Linux Audio

Check our new training course

Loading...
v4.17
 
   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 * Copyright (C) 2017		Facebook Inc.
   8 * Copyright (C) 2017		Dennis Zhou <dennisszhou@gmail.com>
   9 *
  10 * This file is released under the GPLv2 license.
  11 *
  12 * The percpu allocator handles both static and dynamic areas.  Percpu
  13 * areas are allocated in chunks which are divided into units.  There is
  14 * a 1-to-1 mapping for units to possible cpus.  These units are grouped
  15 * based on NUMA properties of the machine.
  16 *
  17 *  c0                           c1                         c2
  18 *  -------------------          -------------------        ------------
  19 * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
  20 *  -------------------  ......  -------------------  ....  ------------
  21 *
  22 * Allocation is done by offsets into a unit's address space.  Ie., an
  23 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
  24 * c1:u1, c1:u2, etc.  On NUMA machines, the mapping may be non-linear
  25 * and even sparse.  Access is handled by configuring percpu base
  26 * registers according to the cpu to unit mappings and offsetting the
  27 * base address using pcpu_unit_size.
  28 *
  29 * There is special consideration for the first chunk which must handle
  30 * the static percpu variables in the kernel image as allocation services
  31 * are not online yet.  In short, the first chunk is structured like so:
  32 *
  33 *                  <Static | [Reserved] | Dynamic>
  34 *
  35 * The static data is copied from the original section managed by the
  36 * linker.  The reserved section, if non-zero, primarily manages static
  37 * percpu variables from kernel modules.  Finally, the dynamic section
  38 * takes care of normal allocations.
  39 *
  40 * The allocator organizes chunks into lists according to free size and
  41 * tries to allocate from the fullest chunk first.  Each chunk is managed
  42 * by a bitmap with metadata blocks.  The allocation map is updated on
  43 * every allocation and free to reflect the current state while the boundary
 
 
 
 
 
  44 * map is only updated on allocation.  Each metadata block contains
  45 * information to help mitigate the need to iterate over large portions
  46 * of the bitmap.  The reverse mapping from page to chunk is stored in
  47 * the page's index.  Lastly, units are lazily backed and grow in unison.
  48 *
  49 * There is a unique conversion that goes on here between bytes and bits.
  50 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE.  The chunk
  51 * tracks the number of pages it is responsible for in nr_pages.  Helper
  52 * functions are used to convert from between the bytes, bits, and blocks.
  53 * All hints are managed in bits unless explicitly stated.
  54 *
  55 * To use this allocator, arch code should do the following:
  56 *
  57 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
  58 *   regular address to percpu pointer and back if they need to be
  59 *   different from the default
  60 *
  61 * - use pcpu_setup_first_chunk() during percpu area initialization to
  62 *   setup the first chunk containing the kernel static percpu area
  63 */
  64
  65#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  66
  67#include <linux/bitmap.h>
  68#include <linux/bootmem.h>
 
  69#include <linux/err.h>
  70#include <linux/lcm.h>
  71#include <linux/list.h>
  72#include <linux/log2.h>
  73#include <linux/mm.h>
  74#include <linux/module.h>
  75#include <linux/mutex.h>
  76#include <linux/percpu.h>
  77#include <linux/pfn.h>
  78#include <linux/slab.h>
  79#include <linux/spinlock.h>
  80#include <linux/vmalloc.h>
  81#include <linux/workqueue.h>
  82#include <linux/kmemleak.h>
  83#include <linux/sched.h>
 
 
  84
  85#include <asm/cacheflush.h>
  86#include <asm/sections.h>
  87#include <asm/tlbflush.h>
  88#include <asm/io.h>
  89
  90#define CREATE_TRACE_POINTS
  91#include <trace/events/percpu.h>
  92
  93#include "percpu-internal.h"
  94
  95/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
 
 
 
  96#define PCPU_SLOT_BASE_SHIFT		5
 
 
  97
  98#define PCPU_EMPTY_POP_PAGES_LOW	2
  99#define PCPU_EMPTY_POP_PAGES_HIGH	4
 100
 101#ifdef CONFIG_SMP
 102/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 103#ifndef __addr_to_pcpu_ptr
 104#define __addr_to_pcpu_ptr(addr)					\
 105	(void __percpu *)((unsigned long)(addr) -			\
 106			  (unsigned long)pcpu_base_addr	+		\
 107			  (unsigned long)__per_cpu_start)
 108#endif
 109#ifndef __pcpu_ptr_to_addr
 110#define __pcpu_ptr_to_addr(ptr)						\
 111	(void __force *)((unsigned long)(ptr) +				\
 112			 (unsigned long)pcpu_base_addr -		\
 113			 (unsigned long)__per_cpu_start)
 114#endif
 115#else	/* CONFIG_SMP */
 116/* on UP, it's always identity mapped */
 117#define __addr_to_pcpu_ptr(addr)	(void __percpu *)(addr)
 118#define __pcpu_ptr_to_addr(ptr)		(void __force *)(ptr)
 119#endif	/* CONFIG_SMP */
 120
 121static int pcpu_unit_pages __ro_after_init;
 122static int pcpu_unit_size __ro_after_init;
 123static int pcpu_nr_units __ro_after_init;
 124static int pcpu_atom_size __ro_after_init;
 125int pcpu_nr_slots __ro_after_init;
 
 
 
 126static size_t pcpu_chunk_struct_size __ro_after_init;
 127
 128/* cpus with the lowest and highest unit addresses */
 129static unsigned int pcpu_low_unit_cpu __ro_after_init;
 130static unsigned int pcpu_high_unit_cpu __ro_after_init;
 131
 132/* the address of the first chunk which starts with the kernel static area */
 133void *pcpu_base_addr __ro_after_init;
 134EXPORT_SYMBOL_GPL(pcpu_base_addr);
 135
 136static const int *pcpu_unit_map __ro_after_init;		/* cpu -> unit */
 137const unsigned long *pcpu_unit_offsets __ro_after_init;	/* cpu -> unit offset */
 138
 139/* group information, used for vm allocation */
 140static int pcpu_nr_groups __ro_after_init;
 141static const unsigned long *pcpu_group_offsets __ro_after_init;
 142static const size_t *pcpu_group_sizes __ro_after_init;
 143
 144/*
 145 * The first chunk which always exists.  Note that unlike other
 146 * chunks, this one can be allocated and mapped in several different
 147 * ways and thus often doesn't live in the vmalloc area.
 148 */
 149struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
 150
 151/*
 152 * Optional reserved chunk.  This chunk reserves part of the first
 153 * chunk and serves it for reserved allocations.  When the reserved
 154 * region doesn't exist, the following variable is NULL.
 155 */
 156struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
 157
 158DEFINE_SPINLOCK(pcpu_lock);	/* all internal data structures */
 159static DEFINE_MUTEX(pcpu_alloc_mutex);	/* chunk create/destroy, [de]pop, map ext */
 160
 161struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
 162
 163/* chunks which need their map areas extended, protected by pcpu_lock */
 164static LIST_HEAD(pcpu_map_extend_chunks);
 165
 166/*
 167 * The number of empty populated pages, protected by pcpu_lock.  The
 168 * reserved chunk doesn't contribute to the count.
 169 */
 170int pcpu_nr_empty_pop_pages;
 171
 172/*
 
 
 
 
 
 
 
 
 173 * Balance work is used to populate or destroy chunks asynchronously.  We
 174 * try to keep the number of populated free pages between
 175 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
 176 * empty chunk.
 177 */
 178static void pcpu_balance_workfn(struct work_struct *work);
 179static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
 180static bool pcpu_async_enabled __read_mostly;
 181static bool pcpu_atomic_alloc_failed;
 182
 183static void pcpu_schedule_balance_work(void)
 184{
 185	if (pcpu_async_enabled)
 186		schedule_work(&pcpu_balance_work);
 187}
 188
 189/**
 190 * pcpu_addr_in_chunk - check if the address is served from this chunk
 191 * @chunk: chunk of interest
 192 * @addr: percpu address
 193 *
 194 * RETURNS:
 195 * True if the address is served from this chunk.
 196 */
 197static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
 198{
 199	void *start_addr, *end_addr;
 200
 201	if (!chunk)
 202		return false;
 203
 204	start_addr = chunk->base_addr + chunk->start_offset;
 205	end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
 206		   chunk->end_offset;
 207
 208	return addr >= start_addr && addr < end_addr;
 209}
 210
 211static int __pcpu_size_to_slot(int size)
 212{
 213	int highbit = fls(size);	/* size is in bytes */
 214	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
 215}
 216
 217static int pcpu_size_to_slot(int size)
 218{
 219	if (size == pcpu_unit_size)
 220		return pcpu_nr_slots - 1;
 221	return __pcpu_size_to_slot(size);
 222}
 223
 224static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
 225{
 226	if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0)
 
 
 
 227		return 0;
 228
 229	return pcpu_size_to_slot(chunk->free_bytes);
 230}
 231
 232/* set the pointer to a chunk in a page struct */
 233static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 234{
 235	page->index = (unsigned long)pcpu;
 236}
 237
 238/* obtain pointer to a chunk from a page struct */
 239static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
 240{
 241	return (struct pcpu_chunk *)page->index;
 242}
 243
 244static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
 245{
 246	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
 247}
 248
 249static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
 250{
 251	return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
 252}
 253
 254static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 255				     unsigned int cpu, int page_idx)
 256{
 257	return (unsigned long)chunk->base_addr +
 258	       pcpu_unit_page_offset(cpu, page_idx);
 259}
 260
 261static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
 262{
 263	*rs = find_next_zero_bit(bitmap, end, *rs);
 264	*re = find_next_bit(bitmap, end, *rs + 1);
 265}
 266
 267static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
 268{
 269	*rs = find_next_bit(bitmap, end, *rs);
 270	*re = find_next_zero_bit(bitmap, end, *rs + 1);
 271}
 272
 273/*
 274 * Bitmap region iterators.  Iterates over the bitmap between
 275 * [@start, @end) in @chunk.  @rs and @re should be integer variables
 276 * and will be set to start and end index of the current free region.
 277 */
 278#define pcpu_for_each_unpop_region(bitmap, rs, re, start, end)		     \
 279	for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
 280	     (rs) < (re);						     \
 281	     (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
 282
 283#define pcpu_for_each_pop_region(bitmap, rs, re, start, end)		     \
 284	for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end));   \
 285	     (rs) < (re);						     \
 286	     (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
 287
 288/*
 289 * The following are helper functions to help access bitmaps and convert
 290 * between bitmap offsets to address offsets.
 291 */
 292static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
 293{
 294	return chunk->alloc_map +
 295	       (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
 296}
 297
 298static unsigned long pcpu_off_to_block_index(int off)
 299{
 300	return off / PCPU_BITMAP_BLOCK_BITS;
 301}
 302
 303static unsigned long pcpu_off_to_block_off(int off)
 304{
 305	return off & (PCPU_BITMAP_BLOCK_BITS - 1);
 306}
 307
 308static unsigned long pcpu_block_off_to_off(int index, int off)
 309{
 310	return index * PCPU_BITMAP_BLOCK_BITS + off;
 311}
 312
 313/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 314 * pcpu_next_md_free_region - finds the next hint free area
 315 * @chunk: chunk of interest
 316 * @bit_off: chunk offset
 317 * @bits: size of free area
 318 *
 319 * Helper function for pcpu_for_each_md_free_region.  It checks
 320 * block->contig_hint and performs aggregation across blocks to find the
 321 * next hint.  It modifies bit_off and bits in-place to be consumed in the
 322 * loop.
 323 */
 324static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
 325				     int *bits)
 326{
 327	int i = pcpu_off_to_block_index(*bit_off);
 328	int block_off = pcpu_off_to_block_off(*bit_off);
 329	struct pcpu_block_md *block;
 330
 331	*bits = 0;
 332	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 333	     block++, i++) {
 334		/* handles contig area across blocks */
 335		if (*bits) {
 336			*bits += block->left_free;
 337			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 338				continue;
 339			return;
 340		}
 341
 342		/*
 343		 * This checks three things.  First is there a contig_hint to
 344		 * check.  Second, have we checked this hint before by
 345		 * comparing the block_off.  Third, is this the same as the
 346		 * right contig hint.  In the last case, it spills over into
 347		 * the next block and should be handled by the contig area
 348		 * across blocks code.
 349		 */
 350		*bits = block->contig_hint;
 351		if (*bits && block->contig_hint_start >= block_off &&
 352		    *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
 353			*bit_off = pcpu_block_off_to_off(i,
 354					block->contig_hint_start);
 355			return;
 356		}
 357		/* reset to satisfy the second predicate above */
 358		block_off = 0;
 359
 360		*bits = block->right_free;
 361		*bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
 362	}
 363}
 364
 365/**
 366 * pcpu_next_fit_region - finds fit areas for a given allocation request
 367 * @chunk: chunk of interest
 368 * @alloc_bits: size of allocation
 369 * @align: alignment of area (max PAGE_SIZE)
 370 * @bit_off: chunk offset
 371 * @bits: size of free area
 372 *
 373 * Finds the next free region that is viable for use with a given size and
 374 * alignment.  This only returns if there is a valid area to be used for this
 375 * allocation.  block->first_free is returned if the allocation request fits
 376 * within the block to see if the request can be fulfilled prior to the contig
 377 * hint.
 378 */
 379static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
 380				 int align, int *bit_off, int *bits)
 381{
 382	int i = pcpu_off_to_block_index(*bit_off);
 383	int block_off = pcpu_off_to_block_off(*bit_off);
 384	struct pcpu_block_md *block;
 385
 386	*bits = 0;
 387	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 388	     block++, i++) {
 389		/* handles contig area across blocks */
 390		if (*bits) {
 391			*bits += block->left_free;
 392			if (*bits >= alloc_bits)
 393				return;
 394			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 395				continue;
 396		}
 397
 398		/* check block->contig_hint */
 399		*bits = ALIGN(block->contig_hint_start, align) -
 400			block->contig_hint_start;
 401		/*
 402		 * This uses the block offset to determine if this has been
 403		 * checked in the prior iteration.
 404		 */
 405		if (block->contig_hint &&
 406		    block->contig_hint_start >= block_off &&
 407		    block->contig_hint >= *bits + alloc_bits) {
 
 
 408			*bits += alloc_bits + block->contig_hint_start -
 409				 block->first_free;
 410			*bit_off = pcpu_block_off_to_off(i, block->first_free);
 411			return;
 412		}
 413		/* reset to satisfy the second predicate above */
 414		block_off = 0;
 415
 416		*bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
 417				 align);
 418		*bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
 419		*bit_off = pcpu_block_off_to_off(i, *bit_off);
 420		if (*bits >= alloc_bits)
 421			return;
 422	}
 423
 424	/* no valid offsets were found - fail condition */
 425	*bit_off = pcpu_chunk_map_bits(chunk);
 426}
 427
 428/*
 429 * Metadata free area iterators.  These perform aggregation of free areas
 430 * based on the metadata blocks and return the offset @bit_off and size in
 431 * bits of the free area @bits.  pcpu_for_each_fit_region only returns when
 432 * a fit is found for the allocation request.
 433 */
 434#define pcpu_for_each_md_free_region(chunk, bit_off, bits)		\
 435	for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits));	\
 436	     (bit_off) < pcpu_chunk_map_bits((chunk));			\
 437	     (bit_off) += (bits) + 1,					\
 438	     pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
 439
 440#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits)     \
 441	for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 442				  &(bits));				      \
 443	     (bit_off) < pcpu_chunk_map_bits((chunk));			      \
 444	     (bit_off) += (bits),					      \
 445	     pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 446				  &(bits)))
 447
 448/**
 449 * pcpu_mem_zalloc - allocate memory
 450 * @size: bytes to allocate
 451 * @gfp: allocation flags
 452 *
 453 * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
 454 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
 455 * This is to facilitate passing through whitelisted flags.  The
 456 * returned memory is always zeroed.
 457 *
 458 * RETURNS:
 459 * Pointer to the allocated area on success, NULL on failure.
 460 */
 461static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
 462{
 463	if (WARN_ON_ONCE(!slab_is_available()))
 464		return NULL;
 465
 466	if (size <= PAGE_SIZE)
 467		return kzalloc(size, gfp);
 468	else
 469		return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL);
 470}
 471
 472/**
 473 * pcpu_mem_free - free memory
 474 * @ptr: memory to free
 475 *
 476 * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
 477 */
 478static void pcpu_mem_free(void *ptr)
 479{
 480	kvfree(ptr);
 481}
 482
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 483/**
 484 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
 485 * @chunk: chunk of interest
 486 * @oslot: the previous slot it was on
 487 *
 488 * This function is called after an allocation or free changed @chunk.
 489 * New slot according to the changed state is determined and @chunk is
 490 * moved to the slot.  Note that the reserved chunk is never put on
 491 * chunk slots.
 492 *
 493 * CONTEXT:
 494 * pcpu_lock.
 495 */
 496static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 497{
 498	int nslot = pcpu_chunk_slot(chunk);
 499
 500	if (chunk != pcpu_reserved_chunk && oslot != nslot) {
 501		if (oslot < nslot)
 502			list_move(&chunk->list, &pcpu_slot[nslot]);
 503		else
 504			list_move_tail(&chunk->list, &pcpu_slot[nslot]);
 505	}
 506}
 507
 508/**
 509 * pcpu_cnt_pop_pages- counts populated backing pages in range
 510 * @chunk: chunk of interest
 511 * @bit_off: start offset
 512 * @bits: size of area to check
 513 *
 514 * Calculates the number of populated pages in the region
 515 * [page_start, page_end).  This keeps track of how many empty populated
 516 * pages are available and decide if async work should be scheduled.
 517 *
 518 * RETURNS:
 519 * The nr of populated pages.
 520 */
 521static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off,
 522				     int bits)
 523{
 524	int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE);
 525	int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
 526
 527	if (page_start >= page_end)
 528		return 0;
 
 
 
 
 529
 530	/*
 531	 * bitmap_weight counts the number of bits set in a bitmap up to
 532	 * the specified number of bits.  This is counting the populated
 533	 * pages up to page_end and then subtracting the populated pages
 534	 * up to page_start to count the populated pages in
 535	 * [page_start, page_end).
 536	 */
 537	return bitmap_weight(chunk->populated, page_end) -
 538	       bitmap_weight(chunk->populated, page_start);
 539}
 540
 541/**
 542 * pcpu_chunk_update - updates the chunk metadata given a free area
 543 * @chunk: chunk of interest
 544 * @bit_off: chunk offset
 545 * @bits: size of free area
 546 *
 547 * This updates the chunk's contig hint and starting offset given a free area.
 548 * Choose the best starting offset if the contig hint is equal.
 
 549 */
 550static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits)
 551{
 552	if (bits > chunk->contig_bits) {
 553		chunk->contig_bits_start = bit_off;
 554		chunk->contig_bits = bits;
 555	} else if (bits == chunk->contig_bits && chunk->contig_bits_start &&
 556		   (!bit_off ||
 557		    __ffs(bit_off) > __ffs(chunk->contig_bits_start))) {
 558		/* use the start with the best alignment */
 559		chunk->contig_bits_start = bit_off;
 560	}
 561}
 562
 563/**
 564 * pcpu_chunk_refresh_hint - updates metadata about a chunk
 565 * @chunk: chunk of interest
 566 *
 567 * Iterates over the metadata blocks to find the largest contig area.
 568 * It also counts the populated pages and uses the delta to update the
 569 * global count.
 570 *
 571 * Updates:
 572 *      chunk->contig_bits
 573 *      chunk->contig_bits_start
 574 *      nr_empty_pop_pages (chunk and global)
 575 */
 576static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk)
 577{
 578	int bit_off, bits, nr_empty_pop_pages;
 579
 580	/* clear metadata */
 581	chunk->contig_bits = 0;
 582
 583	bit_off = chunk->first_bit;
 584	bits = nr_empty_pop_pages = 0;
 585	pcpu_for_each_md_free_region(chunk, bit_off, bits) {
 586		pcpu_chunk_update(chunk, bit_off, bits);
 587
 588		nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits);
 589	}
 590
 591	/*
 592	 * Keep track of nr_empty_pop_pages.
 593	 *
 594	 * The chunk maintains the previous number of free pages it held,
 595	 * so the delta is used to update the global counter.  The reserved
 596	 * chunk is not part of the free page count as they are populated
 597	 * at init and are special to serving reserved allocations.
 598	 */
 599	if (chunk != pcpu_reserved_chunk)
 600		pcpu_nr_empty_pop_pages +=
 601			(nr_empty_pop_pages - chunk->nr_empty_pop_pages);
 602
 603	chunk->nr_empty_pop_pages = nr_empty_pop_pages;
 604}
 605
 606/**
 607 * pcpu_block_update - updates a block given a free area
 608 * @block: block of interest
 609 * @start: start offset in block
 610 * @end: end offset in block
 611 *
 612 * Updates a block given a known free area.  The region [start, end) is
 613 * expected to be the entirety of the free area within a block.  Chooses
 614 * the best starting offset if the contig hints are equal.
 615 */
 616static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
 617{
 618	int contig = end - start;
 619
 620	block->first_free = min(block->first_free, start);
 621	if (start == 0)
 622		block->left_free = contig;
 623
 624	if (end == PCPU_BITMAP_BLOCK_BITS)
 625		block->right_free = contig;
 626
 627	if (contig > block->contig_hint) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 628		block->contig_hint_start = start;
 629		block->contig_hint = contig;
 630	} else if (block->contig_hint_start && contig == block->contig_hint &&
 631		   (!start || __ffs(start) > __ffs(block->contig_hint_start))) {
 632		/* use the start with the best alignment */
 633		block->contig_hint_start = start;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 634	}
 635}
 636
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 637/**
 638 * pcpu_block_refresh_hint
 639 * @chunk: chunk of interest
 640 * @index: index of the metadata block
 641 *
 642 * Scans over the block beginning at first_free and updates the block
 643 * metadata accordingly.
 644 */
 645static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
 646{
 647	struct pcpu_block_md *block = chunk->md_blocks + index;
 648	unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
 649	int rs, re;	/* region start, region end */
 
 
 
 
 
 
 
 
 
 
 
 650
 651	/* clear hints */
 652	block->contig_hint = 0;
 653	block->left_free = block->right_free = 0;
 654
 655	/* iterate over free areas and update the contig hints */
 656	pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free,
 657				   PCPU_BITMAP_BLOCK_BITS) {
 658		pcpu_block_update(block, rs, re);
 659	}
 660}
 661
 662/**
 663 * pcpu_block_update_hint_alloc - update hint on allocation path
 664 * @chunk: chunk of interest
 665 * @bit_off: chunk offset
 666 * @bits: size of request
 667 *
 668 * Updates metadata for the allocation path.  The metadata only has to be
 669 * refreshed by a full scan iff the chunk's contig hint is broken.  Block level
 670 * scans are required if the block's contig hint is broken.
 671 */
 672static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
 673					 int bits)
 674{
 
 
 675	struct pcpu_block_md *s_block, *e_block, *block;
 676	int s_index, e_index;	/* block indexes of the freed allocation */
 677	int s_off, e_off;	/* block offsets of the freed allocation */
 678
 679	/*
 680	 * Calculate per block offsets.
 681	 * The calculation uses an inclusive range, but the resulting offsets
 682	 * are [start, end).  e_index always points to the last block in the
 683	 * range.
 684	 */
 685	s_index = pcpu_off_to_block_index(bit_off);
 686	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 687	s_off = pcpu_off_to_block_off(bit_off);
 688	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 689
 690	s_block = chunk->md_blocks + s_index;
 691	e_block = chunk->md_blocks + e_index;
 692
 693	/*
 694	 * Update s_block.
 
