Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
v4.6
 
   1/*
   2 * mm/percpu.c - percpu memory allocator
   3 *
   4 * Copyright (C) 2009		SUSE Linux Products GmbH
   5 * Copyright (C) 2009		Tejun Heo <tj@kernel.org>
   6 *
   7 * This file is released under the GPLv2.
 
   8 *
   9 * This is percpu allocator which can handle both static and dynamic
  10 * areas.  Percpu areas are allocated in chunks.  Each chunk is
  11 * consisted of boot-time determined number of units and the first
  12 * chunk is used for static percpu variables in the kernel image
  13 * (special boot time alloc/init handling necessary as these areas
  14 * need to be brought up before allocation services are running).
  15 * Unit grows as necessary and all units grow or shrink in unison.
  16 * When a chunk is filled up, another chunk is allocated.
  17 *
  18 *  c0                           c1                         c2
  19 *  -------------------          -------------------        ------------
  20 * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
  21 *  -------------------  ......  -------------------  ....  ------------
  22 *
  23 * Allocation is done in offset-size areas of single unit space.  Ie,
  24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
  25 * c1:u1, c1:u2 and c1:u3.  On UMA, units corresponds directly to
  26 * cpus.  On NUMA, the mapping can be non-linear and even sparse.
  27 * Percpu access can be done by configuring percpu base registers
  28 * according to cpu to unit mapping and pcpu_unit_size.
  29 *
  30 * There are usually many small percpu allocations many of them being
  31 * as small as 4 bytes.  The allocator organizes chunks into lists
  32 * according to free size and tries to allocate from the fullest one.
  33 * Each chunk keeps the maximum contiguous area size hint which is
  34 * guaranteed to be equal to or larger than the maximum contiguous
  35 * area in the chunk.  This helps the allocator not to iterate the
  36 * chunk maps unnecessarily.
  37 *
  38 * Allocation state in each chunk is kept using an array of integers
  39 * on chunk->map.  A positive value in the map represents a free
  40 * region and negative allocated.  Allocation inside a chunk is done
  41 * by scanning this map sequentially and serving the first matching
  42 * entry.  This is mostly copied from the percpu_modalloc() allocator.
  43 * Chunks can be determined from the address using the index field
  44 * in the page struct. The index field contains a pointer to the chunk.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  45 *
  46 * To use this allocator, arch code should do the followings.
  47 *
  48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
  49 *   regular address to percpu pointer and back if they need to be
  50 *   different from the default
  51 *
  52 * - use pcpu_setup_first_chunk() during percpu area initialization to
  53 *   setup the first chunk containing the kernel static percpu area
  54 */
  55
  56#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  57
  58#include <linux/bitmap.h>
  59#include <linux/bootmem.h>
 
  60#include <linux/err.h>
 
  61#include <linux/list.h>
  62#include <linux/log2.h>
  63#include <linux/mm.h>
  64#include <linux/module.h>
  65#include <linux/mutex.h>
  66#include <linux/percpu.h>
  67#include <linux/pfn.h>
  68#include <linux/slab.h>
  69#include <linux/spinlock.h>
  70#include <linux/vmalloc.h>
  71#include <linux/workqueue.h>
  72#include <linux/kmemleak.h>
 
 
 
  73
  74#include <asm/cacheflush.h>
  75#include <asm/sections.h>
  76#include <asm/tlbflush.h>
  77#include <asm/io.h>
  78
  79#define PCPU_SLOT_BASE_SHIFT		5	/* 1-31 shares the same slot */
  80#define PCPU_DFL_MAP_ALLOC		16	/* start a map with 16 ents */
  81#define PCPU_ATOMIC_MAP_MARGIN_LOW	32
  82#define PCPU_ATOMIC_MAP_MARGIN_HIGH	64
 
 
 
 
 
 
 
 
 
  83#define PCPU_EMPTY_POP_PAGES_LOW	2
  84#define PCPU_EMPTY_POP_PAGES_HIGH	4
  85
  86#ifdef CONFIG_SMP
  87/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
  88#ifndef __addr_to_pcpu_ptr
  89#define __addr_to_pcpu_ptr(addr)					\
  90	(void __percpu *)((unsigned long)(addr) -			\
  91			  (unsigned long)pcpu_base_addr	+		\
  92			  (unsigned long)__per_cpu_start)
  93#endif
  94#ifndef __pcpu_ptr_to_addr
  95#define __pcpu_ptr_to_addr(ptr)						\
  96	(void __force *)((unsigned long)(ptr) +				\
  97			 (unsigned long)pcpu_base_addr -		\
  98			 (unsigned long)__per_cpu_start)
  99#endif
 100#else	/* CONFIG_SMP */
 101/* on UP, it's always identity mapped */
 102#define __addr_to_pcpu_ptr(addr)	(void __percpu *)(addr)
 103#define __pcpu_ptr_to_addr(ptr)		(void __force *)(ptr)
 104#endif	/* CONFIG_SMP */
 105
 106struct pcpu_chunk {
 107	struct list_head	list;		/* linked to pcpu_slot lists */
 108	int			free_size;	/* free bytes in the chunk */
 109	int			contig_hint;	/* max contiguous size hint */
 110	void			*base_addr;	/* base address of this chunk */
 111
 112	int			map_used;	/* # of map entries used before the sentry */
 113	int			map_alloc;	/* # of map entries allocated */
 114	int			*map;		/* allocation map */
 115	struct work_struct	map_extend_work;/* async ->map[] extension */
 116
 117	void			*data;		/* chunk data */
 118	int			first_free;	/* no free below this */
 119	bool			immutable;	/* no [de]population allowed */
 120	int			nr_populated;	/* # of populated pages */
 121	unsigned long		populated[];	/* populated bitmap */
 122};
 123
 124static int pcpu_unit_pages __read_mostly;
 125static int pcpu_unit_size __read_mostly;
 126static int pcpu_nr_units __read_mostly;
 127static int pcpu_atom_size __read_mostly;
 128static int pcpu_nr_slots __read_mostly;
 129static size_t pcpu_chunk_struct_size __read_mostly;
 130
 131/* cpus with the lowest and highest unit addresses */
 132static unsigned int pcpu_low_unit_cpu __read_mostly;
 133static unsigned int pcpu_high_unit_cpu __read_mostly;
 134
 135/* the address of the first chunk which starts with the kernel static area */
 136void *pcpu_base_addr __read_mostly;
 137EXPORT_SYMBOL_GPL(pcpu_base_addr);
 138
 139static const int *pcpu_unit_map __read_mostly;		/* cpu -> unit */
 140const unsigned long *pcpu_unit_offsets __read_mostly;	/* cpu -> unit offset */
 141
 142/* group information, used for vm allocation */
 143static int pcpu_nr_groups __read_mostly;
 144static const unsigned long *pcpu_group_offsets __read_mostly;
 145static const size_t *pcpu_group_sizes __read_mostly;
 146
 147/*
 148 * The first chunk which always exists.  Note that unlike other
 149 * chunks, this one can be allocated and mapped in several different
 150 * ways and thus often doesn't live in the vmalloc area.
 151 */
 152static struct pcpu_chunk *pcpu_first_chunk;
 153
 154/*
 155 * Optional reserved chunk.  This chunk reserves part of the first
 156 * chunk and serves it for reserved allocations.  The amount of
 157 * reserved offset is in pcpu_reserved_chunk_limit.  When reserved
 158 * area doesn't exist, the following variables contain NULL and 0
 159 * respectively.
 160 */
 161static struct pcpu_chunk *pcpu_reserved_chunk;
 162static int pcpu_reserved_chunk_limit;
 
 
 
 
 163
 164static DEFINE_SPINLOCK(pcpu_lock);	/* all internal data structures */
 165static DEFINE_MUTEX(pcpu_alloc_mutex);	/* chunk create/destroy, [de]pop */
 166
 167static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
 
 
 
 
 168
 169/*
 170 * The number of empty populated pages, protected by pcpu_lock.  The
 171 * reserved chunk doesn't contribute to the count.
 
 
 172 */
 173static int pcpu_nr_empty_pop_pages;
 174
 175/*
 176 * Balance work is used to populate or destroy chunks asynchronously.  We
 177 * try to keep the number of populated free pages between
 178 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
 179 * empty chunk.
 180 */
 181static void pcpu_balance_workfn(struct work_struct *work);
 182static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
 183static bool pcpu_async_enabled __read_mostly;
 184static bool pcpu_atomic_alloc_failed;
 185
 186static void pcpu_schedule_balance_work(void)
 187{
 188	if (pcpu_async_enabled)
 189		schedule_work(&pcpu_balance_work);
 190}
 191
 192static bool pcpu_addr_in_first_chunk(void *addr)
 
 
 
 
 
 
 
 
 193{
 194	void *first_start = pcpu_first_chunk->base_addr;
 195
 196	return addr >= first_start && addr < first_start + pcpu_unit_size;
 197}
 198
 199static bool pcpu_addr_in_reserved_chunk(void *addr)
 200{
 201	void *first_start = pcpu_first_chunk->base_addr;
 202
 203	return addr >= first_start &&
 204		addr < first_start + pcpu_reserved_chunk_limit;
 205}
 206
 207static int __pcpu_size_to_slot(int size)
 208{
 209	int highbit = fls(size);	/* size is in bytes */
 210	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
 211}
 212
 213static int pcpu_size_to_slot(int size)
 214{
 215	if (size == pcpu_unit_size)
 216		return pcpu_nr_slots - 1;
 217	return __pcpu_size_to_slot(size);
 218}
 219
 220static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
 221{
 222	if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
 
 
 
 223		return 0;
 224
 225	return pcpu_size_to_slot(chunk->free_size);
 226}
 227
 228/* set the pointer to a chunk in a page struct */
 229static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 230{
 231	page->index = (unsigned long)pcpu;
 232}
 233
 234/* obtain pointer to a chunk from a page struct */
 235static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
 236{
 237	return (struct pcpu_chunk *)page->index;
 238}
 239
 240static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
 241{
 242	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
 243}
 244
 
 
 
 
 
 245static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 246				     unsigned int cpu, int page_idx)
 247{
 248	return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
 249		(page_idx << PAGE_SHIFT);
 
 
 
 
 
 
 
 
 
 
 250}
 251
 252static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
 253					   int *rs, int *re, int end)
 254{
 255	*rs = find_next_zero_bit(chunk->populated, end, *rs);
 256	*re = find_next_bit(chunk->populated, end, *rs + 1);
 257}
 258
 259static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
 260					 int *rs, int *re, int end)
 261{
 262	*rs = find_next_bit(chunk->populated, end, *rs);
 263	*re = find_next_zero_bit(chunk->populated, end, *rs + 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 264}
 265
 266/*
 267 * (Un)populated page region iterators.  Iterate over (un)populated
 268 * page regions between @start and @end in @chunk.  @rs and @re should
 269 * be integer variables and will be set to start and end page index of
 270 * the current region.
 271 */
 272#define pcpu_for_each_unpop_region(chunk, rs, re, start, end)		    \
 273	for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
 274	     (rs) < (re);						    \
 275	     (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
 276
 277#define pcpu_for_each_pop_region(chunk, rs, re, start, end)		    \
 278	for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end));   \
 279	     (rs) < (re);						    \
 280	     (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 281
 282/**
 283 * pcpu_mem_zalloc - allocate memory
 284 * @size: bytes to allocate
 
 285 *
 286 * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
 287 * kzalloc() is used; otherwise, vzalloc() is used.  The returned
 288 * memory is always zeroed.
 289 *
 290 * CONTEXT:
 291 * Does GFP_KERNEL allocation.
 292 *
 293 * RETURNS:
 294 * Pointer to the allocated area on success, NULL on failure.
 295 */
 296static void *pcpu_mem_zalloc(size_t size)
 297{
 298	if (WARN_ON_ONCE(!slab_is_available()))
 299		return NULL;
 300
 301	if (size <= PAGE_SIZE)
 302		return kzalloc(size, GFP_KERNEL);
 303	else
 304		return vzalloc(size);
 305}
 306
 307/**
 308 * pcpu_mem_free - free memory
 309 * @ptr: memory to free
 310 *
 311 * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
 312 */
 313static void pcpu_mem_free(void *ptr)
 314{
 315	kvfree(ptr);
 316}
 317
 318/**
 319 * pcpu_count_occupied_pages - count the number of pages an area occupies
 320 * @chunk: chunk of interest
 321 * @i: index of the area in question
 322 *
 323 * Count the number of pages chunk's @i'th area occupies.  When the area's
 324 * start and/or end address isn't aligned to page boundary, the straddled
 325 * page is included in the count iff the rest of the page is free.
 326 */
 327static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
 328{
 329	int off = chunk->map[i] & ~1;
 330	int end = chunk->map[i + 1] & ~1;
 331
 332	if (!PAGE_ALIGNED(off) && i > 0) {
 333		int prev = chunk->map[i - 1];
 334
 335		if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
 336			off = round_down(off, PAGE_SIZE);
 337	}
 338
 339	if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
 340		int next = chunk->map[i + 1];
 341		int nend = chunk->map[i + 2] & ~1;
 342
 343		if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
 344			end = round_up(end, PAGE_SIZE);
 345	}
 
 346
 347	return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
 
 
 348}
 349
 350/**
 351 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
 352 * @chunk: chunk of interest
 353 * @oslot: the previous slot it was on
 354 *
 355 * This function is called after an allocation or free changed @chunk.
 356 * New slot according to the changed state is determined and @chunk is
 357 * moved to the slot.  Note that the reserved chunk is never put on
 358 * chunk slots.
 359 *
 360 * CONTEXT:
 361 * pcpu_lock.
 362 */
 363static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 364{
 365	int nslot = pcpu_chunk_slot(chunk);
 366
 367	if (chunk != pcpu_reserved_chunk && oslot != nslot) {
 368		if (oslot < nslot)
 369			list_move(&chunk->list, &pcpu_slot[nslot]);
 370		else
 371			list_move_tail(&chunk->list, &pcpu_slot[nslot]);
 
 
 
 
 
 
 
 
 
 
 372	}
 
 373}
 374
 375/**
 376 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
 
 
 
 
 
 
 
 
 
 
 
 377 * @chunk: chunk of interest
 378 * @is_atomic: the allocation context
 379 *
 380 * Determine whether area map of @chunk needs to be extended.  If
 381 * @is_atomic, only the amount necessary for a new allocation is
 382 * considered; however, async extension is scheduled if the left amount is
 383 * low.  If !@is_atomic, it aims for more empty space.  Combined, this
 384 * ensures that the map is likely to have enough available space to
 385 * accomodate atomic allocations which can't extend maps directly.
 
 
 
 
 
 
 
 
 
 
 
 386 *
 387 * CONTEXT:
 388 * pcpu_lock.
 
 
 
 
 
 
 
 
 
 
 
 389 *
 390 * RETURNS:
 391 * New target map allocation length if extension is necessary, 0
 392 * otherwise.
 393 */
 394static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
 395{
 396	int margin, new_alloc;
 397
 398	if (is_atomic) {
 399		margin = 3;
 
 400
 401		if (chunk->map_alloc <
 402		    chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
 403		    pcpu_async_enabled)
 404			schedule_work(&chunk->map_extend_work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 405	} else {
 406		margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
 
 
 
 
 
 
 
 
 
 
 
 407	}
 
 408
 409	if (chunk->map_alloc >= chunk->map_used + margin)
 410		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 411
 412	new_alloc = PCPU_DFL_MAP_ALLOC;
 413	while (new_alloc < chunk->map_used + margin)
 414		new_alloc *= 2;
 415
 416	return new_alloc;
 417}
 418
 419/**
 420 * pcpu_extend_area_map - extend area map of a chunk
 421 * @chunk: chunk of interest
 422 * @new_alloc: new target allocation length of the area map
 423 *
 424 * Extend area map of @chunk to have @new_alloc entries.
 425 *
 426 * CONTEXT:
 427 * Does GFP_KERNEL allocation.  Grabs and releases pcpu_lock.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 428 *
 429 * RETURNS:
 430 * 0 on success, -errno on failure.
 431 */
 432static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
 433{
 434	int *old = NULL, *new = NULL;
 435	size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
 436	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 437
 438	new = pcpu_mem_zalloc(new_size);
 439	if (!new)
 440		return -ENOMEM;
 441
 442	/* acquire pcpu_lock and switch to new area map */
 443	spin_lock_irqsave(&pcpu_lock, flags);
 
 
 
 444
 445	if (new_alloc <= chunk->map_alloc)
 446		goto out_unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 447
 448	old_size = chunk->map_alloc * sizeof(chunk->map[0]);
 449	old = chunk->map;
 
 
 
 
 
 
 
 
 450
 451	memcpy(new, old, old_size);
 
 452
 453	chunk->map_alloc = new_alloc;
 454	chunk->map = new;
 455	new = NULL;
 
 
 
 
 
 456
 457out_unlock:
 458	spin_unlock_irqrestore(&pcpu_lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 459
 460	/*
 461	 * pcpu_mem_free() might end up calling vfree() which uses
 462	 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
 463	 */
 464	pcpu_mem_free(old);
 465	pcpu_mem_free(new);
 
 466
 467	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 468}
 469
 470static void pcpu_map_extend_workfn(struct work_struct *work)
 471{
 472	struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk,
 473						map_extend_work);
 474	int new_alloc;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 475
 476	spin_lock_irq(&pcpu_lock);
 477	new_alloc = pcpu_need_to_extend(chunk, false);
 478	spin_unlock_irq(&pcpu_lock);
 
 
 
 
 
 
 
 479
 480	if (new_alloc)
 481		pcpu_extend_area_map(chunk, new_alloc);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 482}
 483
 484/**
 485 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
 486 * @chunk: chunk the candidate area belongs to
 487 * @off: the offset to the start of the candidate area
 488 * @this_size: the size of the candidate area
 489 * @size: the size of the target allocation
 490 * @align: the alignment of the target allocation
 491 * @pop_only: only allocate from already populated region
 492 *
 493 * We're trying to allocate @size bytes aligned at @align.  @chunk's area
 494 * at @off sized @this_size is a candidate.  This function determines
 495 * whether the target allocation fits in the candidate area and returns the
 496 * number of bytes to pad after @off.  If the target area doesn't fit, -1
 497 * is returned.
 498 *
 499 * If @pop_only is %true, this function only considers the already
 500 * populated part of the candidate area.
 
 501 */
 502static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
 503			    int size, int align, bool pop_only)
 504{
 505	int cand_off = off;
 506
 507	while (true) {
 508		int head = ALIGN(cand_off, align) - off;
 509		int page_start, page_end, rs, re;
 510
 511		if (this_size < head + size)
 512			return -1;
 
 
 513
 514		if (!pop_only)
 515			return head;
 
 516
 517		/*
 518		 * If the first unpopulated page is beyond the end of the
 519		 * allocation, the whole allocation is populated;
 520		 * otherwise, retry from the end of the unpopulated area.
 521		 */
 522		page_start = PFN_DOWN(head + off);
 523		page_end = PFN_UP(head + off + size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 524
 525		rs = page_start;
 526		pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
 527		if (rs >= page_end)
 528			return head;
 529		cand_off = re * PAGE_SIZE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 530	}
 
 531}
 532
 533/**
 534 * pcpu_alloc_area - allocate area from a pcpu_chunk
 535 * @chunk: chunk of interest
 536 * @size: wanted size in bytes
 537 * @align: wanted align
 538 * @pop_only: allocate only from the populated area
 539 * @occ_pages_p: out param for the number of pages the area occupies
 540 *
 541 * Try to allocate @size bytes area aligned at @align from @chunk.
 542 * Note that this function only allocates the offset.  It doesn't
 543 * populate or map the area.
 544 *
 545 * @chunk->map must have at least two free slots.
 
 
 
 
 
 
 546 *
 547 * CONTEXT:
 548 * pcpu_lock.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 549 *
 550 * RETURNS:
 551 * Allocated offset in @chunk on success, -1 if no matching area is
 552 * found.
 553 */
 554static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
 555			   bool pop_only, int *occ_pages_p)
 556{
 557	int oslot = pcpu_chunk_slot(chunk);
 558	int max_contig = 0;
 559	int i, off;
 560	bool seen_free = false;
 561	int *p;
 562
 563	for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
 564		int head, tail;
 565		int this_size;
 566
 567		off = *p;
 568		if (off & 1)
 569			continue;
 570
 571		this_size = (p[1] & ~1) - off;
 572
 573		head = pcpu_fit_in_area(chunk, off, this_size, size, align,
 574					pop_only);
 575		if (head < 0) {
 576			if (!seen_free) {
 577				chunk->first_free = i;
 578				seen_free = true;
 579			}
 580			max_contig = max(this_size, max_contig);
 581			continue;
 582		}
 583
 584		/*
 585		 * If head is small or the previous block is free,
 586		 * merge'em.  Note that 'small' is defined as smaller
 587		 * than sizeof(int), which is very small but isn't too
 588		 * uncommon for percpu allocations.
 589		 */
 590		if (head && (head < sizeof(int) || !(p[-1] & 1))) {
 591			*p = off += head;
 592			if (p[-1] & 1)
 593				chunk->free_size -= head;
 594			else
 595				max_contig = max(*p - p[-1], max_contig);
 596			this_size -= head;
 597			head = 0;
 598		}
 599
 600		/* if tail is small, just keep it around */
 601		tail = this_size - head - size;
 602		if (tail < sizeof(int)) {
 603			tail = 0;
 604			size = this_size - head;
 605		}
 606
 607		/* split if warranted */
 608		if (head || tail) {
 609			int nr_extra = !!head + !!tail;
 610
 611			/* insert new subblocks */
 612			memmove(p + nr_extra + 1, p + 1,
 613				sizeof(chunk->map[0]) * (chunk->map_used - i));
 614			chunk->map_used += nr_extra;
 615
 616			if (head) {
 617				if (!seen_free) {
 618					chunk->first_free = i;
 619					seen_free = true;
 620				}
 621				*++p = off += head;
 622				++i;
 623				max_contig = max(head, max_contig);
 624			}
 625			if (tail) {
 626				p[1] = off + size;
 627				max_contig = max(tail, max_contig);
 628			}
 629		}
 630
 631		if (!seen_free)
 632			chunk->first_free = i + 1;
 633
 634		/* update hint and mark allocated */
 635		if (i + 1 == chunk->map_used)
 636			chunk->contig_hint = max_contig; /* fully scanned */
 637		else
 638			chunk->contig_hint = max(chunk->contig_hint,
 639						 max_contig);
 640
 641		chunk->free_size -= size;
 642		*p |= 1;
 643
 644		*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
 645		pcpu_chunk_relocate(chunk, oslot);
 646		return off;
 647	}
 648
 649	chunk->contig_hint = max_contig;	/* fully scanned */
 650	pcpu_chunk_relocate(chunk, oslot);
 651
 652	/* tell the upper layer that this chunk has no matching area */
 653	return -1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 654}
 655
 656/**
 657 * pcpu_free_area - free area to a pcpu_chunk
 658 * @chunk: chunk of interest
 659 * @freeme: offset of area to free
 660 * @occ_pages_p: out param for the number of pages the area occupies
 661 *
 662 * Free area starting from @freeme to @chunk.  Note that this function
 663 * only modifies the allocation map.  It doesn't depopulate or unmap
 664 * the area.
 
