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);