 
 
 
 
 695	 * block->first_free must be updated if the allocation takes its place.
 696	 * If the allocation breaks the contig_hint, a scan is required to
 697	 * restore this hint.
 698	 */
 699	if (s_off == s_block->first_free)
 700		s_block->first_free = find_next_zero_bit(
 701					pcpu_index_alloc_map(chunk, s_index),
 702					PCPU_BITMAP_BLOCK_BITS,
 703					s_off + bits);
 704
 705	if (s_off >= s_block->contig_hint_start &&
 706	    s_off < s_block->contig_hint_start + s_block->contig_hint) {
 
 
 
 
 
 
 
 
 
 707		/* block contig hint is broken - scan to fix it */
 
 
 708		pcpu_block_refresh_hint(chunk, s_index);
 709	} else {
 710		/* update left and right contig manually */
 711		s_block->left_free = min(s_block->left_free, s_off);
 712		if (s_index == e_index)
 713			s_block->right_free = min_t(int, s_block->right_free,
 714					PCPU_BITMAP_BLOCK_BITS - e_off);
 715		else
 716			s_block->right_free = 0;
 717	}
 718
 719	/*
 720	 * Update e_block.
 721	 */
 722	if (s_index != e_index) {
 
 
 
 723		/*
 724		 * When the allocation is across blocks, the end is along
 725		 * the left part of the e_block.
 726		 */
 727		e_block->first_free = find_next_zero_bit(
 728				pcpu_index_alloc_map(chunk, e_index),
 729				PCPU_BITMAP_BLOCK_BITS, e_off);
 730
 731		if (e_off == PCPU_BITMAP_BLOCK_BITS) {
 732			/* reset the block */
 733			e_block++;
 734		} else {
 
 
 
 
 735			if (e_off > e_block->contig_hint_start) {
 736				/* contig hint is broken - scan to fix it */
 737				pcpu_block_refresh_hint(chunk, e_index);
 738			} else {
 739				e_block->left_free = 0;
 740				e_block->right_free =
 741					min_t(int, e_block->right_free,
 742					      PCPU_BITMAP_BLOCK_BITS - e_off);
 743			}
 744		}
 745
 746		/* update in-between md_blocks */
 
 747		for (block = s_block + 1; block < e_block; block++) {
 
 748			block->contig_hint = 0;
 749			block->left_free = 0;
 750			block->right_free = 0;
 751		}
 752	}
 753
 754	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 755	 * The only time a full chunk scan is required is if the chunk
 756	 * contig hint is broken.  Otherwise, it means a smaller space
 757	 * was used and therefore the chunk contig hint is still correct.
 758	 */
 759	if (bit_off >= chunk->contig_bits_start  &&
 760	    bit_off < chunk->contig_bits_start + chunk->contig_bits)
 761		pcpu_chunk_refresh_hint(chunk);
 
 
 
 762}
 763
 764/**
 765 * pcpu_block_update_hint_free - updates the block hints on the free path
 766 * @chunk: chunk of interest
 767 * @bit_off: chunk offset
 768 * @bits: size of request
 769 *
 770 * Updates metadata for the allocation path.  This avoids a blind block
 771 * refresh by making use of the block contig hints.  If this fails, it scans
 772 * forward and backward to determine the extent of the free area.  This is
 773 * capped at the boundary of blocks.
 774 *
 775 * A chunk update is triggered if a page becomes free, a block becomes free,
 776 * or the free spans across blocks.  This tradeoff is to minimize iterating
 777 * over the block metadata to update chunk->contig_bits.  chunk->contig_bits
 778 * may be off by up to a page, but it will never be more than the available
 779 * space.  If the contig hint is contained in one block, it will be accurate.
 
 780 */
 781static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
 782					int bits)
 783{
 
 784	struct pcpu_block_md *s_block, *e_block, *block;
 785	int s_index, e_index;	/* block indexes of the freed allocation */
 786	int s_off, e_off;	/* block offsets of the freed allocation */
 787	int start, end;		/* start and end of the whole free area */
 788
 789	/*
 790	 * Calculate per block offsets.
 791	 * The calculation uses an inclusive range, but the resulting offsets
 792	 * are [start, end).  e_index always points to the last block in the
 793	 * range.
 794	 */
 795	s_index = pcpu_off_to_block_index(bit_off);
 796	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 797	s_off = pcpu_off_to_block_off(bit_off);
 798	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 799
 800	s_block = chunk->md_blocks + s_index;
 801	e_block = chunk->md_blocks + e_index;
 802
 803	/*
 804	 * Check if the freed area aligns with the block->contig_hint.
 805	 * If it does, then the scan to find the beginning/end of the
 806	 * larger free area can be avoided.
 807	 *
 808	 * start and end refer to beginning and end of the free area
 809	 * within each their respective blocks.  This is not necessarily
 810	 * the entire free area as it may span blocks past the beginning
 811	 * or end of the block.
 812	 */
 813	start = s_off;
 814	if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
 815		start = s_block->contig_hint_start;
 816	} else {
 817		/*
 818		 * Scan backwards to find the extent of the free area.
 819		 * find_last_bit returns the starting bit, so if the start bit
 820		 * is returned, that means there was no last bit and the
 821		 * remainder of the chunk is free.
 822		 */
 823		int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
 824					  start);
 825		start = (start == l_bit) ? 0 : l_bit + 1;
 826	}
 827
 828	end = e_off;
 829	if (e_off == e_block->contig_hint_start)
 830		end = e_block->contig_hint_start + e_block->contig_hint;
 831	else
 832		end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
 833				    PCPU_BITMAP_BLOCK_BITS, end);
 834
 835	/* update s_block */
 836	e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
 
 
 837	pcpu_block_update(s_block, start, e_off);
 838
 839	/* freeing in the same block */
 840	if (s_index != e_index) {
 841		/* update e_block */
 
 
 842		pcpu_block_update(e_block, 0, end);
 843
 844		/* reset md_blocks in the middle */
 
 845		for (block = s_block + 1; block < e_block; block++) {
 846			block->first_free = 0;
 
 847			block->contig_hint_start = 0;
 848			block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
 849			block->left_free = PCPU_BITMAP_BLOCK_BITS;
 850			block->right_free = PCPU_BITMAP_BLOCK_BITS;
 851		}
 852	}
 853
 
 
 
 854	/*
 855	 * Refresh chunk metadata when the free makes a page free, a block
 856	 * free, or spans across blocks.  The contig hint may be off by up to
 857	 * a page, but if the hint is contained in a block, it will be accurate
 858	 * with the else condition below.
 859	 */
 860	if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) >
 861	     ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) ||
 862	    s_index != e_index)
 863		pcpu_chunk_refresh_hint(chunk);
 864	else
 865		pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start),
 866				  s_block->contig_hint);
 
 867}
 868
 869/**
 870 * pcpu_is_populated - determines if the region is populated
 871 * @chunk: chunk of interest
 872 * @bit_off: chunk offset
 873 * @bits: size of area
 874 * @next_off: return value for the next offset to start searching
 875 *
 876 * For atomic allocations, check if the backing pages are populated.
 877 *
 878 * RETURNS:
 879 * Bool if the backing pages are populated.
 880 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
 881 */
 882static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
 883			      int *next_off)
 884{
 885	int page_start, page_end, rs, re;
 886
 887	page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
 888	page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
 889
 890	rs = page_start;
 891	pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
 892	if (rs >= page_end)
 893		return true;
 894
 895	*next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
 
 
 896	return false;
 897}
 898
 899/**
 900 * pcpu_find_block_fit - finds the block index to start searching
 901 * @chunk: chunk of interest
 902 * @alloc_bits: size of request in allocation units
 903 * @align: alignment of area (max PAGE_SIZE bytes)
 904 * @pop_only: use populated regions only
 905 *
 906 * Given a chunk and an allocation spec, find the offset to begin searching
 907 * for a free region.  This iterates over the bitmap metadata blocks to
 908 * find an offset that will be guaranteed to fit the requirements.  It is
 909 * not quite first fit as if the allocation does not fit in the contig hint
 910 * of a block or chunk, it is skipped.  This errs on the side of caution
 911 * to prevent excess iteration.  Poor alignment can cause the allocator to
 912 * skip over blocks and chunks that have valid free areas.
 913 *
 914 * RETURNS:
 915 * The offset in the bitmap to begin searching.
 916 * -1 if no offset is found.
 917 */
 918static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
 919			       size_t align, bool pop_only)
 920{
 
 921	int bit_off, bits, next_off;
 922
 923	/*
 924	 * Check to see if the allocation can fit in the chunk's contig hint.
 925	 * This is an optimization to prevent scanning by assuming if it
 926	 * cannot fit in the global hint, there is memory pressure and creating
 927	 * a new chunk would happen soon.
 928	 */
 929	bit_off = ALIGN(chunk->contig_bits_start, align) -
 930		  chunk->contig_bits_start;
 931	if (bit_off + alloc_bits > chunk->contig_bits)
 932		return -1;
 933
 934	bit_off = chunk->first_bit;
 935	bits = 0;
 936	pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
 937		if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
 938						   &next_off))
 939			break;
 940
 941		bit_off = next_off;
 942		bits = 0;
 943	}
 944
 945	if (bit_off == pcpu_chunk_map_bits(chunk))
 946		return -1;
 947
 948	return bit_off;
 949}
 950
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 951/**
 952 * pcpu_alloc_area - allocates an area from a pcpu_chunk
 953 * @chunk: chunk of interest
 954 * @alloc_bits: size of request in allocation units
 955 * @align: alignment of area (max PAGE_SIZE)
 956 * @start: bit_off to start searching
 957 *
 958 * This function takes in a @start offset to begin searching to fit an
 959 * allocation of @alloc_bits with alignment @align.  It needs to scan
 960 * the allocation map because if it fits within the block's contig hint,
 961 * @start will be block->first_free. This is an attempt to fill the
 962 * allocation prior to breaking the contig hint.  The allocation and
 963 * boundary maps are updated accordingly if it confirms a valid
 964 * free area.
 965 *
 966 * RETURNS:
 967 * Allocated addr offset in @chunk on success.
 968 * -1 if no matching area is found.
 969 */
 970static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
 971			   size_t align, int start)
 972{
 
 973	size_t align_mask = (align) ? (align - 1) : 0;
 
 974	int bit_off, end, oslot;
 975
 976	lockdep_assert_held(&pcpu_lock);
 977
 978	oslot = pcpu_chunk_slot(chunk);
 979
 980	/*
 981	 * Search to find a fit.
 982	 */
 983	end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS;
 984	bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start,
 985					     alloc_bits, align_mask);
 
 986	if (bit_off >= end)
 987		return -1;
 988
 
 
 
 989	/* update alloc map */
 990	bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
 991
 992	/* update boundary map */
 993	set_bit(bit_off, chunk->bound_map);
 994	bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
 995	set_bit(bit_off + alloc_bits, chunk->bound_map);
 996
 997	chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
 998
 999	/* update first free bit */
1000	if (bit_off == chunk->first_bit)
1001		chunk->first_bit = find_next_zero_bit(
1002					chunk->alloc_map,
1003					pcpu_chunk_map_bits(chunk),
1004					bit_off + alloc_bits);
1005
1006	pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1007
1008	pcpu_chunk_relocate(chunk, oslot);
1009
1010	return bit_off * PCPU_MIN_ALLOC_SIZE;
1011}
1012
1013/**
1014 * pcpu_free_area - frees the corresponding offset
1015 * @chunk: chunk of interest
1016 * @off: addr offset into chunk
1017 *
1018 * This function determines the size of an allocation to free using
1019 * the boundary bitmap and clears the allocation map.
 
 
 
1020 */
1021static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
1022{
1023	int bit_off, bits, end, oslot;
 
1024
1025	lockdep_assert_held(&pcpu_lock);
1026	pcpu_stats_area_dealloc(chunk);
1027
1028	oslot = pcpu_chunk_slot(chunk);
1029
1030	bit_off = off / PCPU_MIN_ALLOC_SIZE;
1031
1032	/* find end index */
1033	end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1034			    bit_off + 1);
1035	bits = end - bit_off;
1036	bitmap_clear(chunk->alloc_map, bit_off, bits);
1037
 
 
1038	/* update metadata */
1039	chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
1040
1041	/* update first free bit */
1042	chunk->first_bit = min(chunk->first_bit, bit_off);
1043
1044	pcpu_block_update_hint_free(chunk, bit_off, bits);
1045
1046	pcpu_chunk_relocate(chunk, oslot);
 
 
 
 
 
 
 
 
 
 
 
 
1047}
1048
1049static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1050{
1051	struct pcpu_block_md *md_block;
1052
 
 
 
1053	for (md_block = chunk->md_blocks;
1054	     md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1055	     md_block++) {
1056		md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1057		md_block->left_free = PCPU_BITMAP_BLOCK_BITS;
1058		md_block->right_free = PCPU_BITMAP_BLOCK_BITS;
1059	}
1060}
1061
1062/**
1063 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1064 * @tmp_addr: the start of the region served
1065 * @map_size: size of the region served
1066 *
1067 * This is responsible for creating the chunks that serve the first chunk.  The
1068 * base_addr is page aligned down of @tmp_addr while the region end is page
1069 * aligned up.  Offsets are kept track of to determine the region served. All
1070 * this is done to appease the bitmap allocator in avoiding partial blocks.
1071 *
1072 * RETURNS:
1073 * Chunk serving the region at @tmp_addr of @map_size.
1074 */
1075static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1076							 int map_size)
1077{
1078	struct pcpu_chunk *chunk;
1079	unsigned long aligned_addr, lcm_align;
1080	int start_offset, offset_bits, region_size, region_bits;
 
1081
1082	/* region calculations */
1083	aligned_addr = tmp_addr & PAGE_MASK;
1084
1085	start_offset = tmp_addr - aligned_addr;
1086
1087	/*
1088	 * Align the end of the region with the LCM of PAGE_SIZE and
1089	 * PCPU_BITMAP_BLOCK_SIZE.  One of these constants is a multiple of
1090	 * the other.
1091	 */
1092	lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1093	region_size = ALIGN(start_offset + map_size, lcm_align);
1094
1095	/* allocate chunk */
1096	chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) +
1097				    BITS_TO_LONGS(region_size >> PAGE_SHIFT),
1098				    0);
 
 
 
1099
1100	INIT_LIST_HEAD(&chunk->list);
1101
1102	chunk->base_addr = (void *)aligned_addr;
1103	chunk->start_offset = start_offset;
1104	chunk->end_offset = region_size - chunk->start_offset - map_size;
1105
1106	chunk->nr_pages = region_size >> PAGE_SHIFT;
1107	region_bits = pcpu_chunk_map_bits(chunk);
1108
1109	chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) *
1110					       sizeof(chunk->alloc_map[0]), 0);
1111	chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) *
1112					       sizeof(chunk->bound_map[0]), 0);
1113	chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) *
1114					       sizeof(chunk->md_blocks[0]), 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1115	pcpu_init_md_blocks(chunk);
1116
1117	/* manage populated page bitmap */
1118	chunk->immutable = true;
1119	bitmap_fill(chunk->populated, chunk->nr_pages);
1120	chunk->nr_populated = chunk->nr_pages;
1121	chunk->nr_empty_pop_pages =
1122		pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE,
1123				   map_size / PCPU_MIN_ALLOC_SIZE);
1124
1125	chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE;
1126	chunk->free_bytes = map_size;
1127
1128	if (chunk->start_offset) {
1129		/* hide the beginning of the bitmap */
1130		offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1131		bitmap_set(chunk->alloc_map, 0, offset_bits);
1132		set_bit(0, chunk->bound_map);
1133		set_bit(offset_bits, chunk->bound_map);
1134
1135		chunk->first_bit = offset_bits;
1136
1137		pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1138	}
1139
1140	if (chunk->end_offset) {
1141		/* hide the end of the bitmap */
1142		offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1143		bitmap_set(chunk->alloc_map,
1144			   pcpu_chunk_map_bits(chunk) - offset_bits,
1145			   offset_bits);
1146		set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1147			chunk->bound_map);
1148		set_bit(region_bits, chunk->bound_map);
1149
1150		pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1151					     - offset_bits, offset_bits);
1152	}
1153
1154	return chunk;
1155}
1156
1157static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1158{
1159	struct pcpu_chunk *chunk;
1160	int region_bits;
1161
1162	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1163	if (!chunk)
1164		return NULL;
1165
1166	INIT_LIST_HEAD(&chunk->list);
1167	chunk->nr_pages = pcpu_unit_pages;
1168	region_bits = pcpu_chunk_map_bits(chunk);
1169
1170	chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1171					   sizeof(chunk->alloc_map[0]), gfp);
1172	if (!chunk->alloc_map)
1173		goto alloc_map_fail;
1174
1175	chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1176					   sizeof(chunk->bound_map[0]), gfp);
1177	if (!chunk->bound_map)
1178		goto bound_map_fail;
1179
1180	chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1181					   sizeof(chunk->md_blocks[0]), gfp);
1182	if (!chunk->md_blocks)
1183		goto md_blocks_fail;
1184
 
 
 
 
 
 
 
 
 
 
1185	pcpu_init_md_blocks(chunk);
1186
1187	/* init metadata */
1188	chunk->contig_bits = region_bits;
1189	chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1190
1191	return chunk;
1192
 
 
 
 
1193md_blocks_fail:
1194	pcpu_mem_free(chunk->bound_map);
1195bound_map_fail:
1196	pcpu_mem_free(chunk->alloc_map);
1197alloc_map_fail:
1198	pcpu_mem_free(chunk);
1199
1200	return NULL;
1201}
1202
1203static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1204{
1205	if (!chunk)
1206		return;
 
 
 
 
1207	pcpu_mem_free(chunk->bound_map);
1208	pcpu_mem_free(chunk->alloc_map);
1209	pcpu_mem_free(chunk);
1210}
1211
1212/**
1213 * pcpu_chunk_populated - post-population bookkeeping
1214 * @chunk: pcpu_chunk which got populated
1215 * @page_start: the start page
1216 * @page_end: the end page
1217 * @for_alloc: if this is to populate for allocation
1218 *
1219 * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
1220 * the bookkeeping information accordingly.  Must be called after each
1221 * successful population.
1222 *
1223 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1224 * is to serve an allocation in that area.
1225 */
1226static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1227				 int page_end, bool for_alloc)
1228{
1229	int nr = page_end - page_start;
1230
1231	lockdep_assert_held(&pcpu_lock);
1232
1233	bitmap_set(chunk->populated, page_start, nr);
1234	chunk->nr_populated += nr;
 
1235
1236	if (!for_alloc) {
1237		chunk->nr_empty_pop_pages += nr;
1238		pcpu_nr_empty_pop_pages += nr;
1239	}
1240}
1241
1242/**
1243 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1244 * @chunk: pcpu_chunk which got depopulated
1245 * @page_start: the start page
1246 * @page_end: the end page
1247 *
1248 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1249 * Update the bookkeeping information accordingly.  Must be called after
1250 * each successful depopulation.
1251 */
1252static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1253				   int page_start, int page_end)
1254{
1255	int nr = page_end - page_start;
1256
1257	lockdep_assert_held(&pcpu_lock);
1258
1259	bitmap_clear(chunk->populated, page_start, nr);
1260	chunk->nr_populated -= nr;
1261	chunk->nr_empty_pop_pages -= nr;
1262	pcpu_nr_empty_pop_pages -= nr;
 
1263}
1264
1265/*
1266 * Chunk management implementation.
1267 *
1268 * To allow different implementations, chunk alloc/free and
1269 * [de]population are implemented in a separate file which is pulled
1270 * into this file and compiled together.  The following functions
1271 * should be implemented.
1272 *
1273 * pcpu_populate_chunk		- populate the specified range of a chunk
1274 * pcpu_depopulate_chunk	- depopulate the specified range of a chunk
 
1275 * pcpu_create_chunk		- create a new chunk
1276 * pcpu_destroy_chunk		- destroy a chunk, always preceded by full depop
1277 * pcpu_addr_to_page		- translate address to physical address
1278 * pcpu_verify_alloc_info	- check alloc_info is acceptable during init
1279 */
1280static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1281			       int page_start, int page_end, gfp_t gfp);
1282static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1283				  int page_start, int page_end);
 
 
1284static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1285static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1286static struct page *pcpu_addr_to_page(void *addr);
1287static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1288
1289#ifdef CONFIG_NEED_PER_CPU_KM
1290#include "percpu-km.c"
1291#else
1292#include "percpu-vm.c"
1293#endif
1294
1295/**
1296 * pcpu_chunk_addr_search - determine chunk containing specified address
1297 * @addr: address for which the chunk needs to be determined.
1298 *
1299 * This is an internal function that handles all but static allocations.
1300 * Static percpu address values should never be passed into the allocator.
1301 *
1302 * RETURNS:
1303 * The address of the found chunk.
1304 */
1305static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1306{
1307	/* is it in the dynamic region (first chunk)? */
1308	if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1309		return pcpu_first_chunk;
1310
1311	/* is it in the reserved region? */
1312	if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1313		return pcpu_reserved_chunk;
1314
1315	/*
1316	 * The address is relative to unit0 which might be unused and
1317	 * thus unmapped.  Offset the address to the unit space of the
1318	 * current processor before looking it up in the vmalloc
1319	 * space.  Note that any possible cpu id can be used here, so
1320	 * there's no need to worry about preemption or cpu hotplug.
1321	 */
1322	addr += pcpu_unit_offsets[raw_smp_processor_id()];
1323	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1324}
1325
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1326/**
1327 * pcpu_alloc - the percpu allocator
1328 * @size: size of area to allocate in bytes
1329 * @align: alignment of area (max PAGE_SIZE)
1330 * @reserved: allocate from the reserved chunk if available
1331 * @gfp: allocation flags
1332 *
1333 * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
1334 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1335 * then no warning will be triggered on invalid or failed allocation
1336 * requests.
1337 *
1338 * RETURNS:
1339 * Percpu pointer to the allocated area on success, NULL on failure.
1340 */
1341static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1342				 gfp_t gfp)
1343{
1344	/* whitelisted flags that can be passed to the backing allocators */
1345	gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1346	bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1347	bool do_warn = !(gfp & __GFP_NOWARN);
1348	static int warn_limit = 10;
1349	struct pcpu_chunk *chunk;
1350	const char *err;
1351	int slot, off, cpu, ret;
1352	unsigned long flags;
1353	void __percpu *ptr;
1354	size_t bits, bit_align;
1355
 
 
 
 
 
 
1356	/*
1357	 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1358	 * therefore alignment must be a minimum of that many bytes.
1359	 * An allocation may have internal fragmentation from rounding up
1360	 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1361	 */
1362	if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1363		align = PCPU_MIN_ALLOC_SIZE;
1364
1365	size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1366	bits = size >> PCPU_MIN_ALLOC_SHIFT;
1367	bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1368
1369	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1370		     !is_power_of_2(align))) {
1371		WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1372		     size, align);
1373		return NULL;
1374	}
1375
 
 
 
1376	if (!is_atomic) {
1377		/*
1378		 * pcpu_balance_workfn() allocates memory under this mutex,
1379		 * and it may wait for memory reclaim. Allow current task
1380		 * to become OOM victim, in case of memory pressure.
1381		 */
1382		if (gfp & __GFP_NOFAIL)
1383			mutex_lock(&pcpu_alloc_mutex);
1384		else if (mutex_lock_killable(&pcpu_alloc_mutex))
 
1385			return NULL;
 
1386	}
1387
1388	spin_lock_irqsave(&pcpu_lock, flags);
1389
1390	/* serve reserved allocations from the reserved chunk if available */
1391	if (reserved && pcpu_reserved_chunk) {
1392		chunk = pcpu_reserved_chunk;
1393
1394		off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1395		if (off < 0) {
1396			err = "alloc from reserved chunk failed";
1397			goto fail_unlock;
1398		}
1399
1400		off = pcpu_alloc_area(chunk, bits, bit_align, off);
1401		if (off >= 0)
1402			goto area_found;
1403
1404		err = "alloc from reserved chunk failed";
1405		goto fail_unlock;
1406	}
1407
1408restart:
1409	/* search through normal chunks */
1410	for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1411		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
 
1412			off = pcpu_find_block_fit(chunk, bits, bit_align,
1413						  is_atomic);
1414			if (off < 0)
 
 
1415				continue;
 
1416
1417			off = pcpu_alloc_area(chunk, bits, bit_align, off);
1418			if (off >= 0)
 
1419				goto area_found;
1420
1421		}
1422	}
1423
1424	spin_unlock_irqrestore(&pcpu_lock, flags);
1425
1426	/*
1427	 * No space left.  Create a new chunk.  We don't want multiple
1428	 * tasks to create chunks simultaneously.  Serialize and create iff
1429	 * there's still no empty chunk after grabbing the mutex.
1430	 */
1431	if (is_atomic) {
1432		err = "atomic alloc failed, no space left";
1433		goto fail;
1434	}
1435
1436	if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
 
1437		chunk = pcpu_create_chunk(pcpu_gfp);
1438		if (!chunk) {
1439			err = "failed to allocate new chunk";
1440			goto fail;
1441		}
1442
1443		spin_lock_irqsave(&pcpu_lock, flags);
1444		pcpu_chunk_relocate(chunk, -1);
1445	} else {
1446		spin_lock_irqsave(&pcpu_lock, flags);
1447	}
1448
1449	goto restart;
1450
1451area_found:
1452	pcpu_stats_area_alloc(chunk, size);
1453	spin_unlock_irqrestore(&pcpu_lock, flags);
1454
1455	/* populate if not all pages are already there */
1456	if (!is_atomic) {
1457		int page_start, page_end, rs, re;
1458
1459		page_start = PFN_DOWN(off);
1460		page_end = PFN_UP(off + size);
1461
1462		pcpu_for_each_unpop_region(chunk->populated, rs, re,
1463					   page_start, page_end) {
1464			WARN_ON(chunk->immutable);
1465
1466			ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1467
1468			spin_lock_irqsave(&pcpu_lock, flags);
1469			if (ret) {
1470				pcpu_free_area(chunk, off);
1471				err = "failed to populate";
1472				goto fail_unlock;
1473			}
1474			pcpu_chunk_populated(chunk, rs, re, true);
1475			spin_unlock_irqrestore(&pcpu_lock, flags);
1476		}
1477
1478		mutex_unlock(&pcpu_alloc_mutex);
1479	}
1480
1481	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1482		pcpu_schedule_balance_work();
1483
1484	/* clear the areas and return address relative to base address */
1485	for_each_possible_cpu(cpu)
1486		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1487
1488	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1489	kmemleak_alloc_percpu(ptr, size, gfp);
1490
1491	trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1492			chunk->base_addr, off, ptr);
 
 
 