 665 *
 666 * CONTEXT:
 667 * pcpu_lock.
 668 */
 669static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
 670			   int *occ_pages_p)
 671{
 672	int oslot = pcpu_chunk_slot(chunk);
 673	int off = 0;
 674	unsigned i, j;
 675	int to_free = 0;
 676	int *p;
 677
 678	freeme |= 1;	/* we are searching for <given offset, in use> pair */
 679
 680	i = 0;
 681	j = chunk->map_used;
 682	while (i != j) {
 683		unsigned k = (i + j) / 2;
 684		off = chunk->map[k];
 685		if (off < freeme)
 686			i = k + 1;
 687		else if (off > freeme)
 688			j = k;
 689		else
 690			i = j = k;
 691	}
 692	BUG_ON(off != freeme);
 693
 694	if (i < chunk->first_free)
 695		chunk->first_free = i;
 696
 697	p = chunk->map + i;
 698	*p = off &= ~1;
 699	chunk->free_size += (p[1] & ~1) - off;
 
 
 
 
 
 
 700
 701	*occ_pages_p = pcpu_count_occupied_pages(chunk, i);
 
 
 
 
 
 
 702
 703	/* merge with next? */
 704	if (!(p[1] & 1))
 705		to_free++;
 706	/* merge with previous? */
 707	if (i > 0 && !(p[-1] & 1)) {
 708		to_free++;
 709		i--;
 710		p--;
 711	}
 712	if (to_free) {
 713		chunk->map_used -= to_free;
 714		memmove(p + 1, p + 1 + to_free,
 715			(chunk->map_used - i) * sizeof(chunk->map[0]));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 716	}
 717
 718	chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
 719	pcpu_chunk_relocate(chunk, oslot);
 720}
 721
 722static struct pcpu_chunk *pcpu_alloc_chunk(void)
 723{
 724	struct pcpu_chunk *chunk;
 
 725
 726	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
 727	if (!chunk)
 728		return NULL;
 729
 730	chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
 731						sizeof(chunk->map[0]));
 732	if (!chunk->map) {
 733		pcpu_mem_free(chunk);
 734		return NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 735	}
 
 736
 737	chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
 738	chunk->map[0] = 0;
 739	chunk->map[1] = pcpu_unit_size | 1;
 740	chunk->map_used = 1;
 741
 742	INIT_LIST_HEAD(&chunk->list);
 743	INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn);
 744	chunk->free_size = pcpu_unit_size;
 745	chunk->contig_hint = pcpu_unit_size;
 746
 747	return chunk;
 
 
 
 
 
 
 
 
 
 
 
 
 
 748}
 749
 750static void pcpu_free_chunk(struct pcpu_chunk *chunk)
 751{
 752	if (!chunk)
 753		return;
 754	pcpu_mem_free(chunk->map);
 
 
 
 
 
 755	pcpu_mem_free(chunk);
 756}
 757
 758/**
 759 * pcpu_chunk_populated - post-population bookkeeping
 760 * @chunk: pcpu_chunk which got populated
 761 * @page_start: the start page
 762 * @page_end: the end page
 763 *
 764 * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
 765 * the bookkeeping information accordingly.  Must be called after each
 766 * successful population.
 
 
 
 767 */
 768static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
 769				 int page_start, int page_end)
 770{
 771	int nr = page_end - page_start;
 772
 773	lockdep_assert_held(&pcpu_lock);
 774
 775	bitmap_set(chunk->populated, page_start, nr);
 776	chunk->nr_populated += nr;
 777	pcpu_nr_empty_pop_pages += nr;
 
 
 778}
 779
 780/**
 781 * pcpu_chunk_depopulated - post-depopulation bookkeeping
 782 * @chunk: pcpu_chunk which got depopulated
 783 * @page_start: the start page
 784 * @page_end: the end page
 785 *
 786 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
 787 * Update the bookkeeping information accordingly.  Must be called after
 788 * each successful depopulation.
 789 */
 790static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
 791				   int page_start, int page_end)
 792{
 793	int nr = page_end - page_start;
 794
 795	lockdep_assert_held(&pcpu_lock);
 796
 797	bitmap_clear(chunk->populated, page_start, nr);
 798	chunk->nr_populated -= nr;
 799	pcpu_nr_empty_pop_pages -= nr;
 
 
 800}
 801
 802/*
 803 * Chunk management implementation.
 804 *
 805 * To allow different implementations, chunk alloc/free and
 806 * [de]population are implemented in a separate file which is pulled
 807 * into this file and compiled together.  The following functions
 808 * should be implemented.
 809 *
 810 * pcpu_populate_chunk		- populate the specified range of a chunk
 811 * pcpu_depopulate_chunk	- depopulate the specified range of a chunk
 
 812 * pcpu_create_chunk		- create a new chunk
 813 * pcpu_destroy_chunk		- destroy a chunk, always preceded by full depop
 814 * pcpu_addr_to_page		- translate address to physical address
 815 * pcpu_verify_alloc_info	- check alloc_info is acceptable during init
 816 */
 817static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
 818static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
 819static struct pcpu_chunk *pcpu_create_chunk(void);
 
 
 
 
 820static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
 821static struct page *pcpu_addr_to_page(void *addr);
 822static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
 823
 824#ifdef CONFIG_NEED_PER_CPU_KM
 825#include "percpu-km.c"
 826#else
 827#include "percpu-vm.c"
 828#endif
 829
 830/**
 831 * pcpu_chunk_addr_search - determine chunk containing specified address
 832 * @addr: address for which the chunk needs to be determined.
 833 *
 
 
 
 834 * RETURNS:
 835 * The address of the found chunk.
 836 */
 837static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
 838{
 839	/* is it in the first chunk? */
 840	if (pcpu_addr_in_first_chunk(addr)) {
 841		/* is it in the reserved area? */
 842		if (pcpu_addr_in_reserved_chunk(addr))
 843			return pcpu_reserved_chunk;
 844		return pcpu_first_chunk;
 845	}
 
 
 
 846
 847	/*
 848	 * The address is relative to unit0 which might be unused and
 849	 * thus unmapped.  Offset the address to the unit space of the
 850	 * current processor before looking it up in the vmalloc
 851	 * space.  Note that any possible cpu id can be used here, so
 852	 * there's no need to worry about preemption or cpu hotplug.
 853	 */
 854	addr += pcpu_unit_offsets[raw_smp_processor_id()];
 855	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
 856}
 857
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 858/**
 859 * pcpu_alloc - the percpu allocator
 860 * @size: size of area to allocate in bytes
 861 * @align: alignment of area (max PAGE_SIZE)
 862 * @reserved: allocate from the reserved chunk if available
 863 * @gfp: allocation flags
 864 *
 865 * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
 866 * contain %GFP_KERNEL, the allocation is atomic.
 
 
 867 *
 868 * RETURNS:
 869 * Percpu pointer to the allocated area on success, NULL on failure.
 870 */
 871static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
 872				 gfp_t gfp)
 873{
 
 
 
 
 874	static int warn_limit = 10;
 875	struct pcpu_chunk *chunk;
 876	const char *err;
 877	bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
 878	int occ_pages = 0;
 879	int slot, off, new_alloc, cpu, ret;
 880	unsigned long flags;
 881	void __percpu *ptr;
 
 
 
 
 
 
 
 882
 883	/*
 884	 * We want the lowest bit of offset available for in-use/free
 885	 * indicator, so force >= 16bit alignment and make size even.
 
 
 886	 */
 887	if (unlikely(align < 2))
 888		align = 2;
 889
 890	size = ALIGN(size, 2);
 891
 892	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
 893		WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
 
 
 
 
 
 894		     size, align);
 895		return NULL;
 896	}
 897
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 898	spin_lock_irqsave(&pcpu_lock, flags);
 899
 900	/* serve reserved allocations from the reserved chunk if available */
 901	if (reserved && pcpu_reserved_chunk) {
 902		chunk = pcpu_reserved_chunk;
 903
 904		if (size > chunk->contig_hint) {
 
 905			err = "alloc from reserved chunk failed";
 906			goto fail_unlock;
 907		}
 908
 909		while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
 910			spin_unlock_irqrestore(&pcpu_lock, flags);
 911			if (is_atomic ||
 912			    pcpu_extend_area_map(chunk, new_alloc) < 0) {
 913				err = "failed to extend area map of reserved chunk";
 914				goto fail;
 915			}
 916			spin_lock_irqsave(&pcpu_lock, flags);
 917		}
 918
 919		off = pcpu_alloc_area(chunk, size, align, is_atomic,
 920				      &occ_pages);
 921		if (off >= 0)
 922			goto area_found;
 923
 924		err = "alloc from reserved chunk failed";
 925		goto fail_unlock;
 926	}
 927
 928restart:
 929	/* search through normal chunks */
 930	for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
 931		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
 932			if (size > chunk->contig_hint)
 
 
 
 
 
 933				continue;
 934
 935			new_alloc = pcpu_need_to_extend(chunk, is_atomic);
 936			if (new_alloc) {
 937				if (is_atomic)
 938					continue;
 939				spin_unlock_irqrestore(&pcpu_lock, flags);
 940				if (pcpu_extend_area_map(chunk,
 941							 new_alloc) < 0) {
 942					err = "failed to extend area map";
 943					goto fail;
 944				}
 945				spin_lock_irqsave(&pcpu_lock, flags);
 946				/*
 947				 * pcpu_lock has been dropped, need to
 948				 * restart cpu_slot list walking.
 949				 */
 950				goto restart;
 951			}
 952
 953			off = pcpu_alloc_area(chunk, size, align, is_atomic,
 954					      &occ_pages);
 955			if (off >= 0)
 956				goto area_found;
 
 957		}
 958	}
 959
 960	spin_unlock_irqrestore(&pcpu_lock, flags);
 961
 962	/*
 963	 * No space left.  Create a new chunk.  We don't want multiple
 964	 * tasks to create chunks simultaneously.  Serialize and create iff
 965	 * there's still no empty chunk after grabbing the mutex.
 966	 */
 967	if (is_atomic)
 
 968		goto fail;
 
 969
 970	mutex_lock(&pcpu_alloc_mutex);
 971
 972	if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
 973		chunk = pcpu_create_chunk();
 974		if (!chunk) {
 975			mutex_unlock(&pcpu_alloc_mutex);
 976			err = "failed to allocate new chunk";
 977			goto fail;
 978		}
 979
 980		spin_lock_irqsave(&pcpu_lock, flags);
 981		pcpu_chunk_relocate(chunk, -1);
 982	} else {
 983		spin_lock_irqsave(&pcpu_lock, flags);
 984	}
 985
 986	mutex_unlock(&pcpu_alloc_mutex);
 987	goto restart;
 988
 989area_found:
 
 990	spin_unlock_irqrestore(&pcpu_lock, flags);
 991
 992	/* populate if not all pages are already there */
 993	if (!is_atomic) {
 994		int page_start, page_end, rs, re;
 995
 996		mutex_lock(&pcpu_alloc_mutex);
 997
 998		page_start = PFN_DOWN(off);
 999		page_end = PFN_UP(off + size);
1000
1001		pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
 
1002			WARN_ON(chunk->immutable);
1003
1004			ret = pcpu_populate_chunk(chunk, rs, re);
1005
1006			spin_lock_irqsave(&pcpu_lock, flags);
1007			if (ret) {
1008				mutex_unlock(&pcpu_alloc_mutex);
1009				pcpu_free_area(chunk, off, &occ_pages);
1010				err = "failed to populate";
1011				goto fail_unlock;
1012			}
1013			pcpu_chunk_populated(chunk, rs, re);
1014			spin_unlock_irqrestore(&pcpu_lock, flags);
1015		}
1016
1017		mutex_unlock(&pcpu_alloc_mutex);
1018	}
1019
1020	if (chunk != pcpu_reserved_chunk)
1021		pcpu_nr_empty_pop_pages -= occ_pages;
1022
1023	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1024		pcpu_schedule_balance_work();
1025
1026	/* clear the areas and return address relative to base address */
1027	for_each_possible_cpu(cpu)
1028		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1029
1030	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1031	kmemleak_alloc_percpu(ptr, size, gfp);
 
 
 
 
 
 
1032	return ptr;
1033
1034fail_unlock:
1035	spin_unlock_irqrestore(&pcpu_lock, flags);
1036fail:
1037	if (!is_atomic && warn_limit) {
 
 
1038		pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1039			size, align, is_atomic, err);
1040		dump_stack();
1041		if (!--warn_limit)
1042			pr_info("limit reached, disable warning\n");
1043	}
1044	if (is_atomic) {
1045		/* see the flag handling in pcpu_blance_workfn() */
1046		pcpu_atomic_alloc_failed = true;
1047		pcpu_schedule_balance_work();
 
 
1048	}
 
 
 
1049	return NULL;
1050}
1051
1052/**
1053 * __alloc_percpu_gfp - allocate dynamic percpu area
1054 * @size: size of area to allocate in bytes
1055 * @align: alignment of area (max PAGE_SIZE)
1056 * @gfp: allocation flags
1057 *
1058 * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1059 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1060 * be called from any context but is a lot more likely to fail.
 
 
1061 *
1062 * RETURNS:
1063 * Percpu pointer to the allocated area on success, NULL on failure.
1064 */
1065void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1066{
1067	return pcpu_alloc(size, align, false, gfp);
1068}
1069EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1070
1071/**
1072 * __alloc_percpu - allocate dynamic percpu area
1073 * @size: size of area to allocate in bytes
1074 * @align: alignment of area (max PAGE_SIZE)
1075 *
1076 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1077 */
1078void __percpu *__alloc_percpu(size_t size, size_t align)
1079{
1080	return pcpu_alloc(size, align, false, GFP_KERNEL);
1081}
1082EXPORT_SYMBOL_GPL(__alloc_percpu);
1083
1084/**
1085 * __alloc_reserved_percpu - allocate reserved percpu area
1086 * @size: size of area to allocate in bytes
1087 * @align: alignment of area (max PAGE_SIZE)
1088 *
1089 * Allocate zero-filled percpu area of @size bytes aligned at @align
1090 * from reserved percpu area if arch has set it up; otherwise,
1091 * allocation is served from the same dynamic area.  Might sleep.
1092 * Might trigger writeouts.
1093 *
1094 * CONTEXT:
1095 * Does GFP_KERNEL allocation.
1096 *
1097 * RETURNS:
1098 * Percpu pointer to the allocated area on success, NULL on failure.
1099 */
1100void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1101{
1102	return pcpu_alloc(size, align, true, GFP_KERNEL);
1103}
1104
1105/**
1106 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1107 * @work: unused
1108 *
1109 * Reclaim all fully free chunks except for the first one.
 
 
 
 
 
1110 */
1111static void pcpu_balance_workfn(struct work_struct *work)
1112{
1113	LIST_HEAD(to_free);
1114	struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1115	struct pcpu_chunk *chunk, *next;
1116	int slot, nr_to_pop, ret;
 
1117
1118	/*
1119	 * There's no reason to keep around multiple unused chunks and VM
1120	 * areas can be scarce.  Destroy all free chunks except for one.
1121	 */
1122	mutex_lock(&pcpu_alloc_mutex);
1123	spin_lock_irq(&pcpu_lock);
1124
1125	list_for_each_entry_safe(chunk, next, free_head, list) {
1126		WARN_ON(chunk->immutable);
1127
1128		/* spare the first one */
1129		if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1130			continue;
1131
1132		list_move(&chunk->list, &to_free);
 
1133	}
1134
1135	spin_unlock_irq(&pcpu_lock);
 
1136
 
1137	list_for_each_entry_safe(chunk, next, &to_free, list) {
1138		int rs, re;
1139
1140		pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
 
1141			pcpu_depopulate_chunk(chunk, rs, re);
1142			spin_lock_irq(&pcpu_lock);
1143			pcpu_chunk_depopulated(chunk, rs, re);
1144			spin_unlock_irq(&pcpu_lock);
1145		}
1146		pcpu_destroy_chunk(chunk);
 
1147	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1148
1149	/*
1150	 * Ensure there are certain number of free populated pages for
1151	 * atomic allocs.  Fill up from the most packed so that atomic
1152	 * allocs don't increase fragmentation.  If atomic allocation
1153	 * failed previously, always populate the maximum amount.  This
1154	 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1155	 * failing indefinitely; however, large atomic allocs are not
1156	 * something we support properly and can be highly unreliable and
1157	 * inefficient.
1158	 */
1159retry_pop:
1160	if (pcpu_atomic_alloc_failed) {
1161		nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1162		/* best effort anyway, don't worry about synchronization */
1163		pcpu_atomic_alloc_failed = false;
1164	} else {
1165		nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1166				  pcpu_nr_empty_pop_pages,
1167				  0, PCPU_EMPTY_POP_PAGES_HIGH);
1168	}
1169
1170	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1171		int nr_unpop = 0, rs, re;
1172
1173		if (!nr_to_pop)
1174			break;
1175
1176		spin_lock_irq(&pcpu_lock);
1177		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1178			nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1179			if (nr_unpop)
1180				break;
1181		}
1182		spin_unlock_irq(&pcpu_lock);
1183
1184		if (!nr_unpop)
1185			continue;
1186
1187		/* @chunk can't go away while pcpu_alloc_mutex is held */
1188		pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1189			int nr = min(re - rs, nr_to_pop);
 
1190
1191			ret = pcpu_populate_chunk(chunk, rs, rs + nr);
 
 
 
1192			if (!ret) {
1193				nr_to_pop -= nr;
1194				spin_lock_irq(&pcpu_lock);
1195				pcpu_chunk_populated(chunk, rs, rs + nr);
1196				spin_unlock_irq(&pcpu_lock);
1197			} else {
1198				nr_to_pop = 0;
1199			}
1200
1201			if (!nr_to_pop)
1202				break;
1203		}
1204	}
1205
1206	if (nr_to_pop) {
1207		/* ran out of chunks to populate, create a new one and retry */
1208		chunk = pcpu_create_chunk();
 
 
 
1209		if (chunk) {
1210			spin_lock_irq(&pcpu_lock);
1211			pcpu_chunk_relocate(chunk, -1);
1212			spin_unlock_irq(&pcpu_lock);
1213			goto retry_pop;
1214		}
1215	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1216
 
1217	mutex_unlock(&pcpu_alloc_mutex);
1218}
1219
1220/**
1221 * free_percpu - free percpu area
1222 * @ptr: pointer to area to free
1223 *
1224 * Free percpu area @ptr.
1225 *
1226 * CONTEXT:
1227 * Can be called from atomic context.
1228 */
1229void free_percpu(void __percpu *ptr)
1230{
1231	void *addr;
1232	struct pcpu_chunk *chunk;
1233	unsigned long flags;
1234	int off, occ_pages;
 
1235
1236	if (!ptr)
1237		return;
1238
1239	kmemleak_free_percpu(ptr);
1240
1241	addr = __pcpu_ptr_to_addr(ptr);
1242
1243	spin_lock_irqsave(&pcpu_lock, flags);
1244
1245	chunk = pcpu_chunk_addr_search(addr);
1246	off = addr - chunk->base_addr;
1247
1248	pcpu_free_area(chunk, off, &occ_pages);
1249
1250	if (chunk != pcpu_reserved_chunk)
1251		pcpu_nr_empty_pop_pages += occ_pages;
1252
1253	/* if there are more than one fully free chunks, wake up grim reaper */
1254	if (chunk->free_size == pcpu_unit_size) {
 
 
 
 
1255		struct pcpu_chunk *pos;
1256
1257		list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1258			if (pos != chunk) {
1259				pcpu_schedule_balance_work();
1260				break;
1261			}
 
 
 
1262	}
1263
 
 
1264	spin_unlock_irqrestore(&pcpu_lock, flags);
 
 
 
1265}
1266EXPORT_SYMBOL_GPL(free_percpu);
1267
1268/**
1269 * is_kernel_percpu_address - test whether address is from static percpu area
1270 * @addr: address to test
1271 *
1272 * Test whether @addr belongs to in-kernel static percpu area.  Module
1273 * static percpu areas are not considered.  For those, use
1274 * is_module_percpu_address().
1275 *
1276 * RETURNS:
1277 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1278 */
1279bool is_kernel_percpu_address(unsigned long addr)
1280{
1281#ifdef CONFIG_SMP
1282	const size_t static_size = __per_cpu_end - __per_cpu_start;
1283	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1284	unsigned int cpu;
1285
1286	for_each_possible_cpu(cpu) {
1287		void *start = per_cpu_ptr(base, cpu);
 
1288
1289		if ((void *)addr >= start && (void *)addr < start + static_size)
 
 
 
 
 
1290			return true;
1291        }
 
1292#endif
1293	/* on UP, can't distinguish from other static vars, always false */
1294	return false;
1295}
1296
1297/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1298 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1299 * @addr: the address to be converted to physical address
1300 *
1301 * Given @addr which is dereferenceable address obtained via one of
1302 * percpu access macros, this function translates it into its physical
1303 * address.  The caller is responsible for ensuring @addr stays valid
1304 * until this function finishes.
1305 *
1306 * percpu allocator has special setup for the first chunk, which currently
1307 * supports either embedding in linear address space or vmalloc mapping,
1308 * and, from the second one, the backing allocator (currently either vm or
1309 * km) provides translation.
1310 *
1311 * The addr can be translated simply without checking if it falls into the
1312 * first chunk. But the current code reflects better how percpu allocator
1313 * actually works, and the verification can discover both bugs in percpu
1314 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1315 * code.
1316 *
1317 * RETURNS:
1318 * The physical address for @addr.
1319 */
1320phys_addr_t per_cpu_ptr_to_phys(void *addr)
1321{
1322	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1323	bool in_first_chunk = false;
1324	unsigned long first_low, first_high;
1325	unsigned int cpu;
1326
1327	/*
1328	 * The following test on unit_low/high isn't strictly
1329	 * necessary but will speed up lookups of addresses which
1330	 * aren't in the first chunk.
 
