area setup. 2484bbddff05STejun Heo * 2485bbddff05STejun Heo * UP always uses km-based percpu allocator with identity mapping. 2486bbddff05STejun Heo * Static percpu variables are indistinguishable from the usual static 2487bbddff05STejun Heo * variables and don't require any special preparation. 2488bbddff05STejun Heo */ 2489bbddff05STejun Heo void __init setup_per_cpu_areas(void) 2490bbddff05STejun Heo { 2491bbddff05STejun Heo const size_t unit_size = 2492bbddff05STejun Heo roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, 2493bbddff05STejun Heo PERCPU_DYNAMIC_RESERVE)); 2494bbddff05STejun Heo struct pcpu_alloc_info *ai; 2495bbddff05STejun Heo void *fc; 2496bbddff05STejun Heo 2497bbddff05STejun Heo ai = pcpu_alloc_alloc_info(1, 1); 2498999c17e3SSantosh Shilimkar fc = memblock_virt_alloc_from_nopanic(unit_size, 2499999c17e3SSantosh Shilimkar PAGE_SIZE, 2500999c17e3SSantosh Shilimkar __pa(MAX_DMA_ADDRESS)); 2501bbddff05STejun Heo if (!ai || !fc) 2502bbddff05STejun Heo panic("Failed to allocate memory for percpu areas."); 2503100d13c3SCatalin Marinas /* kmemleak tracks the percpu allocations separately */ 2504100d13c3SCatalin Marinas kmemleak_free(fc); 2505bbddff05STejun Heo 2506bbddff05STejun Heo ai->dyn_size = unit_size; 2507bbddff05STejun Heo ai->unit_size = unit_size; 2508bbddff05STejun Heo ai->atom_size = unit_size; 2509bbddff05STejun Heo ai->alloc_size = unit_size; 2510bbddff05STejun Heo ai->groups[0].nr_units = 1; 2511bbddff05STejun Heo ai->groups[0].cpu_map[0] = 0; 2512bbddff05STejun Heo 2513bbddff05STejun Heo if (pcpu_setup_first_chunk(ai, fc) < 0) 2514bbddff05STejun Heo panic("Failed to initialize percpu areas."); 2515bbddff05STejun Heo } 2516bbddff05STejun Heo 2517bbddff05STejun Heo #endif /* CONFIG_SMP */ 2518bbddff05STejun Heo 2519099a19d9STejun Heo /* 25201a4d7607STejun Heo * Percpu allocator is initialized early during boot when neither slab or 25211a4d7607STejun Heo * workqueue is available. Plug async management until everything is up 25221a4d7607STejun Heo * and running. 25231a4d7607STejun Heo */ 25241a4d7607STejun Heo static int __init percpu_enable_async(void) 25251a4d7607STejun Heo { 25261a4d7607STejun Heo pcpu_async_enabled = true; 25271a4d7607STejun Heo return 0; 25281a4d7607STejun Heo } 25291a4d7607STejun Heo subsys_initcall(percpu_enable_async); 2530