1493
1494	return ptr;
1495
1496fail_unlock:
1497	spin_unlock_irqrestore(&pcpu_lock, flags);
1498fail:
1499	trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1500
1501	if (!is_atomic && do_warn && warn_limit) {
1502		pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1503			size, align, is_atomic, err);
1504		dump_stack();
1505		if (!--warn_limit)
1506			pr_info("limit reached, disable warning\n");
1507	}
1508	if (is_atomic) {
1509		/* see the flag handling in pcpu_blance_workfn() */
1510		pcpu_atomic_alloc_failed = true;
1511		pcpu_schedule_balance_work();
1512	} else {
1513		mutex_unlock(&pcpu_alloc_mutex);
1514	}
 
 
 
1515	return NULL;
1516}
1517
1518/**
1519 * __alloc_percpu_gfp - allocate dynamic percpu area
1520 * @size: size of area to allocate in bytes
1521 * @align: alignment of area (max PAGE_SIZE)
1522 * @gfp: allocation flags
1523 *
1524 * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1525 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1526 * be called from any context but is a lot more likely to fail. If @gfp
1527 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1528 * allocation requests.
1529 *
1530 * RETURNS:
1531 * Percpu pointer to the allocated area on success, NULL on failure.
1532 */
1533void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1534{
1535	return pcpu_alloc(size, align, false, gfp);
1536}
1537EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1538
1539/**
1540 * __alloc_percpu - allocate dynamic percpu area
1541 * @size: size of area to allocate in bytes
1542 * @align: alignment of area (max PAGE_SIZE)
1543 *
1544 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1545 */
1546void __percpu *__alloc_percpu(size_t size, size_t align)
1547{
1548	return pcpu_alloc(size, align, false, GFP_KERNEL);
1549}
1550EXPORT_SYMBOL_GPL(__alloc_percpu);
1551
1552/**
1553 * __alloc_reserved_percpu - allocate reserved percpu area
1554 * @size: size of area to allocate in bytes
1555 * @align: alignment of area (max PAGE_SIZE)
1556 *
1557 * Allocate zero-filled percpu area of @size bytes aligned at @align
1558 * from reserved percpu area if arch has set it up; otherwise,
1559 * allocation is served from the same dynamic area.  Might sleep.
1560 * Might trigger writeouts.
1561 *
1562 * CONTEXT:
1563 * Does GFP_KERNEL allocation.
1564 *
1565 * RETURNS:
1566 * Percpu pointer to the allocated area on success, NULL on failure.
1567 */
1568void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1569{
1570	return pcpu_alloc(size, align, true, GFP_KERNEL);
1571}
1572
1573/**
1574 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1575 * @work: unused
1576 *
1577 * Reclaim all fully free chunks except for the first one.  This is also
1578 * responsible for maintaining the pool of empty populated pages.  However,
1579 * it is possible that this is called when physical memory is scarce causing
1580 * OOM killer to be triggered.  We should avoid doing so until an actual
1581 * allocation causes the failure as it is possible that requests can be
1582 * serviced from already backed regions.
1583 */
1584static void pcpu_balance_workfn(struct work_struct *work)
1585{
1586	/* gfp flags passed to underlying allocators */
1587	const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1588	LIST_HEAD(to_free);
1589	struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1590	struct pcpu_chunk *chunk, *next;
1591	int slot, nr_to_pop, ret;
 
1592
1593	/*
1594	 * There's no reason to keep around multiple unused chunks and VM
1595	 * areas can be scarce.  Destroy all free chunks except for one.
1596	 */
1597	mutex_lock(&pcpu_alloc_mutex);
1598	spin_lock_irq(&pcpu_lock);
1599
1600	list_for_each_entry_safe(chunk, next, free_head, list) {
1601		WARN_ON(chunk->immutable);
1602
1603		/* spare the first one */
1604		if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1605			continue;
1606
1607		list_move(&chunk->list, &to_free);
 
1608	}
1609
1610	spin_unlock_irq(&pcpu_lock);
 
1611
 
1612	list_for_each_entry_safe(chunk, next, &to_free, list) {
1613		int rs, re;
1614
1615		pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1616					 chunk->nr_pages) {
1617			pcpu_depopulate_chunk(chunk, rs, re);
1618			spin_lock_irq(&pcpu_lock);
1619			pcpu_chunk_depopulated(chunk, rs, re);
1620			spin_unlock_irq(&pcpu_lock);
1621		}
1622		pcpu_destroy_chunk(chunk);
1623		cond_resched();
1624	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1625
1626	/*
1627	 * Ensure there are certain number of free populated pages for
1628	 * atomic allocs.  Fill up from the most packed so that atomic
1629	 * allocs don't increase fragmentation.  If atomic allocation
1630	 * failed previously, always populate the maximum amount.  This
1631	 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1632	 * failing indefinitely; however, large atomic allocs are not
1633	 * something we support properly and can be highly unreliable and
1634	 * inefficient.
1635	 */
1636retry_pop:
1637	if (pcpu_atomic_alloc_failed) {
1638		nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1639		/* best effort anyway, don't worry about synchronization */
1640		pcpu_atomic_alloc_failed = false;
1641	} else {
1642		nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1643				  pcpu_nr_empty_pop_pages,
1644				  0, PCPU_EMPTY_POP_PAGES_HIGH);
1645	}
1646
1647	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1648		int nr_unpop = 0, rs, re;
1649
1650		if (!nr_to_pop)
1651			break;
1652
1653		spin_lock_irq(&pcpu_lock);
1654		list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1655			nr_unpop = chunk->nr_pages - chunk->nr_populated;
1656			if (nr_unpop)
1657				break;
1658		}
1659		spin_unlock_irq(&pcpu_lock);
1660
1661		if (!nr_unpop)
1662			continue;
1663
1664		/* @chunk can't go away while pcpu_alloc_mutex is held */
1665		pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1666					   chunk->nr_pages) {
1667			int nr = min(re - rs, nr_to_pop);
1668
 
1669			ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
 
 
1670			if (!ret) {
1671				nr_to_pop -= nr;
1672				spin_lock_irq(&pcpu_lock);
1673				pcpu_chunk_populated(chunk, rs, rs + nr, false);
1674				spin_unlock_irq(&pcpu_lock);
1675			} else {
1676				nr_to_pop = 0;
1677			}
1678
1679			if (!nr_to_pop)
1680				break;
1681		}
1682	}
1683
1684	if (nr_to_pop) {
1685		/* ran out of chunks to populate, create a new one and retry */
 
1686		chunk = pcpu_create_chunk(gfp);
 
 
1687		if (chunk) {
1688			spin_lock_irq(&pcpu_lock);
1689			pcpu_chunk_relocate(chunk, -1);
1690			spin_unlock_irq(&pcpu_lock);
1691			goto retry_pop;
1692		}
1693	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1694
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1695	mutex_unlock(&pcpu_alloc_mutex);
1696}
1697
1698/**
1699 * free_percpu - free percpu area
1700 * @ptr: pointer to area to free
1701 *
1702 * Free percpu area @ptr.
1703 *
1704 * CONTEXT:
1705 * Can be called from atomic context.
1706 */
1707void free_percpu(void __percpu *ptr)
1708{
1709	void *addr;
1710	struct pcpu_chunk *chunk;
1711	unsigned long flags;
1712	int off;
 
1713
1714	if (!ptr)
1715		return;
1716
1717	kmemleak_free_percpu(ptr);
1718
1719	addr = __pcpu_ptr_to_addr(ptr);
1720
1721	spin_lock_irqsave(&pcpu_lock, flags);
1722
1723	chunk = pcpu_chunk_addr_search(addr);
1724	off = addr - chunk->base_addr;
1725
1726	pcpu_free_area(chunk, off);
 
 
1727
1728	/* if there are more than one fully free chunks, wake up grim reaper */
1729	if (chunk->free_bytes == pcpu_unit_size) {
 
 
 
 
1730		struct pcpu_chunk *pos;
1731
1732		list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1733			if (pos != chunk) {
1734				pcpu_schedule_balance_work();
1735				break;
1736			}
 
 
 
1737	}
1738
1739	trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1740
1741	spin_unlock_irqrestore(&pcpu_lock, flags);
 
 
 
1742}
1743EXPORT_SYMBOL_GPL(free_percpu);
1744
1745bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1746{
1747#ifdef CONFIG_SMP
1748	const size_t static_size = __per_cpu_end - __per_cpu_start;
1749	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1750	unsigned int cpu;
1751
1752	for_each_possible_cpu(cpu) {
1753		void *start = per_cpu_ptr(base, cpu);
1754		void *va = (void *)addr;
1755
1756		if (va >= start && va < start + static_size) {
1757			if (can_addr) {
1758				*can_addr = (unsigned long) (va - start);
1759				*can_addr += (unsigned long)
1760					per_cpu_ptr(base, get_boot_cpu_id());
1761			}
1762			return true;
1763		}
1764	}
1765#endif
1766	/* on UP, can't distinguish from other static vars, always false */
1767	return false;
1768}
1769
1770/**
1771 * is_kernel_percpu_address - test whether address is from static percpu area
1772 * @addr: address to test
1773 *
1774 * Test whether @addr belongs to in-kernel static percpu area.  Module
1775 * static percpu areas are not considered.  For those, use
1776 * is_module_percpu_address().
1777 *
1778 * RETURNS:
1779 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1780 */
1781bool is_kernel_percpu_address(unsigned long addr)
1782{
1783	return __is_kernel_percpu_address(addr, NULL);
1784}
1785
1786/**
1787 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1788 * @addr: the address to be converted to physical address
1789 *
1790 * Given @addr which is dereferenceable address obtained via one of
1791 * percpu access macros, this function translates it into its physical
1792 * address.  The caller is responsible for ensuring @addr stays valid
1793 * until this function finishes.
1794 *
1795 * percpu allocator has special setup for the first chunk, which currently
1796 * supports either embedding in linear address space or vmalloc mapping,
1797 * and, from the second one, the backing allocator (currently either vm or
1798 * km) provides translation.
1799 *
1800 * The addr can be translated simply without checking if it falls into the
1801 * first chunk. But the current code reflects better how percpu allocator
1802 * actually works, and the verification can discover both bugs in percpu
1803 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1804 * code.
1805 *
1806 * RETURNS:
1807 * The physical address for @addr.
1808 */
1809phys_addr_t per_cpu_ptr_to_phys(void *addr)
1810{
1811	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1812	bool in_first_chunk = false;
1813	unsigned long first_low, first_high;
1814	unsigned int cpu;
1815
1816	/*
1817	 * The following test on unit_low/high isn't strictly
1818	 * necessary but will speed up lookups of addresses which
1819	 * aren't in the first chunk.
1820	 *
1821	 * The address check is against full chunk sizes.  pcpu_base_addr
1822	 * points to the beginning of the first chunk including the
1823	 * static region.  Assumes good intent as the first chunk may
1824	 * not be full (ie. < pcpu_unit_pages in size).
1825	 */
1826	first_low = (unsigned long)pcpu_base_addr +
1827		    pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
1828	first_high = (unsigned long)pcpu_base_addr +
1829		     pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
1830	if ((unsigned long)addr >= first_low &&
1831	    (unsigned long)addr < first_high) {
1832		for_each_possible_cpu(cpu) {
1833			void *start = per_cpu_ptr(base, cpu);
1834
1835			if (addr >= start && addr < start + pcpu_unit_size) {
1836				in_first_chunk = true;
1837				break;
1838			}
1839		}
1840	}
1841
1842	if (in_first_chunk) {
1843		if (!is_vmalloc_addr(addr))
1844			return __pa(addr);
1845		else
1846			return page_to_phys(vmalloc_to_page(addr)) +
1847			       offset_in_page(addr);
1848	} else
1849		return page_to_phys(pcpu_addr_to_page(addr)) +
1850		       offset_in_page(addr);
1851}
1852
1853/**
1854 * pcpu_alloc_alloc_info - allocate percpu allocation info
1855 * @nr_groups: the number of groups
1856 * @nr_units: the number of units
1857 *
1858 * Allocate ai which is large enough for @nr_groups groups containing
1859 * @nr_units units.  The returned ai's groups[0].cpu_map points to the
1860 * cpu_map array which is long enough for @nr_units and filled with
1861 * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
1862 * pointer of other groups.
1863 *
1864 * RETURNS:
1865 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1866 * failure.
1867 */
1868struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1869						      int nr_units)
1870{
1871	struct pcpu_alloc_info *ai;
1872	size_t base_size, ai_size;
1873	void *ptr;
1874	int unit;
1875
1876	base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1877			  __alignof__(ai->groups[0].cpu_map[0]));
1878	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1879
1880	ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), PAGE_SIZE);
1881	if (!ptr)
1882		return NULL;
1883	ai = ptr;
1884	ptr += base_size;
1885
1886	ai->groups[0].cpu_map = ptr;
1887
1888	for (unit = 0; unit < nr_units; unit++)
1889		ai->groups[0].cpu_map[unit] = NR_CPUS;
1890
1891	ai->nr_groups = nr_groups;
1892	ai->__ai_size = PFN_ALIGN(ai_size);
1893
1894	return ai;
1895}
1896
1897/**
1898 * pcpu_free_alloc_info - free percpu allocation info
1899 * @ai: pcpu_alloc_info to free
1900 *
1901 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1902 */
1903void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1904{
1905	memblock_free_early(__pa(ai), ai->__ai_size);
1906}
1907
1908/**
1909 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1910 * @lvl: loglevel
1911 * @ai: allocation info to dump
1912 *
1913 * Print out information about @ai using loglevel @lvl.
1914 */
1915static void pcpu_dump_alloc_info(const char *lvl,
1916				 const struct pcpu_alloc_info *ai)
1917{
1918	int group_width = 1, cpu_width = 1, width;
1919	char empty_str[] = "--------";
1920	int alloc = 0, alloc_end = 0;
1921	int group, v;
1922	int upa, apl;	/* units per alloc, allocs per line */
1923
1924	v = ai->nr_groups;
1925	while (v /= 10)
1926		group_width++;
1927
1928	v = num_possible_cpus();
1929	while (v /= 10)
1930		cpu_width++;
1931	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1932
1933	upa = ai->alloc_size / ai->unit_size;
1934	width = upa * (cpu_width + 1) + group_width + 3;
1935	apl = rounddown_pow_of_two(max(60 / width, 1));
1936
1937	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1938	       lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1939	       ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1940
1941	for (group = 0; group < ai->nr_groups; group++) {
1942		const struct pcpu_group_info *gi = &ai->groups[group];
1943		int unit = 0, unit_end = 0;
1944
1945		BUG_ON(gi->nr_units % upa);
1946		for (alloc_end += gi->nr_units / upa;
1947		     alloc < alloc_end; alloc++) {
1948			if (!(alloc % apl)) {
1949				pr_cont("\n");
1950				printk("%spcpu-alloc: ", lvl);
1951			}
1952			pr_cont("[%0*d] ", group_width, group);
1953
1954			for (unit_end += upa; unit < unit_end; unit++)
1955				if (gi->cpu_map[unit] != NR_CPUS)
1956					pr_cont("%0*d ",
1957						cpu_width, gi->cpu_map[unit]);
1958				else
1959					pr_cont("%s ", empty_str);
1960		}
1961	}
1962	pr_cont("\n");
1963}
1964
1965/**
1966 * pcpu_setup_first_chunk - initialize the first percpu chunk
1967 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1968 * @base_addr: mapped address
1969 *
1970 * Initialize the first percpu chunk which contains the kernel static
1971 * perpcu area.  This function is to be called from arch percpu area
1972 * setup path.
1973 *
1974 * @ai contains all information necessary to initialize the first
1975 * chunk and prime the dynamic percpu allocator.
1976 *
1977 * @ai->static_size is the size of static percpu area.
1978 *
1979 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1980 * reserve after the static area in the first chunk.  This reserves
1981 * the first chunk such that it's available only through reserved
1982 * percpu allocation.  This is primarily used to serve module percpu
1983 * static areas on architectures where the addressing model has
1984 * limited offset range for symbol relocations to guarantee module
1985 * percpu symbols fall inside the relocatable range.
1986 *
1987 * @ai->dyn_size determines the number of bytes available for dynamic
1988 * allocation in the first chunk.  The area between @ai->static_size +
1989 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1990 *
1991 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1992 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1993 * @ai->dyn_size.
1994 *
1995 * @ai->atom_size is the allocation atom size and used as alignment
1996 * for vm areas.
1997 *
1998 * @ai->alloc_size is the allocation size and always multiple of
1999 * @ai->atom_size.  This is larger than @ai->atom_size if
2000 * @ai->unit_size is larger than @ai->atom_size.
2001 *
2002 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2003 * percpu areas.  Units which should be colocated are put into the
2004 * same group.  Dynamic VM areas will be allocated according to these
2005 * groupings.  If @ai->nr_groups is zero, a single group containing
2006 * all units is assumed.
2007 *
2008 * The caller should have mapped the first chunk at @base_addr and
2009 * copied static data to each unit.
2010 *
2011 * The first chunk will always contain a static and a dynamic region.
2012 * However, the static region is not managed by any chunk.  If the first
2013 * chunk also contains a reserved region, it is served by two chunks -
2014 * one for the reserved region and one for the dynamic region.  They
2015 * share the same vm, but use offset regions in the area allocation map.
2016 * The chunk serving the dynamic region is circulated in the chunk slots
2017 * and available for dynamic allocation like any other chunk.
2018 *
2019 * RETURNS:
2020 * 0 on success, -errno on failure.
2021 */
2022int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2023				  void *base_addr)
2024{
2025	size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2026	size_t static_size, dyn_size;
2027	struct pcpu_chunk *chunk;
2028	unsigned long *group_offsets;
2029	size_t *group_sizes;
2030	unsigned long *unit_off;
2031	unsigned int cpu;
2032	int *unit_map;
2033	int group, unit, i;
2034	int map_size;
2035	unsigned long tmp_addr;
 
2036
2037#define PCPU_SETUP_BUG_ON(cond)	do {					\
2038	if (unlikely(cond)) {						\
2039		pr_emerg("failed to initialize, %s\n", #cond);		\
2040		pr_emerg("cpu_possible_mask=%*pb\n",			\
2041			 cpumask_pr_args(cpu_possible_mask));		\
2042		pcpu_dump_alloc_info(KERN_EMERG, ai);			\
2043		BUG();							\
2044	}								\
2045} while (0)
2046
2047	/* sanity checks */
2048	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2049#ifdef CONFIG_SMP
2050	PCPU_SETUP_BUG_ON(!ai->static_size);
2051	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2052#endif
2053	PCPU_SETUP_BUG_ON(!base_addr);
2054	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2055	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2056	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2057	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2058	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2059	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2060	PCPU_SETUP_BUG_ON(!ai->dyn_size);
2061	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2062	PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2063			    IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2064	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2065
2066	/* process group information and build config tables accordingly */
2067	group_offsets = memblock_virt_alloc(ai->nr_groups *
2068					     sizeof(group_offsets[0]), 0);
2069	group_sizes = memblock_virt_alloc(ai->nr_groups *
2070					   sizeof(group_sizes[0]), 0);
2071	unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
2072	unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2073
2074	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2075		unit_map[cpu] = UINT_MAX;
2076
2077	pcpu_low_unit_cpu = NR_CPUS;
2078	pcpu_high_unit_cpu = NR_CPUS;
2079
2080	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2081		const struct pcpu_group_info *gi = &ai->groups[group];
2082
2083		group_offsets[group] = gi->base_offset;
2084		group_sizes[group] = gi->nr_units * ai->unit_size;
2085
2086		for (i = 0; i < gi->nr_units; i++) {
2087			cpu = gi->cpu_map[i];
2088			if (cpu == NR_CPUS)
2089				continue;
2090
2091			PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2092			PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2093			PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2094
2095			unit_map[cpu] = unit + i;
2096			unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2097
2098			/* determine low/high unit_cpu */
2099			if (pcpu_low_unit_cpu == NR_CPUS ||
2100			    unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2101				pcpu_low_unit_cpu = cpu;
2102			if (pcpu_high_unit_cpu == NR_CPUS ||
2103			    unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2104				pcpu_high_unit_cpu = cpu;
2105		}
2106	}
2107	pcpu_nr_units = unit;
2108
2109	for_each_possible_cpu(cpu)
2110		PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2111
2112	/* we're done parsing the input, undefine BUG macro and dump config */
2113#undef PCPU_SETUP_BUG_ON
2114	pcpu_dump_alloc_info(KERN_DEBUG, ai);
2115
2116	pcpu_nr_groups = ai->nr_groups;
2117	pcpu_group_offsets = group_offsets;
2118	pcpu_group_sizes = group_sizes;
2119	pcpu_unit_map = unit_map;
2120	pcpu_unit_offsets = unit_off;
2121
2122	/* determine basic parameters */
2123	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2124	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2125	pcpu_atom_size = ai->atom_size;
2126	pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2127		BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2128
2129	pcpu_stats_save_ai(ai);
2130
2131	/*
2132	 * Allocate chunk slots.  The additional last slot is for
2133	 * empty chunks.
2134	 */
2135	pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2136	pcpu_slot = memblock_virt_alloc(
2137			pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
 
 
 
 
 
 
 
 
 
 
2138	for (i = 0; i < pcpu_nr_slots; i++)
2139		INIT_LIST_HEAD(&pcpu_slot[i]);
2140
2141	/*
2142	 * The end of the static region needs to be aligned with the
2143	 * minimum allocation size as this offsets the reserved and
2144	 * dynamic region.  The first chunk ends page aligned by
2145	 * expanding the dynamic region, therefore the dynamic region
2146	 * can be shrunk to compensate while still staying above the
2147	 * configured sizes.
2148	 */
2149	static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2150	dyn_size = ai->dyn_size - (static_size - ai->static_size);
2151
2152	/*
2153	 * Initialize first chunk.
2154	 * If the reserved_size is non-zero, this initializes the reserved
2155	 * chunk.  If the reserved_size is zero, the reserved chunk is NULL
2156	 * and the dynamic region is initialized here.  The first chunk,
2157	 * pcpu_first_chunk, will always point to the chunk that serves
2158	 * the dynamic region.
2159	 */
2160	tmp_addr = (unsigned long)base_addr + static_size;
2161	map_size = ai->reserved_size ?: dyn_size;
2162	chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2163
2164	/* init dynamic chunk if necessary */
2165	if (ai->reserved_size) {
2166		pcpu_reserved_chunk = chunk;
2167
2168		tmp_addr = (unsigned long)base_addr + static_size +
2169			   ai->reserved_size;
2170		map_size = dyn_size;
2171		chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2172	}
2173
2174	/* link the first chunk in */
2175	pcpu_first_chunk = chunk;
2176	pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2177	pcpu_chunk_relocate(pcpu_first_chunk, -1);
2178
 
 
 
2179	pcpu_stats_chunk_alloc();
2180	trace_percpu_create_chunk(base_addr);
2181
2182	/* we're done */
2183	pcpu_base_addr = base_addr;
2184	return 0;
2185}
2186
2187#ifdef CONFIG_SMP
2188
2189const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2190	[PCPU_FC_AUTO]	= "auto",
2191	[PCPU_FC_EMBED]	= "embed",
2192	[PCPU_FC_PAGE]	= "page",
2193};
2194
2195enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2196
2197static int __init percpu_alloc_setup(char *str)
2198{
2199	if (!str)
2200		return -EINVAL;
2201
2202	if (0)
2203		/* nada */;
2204#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2205	else if (!strcmp(str, "embed"))
2206		pcpu_chosen_fc = PCPU_FC_EMBED;
2207#endif
2208#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2209	else if (!strcmp(str, "page"))
2210		pcpu_chosen_fc = PCPU_FC_PAGE;
2211#endif
2212	else
2213		pr_warn("unknown allocator %s specified\n", str);
2214
2215	return 0;
2216}
2217early_param("percpu_alloc", percpu_alloc_setup);
2218
2219/*
2220 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2221 * Build it if needed by the arch config or the generic setup is going
2222 * to be used.
2223 */
2224#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2225	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2226#define BUILD_EMBED_FIRST_CHUNK
2227#endif
2228
2229/* build pcpu_page_first_chunk() iff needed by the arch config */
2230#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2231#define BUILD_PAGE_FIRST_CHUNK
2232#endif
2233
2234/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2235#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2236/**
2237 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2238 * @reserved_size: the size of reserved percpu area in bytes
2239 * @dyn_size: minimum free size for dynamic allocation in bytes
2240 * @atom_size: allocation atom size
2241 * @cpu_distance_fn: callback to determine distance between cpus, optional
2242 *
2243 * This function determines grouping of units, their mappings to cpus
2244 * and other parameters considering needed percpu size, allocation
2245 * atom size and distances between CPUs.
2246 *
2247 * Groups are always multiples of atom size and CPUs which are of
2248 * LOCAL_DISTANCE both ways are grouped together and share space for
2249 * units in the same group.  The returned configuration is guaranteed
2250 * to have CPUs on different nodes on different groups and >=75% usage
2251 * of allocated virtual address space.
2252 *
2253 * RETURNS:
2254 * On success, pointer to the new allocation_info is returned.  On
2255 * failure, ERR_PTR value is returned.
2256 */
2257static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2258				size_t reserved_size, size_t dyn_size,
2259				size_t atom_size,
2260				pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2261{
2262	static int group_map[NR_CPUS] __initdata;
2263	static int group_cnt[NR_CPUS] __initdata;
 