 
 
 
 
1331	 */
1332	first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1333	first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1334				     pcpu_unit_pages);
 
1335	if ((unsigned long)addr >= first_low &&
1336	    (unsigned long)addr < first_high) {
1337		for_each_possible_cpu(cpu) {
1338			void *start = per_cpu_ptr(base, cpu);
1339
1340			if (addr >= start && addr < start + pcpu_unit_size) {
1341				in_first_chunk = true;
1342				break;
1343			}
1344		}
1345	}
1346
1347	if (in_first_chunk) {
1348		if (!is_vmalloc_addr(addr))
1349			return __pa(addr);
1350		else
1351			return page_to_phys(vmalloc_to_page(addr)) +
1352			       offset_in_page(addr);
1353	} else
1354		return page_to_phys(pcpu_addr_to_page(addr)) +
1355		       offset_in_page(addr);
1356}
1357
1358/**
1359 * pcpu_alloc_alloc_info - allocate percpu allocation info
1360 * @nr_groups: the number of groups
1361 * @nr_units: the number of units
1362 *
1363 * Allocate ai which is large enough for @nr_groups groups containing
1364 * @nr_units units.  The returned ai's groups[0].cpu_map points to the
1365 * cpu_map array which is long enough for @nr_units and filled with
1366 * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
1367 * pointer of other groups.
1368 *
1369 * RETURNS:
1370 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1371 * failure.
1372 */
1373struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1374						      int nr_units)
1375{
1376	struct pcpu_alloc_info *ai;
1377	size_t base_size, ai_size;
1378	void *ptr;
1379	int unit;
1380
1381	base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1382			  __alignof__(ai->groups[0].cpu_map[0]));
1383	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1384
1385	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1386	if (!ptr)
1387		return NULL;
1388	ai = ptr;
1389	ptr += base_size;
1390
1391	ai->groups[0].cpu_map = ptr;
1392
1393	for (unit = 0; unit < nr_units; unit++)
1394		ai->groups[0].cpu_map[unit] = NR_CPUS;
1395
1396	ai->nr_groups = nr_groups;
1397	ai->__ai_size = PFN_ALIGN(ai_size);
1398
1399	return ai;
1400}
1401
1402/**
1403 * pcpu_free_alloc_info - free percpu allocation info
1404 * @ai: pcpu_alloc_info to free
1405 *
1406 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1407 */
1408void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1409{
1410	memblock_free_early(__pa(ai), ai->__ai_size);
1411}
1412
1413/**
1414 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1415 * @lvl: loglevel
1416 * @ai: allocation info to dump
1417 *
1418 * Print out information about @ai using loglevel @lvl.
1419 */
1420static void pcpu_dump_alloc_info(const char *lvl,
1421				 const struct pcpu_alloc_info *ai)
1422{
1423	int group_width = 1, cpu_width = 1, width;
1424	char empty_str[] = "--------";
1425	int alloc = 0, alloc_end = 0;
1426	int group, v;
1427	int upa, apl;	/* units per alloc, allocs per line */
1428
1429	v = ai->nr_groups;
1430	while (v /= 10)
1431		group_width++;
1432
1433	v = num_possible_cpus();
1434	while (v /= 10)
1435		cpu_width++;
1436	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1437
1438	upa = ai->alloc_size / ai->unit_size;
1439	width = upa * (cpu_width + 1) + group_width + 3;
1440	apl = rounddown_pow_of_two(max(60 / width, 1));
1441
1442	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1443	       lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1444	       ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1445
1446	for (group = 0; group < ai->nr_groups; group++) {
1447		const struct pcpu_group_info *gi = &ai->groups[group];
1448		int unit = 0, unit_end = 0;
1449
1450		BUG_ON(gi->nr_units % upa);
1451		for (alloc_end += gi->nr_units / upa;
1452		     alloc < alloc_end; alloc++) {
1453			if (!(alloc % apl)) {
1454				pr_cont("\n");
1455				printk("%spcpu-alloc: ", lvl);
1456			}
1457			pr_cont("[%0*d] ", group_width, group);
1458
1459			for (unit_end += upa; unit < unit_end; unit++)
1460				if (gi->cpu_map[unit] != NR_CPUS)
1461					pr_cont("%0*d ",
1462						cpu_width, gi->cpu_map[unit]);
1463				else
1464					pr_cont("%s ", empty_str);
1465		}
1466	}
1467	pr_cont("\n");
1468}
1469
1470/**
1471 * pcpu_setup_first_chunk - initialize the first percpu chunk
1472 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1473 * @base_addr: mapped address
1474 *
1475 * Initialize the first percpu chunk which contains the kernel static
1476 * perpcu area.  This function is to be called from arch percpu area
1477 * setup path.
1478 *
1479 * @ai contains all information necessary to initialize the first
1480 * chunk and prime the dynamic percpu allocator.
1481 *
1482 * @ai->static_size is the size of static percpu area.
1483 *
1484 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1485 * reserve after the static area in the first chunk.  This reserves
1486 * the first chunk such that it's available only through reserved
1487 * percpu allocation.  This is primarily used to serve module percpu
1488 * static areas on architectures where the addressing model has
1489 * limited offset range for symbol relocations to guarantee module
1490 * percpu symbols fall inside the relocatable range.
1491 *
1492 * @ai->dyn_size determines the number of bytes available for dynamic
1493 * allocation in the first chunk.  The area between @ai->static_size +
1494 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1495 *
1496 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1497 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1498 * @ai->dyn_size.
1499 *
1500 * @ai->atom_size is the allocation atom size and used as alignment
1501 * for vm areas.
1502 *
1503 * @ai->alloc_size is the allocation size and always multiple of
1504 * @ai->atom_size.  This is larger than @ai->atom_size if
1505 * @ai->unit_size is larger than @ai->atom_size.
1506 *
1507 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1508 * percpu areas.  Units which should be colocated are put into the
1509 * same group.  Dynamic VM areas will be allocated according to these
1510 * groupings.  If @ai->nr_groups is zero, a single group containing
1511 * all units is assumed.
1512 *
1513 * The caller should have mapped the first chunk at @base_addr and
1514 * copied static data to each unit.
1515 *
1516 * If the first chunk ends up with both reserved and dynamic areas, it
1517 * is served by two chunks - one to serve the core static and reserved
1518 * areas and the other for the dynamic area.  They share the same vm
1519 * and page map but uses different area allocation map to stay away
1520 * from each other.  The latter chunk is circulated in the chunk slots
1521 * and available for dynamic allocation like any other chunks.
1522 *
1523 * RETURNS:
1524 * 0 on success, -errno on failure.
1525 */
1526int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1527				  void *base_addr)
1528{
1529	static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1530	static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1531	size_t dyn_size = ai->dyn_size;
1532	size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1533	struct pcpu_chunk *schunk, *dchunk = NULL;
1534	unsigned long *group_offsets;
1535	size_t *group_sizes;
1536	unsigned long *unit_off;
1537	unsigned int cpu;
1538	int *unit_map;
1539	int group, unit, i;
 
 
 
1540
1541#define PCPU_SETUP_BUG_ON(cond)	do {					\
1542	if (unlikely(cond)) {						\
1543		pr_emerg("failed to initialize, %s\n", #cond);		\
1544		pr_emerg("cpu_possible_mask=%*pb\n",			\
1545			 cpumask_pr_args(cpu_possible_mask));		\
1546		pcpu_dump_alloc_info(KERN_EMERG, ai);			\
1547		BUG();							\
1548	}								\
1549} while (0)
1550
1551	/* sanity checks */
1552	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1553#ifdef CONFIG_SMP
1554	PCPU_SETUP_BUG_ON(!ai->static_size);
1555	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1556#endif
1557	PCPU_SETUP_BUG_ON(!base_addr);
1558	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1559	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1560	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1561	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
 
1562	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
 
 
 
 
1563	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1564
1565	/* process group information and build config tables accordingly */
1566	group_offsets = memblock_virt_alloc(ai->nr_groups *
1567					     sizeof(group_offsets[0]), 0);
1568	group_sizes = memblock_virt_alloc(ai->nr_groups *
1569					   sizeof(group_sizes[0]), 0);
1570	unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1571	unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1572
1573	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1574		unit_map[cpu] = UINT_MAX;
1575
1576	pcpu_low_unit_cpu = NR_CPUS;
1577	pcpu_high_unit_cpu = NR_CPUS;
1578
1579	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1580		const struct pcpu_group_info *gi = &ai->groups[group];
1581
1582		group_offsets[group] = gi->base_offset;
1583		group_sizes[group] = gi->nr_units * ai->unit_size;
1584
1585		for (i = 0; i < gi->nr_units; i++) {
1586			cpu = gi->cpu_map[i];
1587			if (cpu == NR_CPUS)
1588				continue;
1589
1590			PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1591			PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1592			PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1593
1594			unit_map[cpu] = unit + i;
1595			unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1596
1597			/* determine low/high unit_cpu */
1598			if (pcpu_low_unit_cpu == NR_CPUS ||
1599			    unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1600				pcpu_low_unit_cpu = cpu;
1601			if (pcpu_high_unit_cpu == NR_CPUS ||
1602			    unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1603				pcpu_high_unit_cpu = cpu;
1604		}
1605	}
1606	pcpu_nr_units = unit;
1607
1608	for_each_possible_cpu(cpu)
1609		PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1610
1611	/* we're done parsing the input, undefine BUG macro and dump config */
1612#undef PCPU_SETUP_BUG_ON
1613	pcpu_dump_alloc_info(KERN_DEBUG, ai);
1614
1615	pcpu_nr_groups = ai->nr_groups;
1616	pcpu_group_offsets = group_offsets;
1617	pcpu_group_sizes = group_sizes;
1618	pcpu_unit_map = unit_map;
1619	pcpu_unit_offsets = unit_off;
1620
1621	/* determine basic parameters */
1622	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1623	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1624	pcpu_atom_size = ai->atom_size;
1625	pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1626		BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
 
 
1627
1628	/*
1629	 * Allocate chunk slots.  The additional last slot is for
1630	 * empty chunks.
 
 
1631	 */
1632	pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1633	pcpu_slot = memblock_virt_alloc(
1634			pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
 
 
 
 
 
 
 
 
1635	for (i = 0; i < pcpu_nr_slots; i++)
1636		INIT_LIST_HEAD(&pcpu_slot[i]);
1637
1638	/*
1639	 * Initialize static chunk.  If reserved_size is zero, the
1640	 * static chunk covers static area + dynamic allocation area
1641	 * in the first chunk.  If reserved_size is not zero, it
1642	 * covers static area + reserved area (mostly used for module
1643	 * static percpu allocation).
1644	 */
1645	schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1646	INIT_LIST_HEAD(&schunk->list);
1647	INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn);
1648	schunk->base_addr = base_addr;
1649	schunk->map = smap;
1650	schunk->map_alloc = ARRAY_SIZE(smap);
1651	schunk->immutable = true;
1652	bitmap_fill(schunk->populated, pcpu_unit_pages);
1653	schunk->nr_populated = pcpu_unit_pages;
1654
1655	if (ai->reserved_size) {
1656		schunk->free_size = ai->reserved_size;
1657		pcpu_reserved_chunk = schunk;
1658		pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1659	} else {
1660		schunk->free_size = dyn_size;
1661		dyn_size = 0;			/* dynamic area covered */
1662	}
1663	schunk->contig_hint = schunk->free_size;
1664
1665	schunk->map[0] = 1;
1666	schunk->map[1] = ai->static_size;
1667	schunk->map_used = 1;
1668	if (schunk->free_size)
1669		schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1670	schunk->map[schunk->map_used] |= 1;
 
 
 
 
 
1671
1672	/* init dynamic chunk if necessary */
1673	if (dyn_size) {
1674		dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1675		INIT_LIST_HEAD(&dchunk->list);
1676		INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn);
1677		dchunk->base_addr = base_addr;
1678		dchunk->map = dmap;
1679		dchunk->map_alloc = ARRAY_SIZE(dmap);
1680		dchunk->immutable = true;
1681		bitmap_fill(dchunk->populated, pcpu_unit_pages);
1682		dchunk->nr_populated = pcpu_unit_pages;
1683
1684		dchunk->contig_hint = dchunk->free_size = dyn_size;
1685		dchunk->map[0] = 1;
1686		dchunk->map[1] = pcpu_reserved_chunk_limit;
1687		dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1688		dchunk->map_used = 2;
1689	}
1690
1691	/* link the first chunk in */
1692	pcpu_first_chunk = dchunk ?: schunk;
1693	pcpu_nr_empty_pop_pages +=
1694		pcpu_count_occupied_pages(pcpu_first_chunk, 1);
1695	pcpu_chunk_relocate(pcpu_first_chunk, -1);
1696
 
 
 
 
 
 
1697	/* we're done */
1698	pcpu_base_addr = base_addr;
1699	return 0;
1700}
1701
1702#ifdef CONFIG_SMP
1703
1704const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1705	[PCPU_FC_AUTO]	= "auto",
1706	[PCPU_FC_EMBED]	= "embed",
1707	[PCPU_FC_PAGE]	= "page",
1708};
1709
1710enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1711
1712static int __init percpu_alloc_setup(char *str)
1713{
1714	if (!str)
1715		return -EINVAL;
1716
1717	if (0)
1718		/* nada */;
1719#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1720	else if (!strcmp(str, "embed"))
1721		pcpu_chosen_fc = PCPU_FC_EMBED;
1722#endif
1723#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1724	else if (!strcmp(str, "page"))
1725		pcpu_chosen_fc = PCPU_FC_PAGE;
1726#endif
1727	else
1728		pr_warn("unknown allocator %s specified\n", str);
1729
1730	return 0;
1731}
1732early_param("percpu_alloc", percpu_alloc_setup);
1733
1734/*
1735 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1736 * Build it if needed by the arch config or the generic setup is going
1737 * to be used.
1738 */
1739#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1740	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1741#define BUILD_EMBED_FIRST_CHUNK
1742#endif
1743
1744/* build pcpu_page_first_chunk() iff needed by the arch config */
1745#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1746#define BUILD_PAGE_FIRST_CHUNK
1747#endif
1748
1749/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1750#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1751/**
1752 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1753 * @reserved_size: the size of reserved percpu area in bytes
1754 * @dyn_size: minimum free size for dynamic allocation in bytes
1755 * @atom_size: allocation atom size
1756 * @cpu_distance_fn: callback to determine distance between cpus, optional
1757 *
1758 * This function determines grouping of units, their mappings to cpus
1759 * and other parameters considering needed percpu size, allocation
1760 * atom size and distances between CPUs.
1761 *
1762 * Groups are always multiples of atom size and CPUs which are of
1763 * LOCAL_DISTANCE both ways are grouped together and share space for
1764 * units in the same group.  The returned configuration is guaranteed
1765 * to have CPUs on different nodes on different groups and >=75% usage
1766 * of allocated virtual address space.
1767 *
1768 * RETURNS:
1769 * On success, pointer to the new allocation_info is returned.  On
1770 * failure, ERR_PTR value is returned.
1771 */
1772static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1773				size_t reserved_size, size_t dyn_size,
1774				size_t atom_size,
1775				pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1776{
1777	static int group_map[NR_CPUS] __initdata;
1778	static int group_cnt[NR_CPUS] __initdata;
 
1779	const size_t static_size = __per_cpu_end - __per_cpu_start;
1780	int nr_groups = 1, nr_units = 0;
1781	size_t size_sum, min_unit_size, alloc_size;
1782	int upa, max_upa, uninitialized_var(best_upa);	/* units_per_alloc */
1783	int last_allocs, group, unit;
1784	unsigned int cpu, tcpu;
1785	struct pcpu_alloc_info *ai;
1786	unsigned int *cpu_map;
1787
1788	/* this function may be called multiple times */
1789	memset(group_map, 0, sizeof(group_map));
1790	memset(group_cnt, 0, sizeof(group_cnt));
 
1791
1792	/* calculate size_sum and ensure dyn_size is enough for early alloc */
1793	size_sum = PFN_ALIGN(static_size + reserved_size +
1794			    max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1795	dyn_size = size_sum - static_size - reserved_size;
1796
1797	/*
1798	 * Determine min_unit_size, alloc_size and max_upa such that
1799	 * alloc_size is multiple of atom_size and is the smallest
1800	 * which can accommodate 4k aligned segments which are equal to
1801	 * or larger than min_unit_size.
1802	 */
1803	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1804
 
1805	alloc_size = roundup(min_unit_size, atom_size);
1806	upa = alloc_size / min_unit_size;
1807	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1808		upa--;
1809	max_upa = upa;
1810
 
 
1811	/* group cpus according to their proximity */
1812	for_each_possible_cpu(cpu) {
1813		group = 0;
1814	next_group:
1815		for_each_possible_cpu(tcpu) {
1816			if (cpu == tcpu)
1817				break;
1818			if (group_map[tcpu] == group && cpu_distance_fn &&
1819			    (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1820			     cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1821				group++;
1822				nr_groups = max(nr_groups, group + 1);
1823				goto next_group;
1824			}
1825		}
1826		group_map[cpu] = group;
1827		group_cnt[group]++;
 
 
 
 
 
 
 
 
 
 
 
1828	}
 
1829
1830	/*
1831	 * Expand unit size until address space usage goes over 75%
1832	 * and then as much as possible without using more address
1833	 * space.
1834	 */
1835	last_allocs = INT_MAX;
 
1836	for (upa = max_upa; upa; upa--) {
1837		int allocs = 0, wasted = 0;
1838
1839		if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1840			continue;
1841
1842		for (group = 0; group < nr_groups; group++) {
1843			int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1844			allocs += this_allocs;
1845			wasted += this_allocs * upa - group_cnt[group];
1846		}
1847
1848		/*
1849		 * Don't accept if wastage is over 1/3.  The
1850		 * greater-than comparison ensures upa==1 always
1851		 * passes the following check.
1852		 */
1853		if (wasted > num_possible_cpus() / 3)
1854			continue;
1855
1856		/* and then don't consume more memory */
1857		if (allocs > last_allocs)
1858			break;
1859		last_allocs = allocs;
1860		best_upa = upa;
1861	}
 
1862	upa = best_upa;
1863
1864	/* allocate and fill alloc_info */
1865	for (group = 0; group < nr_groups; group++)
1866		nr_units += roundup(group_cnt[group], upa);
1867
1868	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1869	if (!ai)
1870		return ERR_PTR(-ENOMEM);
1871	cpu_map = ai->groups[0].cpu_map;
1872
1873	for (group = 0; group < nr_groups; group++) {
1874		ai->groups[group].cpu_map = cpu_map;
1875		cpu_map += roundup(group_cnt[group], upa);
1876	}
1877
1878	ai->static_size = static_size;
1879	ai->reserved_size = reserved_size;
1880	ai->dyn_size = dyn_size;
1881	ai->unit_size = alloc_size / upa;
1882	ai->atom_size = atom_size;
1883	ai->alloc_size = alloc_size;
1884
1885	for (group = 0, unit = 0; group_cnt[group]; group++) {
1886		struct pcpu_group_info *gi = &ai->groups[group];
1887
1888		/*
1889		 * Initialize base_offset as if all groups are located
1890		 * back-to-back.  The caller should update this to
1891		 * reflect actual allocation.
1892		 */
1893		gi->base_offset = unit * ai->unit_size;
1894
1895		for_each_possible_cpu(cpu)
1896			if (group_map[cpu] == group)
1897				gi->cpu_map[gi->nr_units++] = cpu;
1898		gi->nr_units = roundup(gi->nr_units, upa);
1899		unit += gi->nr_units;
1900	}
1901	BUG_ON(unit != nr_units);
1902
1903	return ai;
1904}
1905#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1906
1907#if defined(BUILD_EMBED_FIRST_CHUNK)
1908/**
1909 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1910 * @reserved_size: the size of reserved percpu area in bytes
1911 * @dyn_size: minimum free size for dynamic allocation in bytes
1912 * @atom_size: allocation atom size
1913 * @cpu_distance_fn: callback to determine distance between cpus, optional
1914 * @alloc_fn: function to allocate percpu page
1915 * @free_fn: function to free percpu page
1916 *
1917 * This is a helper to ease setting up embedded first percpu chunk and
1918 * can be called where pcpu_setup_first_chunk() is expected.
1919 *
1920 * If this function is used to setup the first chunk, it is allocated
1921 * by calling @alloc_fn and used as-is without being mapped into
1922 * vmalloc area.  Allocations are always whole multiples of @atom_size
1923 * aligned to @atom_size.
1924 *
1925 * This enables the first chunk to piggy back on the linear physical
1926 * mapping which often uses larger page size.  Please note that this
1927 * can result in very sparse cpu->unit mapping on NUMA machines thus
1928 * requiring large vmalloc address space.  Don't use this allocator if
1929 * vmalloc space is not orders of magnitude larger than distances
1930 * between node memory addresses (ie. 32bit NUMA machines).
1931 *
1932 * @dyn_size specifies the minimum dynamic area size.
1933 *
1934 * If the needed size is smaller than the minimum or specified unit
1935 * size, the leftover is returned using @free_fn.
1936 *
1937 * RETURNS:
1938 * 0 on success, -errno on failure.
1939 */
1940int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1941				  size_t atom_size,
1942				  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1943				  pcpu_fc_alloc_fn_t alloc_fn,
1944				  pcpu_fc_free_fn_t free_fn)
1945{
1946	void *base = (void *)ULONG_MAX;
1947	void **areas = NULL;
1948	struct pcpu_alloc_info *ai;
1949	size_t size_sum, areas_size, max_distance;
1950	int group, i, rc;
 
1951
1952	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1953				   cpu_distance_fn);
1954	if (IS_ERR(ai))
1955		return PTR_ERR(ai);
1956
1957	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1958	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1959
1960	areas = memblock_virt_alloc_nopanic(areas_size, 0);
1961	if (!areas) {
1962		rc = -ENOMEM;
1963		goto out_free;
1964	}
1965
1966	/* allocate, copy and determine base address */
 
1967	for (group = 0; group < ai->nr_groups; group++) {
1968		struct pcpu_group_info *gi = &ai->groups[group];
1969		unsigned int cpu = NR_CPUS;
1970		void *ptr;
1971
1972		for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1973			cpu = gi->cpu_map[i];
1974		BUG_ON(cpu == NR_CPUS);
1975
1976		/* allocate space for the whole group */
1977		ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1978		if (!ptr) {
1979			rc = -ENOMEM;
1980			goto out_free_areas;
1981		}
1982		/* kmemleak tracks the percpu allocations separately */
1983		kmemleak_free(ptr);
1984		areas[group] = ptr;
1985
1986		base = min(ptr, base);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1987	}
1988
1989	/*
1990	 * Copy data and free unused parts.  This should happen after all
1991	 * allocations are complete; otherwise, we may end up with
1992	 * overlapping groups.
1993	 */
1994	for (group = 0; group < ai->nr_groups; group++) {
1995		struct pcpu_group_info *gi = &ai->groups[group];
1996		void *ptr = areas[group];
1997
1998		for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1999			if (gi->cpu_map[i] == NR_CPUS) {
2000				/* unused unit, free whole */
2001				free_fn(ptr, ai->unit_size);
2002				continue;
2003			}
2004			/* copy and return the unused part */
2005			memcpy(ptr, __per_cpu_load, ai->static_size);
2006			free_fn(ptr + size_sum, ai->unit_size - size_sum);
2007		}
2008	}
2009
2010	/* base address is now known, determine group base offsets */
2011	max_distance = 0;
2012	for (group = 0; group < ai->nr_groups; group++) {
2013		ai->groups[group].base_offset = areas[group] - base;
2014		max_distance = max_t(size_t, max_distance,
2015				     ai->groups[group].base_offset);
2016	}
2017	max_distance += ai->unit_size;
2018
2019	/* warn if maximum distance is further than 75% of vmalloc space */
2020	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2021		pr_warn("max_distance=0x%zx too large for vmalloc space 0x%lx\n",
2022			max_distance, VMALLOC_TOTAL);
2023#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2024		/* and fail if we have fallback */
2025		rc = -EINVAL;
2026		goto out_free;
2027#endif
2028	}
2029
2030	pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2031		PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2032		ai->dyn_size, ai->unit_size);
2033
2034	rc = pcpu_setup_first_chunk(ai, base);
2035	goto out_free;
2036
2037out_free_areas:
2038	for (group = 0; group < ai->nr_groups; group++)
2039		if (areas[group])
2040			free_fn(areas[group],
2041				ai->groups[group].nr_units * ai->unit_size);
2042out_free:
2043	pcpu_free_alloc_info(ai);
2044	if (areas)
2045		memblock_free_early(__pa(areas), areas_size);
2046	return rc;
2047}
2048#endif /* BUILD_EMBED_FIRST_CHUNK */
2049
2050#ifdef BUILD_PAGE_FIRST_CHUNK
2051/**
2052 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2053 * @reserved_size: the size of reserved percpu area in bytes
2054 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2055 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2056 * @populate_pte_fn: function to populate pte
2057 *
2058 * This is a helper to ease setting up page-remapped first percpu
2059 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2060 *
2061 * This is the basic allocator.  Static percpu area is allocated
2062 * page-by-page into vmalloc area.
2063 *
2064 * RETURNS:
2065 * 0 on success, -errno on failure.
2066 */
2067int __init pcpu_page_first_chunk(size_t reserved_size,
2068				 pcpu_fc_alloc_fn_t alloc_fn,
2069				 pcpu_fc_free_fn_t free_fn,
2070				 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2071{
2072	static struct vm_struct vm;
2073	struct pcpu_alloc_info *ai;
2074	char psize_str[16];
2075	int unit_pages;
2076	size_t pages_size;
2077	struct page **pages;
2078	int unit, i, j, rc;
 
 
2079
2080	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2081
2082	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2083	if (IS_ERR(ai))
2084		return PTR_ERR(ai);
2085	BUG_ON(ai->nr_groups != 1);
2086	BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
 
 
 
 
 
2087
2088	unit_pages = ai->unit_size >> PAGE_SHIFT;
2089
2090	/* unaligned allocations can't be freed, round up to page size */
2091	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2092			       sizeof(pages[0]));
2093	pages = memblock_virt_alloc(pages_size, 0);
 
 
 
2094
2095	/* allocate pages */
2096	j = 0;
2097	for (unit = 0; unit < num_possible_cpus(); unit++)
 
2098		for (i = 0; i < unit_pages; i++) {
2099			unsigned int cpu = ai->groups[0].cpu_map[unit];
2100			void *ptr;
2101
2102			ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2103			if (!ptr) {
2104				pr_warn("failed to allocate %s page for cpu%u\n",
2105					psize_str, cpu);
2106				goto enomem;
2107			}
2108			/* kmemleak tracks the percpu allocations separately */
2109			kmemleak_free(ptr);
2110			pages[j++] = virt_to_page(ptr);
2111		}
 
2112
2113	/* allocate vm area, map the pages and copy static data */
2114	vm.flags = VM_ALLOC;
2115	vm.size = num_possible_cpus() * ai->unit_size;
2116	vm_area_register_early(&vm, PAGE_SIZE);
2117
2118	for (unit = 0; unit < num_possible_cpus(); unit++) {
2119		unsigned long unit_addr =
2120			(unsigned long)vm.addr + unit * ai->unit_size;
2121
2122		for (i = 0; i < unit_pages; i++)
2123			populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2124
2125		/* pte already populated, the following shouldn't fail */
2126		rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2127				      unit_pages);
2128		if (rc < 0)
2129			panic("failed to map percpu area, err=%d\n", rc);
2130
2131		/*
2132		 * FIXME: Archs with virtual cache should flush local
2133		 * cache for the linear mapping here - something
2134		 * equivalent to flush_cache_vmap() on the local cpu.
2135		 * flush_cache_vmap() can't be used as most supporting
2136		 * data structures are not set up yet.
2137		 */
2138
2139		/* copy static data */
2140		memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2141	}
2142
2143	/* we're ready, commit */
2144	pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2145		unit_pages, psize_str, vm.addr, ai->static_size,
2146		ai->reserved_size, ai->dyn_size);
2147
2148	rc = pcpu_setup_first_chunk(ai, vm.addr);
2149	goto out_free_ar;
2150
2151enomem:
2152	while (--j >= 0)
2153		free_fn(page_address(pages[j]), PAGE_SIZE);
2154	rc = -ENOMEM;
2155out_free_ar:
2156	memblock_free_early(__pa(pages), pages_size);
2157	pcpu_free_alloc_info(ai);
2158	return rc;
2159}
2160#endif /* BUILD_PAGE_FIRST_CHUNK */
2161
2162#ifndef	CONFIG_HAVE_SETUP_PER_CPU_AREA
2163/*
2164 * Generic SMP percpu area setup.
2165 *
2166 * The embedding helper is used because its behavior closely resembles
2167 * the original non-dynamic generic percpu area setup.  This is
2168 * important because many archs have addressing restrictions and might
2169 * fail if the percpu area is located far away from the previous
2170 * location.  As an added bonus, in non-NUMA cases, embedding is
2171 * generally a good idea TLB-wise because percpu area can piggy back
2172 * on the physical linear memory mapping which uses large page
2173 * mappings on applicable archs.
2174 */
2175unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2176EXPORT_SYMBOL(__per_cpu_offset);
2177
2178static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2179				       size_t align)
2180{
2181	return  memblock_virt_alloc_from_nopanic(
2182			size, align, __pa(MAX_DMA_ADDRESS));
2183}
2184
2185static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2186{
2187	memblock_free_early(__pa(ptr), size);
2188}
2189
2190void __init setup_per_cpu_areas(void)
2191{
2192	unsigned long delta;
2193	unsigned int cpu;
2194	int rc;
2195
2196	/*
2197	 * Always reserve area for module percpu variables.  That's
2198	 * what the legacy allocator did.
2199	 */
2200	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2201				    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2202				    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2203	if (rc < 0)
2204		panic("Failed to initialize percpu areas.");
2205
2206	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2207	for_each_possible_cpu(cpu)
2208		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2209}
2210#endif	/* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2211
2212#else	/* CONFIG_SMP */
2213
2214/*
2215 * UP percpu area setup.
2216 *
2217 * UP always uses km-based percpu allocator with identity mapping.
2218 * Static percpu variables are indistinguishable from the usual static
2219 * variables and don't require any special preparation.
2220 */
2221void __init setup_per_cpu_areas(void)
2222{
2223	const size_t unit_size =
2224		roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2225					 PERCPU_DYNAMIC_RESERVE));
2226	struct pcpu_alloc_info *ai;
2227	void *fc;
2228
2229	ai = pcpu_alloc_alloc_info(1, 1);
2230	fc = memblock_virt_alloc_from_nopanic(unit_size,
2231					      PAGE_SIZE,
2232					      __pa(MAX_DMA_ADDRESS));
2233	if (!ai || !fc)
2234		panic("Failed to allocate memory for percpu areas.");
2235	/* kmemleak tracks the percpu allocations separately */
2236	kmemleak_free(fc);
2237
2238	ai->dyn_size = unit_size;
2239	ai->unit_size = unit_size;
2240	ai->atom_size = unit_size;
2241	ai->alloc_size = unit_size;
2242	ai->groups[0].nr_units = 1;
2243	ai->groups[0].cpu_map[0] = 0;
2244
2245	if (pcpu_setup_first_chunk(ai, fc) < 0)
2246		panic("Failed to initialize percpu areas.");
2247}
2248
2249#endif	/* CONFIG_SMP */
2250
2251/*
2252 * First and reserved chunks are initialized with temporary allocation
2253 * map in initdata so that they can be used before slab is online.
2254 * This function is called after slab is brought up and replaces those
2255 * with properly allocated maps.
 
 
 
 
 
2256 */
2257void __init percpu_init_late(void)
2258{
2259	struct pcpu_chunk *target_chunks[] =
2260		{ pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2261	struct pcpu_chunk *chunk;
2262	unsigned long flags;
2263	int i;
2264
2265	for (i = 0; (chunk = target_chunks[i]); i++) {
2266		int *map;
2267		const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2268
2269		BUILD_BUG_ON(size > PAGE_SIZE);
2270
2271		map = pcpu_mem_zalloc(size);
2272		BUG_ON(!map);
2273
2274		spin_lock_irqsave(&pcpu_lock, flags);
2275		memcpy(map, chunk->map, size);
2276		chunk->map = map;
2277		spin_unlock_irqrestore(&pcpu_lock, flags);
2278	}
2279}
2280
2281/*
2282 * Percpu allocator is initialized early during boot when neither slab or
2283 * workqueue is available.  Plug async management until everything is up
2284 * and running.
2285 */
2286static int __init percpu_enable_async(void)
2287{
2288	pcpu_async_enabled = true;
2289	return 0;
2290}
2291subsys_initcall(percpu_enable_async);
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * mm/percpu.c - percpu memory allocator
   4 *
   5 * Copyright (C) 2009		SUSE Linux Products GmbH
   6 * Copyright (C) 2009		Tejun Heo <tj@kernel.org>
   7 *
   8 * Copyright (C) 2017		Facebook Inc.
   9 * Copyright (C) 2017		Dennis Zhou <dennis@kernel.org>
  10 *
  11 * The percpu allocator handles both static and dynamic areas.  Percpu
  12 * areas are allocated in chunks which are divided into units.  There is
  13 * a 1-to-1 mapping for units to possible cpus.  These units are grouped
  14 * based on NUMA properties of the machine.
 
 
 
 
  15 *
  16 *  c0                           c1                         c2
  17 *  -------------------          -------------------        ------------
  18 * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
  19 *  -------------------  ......  -------------------  ....  ------------
  20 *
  21 * Allocation is done by offsets into a unit's address space.  Ie., an
  22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
  23 * c1:u1, c1:u2, etc.  On NUMA machines, the mapping may be non-linear
  24 * and even sparse.  Access is handled by configuring percpu base
  25 * registers according to the cpu to unit mappings and offsetting the
  26 * base address using pcpu_unit_size.
  27 *
  28 * There is special consideration for the first chunk which must handle
  29 * the static percpu variables in the kernel image as allocation services
  30 * are not online yet.  In short, the first chunk is structured like so:
  31 *
  32 *                  <Static | [Reserved] | Dynamic>
  33 *
  34 * The static data is copied from the original section managed by the
  35 * linker.  The reserved section, if non-zero, primarily manages static
  36 * percpu variables from kernel modules.  Finally, the dynamic section
  37 * takes care of normal allocations.
  38 *
  39 * The allocator organizes chunks into lists according to free size and
  40 * memcg-awareness.  To make a percpu allocation memcg-aware the __GFP_ACCOUNT
  41 * flag should be passed.  All memcg-aware allocations are sharing one set
  42 * of chunks and all unaccounted allocations and allocations performed
  43 * by processes belonging to the root memory cgroup are using the second set.
  44 *
  45 * The allocator tries to allocate from the fullest chunk first. Each chunk
  46 * is managed by a bitmap with metadata blocks.  The allocation map is updated
  47 * on every allocation and free to reflect the current state while the boundary
  48 * map is only updated on allocation.  Each metadata block contains
  49 * information to help mitigate the need to iterate over large portions
  50 * of the bitmap.  The reverse mapping from page to chunk is stored in
  51 * the page's index.  Lastly, units are lazily backed and grow in unison.
  52 *
  53 * There is a unique conversion that goes on here between bytes and bits.
  54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE.  The chunk
  55 * tracks the number of pages it is responsible for in nr_pages.  Helper
  56 * functions are used to convert from between the bytes, bits, and blocks.
  57 * All hints are managed in bits unless explicitly stated.
  58 *
  59 * To use this allocator, arch code should do the following:
  60 *
  61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
  62 *   regular address to percpu pointer and back if they need to be
  63 *   different from the default
  64 *
  65 * - use pcpu_setup_first_chunk() during percpu area initialization to
  66 *   setup the first chunk containing the kernel static percpu area
  67 */
  68
  69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  70
  71#include <linux/bitmap.h>
  72#include <linux/cpumask.h>
  73#include <linux/memblock.h>
  74#include <linux/err.h>
  75#include <linux/lcm.h>
  76#include <linux/list.h>
  77#include <linux/log2.h>
  78#include <linux/mm.h>
  79#include <linux/module.h>
  80#include <linux/mutex.h>
  81#include <linux/percpu.h>
  82#include <linux/pfn.h>
  83#include <linux/slab.h>
  84#include <linux/spinlock.h>
  85#include <linux/vmalloc.h>
  86#include <linux/workqueue.h>
  87#include <linux/kmemleak.h>
  88#include <linux/sched.h>
  89#include <linux/sched/mm.h>
  90#include <linux/memcontrol.h>
  91
  92#include <asm/cacheflush.h>
  93#include <asm/sections.h>
  94#include <asm/tlbflush.h>
  95#include <asm/io.h>
  96
  97#define CREATE_TRACE_POINTS
  98#include <trace/events/percpu.h>
  99
 100#include "percpu-internal.h"
 101
 102/*
 103 * The slots are sorted by the size of the biggest continuous free area.
 104 * 1-31 bytes share the same slot.
 105 */
 106#define PCPU_SLOT_BASE_SHIFT		5
 107/* chunks in slots below this are subject to being sidelined on failed alloc */
 108#define PCPU_SLOT_FAIL_THRESHOLD	3
 109
 110#define PCPU_EMPTY_POP_PAGES_LOW	2
 111#define PCPU_EMPTY_POP_PAGES_HIGH	4
 112
 113#ifdef CONFIG_SMP
 114/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 115#ifndef __addr_to_pcpu_ptr
 116#define __addr_to_pcpu_ptr(addr)					\
 117	(void __percpu *)((unsigned long)(addr) -			\
 118			  (unsigned long)pcpu_base_addr	+		\
 119			  (unsigned long)__per_cpu_start)
 120#endif
 121#ifndef __pcpu_ptr_to_addr
 122#define __pcpu_ptr_to_addr(ptr)						\
 123	(void __force *)((unsigned long)(ptr) +				\
 124			 (unsigned long)pcpu_base_addr -		\
 125			 (unsigned long)__per_cpu_start)
 126#endif
 127#else	/* CONFIG_SMP */
 128/* on UP, it's always identity mapped */
 129#define __addr_to_pcpu_ptr(addr)	(void __percpu *)(addr)
 130#define __pcpu_ptr_to_addr(ptr)		(void __force *)(ptr)
 131#endif	/* CONFIG_SMP */
 132
 133static int pcpu_unit_pages __ro_after_init;
 134static int pcpu_unit_size __ro_after_init;
 135static int pcpu_nr_units __ro_after_init;
 136static int pcpu_atom_size __ro_after_init;
 137int pcpu_nr_slots __ro_after_init;
 138static int pcpu_free_slot __ro_after_init;
 139int pcpu_sidelined_slot __ro_after_init;
 140int pcpu_to_depopulate_slot __ro_after_init;
 141static size_t pcpu_chunk_struct_size __ro_after_init;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 142
 143/* cpus with the lowest and highest unit addresses */
 144static unsigned int pcpu_low_unit_cpu __ro_after_init;
 145static unsigned int pcpu_high_unit_cpu __ro_after_init;
 146
 147/* the address of the first chunk which starts with the kernel static area */
 148void *pcpu_base_addr __ro_after_init;
 149EXPORT_SYMBOL_GPL(pcpu_base_addr);
 150
 151static const int *pcpu_unit_map __ro_after_init;		/* cpu -> unit */
 152const unsigned long *pcpu_unit_offsets __ro_after_init;	/* cpu -> unit offset */
 153
 154/* group information, used for vm allocation */
 155static int pcpu_nr_groups __ro_after_init;
 156static const unsigned long *pcpu_group_offsets __ro_after_init;
 157static const size_t *pcpu_group_sizes __ro_after_init;
 158
 159/*
 160 * The first chunk which always exists.  Note that unlike other
 161 * chunks, this one can be allocated and mapped in several different
 162 * ways and thus often doesn't live in the vmalloc area.
 163 */
 164struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
 165
 166/*
 167 * Optional reserved chunk.  This chunk reserves part of the first
 168 * chunk and serves it for reserved allocations.  When the reserved
 169 * region doesn't exist, the following variable is NULL.
 
 
 170 */
 171struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
 172
 173DEFINE_SPINLOCK(pcpu_lock);	/* all internal data structures */
 174static DEFINE_MUTEX(pcpu_alloc_mutex);	/* chunk create/destroy, [de]pop, map ext */
 175
 176struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
 177
 178/* chunks which need their map areas extended, protected by pcpu_lock */
 179static LIST_HEAD(pcpu_map_extend_chunks);
 180
 181/*
 182 * The number of empty populated pages, protected by pcpu_lock.
 183 * The reserved chunk doesn't contribute to the count.
 184 */
 185int pcpu_nr_empty_pop_pages;
 186
 187/*
 188 * The number of populated pages in use by the allocator, protected by
 189 * pcpu_lock.  This number is kept per a unit per chunk (i.e. when a page gets
 190 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
 191 * and increments/decrements this count by 1).
 192 */
 193static unsigned long pcpu_nr_populated;
 194
 195/*
 196 * Balance work is used to populate or destroy chunks asynchronously.  We
 197 * try to keep the number of populated free pages between
 198 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
 199 * empty chunk.
 200 */
 201static void pcpu_balance_workfn(struct work_struct *work);
 202static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
 203static bool pcpu_async_enabled __read_mostly;
 204static bool pcpu_atomic_alloc_failed;
 205
 206static void pcpu_schedule_balance_work(void)
 207{
 208	if (pcpu_async_enabled)
 209		schedule_work(&pcpu_balance_work);
 210}
 211
 212/**
 213 * pcpu_addr_in_chunk - check if the address is served from this chunk
 214 * @chunk: chunk of interest
 215 * @addr: percpu address
 216 *
 217 * RETURNS:
 218 * True if the address is served from this chunk.
 219 */
 220static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
 221{
 222	void *start_addr, *end_addr;
 223
 224	if (!chunk)
 225		return false;
 226
 227	start_addr = chunk->base_addr + chunk->start_offset;
 228	end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
 229		   chunk->end_offset;
 230
 231	return addr >= start_addr && addr < end_addr;
 
 232}
 233
 234static int __pcpu_size_to_slot(int size)
 235{
 236	int highbit = fls(size);	/* size is in bytes */
 237	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
 238}
 239
 240static int pcpu_size_to_slot(int size)
 241{
 242	if (size == pcpu_unit_size)
 243		return pcpu_free_slot;
 244	return __pcpu_size_to_slot(size);
 245}
 246
 247static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
 248{
 249	const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 250
 251	if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
 252	    chunk_md->contig_hint == 0)
 253		return 0;
 254
 255	return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
 256}
 257
 258/* set the pointer to a chunk in a page struct */
 259static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 260{
 261	page->index = (unsigned long)pcpu;
 262}
 263
 264/* obtain pointer to a chunk from a page struct */
 265static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
 266{
 267	return (struct pcpu_chunk *)page->index;
 268}
 269
 270static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
 271{
 272	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
 273}
 274
 275static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
 276{
 277	return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
 278}
 279
 280static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 281				     unsigned int cpu, int page_idx)
 282{
 283	return (unsigned long)chunk->base_addr +
 284	       pcpu_unit_page_offset(cpu, page_idx);
 285}
 286
 287/*
 288 * The following are helper functions to help access bitmaps and convert
 289 * between bitmap offsets to address offsets.
 290 */
 291static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
 292{
 293	return chunk->alloc_map +
 294	       (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
 295}
 296
 297static unsigned long pcpu_off_to_block_index(int off)
 
 298{
 299	return off / PCPU_BITMAP_BLOCK_BITS;
 
 300}
 301
 302static unsigned long pcpu_off_to_block_off(int off)
 
 303{
 304	return off & (PCPU_BITMAP_BLOCK_BITS - 1);
 305}
 306
 307static unsigned long pcpu_block_off_to_off(int index, int off)
 308{
 309	return index * PCPU_BITMAP_BLOCK_BITS + off;
 310}
 311
 312/**
 313 * pcpu_check_block_hint - check against the contig hint
 314 * @block: block of interest
 315 * @bits: size of allocation
 316 * @align: alignment of area (max PAGE_SIZE)
 317 *
 318 * Check to see if the allocation can fit in the block's contig hint.
 319 * Note, a chunk uses the same hints as a block so this can also check against
 320 * the chunk's contig hint.
 321 */
 322static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
 323				  size_t align)
 324{
 325	int bit_off = ALIGN(block->contig_hint_start, align) -
 326		block->contig_hint_start;
 327
 328	return bit_off + bits <= block->contig_hint;
 329}
 330
 331/*
 332 * pcpu_next_hint - determine which hint to use
 333 * @block: block of interest
 334 * @alloc_bits: size of allocation
 335 *
 336 * This determines if we should scan based on the scan_hint or first_free.
 337 * In general, we want to scan from first_free to fulfill allocations by
 338 * first fit.  However, if we know a scan_hint at position scan_hint_start
 339 * cannot fulfill an allocation, we can begin scanning from there knowing
 340 * the contig_hint will be our fallback.
 341 */
 342static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
 343{
 344	/*
 345	 * The three conditions below determine if we can skip past the
 346	 * scan_hint.  First, does the scan hint exist.  Second, is the
 347	 * contig_hint after the scan_hint (possibly not true iff
 348	 * contig_hint == scan_hint).  Third, is the allocation request
 349	 * larger than the scan_hint.
 350	 */
 351	if (block->scan_hint &&
 352	    block->contig_hint_start > block->scan_hint_start &&
 353	    alloc_bits > block->scan_hint)
 354		return block->scan_hint_start + block->scan_hint;
 355
 356	return block->first_free;
 357}
 358
 359/**
 360 * pcpu_next_md_free_region - finds the next hint free area
 361 * @chunk: chunk of interest
 362 * @bit_off: chunk offset
 363 * @bits: size of free area
 364 *
 365 * Helper function for pcpu_for_each_md_free_region.  It checks
 366 * block->contig_hint and performs aggregation across blocks to find the
 367 * next hint.  It modifies bit_off and bits in-place to be consumed in the
 368 * loop.
 369 */
 370static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
 371				     int *bits)
 372{
 373	int i = pcpu_off_to_block_index(*bit_off);
 374	int block_off = pcpu_off_to_block_off(*bit_off);
 375	struct pcpu_block_md *block;
 376
 377	*bits = 0;
 378	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 379	     block++, i++) {
 380		/* handles contig area across blocks */
 381		if (*bits) {
 382			*bits += block->left_free;
 383			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 384				continue;
 385			return;
 386		}
 387
 388		/*
 389		 * This checks three things.  First is there a contig_hint to
 390		 * check.  Second, have we checked this hint before by
 391		 * comparing the block_off.  Third, is this the same as the
 392		 * right contig hint.  In the last case, it spills over into
 393		 * the next block and should be handled by the contig area
 394		 * across blocks code.
 395		 */
 396		*bits = block->contig_hint;
 397		if (*bits && block->contig_hint_start >= block_off &&
 398		    *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
 399			*bit_off = pcpu_block_off_to_off(i,
 400					block->contig_hint_start);
 401			return;
 402		}
 403		/* reset to satisfy the second predicate above */
 404		block_off = 0;
 405
 406		*bits = block->right_free;
 407		*bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
 408	}
 409}
 410
 411/**
 412 * pcpu_next_fit_region - finds fit areas for a given allocation request
 413 * @chunk: chunk of interest
 414 * @alloc_bits: size of allocation
 415 * @align: alignment of area (max PAGE_SIZE)
 416 * @bit_off: chunk offset
 417 * @bits: size of free area
 418 *
 419 * Finds the next free region that is viable for use with a given size and
 420 * alignment.  This only returns if there is a valid area to be used for this
 421 * allocation.  block->first_free is returned if the allocation request fits
 422 * within the block to see if the request can be fulfilled prior to the contig
 423 * hint.
 424 */
 425static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
 426				 int align, int *bit_off, int *bits)
 427{
 428	int i = pcpu_off_to_block_index(*bit_off);
 429	int block_off = pcpu_off_to_block_off(*bit_off);
 430	struct pcpu_block_md *block;
 431
 432	*bits = 0;
 433	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 434	     block++, i++) {
 435		/* handles contig area across blocks */
 436		if (*bits) {
 437			*bits += block->left_free;
 438			if (*bits >= alloc_bits)
 439				return;
 440			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 441				continue;
 442		}
 443
 444		/* check block->contig_hint */
 445		*bits = ALIGN(block->contig_hint_start, align) -
 446			block->contig_hint_start;
 447		/*
 448		 * This uses the block offset to determine if this has been
 449		 * checked in the prior iteration.
 450		 */
 451		if (block->contig_hint &&
 452		    block->contig_hint_start >= block_off &&
 453		    block->contig_hint >= *bits + alloc_bits) {
 454			int start = pcpu_next_hint(block, alloc_bits);
 455
 456			*bits += alloc_bits + block->contig_hint_start -
 457				 start;
 458			*bit_off = pcpu_block_off_to_off(i, start);
 459			return;
 460		}
 461		/* reset to satisfy the second predicate above */
 462		block_off = 0;
 463
 464		*bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
 465				 align);
 466		*bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
 467		*bit_off = pcpu_block_off_to_off(i, *bit_off);
 468		if (*bits >= alloc_bits)
 469			return;
 470	}
 471
 472	/* no valid offsets were found - fail condition */
 473	*bit_off = pcpu_chunk_map_bits(chunk);
 474}
 475
 476/*
 477 * Metadata free area iterators.  These perform aggregation of free areas
 478 * based on the metadata blocks and return the offset @bit_off and size in
 479 * bits of the free area @bits.  pcpu_for_each_fit_region only returns when
 480 * a fit is found for the allocation request.
 481 */
 482#define pcpu_for_each_md_free_region(chunk, bit_off, bits)		\
 483	for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits));	\
 484	     (bit_off) < pcpu_chunk_map_bits((chunk));			\
 485	     (bit_off) += (bits) + 1,					\
 486	     pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
 487
 488#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits)     \
 489	for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 490				  &(bits));				      \
 491	     (bit_off) < pcpu_chunk_map_bits((chunk));			      \
 492	     (bit_off) += (bits),					      \
 493	     pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 494				  &(bits)))
 495
 496/**
 497 * pcpu_mem_zalloc - allocate memory
 498 * @size: bytes to allocate
 499 * @gfp: allocation flags
 500 *
 501 * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
 502 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
 503 * This is to facilitate passing through whitelisted flags.  The
 504 * returned memory is always zeroed.
 