2264	const size_t static_size = __per_cpu_end - __per_cpu_start;
2265	int nr_groups = 1, nr_units = 0;
2266	size_t size_sum, min_unit_size, alloc_size;
2267	int upa, max_upa, uninitialized_var(best_upa);	/* units_per_alloc */
2268	int last_allocs, group, unit;
2269	unsigned int cpu, tcpu;
2270	struct pcpu_alloc_info *ai;
2271	unsigned int *cpu_map;
2272
2273	/* this function may be called multiple times */
2274	memset(group_map, 0, sizeof(group_map));
2275	memset(group_cnt, 0, sizeof(group_cnt));
 
2276
2277	/* calculate size_sum and ensure dyn_size is enough for early alloc */
2278	size_sum = PFN_ALIGN(static_size + reserved_size +
2279			    max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2280	dyn_size = size_sum - static_size - reserved_size;
2281
2282	/*
2283	 * Determine min_unit_size, alloc_size and max_upa such that
2284	 * alloc_size is multiple of atom_size and is the smallest
2285	 * which can accommodate 4k aligned segments which are equal to
2286	 * or larger than min_unit_size.
2287	 */
2288	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2289
2290	/* determine the maximum # of units that can fit in an allocation */
2291	alloc_size = roundup(min_unit_size, atom_size);
2292	upa = alloc_size / min_unit_size;
2293	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2294		upa--;
2295	max_upa = upa;
2296
 
 
2297	/* group cpus according to their proximity */
2298	for_each_possible_cpu(cpu) {
2299		group = 0;
2300	next_group:
2301		for_each_possible_cpu(tcpu) {
2302			if (cpu == tcpu)
2303				break;
2304			if (group_map[tcpu] == group && cpu_distance_fn &&
2305			    (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2306			     cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2307				group++;
2308				nr_groups = max(nr_groups, group + 1);
2309				goto next_group;
2310			}
2311		}
2312		group_map[cpu] = group;
2313		group_cnt[group]++;
 
 
 
 
 
 
 
 
 
 
 
2314	}
 
2315
2316	/*
2317	 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2318	 * Expand the unit_size until we use >= 75% of the units allocated.
2319	 * Related to atom_size, which could be much larger than the unit_size.
2320	 */
2321	last_allocs = INT_MAX;
 
2322	for (upa = max_upa; upa; upa--) {
2323		int allocs = 0, wasted = 0;
2324
2325		if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2326			continue;
2327
2328		for (group = 0; group < nr_groups; group++) {
2329			int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2330			allocs += this_allocs;
2331			wasted += this_allocs * upa - group_cnt[group];
2332		}
2333
2334		/*
2335		 * Don't accept if wastage is over 1/3.  The
2336		 * greater-than comparison ensures upa==1 always
2337		 * passes the following check.
2338		 */
2339		if (wasted > num_possible_cpus() / 3)
2340			continue;
2341
2342		/* and then don't consume more memory */
2343		if (allocs > last_allocs)
2344			break;
2345		last_allocs = allocs;
2346		best_upa = upa;
2347	}
 
2348	upa = best_upa;
2349
2350	/* allocate and fill alloc_info */
2351	for (group = 0; group < nr_groups; group++)
2352		nr_units += roundup(group_cnt[group], upa);
2353
2354	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2355	if (!ai)
2356		return ERR_PTR(-ENOMEM);
2357	cpu_map = ai->groups[0].cpu_map;
2358
2359	for (group = 0; group < nr_groups; group++) {
2360		ai->groups[group].cpu_map = cpu_map;
2361		cpu_map += roundup(group_cnt[group], upa);
2362	}
2363
2364	ai->static_size = static_size;
2365	ai->reserved_size = reserved_size;
2366	ai->dyn_size = dyn_size;
2367	ai->unit_size = alloc_size / upa;
2368	ai->atom_size = atom_size;
2369	ai->alloc_size = alloc_size;
2370
2371	for (group = 0, unit = 0; group_cnt[group]; group++) {
2372		struct pcpu_group_info *gi = &ai->groups[group];
2373
2374		/*
2375		 * Initialize base_offset as if all groups are located
2376		 * back-to-back.  The caller should update this to
2377		 * reflect actual allocation.
2378		 */
2379		gi->base_offset = unit * ai->unit_size;
2380
2381		for_each_possible_cpu(cpu)
2382			if (group_map[cpu] == group)
2383				gi->cpu_map[gi->nr_units++] = cpu;
2384		gi->nr_units = roundup(gi->nr_units, upa);
2385		unit += gi->nr_units;
2386	}
2387	BUG_ON(unit != nr_units);
2388
2389	return ai;
2390}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2391#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2392
2393#if defined(BUILD_EMBED_FIRST_CHUNK)
2394/**
2395 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2396 * @reserved_size: the size of reserved percpu area in bytes
2397 * @dyn_size: minimum free size for dynamic allocation in bytes
2398 * @atom_size: allocation atom size
2399 * @cpu_distance_fn: callback to determine distance between cpus, optional
2400 * @alloc_fn: function to allocate percpu page
2401 * @free_fn: function to free percpu page
2402 *
2403 * This is a helper to ease setting up embedded first percpu chunk and
2404 * can be called where pcpu_setup_first_chunk() is expected.
2405 *
2406 * If this function is used to setup the first chunk, it is allocated
2407 * by calling @alloc_fn and used as-is without being mapped into
2408 * vmalloc area.  Allocations are always whole multiples of @atom_size
2409 * aligned to @atom_size.
2410 *
2411 * This enables the first chunk to piggy back on the linear physical
2412 * mapping which often uses larger page size.  Please note that this
2413 * can result in very sparse cpu->unit mapping on NUMA machines thus
2414 * requiring large vmalloc address space.  Don't use this allocator if
2415 * vmalloc space is not orders of magnitude larger than distances
2416 * between node memory addresses (ie. 32bit NUMA machines).
2417 *
2418 * @dyn_size specifies the minimum dynamic area size.
2419 *
2420 * If the needed size is smaller than the minimum or specified unit
2421 * size, the leftover is returned using @free_fn.
2422 *
2423 * RETURNS:
2424 * 0 on success, -errno on failure.
2425 */
2426int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2427				  size_t atom_size,
2428				  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2429				  pcpu_fc_alloc_fn_t alloc_fn,
2430				  pcpu_fc_free_fn_t free_fn)
2431{
2432	void *base = (void *)ULONG_MAX;
2433	void **areas = NULL;
2434	struct pcpu_alloc_info *ai;
2435	size_t size_sum, areas_size;
2436	unsigned long max_distance;
2437	int group, i, highest_group, rc;
2438
2439	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2440				   cpu_distance_fn);
2441	if (IS_ERR(ai))
2442		return PTR_ERR(ai);
2443
2444	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2445	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2446
2447	areas = memblock_virt_alloc_nopanic(areas_size, 0);
2448	if (!areas) {
2449		rc = -ENOMEM;
2450		goto out_free;
2451	}
2452
2453	/* allocate, copy and determine base address & max_distance */
2454	highest_group = 0;
2455	for (group = 0; group < ai->nr_groups; group++) {
2456		struct pcpu_group_info *gi = &ai->groups[group];
2457		unsigned int cpu = NR_CPUS;
2458		void *ptr;
2459
2460		for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2461			cpu = gi->cpu_map[i];
2462		BUG_ON(cpu == NR_CPUS);
2463
2464		/* allocate space for the whole group */
2465		ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2466		if (!ptr) {
2467			rc = -ENOMEM;
2468			goto out_free_areas;
2469		}
2470		/* kmemleak tracks the percpu allocations separately */
2471		kmemleak_free(ptr);
2472		areas[group] = ptr;
2473
2474		base = min(ptr, base);
2475		if (ptr > areas[highest_group])
2476			highest_group = group;
2477	}
2478	max_distance = areas[highest_group] - base;
2479	max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2480
2481	/* warn if maximum distance is further than 75% of vmalloc space */
2482	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2483		pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2484				max_distance, VMALLOC_TOTAL);
2485#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2486		/* and fail if we have fallback */
2487		rc = -EINVAL;
2488		goto out_free_areas;
2489#endif
2490	}
2491
2492	/*
2493	 * Copy data and free unused parts.  This should happen after all
2494	 * allocations are complete; otherwise, we may end up with
2495	 * overlapping groups.
2496	 */
2497	for (group = 0; group < ai->nr_groups; group++) {
2498		struct pcpu_group_info *gi = &ai->groups[group];
2499		void *ptr = areas[group];
2500
2501		for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2502			if (gi->cpu_map[i] == NR_CPUS) {
2503				/* unused unit, free whole */
2504				free_fn(ptr, ai->unit_size);
2505				continue;
2506			}
2507			/* copy and return the unused part */
2508			memcpy(ptr, __per_cpu_load, ai->static_size);
2509			free_fn(ptr + size_sum, ai->unit_size - size_sum);
2510		}
2511	}
2512
2513	/* base address is now known, determine group base offsets */
2514	for (group = 0; group < ai->nr_groups; group++) {
2515		ai->groups[group].base_offset = areas[group] - base;
2516	}
2517
2518	pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2519		PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2520		ai->dyn_size, ai->unit_size);
2521
2522	rc = pcpu_setup_first_chunk(ai, base);
2523	goto out_free;
2524
2525out_free_areas:
2526	for (group = 0; group < ai->nr_groups; group++)
2527		if (areas[group])
2528			free_fn(areas[group],
2529				ai->groups[group].nr_units * ai->unit_size);
2530out_free:
2531	pcpu_free_alloc_info(ai);
2532	if (areas)
2533		memblock_free_early(__pa(areas), areas_size);
2534	return rc;
2535}
2536#endif /* BUILD_EMBED_FIRST_CHUNK */
2537
2538#ifdef BUILD_PAGE_FIRST_CHUNK
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2539/**
2540 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2541 * @reserved_size: the size of reserved percpu area in bytes
2542 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2543 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2544 * @populate_pte_fn: function to populate pte
2545 *
2546 * This is a helper to ease setting up page-remapped first percpu
2547 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2548 *
2549 * This is the basic allocator.  Static percpu area is allocated
2550 * page-by-page into vmalloc area.
2551 *
2552 * RETURNS:
2553 * 0 on success, -errno on failure.
2554 */
2555int __init pcpu_page_first_chunk(size_t reserved_size,
2556				 pcpu_fc_alloc_fn_t alloc_fn,
2557				 pcpu_fc_free_fn_t free_fn,
2558				 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2559{
2560	static struct vm_struct vm;
2561	struct pcpu_alloc_info *ai;
2562	char psize_str[16];
2563	int unit_pages;
2564	size_t pages_size;
2565	struct page **pages;
2566	int unit, i, j, rc;
2567	int upa;
2568	int nr_g0_units;
2569
2570	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2571
2572	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2573	if (IS_ERR(ai))
2574		return PTR_ERR(ai);
2575	BUG_ON(ai->nr_groups != 1);
2576	upa = ai->alloc_size/ai->unit_size;
2577	nr_g0_units = roundup(num_possible_cpus(), upa);
2578	if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2579		pcpu_free_alloc_info(ai);
2580		return -EINVAL;
2581	}
2582
2583	unit_pages = ai->unit_size >> PAGE_SHIFT;
2584
2585	/* unaligned allocations can't be freed, round up to page size */
2586	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2587			       sizeof(pages[0]));
2588	pages = memblock_virt_alloc(pages_size, 0);
 
 
 
2589
2590	/* allocate pages */
2591	j = 0;
2592	for (unit = 0; unit < num_possible_cpus(); unit++) {
2593		unsigned int cpu = ai->groups[0].cpu_map[unit];
2594		for (i = 0; i < unit_pages; i++) {
2595			void *ptr;
2596
2597			ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2598			if (!ptr) {
2599				pr_warn("failed to allocate %s page for cpu%u\n",
2600						psize_str, cpu);
2601				goto enomem;
2602			}
2603			/* kmemleak tracks the percpu allocations separately */
2604			kmemleak_free(ptr);
2605			pages[j++] = virt_to_page(ptr);
2606		}
2607	}
2608
2609	/* allocate vm area, map the pages and copy static data */
2610	vm.flags = VM_ALLOC;
2611	vm.size = num_possible_cpus() * ai->unit_size;
2612	vm_area_register_early(&vm, PAGE_SIZE);
2613
2614	for (unit = 0; unit < num_possible_cpus(); unit++) {
2615		unsigned long unit_addr =
2616			(unsigned long)vm.addr + unit * ai->unit_size;
2617
2618		for (i = 0; i < unit_pages; i++)
2619			populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2620
2621		/* pte already populated, the following shouldn't fail */
2622		rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2623				      unit_pages);
2624		if (rc < 0)
2625			panic("failed to map percpu area, err=%d\n", rc);
2626
2627		/*
2628		 * FIXME: Archs with virtual cache should flush local
2629		 * cache for the linear mapping here - something
2630		 * equivalent to flush_cache_vmap() on the local cpu.
2631		 * flush_cache_vmap() can't be used as most supporting
2632		 * data structures are not set up yet.
2633		 */
2634
2635		/* copy static data */
2636		memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2637	}
2638
2639	/* we're ready, commit */
2640	pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2641		unit_pages, psize_str, vm.addr, ai->static_size,
2642		ai->reserved_size, ai->dyn_size);
2643
2644	rc = pcpu_setup_first_chunk(ai, vm.addr);
2645	goto out_free_ar;
2646
2647enomem:
2648	while (--j >= 0)
2649		free_fn(page_address(pages[j]), PAGE_SIZE);
2650	rc = -ENOMEM;
2651out_free_ar:
2652	memblock_free_early(__pa(pages), pages_size);
2653	pcpu_free_alloc_info(ai);
2654	return rc;
2655}
2656#endif /* BUILD_PAGE_FIRST_CHUNK */
2657
2658#ifndef	CONFIG_HAVE_SETUP_PER_CPU_AREA
2659/*
2660 * Generic SMP percpu area setup.
2661 *
2662 * The embedding helper is used because its behavior closely resembles
2663 * the original non-dynamic generic percpu area setup.  This is
2664 * important because many archs have addressing restrictions and might
2665 * fail if the percpu area is located far away from the previous
2666 * location.  As an added bonus, in non-NUMA cases, embedding is
2667 * generally a good idea TLB-wise because percpu area can piggy back
2668 * on the physical linear memory mapping which uses large page
2669 * mappings on applicable archs.
2670 */
2671unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2672EXPORT_SYMBOL(__per_cpu_offset);
2673
2674static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2675				       size_t align)
2676{
2677	return  memblock_virt_alloc_from_nopanic(
2678			size, align, __pa(MAX_DMA_ADDRESS));
2679}
2680
2681static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2682{
2683	memblock_free_early(__pa(ptr), size);
2684}
2685
2686void __init setup_per_cpu_areas(void)
2687{
2688	unsigned long delta;
2689	unsigned int cpu;
2690	int rc;
2691
2692	/*
2693	 * Always reserve area for module percpu variables.  That's
2694	 * what the legacy allocator did.
2695	 */
2696	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2697				    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2698				    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2699	if (rc < 0)
2700		panic("Failed to initialize percpu areas.");
2701
2702	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2703	for_each_possible_cpu(cpu)
2704		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2705}
2706#endif	/* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2707
2708#else	/* CONFIG_SMP */
2709
2710/*
2711 * UP percpu area setup.
2712 *
2713 * UP always uses km-based percpu allocator with identity mapping.
2714 * Static percpu variables are indistinguishable from the usual static
2715 * variables and don't require any special preparation.
2716 */
2717void __init setup_per_cpu_areas(void)
2718{
2719	const size_t unit_size =
2720		roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2721					 PERCPU_DYNAMIC_RESERVE));
2722	struct pcpu_alloc_info *ai;
2723	void *fc;
2724
2725	ai = pcpu_alloc_alloc_info(1, 1);
2726	fc = memblock_virt_alloc_from_nopanic(unit_size,
2727					      PAGE_SIZE,
2728					      __pa(MAX_DMA_ADDRESS));
2729	if (!ai || !fc)
2730		panic("Failed to allocate memory for percpu areas.");
2731	/* kmemleak tracks the percpu allocations separately */
2732	kmemleak_free(fc);
2733
2734	ai->dyn_size = unit_size;
2735	ai->unit_size = unit_size;
2736	ai->atom_size = unit_size;
2737	ai->alloc_size = unit_size;
2738	ai->groups[0].nr_units = 1;
2739	ai->groups[0].cpu_map[0] = 0;
2740
2741	if (pcpu_setup_first_chunk(ai, fc) < 0)
2742		panic("Failed to initialize percpu areas.");
2743	pcpu_free_alloc_info(ai);
2744}
2745
2746#endif	/* CONFIG_SMP */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2747
2748/*
2749 * Percpu allocator is initialized early during boot when neither slab or
2750 * workqueue is available.  Plug async management until everything is up
2751 * and running.
2752 */
2753static int __init percpu_enable_async(void)
2754{
2755	pcpu_async_enabled = true;
2756	return 0;
2757}
2758subsys_initcall(percpu_enable_async);
v6.2
   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/list.h>
  76#include <linux/log2.h>
  77#include <linux/mm.h>
  78#include <linux/module.h>
  79#include <linux/mutex.h>
  80#include <linux/percpu.h>
  81#include <linux/pfn.h>
  82#include <linux/slab.h>
  83#include <linux/spinlock.h>
  84#include <linux/vmalloc.h>
  85#include <linux/workqueue.h>
  86#include <linux/kmemleak.h>
  87#include <linux/sched.h>
  88#include <linux/sched/mm.h>
  89#include <linux/memcontrol.h>
  90
  91#include <asm/cacheflush.h>
  92#include <asm/sections.h>
  93#include <asm/tlbflush.h>
  94#include <asm/io.h>
  95
  96#define CREATE_TRACE_POINTS
  97#include <trace/events/percpu.h>
  98
  99#include "percpu-internal.h"
 100
 101/*
 102 * The slots are sorted by the size of the biggest continuous free area.
 103 * 1-31 bytes share the same slot.
 104 */
 105#define PCPU_SLOT_BASE_SHIFT		5
 106/* chunks in slots below this are subject to being sidelined on failed alloc */
 107#define PCPU_SLOT_FAIL_THRESHOLD	3
 108
 109#define PCPU_EMPTY_POP_PAGES_LOW	2
 110#define PCPU_EMPTY_POP_PAGES_HIGH	4
 111
 112#ifdef CONFIG_SMP
 113/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 114#ifndef __addr_to_pcpu_ptr
 115#define __addr_to_pcpu_ptr(addr)					\
 116	(void __percpu *)((unsigned long)(addr) -			\
 117			  (unsigned long)pcpu_base_addr	+		\
 118			  (unsigned long)__per_cpu_start)
 119#endif
 120#ifndef __pcpu_ptr_to_addr
 121#define __pcpu_ptr_to_addr(ptr)						\
 122	(void __force *)((unsigned long)(ptr) +				\
 123			 (unsigned long)pcpu_base_addr -		\
 124			 (unsigned long)__per_cpu_start)
 125#endif
 126#else	/* CONFIG_SMP */
 127/* on UP, it's always identity mapped */
 128#define __addr_to_pcpu_ptr(addr)	(void __percpu *)(addr)
 129#define __pcpu_ptr_to_addr(ptr)		(void __force *)(ptr)
 130#endif	/* CONFIG_SMP */
 131
 132static int pcpu_unit_pages __ro_after_init;
 133static int pcpu_unit_size __ro_after_init;
 134static int pcpu_nr_units __ro_after_init;
 135static int pcpu_atom_size __ro_after_init;
 136int pcpu_nr_slots __ro_after_init;
 137static int pcpu_free_slot __ro_after_init;
 138int pcpu_sidelined_slot __ro_after_init;
 139int pcpu_to_depopulate_slot __ro_after_init;
 140static size_t pcpu_chunk_struct_size __ro_after_init;
 141
 142/* cpus with the lowest and highest unit addresses */
 143static unsigned int pcpu_low_unit_cpu __ro_after_init;
 144static unsigned int pcpu_high_unit_cpu __ro_after_init;
 145
 146/* the address of the first chunk which starts with the kernel static area */
 147void *pcpu_base_addr __ro_after_init;
 
 148
 149static const int *pcpu_unit_map __ro_after_init;		/* cpu -> unit */
 150const unsigned long *pcpu_unit_offsets __ro_after_init;	/* cpu -> unit offset */
 151
 152/* group information, used for vm allocation */
 153static int pcpu_nr_groups __ro_after_init;
 154static const unsigned long *pcpu_group_offsets __ro_after_init;
 155static const size_t *pcpu_group_sizes __ro_after_init;
 156
 157/*
 158 * The first chunk which always exists.  Note that unlike other
 159 * chunks, this one can be allocated and mapped in several different
 160 * ways and thus often doesn't live in the vmalloc area.
 161 */
 162struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
 163
 164/*
 165 * Optional reserved chunk.  This chunk reserves part of the first
 166 * chunk and serves it for reserved allocations.  When the reserved
 167 * region doesn't exist, the following variable is NULL.
 168 */
 169struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
 170
 171DEFINE_SPINLOCK(pcpu_lock);	/* all internal data structures */
 172static DEFINE_MUTEX(pcpu_alloc_mutex);	/* chunk create/destroy, [de]pop, map ext */
 173
 174struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
 
 
 