 
 505 *
 506 * RETURNS:
 507 * Pointer to the allocated area on success, NULL on failure.
 508 */
 509static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
 510{
 511	if (WARN_ON_ONCE(!slab_is_available()))
 512		return NULL;
 513
 514	if (size <= PAGE_SIZE)
 515		return kzalloc(size, gfp);
 516	else
 517		return __vmalloc(size, gfp | __GFP_ZERO);
 518}
 519
 520/**
 521 * pcpu_mem_free - free memory
 522 * @ptr: memory to free
 523 *
 524 * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
 525 */
 526static void pcpu_mem_free(void *ptr)
 527{
 528	kvfree(ptr);
 529}
 530
 531static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
 532			      bool move_front)
 
 
 
 
 
 
 
 
 533{
 534	if (chunk != pcpu_reserved_chunk) {
 535		if (move_front)
 536			list_move(&chunk->list, &pcpu_chunk_lists[slot]);
 537		else
 538			list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
 
 
 
 
 
 
 
 
 
 
 
 539	}
 540}
 541
 542static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
 543{
 544	__pcpu_chunk_move(chunk, slot, true);
 545}
 546
 547/**
 548 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
 549 * @chunk: chunk of interest
 550 * @oslot: the previous slot it was on
 551 *
 552 * This function is called after an allocation or free changed @chunk.
 553 * New slot according to the changed state is determined and @chunk is
 554 * moved to the slot.  Note that the reserved chunk is never put on
 555 * chunk slots.
 556 *
 557 * CONTEXT:
 558 * pcpu_lock.
 559 */
 560static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 561{
 562	int nslot = pcpu_chunk_slot(chunk);
 563
 564	/* leave isolated chunks in-place */
 565	if (chunk->isolated)
 566		return;
 567
 568	if (oslot != nslot)
 569		__pcpu_chunk_move(chunk, nslot, oslot < nslot);
 570}
 571
 572static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
 573{
 574	lockdep_assert_held(&pcpu_lock);
 575
 576	if (!chunk->isolated) {
 577		chunk->isolated = true;
 578		pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
 579	}
 580	list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
 581}
 582
 583static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
 584{
 585	lockdep_assert_held(&pcpu_lock);
 586
 587	if (chunk->isolated) {
 588		chunk->isolated = false;
 589		pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
 590		pcpu_chunk_relocate(chunk, -1);
 591	}
 592}
 593
 594/*
 595 * pcpu_update_empty_pages - update empty page counters
 596 * @chunk: chunk of interest
 597 * @nr: nr of empty pages
 598 *
 599 * This is used to keep track of the empty pages now based on the premise
 600 * a md_block covers a page.  The hint update functions recognize if a block
 601 * is made full or broken to calculate deltas for keeping track of free pages.
 602 */
 603static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
 604{
 605	chunk->nr_empty_pop_pages += nr;
 606	if (chunk != pcpu_reserved_chunk && !chunk->isolated)
 607		pcpu_nr_empty_pop_pages += nr;
 608}
 609
 610/*
 611 * pcpu_region_overlap - determines if two regions overlap
 612 * @a: start of first region, inclusive
 613 * @b: end of first region, exclusive
 614 * @x: start of second region, inclusive
 615 * @y: end of second region, exclusive
 616 *
 617 * This is used to determine if the hint region [a, b) overlaps with the
 618 * allocated region [x, y).
 619 */
 620static inline bool pcpu_region_overlap(int a, int b, int x, int y)
 621{
 622	return (a < y) && (x < b);
 623}
 624
 625/**
 626 * pcpu_block_update - updates a block given a free area
 627 * @block: block of interest
 628 * @start: start offset in block
 629 * @end: end offset in block
 630 *
 631 * Updates a block given a known free area.  The region [start, end) is
 632 * expected to be the entirety of the free area within a block.  Chooses
 633 * the best starting offset if the contig hints are equal.
 634 */
 635static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
 636{
 637	int contig = end - start;
 638
 639	block->first_free = min(block->first_free, start);
 640	if (start == 0)
 641		block->left_free = contig;
 642
 643	if (end == block->nr_bits)
 644		block->right_free = contig;
 645
 646	if (contig > block->contig_hint) {
 647		/* promote the old contig_hint to be the new scan_hint */
 648		if (start > block->contig_hint_start) {
 649			if (block->contig_hint > block->scan_hint) {
 650				block->scan_hint_start =
 651					block->contig_hint_start;
 652				block->scan_hint = block->contig_hint;
 653			} else if (start < block->scan_hint_start) {
 654				/*
 655				 * The old contig_hint == scan_hint.  But, the
 656				 * new contig is larger so hold the invariant
 657				 * scan_hint_start < contig_hint_start.
 658				 */
 659				block->scan_hint = 0;
 660			}
 661		} else {
 662			block->scan_hint = 0;
 663		}
 664		block->contig_hint_start = start;
 665		block->contig_hint = contig;
 666	} else if (contig == block->contig_hint) {
 667		if (block->contig_hint_start &&
 668		    (!start ||
 669		     __ffs(start) > __ffs(block->contig_hint_start))) {
 670			/* start has a better alignment so use it */
 671			block->contig_hint_start = start;
 672			if (start < block->scan_hint_start &&
 673			    block->contig_hint > block->scan_hint)
 674				block->scan_hint = 0;
 675		} else if (start > block->scan_hint_start ||
 676			   block->contig_hint > block->scan_hint) {
 677			/*
 678			 * Knowing contig == contig_hint, update the scan_hint
 679			 * if it is farther than or larger than the current
 680			 * scan_hint.
 681			 */
 682			block->scan_hint_start = start;
 683			block->scan_hint = contig;
 684		}
 685	} else {
 686		/*
 687		 * The region is smaller than the contig_hint.  So only update
 688		 * the scan_hint if it is larger than or equal and farther than
 689		 * the current scan_hint.
 690		 */
 691		if ((start < block->contig_hint_start &&
 692		     (contig > block->scan_hint ||
 693		      (contig == block->scan_hint &&
 694		       start > block->scan_hint_start)))) {
 695			block->scan_hint_start = start;
 696			block->scan_hint = contig;
 697		}
 698	}
 699}
 700
 701/*
 702 * pcpu_block_update_scan - update a block given a free area from a scan
 703 * @chunk: chunk of interest
 704 * @bit_off: chunk offset
 705 * @bits: size of free area
 706 *
 707 * Finding the final allocation spot first goes through pcpu_find_block_fit()
 708 * to find a block that can hold the allocation and then pcpu_alloc_area()
 709 * where a scan is used.  When allocations require specific alignments,
 710 * we can inadvertently create holes which will not be seen in the alloc
 711 * or free paths.
 712 *
 713 * This takes a given free area hole and updates a block as it may change the
 714 * scan_hint.  We need to scan backwards to ensure we don't miss free bits
 715 * from alignment.
 716 */
 717static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
 718				   int bits)
 719{
 720	int s_off = pcpu_off_to_block_off(bit_off);
 721	int e_off = s_off + bits;
 722	int s_index, l_bit;
 723	struct pcpu_block_md *block;
 724
 725	if (e_off > PCPU_BITMAP_BLOCK_BITS)
 726		return;
 727
 728	s_index = pcpu_off_to_block_index(bit_off);
 729	block = chunk->md_blocks + s_index;
 730
 731	/* scan backwards in case of alignment skipping free bits */
 732	l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
 733	s_off = (s_off == l_bit) ? 0 : l_bit + 1;
 734
 735	pcpu_block_update(block, s_off, e_off);
 736}
 737
 738/**
 739 * pcpu_chunk_refresh_hint - updates metadata about a chunk
 740 * @chunk: chunk of interest
 741 * @full_scan: if we should scan from the beginning
 742 *
 743 * Iterates over the metadata blocks to find the largest contig area.
 744 * A full scan can be avoided on the allocation path as this is triggered
 745 * if we broke the contig_hint.  In doing so, the scan_hint will be before
 746 * the contig_hint or after if the scan_hint == contig_hint.  This cannot
 747 * be prevented on freeing as we want to find the largest area possibly
 748 * spanning blocks.
 749 */
 750static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
 751{
 752	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 753	int bit_off, bits;
 754
 755	/* promote scan_hint to contig_hint */
 756	if (!full_scan && chunk_md->scan_hint) {
 757		bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
 758		chunk_md->contig_hint_start = chunk_md->scan_hint_start;
 759		chunk_md->contig_hint = chunk_md->scan_hint;
 760		chunk_md->scan_hint = 0;
 761	} else {
 762		bit_off = chunk_md->first_free;
 763		chunk_md->contig_hint = 0;
 764	}
 765
 766	bits = 0;
 767	pcpu_for_each_md_free_region(chunk, bit_off, bits)
 768		pcpu_block_update(chunk_md, bit_off, bit_off + bits);
 769}
 770
 771/**
 772 * pcpu_block_refresh_hint
 773 * @chunk: chunk of interest
 774 * @index: index of the metadata block
 775 *
 776 * Scans over the block beginning at first_free and updates the block
 777 * metadata accordingly.
 778 */
 779static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
 780{
 781	struct pcpu_block_md *block = chunk->md_blocks + index;
 782	unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
 783	unsigned int rs, re, start;	/* region start, region end */
 784
 785	/* promote scan_hint to contig_hint */
 786	if (block->scan_hint) {
 787		start = block->scan_hint_start + block->scan_hint;
 788		block->contig_hint_start = block->scan_hint_start;
 789		block->contig_hint = block->scan_hint;
 790		block->scan_hint = 0;
 791	} else {
 792		start = block->first_free;
 793		block->contig_hint = 0;
 794	}
 795
 796	block->right_free = 0;
 
 
 797
 798	/* iterate over free areas and update the contig hints */
 799	bitmap_for_each_clear_region(alloc_map, rs, re, start,
 800				     PCPU_BITMAP_BLOCK_BITS)
 801		pcpu_block_update(block, rs, re);
 802}
 803
 804/**
 805 * pcpu_block_update_hint_alloc - update hint on allocation path
 806 * @chunk: chunk of interest
 807 * @bit_off: chunk offset
 808 * @bits: size of request
 809 *
 810 * Updates metadata for the allocation path.  The metadata only has to be
 811 * refreshed by a full scan iff the chunk's contig hint is broken.  Block level
 812 * scans are required if the block's contig hint is broken.
 813 */
 814static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
 815					 int bits)
 816{
 817	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 818	int nr_empty_pages = 0;
 819	struct pcpu_block_md *s_block, *e_block, *block;
 820	int s_index, e_index;	/* block indexes of the freed allocation */
 821	int s_off, e_off;	/* block offsets of the freed allocation */
 822
 823	/*
 824	 * Calculate per block offsets.
 825	 * The calculation uses an inclusive range, but the resulting offsets
 826	 * are [start, end).  e_index always points to the last block in the
 827	 * range.
 828	 */
 829	s_index = pcpu_off_to_block_index(bit_off);
 830	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 831	s_off = pcpu_off_to_block_off(bit_off);
 832	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 833
 834	s_block = chunk->md_blocks + s_index;
 835	e_block = chunk->md_blocks + e_index;
 836
 837	/*
 838	 * Update s_block.
 839	 * block->first_free must be updated if the allocation takes its place.
 840	 * If the allocation breaks the contig_hint, a scan is required to
 841	 * restore this hint.
 842	 */
 843	if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 844		nr_empty_pages++;
 845
 846	if (s_off == s_block->first_free)
 847		s_block->first_free = find_next_zero_bit(
 848					pcpu_index_alloc_map(chunk, s_index),
 849					PCPU_BITMAP_BLOCK_BITS,
 850					s_off + bits);
 851
 852	if (pcpu_region_overlap(s_block->scan_hint_start,
 853				s_block->scan_hint_start + s_block->scan_hint,
 854				s_off,
 855				s_off + bits))
 856		s_block->scan_hint = 0;
 857
 858	if (pcpu_region_overlap(s_block->contig_hint_start,
 859				s_block->contig_hint_start +
 860				s_block->contig_hint,
 861				s_off,
 862				s_off + bits)) {
 863		/* block contig hint is broken - scan to fix it */
 864		if (!s_off)
 865			s_block->left_free = 0;
 866		pcpu_block_refresh_hint(chunk, s_index);
 867	} else {
 868		/* update left and right contig manually */
 869		s_block->left_free = min(s_block->left_free, s_off);
 870		if (s_index == e_index)
 871			s_block->right_free = min_t(int, s_block->right_free,
 872					PCPU_BITMAP_BLOCK_BITS - e_off);
 873		else
 874			s_block->right_free = 0;
 875	}
 876
 877	/*
 878	 * Update e_block.
 
 879	 */
 880	if (s_index != e_index) {
 881		if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 882			nr_empty_pages++;
 883
 884		/*
 885		 * When the allocation is across blocks, the end is along
 886		 * the left part of the e_block.
 887		 */
 888		e_block->first_free = find_next_zero_bit(
 889				pcpu_index_alloc_map(chunk, e_index),
 890				PCPU_BITMAP_BLOCK_BITS, e_off);
 891
 892		if (e_off == PCPU_BITMAP_BLOCK_BITS) {
 893			/* reset the block */
 894			e_block++;
 895		} else {
 896			if (e_off > e_block->scan_hint_start)
 897				e_block->scan_hint = 0;
 898
 899			e_block->left_free = 0;
 900			if (e_off > e_block->contig_hint_start) {
 901				/* contig hint is broken - scan to fix it */
 902				pcpu_block_refresh_hint(chunk, e_index);
 903			} else {
 904				e_block->right_free =
 905					min_t(int, e_block->right_free,
 906					      PCPU_BITMAP_BLOCK_BITS - e_off);
 907			}
 908		}
 909
 910		/* update in-between md_blocks */
 911		nr_empty_pages += (e_index - s_index - 1);
 912		for (block = s_block + 1; block < e_block; block++) {
 913			block->scan_hint = 0;
 914			block->contig_hint = 0;
 915			block->left_free = 0;
 916			block->right_free = 0;
 917		}
 918	}
 919
 920	if (nr_empty_pages)
 921		pcpu_update_empty_pages(chunk, -nr_empty_pages);
 922
 923	if (pcpu_region_overlap(chunk_md->scan_hint_start,
 924				chunk_md->scan_hint_start +
 925				chunk_md->scan_hint,
 926				bit_off,
 927				bit_off + bits))
 928		chunk_md->scan_hint = 0;
 929
 930	/*
 931	 * The only time a full chunk scan is required is if the chunk
 932	 * contig hint is broken.  Otherwise, it means a smaller space
 933	 * was used and therefore the chunk contig hint is still correct.
 934	 */
 935	if (pcpu_region_overlap(chunk_md->contig_hint_start,
 936				chunk_md->contig_hint_start +
 937				chunk_md->contig_hint,
 938				bit_off,
 939				bit_off + bits))
 940		pcpu_chunk_refresh_hint(chunk, false);
 941}
 942
 943/**
 944 * pcpu_block_update_hint_free - updates the block hints on the free path
 945 * @chunk: chunk of interest
 946 * @bit_off: chunk offset
 947 * @bits: size of request
 948 *
 949 * Updates metadata for the allocation path.  This avoids a blind block
 950 * refresh by making use of the block contig hints.  If this fails, it scans
 951 * forward and backward to determine the extent of the free area.  This is
 952 * capped at the boundary of blocks.
 953 *
 954 * A chunk update is triggered if a page becomes free, a block becomes free,
 955 * or the free spans across blocks.  This tradeoff is to minimize iterating
 956 * over the block metadata to update chunk_md->contig_hint.
 957 * chunk_md->contig_hint may be off by up to a page, but it will never be more
 958 * than the available space.  If the contig hint is contained in one block, it
 959 * will be accurate.
 960 */
 961static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
 962					int bits)
 963{
 964	int nr_empty_pages = 0;
 965	struct pcpu_block_md *s_block, *e_block, *block;
 966	int s_index, e_index;	/* block indexes of the freed allocation */
 967	int s_off, e_off;	/* block offsets of the freed allocation */
 968	int start, end;		/* start and end of the whole free area */
 969
 970	/*
 971	 * Calculate per block offsets.
 972	 * The calculation uses an inclusive range, but the resulting offsets
 973	 * are [start, end).  e_index always points to the last block in the
 974	 * range.
 975	 */
 976	s_index = pcpu_off_to_block_index(bit_off);
 977	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 978	s_off = pcpu_off_to_block_off(bit_off);
 979	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 980
 981	s_block = chunk->md_blocks + s_index;
 982	e_block = chunk->md_blocks + e_index;
 983
 984	/*
 985	 * Check if the freed area aligns with the block->contig_hint.
 986	 * If it does, then the scan to find the beginning/end of the
 987	 * larger free area can be avoided.
 988	 *
 989	 * start and end refer to beginning and end of the free area
 990	 * within each their respective blocks.  This is not necessarily
 991	 * the entire free area as it may span blocks past the beginning
 992	 * or end of the block.
 993	 */
 994	start = s_off;
 995	if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
 996		start = s_block->contig_hint_start;
 997	} else {
 998		/*
 999		 * Scan backwards to find the extent of the free area.
1000		 * find_last_bit returns the starting bit, so if the start bit
1001		 * is returned, that means there was no last bit and the
1002		 * remainder of the chunk is free.
1003		 */
1004		int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1005					  start);
1006		start = (start == l_bit) ? 0 : l_bit + 1;
1007	}
1008
1009	end = e_off;
1010	if (e_off == e_block->contig_hint_start)
1011		end = e_block->contig_hint_start + e_block->contig_hint;
1012	else
1013		end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1014				    PCPU_BITMAP_BLOCK_BITS, end);
1015
1016	/* update s_block */
1017	e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1018	if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1019		nr_empty_pages++;
1020	pcpu_block_update(s_block, start, e_off);
1021
1022	/* freeing in the same block */
1023	if (s_index != e_index) {
1024		/* update e_block */
1025		if (end == PCPU_BITMAP_BLOCK_BITS)
1026			nr_empty_pages++;
1027		pcpu_block_update(e_block, 0, end);
1028
1029		/* reset md_blocks in the middle */
1030		nr_empty_pages += (e_index - s_index - 1);
1031		for (block = s_block + 1; block < e_block; block++) {
1032			block->first_free = 0;
1033			block->scan_hint = 0;
1034			block->contig_hint_start = 0;
1035			block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1036			block->left_free = PCPU_BITMAP_BLOCK_BITS;
1037			block->right_free = PCPU_BITMAP_BLOCK_BITS;
1038		}
1039	}
1040
1041	if (nr_empty_pages)
1042		pcpu_update_empty_pages(chunk, nr_empty_pages);
1043
1044	/*
1045	 * Refresh chunk metadata when the free makes a block free or spans
1046	 * across blocks.  The contig_hint may be off by up to a page, but if
1047	 * the contig_hint is contained in a block, it will be accurate with
1048	 * the else condition below.
1049	 */
1050	if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1051		pcpu_chunk_refresh_hint(chunk, true);
1052	else
1053		pcpu_block_update(&chunk->chunk_md,
1054				  pcpu_block_off_to_off(s_index, start),
1055				  end);
1056}
1057
1058/**
1059 * pcpu_is_populated - determines if the region is populated
1060 * @chunk: chunk of interest
1061 * @bit_off: chunk offset
1062 * @bits: size of area
1063 * @next_off: return value for the next offset to start searching
 
 
1064 *
1065 * For atomic allocations, check if the backing pages are populated.
 
 
 
 
1066 *
1067 * RETURNS:
1068 * Bool if the backing pages are populated.
1069 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1070 */
1071static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1072			      int *next_off)
1073{
1074	unsigned int page_start, page_end, rs, re;
1075
1076	page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1077	page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
 
1078
1079	rs = page_start;
1080	bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1081	if (rs >= page_end)
1082		return true;
1083
1084	*next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1085	return false;
1086}
1087
1088/**
1089 * pcpu_find_block_fit - finds the block index to start searching
1090 * @chunk: chunk of interest
1091 * @alloc_bits: size of request in allocation units
1092 * @align: alignment of area (max PAGE_SIZE bytes)
1093 * @pop_only: use populated regions only
1094 *
1095 * Given a chunk and an allocation spec, find the offset to begin searching
1096 * for a free region.  This iterates over the bitmap metadata blocks to
1097 * find an offset that will be guaranteed to fit the requirements.  It is
1098 * not quite first fit as if the allocation does not fit in the contig hint
1099 * of a block or chunk, it is skipped.  This errs on the side of caution
1100 * to prevent excess iteration.  Poor alignment can cause the allocator to
1101 * skip over blocks and chunks that have valid free areas.
1102 *
1103 * RETURNS:
1104 * The offset in the bitmap to begin searching.
1105 * -1 if no offset is found.
1106 */
1107static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1108			       size_t align, bool pop_only)
1109{
1110	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1111	int bit_off, bits, next_off;
1112
1113	/*
1114	 * This is an optimization to prevent scanning by assuming if the
1115	 * allocation cannot fit in the global hint, there is memory pressure
1116	 * and creating a new chunk would happen soon.
1117	 */
1118	if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1119		return -1;
1120
1121	bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1122	bits = 0;
1123	pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1124		if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1125						   &next_off))
1126			break;
1127
1128		bit_off = next_off;
1129		bits = 0;
1130	}
1131
1132	if (bit_off == pcpu_chunk_map_bits(chunk))
1133		return -1;
1134
1135	return bit_off;
1136}
1137
1138/*
1139 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1140 * @map: the address to base the search on
1141 * @size: the bitmap size in bits
1142 * @start: the bitnumber to start searching at
1143 * @nr: the number of zeroed bits we're looking for
1144 * @align_mask: alignment mask for zero area
1145 * @largest_off: offset of the largest area skipped
1146 * @largest_bits: size of the largest area skipped
1147 *
1148 * The @align_mask should be one less than a power of 2.
1149 *
1150 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1151 * the largest area that was skipped.  This is imperfect, but in general is
1152 * good enough.  The largest remembered region is the largest failed region
1153 * seen.  This does not include anything we possibly skipped due to alignment.
1154 * pcpu_block_update_scan() does scan backwards to try and recover what was
1155 * lost to alignment.  While this can cause scanning to miss earlier possible
1156 * free areas, smaller allocations will eventually fill those holes.
1157 */
1158static unsigned long pcpu_find_zero_area(unsigned long *map,
1159					 unsigned long size,
1160					 unsigned long start,
1161					 unsigned long nr,
1162					 unsigned long align_mask,
1163					 unsigned long *largest_off,
1164					 unsigned long *largest_bits)
1165{
1166	unsigned long index, end, i, area_off, area_bits;
1167again:
1168	index = find_next_zero_bit(map, size, start);
1169
1170	/* Align allocation */
1171	index = __ALIGN_MASK(index, align_mask);
1172	area_off = index;
1173
1174	end = index + nr;
1175	if (end > size)
1176		return end;
1177	i = find_next_bit(map, end, index);
1178	if (i < end) {
1179		area_bits = i - area_off;
1180		/* remember largest unused area with best alignment */
1181		if (area_bits > *largest_bits ||
1182		    (area_bits == *largest_bits && *largest_off &&
1183		     (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1184			*largest_off = area_off;
1185			*largest_bits = area_bits;
1186		}
1187
1188		start = i + 1;
1189		goto again;
1190	}
1191	return index;
1192}
1193
1194/**
1195 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1196 * @chunk: chunk of interest
1197 * @alloc_bits: size of request in allocation units
1198 * @align: alignment of area (max PAGE_SIZE)
1199 * @start: bit_off to start searching
 
 
 
 
 
1200 *
1201 * This function takes in a @start offset to begin searching to fit an
1202 * allocation of @alloc_bits with alignment @align.  It needs to scan
1203 * the allocation map because if it fits within the block's contig hint,
1204 * @start will be block->first_free. This is an attempt to fill the
1205 * allocation prior to breaking the contig hint.  The allocation and
1206 * boundary maps are updated accordingly if it confirms a valid
1207 * free area.
1208 *
1209 * RETURNS:
1210 * Allocated addr offset in @chunk on success.
1211 * -1 if no matching area is found.
1212 */
1213static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1214			   size_t align, int start)
1215{
1216	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1217	size_t align_mask = (align) ? (align - 1) : 0;
1218	unsigned long area_off = 0, area_bits = 0;
1219	int bit_off, end, oslot;
1220
1221	lockdep_assert_held(&pcpu_lock);
1222
1223	oslot = pcpu_chunk_slot(chunk);
1224
1225	/*
1226	 * Search to find a fit.
1227	 */
1228	end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1229		    pcpu_chunk_map_bits(chunk));
1230	bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1231				      align_mask, &area_off, &area_bits);
1232	if (bit_off >= end)
1233		return -1;
1234
1235	if (area_bits)
1236		pcpu_block_update_scan(chunk, area_off, area_bits);
1237
1238	/* update alloc map */
1239	bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1240
1241	/* update boundary map */
1242	set_bit(bit_off, chunk->bound_map);
1243	bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1244	set_bit(bit_off + alloc_bits, chunk->bound_map);
1245
1246	chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1247
1248	/* update first free bit */
1249	if (bit_off == chunk_md->first_free)
1250		chunk_md->first_free = find_next_zero_bit(
1251					chunk->alloc_map,
1252					pcpu_chunk_map_bits(chunk),
1253					bit_off + alloc_bits);
1254
1255	pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1256
1257	pcpu_chunk_relocate(chunk, oslot);
1258
1259	return bit_off * PCPU_MIN_ALLOC_SIZE;
1260}
1261
1262/**
1263 * pcpu_free_area - frees the corresponding offset
1264 * @chunk: chunk of interest
1265 * @off: addr offset into chunk
1266 *
1267 * This function determines the size of an allocation to free using
1268 * the boundary bitmap and clears the allocation map.
1269 *
1270 * RETURNS:
1271 * Number of freed bytes.
 
1272 */
1273static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
 
1274{
1275	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1276	int bit_off, bits, end, oslot, freed;
 
 
 
 
 
 
 
1277
1278	lockdep_assert_held(&pcpu_lock);
1279	pcpu_stats_area_dealloc(chunk);
 
1280
1281	oslot = pcpu_chunk_slot(chunk);
1282
1283	bit_off = off / PCPU_MIN_ALLOC_SIZE;
 
 
 
 
 
 
 
 
 
1284
1285	/* find end index */
1286	end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1287			    bit_off + 1);
1288	bits = end - bit_off;
1289	bitmap_clear(chunk->alloc_map, bit_off, bits);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1290
1291	freed = bits * PCPU_MIN_ALLOC_SIZE;
 
1292
1293	/* update metadata */
1294	chunk->free_bytes += freed;
 
 
 
 
1295
1296	/* update first free bit */
1297	chunk_md->first_free = min(chunk_md->first_free, bit_off);
1298
1299	pcpu_block_update_hint_free(chunk, bit_off, bits);
 
 
 
1300
 
1301	pcpu_chunk_relocate(chunk, oslot);
1302
1303	return freed;
1304}
1305
1306static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1307{
1308	block->scan_hint = 0;
1309	block->contig_hint = nr_bits;
1310	block->left_free = nr_bits;
1311	block->right_free = nr_bits;
1312	block->first_free = 0;
1313	block->nr_bits = nr_bits;
1314}
1315
1316static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1317{
1318	struct pcpu_block_md *md_block;
1319
1320	/* init the chunk's block */
1321	pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1322
1323	for (md_block = chunk->md_blocks;
1324	     md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1325	     md_block++)
1326		pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1327}
1328
1329/**
1330 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1331 * @tmp_addr: the start of the region served
1332 * @map_size: size of the region served
 
1333 *
1334 * This is responsible for creating the chunks that serve the first chunk.  The
1335 * base_addr is page aligned down of @tmp_addr while the region end is page
1336 * aligned up.  Offsets are kept track of to determine the region served. All
1337 * this is done to appease the bitmap allocator in avoiding partial blocks.
1338 *
1339 * RETURNS:
1340 * Chunk serving the region at @tmp_addr of @map_size.
1341 */
1342static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1343							 int map_size)
1344{
1345	struct pcpu_chunk *chunk;
1346	unsigned long aligned_addr, lcm_align;
1347	int start_offset, offset_bits, region_size, region_bits;
1348	size_t alloc_size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1349
1350	/* region calculations */
1351	aligned_addr = tmp_addr & PAGE_MASK;
1352
1353	start_offset = tmp_addr - aligned_addr;
1354
1355	/*
1356	 * Align the end of the region with the LCM of PAGE_SIZE and
1357	 * PCPU_BITMAP_BLOCK_SIZE.  One of these constants is a multiple of
1358	 * the other.
1359	 */
1360	lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1361	region_size = ALIGN(start_offset + map_size, lcm_align);
1362
1363	/* allocate chunk */
1364	alloc_size = struct_size(chunk, populated,
1365				 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1366	chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1367	if (!chunk)
1368		panic("%s: Failed to allocate %zu bytes\n", __func__,
1369		      alloc_size);
1370
1371	INIT_LIST_HEAD(&chunk->list);
1372
1373	chunk->base_addr = (void *)aligned_addr;
1374	chunk->start_offset = start_offset;
1375	chunk->end_offset = region_size - chunk->start_offset - map_size;
1376
1377	chunk->nr_pages = region_size >> PAGE_SHIFT;
1378	region_bits = pcpu_chunk_map_bits(chunk);
1379
1380	alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1381	chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1382	if (!chunk->alloc_map)
1383		panic("%s: Failed to allocate %zu bytes\n", __func__,
1384		      alloc_size);
1385
1386	alloc_size =
1387		BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1388	chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1389	if (!chunk->bound_map)
1390		panic("%s: Failed to allocate %zu bytes\n", __func__,
1391		      alloc_size);
1392
1393	alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1394	chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1395	if (!chunk->md_blocks)
1396		panic("%s: Failed to allocate %zu bytes\n", __func__,
1397		      alloc_size);
1398
1399#ifdef CONFIG_MEMCG_KMEM
1400	/* first chunk is free to use */
1401	chunk->obj_cgroups = NULL;
1402#endif
1403	pcpu_init_md_blocks(chunk);
1404
1405	/* manage populated page bitmap */
1406	chunk->immutable = true;
1407	bitmap_fill(chunk->populated, chunk->nr_pages);
1408	chunk->nr_populated = chunk->nr_pages;
1409	chunk->nr_empty_pop_pages = chunk->nr_pages;
1410
1411	chunk->free_bytes = map_size;
1412
1413	if (chunk->start_offset) {
1414		/* hide the beginning of the bitmap */
1415		offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1416		bitmap_set(chunk->alloc_map, 0, offset_bits);
1417		set_bit(0, chunk->bound_map);
1418		set_bit(offset_bits, chunk->bound_map);
1419
1420		chunk->chunk_md.first_free = offset_bits;
1421
1422		pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1423	}
1424
1425	if (chunk->end_offset) {
1426		/* hide the end of the bitmap */
1427		offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1428		bitmap_set(chunk->alloc_map,
1429			   pcpu_chunk_map_bits(chunk) - offset_bits,
1430			   offset_bits);
1431		set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1432			chunk->bound_map);
1433		set_bit(region_bits, chunk->bound_map);
1434
1435		pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1436					     - offset_bits, offset_bits);
1437	}
1438
1439	return chunk;
 
1440}
1441
1442static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1443{
1444	struct pcpu_chunk *chunk;
1445	int region_bits;
1446
1447	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1448	if (!chunk)
1449		return NULL;
1450
1451	INIT_LIST_HEAD(&chunk->list);
1452	chunk->nr_pages = pcpu_unit_pages;
1453	region_bits = pcpu_chunk_map_bits(chunk);
1454
1455	chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1456					   sizeof(chunk->alloc_map[0]), gfp);
1457	if (!chunk->alloc_map)
1458		goto alloc_map_fail;
1459
1460	chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1461					   sizeof(chunk->bound_map[0]), gfp);
1462	if (!chunk->bound_map)
1463		goto bound_map_fail;
1464
1465	chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1466					   sizeof(chunk->md_blocks[0]), gfp);
1467	if (!chunk->md_blocks)
1468		goto md_blocks_fail;
1469
1470#ifdef CONFIG_MEMCG_KMEM
1471	if (!mem_cgroup_kmem_disabled()) {
1472		chunk->obj_cgroups =
1473			pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1474					sizeof(struct obj_cgroup *), gfp);
1475		if (!chunk->obj_cgroups)
1476			goto objcg_fail;
1477	}
1478#endif
1479
1480	pcpu_init_md_blocks(chunk);
 
 
 
1481
1482	/* init metadata */
1483	chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
 
 
1484
1485	return chunk;
1486
1487#ifdef CONFIG_MEMCG_KMEM
1488objcg_fail:
1489	pcpu_mem_free(chunk->md_blocks);
1490#endif
1491md_blocks_fail:
1492	pcpu_mem_free(chunk->bound_map);
1493bound_map_fail:
1494	pcpu_mem_free(chunk->alloc_map);
1495alloc_map_fail:
1496	pcpu_mem_free(chunk);
1497
1498	return NULL;
1499}
1500
1501static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1502{
1503	if (!chunk)
1504		return;
1505#ifdef CONFIG_MEMCG_KMEM
1506	pcpu_mem_free(chunk->obj_cgroups);
1507#endif
1508	pcpu_mem_free(chunk->md_blocks);
1509	pcpu_mem_free(chunk->bound_map);
1510	pcpu_mem_free(chunk->alloc_map);
1511	pcpu_mem_free(chunk);
1512}
1513
1514/**
1515 * pcpu_chunk_populated - post-population bookkeeping
1516 * @chunk: pcpu_chunk which got populated
1517 * @page_start: the start page
1518 * @page_end: the end page
1519 *
1520 * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
1521 * the bookkeeping information accordingly.  Must be called after each
1522 * successful population.
1523 *
1524 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1525 * is to serve an allocation in that area.
1526 */
1527static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1528				 int page_end)
1529{
1530	int nr = page_end - page_start;
1531
1532	lockdep_assert_held(&pcpu_lock);
1533
1534	bitmap_set(chunk->populated, page_start, nr);
1535	chunk->nr_populated += nr;
1536	pcpu_nr_populated += nr;
1537
1538	pcpu_update_empty_pages(chunk, nr);
1539}
1540
1541/**
1542 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1543 * @chunk: pcpu_chunk which got depopulated
1544 * @page_start: the start page
1545 * @page_end: the end page
1546 *
1547 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1548 * Update the bookkeeping information accordingly.  Must be called after
1549 * each successful depopulation.
1550 */
1551static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1552				   int page_start, int page_end)
1553{
1554	int nr = page_end - page_start;
1555
1556	lockdep_assert_held(&pcpu_lock);
1557
1558	bitmap_clear(chunk->populated, page_start, nr);
1559	chunk->nr_populated -= nr;
1560	pcpu_nr_populated -= nr;
1561
1562	pcpu_update_empty_pages(chunk, -nr);
1563}
1564
1565/*
1566 * Chunk management implementation.
1567 *
1568 * To allow different implementations, chunk alloc/free and
1569 * [de]population are implemented in a separate file which is pulled
1570 * into this file and compiled together.  The following functions
1571 * should be implemented.
1572 *
1573 * pcpu_populate_chunk		- populate the specified range of a chunk
1574 * pcpu_depopulate_chunk	- depopulate the specified range of a chunk
1575 * pcpu_post_unmap_tlb_flush	- flush tlb for the specified range of a chunk
1576 * pcpu_create_chunk		- create a new chunk
1577 * pcpu_destroy_chunk		- destroy a chunk, always preceded by full depop
1578 * pcpu_addr_to_page		- translate address to physical address
1579 * pcpu_verify_alloc_info	- check alloc_info is acceptable during init
1580 */
1581static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1582			       int page_start, int page_end, gfp_t gfp);
1583static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1584				  int page_start, int page_end);
1585static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1586				      int page_start, int page_end);
1587static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1588static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1589static struct page *pcpu_addr_to_page(void *addr);
1590static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1591
1592#ifdef CONFIG_NEED_PER_CPU_KM
1593#include "percpu-km.c"
1594#else
1595#include "percpu-vm.c"
1596#endif
1597
1598/**
1599 * pcpu_chunk_addr_search - determine chunk containing specified address
1600 * @addr: address for which the chunk needs to be determined.
1601 *
1602 * This is an internal function that handles all but static allocations.
1603 * Static percpu address values should never be passed into the allocator.
1604 *
1605 * RETURNS:
1606 * The address of the found chunk.
1607 */
1608static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1609{
1610	/* is it in the dynamic region (first chunk)? */
1611	if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
 
 
 
1612		return pcpu_first_chunk;
1613
1614	/* is it in the reserved region? */
1615	if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1616		return pcpu_reserved_chunk;
1617
1618	/*
1619	 * The address is relative to unit0 which might be unused and
1620	 * thus unmapped.  Offset the address to the unit space of the
1621	 * current processor before looking it up in the vmalloc
1622	 * space.  Note that any possible cpu id can be used here, so
1623	 * there's no need to worry about preemption or cpu hotplug.
1624	 */
1625	addr += pcpu_unit_offsets[raw_smp_processor_id()];
1626	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1627}
1628
1629#ifdef CONFIG_MEMCG_KMEM
1630static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1631				      struct obj_cgroup **objcgp)
1632{
1633	struct obj_cgroup *objcg;
1634
1635	if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT))
1636		return true;
1637
1638	objcg = get_obj_cgroup_from_current();
1639	if (!objcg)
1640		return true;
1641
1642	if (obj_cgroup_charge(objcg, gfp, size * num_possible_cpus())) {
1643		obj_cgroup_put(objcg);
1644		return false;
1645	}
1646
1647	*objcgp = objcg;
1648	return true;
1649}
1650
1651static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1652				       struct pcpu_chunk *chunk, int off,
1653				       size_t size)
1654{
1655	if (!objcg)
1656		return;
1657
1658	if (likely(chunk && chunk->obj_cgroups)) {
1659		chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1660
1661		rcu_read_lock();
1662		mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1663				size * num_possible_cpus());
1664		rcu_read_unlock();
1665	} else {
1666		obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1667		obj_cgroup_put(objcg);
1668	}
1669}
1670
1671static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1672{
1673	struct obj_cgroup *objcg;
1674
1675	if (unlikely(!chunk->obj_cgroups))
1676		return;
1677
1678	objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1679	if (!objcg)
1680		return;
1681	chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1682
1683	obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1684
1685	rcu_read_lock();
1686	mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1687			-(size * num_possible_cpus()));
1688	rcu_read_unlock();
1689
1690	obj_cgroup_put(objcg);
1691}
1692
1693#else /* CONFIG_MEMCG_KMEM */
1694static bool
1695pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1696{
1697	return true;
1698}
1699
1700static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1701				       struct pcpu_chunk *chunk, int off,
1702				       size_t size)
1703{
1704}
1705
1706static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1707{
1708}
1709#endif /* CONFIG_MEMCG_KMEM */
1710
1711/**
1712 * pcpu_alloc - the percpu allocator
1713 * @size: size of area to allocate in bytes
1714 * @align: alignment of area (max PAGE_SIZE)
1715 * @reserved: allocate from the reserved chunk if available
1716 * @gfp: allocation flags
1717 *
1718 * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
1719 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1720 * then no warning will be triggered on invalid or failed allocation
1721 * requests.
1722 *
1723 * RETURNS:
1724 * Percpu pointer to the allocated area on success, NULL on failure.
1725 */
1726static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1727				 gfp_t gfp)
1728{
1729	gfp_t pcpu_gfp;
1730	bool is_atomic;
1731	bool do_warn;
1732	struct obj_cgroup *objcg = NULL;
1733	static int warn_limit = 10;
1734	struct pcpu_chunk *chunk, *next;
1735	const char *err;
1736	int slot, off, cpu, ret;
 
 
1737	unsigned long flags;
1738	void __percpu *ptr;
1739	size_t bits, bit_align;
1740
1741	gfp = current_gfp_context(gfp);
1742	/* whitelisted flags that can be passed to the backing allocators */
1743	pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1744	is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1745	do_warn = !(gfp & __GFP_NOWARN);
1746
1747	/*
1748	 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1749	 * therefore alignment must be a minimum of that many bytes.
1750	 * An allocation may have internal fragmentation from rounding up
1751	 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1752	 */
1753	if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1754		align = PCPU_MIN_ALLOC_SIZE;
 
 
1755
1756	size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1757	bits = size >> PCPU_MIN_ALLOC_SHIFT;
1758	bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1759
1760	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1761		     !is_power_of_2(align))) {
1762		WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1763		     size, align);
1764		return NULL;
1765	}
1766
1767	if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1768		return NULL;
1769
1770	if (!is_atomic) {
1771		/*
1772		 * pcpu_balance_workfn() allocates memory under this mutex,
1773		 * and it may wait for memory reclaim. Allow current task
1774		 * to become OOM victim, in case of memory pressure.
1775		 */
1776		if (gfp & __GFP_NOFAIL) {
1777			mutex_lock(&pcpu_alloc_mutex);
1778		} else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1779			pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1780			return NULL;
1781		}
1782	}
1783
1784	spin_lock_irqsave(&pcpu_lock, flags);
1785
1786	/* serve reserved allocations from the reserved chunk if available */
1787	if (reserved && pcpu_reserved_chunk) {
1788		chunk = pcpu_reserved_chunk;
1789
1790		off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1791		if (off < 0) {
1792			err = "alloc from reserved chunk failed";
1793			goto fail_unlock;
1794		}
1795
1796		off = pcpu_alloc_area(chunk, bits, bit_align, off);
 
 
 
 
 
 
 
 
 
 
 
1797		if (off >= 0)
1798			goto area_found;
1799
1800		err = "alloc from reserved chunk failed";
1801		goto fail_unlock;
1802	}
1803
1804restart:
1805	/* search through normal chunks */
1806	for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1807		list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1808					 list) {
1809			off = pcpu_find_block_fit(chunk, bits, bit_align,
1810						  is_atomic);
1811			if (off < 0) {
1812				if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1813					pcpu_chunk_move(chunk, 0);
1814				continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1815			}
1816
1817			off = pcpu_alloc_area(chunk, bits, bit_align, off);
1818			if (off >= 0) {
1819				pcpu_reintegrate_chunk(chunk);
1820				goto area_found;
1821			}
1822		}
1823	}
1824
1825	spin_unlock_irqrestore(&pcpu_lock, flags);
1826
1827	/*
1828	 * No space left.  Create a new chunk.  We don't want multiple
1829	 * tasks to create chunks simultaneously.  Serialize and create iff
1830	 * there's still no empty chunk after grabbing the mutex.
1831	 */
1832	if (is_atomic) {
1833		err = "atomic alloc failed, no space left";
1834		goto fail;
1835	}
1836
1837	if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1838		chunk = pcpu_create_chunk(pcpu_gfp);
 
 
1839		if (!chunk) {
 
1840			err = "failed to allocate new chunk";
1841			goto fail;
1842		}
1843
1844		spin_lock_irqsave(&pcpu_lock, flags);
1845		pcpu_chunk_relocate(chunk, -1);
1846	} else {
1847		spin_lock_irqsave(&pcpu_lock, flags);
1848	}
1849
 
1850	goto restart;
1851
1852area_found:
1853	pcpu_stats_area_alloc(chunk, size);
1854	spin_unlock_irqrestore(&pcpu_lock, flags);
1855
1856	/* populate if not all pages are already there */
1857	if (!is_atomic) {
1858		unsigned int page_start, page_end, rs, re;
 
 
1859
1860		page_start = PFN_DOWN(off);
1861		page_end = PFN_UP(off + size);
1862
1863		bitmap_for_each_clear_region(chunk->populated, rs, re,
1864					     page_start, page_end) {
1865			WARN_ON(chunk->immutable);
1866
1867			ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1868
1869			spin_lock_irqsave(&pcpu_lock, flags);
1870			if (ret) {
1871				pcpu_free_area(chunk, off);
 
1872				err = "failed to populate";
1873				goto fail_unlock;
1874			}
1875			pcpu_chunk_populated(chunk, rs, re);
1876			spin_unlock_irqrestore(&pcpu_lock, flags);
1877		}
1878
1879		mutex_unlock(&pcpu_alloc_mutex);
1880	}
1881
 
 
 
1882	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1883		pcpu_schedule_balance_work();
1884
1885	/* clear the areas and return address relative to base address */
1886	for_each_possible_cpu(cpu)
1887		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1888
1889	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1890	kmemleak_alloc_percpu(ptr, size, gfp);
1891
1892	trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1893			chunk->base_addr, off, ptr);
1894
1895	pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1896
1897	return ptr;
1898
1899fail_unlock:
1900	spin_unlock_irqrestore(&pcpu_lock, flags);
1901fail:
1902	trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1903
1904	if (!is_atomic && do_warn && warn_limit) {
1905		pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1906			size, align, is_atomic, err);
1907		dump_stack();
1908		if (!--warn_limit)
1909			pr_info("limit reached, disable warning\n");
1910	}
1911	if (is_atomic) {
1912		/* see the flag handling in pcpu_balance_workfn() */
1913		pcpu_atomic_alloc_failed = true;
1914		pcpu_schedule_balance_work();
1915	} else {
1916		mutex_unlock(&pcpu_alloc_mutex);
1917	}
1918
1919	pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1920
1921	return NULL;
1922}
1923
1924/**
1925 * __alloc_percpu_gfp - allocate dynamic percpu area
1926 * @size: size of area to allocate in bytes
1927 * @align: alignment of area (max PAGE_SIZE)
1928 * @gfp: allocation flags
1929 *
1930 * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1931 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1932 * be called from any context but is a lot more likely to fail. If @gfp
1933 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1934 * allocation requests.
1935 *
1936 * RETURNS:
1937 * Percpu pointer to the allocated area on success, NULL on failure.
1938 */
1939void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1940{
1941	return pcpu_alloc(size, align, false, gfp);
1942}
1943EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1944
1945/**
1946 * __alloc_percpu - allocate dynamic percpu area
1947 * @size: size of area to allocate in bytes
1948 * @align: alignment of area (max PAGE_SIZE)
1949 *
1950 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1951 */
1952void __percpu *__alloc_percpu(size_t size, size_t align)
1953{
1954	return pcpu_alloc(size, align, false, GFP_KERNEL);
1955}
1956EXPORT_SYMBOL_GPL(__alloc_percpu);
1957
1958/**
1959 * __alloc_reserved_percpu - allocate reserved percpu area
1960 * @size: size of area to allocate in bytes
1961 * @align: alignment of area (max PAGE_SIZE)
1962 *
1963 * Allocate zero-filled percpu area of @size bytes aligned at @align
1964 * from reserved percpu area if arch has set it up; otherwise,
1965 * allocation is served from the same dynamic area.  Might sleep.
1966 * Might trigger writeouts.
1967 *
1968 * CONTEXT:
1969 * Does GFP_KERNEL allocation.
1970 *
1971 * RETURNS:
1972 * Percpu pointer to the allocated area on success, NULL on failure.
1973 */
1974void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1975{
1976	return pcpu_alloc(size, align, true, GFP_KERNEL);
1977}
1978
1979/**
1980 * pcpu_balance_free - manage the amount of free chunks
1981 * @empty_only: free chunks only if there are no populated pages
1982 *
1983 * If empty_only is %false, reclaim all fully free chunks regardless of the
1984 * number of populated pages.  Otherwise, only reclaim chunks that have no
1985 * populated pages.
1986 *
1987 * CONTEXT:
1988 * pcpu_lock (can be dropped temporarily)
1989 */
1990static void pcpu_balance_free(bool empty_only)
1991{
1992	LIST_HEAD(to_free);
1993	struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1994	struct pcpu_chunk *chunk, *next;
1995
1996	lockdep_assert_held(&pcpu_lock);
1997
1998	/*
1999	 * There's no reason to keep around multiple unused chunks and VM
2000	 * areas can be scarce.  Destroy all free chunks except for one.
2001	 */
 