 175
 176/*
 177 * The number of empty populated pages, protected by pcpu_lock.
 178 * The reserved chunk doesn't contribute to the count.
 179 */
 180int pcpu_nr_empty_pop_pages;
 181
 182/*
 183 * The number of populated pages in use by the allocator, protected by
 184 * pcpu_lock.  This number is kept per a unit per chunk (i.e. when a page gets
 185 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
 186 * and increments/decrements this count by 1).
 187 */
 188static unsigned long pcpu_nr_populated;
 189
 190/*
 191 * Balance work is used to populate or destroy chunks asynchronously.  We
 192 * try to keep the number of populated free pages between
 193 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
 194 * empty chunk.
 195 */
 196static void pcpu_balance_workfn(struct work_struct *work);
 197static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
 198static bool pcpu_async_enabled __read_mostly;
 199static bool pcpu_atomic_alloc_failed;
 200
 201static void pcpu_schedule_balance_work(void)
 202{
 203	if (pcpu_async_enabled)
 204		schedule_work(&pcpu_balance_work);
 205}
 206
 207/**
 208 * pcpu_addr_in_chunk - check if the address is served from this chunk
 209 * @chunk: chunk of interest
 210 * @addr: percpu address
 211 *
 212 * RETURNS:
 213 * True if the address is served from this chunk.
 214 */
 215static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
 216{
 217	void *start_addr, *end_addr;
 218
 219	if (!chunk)
 220		return false;
 221
 222	start_addr = chunk->base_addr + chunk->start_offset;
 223	end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
 224		   chunk->end_offset;
 225
 226	return addr >= start_addr && addr < end_addr;
 227}
 228
 229static int __pcpu_size_to_slot(int size)
 230{
 231	int highbit = fls(size);	/* size is in bytes */
 232	return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
 233}
 234
 235static int pcpu_size_to_slot(int size)
 236{
 237	if (size == pcpu_unit_size)
 238		return pcpu_free_slot;
 239	return __pcpu_size_to_slot(size);
 240}
 241
 242static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
 243{
 244	const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 245
 246	if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
 247	    chunk_md->contig_hint == 0)
 248		return 0;
 249
 250	return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
 251}
 252
 253/* set the pointer to a chunk in a page struct */
 254static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
 255{
 256	page->index = (unsigned long)pcpu;
 257}
 258
 259/* obtain pointer to a chunk from a page struct */
 260static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
 261{
 262	return (struct pcpu_chunk *)page->index;
 263}
 264
 265static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
 266{
 267	return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
 268}
 269
 270static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
 271{
 272	return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
 273}
 274
 275static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
 276				     unsigned int cpu, int page_idx)
 277{
 278	return (unsigned long)chunk->base_addr +
 279	       pcpu_unit_page_offset(cpu, page_idx);
 280}
 281
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 282/*
 283 * The following are helper functions to help access bitmaps and convert
 284 * between bitmap offsets to address offsets.
 285 */
 286static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
 287{
 288	return chunk->alloc_map +
 289	       (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
 290}
 291
 292static unsigned long pcpu_off_to_block_index(int off)
 293{
 294	return off / PCPU_BITMAP_BLOCK_BITS;
 295}
 296
 297static unsigned long pcpu_off_to_block_off(int off)
 298{
 299	return off & (PCPU_BITMAP_BLOCK_BITS - 1);
 300}
 301
 302static unsigned long pcpu_block_off_to_off(int index, int off)
 303{
 304	return index * PCPU_BITMAP_BLOCK_BITS + off;
 305}
 306
 307/**
 308 * pcpu_check_block_hint - check against the contig hint
 309 * @block: block of interest
 310 * @bits: size of allocation
 311 * @align: alignment of area (max PAGE_SIZE)
 312 *
 313 * Check to see if the allocation can fit in the block's contig hint.
 314 * Note, a chunk uses the same hints as a block so this can also check against
 315 * the chunk's contig hint.
 316 */
 317static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
 318				  size_t align)
 319{
 320	int bit_off = ALIGN(block->contig_hint_start, align) -
 321		block->contig_hint_start;
 322
 323	return bit_off + bits <= block->contig_hint;
 324}
 325
 326/*
 327 * pcpu_next_hint - determine which hint to use
 328 * @block: block of interest
 329 * @alloc_bits: size of allocation
 330 *
 331 * This determines if we should scan based on the scan_hint or first_free.
 332 * In general, we want to scan from first_free to fulfill allocations by
 333 * first fit.  However, if we know a scan_hint at position scan_hint_start
 334 * cannot fulfill an allocation, we can begin scanning from there knowing
 335 * the contig_hint will be our fallback.
 336 */
 337static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
 338{
 339	/*
 340	 * The three conditions below determine if we can skip past the
 341	 * scan_hint.  First, does the scan hint exist.  Second, is the
 342	 * contig_hint after the scan_hint (possibly not true iff
 343	 * contig_hint == scan_hint).  Third, is the allocation request
 344	 * larger than the scan_hint.
 345	 */
 346	if (block->scan_hint &&
 347	    block->contig_hint_start > block->scan_hint_start &&
 348	    alloc_bits > block->scan_hint)
 349		return block->scan_hint_start + block->scan_hint;
 350
 351	return block->first_free;
 352}
 353
 354/**
 355 * pcpu_next_md_free_region - finds the next hint free area
 356 * @chunk: chunk of interest
 357 * @bit_off: chunk offset
 358 * @bits: size of free area
 359 *
 360 * Helper function for pcpu_for_each_md_free_region.  It checks
 361 * block->contig_hint and performs aggregation across blocks to find the
 362 * next hint.  It modifies bit_off and bits in-place to be consumed in the
 363 * loop.
 364 */
 365static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
 366				     int *bits)
 367{
 368	int i = pcpu_off_to_block_index(*bit_off);
 369	int block_off = pcpu_off_to_block_off(*bit_off);
 370	struct pcpu_block_md *block;
 371
 372	*bits = 0;
 373	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 374	     block++, i++) {
 375		/* handles contig area across blocks */
 376		if (*bits) {
 377			*bits += block->left_free;
 378			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 379				continue;
 380			return;
 381		}
 382
 383		/*
 384		 * This checks three things.  First is there a contig_hint to
 385		 * check.  Second, have we checked this hint before by
 386		 * comparing the block_off.  Third, is this the same as the
 387		 * right contig hint.  In the last case, it spills over into
 388		 * the next block and should be handled by the contig area
 389		 * across blocks code.
 390		 */
 391		*bits = block->contig_hint;
 392		if (*bits && block->contig_hint_start >= block_off &&
 393		    *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
 394			*bit_off = pcpu_block_off_to_off(i,
 395					block->contig_hint_start);
 396			return;
 397		}
 398		/* reset to satisfy the second predicate above */
 399		block_off = 0;
 400
 401		*bits = block->right_free;
 402		*bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
 403	}
 404}
 405
 406/**
 407 * pcpu_next_fit_region - finds fit areas for a given allocation request
 408 * @chunk: chunk of interest
 409 * @alloc_bits: size of allocation
 410 * @align: alignment of area (max PAGE_SIZE)
 411 * @bit_off: chunk offset
 412 * @bits: size of free area
 413 *
 414 * Finds the next free region that is viable for use with a given size and
 415 * alignment.  This only returns if there is a valid area to be used for this
 416 * allocation.  block->first_free is returned if the allocation request fits
 417 * within the block to see if the request can be fulfilled prior to the contig
 418 * hint.
 419 */
 420static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
 421				 int align, int *bit_off, int *bits)
 422{
 423	int i = pcpu_off_to_block_index(*bit_off);
 424	int block_off = pcpu_off_to_block_off(*bit_off);
 425	struct pcpu_block_md *block;
 426
 427	*bits = 0;
 428	for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
 429	     block++, i++) {
 430		/* handles contig area across blocks */
 431		if (*bits) {
 432			*bits += block->left_free;
 433			if (*bits >= alloc_bits)
 434				return;
 435			if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
 436				continue;
 437		}
 438
 439		/* check block->contig_hint */
 440		*bits = ALIGN(block->contig_hint_start, align) -
 441			block->contig_hint_start;
 442		/*
 443		 * This uses the block offset to determine if this has been
 444		 * checked in the prior iteration.
 445		 */
 446		if (block->contig_hint &&
 447		    block->contig_hint_start >= block_off &&
 448		    block->contig_hint >= *bits + alloc_bits) {
 449			int start = pcpu_next_hint(block, alloc_bits);
 450
 451			*bits += alloc_bits + block->contig_hint_start -
 452				 start;
 453			*bit_off = pcpu_block_off_to_off(i, start);
 454			return;
 455		}
 456		/* reset to satisfy the second predicate above */
 457		block_off = 0;
 458
 459		*bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
 460				 align);
 461		*bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
 462		*bit_off = pcpu_block_off_to_off(i, *bit_off);
 463		if (*bits >= alloc_bits)
 464			return;
 465	}
 466
 467	/* no valid offsets were found - fail condition */
 468	*bit_off = pcpu_chunk_map_bits(chunk);
 469}
 470
 471/*
 472 * Metadata free area iterators.  These perform aggregation of free areas
 473 * based on the metadata blocks and return the offset @bit_off and size in
 474 * bits of the free area @bits.  pcpu_for_each_fit_region only returns when
 475 * a fit is found for the allocation request.
 476 */
 477#define pcpu_for_each_md_free_region(chunk, bit_off, bits)		\
 478	for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits));	\
 479	     (bit_off) < pcpu_chunk_map_bits((chunk));			\
 480	     (bit_off) += (bits) + 1,					\
 481	     pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
 482
 483#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits)     \
 484	for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 485				  &(bits));				      \
 486	     (bit_off) < pcpu_chunk_map_bits((chunk));			      \
 487	     (bit_off) += (bits),					      \
 488	     pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
 489				  &(bits)))
 490
 491/**
 492 * pcpu_mem_zalloc - allocate memory
 493 * @size: bytes to allocate
 494 * @gfp: allocation flags
 495 *
 496 * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
 497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
 498 * This is to facilitate passing through whitelisted flags.  The
 499 * returned memory is always zeroed.
 500 *
 501 * RETURNS:
 502 * Pointer to the allocated area on success, NULL on failure.
 503 */
 504static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
 505{
 506	if (WARN_ON_ONCE(!slab_is_available()))
 507		return NULL;
 508
 509	if (size <= PAGE_SIZE)
 510		return kzalloc(size, gfp);
 511	else
 512		return __vmalloc(size, gfp | __GFP_ZERO);
 513}
 514
 515/**
 516 * pcpu_mem_free - free memory
 517 * @ptr: memory to free
 518 *
 519 * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
 520 */
 521static void pcpu_mem_free(void *ptr)
 522{
 523	kvfree(ptr);
 524}
 525
 526static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
 527			      bool move_front)
 528{
 529	if (chunk != pcpu_reserved_chunk) {
 530		if (move_front)
 531			list_move(&chunk->list, &pcpu_chunk_lists[slot]);
 532		else
 533			list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
 534	}
 535}
 536
 537static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
 538{
 539	__pcpu_chunk_move(chunk, slot, true);
 540}
 541
 542/**
 543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
 544 * @chunk: chunk of interest
 545 * @oslot: the previous slot it was on
 546 *
 547 * This function is called after an allocation or free changed @chunk.
 548 * New slot according to the changed state is determined and @chunk is
 549 * moved to the slot.  Note that the reserved chunk is never put on
 550 * chunk slots.
 551 *
 552 * CONTEXT:
 553 * pcpu_lock.
 554 */
 555static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
 556{
 557	int nslot = pcpu_chunk_slot(chunk);
 558
 559	/* leave isolated chunks in-place */
 560	if (chunk->isolated)
 561		return;
 562
 563	if (oslot != nslot)
 564		__pcpu_chunk_move(chunk, nslot, oslot < nslot);
 565}
 566
 567static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 568{
 569	lockdep_assert_held(&pcpu_lock);
 
 570
 571	if (!chunk->isolated) {
 572		chunk->isolated = true;
 573		pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
 574	}
 575	list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
 576}
 577
 578static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
 579{
 580	lockdep_assert_held(&pcpu_lock);
 581
 582	if (chunk->isolated) {
 583		chunk->isolated = false;
 584		pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
 585		pcpu_chunk_relocate(chunk, -1);
 586	}
 587}
 588
 589/*
 590 * pcpu_update_empty_pages - update empty page counters
 591 * @chunk: chunk of interest
 592 * @nr: nr of empty pages
 
 593 *
 594 * This is used to keep track of the empty pages now based on the premise
 595 * a md_block covers a page.  The hint update functions recognize if a block
 596 * is made full or broken to calculate deltas for keeping track of free pages.
 597 */
 598static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
 599{
 600	chunk->nr_empty_pop_pages += nr;
 601	if (chunk != pcpu_reserved_chunk && !chunk->isolated)
 602		pcpu_nr_empty_pop_pages += nr;
 
 
 
 
 
 
 603}
 604
 605/*
 606 * pcpu_region_overlap - determines if two regions overlap
 607 * @a: start of first region, inclusive
 608 * @b: end of first region, exclusive
 609 * @x: start of second region, inclusive
 610 * @y: end of second region, exclusive
 
 611 *
 612 * This is used to determine if the hint region [a, b) overlaps with the
 613 * allocated region [x, y).
 
 
 614 */
 615static inline bool pcpu_region_overlap(int a, int b, int x, int y)
 616{
 617	return (a < y) && (x < b);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 618}
 619
 620/**
 621 * pcpu_block_update - updates a block given a free area
 622 * @block: block of interest
 623 * @start: start offset in block
 624 * @end: end offset in block
 625 *
 626 * Updates a block given a known free area.  The region [start, end) is
 627 * expected to be the entirety of the free area within a block.  Chooses
 628 * the best starting offset if the contig hints are equal.
 629 */
 630static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
 631{
 632	int contig = end - start;
 633
 634	block->first_free = min(block->first_free, start);
 635	if (start == 0)
 636		block->left_free = contig;
 637
 638	if (end == block->nr_bits)
 639		block->right_free = contig;
 640
 641	if (contig > block->contig_hint) {
 642		/* promote the old contig_hint to be the new scan_hint */
 643		if (start > block->contig_hint_start) {
 644			if (block->contig_hint > block->scan_hint) {
 645				block->scan_hint_start =
 646					block->contig_hint_start;
 647				block->scan_hint = block->contig_hint;
 648			} else if (start < block->scan_hint_start) {
 649				/*
 650				 * The old contig_hint == scan_hint.  But, the
 651				 * new contig is larger so hold the invariant
 652				 * scan_hint_start < contig_hint_start.
 653				 */
 654				block->scan_hint = 0;
 655			}
 656		} else {
 657			block->scan_hint = 0;
 658		}
 659		block->contig_hint_start = start;
 660		block->contig_hint = contig;
 661	} else if (contig == block->contig_hint) {
 662		if (block->contig_hint_start &&
 663		    (!start ||
 664		     __ffs(start) > __ffs(block->contig_hint_start))) {
 665			/* start has a better alignment so use it */
 666			block->contig_hint_start = start;
 667			if (start < block->scan_hint_start &&
 668			    block->contig_hint > block->scan_hint)
 669				block->scan_hint = 0;
 670		} else if (start > block->scan_hint_start ||
 671			   block->contig_hint > block->scan_hint) {
 672			/*
 673			 * Knowing contig == contig_hint, update the scan_hint
 674			 * if it is farther than or larger than the current
 675			 * scan_hint.
 676			 */
 677			block->scan_hint_start = start;
 678			block->scan_hint = contig;
 679		}
 680	} else {
 681		/*
 682		 * The region is smaller than the contig_hint.  So only update
 683		 * the scan_hint if it is larger than or equal and farther than
 684		 * the current scan_hint.
 685		 */
 686		if ((start < block->contig_hint_start &&
 687		     (contig > block->scan_hint ||
 688		      (contig == block->scan_hint &&
 689		       start > block->scan_hint_start)))) {
 690			block->scan_hint_start = start;
 691			block->scan_hint = contig;
 692		}
 693	}
 694}
 695
 696/*
 697 * pcpu_block_update_scan - update a block given a free area from a scan
 698 * @chunk: chunk of interest
 699 * @bit_off: chunk offset
 700 * @bits: size of free area
 701 *
 702 * Finding the final allocation spot first goes through pcpu_find_block_fit()
 703 * to find a block that can hold the allocation and then pcpu_alloc_area()
 704 * where a scan is used.  When allocations require specific alignments,
 705 * we can inadvertently create holes which will not be seen in the alloc
 706 * or free paths.
 707 *
 708 * This takes a given free area hole and updates a block as it may change the
 709 * scan_hint.  We need to scan backwards to ensure we don't miss free bits
 710 * from alignment.
 711 */
 712static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
 713				   int bits)
 714{
 715	int s_off = pcpu_off_to_block_off(bit_off);
 716	int e_off = s_off + bits;
 717	int s_index, l_bit;
 718	struct pcpu_block_md *block;
 719
 720	if (e_off > PCPU_BITMAP_BLOCK_BITS)
 721		return;
 722
 723	s_index = pcpu_off_to_block_index(bit_off);
 724	block = chunk->md_blocks + s_index;
 725
 726	/* scan backwards in case of alignment skipping free bits */
 727	l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
 728	s_off = (s_off == l_bit) ? 0 : l_bit + 1;
 729
 730	pcpu_block_update(block, s_off, e_off);
 731}
 732
 733/**
 734 * pcpu_chunk_refresh_hint - updates metadata about a chunk
 735 * @chunk: chunk of interest
 736 * @full_scan: if we should scan from the beginning
 737 *
 738 * Iterates over the metadata blocks to find the largest contig area.
 739 * A full scan can be avoided on the allocation path as this is triggered
 740 * if we broke the contig_hint.  In doing so, the scan_hint will be before
 741 * the contig_hint or after if the scan_hint == contig_hint.  This cannot
 742 * be prevented on freeing as we want to find the largest area possibly
 743 * spanning blocks.
 744 */
 745static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
 746{
 747	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 748	int bit_off, bits;
 749
 750	/* promote scan_hint to contig_hint */
 751	if (!full_scan && chunk_md->scan_hint) {
 752		bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
 753		chunk_md->contig_hint_start = chunk_md->scan_hint_start;
 754		chunk_md->contig_hint = chunk_md->scan_hint;
 755		chunk_md->scan_hint = 0;
 756	} else {
 757		bit_off = chunk_md->first_free;
 758		chunk_md->contig_hint = 0;
 759	}
 760
 761	bits = 0;
 762	pcpu_for_each_md_free_region(chunk, bit_off, bits)
 763		pcpu_block_update(chunk_md, bit_off, bit_off + bits);
 764}
 765
 766/**
 767 * pcpu_block_refresh_hint
 768 * @chunk: chunk of interest
 769 * @index: index of the metadata block
 770 *
 771 * Scans over the block beginning at first_free and updates the block
 772 * metadata accordingly.
 773 */
 774static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
 775{
 776	struct pcpu_block_md *block = chunk->md_blocks + index;
 777	unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
 778	unsigned int start, end;	/* region start, region end */
 779
 780	/* promote scan_hint to contig_hint */
 781	if (block->scan_hint) {
 782		start = block->scan_hint_start + block->scan_hint;
 783		block->contig_hint_start = block->scan_hint_start;
 784		block->contig_hint = block->scan_hint;
 785		block->scan_hint = 0;
 786	} else {
 787		start = block->first_free;
 788		block->contig_hint = 0;
 789	}
 790
 791	block->right_free = 0;
 
 
 792
 793	/* iterate over free areas and update the contig hints */
 794	for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
 795		pcpu_block_update(block, start, end);
 
 
 796}
 797
 798/**
 799 * pcpu_block_update_hint_alloc - update hint on allocation path
 800 * @chunk: chunk of interest
 801 * @bit_off: chunk offset
 802 * @bits: size of request
 803 *
 804 * Updates metadata for the allocation path.  The metadata only has to be
 805 * refreshed by a full scan iff the chunk's contig hint is broken.  Block level
 806 * scans are required if the block's contig hint is broken.
 807 */
 808static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
 809					 int bits)
 810{
 811	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
 812	int nr_empty_pages = 0;
 813	struct pcpu_block_md *s_block, *e_block, *block;
 814	int s_index, e_index;	/* block indexes of the freed allocation */
 815	int s_off, e_off;	/* block offsets of the freed allocation */
 816
 817	/*
 818	 * Calculate per block offsets.
 819	 * The calculation uses an inclusive range, but the resulting offsets
 820	 * are [start, end).  e_index always points to the last block in the
 821	 * range.
 822	 */
 823	s_index = pcpu_off_to_block_index(bit_off);
 824	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 825	s_off = pcpu_off_to_block_off(bit_off);
 826	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 827
 828	s_block = chunk->md_blocks + s_index;
 829	e_block = chunk->md_blocks + e_index;
 830
 831	/*
 832	 * Update s_block.
 833	 */
 834	if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 835		nr_empty_pages++;
 836
 837	/*
 838	 * block->first_free must be updated if the allocation takes its place.
 839	 * If the allocation breaks the contig_hint, a scan is required to
 840	 * restore this hint.
 841	 */
 842	if (s_off == s_block->first_free)
 843		s_block->first_free = find_next_zero_bit(
 844					pcpu_index_alloc_map(chunk, s_index),
 845					PCPU_BITMAP_BLOCK_BITS,
 846					s_off + bits);
 847
 848	if (pcpu_region_overlap(s_block->scan_hint_start,
 849				s_block->scan_hint_start + s_block->scan_hint,
 850				s_off,
 851				s_off + bits))
 852		s_block->scan_hint = 0;
 853
 854	if (pcpu_region_overlap(s_block->contig_hint_start,
 855				s_block->contig_hint_start +
 856				s_block->contig_hint,
 857				s_off,
 858				s_off + bits)) {
 859		/* block contig hint is broken - scan to fix it */
 860		if (!s_off)
 861			s_block->left_free = 0;
 862		pcpu_block_refresh_hint(chunk, s_index);
 863	} else {
 864		/* update left and right contig manually */
 865		s_block->left_free = min(s_block->left_free, s_off);
 866		if (s_index == e_index)
 867			s_block->right_free = min_t(int, s_block->right_free,
 868					PCPU_BITMAP_BLOCK_BITS - e_off);
 869		else
 870			s_block->right_free = 0;
 871	}
 872
 873	/*
 874	 * Update e_block.
 875	 */
 876	if (s_index != e_index) {
 877		if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
 878			nr_empty_pages++;
 879
 880		/*
 881		 * When the allocation is across blocks, the end is along
 882		 * the left part of the e_block.
 883		 */
 884		e_block->first_free = find_next_zero_bit(
 885				pcpu_index_alloc_map(chunk, e_index),
 886				PCPU_BITMAP_BLOCK_BITS, e_off);
 887
 888		if (e_off == PCPU_BITMAP_BLOCK_BITS) {
 889			/* reset the block */
 890			e_block++;
 891		} else {
 892			if (e_off > e_block->scan_hint_start)
 893				e_block->scan_hint = 0;
 894
 895			e_block->left_free = 0;
 896			if (e_off > e_block->contig_hint_start) {
 897				/* contig hint is broken - scan to fix it */
 898				pcpu_block_refresh_hint(chunk, e_index);
 899			} else {
 
 900				e_block->right_free =
 901					min_t(int, e_block->right_free,
 902					      PCPU_BITMAP_BLOCK_BITS - e_off);
 903			}
 904		}
 905
 906		/* update in-between md_blocks */
 907		nr_empty_pages += (e_index - s_index - 1);
 908		for (block = s_block + 1; block < e_block; block++) {
 909			block->scan_hint = 0;
 910			block->contig_hint = 0;
 911			block->left_free = 0;
 912			block->right_free = 0;
 913		}
 914	}
 915
 916	/*
 917	 * If the allocation is not atomic, some blocks may not be
 918	 * populated with pages, while we account it here.  The number
 919	 * of pages will be added back with pcpu_chunk_populated()
 920	 * when populating pages.
 921	 */
 922	if (nr_empty_pages)
 923		pcpu_update_empty_pages(chunk, -nr_empty_pages);
 924
 925	if (pcpu_region_overlap(chunk_md->scan_hint_start,
 926				chunk_md->scan_hint_start +
 927				chunk_md->scan_hint,
 928				bit_off,
 929				bit_off + bits))
 930		chunk_md->scan_hint = 0;
 931
 932	/*
 933	 * The only time a full chunk scan is required is if the chunk
 934	 * contig hint is broken.  Otherwise, it means a smaller space
 935	 * was used and therefore the chunk contig hint is still correct.
 936	 */
 937	if (pcpu_region_overlap(chunk_md->contig_hint_start,
 938				chunk_md->contig_hint_start +
 939				chunk_md->contig_hint,
 940				bit_off,
 941				bit_off + bits))
 942		pcpu_chunk_refresh_hint(chunk, false);
 943}
 944
 945/**
 946 * pcpu_block_update_hint_free - updates the block hints on the free path
 947 * @chunk: chunk of interest
 948 * @bit_off: chunk offset
 949 * @bits: size of request
 950 *
 951 * Updates metadata for the allocation path.  This avoids a blind block
 952 * refresh by making use of the block contig hints.  If this fails, it scans
 953 * forward and backward to determine the extent of the free area.  This is
 954 * capped at the boundary of blocks.
 955 *
 956 * A chunk update is triggered if a page becomes free, a block becomes free,
 957 * or the free spans across blocks.  This tradeoff is to minimize iterating
 958 * over the block metadata to update chunk_md->contig_hint.
 959 * chunk_md->contig_hint may be off by up to a page, but it will never be more
 960 * than the available space.  If the contig hint is contained in one block, it
 961 * will be accurate.
 962 */
 963static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
 964					int bits)
 965{
 966	int nr_empty_pages = 0;
 967	struct pcpu_block_md *s_block, *e_block, *block;
 968	int s_index, e_index;	/* block indexes of the freed allocation */
 969	int s_off, e_off;	/* block offsets of the freed allocation */
 970	int start, end;		/* start and end of the whole free area */
 971
 972	/*
 973	 * Calculate per block offsets.
 974	 * The calculation uses an inclusive range, but the resulting offsets
 975	 * are [start, end).  e_index always points to the last block in the
 976	 * range.
 977	 */
 978	s_index = pcpu_off_to_block_index(bit_off);
 979	e_index = pcpu_off_to_block_index(bit_off + bits - 1);
 980	s_off = pcpu_off_to_block_off(bit_off);
 981	e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
 982
 983	s_block = chunk->md_blocks + s_index;
 984	e_block = chunk->md_blocks + e_index;
 985
 986	/*
 987	 * Check if the freed area aligns with the block->contig_hint.
 988	 * If it does, then the scan to find the beginning/end of the
 989	 * larger free area can be avoided.
 990	 *
 991	 * start and end refer to beginning and end of the free area
 992	 * within each their respective blocks.  This is not necessarily
 993	 * the entire free area as it may span blocks past the beginning
 994	 * or end of the block.
 995	 */
 996	start = s_off;
 997	if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
 998		start = s_block->contig_hint_start;
 999	} else {
1000		/*
1001		 * Scan backwards to find the extent of the free area.
1002		 * find_last_bit returns the starting bit, so if the start bit
1003		 * is returned, that means there was no last bit and the
1004		 * remainder of the chunk is free.
1005		 */
1006		int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1007					  start);
1008		start = (start == l_bit) ? 0 : l_bit + 1;
1009	}
1010
1011	end = e_off;
1012	if (e_off == e_block->contig_hint_start)
1013		end = e_block->contig_hint_start + e_block->contig_hint;
1014	else
1015		end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1016				    PCPU_BITMAP_BLOCK_BITS, end);
1017
1018	/* update s_block */
1019	e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1020	if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1021		nr_empty_pages++;
1022	pcpu_block_update(s_block, start, e_off);
1023
1024	/* freeing in the same block */
1025	if (s_index != e_index) {
1026		/* update e_block */
1027		if (end == PCPU_BITMAP_BLOCK_BITS)
1028			nr_empty_pages++;
1029		pcpu_block_update(e_block, 0, end);
1030
1031		/* reset md_blocks in the middle */
1032		nr_empty_pages += (e_index - s_index - 1);
1033		for (block = s_block + 1; block < e_block; block++) {
1034			block->first_free = 0;
1035			block->scan_hint = 0;
1036			block->contig_hint_start = 0;
1037			block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1038			block->left_free = PCPU_BITMAP_BLOCK_BITS;
1039			block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040		}
1041	}
1042
1043	if (nr_empty_pages)
1044		pcpu_update_empty_pages(chunk, nr_empty_pages);
1045
1046	/*
1047	 * Refresh chunk metadata when the free makes a block free or spans
1048	 * across blocks.  The contig_hint may be off by up to a page, but if
1049	 * the contig_hint is contained in a block, it will be accurate with
1050	 * the else condition below.
1051	 */
1052	if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1053		pcpu_chunk_refresh_hint(chunk, true);
 
 
1054	else
1055		pcpu_block_update(&chunk->chunk_md,
1056				  pcpu_block_off_to_off(s_index, start),
1057				  end);
1058}
1059
1060/**
1061 * pcpu_is_populated - determines if the region is populated
1062 * @chunk: chunk of interest
1063 * @bit_off: chunk offset
1064 * @bits: size of area
1065 * @next_off: return value for the next offset to start searching
1066 *
1067 * For atomic allocations, check if the backing pages are populated.
1068 *
1069 * RETURNS:
1070 * Bool if the backing pages are populated.
1071 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1072 */
1073static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1074			      int *next_off)
1075{
1076	unsigned int start, end;
1077
1078	start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1079	end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1080
1081	start = find_next_zero_bit(chunk->populated, end, start);
1082	if (start >= end)
 
1083		return true;
1084
1085	end = find_next_bit(chunk->populated, end, start + 1);
1086
1087	*next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088	return false;
1089}
1090
1091/**
1092 * pcpu_find_block_fit - finds the block index to start searching
1093 * @chunk: chunk of interest
1094 * @alloc_bits: size of request in allocation units
1095 * @align: alignment of area (max PAGE_SIZE bytes)
1096 * @pop_only: use populated regions only
1097 *
1098 * Given a chunk and an allocation spec, find the offset to begin searching
1099 * for a free region.  This iterates over the bitmap metadata blocks to
1100 * find an offset that will be guaranteed to fit the requirements.  It is
1101 * not quite first fit as if the allocation does not fit in the contig hint
1102 * of a block or chunk, it is skipped.  This errs on the side of caution
1103 * to prevent excess iteration.  Poor alignment can cause the allocator to
1104 * skip over blocks and chunks that have valid free areas.
1105 *
1106 * RETURNS:
1107 * The offset in the bitmap to begin searching.
1108 * -1 if no offset is found.
1109 */
1110static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1111			       size_t align, bool pop_only)
1112{
1113	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1114	int bit_off, bits, next_off;
1115
1116	/*
1117	 * This is an optimization to prevent scanning by assuming if the
1118	 * allocation cannot fit in the global hint, there is memory pressure
1119	 * and creating a new chunk would happen soon.
1120	 */
1121	if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
 
 
 
1122		return -1;
1123
1124	bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1125	bits = 0;
1126	pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1127		if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1128						   &next_off))
1129			break;
1130
1131		bit_off = next_off;
1132		bits = 0;
1133	}
1134
1135	if (bit_off == pcpu_chunk_map_bits(chunk))
1136		return -1;
1137
1138	return bit_off;
1139}
1140
1141/*
1142 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1143 * @map: the address to base the search on
1144 * @size: the bitmap size in bits
1145 * @start: the bitnumber to start searching at
1146 * @nr: the number of zeroed bits we're looking for
1147 * @align_mask: alignment mask for zero area
1148 * @largest_off: offset of the largest area skipped
1149 * @largest_bits: size of the largest area skipped
1150 *
1151 * The @align_mask should be one less than a power of 2.
1152 *
1153 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1154 * the largest area that was skipped.  This is imperfect, but in general is
1155 * good enough.  The largest remembered region is the largest failed region
1156 * seen.  This does not include anything we possibly skipped due to alignment.
1157 * pcpu_block_update_scan() does scan backwards to try and recover what was
1158 * lost to alignment.  While this can cause scanning to miss earlier possible
1159 * free areas, smaller allocations will eventually fill those holes.
1160 */
1161static unsigned long pcpu_find_zero_area(unsigned long *map,
1162					 unsigned long size,
1163					 unsigned long start,
1164					 unsigned long nr,
1165					 unsigned long align_mask,
1166					 unsigned long *largest_off,
1167					 unsigned long *largest_bits)
1168{
1169	unsigned long index, end, i, area_off, area_bits;
1170again:
1171	index = find_next_zero_bit(map, size, start);
1172
1173	/* Align allocation */
1174	index = __ALIGN_MASK(index, align_mask);
1175	area_off = index;
1176
1177	end = index + nr;
1178	if (end > size)
1179		return end;
1180	i = find_next_bit(map, end, index);
1181	if (i < end) {
1182		area_bits = i - area_off;
1183		/* remember largest unused area with best alignment */
1184		if (area_bits > *largest_bits ||
1185		    (area_bits == *largest_bits && *largest_off &&
1186		     (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1187			*largest_off = area_off;
1188			*largest_bits = area_bits;
1189		}
1190
1191		start = i + 1;
1192		goto again;
1193	}
1194	return index;
1195}
1196
1197/**
1198 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1199 * @chunk: chunk of interest
1200 * @alloc_bits: size of request in allocation units
1201 * @align: alignment of area (max PAGE_SIZE)
1202 * @start: bit_off to start searching
1203 *
1204 * This function takes in a @start offset to begin searching to fit an
1205 * allocation of @alloc_bits with alignment @align.  It needs to scan
1206 * the allocation map because if it fits within the block's contig hint,
1207 * @start will be block->first_free. This is an attempt to fill the
1208 * allocation prior to breaking the contig hint.  The allocation and
1209 * boundary maps are updated accordingly if it confirms a valid
1210 * free area.
1211 *
1212 * RETURNS:
1213 * Allocated addr offset in @chunk on success.
1214 * -1 if no matching area is found.
1215 */
1216static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1217			   size_t align, int start)
1218{
1219	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1220	size_t align_mask = (align) ? (align - 1) : 0;
1221	unsigned long area_off = 0, area_bits = 0;
1222	int bit_off, end, oslot;
1223
1224	lockdep_assert_held(&pcpu_lock);
1225
1226	oslot = pcpu_chunk_slot(chunk);
1227
1228	/*
1229	 * Search to find a fit.
1230	 */
1231	end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1232		    pcpu_chunk_map_bits(chunk));
1233	bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1234				      align_mask, &area_off, &area_bits);
1235	if (bit_off >= end)
1236		return -1;
1237
1238	if (area_bits)
1239		pcpu_block_update_scan(chunk, area_off, area_bits);
1240
1241	/* update alloc map */
1242	bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1243
1244	/* update boundary map */
1245	set_bit(bit_off, chunk->bound_map);
1246	bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1247	set_bit(bit_off + alloc_bits, chunk->bound_map);
1248
1249	chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1250
1251	/* update first free bit */
1252	if (bit_off == chunk_md->first_free)
1253		chunk_md->first_free = find_next_zero_bit(
1254					chunk->alloc_map,
1255					pcpu_chunk_map_bits(chunk),
1256					bit_off + alloc_bits);
1257
1258	pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1259
1260	pcpu_chunk_relocate(chunk, oslot);
1261
1262	return bit_off * PCPU_MIN_ALLOC_SIZE;
1263}
1264
1265/**
1266 * pcpu_free_area - frees the corresponding offset
1267 * @chunk: chunk of interest
1268 * @off: addr offset into chunk
1269 *
1270 * This function determines the size of an allocation to free using
1271 * the boundary bitmap and clears the allocation map.
1272 *
1273 * RETURNS:
1274 * Number of freed bytes.
1275 */
1276static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1277{
1278	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1279	int bit_off, bits, end, oslot, freed;
1280
1281	lockdep_assert_held(&pcpu_lock);
1282	pcpu_stats_area_dealloc(chunk);
1283
1284	oslot = pcpu_chunk_slot(chunk);
1285
1286	bit_off = off / PCPU_MIN_ALLOC_SIZE;
1287
1288	/* find end index */
1289	end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1290			    bit_off + 1);
1291	bits = end - bit_off;
1292	bitmap_clear(chunk->alloc_map, bit_off, bits);
1293
1294	freed = bits * PCPU_MIN_ALLOC_SIZE;
1295
1296	/* update metadata */
1297	chunk->free_bytes += freed;
1298
1299	/* update first free bit */
1300	chunk_md->first_free = min(chunk_md->first_free, bit_off);
1301
1302	pcpu_block_update_hint_free(chunk, bit_off, bits);
1303
1304	pcpu_chunk_relocate(chunk, oslot);
1305
1306	return freed;
1307}
1308
1309static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1310{
1311	block->scan_hint = 0;
1312	block->contig_hint = nr_bits;
1313	block->left_free = nr_bits;
1314	block->right_free = nr_bits;
1315	block->first_free = 0;
1316	block->nr_bits = nr_bits;
1317}
1318
1319static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1320{
1321	struct pcpu_block_md *md_block;
1322
1323	/* init the chunk's block */
1324	pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1325
1326	for (md_block = chunk->md_blocks;
1327	     md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1328	     md_block++)
1329		pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
 
 
 
1330}
1331
1332/**
1333 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1334 * @tmp_addr: the start of the region served
1335 * @map_size: size of the region served
1336 *
1337 * This is responsible for creating the chunks that serve the first chunk.  The
1338 * base_addr is page aligned down of @tmp_addr while the region end is page
1339 * aligned up.  Offsets are kept track of to determine the region served. All
1340 * this is done to appease the bitmap allocator in avoiding partial blocks.
1341 *
1342 * RETURNS:
1343 * Chunk serving the region at @tmp_addr of @map_size.
1344 */
1345static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1346							 int map_size)
1347{
1348	struct pcpu_chunk *chunk;
1349	unsigned long aligned_addr;
1350	int start_offset, offset_bits, region_size, region_bits;
1351	size_t alloc_size;
1352
1353	/* region calculations */
1354	aligned_addr = tmp_addr & PAGE_MASK;
1355
1356	start_offset = tmp_addr - aligned_addr;
1357	region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
 
 
 
 
 
 
 
1358
1359	/* allocate chunk */
1360	alloc_size = struct_size(chunk, populated,
1361				 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1362	chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1363	if (!chunk)
1364		panic("%s: Failed to allocate %zu bytes\n", __func__,
1365		      alloc_size);
1366
1367	INIT_LIST_HEAD(&chunk->list);
1368
1369	chunk->base_addr = (void *)aligned_addr;
1370	chunk->start_offset = start_offset;
1371	chunk->end_offset = region_size - chunk->start_offset - map_size;
1372
1373	chunk->nr_pages = region_size >> PAGE_SHIFT;
1374	region_bits = pcpu_chunk_map_bits(chunk);
1375
1376	alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1377	chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1378	if (!chunk->alloc_map)
1379		panic("%s: Failed to allocate %zu bytes\n", __func__,
1380		      alloc_size);
1381
1382	alloc_size =
1383		BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1384	chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1385	if (!chunk->bound_map)
1386		panic("%s: Failed to allocate %zu bytes\n", __func__,
1387		      alloc_size);
1388
1389	alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1390	chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1391	if (!chunk->md_blocks)
1392		panic("%s: Failed to allocate %zu bytes\n", __func__,
1393		      alloc_size);
1394
1395#ifdef CONFIG_MEMCG_KMEM
1396	/* first chunk is free to use */
1397	chunk->obj_cgroups = NULL;
1398#endif
1399	pcpu_init_md_blocks(chunk);
1400
1401	/* manage populated page bitmap */
1402	chunk->immutable = true;
1403	bitmap_fill(chunk->populated, chunk->nr_pages);
1404	chunk->nr_populated = chunk->nr_pages;
1405	chunk->nr_empty_pop_pages = chunk->nr_pages;
 
 
1406
 
1407	chunk->free_bytes = map_size;
1408
1409	if (chunk->start_offset) {
1410		/* hide the beginning of the bitmap */
1411		offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1412		bitmap_set(chunk->alloc_map, 0, offset_bits);
1413		set_bit(0, chunk->bound_map);
1414		set_bit(offset_bits, chunk->bound_map);
1415
1416		chunk->chunk_md.first_free = offset_bits;
1417
1418		pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1419	}
1420
1421	if (chunk->end_offset) {
1422		/* hide the end of the bitmap */
1423		offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1424		bitmap_set(chunk->alloc_map,
1425			   pcpu_chunk_map_bits(chunk) - offset_bits,
1426			   offset_bits);
1427		set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1428			chunk->bound_map);
1429		set_bit(region_bits, chunk->bound_map);
1430
1431		pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1432					     - offset_bits, offset_bits);
1433	}
1434
1435	return chunk;
1436}
1437
1438static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1439{
1440	struct pcpu_chunk *chunk;
1441	int region_bits;
1442
1443	chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1444	if (!chunk)
1445		return NULL;
1446
1447	INIT_LIST_HEAD(&chunk->list);
1448	chunk->nr_pages = pcpu_unit_pages;
1449	region_bits = pcpu_chunk_map_bits(chunk);
1450
1451	chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1452					   sizeof(chunk->alloc_map[0]), gfp);
1453	if (!chunk->alloc_map)
1454		goto alloc_map_fail;
1455
1456	chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1457					   sizeof(chunk->bound_map[0]), gfp);
1458	if (!chunk->bound_map)
1459		goto bound_map_fail;
1460
1461	chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1462					   sizeof(chunk->md_blocks[0]), gfp);
1463	if (!chunk->md_blocks)
1464		goto md_blocks_fail;
1465
1466#ifdef CONFIG_MEMCG_KMEM
1467	if (!mem_cgroup_kmem_disabled()) {
1468		chunk->obj_cgroups =
1469			pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1470					sizeof(struct obj_cgroup *), gfp);
1471		if (!chunk->obj_cgroups)
1472			goto objcg_fail;
1473	}
1474#endif
1475
1476	pcpu_init_md_blocks(chunk);
1477
1478	/* init metadata */
 
1479	chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1480
1481	return chunk;
1482
1483#ifdef CONFIG_MEMCG_KMEM
1484objcg_fail:
1485	pcpu_mem_free(chunk->md_blocks);
1486#endif
1487md_blocks_fail:
1488	pcpu_mem_free(chunk->bound_map);
1489bound_map_fail:
1490	pcpu_mem_free(chunk->alloc_map);
1491alloc_map_fail:
1492	pcpu_mem_free(chunk);
1493
1494	return NULL;
1495}
1496
1497static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1498{
1499	if (!chunk)
1500		return;
1501#ifdef CONFIG_MEMCG_KMEM
1502	pcpu_mem_free(chunk->obj_cgroups);
1503#endif
1504	pcpu_mem_free(chunk->md_blocks);
1505	pcpu_mem_free(chunk->bound_map);
1506	pcpu_mem_free(chunk->alloc_map);
1507	pcpu_mem_free(chunk);
1508}
1509
1510/**
1511 * pcpu_chunk_populated - post-population bookkeeping
1512 * @chunk: pcpu_chunk which got populated
1513 * @page_start: the start page
1514 * @page_end: the end page
 
1515 *
1516 * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
1517 * the bookkeeping information accordingly.  Must be called after each
1518 * successful population.
 
 
 
1519 */
1520static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1521				 int page_end)
1522{
1523	int nr = page_end - page_start;
1524
1525	lockdep_assert_held(&pcpu_lock);
1526
1527	bitmap_set(chunk->populated, page_start, nr);
1528	chunk->nr_populated += nr;
1529	pcpu_nr_populated += nr;
1530
1531	pcpu_update_empty_pages(chunk, nr);
 
 
 
1532}
1533
1534/**
1535 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1536 * @chunk: pcpu_chunk which got depopulated
1537 * @page_start: the start page
1538 * @page_end: the end page
1539 *
1540 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1541 * Update the bookkeeping information accordingly.  Must be called after
1542 * each successful depopulation.
1543 */
1544static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1545				   int page_start, int page_end)
1546{
1547	int nr = page_end - page_start;
1548
1549	lockdep_assert_held(&pcpu_lock);
1550
1551	bitmap_clear(chunk->populated, page_start, nr);
1552	chunk->nr_populated -= nr;
1553	pcpu_nr_populated -= nr;
1554
1555	pcpu_update_empty_pages(chunk, -nr);
1556}
1557
1558/*
1559 * Chunk management implementation.
1560 *
1561 * To allow different implementations, chunk alloc/free and
1562 * [de]population are implemented in a separate file which is pulled
1563 * into this file and compiled together.  The following functions
1564 * should be implemented.
1565 *
1566 * pcpu_populate_chunk		- populate the specified range of a chunk
1567 * pcpu_depopulate_chunk	- depopulate the specified range of a chunk
1568 * pcpu_post_unmap_tlb_flush	- flush tlb for the specified range of a chunk
1569 * pcpu_create_chunk		- create a new chunk
1570 * pcpu_destroy_chunk		- destroy a chunk, always preceded by full depop
1571 * pcpu_addr_to_page		- translate address to physical address
1572 * pcpu_verify_alloc_info	- check alloc_info is acceptable during init
1573 */
1574static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1575			       int page_start, int page_end, gfp_t gfp);
1576static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1577				  int page_start, int page_end);
1578static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1579				      int page_start, int page_end);
1580static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1581static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1582static struct page *pcpu_addr_to_page(void *addr);
1583static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1584
1585#ifdef CONFIG_NEED_PER_CPU_KM
1586#include "percpu-km.c"
1587#else
1588#include "percpu-vm.c"
1589#endif
1590
1591/**
1592 * pcpu_chunk_addr_search - determine chunk containing specified address
1593 * @addr: address for which the chunk needs to be determined.
1594 *
1595 * This is an internal function that handles all but static allocations.
1596 * Static percpu address values should never be passed into the allocator.
1597 *
1598 * RETURNS:
1599 * The address of the found chunk.
1600 */
1601static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1602{
1603	/* is it in the dynamic region (first chunk)? */
1604	if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1605		return pcpu_first_chunk;
1606
1607	/* is it in the reserved region? */
1608	if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1609		return pcpu_reserved_chunk;
1610
1611	/*
1612	 * The address is relative to unit0 which might be unused and
1613	 * thus unmapped.  Offset the address to the unit space of the
1614	 * current processor before looking it up in the vmalloc
1615	 * space.  Note that any possible cpu id can be used here, so
1616	 * there's no need to worry about preemption or cpu hotplug.
1617	 */
1618	addr += pcpu_unit_offsets[raw_smp_processor_id()];
1619	return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1620}
1621
1622#ifdef CONFIG_MEMCG_KMEM
1623static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1624				      struct obj_cgroup **objcgp)
1625{
1626	struct obj_cgroup *objcg;
1627
1628	if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT))
1629		return true;
1630
1631	objcg = get_obj_cgroup_from_current();
1632	if (!objcg)
1633		return true;
1634
1635	if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size))) {
1636		obj_cgroup_put(objcg);
1637		return false;
1638	}
1639
1640	*objcgp = objcg;
1641	return true;
1642}
1643
1644static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1645				       struct pcpu_chunk *chunk, int off,
1646				       size_t size)
1647{
1648	if (!objcg)
1649		return;
1650
1651	if (likely(chunk && chunk->obj_cgroups)) {
1652		chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1653
1654		rcu_read_lock();
1655		mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1656				pcpu_obj_full_size(size));
1657		rcu_read_unlock();
1658	} else {
1659		obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1660		obj_cgroup_put(objcg);
1661	}
1662}
1663
1664static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1665{
1666	struct obj_cgroup *objcg;
1667
1668	if (unlikely(!chunk->obj_cgroups))
1669		return;
1670
1671	objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1672	if (!objcg)
1673		return;
1674	chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1675
1676	obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1677
1678	rcu_read_lock();
1679	mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1680			-pcpu_obj_full_size(size));
1681	rcu_read_unlock();
1682
1683	obj_cgroup_put(objcg);
1684}
1685
1686#else /* CONFIG_MEMCG_KMEM */
1687static bool
1688pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1689{
1690	return true;
1691}
1692
1693static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1694				       struct pcpu_chunk *chunk, int off,
1695				       size_t size)
1696{
1697}
1698
1699static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1700{
1701}
1702#endif /* CONFIG_MEMCG_KMEM */
1703
1704/**
1705 * pcpu_alloc - the percpu allocator
1706 * @size: size of area to allocate in bytes
1707 * @align: alignment of area (max PAGE_SIZE)
1708 * @reserved: allocate from the reserved chunk if available
1709 * @gfp: allocation flags
1710 *
1711 * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
1712 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1713 * then no warning will be triggered on invalid or failed allocation
1714 * requests.
1715 *
1716 * RETURNS:
1717 * Percpu pointer to the allocated area on success, NULL on failure.
1718 */
1719static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1720				 gfp_t gfp)
1721{
1722	gfp_t pcpu_gfp;
1723	bool is_atomic;
1724	bool do_warn;
1725	struct obj_cgroup *objcg = NULL;
1726	static int warn_limit = 10;
1727	struct pcpu_chunk *chunk, *next;
1728	const char *err;
1729	int slot, off, cpu, ret;
1730	unsigned long flags;
1731	void __percpu *ptr;
1732	size_t bits, bit_align;
1733
1734	gfp = current_gfp_context(gfp);
1735	/* whitelisted flags that can be passed to the backing allocators */
1736	pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1737	is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1738	do_warn = !(gfp & __GFP_NOWARN);
1739
1740	/*
1741	 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1742	 * therefore alignment must be a minimum of that many bytes.
1743	 * An allocation may have internal fragmentation from rounding up
1744	 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1745	 */
1746	if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1747		align = PCPU_MIN_ALLOC_SIZE;
1748
1749	size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1750	bits = size >> PCPU_MIN_ALLOC_SHIFT;
1751	bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1752
1753	if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1754		     !is_power_of_2(align))) {
1755		WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1756		     size, align);
1757		return NULL;
1758	}
1759
1760	if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1761		return NULL;
1762
1763	if (!is_atomic) {
1764		/*
1765		 * pcpu_balance_workfn() allocates memory under this mutex,
1766		 * and it may wait for memory reclaim. Allow current task
1767		 * to become OOM victim, in case of memory pressure.
1768		 */
1769		if (gfp & __GFP_NOFAIL) {
1770			mutex_lock(&pcpu_alloc_mutex);
1771		} else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1772			pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1773			return NULL;
1774		}
1775	}
1776
1777	spin_lock_irqsave(&pcpu_lock, flags);
1778
1779	/* serve reserved allocations from the reserved chunk if available */
1780	if (reserved && pcpu_reserved_chunk) {
1781		chunk = pcpu_reserved_chunk;
1782
1783		off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1784		if (off < 0) {
1785			err = "alloc from reserved chunk failed";
1786			goto fail_unlock;
1787		}
1788
1789		off = pcpu_alloc_area(chunk, bits, bit_align, off);
1790		if (off >= 0)
1791			goto area_found;
1792
1793		err = "alloc from reserved chunk failed";
1794		goto fail_unlock;
1795	}
1796
1797restart:
1798	/* search through normal chunks */
1799	for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1800		list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1801					 list) {
1802			off = pcpu_find_block_fit(chunk, bits, bit_align,
1803						  is_atomic);
1804			if (off < 0) {
1805				if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1806					pcpu_chunk_move(chunk, 0);
1807				continue;
1808			}
1809
1810			off = pcpu_alloc_area(chunk, bits, bit_align, off);
1811			if (off >= 0) {
1812				pcpu_reintegrate_chunk(chunk);
1813				goto area_found;
1814			}
1815		}
1816	}
1817
1818	spin_unlock_irqrestore(&pcpu_lock, flags);
1819
 
 
 
 
 
1820	if (is_atomic) {
1821		err = "atomic alloc failed, no space left";
1822		goto fail;
1823	}
1824
1825	/* No space left.  Create a new chunk. */
1826	if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1827		chunk = pcpu_create_chunk(pcpu_gfp);
1828		if (!chunk) {
1829			err = "failed to allocate new chunk";
1830			goto fail;
1831		}
1832
1833		spin_lock_irqsave(&pcpu_lock, flags);
1834		pcpu_chunk_relocate(chunk, -1);
1835	} else {
1836		spin_lock_irqsave(&pcpu_lock, flags);
1837	}
1838
1839	goto restart;
1840
1841area_found:
1842	pcpu_stats_area_alloc(chunk, size);
1843	spin_unlock_irqrestore(&pcpu_lock, flags);
1844
1845	/* populate if not all pages are already there */
1846	if (!is_atomic) {
1847		unsigned int page_end, rs, re;
1848
1849		rs = PFN_DOWN(off);
1850		page_end = PFN_UP(off + size);
1851
1852		for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
 
1853			WARN_ON(chunk->immutable);
1854
1855			ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1856
1857			spin_lock_irqsave(&pcpu_lock, flags);
1858			if (ret) {
1859				pcpu_free_area(chunk, off);
1860				err = "failed to populate";
1861				goto fail_unlock;
1862			}
1863			pcpu_chunk_populated(chunk, rs, re);
1864			spin_unlock_irqrestore(&pcpu_lock, flags);
1865		}
1866
1867		mutex_unlock(&pcpu_alloc_mutex);
1868	}
1869
1870	if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1871		pcpu_schedule_balance_work();
1872
1873	/* clear the areas and return address relative to base address */
1874	for_each_possible_cpu(cpu)
1875		memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1876
1877	ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1878	kmemleak_alloc_percpu(ptr, size, gfp);
1879
1880	trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1881				  chunk->base_addr, off, ptr,
1882				  pcpu_obj_full_size(size), gfp);
1883
1884	pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1885
1886	return ptr;
1887
1888fail_unlock:
1889	spin_unlock_irqrestore(&pcpu_lock, flags);
1890fail:
1891	trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1892
1893	if (!is_atomic && do_warn && warn_limit) {
1894		pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1895			size, align, is_atomic, err);
1896		dump_stack();
1897		if (!--warn_limit)
1898			pr_info("limit reached, disable warning\n");
1899	}
1900	if (is_atomic) {
1901		/* see the flag handling in pcpu_balance_workfn() */
1902		pcpu_atomic_alloc_failed = true;
1903		pcpu_schedule_balance_work();
1904	} else {
1905		mutex_unlock(&pcpu_alloc_mutex);
1906	}
1907
1908	pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1909
1910	return NULL;
1911}
1912
1913/**
1914 * __alloc_percpu_gfp - allocate dynamic percpu area
1915 * @size: size of area to allocate in bytes
1916 * @align: alignment of area (max PAGE_SIZE)
1917 * @gfp: allocation flags
1918 *
1919 * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1920 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1921 * be called from any context but is a lot more likely to fail. If @gfp
1922 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1923 * allocation requests.
1924 *
1925 * RETURNS:
1926 * Percpu pointer to the allocated area on success, NULL on failure.
1927 */
1928void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1929{
1930	return pcpu_alloc(size, align, false, gfp);
1931}
1932EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1933
1934/**
1935 * __alloc_percpu - allocate dynamic percpu area
1936 * @size: size of area to allocate in bytes
1937 * @align: alignment of area (max PAGE_SIZE)
1938 *
1939 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1940 */
1941void __percpu *__alloc_percpu(size_t size, size_t align)
1942{
1943	return pcpu_alloc(size, align, false, GFP_KERNEL);
1944}
1945EXPORT_SYMBOL_GPL(__alloc_percpu);
1946
1947/**
1948 * __alloc_reserved_percpu - allocate reserved percpu area
1949 * @size: size of area to allocate in bytes
1950 * @align: alignment of area (max PAGE_SIZE)
1951 *
1952 * Allocate zero-filled percpu area of @size bytes aligned at @align
1953 * from reserved percpu area if arch has set it up; otherwise,
1954 * allocation is served from the same dynamic area.  Might sleep.
1955 * Might trigger writeouts.
1956 *
1957 * CONTEXT:
1958 * Does GFP_KERNEL allocation.
1959 *
1960 * RETURNS:
1961 * Percpu pointer to the allocated area on success, NULL on failure.
1962 */
1963void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1964{
1965	return pcpu_alloc(size, align, true, GFP_KERNEL);
1966}
1967
1968/**
1969 * pcpu_balance_free - manage the amount of free chunks
1970 * @empty_only: free chunks only if there are no populated pages
1971 *
1972 * If empty_only is %false, reclaim all fully free chunks regardless of the
1973 * number of populated pages.  Otherwise, only reclaim chunks that have no
1974 * populated pages.
1975 *
1976 * CONTEXT:
1977 * pcpu_lock (can be dropped temporarily)
1978 */
1979static void pcpu_balance_free(bool empty_only)
1980{
 