 
 
2002	list_for_each_entry_safe(chunk, next, free_head, list) {
2003		WARN_ON(chunk->immutable);
2004
2005		/* spare the first one */
2006		if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
2007			continue;
2008
2009		if (!empty_only || chunk->nr_empty_pop_pages == 0)
2010			list_move(&chunk->list, &to_free);
2011	}
2012
2013	if (list_empty(&to_free))
2014		return;
2015
2016	spin_unlock_irq(&pcpu_lock);
2017	list_for_each_entry_safe(chunk, next, &to_free, list) {
2018		unsigned int rs, re;
2019
2020		bitmap_for_each_set_region(chunk->populated, rs, re, 0,
2021					   chunk->nr_pages) {
2022			pcpu_depopulate_chunk(chunk, rs, re);
2023			spin_lock_irq(&pcpu_lock);
2024			pcpu_chunk_depopulated(chunk, rs, re);
2025			spin_unlock_irq(&pcpu_lock);
2026		}
2027		pcpu_destroy_chunk(chunk);
2028		cond_resched();
2029	}
2030	spin_lock_irq(&pcpu_lock);
2031}
2032
2033/**
2034 * pcpu_balance_populated - manage the amount of populated pages
2035 *
2036 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2037 * It is possible that this is called when physical memory is scarce causing
2038 * OOM killer to be triggered.  We should avoid doing so until an actual
2039 * allocation causes the failure as it is possible that requests can be
2040 * serviced from already backed regions.
2041 *
2042 * CONTEXT:
2043 * pcpu_lock (can be dropped temporarily)
2044 */
2045static void pcpu_balance_populated(void)
2046{
2047	/* gfp flags passed to underlying allocators */
2048	const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2049	struct pcpu_chunk *chunk;
2050	int slot, nr_to_pop, ret;
2051
2052	lockdep_assert_held(&pcpu_lock);
2053
2054	/*
2055	 * Ensure there are certain number of free populated pages for
2056	 * atomic allocs.  Fill up from the most packed so that atomic
2057	 * allocs don't increase fragmentation.  If atomic allocation
2058	 * failed previously, always populate the maximum amount.  This
2059	 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2060	 * failing indefinitely; however, large atomic allocs are not
2061	 * something we support properly and can be highly unreliable and
2062	 * inefficient.
2063	 */
2064retry_pop:
2065	if (pcpu_atomic_alloc_failed) {
2066		nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2067		/* best effort anyway, don't worry about synchronization */
2068		pcpu_atomic_alloc_failed = false;
2069	} else {
2070		nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2071				  pcpu_nr_empty_pop_pages,
2072				  0, PCPU_EMPTY_POP_PAGES_HIGH);
2073	}
2074
2075	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2076		unsigned int nr_unpop = 0, rs, re;
2077
2078		if (!nr_to_pop)
2079			break;
2080
2081		list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2082			nr_unpop = chunk->nr_pages - chunk->nr_populated;
 
2083			if (nr_unpop)
2084				break;
2085		}
 
2086
2087		if (!nr_unpop)
2088			continue;
2089
2090		/* @chunk can't go away while pcpu_alloc_mutex is held */
2091		bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
2092					     chunk->nr_pages) {
2093			int nr = min_t(int, re - rs, nr_to_pop);
2094
2095			spin_unlock_irq(&pcpu_lock);
2096			ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2097			cond_resched();
2098			spin_lock_irq(&pcpu_lock);
2099			if (!ret) {
2100				nr_to_pop -= nr;
 
2101				pcpu_chunk_populated(chunk, rs, rs + nr);
 
2102			} else {
2103				nr_to_pop = 0;
2104			}
2105
2106			if (!nr_to_pop)
2107				break;
2108		}
2109	}
2110
2111	if (nr_to_pop) {
2112		/* ran out of chunks to populate, create a new one and retry */
2113		spin_unlock_irq(&pcpu_lock);
2114		chunk = pcpu_create_chunk(gfp);
2115		cond_resched();
2116		spin_lock_irq(&pcpu_lock);
2117		if (chunk) {
 
2118			pcpu_chunk_relocate(chunk, -1);
 
2119			goto retry_pop;
2120		}
2121	}
2122}
2123
2124/**
2125 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2126 *
2127 * Scan over chunks in the depopulate list and try to release unused populated
2128 * pages back to the system.  Depopulated chunks are sidelined to prevent
2129 * repopulating these pages unless required.  Fully free chunks are reintegrated
2130 * and freed accordingly (1 is kept around).  If we drop below the empty
2131 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2132 * Each chunk is scanned in the reverse order to keep populated pages close to
2133 * the beginning of the chunk.
2134 *
2135 * CONTEXT:
2136 * pcpu_lock (can be dropped temporarily)
2137 *
2138 */
2139static void pcpu_reclaim_populated(void)
2140{
2141	struct pcpu_chunk *chunk;
2142	struct pcpu_block_md *block;
2143	int freed_page_start, freed_page_end;
2144	int i, end;
2145	bool reintegrate;
2146
2147	lockdep_assert_held(&pcpu_lock);
2148
2149	/*
2150	 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2151	 * longer discoverable to allocations whom may populate pages.  The only
2152	 * other accessor is the free path which only returns area back to the
2153	 * allocator not touching the populated bitmap.
2154	 */
2155	while (!list_empty(&pcpu_chunk_lists[pcpu_to_depopulate_slot])) {
2156		chunk = list_first_entry(&pcpu_chunk_lists[pcpu_to_depopulate_slot],
2157					 struct pcpu_chunk, list);
2158		WARN_ON(chunk->immutable);
2159
2160		/*
2161		 * Scan chunk's pages in the reverse order to keep populated
2162		 * pages close to the beginning of the chunk.
2163		 */
2164		freed_page_start = chunk->nr_pages;
2165		freed_page_end = 0;
2166		reintegrate = false;
2167		for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2168			/* no more work to do */
2169			if (chunk->nr_empty_pop_pages == 0)
2170				break;
2171
2172			/* reintegrate chunk to prevent atomic alloc failures */
2173			if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2174				reintegrate = true;
2175				goto end_chunk;
2176			}
2177
2178			/*
2179			 * If the page is empty and populated, start or
2180			 * extend the (i, end) range.  If i == 0, decrease
2181			 * i and perform the depopulation to cover the last
2182			 * (first) page in the chunk.
2183			 */
2184			block = chunk->md_blocks + i;
2185			if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2186			    test_bit(i, chunk->populated)) {
2187				if (end == -1)
2188					end = i;
2189				if (i > 0)
2190					continue;
2191				i--;
2192			}
2193
2194			/* depopulate if there is an active range */
2195			if (end == -1)
2196				continue;
2197
2198			spin_unlock_irq(&pcpu_lock);
2199			pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2200			cond_resched();
2201			spin_lock_irq(&pcpu_lock);
2202
2203			pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2204			freed_page_start = min(freed_page_start, i + 1);
2205			freed_page_end = max(freed_page_end, end + 1);
2206
2207			/* reset the range and continue */
2208			end = -1;
2209		}
2210
2211end_chunk:
2212		/* batch tlb flush per chunk to amortize cost */
2213		if (freed_page_start < freed_page_end) {
2214			spin_unlock_irq(&pcpu_lock);
2215			pcpu_post_unmap_tlb_flush(chunk,
2216						  freed_page_start,
2217						  freed_page_end);
2218			cond_resched();
2219			spin_lock_irq(&pcpu_lock);
2220		}
2221
2222		if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2223			pcpu_reintegrate_chunk(chunk);
2224		else
2225			list_move_tail(&chunk->list,
2226				       &pcpu_chunk_lists[pcpu_sidelined_slot]);
2227	}
2228}
2229
2230/**
2231 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2232 * @work: unused
2233 *
2234 * For each chunk type, manage the number of fully free chunks and the number of
2235 * populated pages.  An important thing to consider is when pages are freed and
2236 * how they contribute to the global counts.
2237 */
2238static void pcpu_balance_workfn(struct work_struct *work)
2239{
2240	/*
2241	 * pcpu_balance_free() is called twice because the first time we may
2242	 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2243	 * to grow other chunks.  This then gives pcpu_reclaim_populated() time
2244	 * to move fully free chunks to the active list to be freed if
2245	 * appropriate.
2246	 */
2247	mutex_lock(&pcpu_alloc_mutex);
2248	spin_lock_irq(&pcpu_lock);
2249
2250	pcpu_balance_free(false);
2251	pcpu_reclaim_populated();
2252	pcpu_balance_populated();
2253	pcpu_balance_free(true);
2254
2255	spin_unlock_irq(&pcpu_lock);
2256	mutex_unlock(&pcpu_alloc_mutex);
2257}
2258
2259/**
2260 * free_percpu - free percpu area
2261 * @ptr: pointer to area to free
2262 *
2263 * Free percpu area @ptr.
2264 *
2265 * CONTEXT:
2266 * Can be called from atomic context.
2267 */
2268void free_percpu(void __percpu *ptr)
2269{
2270	void *addr;
2271	struct pcpu_chunk *chunk;
2272	unsigned long flags;
2273	int size, off;
2274	bool need_balance = false;
2275
2276	if (!ptr)
2277		return;
2278
2279	kmemleak_free_percpu(ptr);
2280
2281	addr = __pcpu_ptr_to_addr(ptr);
2282
2283	spin_lock_irqsave(&pcpu_lock, flags);
2284
2285	chunk = pcpu_chunk_addr_search(addr);
2286	off = addr - chunk->base_addr;
2287
2288	size = pcpu_free_area(chunk, off);
2289
2290	pcpu_memcg_free_hook(chunk, off, size);
 
2291
2292	/*
2293	 * If there are more than one fully free chunks, wake up grim reaper.
2294	 * If the chunk is isolated, it may be in the process of being
2295	 * reclaimed.  Let reclaim manage cleaning up of that chunk.
2296	 */
2297	if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2298		struct pcpu_chunk *pos;
2299
2300		list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2301			if (pos != chunk) {
2302				need_balance = true;
2303				break;
2304			}
2305	} else if (pcpu_should_reclaim_chunk(chunk)) {
2306		pcpu_isolate_chunk(chunk);
2307		need_balance = true;
2308	}
2309
2310	trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2311
2312	spin_unlock_irqrestore(&pcpu_lock, flags);
2313
2314	if (need_balance)
2315		pcpu_schedule_balance_work();
2316}
2317EXPORT_SYMBOL_GPL(free_percpu);
2318
2319bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
 
 
 
 
 
 
 
 
 
 
 
2320{
2321#ifdef CONFIG_SMP
2322	const size_t static_size = __per_cpu_end - __per_cpu_start;
2323	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2324	unsigned int cpu;
2325
2326	for_each_possible_cpu(cpu) {
2327		void *start = per_cpu_ptr(base, cpu);
2328		void *va = (void *)addr;
2329
2330		if (va >= start && va < start + static_size) {
2331			if (can_addr) {
2332				*can_addr = (unsigned long) (va - start);
2333				*can_addr += (unsigned long)
2334					per_cpu_ptr(base, get_boot_cpu_id());
2335			}
2336			return true;
2337		}
2338	}
2339#endif
2340	/* on UP, can't distinguish from other static vars, always false */
2341	return false;
2342}
2343
2344/**
2345 * is_kernel_percpu_address - test whether address is from static percpu area
2346 * @addr: address to test
2347 *
2348 * Test whether @addr belongs to in-kernel static percpu area.  Module
2349 * static percpu areas are not considered.  For those, use
2350 * is_module_percpu_address().
2351 *
2352 * RETURNS:
2353 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2354 */
2355bool is_kernel_percpu_address(unsigned long addr)
2356{
2357	return __is_kernel_percpu_address(addr, NULL);
2358}
2359
2360/**
2361 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2362 * @addr: the address to be converted to physical address
2363 *
2364 * Given @addr which is dereferenceable address obtained via one of
2365 * percpu access macros, this function translates it into its physical
2366 * address.  The caller is responsible for ensuring @addr stays valid
2367 * until this function finishes.
2368 *
2369 * percpu allocator has special setup for the first chunk, which currently
2370 * supports either embedding in linear address space or vmalloc mapping,
2371 * and, from the second one, the backing allocator (currently either vm or
2372 * km) provides translation.
2373 *
2374 * The addr can be translated simply without checking if it falls into the
2375 * first chunk. But the current code reflects better how percpu allocator
2376 * actually works, and the verification can discover both bugs in percpu
2377 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2378 * code.
2379 *
2380 * RETURNS:
2381 * The physical address for @addr.
2382 */
2383phys_addr_t per_cpu_ptr_to_phys(void *addr)
2384{
2385	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2386	bool in_first_chunk = false;
2387	unsigned long first_low, first_high;
2388	unsigned int cpu;
2389
2390	/*
2391	 * The following test on unit_low/high isn't strictly
2392	 * necessary but will speed up lookups of addresses which
2393	 * aren't in the first chunk.
2394	 *
2395	 * The address check is against full chunk sizes.  pcpu_base_addr
2396	 * points to the beginning of the first chunk including the
2397	 * static region.  Assumes good intent as the first chunk may
2398	 * not be full (ie. < pcpu_unit_pages in size).
2399	 */
2400	first_low = (unsigned long)pcpu_base_addr +
2401		    pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2402	first_high = (unsigned long)pcpu_base_addr +
2403		     pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2404	if ((unsigned long)addr >= first_low &&
2405	    (unsigned long)addr < first_high) {
2406		for_each_possible_cpu(cpu) {
2407			void *start = per_cpu_ptr(base, cpu);
2408
2409			if (addr >= start && addr < start + pcpu_unit_size) {
2410				in_first_chunk = true;
2411				break;
2412			}
2413		}
2414	}
2415
2416	if (in_first_chunk) {
2417		if (!is_vmalloc_addr(addr))
2418			return __pa(addr);
2419		else
2420			return page_to_phys(vmalloc_to_page(addr)) +
2421			       offset_in_page(addr);
2422	} else
2423		return page_to_phys(pcpu_addr_to_page(addr)) +
2424		       offset_in_page(addr);
2425}
2426
2427/**
2428 * pcpu_alloc_alloc_info - allocate percpu allocation info
2429 * @nr_groups: the number of groups
2430 * @nr_units: the number of units
2431 *
2432 * Allocate ai which is large enough for @nr_groups groups containing
2433 * @nr_units units.  The returned ai's groups[0].cpu_map points to the
2434 * cpu_map array which is long enough for @nr_units and filled with
2435 * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
2436 * pointer of other groups.
2437 *
2438 * RETURNS:
2439 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2440 * failure.
2441 */
2442struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2443						      int nr_units)
2444{
2445	struct pcpu_alloc_info *ai;
2446	size_t base_size, ai_size;
2447	void *ptr;
2448	int unit;
2449
2450	base_size = ALIGN(struct_size(ai, groups, nr_groups),
2451			  __alignof__(ai->groups[0].cpu_map[0]));
2452	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2453
2454	ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2455	if (!ptr)
2456		return NULL;
2457	ai = ptr;
2458	ptr += base_size;
2459
2460	ai->groups[0].cpu_map = ptr;
2461
2462	for (unit = 0; unit < nr_units; unit++)
2463		ai->groups[0].cpu_map[unit] = NR_CPUS;
2464
2465	ai->nr_groups = nr_groups;
2466	ai->__ai_size = PFN_ALIGN(ai_size);
2467
2468	return ai;
2469}
2470
2471/**
2472 * pcpu_free_alloc_info - free percpu allocation info
2473 * @ai: pcpu_alloc_info to free
2474 *
2475 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2476 */
2477void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2478{
2479	memblock_free_early(__pa(ai), ai->__ai_size);
2480}
2481
2482/**
2483 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2484 * @lvl: loglevel
2485 * @ai: allocation info to dump
2486 *
2487 * Print out information about @ai using loglevel @lvl.
2488 */
2489static void pcpu_dump_alloc_info(const char *lvl,
2490				 const struct pcpu_alloc_info *ai)
2491{
2492	int group_width = 1, cpu_width = 1, width;
2493	char empty_str[] = "--------";
2494	int alloc = 0, alloc_end = 0;
2495	int group, v;
2496	int upa, apl;	/* units per alloc, allocs per line */
2497
2498	v = ai->nr_groups;
2499	while (v /= 10)
2500		group_width++;
2501
2502	v = num_possible_cpus();
2503	while (v /= 10)
2504		cpu_width++;
2505	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2506
2507	upa = ai->alloc_size / ai->unit_size;
2508	width = upa * (cpu_width + 1) + group_width + 3;
2509	apl = rounddown_pow_of_two(max(60 / width, 1));
2510
2511	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2512	       lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2513	       ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2514
2515	for (group = 0; group < ai->nr_groups; group++) {
2516		const struct pcpu_group_info *gi = &ai->groups[group];
2517		int unit = 0, unit_end = 0;
2518
2519		BUG_ON(gi->nr_units % upa);
2520		for (alloc_end += gi->nr_units / upa;
2521		     alloc < alloc_end; alloc++) {
2522			if (!(alloc % apl)) {
2523				pr_cont("\n");
2524				printk("%spcpu-alloc: ", lvl);
2525			}
2526			pr_cont("[%0*d] ", group_width, group);
2527
2528			for (unit_end += upa; unit < unit_end; unit++)
2529				if (gi->cpu_map[unit] != NR_CPUS)
2530					pr_cont("%0*d ",
2531						cpu_width, gi->cpu_map[unit]);
2532				else
2533					pr_cont("%s ", empty_str);
2534		}
2535	}
2536	pr_cont("\n");
2537}
2538
2539/**
2540 * pcpu_setup_first_chunk - initialize the first percpu chunk
2541 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2542 * @base_addr: mapped address
2543 *
2544 * Initialize the first percpu chunk which contains the kernel static
2545 * percpu area.  This function is to be called from arch percpu area
2546 * setup path.
2547 *
2548 * @ai contains all information necessary to initialize the first
2549 * chunk and prime the dynamic percpu allocator.
2550 *
2551 * @ai->static_size is the size of static percpu area.
2552 *
2553 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2554 * reserve after the static area in the first chunk.  This reserves
2555 * the first chunk such that it's available only through reserved
2556 * percpu allocation.  This is primarily used to serve module percpu
2557 * static areas on architectures where the addressing model has
2558 * limited offset range for symbol relocations to guarantee module
2559 * percpu symbols fall inside the relocatable range.
2560 *
2561 * @ai->dyn_size determines the number of bytes available for dynamic
2562 * allocation in the first chunk.  The area between @ai->static_size +
2563 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2564 *
2565 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2566 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2567 * @ai->dyn_size.
2568 *
2569 * @ai->atom_size is the allocation atom size and used as alignment
2570 * for vm areas.
2571 *
2572 * @ai->alloc_size is the allocation size and always multiple of
2573 * @ai->atom_size.  This is larger than @ai->atom_size if
2574 * @ai->unit_size is larger than @ai->atom_size.
2575 *
2576 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2577 * percpu areas.  Units which should be colocated are put into the
2578 * same group.  Dynamic VM areas will be allocated according to these
2579 * groupings.  If @ai->nr_groups is zero, a single group containing
2580 * all units is assumed.
2581 *
2582 * The caller should have mapped the first chunk at @base_addr and
2583 * copied static data to each unit.
2584 *
2585 * The first chunk will always contain a static and a dynamic region.
2586 * However, the static region is not managed by any chunk.  If the first
2587 * chunk also contains a reserved region, it is served by two chunks -
2588 * one for the reserved region and one for the dynamic region.  They
2589 * share the same vm, but use offset regions in the area allocation map.
2590 * The chunk serving the dynamic region is circulated in the chunk slots
2591 * and available for dynamic allocation like any other chunk.
 