 
1981	LIST_HEAD(to_free);
1982	struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1983	struct pcpu_chunk *chunk, *next;
1984
1985	lockdep_assert_held(&pcpu_lock);
1986
1987	/*
1988	 * There's no reason to keep around multiple unused chunks and VM
1989	 * areas can be scarce.  Destroy all free chunks except for one.
1990	 */
 
 
 
1991	list_for_each_entry_safe(chunk, next, free_head, list) {
1992		WARN_ON(chunk->immutable);
1993
1994		/* spare the first one */
1995		if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1996			continue;
1997
1998		if (!empty_only || chunk->nr_empty_pop_pages == 0)
1999			list_move(&chunk->list, &to_free);
2000	}
2001
2002	if (list_empty(&to_free))
2003		return;
2004
2005	spin_unlock_irq(&pcpu_lock);
2006	list_for_each_entry_safe(chunk, next, &to_free, list) {
2007		unsigned int rs, re;
2008
2009		for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
 
2010			pcpu_depopulate_chunk(chunk, rs, re);
2011			spin_lock_irq(&pcpu_lock);
2012			pcpu_chunk_depopulated(chunk, rs, re);
2013			spin_unlock_irq(&pcpu_lock);
2014		}
2015		pcpu_destroy_chunk(chunk);
2016		cond_resched();
2017	}
2018	spin_lock_irq(&pcpu_lock);
2019}
2020
2021/**
2022 * pcpu_balance_populated - manage the amount of populated pages
2023 *
2024 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2025 * It is possible that this is called when physical memory is scarce causing
2026 * OOM killer to be triggered.  We should avoid doing so until an actual
2027 * allocation causes the failure as it is possible that requests can be
2028 * serviced from already backed regions.
2029 *
2030 * CONTEXT:
2031 * pcpu_lock (can be dropped temporarily)
2032 */
2033static void pcpu_balance_populated(void)
2034{
2035	/* gfp flags passed to underlying allocators */
2036	const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2037	struct pcpu_chunk *chunk;
2038	int slot, nr_to_pop, ret;
2039
2040	lockdep_assert_held(&pcpu_lock);
2041
2042	/*
2043	 * Ensure there are certain number of free populated pages for
2044	 * atomic allocs.  Fill up from the most packed so that atomic
2045	 * allocs don't increase fragmentation.  If atomic allocation
2046	 * failed previously, always populate the maximum amount.  This
2047	 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2048	 * failing indefinitely; however, large atomic allocs are not
2049	 * something we support properly and can be highly unreliable and
2050	 * inefficient.
2051	 */
2052retry_pop:
2053	if (pcpu_atomic_alloc_failed) {
2054		nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2055		/* best effort anyway, don't worry about synchronization */
2056		pcpu_atomic_alloc_failed = false;
2057	} else {
2058		nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2059				  pcpu_nr_empty_pop_pages,
2060				  0, PCPU_EMPTY_POP_PAGES_HIGH);
2061	}
2062
2063	for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2064		unsigned int nr_unpop = 0, rs, re;
2065
2066		if (!nr_to_pop)
2067			break;
2068
2069		list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
 
2070			nr_unpop = chunk->nr_pages - chunk->nr_populated;
2071			if (nr_unpop)
2072				break;
2073		}
 
2074
2075		if (!nr_unpop)
2076			continue;
2077
2078		/* @chunk can't go away while pcpu_alloc_mutex is held */
2079		for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2080			int nr = min_t(int, re - rs, nr_to_pop);
 
2081
2082			spin_unlock_irq(&pcpu_lock);
2083			ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2084			cond_resched();
2085			spin_lock_irq(&pcpu_lock);
2086			if (!ret) {
2087				nr_to_pop -= nr;
2088				pcpu_chunk_populated(chunk, rs, rs + nr);
 
 
2089			} else {
2090				nr_to_pop = 0;
2091			}
2092
2093			if (!nr_to_pop)
2094				break;
2095		}
2096	}
2097
2098	if (nr_to_pop) {
2099		/* ran out of chunks to populate, create a new one and retry */
2100		spin_unlock_irq(&pcpu_lock);
2101		chunk = pcpu_create_chunk(gfp);
2102		cond_resched();
2103		spin_lock_irq(&pcpu_lock);
2104		if (chunk) {
 
2105			pcpu_chunk_relocate(chunk, -1);
 
2106			goto retry_pop;
2107		}
2108	}
2109}
2110
2111/**
2112 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2113 *
2114 * Scan over chunks in the depopulate list and try to release unused populated
2115 * pages back to the system.  Depopulated chunks are sidelined to prevent
2116 * repopulating these pages unless required.  Fully free chunks are reintegrated
2117 * and freed accordingly (1 is kept around).  If we drop below the empty
2118 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2119 * Each chunk is scanned in the reverse order to keep populated pages close to
2120 * the beginning of the chunk.
2121 *
2122 * CONTEXT:
2123 * pcpu_lock (can be dropped temporarily)
2124 *
2125 */
2126static void pcpu_reclaim_populated(void)
2127{
2128	struct pcpu_chunk *chunk;
2129	struct pcpu_block_md *block;
2130	int freed_page_start, freed_page_end;
2131	int i, end;
2132	bool reintegrate;
2133
2134	lockdep_assert_held(&pcpu_lock);
2135
2136	/*
2137	 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2138	 * longer discoverable to allocations whom may populate pages.  The only
2139	 * other accessor is the free path which only returns area back to the
2140	 * allocator not touching the populated bitmap.
2141	 */
2142	while ((chunk = list_first_entry_or_null(
2143			&pcpu_chunk_lists[pcpu_to_depopulate_slot],
2144			struct pcpu_chunk, list))) {
2145		WARN_ON(chunk->immutable);
2146
2147		/*
2148		 * Scan chunk's pages in the reverse order to keep populated
2149		 * pages close to the beginning of the chunk.
2150		 */
2151		freed_page_start = chunk->nr_pages;
2152		freed_page_end = 0;
2153		reintegrate = false;
2154		for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2155			/* no more work to do */
2156			if (chunk->nr_empty_pop_pages == 0)
2157				break;
2158
2159			/* reintegrate chunk to prevent atomic alloc failures */
2160			if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2161				reintegrate = true;
2162				break;
2163			}
2164
2165			/*
2166			 * If the page is empty and populated, start or
2167			 * extend the (i, end) range.  If i == 0, decrease
2168			 * i and perform the depopulation to cover the last
2169			 * (first) page in the chunk.
2170			 */
2171			block = chunk->md_blocks + i;
2172			if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2173			    test_bit(i, chunk->populated)) {
2174				if (end == -1)
2175					end = i;
2176				if (i > 0)
2177					continue;
2178				i--;
2179			}
2180
2181			/* depopulate if there is an active range */
2182			if (end == -1)
2183				continue;
2184
2185			spin_unlock_irq(&pcpu_lock);
2186			pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2187			cond_resched();
2188			spin_lock_irq(&pcpu_lock);
2189
2190			pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2191			freed_page_start = min(freed_page_start, i + 1);
2192			freed_page_end = max(freed_page_end, end + 1);
2193
2194			/* reset the range and continue */
2195			end = -1;
2196		}
2197
2198		/* batch tlb flush per chunk to amortize cost */
2199		if (freed_page_start < freed_page_end) {
2200			spin_unlock_irq(&pcpu_lock);
2201			pcpu_post_unmap_tlb_flush(chunk,
2202						  freed_page_start,
2203						  freed_page_end);
2204			cond_resched();
2205			spin_lock_irq(&pcpu_lock);
2206		}
2207
2208		if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2209			pcpu_reintegrate_chunk(chunk);
2210		else
2211			list_move_tail(&chunk->list,
2212				       &pcpu_chunk_lists[pcpu_sidelined_slot]);
2213	}
2214}
2215
2216/**
2217 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2218 * @work: unused
2219 *
2220 * For each chunk type, manage the number of fully free chunks and the number of
2221 * populated pages.  An important thing to consider is when pages are freed and
2222 * how they contribute to the global counts.
2223 */
2224static void pcpu_balance_workfn(struct work_struct *work)
2225{
2226	/*
2227	 * pcpu_balance_free() is called twice because the first time we may
2228	 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2229	 * to grow other chunks.  This then gives pcpu_reclaim_populated() time
2230	 * to move fully free chunks to the active list to be freed if
2231	 * appropriate.
2232	 */
2233	mutex_lock(&pcpu_alloc_mutex);
2234	spin_lock_irq(&pcpu_lock);
2235
2236	pcpu_balance_free(false);
2237	pcpu_reclaim_populated();
2238	pcpu_balance_populated();
2239	pcpu_balance_free(true);
2240
2241	spin_unlock_irq(&pcpu_lock);
2242	mutex_unlock(&pcpu_alloc_mutex);
2243}
2244
2245/**
2246 * free_percpu - free percpu area
2247 * @ptr: pointer to area to free
2248 *
2249 * Free percpu area @ptr.
2250 *
2251 * CONTEXT:
2252 * Can be called from atomic context.
2253 */
2254void free_percpu(void __percpu *ptr)
2255{
2256	void *addr;
2257	struct pcpu_chunk *chunk;
2258	unsigned long flags;
2259	int size, off;
2260	bool need_balance = false;
2261
2262	if (!ptr)
2263		return;
2264
2265	kmemleak_free_percpu(ptr);
2266
2267	addr = __pcpu_ptr_to_addr(ptr);
2268
2269	spin_lock_irqsave(&pcpu_lock, flags);
2270
2271	chunk = pcpu_chunk_addr_search(addr);
2272	off = addr - chunk->base_addr;
2273
2274	size = pcpu_free_area(chunk, off);
2275
2276	pcpu_memcg_free_hook(chunk, off, size);
2277
2278	/*
2279	 * If there are more than one fully free chunks, wake up grim reaper.
2280	 * If the chunk is isolated, it may be in the process of being
2281	 * reclaimed.  Let reclaim manage cleaning up of that chunk.
2282	 */
2283	if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2284		struct pcpu_chunk *pos;
2285
2286		list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2287			if (pos != chunk) {
2288				need_balance = true;
2289				break;
2290			}
2291	} else if (pcpu_should_reclaim_chunk(chunk)) {
2292		pcpu_isolate_chunk(chunk);
2293		need_balance = true;
2294	}
2295
2296	trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2297
2298	spin_unlock_irqrestore(&pcpu_lock, flags);
2299
2300	if (need_balance)
2301		pcpu_schedule_balance_work();
2302}
2303EXPORT_SYMBOL_GPL(free_percpu);
2304
2305bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2306{
2307#ifdef CONFIG_SMP
2308	const size_t static_size = __per_cpu_end - __per_cpu_start;
2309	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2310	unsigned int cpu;
2311
2312	for_each_possible_cpu(cpu) {
2313		void *start = per_cpu_ptr(base, cpu);
2314		void *va = (void *)addr;
2315
2316		if (va >= start && va < start + static_size) {
2317			if (can_addr) {
2318				*can_addr = (unsigned long) (va - start);
2319				*can_addr += (unsigned long)
2320					per_cpu_ptr(base, get_boot_cpu_id());
2321			}
2322			return true;
2323		}
2324	}
2325#endif
2326	/* on UP, can't distinguish from other static vars, always false */
2327	return false;
2328}
2329
2330/**
2331 * is_kernel_percpu_address - test whether address is from static percpu area
2332 * @addr: address to test
2333 *
2334 * Test whether @addr belongs to in-kernel static percpu area.  Module
2335 * static percpu areas are not considered.  For those, use
2336 * is_module_percpu_address().
2337 *
2338 * RETURNS:
2339 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2340 */
2341bool is_kernel_percpu_address(unsigned long addr)
2342{
2343	return __is_kernel_percpu_address(addr, NULL);
2344}
2345
2346/**
2347 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2348 * @addr: the address to be converted to physical address
2349 *
2350 * Given @addr which is dereferenceable address obtained via one of
2351 * percpu access macros, this function translates it into its physical
2352 * address.  The caller is responsible for ensuring @addr stays valid
2353 * until this function finishes.
2354 *
2355 * percpu allocator has special setup for the first chunk, which currently
2356 * supports either embedding in linear address space or vmalloc mapping,
2357 * and, from the second one, the backing allocator (currently either vm or
2358 * km) provides translation.
2359 *
2360 * The addr can be translated simply without checking if it falls into the
2361 * first chunk. But the current code reflects better how percpu allocator
2362 * actually works, and the verification can discover both bugs in percpu
2363 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2364 * code.
2365 *
2366 * RETURNS:
2367 * The physical address for @addr.
2368 */
2369phys_addr_t per_cpu_ptr_to_phys(void *addr)
2370{
2371	void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2372	bool in_first_chunk = false;
2373	unsigned long first_low, first_high;
2374	unsigned int cpu;
2375
2376	/*
2377	 * The following test on unit_low/high isn't strictly
2378	 * necessary but will speed up lookups of addresses which
2379	 * aren't in the first chunk.
2380	 *
2381	 * The address check is against full chunk sizes.  pcpu_base_addr
2382	 * points to the beginning of the first chunk including the
2383	 * static region.  Assumes good intent as the first chunk may
2384	 * not be full (ie. < pcpu_unit_pages in size).
2385	 */
2386	first_low = (unsigned long)pcpu_base_addr +
2387		    pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2388	first_high = (unsigned long)pcpu_base_addr +
2389		     pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2390	if ((unsigned long)addr >= first_low &&
2391	    (unsigned long)addr < first_high) {
2392		for_each_possible_cpu(cpu) {
2393			void *start = per_cpu_ptr(base, cpu);
2394
2395			if (addr >= start && addr < start + pcpu_unit_size) {
2396				in_first_chunk = true;
2397				break;
2398			}
2399		}
2400	}
2401
2402	if (in_first_chunk) {
2403		if (!is_vmalloc_addr(addr))
2404			return __pa(addr);
2405		else
2406			return page_to_phys(vmalloc_to_page(addr)) +
2407			       offset_in_page(addr);
2408	} else
2409		return page_to_phys(pcpu_addr_to_page(addr)) +
2410		       offset_in_page(addr);
2411}
2412
2413/**
2414 * pcpu_alloc_alloc_info - allocate percpu allocation info
2415 * @nr_groups: the number of groups
2416 * @nr_units: the number of units
2417 *
2418 * Allocate ai which is large enough for @nr_groups groups containing
2419 * @nr_units units.  The returned ai's groups[0].cpu_map points to the
2420 * cpu_map array which is long enough for @nr_units and filled with
2421 * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
2422 * pointer of other groups.
2423 *
2424 * RETURNS:
2425 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2426 * failure.
2427 */
2428struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2429						      int nr_units)
2430{
2431	struct pcpu_alloc_info *ai;
2432	size_t base_size, ai_size;
2433	void *ptr;
2434	int unit;
2435
2436	base_size = ALIGN(struct_size(ai, groups, nr_groups),
2437			  __alignof__(ai->groups[0].cpu_map[0]));
2438	ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2439
2440	ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2441	if (!ptr)
2442		return NULL;
2443	ai = ptr;
2444	ptr += base_size;
2445
2446	ai->groups[0].cpu_map = ptr;
2447
2448	for (unit = 0; unit < nr_units; unit++)
2449		ai->groups[0].cpu_map[unit] = NR_CPUS;
2450
2451	ai->nr_groups = nr_groups;
2452	ai->__ai_size = PFN_ALIGN(ai_size);
2453
2454	return ai;
2455}
2456
2457/**
2458 * pcpu_free_alloc_info - free percpu allocation info
2459 * @ai: pcpu_alloc_info to free
2460 *
2461 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2462 */
2463void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2464{
2465	memblock_free(ai, ai->__ai_size);
2466}
2467
2468/**
2469 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2470 * @lvl: loglevel
2471 * @ai: allocation info to dump
2472 *
2473 * Print out information about @ai using loglevel @lvl.
2474 */
2475static void pcpu_dump_alloc_info(const char *lvl,
2476				 const struct pcpu_alloc_info *ai)
2477{
2478	int group_width = 1, cpu_width = 1, width;
2479	char empty_str[] = "--------";
2480	int alloc = 0, alloc_end = 0;
2481	int group, v;
2482	int upa, apl;	/* units per alloc, allocs per line */
2483
2484	v = ai->nr_groups;
2485	while (v /= 10)
2486		group_width++;
2487
2488	v = num_possible_cpus();
2489	while (v /= 10)
2490		cpu_width++;
2491	empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2492
2493	upa = ai->alloc_size / ai->unit_size;
2494	width = upa * (cpu_width + 1) + group_width + 3;
2495	apl = rounddown_pow_of_two(max(60 / width, 1));
2496
2497	printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2498	       lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2499	       ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2500
2501	for (group = 0; group < ai->nr_groups; group++) {
2502		const struct pcpu_group_info *gi = &ai->groups[group];
2503		int unit = 0, unit_end = 0;
2504
2505		BUG_ON(gi->nr_units % upa);
2506		for (alloc_end += gi->nr_units / upa;
2507		     alloc < alloc_end; alloc++) {
2508			if (!(alloc % apl)) {
2509				pr_cont("\n");
2510				printk("%spcpu-alloc: ", lvl);
2511			}
2512			pr_cont("[%0*d] ", group_width, group);
2513
2514			for (unit_end += upa; unit < unit_end; unit++)
2515				if (gi->cpu_map[unit] != NR_CPUS)
2516					pr_cont("%0*d ",
2517						cpu_width, gi->cpu_map[unit]);
2518				else
2519					pr_cont("%s ", empty_str);
2520		}
2521	}
2522	pr_cont("\n");
2523}
2524
2525/**
2526 * pcpu_setup_first_chunk - initialize the first percpu chunk
2527 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2528 * @base_addr: mapped address
2529 *
2530 * Initialize the first percpu chunk which contains the kernel static
2531 * percpu area.  This function is to be called from arch percpu area
2532 * setup path.
2533 *
2534 * @ai contains all information necessary to initialize the first
2535 * chunk and prime the dynamic percpu allocator.
2536 *
2537 * @ai->static_size is the size of static percpu area.
2538 *
2539 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2540 * reserve after the static area in the first chunk.  This reserves
2541 * the first chunk such that it's available only through reserved
2542 * percpu allocation.  This is primarily used to serve module percpu
2543 * static areas on architectures where the addressing model has
2544 * limited offset range for symbol relocations to guarantee module
2545 * percpu symbols fall inside the relocatable range.
2546 *
2547 * @ai->dyn_size determines the number of bytes available for dynamic
2548 * allocation in the first chunk.  The area between @ai->static_size +
2549 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2550 *
2551 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2552 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2553 * @ai->dyn_size.
2554 *
2555 * @ai->atom_size is the allocation atom size and used as alignment
2556 * for vm areas.
2557 *
2558 * @ai->alloc_size is the allocation size and always multiple of
2559 * @ai->atom_size.  This is larger than @ai->atom_size if
2560 * @ai->unit_size is larger than @ai->atom_size.
2561 *
2562 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2563 * percpu areas.  Units which should be colocated are put into the
2564 * same group.  Dynamic VM areas will be allocated according to these
2565 * groupings.  If @ai->nr_groups is zero, a single group containing
2566 * all units is assumed.
2567 *
2568 * The caller should have mapped the first chunk at @base_addr and
2569 * copied static data to each unit.
2570 *
2571 * The first chunk will always contain a static and a dynamic region.
2572 * However, the static region is not managed by any chunk.  If the first
2573 * chunk also contains a reserved region, it is served by two chunks -
2574 * one for the reserved region and one for the dynamic region.  They
2575 * share the same vm, but use offset regions in the area allocation map.
2576 * The chunk serving the dynamic region is circulated in the chunk slots
2577 * and available for dynamic allocation like any other chunk.
 
 
 
2578 */
2579void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2580				   void *base_addr)
2581{
2582	size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2583	size_t static_size, dyn_size;
2584	struct pcpu_chunk *chunk;
2585	unsigned long *group_offsets;
2586	size_t *group_sizes;
2587	unsigned long *unit_off;
2588	unsigned int cpu;
2589	int *unit_map;
2590	int group, unit, i;
2591	int map_size;
2592	unsigned long tmp_addr;
2593	size_t alloc_size;
2594
2595#define PCPU_SETUP_BUG_ON(cond)	do {					\
2596	if (unlikely(cond)) {						\
2597		pr_emerg("failed to initialize, %s\n", #cond);		\
2598		pr_emerg("cpu_possible_mask=%*pb\n",			\
2599			 cpumask_pr_args(cpu_possible_mask));		\
2600		pcpu_dump_alloc_info(KERN_EMERG, ai);			\
2601		BUG();							\
2602	}								\
2603} while (0)
2604
2605	/* sanity checks */
2606	PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2607#ifdef CONFIG_SMP
2608	PCPU_SETUP_BUG_ON(!ai->static_size);
2609	PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2610#endif
2611	PCPU_SETUP_BUG_ON(!base_addr);
2612	PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2613	PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2614	PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2615	PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2616	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2617	PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2618	PCPU_SETUP_BUG_ON(!ai->dyn_size);
2619	PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2620	PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2621			    IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2622	PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2623
2624	/* process group information and build config tables accordingly */
2625	alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2626	group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2627	if (!group_offsets)
2628		panic("%s: Failed to allocate %zu bytes\n", __func__,
2629		      alloc_size);
2630
2631	alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2632	group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2633	if (!group_sizes)
2634		panic("%s: Failed to allocate %zu bytes\n", __func__,
2635		      alloc_size);
2636
2637	alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2638	unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2639	if (!unit_map)
2640		panic("%s: Failed to allocate %zu bytes\n", __func__,
2641		      alloc_size);
2642
2643	alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2644	unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2645	if (!unit_off)
2646		panic("%s: Failed to allocate %zu bytes\n", __func__,
2647		      alloc_size);
2648
2649	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2650		unit_map[cpu] = UINT_MAX;
2651
2652	pcpu_low_unit_cpu = NR_CPUS;
2653	pcpu_high_unit_cpu = NR_CPUS;
2654
2655	for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2656		const struct pcpu_group_info *gi = &ai->groups[group];
2657
2658		group_offsets[group] = gi->base_offset;
2659		group_sizes[group] = gi->nr_units * ai->unit_size;
2660
2661		for (i = 0; i < gi->nr_units; i++) {
2662			cpu = gi->cpu_map[i];
2663			if (cpu == NR_CPUS)
2664				continue;
2665
2666			PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2667			PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2668			PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2669
2670			unit_map[cpu] = unit + i;
2671			unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2672
2673			/* determine low/high unit_cpu */
2674			if (pcpu_low_unit_cpu == NR_CPUS ||
2675			    unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2676				pcpu_low_unit_cpu = cpu;
2677			if (pcpu_high_unit_cpu == NR_CPUS ||
2678			    unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2679				pcpu_high_unit_cpu = cpu;
2680		}
2681	}
2682	pcpu_nr_units = unit;
2683
2684	for_each_possible_cpu(cpu)
2685		PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2686
2687	/* we're done parsing the input, undefine BUG macro and dump config */
2688#undef PCPU_SETUP_BUG_ON
2689	pcpu_dump_alloc_info(KERN_DEBUG, ai);
2690
2691	pcpu_nr_groups = ai->nr_groups;
2692	pcpu_group_offsets = group_offsets;
2693	pcpu_group_sizes = group_sizes;
2694	pcpu_unit_map = unit_map;
2695	pcpu_unit_offsets = unit_off;
2696
2697	/* determine basic parameters */
2698	pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2699	pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2700	pcpu_atom_size = ai->atom_size;
2701	pcpu_chunk_struct_size = struct_size(chunk, populated,
2702					     BITS_TO_LONGS(pcpu_unit_pages));
2703
2704	pcpu_stats_save_ai(ai);
2705
2706	/*
2707	 * Allocate chunk slots.  The slots after the active slots are:
2708	 *   sidelined_slot - isolated, depopulated chunks
2709	 *   free_slot - fully free chunks
2710	 *   to_depopulate_slot - isolated, chunks to depopulate
2711	 */
2712	pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2713	pcpu_free_slot = pcpu_sidelined_slot + 1;
2714	pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2715	pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2716	pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2717					  sizeof(pcpu_chunk_lists[0]),
2718					  SMP_CACHE_BYTES);
2719	if (!pcpu_chunk_lists)
2720		panic("%s: Failed to allocate %zu bytes\n", __func__,
2721		      pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2722
2723	for (i = 0; i < pcpu_nr_slots; i++)
2724		INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2725
2726	/*
2727	 * The end of the static region needs to be aligned with the
2728	 * minimum allocation size as this offsets the reserved and
2729	 * dynamic region.  The first chunk ends page aligned by
2730	 * expanding the dynamic region, therefore the dynamic region
2731	 * can be shrunk to compensate while still staying above the
2732	 * configured sizes.
2733	 */
2734	static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2735	dyn_size = ai->dyn_size - (static_size - ai->static_size);
2736
2737	/*
2738	 * Initialize first chunk.
2739	 * If the reserved_size is non-zero, this initializes the reserved
2740	 * chunk.  If the reserved_size is zero, the reserved chunk is NULL
2741	 * and the dynamic region is initialized here.  The first chunk,
2742	 * pcpu_first_chunk, will always point to the chunk that serves
2743	 * the dynamic region.
2744	 */
2745	tmp_addr = (unsigned long)base_addr + static_size;
2746	map_size = ai->reserved_size ?: dyn_size;
2747	chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2748
2749	/* init dynamic chunk if necessary */
2750	if (ai->reserved_size) {
2751		pcpu_reserved_chunk = chunk;
2752
2753		tmp_addr = (unsigned long)base_addr + static_size +
2754			   ai->reserved_size;
2755		map_size = dyn_size;
2756		chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2757	}
2758
2759	/* link the first chunk in */
2760	pcpu_first_chunk = chunk;
2761	pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2762	pcpu_chunk_relocate(pcpu_first_chunk, -1);
2763
2764	/* include all regions of the first chunk */
2765	pcpu_nr_populated += PFN_DOWN(size_sum);
2766
2767	pcpu_stats_chunk_alloc();
2768	trace_percpu_create_chunk(base_addr);
2769
2770	/* we're done */
2771	pcpu_base_addr = base_addr;
 
2772}
2773
2774#ifdef CONFIG_SMP
2775
2776const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2777	[PCPU_FC_AUTO]	= "auto",
2778	[PCPU_FC_EMBED]	= "embed",
2779	[PCPU_FC_PAGE]	= "page",
2780};
2781
2782enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2783
2784static int __init percpu_alloc_setup(char *str)
2785{
2786	if (!str)
2787		return -EINVAL;
2788
2789	if (0)
2790		/* nada */;
2791#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2792	else if (!strcmp(str, "embed"))
2793		pcpu_chosen_fc = PCPU_FC_EMBED;
2794#endif
2795#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2796	else if (!strcmp(str, "page"))
2797		pcpu_chosen_fc = PCPU_FC_PAGE;
2798#endif
2799	else
2800		pr_warn("unknown allocator %s specified\n", str);
2801
2802	return 0;
2803}
2804early_param("percpu_alloc", percpu_alloc_setup);
2805
2806/*
2807 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2808 * Build it if needed by the arch config or the generic setup is going
2809 * to be used.
2810 */
2811#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2812	!defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2813#define BUILD_EMBED_FIRST_CHUNK
2814#endif
2815
2816/* build pcpu_page_first_chunk() iff needed by the arch config */
2817#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2818#define BUILD_PAGE_FIRST_CHUNK
2819#endif
2820
2821/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2822#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2823/**
2824 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2825 * @reserved_size: the size of reserved percpu area in bytes
2826 * @dyn_size: minimum free size for dynamic allocation in bytes
2827 * @atom_size: allocation atom size
2828 * @cpu_distance_fn: callback to determine distance between cpus, optional
2829 *
2830 * This function determines grouping of units, their mappings to cpus
2831 * and other parameters considering needed percpu size, allocation
2832 * atom size and distances between CPUs.
2833 *
2834 * Groups are always multiples of atom size and CPUs which are of
2835 * LOCAL_DISTANCE both ways are grouped together and share space for
2836 * units in the same group.  The returned configuration is guaranteed
2837 * to have CPUs on different nodes on different groups and >=75% usage
2838 * of allocated virtual address space.
2839 *
2840 * RETURNS:
2841 * On success, pointer to the new allocation_info is returned.  On
2842 * failure, ERR_PTR value is returned.
2843 */
2844static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2845				size_t reserved_size, size_t dyn_size,
2846				size_t atom_size,
2847				pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2848{
2849	static int group_map[NR_CPUS] __initdata;
2850	static int group_cnt[NR_CPUS] __initdata;
2851	static struct cpumask mask __initdata;
2852	const size_t static_size = __per_cpu_end - __per_cpu_start;
2853	int nr_groups = 1, nr_units = 0;
2854	size_t size_sum, min_unit_size, alloc_size;
2855	int upa, max_upa, best_upa;	/* units_per_alloc */
2856	int last_allocs, group, unit;
2857	unsigned int cpu, tcpu;
2858	struct pcpu_alloc_info *ai;
2859	unsigned int *cpu_map;
2860
2861	/* this function may be called multiple times */
2862	memset(group_map, 0, sizeof(group_map));
2863	memset(group_cnt, 0, sizeof(group_cnt));
2864	cpumask_clear(&mask);
2865
2866	/* calculate size_sum and ensure dyn_size is enough for early alloc */
2867	size_sum = PFN_ALIGN(static_size + reserved_size +
2868			    max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2869	dyn_size = size_sum - static_size - reserved_size;
2870
2871	/*
2872	 * Determine min_unit_size, alloc_size and max_upa such that
2873	 * alloc_size is multiple of atom_size and is the smallest
2874	 * which can accommodate 4k aligned segments which are equal to
2875	 * or larger than min_unit_size.
2876	 */
2877	min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2878
2879	/* determine the maximum # of units that can fit in an allocation */
2880	alloc_size = roundup(min_unit_size, atom_size);
2881	upa = alloc_size / min_unit_size;
2882	while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2883		upa--;
2884	max_upa = upa;
2885
2886	cpumask_copy(&mask, cpu_possible_mask);
2887
2888	/* group cpus according to their proximity */
2889	for (group = 0; !cpumask_empty(&mask); group++) {
2890		/* pop the group's first cpu */
2891		cpu = cpumask_first(&mask);
 
 
 
 
 
 
 
 
 
 
 
2892		group_map[cpu] = group;
2893		group_cnt[group]++;
2894		cpumask_clear_cpu(cpu, &mask);
2895
2896		for_each_cpu(tcpu, &mask) {
2897			if (!cpu_distance_fn ||
2898			    (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2899			     cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2900				group_map[tcpu] = group;
2901				group_cnt[group]++;
2902				cpumask_clear_cpu(tcpu, &mask);
2903			}
2904		}
2905	}
2906	nr_groups = group;
2907
2908	/*
2909	 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2910	 * Expand the unit_size until we use >= 75% of the units allocated.
2911	 * Related to atom_size, which could be much larger than the unit_size.
2912	 */
2913	last_allocs = INT_MAX;
2914	best_upa = 0;
2915	for (upa = max_upa; upa; upa--) {
2916		int allocs = 0, wasted = 0;
2917
2918		if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2919			continue;
2920
2921		for (group = 0; group < nr_groups; group++) {
2922			int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2923			allocs += this_allocs;
2924			wasted += this_allocs * upa - group_cnt[group];
2925		}
2926
2927		/*
2928		 * Don't accept if wastage is over 1/3.  The
2929		 * greater-than comparison ensures upa==1 always
2930		 * passes the following check.
2931		 */
2932		if (wasted > num_possible_cpus() / 3)
2933			continue;
2934
2935		/* and then don't consume more memory */
2936		if (allocs > last_allocs)
2937			break;
2938		last_allocs = allocs;
2939		best_upa = upa;
2940	}
2941	BUG_ON(!best_upa);
2942	upa = best_upa;
2943
2944	/* allocate and fill alloc_info */
2945	for (group = 0; group < nr_groups; group++)
2946		nr_units += roundup(group_cnt[group], upa);
2947
2948	ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2949	if (!ai)
2950		return ERR_PTR(-ENOMEM);
2951	cpu_map = ai->groups[0].cpu_map;
2952
2953	for (group = 0; group < nr_groups; group++) {
2954		ai->groups[group].cpu_map = cpu_map;
2955		cpu_map += roundup(group_cnt[group], upa);
2956	}
2957
2958	ai->static_size = static_size;
2959	ai->reserved_size = reserved_size;
2960	ai->dyn_size = dyn_size;
2961	ai->unit_size = alloc_size / upa;
2962	ai->atom_size = atom_size;
2963	ai->alloc_size = alloc_size;
2964
2965	for (group = 0, unit = 0; group < nr_groups; group++) {
2966		struct pcpu_group_info *gi = &ai->groups[group];
2967
2968		/*
2969		 * Initialize base_offset as if all groups are located
2970		 * back-to-back.  The caller should update this to
2971		 * reflect actual allocation.
2972		 */
2973		gi->base_offset = unit * ai->unit_size;
2974
2975		for_each_possible_cpu(cpu)
2976			if (group_map[cpu] == group)
2977				gi->cpu_map[gi->nr_units++] = cpu;
2978		gi->nr_units = roundup(gi->nr_units, upa);
2979		unit += gi->nr_units;
2980	}
2981	BUG_ON(unit != nr_units);
2982
2983	return ai;
2984}
2985
2986static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
2987				   pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
2988{
2989	const unsigned long goal = __pa(MAX_DMA_ADDRESS);
2990#ifdef CONFIG_NUMA
2991	int node = NUMA_NO_NODE;
2992	void *ptr;
2993
2994	if (cpu_to_nd_fn)
2995		node = cpu_to_nd_fn(cpu);
2996
2997	if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
2998		ptr = memblock_alloc_from(size, align, goal);
2999		pr_info("cpu %d has no node %d or node-local memory\n",
3000			cpu, node);
3001		pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
3002			 cpu, size, (u64)__pa(ptr));
3003	} else {
3004		ptr = memblock_alloc_try_nid(size, align, goal,
3005					     MEMBLOCK_ALLOC_ACCESSIBLE,
3006					     node);
3007
3008		pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
3009			 cpu, size, node, (u64)__pa(ptr));
3010	}
3011	return ptr;
3012#else
3013	return memblock_alloc_from(size, align, goal);
3014#endif
3015}
3016
3017static void __init pcpu_fc_free(void *ptr, size_t size)
3018{
3019	memblock_free(ptr, size);
3020}
3021#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3022
3023#if defined(BUILD_EMBED_FIRST_CHUNK)
3024/**
3025 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3026 * @reserved_size: the size of reserved percpu area in bytes
3027 * @dyn_size: minimum free size for dynamic allocation in bytes
3028 * @atom_size: allocation atom size
3029 * @cpu_distance_fn: callback to determine distance between cpus, optional
3030 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
 
3031 *
3032 * This is a helper to ease setting up embedded first percpu chunk and
3033 * can be called where pcpu_setup_first_chunk() is expected.
3034 *
3035 * If this function is used to setup the first chunk, it is allocated
3036 * by calling pcpu_fc_alloc and used as-is without being mapped into
3037 * vmalloc area.  Allocations are always whole multiples of @atom_size
3038 * aligned to @atom_size.
3039 *
3040 * This enables the first chunk to piggy back on the linear physical
3041 * mapping which often uses larger page size.  Please note that this
3042 * can result in very sparse cpu->unit mapping on NUMA machines thus
3043 * requiring large vmalloc address space.  Don't use this allocator if
3044 * vmalloc space is not orders of magnitude larger than distances
3045 * between node memory addresses (ie. 32bit NUMA machines).
3046 *
3047 * @dyn_size specifies the minimum dynamic area size.
3048 *
3049 * If the needed size is smaller than the minimum or specified unit
3050 * size, the leftover is returned using pcpu_fc_free.
3051 *
3052 * RETURNS:
3053 * 0 on success, -errno on failure.
3054 */
3055int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3056				  size_t atom_size,
3057				  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3058				  pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
 
3059{
3060	void *base = (void *)ULONG_MAX;
3061	void **areas = NULL;
3062	struct pcpu_alloc_info *ai;
3063	size_t size_sum, areas_size;
3064	unsigned long max_distance;
3065	int group, i, highest_group, rc = 0;
3066
3067	ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3068				   cpu_distance_fn);
3069	if (IS_ERR(ai))
3070		return PTR_ERR(ai);
3071
3072	size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3073	areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3074
3075	areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3076	if (!areas) {
3077		rc = -ENOMEM;
3078		goto out_free;
3079	}
3080
3081	/* allocate, copy and determine base address & max_distance */
3082	highest_group = 0;
3083	for (group = 0; group < ai->nr_groups; group++) {
3084		struct pcpu_group_info *gi = &ai->groups[group];
3085		unsigned int cpu = NR_CPUS;
3086		void *ptr;
3087
3088		for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3089			cpu = gi->cpu_map[i];
3090		BUG_ON(cpu == NR_CPUS);
3091
3092		/* allocate space for the whole group */
3093		ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn);
3094		if (!ptr) {
3095			rc = -ENOMEM;
3096			goto out_free_areas;
3097		}
3098		/* kmemleak tracks the percpu allocations separately */
3099		kmemleak_ignore_phys(__pa(ptr));
3100		areas[group] = ptr;
3101
3102		base = min(ptr, base);
3103		if (ptr > areas[highest_group])
3104			highest_group = group;
3105	}
3106	max_distance = areas[highest_group] - base;
3107	max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3108
3109	/* warn if maximum distance is further than 75% of vmalloc space */
3110	if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3111		pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3112				max_distance, VMALLOC_TOTAL);
3113#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3114		/* and fail if we have fallback */
3115		rc = -EINVAL;
3116		goto out_free_areas;
3117#endif
3118	}
3119
3120	/*
3121	 * Copy data and free unused parts.  This should happen after all
3122	 * allocations are complete; otherwise, we may end up with
3123	 * overlapping groups.
3124	 */
3125	for (group = 0; group < ai->nr_groups; group++) {
3126		struct pcpu_group_info *gi = &ai->groups[group];
3127		void *ptr = areas[group];
3128
3129		for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3130			if (gi->cpu_map[i] == NR_CPUS) {
3131				/* unused unit, free whole */
3132				pcpu_fc_free(ptr, ai->unit_size);
3133				continue;
3134			}
3135			/* copy and return the unused part */
3136			memcpy(ptr, __per_cpu_load, ai->static_size);
3137			pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum);
3138		}
3139	}
3140
3141	/* base address is now known, determine group base offsets */
3142	for (group = 0; group < ai->nr_groups; group++) {
3143		ai->groups[group].base_offset = areas[group] - base;
3144	}
3145
3146	pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3147		PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3148		ai->dyn_size, ai->unit_size);
3149
3150	pcpu_setup_first_chunk(ai, base);
3151	goto out_free;
3152
3153out_free_areas:
3154	for (group = 0; group < ai->nr_groups; group++)
3155		if (areas[group])
3156			pcpu_fc_free(areas[group],
3157				ai->groups[group].nr_units * ai->unit_size);
3158out_free:
3159	pcpu_free_alloc_info(ai);
3160	if (areas)
3161		memblock_free(areas, areas_size);
3162	return rc;
3163}
3164#endif /* BUILD_EMBED_FIRST_CHUNK */
3165
3166#ifdef BUILD_PAGE_FIRST_CHUNK
3167#include <asm/pgalloc.h>
3168
3169#ifndef P4D_TABLE_SIZE
3170#define P4D_TABLE_SIZE PAGE_SIZE
3171#endif
3172
3173#ifndef PUD_TABLE_SIZE
3174#define PUD_TABLE_SIZE PAGE_SIZE
3175#endif
3176
3177#ifndef PMD_TABLE_SIZE
3178#define PMD_TABLE_SIZE PAGE_SIZE
3179#endif
3180
3181#ifndef PTE_TABLE_SIZE
3182#define PTE_TABLE_SIZE PAGE_SIZE
3183#endif
3184void __init __weak pcpu_populate_pte(unsigned long addr)
3185{
3186	pgd_t *pgd = pgd_offset_k(addr);
3187	p4d_t *p4d;
3188	pud_t *pud;
3189	pmd_t *pmd;
3190
3191	if (pgd_none(*pgd)) {
3192		p4d_t *new;
3193
3194		new = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3195		if (!new)
3196			goto err_alloc;
3197		pgd_populate(&init_mm, pgd, new);
3198	}
3199
3200	p4d = p4d_offset(pgd, addr);
3201	if (p4d_none(*p4d)) {
3202		pud_t *new;
3203
3204		new = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3205		if (!new)
3206			goto err_alloc;
3207		p4d_populate(&init_mm, p4d, new);
3208	}
3209
3210	pud = pud_offset(p4d, addr);
3211	if (pud_none(*pud)) {
3212		pmd_t *new;
3213
3214		new = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3215		if (!new)
3216			goto err_alloc;
3217		pud_populate(&init_mm, pud, new);
3218	}
3219
3220	pmd = pmd_offset(pud, addr);
3221	if (!pmd_present(*pmd)) {
3222		pte_t *new;
3223
3224		new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3225		if (!new)
3226			goto err_alloc;
3227		pmd_populate_kernel(&init_mm, pmd, new);
3228	}
3229
3230	return;
3231
3232err_alloc:
3233	panic("%s: Failed to allocate memory\n", __func__);
3234}
3235
3236/**
3237 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3238 * @reserved_size: the size of reserved percpu area in bytes
3239 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
 
 
3240 *
3241 * This is a helper to ease setting up page-remapped first percpu
3242 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3243 *
3244 * This is the basic allocator.  Static percpu area is allocated
3245 * page-by-page into vmalloc area.
3246 *
3247 * RETURNS:
3248 * 0 on success, -errno on failure.
3249 */
3250int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
 
 
 
3251{
3252	static struct vm_struct vm;
3253	struct pcpu_alloc_info *ai;
3254	char psize_str[16];
3255	int unit_pages;
3256	size_t pages_size;
3257	struct page **pages;
3258	int unit, i, j, rc = 0;
3259	int upa;
3260	int nr_g0_units;
3261
3262	snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3263
3264	ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3265	if (IS_ERR(ai))
3266		return PTR_ERR(ai);
3267	BUG_ON(ai->nr_groups != 1);
3268	upa = ai->alloc_size/ai->unit_size;
3269	nr_g0_units = roundup(num_possible_cpus(), upa);
3270	if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3271		pcpu_free_alloc_info(ai);
3272		return -EINVAL;
3273	}
3274
3275	unit_pages = ai->unit_size >> PAGE_SHIFT;
3276
3277	/* unaligned allocations can't be freed, round up to page size */
3278	pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3279			       sizeof(pages[0]));
3280	pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3281	if (!pages)
3282		panic("%s: Failed to allocate %zu bytes\n", __func__,
3283		      pages_size);
3284
3285	/* allocate pages */
3286	j = 0;
3287	for (unit = 0; unit < num_possible_cpus(); unit++) {
3288		unsigned int cpu = ai->groups[0].cpu_map[unit];
3289		for (i = 0; i < unit_pages; i++) {
3290			void *ptr;
3291
3292			ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3293			if (!ptr) {
3294				pr_warn("failed to allocate %s page for cpu%u\n",
3295						psize_str, cpu);
3296				goto enomem;
3297			}
3298			/* kmemleak tracks the percpu allocations separately */
3299			kmemleak_ignore_phys(__pa(ptr));
3300			pages[j++] = virt_to_page(ptr);
3301		}
3302	}
3303
3304	/* allocate vm area, map the pages and copy static data */
3305	vm.flags = VM_ALLOC;
3306	vm.size = num_possible_cpus() * ai->unit_size;
3307	vm_area_register_early(&vm, PAGE_SIZE);
3308
3309	for (unit = 0; unit < num_possible_cpus(); unit++) {
3310		unsigned long unit_addr =
3311			(unsigned long)vm.addr + unit * ai->unit_size;
3312
3313		for (i = 0; i < unit_pages; i++)
3314			pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT));
3315
3316		/* pte already populated, the following shouldn't fail */
3317		rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3318				      unit_pages);
3319		if (rc < 0)
3320			panic("failed to map percpu area, err=%d\n", rc);
3321
3322		/*
3323		 * FIXME: Archs with virtual cache should flush local
3324		 * cache for the linear mapping here - something
3325		 * equivalent to flush_cache_vmap() on the local cpu.
3326		 * flush_cache_vmap() can't be used as most supporting
3327		 * data structures are not set up yet.
3328		 */
3329
3330		/* copy static data */
3331		memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3332	}
3333
3334	/* we're ready, commit */
3335	pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3336		unit_pages, psize_str, ai->static_size,
3337		ai->reserved_size, ai->dyn_size);
3338
3339	pcpu_setup_first_chunk(ai, vm.addr);
3340	goto out_free_ar;
3341
3342enomem:
3343	while (--j >= 0)
3344		pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3345	rc = -ENOMEM;
3346out_free_ar:
3347	memblock_free(pages, pages_size);
3348	pcpu_free_alloc_info(ai);
3349	return rc;
3350}
3351#endif /* BUILD_PAGE_FIRST_CHUNK */
3352
3353#ifndef	CONFIG_HAVE_SETUP_PER_CPU_AREA
3354/*
3355 * Generic SMP percpu area setup.
3356 *
3357 * The embedding helper is used because its behavior closely resembles
3358 * the original non-dynamic generic percpu area setup.  This is
3359 * important because many archs have addressing restrictions and might
3360 * fail if the percpu area is located far away from the previous
3361 * location.  As an added bonus, in non-NUMA cases, embedding is
3362 * generally a good idea TLB-wise because percpu area can piggy back
3363 * on the physical linear memory mapping which uses large page
3364 * mappings on applicable archs.
3365 */
3366unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3367EXPORT_SYMBOL(__per_cpu_offset);
3368
 
 
 
 
 
 
 
 
 
 
 
 
3369void __init setup_per_cpu_areas(void)
3370{
3371	unsigned long delta;
3372	unsigned int cpu;
3373	int rc;
3374
3375	/*
3376	 * Always reserve area for module percpu variables.  That's
3377	 * what the legacy allocator did.
3378	 */
3379	rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3380				    PAGE_SIZE, NULL, NULL);
 
3381	if (rc < 0)
3382		panic("Failed to initialize percpu areas.");
3383
3384	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3385	for_each_possible_cpu(cpu)
3386		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3387}
3388#endif	/* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3389
3390#else	/* CONFIG_SMP */
3391
3392/*
3393 * UP percpu area setup.
3394 *
3395 * UP always uses km-based percpu allocator with identity mapping.
3396 * Static percpu variables are indistinguishable from the usual static
3397 * variables and don't require any special preparation.
3398 */
3399void __init setup_per_cpu_areas(void)
3400{
3401	const size_t unit_size =
3402		roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3403					 PERCPU_DYNAMIC_RESERVE));
3404	struct pcpu_alloc_info *ai;
3405	void *fc;
3406
3407	ai = pcpu_alloc_alloc_info(1, 1);
3408	fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 
 
3409	if (!ai || !fc)
3410		panic("Failed to allocate memory for percpu areas.");
3411	/* kmemleak tracks the percpu allocations separately */
3412	kmemleak_ignore_phys(__pa(fc));
3413
3414	ai->dyn_size = unit_size;
3415	ai->unit_size = unit_size;
3416	ai->atom_size = unit_size;
3417	ai->alloc_size = unit_size;
3418	ai->groups[0].nr_units = 1;
3419	ai->groups[0].cpu_map[0] = 0;
3420
3421	pcpu_setup_first_chunk(ai, fc);
 
3422	pcpu_free_alloc_info(ai);
3423}
3424
3425#endif	/* CONFIG_SMP */
3426
3427/*
3428 * pcpu_nr_pages - calculate total number of populated backing pages
3429 *
3430 * This reflects the number of pages populated to back chunks.  Metadata is
3431 * excluded in the number exposed in meminfo as the number of backing pages
3432 * scales with the number of cpus and can quickly outweigh the memory used for
3433 * metadata.  It also keeps this calculation nice and simple.
3434 *
3435 * RETURNS:
3436 * Total number of populated backing pages in use by the allocator.
3437 */
3438unsigned long pcpu_nr_pages(void)
3439{
3440	return pcpu_nr_populated * pcpu_nr_units;
3441}
3442
3443/*
3444 * Percpu allocator is initialized early during boot when neither slab or
3445 * workqueue is available.  Plug async management until everything is up
3446 * and running.
3447 */
3448static int __init percpu_enable_async(void)
3449{
3450	pcpu_async_enabled = true;
3451	return 0;
3452}
3453subsys_initcall(percpu_enable_async);