 
2592 */
2593void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2594				   void *base_addr)
2595{
2596	size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2597	size_t static_size, dyn_size;
2598	struct pcpu_chunk *chunk;
 
 
2599	unsigned long *group_offsets;
2600	size_t *group_sizes;
2601	unsigned long *unit_off;
2602	unsigned int cpu;
2603	int *unit_map;
2604	int group, unit, i;
2605	int map_size;
2606	unsigned long tmp_addr;
2607	size_t alloc_size;
2608
2609#define PCPU_SETUP_BUG_ON(cond)	do {					\
2610	if (unlikely(cond)) {						\
2611		pr_emerg("failed to initialize, %s\n", #cond);		\
2612		pr_emerg("cpu_possible_mask=%*pb\n",			\
2613			 cpumask_pr_args(cpu_possible_mask));		\
2614		pcpu_dump_alloc_info(KERN_EMERG, ai);			\
2615		BUG();							\
2616	}								\
2617} while (0)
2618
2619	/* sanity checks */
2620	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2621#ifdef CONFIG_SMP
2622	PCPU_SETUP_BUG_ON(!ai->static_size);
2623	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2624#endif
2625	PCPU_SETUP_BUG_ON(!base_addr);
2626	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2627	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2628	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2629	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2630	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2631	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2632	PCPU_SETUP_BUG_ON(!ai->dyn_size);
2633	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2634	PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2635			    IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2636	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2637
2638	/* process group information and build config tables accordingly */
2639	alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2640	group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2641	if (!group_offsets)
2642		panic("%s: Failed to allocate %zu bytes\n", __func__,
2643		      alloc_size);
2644
2645	alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2646	group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2647	if (!group_sizes)
2648		panic("%s: Failed to allocate %zu bytes\n", __func__,
2649		      alloc_size);
2650
2651	alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2652	unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2653	if (!unit_map)
2654		panic("%s: Failed to allocate %zu bytes\n", __func__,
2655		      alloc_size);
2656
2657	alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2658	unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2659	if (!unit_off)
2660		panic("%s: Failed to allocate %zu bytes\n", __func__,
2661		      alloc_size);
2662
2663	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2664		unit_map[cpu] = UINT_MAX;
2665
2666	pcpu_low_unit_cpu = NR_CPUS;
2667	pcpu_high_unit_cpu = NR_CPUS;
2668
2669	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2670		const struct pcpu_group_info *gi = &ai->groups[group];
2671
2672		group_offsets[group] = gi->base_offset;
2673		group_sizes[group] = gi->nr_units * ai->unit_size;
2674
2675		for (i = 0; i < gi->nr_units; i++) {
2676			cpu = gi->cpu_map[i];
2677			if (cpu == NR_CPUS)
2678				continue;
2679
2680			PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2681			PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2682			PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2683
2684			unit_map[cpu] = unit + i;
2685			unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2686
2687			/* determine low/high unit_cpu */
2688			if (pcpu_low_unit_cpu == NR_CPUS ||
2689			    unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2690				pcpu_low_unit_cpu = cpu;
2691			if (pcpu_high_unit_cpu == NR_CPUS ||
2692			    unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2693				pcpu_high_unit_cpu = cpu;
2694		}
2695	}
2696	pcpu_nr_units = unit;
2697
2698	for_each_possible_cpu(cpu)
2699		PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2700
2701	/* we're done parsing the input, undefine BUG macro and dump config */
2702#undef PCPU_SETUP_BUG_ON
2703	pcpu_dump_alloc_info(KERN_DEBUG, ai);
2704
2705	pcpu_nr_groups = ai->nr_groups;
2706	pcpu_group_offsets = group_offsets;
2707	pcpu_group_sizes = group_sizes;
2708	pcpu_unit_map = unit_map;
2709	pcpu_unit_offsets = unit_off;
2710
2711	/* determine basic parameters */
2712	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2713	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2714	pcpu_atom_size = ai->atom_size;
2715	pcpu_chunk_struct_size = struct_size(chunk, populated,
2716					     BITS_TO_LONGS(pcpu_unit_pages));
2717
2718	pcpu_stats_save_ai(ai);
2719
2720	/*
2721	 * Allocate chunk slots.  The slots after the active slots are:
2722	 *   sidelined_slot - isolated, depopulated chunks
2723	 *   free_slot - fully free chunks
2724	 *   to_depopulate_slot - isolated, chunks to depopulate
2725	 */
2726	pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2727	pcpu_free_slot = pcpu_sidelined_slot + 1;
2728	pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2729	pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2730	pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2731					  sizeof(pcpu_chunk_lists[0]),
2732					  SMP_CACHE_BYTES);
2733	if (!pcpu_chunk_lists)
2734		panic("%s: Failed to allocate %zu bytes\n", __func__,
2735		      pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2736
2737	for (i = 0; i < pcpu_nr_slots; i++)
2738		INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2739
2740	/*
2741	 * The end of the static region needs to be aligned with the
2742	 * minimum allocation size as this offsets the reserved and
2743	 * dynamic region.  The first chunk ends page aligned by
2744	 * expanding the dynamic region, therefore the dynamic region
2745	 * can be shrunk to compensate while still staying above the
2746	 * configured sizes.
2747	 */
2748	static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2749	dyn_size = ai->dyn_size - (static_size - ai->static_size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2750
2751	/*
2752	 * Initialize first chunk.
2753	 * If the reserved_size is non-zero, this initializes the reserved
2754	 * chunk.  If the reserved_size is zero, the reserved chunk is NULL
2755	 * and the dynamic region is initialized here.  The first chunk,
2756	 * pcpu_first_chunk, will always point to the chunk that serves
2757	 * the dynamic region.
2758	 */
2759	tmp_addr = (unsigned long)base_addr + static_size;
2760	map_size = ai->reserved_size ?: dyn_size;
2761	chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2762
2763	/* init dynamic chunk if necessary */
2764	if (ai->reserved_size) {
2765		pcpu_reserved_chunk = chunk;
2766
2767		tmp_addr = (unsigned long)base_addr + static_size +
2768			   ai->reserved_size;
2769		map_size = dyn_size;
2770		chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
 
 
 
 
 
 
 
 
 
2771	}
2772
2773	/* link the first chunk in */
2774	pcpu_first_chunk = chunk;
2775	pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
 
2776	pcpu_chunk_relocate(pcpu_first_chunk, -1);
2777
2778	/* include all regions of the first chunk */
2779	pcpu_nr_populated += PFN_DOWN(size_sum);
2780
2781	pcpu_stats_chunk_alloc();
2782	trace_percpu_create_chunk(base_addr);
2783
2784	/* we're done */
2785	pcpu_base_addr = base_addr;
 
2786}
2787
2788#ifdef CONFIG_SMP
2789
2790const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2791	[PCPU_FC_AUTO]	= "auto",
2792	[PCPU_FC_EMBED]	= "embed",
2793	[PCPU_FC_PAGE]	= "page",
2794};
2795
2796enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2797
2798static int __init percpu_alloc_setup(char *str)
2799{
2800	if (!str)
2801		return -EINVAL;
2802
2803	if (0)
2804		/* nada */;
2805#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2806	else if (!strcmp(str, "embed"))
2807		pcpu_chosen_fc = PCPU_FC_EMBED;
2808#endif
2809#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2810	else if (!strcmp(str, "page"))
2811		pcpu_chosen_fc = PCPU_FC_PAGE;
2812#endif
2813	else
2814		pr_warn("unknown allocator %s specified\n", str);
2815
2816	return 0;
2817}
2818early_param("percpu_alloc", percpu_alloc_setup);
2819
2820/*
2821 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2822 * Build it if needed by the arch config or the generic setup is going
2823 * to be used.
2824 */
2825#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2826	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2827#define BUILD_EMBED_FIRST_CHUNK
2828#endif
2829
2830/* build pcpu_page_first_chunk() iff needed by the arch config */
2831#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2832#define BUILD_PAGE_FIRST_CHUNK
2833#endif
2834
2835/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2836#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2837/**
2838 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2839 * @reserved_size: the size of reserved percpu area in bytes
2840 * @dyn_size: minimum free size for dynamic allocation in bytes
2841 * @atom_size: allocation atom size
2842 * @cpu_distance_fn: callback to determine distance between cpus, optional
2843 *
2844 * This function determines grouping of units, their mappings to cpus
2845 * and other parameters considering needed percpu size, allocation
2846 * atom size and distances between CPUs.
2847 *
2848 * Groups are always multiples of atom size and CPUs which are of
2849 * LOCAL_DISTANCE both ways are grouped together and share space for
2850 * units in the same group.  The returned configuration is guaranteed
2851 * to have CPUs on different nodes on different groups and >=75% usage
2852 * of allocated virtual address space.
2853 *
2854 * RETURNS:
2855 * On success, pointer to the new allocation_info is returned.  On
2856 * failure, ERR_PTR value is returned.
2857 */
2858static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2859				size_t reserved_size, size_t dyn_size,
2860				size_t atom_size,
2861				pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2862{
2863	static int group_map[NR_CPUS] __initdata;
2864	static int group_cnt[NR_CPUS] __initdata;
2865	static struct cpumask mask __initdata;
2866	const size_t static_size = __per_cpu_end - __per_cpu_start;
2867	int nr_groups = 1, nr_units = 0;
2868	size_t size_sum, min_unit_size, alloc_size;
2869	int upa, max_upa, best_upa;	/* units_per_alloc */
2870	int last_allocs, group, unit;
2871	unsigned int cpu, tcpu;
2872	struct pcpu_alloc_info *ai;
2873	unsigned int *cpu_map;
2874
2875	/* this function may be called multiple times */
2876	memset(group_map, 0, sizeof(group_map));
2877	memset(group_cnt, 0, sizeof(group_cnt));
2878	cpumask_clear(&mask);
2879
2880	/* calculate size_sum and ensure dyn_size is enough for early alloc */
2881	size_sum = PFN_ALIGN(static_size + reserved_size +
2882			    max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2883	dyn_size = size_sum - static_size - reserved_size;
2884
2885	/*
2886	 * Determine min_unit_size, alloc_size and max_upa such that
2887	 * alloc_size is multiple of atom_size and is the smallest
2888	 * which can accommodate 4k aligned segments which are equal to
2889	 * or larger than min_unit_size.
2890	 */
2891	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2892
2893	/* determine the maximum # of units that can fit in an allocation */
2894	alloc_size = roundup(min_unit_size, atom_size);
2895	upa = alloc_size / min_unit_size;
2896	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2897		upa--;
2898	max_upa = upa;
2899
2900	cpumask_copy(&mask, cpu_possible_mask);
2901
2902	/* group cpus according to their proximity */
2903	for (group = 0; !cpumask_empty(&mask); group++) {
2904		/* pop the group's first cpu */
2905		cpu = cpumask_first(&mask);
 
 
 
 
 
 
 
 
 
 
 
2906		group_map[cpu] = group;
2907		group_cnt[group]++;
2908		cpumask_clear_cpu(cpu, &mask);
2909
2910		for_each_cpu(tcpu, &mask) {
2911			if (!cpu_distance_fn ||
2912			    (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2913			     cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2914				group_map[tcpu] = group;
2915				group_cnt[group]++;
2916				cpumask_clear_cpu(tcpu, &mask);
2917			}
2918		}
2919	}
2920	nr_groups = group;
2921
2922	/*
2923	 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2924	 * Expand the unit_size until we use >= 75% of the units allocated.
2925	 * Related to atom_size, which could be much larger than the unit_size.
2926	 */
2927	last_allocs = INT_MAX;
2928	best_upa = 0;
2929	for (upa = max_upa; upa; upa--) {
2930		int allocs = 0, wasted = 0;
2931
2932		if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2933			continue;
2934
2935		for (group = 0; group < nr_groups; group++) {
2936			int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2937			allocs += this_allocs;
2938			wasted += this_allocs * upa - group_cnt[group];
2939		}
2940
2941		/*
2942		 * Don't accept if wastage is over 1/3.  The
2943		 * greater-than comparison ensures upa==1 always
2944		 * passes the following check.
2945		 */
2946		if (wasted > num_possible_cpus() / 3)
2947			continue;
2948
2949		/* and then don't consume more memory */
2950		if (allocs > last_allocs)
2951			break;
2952		last_allocs = allocs;
2953		best_upa = upa;
2954	}
2955	BUG_ON(!best_upa);
2956	upa = best_upa;
2957
2958	/* allocate and fill alloc_info */
2959	for (group = 0; group < nr_groups; group++)
2960		nr_units += roundup(group_cnt[group], upa);
2961
2962	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2963	if (!ai)
2964		return ERR_PTR(-ENOMEM);
2965	cpu_map = ai->groups[0].cpu_map;
2966
2967	for (group = 0; group < nr_groups; group++) {
2968		ai->groups[group].cpu_map = cpu_map;
2969		cpu_map += roundup(group_cnt[group], upa);
2970	}
2971
2972	ai->static_size = static_size;
2973	ai->reserved_size = reserved_size;
2974	ai->dyn_size = dyn_size;
2975	ai->unit_size = alloc_size / upa;
2976	ai->atom_size = atom_size;
2977	ai->alloc_size = alloc_size;
2978
2979	for (group = 0, unit = 0; group < nr_groups; group++) {
2980		struct pcpu_group_info *gi = &ai->groups[group];
2981
2982		/*
2983		 * Initialize base_offset as if all groups are located
2984		 * back-to-back.  The caller should update this to
2985		 * reflect actual allocation.
2986		 */
2987		gi->base_offset = unit * ai->unit_size;
2988
2989		for_each_possible_cpu(cpu)
2990			if (group_map[cpu] == group)
2991				gi->cpu_map[gi->nr_units++] = cpu;
2992		gi->nr_units = roundup(gi->nr_units, upa);
2993		unit += gi->nr_units;
2994	}
2995	BUG_ON(unit != nr_units);
2996
2997	return ai;
2998}
2999#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3000
3001#if defined(BUILD_EMBED_FIRST_CHUNK)
3002/**
3003 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3004 * @reserved_size: the size of reserved percpu area in bytes
3005 * @dyn_size: minimum free size for dynamic allocation in bytes
3006 * @atom_size: allocation atom size
3007 * @cpu_distance_fn: callback to determine distance between cpus, optional
3008 * @alloc_fn: function to allocate percpu page
3009 * @free_fn: function to free percpu page
3010 *
3011 * This is a helper to ease setting up embedded first percpu chunk and
3012 * can be called where pcpu_setup_first_chunk() is expected.
3013 *
3014 * If this function is used to setup the first chunk, it is allocated
3015 * by calling @alloc_fn and used as-is without being mapped into
3016 * vmalloc area.  Allocations are always whole multiples of @atom_size
3017 * aligned to @atom_size.
3018 *
3019 * This enables the first chunk to piggy back on the linear physical
3020 * mapping which often uses larger page size.  Please note that this
3021 * can result in very sparse cpu->unit mapping on NUMA machines thus
3022 * requiring large vmalloc address space.  Don't use this allocator if
3023 * vmalloc space is not orders of magnitude larger than distances
3024 * between node memory addresses (ie. 32bit NUMA machines).
3025 *
3026 * @dyn_size specifies the minimum dynamic area size.
3027 *
3028 * If the needed size is smaller than the minimum or specified unit
3029 * size, the leftover is returned using @free_fn.
3030 *
3031 * RETURNS:
3032 * 0 on success, -errno on failure.
3033 */
3034int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3035				  size_t atom_size,
3036				  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3037				  pcpu_fc_alloc_fn_t alloc_fn,
3038				  pcpu_fc_free_fn_t free_fn)
3039{
3040	void *base = (void *)ULONG_MAX;
3041	void **areas = NULL;
3042	struct pcpu_alloc_info *ai;
3043	size_t size_sum, areas_size;
3044	unsigned long max_distance;
3045	int group, i, highest_group, rc = 0;
3046
3047	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3048				   cpu_distance_fn);
3049	if (IS_ERR(ai))
3050		return PTR_ERR(ai);
3051
3052	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3053	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3054
3055	areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3056	if (!areas) {
3057		rc = -ENOMEM;
3058		goto out_free;
3059	}
3060
3061	/* allocate, copy and determine base address & max_distance */
3062	highest_group = 0;
3063	for (group = 0; group < ai->nr_groups; group++) {
3064		struct pcpu_group_info *gi = &ai->groups[group];
3065		unsigned int cpu = NR_CPUS;
3066		void *ptr;
3067
3068		for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3069			cpu = gi->cpu_map[i];
3070		BUG_ON(cpu == NR_CPUS);
3071
3072		/* allocate space for the whole group */
3073		ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
3074		if (!ptr) {
3075			rc = -ENOMEM;
3076			goto out_free_areas;
3077		}
3078		/* kmemleak tracks the percpu allocations separately */
3079		kmemleak_free(ptr);
3080		areas[group] = ptr;
3081
3082		base = min(ptr, base);
3083		if (ptr > areas[highest_group])
3084			highest_group = group;
3085	}
3086	max_distance = areas[highest_group] - base;
3087	max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3088
3089	/* warn if maximum distance is further than 75% of vmalloc space */
3090	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3091		pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3092				max_distance, VMALLOC_TOTAL);
3093#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3094		/* and fail if we have fallback */
3095		rc = -EINVAL;
3096		goto out_free_areas;
3097#endif
3098	}
3099
3100	/*
3101	 * Copy data and free unused parts.  This should happen after all
3102	 * allocations are complete; otherwise, we may end up with
3103	 * overlapping groups.
3104	 */
3105	for (group = 0; group < ai->nr_groups; group++) {
3106		struct pcpu_group_info *gi = &ai->groups[group];
3107		void *ptr = areas[group];
3108
3109		for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3110			if (gi->cpu_map[i] == NR_CPUS) {
3111				/* unused unit, free whole */
3112				free_fn(ptr, ai->unit_size);
3113				continue;
3114			}
3115			/* copy and return the unused part */
3116			memcpy(ptr, __per_cpu_load, ai->static_size);
3117			free_fn(ptr + size_sum, ai->unit_size - size_sum);
3118		}
3119	}
3120
3121	/* base address is now known, determine group base offsets */
 
3122	for (group = 0; group < ai->nr_groups; group++) {
3123		ai->groups[group].base_offset = areas[group] - base;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3124	}
3125
3126	pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3127		PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3128		ai->dyn_size, ai->unit_size);
3129
3130	pcpu_setup_first_chunk(ai, base);
3131	goto out_free;
3132
3133out_free_areas:
3134	for (group = 0; group < ai->nr_groups; group++)
3135		if (areas[group])
3136			free_fn(areas[group],
3137				ai->groups[group].nr_units * ai->unit_size);
3138out_free:
3139	pcpu_free_alloc_info(ai);
3140	if (areas)
3141		memblock_free_early(__pa(areas), areas_size);
3142	return rc;
3143}
3144#endif /* BUILD_EMBED_FIRST_CHUNK */
3145
3146#ifdef BUILD_PAGE_FIRST_CHUNK
3147/**
3148 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3149 * @reserved_size: the size of reserved percpu area in bytes
3150 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
3151 * @free_fn: function to free percpu page, always called with PAGE_SIZE
3152 * @populate_pte_fn: function to populate pte
3153 *
3154 * This is a helper to ease setting up page-remapped first percpu
3155 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3156 *
3157 * This is the basic allocator.  Static percpu area is allocated
3158 * page-by-page into vmalloc area.
3159 *
3160 * RETURNS:
3161 * 0 on success, -errno on failure.
3162 */
3163int __init pcpu_page_first_chunk(size_t reserved_size,
3164				 pcpu_fc_alloc_fn_t alloc_fn,
3165				 pcpu_fc_free_fn_t free_fn,
3166				 pcpu_fc_populate_pte_fn_t populate_pte_fn)
3167{
3168	static struct vm_struct vm;
3169	struct pcpu_alloc_info *ai;
3170	char psize_str[16];
3171	int unit_pages;
3172	size_t pages_size;
3173	struct page **pages;
3174	int unit, i, j, rc = 0;
3175	int upa;
3176	int nr_g0_units;
3177
3178	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3179
3180	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3181	if (IS_ERR(ai))
3182		return PTR_ERR(ai);
3183	BUG_ON(ai->nr_groups != 1);
3184	upa = ai->alloc_size/ai->unit_size;
3185	nr_g0_units = roundup(num_possible_cpus(), upa);
3186	if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3187		pcpu_free_alloc_info(ai);
3188		return -EINVAL;
3189	}
3190
3191	unit_pages = ai->unit_size >> PAGE_SHIFT;
3192
3193	/* unaligned allocations can't be freed, round up to page size */
3194	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3195			       sizeof(pages[0]));
3196	pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3197	if (!pages)
3198		panic("%s: Failed to allocate %zu bytes\n", __func__,
3199		      pages_size);
3200
3201	/* allocate pages */
3202	j = 0;
3203	for (unit = 0; unit < num_possible_cpus(); unit++) {
3204		unsigned int cpu = ai->groups[0].cpu_map[unit];
3205		for (i = 0; i < unit_pages; i++) {
 
3206			void *ptr;
3207
3208			ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
3209			if (!ptr) {
3210				pr_warn("failed to allocate %s page for cpu%u\n",
3211						psize_str, cpu);
3212				goto enomem;
3213			}
3214			/* kmemleak tracks the percpu allocations separately */
3215			kmemleak_free(ptr);
3216			pages[j++] = virt_to_page(ptr);
3217		}
3218	}
3219
3220	/* allocate vm area, map the pages and copy static data */
3221	vm.flags = VM_ALLOC;
3222	vm.size = num_possible_cpus() * ai->unit_size;
3223	vm_area_register_early(&vm, PAGE_SIZE);
3224
3225	for (unit = 0; unit < num_possible_cpus(); unit++) {
3226		unsigned long unit_addr =
3227			(unsigned long)vm.addr + unit * ai->unit_size;
3228
3229		for (i = 0; i < unit_pages; i++)
3230			populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
3231
3232		/* pte already populated, the following shouldn't fail */
3233		rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3234				      unit_pages);
3235		if (rc < 0)
3236			panic("failed to map percpu area, err=%d\n", rc);
3237
3238		/*
3239		 * FIXME: Archs with virtual cache should flush local
3240		 * cache for the linear mapping here - something
3241		 * equivalent to flush_cache_vmap() on the local cpu.
3242		 * flush_cache_vmap() can't be used as most supporting
3243		 * data structures are not set up yet.
3244		 */
3245
3246		/* copy static data */
3247		memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3248	}
3249
3250	/* we're ready, commit */
3251	pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3252		unit_pages, psize_str, ai->static_size,
3253		ai->reserved_size, ai->dyn_size);
3254
3255	pcpu_setup_first_chunk(ai, vm.addr);
3256	goto out_free_ar;
3257
3258enomem:
3259	while (--j >= 0)
3260		free_fn(page_address(pages[j]), PAGE_SIZE);
3261	rc = -ENOMEM;
3262out_free_ar:
3263	memblock_free_early(__pa(pages), pages_size);
3264	pcpu_free_alloc_info(ai);
3265	return rc;
3266}
3267#endif /* BUILD_PAGE_FIRST_CHUNK */
3268
3269#ifndef	CONFIG_HAVE_SETUP_PER_CPU_AREA
3270/*
3271 * Generic SMP percpu area setup.
3272 *
3273 * The embedding helper is used because its behavior closely resembles
3274 * the original non-dynamic generic percpu area setup.  This is
3275 * important because many archs have addressing restrictions and might
3276 * fail if the percpu area is located far away from the previous
3277 * location.  As an added bonus, in non-NUMA cases, embedding is
3278 * generally a good idea TLB-wise because percpu area can piggy back
3279 * on the physical linear memory mapping which uses large page
3280 * mappings on applicable archs.
3281 */
3282unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3283EXPORT_SYMBOL(__per_cpu_offset);
3284
3285static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
3286				       size_t align)
3287{
3288	return  memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
 
3289}
3290
3291static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
3292{
3293	memblock_free_early(__pa(ptr), size);
3294}
3295
3296void __init setup_per_cpu_areas(void)
3297{
3298	unsigned long delta;
3299	unsigned int cpu;
3300	int rc;
3301
3302	/*
3303	 * Always reserve area for module percpu variables.  That's
3304	 * what the legacy allocator did.
3305	 */
3306	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
3307				    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
3308				    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
3309	if (rc < 0)
3310		panic("Failed to initialize percpu areas.");
3311
3312	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3313	for_each_possible_cpu(cpu)
3314		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3315}
3316#endif	/* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3317
3318#else	/* CONFIG_SMP */
3319
3320/*
3321 * UP percpu area setup.
3322 *
3323 * UP always uses km-based percpu allocator with identity mapping.
3324 * Static percpu variables are indistinguishable from the usual static
3325 * variables and don't require any special preparation.
3326 */
3327void __init setup_per_cpu_areas(void)
3328{
3329	const size_t unit_size =
3330		roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3331					 PERCPU_DYNAMIC_RESERVE));
3332	struct pcpu_alloc_info *ai;
3333	void *fc;
3334
3335	ai = pcpu_alloc_alloc_info(1, 1);
3336	fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 
 
3337	if (!ai || !fc)
3338		panic("Failed to allocate memory for percpu areas.");
3339	/* kmemleak tracks the percpu allocations separately */
3340	kmemleak_free(fc);
3341
3342	ai->dyn_size = unit_size;
3343	ai->unit_size = unit_size;
3344	ai->atom_size = unit_size;
3345	ai->alloc_size = unit_size;
3346	ai->groups[0].nr_units = 1;
3347	ai->groups[0].cpu_map[0] = 0;
3348
3349	pcpu_setup_first_chunk(ai, fc);
3350	pcpu_free_alloc_info(ai);
3351}
3352
3353#endif	/* CONFIG_SMP */
3354
3355/*
3356 * pcpu_nr_pages - calculate total number of populated backing pages
3357 *
3358 * This reflects the number of pages populated to back chunks.  Metadata is
3359 * excluded in the number exposed in meminfo as the number of backing pages
3360 * scales with the number of cpus and can quickly outweigh the memory used for
3361 * metadata.  It also keeps this calculation nice and simple.
3362 *
3363 * RETURNS:
3364 * Total number of populated backing pages in use by the allocator.
3365 */
3366unsigned long pcpu_nr_pages(void)
3367{
3368	return pcpu_nr_populated * pcpu_nr_units;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3369}
3370
3371/*
3372 * Percpu allocator is initialized early during boot when neither slab or
3373 * workqueue is available.  Plug async management until everything is up
3374 * and running.
3375 */
3376static int __init percpu_enable_async(void)
3377{
3378	pcpu_async_enabled = true;
3379	return 0;
3380}
3381subsys_initcall(percpu_enable_async);